4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
10 #include <linux/nospec.h>
12 #include <linux/kcov.h>
14 #include <asm/switch_to.h>
17 #include "../workqueue_internal.h"
18 #include "../smpboot.h"
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/sched.h>
25 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
27 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
29 * Debugging: various feature bits
31 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
32 * sysctl_sched_features, defined in sched.h, to allow constants propagation
33 * at compile time and compiler optimization based on features default.
35 #define SCHED_FEAT(name, enabled) \
36 (1UL << __SCHED_FEAT_##name) * enabled |
37 const_debug
unsigned int sysctl_sched_features
=
44 * Number of tasks to iterate in a single balance run.
45 * Limited because this is done with IRQs disabled.
47 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
50 * period over which we measure -rt task CPU usage in us.
53 unsigned int sysctl_sched_rt_period
= 1000000;
55 __read_mostly
int scheduler_running
;
58 * part of the period that we allow rt tasks to run in us.
61 int sysctl_sched_rt_runtime
= 950000;
64 * __task_rq_lock - lock the rq @p resides on.
66 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
71 lockdep_assert_held(&p
->pi_lock
);
75 raw_spin_lock(&rq
->lock
);
76 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
80 raw_spin_unlock(&rq
->lock
);
82 while (unlikely(task_on_rq_migrating(p
)))
88 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
90 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
91 __acquires(p
->pi_lock
)
97 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
99 raw_spin_lock(&rq
->lock
);
101 * move_queued_task() task_rq_lock()
104 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
105 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
106 * [S] ->cpu = new_cpu [L] task_rq()
110 * If we observe the old CPU in task_rq_lock, the acquire of
111 * the old rq->lock will fully serialize against the stores.
113 * If we observe the new CPU in task_rq_lock, the acquire will
114 * pair with the WMB to ensure we must then also see migrating.
116 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
120 raw_spin_unlock(&rq
->lock
);
121 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
123 while (unlikely(task_on_rq_migrating(p
)))
129 * RQ-clock updating methods:
132 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
135 * In theory, the compile should just see 0 here, and optimize out the call
136 * to sched_rt_avg_update. But I don't trust it...
138 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
139 s64 steal
= 0, irq_delta
= 0;
141 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
142 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
145 * Since irq_time is only updated on {soft,}irq_exit, we might run into
146 * this case when a previous update_rq_clock() happened inside a
149 * When this happens, we stop ->clock_task and only update the
150 * prev_irq_time stamp to account for the part that fit, so that a next
151 * update will consume the rest. This ensures ->clock_task is
154 * It does however cause some slight miss-attribution of {soft,}irq
155 * time, a more accurate solution would be to update the irq_time using
156 * the current rq->clock timestamp, except that would require using
159 if (irq_delta
> delta
)
162 rq
->prev_irq_time
+= irq_delta
;
165 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
166 if (static_key_false((¶virt_steal_rq_enabled
))) {
167 steal
= paravirt_steal_clock(cpu_of(rq
));
168 steal
-= rq
->prev_steal_time_rq
;
170 if (unlikely(steal
> delta
))
173 rq
->prev_steal_time_rq
+= steal
;
178 rq
->clock_task
+= delta
;
180 #ifdef HAVE_SCHED_AVG_IRQ
181 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
182 update_irq_load_avg(rq
, irq_delta
+ steal
);
186 void update_rq_clock(struct rq
*rq
)
190 lockdep_assert_held(&rq
->lock
);
192 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
195 #ifdef CONFIG_SCHED_DEBUG
196 if (sched_feat(WARN_DOUBLE_CLOCK
))
197 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
198 rq
->clock_update_flags
|= RQCF_UPDATED
;
201 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
205 update_rq_clock_task(rq
, delta
);
209 #ifdef CONFIG_SCHED_HRTICK
211 * Use HR-timers to deliver accurate preemption points.
214 static void hrtick_clear(struct rq
*rq
)
216 if (hrtimer_active(&rq
->hrtick_timer
))
217 hrtimer_cancel(&rq
->hrtick_timer
);
221 * High-resolution timer tick.
222 * Runs from hardirq context with interrupts disabled.
224 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
226 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
229 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
233 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
236 return HRTIMER_NORESTART
;
241 static void __hrtick_restart(struct rq
*rq
)
243 struct hrtimer
*timer
= &rq
->hrtick_timer
;
245 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
249 * called from hardirq (IPI) context
251 static void __hrtick_start(void *arg
)
257 __hrtick_restart(rq
);
258 rq
->hrtick_csd_pending
= 0;
263 * Called to set the hrtick timer state.
265 * called with rq->lock held and irqs disabled
267 void hrtick_start(struct rq
*rq
, u64 delay
)
269 struct hrtimer
*timer
= &rq
->hrtick_timer
;
274 * Don't schedule slices shorter than 10000ns, that just
275 * doesn't make sense and can cause timer DoS.
277 delta
= max_t(s64
, delay
, 10000LL);
278 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
280 hrtimer_set_expires(timer
, time
);
282 if (rq
== this_rq()) {
283 __hrtick_restart(rq
);
284 } else if (!rq
->hrtick_csd_pending
) {
285 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
286 rq
->hrtick_csd_pending
= 1;
292 * Called to set the hrtick timer state.
294 * called with rq->lock held and irqs disabled
296 void hrtick_start(struct rq
*rq
, u64 delay
)
299 * Don't schedule slices shorter than 10000ns, that just
300 * doesn't make sense. Rely on vruntime for fairness.
302 delay
= max_t(u64
, delay
, 10000LL);
303 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
304 HRTIMER_MODE_REL_PINNED
);
306 #endif /* CONFIG_SMP */
308 static void hrtick_rq_init(struct rq
*rq
)
311 rq
->hrtick_csd_pending
= 0;
313 rq
->hrtick_csd
.flags
= 0;
314 rq
->hrtick_csd
.func
= __hrtick_start
;
315 rq
->hrtick_csd
.info
= rq
;
318 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
319 rq
->hrtick_timer
.function
= hrtick
;
321 #else /* CONFIG_SCHED_HRTICK */
322 static inline void hrtick_clear(struct rq
*rq
)
326 static inline void hrtick_rq_init(struct rq
*rq
)
329 #endif /* CONFIG_SCHED_HRTICK */
332 * cmpxchg based fetch_or, macro so it works for different integer types
334 #define fetch_or(ptr, mask) \
336 typeof(ptr) _ptr = (ptr); \
337 typeof(mask) _mask = (mask); \
338 typeof(*_ptr) _old, _val = *_ptr; \
341 _old = cmpxchg(_ptr, _val, _val | _mask); \
349 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
351 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
352 * this avoids any races wrt polling state changes and thereby avoids
355 static bool set_nr_and_not_polling(struct task_struct
*p
)
357 struct thread_info
*ti
= task_thread_info(p
);
358 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
362 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
364 * If this returns true, then the idle task promises to call
365 * sched_ttwu_pending() and reschedule soon.
367 static bool set_nr_if_polling(struct task_struct
*p
)
369 struct thread_info
*ti
= task_thread_info(p
);
370 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
373 if (!(val
& _TIF_POLLING_NRFLAG
))
375 if (val
& _TIF_NEED_RESCHED
)
377 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
386 static bool set_nr_and_not_polling(struct task_struct
*p
)
388 set_tsk_need_resched(p
);
393 static bool set_nr_if_polling(struct task_struct
*p
)
400 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
402 struct wake_q_node
*node
= &task
->wake_q
;
405 * Atomically grab the task, if ->wake_q is !nil already it means
406 * its already queued (either by us or someone else) and will get the
407 * wakeup due to that.
409 * This cmpxchg() executes a full barrier, which pairs with the full
410 * barrier executed by the wakeup in wake_up_q().
412 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
415 get_task_struct(task
);
418 * The head is context local, there can be no concurrency.
421 head
->lastp
= &node
->next
;
424 void wake_up_q(struct wake_q_head
*head
)
426 struct wake_q_node
*node
= head
->first
;
428 while (node
!= WAKE_Q_TAIL
) {
429 struct task_struct
*task
;
431 task
= container_of(node
, struct task_struct
, wake_q
);
433 /* Task can safely be re-inserted now: */
435 task
->wake_q
.next
= NULL
;
438 * wake_up_process() executes a full barrier, which pairs with
439 * the queueing in wake_q_add() so as not to miss wakeups.
441 wake_up_process(task
);
442 put_task_struct(task
);
447 * resched_curr - mark rq's current task 'to be rescheduled now'.
449 * On UP this means the setting of the need_resched flag, on SMP it
450 * might also involve a cross-CPU call to trigger the scheduler on
453 void resched_curr(struct rq
*rq
)
455 struct task_struct
*curr
= rq
->curr
;
458 lockdep_assert_held(&rq
->lock
);
460 if (test_tsk_need_resched(curr
))
465 if (cpu
== smp_processor_id()) {
466 set_tsk_need_resched(curr
);
467 set_preempt_need_resched();
471 if (set_nr_and_not_polling(curr
))
472 smp_send_reschedule(cpu
);
474 trace_sched_wake_idle_without_ipi(cpu
);
477 void resched_cpu(int cpu
)
479 struct rq
*rq
= cpu_rq(cpu
);
482 raw_spin_lock_irqsave(&rq
->lock
, flags
);
483 if (cpu_online(cpu
) || cpu
== smp_processor_id())
485 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
489 #ifdef CONFIG_NO_HZ_COMMON
491 * In the semi idle case, use the nearest busy CPU for migrating timers
492 * from an idle CPU. This is good for power-savings.
494 * We don't do similar optimization for completely idle system, as
495 * selecting an idle CPU will add more delays to the timers than intended
496 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
498 int get_nohz_timer_target(void)
500 int i
, cpu
= smp_processor_id();
501 struct sched_domain
*sd
;
503 if (!idle_cpu(cpu
) && housekeeping_cpu(cpu
, HK_FLAG_TIMER
))
507 for_each_domain(cpu
, sd
) {
508 for_each_cpu(i
, sched_domain_span(sd
)) {
512 if (!idle_cpu(i
) && housekeeping_cpu(i
, HK_FLAG_TIMER
)) {
519 if (!housekeeping_cpu(cpu
, HK_FLAG_TIMER
))
520 cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
527 * When add_timer_on() enqueues a timer into the timer wheel of an
528 * idle CPU then this timer might expire before the next timer event
529 * which is scheduled to wake up that CPU. In case of a completely
530 * idle system the next event might even be infinite time into the
531 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
532 * leaves the inner idle loop so the newly added timer is taken into
533 * account when the CPU goes back to idle and evaluates the timer
534 * wheel for the next timer event.
536 static void wake_up_idle_cpu(int cpu
)
538 struct rq
*rq
= cpu_rq(cpu
);
540 if (cpu
== smp_processor_id())
543 if (set_nr_and_not_polling(rq
->idle
))
544 smp_send_reschedule(cpu
);
546 trace_sched_wake_idle_without_ipi(cpu
);
549 static bool wake_up_full_nohz_cpu(int cpu
)
552 * We just need the target to call irq_exit() and re-evaluate
553 * the next tick. The nohz full kick at least implies that.
554 * If needed we can still optimize that later with an
557 if (cpu_is_offline(cpu
))
558 return true; /* Don't try to wake offline CPUs. */
559 if (tick_nohz_full_cpu(cpu
)) {
560 if (cpu
!= smp_processor_id() ||
561 tick_nohz_tick_stopped())
562 tick_nohz_full_kick_cpu(cpu
);
570 * Wake up the specified CPU. If the CPU is going offline, it is the
571 * caller's responsibility to deal with the lost wakeup, for example,
572 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
574 void wake_up_nohz_cpu(int cpu
)
576 if (!wake_up_full_nohz_cpu(cpu
))
577 wake_up_idle_cpu(cpu
);
580 static inline bool got_nohz_idle_kick(void)
582 int cpu
= smp_processor_id();
584 if (!(atomic_read(nohz_flags(cpu
)) & NOHZ_KICK_MASK
))
587 if (idle_cpu(cpu
) && !need_resched())
591 * We can't run Idle Load Balance on this CPU for this time so we
592 * cancel it and clear NOHZ_BALANCE_KICK
594 atomic_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
598 #else /* CONFIG_NO_HZ_COMMON */
600 static inline bool got_nohz_idle_kick(void)
605 #endif /* CONFIG_NO_HZ_COMMON */
607 #ifdef CONFIG_NO_HZ_FULL
608 bool sched_can_stop_tick(struct rq
*rq
)
612 /* Deadline tasks, even if single, need the tick */
613 if (rq
->dl
.dl_nr_running
)
617 * If there are more than one RR tasks, we need the tick to effect the
618 * actual RR behaviour.
620 if (rq
->rt
.rr_nr_running
) {
621 if (rq
->rt
.rr_nr_running
== 1)
628 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
629 * forced preemption between FIFO tasks.
631 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
636 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
637 * if there's more than one we need the tick for involuntary
640 if (rq
->nr_running
> 1)
645 #endif /* CONFIG_NO_HZ_FULL */
646 #endif /* CONFIG_SMP */
648 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
649 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
651 * Iterate task_group tree rooted at *from, calling @down when first entering a
652 * node and @up when leaving it for the final time.
654 * Caller must hold rcu_lock or sufficient equivalent.
656 int walk_tg_tree_from(struct task_group
*from
,
657 tg_visitor down
, tg_visitor up
, void *data
)
659 struct task_group
*parent
, *child
;
665 ret
= (*down
)(parent
, data
);
668 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
675 ret
= (*up
)(parent
, data
);
676 if (ret
|| parent
== from
)
680 parent
= parent
->parent
;
687 int tg_nop(struct task_group
*tg
, void *data
)
693 static void set_load_weight(struct task_struct
*p
, bool update_load
)
695 int prio
= p
->static_prio
- MAX_RT_PRIO
;
696 struct load_weight
*load
= &p
->se
.load
;
699 * SCHED_IDLE tasks get minimal weight:
701 if (idle_policy(p
->policy
)) {
702 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
703 load
->inv_weight
= WMULT_IDLEPRIO
;
708 * SCHED_OTHER tasks have to update their load when changing their
711 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
712 reweight_task(p
, prio
);
714 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
715 load
->inv_weight
= sched_prio_to_wmult
[prio
];
719 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
721 if (!(flags
& ENQUEUE_NOCLOCK
))
724 if (!(flags
& ENQUEUE_RESTORE
))
725 sched_info_queued(rq
, p
);
727 p
->sched_class
->enqueue_task(rq
, p
, flags
);
730 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
732 if (!(flags
& DEQUEUE_NOCLOCK
))
735 if (!(flags
& DEQUEUE_SAVE
))
736 sched_info_dequeued(rq
, p
);
738 p
->sched_class
->dequeue_task(rq
, p
, flags
);
741 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
743 if (task_contributes_to_load(p
))
744 rq
->nr_uninterruptible
--;
746 enqueue_task(rq
, p
, flags
);
749 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
751 if (task_contributes_to_load(p
))
752 rq
->nr_uninterruptible
++;
754 dequeue_task(rq
, p
, flags
);
758 * __normal_prio - return the priority that is based on the static prio
760 static inline int __normal_prio(struct task_struct
*p
)
762 return p
->static_prio
;
766 * Calculate the expected normal priority: i.e. priority
767 * without taking RT-inheritance into account. Might be
768 * boosted by interactivity modifiers. Changes upon fork,
769 * setprio syscalls, and whenever the interactivity
770 * estimator recalculates.
772 static inline int normal_prio(struct task_struct
*p
)
776 if (task_has_dl_policy(p
))
777 prio
= MAX_DL_PRIO
-1;
778 else if (task_has_rt_policy(p
))
779 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
781 prio
= __normal_prio(p
);
786 * Calculate the current priority, i.e. the priority
787 * taken into account by the scheduler. This value might
788 * be boosted by RT tasks, or might be boosted by
789 * interactivity modifiers. Will be RT if the task got
790 * RT-boosted. If not then it returns p->normal_prio.
792 static int effective_prio(struct task_struct
*p
)
794 p
->normal_prio
= normal_prio(p
);
796 * If we are RT tasks or we were boosted to RT priority,
797 * keep the priority unchanged. Otherwise, update priority
798 * to the normal priority:
800 if (!rt_prio(p
->prio
))
801 return p
->normal_prio
;
806 * task_curr - is this task currently executing on a CPU?
807 * @p: the task in question.
809 * Return: 1 if the task is currently executing. 0 otherwise.
811 inline int task_curr(const struct task_struct
*p
)
813 return cpu_curr(task_cpu(p
)) == p
;
817 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
818 * use the balance_callback list if you want balancing.
820 * this means any call to check_class_changed() must be followed by a call to
821 * balance_callback().
823 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
824 const struct sched_class
*prev_class
,
827 if (prev_class
!= p
->sched_class
) {
828 if (prev_class
->switched_from
)
829 prev_class
->switched_from(rq
, p
);
831 p
->sched_class
->switched_to(rq
, p
);
832 } else if (oldprio
!= p
->prio
|| dl_task(p
))
833 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
836 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
838 const struct sched_class
*class;
840 if (p
->sched_class
== rq
->curr
->sched_class
) {
841 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
843 for_each_class(class) {
844 if (class == rq
->curr
->sched_class
)
846 if (class == p
->sched_class
) {
854 * A queue event has occurred, and we're going to schedule. In
855 * this case, we can save a useless back to back clock update.
857 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
858 rq_clock_skip_update(rq
);
863 static inline bool is_per_cpu_kthread(struct task_struct
*p
)
865 if (!(p
->flags
& PF_KTHREAD
))
868 if (p
->nr_cpus_allowed
!= 1)
875 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
876 * __set_cpus_allowed_ptr() and select_fallback_rq().
878 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
880 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
883 if (is_per_cpu_kthread(p
))
884 return cpu_online(cpu
);
886 return cpu_active(cpu
);
890 * This is how migration works:
892 * 1) we invoke migration_cpu_stop() on the target CPU using
894 * 2) stopper starts to run (implicitly forcing the migrated thread
896 * 3) it checks whether the migrated task is still in the wrong runqueue.
897 * 4) if it's in the wrong runqueue then the migration thread removes
898 * it and puts it into the right queue.
899 * 5) stopper completes and stop_one_cpu() returns and the migration
904 * move_queued_task - move a queued task to new rq.
906 * Returns (locked) new rq. Old rq's lock is released.
908 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
909 struct task_struct
*p
, int new_cpu
)
911 lockdep_assert_held(&rq
->lock
);
913 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
914 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
915 set_task_cpu(p
, new_cpu
);
918 rq
= cpu_rq(new_cpu
);
921 BUG_ON(task_cpu(p
) != new_cpu
);
922 enqueue_task(rq
, p
, 0);
923 p
->on_rq
= TASK_ON_RQ_QUEUED
;
924 check_preempt_curr(rq
, p
, 0);
929 struct migration_arg
{
930 struct task_struct
*task
;
935 * Move (not current) task off this CPU, onto the destination CPU. We're doing
936 * this because either it can't run here any more (set_cpus_allowed()
937 * away from this CPU, or CPU going down), or because we're
938 * attempting to rebalance this task on exec (sched_exec).
940 * So we race with normal scheduler movements, but that's OK, as long
941 * as the task is no longer on this CPU.
943 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
944 struct task_struct
*p
, int dest_cpu
)
946 /* Affinity changed (again). */
947 if (!is_cpu_allowed(p
, dest_cpu
))
951 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
957 * migration_cpu_stop - this will be executed by a highprio stopper thread
958 * and performs thread migration by bumping thread off CPU then
959 * 'pushing' onto another runqueue.
961 static int migration_cpu_stop(void *data
)
963 struct migration_arg
*arg
= data
;
964 struct task_struct
*p
= arg
->task
;
965 struct rq
*rq
= this_rq();
969 * The original target CPU might have gone down and we might
970 * be on another CPU but it doesn't matter.
974 * We need to explicitly wake pending tasks before running
975 * __migrate_task() such that we will not miss enforcing cpus_allowed
976 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
978 sched_ttwu_pending();
980 raw_spin_lock(&p
->pi_lock
);
983 * If task_rq(p) != rq, it cannot be migrated here, because we're
984 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
985 * we're holding p->pi_lock.
987 if (task_rq(p
) == rq
) {
988 if (task_on_rq_queued(p
))
989 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
991 p
->wake_cpu
= arg
->dest_cpu
;
994 raw_spin_unlock(&p
->pi_lock
);
1001 * sched_class::set_cpus_allowed must do the below, but is not required to
1002 * actually call this function.
1004 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1006 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1007 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1010 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1012 struct rq
*rq
= task_rq(p
);
1013 bool queued
, running
;
1015 lockdep_assert_held(&p
->pi_lock
);
1017 queued
= task_on_rq_queued(p
);
1018 running
= task_current(rq
, p
);
1022 * Because __kthread_bind() calls this on blocked tasks without
1025 lockdep_assert_held(&rq
->lock
);
1026 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1029 put_prev_task(rq
, p
);
1031 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1034 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1036 set_curr_task(rq
, p
);
1040 * Change a given task's CPU affinity. Migrate the thread to a
1041 * proper CPU and schedule it away if the CPU it's executing on
1042 * is removed from the allowed bitmask.
1044 * NOTE: the caller must have a valid reference to the task, the
1045 * task must not exit() & deallocate itself prematurely. The
1046 * call is not atomic; no spinlocks may be held.
1048 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1049 const struct cpumask
*new_mask
, bool check
)
1051 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1052 unsigned int dest_cpu
;
1057 rq
= task_rq_lock(p
, &rf
);
1058 update_rq_clock(rq
);
1060 if (p
->flags
& PF_KTHREAD
) {
1062 * Kernel threads are allowed on online && !active CPUs
1064 cpu_valid_mask
= cpu_online_mask
;
1068 * Must re-check here, to close a race against __kthread_bind(),
1069 * sched_setaffinity() is not guaranteed to observe the flag.
1071 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1076 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1079 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1084 do_set_cpus_allowed(p
, new_mask
);
1086 if (p
->flags
& PF_KTHREAD
) {
1088 * For kernel threads that do indeed end up on online &&
1089 * !active we want to ensure they are strict per-CPU threads.
1091 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1092 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1093 p
->nr_cpus_allowed
!= 1);
1096 /* Can the task run on the task's current CPU? If so, we're done */
1097 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1100 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1101 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1102 struct migration_arg arg
= { p
, dest_cpu
};
1103 /* Need help from migration thread: drop lock and wait. */
1104 task_rq_unlock(rq
, p
, &rf
);
1105 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1106 tlb_migrate_finish(p
->mm
);
1108 } else if (task_on_rq_queued(p
)) {
1110 * OK, since we're going to drop the lock immediately
1111 * afterwards anyway.
1113 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1116 task_rq_unlock(rq
, p
, &rf
);
1121 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1123 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1125 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1127 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1129 #ifdef CONFIG_SCHED_DEBUG
1131 * We should never call set_task_cpu() on a blocked task,
1132 * ttwu() will sort out the placement.
1134 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1138 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1139 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1140 * time relying on p->on_rq.
1142 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1143 p
->sched_class
== &fair_sched_class
&&
1144 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1146 #ifdef CONFIG_LOCKDEP
1148 * The caller should hold either p->pi_lock or rq->lock, when changing
1149 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1151 * sched_move_task() holds both and thus holding either pins the cgroup,
1154 * Furthermore, all task_rq users should acquire both locks, see
1157 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1158 lockdep_is_held(&task_rq(p
)->lock
)));
1161 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1163 WARN_ON_ONCE(!cpu_online(new_cpu
));
1166 trace_sched_migrate_task(p
, new_cpu
);
1168 if (task_cpu(p
) != new_cpu
) {
1169 if (p
->sched_class
->migrate_task_rq
)
1170 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1171 p
->se
.nr_migrations
++;
1173 perf_event_task_migrate(p
);
1176 __set_task_cpu(p
, new_cpu
);
1179 #ifdef CONFIG_NUMA_BALANCING
1180 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1182 if (task_on_rq_queued(p
)) {
1183 struct rq
*src_rq
, *dst_rq
;
1184 struct rq_flags srf
, drf
;
1186 src_rq
= task_rq(p
);
1187 dst_rq
= cpu_rq(cpu
);
1189 rq_pin_lock(src_rq
, &srf
);
1190 rq_pin_lock(dst_rq
, &drf
);
1192 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1193 deactivate_task(src_rq
, p
, 0);
1194 set_task_cpu(p
, cpu
);
1195 activate_task(dst_rq
, p
, 0);
1196 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1197 check_preempt_curr(dst_rq
, p
, 0);
1199 rq_unpin_lock(dst_rq
, &drf
);
1200 rq_unpin_lock(src_rq
, &srf
);
1204 * Task isn't running anymore; make it appear like we migrated
1205 * it before it went to sleep. This means on wakeup we make the
1206 * previous CPU our target instead of where it really is.
1212 struct migration_swap_arg
{
1213 struct task_struct
*src_task
, *dst_task
;
1214 int src_cpu
, dst_cpu
;
1217 static int migrate_swap_stop(void *data
)
1219 struct migration_swap_arg
*arg
= data
;
1220 struct rq
*src_rq
, *dst_rq
;
1223 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1226 src_rq
= cpu_rq(arg
->src_cpu
);
1227 dst_rq
= cpu_rq(arg
->dst_cpu
);
1229 double_raw_lock(&arg
->src_task
->pi_lock
,
1230 &arg
->dst_task
->pi_lock
);
1231 double_rq_lock(src_rq
, dst_rq
);
1233 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1236 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1239 if (!cpumask_test_cpu(arg
->dst_cpu
, &arg
->src_task
->cpus_allowed
))
1242 if (!cpumask_test_cpu(arg
->src_cpu
, &arg
->dst_task
->cpus_allowed
))
1245 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1246 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1251 double_rq_unlock(src_rq
, dst_rq
);
1252 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1253 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1259 * Cross migrate two tasks
1261 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
1262 int target_cpu
, int curr_cpu
)
1264 struct migration_swap_arg arg
;
1267 arg
= (struct migration_swap_arg
){
1269 .src_cpu
= curr_cpu
,
1271 .dst_cpu
= target_cpu
,
1274 if (arg
.src_cpu
== arg
.dst_cpu
)
1278 * These three tests are all lockless; this is OK since all of them
1279 * will be re-checked with proper locks held further down the line.
1281 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1284 if (!cpumask_test_cpu(arg
.dst_cpu
, &arg
.src_task
->cpus_allowed
))
1287 if (!cpumask_test_cpu(arg
.src_cpu
, &arg
.dst_task
->cpus_allowed
))
1290 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1291 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1296 #endif /* CONFIG_NUMA_BALANCING */
1299 * wait_task_inactive - wait for a thread to unschedule.
1301 * If @match_state is nonzero, it's the @p->state value just checked and
1302 * not expected to change. If it changes, i.e. @p might have woken up,
1303 * then return zero. When we succeed in waiting for @p to be off its CPU,
1304 * we return a positive number (its total switch count). If a second call
1305 * a short while later returns the same number, the caller can be sure that
1306 * @p has remained unscheduled the whole time.
1308 * The caller must ensure that the task *will* unschedule sometime soon,
1309 * else this function might spin for a *long* time. This function can't
1310 * be called with interrupts off, or it may introduce deadlock with
1311 * smp_call_function() if an IPI is sent by the same process we are
1312 * waiting to become inactive.
1314 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1316 int running
, queued
;
1323 * We do the initial early heuristics without holding
1324 * any task-queue locks at all. We'll only try to get
1325 * the runqueue lock when things look like they will
1331 * If the task is actively running on another CPU
1332 * still, just relax and busy-wait without holding
1335 * NOTE! Since we don't hold any locks, it's not
1336 * even sure that "rq" stays as the right runqueue!
1337 * But we don't care, since "task_running()" will
1338 * return false if the runqueue has changed and p
1339 * is actually now running somewhere else!
1341 while (task_running(rq
, p
)) {
1342 if (match_state
&& unlikely(p
->state
!= match_state
))
1348 * Ok, time to look more closely! We need the rq
1349 * lock now, to be *sure*. If we're wrong, we'll
1350 * just go back and repeat.
1352 rq
= task_rq_lock(p
, &rf
);
1353 trace_sched_wait_task(p
);
1354 running
= task_running(rq
, p
);
1355 queued
= task_on_rq_queued(p
);
1357 if (!match_state
|| p
->state
== match_state
)
1358 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1359 task_rq_unlock(rq
, p
, &rf
);
1362 * If it changed from the expected state, bail out now.
1364 if (unlikely(!ncsw
))
1368 * Was it really running after all now that we
1369 * checked with the proper locks actually held?
1371 * Oops. Go back and try again..
1373 if (unlikely(running
)) {
1379 * It's not enough that it's not actively running,
1380 * it must be off the runqueue _entirely_, and not
1383 * So if it was still runnable (but just not actively
1384 * running right now), it's preempted, and we should
1385 * yield - it could be a while.
1387 if (unlikely(queued
)) {
1388 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1390 set_current_state(TASK_UNINTERRUPTIBLE
);
1391 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1396 * Ahh, all good. It wasn't running, and it wasn't
1397 * runnable, which means that it will never become
1398 * running in the future either. We're all done!
1407 * kick_process - kick a running thread to enter/exit the kernel
1408 * @p: the to-be-kicked thread
1410 * Cause a process which is running on another CPU to enter
1411 * kernel-mode, without any delay. (to get signals handled.)
1413 * NOTE: this function doesn't have to take the runqueue lock,
1414 * because all it wants to ensure is that the remote task enters
1415 * the kernel. If the IPI races and the task has been migrated
1416 * to another CPU then no harm is done and the purpose has been
1419 void kick_process(struct task_struct
*p
)
1425 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1426 smp_send_reschedule(cpu
);
1429 EXPORT_SYMBOL_GPL(kick_process
);
1432 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1434 * A few notes on cpu_active vs cpu_online:
1436 * - cpu_active must be a subset of cpu_online
1438 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1439 * see __set_cpus_allowed_ptr(). At this point the newly online
1440 * CPU isn't yet part of the sched domains, and balancing will not
1443 * - on CPU-down we clear cpu_active() to mask the sched domains and
1444 * avoid the load balancer to place new tasks on the to be removed
1445 * CPU. Existing tasks will remain running there and will be taken
1448 * This means that fallback selection must not select !active CPUs.
1449 * And can assume that any active CPU must be online. Conversely
1450 * select_task_rq() below may allow selection of !active CPUs in order
1451 * to satisfy the above rules.
1453 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1455 int nid
= cpu_to_node(cpu
);
1456 const struct cpumask
*nodemask
= NULL
;
1457 enum { cpuset
, possible
, fail
} state
= cpuset
;
1461 * If the node that the CPU is on has been offlined, cpu_to_node()
1462 * will return -1. There is no CPU on the node, and we should
1463 * select the CPU on the other node.
1466 nodemask
= cpumask_of_node(nid
);
1468 /* Look for allowed, online CPU in same node. */
1469 for_each_cpu(dest_cpu
, nodemask
) {
1470 if (!cpu_active(dest_cpu
))
1472 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
1478 /* Any allowed, online CPU? */
1479 for_each_cpu(dest_cpu
, &p
->cpus_allowed
) {
1480 if (!is_cpu_allowed(p
, dest_cpu
))
1486 /* No more Mr. Nice Guy. */
1489 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1490 cpuset_cpus_allowed_fallback(p
);
1496 do_set_cpus_allowed(p
, cpu_possible_mask
);
1507 if (state
!= cpuset
) {
1509 * Don't tell them about moving exiting tasks or
1510 * kernel threads (both mm NULL), since they never
1513 if (p
->mm
&& printk_ratelimit()) {
1514 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1515 task_pid_nr(p
), p
->comm
, cpu
);
1523 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1526 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1528 lockdep_assert_held(&p
->pi_lock
);
1530 if (p
->nr_cpus_allowed
> 1)
1531 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1533 cpu
= cpumask_any(&p
->cpus_allowed
);
1536 * In order not to call set_task_cpu() on a blocking task we need
1537 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1540 * Since this is common to all placement strategies, this lives here.
1542 * [ this allows ->select_task() to simply return task_cpu(p) and
1543 * not worry about this generic constraint ]
1545 if (unlikely(!is_cpu_allowed(p
, cpu
)))
1546 cpu
= select_fallback_rq(task_cpu(p
), p
);
1551 static void update_avg(u64
*avg
, u64 sample
)
1553 s64 diff
= sample
- *avg
;
1557 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1559 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1560 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1564 * Make it appear like a SCHED_FIFO task, its something
1565 * userspace knows about and won't get confused about.
1567 * Also, it will make PI more or less work without too
1568 * much confusion -- but then, stop work should not
1569 * rely on PI working anyway.
1571 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1573 stop
->sched_class
= &stop_sched_class
;
1576 cpu_rq(cpu
)->stop
= stop
;
1580 * Reset it back to a normal scheduling class so that
1581 * it can die in pieces.
1583 old_stop
->sched_class
= &rt_sched_class
;
1589 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1590 const struct cpumask
*new_mask
, bool check
)
1592 return set_cpus_allowed_ptr(p
, new_mask
);
1595 #endif /* CONFIG_SMP */
1598 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1602 if (!schedstat_enabled())
1608 if (cpu
== rq
->cpu
) {
1609 __schedstat_inc(rq
->ttwu_local
);
1610 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1612 struct sched_domain
*sd
;
1614 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1616 for_each_domain(rq
->cpu
, sd
) {
1617 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1618 __schedstat_inc(sd
->ttwu_wake_remote
);
1625 if (wake_flags
& WF_MIGRATED
)
1626 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1627 #endif /* CONFIG_SMP */
1629 __schedstat_inc(rq
->ttwu_count
);
1630 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1632 if (wake_flags
& WF_SYNC
)
1633 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1636 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1638 activate_task(rq
, p
, en_flags
);
1639 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1641 /* If a worker is waking up, notify the workqueue: */
1642 if (p
->flags
& PF_WQ_WORKER
)
1643 wq_worker_waking_up(p
, cpu_of(rq
));
1647 * Mark the task runnable and perform wakeup-preemption.
1649 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1650 struct rq_flags
*rf
)
1652 check_preempt_curr(rq
, p
, wake_flags
);
1653 p
->state
= TASK_RUNNING
;
1654 trace_sched_wakeup(p
);
1657 if (p
->sched_class
->task_woken
) {
1659 * Our task @p is fully woken up and running; so its safe to
1660 * drop the rq->lock, hereafter rq is only used for statistics.
1662 rq_unpin_lock(rq
, rf
);
1663 p
->sched_class
->task_woken(rq
, p
);
1664 rq_repin_lock(rq
, rf
);
1667 if (rq
->idle_stamp
) {
1668 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1669 u64 max
= 2*rq
->max_idle_balance_cost
;
1671 update_avg(&rq
->avg_idle
, delta
);
1673 if (rq
->avg_idle
> max
)
1682 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1683 struct rq_flags
*rf
)
1685 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
1687 lockdep_assert_held(&rq
->lock
);
1690 if (p
->sched_contributes_to_load
)
1691 rq
->nr_uninterruptible
--;
1693 if (wake_flags
& WF_MIGRATED
)
1694 en_flags
|= ENQUEUE_MIGRATED
;
1697 ttwu_activate(rq
, p
, en_flags
);
1698 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1702 * Called in case the task @p isn't fully descheduled from its runqueue,
1703 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1704 * since all we need to do is flip p->state to TASK_RUNNING, since
1705 * the task is still ->on_rq.
1707 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1713 rq
= __task_rq_lock(p
, &rf
);
1714 if (task_on_rq_queued(p
)) {
1715 /* check_preempt_curr() may use rq clock */
1716 update_rq_clock(rq
);
1717 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1720 __task_rq_unlock(rq
, &rf
);
1726 void sched_ttwu_pending(void)
1728 struct rq
*rq
= this_rq();
1729 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1730 struct task_struct
*p
, *t
;
1736 rq_lock_irqsave(rq
, &rf
);
1737 update_rq_clock(rq
);
1739 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
1740 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
1742 rq_unlock_irqrestore(rq
, &rf
);
1745 void scheduler_ipi(void)
1748 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1749 * TIF_NEED_RESCHED remotely (for the first time) will also send
1752 preempt_fold_need_resched();
1754 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1758 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1759 * traditionally all their work was done from the interrupt return
1760 * path. Now that we actually do some work, we need to make sure
1763 * Some archs already do call them, luckily irq_enter/exit nest
1766 * Arguably we should visit all archs and update all handlers,
1767 * however a fair share of IPIs are still resched only so this would
1768 * somewhat pessimize the simple resched case.
1771 sched_ttwu_pending();
1774 * Check if someone kicked us for doing the nohz idle load balance.
1776 if (unlikely(got_nohz_idle_kick())) {
1777 this_rq()->idle_balance
= 1;
1778 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1783 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1785 struct rq
*rq
= cpu_rq(cpu
);
1787 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1789 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1790 if (!set_nr_if_polling(rq
->idle
))
1791 smp_send_reschedule(cpu
);
1793 trace_sched_wake_idle_without_ipi(cpu
);
1797 void wake_up_if_idle(int cpu
)
1799 struct rq
*rq
= cpu_rq(cpu
);
1804 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1807 if (set_nr_if_polling(rq
->idle
)) {
1808 trace_sched_wake_idle_without_ipi(cpu
);
1810 rq_lock_irqsave(rq
, &rf
);
1811 if (is_idle_task(rq
->curr
))
1812 smp_send_reschedule(cpu
);
1813 /* Else CPU is not idle, do nothing here: */
1814 rq_unlock_irqrestore(rq
, &rf
);
1821 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1823 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1825 #endif /* CONFIG_SMP */
1827 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1829 struct rq
*rq
= cpu_rq(cpu
);
1832 #if defined(CONFIG_SMP)
1833 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1834 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
1835 ttwu_queue_remote(p
, cpu
, wake_flags
);
1841 update_rq_clock(rq
);
1842 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1847 * Notes on Program-Order guarantees on SMP systems.
1851 * The basic program-order guarantee on SMP systems is that when a task [t]
1852 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1853 * execution on its new CPU [c1].
1855 * For migration (of runnable tasks) this is provided by the following means:
1857 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1858 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1859 * rq(c1)->lock (if not at the same time, then in that order).
1860 * C) LOCK of the rq(c1)->lock scheduling in task
1862 * Release/acquire chaining guarantees that B happens after A and C after B.
1863 * Note: the CPU doing B need not be c0 or c1
1872 * UNLOCK rq(0)->lock
1874 * LOCK rq(0)->lock // orders against CPU0
1876 * UNLOCK rq(0)->lock
1880 * UNLOCK rq(1)->lock
1882 * LOCK rq(1)->lock // orders against CPU2
1885 * UNLOCK rq(1)->lock
1888 * BLOCKING -- aka. SLEEP + WAKEUP
1890 * For blocking we (obviously) need to provide the same guarantee as for
1891 * migration. However the means are completely different as there is no lock
1892 * chain to provide order. Instead we do:
1894 * 1) smp_store_release(X->on_cpu, 0)
1895 * 2) smp_cond_load_acquire(!X->on_cpu)
1899 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1901 * LOCK rq(0)->lock LOCK X->pi_lock
1904 * smp_store_release(X->on_cpu, 0);
1906 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1912 * X->state = RUNNING
1913 * UNLOCK rq(2)->lock
1915 * LOCK rq(2)->lock // orders against CPU1
1918 * UNLOCK rq(2)->lock
1921 * UNLOCK rq(0)->lock
1924 * However, for wakeups there is a second guarantee we must provide, namely we
1925 * must ensure that CONDITION=1 done by the caller can not be reordered with
1926 * accesses to the task state; see try_to_wake_up() and set_current_state().
1930 * try_to_wake_up - wake up a thread
1931 * @p: the thread to be awakened
1932 * @state: the mask of task states that can be woken
1933 * @wake_flags: wake modifier flags (WF_*)
1935 * If (@state & @p->state) @p->state = TASK_RUNNING.
1937 * If the task was not queued/runnable, also place it back on a runqueue.
1939 * Atomic against schedule() which would dequeue a task, also see
1940 * set_current_state().
1942 * This function executes a full memory barrier before accessing the task
1943 * state; see set_current_state().
1945 * Return: %true if @p->state changes (an actual wakeup was done),
1949 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1951 unsigned long flags
;
1952 int cpu
, success
= 0;
1955 * If we are going to wake up a thread waiting for CONDITION we
1956 * need to ensure that CONDITION=1 done by the caller can not be
1957 * reordered with p->state check below. This pairs with mb() in
1958 * set_current_state() the waiting thread does.
1960 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1961 smp_mb__after_spinlock();
1962 if (!(p
->state
& state
))
1965 trace_sched_waking(p
);
1967 /* We're going to change ->state: */
1972 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1973 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1974 * in smp_cond_load_acquire() below.
1976 * sched_ttwu_pending() try_to_wake_up()
1977 * STORE p->on_rq = 1 LOAD p->state
1980 * __schedule() (switch to task 'p')
1981 * LOCK rq->lock smp_rmb();
1982 * smp_mb__after_spinlock();
1986 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
1988 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1989 * __schedule(). See the comment for smp_mb__after_spinlock().
1992 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1997 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1998 * possible to, falsely, observe p->on_cpu == 0.
2000 * One must be running (->on_cpu == 1) in order to remove oneself
2001 * from the runqueue.
2003 * __schedule() (switch to task 'p') try_to_wake_up()
2004 * STORE p->on_cpu = 1 LOAD p->on_rq
2007 * __schedule() (put 'p' to sleep)
2008 * LOCK rq->lock smp_rmb();
2009 * smp_mb__after_spinlock();
2010 * STORE p->on_rq = 0 LOAD p->on_cpu
2012 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2013 * __schedule(). See the comment for smp_mb__after_spinlock().
2018 * If the owning (remote) CPU is still in the middle of schedule() with
2019 * this task as prev, wait until its done referencing the task.
2021 * Pairs with the smp_store_release() in finish_task().
2023 * This ensures that tasks getting woken will be fully ordered against
2024 * their previous state and preserve Program Order.
2026 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2028 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2029 p
->state
= TASK_WAKING
;
2032 delayacct_blkio_end(p
);
2033 atomic_dec(&task_rq(p
)->nr_iowait
);
2036 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2037 if (task_cpu(p
) != cpu
) {
2038 wake_flags
|= WF_MIGRATED
;
2039 set_task_cpu(p
, cpu
);
2042 #else /* CONFIG_SMP */
2045 delayacct_blkio_end(p
);
2046 atomic_dec(&task_rq(p
)->nr_iowait
);
2049 #endif /* CONFIG_SMP */
2051 ttwu_queue(p
, cpu
, wake_flags
);
2053 ttwu_stat(p
, cpu
, wake_flags
);
2055 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2061 * try_to_wake_up_local - try to wake up a local task with rq lock held
2062 * @p: the thread to be awakened
2063 * @rf: request-queue flags for pinning
2065 * Put @p on the run-queue if it's not already there. The caller must
2066 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2069 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2071 struct rq
*rq
= task_rq(p
);
2073 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2074 WARN_ON_ONCE(p
== current
))
2077 lockdep_assert_held(&rq
->lock
);
2079 if (!raw_spin_trylock(&p
->pi_lock
)) {
2081 * This is OK, because current is on_cpu, which avoids it being
2082 * picked for load-balance and preemption/IRQs are still
2083 * disabled avoiding further scheduler activity on it and we've
2084 * not yet picked a replacement task.
2087 raw_spin_lock(&p
->pi_lock
);
2091 if (!(p
->state
& TASK_NORMAL
))
2094 trace_sched_waking(p
);
2096 if (!task_on_rq_queued(p
)) {
2098 delayacct_blkio_end(p
);
2099 atomic_dec(&rq
->nr_iowait
);
2101 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
);
2104 ttwu_do_wakeup(rq
, p
, 0, rf
);
2105 ttwu_stat(p
, smp_processor_id(), 0);
2107 raw_spin_unlock(&p
->pi_lock
);
2111 * wake_up_process - Wake up a specific process
2112 * @p: The process to be woken up.
2114 * Attempt to wake up the nominated process and move it to the set of runnable
2117 * Return: 1 if the process was woken up, 0 if it was already running.
2119 * This function executes a full memory barrier before accessing the task state.
2121 int wake_up_process(struct task_struct
*p
)
2123 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2125 EXPORT_SYMBOL(wake_up_process
);
2127 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2129 return try_to_wake_up(p
, state
, 0);
2133 * Perform scheduler related setup for a newly forked process p.
2134 * p is forked by current.
2136 * __sched_fork() is basic setup used by init_idle() too:
2138 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2143 p
->se
.exec_start
= 0;
2144 p
->se
.sum_exec_runtime
= 0;
2145 p
->se
.prev_sum_exec_runtime
= 0;
2146 p
->se
.nr_migrations
= 0;
2148 INIT_LIST_HEAD(&p
->se
.group_node
);
2150 #ifdef CONFIG_FAIR_GROUP_SCHED
2151 p
->se
.cfs_rq
= NULL
;
2154 #ifdef CONFIG_SCHEDSTATS
2155 /* Even if schedstat is disabled, there should not be garbage */
2156 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2159 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2160 init_dl_task_timer(&p
->dl
);
2161 init_dl_inactive_task_timer(&p
->dl
);
2162 __dl_clear_params(p
);
2164 INIT_LIST_HEAD(&p
->rt
.run_list
);
2166 p
->rt
.time_slice
= sched_rr_timeslice
;
2170 #ifdef CONFIG_PREEMPT_NOTIFIERS
2171 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2174 init_numa_balancing(clone_flags
, p
);
2177 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2179 #ifdef CONFIG_NUMA_BALANCING
2181 void set_numabalancing_state(bool enabled
)
2184 static_branch_enable(&sched_numa_balancing
);
2186 static_branch_disable(&sched_numa_balancing
);
2189 #ifdef CONFIG_PROC_SYSCTL
2190 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2191 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2195 int state
= static_branch_likely(&sched_numa_balancing
);
2197 if (write
&& !capable(CAP_SYS_ADMIN
))
2202 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2206 set_numabalancing_state(state
);
2212 #ifdef CONFIG_SCHEDSTATS
2214 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2215 static bool __initdata __sched_schedstats
= false;
2217 static void set_schedstats(bool enabled
)
2220 static_branch_enable(&sched_schedstats
);
2222 static_branch_disable(&sched_schedstats
);
2225 void force_schedstat_enabled(void)
2227 if (!schedstat_enabled()) {
2228 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2229 static_branch_enable(&sched_schedstats
);
2233 static int __init
setup_schedstats(char *str
)
2240 * This code is called before jump labels have been set up, so we can't
2241 * change the static branch directly just yet. Instead set a temporary
2242 * variable so init_schedstats() can do it later.
2244 if (!strcmp(str
, "enable")) {
2245 __sched_schedstats
= true;
2247 } else if (!strcmp(str
, "disable")) {
2248 __sched_schedstats
= false;
2253 pr_warn("Unable to parse schedstats=\n");
2257 __setup("schedstats=", setup_schedstats
);
2259 static void __init
init_schedstats(void)
2261 set_schedstats(__sched_schedstats
);
2264 #ifdef CONFIG_PROC_SYSCTL
2265 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2266 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2270 int state
= static_branch_likely(&sched_schedstats
);
2272 if (write
&& !capable(CAP_SYS_ADMIN
))
2277 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2281 set_schedstats(state
);
2284 #endif /* CONFIG_PROC_SYSCTL */
2285 #else /* !CONFIG_SCHEDSTATS */
2286 static inline void init_schedstats(void) {}
2287 #endif /* CONFIG_SCHEDSTATS */
2290 * fork()/clone()-time setup:
2292 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2294 unsigned long flags
;
2296 __sched_fork(clone_flags
, p
);
2298 * We mark the process as NEW here. This guarantees that
2299 * nobody will actually run it, and a signal or other external
2300 * event cannot wake it up and insert it on the runqueue either.
2302 p
->state
= TASK_NEW
;
2305 * Make sure we do not leak PI boosting priority to the child.
2307 p
->prio
= current
->normal_prio
;
2310 * Revert to default priority/policy on fork if requested.
2312 if (unlikely(p
->sched_reset_on_fork
)) {
2313 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2314 p
->policy
= SCHED_NORMAL
;
2315 p
->static_prio
= NICE_TO_PRIO(0);
2317 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2318 p
->static_prio
= NICE_TO_PRIO(0);
2320 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2321 set_load_weight(p
, false);
2324 * We don't need the reset flag anymore after the fork. It has
2325 * fulfilled its duty:
2327 p
->sched_reset_on_fork
= 0;
2330 if (dl_prio(p
->prio
))
2332 else if (rt_prio(p
->prio
))
2333 p
->sched_class
= &rt_sched_class
;
2335 p
->sched_class
= &fair_sched_class
;
2337 init_entity_runnable_average(&p
->se
);
2340 * The child is not yet in the pid-hash so no cgroup attach races,
2341 * and the cgroup is pinned to this child due to cgroup_fork()
2342 * is ran before sched_fork().
2344 * Silence PROVE_RCU.
2346 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2348 * We're setting the CPU for the first time, we don't migrate,
2349 * so use __set_task_cpu().
2351 __set_task_cpu(p
, smp_processor_id());
2352 if (p
->sched_class
->task_fork
)
2353 p
->sched_class
->task_fork(p
);
2354 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2356 #ifdef CONFIG_SCHED_INFO
2357 if (likely(sched_info_on()))
2358 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2360 #if defined(CONFIG_SMP)
2363 init_task_preempt_count(p
);
2365 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2366 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2371 unsigned long to_ratio(u64 period
, u64 runtime
)
2373 if (runtime
== RUNTIME_INF
)
2377 * Doing this here saves a lot of checks in all
2378 * the calling paths, and returning zero seems
2379 * safe for them anyway.
2384 return div64_u64(runtime
<< BW_SHIFT
, period
);
2388 * wake_up_new_task - wake up a newly created task for the first time.
2390 * This function will do some initial scheduler statistics housekeeping
2391 * that must be done for every newly created context, then puts the task
2392 * on the runqueue and wakes it.
2394 void wake_up_new_task(struct task_struct
*p
)
2399 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2400 p
->state
= TASK_RUNNING
;
2403 * Fork balancing, do it here and not earlier because:
2404 * - cpus_allowed can change in the fork path
2405 * - any previously selected CPU might disappear through hotplug
2407 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2408 * as we're not fully set-up yet.
2410 p
->recent_used_cpu
= task_cpu(p
);
2411 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2413 rq
= __task_rq_lock(p
, &rf
);
2414 update_rq_clock(rq
);
2415 post_init_entity_util_avg(&p
->se
);
2417 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2418 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2419 trace_sched_wakeup_new(p
);
2420 check_preempt_curr(rq
, p
, WF_FORK
);
2422 if (p
->sched_class
->task_woken
) {
2424 * Nothing relies on rq->lock after this, so its fine to
2427 rq_unpin_lock(rq
, &rf
);
2428 p
->sched_class
->task_woken(rq
, p
);
2429 rq_repin_lock(rq
, &rf
);
2432 task_rq_unlock(rq
, p
, &rf
);
2435 #ifdef CONFIG_PREEMPT_NOTIFIERS
2437 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
2439 void preempt_notifier_inc(void)
2441 static_branch_inc(&preempt_notifier_key
);
2443 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2445 void preempt_notifier_dec(void)
2447 static_branch_dec(&preempt_notifier_key
);
2449 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2452 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2453 * @notifier: notifier struct to register
2455 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2457 if (!static_branch_unlikely(&preempt_notifier_key
))
2458 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2460 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2462 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2465 * preempt_notifier_unregister - no longer interested in preemption notifications
2466 * @notifier: notifier struct to unregister
2468 * This is *not* safe to call from within a preemption notifier.
2470 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2472 hlist_del(¬ifier
->link
);
2474 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2476 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2478 struct preempt_notifier
*notifier
;
2480 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2481 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2484 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2486 if (static_branch_unlikely(&preempt_notifier_key
))
2487 __fire_sched_in_preempt_notifiers(curr
);
2491 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2492 struct task_struct
*next
)
2494 struct preempt_notifier
*notifier
;
2496 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2497 notifier
->ops
->sched_out(notifier
, next
);
2500 static __always_inline
void
2501 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2502 struct task_struct
*next
)
2504 if (static_branch_unlikely(&preempt_notifier_key
))
2505 __fire_sched_out_preempt_notifiers(curr
, next
);
2508 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2510 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2515 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2516 struct task_struct
*next
)
2520 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2522 static inline void prepare_task(struct task_struct
*next
)
2526 * Claim the task as running, we do this before switching to it
2527 * such that any running task will have this set.
2533 static inline void finish_task(struct task_struct
*prev
)
2537 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2538 * We must ensure this doesn't happen until the switch is completely
2541 * In particular, the load of prev->state in finish_task_switch() must
2542 * happen before this.
2544 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2546 smp_store_release(&prev
->on_cpu
, 0);
2551 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
2554 * Since the runqueue lock will be released by the next
2555 * task (which is an invalid locking op but in the case
2556 * of the scheduler it's an obvious special-case), so we
2557 * do an early lockdep release here:
2559 rq_unpin_lock(rq
, rf
);
2560 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2561 #ifdef CONFIG_DEBUG_SPINLOCK
2562 /* this is a valid case when another task releases the spinlock */
2563 rq
->lock
.owner
= next
;
2567 static inline void finish_lock_switch(struct rq
*rq
)
2570 * If we are tracking spinlock dependencies then we have to
2571 * fix up the runqueue lock - which gets 'carried over' from
2572 * prev into current:
2574 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
2575 raw_spin_unlock_irq(&rq
->lock
);
2579 * NOP if the arch has not defined these:
2582 #ifndef prepare_arch_switch
2583 # define prepare_arch_switch(next) do { } while (0)
2586 #ifndef finish_arch_post_lock_switch
2587 # define finish_arch_post_lock_switch() do { } while (0)
2591 * prepare_task_switch - prepare to switch tasks
2592 * @rq: the runqueue preparing to switch
2593 * @prev: the current task that is being switched out
2594 * @next: the task we are going to switch to.
2596 * This is called with the rq lock held and interrupts off. It must
2597 * be paired with a subsequent finish_task_switch after the context
2600 * prepare_task_switch sets up locking and calls architecture specific
2604 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2605 struct task_struct
*next
)
2607 kcov_prepare_switch(prev
);
2608 sched_info_switch(rq
, prev
, next
);
2609 perf_event_task_sched_out(prev
, next
);
2611 fire_sched_out_preempt_notifiers(prev
, next
);
2613 prepare_arch_switch(next
);
2617 * finish_task_switch - clean up after a task-switch
2618 * @prev: the thread we just switched away from.
2620 * finish_task_switch must be called after the context switch, paired
2621 * with a prepare_task_switch call before the context switch.
2622 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2623 * and do any other architecture-specific cleanup actions.
2625 * Note that we may have delayed dropping an mm in context_switch(). If
2626 * so, we finish that here outside of the runqueue lock. (Doing it
2627 * with the lock held can cause deadlocks; see schedule() for
2630 * The context switch have flipped the stack from under us and restored the
2631 * local variables which were saved when this task called schedule() in the
2632 * past. prev == current is still correct but we need to recalculate this_rq
2633 * because prev may have moved to another CPU.
2635 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2636 __releases(rq
->lock
)
2638 struct rq
*rq
= this_rq();
2639 struct mm_struct
*mm
= rq
->prev_mm
;
2643 * The previous task will have left us with a preempt_count of 2
2644 * because it left us after:
2647 * preempt_disable(); // 1
2649 * raw_spin_lock_irq(&rq->lock) // 2
2651 * Also, see FORK_PREEMPT_COUNT.
2653 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2654 "corrupted preempt_count: %s/%d/0x%x\n",
2655 current
->comm
, current
->pid
, preempt_count()))
2656 preempt_count_set(FORK_PREEMPT_COUNT
);
2661 * A task struct has one reference for the use as "current".
2662 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2663 * schedule one last time. The schedule call will never return, and
2664 * the scheduled task must drop that reference.
2666 * We must observe prev->state before clearing prev->on_cpu (in
2667 * finish_task), otherwise a concurrent wakeup can get prev
2668 * running on another CPU and we could rave with its RUNNING -> DEAD
2669 * transition, resulting in a double drop.
2671 prev_state
= prev
->state
;
2672 vtime_task_switch(prev
);
2673 perf_event_task_sched_in(prev
, current
);
2675 finish_lock_switch(rq
);
2676 finish_arch_post_lock_switch();
2677 kcov_finish_switch(current
);
2679 fire_sched_in_preempt_notifiers(current
);
2681 * When switching through a kernel thread, the loop in
2682 * membarrier_{private,global}_expedited() may have observed that
2683 * kernel thread and not issued an IPI. It is therefore possible to
2684 * schedule between user->kernel->user threads without passing though
2685 * switch_mm(). Membarrier requires a barrier after storing to
2686 * rq->curr, before returning to userspace, so provide them here:
2688 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2689 * provided by mmdrop(),
2690 * - a sync_core for SYNC_CORE.
2693 membarrier_mm_sync_core_before_usermode(mm
);
2696 if (unlikely(prev_state
== TASK_DEAD
)) {
2697 if (prev
->sched_class
->task_dead
)
2698 prev
->sched_class
->task_dead(prev
);
2701 * Remove function-return probe instances associated with this
2702 * task and put them back on the free list.
2704 kprobe_flush_task(prev
);
2706 /* Task is done with its stack. */
2707 put_task_stack(prev
);
2709 put_task_struct(prev
);
2712 tick_nohz_task_switch();
2718 /* rq->lock is NOT held, but preemption is disabled */
2719 static void __balance_callback(struct rq
*rq
)
2721 struct callback_head
*head
, *next
;
2722 void (*func
)(struct rq
*rq
);
2723 unsigned long flags
;
2725 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2726 head
= rq
->balance_callback
;
2727 rq
->balance_callback
= NULL
;
2729 func
= (void (*)(struct rq
*))head
->func
;
2736 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2739 static inline void balance_callback(struct rq
*rq
)
2741 if (unlikely(rq
->balance_callback
))
2742 __balance_callback(rq
);
2747 static inline void balance_callback(struct rq
*rq
)
2754 * schedule_tail - first thing a freshly forked thread must call.
2755 * @prev: the thread we just switched away from.
2757 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2758 __releases(rq
->lock
)
2763 * New tasks start with FORK_PREEMPT_COUNT, see there and
2764 * finish_task_switch() for details.
2766 * finish_task_switch() will drop rq->lock() and lower preempt_count
2767 * and the preempt_enable() will end up enabling preemption (on
2768 * PREEMPT_COUNT kernels).
2771 rq
= finish_task_switch(prev
);
2772 balance_callback(rq
);
2775 if (current
->set_child_tid
)
2776 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2778 calculate_sigpending();
2782 * context_switch - switch to the new MM and the new thread's register state.
2784 static __always_inline
struct rq
*
2785 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2786 struct task_struct
*next
, struct rq_flags
*rf
)
2788 struct mm_struct
*mm
, *oldmm
;
2790 prepare_task_switch(rq
, prev
, next
);
2793 oldmm
= prev
->active_mm
;
2795 * For paravirt, this is coupled with an exit in switch_to to
2796 * combine the page table reload and the switch backend into
2799 arch_start_context_switch(prev
);
2802 * If mm is non-NULL, we pass through switch_mm(). If mm is
2803 * NULL, we will pass through mmdrop() in finish_task_switch().
2804 * Both of these contain the full memory barrier required by
2805 * membarrier after storing to rq->curr, before returning to
2809 next
->active_mm
= oldmm
;
2811 enter_lazy_tlb(oldmm
, next
);
2813 switch_mm_irqs_off(oldmm
, mm
, next
);
2816 prev
->active_mm
= NULL
;
2817 rq
->prev_mm
= oldmm
;
2820 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
2822 prepare_lock_switch(rq
, next
, rf
);
2824 /* Here we just switch the register state and the stack. */
2825 switch_to(prev
, next
, prev
);
2828 return finish_task_switch(prev
);
2832 * nr_running and nr_context_switches:
2834 * externally visible scheduler statistics: current number of runnable
2835 * threads, total number of context switches performed since bootup.
2837 unsigned long nr_running(void)
2839 unsigned long i
, sum
= 0;
2841 for_each_online_cpu(i
)
2842 sum
+= cpu_rq(i
)->nr_running
;
2848 * Check if only the current task is running on the CPU.
2850 * Caution: this function does not check that the caller has disabled
2851 * preemption, thus the result might have a time-of-check-to-time-of-use
2852 * race. The caller is responsible to use it correctly, for example:
2854 * - from a non-preemptable section (of course)
2856 * - from a thread that is bound to a single CPU
2858 * - in a loop with very short iterations (e.g. a polling loop)
2860 bool single_task_running(void)
2862 return raw_rq()->nr_running
== 1;
2864 EXPORT_SYMBOL(single_task_running
);
2866 unsigned long long nr_context_switches(void)
2869 unsigned long long sum
= 0;
2871 for_each_possible_cpu(i
)
2872 sum
+= cpu_rq(i
)->nr_switches
;
2878 * IO-wait accounting, and how its mostly bollocks (on SMP).
2880 * The idea behind IO-wait account is to account the idle time that we could
2881 * have spend running if it were not for IO. That is, if we were to improve the
2882 * storage performance, we'd have a proportional reduction in IO-wait time.
2884 * This all works nicely on UP, where, when a task blocks on IO, we account
2885 * idle time as IO-wait, because if the storage were faster, it could've been
2886 * running and we'd not be idle.
2888 * This has been extended to SMP, by doing the same for each CPU. This however
2891 * Imagine for instance the case where two tasks block on one CPU, only the one
2892 * CPU will have IO-wait accounted, while the other has regular idle. Even
2893 * though, if the storage were faster, both could've ran at the same time,
2894 * utilising both CPUs.
2896 * This means, that when looking globally, the current IO-wait accounting on
2897 * SMP is a lower bound, by reason of under accounting.
2899 * Worse, since the numbers are provided per CPU, they are sometimes
2900 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2901 * associated with any one particular CPU, it can wake to another CPU than it
2902 * blocked on. This means the per CPU IO-wait number is meaningless.
2904 * Task CPU affinities can make all that even more 'interesting'.
2907 unsigned long nr_iowait(void)
2909 unsigned long i
, sum
= 0;
2911 for_each_possible_cpu(i
)
2912 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2918 * Consumers of these two interfaces, like for example the cpufreq menu
2919 * governor are using nonsensical data. Boosting frequency for a CPU that has
2920 * IO-wait which might not even end up running the task when it does become
2924 unsigned long nr_iowait_cpu(int cpu
)
2926 struct rq
*this = cpu_rq(cpu
);
2927 return atomic_read(&this->nr_iowait
);
2930 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2932 struct rq
*rq
= this_rq();
2933 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2934 *load
= rq
->load
.weight
;
2940 * sched_exec - execve() is a valuable balancing opportunity, because at
2941 * this point the task has the smallest effective memory and cache footprint.
2943 void sched_exec(void)
2945 struct task_struct
*p
= current
;
2946 unsigned long flags
;
2949 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2950 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2951 if (dest_cpu
== smp_processor_id())
2954 if (likely(cpu_active(dest_cpu
))) {
2955 struct migration_arg arg
= { p
, dest_cpu
};
2957 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2958 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2962 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2967 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2968 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2970 EXPORT_PER_CPU_SYMBOL(kstat
);
2971 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2974 * The function fair_sched_class.update_curr accesses the struct curr
2975 * and its field curr->exec_start; when called from task_sched_runtime(),
2976 * we observe a high rate of cache misses in practice.
2977 * Prefetching this data results in improved performance.
2979 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
2981 #ifdef CONFIG_FAIR_GROUP_SCHED
2982 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
2984 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
2987 prefetch(&curr
->exec_start
);
2991 * Return accounted runtime for the task.
2992 * In case the task is currently running, return the runtime plus current's
2993 * pending runtime that have not been accounted yet.
2995 unsigned long long task_sched_runtime(struct task_struct
*p
)
3001 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3003 * 64-bit doesn't need locks to atomically read a 64-bit value.
3004 * So we have a optimization chance when the task's delta_exec is 0.
3005 * Reading ->on_cpu is racy, but this is ok.
3007 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3008 * If we race with it entering CPU, unaccounted time is 0. This is
3009 * indistinguishable from the read occurring a few cycles earlier.
3010 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3011 * been accounted, so we're correct here as well.
3013 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3014 return p
->se
.sum_exec_runtime
;
3017 rq
= task_rq_lock(p
, &rf
);
3019 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3020 * project cycles that may never be accounted to this
3021 * thread, breaking clock_gettime().
3023 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3024 prefetch_curr_exec_start(p
);
3025 update_rq_clock(rq
);
3026 p
->sched_class
->update_curr(rq
);
3028 ns
= p
->se
.sum_exec_runtime
;
3029 task_rq_unlock(rq
, p
, &rf
);
3035 * This function gets called by the timer code, with HZ frequency.
3036 * We call it with interrupts disabled.
3038 void scheduler_tick(void)
3040 int cpu
= smp_processor_id();
3041 struct rq
*rq
= cpu_rq(cpu
);
3042 struct task_struct
*curr
= rq
->curr
;
3049 update_rq_clock(rq
);
3050 curr
->sched_class
->task_tick(rq
, curr
, 0);
3051 cpu_load_update_active(rq
);
3052 calc_global_load_tick(rq
);
3056 perf_event_task_tick();
3059 rq
->idle_balance
= idle_cpu(cpu
);
3060 trigger_load_balance(rq
);
3064 #ifdef CONFIG_NO_HZ_FULL
3068 struct delayed_work work
;
3071 static struct tick_work __percpu
*tick_work_cpu
;
3073 static void sched_tick_remote(struct work_struct
*work
)
3075 struct delayed_work
*dwork
= to_delayed_work(work
);
3076 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3077 int cpu
= twork
->cpu
;
3078 struct rq
*rq
= cpu_rq(cpu
);
3079 struct task_struct
*curr
;
3084 * Handle the tick only if it appears the remote CPU is running in full
3085 * dynticks mode. The check is racy by nature, but missing a tick or
3086 * having one too much is no big deal because the scheduler tick updates
3087 * statistics and checks timeslices in a time-independent way, regardless
3088 * of when exactly it is running.
3090 if (idle_cpu(cpu
) || !tick_nohz_tick_stopped_cpu(cpu
))
3093 rq_lock_irq(rq
, &rf
);
3095 if (is_idle_task(curr
))
3098 update_rq_clock(rq
);
3099 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3102 * Make sure the next tick runs within a reasonable
3105 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3106 curr
->sched_class
->task_tick(rq
, curr
, 0);
3109 rq_unlock_irq(rq
, &rf
);
3113 * Run the remote tick once per second (1Hz). This arbitrary
3114 * frequency is large enough to avoid overload but short enough
3115 * to keep scheduler internal stats reasonably up to date.
3117 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3120 static void sched_tick_start(int cpu
)
3122 struct tick_work
*twork
;
3124 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3127 WARN_ON_ONCE(!tick_work_cpu
);
3129 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3131 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3132 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3135 #ifdef CONFIG_HOTPLUG_CPU
3136 static void sched_tick_stop(int cpu
)
3138 struct tick_work
*twork
;
3140 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3143 WARN_ON_ONCE(!tick_work_cpu
);
3145 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3146 cancel_delayed_work_sync(&twork
->work
);
3148 #endif /* CONFIG_HOTPLUG_CPU */
3150 int __init
sched_tick_offload_init(void)
3152 tick_work_cpu
= alloc_percpu(struct tick_work
);
3153 BUG_ON(!tick_work_cpu
);
3158 #else /* !CONFIG_NO_HZ_FULL */
3159 static inline void sched_tick_start(int cpu
) { }
3160 static inline void sched_tick_stop(int cpu
) { }
3163 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3164 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3166 * If the value passed in is equal to the current preempt count
3167 * then we just disabled preemption. Start timing the latency.
3169 static inline void preempt_latency_start(int val
)
3171 if (preempt_count() == val
) {
3172 unsigned long ip
= get_lock_parent_ip();
3173 #ifdef CONFIG_DEBUG_PREEMPT
3174 current
->preempt_disable_ip
= ip
;
3176 trace_preempt_off(CALLER_ADDR0
, ip
);
3180 void preempt_count_add(int val
)
3182 #ifdef CONFIG_DEBUG_PREEMPT
3186 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3189 __preempt_count_add(val
);
3190 #ifdef CONFIG_DEBUG_PREEMPT
3192 * Spinlock count overflowing soon?
3194 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3197 preempt_latency_start(val
);
3199 EXPORT_SYMBOL(preempt_count_add
);
3200 NOKPROBE_SYMBOL(preempt_count_add
);
3203 * If the value passed in equals to the current preempt count
3204 * then we just enabled preemption. Stop timing the latency.
3206 static inline void preempt_latency_stop(int val
)
3208 if (preempt_count() == val
)
3209 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3212 void preempt_count_sub(int val
)
3214 #ifdef CONFIG_DEBUG_PREEMPT
3218 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3221 * Is the spinlock portion underflowing?
3223 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3224 !(preempt_count() & PREEMPT_MASK
)))
3228 preempt_latency_stop(val
);
3229 __preempt_count_sub(val
);
3231 EXPORT_SYMBOL(preempt_count_sub
);
3232 NOKPROBE_SYMBOL(preempt_count_sub
);
3235 static inline void preempt_latency_start(int val
) { }
3236 static inline void preempt_latency_stop(int val
) { }
3239 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3241 #ifdef CONFIG_DEBUG_PREEMPT
3242 return p
->preempt_disable_ip
;
3249 * Print scheduling while atomic bug:
3251 static noinline
void __schedule_bug(struct task_struct
*prev
)
3253 /* Save this before calling printk(), since that will clobber it */
3254 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3256 if (oops_in_progress
)
3259 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3260 prev
->comm
, prev
->pid
, preempt_count());
3262 debug_show_held_locks(prev
);
3264 if (irqs_disabled())
3265 print_irqtrace_events(prev
);
3266 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3267 && in_atomic_preempt_off()) {
3268 pr_err("Preemption disabled at:");
3269 print_ip_sym(preempt_disable_ip
);
3273 panic("scheduling while atomic\n");
3276 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3280 * Various schedule()-time debugging checks and statistics:
3282 static inline void schedule_debug(struct task_struct
*prev
)
3284 #ifdef CONFIG_SCHED_STACK_END_CHECK
3285 if (task_stack_end_corrupted(prev
))
3286 panic("corrupted stack end detected inside scheduler\n");
3289 if (unlikely(in_atomic_preempt_off())) {
3290 __schedule_bug(prev
);
3291 preempt_count_set(PREEMPT_DISABLED
);
3295 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3297 schedstat_inc(this_rq()->sched_count
);
3301 * Pick up the highest-prio task:
3303 static inline struct task_struct
*
3304 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3306 const struct sched_class
*class;
3307 struct task_struct
*p
;
3310 * Optimization: we know that if all tasks are in the fair class we can
3311 * call that function directly, but only if the @prev task wasn't of a
3312 * higher scheduling class, because otherwise those loose the
3313 * opportunity to pull in more work from other CPUs.
3315 if (likely((prev
->sched_class
== &idle_sched_class
||
3316 prev
->sched_class
== &fair_sched_class
) &&
3317 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3319 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3320 if (unlikely(p
== RETRY_TASK
))
3323 /* Assumes fair_sched_class->next == idle_sched_class */
3325 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3331 for_each_class(class) {
3332 p
= class->pick_next_task(rq
, prev
, rf
);
3334 if (unlikely(p
== RETRY_TASK
))
3340 /* The idle class should always have a runnable task: */
3345 * __schedule() is the main scheduler function.
3347 * The main means of driving the scheduler and thus entering this function are:
3349 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3351 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3352 * paths. For example, see arch/x86/entry_64.S.
3354 * To drive preemption between tasks, the scheduler sets the flag in timer
3355 * interrupt handler scheduler_tick().
3357 * 3. Wakeups don't really cause entry into schedule(). They add a
3358 * task to the run-queue and that's it.
3360 * Now, if the new task added to the run-queue preempts the current
3361 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3362 * called on the nearest possible occasion:
3364 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3366 * - in syscall or exception context, at the next outmost
3367 * preempt_enable(). (this might be as soon as the wake_up()'s
3370 * - in IRQ context, return from interrupt-handler to
3371 * preemptible context
3373 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3376 * - cond_resched() call
3377 * - explicit schedule() call
3378 * - return from syscall or exception to user-space
3379 * - return from interrupt-handler to user-space
3381 * WARNING: must be called with preemption disabled!
3383 static void __sched notrace
__schedule(bool preempt
)
3385 struct task_struct
*prev
, *next
;
3386 unsigned long *switch_count
;
3391 cpu
= smp_processor_id();
3395 schedule_debug(prev
);
3397 if (sched_feat(HRTICK
))
3400 local_irq_disable();
3401 rcu_note_context_switch(preempt
);
3404 * Make sure that signal_pending_state()->signal_pending() below
3405 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3406 * done by the caller to avoid the race with signal_wake_up().
3408 * The membarrier system call requires a full memory barrier
3409 * after coming from user-space, before storing to rq->curr.
3412 smp_mb__after_spinlock();
3414 /* Promote REQ to ACT */
3415 rq
->clock_update_flags
<<= 1;
3416 update_rq_clock(rq
);
3418 switch_count
= &prev
->nivcsw
;
3419 if (!preempt
&& prev
->state
) {
3420 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3421 prev
->state
= TASK_RUNNING
;
3423 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
3426 if (prev
->in_iowait
) {
3427 atomic_inc(&rq
->nr_iowait
);
3428 delayacct_blkio_start();
3432 * If a worker went to sleep, notify and ask workqueue
3433 * whether it wants to wake up a task to maintain
3436 if (prev
->flags
& PF_WQ_WORKER
) {
3437 struct task_struct
*to_wakeup
;
3439 to_wakeup
= wq_worker_sleeping(prev
);
3441 try_to_wake_up_local(to_wakeup
, &rf
);
3444 switch_count
= &prev
->nvcsw
;
3447 next
= pick_next_task(rq
, prev
, &rf
);
3448 clear_tsk_need_resched(prev
);
3449 clear_preempt_need_resched();
3451 if (likely(prev
!= next
)) {
3455 * The membarrier system call requires each architecture
3456 * to have a full memory barrier after updating
3457 * rq->curr, before returning to user-space.
3459 * Here are the schemes providing that barrier on the
3460 * various architectures:
3461 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3462 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3463 * - finish_lock_switch() for weakly-ordered
3464 * architectures where spin_unlock is a full barrier,
3465 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3466 * is a RELEASE barrier),
3470 trace_sched_switch(preempt
, prev
, next
);
3472 /* Also unlocks the rq: */
3473 rq
= context_switch(rq
, prev
, next
, &rf
);
3475 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3476 rq_unlock_irq(rq
, &rf
);
3479 balance_callback(rq
);
3482 void __noreturn
do_task_dead(void)
3484 /* Causes final put_task_struct in finish_task_switch(): */
3485 set_special_state(TASK_DEAD
);
3487 /* Tell freezer to ignore us: */
3488 current
->flags
|= PF_NOFREEZE
;
3493 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3498 static inline void sched_submit_work(struct task_struct
*tsk
)
3500 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3503 * If we are going to sleep and we have plugged IO queued,
3504 * make sure to submit it to avoid deadlocks.
3506 if (blk_needs_flush_plug(tsk
))
3507 blk_schedule_flush_plug(tsk
);
3510 asmlinkage __visible
void __sched
schedule(void)
3512 struct task_struct
*tsk
= current
;
3514 sched_submit_work(tsk
);
3518 sched_preempt_enable_no_resched();
3519 } while (need_resched());
3521 EXPORT_SYMBOL(schedule
);
3524 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3525 * state (have scheduled out non-voluntarily) by making sure that all
3526 * tasks have either left the run queue or have gone into user space.
3527 * As idle tasks do not do either, they must not ever be preempted
3528 * (schedule out non-voluntarily).
3530 * schedule_idle() is similar to schedule_preempt_disable() except that it
3531 * never enables preemption because it does not call sched_submit_work().
3533 void __sched
schedule_idle(void)
3536 * As this skips calling sched_submit_work(), which the idle task does
3537 * regardless because that function is a nop when the task is in a
3538 * TASK_RUNNING state, make sure this isn't used someplace that the
3539 * current task can be in any other state. Note, idle is always in the
3540 * TASK_RUNNING state.
3542 WARN_ON_ONCE(current
->state
);
3545 } while (need_resched());
3548 #ifdef CONFIG_CONTEXT_TRACKING
3549 asmlinkage __visible
void __sched
schedule_user(void)
3552 * If we come here after a random call to set_need_resched(),
3553 * or we have been woken up remotely but the IPI has not yet arrived,
3554 * we haven't yet exited the RCU idle mode. Do it here manually until
3555 * we find a better solution.
3557 * NB: There are buggy callers of this function. Ideally we
3558 * should warn if prev_state != CONTEXT_USER, but that will trigger
3559 * too frequently to make sense yet.
3561 enum ctx_state prev_state
= exception_enter();
3563 exception_exit(prev_state
);
3568 * schedule_preempt_disabled - called with preemption disabled
3570 * Returns with preemption disabled. Note: preempt_count must be 1
3572 void __sched
schedule_preempt_disabled(void)
3574 sched_preempt_enable_no_resched();
3579 static void __sched notrace
preempt_schedule_common(void)
3583 * Because the function tracer can trace preempt_count_sub()
3584 * and it also uses preempt_enable/disable_notrace(), if
3585 * NEED_RESCHED is set, the preempt_enable_notrace() called
3586 * by the function tracer will call this function again and
3587 * cause infinite recursion.
3589 * Preemption must be disabled here before the function
3590 * tracer can trace. Break up preempt_disable() into two
3591 * calls. One to disable preemption without fear of being
3592 * traced. The other to still record the preemption latency,
3593 * which can also be traced by the function tracer.
3595 preempt_disable_notrace();
3596 preempt_latency_start(1);
3598 preempt_latency_stop(1);
3599 preempt_enable_no_resched_notrace();
3602 * Check again in case we missed a preemption opportunity
3603 * between schedule and now.
3605 } while (need_resched());
3608 #ifdef CONFIG_PREEMPT
3610 * this is the entry point to schedule() from in-kernel preemption
3611 * off of preempt_enable. Kernel preemptions off return from interrupt
3612 * occur there and call schedule directly.
3614 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3617 * If there is a non-zero preempt_count or interrupts are disabled,
3618 * we do not want to preempt the current task. Just return..
3620 if (likely(!preemptible()))
3623 preempt_schedule_common();
3625 NOKPROBE_SYMBOL(preempt_schedule
);
3626 EXPORT_SYMBOL(preempt_schedule
);
3629 * preempt_schedule_notrace - preempt_schedule called by tracing
3631 * The tracing infrastructure uses preempt_enable_notrace to prevent
3632 * recursion and tracing preempt enabling caused by the tracing
3633 * infrastructure itself. But as tracing can happen in areas coming
3634 * from userspace or just about to enter userspace, a preempt enable
3635 * can occur before user_exit() is called. This will cause the scheduler
3636 * to be called when the system is still in usermode.
3638 * To prevent this, the preempt_enable_notrace will use this function
3639 * instead of preempt_schedule() to exit user context if needed before
3640 * calling the scheduler.
3642 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3644 enum ctx_state prev_ctx
;
3646 if (likely(!preemptible()))
3651 * Because the function tracer can trace preempt_count_sub()
3652 * and it also uses preempt_enable/disable_notrace(), if
3653 * NEED_RESCHED is set, the preempt_enable_notrace() called
3654 * by the function tracer will call this function again and
3655 * cause infinite recursion.
3657 * Preemption must be disabled here before the function
3658 * tracer can trace. Break up preempt_disable() into two
3659 * calls. One to disable preemption without fear of being
3660 * traced. The other to still record the preemption latency,
3661 * which can also be traced by the function tracer.
3663 preempt_disable_notrace();
3664 preempt_latency_start(1);
3666 * Needs preempt disabled in case user_exit() is traced
3667 * and the tracer calls preempt_enable_notrace() causing
3668 * an infinite recursion.
3670 prev_ctx
= exception_enter();
3672 exception_exit(prev_ctx
);
3674 preempt_latency_stop(1);
3675 preempt_enable_no_resched_notrace();
3676 } while (need_resched());
3678 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3680 #endif /* CONFIG_PREEMPT */
3683 * this is the entry point to schedule() from kernel preemption
3684 * off of irq context.
3685 * Note, that this is called and return with irqs disabled. This will
3686 * protect us against recursive calling from irq.
3688 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3690 enum ctx_state prev_state
;
3692 /* Catch callers which need to be fixed */
3693 BUG_ON(preempt_count() || !irqs_disabled());
3695 prev_state
= exception_enter();
3701 local_irq_disable();
3702 sched_preempt_enable_no_resched();
3703 } while (need_resched());
3705 exception_exit(prev_state
);
3708 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
3711 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3713 EXPORT_SYMBOL(default_wake_function
);
3715 #ifdef CONFIG_RT_MUTEXES
3717 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
3720 prio
= min(prio
, pi_task
->prio
);
3725 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3727 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3729 return __rt_effective_prio(pi_task
, prio
);
3733 * rt_mutex_setprio - set the current priority of a task
3735 * @pi_task: donor task
3737 * This function changes the 'effective' priority of a task. It does
3738 * not touch ->normal_prio like __setscheduler().
3740 * Used by the rt_mutex code to implement priority inheritance
3741 * logic. Call site only calls if the priority of the task changed.
3743 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
3745 int prio
, oldprio
, queued
, running
, queue_flag
=
3746 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
3747 const struct sched_class
*prev_class
;
3751 /* XXX used to be waiter->prio, not waiter->task->prio */
3752 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
3755 * If nothing changed; bail early.
3757 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
3760 rq
= __task_rq_lock(p
, &rf
);
3761 update_rq_clock(rq
);
3763 * Set under pi_lock && rq->lock, such that the value can be used under
3766 * Note that there is loads of tricky to make this pointer cache work
3767 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3768 * ensure a task is de-boosted (pi_task is set to NULL) before the
3769 * task is allowed to run again (and can exit). This ensures the pointer
3770 * points to a blocked task -- which guaratees the task is present.
3772 p
->pi_top_task
= pi_task
;
3775 * For FIFO/RR we only need to set prio, if that matches we're done.
3777 if (prio
== p
->prio
&& !dl_prio(prio
))
3781 * Idle task boosting is a nono in general. There is one
3782 * exception, when PREEMPT_RT and NOHZ is active:
3784 * The idle task calls get_next_timer_interrupt() and holds
3785 * the timer wheel base->lock on the CPU and another CPU wants
3786 * to access the timer (probably to cancel it). We can safely
3787 * ignore the boosting request, as the idle CPU runs this code
3788 * with interrupts disabled and will complete the lock
3789 * protected section without being interrupted. So there is no
3790 * real need to boost.
3792 if (unlikely(p
== rq
->idle
)) {
3793 WARN_ON(p
!= rq
->curr
);
3794 WARN_ON(p
->pi_blocked_on
);
3798 trace_sched_pi_setprio(p
, pi_task
);
3801 if (oldprio
== prio
)
3802 queue_flag
&= ~DEQUEUE_MOVE
;
3804 prev_class
= p
->sched_class
;
3805 queued
= task_on_rq_queued(p
);
3806 running
= task_current(rq
, p
);
3808 dequeue_task(rq
, p
, queue_flag
);
3810 put_prev_task(rq
, p
);
3813 * Boosting condition are:
3814 * 1. -rt task is running and holds mutex A
3815 * --> -dl task blocks on mutex A
3817 * 2. -dl task is running and holds mutex A
3818 * --> -dl task blocks on mutex A and could preempt the
3821 if (dl_prio(prio
)) {
3822 if (!dl_prio(p
->normal_prio
) ||
3823 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3824 p
->dl
.dl_boosted
= 1;
3825 queue_flag
|= ENQUEUE_REPLENISH
;
3827 p
->dl
.dl_boosted
= 0;
3828 p
->sched_class
= &dl_sched_class
;
3829 } else if (rt_prio(prio
)) {
3830 if (dl_prio(oldprio
))
3831 p
->dl
.dl_boosted
= 0;
3833 queue_flag
|= ENQUEUE_HEAD
;
3834 p
->sched_class
= &rt_sched_class
;
3836 if (dl_prio(oldprio
))
3837 p
->dl
.dl_boosted
= 0;
3838 if (rt_prio(oldprio
))
3840 p
->sched_class
= &fair_sched_class
;
3846 enqueue_task(rq
, p
, queue_flag
);
3848 set_curr_task(rq
, p
);
3850 check_class_changed(rq
, p
, prev_class
, oldprio
);
3852 /* Avoid rq from going away on us: */
3854 __task_rq_unlock(rq
, &rf
);
3856 balance_callback(rq
);
3860 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3866 void set_user_nice(struct task_struct
*p
, long nice
)
3868 bool queued
, running
;
3869 int old_prio
, delta
;
3873 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3876 * We have to be careful, if called from sys_setpriority(),
3877 * the task might be in the middle of scheduling on another CPU.
3879 rq
= task_rq_lock(p
, &rf
);
3880 update_rq_clock(rq
);
3883 * The RT priorities are set via sched_setscheduler(), but we still
3884 * allow the 'normal' nice value to be set - but as expected
3885 * it wont have any effect on scheduling until the task is
3886 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3888 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3889 p
->static_prio
= NICE_TO_PRIO(nice
);
3892 queued
= task_on_rq_queued(p
);
3893 running
= task_current(rq
, p
);
3895 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
3897 put_prev_task(rq
, p
);
3899 p
->static_prio
= NICE_TO_PRIO(nice
);
3900 set_load_weight(p
, true);
3902 p
->prio
= effective_prio(p
);
3903 delta
= p
->prio
- old_prio
;
3906 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
3908 * If the task increased its priority or is running and
3909 * lowered its priority, then reschedule its CPU:
3911 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3915 set_curr_task(rq
, p
);
3917 task_rq_unlock(rq
, p
, &rf
);
3919 EXPORT_SYMBOL(set_user_nice
);
3922 * can_nice - check if a task can reduce its nice value
3926 int can_nice(const struct task_struct
*p
, const int nice
)
3928 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3929 int nice_rlim
= nice_to_rlimit(nice
);
3931 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3932 capable(CAP_SYS_NICE
));
3935 #ifdef __ARCH_WANT_SYS_NICE
3938 * sys_nice - change the priority of the current process.
3939 * @increment: priority increment
3941 * sys_setpriority is a more generic, but much slower function that
3942 * does similar things.
3944 SYSCALL_DEFINE1(nice
, int, increment
)
3949 * Setpriority might change our priority at the same moment.
3950 * We don't have to worry. Conceptually one call occurs first
3951 * and we have a single winner.
3953 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3954 nice
= task_nice(current
) + increment
;
3956 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3957 if (increment
< 0 && !can_nice(current
, nice
))
3960 retval
= security_task_setnice(current
, nice
);
3964 set_user_nice(current
, nice
);
3971 * task_prio - return the priority value of a given task.
3972 * @p: the task in question.
3974 * Return: The priority value as seen by users in /proc.
3975 * RT tasks are offset by -200. Normal tasks are centered
3976 * around 0, value goes from -16 to +15.
3978 int task_prio(const struct task_struct
*p
)
3980 return p
->prio
- MAX_RT_PRIO
;
3984 * idle_cpu - is a given CPU idle currently?
3985 * @cpu: the processor in question.
3987 * Return: 1 if the CPU is currently idle. 0 otherwise.
3989 int idle_cpu(int cpu
)
3991 struct rq
*rq
= cpu_rq(cpu
);
3993 if (rq
->curr
!= rq
->idle
)
4000 if (!llist_empty(&rq
->wake_list
))
4008 * available_idle_cpu - is a given CPU idle for enqueuing work.
4009 * @cpu: the CPU in question.
4011 * Return: 1 if the CPU is currently idle. 0 otherwise.
4013 int available_idle_cpu(int cpu
)
4018 if (vcpu_is_preempted(cpu
))
4025 * idle_task - return the idle task for a given CPU.
4026 * @cpu: the processor in question.
4028 * Return: The idle task for the CPU @cpu.
4030 struct task_struct
*idle_task(int cpu
)
4032 return cpu_rq(cpu
)->idle
;
4036 * find_process_by_pid - find a process with a matching PID value.
4037 * @pid: the pid in question.
4039 * The task of @pid, if found. %NULL otherwise.
4041 static struct task_struct
*find_process_by_pid(pid_t pid
)
4043 return pid
? find_task_by_vpid(pid
) : current
;
4047 * sched_setparam() passes in -1 for its policy, to let the functions
4048 * it calls know not to change it.
4050 #define SETPARAM_POLICY -1
4052 static void __setscheduler_params(struct task_struct
*p
,
4053 const struct sched_attr
*attr
)
4055 int policy
= attr
->sched_policy
;
4057 if (policy
== SETPARAM_POLICY
)
4062 if (dl_policy(policy
))
4063 __setparam_dl(p
, attr
);
4064 else if (fair_policy(policy
))
4065 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4068 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4069 * !rt_policy. Always setting this ensures that things like
4070 * getparam()/getattr() don't report silly values for !rt tasks.
4072 p
->rt_priority
= attr
->sched_priority
;
4073 p
->normal_prio
= normal_prio(p
);
4074 set_load_weight(p
, true);
4077 /* Actually do priority change: must hold pi & rq lock. */
4078 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4079 const struct sched_attr
*attr
, bool keep_boost
)
4081 __setscheduler_params(p
, attr
);
4084 * Keep a potential priority boosting if called from
4085 * sched_setscheduler().
4087 p
->prio
= normal_prio(p
);
4089 p
->prio
= rt_effective_prio(p
, p
->prio
);
4091 if (dl_prio(p
->prio
))
4092 p
->sched_class
= &dl_sched_class
;
4093 else if (rt_prio(p
->prio
))
4094 p
->sched_class
= &rt_sched_class
;
4096 p
->sched_class
= &fair_sched_class
;
4100 * Check the target process has a UID that matches the current process's:
4102 static bool check_same_owner(struct task_struct
*p
)
4104 const struct cred
*cred
= current_cred(), *pcred
;
4108 pcred
= __task_cred(p
);
4109 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4110 uid_eq(cred
->euid
, pcred
->uid
));
4115 static int __sched_setscheduler(struct task_struct
*p
,
4116 const struct sched_attr
*attr
,
4119 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4120 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4121 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4122 int new_effective_prio
, policy
= attr
->sched_policy
;
4123 const struct sched_class
*prev_class
;
4126 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4129 /* The pi code expects interrupts enabled */
4130 BUG_ON(pi
&& in_interrupt());
4132 /* Double check policy once rq lock held: */
4134 reset_on_fork
= p
->sched_reset_on_fork
;
4135 policy
= oldpolicy
= p
->policy
;
4137 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4139 if (!valid_policy(policy
))
4143 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4147 * Valid priorities for SCHED_FIFO and SCHED_RR are
4148 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4149 * SCHED_BATCH and SCHED_IDLE is 0.
4151 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4152 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4154 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4155 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4159 * Allow unprivileged RT tasks to decrease priority:
4161 if (user
&& !capable(CAP_SYS_NICE
)) {
4162 if (fair_policy(policy
)) {
4163 if (attr
->sched_nice
< task_nice(p
) &&
4164 !can_nice(p
, attr
->sched_nice
))
4168 if (rt_policy(policy
)) {
4169 unsigned long rlim_rtprio
=
4170 task_rlimit(p
, RLIMIT_RTPRIO
);
4172 /* Can't set/change the rt policy: */
4173 if (policy
!= p
->policy
&& !rlim_rtprio
)
4176 /* Can't increase priority: */
4177 if (attr
->sched_priority
> p
->rt_priority
&&
4178 attr
->sched_priority
> rlim_rtprio
)
4183 * Can't set/change SCHED_DEADLINE policy at all for now
4184 * (safest behavior); in the future we would like to allow
4185 * unprivileged DL tasks to increase their relative deadline
4186 * or reduce their runtime (both ways reducing utilization)
4188 if (dl_policy(policy
))
4192 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4193 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4195 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4196 if (!can_nice(p
, task_nice(p
)))
4200 /* Can't change other user's priorities: */
4201 if (!check_same_owner(p
))
4204 /* Normal users shall not reset the sched_reset_on_fork flag: */
4205 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4210 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4213 retval
= security_task_setscheduler(p
);
4219 * Make sure no PI-waiters arrive (or leave) while we are
4220 * changing the priority of the task:
4222 * To be able to change p->policy safely, the appropriate
4223 * runqueue lock must be held.
4225 rq
= task_rq_lock(p
, &rf
);
4226 update_rq_clock(rq
);
4229 * Changing the policy of the stop threads its a very bad idea:
4231 if (p
== rq
->stop
) {
4232 task_rq_unlock(rq
, p
, &rf
);
4237 * If not changing anything there's no need to proceed further,
4238 * but store a possible modification of reset_on_fork.
4240 if (unlikely(policy
== p
->policy
)) {
4241 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4243 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4245 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4248 p
->sched_reset_on_fork
= reset_on_fork
;
4249 task_rq_unlock(rq
, p
, &rf
);
4255 #ifdef CONFIG_RT_GROUP_SCHED
4257 * Do not allow realtime tasks into groups that have no runtime
4260 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4261 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4262 !task_group_is_autogroup(task_group(p
))) {
4263 task_rq_unlock(rq
, p
, &rf
);
4268 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
4269 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
4270 cpumask_t
*span
= rq
->rd
->span
;
4273 * Don't allow tasks with an affinity mask smaller than
4274 * the entire root_domain to become SCHED_DEADLINE. We
4275 * will also fail if there's no bandwidth available.
4277 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4278 rq
->rd
->dl_bw
.bw
== 0) {
4279 task_rq_unlock(rq
, p
, &rf
);
4286 /* Re-check policy now with rq lock held: */
4287 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4288 policy
= oldpolicy
= -1;
4289 task_rq_unlock(rq
, p
, &rf
);
4294 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4295 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4298 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
4299 task_rq_unlock(rq
, p
, &rf
);
4303 p
->sched_reset_on_fork
= reset_on_fork
;
4308 * Take priority boosted tasks into account. If the new
4309 * effective priority is unchanged, we just store the new
4310 * normal parameters and do not touch the scheduler class and
4311 * the runqueue. This will be done when the task deboost
4314 new_effective_prio
= rt_effective_prio(p
, newprio
);
4315 if (new_effective_prio
== oldprio
)
4316 queue_flags
&= ~DEQUEUE_MOVE
;
4319 queued
= task_on_rq_queued(p
);
4320 running
= task_current(rq
, p
);
4322 dequeue_task(rq
, p
, queue_flags
);
4324 put_prev_task(rq
, p
);
4326 prev_class
= p
->sched_class
;
4327 __setscheduler(rq
, p
, attr
, pi
);
4331 * We enqueue to tail when the priority of a task is
4332 * increased (user space view).
4334 if (oldprio
< p
->prio
)
4335 queue_flags
|= ENQUEUE_HEAD
;
4337 enqueue_task(rq
, p
, queue_flags
);
4340 set_curr_task(rq
, p
);
4342 check_class_changed(rq
, p
, prev_class
, oldprio
);
4344 /* Avoid rq from going away on us: */
4346 task_rq_unlock(rq
, p
, &rf
);
4349 rt_mutex_adjust_pi(p
);
4351 /* Run balance callbacks after we've adjusted the PI chain: */
4352 balance_callback(rq
);
4358 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4359 const struct sched_param
*param
, bool check
)
4361 struct sched_attr attr
= {
4362 .sched_policy
= policy
,
4363 .sched_priority
= param
->sched_priority
,
4364 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4367 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4368 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4369 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4370 policy
&= ~SCHED_RESET_ON_FORK
;
4371 attr
.sched_policy
= policy
;
4374 return __sched_setscheduler(p
, &attr
, check
, true);
4377 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4378 * @p: the task in question.
4379 * @policy: new policy.
4380 * @param: structure containing the new RT priority.
4382 * Return: 0 on success. An error code otherwise.
4384 * NOTE that the task may be already dead.
4386 int sched_setscheduler(struct task_struct
*p
, int policy
,
4387 const struct sched_param
*param
)
4389 return _sched_setscheduler(p
, policy
, param
, true);
4391 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4393 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4395 return __sched_setscheduler(p
, attr
, true, true);
4397 EXPORT_SYMBOL_GPL(sched_setattr
);
4399 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
4401 return __sched_setscheduler(p
, attr
, false, true);
4405 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4406 * @p: the task in question.
4407 * @policy: new policy.
4408 * @param: structure containing the new RT priority.
4410 * Just like sched_setscheduler, only don't bother checking if the
4411 * current context has permission. For example, this is needed in
4412 * stop_machine(): we create temporary high priority worker threads,
4413 * but our caller might not have that capability.
4415 * Return: 0 on success. An error code otherwise.
4417 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4418 const struct sched_param
*param
)
4420 return _sched_setscheduler(p
, policy
, param
, false);
4422 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4425 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4427 struct sched_param lparam
;
4428 struct task_struct
*p
;
4431 if (!param
|| pid
< 0)
4433 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4438 p
= find_process_by_pid(pid
);
4440 retval
= sched_setscheduler(p
, policy
, &lparam
);
4447 * Mimics kernel/events/core.c perf_copy_attr().
4449 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
4454 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4457 /* Zero the full structure, so that a short copy will be nice: */
4458 memset(attr
, 0, sizeof(*attr
));
4460 ret
= get_user(size
, &uattr
->size
);
4464 /* Bail out on silly large: */
4465 if (size
> PAGE_SIZE
)
4468 /* ABI compatibility quirk: */
4470 size
= SCHED_ATTR_SIZE_VER0
;
4472 if (size
< SCHED_ATTR_SIZE_VER0
)
4476 * If we're handed a bigger struct than we know of,
4477 * ensure all the unknown bits are 0 - i.e. new
4478 * user-space does not rely on any kernel feature
4479 * extensions we dont know about yet.
4481 if (size
> sizeof(*attr
)) {
4482 unsigned char __user
*addr
;
4483 unsigned char __user
*end
;
4486 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4487 end
= (void __user
*)uattr
+ size
;
4489 for (; addr
< end
; addr
++) {
4490 ret
= get_user(val
, addr
);
4496 size
= sizeof(*attr
);
4499 ret
= copy_from_user(attr
, uattr
, size
);
4504 * XXX: Do we want to be lenient like existing syscalls; or do we want
4505 * to be strict and return an error on out-of-bounds values?
4507 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4512 put_user(sizeof(*attr
), &uattr
->size
);
4517 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4518 * @pid: the pid in question.
4519 * @policy: new policy.
4520 * @param: structure containing the new RT priority.
4522 * Return: 0 on success. An error code otherwise.
4524 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
4529 return do_sched_setscheduler(pid
, policy
, param
);
4533 * sys_sched_setparam - set/change the RT priority of a thread
4534 * @pid: the pid in question.
4535 * @param: structure containing the new RT priority.
4537 * Return: 0 on success. An error code otherwise.
4539 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4541 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4545 * sys_sched_setattr - same as above, but with extended sched_attr
4546 * @pid: the pid in question.
4547 * @uattr: structure containing the extended parameters.
4548 * @flags: for future extension.
4550 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4551 unsigned int, flags
)
4553 struct sched_attr attr
;
4554 struct task_struct
*p
;
4557 if (!uattr
|| pid
< 0 || flags
)
4560 retval
= sched_copy_attr(uattr
, &attr
);
4564 if ((int)attr
.sched_policy
< 0)
4569 p
= find_process_by_pid(pid
);
4571 retval
= sched_setattr(p
, &attr
);
4578 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4579 * @pid: the pid in question.
4581 * Return: On success, the policy of the thread. Otherwise, a negative error
4584 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4586 struct task_struct
*p
;
4594 p
= find_process_by_pid(pid
);
4596 retval
= security_task_getscheduler(p
);
4599 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4606 * sys_sched_getparam - get the RT priority of a thread
4607 * @pid: the pid in question.
4608 * @param: structure containing the RT priority.
4610 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4613 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4615 struct sched_param lp
= { .sched_priority
= 0 };
4616 struct task_struct
*p
;
4619 if (!param
|| pid
< 0)
4623 p
= find_process_by_pid(pid
);
4628 retval
= security_task_getscheduler(p
);
4632 if (task_has_rt_policy(p
))
4633 lp
.sched_priority
= p
->rt_priority
;
4637 * This one might sleep, we cannot do it with a spinlock held ...
4639 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4648 static int sched_read_attr(struct sched_attr __user
*uattr
,
4649 struct sched_attr
*attr
,
4654 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4658 * If we're handed a smaller struct than we know of,
4659 * ensure all the unknown bits are 0 - i.e. old
4660 * user-space does not get uncomplete information.
4662 if (usize
< sizeof(*attr
)) {
4663 unsigned char *addr
;
4666 addr
= (void *)attr
+ usize
;
4667 end
= (void *)attr
+ sizeof(*attr
);
4669 for (; addr
< end
; addr
++) {
4677 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4685 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4686 * @pid: the pid in question.
4687 * @uattr: structure containing the extended parameters.
4688 * @size: sizeof(attr) for fwd/bwd comp.
4689 * @flags: for future extension.
4691 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4692 unsigned int, size
, unsigned int, flags
)
4694 struct sched_attr attr
= {
4695 .size
= sizeof(struct sched_attr
),
4697 struct task_struct
*p
;
4700 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4701 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4705 p
= find_process_by_pid(pid
);
4710 retval
= security_task_getscheduler(p
);
4714 attr
.sched_policy
= p
->policy
;
4715 if (p
->sched_reset_on_fork
)
4716 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4717 if (task_has_dl_policy(p
))
4718 __getparam_dl(p
, &attr
);
4719 else if (task_has_rt_policy(p
))
4720 attr
.sched_priority
= p
->rt_priority
;
4722 attr
.sched_nice
= task_nice(p
);
4726 retval
= sched_read_attr(uattr
, &attr
, size
);
4734 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4736 cpumask_var_t cpus_allowed
, new_mask
;
4737 struct task_struct
*p
;
4742 p
= find_process_by_pid(pid
);
4748 /* Prevent p going away */
4752 if (p
->flags
& PF_NO_SETAFFINITY
) {
4756 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4760 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4762 goto out_free_cpus_allowed
;
4765 if (!check_same_owner(p
)) {
4767 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4769 goto out_free_new_mask
;
4774 retval
= security_task_setscheduler(p
);
4776 goto out_free_new_mask
;
4779 cpuset_cpus_allowed(p
, cpus_allowed
);
4780 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4783 * Since bandwidth control happens on root_domain basis,
4784 * if admission test is enabled, we only admit -deadline
4785 * tasks allowed to run on all the CPUs in the task's
4789 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4791 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4794 goto out_free_new_mask
;
4800 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4803 cpuset_cpus_allowed(p
, cpus_allowed
);
4804 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4806 * We must have raced with a concurrent cpuset
4807 * update. Just reset the cpus_allowed to the
4808 * cpuset's cpus_allowed
4810 cpumask_copy(new_mask
, cpus_allowed
);
4815 free_cpumask_var(new_mask
);
4816 out_free_cpus_allowed
:
4817 free_cpumask_var(cpus_allowed
);
4823 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4824 struct cpumask
*new_mask
)
4826 if (len
< cpumask_size())
4827 cpumask_clear(new_mask
);
4828 else if (len
> cpumask_size())
4829 len
= cpumask_size();
4831 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4835 * sys_sched_setaffinity - set the CPU affinity of a process
4836 * @pid: pid of the process
4837 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4838 * @user_mask_ptr: user-space pointer to the new CPU mask
4840 * Return: 0 on success. An error code otherwise.
4842 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4843 unsigned long __user
*, user_mask_ptr
)
4845 cpumask_var_t new_mask
;
4848 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4851 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4853 retval
= sched_setaffinity(pid
, new_mask
);
4854 free_cpumask_var(new_mask
);
4858 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4860 struct task_struct
*p
;
4861 unsigned long flags
;
4867 p
= find_process_by_pid(pid
);
4871 retval
= security_task_getscheduler(p
);
4875 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4876 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4877 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4886 * sys_sched_getaffinity - get the CPU affinity of a process
4887 * @pid: pid of the process
4888 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4889 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4891 * Return: size of CPU mask copied to user_mask_ptr on success. An
4892 * error code otherwise.
4894 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4895 unsigned long __user
*, user_mask_ptr
)
4900 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4902 if (len
& (sizeof(unsigned long)-1))
4905 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4908 ret
= sched_getaffinity(pid
, mask
);
4910 unsigned int retlen
= min(len
, cpumask_size());
4912 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4917 free_cpumask_var(mask
);
4923 * sys_sched_yield - yield the current processor to other threads.
4925 * This function yields the current CPU to other tasks. If there are no
4926 * other threads running on this CPU then this function will return.
4930 static void do_sched_yield(void)
4935 local_irq_disable();
4939 schedstat_inc(rq
->yld_count
);
4940 current
->sched_class
->yield_task(rq
);
4943 * Since we are going to call schedule() anyway, there's
4944 * no need to preempt or enable interrupts:
4948 sched_preempt_enable_no_resched();
4953 SYSCALL_DEFINE0(sched_yield
)
4959 #ifndef CONFIG_PREEMPT
4960 int __sched
_cond_resched(void)
4962 if (should_resched(0)) {
4963 preempt_schedule_common();
4969 EXPORT_SYMBOL(_cond_resched
);
4973 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4974 * call schedule, and on return reacquire the lock.
4976 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4977 * operations here to prevent schedule() from being called twice (once via
4978 * spin_unlock(), once by hand).
4980 int __cond_resched_lock(spinlock_t
*lock
)
4982 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4985 lockdep_assert_held(lock
);
4987 if (spin_needbreak(lock
) || resched
) {
4990 preempt_schedule_common();
4998 EXPORT_SYMBOL(__cond_resched_lock
);
5001 * yield - yield the current processor to other threads.
5003 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5005 * The scheduler is at all times free to pick the calling task as the most
5006 * eligible task to run, if removing the yield() call from your code breaks
5007 * it, its already broken.
5009 * Typical broken usage is:
5014 * where one assumes that yield() will let 'the other' process run that will
5015 * make event true. If the current task is a SCHED_FIFO task that will never
5016 * happen. Never use yield() as a progress guarantee!!
5018 * If you want to use yield() to wait for something, use wait_event().
5019 * If you want to use yield() to be 'nice' for others, use cond_resched().
5020 * If you still want to use yield(), do not!
5022 void __sched
yield(void)
5024 set_current_state(TASK_RUNNING
);
5027 EXPORT_SYMBOL(yield
);
5030 * yield_to - yield the current processor to another thread in
5031 * your thread group, or accelerate that thread toward the
5032 * processor it's on.
5034 * @preempt: whether task preemption is allowed or not
5036 * It's the caller's job to ensure that the target task struct
5037 * can't go away on us before we can do any checks.
5040 * true (>0) if we indeed boosted the target task.
5041 * false (0) if we failed to boost the target.
5042 * -ESRCH if there's no task to yield to.
5044 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5046 struct task_struct
*curr
= current
;
5047 struct rq
*rq
, *p_rq
;
5048 unsigned long flags
;
5051 local_irq_save(flags
);
5057 * If we're the only runnable task on the rq and target rq also
5058 * has only one task, there's absolutely no point in yielding.
5060 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5065 double_rq_lock(rq
, p_rq
);
5066 if (task_rq(p
) != p_rq
) {
5067 double_rq_unlock(rq
, p_rq
);
5071 if (!curr
->sched_class
->yield_to_task
)
5074 if (curr
->sched_class
!= p
->sched_class
)
5077 if (task_running(p_rq
, p
) || p
->state
)
5080 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5082 schedstat_inc(rq
->yld_count
);
5084 * Make p's CPU reschedule; pick_next_entity takes care of
5087 if (preempt
&& rq
!= p_rq
)
5092 double_rq_unlock(rq
, p_rq
);
5094 local_irq_restore(flags
);
5101 EXPORT_SYMBOL_GPL(yield_to
);
5103 int io_schedule_prepare(void)
5105 int old_iowait
= current
->in_iowait
;
5107 current
->in_iowait
= 1;
5108 blk_schedule_flush_plug(current
);
5113 void io_schedule_finish(int token
)
5115 current
->in_iowait
= token
;
5119 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5120 * that process accounting knows that this is a task in IO wait state.
5122 long __sched
io_schedule_timeout(long timeout
)
5127 token
= io_schedule_prepare();
5128 ret
= schedule_timeout(timeout
);
5129 io_schedule_finish(token
);
5133 EXPORT_SYMBOL(io_schedule_timeout
);
5135 void io_schedule(void)
5139 token
= io_schedule_prepare();
5141 io_schedule_finish(token
);
5143 EXPORT_SYMBOL(io_schedule
);
5146 * sys_sched_get_priority_max - return maximum RT priority.
5147 * @policy: scheduling class.
5149 * Return: On success, this syscall returns the maximum
5150 * rt_priority that can be used by a given scheduling class.
5151 * On failure, a negative error code is returned.
5153 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5160 ret
= MAX_USER_RT_PRIO
-1;
5162 case SCHED_DEADLINE
:
5173 * sys_sched_get_priority_min - return minimum RT priority.
5174 * @policy: scheduling class.
5176 * Return: On success, this syscall returns the minimum
5177 * rt_priority that can be used by a given scheduling class.
5178 * On failure, a negative error code is returned.
5180 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5189 case SCHED_DEADLINE
:
5198 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5200 struct task_struct
*p
;
5201 unsigned int time_slice
;
5211 p
= find_process_by_pid(pid
);
5215 retval
= security_task_getscheduler(p
);
5219 rq
= task_rq_lock(p
, &rf
);
5221 if (p
->sched_class
->get_rr_interval
)
5222 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5223 task_rq_unlock(rq
, p
, &rf
);
5226 jiffies_to_timespec64(time_slice
, t
);
5235 * sys_sched_rr_get_interval - return the default timeslice of a process.
5236 * @pid: pid of the process.
5237 * @interval: userspace pointer to the timeslice value.
5239 * this syscall writes the default timeslice value of a given process
5240 * into the user-space timespec buffer. A value of '0' means infinity.
5242 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5245 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5246 struct timespec __user
*, interval
)
5248 struct timespec64 t
;
5249 int retval
= sched_rr_get_interval(pid
, &t
);
5252 retval
= put_timespec64(&t
, interval
);
5257 #ifdef CONFIG_COMPAT
5258 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval
,
5260 struct compat_timespec __user
*, interval
)
5262 struct timespec64 t
;
5263 int retval
= sched_rr_get_interval(pid
, &t
);
5266 retval
= compat_put_timespec64(&t
, interval
);
5271 void sched_show_task(struct task_struct
*p
)
5273 unsigned long free
= 0;
5276 if (!try_get_task_stack(p
))
5279 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5281 if (p
->state
== TASK_RUNNING
)
5282 printk(KERN_CONT
" running task ");
5283 #ifdef CONFIG_DEBUG_STACK_USAGE
5284 free
= stack_not_used(p
);
5289 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5291 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5292 task_pid_nr(p
), ppid
,
5293 (unsigned long)task_thread_info(p
)->flags
);
5295 print_worker_info(KERN_INFO
, p
);
5296 show_stack(p
, NULL
);
5299 EXPORT_SYMBOL_GPL(sched_show_task
);
5302 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
5304 /* no filter, everything matches */
5308 /* filter, but doesn't match */
5309 if (!(p
->state
& state_filter
))
5313 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5316 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
5323 void show_state_filter(unsigned long state_filter
)
5325 struct task_struct
*g
, *p
;
5327 #if BITS_PER_LONG == 32
5329 " task PC stack pid father\n");
5332 " task PC stack pid father\n");
5335 for_each_process_thread(g
, p
) {
5337 * reset the NMI-timeout, listing all files on a slow
5338 * console might take a lot of time:
5339 * Also, reset softlockup watchdogs on all CPUs, because
5340 * another CPU might be blocked waiting for us to process
5343 touch_nmi_watchdog();
5344 touch_all_softlockup_watchdogs();
5345 if (state_filter_match(state_filter
, p
))
5349 #ifdef CONFIG_SCHED_DEBUG
5351 sysrq_sched_debug_show();
5355 * Only show locks if all tasks are dumped:
5358 debug_show_all_locks();
5362 * init_idle - set up an idle thread for a given CPU
5363 * @idle: task in question
5364 * @cpu: CPU the idle task belongs to
5366 * NOTE: this function does not set the idle thread's NEED_RESCHED
5367 * flag, to make booting more robust.
5369 void init_idle(struct task_struct
*idle
, int cpu
)
5371 struct rq
*rq
= cpu_rq(cpu
);
5372 unsigned long flags
;
5374 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5375 raw_spin_lock(&rq
->lock
);
5377 __sched_fork(0, idle
);
5378 idle
->state
= TASK_RUNNING
;
5379 idle
->se
.exec_start
= sched_clock();
5380 idle
->flags
|= PF_IDLE
;
5382 kasan_unpoison_task_stack(idle
);
5386 * Its possible that init_idle() gets called multiple times on a task,
5387 * in that case do_set_cpus_allowed() will not do the right thing.
5389 * And since this is boot we can forgo the serialization.
5391 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5394 * We're having a chicken and egg problem, even though we are
5395 * holding rq->lock, the CPU isn't yet set to this CPU so the
5396 * lockdep check in task_group() will fail.
5398 * Similar case to sched_fork(). / Alternatively we could
5399 * use task_rq_lock() here and obtain the other rq->lock.
5404 __set_task_cpu(idle
, cpu
);
5407 rq
->curr
= rq
->idle
= idle
;
5408 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5412 raw_spin_unlock(&rq
->lock
);
5413 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5415 /* Set the preempt count _outside_ the spinlocks! */
5416 init_idle_preempt_count(idle
, cpu
);
5419 * The idle tasks have their own, simple scheduling class:
5421 idle
->sched_class
= &idle_sched_class
;
5422 ftrace_graph_init_idle_task(idle
, cpu
);
5423 vtime_init_idle(idle
, cpu
);
5425 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5431 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5432 const struct cpumask
*trial
)
5436 if (!cpumask_weight(cur
))
5439 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
5444 int task_can_attach(struct task_struct
*p
,
5445 const struct cpumask
*cs_cpus_allowed
)
5450 * Kthreads which disallow setaffinity shouldn't be moved
5451 * to a new cpuset; we don't want to change their CPU
5452 * affinity and isolating such threads by their set of
5453 * allowed nodes is unnecessary. Thus, cpusets are not
5454 * applicable for such threads. This prevents checking for
5455 * success of set_cpus_allowed_ptr() on all attached tasks
5456 * before cpus_allowed may be changed.
5458 if (p
->flags
& PF_NO_SETAFFINITY
) {
5463 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5465 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
5471 bool sched_smp_initialized __read_mostly
;
5473 #ifdef CONFIG_NUMA_BALANCING
5474 /* Migrate current task p to target_cpu */
5475 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5477 struct migration_arg arg
= { p
, target_cpu
};
5478 int curr_cpu
= task_cpu(p
);
5480 if (curr_cpu
== target_cpu
)
5483 if (!cpumask_test_cpu(target_cpu
, &p
->cpus_allowed
))
5486 /* TODO: This is not properly updating schedstats */
5488 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5489 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5493 * Requeue a task on a given node and accurately track the number of NUMA
5494 * tasks on the runqueues
5496 void sched_setnuma(struct task_struct
*p
, int nid
)
5498 bool queued
, running
;
5502 rq
= task_rq_lock(p
, &rf
);
5503 queued
= task_on_rq_queued(p
);
5504 running
= task_current(rq
, p
);
5507 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5509 put_prev_task(rq
, p
);
5511 p
->numa_preferred_nid
= nid
;
5514 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5516 set_curr_task(rq
, p
);
5517 task_rq_unlock(rq
, p
, &rf
);
5519 #endif /* CONFIG_NUMA_BALANCING */
5521 #ifdef CONFIG_HOTPLUG_CPU
5523 * Ensure that the idle task is using init_mm right before its CPU goes
5526 void idle_task_exit(void)
5528 struct mm_struct
*mm
= current
->active_mm
;
5530 BUG_ON(cpu_online(smp_processor_id()));
5532 if (mm
!= &init_mm
) {
5533 switch_mm(mm
, &init_mm
, current
);
5534 current
->active_mm
= &init_mm
;
5535 finish_arch_post_lock_switch();
5541 * Since this CPU is going 'away' for a while, fold any nr_active delta
5542 * we might have. Assumes we're called after migrate_tasks() so that the
5543 * nr_active count is stable. We need to take the teardown thread which
5544 * is calling this into account, so we hand in adjust = 1 to the load
5547 * Also see the comment "Global load-average calculations".
5549 static void calc_load_migrate(struct rq
*rq
)
5551 long delta
= calc_load_fold_active(rq
, 1);
5553 atomic_long_add(delta
, &calc_load_tasks
);
5556 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5560 static const struct sched_class fake_sched_class
= {
5561 .put_prev_task
= put_prev_task_fake
,
5564 static struct task_struct fake_task
= {
5566 * Avoid pull_{rt,dl}_task()
5568 .prio
= MAX_PRIO
+ 1,
5569 .sched_class
= &fake_sched_class
,
5573 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5574 * try_to_wake_up()->select_task_rq().
5576 * Called with rq->lock held even though we'er in stop_machine() and
5577 * there's no concurrency possible, we hold the required locks anyway
5578 * because of lock validation efforts.
5580 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
5582 struct rq
*rq
= dead_rq
;
5583 struct task_struct
*next
, *stop
= rq
->stop
;
5584 struct rq_flags orf
= *rf
;
5588 * Fudge the rq selection such that the below task selection loop
5589 * doesn't get stuck on the currently eligible stop task.
5591 * We're currently inside stop_machine() and the rq is either stuck
5592 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5593 * either way we should never end up calling schedule() until we're
5599 * put_prev_task() and pick_next_task() sched
5600 * class method both need to have an up-to-date
5601 * value of rq->clock[_task]
5603 update_rq_clock(rq
);
5607 * There's this thread running, bail when that's the only
5610 if (rq
->nr_running
== 1)
5614 * pick_next_task() assumes pinned rq->lock:
5616 next
= pick_next_task(rq
, &fake_task
, rf
);
5618 put_prev_task(rq
, next
);
5621 * Rules for changing task_struct::cpus_allowed are holding
5622 * both pi_lock and rq->lock, such that holding either
5623 * stabilizes the mask.
5625 * Drop rq->lock is not quite as disastrous as it usually is
5626 * because !cpu_active at this point, which means load-balance
5627 * will not interfere. Also, stop-machine.
5630 raw_spin_lock(&next
->pi_lock
);
5634 * Since we're inside stop-machine, _nothing_ should have
5635 * changed the task, WARN if weird stuff happened, because in
5636 * that case the above rq->lock drop is a fail too.
5638 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5639 raw_spin_unlock(&next
->pi_lock
);
5643 /* Find suitable destination for @next, with force if needed. */
5644 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5645 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
5646 if (rq
!= dead_rq
) {
5652 raw_spin_unlock(&next
->pi_lock
);
5657 #endif /* CONFIG_HOTPLUG_CPU */
5659 void set_rq_online(struct rq
*rq
)
5662 const struct sched_class
*class;
5664 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5667 for_each_class(class) {
5668 if (class->rq_online
)
5669 class->rq_online(rq
);
5674 void set_rq_offline(struct rq
*rq
)
5677 const struct sched_class
*class;
5679 for_each_class(class) {
5680 if (class->rq_offline
)
5681 class->rq_offline(rq
);
5684 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5690 * used to mark begin/end of suspend/resume:
5692 static int num_cpus_frozen
;
5695 * Update cpusets according to cpu_active mask. If cpusets are
5696 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5697 * around partition_sched_domains().
5699 * If we come here as part of a suspend/resume, don't touch cpusets because we
5700 * want to restore it back to its original state upon resume anyway.
5702 static void cpuset_cpu_active(void)
5704 if (cpuhp_tasks_frozen
) {
5706 * num_cpus_frozen tracks how many CPUs are involved in suspend
5707 * resume sequence. As long as this is not the last online
5708 * operation in the resume sequence, just build a single sched
5709 * domain, ignoring cpusets.
5711 partition_sched_domains(1, NULL
, NULL
);
5712 if (--num_cpus_frozen
)
5715 * This is the last CPU online operation. So fall through and
5716 * restore the original sched domains by considering the
5717 * cpuset configurations.
5719 cpuset_force_rebuild();
5721 cpuset_update_active_cpus();
5724 static int cpuset_cpu_inactive(unsigned int cpu
)
5726 if (!cpuhp_tasks_frozen
) {
5727 if (dl_cpu_busy(cpu
))
5729 cpuset_update_active_cpus();
5732 partition_sched_domains(1, NULL
, NULL
);
5737 int sched_cpu_activate(unsigned int cpu
)
5739 struct rq
*rq
= cpu_rq(cpu
);
5742 #ifdef CONFIG_SCHED_SMT
5744 * The sched_smt_present static key needs to be evaluated on every
5745 * hotplug event because at boot time SMT might be disabled when
5746 * the number of booted CPUs is limited.
5748 * If then later a sibling gets hotplugged, then the key would stay
5749 * off and SMT scheduling would never be functional.
5751 if (cpumask_weight(cpu_smt_mask(cpu
)) > 1)
5752 static_branch_enable_cpuslocked(&sched_smt_present
);
5754 set_cpu_active(cpu
, true);
5756 if (sched_smp_initialized
) {
5757 sched_domains_numa_masks_set(cpu
);
5758 cpuset_cpu_active();
5762 * Put the rq online, if not already. This happens:
5764 * 1) In the early boot process, because we build the real domains
5765 * after all CPUs have been brought up.
5767 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5770 rq_lock_irqsave(rq
, &rf
);
5772 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5775 rq_unlock_irqrestore(rq
, &rf
);
5777 update_max_interval();
5782 int sched_cpu_deactivate(unsigned int cpu
)
5786 set_cpu_active(cpu
, false);
5788 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5789 * users of this state to go away such that all new such users will
5792 * Do sync before park smpboot threads to take care the rcu boost case.
5794 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
5796 if (!sched_smp_initialized
)
5799 ret
= cpuset_cpu_inactive(cpu
);
5801 set_cpu_active(cpu
, true);
5804 sched_domains_numa_masks_clear(cpu
);
5808 static void sched_rq_cpu_starting(unsigned int cpu
)
5810 struct rq
*rq
= cpu_rq(cpu
);
5812 rq
->calc_load_update
= calc_load_update
;
5813 update_max_interval();
5816 int sched_cpu_starting(unsigned int cpu
)
5818 sched_rq_cpu_starting(cpu
);
5819 sched_tick_start(cpu
);
5823 #ifdef CONFIG_HOTPLUG_CPU
5824 int sched_cpu_dying(unsigned int cpu
)
5826 struct rq
*rq
= cpu_rq(cpu
);
5829 /* Handle pending wakeups and then migrate everything off */
5830 sched_ttwu_pending();
5831 sched_tick_stop(cpu
);
5833 rq_lock_irqsave(rq
, &rf
);
5835 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5838 migrate_tasks(rq
, &rf
);
5839 BUG_ON(rq
->nr_running
!= 1);
5840 rq_unlock_irqrestore(rq
, &rf
);
5842 calc_load_migrate(rq
);
5843 update_max_interval();
5844 nohz_balance_exit_idle(rq
);
5850 void __init
sched_init_smp(void)
5855 * There's no userspace yet to cause hotplug operations; hence all the
5856 * CPU masks are stable and all blatant races in the below code cannot
5859 mutex_lock(&sched_domains_mutex
);
5860 sched_init_domains(cpu_active_mask
);
5861 mutex_unlock(&sched_domains_mutex
);
5863 /* Move init over to a non-isolated CPU */
5864 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
5866 sched_init_granularity();
5868 init_sched_rt_class();
5869 init_sched_dl_class();
5871 sched_smp_initialized
= true;
5874 static int __init
migration_init(void)
5876 sched_rq_cpu_starting(smp_processor_id());
5879 early_initcall(migration_init
);
5882 void __init
sched_init_smp(void)
5884 sched_init_granularity();
5886 #endif /* CONFIG_SMP */
5888 int in_sched_functions(unsigned long addr
)
5890 return in_lock_functions(addr
) ||
5891 (addr
>= (unsigned long)__sched_text_start
5892 && addr
< (unsigned long)__sched_text_end
);
5895 #ifdef CONFIG_CGROUP_SCHED
5897 * Default task group.
5898 * Every task in system belongs to this group at bootup.
5900 struct task_group root_task_group
;
5901 LIST_HEAD(task_groups
);
5903 /* Cacheline aligned slab cache for task_group */
5904 static struct kmem_cache
*task_group_cache __read_mostly
;
5907 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5908 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5910 void __init
sched_init(void)
5913 unsigned long alloc_size
= 0, ptr
;
5917 #ifdef CONFIG_FAIR_GROUP_SCHED
5918 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5920 #ifdef CONFIG_RT_GROUP_SCHED
5921 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5924 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
5926 #ifdef CONFIG_FAIR_GROUP_SCHED
5927 root_task_group
.se
= (struct sched_entity
**)ptr
;
5928 ptr
+= nr_cpu_ids
* sizeof(void **);
5930 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
5931 ptr
+= nr_cpu_ids
* sizeof(void **);
5933 #endif /* CONFIG_FAIR_GROUP_SCHED */
5934 #ifdef CONFIG_RT_GROUP_SCHED
5935 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
5936 ptr
+= nr_cpu_ids
* sizeof(void **);
5938 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
5939 ptr
+= nr_cpu_ids
* sizeof(void **);
5941 #endif /* CONFIG_RT_GROUP_SCHED */
5943 #ifdef CONFIG_CPUMASK_OFFSTACK
5944 for_each_possible_cpu(i
) {
5945 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5946 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5947 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5948 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5950 #endif /* CONFIG_CPUMASK_OFFSTACK */
5952 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
5953 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
5956 init_defrootdomain();
5959 #ifdef CONFIG_RT_GROUP_SCHED
5960 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
5961 global_rt_period(), global_rt_runtime());
5962 #endif /* CONFIG_RT_GROUP_SCHED */
5964 #ifdef CONFIG_CGROUP_SCHED
5965 task_group_cache
= KMEM_CACHE(task_group
, 0);
5967 list_add(&root_task_group
.list
, &task_groups
);
5968 INIT_LIST_HEAD(&root_task_group
.children
);
5969 INIT_LIST_HEAD(&root_task_group
.siblings
);
5970 autogroup_init(&init_task
);
5971 #endif /* CONFIG_CGROUP_SCHED */
5973 for_each_possible_cpu(i
) {
5977 raw_spin_lock_init(&rq
->lock
);
5979 rq
->calc_load_active
= 0;
5980 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
5981 init_cfs_rq(&rq
->cfs
);
5982 init_rt_rq(&rq
->rt
);
5983 init_dl_rq(&rq
->dl
);
5984 #ifdef CONFIG_FAIR_GROUP_SCHED
5985 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
5986 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
5987 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
5989 * How much CPU bandwidth does root_task_group get?
5991 * In case of task-groups formed thr' the cgroup filesystem, it
5992 * gets 100% of the CPU resources in the system. This overall
5993 * system CPU resource is divided among the tasks of
5994 * root_task_group and its child task-groups in a fair manner,
5995 * based on each entity's (task or task-group's) weight
5996 * (se->load.weight).
5998 * In other words, if root_task_group has 10 tasks of weight
5999 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6000 * then A0's share of the CPU resource is:
6002 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6004 * We achieve this by letting root_task_group's tasks sit
6005 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6007 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6008 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6009 #endif /* CONFIG_FAIR_GROUP_SCHED */
6011 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6012 #ifdef CONFIG_RT_GROUP_SCHED
6013 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6016 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6017 rq
->cpu_load
[j
] = 0;
6022 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6023 rq
->balance_callback
= NULL
;
6024 rq
->active_balance
= 0;
6025 rq
->next_balance
= jiffies
;
6030 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6031 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6033 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6035 rq_attach_root(rq
, &def_root_domain
);
6036 #ifdef CONFIG_NO_HZ_COMMON
6037 rq
->last_load_update_tick
= jiffies
;
6038 rq
->last_blocked_load_update_tick
= jiffies
;
6039 atomic_set(&rq
->nohz_flags
, 0);
6041 #endif /* CONFIG_SMP */
6043 atomic_set(&rq
->nr_iowait
, 0);
6046 set_load_weight(&init_task
, false);
6049 * The boot idle thread does lazy MMU switching as well:
6052 enter_lazy_tlb(&init_mm
, current
);
6055 * Make us the idle thread. Technically, schedule() should not be
6056 * called from this thread, however somewhere below it might be,
6057 * but because we are the idle thread, we just pick up running again
6058 * when this runqueue becomes "idle".
6060 init_idle(current
, smp_processor_id());
6062 calc_load_update
= jiffies
+ LOAD_FREQ
;
6065 idle_thread_set_boot_cpu();
6067 init_sched_fair_class();
6071 scheduler_running
= 1;
6074 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6075 static inline int preempt_count_equals(int preempt_offset
)
6077 int nested
= preempt_count() + rcu_preempt_depth();
6079 return (nested
== preempt_offset
);
6082 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6085 * Blocking primitives will set (and therefore destroy) current->state,
6086 * since we will exit with TASK_RUNNING make sure we enter with it,
6087 * otherwise we will destroy state.
6089 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6090 "do not call blocking ops when !TASK_RUNNING; "
6091 "state=%lx set at [<%p>] %pS\n",
6093 (void *)current
->task_state_change
,
6094 (void *)current
->task_state_change
);
6096 ___might_sleep(file
, line
, preempt_offset
);
6098 EXPORT_SYMBOL(__might_sleep
);
6100 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6102 /* Ratelimiting timestamp: */
6103 static unsigned long prev_jiffy
;
6105 unsigned long preempt_disable_ip
;
6107 /* WARN_ON_ONCE() by default, no rate limit required: */
6110 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6111 !is_idle_task(current
)) ||
6112 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6116 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6118 prev_jiffy
= jiffies
;
6120 /* Save this before calling printk(), since that will clobber it: */
6121 preempt_disable_ip
= get_preempt_disable_ip(current
);
6124 "BUG: sleeping function called from invalid context at %s:%d\n",
6127 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6128 in_atomic(), irqs_disabled(),
6129 current
->pid
, current
->comm
);
6131 if (task_stack_end_corrupted(current
))
6132 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6134 debug_show_held_locks(current
);
6135 if (irqs_disabled())
6136 print_irqtrace_events(current
);
6137 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6138 && !preempt_count_equals(preempt_offset
)) {
6139 pr_err("Preemption disabled at:");
6140 print_ip_sym(preempt_disable_ip
);
6144 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6146 EXPORT_SYMBOL(___might_sleep
);
6149 #ifdef CONFIG_MAGIC_SYSRQ
6150 void normalize_rt_tasks(void)
6152 struct task_struct
*g
, *p
;
6153 struct sched_attr attr
= {
6154 .sched_policy
= SCHED_NORMAL
,
6157 read_lock(&tasklist_lock
);
6158 for_each_process_thread(g
, p
) {
6160 * Only normalize user tasks:
6162 if (p
->flags
& PF_KTHREAD
)
6165 p
->se
.exec_start
= 0;
6166 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6167 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6168 schedstat_set(p
->se
.statistics
.block_start
, 0);
6170 if (!dl_task(p
) && !rt_task(p
)) {
6172 * Renice negative nice level userspace
6175 if (task_nice(p
) < 0)
6176 set_user_nice(p
, 0);
6180 __sched_setscheduler(p
, &attr
, false, false);
6182 read_unlock(&tasklist_lock
);
6185 #endif /* CONFIG_MAGIC_SYSRQ */
6187 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6189 * These functions are only useful for the IA64 MCA handling, or kdb.
6191 * They can only be called when the whole system has been
6192 * stopped - every CPU needs to be quiescent, and no scheduling
6193 * activity can take place. Using them for anything else would
6194 * be a serious bug, and as a result, they aren't even visible
6195 * under any other configuration.
6199 * curr_task - return the current task for a given CPU.
6200 * @cpu: the processor in question.
6202 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6204 * Return: The current task for @cpu.
6206 struct task_struct
*curr_task(int cpu
)
6208 return cpu_curr(cpu
);
6211 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6215 * set_curr_task - set the current task for a given CPU.
6216 * @cpu: the processor in question.
6217 * @p: the task pointer to set.
6219 * Description: This function must only be used when non-maskable interrupts
6220 * are serviced on a separate stack. It allows the architecture to switch the
6221 * notion of the current task on a CPU in a non-blocking manner. This function
6222 * must be called with all CPU's synchronized, and interrupts disabled, the
6223 * and caller must save the original value of the current task (see
6224 * curr_task() above) and restore that value before reenabling interrupts and
6225 * re-starting the system.
6227 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6229 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6236 #ifdef CONFIG_CGROUP_SCHED
6237 /* task_group_lock serializes the addition/removal of task groups */
6238 static DEFINE_SPINLOCK(task_group_lock
);
6240 static void sched_free_group(struct task_group
*tg
)
6242 free_fair_sched_group(tg
);
6243 free_rt_sched_group(tg
);
6245 kmem_cache_free(task_group_cache
, tg
);
6248 /* allocate runqueue etc for a new task group */
6249 struct task_group
*sched_create_group(struct task_group
*parent
)
6251 struct task_group
*tg
;
6253 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6255 return ERR_PTR(-ENOMEM
);
6257 if (!alloc_fair_sched_group(tg
, parent
))
6260 if (!alloc_rt_sched_group(tg
, parent
))
6266 sched_free_group(tg
);
6267 return ERR_PTR(-ENOMEM
);
6270 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6272 unsigned long flags
;
6274 spin_lock_irqsave(&task_group_lock
, flags
);
6275 list_add_rcu(&tg
->list
, &task_groups
);
6277 /* Root should already exist: */
6280 tg
->parent
= parent
;
6281 INIT_LIST_HEAD(&tg
->children
);
6282 list_add_rcu(&tg
->siblings
, &parent
->children
);
6283 spin_unlock_irqrestore(&task_group_lock
, flags
);
6285 online_fair_sched_group(tg
);
6288 /* rcu callback to free various structures associated with a task group */
6289 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6291 /* Now it should be safe to free those cfs_rqs: */
6292 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6295 void sched_destroy_group(struct task_group
*tg
)
6297 /* Wait for possible concurrent references to cfs_rqs complete: */
6298 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6301 void sched_offline_group(struct task_group
*tg
)
6303 unsigned long flags
;
6305 /* End participation in shares distribution: */
6306 unregister_fair_sched_group(tg
);
6308 spin_lock_irqsave(&task_group_lock
, flags
);
6309 list_del_rcu(&tg
->list
);
6310 list_del_rcu(&tg
->siblings
);
6311 spin_unlock_irqrestore(&task_group_lock
, flags
);
6314 static void sched_change_group(struct task_struct
*tsk
, int type
)
6316 struct task_group
*tg
;
6319 * All callers are synchronized by task_rq_lock(); we do not use RCU
6320 * which is pointless here. Thus, we pass "true" to task_css_check()
6321 * to prevent lockdep warnings.
6323 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
6324 struct task_group
, css
);
6325 tg
= autogroup_task_group(tsk
, tg
);
6326 tsk
->sched_task_group
= tg
;
6328 #ifdef CONFIG_FAIR_GROUP_SCHED
6329 if (tsk
->sched_class
->task_change_group
)
6330 tsk
->sched_class
->task_change_group(tsk
, type
);
6333 set_task_rq(tsk
, task_cpu(tsk
));
6337 * Change task's runqueue when it moves between groups.
6339 * The caller of this function should have put the task in its new group by
6340 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6343 void sched_move_task(struct task_struct
*tsk
)
6345 int queued
, running
, queue_flags
=
6346 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
6350 rq
= task_rq_lock(tsk
, &rf
);
6351 update_rq_clock(rq
);
6353 running
= task_current(rq
, tsk
);
6354 queued
= task_on_rq_queued(tsk
);
6357 dequeue_task(rq
, tsk
, queue_flags
);
6359 put_prev_task(rq
, tsk
);
6361 sched_change_group(tsk
, TASK_MOVE_GROUP
);
6364 enqueue_task(rq
, tsk
, queue_flags
);
6366 set_curr_task(rq
, tsk
);
6368 task_rq_unlock(rq
, tsk
, &rf
);
6371 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
6373 return css
? container_of(css
, struct task_group
, css
) : NULL
;
6376 static struct cgroup_subsys_state
*
6377 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6379 struct task_group
*parent
= css_tg(parent_css
);
6380 struct task_group
*tg
;
6383 /* This is early initialization for the top cgroup */
6384 return &root_task_group
.css
;
6387 tg
= sched_create_group(parent
);
6389 return ERR_PTR(-ENOMEM
);
6394 /* Expose task group only after completing cgroup initialization */
6395 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
6397 struct task_group
*tg
= css_tg(css
);
6398 struct task_group
*parent
= css_tg(css
->parent
);
6401 sched_online_group(tg
, parent
);
6405 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
6407 struct task_group
*tg
= css_tg(css
);
6409 sched_offline_group(tg
);
6412 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
6414 struct task_group
*tg
= css_tg(css
);
6417 * Relies on the RCU grace period between css_released() and this.
6419 sched_free_group(tg
);
6423 * This is called before wake_up_new_task(), therefore we really only
6424 * have to set its group bits, all the other stuff does not apply.
6426 static void cpu_cgroup_fork(struct task_struct
*task
)
6431 rq
= task_rq_lock(task
, &rf
);
6433 update_rq_clock(rq
);
6434 sched_change_group(task
, TASK_SET_GROUP
);
6436 task_rq_unlock(rq
, task
, &rf
);
6439 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
6441 struct task_struct
*task
;
6442 struct cgroup_subsys_state
*css
;
6445 cgroup_taskset_for_each(task
, css
, tset
) {
6446 #ifdef CONFIG_RT_GROUP_SCHED
6447 if (!sched_rt_can_attach(css_tg(css
), task
))
6450 /* We don't support RT-tasks being in separate groups */
6451 if (task
->sched_class
!= &fair_sched_class
)
6455 * Serialize against wake_up_new_task() such that if its
6456 * running, we're sure to observe its full state.
6458 raw_spin_lock_irq(&task
->pi_lock
);
6460 * Avoid calling sched_move_task() before wake_up_new_task()
6461 * has happened. This would lead to problems with PELT, due to
6462 * move wanting to detach+attach while we're not attached yet.
6464 if (task
->state
== TASK_NEW
)
6466 raw_spin_unlock_irq(&task
->pi_lock
);
6474 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
6476 struct task_struct
*task
;
6477 struct cgroup_subsys_state
*css
;
6479 cgroup_taskset_for_each(task
, css
, tset
)
6480 sched_move_task(task
);
6483 #ifdef CONFIG_FAIR_GROUP_SCHED
6484 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
6485 struct cftype
*cftype
, u64 shareval
)
6487 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
6490 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
6493 struct task_group
*tg
= css_tg(css
);
6495 return (u64
) scale_load_down(tg
->shares
);
6498 #ifdef CONFIG_CFS_BANDWIDTH
6499 static DEFINE_MUTEX(cfs_constraints_mutex
);
6501 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
6502 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
6504 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
6506 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
6508 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
6509 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6511 if (tg
== &root_task_group
)
6515 * Ensure we have at some amount of bandwidth every period. This is
6516 * to prevent reaching a state of large arrears when throttled via
6517 * entity_tick() resulting in prolonged exit starvation.
6519 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
6523 * Likewise, bound things on the otherside by preventing insane quota
6524 * periods. This also allows us to normalize in computing quota
6527 if (period
> max_cfs_quota_period
)
6531 * Prevent race between setting of cfs_rq->runtime_enabled and
6532 * unthrottle_offline_cfs_rqs().
6535 mutex_lock(&cfs_constraints_mutex
);
6536 ret
= __cfs_schedulable(tg
, period
, quota
);
6540 runtime_enabled
= quota
!= RUNTIME_INF
;
6541 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
6543 * If we need to toggle cfs_bandwidth_used, off->on must occur
6544 * before making related changes, and on->off must occur afterwards
6546 if (runtime_enabled
&& !runtime_was_enabled
)
6547 cfs_bandwidth_usage_inc();
6548 raw_spin_lock_irq(&cfs_b
->lock
);
6549 cfs_b
->period
= ns_to_ktime(period
);
6550 cfs_b
->quota
= quota
;
6552 __refill_cfs_bandwidth_runtime(cfs_b
);
6554 /* Restart the period timer (if active) to handle new period expiry: */
6555 if (runtime_enabled
)
6556 start_cfs_bandwidth(cfs_b
);
6558 raw_spin_unlock_irq(&cfs_b
->lock
);
6560 for_each_online_cpu(i
) {
6561 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
6562 struct rq
*rq
= cfs_rq
->rq
;
6565 rq_lock_irq(rq
, &rf
);
6566 cfs_rq
->runtime_enabled
= runtime_enabled
;
6567 cfs_rq
->runtime_remaining
= 0;
6569 if (cfs_rq
->throttled
)
6570 unthrottle_cfs_rq(cfs_rq
);
6571 rq_unlock_irq(rq
, &rf
);
6573 if (runtime_was_enabled
&& !runtime_enabled
)
6574 cfs_bandwidth_usage_dec();
6576 mutex_unlock(&cfs_constraints_mutex
);
6582 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
6586 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6587 if (cfs_quota_us
< 0)
6588 quota
= RUNTIME_INF
;
6590 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
6592 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6595 long tg_get_cfs_quota(struct task_group
*tg
)
6599 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
6602 quota_us
= tg
->cfs_bandwidth
.quota
;
6603 do_div(quota_us
, NSEC_PER_USEC
);
6608 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
6612 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
6613 quota
= tg
->cfs_bandwidth
.quota
;
6615 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6618 long tg_get_cfs_period(struct task_group
*tg
)
6622 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6623 do_div(cfs_period_us
, NSEC_PER_USEC
);
6625 return cfs_period_us
;
6628 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
6631 return tg_get_cfs_quota(css_tg(css
));
6634 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
6635 struct cftype
*cftype
, s64 cfs_quota_us
)
6637 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
6640 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
6643 return tg_get_cfs_period(css_tg(css
));
6646 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
6647 struct cftype
*cftype
, u64 cfs_period_us
)
6649 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
6652 struct cfs_schedulable_data
{
6653 struct task_group
*tg
;
6658 * normalize group quota/period to be quota/max_period
6659 * note: units are usecs
6661 static u64
normalize_cfs_quota(struct task_group
*tg
,
6662 struct cfs_schedulable_data
*d
)
6670 period
= tg_get_cfs_period(tg
);
6671 quota
= tg_get_cfs_quota(tg
);
6674 /* note: these should typically be equivalent */
6675 if (quota
== RUNTIME_INF
|| quota
== -1)
6678 return to_ratio(period
, quota
);
6681 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
6683 struct cfs_schedulable_data
*d
= data
;
6684 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6685 s64 quota
= 0, parent_quota
= -1;
6688 quota
= RUNTIME_INF
;
6690 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
6692 quota
= normalize_cfs_quota(tg
, d
);
6693 parent_quota
= parent_b
->hierarchical_quota
;
6696 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6697 * always take the min. On cgroup1, only inherit when no
6700 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
6701 quota
= min(quota
, parent_quota
);
6703 if (quota
== RUNTIME_INF
)
6704 quota
= parent_quota
;
6705 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
6709 cfs_b
->hierarchical_quota
= quota
;
6714 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
6717 struct cfs_schedulable_data data
= {
6723 if (quota
!= RUNTIME_INF
) {
6724 do_div(data
.period
, NSEC_PER_USEC
);
6725 do_div(data
.quota
, NSEC_PER_USEC
);
6729 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
6735 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
6737 struct task_group
*tg
= css_tg(seq_css(sf
));
6738 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6740 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
6741 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
6742 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
6744 if (schedstat_enabled() && tg
!= &root_task_group
) {
6748 for_each_possible_cpu(i
)
6749 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
6751 seq_printf(sf
, "wait_sum %llu\n", ws
);
6756 #endif /* CONFIG_CFS_BANDWIDTH */
6757 #endif /* CONFIG_FAIR_GROUP_SCHED */
6759 #ifdef CONFIG_RT_GROUP_SCHED
6760 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
6761 struct cftype
*cft
, s64 val
)
6763 return sched_group_set_rt_runtime(css_tg(css
), val
);
6766 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
6769 return sched_group_rt_runtime(css_tg(css
));
6772 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
6773 struct cftype
*cftype
, u64 rt_period_us
)
6775 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
6778 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
6781 return sched_group_rt_period(css_tg(css
));
6783 #endif /* CONFIG_RT_GROUP_SCHED */
6785 static struct cftype cpu_legacy_files
[] = {
6786 #ifdef CONFIG_FAIR_GROUP_SCHED
6789 .read_u64
= cpu_shares_read_u64
,
6790 .write_u64
= cpu_shares_write_u64
,
6793 #ifdef CONFIG_CFS_BANDWIDTH
6795 .name
= "cfs_quota_us",
6796 .read_s64
= cpu_cfs_quota_read_s64
,
6797 .write_s64
= cpu_cfs_quota_write_s64
,
6800 .name
= "cfs_period_us",
6801 .read_u64
= cpu_cfs_period_read_u64
,
6802 .write_u64
= cpu_cfs_period_write_u64
,
6806 .seq_show
= cpu_cfs_stat_show
,
6809 #ifdef CONFIG_RT_GROUP_SCHED
6811 .name
= "rt_runtime_us",
6812 .read_s64
= cpu_rt_runtime_read
,
6813 .write_s64
= cpu_rt_runtime_write
,
6816 .name
= "rt_period_us",
6817 .read_u64
= cpu_rt_period_read_uint
,
6818 .write_u64
= cpu_rt_period_write_uint
,
6824 static int cpu_extra_stat_show(struct seq_file
*sf
,
6825 struct cgroup_subsys_state
*css
)
6827 #ifdef CONFIG_CFS_BANDWIDTH
6829 struct task_group
*tg
= css_tg(css
);
6830 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6833 throttled_usec
= cfs_b
->throttled_time
;
6834 do_div(throttled_usec
, NSEC_PER_USEC
);
6836 seq_printf(sf
, "nr_periods %d\n"
6838 "throttled_usec %llu\n",
6839 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
6846 #ifdef CONFIG_FAIR_GROUP_SCHED
6847 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
6850 struct task_group
*tg
= css_tg(css
);
6851 u64 weight
= scale_load_down(tg
->shares
);
6853 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
6856 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
6857 struct cftype
*cft
, u64 weight
)
6860 * cgroup weight knobs should use the common MIN, DFL and MAX
6861 * values which are 1, 100 and 10000 respectively. While it loses
6862 * a bit of range on both ends, it maps pretty well onto the shares
6863 * value used by scheduler and the round-trip conversions preserve
6864 * the original value over the entire range.
6866 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
6869 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
6871 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
6874 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
6877 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
6878 int last_delta
= INT_MAX
;
6881 /* find the closest nice value to the current weight */
6882 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
6883 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
6884 if (delta
>= last_delta
)
6889 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
6892 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
6893 struct cftype
*cft
, s64 nice
)
6895 unsigned long weight
;
6898 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
6901 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
6902 idx
= array_index_nospec(idx
, 40);
6903 weight
= sched_prio_to_weight
[idx
];
6905 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
6909 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
6910 long period
, long quota
)
6913 seq_puts(sf
, "max");
6915 seq_printf(sf
, "%ld", quota
);
6917 seq_printf(sf
, " %ld\n", period
);
6920 /* caller should put the current value in *@periodp before calling */
6921 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
6922 u64
*periodp
, u64
*quotap
)
6924 char tok
[21]; /* U64_MAX */
6926 if (!sscanf(buf
, "%s %llu", tok
, periodp
))
6929 *periodp
*= NSEC_PER_USEC
;
6931 if (sscanf(tok
, "%llu", quotap
))
6932 *quotap
*= NSEC_PER_USEC
;
6933 else if (!strcmp(tok
, "max"))
6934 *quotap
= RUNTIME_INF
;
6941 #ifdef CONFIG_CFS_BANDWIDTH
6942 static int cpu_max_show(struct seq_file
*sf
, void *v
)
6944 struct task_group
*tg
= css_tg(seq_css(sf
));
6946 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
6950 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
6951 char *buf
, size_t nbytes
, loff_t off
)
6953 struct task_group
*tg
= css_tg(of_css(of
));
6954 u64 period
= tg_get_cfs_period(tg
);
6958 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
6960 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
6961 return ret
?: nbytes
;
6965 static struct cftype cpu_files
[] = {
6966 #ifdef CONFIG_FAIR_GROUP_SCHED
6969 .flags
= CFTYPE_NOT_ON_ROOT
,
6970 .read_u64
= cpu_weight_read_u64
,
6971 .write_u64
= cpu_weight_write_u64
,
6974 .name
= "weight.nice",
6975 .flags
= CFTYPE_NOT_ON_ROOT
,
6976 .read_s64
= cpu_weight_nice_read_s64
,
6977 .write_s64
= cpu_weight_nice_write_s64
,
6980 #ifdef CONFIG_CFS_BANDWIDTH
6983 .flags
= CFTYPE_NOT_ON_ROOT
,
6984 .seq_show
= cpu_max_show
,
6985 .write
= cpu_max_write
,
6991 struct cgroup_subsys cpu_cgrp_subsys
= {
6992 .css_alloc
= cpu_cgroup_css_alloc
,
6993 .css_online
= cpu_cgroup_css_online
,
6994 .css_released
= cpu_cgroup_css_released
,
6995 .css_free
= cpu_cgroup_css_free
,
6996 .css_extra_stat_show
= cpu_extra_stat_show
,
6997 .fork
= cpu_cgroup_fork
,
6998 .can_attach
= cpu_cgroup_can_attach
,
6999 .attach
= cpu_cgroup_attach
,
7000 .legacy_cftypes
= cpu_legacy_files
,
7001 .dfl_cftypes
= cpu_files
,
7006 #endif /* CONFIG_CGROUP_SCHED */
7008 void dump_cpu_task(int cpu
)
7010 pr_info("Task dump for CPU %d:\n", cpu
);
7011 sched_show_task(cpu_curr(cpu
));
7015 * Nice levels are multiplicative, with a gentle 10% change for every
7016 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7017 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7018 * that remained on nice 0.
7020 * The "10% effect" is relative and cumulative: from _any_ nice level,
7021 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7022 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7023 * If a task goes up by ~10% and another task goes down by ~10% then
7024 * the relative distance between them is ~25%.)
7026 const int sched_prio_to_weight
[40] = {
7027 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7028 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7029 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7030 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7031 /* 0 */ 1024, 820, 655, 526, 423,
7032 /* 5 */ 335, 272, 215, 172, 137,
7033 /* 10 */ 110, 87, 70, 56, 45,
7034 /* 15 */ 36, 29, 23, 18, 15,
7038 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7040 * In cases where the weight does not change often, we can use the
7041 * precalculated inverse to speed up arithmetics by turning divisions
7042 * into multiplications:
7044 const u32 sched_prio_to_wmult
[40] = {
7045 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7046 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7047 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7048 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7049 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7050 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7051 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7052 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7055 #undef CREATE_TRACE_POINTS