4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <uapi/linux/sched/types.h>
11 #include <linux/sched/loadavg.h>
12 #include <linux/sched/hotplug.h>
13 #include <linux/wait_bit.h>
14 #include <linux/cpuset.h>
15 #include <linux/delayacct.h>
16 #include <linux/init_task.h>
17 #include <linux/context_tracking.h>
18 #include <linux/rcupdate_wait.h>
20 #include <linux/blkdev.h>
21 #include <linux/kprobes.h>
22 #include <linux/mmu_context.h>
23 #include <linux/module.h>
24 #include <linux/nmi.h>
25 #include <linux/prefetch.h>
26 #include <linux/profile.h>
27 #include <linux/security.h>
28 #include <linux/syscalls.h>
30 #include <asm/switch_to.h>
32 #ifdef CONFIG_PARAVIRT
33 #include <asm/paravirt.h>
37 #include "../workqueue_internal.h"
38 #include "../smpboot.h"
40 #define CREATE_TRACE_POINTS
41 #include <trace/events/sched.h>
43 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
46 * Debugging: various feature bits
49 #define SCHED_FEAT(name, enabled) \
50 (1UL << __SCHED_FEAT_##name) * enabled |
52 const_debug
unsigned int sysctl_sched_features
=
59 * Number of tasks to iterate in a single balance run.
60 * Limited because this is done with IRQs disabled.
62 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
65 * period over which we average the RT time consumption, measured
70 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
73 * period over which we measure -rt task CPU usage in us.
76 unsigned int sysctl_sched_rt_period
= 1000000;
78 __read_mostly
int scheduler_running
;
81 * part of the period that we allow rt tasks to run in us.
84 int sysctl_sched_rt_runtime
= 950000;
86 /* CPUs with isolated domains */
87 cpumask_var_t cpu_isolated_map
;
90 * __task_rq_lock - lock the rq @p resides on.
92 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
97 lockdep_assert_held(&p
->pi_lock
);
101 raw_spin_lock(&rq
->lock
);
102 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
106 raw_spin_unlock(&rq
->lock
);
108 while (unlikely(task_on_rq_migrating(p
)))
114 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
116 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
117 __acquires(p
->pi_lock
)
123 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
125 raw_spin_lock(&rq
->lock
);
127 * move_queued_task() task_rq_lock()
130 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
131 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
132 * [S] ->cpu = new_cpu [L] task_rq()
136 * If we observe the old cpu in task_rq_lock, the acquire of
137 * the old rq->lock will fully serialize against the stores.
139 * If we observe the new CPU in task_rq_lock, the acquire will
140 * pair with the WMB to ensure we must then also see migrating.
142 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
146 raw_spin_unlock(&rq
->lock
);
147 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
149 while (unlikely(task_on_rq_migrating(p
)))
155 * RQ-clock updating methods:
158 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
161 * In theory, the compile should just see 0 here, and optimize out the call
162 * to sched_rt_avg_update. But I don't trust it...
164 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
165 s64 steal
= 0, irq_delta
= 0;
167 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
168 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
171 * Since irq_time is only updated on {soft,}irq_exit, we might run into
172 * this case when a previous update_rq_clock() happened inside a
175 * When this happens, we stop ->clock_task and only update the
176 * prev_irq_time stamp to account for the part that fit, so that a next
177 * update will consume the rest. This ensures ->clock_task is
180 * It does however cause some slight miss-attribution of {soft,}irq
181 * time, a more accurate solution would be to update the irq_time using
182 * the current rq->clock timestamp, except that would require using
185 if (irq_delta
> delta
)
188 rq
->prev_irq_time
+= irq_delta
;
191 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
192 if (static_key_false((¶virt_steal_rq_enabled
))) {
193 steal
= paravirt_steal_clock(cpu_of(rq
));
194 steal
-= rq
->prev_steal_time_rq
;
196 if (unlikely(steal
> delta
))
199 rq
->prev_steal_time_rq
+= steal
;
204 rq
->clock_task
+= delta
;
206 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
207 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
208 sched_rt_avg_update(rq
, irq_delta
+ steal
);
212 void update_rq_clock(struct rq
*rq
)
216 lockdep_assert_held(&rq
->lock
);
218 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
221 #ifdef CONFIG_SCHED_DEBUG
222 if (sched_feat(WARN_DOUBLE_CLOCK
))
223 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
224 rq
->clock_update_flags
|= RQCF_UPDATED
;
227 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
231 update_rq_clock_task(rq
, delta
);
235 #ifdef CONFIG_SCHED_HRTICK
237 * Use HR-timers to deliver accurate preemption points.
240 static void hrtick_clear(struct rq
*rq
)
242 if (hrtimer_active(&rq
->hrtick_timer
))
243 hrtimer_cancel(&rq
->hrtick_timer
);
247 * High-resolution timer tick.
248 * Runs from hardirq context with interrupts disabled.
250 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
252 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
255 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
259 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
262 return HRTIMER_NORESTART
;
267 static void __hrtick_restart(struct rq
*rq
)
269 struct hrtimer
*timer
= &rq
->hrtick_timer
;
271 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
275 * called from hardirq (IPI) context
277 static void __hrtick_start(void *arg
)
283 __hrtick_restart(rq
);
284 rq
->hrtick_csd_pending
= 0;
289 * Called to set the hrtick timer state.
291 * called with rq->lock held and irqs disabled
293 void hrtick_start(struct rq
*rq
, u64 delay
)
295 struct hrtimer
*timer
= &rq
->hrtick_timer
;
300 * Don't schedule slices shorter than 10000ns, that just
301 * doesn't make sense and can cause timer DoS.
303 delta
= max_t(s64
, delay
, 10000LL);
304 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
306 hrtimer_set_expires(timer
, time
);
308 if (rq
== this_rq()) {
309 __hrtick_restart(rq
);
310 } else if (!rq
->hrtick_csd_pending
) {
311 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
312 rq
->hrtick_csd_pending
= 1;
318 * Called to set the hrtick timer state.
320 * called with rq->lock held and irqs disabled
322 void hrtick_start(struct rq
*rq
, u64 delay
)
325 * Don't schedule slices shorter than 10000ns, that just
326 * doesn't make sense. Rely on vruntime for fairness.
328 delay
= max_t(u64
, delay
, 10000LL);
329 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
330 HRTIMER_MODE_REL_PINNED
);
332 #endif /* CONFIG_SMP */
334 static void init_rq_hrtick(struct rq
*rq
)
337 rq
->hrtick_csd_pending
= 0;
339 rq
->hrtick_csd
.flags
= 0;
340 rq
->hrtick_csd
.func
= __hrtick_start
;
341 rq
->hrtick_csd
.info
= rq
;
344 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
345 rq
->hrtick_timer
.function
= hrtick
;
347 #else /* CONFIG_SCHED_HRTICK */
348 static inline void hrtick_clear(struct rq
*rq
)
352 static inline void init_rq_hrtick(struct rq
*rq
)
355 #endif /* CONFIG_SCHED_HRTICK */
358 * cmpxchg based fetch_or, macro so it works for different integer types
360 #define fetch_or(ptr, mask) \
362 typeof(ptr) _ptr = (ptr); \
363 typeof(mask) _mask = (mask); \
364 typeof(*_ptr) _old, _val = *_ptr; \
367 _old = cmpxchg(_ptr, _val, _val | _mask); \
375 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
377 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
378 * this avoids any races wrt polling state changes and thereby avoids
381 static bool set_nr_and_not_polling(struct task_struct
*p
)
383 struct thread_info
*ti
= task_thread_info(p
);
384 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
388 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
390 * If this returns true, then the idle task promises to call
391 * sched_ttwu_pending() and reschedule soon.
393 static bool set_nr_if_polling(struct task_struct
*p
)
395 struct thread_info
*ti
= task_thread_info(p
);
396 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
399 if (!(val
& _TIF_POLLING_NRFLAG
))
401 if (val
& _TIF_NEED_RESCHED
)
403 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
412 static bool set_nr_and_not_polling(struct task_struct
*p
)
414 set_tsk_need_resched(p
);
419 static bool set_nr_if_polling(struct task_struct
*p
)
426 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
428 struct wake_q_node
*node
= &task
->wake_q
;
431 * Atomically grab the task, if ->wake_q is !nil already it means
432 * its already queued (either by us or someone else) and will get the
433 * wakeup due to that.
435 * This cmpxchg() implies a full barrier, which pairs with the write
436 * barrier implied by the wakeup in wake_up_q().
438 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
441 get_task_struct(task
);
444 * The head is context local, there can be no concurrency.
447 head
->lastp
= &node
->next
;
450 void wake_up_q(struct wake_q_head
*head
)
452 struct wake_q_node
*node
= head
->first
;
454 while (node
!= WAKE_Q_TAIL
) {
455 struct task_struct
*task
;
457 task
= container_of(node
, struct task_struct
, wake_q
);
459 /* Task can safely be re-inserted now: */
461 task
->wake_q
.next
= NULL
;
464 * wake_up_process() implies a wmb() to pair with the queueing
465 * in wake_q_add() so as not to miss wakeups.
467 wake_up_process(task
);
468 put_task_struct(task
);
473 * resched_curr - mark rq's current task 'to be rescheduled now'.
475 * On UP this means the setting of the need_resched flag, on SMP it
476 * might also involve a cross-CPU call to trigger the scheduler on
479 void resched_curr(struct rq
*rq
)
481 struct task_struct
*curr
= rq
->curr
;
484 lockdep_assert_held(&rq
->lock
);
486 if (test_tsk_need_resched(curr
))
491 if (cpu
== smp_processor_id()) {
492 set_tsk_need_resched(curr
);
493 set_preempt_need_resched();
497 if (set_nr_and_not_polling(curr
))
498 smp_send_reschedule(cpu
);
500 trace_sched_wake_idle_without_ipi(cpu
);
503 void resched_cpu(int cpu
)
505 struct rq
*rq
= cpu_rq(cpu
);
508 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
511 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
515 #ifdef CONFIG_NO_HZ_COMMON
517 * In the semi idle case, use the nearest busy CPU for migrating timers
518 * from an idle CPU. This is good for power-savings.
520 * We don't do similar optimization for completely idle system, as
521 * selecting an idle CPU will add more delays to the timers than intended
522 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
524 int get_nohz_timer_target(void)
526 int i
, cpu
= smp_processor_id();
527 struct sched_domain
*sd
;
529 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
533 for_each_domain(cpu
, sd
) {
534 for_each_cpu(i
, sched_domain_span(sd
)) {
538 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
545 if (!is_housekeeping_cpu(cpu
))
546 cpu
= housekeeping_any_cpu();
553 * When add_timer_on() enqueues a timer into the timer wheel of an
554 * idle CPU then this timer might expire before the next timer event
555 * which is scheduled to wake up that CPU. In case of a completely
556 * idle system the next event might even be infinite time into the
557 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
558 * leaves the inner idle loop so the newly added timer is taken into
559 * account when the CPU goes back to idle and evaluates the timer
560 * wheel for the next timer event.
562 static void wake_up_idle_cpu(int cpu
)
564 struct rq
*rq
= cpu_rq(cpu
);
566 if (cpu
== smp_processor_id())
569 if (set_nr_and_not_polling(rq
->idle
))
570 smp_send_reschedule(cpu
);
572 trace_sched_wake_idle_without_ipi(cpu
);
575 static bool wake_up_full_nohz_cpu(int cpu
)
578 * We just need the target to call irq_exit() and re-evaluate
579 * the next tick. The nohz full kick at least implies that.
580 * If needed we can still optimize that later with an
583 if (cpu_is_offline(cpu
))
584 return true; /* Don't try to wake offline CPUs. */
585 if (tick_nohz_full_cpu(cpu
)) {
586 if (cpu
!= smp_processor_id() ||
587 tick_nohz_tick_stopped())
588 tick_nohz_full_kick_cpu(cpu
);
596 * Wake up the specified CPU. If the CPU is going offline, it is the
597 * caller's responsibility to deal with the lost wakeup, for example,
598 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
600 void wake_up_nohz_cpu(int cpu
)
602 if (!wake_up_full_nohz_cpu(cpu
))
603 wake_up_idle_cpu(cpu
);
606 static inline bool got_nohz_idle_kick(void)
608 int cpu
= smp_processor_id();
610 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
613 if (idle_cpu(cpu
) && !need_resched())
617 * We can't run Idle Load Balance on this CPU for this time so we
618 * cancel it and clear NOHZ_BALANCE_KICK
620 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
624 #else /* CONFIG_NO_HZ_COMMON */
626 static inline bool got_nohz_idle_kick(void)
631 #endif /* CONFIG_NO_HZ_COMMON */
633 #ifdef CONFIG_NO_HZ_FULL
634 bool sched_can_stop_tick(struct rq
*rq
)
638 /* Deadline tasks, even if single, need the tick */
639 if (rq
->dl
.dl_nr_running
)
643 * If there are more than one RR tasks, we need the tick to effect the
644 * actual RR behaviour.
646 if (rq
->rt
.rr_nr_running
) {
647 if (rq
->rt
.rr_nr_running
== 1)
654 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
655 * forced preemption between FIFO tasks.
657 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
662 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
663 * if there's more than one we need the tick for involuntary
666 if (rq
->nr_running
> 1)
671 #endif /* CONFIG_NO_HZ_FULL */
673 void sched_avg_update(struct rq
*rq
)
675 s64 period
= sched_avg_period();
677 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
679 * Inline assembly required to prevent the compiler
680 * optimising this loop into a divmod call.
681 * See __iter_div_u64_rem() for another example of this.
683 asm("" : "+rm" (rq
->age_stamp
));
684 rq
->age_stamp
+= period
;
689 #endif /* CONFIG_SMP */
691 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
692 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
694 * Iterate task_group tree rooted at *from, calling @down when first entering a
695 * node and @up when leaving it for the final time.
697 * Caller must hold rcu_lock or sufficient equivalent.
699 int walk_tg_tree_from(struct task_group
*from
,
700 tg_visitor down
, tg_visitor up
, void *data
)
702 struct task_group
*parent
, *child
;
708 ret
= (*down
)(parent
, data
);
711 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
718 ret
= (*up
)(parent
, data
);
719 if (ret
|| parent
== from
)
723 parent
= parent
->parent
;
730 int tg_nop(struct task_group
*tg
, void *data
)
736 static void set_load_weight(struct task_struct
*p
)
738 int prio
= p
->static_prio
- MAX_RT_PRIO
;
739 struct load_weight
*load
= &p
->se
.load
;
742 * SCHED_IDLE tasks get minimal weight:
744 if (idle_policy(p
->policy
)) {
745 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
746 load
->inv_weight
= WMULT_IDLEPRIO
;
750 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
751 load
->inv_weight
= sched_prio_to_wmult
[prio
];
754 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
756 if (!(flags
& ENQUEUE_NOCLOCK
))
759 if (!(flags
& ENQUEUE_RESTORE
))
760 sched_info_queued(rq
, p
);
762 p
->sched_class
->enqueue_task(rq
, p
, flags
);
765 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
767 if (!(flags
& DEQUEUE_NOCLOCK
))
770 if (!(flags
& DEQUEUE_SAVE
))
771 sched_info_dequeued(rq
, p
);
773 p
->sched_class
->dequeue_task(rq
, p
, flags
);
776 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
778 if (task_contributes_to_load(p
))
779 rq
->nr_uninterruptible
--;
781 enqueue_task(rq
, p
, flags
);
784 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
786 if (task_contributes_to_load(p
))
787 rq
->nr_uninterruptible
++;
789 dequeue_task(rq
, p
, flags
);
793 * __normal_prio - return the priority that is based on the static prio
795 static inline int __normal_prio(struct task_struct
*p
)
797 return p
->static_prio
;
801 * Calculate the expected normal priority: i.e. priority
802 * without taking RT-inheritance into account. Might be
803 * boosted by interactivity modifiers. Changes upon fork,
804 * setprio syscalls, and whenever the interactivity
805 * estimator recalculates.
807 static inline int normal_prio(struct task_struct
*p
)
811 if (task_has_dl_policy(p
))
812 prio
= MAX_DL_PRIO
-1;
813 else if (task_has_rt_policy(p
))
814 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
816 prio
= __normal_prio(p
);
821 * Calculate the current priority, i.e. the priority
822 * taken into account by the scheduler. This value might
823 * be boosted by RT tasks, or might be boosted by
824 * interactivity modifiers. Will be RT if the task got
825 * RT-boosted. If not then it returns p->normal_prio.
827 static int effective_prio(struct task_struct
*p
)
829 p
->normal_prio
= normal_prio(p
);
831 * If we are RT tasks or we were boosted to RT priority,
832 * keep the priority unchanged. Otherwise, update priority
833 * to the normal priority:
835 if (!rt_prio(p
->prio
))
836 return p
->normal_prio
;
841 * task_curr - is this task currently executing on a CPU?
842 * @p: the task in question.
844 * Return: 1 if the task is currently executing. 0 otherwise.
846 inline int task_curr(const struct task_struct
*p
)
848 return cpu_curr(task_cpu(p
)) == p
;
852 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
853 * use the balance_callback list if you want balancing.
855 * this means any call to check_class_changed() must be followed by a call to
856 * balance_callback().
858 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
859 const struct sched_class
*prev_class
,
862 if (prev_class
!= p
->sched_class
) {
863 if (prev_class
->switched_from
)
864 prev_class
->switched_from(rq
, p
);
866 p
->sched_class
->switched_to(rq
, p
);
867 } else if (oldprio
!= p
->prio
|| dl_task(p
))
868 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
871 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
873 const struct sched_class
*class;
875 if (p
->sched_class
== rq
->curr
->sched_class
) {
876 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
878 for_each_class(class) {
879 if (class == rq
->curr
->sched_class
)
881 if (class == p
->sched_class
) {
889 * A queue event has occurred, and we're going to schedule. In
890 * this case, we can save a useless back to back clock update.
892 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
893 rq_clock_skip_update(rq
, true);
898 * This is how migration works:
900 * 1) we invoke migration_cpu_stop() on the target CPU using
902 * 2) stopper starts to run (implicitly forcing the migrated thread
904 * 3) it checks whether the migrated task is still in the wrong runqueue.
905 * 4) if it's in the wrong runqueue then the migration thread removes
906 * it and puts it into the right queue.
907 * 5) stopper completes and stop_one_cpu() returns and the migration
912 * move_queued_task - move a queued task to new rq.
914 * Returns (locked) new rq. Old rq's lock is released.
916 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
917 struct task_struct
*p
, int new_cpu
)
919 lockdep_assert_held(&rq
->lock
);
921 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
922 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
923 set_task_cpu(p
, new_cpu
);
926 rq
= cpu_rq(new_cpu
);
929 BUG_ON(task_cpu(p
) != new_cpu
);
930 enqueue_task(rq
, p
, 0);
931 p
->on_rq
= TASK_ON_RQ_QUEUED
;
932 check_preempt_curr(rq
, p
, 0);
937 struct migration_arg
{
938 struct task_struct
*task
;
943 * Move (not current) task off this CPU, onto the destination CPU. We're doing
944 * this because either it can't run here any more (set_cpus_allowed()
945 * away from this CPU, or CPU going down), or because we're
946 * attempting to rebalance this task on exec (sched_exec).
948 * So we race with normal scheduler movements, but that's OK, as long
949 * as the task is no longer on this CPU.
951 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
952 struct task_struct
*p
, int dest_cpu
)
954 if (p
->flags
& PF_KTHREAD
) {
955 if (unlikely(!cpu_online(dest_cpu
)))
958 if (unlikely(!cpu_active(dest_cpu
)))
962 /* Affinity changed (again). */
963 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
967 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
973 * migration_cpu_stop - this will be executed by a highprio stopper thread
974 * and performs thread migration by bumping thread off CPU then
975 * 'pushing' onto another runqueue.
977 static int migration_cpu_stop(void *data
)
979 struct migration_arg
*arg
= data
;
980 struct task_struct
*p
= arg
->task
;
981 struct rq
*rq
= this_rq();
985 * The original target CPU might have gone down and we might
986 * be on another CPU but it doesn't matter.
990 * We need to explicitly wake pending tasks before running
991 * __migrate_task() such that we will not miss enforcing cpus_allowed
992 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
994 sched_ttwu_pending();
996 raw_spin_lock(&p
->pi_lock
);
999 * If task_rq(p) != rq, it cannot be migrated here, because we're
1000 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1001 * we're holding p->pi_lock.
1003 if (task_rq(p
) == rq
) {
1004 if (task_on_rq_queued(p
))
1005 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1007 p
->wake_cpu
= arg
->dest_cpu
;
1010 raw_spin_unlock(&p
->pi_lock
);
1017 * sched_class::set_cpus_allowed must do the below, but is not required to
1018 * actually call this function.
1020 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1022 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1023 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1026 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1028 struct rq
*rq
= task_rq(p
);
1029 bool queued
, running
;
1031 lockdep_assert_held(&p
->pi_lock
);
1033 queued
= task_on_rq_queued(p
);
1034 running
= task_current(rq
, p
);
1038 * Because __kthread_bind() calls this on blocked tasks without
1041 lockdep_assert_held(&rq
->lock
);
1042 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1045 put_prev_task(rq
, p
);
1047 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1050 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1052 set_curr_task(rq
, p
);
1056 * Change a given task's CPU affinity. Migrate the thread to a
1057 * proper CPU and schedule it away if the CPU it's executing on
1058 * is removed from the allowed bitmask.
1060 * NOTE: the caller must have a valid reference to the task, the
1061 * task must not exit() & deallocate itself prematurely. The
1062 * call is not atomic; no spinlocks may be held.
1064 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1065 const struct cpumask
*new_mask
, bool check
)
1067 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1068 unsigned int dest_cpu
;
1073 rq
= task_rq_lock(p
, &rf
);
1074 update_rq_clock(rq
);
1076 if (p
->flags
& PF_KTHREAD
) {
1078 * Kernel threads are allowed on online && !active CPUs
1080 cpu_valid_mask
= cpu_online_mask
;
1084 * Must re-check here, to close a race against __kthread_bind(),
1085 * sched_setaffinity() is not guaranteed to observe the flag.
1087 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1092 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1095 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1100 do_set_cpus_allowed(p
, new_mask
);
1102 if (p
->flags
& PF_KTHREAD
) {
1104 * For kernel threads that do indeed end up on online &&
1105 * !active we want to ensure they are strict per-CPU threads.
1107 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1108 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1109 p
->nr_cpus_allowed
!= 1);
1112 /* Can the task run on the task's current CPU? If so, we're done */
1113 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1116 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1117 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1118 struct migration_arg arg
= { p
, dest_cpu
};
1119 /* Need help from migration thread: drop lock and wait. */
1120 task_rq_unlock(rq
, p
, &rf
);
1121 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1122 tlb_migrate_finish(p
->mm
);
1124 } else if (task_on_rq_queued(p
)) {
1126 * OK, since we're going to drop the lock immediately
1127 * afterwards anyway.
1129 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1132 task_rq_unlock(rq
, p
, &rf
);
1137 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1139 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1141 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1143 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1145 #ifdef CONFIG_SCHED_DEBUG
1147 * We should never call set_task_cpu() on a blocked task,
1148 * ttwu() will sort out the placement.
1150 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1154 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1155 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1156 * time relying on p->on_rq.
1158 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1159 p
->sched_class
== &fair_sched_class
&&
1160 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1162 #ifdef CONFIG_LOCKDEP
1164 * The caller should hold either p->pi_lock or rq->lock, when changing
1165 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1167 * sched_move_task() holds both and thus holding either pins the cgroup,
1170 * Furthermore, all task_rq users should acquire both locks, see
1173 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1174 lockdep_is_held(&task_rq(p
)->lock
)));
1177 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1179 WARN_ON_ONCE(!cpu_online(new_cpu
));
1182 trace_sched_migrate_task(p
, new_cpu
);
1184 if (task_cpu(p
) != new_cpu
) {
1185 if (p
->sched_class
->migrate_task_rq
)
1186 p
->sched_class
->migrate_task_rq(p
);
1187 p
->se
.nr_migrations
++;
1188 perf_event_task_migrate(p
);
1191 __set_task_cpu(p
, new_cpu
);
1194 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1196 if (task_on_rq_queued(p
)) {
1197 struct rq
*src_rq
, *dst_rq
;
1198 struct rq_flags srf
, drf
;
1200 src_rq
= task_rq(p
);
1201 dst_rq
= cpu_rq(cpu
);
1203 rq_pin_lock(src_rq
, &srf
);
1204 rq_pin_lock(dst_rq
, &drf
);
1206 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1207 deactivate_task(src_rq
, p
, 0);
1208 set_task_cpu(p
, cpu
);
1209 activate_task(dst_rq
, p
, 0);
1210 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1211 check_preempt_curr(dst_rq
, p
, 0);
1213 rq_unpin_lock(dst_rq
, &drf
);
1214 rq_unpin_lock(src_rq
, &srf
);
1218 * Task isn't running anymore; make it appear like we migrated
1219 * it before it went to sleep. This means on wakeup we make the
1220 * previous CPU our target instead of where it really is.
1226 struct migration_swap_arg
{
1227 struct task_struct
*src_task
, *dst_task
;
1228 int src_cpu
, dst_cpu
;
1231 static int migrate_swap_stop(void *data
)
1233 struct migration_swap_arg
*arg
= data
;
1234 struct rq
*src_rq
, *dst_rq
;
1237 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1240 src_rq
= cpu_rq(arg
->src_cpu
);
1241 dst_rq
= cpu_rq(arg
->dst_cpu
);
1243 double_raw_lock(&arg
->src_task
->pi_lock
,
1244 &arg
->dst_task
->pi_lock
);
1245 double_rq_lock(src_rq
, dst_rq
);
1247 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1250 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1253 if (!cpumask_test_cpu(arg
->dst_cpu
, &arg
->src_task
->cpus_allowed
))
1256 if (!cpumask_test_cpu(arg
->src_cpu
, &arg
->dst_task
->cpus_allowed
))
1259 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1260 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1265 double_rq_unlock(src_rq
, dst_rq
);
1266 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1267 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1273 * Cross migrate two tasks
1275 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1277 struct migration_swap_arg arg
;
1280 arg
= (struct migration_swap_arg
){
1282 .src_cpu
= task_cpu(cur
),
1284 .dst_cpu
= task_cpu(p
),
1287 if (arg
.src_cpu
== arg
.dst_cpu
)
1291 * These three tests are all lockless; this is OK since all of them
1292 * will be re-checked with proper locks held further down the line.
1294 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1297 if (!cpumask_test_cpu(arg
.dst_cpu
, &arg
.src_task
->cpus_allowed
))
1300 if (!cpumask_test_cpu(arg
.src_cpu
, &arg
.dst_task
->cpus_allowed
))
1303 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1304 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1311 * wait_task_inactive - wait for a thread to unschedule.
1313 * If @match_state is nonzero, it's the @p->state value just checked and
1314 * not expected to change. If it changes, i.e. @p might have woken up,
1315 * then return zero. When we succeed in waiting for @p to be off its CPU,
1316 * we return a positive number (its total switch count). If a second call
1317 * a short while later returns the same number, the caller can be sure that
1318 * @p has remained unscheduled the whole time.
1320 * The caller must ensure that the task *will* unschedule sometime soon,
1321 * else this function might spin for a *long* time. This function can't
1322 * be called with interrupts off, or it may introduce deadlock with
1323 * smp_call_function() if an IPI is sent by the same process we are
1324 * waiting to become inactive.
1326 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1328 int running
, queued
;
1335 * We do the initial early heuristics without holding
1336 * any task-queue locks at all. We'll only try to get
1337 * the runqueue lock when things look like they will
1343 * If the task is actively running on another CPU
1344 * still, just relax and busy-wait without holding
1347 * NOTE! Since we don't hold any locks, it's not
1348 * even sure that "rq" stays as the right runqueue!
1349 * But we don't care, since "task_running()" will
1350 * return false if the runqueue has changed and p
1351 * is actually now running somewhere else!
1353 while (task_running(rq
, p
)) {
1354 if (match_state
&& unlikely(p
->state
!= match_state
))
1360 * Ok, time to look more closely! We need the rq
1361 * lock now, to be *sure*. If we're wrong, we'll
1362 * just go back and repeat.
1364 rq
= task_rq_lock(p
, &rf
);
1365 trace_sched_wait_task(p
);
1366 running
= task_running(rq
, p
);
1367 queued
= task_on_rq_queued(p
);
1369 if (!match_state
|| p
->state
== match_state
)
1370 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1371 task_rq_unlock(rq
, p
, &rf
);
1374 * If it changed from the expected state, bail out now.
1376 if (unlikely(!ncsw
))
1380 * Was it really running after all now that we
1381 * checked with the proper locks actually held?
1383 * Oops. Go back and try again..
1385 if (unlikely(running
)) {
1391 * It's not enough that it's not actively running,
1392 * it must be off the runqueue _entirely_, and not
1395 * So if it was still runnable (but just not actively
1396 * running right now), it's preempted, and we should
1397 * yield - it could be a while.
1399 if (unlikely(queued
)) {
1400 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1402 set_current_state(TASK_UNINTERRUPTIBLE
);
1403 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1408 * Ahh, all good. It wasn't running, and it wasn't
1409 * runnable, which means that it will never become
1410 * running in the future either. We're all done!
1419 * kick_process - kick a running thread to enter/exit the kernel
1420 * @p: the to-be-kicked thread
1422 * Cause a process which is running on another CPU to enter
1423 * kernel-mode, without any delay. (to get signals handled.)
1425 * NOTE: this function doesn't have to take the runqueue lock,
1426 * because all it wants to ensure is that the remote task enters
1427 * the kernel. If the IPI races and the task has been migrated
1428 * to another CPU then no harm is done and the purpose has been
1431 void kick_process(struct task_struct
*p
)
1437 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1438 smp_send_reschedule(cpu
);
1441 EXPORT_SYMBOL_GPL(kick_process
);
1444 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1446 * A few notes on cpu_active vs cpu_online:
1448 * - cpu_active must be a subset of cpu_online
1450 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1451 * see __set_cpus_allowed_ptr(). At this point the newly online
1452 * CPU isn't yet part of the sched domains, and balancing will not
1455 * - on CPU-down we clear cpu_active() to mask the sched domains and
1456 * avoid the load balancer to place new tasks on the to be removed
1457 * CPU. Existing tasks will remain running there and will be taken
1460 * This means that fallback selection must not select !active CPUs.
1461 * And can assume that any active CPU must be online. Conversely
1462 * select_task_rq() below may allow selection of !active CPUs in order
1463 * to satisfy the above rules.
1465 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1467 int nid
= cpu_to_node(cpu
);
1468 const struct cpumask
*nodemask
= NULL
;
1469 enum { cpuset
, possible
, fail
} state
= cpuset
;
1473 * If the node that the CPU is on has been offlined, cpu_to_node()
1474 * will return -1. There is no CPU on the node, and we should
1475 * select the CPU on the other node.
1478 nodemask
= cpumask_of_node(nid
);
1480 /* Look for allowed, online CPU in same node. */
1481 for_each_cpu(dest_cpu
, nodemask
) {
1482 if (!cpu_active(dest_cpu
))
1484 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
1490 /* Any allowed, online CPU? */
1491 for_each_cpu(dest_cpu
, &p
->cpus_allowed
) {
1492 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1494 if (!cpu_online(dest_cpu
))
1499 /* No more Mr. Nice Guy. */
1502 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1503 cpuset_cpus_allowed_fallback(p
);
1509 do_set_cpus_allowed(p
, cpu_possible_mask
);
1520 if (state
!= cpuset
) {
1522 * Don't tell them about moving exiting tasks or
1523 * kernel threads (both mm NULL), since they never
1526 if (p
->mm
&& printk_ratelimit()) {
1527 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1528 task_pid_nr(p
), p
->comm
, cpu
);
1536 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1539 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1541 lockdep_assert_held(&p
->pi_lock
);
1543 if (p
->nr_cpus_allowed
> 1)
1544 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1546 cpu
= cpumask_any(&p
->cpus_allowed
);
1549 * In order not to call set_task_cpu() on a blocking task we need
1550 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1553 * Since this is common to all placement strategies, this lives here.
1555 * [ this allows ->select_task() to simply return task_cpu(p) and
1556 * not worry about this generic constraint ]
1558 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
1560 cpu
= select_fallback_rq(task_cpu(p
), p
);
1565 static void update_avg(u64
*avg
, u64 sample
)
1567 s64 diff
= sample
- *avg
;
1571 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1573 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1574 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1578 * Make it appear like a SCHED_FIFO task, its something
1579 * userspace knows about and won't get confused about.
1581 * Also, it will make PI more or less work without too
1582 * much confusion -- but then, stop work should not
1583 * rely on PI working anyway.
1585 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1587 stop
->sched_class
= &stop_sched_class
;
1590 cpu_rq(cpu
)->stop
= stop
;
1594 * Reset it back to a normal scheduling class so that
1595 * it can die in pieces.
1597 old_stop
->sched_class
= &rt_sched_class
;
1603 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1604 const struct cpumask
*new_mask
, bool check
)
1606 return set_cpus_allowed_ptr(p
, new_mask
);
1609 #endif /* CONFIG_SMP */
1612 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1616 if (!schedstat_enabled())
1622 if (cpu
== rq
->cpu
) {
1623 schedstat_inc(rq
->ttwu_local
);
1624 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1626 struct sched_domain
*sd
;
1628 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1630 for_each_domain(rq
->cpu
, sd
) {
1631 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1632 schedstat_inc(sd
->ttwu_wake_remote
);
1639 if (wake_flags
& WF_MIGRATED
)
1640 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1641 #endif /* CONFIG_SMP */
1643 schedstat_inc(rq
->ttwu_count
);
1644 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1646 if (wake_flags
& WF_SYNC
)
1647 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1650 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1652 activate_task(rq
, p
, en_flags
);
1653 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1655 /* If a worker is waking up, notify the workqueue: */
1656 if (p
->flags
& PF_WQ_WORKER
)
1657 wq_worker_waking_up(p
, cpu_of(rq
));
1661 * Mark the task runnable and perform wakeup-preemption.
1663 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1664 struct rq_flags
*rf
)
1666 check_preempt_curr(rq
, p
, wake_flags
);
1667 p
->state
= TASK_RUNNING
;
1668 trace_sched_wakeup(p
);
1671 if (p
->sched_class
->task_woken
) {
1673 * Our task @p is fully woken up and running; so its safe to
1674 * drop the rq->lock, hereafter rq is only used for statistics.
1676 rq_unpin_lock(rq
, rf
);
1677 p
->sched_class
->task_woken(rq
, p
);
1678 rq_repin_lock(rq
, rf
);
1681 if (rq
->idle_stamp
) {
1682 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1683 u64 max
= 2*rq
->max_idle_balance_cost
;
1685 update_avg(&rq
->avg_idle
, delta
);
1687 if (rq
->avg_idle
> max
)
1696 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1697 struct rq_flags
*rf
)
1699 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
1701 lockdep_assert_held(&rq
->lock
);
1704 if (p
->sched_contributes_to_load
)
1705 rq
->nr_uninterruptible
--;
1707 if (wake_flags
& WF_MIGRATED
)
1708 en_flags
|= ENQUEUE_MIGRATED
;
1711 ttwu_activate(rq
, p
, en_flags
);
1712 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1716 * Called in case the task @p isn't fully descheduled from its runqueue,
1717 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1718 * since all we need to do is flip p->state to TASK_RUNNING, since
1719 * the task is still ->on_rq.
1721 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1727 rq
= __task_rq_lock(p
, &rf
);
1728 if (task_on_rq_queued(p
)) {
1729 /* check_preempt_curr() may use rq clock */
1730 update_rq_clock(rq
);
1731 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1734 __task_rq_unlock(rq
, &rf
);
1740 void sched_ttwu_pending(void)
1742 struct rq
*rq
= this_rq();
1743 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1744 struct task_struct
*p
, *t
;
1750 rq_lock_irqsave(rq
, &rf
);
1751 update_rq_clock(rq
);
1753 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
1754 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
1756 rq_unlock_irqrestore(rq
, &rf
);
1759 void scheduler_ipi(void)
1762 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1763 * TIF_NEED_RESCHED remotely (for the first time) will also send
1766 preempt_fold_need_resched();
1768 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1772 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1773 * traditionally all their work was done from the interrupt return
1774 * path. Now that we actually do some work, we need to make sure
1777 * Some archs already do call them, luckily irq_enter/exit nest
1780 * Arguably we should visit all archs and update all handlers,
1781 * however a fair share of IPIs are still resched only so this would
1782 * somewhat pessimize the simple resched case.
1785 sched_ttwu_pending();
1788 * Check if someone kicked us for doing the nohz idle load balance.
1790 if (unlikely(got_nohz_idle_kick())) {
1791 this_rq()->idle_balance
= 1;
1792 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1797 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1799 struct rq
*rq
= cpu_rq(cpu
);
1801 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1803 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1804 if (!set_nr_if_polling(rq
->idle
))
1805 smp_send_reschedule(cpu
);
1807 trace_sched_wake_idle_without_ipi(cpu
);
1811 void wake_up_if_idle(int cpu
)
1813 struct rq
*rq
= cpu_rq(cpu
);
1818 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1821 if (set_nr_if_polling(rq
->idle
)) {
1822 trace_sched_wake_idle_without_ipi(cpu
);
1824 rq_lock_irqsave(rq
, &rf
);
1825 if (is_idle_task(rq
->curr
))
1826 smp_send_reschedule(cpu
);
1827 /* Else CPU is not idle, do nothing here: */
1828 rq_unlock_irqrestore(rq
, &rf
);
1835 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1837 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1839 #endif /* CONFIG_SMP */
1841 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1843 struct rq
*rq
= cpu_rq(cpu
);
1846 #if defined(CONFIG_SMP)
1847 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1848 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
1849 ttwu_queue_remote(p
, cpu
, wake_flags
);
1855 update_rq_clock(rq
);
1856 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1861 * Notes on Program-Order guarantees on SMP systems.
1865 * The basic program-order guarantee on SMP systems is that when a task [t]
1866 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1867 * execution on its new CPU [c1].
1869 * For migration (of runnable tasks) this is provided by the following means:
1871 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1872 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1873 * rq(c1)->lock (if not at the same time, then in that order).
1874 * C) LOCK of the rq(c1)->lock scheduling in task
1876 * Transitivity guarantees that B happens after A and C after B.
1877 * Note: we only require RCpc transitivity.
1878 * Note: the CPU doing B need not be c0 or c1
1887 * UNLOCK rq(0)->lock
1889 * LOCK rq(0)->lock // orders against CPU0
1891 * UNLOCK rq(0)->lock
1895 * UNLOCK rq(1)->lock
1897 * LOCK rq(1)->lock // orders against CPU2
1900 * UNLOCK rq(1)->lock
1903 * BLOCKING -- aka. SLEEP + WAKEUP
1905 * For blocking we (obviously) need to provide the same guarantee as for
1906 * migration. However the means are completely different as there is no lock
1907 * chain to provide order. Instead we do:
1909 * 1) smp_store_release(X->on_cpu, 0)
1910 * 2) smp_cond_load_acquire(!X->on_cpu)
1914 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1916 * LOCK rq(0)->lock LOCK X->pi_lock
1919 * smp_store_release(X->on_cpu, 0);
1921 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1927 * X->state = RUNNING
1928 * UNLOCK rq(2)->lock
1930 * LOCK rq(2)->lock // orders against CPU1
1933 * UNLOCK rq(2)->lock
1936 * UNLOCK rq(0)->lock
1939 * However; for wakeups there is a second guarantee we must provide, namely we
1940 * must observe the state that lead to our wakeup. That is, not only must our
1941 * task observe its own prior state, it must also observe the stores prior to
1944 * This means that any means of doing remote wakeups must order the CPU doing
1945 * the wakeup against the CPU the task is going to end up running on. This,
1946 * however, is already required for the regular Program-Order guarantee above,
1947 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1952 * try_to_wake_up - wake up a thread
1953 * @p: the thread to be awakened
1954 * @state: the mask of task states that can be woken
1955 * @wake_flags: wake modifier flags (WF_*)
1957 * If (@state & @p->state) @p->state = TASK_RUNNING.
1959 * If the task was not queued/runnable, also place it back on a runqueue.
1961 * Atomic against schedule() which would dequeue a task, also see
1962 * set_current_state().
1964 * Return: %true if @p->state changes (an actual wakeup was done),
1968 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1970 unsigned long flags
;
1971 int cpu
, success
= 0;
1974 * If we are going to wake up a thread waiting for CONDITION we
1975 * need to ensure that CONDITION=1 done by the caller can not be
1976 * reordered with p->state check below. This pairs with mb() in
1977 * set_current_state() the waiting thread does.
1979 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1980 smp_mb__after_spinlock();
1981 if (!(p
->state
& state
))
1984 trace_sched_waking(p
);
1986 /* We're going to change ->state: */
1991 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1992 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1993 * in smp_cond_load_acquire() below.
1995 * sched_ttwu_pending() try_to_wake_up()
1996 * [S] p->on_rq = 1; [L] P->state
1997 * UNLOCK rq->lock -----.
2001 * LOCK rq->lock -----'
2005 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2007 * Pairs with the UNLOCK+LOCK on rq->lock from the
2008 * last wakeup of our task and the schedule that got our task
2012 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2017 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2018 * possible to, falsely, observe p->on_cpu == 0.
2020 * One must be running (->on_cpu == 1) in order to remove oneself
2021 * from the runqueue.
2023 * [S] ->on_cpu = 1; [L] ->on_rq
2027 * [S] ->on_rq = 0; [L] ->on_cpu
2029 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2030 * from the consecutive calls to schedule(); the first switching to our
2031 * task, the second putting it to sleep.
2036 * If the owning (remote) CPU is still in the middle of schedule() with
2037 * this task as prev, wait until its done referencing the task.
2039 * Pairs with the smp_store_release() in finish_lock_switch().
2041 * This ensures that tasks getting woken will be fully ordered against
2042 * their previous state and preserve Program Order.
2044 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2046 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2047 p
->state
= TASK_WAKING
;
2050 delayacct_blkio_end();
2051 atomic_dec(&task_rq(p
)->nr_iowait
);
2054 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2055 if (task_cpu(p
) != cpu
) {
2056 wake_flags
|= WF_MIGRATED
;
2057 set_task_cpu(p
, cpu
);
2060 #else /* CONFIG_SMP */
2063 delayacct_blkio_end();
2064 atomic_dec(&task_rq(p
)->nr_iowait
);
2067 #endif /* CONFIG_SMP */
2069 ttwu_queue(p
, cpu
, wake_flags
);
2071 ttwu_stat(p
, cpu
, wake_flags
);
2073 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2079 * try_to_wake_up_local - try to wake up a local task with rq lock held
2080 * @p: the thread to be awakened
2081 * @rf: request-queue flags for pinning
2083 * Put @p on the run-queue if it's not already there. The caller must
2084 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2087 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2089 struct rq
*rq
= task_rq(p
);
2091 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2092 WARN_ON_ONCE(p
== current
))
2095 lockdep_assert_held(&rq
->lock
);
2097 if (!raw_spin_trylock(&p
->pi_lock
)) {
2099 * This is OK, because current is on_cpu, which avoids it being
2100 * picked for load-balance and preemption/IRQs are still
2101 * disabled avoiding further scheduler activity on it and we've
2102 * not yet picked a replacement task.
2105 raw_spin_lock(&p
->pi_lock
);
2109 if (!(p
->state
& TASK_NORMAL
))
2112 trace_sched_waking(p
);
2114 if (!task_on_rq_queued(p
)) {
2116 delayacct_blkio_end();
2117 atomic_dec(&rq
->nr_iowait
);
2119 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
);
2122 ttwu_do_wakeup(rq
, p
, 0, rf
);
2123 ttwu_stat(p
, smp_processor_id(), 0);
2125 raw_spin_unlock(&p
->pi_lock
);
2129 * wake_up_process - Wake up a specific process
2130 * @p: The process to be woken up.
2132 * Attempt to wake up the nominated process and move it to the set of runnable
2135 * Return: 1 if the process was woken up, 0 if it was already running.
2137 * It may be assumed that this function implies a write memory barrier before
2138 * changing the task state if and only if any tasks are woken up.
2140 int wake_up_process(struct task_struct
*p
)
2142 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2144 EXPORT_SYMBOL(wake_up_process
);
2146 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2148 return try_to_wake_up(p
, state
, 0);
2152 * Perform scheduler related setup for a newly forked process p.
2153 * p is forked by current.
2155 * __sched_fork() is basic setup used by init_idle() too:
2157 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2162 p
->se
.exec_start
= 0;
2163 p
->se
.sum_exec_runtime
= 0;
2164 p
->se
.prev_sum_exec_runtime
= 0;
2165 p
->se
.nr_migrations
= 0;
2167 INIT_LIST_HEAD(&p
->se
.group_node
);
2169 #ifdef CONFIG_FAIR_GROUP_SCHED
2170 p
->se
.cfs_rq
= NULL
;
2173 #ifdef CONFIG_SCHEDSTATS
2174 /* Even if schedstat is disabled, there should not be garbage */
2175 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2178 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2179 init_dl_task_timer(&p
->dl
);
2180 init_dl_inactive_task_timer(&p
->dl
);
2181 __dl_clear_params(p
);
2183 INIT_LIST_HEAD(&p
->rt
.run_list
);
2185 p
->rt
.time_slice
= sched_rr_timeslice
;
2189 #ifdef CONFIG_PREEMPT_NOTIFIERS
2190 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2193 #ifdef CONFIG_NUMA_BALANCING
2194 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2195 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2196 p
->mm
->numa_scan_seq
= 0;
2199 if (clone_flags
& CLONE_VM
)
2200 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2202 p
->numa_preferred_nid
= -1;
2204 p
->node_stamp
= 0ULL;
2205 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2206 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2207 p
->numa_work
.next
= &p
->numa_work
;
2208 p
->numa_faults
= NULL
;
2209 p
->last_task_numa_placement
= 0;
2210 p
->last_sum_exec_runtime
= 0;
2212 p
->numa_group
= NULL
;
2213 #endif /* CONFIG_NUMA_BALANCING */
2216 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2218 #ifdef CONFIG_NUMA_BALANCING
2220 void set_numabalancing_state(bool enabled
)
2223 static_branch_enable(&sched_numa_balancing
);
2225 static_branch_disable(&sched_numa_balancing
);
2228 #ifdef CONFIG_PROC_SYSCTL
2229 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2230 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2234 int state
= static_branch_likely(&sched_numa_balancing
);
2236 if (write
&& !capable(CAP_SYS_ADMIN
))
2241 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2245 set_numabalancing_state(state
);
2251 #ifdef CONFIG_SCHEDSTATS
2253 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2254 static bool __initdata __sched_schedstats
= false;
2256 static void set_schedstats(bool enabled
)
2259 static_branch_enable(&sched_schedstats
);
2261 static_branch_disable(&sched_schedstats
);
2264 void force_schedstat_enabled(void)
2266 if (!schedstat_enabled()) {
2267 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2268 static_branch_enable(&sched_schedstats
);
2272 static int __init
setup_schedstats(char *str
)
2279 * This code is called before jump labels have been set up, so we can't
2280 * change the static branch directly just yet. Instead set a temporary
2281 * variable so init_schedstats() can do it later.
2283 if (!strcmp(str
, "enable")) {
2284 __sched_schedstats
= true;
2286 } else if (!strcmp(str
, "disable")) {
2287 __sched_schedstats
= false;
2292 pr_warn("Unable to parse schedstats=\n");
2296 __setup("schedstats=", setup_schedstats
);
2298 static void __init
init_schedstats(void)
2300 set_schedstats(__sched_schedstats
);
2303 #ifdef CONFIG_PROC_SYSCTL
2304 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2305 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2309 int state
= static_branch_likely(&sched_schedstats
);
2311 if (write
&& !capable(CAP_SYS_ADMIN
))
2316 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2320 set_schedstats(state
);
2323 #endif /* CONFIG_PROC_SYSCTL */
2324 #else /* !CONFIG_SCHEDSTATS */
2325 static inline void init_schedstats(void) {}
2326 #endif /* CONFIG_SCHEDSTATS */
2329 * fork()/clone()-time setup:
2331 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2333 unsigned long flags
;
2334 int cpu
= get_cpu();
2336 __sched_fork(clone_flags
, p
);
2338 * We mark the process as NEW here. This guarantees that
2339 * nobody will actually run it, and a signal or other external
2340 * event cannot wake it up and insert it on the runqueue either.
2342 p
->state
= TASK_NEW
;
2345 * Make sure we do not leak PI boosting priority to the child.
2347 p
->prio
= current
->normal_prio
;
2350 * Revert to default priority/policy on fork if requested.
2352 if (unlikely(p
->sched_reset_on_fork
)) {
2353 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2354 p
->policy
= SCHED_NORMAL
;
2355 p
->static_prio
= NICE_TO_PRIO(0);
2357 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2358 p
->static_prio
= NICE_TO_PRIO(0);
2360 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2364 * We don't need the reset flag anymore after the fork. It has
2365 * fulfilled its duty:
2367 p
->sched_reset_on_fork
= 0;
2370 if (dl_prio(p
->prio
)) {
2373 } else if (rt_prio(p
->prio
)) {
2374 p
->sched_class
= &rt_sched_class
;
2376 p
->sched_class
= &fair_sched_class
;
2379 init_entity_runnable_average(&p
->se
);
2382 * The child is not yet in the pid-hash so no cgroup attach races,
2383 * and the cgroup is pinned to this child due to cgroup_fork()
2384 * is ran before sched_fork().
2386 * Silence PROVE_RCU.
2388 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2390 * We're setting the CPU for the first time, we don't migrate,
2391 * so use __set_task_cpu().
2393 __set_task_cpu(p
, cpu
);
2394 if (p
->sched_class
->task_fork
)
2395 p
->sched_class
->task_fork(p
);
2396 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2398 #ifdef CONFIG_SCHED_INFO
2399 if (likely(sched_info_on()))
2400 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2402 #if defined(CONFIG_SMP)
2405 init_task_preempt_count(p
);
2407 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2408 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2415 unsigned long to_ratio(u64 period
, u64 runtime
)
2417 if (runtime
== RUNTIME_INF
)
2421 * Doing this here saves a lot of checks in all
2422 * the calling paths, and returning zero seems
2423 * safe for them anyway.
2428 return div64_u64(runtime
<< BW_SHIFT
, period
);
2432 * wake_up_new_task - wake up a newly created task for the first time.
2434 * This function will do some initial scheduler statistics housekeeping
2435 * that must be done for every newly created context, then puts the task
2436 * on the runqueue and wakes it.
2438 void wake_up_new_task(struct task_struct
*p
)
2443 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2444 p
->state
= TASK_RUNNING
;
2447 * Fork balancing, do it here and not earlier because:
2448 * - cpus_allowed can change in the fork path
2449 * - any previously selected CPU might disappear through hotplug
2451 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2452 * as we're not fully set-up yet.
2454 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2456 rq
= __task_rq_lock(p
, &rf
);
2457 update_rq_clock(rq
);
2458 post_init_entity_util_avg(&p
->se
);
2460 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2461 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2462 trace_sched_wakeup_new(p
);
2463 check_preempt_curr(rq
, p
, WF_FORK
);
2465 if (p
->sched_class
->task_woken
) {
2467 * Nothing relies on rq->lock after this, so its fine to
2470 rq_unpin_lock(rq
, &rf
);
2471 p
->sched_class
->task_woken(rq
, p
);
2472 rq_repin_lock(rq
, &rf
);
2475 task_rq_unlock(rq
, p
, &rf
);
2478 #ifdef CONFIG_PREEMPT_NOTIFIERS
2480 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2482 void preempt_notifier_inc(void)
2484 static_key_slow_inc(&preempt_notifier_key
);
2486 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2488 void preempt_notifier_dec(void)
2490 static_key_slow_dec(&preempt_notifier_key
);
2492 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2495 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2496 * @notifier: notifier struct to register
2498 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2500 if (!static_key_false(&preempt_notifier_key
))
2501 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2503 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2505 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2508 * preempt_notifier_unregister - no longer interested in preemption notifications
2509 * @notifier: notifier struct to unregister
2511 * This is *not* safe to call from within a preemption notifier.
2513 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2515 hlist_del(¬ifier
->link
);
2517 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2519 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2521 struct preempt_notifier
*notifier
;
2523 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2524 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2527 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2529 if (static_key_false(&preempt_notifier_key
))
2530 __fire_sched_in_preempt_notifiers(curr
);
2534 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2535 struct task_struct
*next
)
2537 struct preempt_notifier
*notifier
;
2539 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2540 notifier
->ops
->sched_out(notifier
, next
);
2543 static __always_inline
void
2544 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2545 struct task_struct
*next
)
2547 if (static_key_false(&preempt_notifier_key
))
2548 __fire_sched_out_preempt_notifiers(curr
, next
);
2551 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2553 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2558 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2559 struct task_struct
*next
)
2563 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2566 * prepare_task_switch - prepare to switch tasks
2567 * @rq: the runqueue preparing to switch
2568 * @prev: the current task that is being switched out
2569 * @next: the task we are going to switch to.
2571 * This is called with the rq lock held and interrupts off. It must
2572 * be paired with a subsequent finish_task_switch after the context
2575 * prepare_task_switch sets up locking and calls architecture specific
2579 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2580 struct task_struct
*next
)
2582 sched_info_switch(rq
, prev
, next
);
2583 perf_event_task_sched_out(prev
, next
);
2584 fire_sched_out_preempt_notifiers(prev
, next
);
2585 prepare_lock_switch(rq
, next
);
2586 prepare_arch_switch(next
);
2590 * finish_task_switch - clean up after a task-switch
2591 * @prev: the thread we just switched away from.
2593 * finish_task_switch must be called after the context switch, paired
2594 * with a prepare_task_switch call before the context switch.
2595 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2596 * and do any other architecture-specific cleanup actions.
2598 * Note that we may have delayed dropping an mm in context_switch(). If
2599 * so, we finish that here outside of the runqueue lock. (Doing it
2600 * with the lock held can cause deadlocks; see schedule() for
2603 * The context switch have flipped the stack from under us and restored the
2604 * local variables which were saved when this task called schedule() in the
2605 * past. prev == current is still correct but we need to recalculate this_rq
2606 * because prev may have moved to another CPU.
2608 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2609 __releases(rq
->lock
)
2611 struct rq
*rq
= this_rq();
2612 struct mm_struct
*mm
= rq
->prev_mm
;
2616 * The previous task will have left us with a preempt_count of 2
2617 * because it left us after:
2620 * preempt_disable(); // 1
2622 * raw_spin_lock_irq(&rq->lock) // 2
2624 * Also, see FORK_PREEMPT_COUNT.
2626 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2627 "corrupted preempt_count: %s/%d/0x%x\n",
2628 current
->comm
, current
->pid
, preempt_count()))
2629 preempt_count_set(FORK_PREEMPT_COUNT
);
2634 * A task struct has one reference for the use as "current".
2635 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2636 * schedule one last time. The schedule call will never return, and
2637 * the scheduled task must drop that reference.
2639 * We must observe prev->state before clearing prev->on_cpu (in
2640 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2641 * running on another CPU and we could rave with its RUNNING -> DEAD
2642 * transition, resulting in a double drop.
2644 prev_state
= prev
->state
;
2645 vtime_task_switch(prev
);
2646 perf_event_task_sched_in(prev
, current
);
2648 * The membarrier system call requires a full memory barrier
2649 * after storing to rq->curr, before going back to user-space.
2651 * TODO: This smp_mb__after_unlock_lock can go away if PPC end
2652 * up adding a full barrier to switch_mm(), or we should figure
2653 * out if a smp_mb__after_unlock_lock is really the proper API
2656 smp_mb__after_unlock_lock();
2657 finish_lock_switch(rq
, prev
);
2658 finish_arch_post_lock_switch();
2660 fire_sched_in_preempt_notifiers(current
);
2663 if (unlikely(prev_state
== TASK_DEAD
)) {
2664 if (prev
->sched_class
->task_dead
)
2665 prev
->sched_class
->task_dead(prev
);
2668 * Remove function-return probe instances associated with this
2669 * task and put them back on the free list.
2671 kprobe_flush_task(prev
);
2673 /* Task is done with its stack. */
2674 put_task_stack(prev
);
2676 put_task_struct(prev
);
2679 tick_nohz_task_switch();
2685 /* rq->lock is NOT held, but preemption is disabled */
2686 static void __balance_callback(struct rq
*rq
)
2688 struct callback_head
*head
, *next
;
2689 void (*func
)(struct rq
*rq
);
2690 unsigned long flags
;
2692 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2693 head
= rq
->balance_callback
;
2694 rq
->balance_callback
= NULL
;
2696 func
= (void (*)(struct rq
*))head
->func
;
2703 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2706 static inline void balance_callback(struct rq
*rq
)
2708 if (unlikely(rq
->balance_callback
))
2709 __balance_callback(rq
);
2714 static inline void balance_callback(struct rq
*rq
)
2721 * schedule_tail - first thing a freshly forked thread must call.
2722 * @prev: the thread we just switched away from.
2724 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2725 __releases(rq
->lock
)
2730 * New tasks start with FORK_PREEMPT_COUNT, see there and
2731 * finish_task_switch() for details.
2733 * finish_task_switch() will drop rq->lock() and lower preempt_count
2734 * and the preempt_enable() will end up enabling preemption (on
2735 * PREEMPT_COUNT kernels).
2738 rq
= finish_task_switch(prev
);
2739 balance_callback(rq
);
2742 if (current
->set_child_tid
)
2743 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2747 * context_switch - switch to the new MM and the new thread's register state.
2749 static __always_inline
struct rq
*
2750 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2751 struct task_struct
*next
, struct rq_flags
*rf
)
2753 struct mm_struct
*mm
, *oldmm
;
2755 prepare_task_switch(rq
, prev
, next
);
2758 oldmm
= prev
->active_mm
;
2760 * For paravirt, this is coupled with an exit in switch_to to
2761 * combine the page table reload and the switch backend into
2764 arch_start_context_switch(prev
);
2767 next
->active_mm
= oldmm
;
2769 enter_lazy_tlb(oldmm
, next
);
2771 switch_mm_irqs_off(oldmm
, mm
, next
);
2774 prev
->active_mm
= NULL
;
2775 rq
->prev_mm
= oldmm
;
2778 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
2781 * Since the runqueue lock will be released by the next
2782 * task (which is an invalid locking op but in the case
2783 * of the scheduler it's an obvious special-case), so we
2784 * do an early lockdep release here:
2786 rq_unpin_lock(rq
, rf
);
2787 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2789 /* Here we just switch the register state and the stack. */
2790 switch_to(prev
, next
, prev
);
2793 return finish_task_switch(prev
);
2797 * nr_running and nr_context_switches:
2799 * externally visible scheduler statistics: current number of runnable
2800 * threads, total number of context switches performed since bootup.
2802 unsigned long nr_running(void)
2804 unsigned long i
, sum
= 0;
2806 for_each_online_cpu(i
)
2807 sum
+= cpu_rq(i
)->nr_running
;
2813 * Check if only the current task is running on the CPU.
2815 * Caution: this function does not check that the caller has disabled
2816 * preemption, thus the result might have a time-of-check-to-time-of-use
2817 * race. The caller is responsible to use it correctly, for example:
2819 * - from a non-preemptable section (of course)
2821 * - from a thread that is bound to a single CPU
2823 * - in a loop with very short iterations (e.g. a polling loop)
2825 bool single_task_running(void)
2827 return raw_rq()->nr_running
== 1;
2829 EXPORT_SYMBOL(single_task_running
);
2831 unsigned long long nr_context_switches(void)
2834 unsigned long long sum
= 0;
2836 for_each_possible_cpu(i
)
2837 sum
+= cpu_rq(i
)->nr_switches
;
2843 * IO-wait accounting, and how its mostly bollocks (on SMP).
2845 * The idea behind IO-wait account is to account the idle time that we could
2846 * have spend running if it were not for IO. That is, if we were to improve the
2847 * storage performance, we'd have a proportional reduction in IO-wait time.
2849 * This all works nicely on UP, where, when a task blocks on IO, we account
2850 * idle time as IO-wait, because if the storage were faster, it could've been
2851 * running and we'd not be idle.
2853 * This has been extended to SMP, by doing the same for each CPU. This however
2856 * Imagine for instance the case where two tasks block on one CPU, only the one
2857 * CPU will have IO-wait accounted, while the other has regular idle. Even
2858 * though, if the storage were faster, both could've ran at the same time,
2859 * utilising both CPUs.
2861 * This means, that when looking globally, the current IO-wait accounting on
2862 * SMP is a lower bound, by reason of under accounting.
2864 * Worse, since the numbers are provided per CPU, they are sometimes
2865 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2866 * associated with any one particular CPU, it can wake to another CPU than it
2867 * blocked on. This means the per CPU IO-wait number is meaningless.
2869 * Task CPU affinities can make all that even more 'interesting'.
2872 unsigned long nr_iowait(void)
2874 unsigned long i
, sum
= 0;
2876 for_each_possible_cpu(i
)
2877 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2883 * Consumers of these two interfaces, like for example the cpufreq menu
2884 * governor are using nonsensical data. Boosting frequency for a CPU that has
2885 * IO-wait which might not even end up running the task when it does become
2889 unsigned long nr_iowait_cpu(int cpu
)
2891 struct rq
*this = cpu_rq(cpu
);
2892 return atomic_read(&this->nr_iowait
);
2895 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2897 struct rq
*rq
= this_rq();
2898 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2899 *load
= rq
->load
.weight
;
2905 * sched_exec - execve() is a valuable balancing opportunity, because at
2906 * this point the task has the smallest effective memory and cache footprint.
2908 void sched_exec(void)
2910 struct task_struct
*p
= current
;
2911 unsigned long flags
;
2914 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2915 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2916 if (dest_cpu
== smp_processor_id())
2919 if (likely(cpu_active(dest_cpu
))) {
2920 struct migration_arg arg
= { p
, dest_cpu
};
2922 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2923 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2927 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2932 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2933 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2935 EXPORT_PER_CPU_SYMBOL(kstat
);
2936 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2939 * The function fair_sched_class.update_curr accesses the struct curr
2940 * and its field curr->exec_start; when called from task_sched_runtime(),
2941 * we observe a high rate of cache misses in practice.
2942 * Prefetching this data results in improved performance.
2944 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
2946 #ifdef CONFIG_FAIR_GROUP_SCHED
2947 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
2949 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
2952 prefetch(&curr
->exec_start
);
2956 * Return accounted runtime for the task.
2957 * In case the task is currently running, return the runtime plus current's
2958 * pending runtime that have not been accounted yet.
2960 unsigned long long task_sched_runtime(struct task_struct
*p
)
2966 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2968 * 64-bit doesn't need locks to atomically read a 64bit value.
2969 * So we have a optimization chance when the task's delta_exec is 0.
2970 * Reading ->on_cpu is racy, but this is ok.
2972 * If we race with it leaving CPU, we'll take a lock. So we're correct.
2973 * If we race with it entering CPU, unaccounted time is 0. This is
2974 * indistinguishable from the read occurring a few cycles earlier.
2975 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2976 * been accounted, so we're correct here as well.
2978 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2979 return p
->se
.sum_exec_runtime
;
2982 rq
= task_rq_lock(p
, &rf
);
2984 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2985 * project cycles that may never be accounted to this
2986 * thread, breaking clock_gettime().
2988 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2989 prefetch_curr_exec_start(p
);
2990 update_rq_clock(rq
);
2991 p
->sched_class
->update_curr(rq
);
2993 ns
= p
->se
.sum_exec_runtime
;
2994 task_rq_unlock(rq
, p
, &rf
);
3000 * This function gets called by the timer code, with HZ frequency.
3001 * We call it with interrupts disabled.
3003 void scheduler_tick(void)
3005 int cpu
= smp_processor_id();
3006 struct rq
*rq
= cpu_rq(cpu
);
3007 struct task_struct
*curr
= rq
->curr
;
3014 update_rq_clock(rq
);
3015 curr
->sched_class
->task_tick(rq
, curr
, 0);
3016 cpu_load_update_active(rq
);
3017 calc_global_load_tick(rq
);
3021 perf_event_task_tick();
3024 rq
->idle_balance
= idle_cpu(cpu
);
3025 trigger_load_balance(rq
);
3027 rq_last_tick_reset(rq
);
3030 #ifdef CONFIG_NO_HZ_FULL
3032 * scheduler_tick_max_deferment
3034 * Keep at least one tick per second when a single
3035 * active task is running because the scheduler doesn't
3036 * yet completely support full dynticks environment.
3038 * This makes sure that uptime, CFS vruntime, load
3039 * balancing, etc... continue to move forward, even
3040 * with a very low granularity.
3042 * Return: Maximum deferment in nanoseconds.
3044 u64
scheduler_tick_max_deferment(void)
3046 struct rq
*rq
= this_rq();
3047 unsigned long next
, now
= READ_ONCE(jiffies
);
3049 next
= rq
->last_sched_tick
+ HZ
;
3051 if (time_before_eq(next
, now
))
3054 return jiffies_to_nsecs(next
- now
);
3058 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3059 defined(CONFIG_PREEMPT_TRACER))
3061 * If the value passed in is equal to the current preempt count
3062 * then we just disabled preemption. Start timing the latency.
3064 static inline void preempt_latency_start(int val
)
3066 if (preempt_count() == val
) {
3067 unsigned long ip
= get_lock_parent_ip();
3068 #ifdef CONFIG_DEBUG_PREEMPT
3069 current
->preempt_disable_ip
= ip
;
3071 trace_preempt_off(CALLER_ADDR0
, ip
);
3075 void preempt_count_add(int val
)
3077 #ifdef CONFIG_DEBUG_PREEMPT
3081 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3084 __preempt_count_add(val
);
3085 #ifdef CONFIG_DEBUG_PREEMPT
3087 * Spinlock count overflowing soon?
3089 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3092 preempt_latency_start(val
);
3094 EXPORT_SYMBOL(preempt_count_add
);
3095 NOKPROBE_SYMBOL(preempt_count_add
);
3098 * If the value passed in equals to the current preempt count
3099 * then we just enabled preemption. Stop timing the latency.
3101 static inline void preempt_latency_stop(int val
)
3103 if (preempt_count() == val
)
3104 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3107 void preempt_count_sub(int val
)
3109 #ifdef CONFIG_DEBUG_PREEMPT
3113 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3116 * Is the spinlock portion underflowing?
3118 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3119 !(preempt_count() & PREEMPT_MASK
)))
3123 preempt_latency_stop(val
);
3124 __preempt_count_sub(val
);
3126 EXPORT_SYMBOL(preempt_count_sub
);
3127 NOKPROBE_SYMBOL(preempt_count_sub
);
3130 static inline void preempt_latency_start(int val
) { }
3131 static inline void preempt_latency_stop(int val
) { }
3134 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3136 #ifdef CONFIG_DEBUG_PREEMPT
3137 return p
->preempt_disable_ip
;
3144 * Print scheduling while atomic bug:
3146 static noinline
void __schedule_bug(struct task_struct
*prev
)
3148 /* Save this before calling printk(), since that will clobber it */
3149 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3151 if (oops_in_progress
)
3154 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3155 prev
->comm
, prev
->pid
, preempt_count());
3157 debug_show_held_locks(prev
);
3159 if (irqs_disabled())
3160 print_irqtrace_events(prev
);
3161 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3162 && in_atomic_preempt_off()) {
3163 pr_err("Preemption disabled at:");
3164 print_ip_sym(preempt_disable_ip
);
3168 panic("scheduling while atomic\n");
3171 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3175 * Various schedule()-time debugging checks and statistics:
3177 static inline void schedule_debug(struct task_struct
*prev
)
3179 #ifdef CONFIG_SCHED_STACK_END_CHECK
3180 if (task_stack_end_corrupted(prev
))
3181 panic("corrupted stack end detected inside scheduler\n");
3184 if (unlikely(in_atomic_preempt_off())) {
3185 __schedule_bug(prev
);
3186 preempt_count_set(PREEMPT_DISABLED
);
3190 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3192 schedstat_inc(this_rq()->sched_count
);
3196 * Pick up the highest-prio task:
3198 static inline struct task_struct
*
3199 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3201 const struct sched_class
*class;
3202 struct task_struct
*p
;
3205 * Optimization: we know that if all tasks are in the fair class we can
3206 * call that function directly, but only if the @prev task wasn't of a
3207 * higher scheduling class, because otherwise those loose the
3208 * opportunity to pull in more work from other CPUs.
3210 if (likely((prev
->sched_class
== &idle_sched_class
||
3211 prev
->sched_class
== &fair_sched_class
) &&
3212 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3214 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3215 if (unlikely(p
== RETRY_TASK
))
3218 /* Assumes fair_sched_class->next == idle_sched_class */
3220 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3226 for_each_class(class) {
3227 p
= class->pick_next_task(rq
, prev
, rf
);
3229 if (unlikely(p
== RETRY_TASK
))
3235 /* The idle class should always have a runnable task: */
3240 * __schedule() is the main scheduler function.
3242 * The main means of driving the scheduler and thus entering this function are:
3244 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3246 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3247 * paths. For example, see arch/x86/entry_64.S.
3249 * To drive preemption between tasks, the scheduler sets the flag in timer
3250 * interrupt handler scheduler_tick().
3252 * 3. Wakeups don't really cause entry into schedule(). They add a
3253 * task to the run-queue and that's it.
3255 * Now, if the new task added to the run-queue preempts the current
3256 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3257 * called on the nearest possible occasion:
3259 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3261 * - in syscall or exception context, at the next outmost
3262 * preempt_enable(). (this might be as soon as the wake_up()'s
3265 * - in IRQ context, return from interrupt-handler to
3266 * preemptible context
3268 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3271 * - cond_resched() call
3272 * - explicit schedule() call
3273 * - return from syscall or exception to user-space
3274 * - return from interrupt-handler to user-space
3276 * WARNING: must be called with preemption disabled!
3278 static void __sched notrace
__schedule(bool preempt
)
3280 struct task_struct
*prev
, *next
;
3281 unsigned long *switch_count
;
3286 cpu
= smp_processor_id();
3290 schedule_debug(prev
);
3292 if (sched_feat(HRTICK
))
3295 local_irq_disable();
3296 rcu_note_context_switch(preempt
);
3299 * Make sure that signal_pending_state()->signal_pending() below
3300 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3301 * done by the caller to avoid the race with signal_wake_up().
3304 smp_mb__after_spinlock();
3306 /* Promote REQ to ACT */
3307 rq
->clock_update_flags
<<= 1;
3308 update_rq_clock(rq
);
3310 switch_count
= &prev
->nivcsw
;
3311 if (!preempt
&& prev
->state
) {
3312 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3313 prev
->state
= TASK_RUNNING
;
3315 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
3318 if (prev
->in_iowait
) {
3319 atomic_inc(&rq
->nr_iowait
);
3320 delayacct_blkio_start();
3324 * If a worker went to sleep, notify and ask workqueue
3325 * whether it wants to wake up a task to maintain
3328 if (prev
->flags
& PF_WQ_WORKER
) {
3329 struct task_struct
*to_wakeup
;
3331 to_wakeup
= wq_worker_sleeping(prev
);
3333 try_to_wake_up_local(to_wakeup
, &rf
);
3336 switch_count
= &prev
->nvcsw
;
3339 next
= pick_next_task(rq
, prev
, &rf
);
3340 clear_tsk_need_resched(prev
);
3341 clear_preempt_need_resched();
3343 if (likely(prev
!= next
)) {
3347 * The membarrier system call requires each architecture
3348 * to have a full memory barrier after updating
3349 * rq->curr, before returning to user-space. For TSO
3350 * (e.g. x86), the architecture must provide its own
3351 * barrier in switch_mm(). For weakly ordered machines
3352 * for which spin_unlock() acts as a full memory
3353 * barrier, finish_lock_switch() in common code takes
3354 * care of this barrier. For weakly ordered machines for
3355 * which spin_unlock() acts as a RELEASE barrier (only
3356 * arm64 and PowerPC), arm64 has a full barrier in
3357 * switch_to(), and PowerPC has
3358 * smp_mb__after_unlock_lock() before
3359 * finish_lock_switch().
3363 trace_sched_switch(preempt
, prev
, next
);
3365 /* Also unlocks the rq: */
3366 rq
= context_switch(rq
, prev
, next
, &rf
);
3368 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3369 rq_unlock_irq(rq
, &rf
);
3372 balance_callback(rq
);
3375 void __noreturn
do_task_dead(void)
3378 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3379 * when the following two conditions become true.
3380 * - There is race condition of mmap_sem (It is acquired by
3382 * - SMI occurs before setting TASK_RUNINNG.
3383 * (or hypervisor of virtual machine switches to other guest)
3384 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3386 * To avoid it, we have to wait for releasing tsk->pi_lock which
3387 * is held by try_to_wake_up()
3389 raw_spin_lock_irq(¤t
->pi_lock
);
3390 raw_spin_unlock_irq(¤t
->pi_lock
);
3392 /* Causes final put_task_struct in finish_task_switch(): */
3393 __set_current_state(TASK_DEAD
);
3395 /* Tell freezer to ignore us: */
3396 current
->flags
|= PF_NOFREEZE
;
3401 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3406 static inline void sched_submit_work(struct task_struct
*tsk
)
3408 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3411 * If we are going to sleep and we have plugged IO queued,
3412 * make sure to submit it to avoid deadlocks.
3414 if (blk_needs_flush_plug(tsk
))
3415 blk_schedule_flush_plug(tsk
);
3418 asmlinkage __visible
void __sched
schedule(void)
3420 struct task_struct
*tsk
= current
;
3422 sched_submit_work(tsk
);
3426 sched_preempt_enable_no_resched();
3427 } while (need_resched());
3429 EXPORT_SYMBOL(schedule
);
3432 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3433 * state (have scheduled out non-voluntarily) by making sure that all
3434 * tasks have either left the run queue or have gone into user space.
3435 * As idle tasks do not do either, they must not ever be preempted
3436 * (schedule out non-voluntarily).
3438 * schedule_idle() is similar to schedule_preempt_disable() except that it
3439 * never enables preemption because it does not call sched_submit_work().
3441 void __sched
schedule_idle(void)
3444 * As this skips calling sched_submit_work(), which the idle task does
3445 * regardless because that function is a nop when the task is in a
3446 * TASK_RUNNING state, make sure this isn't used someplace that the
3447 * current task can be in any other state. Note, idle is always in the
3448 * TASK_RUNNING state.
3450 WARN_ON_ONCE(current
->state
);
3453 } while (need_resched());
3456 #ifdef CONFIG_CONTEXT_TRACKING
3457 asmlinkage __visible
void __sched
schedule_user(void)
3460 * If we come here after a random call to set_need_resched(),
3461 * or we have been woken up remotely but the IPI has not yet arrived,
3462 * we haven't yet exited the RCU idle mode. Do it here manually until
3463 * we find a better solution.
3465 * NB: There are buggy callers of this function. Ideally we
3466 * should warn if prev_state != CONTEXT_USER, but that will trigger
3467 * too frequently to make sense yet.
3469 enum ctx_state prev_state
= exception_enter();
3471 exception_exit(prev_state
);
3476 * schedule_preempt_disabled - called with preemption disabled
3478 * Returns with preemption disabled. Note: preempt_count must be 1
3480 void __sched
schedule_preempt_disabled(void)
3482 sched_preempt_enable_no_resched();
3487 static void __sched notrace
preempt_schedule_common(void)
3491 * Because the function tracer can trace preempt_count_sub()
3492 * and it also uses preempt_enable/disable_notrace(), if
3493 * NEED_RESCHED is set, the preempt_enable_notrace() called
3494 * by the function tracer will call this function again and
3495 * cause infinite recursion.
3497 * Preemption must be disabled here before the function
3498 * tracer can trace. Break up preempt_disable() into two
3499 * calls. One to disable preemption without fear of being
3500 * traced. The other to still record the preemption latency,
3501 * which can also be traced by the function tracer.
3503 preempt_disable_notrace();
3504 preempt_latency_start(1);
3506 preempt_latency_stop(1);
3507 preempt_enable_no_resched_notrace();
3510 * Check again in case we missed a preemption opportunity
3511 * between schedule and now.
3513 } while (need_resched());
3516 #ifdef CONFIG_PREEMPT
3518 * this is the entry point to schedule() from in-kernel preemption
3519 * off of preempt_enable. Kernel preemptions off return from interrupt
3520 * occur there and call schedule directly.
3522 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3525 * If there is a non-zero preempt_count or interrupts are disabled,
3526 * we do not want to preempt the current task. Just return..
3528 if (likely(!preemptible()))
3531 preempt_schedule_common();
3533 NOKPROBE_SYMBOL(preempt_schedule
);
3534 EXPORT_SYMBOL(preempt_schedule
);
3537 * preempt_schedule_notrace - preempt_schedule called by tracing
3539 * The tracing infrastructure uses preempt_enable_notrace to prevent
3540 * recursion and tracing preempt enabling caused by the tracing
3541 * infrastructure itself. But as tracing can happen in areas coming
3542 * from userspace or just about to enter userspace, a preempt enable
3543 * can occur before user_exit() is called. This will cause the scheduler
3544 * to be called when the system is still in usermode.
3546 * To prevent this, the preempt_enable_notrace will use this function
3547 * instead of preempt_schedule() to exit user context if needed before
3548 * calling the scheduler.
3550 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3552 enum ctx_state prev_ctx
;
3554 if (likely(!preemptible()))
3559 * Because the function tracer can trace preempt_count_sub()
3560 * and it also uses preempt_enable/disable_notrace(), if
3561 * NEED_RESCHED is set, the preempt_enable_notrace() called
3562 * by the function tracer will call this function again and
3563 * cause infinite recursion.
3565 * Preemption must be disabled here before the function
3566 * tracer can trace. Break up preempt_disable() into two
3567 * calls. One to disable preemption without fear of being
3568 * traced. The other to still record the preemption latency,
3569 * which can also be traced by the function tracer.
3571 preempt_disable_notrace();
3572 preempt_latency_start(1);
3574 * Needs preempt disabled in case user_exit() is traced
3575 * and the tracer calls preempt_enable_notrace() causing
3576 * an infinite recursion.
3578 prev_ctx
= exception_enter();
3580 exception_exit(prev_ctx
);
3582 preempt_latency_stop(1);
3583 preempt_enable_no_resched_notrace();
3584 } while (need_resched());
3586 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3588 #endif /* CONFIG_PREEMPT */
3591 * this is the entry point to schedule() from kernel preemption
3592 * off of irq context.
3593 * Note, that this is called and return with irqs disabled. This will
3594 * protect us against recursive calling from irq.
3596 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3598 enum ctx_state prev_state
;
3600 /* Catch callers which need to be fixed */
3601 BUG_ON(preempt_count() || !irqs_disabled());
3603 prev_state
= exception_enter();
3609 local_irq_disable();
3610 sched_preempt_enable_no_resched();
3611 } while (need_resched());
3613 exception_exit(prev_state
);
3616 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
3619 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3621 EXPORT_SYMBOL(default_wake_function
);
3623 #ifdef CONFIG_RT_MUTEXES
3625 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
3628 prio
= min(prio
, pi_task
->prio
);
3633 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3635 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3637 return __rt_effective_prio(pi_task
, prio
);
3641 * rt_mutex_setprio - set the current priority of a task
3643 * @pi_task: donor task
3645 * This function changes the 'effective' priority of a task. It does
3646 * not touch ->normal_prio like __setscheduler().
3648 * Used by the rt_mutex code to implement priority inheritance
3649 * logic. Call site only calls if the priority of the task changed.
3651 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
3653 int prio
, oldprio
, queued
, running
, queue_flag
=
3654 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
3655 const struct sched_class
*prev_class
;
3659 /* XXX used to be waiter->prio, not waiter->task->prio */
3660 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
3663 * If nothing changed; bail early.
3665 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
3668 rq
= __task_rq_lock(p
, &rf
);
3669 update_rq_clock(rq
);
3671 * Set under pi_lock && rq->lock, such that the value can be used under
3674 * Note that there is loads of tricky to make this pointer cache work
3675 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3676 * ensure a task is de-boosted (pi_task is set to NULL) before the
3677 * task is allowed to run again (and can exit). This ensures the pointer
3678 * points to a blocked task -- which guaratees the task is present.
3680 p
->pi_top_task
= pi_task
;
3683 * For FIFO/RR we only need to set prio, if that matches we're done.
3685 if (prio
== p
->prio
&& !dl_prio(prio
))
3689 * Idle task boosting is a nono in general. There is one
3690 * exception, when PREEMPT_RT and NOHZ is active:
3692 * The idle task calls get_next_timer_interrupt() and holds
3693 * the timer wheel base->lock on the CPU and another CPU wants
3694 * to access the timer (probably to cancel it). We can safely
3695 * ignore the boosting request, as the idle CPU runs this code
3696 * with interrupts disabled and will complete the lock
3697 * protected section without being interrupted. So there is no
3698 * real need to boost.
3700 if (unlikely(p
== rq
->idle
)) {
3701 WARN_ON(p
!= rq
->curr
);
3702 WARN_ON(p
->pi_blocked_on
);
3706 trace_sched_pi_setprio(p
, pi_task
);
3709 if (oldprio
== prio
)
3710 queue_flag
&= ~DEQUEUE_MOVE
;
3712 prev_class
= p
->sched_class
;
3713 queued
= task_on_rq_queued(p
);
3714 running
= task_current(rq
, p
);
3716 dequeue_task(rq
, p
, queue_flag
);
3718 put_prev_task(rq
, p
);
3721 * Boosting condition are:
3722 * 1. -rt task is running and holds mutex A
3723 * --> -dl task blocks on mutex A
3725 * 2. -dl task is running and holds mutex A
3726 * --> -dl task blocks on mutex A and could preempt the
3729 if (dl_prio(prio
)) {
3730 if (!dl_prio(p
->normal_prio
) ||
3731 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3732 p
->dl
.dl_boosted
= 1;
3733 queue_flag
|= ENQUEUE_REPLENISH
;
3735 p
->dl
.dl_boosted
= 0;
3736 p
->sched_class
= &dl_sched_class
;
3737 } else if (rt_prio(prio
)) {
3738 if (dl_prio(oldprio
))
3739 p
->dl
.dl_boosted
= 0;
3741 queue_flag
|= ENQUEUE_HEAD
;
3742 p
->sched_class
= &rt_sched_class
;
3744 if (dl_prio(oldprio
))
3745 p
->dl
.dl_boosted
= 0;
3746 if (rt_prio(oldprio
))
3748 p
->sched_class
= &fair_sched_class
;
3754 enqueue_task(rq
, p
, queue_flag
);
3756 set_curr_task(rq
, p
);
3758 check_class_changed(rq
, p
, prev_class
, oldprio
);
3760 /* Avoid rq from going away on us: */
3762 __task_rq_unlock(rq
, &rf
);
3764 balance_callback(rq
);
3768 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3774 void set_user_nice(struct task_struct
*p
, long nice
)
3776 bool queued
, running
;
3777 int old_prio
, delta
;
3781 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3784 * We have to be careful, if called from sys_setpriority(),
3785 * the task might be in the middle of scheduling on another CPU.
3787 rq
= task_rq_lock(p
, &rf
);
3788 update_rq_clock(rq
);
3791 * The RT priorities are set via sched_setscheduler(), but we still
3792 * allow the 'normal' nice value to be set - but as expected
3793 * it wont have any effect on scheduling until the task is
3794 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3796 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3797 p
->static_prio
= NICE_TO_PRIO(nice
);
3800 queued
= task_on_rq_queued(p
);
3801 running
= task_current(rq
, p
);
3803 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
3805 put_prev_task(rq
, p
);
3807 p
->static_prio
= NICE_TO_PRIO(nice
);
3810 p
->prio
= effective_prio(p
);
3811 delta
= p
->prio
- old_prio
;
3814 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
3816 * If the task increased its priority or is running and
3817 * lowered its priority, then reschedule its CPU:
3819 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3823 set_curr_task(rq
, p
);
3825 task_rq_unlock(rq
, p
, &rf
);
3827 EXPORT_SYMBOL(set_user_nice
);
3830 * can_nice - check if a task can reduce its nice value
3834 int can_nice(const struct task_struct
*p
, const int nice
)
3836 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3837 int nice_rlim
= nice_to_rlimit(nice
);
3839 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3840 capable(CAP_SYS_NICE
));
3843 #ifdef __ARCH_WANT_SYS_NICE
3846 * sys_nice - change the priority of the current process.
3847 * @increment: priority increment
3849 * sys_setpriority is a more generic, but much slower function that
3850 * does similar things.
3852 SYSCALL_DEFINE1(nice
, int, increment
)
3857 * Setpriority might change our priority at the same moment.
3858 * We don't have to worry. Conceptually one call occurs first
3859 * and we have a single winner.
3861 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3862 nice
= task_nice(current
) + increment
;
3864 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3865 if (increment
< 0 && !can_nice(current
, nice
))
3868 retval
= security_task_setnice(current
, nice
);
3872 set_user_nice(current
, nice
);
3879 * task_prio - return the priority value of a given task.
3880 * @p: the task in question.
3882 * Return: The priority value as seen by users in /proc.
3883 * RT tasks are offset by -200. Normal tasks are centered
3884 * around 0, value goes from -16 to +15.
3886 int task_prio(const struct task_struct
*p
)
3888 return p
->prio
- MAX_RT_PRIO
;
3892 * idle_cpu - is a given CPU idle currently?
3893 * @cpu: the processor in question.
3895 * Return: 1 if the CPU is currently idle. 0 otherwise.
3897 int idle_cpu(int cpu
)
3899 struct rq
*rq
= cpu_rq(cpu
);
3901 if (rq
->curr
!= rq
->idle
)
3908 if (!llist_empty(&rq
->wake_list
))
3916 * idle_task - return the idle task for a given CPU.
3917 * @cpu: the processor in question.
3919 * Return: The idle task for the CPU @cpu.
3921 struct task_struct
*idle_task(int cpu
)
3923 return cpu_rq(cpu
)->idle
;
3927 * find_process_by_pid - find a process with a matching PID value.
3928 * @pid: the pid in question.
3930 * The task of @pid, if found. %NULL otherwise.
3932 static struct task_struct
*find_process_by_pid(pid_t pid
)
3934 return pid
? find_task_by_vpid(pid
) : current
;
3938 * sched_setparam() passes in -1 for its policy, to let the functions
3939 * it calls know not to change it.
3941 #define SETPARAM_POLICY -1
3943 static void __setscheduler_params(struct task_struct
*p
,
3944 const struct sched_attr
*attr
)
3946 int policy
= attr
->sched_policy
;
3948 if (policy
== SETPARAM_POLICY
)
3953 if (dl_policy(policy
))
3954 __setparam_dl(p
, attr
);
3955 else if (fair_policy(policy
))
3956 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3959 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3960 * !rt_policy. Always setting this ensures that things like
3961 * getparam()/getattr() don't report silly values for !rt tasks.
3963 p
->rt_priority
= attr
->sched_priority
;
3964 p
->normal_prio
= normal_prio(p
);
3968 /* Actually do priority change: must hold pi & rq lock. */
3969 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3970 const struct sched_attr
*attr
, bool keep_boost
)
3972 __setscheduler_params(p
, attr
);
3975 * Keep a potential priority boosting if called from
3976 * sched_setscheduler().
3978 p
->prio
= normal_prio(p
);
3980 p
->prio
= rt_effective_prio(p
, p
->prio
);
3982 if (dl_prio(p
->prio
))
3983 p
->sched_class
= &dl_sched_class
;
3984 else if (rt_prio(p
->prio
))
3985 p
->sched_class
= &rt_sched_class
;
3987 p
->sched_class
= &fair_sched_class
;
3991 * Check the target process has a UID that matches the current process's:
3993 static bool check_same_owner(struct task_struct
*p
)
3995 const struct cred
*cred
= current_cred(), *pcred
;
3999 pcred
= __task_cred(p
);
4000 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4001 uid_eq(cred
->euid
, pcred
->uid
));
4006 static int __sched_setscheduler(struct task_struct
*p
,
4007 const struct sched_attr
*attr
,
4010 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4011 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4012 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4013 int new_effective_prio
, policy
= attr
->sched_policy
;
4014 const struct sched_class
*prev_class
;
4017 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4020 /* The pi code expects interrupts enabled */
4021 BUG_ON(pi
&& in_interrupt());
4023 /* Double check policy once rq lock held: */
4025 reset_on_fork
= p
->sched_reset_on_fork
;
4026 policy
= oldpolicy
= p
->policy
;
4028 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4030 if (!valid_policy(policy
))
4034 if (attr
->sched_flags
&
4035 ~(SCHED_FLAG_RESET_ON_FORK
| SCHED_FLAG_RECLAIM
))
4039 * Valid priorities for SCHED_FIFO and SCHED_RR are
4040 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4041 * SCHED_BATCH and SCHED_IDLE is 0.
4043 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4044 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4046 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4047 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4051 * Allow unprivileged RT tasks to decrease priority:
4053 if (user
&& !capable(CAP_SYS_NICE
)) {
4054 if (fair_policy(policy
)) {
4055 if (attr
->sched_nice
< task_nice(p
) &&
4056 !can_nice(p
, attr
->sched_nice
))
4060 if (rt_policy(policy
)) {
4061 unsigned long rlim_rtprio
=
4062 task_rlimit(p
, RLIMIT_RTPRIO
);
4064 /* Can't set/change the rt policy: */
4065 if (policy
!= p
->policy
&& !rlim_rtprio
)
4068 /* Can't increase priority: */
4069 if (attr
->sched_priority
> p
->rt_priority
&&
4070 attr
->sched_priority
> rlim_rtprio
)
4075 * Can't set/change SCHED_DEADLINE policy at all for now
4076 * (safest behavior); in the future we would like to allow
4077 * unprivileged DL tasks to increase their relative deadline
4078 * or reduce their runtime (both ways reducing utilization)
4080 if (dl_policy(policy
))
4084 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4085 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4087 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4088 if (!can_nice(p
, task_nice(p
)))
4092 /* Can't change other user's priorities: */
4093 if (!check_same_owner(p
))
4096 /* Normal users shall not reset the sched_reset_on_fork flag: */
4097 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4102 retval
= security_task_setscheduler(p
);
4108 * Make sure no PI-waiters arrive (or leave) while we are
4109 * changing the priority of the task:
4111 * To be able to change p->policy safely, the appropriate
4112 * runqueue lock must be held.
4114 rq
= task_rq_lock(p
, &rf
);
4115 update_rq_clock(rq
);
4118 * Changing the policy of the stop threads its a very bad idea:
4120 if (p
== rq
->stop
) {
4121 task_rq_unlock(rq
, p
, &rf
);
4126 * If not changing anything there's no need to proceed further,
4127 * but store a possible modification of reset_on_fork.
4129 if (unlikely(policy
== p
->policy
)) {
4130 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4132 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4134 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4137 p
->sched_reset_on_fork
= reset_on_fork
;
4138 task_rq_unlock(rq
, p
, &rf
);
4144 #ifdef CONFIG_RT_GROUP_SCHED
4146 * Do not allow realtime tasks into groups that have no runtime
4149 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4150 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4151 !task_group_is_autogroup(task_group(p
))) {
4152 task_rq_unlock(rq
, p
, &rf
);
4157 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4158 cpumask_t
*span
= rq
->rd
->span
;
4161 * Don't allow tasks with an affinity mask smaller than
4162 * the entire root_domain to become SCHED_DEADLINE. We
4163 * will also fail if there's no bandwidth available.
4165 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4166 rq
->rd
->dl_bw
.bw
== 0) {
4167 task_rq_unlock(rq
, p
, &rf
);
4174 /* Re-check policy now with rq lock held: */
4175 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4176 policy
= oldpolicy
= -1;
4177 task_rq_unlock(rq
, p
, &rf
);
4182 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4183 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4186 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
4187 task_rq_unlock(rq
, p
, &rf
);
4191 p
->sched_reset_on_fork
= reset_on_fork
;
4196 * Take priority boosted tasks into account. If the new
4197 * effective priority is unchanged, we just store the new
4198 * normal parameters and do not touch the scheduler class and
4199 * the runqueue. This will be done when the task deboost
4202 new_effective_prio
= rt_effective_prio(p
, newprio
);
4203 if (new_effective_prio
== oldprio
)
4204 queue_flags
&= ~DEQUEUE_MOVE
;
4207 queued
= task_on_rq_queued(p
);
4208 running
= task_current(rq
, p
);
4210 dequeue_task(rq
, p
, queue_flags
);
4212 put_prev_task(rq
, p
);
4214 prev_class
= p
->sched_class
;
4215 __setscheduler(rq
, p
, attr
, pi
);
4219 * We enqueue to tail when the priority of a task is
4220 * increased (user space view).
4222 if (oldprio
< p
->prio
)
4223 queue_flags
|= ENQUEUE_HEAD
;
4225 enqueue_task(rq
, p
, queue_flags
);
4228 set_curr_task(rq
, p
);
4230 check_class_changed(rq
, p
, prev_class
, oldprio
);
4232 /* Avoid rq from going away on us: */
4234 task_rq_unlock(rq
, p
, &rf
);
4237 rt_mutex_adjust_pi(p
);
4239 /* Run balance callbacks after we've adjusted the PI chain: */
4240 balance_callback(rq
);
4246 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4247 const struct sched_param
*param
, bool check
)
4249 struct sched_attr attr
= {
4250 .sched_policy
= policy
,
4251 .sched_priority
= param
->sched_priority
,
4252 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4255 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4256 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4257 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4258 policy
&= ~SCHED_RESET_ON_FORK
;
4259 attr
.sched_policy
= policy
;
4262 return __sched_setscheduler(p
, &attr
, check
, true);
4265 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4266 * @p: the task in question.
4267 * @policy: new policy.
4268 * @param: structure containing the new RT priority.
4270 * Return: 0 on success. An error code otherwise.
4272 * NOTE that the task may be already dead.
4274 int sched_setscheduler(struct task_struct
*p
, int policy
,
4275 const struct sched_param
*param
)
4277 return _sched_setscheduler(p
, policy
, param
, true);
4279 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4281 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4283 return __sched_setscheduler(p
, attr
, true, true);
4285 EXPORT_SYMBOL_GPL(sched_setattr
);
4288 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4289 * @p: the task in question.
4290 * @policy: new policy.
4291 * @param: structure containing the new RT priority.
4293 * Just like sched_setscheduler, only don't bother checking if the
4294 * current context has permission. For example, this is needed in
4295 * stop_machine(): we create temporary high priority worker threads,
4296 * but our caller might not have that capability.
4298 * Return: 0 on success. An error code otherwise.
4300 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4301 const struct sched_param
*param
)
4303 return _sched_setscheduler(p
, policy
, param
, false);
4305 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4308 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4310 struct sched_param lparam
;
4311 struct task_struct
*p
;
4314 if (!param
|| pid
< 0)
4316 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4321 p
= find_process_by_pid(pid
);
4323 retval
= sched_setscheduler(p
, policy
, &lparam
);
4330 * Mimics kernel/events/core.c perf_copy_attr().
4332 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
4337 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4340 /* Zero the full structure, so that a short copy will be nice: */
4341 memset(attr
, 0, sizeof(*attr
));
4343 ret
= get_user(size
, &uattr
->size
);
4347 /* Bail out on silly large: */
4348 if (size
> PAGE_SIZE
)
4351 /* ABI compatibility quirk: */
4353 size
= SCHED_ATTR_SIZE_VER0
;
4355 if (size
< SCHED_ATTR_SIZE_VER0
)
4359 * If we're handed a bigger struct than we know of,
4360 * ensure all the unknown bits are 0 - i.e. new
4361 * user-space does not rely on any kernel feature
4362 * extensions we dont know about yet.
4364 if (size
> sizeof(*attr
)) {
4365 unsigned char __user
*addr
;
4366 unsigned char __user
*end
;
4369 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4370 end
= (void __user
*)uattr
+ size
;
4372 for (; addr
< end
; addr
++) {
4373 ret
= get_user(val
, addr
);
4379 size
= sizeof(*attr
);
4382 ret
= copy_from_user(attr
, uattr
, size
);
4387 * XXX: Do we want to be lenient like existing syscalls; or do we want
4388 * to be strict and return an error on out-of-bounds values?
4390 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4395 put_user(sizeof(*attr
), &uattr
->size
);
4400 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4401 * @pid: the pid in question.
4402 * @policy: new policy.
4403 * @param: structure containing the new RT priority.
4405 * Return: 0 on success. An error code otherwise.
4407 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
4412 return do_sched_setscheduler(pid
, policy
, param
);
4416 * sys_sched_setparam - set/change the RT priority of a thread
4417 * @pid: the pid in question.
4418 * @param: structure containing the new RT priority.
4420 * Return: 0 on success. An error code otherwise.
4422 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4424 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4428 * sys_sched_setattr - same as above, but with extended sched_attr
4429 * @pid: the pid in question.
4430 * @uattr: structure containing the extended parameters.
4431 * @flags: for future extension.
4433 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4434 unsigned int, flags
)
4436 struct sched_attr attr
;
4437 struct task_struct
*p
;
4440 if (!uattr
|| pid
< 0 || flags
)
4443 retval
= sched_copy_attr(uattr
, &attr
);
4447 if ((int)attr
.sched_policy
< 0)
4452 p
= find_process_by_pid(pid
);
4454 retval
= sched_setattr(p
, &attr
);
4461 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4462 * @pid: the pid in question.
4464 * Return: On success, the policy of the thread. Otherwise, a negative error
4467 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4469 struct task_struct
*p
;
4477 p
= find_process_by_pid(pid
);
4479 retval
= security_task_getscheduler(p
);
4482 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4489 * sys_sched_getparam - get the RT priority of a thread
4490 * @pid: the pid in question.
4491 * @param: structure containing the RT priority.
4493 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4496 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4498 struct sched_param lp
= { .sched_priority
= 0 };
4499 struct task_struct
*p
;
4502 if (!param
|| pid
< 0)
4506 p
= find_process_by_pid(pid
);
4511 retval
= security_task_getscheduler(p
);
4515 if (task_has_rt_policy(p
))
4516 lp
.sched_priority
= p
->rt_priority
;
4520 * This one might sleep, we cannot do it with a spinlock held ...
4522 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4531 static int sched_read_attr(struct sched_attr __user
*uattr
,
4532 struct sched_attr
*attr
,
4537 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4541 * If we're handed a smaller struct than we know of,
4542 * ensure all the unknown bits are 0 - i.e. old
4543 * user-space does not get uncomplete information.
4545 if (usize
< sizeof(*attr
)) {
4546 unsigned char *addr
;
4549 addr
= (void *)attr
+ usize
;
4550 end
= (void *)attr
+ sizeof(*attr
);
4552 for (; addr
< end
; addr
++) {
4560 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4568 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4569 * @pid: the pid in question.
4570 * @uattr: structure containing the extended parameters.
4571 * @size: sizeof(attr) for fwd/bwd comp.
4572 * @flags: for future extension.
4574 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4575 unsigned int, size
, unsigned int, flags
)
4577 struct sched_attr attr
= {
4578 .size
= sizeof(struct sched_attr
),
4580 struct task_struct
*p
;
4583 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4584 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4588 p
= find_process_by_pid(pid
);
4593 retval
= security_task_getscheduler(p
);
4597 attr
.sched_policy
= p
->policy
;
4598 if (p
->sched_reset_on_fork
)
4599 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4600 if (task_has_dl_policy(p
))
4601 __getparam_dl(p
, &attr
);
4602 else if (task_has_rt_policy(p
))
4603 attr
.sched_priority
= p
->rt_priority
;
4605 attr
.sched_nice
= task_nice(p
);
4609 retval
= sched_read_attr(uattr
, &attr
, size
);
4617 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4619 cpumask_var_t cpus_allowed
, new_mask
;
4620 struct task_struct
*p
;
4625 p
= find_process_by_pid(pid
);
4631 /* Prevent p going away */
4635 if (p
->flags
& PF_NO_SETAFFINITY
) {
4639 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4643 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4645 goto out_free_cpus_allowed
;
4648 if (!check_same_owner(p
)) {
4650 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4652 goto out_free_new_mask
;
4657 retval
= security_task_setscheduler(p
);
4659 goto out_free_new_mask
;
4662 cpuset_cpus_allowed(p
, cpus_allowed
);
4663 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4666 * Since bandwidth control happens on root_domain basis,
4667 * if admission test is enabled, we only admit -deadline
4668 * tasks allowed to run on all the CPUs in the task's
4672 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4674 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4677 goto out_free_new_mask
;
4683 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4686 cpuset_cpus_allowed(p
, cpus_allowed
);
4687 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4689 * We must have raced with a concurrent cpuset
4690 * update. Just reset the cpus_allowed to the
4691 * cpuset's cpus_allowed
4693 cpumask_copy(new_mask
, cpus_allowed
);
4698 free_cpumask_var(new_mask
);
4699 out_free_cpus_allowed
:
4700 free_cpumask_var(cpus_allowed
);
4706 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4707 struct cpumask
*new_mask
)
4709 if (len
< cpumask_size())
4710 cpumask_clear(new_mask
);
4711 else if (len
> cpumask_size())
4712 len
= cpumask_size();
4714 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4718 * sys_sched_setaffinity - set the CPU affinity of a process
4719 * @pid: pid of the process
4720 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4721 * @user_mask_ptr: user-space pointer to the new CPU mask
4723 * Return: 0 on success. An error code otherwise.
4725 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4726 unsigned long __user
*, user_mask_ptr
)
4728 cpumask_var_t new_mask
;
4731 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4734 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4736 retval
= sched_setaffinity(pid
, new_mask
);
4737 free_cpumask_var(new_mask
);
4741 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4743 struct task_struct
*p
;
4744 unsigned long flags
;
4750 p
= find_process_by_pid(pid
);
4754 retval
= security_task_getscheduler(p
);
4758 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4759 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4760 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4769 * sys_sched_getaffinity - get the CPU affinity of a process
4770 * @pid: pid of the process
4771 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4772 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4774 * Return: size of CPU mask copied to user_mask_ptr on success. An
4775 * error code otherwise.
4777 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4778 unsigned long __user
*, user_mask_ptr
)
4783 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4785 if (len
& (sizeof(unsigned long)-1))
4788 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4791 ret
= sched_getaffinity(pid
, mask
);
4793 size_t retlen
= min_t(size_t, len
, cpumask_size());
4795 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4800 free_cpumask_var(mask
);
4806 * sys_sched_yield - yield the current processor to other threads.
4808 * This function yields the current CPU to other tasks. If there are no
4809 * other threads running on this CPU then this function will return.
4813 SYSCALL_DEFINE0(sched_yield
)
4818 local_irq_disable();
4822 schedstat_inc(rq
->yld_count
);
4823 current
->sched_class
->yield_task(rq
);
4826 * Since we are going to call schedule() anyway, there's
4827 * no need to preempt or enable interrupts:
4831 sched_preempt_enable_no_resched();
4838 #ifndef CONFIG_PREEMPT
4839 int __sched
_cond_resched(void)
4841 if (should_resched(0)) {
4842 preempt_schedule_common();
4847 EXPORT_SYMBOL(_cond_resched
);
4851 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4852 * call schedule, and on return reacquire the lock.
4854 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4855 * operations here to prevent schedule() from being called twice (once via
4856 * spin_unlock(), once by hand).
4858 int __cond_resched_lock(spinlock_t
*lock
)
4860 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4863 lockdep_assert_held(lock
);
4865 if (spin_needbreak(lock
) || resched
) {
4868 preempt_schedule_common();
4876 EXPORT_SYMBOL(__cond_resched_lock
);
4878 int __sched
__cond_resched_softirq(void)
4880 BUG_ON(!in_softirq());
4882 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4884 preempt_schedule_common();
4890 EXPORT_SYMBOL(__cond_resched_softirq
);
4893 * yield - yield the current processor to other threads.
4895 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4897 * The scheduler is at all times free to pick the calling task as the most
4898 * eligible task to run, if removing the yield() call from your code breaks
4899 * it, its already broken.
4901 * Typical broken usage is:
4906 * where one assumes that yield() will let 'the other' process run that will
4907 * make event true. If the current task is a SCHED_FIFO task that will never
4908 * happen. Never use yield() as a progress guarantee!!
4910 * If you want to use yield() to wait for something, use wait_event().
4911 * If you want to use yield() to be 'nice' for others, use cond_resched().
4912 * If you still want to use yield(), do not!
4914 void __sched
yield(void)
4916 set_current_state(TASK_RUNNING
);
4919 EXPORT_SYMBOL(yield
);
4922 * yield_to - yield the current processor to another thread in
4923 * your thread group, or accelerate that thread toward the
4924 * processor it's on.
4926 * @preempt: whether task preemption is allowed or not
4928 * It's the caller's job to ensure that the target task struct
4929 * can't go away on us before we can do any checks.
4932 * true (>0) if we indeed boosted the target task.
4933 * false (0) if we failed to boost the target.
4934 * -ESRCH if there's no task to yield to.
4936 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4938 struct task_struct
*curr
= current
;
4939 struct rq
*rq
, *p_rq
;
4940 unsigned long flags
;
4943 local_irq_save(flags
);
4949 * If we're the only runnable task on the rq and target rq also
4950 * has only one task, there's absolutely no point in yielding.
4952 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4957 double_rq_lock(rq
, p_rq
);
4958 if (task_rq(p
) != p_rq
) {
4959 double_rq_unlock(rq
, p_rq
);
4963 if (!curr
->sched_class
->yield_to_task
)
4966 if (curr
->sched_class
!= p
->sched_class
)
4969 if (task_running(p_rq
, p
) || p
->state
)
4972 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4974 schedstat_inc(rq
->yld_count
);
4976 * Make p's CPU reschedule; pick_next_entity takes care of
4979 if (preempt
&& rq
!= p_rq
)
4984 double_rq_unlock(rq
, p_rq
);
4986 local_irq_restore(flags
);
4993 EXPORT_SYMBOL_GPL(yield_to
);
4995 int io_schedule_prepare(void)
4997 int old_iowait
= current
->in_iowait
;
4999 current
->in_iowait
= 1;
5000 blk_schedule_flush_plug(current
);
5005 void io_schedule_finish(int token
)
5007 current
->in_iowait
= token
;
5011 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5012 * that process accounting knows that this is a task in IO wait state.
5014 long __sched
io_schedule_timeout(long timeout
)
5019 token
= io_schedule_prepare();
5020 ret
= schedule_timeout(timeout
);
5021 io_schedule_finish(token
);
5025 EXPORT_SYMBOL(io_schedule_timeout
);
5027 void io_schedule(void)
5031 token
= io_schedule_prepare();
5033 io_schedule_finish(token
);
5035 EXPORT_SYMBOL(io_schedule
);
5038 * sys_sched_get_priority_max - return maximum RT priority.
5039 * @policy: scheduling class.
5041 * Return: On success, this syscall returns the maximum
5042 * rt_priority that can be used by a given scheduling class.
5043 * On failure, a negative error code is returned.
5045 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5052 ret
= MAX_USER_RT_PRIO
-1;
5054 case SCHED_DEADLINE
:
5065 * sys_sched_get_priority_min - return minimum RT priority.
5066 * @policy: scheduling class.
5068 * Return: On success, this syscall returns the minimum
5069 * rt_priority that can be used by a given scheduling class.
5070 * On failure, a negative error code is returned.
5072 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5081 case SCHED_DEADLINE
:
5091 * sys_sched_rr_get_interval - return the default timeslice of a process.
5092 * @pid: pid of the process.
5093 * @interval: userspace pointer to the timeslice value.
5095 * this syscall writes the default timeslice value of a given process
5096 * into the user-space timespec buffer. A value of '0' means infinity.
5098 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5101 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5102 struct timespec __user
*, interval
)
5104 struct task_struct
*p
;
5105 unsigned int time_slice
;
5116 p
= find_process_by_pid(pid
);
5120 retval
= security_task_getscheduler(p
);
5124 rq
= task_rq_lock(p
, &rf
);
5126 if (p
->sched_class
->get_rr_interval
)
5127 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5128 task_rq_unlock(rq
, p
, &rf
);
5131 jiffies_to_timespec(time_slice
, &t
);
5132 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5140 void sched_show_task(struct task_struct
*p
)
5142 unsigned long free
= 0;
5145 if (!try_get_task_stack(p
))
5148 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5150 if (p
->state
== TASK_RUNNING
)
5151 printk(KERN_CONT
" running task ");
5152 #ifdef CONFIG_DEBUG_STACK_USAGE
5153 free
= stack_not_used(p
);
5158 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5160 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5161 task_pid_nr(p
), ppid
,
5162 (unsigned long)task_thread_info(p
)->flags
);
5164 print_worker_info(KERN_INFO
, p
);
5165 show_stack(p
, NULL
);
5169 void show_state_filter(unsigned long state_filter
)
5171 struct task_struct
*g
, *p
;
5173 #if BITS_PER_LONG == 32
5175 " task PC stack pid father\n");
5178 " task PC stack pid father\n");
5181 for_each_process_thread(g
, p
) {
5183 * reset the NMI-timeout, listing all files on a slow
5184 * console might take a lot of time:
5185 * Also, reset softlockup watchdogs on all CPUs, because
5186 * another CPU might be blocked waiting for us to process
5189 touch_nmi_watchdog();
5190 touch_all_softlockup_watchdogs();
5191 if (!state_filter
|| (p
->state
& state_filter
))
5195 #ifdef CONFIG_SCHED_DEBUG
5197 sysrq_sched_debug_show();
5201 * Only show locks if all tasks are dumped:
5204 debug_show_all_locks();
5208 * init_idle - set up an idle thread for a given CPU
5209 * @idle: task in question
5210 * @cpu: CPU the idle task belongs to
5212 * NOTE: this function does not set the idle thread's NEED_RESCHED
5213 * flag, to make booting more robust.
5215 void init_idle(struct task_struct
*idle
, int cpu
)
5217 struct rq
*rq
= cpu_rq(cpu
);
5218 unsigned long flags
;
5220 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5221 raw_spin_lock(&rq
->lock
);
5223 __sched_fork(0, idle
);
5224 idle
->state
= TASK_RUNNING
;
5225 idle
->se
.exec_start
= sched_clock();
5226 idle
->flags
|= PF_IDLE
;
5228 kasan_unpoison_task_stack(idle
);
5232 * Its possible that init_idle() gets called multiple times on a task,
5233 * in that case do_set_cpus_allowed() will not do the right thing.
5235 * And since this is boot we can forgo the serialization.
5237 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5240 * We're having a chicken and egg problem, even though we are
5241 * holding rq->lock, the CPU isn't yet set to this CPU so the
5242 * lockdep check in task_group() will fail.
5244 * Similar case to sched_fork(). / Alternatively we could
5245 * use task_rq_lock() here and obtain the other rq->lock.
5250 __set_task_cpu(idle
, cpu
);
5253 rq
->curr
= rq
->idle
= idle
;
5254 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5258 raw_spin_unlock(&rq
->lock
);
5259 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5261 /* Set the preempt count _outside_ the spinlocks! */
5262 init_idle_preempt_count(idle
, cpu
);
5265 * The idle tasks have their own, simple scheduling class:
5267 idle
->sched_class
= &idle_sched_class
;
5268 ftrace_graph_init_idle_task(idle
, cpu
);
5269 vtime_init_idle(idle
, cpu
);
5271 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5277 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5278 const struct cpumask
*trial
)
5282 if (!cpumask_weight(cur
))
5285 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
5290 int task_can_attach(struct task_struct
*p
,
5291 const struct cpumask
*cs_cpus_allowed
)
5296 * Kthreads which disallow setaffinity shouldn't be moved
5297 * to a new cpuset; we don't want to change their CPU
5298 * affinity and isolating such threads by their set of
5299 * allowed nodes is unnecessary. Thus, cpusets are not
5300 * applicable for such threads. This prevents checking for
5301 * success of set_cpus_allowed_ptr() on all attached tasks
5302 * before cpus_allowed may be changed.
5304 if (p
->flags
& PF_NO_SETAFFINITY
) {
5309 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5311 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
5317 bool sched_smp_initialized __read_mostly
;
5319 #ifdef CONFIG_NUMA_BALANCING
5320 /* Migrate current task p to target_cpu */
5321 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5323 struct migration_arg arg
= { p
, target_cpu
};
5324 int curr_cpu
= task_cpu(p
);
5326 if (curr_cpu
== target_cpu
)
5329 if (!cpumask_test_cpu(target_cpu
, &p
->cpus_allowed
))
5332 /* TODO: This is not properly updating schedstats */
5334 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5335 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5339 * Requeue a task on a given node and accurately track the number of NUMA
5340 * tasks on the runqueues
5342 void sched_setnuma(struct task_struct
*p
, int nid
)
5344 bool queued
, running
;
5348 rq
= task_rq_lock(p
, &rf
);
5349 queued
= task_on_rq_queued(p
);
5350 running
= task_current(rq
, p
);
5353 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5355 put_prev_task(rq
, p
);
5357 p
->numa_preferred_nid
= nid
;
5360 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5362 set_curr_task(rq
, p
);
5363 task_rq_unlock(rq
, p
, &rf
);
5365 #endif /* CONFIG_NUMA_BALANCING */
5367 #ifdef CONFIG_HOTPLUG_CPU
5369 * Ensure that the idle task is using init_mm right before its CPU goes
5372 void idle_task_exit(void)
5374 struct mm_struct
*mm
= current
->active_mm
;
5376 BUG_ON(cpu_online(smp_processor_id()));
5378 if (mm
!= &init_mm
) {
5379 switch_mm(mm
, &init_mm
, current
);
5380 finish_arch_post_lock_switch();
5386 * Since this CPU is going 'away' for a while, fold any nr_active delta
5387 * we might have. Assumes we're called after migrate_tasks() so that the
5388 * nr_active count is stable. We need to take the teardown thread which
5389 * is calling this into account, so we hand in adjust = 1 to the load
5392 * Also see the comment "Global load-average calculations".
5394 static void calc_load_migrate(struct rq
*rq
)
5396 long delta
= calc_load_fold_active(rq
, 1);
5398 atomic_long_add(delta
, &calc_load_tasks
);
5401 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5405 static const struct sched_class fake_sched_class
= {
5406 .put_prev_task
= put_prev_task_fake
,
5409 static struct task_struct fake_task
= {
5411 * Avoid pull_{rt,dl}_task()
5413 .prio
= MAX_PRIO
+ 1,
5414 .sched_class
= &fake_sched_class
,
5418 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5419 * try_to_wake_up()->select_task_rq().
5421 * Called with rq->lock held even though we'er in stop_machine() and
5422 * there's no concurrency possible, we hold the required locks anyway
5423 * because of lock validation efforts.
5425 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
5427 struct rq
*rq
= dead_rq
;
5428 struct task_struct
*next
, *stop
= rq
->stop
;
5429 struct rq_flags orf
= *rf
;
5433 * Fudge the rq selection such that the below task selection loop
5434 * doesn't get stuck on the currently eligible stop task.
5436 * We're currently inside stop_machine() and the rq is either stuck
5437 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5438 * either way we should never end up calling schedule() until we're
5444 * put_prev_task() and pick_next_task() sched
5445 * class method both need to have an up-to-date
5446 * value of rq->clock[_task]
5448 update_rq_clock(rq
);
5452 * There's this thread running, bail when that's the only
5455 if (rq
->nr_running
== 1)
5459 * pick_next_task() assumes pinned rq->lock:
5461 next
= pick_next_task(rq
, &fake_task
, rf
);
5463 put_prev_task(rq
, next
);
5466 * Rules for changing task_struct::cpus_allowed are holding
5467 * both pi_lock and rq->lock, such that holding either
5468 * stabilizes the mask.
5470 * Drop rq->lock is not quite as disastrous as it usually is
5471 * because !cpu_active at this point, which means load-balance
5472 * will not interfere. Also, stop-machine.
5475 raw_spin_lock(&next
->pi_lock
);
5479 * Since we're inside stop-machine, _nothing_ should have
5480 * changed the task, WARN if weird stuff happened, because in
5481 * that case the above rq->lock drop is a fail too.
5483 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5484 raw_spin_unlock(&next
->pi_lock
);
5488 /* Find suitable destination for @next, with force if needed. */
5489 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5490 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
5491 if (rq
!= dead_rq
) {
5497 raw_spin_unlock(&next
->pi_lock
);
5502 #endif /* CONFIG_HOTPLUG_CPU */
5504 void set_rq_online(struct rq
*rq
)
5507 const struct sched_class
*class;
5509 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5512 for_each_class(class) {
5513 if (class->rq_online
)
5514 class->rq_online(rq
);
5519 void set_rq_offline(struct rq
*rq
)
5522 const struct sched_class
*class;
5524 for_each_class(class) {
5525 if (class->rq_offline
)
5526 class->rq_offline(rq
);
5529 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5534 static void set_cpu_rq_start_time(unsigned int cpu
)
5536 struct rq
*rq
= cpu_rq(cpu
);
5538 rq
->age_stamp
= sched_clock_cpu(cpu
);
5542 * used to mark begin/end of suspend/resume:
5544 static int num_cpus_frozen
;
5547 * Update cpusets according to cpu_active mask. If cpusets are
5548 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5549 * around partition_sched_domains().
5551 * If we come here as part of a suspend/resume, don't touch cpusets because we
5552 * want to restore it back to its original state upon resume anyway.
5554 static void cpuset_cpu_active(void)
5556 if (cpuhp_tasks_frozen
) {
5558 * num_cpus_frozen tracks how many CPUs are involved in suspend
5559 * resume sequence. As long as this is not the last online
5560 * operation in the resume sequence, just build a single sched
5561 * domain, ignoring cpusets.
5563 partition_sched_domains(1, NULL
, NULL
);
5564 if (--num_cpus_frozen
)
5567 * This is the last CPU online operation. So fall through and
5568 * restore the original sched domains by considering the
5569 * cpuset configurations.
5571 cpuset_force_rebuild();
5573 cpuset_update_active_cpus();
5576 static int cpuset_cpu_inactive(unsigned int cpu
)
5578 if (!cpuhp_tasks_frozen
) {
5579 if (dl_cpu_busy(cpu
))
5581 cpuset_update_active_cpus();
5584 partition_sched_domains(1, NULL
, NULL
);
5589 int sched_cpu_activate(unsigned int cpu
)
5591 struct rq
*rq
= cpu_rq(cpu
);
5594 set_cpu_active(cpu
, true);
5596 if (sched_smp_initialized
) {
5597 sched_domains_numa_masks_set(cpu
);
5598 cpuset_cpu_active();
5602 * Put the rq online, if not already. This happens:
5604 * 1) In the early boot process, because we build the real domains
5605 * after all CPUs have been brought up.
5607 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5610 rq_lock_irqsave(rq
, &rf
);
5612 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5615 rq_unlock_irqrestore(rq
, &rf
);
5617 update_max_interval();
5622 int sched_cpu_deactivate(unsigned int cpu
)
5626 set_cpu_active(cpu
, false);
5628 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5629 * users of this state to go away such that all new such users will
5632 * Do sync before park smpboot threads to take care the rcu boost case.
5634 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
5636 if (!sched_smp_initialized
)
5639 ret
= cpuset_cpu_inactive(cpu
);
5641 set_cpu_active(cpu
, true);
5644 sched_domains_numa_masks_clear(cpu
);
5648 static void sched_rq_cpu_starting(unsigned int cpu
)
5650 struct rq
*rq
= cpu_rq(cpu
);
5652 rq
->calc_load_update
= calc_load_update
;
5653 update_max_interval();
5656 int sched_cpu_starting(unsigned int cpu
)
5658 set_cpu_rq_start_time(cpu
);
5659 sched_rq_cpu_starting(cpu
);
5663 #ifdef CONFIG_HOTPLUG_CPU
5664 int sched_cpu_dying(unsigned int cpu
)
5666 struct rq
*rq
= cpu_rq(cpu
);
5669 /* Handle pending wakeups and then migrate everything off */
5670 sched_ttwu_pending();
5672 rq_lock_irqsave(rq
, &rf
);
5674 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5677 migrate_tasks(rq
, &rf
);
5678 BUG_ON(rq
->nr_running
!= 1);
5679 rq_unlock_irqrestore(rq
, &rf
);
5681 calc_load_migrate(rq
);
5682 update_max_interval();
5683 nohz_balance_exit_idle(cpu
);
5689 #ifdef CONFIG_SCHED_SMT
5690 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
5692 static void sched_init_smt(void)
5695 * We've enumerated all CPUs and will assume that if any CPU
5696 * has SMT siblings, CPU0 will too.
5698 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5699 static_branch_enable(&sched_smt_present
);
5702 static inline void sched_init_smt(void) { }
5705 void __init
sched_init_smp(void)
5707 cpumask_var_t non_isolated_cpus
;
5709 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
5714 * There's no userspace yet to cause hotplug operations; hence all the
5715 * CPU masks are stable and all blatant races in the below code cannot
5718 mutex_lock(&sched_domains_mutex
);
5719 sched_init_domains(cpu_active_mask
);
5720 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
5721 if (cpumask_empty(non_isolated_cpus
))
5722 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
5723 mutex_unlock(&sched_domains_mutex
);
5725 /* Move init over to a non-isolated CPU */
5726 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
5728 sched_init_granularity();
5729 free_cpumask_var(non_isolated_cpus
);
5731 init_sched_rt_class();
5732 init_sched_dl_class();
5736 sched_smp_initialized
= true;
5739 static int __init
migration_init(void)
5741 sched_rq_cpu_starting(smp_processor_id());
5744 early_initcall(migration_init
);
5747 void __init
sched_init_smp(void)
5749 sched_init_granularity();
5751 #endif /* CONFIG_SMP */
5753 int in_sched_functions(unsigned long addr
)
5755 return in_lock_functions(addr
) ||
5756 (addr
>= (unsigned long)__sched_text_start
5757 && addr
< (unsigned long)__sched_text_end
);
5760 #ifdef CONFIG_CGROUP_SCHED
5762 * Default task group.
5763 * Every task in system belongs to this group at bootup.
5765 struct task_group root_task_group
;
5766 LIST_HEAD(task_groups
);
5768 /* Cacheline aligned slab cache for task_group */
5769 static struct kmem_cache
*task_group_cache __read_mostly
;
5772 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5773 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5775 void __init
sched_init(void)
5778 unsigned long alloc_size
= 0, ptr
;
5783 #ifdef CONFIG_FAIR_GROUP_SCHED
5784 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5786 #ifdef CONFIG_RT_GROUP_SCHED
5787 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5790 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
5792 #ifdef CONFIG_FAIR_GROUP_SCHED
5793 root_task_group
.se
= (struct sched_entity
**)ptr
;
5794 ptr
+= nr_cpu_ids
* sizeof(void **);
5796 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
5797 ptr
+= nr_cpu_ids
* sizeof(void **);
5799 #endif /* CONFIG_FAIR_GROUP_SCHED */
5800 #ifdef CONFIG_RT_GROUP_SCHED
5801 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
5802 ptr
+= nr_cpu_ids
* sizeof(void **);
5804 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
5805 ptr
+= nr_cpu_ids
* sizeof(void **);
5807 #endif /* CONFIG_RT_GROUP_SCHED */
5809 #ifdef CONFIG_CPUMASK_OFFSTACK
5810 for_each_possible_cpu(i
) {
5811 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5812 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5813 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5814 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5816 #endif /* CONFIG_CPUMASK_OFFSTACK */
5818 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
5819 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
5822 init_defrootdomain();
5825 #ifdef CONFIG_RT_GROUP_SCHED
5826 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
5827 global_rt_period(), global_rt_runtime());
5828 #endif /* CONFIG_RT_GROUP_SCHED */
5830 #ifdef CONFIG_CGROUP_SCHED
5831 task_group_cache
= KMEM_CACHE(task_group
, 0);
5833 list_add(&root_task_group
.list
, &task_groups
);
5834 INIT_LIST_HEAD(&root_task_group
.children
);
5835 INIT_LIST_HEAD(&root_task_group
.siblings
);
5836 autogroup_init(&init_task
);
5837 #endif /* CONFIG_CGROUP_SCHED */
5839 for_each_possible_cpu(i
) {
5843 raw_spin_lock_init(&rq
->lock
);
5845 rq
->calc_load_active
= 0;
5846 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
5847 init_cfs_rq(&rq
->cfs
);
5848 init_rt_rq(&rq
->rt
);
5849 init_dl_rq(&rq
->dl
);
5850 #ifdef CONFIG_FAIR_GROUP_SCHED
5851 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
5852 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
5853 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
5855 * How much CPU bandwidth does root_task_group get?
5857 * In case of task-groups formed thr' the cgroup filesystem, it
5858 * gets 100% of the CPU resources in the system. This overall
5859 * system CPU resource is divided among the tasks of
5860 * root_task_group and its child task-groups in a fair manner,
5861 * based on each entity's (task or task-group's) weight
5862 * (se->load.weight).
5864 * In other words, if root_task_group has 10 tasks of weight
5865 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5866 * then A0's share of the CPU resource is:
5868 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
5870 * We achieve this by letting root_task_group's tasks sit
5871 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
5873 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
5874 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
5875 #endif /* CONFIG_FAIR_GROUP_SCHED */
5877 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
5878 #ifdef CONFIG_RT_GROUP_SCHED
5879 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
5882 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
5883 rq
->cpu_load
[j
] = 0;
5888 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
5889 rq
->balance_callback
= NULL
;
5890 rq
->active_balance
= 0;
5891 rq
->next_balance
= jiffies
;
5896 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
5897 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
5899 INIT_LIST_HEAD(&rq
->cfs_tasks
);
5901 rq_attach_root(rq
, &def_root_domain
);
5902 #ifdef CONFIG_NO_HZ_COMMON
5903 rq
->last_load_update_tick
= jiffies
;
5906 #ifdef CONFIG_NO_HZ_FULL
5907 rq
->last_sched_tick
= 0;
5909 #endif /* CONFIG_SMP */
5911 atomic_set(&rq
->nr_iowait
, 0);
5914 set_load_weight(&init_task
);
5917 * The boot idle thread does lazy MMU switching as well:
5920 enter_lazy_tlb(&init_mm
, current
);
5923 * Make us the idle thread. Technically, schedule() should not be
5924 * called from this thread, however somewhere below it might be,
5925 * but because we are the idle thread, we just pick up running again
5926 * when this runqueue becomes "idle".
5928 init_idle(current
, smp_processor_id());
5930 calc_load_update
= jiffies
+ LOAD_FREQ
;
5933 /* May be allocated at isolcpus cmdline parse time */
5934 if (cpu_isolated_map
== NULL
)
5935 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
5936 idle_thread_set_boot_cpu();
5937 set_cpu_rq_start_time(smp_processor_id());
5939 init_sched_fair_class();
5943 scheduler_running
= 1;
5946 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5947 static inline int preempt_count_equals(int preempt_offset
)
5949 int nested
= preempt_count() + rcu_preempt_depth();
5951 return (nested
== preempt_offset
);
5954 void __might_sleep(const char *file
, int line
, int preempt_offset
)
5957 * Blocking primitives will set (and therefore destroy) current->state,
5958 * since we will exit with TASK_RUNNING make sure we enter with it,
5959 * otherwise we will destroy state.
5961 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
5962 "do not call blocking ops when !TASK_RUNNING; "
5963 "state=%lx set at [<%p>] %pS\n",
5965 (void *)current
->task_state_change
,
5966 (void *)current
->task_state_change
);
5968 ___might_sleep(file
, line
, preempt_offset
);
5970 EXPORT_SYMBOL(__might_sleep
);
5972 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
5974 /* Ratelimiting timestamp: */
5975 static unsigned long prev_jiffy
;
5977 unsigned long preempt_disable_ip
;
5979 /* WARN_ON_ONCE() by default, no rate limit required: */
5982 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
5983 !is_idle_task(current
)) ||
5984 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
5988 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
5990 prev_jiffy
= jiffies
;
5992 /* Save this before calling printk(), since that will clobber it: */
5993 preempt_disable_ip
= get_preempt_disable_ip(current
);
5996 "BUG: sleeping function called from invalid context at %s:%d\n",
5999 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6000 in_atomic(), irqs_disabled(),
6001 current
->pid
, current
->comm
);
6003 if (task_stack_end_corrupted(current
))
6004 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6006 debug_show_held_locks(current
);
6007 if (irqs_disabled())
6008 print_irqtrace_events(current
);
6009 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6010 && !preempt_count_equals(preempt_offset
)) {
6011 pr_err("Preemption disabled at:");
6012 print_ip_sym(preempt_disable_ip
);
6016 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6018 EXPORT_SYMBOL(___might_sleep
);
6021 #ifdef CONFIG_MAGIC_SYSRQ
6022 void normalize_rt_tasks(void)
6024 struct task_struct
*g
, *p
;
6025 struct sched_attr attr
= {
6026 .sched_policy
= SCHED_NORMAL
,
6029 read_lock(&tasklist_lock
);
6030 for_each_process_thread(g
, p
) {
6032 * Only normalize user tasks:
6034 if (p
->flags
& PF_KTHREAD
)
6037 p
->se
.exec_start
= 0;
6038 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6039 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6040 schedstat_set(p
->se
.statistics
.block_start
, 0);
6042 if (!dl_task(p
) && !rt_task(p
)) {
6044 * Renice negative nice level userspace
6047 if (task_nice(p
) < 0)
6048 set_user_nice(p
, 0);
6052 __sched_setscheduler(p
, &attr
, false, false);
6054 read_unlock(&tasklist_lock
);
6057 #endif /* CONFIG_MAGIC_SYSRQ */
6059 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6061 * These functions are only useful for the IA64 MCA handling, or kdb.
6063 * They can only be called when the whole system has been
6064 * stopped - every CPU needs to be quiescent, and no scheduling
6065 * activity can take place. Using them for anything else would
6066 * be a serious bug, and as a result, they aren't even visible
6067 * under any other configuration.
6071 * curr_task - return the current task for a given CPU.
6072 * @cpu: the processor in question.
6074 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6076 * Return: The current task for @cpu.
6078 struct task_struct
*curr_task(int cpu
)
6080 return cpu_curr(cpu
);
6083 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6087 * set_curr_task - set the current task for a given CPU.
6088 * @cpu: the processor in question.
6089 * @p: the task pointer to set.
6091 * Description: This function must only be used when non-maskable interrupts
6092 * are serviced on a separate stack. It allows the architecture to switch the
6093 * notion of the current task on a CPU in a non-blocking manner. This function
6094 * must be called with all CPU's synchronized, and interrupts disabled, the
6095 * and caller must save the original value of the current task (see
6096 * curr_task() above) and restore that value before reenabling interrupts and
6097 * re-starting the system.
6099 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6101 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6108 #ifdef CONFIG_CGROUP_SCHED
6109 /* task_group_lock serializes the addition/removal of task groups */
6110 static DEFINE_SPINLOCK(task_group_lock
);
6112 static void sched_free_group(struct task_group
*tg
)
6114 free_fair_sched_group(tg
);
6115 free_rt_sched_group(tg
);
6117 kmem_cache_free(task_group_cache
, tg
);
6120 /* allocate runqueue etc for a new task group */
6121 struct task_group
*sched_create_group(struct task_group
*parent
)
6123 struct task_group
*tg
;
6125 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6127 return ERR_PTR(-ENOMEM
);
6129 if (!alloc_fair_sched_group(tg
, parent
))
6132 if (!alloc_rt_sched_group(tg
, parent
))
6138 sched_free_group(tg
);
6139 return ERR_PTR(-ENOMEM
);
6142 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6144 unsigned long flags
;
6146 spin_lock_irqsave(&task_group_lock
, flags
);
6147 list_add_rcu(&tg
->list
, &task_groups
);
6149 /* Root should already exist: */
6152 tg
->parent
= parent
;
6153 INIT_LIST_HEAD(&tg
->children
);
6154 list_add_rcu(&tg
->siblings
, &parent
->children
);
6155 spin_unlock_irqrestore(&task_group_lock
, flags
);
6157 online_fair_sched_group(tg
);
6160 /* rcu callback to free various structures associated with a task group */
6161 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6163 /* Now it should be safe to free those cfs_rqs: */
6164 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6167 void sched_destroy_group(struct task_group
*tg
)
6169 /* Wait for possible concurrent references to cfs_rqs complete: */
6170 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6173 void sched_offline_group(struct task_group
*tg
)
6175 unsigned long flags
;
6177 /* End participation in shares distribution: */
6178 unregister_fair_sched_group(tg
);
6180 spin_lock_irqsave(&task_group_lock
, flags
);
6181 list_del_rcu(&tg
->list
);
6182 list_del_rcu(&tg
->siblings
);
6183 spin_unlock_irqrestore(&task_group_lock
, flags
);
6186 static void sched_change_group(struct task_struct
*tsk
, int type
)
6188 struct task_group
*tg
;
6191 * All callers are synchronized by task_rq_lock(); we do not use RCU
6192 * which is pointless here. Thus, we pass "true" to task_css_check()
6193 * to prevent lockdep warnings.
6195 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
6196 struct task_group
, css
);
6197 tg
= autogroup_task_group(tsk
, tg
);
6198 tsk
->sched_task_group
= tg
;
6200 #ifdef CONFIG_FAIR_GROUP_SCHED
6201 if (tsk
->sched_class
->task_change_group
)
6202 tsk
->sched_class
->task_change_group(tsk
, type
);
6205 set_task_rq(tsk
, task_cpu(tsk
));
6209 * Change task's runqueue when it moves between groups.
6211 * The caller of this function should have put the task in its new group by
6212 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6215 void sched_move_task(struct task_struct
*tsk
)
6217 int queued
, running
, queue_flags
=
6218 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
6222 rq
= task_rq_lock(tsk
, &rf
);
6223 update_rq_clock(rq
);
6225 running
= task_current(rq
, tsk
);
6226 queued
= task_on_rq_queued(tsk
);
6229 dequeue_task(rq
, tsk
, queue_flags
);
6231 put_prev_task(rq
, tsk
);
6233 sched_change_group(tsk
, TASK_MOVE_GROUP
);
6236 enqueue_task(rq
, tsk
, queue_flags
);
6238 set_curr_task(rq
, tsk
);
6240 task_rq_unlock(rq
, tsk
, &rf
);
6243 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
6245 return css
? container_of(css
, struct task_group
, css
) : NULL
;
6248 static struct cgroup_subsys_state
*
6249 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6251 struct task_group
*parent
= css_tg(parent_css
);
6252 struct task_group
*tg
;
6255 /* This is early initialization for the top cgroup */
6256 return &root_task_group
.css
;
6259 tg
= sched_create_group(parent
);
6261 return ERR_PTR(-ENOMEM
);
6266 /* Expose task group only after completing cgroup initialization */
6267 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
6269 struct task_group
*tg
= css_tg(css
);
6270 struct task_group
*parent
= css_tg(css
->parent
);
6273 sched_online_group(tg
, parent
);
6277 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
6279 struct task_group
*tg
= css_tg(css
);
6281 sched_offline_group(tg
);
6284 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
6286 struct task_group
*tg
= css_tg(css
);
6289 * Relies on the RCU grace period between css_released() and this.
6291 sched_free_group(tg
);
6295 * This is called before wake_up_new_task(), therefore we really only
6296 * have to set its group bits, all the other stuff does not apply.
6298 static void cpu_cgroup_fork(struct task_struct
*task
)
6303 rq
= task_rq_lock(task
, &rf
);
6305 update_rq_clock(rq
);
6306 sched_change_group(task
, TASK_SET_GROUP
);
6308 task_rq_unlock(rq
, task
, &rf
);
6311 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
6313 struct task_struct
*task
;
6314 struct cgroup_subsys_state
*css
;
6317 cgroup_taskset_for_each(task
, css
, tset
) {
6318 #ifdef CONFIG_RT_GROUP_SCHED
6319 if (!sched_rt_can_attach(css_tg(css
), task
))
6322 /* We don't support RT-tasks being in separate groups */
6323 if (task
->sched_class
!= &fair_sched_class
)
6327 * Serialize against wake_up_new_task() such that if its
6328 * running, we're sure to observe its full state.
6330 raw_spin_lock_irq(&task
->pi_lock
);
6332 * Avoid calling sched_move_task() before wake_up_new_task()
6333 * has happened. This would lead to problems with PELT, due to
6334 * move wanting to detach+attach while we're not attached yet.
6336 if (task
->state
== TASK_NEW
)
6338 raw_spin_unlock_irq(&task
->pi_lock
);
6346 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
6348 struct task_struct
*task
;
6349 struct cgroup_subsys_state
*css
;
6351 cgroup_taskset_for_each(task
, css
, tset
)
6352 sched_move_task(task
);
6355 #ifdef CONFIG_FAIR_GROUP_SCHED
6356 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
6357 struct cftype
*cftype
, u64 shareval
)
6359 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
6362 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
6365 struct task_group
*tg
= css_tg(css
);
6367 return (u64
) scale_load_down(tg
->shares
);
6370 #ifdef CONFIG_CFS_BANDWIDTH
6371 static DEFINE_MUTEX(cfs_constraints_mutex
);
6373 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
6374 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
6376 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
6378 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
6380 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
6381 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6383 if (tg
== &root_task_group
)
6387 * Ensure we have at some amount of bandwidth every period. This is
6388 * to prevent reaching a state of large arrears when throttled via
6389 * entity_tick() resulting in prolonged exit starvation.
6391 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
6395 * Likewise, bound things on the otherside by preventing insane quota
6396 * periods. This also allows us to normalize in computing quota
6399 if (period
> max_cfs_quota_period
)
6403 * Prevent race between setting of cfs_rq->runtime_enabled and
6404 * unthrottle_offline_cfs_rqs().
6407 mutex_lock(&cfs_constraints_mutex
);
6408 ret
= __cfs_schedulable(tg
, period
, quota
);
6412 runtime_enabled
= quota
!= RUNTIME_INF
;
6413 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
6415 * If we need to toggle cfs_bandwidth_used, off->on must occur
6416 * before making related changes, and on->off must occur afterwards
6418 if (runtime_enabled
&& !runtime_was_enabled
)
6419 cfs_bandwidth_usage_inc();
6420 raw_spin_lock_irq(&cfs_b
->lock
);
6421 cfs_b
->period
= ns_to_ktime(period
);
6422 cfs_b
->quota
= quota
;
6424 __refill_cfs_bandwidth_runtime(cfs_b
);
6426 /* Restart the period timer (if active) to handle new period expiry: */
6427 if (runtime_enabled
)
6428 start_cfs_bandwidth(cfs_b
);
6430 raw_spin_unlock_irq(&cfs_b
->lock
);
6432 for_each_online_cpu(i
) {
6433 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
6434 struct rq
*rq
= cfs_rq
->rq
;
6437 rq_lock_irq(rq
, &rf
);
6438 cfs_rq
->runtime_enabled
= runtime_enabled
;
6439 cfs_rq
->runtime_remaining
= 0;
6441 if (cfs_rq
->throttled
)
6442 unthrottle_cfs_rq(cfs_rq
);
6443 rq_unlock_irq(rq
, &rf
);
6445 if (runtime_was_enabled
&& !runtime_enabled
)
6446 cfs_bandwidth_usage_dec();
6448 mutex_unlock(&cfs_constraints_mutex
);
6454 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
6458 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6459 if (cfs_quota_us
< 0)
6460 quota
= RUNTIME_INF
;
6462 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
6464 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6467 long tg_get_cfs_quota(struct task_group
*tg
)
6471 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
6474 quota_us
= tg
->cfs_bandwidth
.quota
;
6475 do_div(quota_us
, NSEC_PER_USEC
);
6480 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
6484 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
6485 quota
= tg
->cfs_bandwidth
.quota
;
6487 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6490 long tg_get_cfs_period(struct task_group
*tg
)
6494 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6495 do_div(cfs_period_us
, NSEC_PER_USEC
);
6497 return cfs_period_us
;
6500 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
6503 return tg_get_cfs_quota(css_tg(css
));
6506 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
6507 struct cftype
*cftype
, s64 cfs_quota_us
)
6509 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
6512 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
6515 return tg_get_cfs_period(css_tg(css
));
6518 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
6519 struct cftype
*cftype
, u64 cfs_period_us
)
6521 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
6524 struct cfs_schedulable_data
{
6525 struct task_group
*tg
;
6530 * normalize group quota/period to be quota/max_period
6531 * note: units are usecs
6533 static u64
normalize_cfs_quota(struct task_group
*tg
,
6534 struct cfs_schedulable_data
*d
)
6542 period
= tg_get_cfs_period(tg
);
6543 quota
= tg_get_cfs_quota(tg
);
6546 /* note: these should typically be equivalent */
6547 if (quota
== RUNTIME_INF
|| quota
== -1)
6550 return to_ratio(period
, quota
);
6553 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
6555 struct cfs_schedulable_data
*d
= data
;
6556 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6557 s64 quota
= 0, parent_quota
= -1;
6560 quota
= RUNTIME_INF
;
6562 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
6564 quota
= normalize_cfs_quota(tg
, d
);
6565 parent_quota
= parent_b
->hierarchical_quota
;
6568 * Ensure max(child_quota) <= parent_quota, inherit when no
6571 if (quota
== RUNTIME_INF
)
6572 quota
= parent_quota
;
6573 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
6576 cfs_b
->hierarchical_quota
= quota
;
6581 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
6584 struct cfs_schedulable_data data
= {
6590 if (quota
!= RUNTIME_INF
) {
6591 do_div(data
.period
, NSEC_PER_USEC
);
6592 do_div(data
.quota
, NSEC_PER_USEC
);
6596 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
6602 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
6604 struct task_group
*tg
= css_tg(seq_css(sf
));
6605 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6607 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
6608 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
6609 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
6613 #endif /* CONFIG_CFS_BANDWIDTH */
6614 #endif /* CONFIG_FAIR_GROUP_SCHED */
6616 #ifdef CONFIG_RT_GROUP_SCHED
6617 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
6618 struct cftype
*cft
, s64 val
)
6620 return sched_group_set_rt_runtime(css_tg(css
), val
);
6623 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
6626 return sched_group_rt_runtime(css_tg(css
));
6629 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
6630 struct cftype
*cftype
, u64 rt_period_us
)
6632 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
6635 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
6638 return sched_group_rt_period(css_tg(css
));
6640 #endif /* CONFIG_RT_GROUP_SCHED */
6642 static struct cftype cpu_files
[] = {
6643 #ifdef CONFIG_FAIR_GROUP_SCHED
6646 .read_u64
= cpu_shares_read_u64
,
6647 .write_u64
= cpu_shares_write_u64
,
6650 #ifdef CONFIG_CFS_BANDWIDTH
6652 .name
= "cfs_quota_us",
6653 .read_s64
= cpu_cfs_quota_read_s64
,
6654 .write_s64
= cpu_cfs_quota_write_s64
,
6657 .name
= "cfs_period_us",
6658 .read_u64
= cpu_cfs_period_read_u64
,
6659 .write_u64
= cpu_cfs_period_write_u64
,
6663 .seq_show
= cpu_stats_show
,
6666 #ifdef CONFIG_RT_GROUP_SCHED
6668 .name
= "rt_runtime_us",
6669 .read_s64
= cpu_rt_runtime_read
,
6670 .write_s64
= cpu_rt_runtime_write
,
6673 .name
= "rt_period_us",
6674 .read_u64
= cpu_rt_period_read_uint
,
6675 .write_u64
= cpu_rt_period_write_uint
,
6681 struct cgroup_subsys cpu_cgrp_subsys
= {
6682 .css_alloc
= cpu_cgroup_css_alloc
,
6683 .css_online
= cpu_cgroup_css_online
,
6684 .css_released
= cpu_cgroup_css_released
,
6685 .css_free
= cpu_cgroup_css_free
,
6686 .fork
= cpu_cgroup_fork
,
6687 .can_attach
= cpu_cgroup_can_attach
,
6688 .attach
= cpu_cgroup_attach
,
6689 .legacy_cftypes
= cpu_files
,
6693 #endif /* CONFIG_CGROUP_SCHED */
6695 void dump_cpu_task(int cpu
)
6697 pr_info("Task dump for CPU %d:\n", cpu
);
6698 sched_show_task(cpu_curr(cpu
));
6702 * Nice levels are multiplicative, with a gentle 10% change for every
6703 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
6704 * nice 1, it will get ~10% less CPU time than another CPU-bound task
6705 * that remained on nice 0.
6707 * The "10% effect" is relative and cumulative: from _any_ nice level,
6708 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
6709 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
6710 * If a task goes up by ~10% and another task goes down by ~10% then
6711 * the relative distance between them is ~25%.)
6713 const int sched_prio_to_weight
[40] = {
6714 /* -20 */ 88761, 71755, 56483, 46273, 36291,
6715 /* -15 */ 29154, 23254, 18705, 14949, 11916,
6716 /* -10 */ 9548, 7620, 6100, 4904, 3906,
6717 /* -5 */ 3121, 2501, 1991, 1586, 1277,
6718 /* 0 */ 1024, 820, 655, 526, 423,
6719 /* 5 */ 335, 272, 215, 172, 137,
6720 /* 10 */ 110, 87, 70, 56, 45,
6721 /* 15 */ 36, 29, 23, 18, 15,
6725 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
6727 * In cases where the weight does not change often, we can use the
6728 * precalculated inverse to speed up arithmetics by turning divisions
6729 * into multiplications:
6731 const u32 sched_prio_to_wmult
[40] = {
6732 /* -20 */ 48388, 59856, 76040, 92818, 118348,
6733 /* -15 */ 147320, 184698, 229616, 287308, 360437,
6734 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
6735 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
6736 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
6737 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
6738 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
6739 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,