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/cpuset.h>
14 #include <linux/delayacct.h>
15 #include <linux/init_task.h>
16 #include <linux/context_tracking.h>
17 #include <linux/rcupdate_wait.h>
19 #include <linux/blkdev.h>
20 #include <linux/kprobes.h>
21 #include <linux/mmu_context.h>
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/prefetch.h>
25 #include <linux/profile.h>
26 #include <linux/security.h>
27 #include <linux/syscalls.h>
29 #include <asm/switch_to.h>
31 #ifdef CONFIG_PARAVIRT
32 #include <asm/paravirt.h>
36 #include "../workqueue_internal.h"
37 #include "../smpboot.h"
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/sched.h>
42 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
45 * Debugging: various feature bits
48 #define SCHED_FEAT(name, enabled) \
49 (1UL << __SCHED_FEAT_##name) * enabled |
51 const_debug
unsigned int sysctl_sched_features
=
58 * Number of tasks to iterate in a single balance run.
59 * Limited because this is done with IRQs disabled.
61 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
64 * period over which we average the RT time consumption, measured
69 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
72 * period over which we measure -rt task CPU usage in us.
75 unsigned int sysctl_sched_rt_period
= 1000000;
77 __read_mostly
int scheduler_running
;
80 * part of the period that we allow rt tasks to run in us.
83 int sysctl_sched_rt_runtime
= 950000;
85 /* CPUs with isolated domains */
86 cpumask_var_t cpu_isolated_map
;
89 * __task_rq_lock - lock the rq @p resides on.
91 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
96 lockdep_assert_held(&p
->pi_lock
);
100 raw_spin_lock(&rq
->lock
);
101 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
105 raw_spin_unlock(&rq
->lock
);
107 while (unlikely(task_on_rq_migrating(p
)))
113 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
115 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
116 __acquires(p
->pi_lock
)
122 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
124 raw_spin_lock(&rq
->lock
);
126 * move_queued_task() task_rq_lock()
129 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
130 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
131 * [S] ->cpu = new_cpu [L] task_rq()
135 * If we observe the old cpu in task_rq_lock, the acquire of
136 * the old rq->lock will fully serialize against the stores.
138 * If we observe the new CPU in task_rq_lock, the acquire will
139 * pair with the WMB to ensure we must then also see migrating.
141 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
145 raw_spin_unlock(&rq
->lock
);
146 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
148 while (unlikely(task_on_rq_migrating(p
)))
154 * RQ-clock updating methods:
157 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
160 * In theory, the compile should just see 0 here, and optimize out the call
161 * to sched_rt_avg_update. But I don't trust it...
163 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
164 s64 steal
= 0, irq_delta
= 0;
166 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
167 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
170 * Since irq_time is only updated on {soft,}irq_exit, we might run into
171 * this case when a previous update_rq_clock() happened inside a
174 * When this happens, we stop ->clock_task and only update the
175 * prev_irq_time stamp to account for the part that fit, so that a next
176 * update will consume the rest. This ensures ->clock_task is
179 * It does however cause some slight miss-attribution of {soft,}irq
180 * time, a more accurate solution would be to update the irq_time using
181 * the current rq->clock timestamp, except that would require using
184 if (irq_delta
> delta
)
187 rq
->prev_irq_time
+= irq_delta
;
190 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
191 if (static_key_false((¶virt_steal_rq_enabled
))) {
192 steal
= paravirt_steal_clock(cpu_of(rq
));
193 steal
-= rq
->prev_steal_time_rq
;
195 if (unlikely(steal
> delta
))
198 rq
->prev_steal_time_rq
+= steal
;
203 rq
->clock_task
+= delta
;
205 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
206 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
207 sched_rt_avg_update(rq
, irq_delta
+ steal
);
211 void update_rq_clock(struct rq
*rq
)
215 lockdep_assert_held(&rq
->lock
);
217 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
220 #ifdef CONFIG_SCHED_DEBUG
221 if (sched_feat(WARN_DOUBLE_CLOCK
))
222 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
223 rq
->clock_update_flags
|= RQCF_UPDATED
;
226 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
230 update_rq_clock_task(rq
, delta
);
234 #ifdef CONFIG_SCHED_HRTICK
236 * Use HR-timers to deliver accurate preemption points.
239 static void hrtick_clear(struct rq
*rq
)
241 if (hrtimer_active(&rq
->hrtick_timer
))
242 hrtimer_cancel(&rq
->hrtick_timer
);
246 * High-resolution timer tick.
247 * Runs from hardirq context with interrupts disabled.
249 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
251 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
254 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
258 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
261 return HRTIMER_NORESTART
;
266 static void __hrtick_restart(struct rq
*rq
)
268 struct hrtimer
*timer
= &rq
->hrtick_timer
;
270 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
274 * called from hardirq (IPI) context
276 static void __hrtick_start(void *arg
)
282 __hrtick_restart(rq
);
283 rq
->hrtick_csd_pending
= 0;
288 * Called to set the hrtick timer state.
290 * called with rq->lock held and irqs disabled
292 void hrtick_start(struct rq
*rq
, u64 delay
)
294 struct hrtimer
*timer
= &rq
->hrtick_timer
;
299 * Don't schedule slices shorter than 10000ns, that just
300 * doesn't make sense and can cause timer DoS.
302 delta
= max_t(s64
, delay
, 10000LL);
303 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
305 hrtimer_set_expires(timer
, time
);
307 if (rq
== this_rq()) {
308 __hrtick_restart(rq
);
309 } else if (!rq
->hrtick_csd_pending
) {
310 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
311 rq
->hrtick_csd_pending
= 1;
317 * Called to set the hrtick timer state.
319 * called with rq->lock held and irqs disabled
321 void hrtick_start(struct rq
*rq
, u64 delay
)
324 * Don't schedule slices shorter than 10000ns, that just
325 * doesn't make sense. Rely on vruntime for fairness.
327 delay
= max_t(u64
, delay
, 10000LL);
328 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
329 HRTIMER_MODE_REL_PINNED
);
331 #endif /* CONFIG_SMP */
333 static void init_rq_hrtick(struct rq
*rq
)
336 rq
->hrtick_csd_pending
= 0;
338 rq
->hrtick_csd
.flags
= 0;
339 rq
->hrtick_csd
.func
= __hrtick_start
;
340 rq
->hrtick_csd
.info
= rq
;
343 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
344 rq
->hrtick_timer
.function
= hrtick
;
346 #else /* CONFIG_SCHED_HRTICK */
347 static inline void hrtick_clear(struct rq
*rq
)
351 static inline void init_rq_hrtick(struct rq
*rq
)
354 #endif /* CONFIG_SCHED_HRTICK */
357 * cmpxchg based fetch_or, macro so it works for different integer types
359 #define fetch_or(ptr, mask) \
361 typeof(ptr) _ptr = (ptr); \
362 typeof(mask) _mask = (mask); \
363 typeof(*_ptr) _old, _val = *_ptr; \
366 _old = cmpxchg(_ptr, _val, _val | _mask); \
374 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
376 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
377 * this avoids any races wrt polling state changes and thereby avoids
380 static bool set_nr_and_not_polling(struct task_struct
*p
)
382 struct thread_info
*ti
= task_thread_info(p
);
383 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
387 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
389 * If this returns true, then the idle task promises to call
390 * sched_ttwu_pending() and reschedule soon.
392 static bool set_nr_if_polling(struct task_struct
*p
)
394 struct thread_info
*ti
= task_thread_info(p
);
395 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
398 if (!(val
& _TIF_POLLING_NRFLAG
))
400 if (val
& _TIF_NEED_RESCHED
)
402 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
411 static bool set_nr_and_not_polling(struct task_struct
*p
)
413 set_tsk_need_resched(p
);
418 static bool set_nr_if_polling(struct task_struct
*p
)
425 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
427 struct wake_q_node
*node
= &task
->wake_q
;
430 * Atomically grab the task, if ->wake_q is !nil already it means
431 * its already queued (either by us or someone else) and will get the
432 * wakeup due to that.
434 * This cmpxchg() implies a full barrier, which pairs with the write
435 * barrier implied by the wakeup in wake_up_q().
437 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
440 get_task_struct(task
);
443 * The head is context local, there can be no concurrency.
446 head
->lastp
= &node
->next
;
449 void wake_up_q(struct wake_q_head
*head
)
451 struct wake_q_node
*node
= head
->first
;
453 while (node
!= WAKE_Q_TAIL
) {
454 struct task_struct
*task
;
456 task
= container_of(node
, struct task_struct
, wake_q
);
458 /* Task can safely be re-inserted now: */
460 task
->wake_q
.next
= NULL
;
463 * wake_up_process() implies a wmb() to pair with the queueing
464 * in wake_q_add() so as not to miss wakeups.
466 wake_up_process(task
);
467 put_task_struct(task
);
472 * resched_curr - mark rq's current task 'to be rescheduled now'.
474 * On UP this means the setting of the need_resched flag, on SMP it
475 * might also involve a cross-CPU call to trigger the scheduler on
478 void resched_curr(struct rq
*rq
)
480 struct task_struct
*curr
= rq
->curr
;
483 lockdep_assert_held(&rq
->lock
);
485 if (test_tsk_need_resched(curr
))
490 if (cpu
== smp_processor_id()) {
491 set_tsk_need_resched(curr
);
492 set_preempt_need_resched();
496 if (set_nr_and_not_polling(curr
))
497 smp_send_reschedule(cpu
);
499 trace_sched_wake_idle_without_ipi(cpu
);
502 void resched_cpu(int cpu
)
504 struct rq
*rq
= cpu_rq(cpu
);
507 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
510 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
514 #ifdef CONFIG_NO_HZ_COMMON
516 * In the semi idle case, use the nearest busy CPU for migrating timers
517 * from an idle CPU. This is good for power-savings.
519 * We don't do similar optimization for completely idle system, as
520 * selecting an idle CPU will add more delays to the timers than intended
521 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
523 int get_nohz_timer_target(void)
525 int i
, cpu
= smp_processor_id();
526 struct sched_domain
*sd
;
528 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
532 for_each_domain(cpu
, sd
) {
533 for_each_cpu(i
, sched_domain_span(sd
)) {
537 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
544 if (!is_housekeeping_cpu(cpu
))
545 cpu
= housekeeping_any_cpu();
552 * When add_timer_on() enqueues a timer into the timer wheel of an
553 * idle CPU then this timer might expire before the next timer event
554 * which is scheduled to wake up that CPU. In case of a completely
555 * idle system the next event might even be infinite time into the
556 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
557 * leaves the inner idle loop so the newly added timer is taken into
558 * account when the CPU goes back to idle and evaluates the timer
559 * wheel for the next timer event.
561 static void wake_up_idle_cpu(int cpu
)
563 struct rq
*rq
= cpu_rq(cpu
);
565 if (cpu
== smp_processor_id())
568 if (set_nr_and_not_polling(rq
->idle
))
569 smp_send_reschedule(cpu
);
571 trace_sched_wake_idle_without_ipi(cpu
);
574 static bool wake_up_full_nohz_cpu(int cpu
)
577 * We just need the target to call irq_exit() and re-evaluate
578 * the next tick. The nohz full kick at least implies that.
579 * If needed we can still optimize that later with an
582 if (cpu_is_offline(cpu
))
583 return true; /* Don't try to wake offline CPUs. */
584 if (tick_nohz_full_cpu(cpu
)) {
585 if (cpu
!= smp_processor_id() ||
586 tick_nohz_tick_stopped())
587 tick_nohz_full_kick_cpu(cpu
);
595 * Wake up the specified CPU. If the CPU is going offline, it is the
596 * caller's responsibility to deal with the lost wakeup, for example,
597 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
599 void wake_up_nohz_cpu(int cpu
)
601 if (!wake_up_full_nohz_cpu(cpu
))
602 wake_up_idle_cpu(cpu
);
605 static inline bool got_nohz_idle_kick(void)
607 int cpu
= smp_processor_id();
609 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
612 if (idle_cpu(cpu
) && !need_resched())
616 * We can't run Idle Load Balance on this CPU for this time so we
617 * cancel it and clear NOHZ_BALANCE_KICK
619 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
623 #else /* CONFIG_NO_HZ_COMMON */
625 static inline bool got_nohz_idle_kick(void)
630 #endif /* CONFIG_NO_HZ_COMMON */
632 #ifdef CONFIG_NO_HZ_FULL
633 bool sched_can_stop_tick(struct rq
*rq
)
637 /* Deadline tasks, even if single, need the tick */
638 if (rq
->dl
.dl_nr_running
)
642 * If there are more than one RR tasks, we need the tick to effect the
643 * actual RR behaviour.
645 if (rq
->rt
.rr_nr_running
) {
646 if (rq
->rt
.rr_nr_running
== 1)
653 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
654 * forced preemption between FIFO tasks.
656 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
661 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
662 * if there's more than one we need the tick for involuntary
665 if (rq
->nr_running
> 1)
670 #endif /* CONFIG_NO_HZ_FULL */
672 void sched_avg_update(struct rq
*rq
)
674 s64 period
= sched_avg_period();
676 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
678 * Inline assembly required to prevent the compiler
679 * optimising this loop into a divmod call.
680 * See __iter_div_u64_rem() for another example of this.
682 asm("" : "+rm" (rq
->age_stamp
));
683 rq
->age_stamp
+= period
;
688 #endif /* CONFIG_SMP */
690 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
691 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
693 * Iterate task_group tree rooted at *from, calling @down when first entering a
694 * node and @up when leaving it for the final time.
696 * Caller must hold rcu_lock or sufficient equivalent.
698 int walk_tg_tree_from(struct task_group
*from
,
699 tg_visitor down
, tg_visitor up
, void *data
)
701 struct task_group
*parent
, *child
;
707 ret
= (*down
)(parent
, data
);
710 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
717 ret
= (*up
)(parent
, data
);
718 if (ret
|| parent
== from
)
722 parent
= parent
->parent
;
729 int tg_nop(struct task_group
*tg
, void *data
)
735 static void set_load_weight(struct task_struct
*p
)
737 int prio
= p
->static_prio
- MAX_RT_PRIO
;
738 struct load_weight
*load
= &p
->se
.load
;
741 * SCHED_IDLE tasks get minimal weight:
743 if (idle_policy(p
->policy
)) {
744 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
745 load
->inv_weight
= WMULT_IDLEPRIO
;
749 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
750 load
->inv_weight
= sched_prio_to_wmult
[prio
];
753 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
755 if (!(flags
& ENQUEUE_NOCLOCK
))
758 if (!(flags
& ENQUEUE_RESTORE
))
759 sched_info_queued(rq
, p
);
761 p
->sched_class
->enqueue_task(rq
, p
, flags
);
764 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 if (!(flags
& DEQUEUE_NOCLOCK
))
769 if (!(flags
& DEQUEUE_SAVE
))
770 sched_info_dequeued(rq
, p
);
772 p
->sched_class
->dequeue_task(rq
, p
, flags
);
775 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
777 if (task_contributes_to_load(p
))
778 rq
->nr_uninterruptible
--;
780 enqueue_task(rq
, p
, flags
);
783 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
785 if (task_contributes_to_load(p
))
786 rq
->nr_uninterruptible
++;
788 dequeue_task(rq
, p
, flags
);
791 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
793 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
794 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
798 * Make it appear like a SCHED_FIFO task, its something
799 * userspace knows about and won't get confused about.
801 * Also, it will make PI more or less work without too
802 * much confusion -- but then, stop work should not
803 * rely on PI working anyway.
805 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
807 stop
->sched_class
= &stop_sched_class
;
810 cpu_rq(cpu
)->stop
= stop
;
814 * Reset it back to a normal scheduling class so that
815 * it can die in pieces.
817 old_stop
->sched_class
= &rt_sched_class
;
822 * __normal_prio - return the priority that is based on the static prio
824 static inline int __normal_prio(struct task_struct
*p
)
826 return p
->static_prio
;
830 * Calculate the expected normal priority: i.e. priority
831 * without taking RT-inheritance into account. Might be
832 * boosted by interactivity modifiers. Changes upon fork,
833 * setprio syscalls, and whenever the interactivity
834 * estimator recalculates.
836 static inline int normal_prio(struct task_struct
*p
)
840 if (task_has_dl_policy(p
))
841 prio
= MAX_DL_PRIO
-1;
842 else if (task_has_rt_policy(p
))
843 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
845 prio
= __normal_prio(p
);
850 * Calculate the current priority, i.e. the priority
851 * taken into account by the scheduler. This value might
852 * be boosted by RT tasks, or might be boosted by
853 * interactivity modifiers. Will be RT if the task got
854 * RT-boosted. If not then it returns p->normal_prio.
856 static int effective_prio(struct task_struct
*p
)
858 p
->normal_prio
= normal_prio(p
);
860 * If we are RT tasks or we were boosted to RT priority,
861 * keep the priority unchanged. Otherwise, update priority
862 * to the normal priority:
864 if (!rt_prio(p
->prio
))
865 return p
->normal_prio
;
870 * task_curr - is this task currently executing on a CPU?
871 * @p: the task in question.
873 * Return: 1 if the task is currently executing. 0 otherwise.
875 inline int task_curr(const struct task_struct
*p
)
877 return cpu_curr(task_cpu(p
)) == p
;
881 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
882 * use the balance_callback list if you want balancing.
884 * this means any call to check_class_changed() must be followed by a call to
885 * balance_callback().
887 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
888 const struct sched_class
*prev_class
,
891 if (prev_class
!= p
->sched_class
) {
892 if (prev_class
->switched_from
)
893 prev_class
->switched_from(rq
, p
);
895 p
->sched_class
->switched_to(rq
, p
);
896 } else if (oldprio
!= p
->prio
|| dl_task(p
))
897 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
900 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
902 const struct sched_class
*class;
904 if (p
->sched_class
== rq
->curr
->sched_class
) {
905 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
907 for_each_class(class) {
908 if (class == rq
->curr
->sched_class
)
910 if (class == p
->sched_class
) {
918 * A queue event has occurred, and we're going to schedule. In
919 * this case, we can save a useless back to back clock update.
921 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
922 rq_clock_skip_update(rq
, true);
927 * This is how migration works:
929 * 1) we invoke migration_cpu_stop() on the target CPU using
931 * 2) stopper starts to run (implicitly forcing the migrated thread
933 * 3) it checks whether the migrated task is still in the wrong runqueue.
934 * 4) if it's in the wrong runqueue then the migration thread removes
935 * it and puts it into the right queue.
936 * 5) stopper completes and stop_one_cpu() returns and the migration
941 * move_queued_task - move a queued task to new rq.
943 * Returns (locked) new rq. Old rq's lock is released.
945 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
946 struct task_struct
*p
, int new_cpu
)
948 lockdep_assert_held(&rq
->lock
);
950 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
951 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
952 set_task_cpu(p
, new_cpu
);
955 rq
= cpu_rq(new_cpu
);
958 BUG_ON(task_cpu(p
) != new_cpu
);
959 enqueue_task(rq
, p
, 0);
960 p
->on_rq
= TASK_ON_RQ_QUEUED
;
961 check_preempt_curr(rq
, p
, 0);
966 struct migration_arg
{
967 struct task_struct
*task
;
972 * Move (not current) task off this CPU, onto the destination CPU. We're doing
973 * this because either it can't run here any more (set_cpus_allowed()
974 * away from this CPU, or CPU going down), or because we're
975 * attempting to rebalance this task on exec (sched_exec).
977 * So we race with normal scheduler movements, but that's OK, as long
978 * as the task is no longer on this CPU.
980 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
981 struct task_struct
*p
, int dest_cpu
)
983 if (unlikely(!cpu_active(dest_cpu
)))
986 /* Affinity changed (again). */
987 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
991 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
997 * migration_cpu_stop - this will be executed by a highprio stopper thread
998 * and performs thread migration by bumping thread off CPU then
999 * 'pushing' onto another runqueue.
1001 static int migration_cpu_stop(void *data
)
1003 struct migration_arg
*arg
= data
;
1004 struct task_struct
*p
= arg
->task
;
1005 struct rq
*rq
= this_rq();
1009 * The original target CPU might have gone down and we might
1010 * be on another CPU but it doesn't matter.
1012 local_irq_disable();
1014 * We need to explicitly wake pending tasks before running
1015 * __migrate_task() such that we will not miss enforcing cpus_allowed
1016 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1018 sched_ttwu_pending();
1020 raw_spin_lock(&p
->pi_lock
);
1023 * If task_rq(p) != rq, it cannot be migrated here, because we're
1024 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1025 * we're holding p->pi_lock.
1027 if (task_rq(p
) == rq
) {
1028 if (task_on_rq_queued(p
))
1029 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1031 p
->wake_cpu
= arg
->dest_cpu
;
1034 raw_spin_unlock(&p
->pi_lock
);
1041 * sched_class::set_cpus_allowed must do the below, but is not required to
1042 * actually call this function.
1044 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1046 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1047 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1050 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1052 struct rq
*rq
= task_rq(p
);
1053 bool queued
, running
;
1055 lockdep_assert_held(&p
->pi_lock
);
1057 queued
= task_on_rq_queued(p
);
1058 running
= task_current(rq
, p
);
1062 * Because __kthread_bind() calls this on blocked tasks without
1065 lockdep_assert_held(&rq
->lock
);
1066 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1069 put_prev_task(rq
, p
);
1071 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1074 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1076 set_curr_task(rq
, p
);
1080 * Change a given task's CPU affinity. Migrate the thread to a
1081 * proper CPU and schedule it away if the CPU it's executing on
1082 * is removed from the allowed bitmask.
1084 * NOTE: the caller must have a valid reference to the task, the
1085 * task must not exit() & deallocate itself prematurely. The
1086 * call is not atomic; no spinlocks may be held.
1088 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1089 const struct cpumask
*new_mask
, bool check
)
1091 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1092 unsigned int dest_cpu
;
1097 rq
= task_rq_lock(p
, &rf
);
1098 update_rq_clock(rq
);
1100 if (p
->flags
& PF_KTHREAD
) {
1102 * Kernel threads are allowed on online && !active CPUs
1104 cpu_valid_mask
= cpu_online_mask
;
1108 * Must re-check here, to close a race against __kthread_bind(),
1109 * sched_setaffinity() is not guaranteed to observe the flag.
1111 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1116 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1119 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1124 do_set_cpus_allowed(p
, new_mask
);
1126 if (p
->flags
& PF_KTHREAD
) {
1128 * For kernel threads that do indeed end up on online &&
1129 * !active we want to ensure they are strict per-CPU threads.
1131 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1132 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1133 p
->nr_cpus_allowed
!= 1);
1136 /* Can the task run on the task's current CPU? If so, we're done */
1137 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1140 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1141 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1142 struct migration_arg arg
= { p
, dest_cpu
};
1143 /* Need help from migration thread: drop lock and wait. */
1144 task_rq_unlock(rq
, p
, &rf
);
1145 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1146 tlb_migrate_finish(p
->mm
);
1148 } else if (task_on_rq_queued(p
)) {
1150 * OK, since we're going to drop the lock immediately
1151 * afterwards anyway.
1153 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1156 task_rq_unlock(rq
, p
, &rf
);
1161 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1163 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1165 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1167 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1169 #ifdef CONFIG_SCHED_DEBUG
1171 * We should never call set_task_cpu() on a blocked task,
1172 * ttwu() will sort out the placement.
1174 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1178 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1179 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1180 * time relying on p->on_rq.
1182 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1183 p
->sched_class
== &fair_sched_class
&&
1184 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1186 #ifdef CONFIG_LOCKDEP
1188 * The caller should hold either p->pi_lock or rq->lock, when changing
1189 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1191 * sched_move_task() holds both and thus holding either pins the cgroup,
1194 * Furthermore, all task_rq users should acquire both locks, see
1197 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1198 lockdep_is_held(&task_rq(p
)->lock
)));
1202 trace_sched_migrate_task(p
, new_cpu
);
1204 if (task_cpu(p
) != new_cpu
) {
1205 if (p
->sched_class
->migrate_task_rq
)
1206 p
->sched_class
->migrate_task_rq(p
);
1207 p
->se
.nr_migrations
++;
1208 perf_event_task_migrate(p
);
1211 __set_task_cpu(p
, new_cpu
);
1214 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1216 if (task_on_rq_queued(p
)) {
1217 struct rq
*src_rq
, *dst_rq
;
1218 struct rq_flags srf
, drf
;
1220 src_rq
= task_rq(p
);
1221 dst_rq
= cpu_rq(cpu
);
1223 rq_pin_lock(src_rq
, &srf
);
1224 rq_pin_lock(dst_rq
, &drf
);
1226 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1227 deactivate_task(src_rq
, p
, 0);
1228 set_task_cpu(p
, cpu
);
1229 activate_task(dst_rq
, p
, 0);
1230 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1231 check_preempt_curr(dst_rq
, p
, 0);
1233 rq_unpin_lock(dst_rq
, &drf
);
1234 rq_unpin_lock(src_rq
, &srf
);
1238 * Task isn't running anymore; make it appear like we migrated
1239 * it before it went to sleep. This means on wakeup we make the
1240 * previous CPU our target instead of where it really is.
1246 struct migration_swap_arg
{
1247 struct task_struct
*src_task
, *dst_task
;
1248 int src_cpu
, dst_cpu
;
1251 static int migrate_swap_stop(void *data
)
1253 struct migration_swap_arg
*arg
= data
;
1254 struct rq
*src_rq
, *dst_rq
;
1257 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1260 src_rq
= cpu_rq(arg
->src_cpu
);
1261 dst_rq
= cpu_rq(arg
->dst_cpu
);
1263 double_raw_lock(&arg
->src_task
->pi_lock
,
1264 &arg
->dst_task
->pi_lock
);
1265 double_rq_lock(src_rq
, dst_rq
);
1267 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1270 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1273 if (!cpumask_test_cpu(arg
->dst_cpu
, &arg
->src_task
->cpus_allowed
))
1276 if (!cpumask_test_cpu(arg
->src_cpu
, &arg
->dst_task
->cpus_allowed
))
1279 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1280 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1285 double_rq_unlock(src_rq
, dst_rq
);
1286 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1287 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1293 * Cross migrate two tasks
1295 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1297 struct migration_swap_arg arg
;
1300 arg
= (struct migration_swap_arg
){
1302 .src_cpu
= task_cpu(cur
),
1304 .dst_cpu
= task_cpu(p
),
1307 if (arg
.src_cpu
== arg
.dst_cpu
)
1311 * These three tests are all lockless; this is OK since all of them
1312 * will be re-checked with proper locks held further down the line.
1314 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1317 if (!cpumask_test_cpu(arg
.dst_cpu
, &arg
.src_task
->cpus_allowed
))
1320 if (!cpumask_test_cpu(arg
.src_cpu
, &arg
.dst_task
->cpus_allowed
))
1323 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1324 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1331 * wait_task_inactive - wait for a thread to unschedule.
1333 * If @match_state is nonzero, it's the @p->state value just checked and
1334 * not expected to change. If it changes, i.e. @p might have woken up,
1335 * then return zero. When we succeed in waiting for @p to be off its CPU,
1336 * we return a positive number (its total switch count). If a second call
1337 * a short while later returns the same number, the caller can be sure that
1338 * @p has remained unscheduled the whole time.
1340 * The caller must ensure that the task *will* unschedule sometime soon,
1341 * else this function might spin for a *long* time. This function can't
1342 * be called with interrupts off, or it may introduce deadlock with
1343 * smp_call_function() if an IPI is sent by the same process we are
1344 * waiting to become inactive.
1346 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1348 int running
, queued
;
1355 * We do the initial early heuristics without holding
1356 * any task-queue locks at all. We'll only try to get
1357 * the runqueue lock when things look like they will
1363 * If the task is actively running on another CPU
1364 * still, just relax and busy-wait without holding
1367 * NOTE! Since we don't hold any locks, it's not
1368 * even sure that "rq" stays as the right runqueue!
1369 * But we don't care, since "task_running()" will
1370 * return false if the runqueue has changed and p
1371 * is actually now running somewhere else!
1373 while (task_running(rq
, p
)) {
1374 if (match_state
&& unlikely(p
->state
!= match_state
))
1380 * Ok, time to look more closely! We need the rq
1381 * lock now, to be *sure*. If we're wrong, we'll
1382 * just go back and repeat.
1384 rq
= task_rq_lock(p
, &rf
);
1385 trace_sched_wait_task(p
);
1386 running
= task_running(rq
, p
);
1387 queued
= task_on_rq_queued(p
);
1389 if (!match_state
|| p
->state
== match_state
)
1390 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1391 task_rq_unlock(rq
, p
, &rf
);
1394 * If it changed from the expected state, bail out now.
1396 if (unlikely(!ncsw
))
1400 * Was it really running after all now that we
1401 * checked with the proper locks actually held?
1403 * Oops. Go back and try again..
1405 if (unlikely(running
)) {
1411 * It's not enough that it's not actively running,
1412 * it must be off the runqueue _entirely_, and not
1415 * So if it was still runnable (but just not actively
1416 * running right now), it's preempted, and we should
1417 * yield - it could be a while.
1419 if (unlikely(queued
)) {
1420 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1422 set_current_state(TASK_UNINTERRUPTIBLE
);
1423 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1428 * Ahh, all good. It wasn't running, and it wasn't
1429 * runnable, which means that it will never become
1430 * running in the future either. We're all done!
1439 * kick_process - kick a running thread to enter/exit the kernel
1440 * @p: the to-be-kicked thread
1442 * Cause a process which is running on another CPU to enter
1443 * kernel-mode, without any delay. (to get signals handled.)
1445 * NOTE: this function doesn't have to take the runqueue lock,
1446 * because all it wants to ensure is that the remote task enters
1447 * the kernel. If the IPI races and the task has been migrated
1448 * to another CPU then no harm is done and the purpose has been
1451 void kick_process(struct task_struct
*p
)
1457 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1458 smp_send_reschedule(cpu
);
1461 EXPORT_SYMBOL_GPL(kick_process
);
1464 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1466 * A few notes on cpu_active vs cpu_online:
1468 * - cpu_active must be a subset of cpu_online
1470 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1471 * see __set_cpus_allowed_ptr(). At this point the newly online
1472 * CPU isn't yet part of the sched domains, and balancing will not
1475 * - on CPU-down we clear cpu_active() to mask the sched domains and
1476 * avoid the load balancer to place new tasks on the to be removed
1477 * CPU. Existing tasks will remain running there and will be taken
1480 * This means that fallback selection must not select !active CPUs.
1481 * And can assume that any active CPU must be online. Conversely
1482 * select_task_rq() below may allow selection of !active CPUs in order
1483 * to satisfy the above rules.
1485 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1487 int nid
= cpu_to_node(cpu
);
1488 const struct cpumask
*nodemask
= NULL
;
1489 enum { cpuset
, possible
, fail
} state
= cpuset
;
1493 * If the node that the CPU is on has been offlined, cpu_to_node()
1494 * will return -1. There is no CPU on the node, and we should
1495 * select the CPU on the other node.
1498 nodemask
= cpumask_of_node(nid
);
1500 /* Look for allowed, online CPU in same node. */
1501 for_each_cpu(dest_cpu
, nodemask
) {
1502 if (!cpu_active(dest_cpu
))
1504 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
1510 /* Any allowed, online CPU? */
1511 for_each_cpu(dest_cpu
, &p
->cpus_allowed
) {
1512 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1514 if (!cpu_online(dest_cpu
))
1519 /* No more Mr. Nice Guy. */
1522 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1523 cpuset_cpus_allowed_fallback(p
);
1529 do_set_cpus_allowed(p
, cpu_possible_mask
);
1540 if (state
!= cpuset
) {
1542 * Don't tell them about moving exiting tasks or
1543 * kernel threads (both mm NULL), since they never
1546 if (p
->mm
&& printk_ratelimit()) {
1547 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1548 task_pid_nr(p
), p
->comm
, cpu
);
1556 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1559 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1561 lockdep_assert_held(&p
->pi_lock
);
1563 if (p
->nr_cpus_allowed
> 1)
1564 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1566 cpu
= cpumask_any(&p
->cpus_allowed
);
1569 * In order not to call set_task_cpu() on a blocking task we need
1570 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1573 * Since this is common to all placement strategies, this lives here.
1575 * [ this allows ->select_task() to simply return task_cpu(p) and
1576 * not worry about this generic constraint ]
1578 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
1580 cpu
= select_fallback_rq(task_cpu(p
), p
);
1585 static void update_avg(u64
*avg
, u64 sample
)
1587 s64 diff
= sample
- *avg
;
1593 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1594 const struct cpumask
*new_mask
, bool check
)
1596 return set_cpus_allowed_ptr(p
, new_mask
);
1599 #endif /* CONFIG_SMP */
1602 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1606 if (!schedstat_enabled())
1612 if (cpu
== rq
->cpu
) {
1613 schedstat_inc(rq
->ttwu_local
);
1614 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1616 struct sched_domain
*sd
;
1618 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1620 for_each_domain(rq
->cpu
, sd
) {
1621 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1622 schedstat_inc(sd
->ttwu_wake_remote
);
1629 if (wake_flags
& WF_MIGRATED
)
1630 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1631 #endif /* CONFIG_SMP */
1633 schedstat_inc(rq
->ttwu_count
);
1634 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1636 if (wake_flags
& WF_SYNC
)
1637 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1640 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1642 activate_task(rq
, p
, en_flags
);
1643 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1645 /* If a worker is waking up, notify the workqueue: */
1646 if (p
->flags
& PF_WQ_WORKER
)
1647 wq_worker_waking_up(p
, cpu_of(rq
));
1651 * Mark the task runnable and perform wakeup-preemption.
1653 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1654 struct rq_flags
*rf
)
1656 check_preempt_curr(rq
, p
, wake_flags
);
1657 p
->state
= TASK_RUNNING
;
1658 trace_sched_wakeup(p
);
1661 if (p
->sched_class
->task_woken
) {
1663 * Our task @p is fully woken up and running; so its safe to
1664 * drop the rq->lock, hereafter rq is only used for statistics.
1666 rq_unpin_lock(rq
, rf
);
1667 p
->sched_class
->task_woken(rq
, p
);
1668 rq_repin_lock(rq
, rf
);
1671 if (rq
->idle_stamp
) {
1672 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1673 u64 max
= 2*rq
->max_idle_balance_cost
;
1675 update_avg(&rq
->avg_idle
, delta
);
1677 if (rq
->avg_idle
> max
)
1686 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1687 struct rq_flags
*rf
)
1689 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
1691 lockdep_assert_held(&rq
->lock
);
1694 if (p
->sched_contributes_to_load
)
1695 rq
->nr_uninterruptible
--;
1697 if (wake_flags
& WF_MIGRATED
)
1698 en_flags
|= ENQUEUE_MIGRATED
;
1701 ttwu_activate(rq
, p
, en_flags
);
1702 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1706 * Called in case the task @p isn't fully descheduled from its runqueue,
1707 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1708 * since all we need to do is flip p->state to TASK_RUNNING, since
1709 * the task is still ->on_rq.
1711 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1717 rq
= __task_rq_lock(p
, &rf
);
1718 if (task_on_rq_queued(p
)) {
1719 /* check_preempt_curr() may use rq clock */
1720 update_rq_clock(rq
);
1721 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1724 __task_rq_unlock(rq
, &rf
);
1730 void sched_ttwu_pending(void)
1732 struct rq
*rq
= this_rq();
1733 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1734 struct task_struct
*p
;
1740 rq_lock_irqsave(rq
, &rf
);
1741 update_rq_clock(rq
);
1746 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1747 llist
= llist_next(llist
);
1749 if (p
->sched_remote_wakeup
)
1750 wake_flags
= WF_MIGRATED
;
1752 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1755 rq_unlock_irqrestore(rq
, &rf
);
1758 void scheduler_ipi(void)
1761 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1762 * TIF_NEED_RESCHED remotely (for the first time) will also send
1765 preempt_fold_need_resched();
1767 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1771 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1772 * traditionally all their work was done from the interrupt return
1773 * path. Now that we actually do some work, we need to make sure
1776 * Some archs already do call them, luckily irq_enter/exit nest
1779 * Arguably we should visit all archs and update all handlers,
1780 * however a fair share of IPIs are still resched only so this would
1781 * somewhat pessimize the simple resched case.
1784 sched_ttwu_pending();
1787 * Check if someone kicked us for doing the nohz idle load balance.
1789 if (unlikely(got_nohz_idle_kick())) {
1790 this_rq()->idle_balance
= 1;
1791 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1796 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1798 struct rq
*rq
= cpu_rq(cpu
);
1800 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1802 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1803 if (!set_nr_if_polling(rq
->idle
))
1804 smp_send_reschedule(cpu
);
1806 trace_sched_wake_idle_without_ipi(cpu
);
1810 void wake_up_if_idle(int cpu
)
1812 struct rq
*rq
= cpu_rq(cpu
);
1817 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1820 if (set_nr_if_polling(rq
->idle
)) {
1821 trace_sched_wake_idle_without_ipi(cpu
);
1823 rq_lock_irqsave(rq
, &rf
);
1824 if (is_idle_task(rq
->curr
))
1825 smp_send_reschedule(cpu
);
1826 /* Else CPU is not idle, do nothing here: */
1827 rq_unlock_irqrestore(rq
, &rf
);
1834 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1836 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1838 #endif /* CONFIG_SMP */
1840 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1842 struct rq
*rq
= cpu_rq(cpu
);
1845 #if defined(CONFIG_SMP)
1846 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1847 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
1848 ttwu_queue_remote(p
, cpu
, wake_flags
);
1854 update_rq_clock(rq
);
1855 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1860 * Notes on Program-Order guarantees on SMP systems.
1864 * The basic program-order guarantee on SMP systems is that when a task [t]
1865 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1866 * execution on its new CPU [c1].
1868 * For migration (of runnable tasks) this is provided by the following means:
1870 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1871 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1872 * rq(c1)->lock (if not at the same time, then in that order).
1873 * C) LOCK of the rq(c1)->lock scheduling in task
1875 * Transitivity guarantees that B happens after A and C after B.
1876 * Note: we only require RCpc transitivity.
1877 * Note: the CPU doing B need not be c0 or c1
1886 * UNLOCK rq(0)->lock
1888 * LOCK rq(0)->lock // orders against CPU0
1890 * UNLOCK rq(0)->lock
1894 * UNLOCK rq(1)->lock
1896 * LOCK rq(1)->lock // orders against CPU2
1899 * UNLOCK rq(1)->lock
1902 * BLOCKING -- aka. SLEEP + WAKEUP
1904 * For blocking we (obviously) need to provide the same guarantee as for
1905 * migration. However the means are completely different as there is no lock
1906 * chain to provide order. Instead we do:
1908 * 1) smp_store_release(X->on_cpu, 0)
1909 * 2) smp_cond_load_acquire(!X->on_cpu)
1913 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1915 * LOCK rq(0)->lock LOCK X->pi_lock
1918 * smp_store_release(X->on_cpu, 0);
1920 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1926 * X->state = RUNNING
1927 * UNLOCK rq(2)->lock
1929 * LOCK rq(2)->lock // orders against CPU1
1932 * UNLOCK rq(2)->lock
1935 * UNLOCK rq(0)->lock
1938 * However; for wakeups there is a second guarantee we must provide, namely we
1939 * must observe the state that lead to our wakeup. That is, not only must our
1940 * task observe its own prior state, it must also observe the stores prior to
1943 * This means that any means of doing remote wakeups must order the CPU doing
1944 * the wakeup against the CPU the task is going to end up running on. This,
1945 * however, is already required for the regular Program-Order guarantee above,
1946 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1951 * try_to_wake_up - wake up a thread
1952 * @p: the thread to be awakened
1953 * @state: the mask of task states that can be woken
1954 * @wake_flags: wake modifier flags (WF_*)
1956 * If (@state & @p->state) @p->state = TASK_RUNNING.
1958 * If the task was not queued/runnable, also place it back on a runqueue.
1960 * Atomic against schedule() which would dequeue a task, also see
1961 * set_current_state().
1963 * Return: %true if @p->state changes (an actual wakeup was done),
1967 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1969 unsigned long flags
;
1970 int cpu
, success
= 0;
1973 * If we are going to wake up a thread waiting for CONDITION we
1974 * need to ensure that CONDITION=1 done by the caller can not be
1975 * reordered with p->state check below. This pairs with mb() in
1976 * set_current_state() the waiting thread does.
1978 smp_mb__before_spinlock();
1979 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1980 if (!(p
->state
& state
))
1983 trace_sched_waking(p
);
1985 /* We're going to change ->state: */
1990 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1991 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1992 * in smp_cond_load_acquire() below.
1994 * sched_ttwu_pending() try_to_wake_up()
1995 * [S] p->on_rq = 1; [L] P->state
1996 * UNLOCK rq->lock -----.
2000 * LOCK rq->lock -----'
2004 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2006 * Pairs with the UNLOCK+LOCK on rq->lock from the
2007 * last wakeup of our task and the schedule that got our task
2011 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2016 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2017 * possible to, falsely, observe p->on_cpu == 0.
2019 * One must be running (->on_cpu == 1) in order to remove oneself
2020 * from the runqueue.
2022 * [S] ->on_cpu = 1; [L] ->on_rq
2026 * [S] ->on_rq = 0; [L] ->on_cpu
2028 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2029 * from the consecutive calls to schedule(); the first switching to our
2030 * task, the second putting it to sleep.
2035 * If the owning (remote) CPU is still in the middle of schedule() with
2036 * this task as prev, wait until its done referencing the task.
2038 * Pairs with the smp_store_release() in finish_lock_switch().
2040 * This ensures that tasks getting woken will be fully ordered against
2041 * their previous state and preserve Program Order.
2043 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2045 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2046 p
->state
= TASK_WAKING
;
2049 delayacct_blkio_end();
2050 atomic_dec(&task_rq(p
)->nr_iowait
);
2053 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2054 if (task_cpu(p
) != cpu
) {
2055 wake_flags
|= WF_MIGRATED
;
2056 set_task_cpu(p
, cpu
);
2059 #else /* CONFIG_SMP */
2062 delayacct_blkio_end();
2063 atomic_dec(&task_rq(p
)->nr_iowait
);
2066 #endif /* CONFIG_SMP */
2068 ttwu_queue(p
, cpu
, wake_flags
);
2070 ttwu_stat(p
, cpu
, wake_flags
);
2072 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2078 * try_to_wake_up_local - try to wake up a local task with rq lock held
2079 * @p: the thread to be awakened
2080 * @cookie: context's cookie for pinning
2082 * Put @p on the run-queue if it's not already there. The caller must
2083 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2086 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2088 struct rq
*rq
= task_rq(p
);
2090 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2091 WARN_ON_ONCE(p
== current
))
2094 lockdep_assert_held(&rq
->lock
);
2096 if (!raw_spin_trylock(&p
->pi_lock
)) {
2098 * This is OK, because current is on_cpu, which avoids it being
2099 * picked for load-balance and preemption/IRQs are still
2100 * disabled avoiding further scheduler activity on it and we've
2101 * not yet picked a replacement task.
2104 raw_spin_lock(&p
->pi_lock
);
2108 if (!(p
->state
& TASK_NORMAL
))
2111 trace_sched_waking(p
);
2113 if (!task_on_rq_queued(p
)) {
2115 delayacct_blkio_end();
2116 atomic_dec(&rq
->nr_iowait
);
2118 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
);
2121 ttwu_do_wakeup(rq
, p
, 0, rf
);
2122 ttwu_stat(p
, smp_processor_id(), 0);
2124 raw_spin_unlock(&p
->pi_lock
);
2128 * wake_up_process - Wake up a specific process
2129 * @p: The process to be woken up.
2131 * Attempt to wake up the nominated process and move it to the set of runnable
2134 * Return: 1 if the process was woken up, 0 if it was already running.
2136 * It may be assumed that this function implies a write memory barrier before
2137 * changing the task state if and only if any tasks are woken up.
2139 int wake_up_process(struct task_struct
*p
)
2141 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2143 EXPORT_SYMBOL(wake_up_process
);
2145 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2147 return try_to_wake_up(p
, state
, 0);
2151 * This function clears the sched_dl_entity static params.
2153 void __dl_clear_params(struct task_struct
*p
)
2155 struct sched_dl_entity
*dl_se
= &p
->dl
;
2157 dl_se
->dl_runtime
= 0;
2158 dl_se
->dl_deadline
= 0;
2159 dl_se
->dl_period
= 0;
2163 dl_se
->dl_throttled
= 0;
2164 dl_se
->dl_yielded
= 0;
2168 * Perform scheduler related setup for a newly forked process p.
2169 * p is forked by current.
2171 * __sched_fork() is basic setup used by init_idle() too:
2173 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2178 p
->se
.exec_start
= 0;
2179 p
->se
.sum_exec_runtime
= 0;
2180 p
->se
.prev_sum_exec_runtime
= 0;
2181 p
->se
.nr_migrations
= 0;
2183 INIT_LIST_HEAD(&p
->se
.group_node
);
2185 #ifdef CONFIG_FAIR_GROUP_SCHED
2186 p
->se
.cfs_rq
= NULL
;
2189 #ifdef CONFIG_SCHEDSTATS
2190 /* Even if schedstat is disabled, there should not be garbage */
2191 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2194 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2195 init_dl_task_timer(&p
->dl
);
2196 __dl_clear_params(p
);
2198 INIT_LIST_HEAD(&p
->rt
.run_list
);
2200 p
->rt
.time_slice
= sched_rr_timeslice
;
2204 #ifdef CONFIG_PREEMPT_NOTIFIERS
2205 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2208 #ifdef CONFIG_NUMA_BALANCING
2209 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2210 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2211 p
->mm
->numa_scan_seq
= 0;
2214 if (clone_flags
& CLONE_VM
)
2215 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2217 p
->numa_preferred_nid
= -1;
2219 p
->node_stamp
= 0ULL;
2220 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2221 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2222 p
->numa_work
.next
= &p
->numa_work
;
2223 p
->numa_faults
= NULL
;
2224 p
->last_task_numa_placement
= 0;
2225 p
->last_sum_exec_runtime
= 0;
2227 p
->numa_group
= NULL
;
2228 #endif /* CONFIG_NUMA_BALANCING */
2231 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2233 #ifdef CONFIG_NUMA_BALANCING
2235 void set_numabalancing_state(bool enabled
)
2238 static_branch_enable(&sched_numa_balancing
);
2240 static_branch_disable(&sched_numa_balancing
);
2243 #ifdef CONFIG_PROC_SYSCTL
2244 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2245 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2249 int state
= static_branch_likely(&sched_numa_balancing
);
2251 if (write
&& !capable(CAP_SYS_ADMIN
))
2256 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2260 set_numabalancing_state(state
);
2266 #ifdef CONFIG_SCHEDSTATS
2268 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2269 static bool __initdata __sched_schedstats
= false;
2271 static void set_schedstats(bool enabled
)
2274 static_branch_enable(&sched_schedstats
);
2276 static_branch_disable(&sched_schedstats
);
2279 void force_schedstat_enabled(void)
2281 if (!schedstat_enabled()) {
2282 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2283 static_branch_enable(&sched_schedstats
);
2287 static int __init
setup_schedstats(char *str
)
2294 * This code is called before jump labels have been set up, so we can't
2295 * change the static branch directly just yet. Instead set a temporary
2296 * variable so init_schedstats() can do it later.
2298 if (!strcmp(str
, "enable")) {
2299 __sched_schedstats
= true;
2301 } else if (!strcmp(str
, "disable")) {
2302 __sched_schedstats
= false;
2307 pr_warn("Unable to parse schedstats=\n");
2311 __setup("schedstats=", setup_schedstats
);
2313 static void __init
init_schedstats(void)
2315 set_schedstats(__sched_schedstats
);
2318 #ifdef CONFIG_PROC_SYSCTL
2319 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2320 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2324 int state
= static_branch_likely(&sched_schedstats
);
2326 if (write
&& !capable(CAP_SYS_ADMIN
))
2331 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2335 set_schedstats(state
);
2338 #endif /* CONFIG_PROC_SYSCTL */
2339 #else /* !CONFIG_SCHEDSTATS */
2340 static inline void init_schedstats(void) {}
2341 #endif /* CONFIG_SCHEDSTATS */
2344 * fork()/clone()-time setup:
2346 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2348 unsigned long flags
;
2349 int cpu
= get_cpu();
2351 __sched_fork(clone_flags
, p
);
2353 * We mark the process as NEW here. This guarantees that
2354 * nobody will actually run it, and a signal or other external
2355 * event cannot wake it up and insert it on the runqueue either.
2357 p
->state
= TASK_NEW
;
2360 * Make sure we do not leak PI boosting priority to the child.
2362 p
->prio
= current
->normal_prio
;
2365 * Revert to default priority/policy on fork if requested.
2367 if (unlikely(p
->sched_reset_on_fork
)) {
2368 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2369 p
->policy
= SCHED_NORMAL
;
2370 p
->static_prio
= NICE_TO_PRIO(0);
2372 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2373 p
->static_prio
= NICE_TO_PRIO(0);
2375 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2379 * We don't need the reset flag anymore after the fork. It has
2380 * fulfilled its duty:
2382 p
->sched_reset_on_fork
= 0;
2385 if (dl_prio(p
->prio
)) {
2388 } else if (rt_prio(p
->prio
)) {
2389 p
->sched_class
= &rt_sched_class
;
2391 p
->sched_class
= &fair_sched_class
;
2394 init_entity_runnable_average(&p
->se
);
2397 * The child is not yet in the pid-hash so no cgroup attach races,
2398 * and the cgroup is pinned to this child due to cgroup_fork()
2399 * is ran before sched_fork().
2401 * Silence PROVE_RCU.
2403 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2405 * We're setting the CPU for the first time, we don't migrate,
2406 * so use __set_task_cpu().
2408 __set_task_cpu(p
, cpu
);
2409 if (p
->sched_class
->task_fork
)
2410 p
->sched_class
->task_fork(p
);
2411 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2413 #ifdef CONFIG_SCHED_INFO
2414 if (likely(sched_info_on()))
2415 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2417 #if defined(CONFIG_SMP)
2420 init_task_preempt_count(p
);
2422 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2423 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2430 unsigned long to_ratio(u64 period
, u64 runtime
)
2432 if (runtime
== RUNTIME_INF
)
2436 * Doing this here saves a lot of checks in all
2437 * the calling paths, and returning zero seems
2438 * safe for them anyway.
2443 return div64_u64(runtime
<< 20, period
);
2447 inline struct dl_bw
*dl_bw_of(int i
)
2449 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2450 "sched RCU must be held");
2451 return &cpu_rq(i
)->rd
->dl_bw
;
2454 static inline int dl_bw_cpus(int i
)
2456 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2459 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2460 "sched RCU must be held");
2461 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2467 inline struct dl_bw
*dl_bw_of(int i
)
2469 return &cpu_rq(i
)->dl
.dl_bw
;
2472 static inline int dl_bw_cpus(int i
)
2479 * We must be sure that accepting a new task (or allowing changing the
2480 * parameters of an existing one) is consistent with the bandwidth
2481 * constraints. If yes, this function also accordingly updates the currently
2482 * allocated bandwidth to reflect the new situation.
2484 * This function is called while holding p's rq->lock.
2486 * XXX we should delay bw change until the task's 0-lag point, see
2489 static int dl_overflow(struct task_struct
*p
, int policy
,
2490 const struct sched_attr
*attr
)
2493 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2494 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2495 u64 runtime
= attr
->sched_runtime
;
2496 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2499 /* !deadline task may carry old deadline bandwidth */
2500 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2504 * Either if a task, enters, leave, or stays -deadline but changes
2505 * its parameters, we may need to update accordingly the total
2506 * allocated bandwidth of the container.
2508 raw_spin_lock(&dl_b
->lock
);
2509 cpus
= dl_bw_cpus(task_cpu(p
));
2510 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2511 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2512 __dl_add(dl_b
, new_bw
);
2514 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2515 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2516 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2517 __dl_add(dl_b
, new_bw
);
2519 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2520 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2523 raw_spin_unlock(&dl_b
->lock
);
2528 extern void init_dl_bw(struct dl_bw
*dl_b
);
2531 * wake_up_new_task - wake up a newly created task for the first time.
2533 * This function will do some initial scheduler statistics housekeeping
2534 * that must be done for every newly created context, then puts the task
2535 * on the runqueue and wakes it.
2537 void wake_up_new_task(struct task_struct
*p
)
2542 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2543 p
->state
= TASK_RUNNING
;
2546 * Fork balancing, do it here and not earlier because:
2547 * - cpus_allowed can change in the fork path
2548 * - any previously selected CPU might disappear through hotplug
2550 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2551 * as we're not fully set-up yet.
2553 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2555 rq
= __task_rq_lock(p
, &rf
);
2556 update_rq_clock(rq
);
2557 post_init_entity_util_avg(&p
->se
);
2559 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2560 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2561 trace_sched_wakeup_new(p
);
2562 check_preempt_curr(rq
, p
, WF_FORK
);
2564 if (p
->sched_class
->task_woken
) {
2566 * Nothing relies on rq->lock after this, so its fine to
2569 rq_unpin_lock(rq
, &rf
);
2570 p
->sched_class
->task_woken(rq
, p
);
2571 rq_repin_lock(rq
, &rf
);
2574 task_rq_unlock(rq
, p
, &rf
);
2577 #ifdef CONFIG_PREEMPT_NOTIFIERS
2579 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2581 void preempt_notifier_inc(void)
2583 static_key_slow_inc(&preempt_notifier_key
);
2585 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2587 void preempt_notifier_dec(void)
2589 static_key_slow_dec(&preempt_notifier_key
);
2591 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2594 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2595 * @notifier: notifier struct to register
2597 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2599 if (!static_key_false(&preempt_notifier_key
))
2600 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2602 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2604 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2607 * preempt_notifier_unregister - no longer interested in preemption notifications
2608 * @notifier: notifier struct to unregister
2610 * This is *not* safe to call from within a preemption notifier.
2612 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2614 hlist_del(¬ifier
->link
);
2616 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2618 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2620 struct preempt_notifier
*notifier
;
2622 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2623 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2626 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2628 if (static_key_false(&preempt_notifier_key
))
2629 __fire_sched_in_preempt_notifiers(curr
);
2633 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2634 struct task_struct
*next
)
2636 struct preempt_notifier
*notifier
;
2638 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2639 notifier
->ops
->sched_out(notifier
, next
);
2642 static __always_inline
void
2643 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2644 struct task_struct
*next
)
2646 if (static_key_false(&preempt_notifier_key
))
2647 __fire_sched_out_preempt_notifiers(curr
, next
);
2650 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2652 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2657 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2658 struct task_struct
*next
)
2662 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2665 * prepare_task_switch - prepare to switch tasks
2666 * @rq: the runqueue preparing to switch
2667 * @prev: the current task that is being switched out
2668 * @next: the task we are going to switch to.
2670 * This is called with the rq lock held and interrupts off. It must
2671 * be paired with a subsequent finish_task_switch after the context
2674 * prepare_task_switch sets up locking and calls architecture specific
2678 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2679 struct task_struct
*next
)
2681 sched_info_switch(rq
, prev
, next
);
2682 perf_event_task_sched_out(prev
, next
);
2683 fire_sched_out_preempt_notifiers(prev
, next
);
2684 prepare_lock_switch(rq
, next
);
2685 prepare_arch_switch(next
);
2689 * finish_task_switch - clean up after a task-switch
2690 * @prev: the thread we just switched away from.
2692 * finish_task_switch must be called after the context switch, paired
2693 * with a prepare_task_switch call before the context switch.
2694 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2695 * and do any other architecture-specific cleanup actions.
2697 * Note that we may have delayed dropping an mm in context_switch(). If
2698 * so, we finish that here outside of the runqueue lock. (Doing it
2699 * with the lock held can cause deadlocks; see schedule() for
2702 * The context switch have flipped the stack from under us and restored the
2703 * local variables which were saved when this task called schedule() in the
2704 * past. prev == current is still correct but we need to recalculate this_rq
2705 * because prev may have moved to another CPU.
2707 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2708 __releases(rq
->lock
)
2710 struct rq
*rq
= this_rq();
2711 struct mm_struct
*mm
= rq
->prev_mm
;
2715 * The previous task will have left us with a preempt_count of 2
2716 * because it left us after:
2719 * preempt_disable(); // 1
2721 * raw_spin_lock_irq(&rq->lock) // 2
2723 * Also, see FORK_PREEMPT_COUNT.
2725 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2726 "corrupted preempt_count: %s/%d/0x%x\n",
2727 current
->comm
, current
->pid
, preempt_count()))
2728 preempt_count_set(FORK_PREEMPT_COUNT
);
2733 * A task struct has one reference for the use as "current".
2734 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2735 * schedule one last time. The schedule call will never return, and
2736 * the scheduled task must drop that reference.
2738 * We must observe prev->state before clearing prev->on_cpu (in
2739 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2740 * running on another CPU and we could rave with its RUNNING -> DEAD
2741 * transition, resulting in a double drop.
2743 prev_state
= prev
->state
;
2744 vtime_task_switch(prev
);
2745 perf_event_task_sched_in(prev
, current
);
2746 finish_lock_switch(rq
, prev
);
2747 finish_arch_post_lock_switch();
2749 fire_sched_in_preempt_notifiers(current
);
2752 if (unlikely(prev_state
== TASK_DEAD
)) {
2753 if (prev
->sched_class
->task_dead
)
2754 prev
->sched_class
->task_dead(prev
);
2757 * Remove function-return probe instances associated with this
2758 * task and put them back on the free list.
2760 kprobe_flush_task(prev
);
2762 /* Task is done with its stack. */
2763 put_task_stack(prev
);
2765 put_task_struct(prev
);
2768 tick_nohz_task_switch();
2774 /* rq->lock is NOT held, but preemption is disabled */
2775 static void __balance_callback(struct rq
*rq
)
2777 struct callback_head
*head
, *next
;
2778 void (*func
)(struct rq
*rq
);
2779 unsigned long flags
;
2781 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2782 head
= rq
->balance_callback
;
2783 rq
->balance_callback
= NULL
;
2785 func
= (void (*)(struct rq
*))head
->func
;
2792 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2795 static inline void balance_callback(struct rq
*rq
)
2797 if (unlikely(rq
->balance_callback
))
2798 __balance_callback(rq
);
2803 static inline void balance_callback(struct rq
*rq
)
2810 * schedule_tail - first thing a freshly forked thread must call.
2811 * @prev: the thread we just switched away from.
2813 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2814 __releases(rq
->lock
)
2819 * New tasks start with FORK_PREEMPT_COUNT, see there and
2820 * finish_task_switch() for details.
2822 * finish_task_switch() will drop rq->lock() and lower preempt_count
2823 * and the preempt_enable() will end up enabling preemption (on
2824 * PREEMPT_COUNT kernels).
2827 rq
= finish_task_switch(prev
);
2828 balance_callback(rq
);
2831 if (current
->set_child_tid
)
2832 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2836 * context_switch - switch to the new MM and the new thread's register state.
2838 static __always_inline
struct rq
*
2839 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2840 struct task_struct
*next
, struct rq_flags
*rf
)
2842 struct mm_struct
*mm
, *oldmm
;
2844 prepare_task_switch(rq
, prev
, next
);
2847 oldmm
= prev
->active_mm
;
2849 * For paravirt, this is coupled with an exit in switch_to to
2850 * combine the page table reload and the switch backend into
2853 arch_start_context_switch(prev
);
2856 next
->active_mm
= oldmm
;
2858 enter_lazy_tlb(oldmm
, next
);
2860 switch_mm_irqs_off(oldmm
, mm
, next
);
2863 prev
->active_mm
= NULL
;
2864 rq
->prev_mm
= oldmm
;
2867 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
2870 * Since the runqueue lock will be released by the next
2871 * task (which is an invalid locking op but in the case
2872 * of the scheduler it's an obvious special-case), so we
2873 * do an early lockdep release here:
2875 rq_unpin_lock(rq
, rf
);
2876 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2878 /* Here we just switch the register state and the stack. */
2879 switch_to(prev
, next
, prev
);
2882 return finish_task_switch(prev
);
2886 * nr_running and nr_context_switches:
2888 * externally visible scheduler statistics: current number of runnable
2889 * threads, total number of context switches performed since bootup.
2891 unsigned long nr_running(void)
2893 unsigned long i
, sum
= 0;
2895 for_each_online_cpu(i
)
2896 sum
+= cpu_rq(i
)->nr_running
;
2902 * Check if only the current task is running on the CPU.
2904 * Caution: this function does not check that the caller has disabled
2905 * preemption, thus the result might have a time-of-check-to-time-of-use
2906 * race. The caller is responsible to use it correctly, for example:
2908 * - from a non-preemptable section (of course)
2910 * - from a thread that is bound to a single CPU
2912 * - in a loop with very short iterations (e.g. a polling loop)
2914 bool single_task_running(void)
2916 return raw_rq()->nr_running
== 1;
2918 EXPORT_SYMBOL(single_task_running
);
2920 unsigned long long nr_context_switches(void)
2923 unsigned long long sum
= 0;
2925 for_each_possible_cpu(i
)
2926 sum
+= cpu_rq(i
)->nr_switches
;
2932 * IO-wait accounting, and how its mostly bollocks (on SMP).
2934 * The idea behind IO-wait account is to account the idle time that we could
2935 * have spend running if it were not for IO. That is, if we were to improve the
2936 * storage performance, we'd have a proportional reduction in IO-wait time.
2938 * This all works nicely on UP, where, when a task blocks on IO, we account
2939 * idle time as IO-wait, because if the storage were faster, it could've been
2940 * running and we'd not be idle.
2942 * This has been extended to SMP, by doing the same for each CPU. This however
2945 * Imagine for instance the case where two tasks block on one CPU, only the one
2946 * CPU will have IO-wait accounted, while the other has regular idle. Even
2947 * though, if the storage were faster, both could've ran at the same time,
2948 * utilising both CPUs.
2950 * This means, that when looking globally, the current IO-wait accounting on
2951 * SMP is a lower bound, by reason of under accounting.
2953 * Worse, since the numbers are provided per CPU, they are sometimes
2954 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2955 * associated with any one particular CPU, it can wake to another CPU than it
2956 * blocked on. This means the per CPU IO-wait number is meaningless.
2958 * Task CPU affinities can make all that even more 'interesting'.
2961 unsigned long nr_iowait(void)
2963 unsigned long i
, sum
= 0;
2965 for_each_possible_cpu(i
)
2966 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2972 * Consumers of these two interfaces, like for example the cpufreq menu
2973 * governor are using nonsensical data. Boosting frequency for a CPU that has
2974 * IO-wait which might not even end up running the task when it does become
2978 unsigned long nr_iowait_cpu(int cpu
)
2980 struct rq
*this = cpu_rq(cpu
);
2981 return atomic_read(&this->nr_iowait
);
2984 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2986 struct rq
*rq
= this_rq();
2987 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2988 *load
= rq
->load
.weight
;
2994 * sched_exec - execve() is a valuable balancing opportunity, because at
2995 * this point the task has the smallest effective memory and cache footprint.
2997 void sched_exec(void)
2999 struct task_struct
*p
= current
;
3000 unsigned long flags
;
3003 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3004 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
3005 if (dest_cpu
== smp_processor_id())
3008 if (likely(cpu_active(dest_cpu
))) {
3009 struct migration_arg arg
= { p
, dest_cpu
};
3011 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3012 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3016 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3021 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3022 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3024 EXPORT_PER_CPU_SYMBOL(kstat
);
3025 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3028 * The function fair_sched_class.update_curr accesses the struct curr
3029 * and its field curr->exec_start; when called from task_sched_runtime(),
3030 * we observe a high rate of cache misses in practice.
3031 * Prefetching this data results in improved performance.
3033 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3035 #ifdef CONFIG_FAIR_GROUP_SCHED
3036 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3038 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3041 prefetch(&curr
->exec_start
);
3045 * Return accounted runtime for the task.
3046 * In case the task is currently running, return the runtime plus current's
3047 * pending runtime that have not been accounted yet.
3049 unsigned long long task_sched_runtime(struct task_struct
*p
)
3055 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3057 * 64-bit doesn't need locks to atomically read a 64bit value.
3058 * So we have a optimization chance when the task's delta_exec is 0.
3059 * Reading ->on_cpu is racy, but this is ok.
3061 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3062 * If we race with it entering CPU, unaccounted time is 0. This is
3063 * indistinguishable from the read occurring a few cycles earlier.
3064 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3065 * been accounted, so we're correct here as well.
3067 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3068 return p
->se
.sum_exec_runtime
;
3071 rq
= task_rq_lock(p
, &rf
);
3073 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3074 * project cycles that may never be accounted to this
3075 * thread, breaking clock_gettime().
3077 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3078 prefetch_curr_exec_start(p
);
3079 update_rq_clock(rq
);
3080 p
->sched_class
->update_curr(rq
);
3082 ns
= p
->se
.sum_exec_runtime
;
3083 task_rq_unlock(rq
, p
, &rf
);
3089 * This function gets called by the timer code, with HZ frequency.
3090 * We call it with interrupts disabled.
3092 void scheduler_tick(void)
3094 int cpu
= smp_processor_id();
3095 struct rq
*rq
= cpu_rq(cpu
);
3096 struct task_struct
*curr
= rq
->curr
;
3103 update_rq_clock(rq
);
3104 curr
->sched_class
->task_tick(rq
, curr
, 0);
3105 cpu_load_update_active(rq
);
3106 calc_global_load_tick(rq
);
3110 perf_event_task_tick();
3113 rq
->idle_balance
= idle_cpu(cpu
);
3114 trigger_load_balance(rq
);
3116 rq_last_tick_reset(rq
);
3119 #ifdef CONFIG_NO_HZ_FULL
3121 * scheduler_tick_max_deferment
3123 * Keep at least one tick per second when a single
3124 * active task is running because the scheduler doesn't
3125 * yet completely support full dynticks environment.
3127 * This makes sure that uptime, CFS vruntime, load
3128 * balancing, etc... continue to move forward, even
3129 * with a very low granularity.
3131 * Return: Maximum deferment in nanoseconds.
3133 u64
scheduler_tick_max_deferment(void)
3135 struct rq
*rq
= this_rq();
3136 unsigned long next
, now
= READ_ONCE(jiffies
);
3138 next
= rq
->last_sched_tick
+ HZ
;
3140 if (time_before_eq(next
, now
))
3143 return jiffies_to_nsecs(next
- now
);
3147 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3148 defined(CONFIG_PREEMPT_TRACER))
3150 * If the value passed in is equal to the current preempt count
3151 * then we just disabled preemption. Start timing the latency.
3153 static inline void preempt_latency_start(int val
)
3155 if (preempt_count() == val
) {
3156 unsigned long ip
= get_lock_parent_ip();
3157 #ifdef CONFIG_DEBUG_PREEMPT
3158 current
->preempt_disable_ip
= ip
;
3160 trace_preempt_off(CALLER_ADDR0
, ip
);
3164 void preempt_count_add(int val
)
3166 #ifdef CONFIG_DEBUG_PREEMPT
3170 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3173 __preempt_count_add(val
);
3174 #ifdef CONFIG_DEBUG_PREEMPT
3176 * Spinlock count overflowing soon?
3178 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3181 preempt_latency_start(val
);
3183 EXPORT_SYMBOL(preempt_count_add
);
3184 NOKPROBE_SYMBOL(preempt_count_add
);
3187 * If the value passed in equals to the current preempt count
3188 * then we just enabled preemption. Stop timing the latency.
3190 static inline void preempt_latency_stop(int val
)
3192 if (preempt_count() == val
)
3193 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3196 void preempt_count_sub(int val
)
3198 #ifdef CONFIG_DEBUG_PREEMPT
3202 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3205 * Is the spinlock portion underflowing?
3207 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3208 !(preempt_count() & PREEMPT_MASK
)))
3212 preempt_latency_stop(val
);
3213 __preempt_count_sub(val
);
3215 EXPORT_SYMBOL(preempt_count_sub
);
3216 NOKPROBE_SYMBOL(preempt_count_sub
);
3219 static inline void preempt_latency_start(int val
) { }
3220 static inline void preempt_latency_stop(int val
) { }
3223 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3225 #ifdef CONFIG_DEBUG_PREEMPT
3226 return p
->preempt_disable_ip
;
3233 * Print scheduling while atomic bug:
3235 static noinline
void __schedule_bug(struct task_struct
*prev
)
3237 /* Save this before calling printk(), since that will clobber it */
3238 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3240 if (oops_in_progress
)
3243 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3244 prev
->comm
, prev
->pid
, preempt_count());
3246 debug_show_held_locks(prev
);
3248 if (irqs_disabled())
3249 print_irqtrace_events(prev
);
3250 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3251 && in_atomic_preempt_off()) {
3252 pr_err("Preemption disabled at:");
3253 print_ip_sym(preempt_disable_ip
);
3257 panic("scheduling while atomic\n");
3260 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3264 * Various schedule()-time debugging checks and statistics:
3266 static inline void schedule_debug(struct task_struct
*prev
)
3268 #ifdef CONFIG_SCHED_STACK_END_CHECK
3269 if (task_stack_end_corrupted(prev
))
3270 panic("corrupted stack end detected inside scheduler\n");
3273 if (unlikely(in_atomic_preempt_off())) {
3274 __schedule_bug(prev
);
3275 preempt_count_set(PREEMPT_DISABLED
);
3279 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3281 schedstat_inc(this_rq()->sched_count
);
3285 * Pick up the highest-prio task:
3287 static inline struct task_struct
*
3288 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3290 const struct sched_class
*class;
3291 struct task_struct
*p
;
3294 * Optimization: we know that if all tasks are in the fair class we can
3295 * call that function directly, but only if the @prev task wasn't of a
3296 * higher scheduling class, because otherwise those loose the
3297 * opportunity to pull in more work from other CPUs.
3299 if (likely((prev
->sched_class
== &idle_sched_class
||
3300 prev
->sched_class
== &fair_sched_class
) &&
3301 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3303 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3304 if (unlikely(p
== RETRY_TASK
))
3307 /* Assumes fair_sched_class->next == idle_sched_class */
3309 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3315 for_each_class(class) {
3316 p
= class->pick_next_task(rq
, prev
, rf
);
3318 if (unlikely(p
== RETRY_TASK
))
3324 /* The idle class should always have a runnable task: */
3329 * __schedule() is the main scheduler function.
3331 * The main means of driving the scheduler and thus entering this function are:
3333 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3335 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3336 * paths. For example, see arch/x86/entry_64.S.
3338 * To drive preemption between tasks, the scheduler sets the flag in timer
3339 * interrupt handler scheduler_tick().
3341 * 3. Wakeups don't really cause entry into schedule(). They add a
3342 * task to the run-queue and that's it.
3344 * Now, if the new task added to the run-queue preempts the current
3345 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3346 * called on the nearest possible occasion:
3348 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3350 * - in syscall or exception context, at the next outmost
3351 * preempt_enable(). (this might be as soon as the wake_up()'s
3354 * - in IRQ context, return from interrupt-handler to
3355 * preemptible context
3357 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3360 * - cond_resched() call
3361 * - explicit schedule() call
3362 * - return from syscall or exception to user-space
3363 * - return from interrupt-handler to user-space
3365 * WARNING: must be called with preemption disabled!
3367 static void __sched notrace
__schedule(bool preempt
)
3369 struct task_struct
*prev
, *next
;
3370 unsigned long *switch_count
;
3375 cpu
= smp_processor_id();
3379 schedule_debug(prev
);
3381 if (sched_feat(HRTICK
))
3384 local_irq_disable();
3385 rcu_note_context_switch(preempt
);
3388 * Make sure that signal_pending_state()->signal_pending() below
3389 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3390 * done by the caller to avoid the race with signal_wake_up().
3392 smp_mb__before_spinlock();
3395 /* Promote REQ to ACT */
3396 rq
->clock_update_flags
<<= 1;
3397 update_rq_clock(rq
);
3399 switch_count
= &prev
->nivcsw
;
3400 if (!preempt
&& prev
->state
) {
3401 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3402 prev
->state
= TASK_RUNNING
;
3404 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
3407 if (prev
->in_iowait
) {
3408 atomic_inc(&rq
->nr_iowait
);
3409 delayacct_blkio_start();
3413 * If a worker went to sleep, notify and ask workqueue
3414 * whether it wants to wake up a task to maintain
3417 if (prev
->flags
& PF_WQ_WORKER
) {
3418 struct task_struct
*to_wakeup
;
3420 to_wakeup
= wq_worker_sleeping(prev
);
3422 try_to_wake_up_local(to_wakeup
, &rf
);
3425 switch_count
= &prev
->nvcsw
;
3428 next
= pick_next_task(rq
, prev
, &rf
);
3429 clear_tsk_need_resched(prev
);
3430 clear_preempt_need_resched();
3432 if (likely(prev
!= next
)) {
3437 trace_sched_switch(preempt
, prev
, next
);
3439 /* Also unlocks the rq: */
3440 rq
= context_switch(rq
, prev
, next
, &rf
);
3442 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3443 rq_unlock_irq(rq
, &rf
);
3446 balance_callback(rq
);
3449 void __noreturn
do_task_dead(void)
3452 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3453 * when the following two conditions become true.
3454 * - There is race condition of mmap_sem (It is acquired by
3456 * - SMI occurs before setting TASK_RUNINNG.
3457 * (or hypervisor of virtual machine switches to other guest)
3458 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3460 * To avoid it, we have to wait for releasing tsk->pi_lock which
3461 * is held by try_to_wake_up()
3464 raw_spin_unlock_wait(¤t
->pi_lock
);
3466 /* Causes final put_task_struct in finish_task_switch(): */
3467 __set_current_state(TASK_DEAD
);
3469 /* Tell freezer to ignore us: */
3470 current
->flags
|= PF_NOFREEZE
;
3475 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3480 static inline void sched_submit_work(struct task_struct
*tsk
)
3482 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3485 * If we are going to sleep and we have plugged IO queued,
3486 * make sure to submit it to avoid deadlocks.
3488 if (blk_needs_flush_plug(tsk
))
3489 blk_schedule_flush_plug(tsk
);
3492 asmlinkage __visible
void __sched
schedule(void)
3494 struct task_struct
*tsk
= current
;
3496 sched_submit_work(tsk
);
3500 sched_preempt_enable_no_resched();
3501 } while (need_resched());
3503 EXPORT_SYMBOL(schedule
);
3505 #ifdef CONFIG_CONTEXT_TRACKING
3506 asmlinkage __visible
void __sched
schedule_user(void)
3509 * If we come here after a random call to set_need_resched(),
3510 * or we have been woken up remotely but the IPI has not yet arrived,
3511 * we haven't yet exited the RCU idle mode. Do it here manually until
3512 * we find a better solution.
3514 * NB: There are buggy callers of this function. Ideally we
3515 * should warn if prev_state != CONTEXT_USER, but that will trigger
3516 * too frequently to make sense yet.
3518 enum ctx_state prev_state
= exception_enter();
3520 exception_exit(prev_state
);
3525 * schedule_preempt_disabled - called with preemption disabled
3527 * Returns with preemption disabled. Note: preempt_count must be 1
3529 void __sched
schedule_preempt_disabled(void)
3531 sched_preempt_enable_no_resched();
3536 static void __sched notrace
preempt_schedule_common(void)
3540 * Because the function tracer can trace preempt_count_sub()
3541 * and it also uses preempt_enable/disable_notrace(), if
3542 * NEED_RESCHED is set, the preempt_enable_notrace() called
3543 * by the function tracer will call this function again and
3544 * cause infinite recursion.
3546 * Preemption must be disabled here before the function
3547 * tracer can trace. Break up preempt_disable() into two
3548 * calls. One to disable preemption without fear of being
3549 * traced. The other to still record the preemption latency,
3550 * which can also be traced by the function tracer.
3552 preempt_disable_notrace();
3553 preempt_latency_start(1);
3555 preempt_latency_stop(1);
3556 preempt_enable_no_resched_notrace();
3559 * Check again in case we missed a preemption opportunity
3560 * between schedule and now.
3562 } while (need_resched());
3565 #ifdef CONFIG_PREEMPT
3567 * this is the entry point to schedule() from in-kernel preemption
3568 * off of preempt_enable. Kernel preemptions off return from interrupt
3569 * occur there and call schedule directly.
3571 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3574 * If there is a non-zero preempt_count or interrupts are disabled,
3575 * we do not want to preempt the current task. Just return..
3577 if (likely(!preemptible()))
3580 preempt_schedule_common();
3582 NOKPROBE_SYMBOL(preempt_schedule
);
3583 EXPORT_SYMBOL(preempt_schedule
);
3586 * preempt_schedule_notrace - preempt_schedule called by tracing
3588 * The tracing infrastructure uses preempt_enable_notrace to prevent
3589 * recursion and tracing preempt enabling caused by the tracing
3590 * infrastructure itself. But as tracing can happen in areas coming
3591 * from userspace or just about to enter userspace, a preempt enable
3592 * can occur before user_exit() is called. This will cause the scheduler
3593 * to be called when the system is still in usermode.
3595 * To prevent this, the preempt_enable_notrace will use this function
3596 * instead of preempt_schedule() to exit user context if needed before
3597 * calling the scheduler.
3599 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3601 enum ctx_state prev_ctx
;
3603 if (likely(!preemptible()))
3608 * Because the function tracer can trace preempt_count_sub()
3609 * and it also uses preempt_enable/disable_notrace(), if
3610 * NEED_RESCHED is set, the preempt_enable_notrace() called
3611 * by the function tracer will call this function again and
3612 * cause infinite recursion.
3614 * Preemption must be disabled here before the function
3615 * tracer can trace. Break up preempt_disable() into two
3616 * calls. One to disable preemption without fear of being
3617 * traced. The other to still record the preemption latency,
3618 * which can also be traced by the function tracer.
3620 preempt_disable_notrace();
3621 preempt_latency_start(1);
3623 * Needs preempt disabled in case user_exit() is traced
3624 * and the tracer calls preempt_enable_notrace() causing
3625 * an infinite recursion.
3627 prev_ctx
= exception_enter();
3629 exception_exit(prev_ctx
);
3631 preempt_latency_stop(1);
3632 preempt_enable_no_resched_notrace();
3633 } while (need_resched());
3635 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3637 #endif /* CONFIG_PREEMPT */
3640 * this is the entry point to schedule() from kernel preemption
3641 * off of irq context.
3642 * Note, that this is called and return with irqs disabled. This will
3643 * protect us against recursive calling from irq.
3645 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3647 enum ctx_state prev_state
;
3649 /* Catch callers which need to be fixed */
3650 BUG_ON(preempt_count() || !irqs_disabled());
3652 prev_state
= exception_enter();
3658 local_irq_disable();
3659 sched_preempt_enable_no_resched();
3660 } while (need_resched());
3662 exception_exit(prev_state
);
3665 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3668 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3670 EXPORT_SYMBOL(default_wake_function
);
3672 #ifdef CONFIG_RT_MUTEXES
3674 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
3677 prio
= min(prio
, pi_task
->prio
);
3682 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3684 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3686 return __rt_effective_prio(pi_task
, prio
);
3690 * rt_mutex_setprio - set the current priority of a task
3692 * @pi_task: donor task
3694 * This function changes the 'effective' priority of a task. It does
3695 * not touch ->normal_prio like __setscheduler().
3697 * Used by the rt_mutex code to implement priority inheritance
3698 * logic. Call site only calls if the priority of the task changed.
3700 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
3702 int prio
, oldprio
, queued
, running
, queue_flag
=
3703 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
3704 const struct sched_class
*prev_class
;
3708 /* XXX used to be waiter->prio, not waiter->task->prio */
3709 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
3712 * If nothing changed; bail early.
3714 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
3717 rq
= __task_rq_lock(p
, &rf
);
3718 update_rq_clock(rq
);
3720 * Set under pi_lock && rq->lock, such that the value can be used under
3723 * Note that there is loads of tricky to make this pointer cache work
3724 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3725 * ensure a task is de-boosted (pi_task is set to NULL) before the
3726 * task is allowed to run again (and can exit). This ensures the pointer
3727 * points to a blocked task -- which guaratees the task is present.
3729 p
->pi_top_task
= pi_task
;
3732 * For FIFO/RR we only need to set prio, if that matches we're done.
3734 if (prio
== p
->prio
&& !dl_prio(prio
))
3738 * Idle task boosting is a nono in general. There is one
3739 * exception, when PREEMPT_RT and NOHZ is active:
3741 * The idle task calls get_next_timer_interrupt() and holds
3742 * the timer wheel base->lock on the CPU and another CPU wants
3743 * to access the timer (probably to cancel it). We can safely
3744 * ignore the boosting request, as the idle CPU runs this code
3745 * with interrupts disabled and will complete the lock
3746 * protected section without being interrupted. So there is no
3747 * real need to boost.
3749 if (unlikely(p
== rq
->idle
)) {
3750 WARN_ON(p
!= rq
->curr
);
3751 WARN_ON(p
->pi_blocked_on
);
3755 trace_sched_pi_setprio(p
, pi_task
);
3758 if (oldprio
== prio
)
3759 queue_flag
&= ~DEQUEUE_MOVE
;
3761 prev_class
= p
->sched_class
;
3762 queued
= task_on_rq_queued(p
);
3763 running
= task_current(rq
, p
);
3765 dequeue_task(rq
, p
, queue_flag
);
3767 put_prev_task(rq
, p
);
3770 * Boosting condition are:
3771 * 1. -rt task is running and holds mutex A
3772 * --> -dl task blocks on mutex A
3774 * 2. -dl task is running and holds mutex A
3775 * --> -dl task blocks on mutex A and could preempt the
3778 if (dl_prio(prio
)) {
3779 if (!dl_prio(p
->normal_prio
) ||
3780 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3781 p
->dl
.dl_boosted
= 1;
3782 queue_flag
|= ENQUEUE_REPLENISH
;
3784 p
->dl
.dl_boosted
= 0;
3785 p
->sched_class
= &dl_sched_class
;
3786 } else if (rt_prio(prio
)) {
3787 if (dl_prio(oldprio
))
3788 p
->dl
.dl_boosted
= 0;
3790 queue_flag
|= ENQUEUE_HEAD
;
3791 p
->sched_class
= &rt_sched_class
;
3793 if (dl_prio(oldprio
))
3794 p
->dl
.dl_boosted
= 0;
3795 if (rt_prio(oldprio
))
3797 p
->sched_class
= &fair_sched_class
;
3803 enqueue_task(rq
, p
, queue_flag
);
3805 set_curr_task(rq
, p
);
3807 check_class_changed(rq
, p
, prev_class
, oldprio
);
3809 /* Avoid rq from going away on us: */
3811 __task_rq_unlock(rq
, &rf
);
3813 balance_callback(rq
);
3817 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3823 void set_user_nice(struct task_struct
*p
, long nice
)
3825 bool queued
, running
;
3826 int old_prio
, delta
;
3830 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3833 * We have to be careful, if called from sys_setpriority(),
3834 * the task might be in the middle of scheduling on another CPU.
3836 rq
= task_rq_lock(p
, &rf
);
3837 update_rq_clock(rq
);
3840 * The RT priorities are set via sched_setscheduler(), but we still
3841 * allow the 'normal' nice value to be set - but as expected
3842 * it wont have any effect on scheduling until the task is
3843 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3845 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3846 p
->static_prio
= NICE_TO_PRIO(nice
);
3849 queued
= task_on_rq_queued(p
);
3850 running
= task_current(rq
, p
);
3852 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
3854 put_prev_task(rq
, p
);
3856 p
->static_prio
= NICE_TO_PRIO(nice
);
3859 p
->prio
= effective_prio(p
);
3860 delta
= p
->prio
- old_prio
;
3863 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
3865 * If the task increased its priority or is running and
3866 * lowered its priority, then reschedule its CPU:
3868 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3872 set_curr_task(rq
, p
);
3874 task_rq_unlock(rq
, p
, &rf
);
3876 EXPORT_SYMBOL(set_user_nice
);
3879 * can_nice - check if a task can reduce its nice value
3883 int can_nice(const struct task_struct
*p
, const int nice
)
3885 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3886 int nice_rlim
= nice_to_rlimit(nice
);
3888 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3889 capable(CAP_SYS_NICE
));
3892 #ifdef __ARCH_WANT_SYS_NICE
3895 * sys_nice - change the priority of the current process.
3896 * @increment: priority increment
3898 * sys_setpriority is a more generic, but much slower function that
3899 * does similar things.
3901 SYSCALL_DEFINE1(nice
, int, increment
)
3906 * Setpriority might change our priority at the same moment.
3907 * We don't have to worry. Conceptually one call occurs first
3908 * and we have a single winner.
3910 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3911 nice
= task_nice(current
) + increment
;
3913 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3914 if (increment
< 0 && !can_nice(current
, nice
))
3917 retval
= security_task_setnice(current
, nice
);
3921 set_user_nice(current
, nice
);
3928 * task_prio - return the priority value of a given task.
3929 * @p: the task in question.
3931 * Return: The priority value as seen by users in /proc.
3932 * RT tasks are offset by -200. Normal tasks are centered
3933 * around 0, value goes from -16 to +15.
3935 int task_prio(const struct task_struct
*p
)
3937 return p
->prio
- MAX_RT_PRIO
;
3941 * idle_cpu - is a given CPU idle currently?
3942 * @cpu: the processor in question.
3944 * Return: 1 if the CPU is currently idle. 0 otherwise.
3946 int idle_cpu(int cpu
)
3948 struct rq
*rq
= cpu_rq(cpu
);
3950 if (rq
->curr
!= rq
->idle
)
3957 if (!llist_empty(&rq
->wake_list
))
3965 * idle_task - return the idle task for a given CPU.
3966 * @cpu: the processor in question.
3968 * Return: The idle task for the CPU @cpu.
3970 struct task_struct
*idle_task(int cpu
)
3972 return cpu_rq(cpu
)->idle
;
3976 * find_process_by_pid - find a process with a matching PID value.
3977 * @pid: the pid in question.
3979 * The task of @pid, if found. %NULL otherwise.
3981 static struct task_struct
*find_process_by_pid(pid_t pid
)
3983 return pid
? find_task_by_vpid(pid
) : current
;
3987 * This function initializes the sched_dl_entity of a newly becoming
3988 * SCHED_DEADLINE task.
3990 * Only the static values are considered here, the actual runtime and the
3991 * absolute deadline will be properly calculated when the task is enqueued
3992 * for the first time with its new policy.
3995 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3997 struct sched_dl_entity
*dl_se
= &p
->dl
;
3999 dl_se
->dl_runtime
= attr
->sched_runtime
;
4000 dl_se
->dl_deadline
= attr
->sched_deadline
;
4001 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
4002 dl_se
->flags
= attr
->sched_flags
;
4003 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
4006 * Changing the parameters of a task is 'tricky' and we're not doing
4007 * the correct thing -- also see task_dead_dl() and switched_from_dl().
4009 * What we SHOULD do is delay the bandwidth release until the 0-lag
4010 * point. This would include retaining the task_struct until that time
4011 * and change dl_overflow() to not immediately decrement the current
4014 * Instead we retain the current runtime/deadline and let the new
4015 * parameters take effect after the current reservation period lapses.
4016 * This is safe (albeit pessimistic) because the 0-lag point is always
4017 * before the current scheduling deadline.
4019 * We can still have temporary overloads because we do not delay the
4020 * change in bandwidth until that time; so admission control is
4021 * not on the safe side. It does however guarantee tasks will never
4022 * consume more than promised.
4027 * sched_setparam() passes in -1 for its policy, to let the functions
4028 * it calls know not to change it.
4030 #define SETPARAM_POLICY -1
4032 static void __setscheduler_params(struct task_struct
*p
,
4033 const struct sched_attr
*attr
)
4035 int policy
= attr
->sched_policy
;
4037 if (policy
== SETPARAM_POLICY
)
4042 if (dl_policy(policy
))
4043 __setparam_dl(p
, attr
);
4044 else if (fair_policy(policy
))
4045 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4048 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4049 * !rt_policy. Always setting this ensures that things like
4050 * getparam()/getattr() don't report silly values for !rt tasks.
4052 p
->rt_priority
= attr
->sched_priority
;
4053 p
->normal_prio
= normal_prio(p
);
4057 /* Actually do priority change: must hold pi & rq lock. */
4058 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4059 const struct sched_attr
*attr
, bool keep_boost
)
4061 __setscheduler_params(p
, attr
);
4064 * Keep a potential priority boosting if called from
4065 * sched_setscheduler().
4067 p
->prio
= normal_prio(p
);
4069 p
->prio
= rt_effective_prio(p
, p
->prio
);
4071 if (dl_prio(p
->prio
))
4072 p
->sched_class
= &dl_sched_class
;
4073 else if (rt_prio(p
->prio
))
4074 p
->sched_class
= &rt_sched_class
;
4076 p
->sched_class
= &fair_sched_class
;
4080 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
4082 struct sched_dl_entity
*dl_se
= &p
->dl
;
4084 attr
->sched_priority
= p
->rt_priority
;
4085 attr
->sched_runtime
= dl_se
->dl_runtime
;
4086 attr
->sched_deadline
= dl_se
->dl_deadline
;
4087 attr
->sched_period
= dl_se
->dl_period
;
4088 attr
->sched_flags
= dl_se
->flags
;
4092 * This function validates the new parameters of a -deadline task.
4093 * We ask for the deadline not being zero, and greater or equal
4094 * than the runtime, as well as the period of being zero or
4095 * greater than deadline. Furthermore, we have to be sure that
4096 * user parameters are above the internal resolution of 1us (we
4097 * check sched_runtime only since it is always the smaller one) and
4098 * below 2^63 ns (we have to check both sched_deadline and
4099 * sched_period, as the latter can be zero).
4102 __checkparam_dl(const struct sched_attr
*attr
)
4105 if (attr
->sched_deadline
== 0)
4109 * Since we truncate DL_SCALE bits, make sure we're at least
4112 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
4116 * Since we use the MSB for wrap-around and sign issues, make
4117 * sure it's not set (mind that period can be equal to zero).
4119 if (attr
->sched_deadline
& (1ULL << 63) ||
4120 attr
->sched_period
& (1ULL << 63))
4123 /* runtime <= deadline <= period (if period != 0) */
4124 if ((attr
->sched_period
!= 0 &&
4125 attr
->sched_period
< attr
->sched_deadline
) ||
4126 attr
->sched_deadline
< attr
->sched_runtime
)
4133 * Check the target process has a UID that matches the current process's:
4135 static bool check_same_owner(struct task_struct
*p
)
4137 const struct cred
*cred
= current_cred(), *pcred
;
4141 pcred
= __task_cred(p
);
4142 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4143 uid_eq(cred
->euid
, pcred
->uid
));
4148 static bool dl_param_changed(struct task_struct
*p
, const struct sched_attr
*attr
)
4150 struct sched_dl_entity
*dl_se
= &p
->dl
;
4152 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
4153 dl_se
->dl_deadline
!= attr
->sched_deadline
||
4154 dl_se
->dl_period
!= attr
->sched_period
||
4155 dl_se
->flags
!= attr
->sched_flags
)
4161 static int __sched_setscheduler(struct task_struct
*p
,
4162 const struct sched_attr
*attr
,
4165 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4166 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4167 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4168 int new_effective_prio
, policy
= attr
->sched_policy
;
4169 const struct sched_class
*prev_class
;
4172 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4175 /* May grab non-irq protected spin_locks: */
4176 BUG_ON(in_interrupt());
4178 /* Double check policy once rq lock held: */
4180 reset_on_fork
= p
->sched_reset_on_fork
;
4181 policy
= oldpolicy
= p
->policy
;
4183 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4185 if (!valid_policy(policy
))
4189 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4193 * Valid priorities for SCHED_FIFO and SCHED_RR are
4194 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4195 * SCHED_BATCH and SCHED_IDLE is 0.
4197 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4198 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4200 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4201 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4205 * Allow unprivileged RT tasks to decrease priority:
4207 if (user
&& !capable(CAP_SYS_NICE
)) {
4208 if (fair_policy(policy
)) {
4209 if (attr
->sched_nice
< task_nice(p
) &&
4210 !can_nice(p
, attr
->sched_nice
))
4214 if (rt_policy(policy
)) {
4215 unsigned long rlim_rtprio
=
4216 task_rlimit(p
, RLIMIT_RTPRIO
);
4218 /* Can't set/change the rt policy: */
4219 if (policy
!= p
->policy
&& !rlim_rtprio
)
4222 /* Can't increase priority: */
4223 if (attr
->sched_priority
> p
->rt_priority
&&
4224 attr
->sched_priority
> rlim_rtprio
)
4229 * Can't set/change SCHED_DEADLINE policy at all for now
4230 * (safest behavior); in the future we would like to allow
4231 * unprivileged DL tasks to increase their relative deadline
4232 * or reduce their runtime (both ways reducing utilization)
4234 if (dl_policy(policy
))
4238 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4239 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4241 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4242 if (!can_nice(p
, task_nice(p
)))
4246 /* Can't change other user's priorities: */
4247 if (!check_same_owner(p
))
4250 /* Normal users shall not reset the sched_reset_on_fork flag: */
4251 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4256 retval
= security_task_setscheduler(p
);
4262 * Make sure no PI-waiters arrive (or leave) while we are
4263 * changing the priority of the task:
4265 * To be able to change p->policy safely, the appropriate
4266 * runqueue lock must be held.
4268 rq
= task_rq_lock(p
, &rf
);
4269 update_rq_clock(rq
);
4272 * Changing the policy of the stop threads its a very bad idea:
4274 if (p
== rq
->stop
) {
4275 task_rq_unlock(rq
, p
, &rf
);
4280 * If not changing anything there's no need to proceed further,
4281 * but store a possible modification of reset_on_fork.
4283 if (unlikely(policy
== p
->policy
)) {
4284 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4286 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4288 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4291 p
->sched_reset_on_fork
= reset_on_fork
;
4292 task_rq_unlock(rq
, p
, &rf
);
4298 #ifdef CONFIG_RT_GROUP_SCHED
4300 * Do not allow realtime tasks into groups that have no runtime
4303 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4304 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4305 !task_group_is_autogroup(task_group(p
))) {
4306 task_rq_unlock(rq
, p
, &rf
);
4311 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4312 cpumask_t
*span
= rq
->rd
->span
;
4315 * Don't allow tasks with an affinity mask smaller than
4316 * the entire root_domain to become SCHED_DEADLINE. We
4317 * will also fail if there's no bandwidth available.
4319 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4320 rq
->rd
->dl_bw
.bw
== 0) {
4321 task_rq_unlock(rq
, p
, &rf
);
4328 /* Re-check policy now with rq lock held: */
4329 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4330 policy
= oldpolicy
= -1;
4331 task_rq_unlock(rq
, p
, &rf
);
4336 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4337 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4340 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4341 task_rq_unlock(rq
, p
, &rf
);
4345 p
->sched_reset_on_fork
= reset_on_fork
;
4350 * Take priority boosted tasks into account. If the new
4351 * effective priority is unchanged, we just store the new
4352 * normal parameters and do not touch the scheduler class and
4353 * the runqueue. This will be done when the task deboost
4356 new_effective_prio
= rt_effective_prio(p
, newprio
);
4357 if (new_effective_prio
== oldprio
)
4358 queue_flags
&= ~DEQUEUE_MOVE
;
4361 queued
= task_on_rq_queued(p
);
4362 running
= task_current(rq
, p
);
4364 dequeue_task(rq
, p
, queue_flags
);
4366 put_prev_task(rq
, p
);
4368 prev_class
= p
->sched_class
;
4369 __setscheduler(rq
, p
, attr
, pi
);
4373 * We enqueue to tail when the priority of a task is
4374 * increased (user space view).
4376 if (oldprio
< p
->prio
)
4377 queue_flags
|= ENQUEUE_HEAD
;
4379 enqueue_task(rq
, p
, queue_flags
);
4382 set_curr_task(rq
, p
);
4384 check_class_changed(rq
, p
, prev_class
, oldprio
);
4386 /* Avoid rq from going away on us: */
4388 task_rq_unlock(rq
, p
, &rf
);
4391 rt_mutex_adjust_pi(p
);
4393 /* Run balance callbacks after we've adjusted the PI chain: */
4394 balance_callback(rq
);
4400 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4401 const struct sched_param
*param
, bool check
)
4403 struct sched_attr attr
= {
4404 .sched_policy
= policy
,
4405 .sched_priority
= param
->sched_priority
,
4406 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4409 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4410 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4411 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4412 policy
&= ~SCHED_RESET_ON_FORK
;
4413 attr
.sched_policy
= policy
;
4416 return __sched_setscheduler(p
, &attr
, check
, true);
4419 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4420 * @p: the task in question.
4421 * @policy: new policy.
4422 * @param: structure containing the new RT priority.
4424 * Return: 0 on success. An error code otherwise.
4426 * NOTE that the task may be already dead.
4428 int sched_setscheduler(struct task_struct
*p
, int policy
,
4429 const struct sched_param
*param
)
4431 return _sched_setscheduler(p
, policy
, param
, true);
4433 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4435 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4437 return __sched_setscheduler(p
, attr
, true, true);
4439 EXPORT_SYMBOL_GPL(sched_setattr
);
4442 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4443 * @p: the task in question.
4444 * @policy: new policy.
4445 * @param: structure containing the new RT priority.
4447 * Just like sched_setscheduler, only don't bother checking if the
4448 * current context has permission. For example, this is needed in
4449 * stop_machine(): we create temporary high priority worker threads,
4450 * but our caller might not have that capability.
4452 * Return: 0 on success. An error code otherwise.
4454 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4455 const struct sched_param
*param
)
4457 return _sched_setscheduler(p
, policy
, param
, false);
4459 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4462 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4464 struct sched_param lparam
;
4465 struct task_struct
*p
;
4468 if (!param
|| pid
< 0)
4470 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4475 p
= find_process_by_pid(pid
);
4477 retval
= sched_setscheduler(p
, policy
, &lparam
);
4484 * Mimics kernel/events/core.c perf_copy_attr().
4486 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
4491 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4494 /* Zero the full structure, so that a short copy will be nice: */
4495 memset(attr
, 0, sizeof(*attr
));
4497 ret
= get_user(size
, &uattr
->size
);
4501 /* Bail out on silly large: */
4502 if (size
> PAGE_SIZE
)
4505 /* ABI compatibility quirk: */
4507 size
= SCHED_ATTR_SIZE_VER0
;
4509 if (size
< SCHED_ATTR_SIZE_VER0
)
4513 * If we're handed a bigger struct than we know of,
4514 * ensure all the unknown bits are 0 - i.e. new
4515 * user-space does not rely on any kernel feature
4516 * extensions we dont know about yet.
4518 if (size
> sizeof(*attr
)) {
4519 unsigned char __user
*addr
;
4520 unsigned char __user
*end
;
4523 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4524 end
= (void __user
*)uattr
+ size
;
4526 for (; addr
< end
; addr
++) {
4527 ret
= get_user(val
, addr
);
4533 size
= sizeof(*attr
);
4536 ret
= copy_from_user(attr
, uattr
, size
);
4541 * XXX: Do we want to be lenient like existing syscalls; or do we want
4542 * to be strict and return an error on out-of-bounds values?
4544 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4549 put_user(sizeof(*attr
), &uattr
->size
);
4554 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4555 * @pid: the pid in question.
4556 * @policy: new policy.
4557 * @param: structure containing the new RT priority.
4559 * Return: 0 on success. An error code otherwise.
4561 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
4566 return do_sched_setscheduler(pid
, policy
, param
);
4570 * sys_sched_setparam - set/change the RT priority of a thread
4571 * @pid: the pid in question.
4572 * @param: structure containing the new RT priority.
4574 * Return: 0 on success. An error code otherwise.
4576 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4578 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4582 * sys_sched_setattr - same as above, but with extended sched_attr
4583 * @pid: the pid in question.
4584 * @uattr: structure containing the extended parameters.
4585 * @flags: for future extension.
4587 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4588 unsigned int, flags
)
4590 struct sched_attr attr
;
4591 struct task_struct
*p
;
4594 if (!uattr
|| pid
< 0 || flags
)
4597 retval
= sched_copy_attr(uattr
, &attr
);
4601 if ((int)attr
.sched_policy
< 0)
4606 p
= find_process_by_pid(pid
);
4608 retval
= sched_setattr(p
, &attr
);
4615 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4616 * @pid: the pid in question.
4618 * Return: On success, the policy of the thread. Otherwise, a negative error
4621 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4623 struct task_struct
*p
;
4631 p
= find_process_by_pid(pid
);
4633 retval
= security_task_getscheduler(p
);
4636 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4643 * sys_sched_getparam - get the RT priority of a thread
4644 * @pid: the pid in question.
4645 * @param: structure containing the RT priority.
4647 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4650 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4652 struct sched_param lp
= { .sched_priority
= 0 };
4653 struct task_struct
*p
;
4656 if (!param
|| pid
< 0)
4660 p
= find_process_by_pid(pid
);
4665 retval
= security_task_getscheduler(p
);
4669 if (task_has_rt_policy(p
))
4670 lp
.sched_priority
= p
->rt_priority
;
4674 * This one might sleep, we cannot do it with a spinlock held ...
4676 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4685 static int sched_read_attr(struct sched_attr __user
*uattr
,
4686 struct sched_attr
*attr
,
4691 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4695 * If we're handed a smaller struct than we know of,
4696 * ensure all the unknown bits are 0 - i.e. old
4697 * user-space does not get uncomplete information.
4699 if (usize
< sizeof(*attr
)) {
4700 unsigned char *addr
;
4703 addr
= (void *)attr
+ usize
;
4704 end
= (void *)attr
+ sizeof(*attr
);
4706 for (; addr
< end
; addr
++) {
4714 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4722 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4723 * @pid: the pid in question.
4724 * @uattr: structure containing the extended parameters.
4725 * @size: sizeof(attr) for fwd/bwd comp.
4726 * @flags: for future extension.
4728 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4729 unsigned int, size
, unsigned int, flags
)
4731 struct sched_attr attr
= {
4732 .size
= sizeof(struct sched_attr
),
4734 struct task_struct
*p
;
4737 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4738 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4742 p
= find_process_by_pid(pid
);
4747 retval
= security_task_getscheduler(p
);
4751 attr
.sched_policy
= p
->policy
;
4752 if (p
->sched_reset_on_fork
)
4753 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4754 if (task_has_dl_policy(p
))
4755 __getparam_dl(p
, &attr
);
4756 else if (task_has_rt_policy(p
))
4757 attr
.sched_priority
= p
->rt_priority
;
4759 attr
.sched_nice
= task_nice(p
);
4763 retval
= sched_read_attr(uattr
, &attr
, size
);
4771 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4773 cpumask_var_t cpus_allowed
, new_mask
;
4774 struct task_struct
*p
;
4779 p
= find_process_by_pid(pid
);
4785 /* Prevent p going away */
4789 if (p
->flags
& PF_NO_SETAFFINITY
) {
4793 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4797 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4799 goto out_free_cpus_allowed
;
4802 if (!check_same_owner(p
)) {
4804 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4806 goto out_free_new_mask
;
4811 retval
= security_task_setscheduler(p
);
4813 goto out_free_new_mask
;
4816 cpuset_cpus_allowed(p
, cpus_allowed
);
4817 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4820 * Since bandwidth control happens on root_domain basis,
4821 * if admission test is enabled, we only admit -deadline
4822 * tasks allowed to run on all the CPUs in the task's
4826 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4828 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4831 goto out_free_new_mask
;
4837 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4840 cpuset_cpus_allowed(p
, cpus_allowed
);
4841 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4843 * We must have raced with a concurrent cpuset
4844 * update. Just reset the cpus_allowed to the
4845 * cpuset's cpus_allowed
4847 cpumask_copy(new_mask
, cpus_allowed
);
4852 free_cpumask_var(new_mask
);
4853 out_free_cpus_allowed
:
4854 free_cpumask_var(cpus_allowed
);
4860 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4861 struct cpumask
*new_mask
)
4863 if (len
< cpumask_size())
4864 cpumask_clear(new_mask
);
4865 else if (len
> cpumask_size())
4866 len
= cpumask_size();
4868 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4872 * sys_sched_setaffinity - set the CPU affinity of a process
4873 * @pid: pid of the process
4874 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4875 * @user_mask_ptr: user-space pointer to the new CPU mask
4877 * Return: 0 on success. An error code otherwise.
4879 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4880 unsigned long __user
*, user_mask_ptr
)
4882 cpumask_var_t new_mask
;
4885 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4888 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4890 retval
= sched_setaffinity(pid
, new_mask
);
4891 free_cpumask_var(new_mask
);
4895 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4897 struct task_struct
*p
;
4898 unsigned long flags
;
4904 p
= find_process_by_pid(pid
);
4908 retval
= security_task_getscheduler(p
);
4912 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4913 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4914 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4923 * sys_sched_getaffinity - get the CPU affinity of a process
4924 * @pid: pid of the process
4925 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4926 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4928 * Return: size of CPU mask copied to user_mask_ptr on success. An
4929 * error code otherwise.
4931 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4932 unsigned long __user
*, user_mask_ptr
)
4937 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4939 if (len
& (sizeof(unsigned long)-1))
4942 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4945 ret
= sched_getaffinity(pid
, mask
);
4947 size_t retlen
= min_t(size_t, len
, cpumask_size());
4949 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4954 free_cpumask_var(mask
);
4960 * sys_sched_yield - yield the current processor to other threads.
4962 * This function yields the current CPU to other tasks. If there are no
4963 * other threads running on this CPU then this function will return.
4967 SYSCALL_DEFINE0(sched_yield
)
4972 local_irq_disable();
4976 schedstat_inc(rq
->yld_count
);
4977 current
->sched_class
->yield_task(rq
);
4980 * Since we are going to call schedule() anyway, there's
4981 * no need to preempt or enable interrupts:
4985 sched_preempt_enable_no_resched();
4992 #ifndef CONFIG_PREEMPT
4993 int __sched
_cond_resched(void)
4995 if (should_resched(0)) {
4996 preempt_schedule_common();
5001 EXPORT_SYMBOL(_cond_resched
);
5005 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5006 * call schedule, and on return reacquire the lock.
5008 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5009 * operations here to prevent schedule() from being called twice (once via
5010 * spin_unlock(), once by hand).
5012 int __cond_resched_lock(spinlock_t
*lock
)
5014 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5017 lockdep_assert_held(lock
);
5019 if (spin_needbreak(lock
) || resched
) {
5022 preempt_schedule_common();
5030 EXPORT_SYMBOL(__cond_resched_lock
);
5032 int __sched
__cond_resched_softirq(void)
5034 BUG_ON(!in_softirq());
5036 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
5038 preempt_schedule_common();
5044 EXPORT_SYMBOL(__cond_resched_softirq
);
5047 * yield - yield the current processor to other threads.
5049 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5051 * The scheduler is at all times free to pick the calling task as the most
5052 * eligible task to run, if removing the yield() call from your code breaks
5053 * it, its already broken.
5055 * Typical broken usage is:
5060 * where one assumes that yield() will let 'the other' process run that will
5061 * make event true. If the current task is a SCHED_FIFO task that will never
5062 * happen. Never use yield() as a progress guarantee!!
5064 * If you want to use yield() to wait for something, use wait_event().
5065 * If you want to use yield() to be 'nice' for others, use cond_resched().
5066 * If you still want to use yield(), do not!
5068 void __sched
yield(void)
5070 set_current_state(TASK_RUNNING
);
5073 EXPORT_SYMBOL(yield
);
5076 * yield_to - yield the current processor to another thread in
5077 * your thread group, or accelerate that thread toward the
5078 * processor it's on.
5080 * @preempt: whether task preemption is allowed or not
5082 * It's the caller's job to ensure that the target task struct
5083 * can't go away on us before we can do any checks.
5086 * true (>0) if we indeed boosted the target task.
5087 * false (0) if we failed to boost the target.
5088 * -ESRCH if there's no task to yield to.
5090 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5092 struct task_struct
*curr
= current
;
5093 struct rq
*rq
, *p_rq
;
5094 unsigned long flags
;
5097 local_irq_save(flags
);
5103 * If we're the only runnable task on the rq and target rq also
5104 * has only one task, there's absolutely no point in yielding.
5106 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5111 double_rq_lock(rq
, p_rq
);
5112 if (task_rq(p
) != p_rq
) {
5113 double_rq_unlock(rq
, p_rq
);
5117 if (!curr
->sched_class
->yield_to_task
)
5120 if (curr
->sched_class
!= p
->sched_class
)
5123 if (task_running(p_rq
, p
) || p
->state
)
5126 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5128 schedstat_inc(rq
->yld_count
);
5130 * Make p's CPU reschedule; pick_next_entity takes care of
5133 if (preempt
&& rq
!= p_rq
)
5138 double_rq_unlock(rq
, p_rq
);
5140 local_irq_restore(flags
);
5147 EXPORT_SYMBOL_GPL(yield_to
);
5149 int io_schedule_prepare(void)
5151 int old_iowait
= current
->in_iowait
;
5153 current
->in_iowait
= 1;
5154 blk_schedule_flush_plug(current
);
5159 void io_schedule_finish(int token
)
5161 current
->in_iowait
= token
;
5165 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5166 * that process accounting knows that this is a task in IO wait state.
5168 long __sched
io_schedule_timeout(long timeout
)
5173 token
= io_schedule_prepare();
5174 ret
= schedule_timeout(timeout
);
5175 io_schedule_finish(token
);
5179 EXPORT_SYMBOL(io_schedule_timeout
);
5181 void io_schedule(void)
5185 token
= io_schedule_prepare();
5187 io_schedule_finish(token
);
5189 EXPORT_SYMBOL(io_schedule
);
5192 * sys_sched_get_priority_max - return maximum RT priority.
5193 * @policy: scheduling class.
5195 * Return: On success, this syscall returns the maximum
5196 * rt_priority that can be used by a given scheduling class.
5197 * On failure, a negative error code is returned.
5199 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5206 ret
= MAX_USER_RT_PRIO
-1;
5208 case SCHED_DEADLINE
:
5219 * sys_sched_get_priority_min - return minimum RT priority.
5220 * @policy: scheduling class.
5222 * Return: On success, this syscall returns the minimum
5223 * rt_priority that can be used by a given scheduling class.
5224 * On failure, a negative error code is returned.
5226 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5235 case SCHED_DEADLINE
:
5245 * sys_sched_rr_get_interval - return the default timeslice of a process.
5246 * @pid: pid of the process.
5247 * @interval: userspace pointer to the timeslice value.
5249 * this syscall writes the default timeslice value of a given process
5250 * into the user-space timespec buffer. A value of '0' means infinity.
5252 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5255 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5256 struct timespec __user
*, interval
)
5258 struct task_struct
*p
;
5259 unsigned int time_slice
;
5270 p
= find_process_by_pid(pid
);
5274 retval
= security_task_getscheduler(p
);
5278 rq
= task_rq_lock(p
, &rf
);
5280 if (p
->sched_class
->get_rr_interval
)
5281 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5282 task_rq_unlock(rq
, p
, &rf
);
5285 jiffies_to_timespec(time_slice
, &t
);
5286 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5294 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5296 void sched_show_task(struct task_struct
*p
)
5298 unsigned long free
= 0;
5300 unsigned long state
= p
->state
;
5302 /* Make sure the string lines up properly with the number of task states: */
5303 BUILD_BUG_ON(sizeof(TASK_STATE_TO_CHAR_STR
)-1 != ilog2(TASK_STATE_MAX
)+1);
5305 if (!try_get_task_stack(p
))
5308 state
= __ffs(state
) + 1;
5309 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5310 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5311 if (state
== TASK_RUNNING
)
5312 printk(KERN_CONT
" running task ");
5313 #ifdef CONFIG_DEBUG_STACK_USAGE
5314 free
= stack_not_used(p
);
5319 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5321 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5322 task_pid_nr(p
), ppid
,
5323 (unsigned long)task_thread_info(p
)->flags
);
5325 print_worker_info(KERN_INFO
, p
);
5326 show_stack(p
, NULL
);
5330 void show_state_filter(unsigned long state_filter
)
5332 struct task_struct
*g
, *p
;
5334 #if BITS_PER_LONG == 32
5336 " task PC stack pid father\n");
5339 " task PC stack pid father\n");
5342 for_each_process_thread(g
, p
) {
5344 * reset the NMI-timeout, listing all files on a slow
5345 * console might take a lot of time:
5346 * Also, reset softlockup watchdogs on all CPUs, because
5347 * another CPU might be blocked waiting for us to process
5350 touch_nmi_watchdog();
5351 touch_all_softlockup_watchdogs();
5352 if (!state_filter
|| (p
->state
& state_filter
))
5356 #ifdef CONFIG_SCHED_DEBUG
5358 sysrq_sched_debug_show();
5362 * Only show locks if all tasks are dumped:
5365 debug_show_all_locks();
5368 void init_idle_bootup_task(struct task_struct
*idle
)
5370 idle
->sched_class
= &idle_sched_class
;
5374 * init_idle - set up an idle thread for a given CPU
5375 * @idle: task in question
5376 * @cpu: CPU the idle task belongs to
5378 * NOTE: this function does not set the idle thread's NEED_RESCHED
5379 * flag, to make booting more robust.
5381 void init_idle(struct task_struct
*idle
, int cpu
)
5383 struct rq
*rq
= cpu_rq(cpu
);
5384 unsigned long flags
;
5386 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5387 raw_spin_lock(&rq
->lock
);
5389 __sched_fork(0, idle
);
5390 idle
->state
= TASK_RUNNING
;
5391 idle
->se
.exec_start
= sched_clock();
5392 idle
->flags
|= PF_IDLE
;
5394 kasan_unpoison_task_stack(idle
);
5398 * Its possible that init_idle() gets called multiple times on a task,
5399 * in that case do_set_cpus_allowed() will not do the right thing.
5401 * And since this is boot we can forgo the serialization.
5403 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5406 * We're having a chicken and egg problem, even though we are
5407 * holding rq->lock, the CPU isn't yet set to this CPU so the
5408 * lockdep check in task_group() will fail.
5410 * Similar case to sched_fork(). / Alternatively we could
5411 * use task_rq_lock() here and obtain the other rq->lock.
5416 __set_task_cpu(idle
, cpu
);
5419 rq
->curr
= rq
->idle
= idle
;
5420 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5424 raw_spin_unlock(&rq
->lock
);
5425 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5427 /* Set the preempt count _outside_ the spinlocks! */
5428 init_idle_preempt_count(idle
, cpu
);
5431 * The idle tasks have their own, simple scheduling class:
5433 idle
->sched_class
= &idle_sched_class
;
5434 ftrace_graph_init_idle_task(idle
, cpu
);
5435 vtime_init_idle(idle
, cpu
);
5437 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5441 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5442 const struct cpumask
*trial
)
5444 int ret
= 1, trial_cpus
;
5445 struct dl_bw
*cur_dl_b
;
5446 unsigned long flags
;
5448 if (!cpumask_weight(cur
))
5451 rcu_read_lock_sched();
5452 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5453 trial_cpus
= cpumask_weight(trial
);
5455 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5456 if (cur_dl_b
->bw
!= -1 &&
5457 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5459 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5460 rcu_read_unlock_sched();
5465 int task_can_attach(struct task_struct
*p
,
5466 const struct cpumask
*cs_cpus_allowed
)
5471 * Kthreads which disallow setaffinity shouldn't be moved
5472 * to a new cpuset; we don't want to change their CPU
5473 * affinity and isolating such threads by their set of
5474 * allowed nodes is unnecessary. Thus, cpusets are not
5475 * applicable for such threads. This prevents checking for
5476 * success of set_cpus_allowed_ptr() on all attached tasks
5477 * before cpus_allowed may be changed.
5479 if (p
->flags
& PF_NO_SETAFFINITY
) {
5485 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5487 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5492 unsigned long flags
;
5494 rcu_read_lock_sched();
5495 dl_b
= dl_bw_of(dest_cpu
);
5496 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5497 cpus
= dl_bw_cpus(dest_cpu
);
5498 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5503 * We reserve space for this task in the destination
5504 * root_domain, as we can't fail after this point.
5505 * We will free resources in the source root_domain
5506 * later on (see set_cpus_allowed_dl()).
5508 __dl_add(dl_b
, p
->dl
.dl_bw
);
5510 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5511 rcu_read_unlock_sched();
5521 bool sched_smp_initialized __read_mostly
;
5523 #ifdef CONFIG_NUMA_BALANCING
5524 /* Migrate current task p to target_cpu */
5525 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5527 struct migration_arg arg
= { p
, target_cpu
};
5528 int curr_cpu
= task_cpu(p
);
5530 if (curr_cpu
== target_cpu
)
5533 if (!cpumask_test_cpu(target_cpu
, &p
->cpus_allowed
))
5536 /* TODO: This is not properly updating schedstats */
5538 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5539 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5543 * Requeue a task on a given node and accurately track the number of NUMA
5544 * tasks on the runqueues
5546 void sched_setnuma(struct task_struct
*p
, int nid
)
5548 bool queued
, running
;
5552 rq
= task_rq_lock(p
, &rf
);
5553 queued
= task_on_rq_queued(p
);
5554 running
= task_current(rq
, p
);
5557 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5559 put_prev_task(rq
, p
);
5561 p
->numa_preferred_nid
= nid
;
5564 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5566 set_curr_task(rq
, p
);
5567 task_rq_unlock(rq
, p
, &rf
);
5569 #endif /* CONFIG_NUMA_BALANCING */
5571 #ifdef CONFIG_HOTPLUG_CPU
5573 * Ensure that the idle task is using init_mm right before its CPU goes
5576 void idle_task_exit(void)
5578 struct mm_struct
*mm
= current
->active_mm
;
5580 BUG_ON(cpu_online(smp_processor_id()));
5582 if (mm
!= &init_mm
) {
5583 switch_mm_irqs_off(mm
, &init_mm
, current
);
5584 finish_arch_post_lock_switch();
5590 * Since this CPU is going 'away' for a while, fold any nr_active delta
5591 * we might have. Assumes we're called after migrate_tasks() so that the
5592 * nr_active count is stable. We need to take the teardown thread which
5593 * is calling this into account, so we hand in adjust = 1 to the load
5596 * Also see the comment "Global load-average calculations".
5598 static void calc_load_migrate(struct rq
*rq
)
5600 long delta
= calc_load_fold_active(rq
, 1);
5602 atomic_long_add(delta
, &calc_load_tasks
);
5605 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5609 static const struct sched_class fake_sched_class
= {
5610 .put_prev_task
= put_prev_task_fake
,
5613 static struct task_struct fake_task
= {
5615 * Avoid pull_{rt,dl}_task()
5617 .prio
= MAX_PRIO
+ 1,
5618 .sched_class
= &fake_sched_class
,
5622 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5623 * try_to_wake_up()->select_task_rq().
5625 * Called with rq->lock held even though we'er in stop_machine() and
5626 * there's no concurrency possible, we hold the required locks anyway
5627 * because of lock validation efforts.
5629 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
5631 struct rq
*rq
= dead_rq
;
5632 struct task_struct
*next
, *stop
= rq
->stop
;
5633 struct rq_flags orf
= *rf
;
5637 * Fudge the rq selection such that the below task selection loop
5638 * doesn't get stuck on the currently eligible stop task.
5640 * We're currently inside stop_machine() and the rq is either stuck
5641 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5642 * either way we should never end up calling schedule() until we're
5648 * put_prev_task() and pick_next_task() sched
5649 * class method both need to have an up-to-date
5650 * value of rq->clock[_task]
5652 update_rq_clock(rq
);
5656 * There's this thread running, bail when that's the only
5659 if (rq
->nr_running
== 1)
5663 * pick_next_task() assumes pinned rq->lock:
5665 next
= pick_next_task(rq
, &fake_task
, rf
);
5667 next
->sched_class
->put_prev_task(rq
, next
);
5670 * Rules for changing task_struct::cpus_allowed are holding
5671 * both pi_lock and rq->lock, such that holding either
5672 * stabilizes the mask.
5674 * Drop rq->lock is not quite as disastrous as it usually is
5675 * because !cpu_active at this point, which means load-balance
5676 * will not interfere. Also, stop-machine.
5679 raw_spin_lock(&next
->pi_lock
);
5683 * Since we're inside stop-machine, _nothing_ should have
5684 * changed the task, WARN if weird stuff happened, because in
5685 * that case the above rq->lock drop is a fail too.
5687 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5688 raw_spin_unlock(&next
->pi_lock
);
5692 /* Find suitable destination for @next, with force if needed. */
5693 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5694 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
5695 if (rq
!= dead_rq
) {
5701 raw_spin_unlock(&next
->pi_lock
);
5706 #endif /* CONFIG_HOTPLUG_CPU */
5708 void set_rq_online(struct rq
*rq
)
5711 const struct sched_class
*class;
5713 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5716 for_each_class(class) {
5717 if (class->rq_online
)
5718 class->rq_online(rq
);
5723 void set_rq_offline(struct rq
*rq
)
5726 const struct sched_class
*class;
5728 for_each_class(class) {
5729 if (class->rq_offline
)
5730 class->rq_offline(rq
);
5733 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5738 static void set_cpu_rq_start_time(unsigned int cpu
)
5740 struct rq
*rq
= cpu_rq(cpu
);
5742 rq
->age_stamp
= sched_clock_cpu(cpu
);
5746 * used to mark begin/end of suspend/resume:
5748 static int num_cpus_frozen
;
5751 * Update cpusets according to cpu_active mask. If cpusets are
5752 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5753 * around partition_sched_domains().
5755 * If we come here as part of a suspend/resume, don't touch cpusets because we
5756 * want to restore it back to its original state upon resume anyway.
5758 static void cpuset_cpu_active(void)
5760 if (cpuhp_tasks_frozen
) {
5762 * num_cpus_frozen tracks how many CPUs are involved in suspend
5763 * resume sequence. As long as this is not the last online
5764 * operation in the resume sequence, just build a single sched
5765 * domain, ignoring cpusets.
5768 if (likely(num_cpus_frozen
)) {
5769 partition_sched_domains(1, NULL
, NULL
);
5773 * This is the last CPU online operation. So fall through and
5774 * restore the original sched domains by considering the
5775 * cpuset configurations.
5778 cpuset_update_active_cpus();
5781 static int cpuset_cpu_inactive(unsigned int cpu
)
5783 unsigned long flags
;
5788 if (!cpuhp_tasks_frozen
) {
5789 rcu_read_lock_sched();
5790 dl_b
= dl_bw_of(cpu
);
5792 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5793 cpus
= dl_bw_cpus(cpu
);
5794 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5795 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5797 rcu_read_unlock_sched();
5801 cpuset_update_active_cpus();
5804 partition_sched_domains(1, NULL
, NULL
);
5809 int sched_cpu_activate(unsigned int cpu
)
5811 struct rq
*rq
= cpu_rq(cpu
);
5814 set_cpu_active(cpu
, true);
5816 if (sched_smp_initialized
) {
5817 sched_domains_numa_masks_set(cpu
);
5818 cpuset_cpu_active();
5822 * Put the rq online, if not already. This happens:
5824 * 1) In the early boot process, because we build the real domains
5825 * after all CPUs have been brought up.
5827 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5830 rq_lock_irqsave(rq
, &rf
);
5832 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5835 rq_unlock_irqrestore(rq
, &rf
);
5837 update_max_interval();
5842 int sched_cpu_deactivate(unsigned int cpu
)
5846 set_cpu_active(cpu
, false);
5848 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5849 * users of this state to go away such that all new such users will
5852 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5853 * not imply sync_sched(), so wait for both.
5855 * Do sync before park smpboot threads to take care the rcu boost case.
5857 if (IS_ENABLED(CONFIG_PREEMPT
))
5858 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
5862 if (!sched_smp_initialized
)
5865 ret
= cpuset_cpu_inactive(cpu
);
5867 set_cpu_active(cpu
, true);
5870 sched_domains_numa_masks_clear(cpu
);
5874 static void sched_rq_cpu_starting(unsigned int cpu
)
5876 struct rq
*rq
= cpu_rq(cpu
);
5878 rq
->calc_load_update
= calc_load_update
;
5879 update_max_interval();
5882 int sched_cpu_starting(unsigned int cpu
)
5884 set_cpu_rq_start_time(cpu
);
5885 sched_rq_cpu_starting(cpu
);
5889 #ifdef CONFIG_HOTPLUG_CPU
5890 int sched_cpu_dying(unsigned int cpu
)
5892 struct rq
*rq
= cpu_rq(cpu
);
5895 /* Handle pending wakeups and then migrate everything off */
5896 sched_ttwu_pending();
5898 rq_lock_irqsave(rq
, &rf
);
5900 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5903 migrate_tasks(rq
, &rf
);
5904 BUG_ON(rq
->nr_running
!= 1);
5905 rq_unlock_irqrestore(rq
, &rf
);
5907 calc_load_migrate(rq
);
5908 update_max_interval();
5909 nohz_balance_exit_idle(cpu
);
5915 #ifdef CONFIG_SCHED_SMT
5916 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
5918 static void sched_init_smt(void)
5921 * We've enumerated all CPUs and will assume that if any CPU
5922 * has SMT siblings, CPU0 will too.
5924 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5925 static_branch_enable(&sched_smt_present
);
5928 static inline void sched_init_smt(void) { }
5931 void __init
sched_init_smp(void)
5933 cpumask_var_t non_isolated_cpus
;
5935 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
5936 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
5941 * There's no userspace yet to cause hotplug operations; hence all the
5942 * CPU masks are stable and all blatant races in the below code cannot
5945 mutex_lock(&sched_domains_mutex
);
5946 init_sched_domains(cpu_active_mask
);
5947 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
5948 if (cpumask_empty(non_isolated_cpus
))
5949 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
5950 mutex_unlock(&sched_domains_mutex
);
5952 /* Move init over to a non-isolated CPU */
5953 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
5955 sched_init_granularity();
5956 free_cpumask_var(non_isolated_cpus
);
5958 init_sched_rt_class();
5959 init_sched_dl_class();
5962 sched_clock_init_late();
5964 sched_smp_initialized
= true;
5967 static int __init
migration_init(void)
5969 sched_rq_cpu_starting(smp_processor_id());
5972 early_initcall(migration_init
);
5975 void __init
sched_init_smp(void)
5977 sched_init_granularity();
5978 sched_clock_init_late();
5980 #endif /* CONFIG_SMP */
5982 int in_sched_functions(unsigned long addr
)
5984 return in_lock_functions(addr
) ||
5985 (addr
>= (unsigned long)__sched_text_start
5986 && addr
< (unsigned long)__sched_text_end
);
5989 #ifdef CONFIG_CGROUP_SCHED
5991 * Default task group.
5992 * Every task in system belongs to this group at bootup.
5994 struct task_group root_task_group
;
5995 LIST_HEAD(task_groups
);
5997 /* Cacheline aligned slab cache for task_group */
5998 static struct kmem_cache
*task_group_cache __read_mostly
;
6001 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6002 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
6004 #define WAIT_TABLE_BITS 8
6005 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
6006 static wait_queue_head_t bit_wait_table
[WAIT_TABLE_SIZE
] __cacheline_aligned
;
6008 wait_queue_head_t
*bit_waitqueue(void *word
, int bit
)
6010 const int shift
= BITS_PER_LONG
== 32 ? 5 : 6;
6011 unsigned long val
= (unsigned long)word
<< shift
| bit
;
6013 return bit_wait_table
+ hash_long(val
, WAIT_TABLE_BITS
);
6015 EXPORT_SYMBOL(bit_waitqueue
);
6017 void __init
sched_init(void)
6020 unsigned long alloc_size
= 0, ptr
;
6024 for (i
= 0; i
< WAIT_TABLE_SIZE
; i
++)
6025 init_waitqueue_head(bit_wait_table
+ i
);
6027 #ifdef CONFIG_FAIR_GROUP_SCHED
6028 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6030 #ifdef CONFIG_RT_GROUP_SCHED
6031 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6034 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6036 #ifdef CONFIG_FAIR_GROUP_SCHED
6037 root_task_group
.se
= (struct sched_entity
**)ptr
;
6038 ptr
+= nr_cpu_ids
* sizeof(void **);
6040 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6041 ptr
+= nr_cpu_ids
* sizeof(void **);
6043 #endif /* CONFIG_FAIR_GROUP_SCHED */
6044 #ifdef CONFIG_RT_GROUP_SCHED
6045 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6046 ptr
+= nr_cpu_ids
* sizeof(void **);
6048 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6049 ptr
+= nr_cpu_ids
* sizeof(void **);
6051 #endif /* CONFIG_RT_GROUP_SCHED */
6053 #ifdef CONFIG_CPUMASK_OFFSTACK
6054 for_each_possible_cpu(i
) {
6055 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6056 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6057 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6058 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6060 #endif /* CONFIG_CPUMASK_OFFSTACK */
6062 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
6063 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6066 init_defrootdomain();
6069 #ifdef CONFIG_RT_GROUP_SCHED
6070 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6071 global_rt_period(), global_rt_runtime());
6072 #endif /* CONFIG_RT_GROUP_SCHED */
6074 #ifdef CONFIG_CGROUP_SCHED
6075 task_group_cache
= KMEM_CACHE(task_group
, 0);
6077 list_add(&root_task_group
.list
, &task_groups
);
6078 INIT_LIST_HEAD(&root_task_group
.children
);
6079 INIT_LIST_HEAD(&root_task_group
.siblings
);
6080 autogroup_init(&init_task
);
6081 #endif /* CONFIG_CGROUP_SCHED */
6083 for_each_possible_cpu(i
) {
6087 raw_spin_lock_init(&rq
->lock
);
6089 rq
->calc_load_active
= 0;
6090 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6091 init_cfs_rq(&rq
->cfs
);
6092 init_rt_rq(&rq
->rt
);
6093 init_dl_rq(&rq
->dl
);
6094 #ifdef CONFIG_FAIR_GROUP_SCHED
6095 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6096 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6097 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6099 * How much CPU bandwidth does root_task_group get?
6101 * In case of task-groups formed thr' the cgroup filesystem, it
6102 * gets 100% of the CPU resources in the system. This overall
6103 * system CPU resource is divided among the tasks of
6104 * root_task_group and its child task-groups in a fair manner,
6105 * based on each entity's (task or task-group's) weight
6106 * (se->load.weight).
6108 * In other words, if root_task_group has 10 tasks of weight
6109 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6110 * then A0's share of the CPU resource is:
6112 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6114 * We achieve this by letting root_task_group's tasks sit
6115 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6117 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6118 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6119 #endif /* CONFIG_FAIR_GROUP_SCHED */
6121 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6122 #ifdef CONFIG_RT_GROUP_SCHED
6123 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6126 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6127 rq
->cpu_load
[j
] = 0;
6132 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6133 rq
->balance_callback
= NULL
;
6134 rq
->active_balance
= 0;
6135 rq
->next_balance
= jiffies
;
6140 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6141 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6143 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6145 rq_attach_root(rq
, &def_root_domain
);
6146 #ifdef CONFIG_NO_HZ_COMMON
6147 rq
->last_load_update_tick
= jiffies
;
6150 #ifdef CONFIG_NO_HZ_FULL
6151 rq
->last_sched_tick
= 0;
6153 #endif /* CONFIG_SMP */
6155 atomic_set(&rq
->nr_iowait
, 0);
6158 set_load_weight(&init_task
);
6161 * The boot idle thread does lazy MMU switching as well:
6164 enter_lazy_tlb(&init_mm
, current
);
6167 * Make us the idle thread. Technically, schedule() should not be
6168 * called from this thread, however somewhere below it might be,
6169 * but because we are the idle thread, we just pick up running again
6170 * when this runqueue becomes "idle".
6172 init_idle(current
, smp_processor_id());
6174 calc_load_update
= jiffies
+ LOAD_FREQ
;
6177 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6178 /* May be allocated at isolcpus cmdline parse time */
6179 if (cpu_isolated_map
== NULL
)
6180 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6181 idle_thread_set_boot_cpu();
6182 set_cpu_rq_start_time(smp_processor_id());
6184 init_sched_fair_class();
6188 scheduler_running
= 1;
6191 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6192 static inline int preempt_count_equals(int preempt_offset
)
6194 int nested
= preempt_count() + rcu_preempt_depth();
6196 return (nested
== preempt_offset
);
6199 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6202 * Blocking primitives will set (and therefore destroy) current->state,
6203 * since we will exit with TASK_RUNNING make sure we enter with it,
6204 * otherwise we will destroy state.
6206 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6207 "do not call blocking ops when !TASK_RUNNING; "
6208 "state=%lx set at [<%p>] %pS\n",
6210 (void *)current
->task_state_change
,
6211 (void *)current
->task_state_change
);
6213 ___might_sleep(file
, line
, preempt_offset
);
6215 EXPORT_SYMBOL(__might_sleep
);
6217 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6219 /* Ratelimiting timestamp: */
6220 static unsigned long prev_jiffy
;
6222 unsigned long preempt_disable_ip
;
6224 /* WARN_ON_ONCE() by default, no rate limit required: */
6227 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6228 !is_idle_task(current
)) ||
6229 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6231 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6233 prev_jiffy
= jiffies
;
6235 /* Save this before calling printk(), since that will clobber it: */
6236 preempt_disable_ip
= get_preempt_disable_ip(current
);
6239 "BUG: sleeping function called from invalid context at %s:%d\n",
6242 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6243 in_atomic(), irqs_disabled(),
6244 current
->pid
, current
->comm
);
6246 if (task_stack_end_corrupted(current
))
6247 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6249 debug_show_held_locks(current
);
6250 if (irqs_disabled())
6251 print_irqtrace_events(current
);
6252 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6253 && !preempt_count_equals(preempt_offset
)) {
6254 pr_err("Preemption disabled at:");
6255 print_ip_sym(preempt_disable_ip
);
6259 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6261 EXPORT_SYMBOL(___might_sleep
);
6264 #ifdef CONFIG_MAGIC_SYSRQ
6265 void normalize_rt_tasks(void)
6267 struct task_struct
*g
, *p
;
6268 struct sched_attr attr
= {
6269 .sched_policy
= SCHED_NORMAL
,
6272 read_lock(&tasklist_lock
);
6273 for_each_process_thread(g
, p
) {
6275 * Only normalize user tasks:
6277 if (p
->flags
& PF_KTHREAD
)
6280 p
->se
.exec_start
= 0;
6281 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6282 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6283 schedstat_set(p
->se
.statistics
.block_start
, 0);
6285 if (!dl_task(p
) && !rt_task(p
)) {
6287 * Renice negative nice level userspace
6290 if (task_nice(p
) < 0)
6291 set_user_nice(p
, 0);
6295 __sched_setscheduler(p
, &attr
, false, false);
6297 read_unlock(&tasklist_lock
);
6300 #endif /* CONFIG_MAGIC_SYSRQ */
6302 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6304 * These functions are only useful for the IA64 MCA handling, or kdb.
6306 * They can only be called when the whole system has been
6307 * stopped - every CPU needs to be quiescent, and no scheduling
6308 * activity can take place. Using them for anything else would
6309 * be a serious bug, and as a result, they aren't even visible
6310 * under any other configuration.
6314 * curr_task - return the current task for a given CPU.
6315 * @cpu: the processor in question.
6317 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6319 * Return: The current task for @cpu.
6321 struct task_struct
*curr_task(int cpu
)
6323 return cpu_curr(cpu
);
6326 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6330 * set_curr_task - set the current task for a given CPU.
6331 * @cpu: the processor in question.
6332 * @p: the task pointer to set.
6334 * Description: This function must only be used when non-maskable interrupts
6335 * are serviced on a separate stack. It allows the architecture to switch the
6336 * notion of the current task on a CPU in a non-blocking manner. This function
6337 * must be called with all CPU's synchronized, and interrupts disabled, the
6338 * and caller must save the original value of the current task (see
6339 * curr_task() above) and restore that value before reenabling interrupts and
6340 * re-starting the system.
6342 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6344 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6351 #ifdef CONFIG_CGROUP_SCHED
6352 /* task_group_lock serializes the addition/removal of task groups */
6353 static DEFINE_SPINLOCK(task_group_lock
);
6355 static void sched_free_group(struct task_group
*tg
)
6357 free_fair_sched_group(tg
);
6358 free_rt_sched_group(tg
);
6360 kmem_cache_free(task_group_cache
, tg
);
6363 /* allocate runqueue etc for a new task group */
6364 struct task_group
*sched_create_group(struct task_group
*parent
)
6366 struct task_group
*tg
;
6368 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6370 return ERR_PTR(-ENOMEM
);
6372 if (!alloc_fair_sched_group(tg
, parent
))
6375 if (!alloc_rt_sched_group(tg
, parent
))
6381 sched_free_group(tg
);
6382 return ERR_PTR(-ENOMEM
);
6385 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6387 unsigned long flags
;
6389 spin_lock_irqsave(&task_group_lock
, flags
);
6390 list_add_rcu(&tg
->list
, &task_groups
);
6392 /* Root should already exist: */
6395 tg
->parent
= parent
;
6396 INIT_LIST_HEAD(&tg
->children
);
6397 list_add_rcu(&tg
->siblings
, &parent
->children
);
6398 spin_unlock_irqrestore(&task_group_lock
, flags
);
6400 online_fair_sched_group(tg
);
6403 /* rcu callback to free various structures associated with a task group */
6404 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6406 /* Now it should be safe to free those cfs_rqs: */
6407 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6410 void sched_destroy_group(struct task_group
*tg
)
6412 /* Wait for possible concurrent references to cfs_rqs complete: */
6413 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6416 void sched_offline_group(struct task_group
*tg
)
6418 unsigned long flags
;
6420 /* End participation in shares distribution: */
6421 unregister_fair_sched_group(tg
);
6423 spin_lock_irqsave(&task_group_lock
, flags
);
6424 list_del_rcu(&tg
->list
);
6425 list_del_rcu(&tg
->siblings
);
6426 spin_unlock_irqrestore(&task_group_lock
, flags
);
6429 static void sched_change_group(struct task_struct
*tsk
, int type
)
6431 struct task_group
*tg
;
6434 * All callers are synchronized by task_rq_lock(); we do not use RCU
6435 * which is pointless here. Thus, we pass "true" to task_css_check()
6436 * to prevent lockdep warnings.
6438 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
6439 struct task_group
, css
);
6440 tg
= autogroup_task_group(tsk
, tg
);
6441 tsk
->sched_task_group
= tg
;
6443 #ifdef CONFIG_FAIR_GROUP_SCHED
6444 if (tsk
->sched_class
->task_change_group
)
6445 tsk
->sched_class
->task_change_group(tsk
, type
);
6448 set_task_rq(tsk
, task_cpu(tsk
));
6452 * Change task's runqueue when it moves between groups.
6454 * The caller of this function should have put the task in its new group by
6455 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6458 void sched_move_task(struct task_struct
*tsk
)
6460 int queued
, running
, queue_flags
=
6461 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
6465 rq
= task_rq_lock(tsk
, &rf
);
6466 update_rq_clock(rq
);
6468 running
= task_current(rq
, tsk
);
6469 queued
= task_on_rq_queued(tsk
);
6472 dequeue_task(rq
, tsk
, queue_flags
);
6474 put_prev_task(rq
, tsk
);
6476 sched_change_group(tsk
, TASK_MOVE_GROUP
);
6479 enqueue_task(rq
, tsk
, queue_flags
);
6481 set_curr_task(rq
, tsk
);
6483 task_rq_unlock(rq
, tsk
, &rf
);
6485 #endif /* CONFIG_CGROUP_SCHED */
6487 #ifdef CONFIG_RT_GROUP_SCHED
6489 * Ensure that the real time constraints are schedulable.
6491 static DEFINE_MUTEX(rt_constraints_mutex
);
6493 /* Must be called with tasklist_lock held */
6494 static inline int tg_has_rt_tasks(struct task_group
*tg
)
6496 struct task_struct
*g
, *p
;
6499 * Autogroups do not have RT tasks; see autogroup_create().
6501 if (task_group_is_autogroup(tg
))
6504 for_each_process_thread(g
, p
) {
6505 if (rt_task(p
) && task_group(p
) == tg
)
6512 struct rt_schedulable_data
{
6513 struct task_group
*tg
;
6518 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
6520 struct rt_schedulable_data
*d
= data
;
6521 struct task_group
*child
;
6522 unsigned long total
, sum
= 0;
6523 u64 period
, runtime
;
6525 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6526 runtime
= tg
->rt_bandwidth
.rt_runtime
;
6529 period
= d
->rt_period
;
6530 runtime
= d
->rt_runtime
;
6534 * Cannot have more runtime than the period.
6536 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
6540 * Ensure we don't starve existing RT tasks.
6542 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
6545 total
= to_ratio(period
, runtime
);
6548 * Nobody can have more than the global setting allows.
6550 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
6554 * The sum of our children's runtime should not exceed our own.
6556 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
6557 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
6558 runtime
= child
->rt_bandwidth
.rt_runtime
;
6560 if (child
== d
->tg
) {
6561 period
= d
->rt_period
;
6562 runtime
= d
->rt_runtime
;
6565 sum
+= to_ratio(period
, runtime
);
6574 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
6578 struct rt_schedulable_data data
= {
6580 .rt_period
= period
,
6581 .rt_runtime
= runtime
,
6585 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
6591 static int tg_set_rt_bandwidth(struct task_group
*tg
,
6592 u64 rt_period
, u64 rt_runtime
)
6597 * Disallowing the root group RT runtime is BAD, it would disallow the
6598 * kernel creating (and or operating) RT threads.
6600 if (tg
== &root_task_group
&& rt_runtime
== 0)
6603 /* No period doesn't make any sense. */
6607 mutex_lock(&rt_constraints_mutex
);
6608 read_lock(&tasklist_lock
);
6609 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
6613 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6614 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
6615 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
6617 for_each_possible_cpu(i
) {
6618 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
6620 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6621 rt_rq
->rt_runtime
= rt_runtime
;
6622 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6624 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6626 read_unlock(&tasklist_lock
);
6627 mutex_unlock(&rt_constraints_mutex
);
6632 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
6634 u64 rt_runtime
, rt_period
;
6636 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6637 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
6638 if (rt_runtime_us
< 0)
6639 rt_runtime
= RUNTIME_INF
;
6641 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6644 static long sched_group_rt_runtime(struct task_group
*tg
)
6648 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
6651 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
6652 do_div(rt_runtime_us
, NSEC_PER_USEC
);
6653 return rt_runtime_us
;
6656 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
6658 u64 rt_runtime
, rt_period
;
6660 rt_period
= rt_period_us
* NSEC_PER_USEC
;
6661 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
6663 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6666 static long sched_group_rt_period(struct task_group
*tg
)
6670 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6671 do_div(rt_period_us
, NSEC_PER_USEC
);
6672 return rt_period_us
;
6674 #endif /* CONFIG_RT_GROUP_SCHED */
6676 #ifdef CONFIG_RT_GROUP_SCHED
6677 static int sched_rt_global_constraints(void)
6681 mutex_lock(&rt_constraints_mutex
);
6682 read_lock(&tasklist_lock
);
6683 ret
= __rt_schedulable(NULL
, 0, 0);
6684 read_unlock(&tasklist_lock
);
6685 mutex_unlock(&rt_constraints_mutex
);
6690 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
6692 /* Don't accept realtime tasks when there is no way for them to run */
6693 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
6699 #else /* !CONFIG_RT_GROUP_SCHED */
6700 static int sched_rt_global_constraints(void)
6702 unsigned long flags
;
6705 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
6706 for_each_possible_cpu(i
) {
6707 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
6709 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6710 rt_rq
->rt_runtime
= global_rt_runtime();
6711 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6713 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
6717 #endif /* CONFIG_RT_GROUP_SCHED */
6719 static int sched_dl_global_validate(void)
6721 u64 runtime
= global_rt_runtime();
6722 u64 period
= global_rt_period();
6723 u64 new_bw
= to_ratio(period
, runtime
);
6726 unsigned long flags
;
6729 * Here we want to check the bandwidth not being set to some
6730 * value smaller than the currently allocated bandwidth in
6731 * any of the root_domains.
6733 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6734 * cycling on root_domains... Discussion on different/better
6735 * solutions is welcome!
6737 for_each_possible_cpu(cpu
) {
6738 rcu_read_lock_sched();
6739 dl_b
= dl_bw_of(cpu
);
6741 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
6742 if (new_bw
< dl_b
->total_bw
)
6744 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
6746 rcu_read_unlock_sched();
6755 static void sched_dl_do_global(void)
6760 unsigned long flags
;
6762 def_dl_bandwidth
.dl_period
= global_rt_period();
6763 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
6765 if (global_rt_runtime() != RUNTIME_INF
)
6766 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
6769 * FIXME: As above...
6771 for_each_possible_cpu(cpu
) {
6772 rcu_read_lock_sched();
6773 dl_b
= dl_bw_of(cpu
);
6775 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
6777 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
6779 rcu_read_unlock_sched();
6783 static int sched_rt_global_validate(void)
6785 if (sysctl_sched_rt_period
<= 0)
6788 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
6789 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
6795 static void sched_rt_do_global(void)
6797 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
6798 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
6801 int sched_rt_handler(struct ctl_table
*table
, int write
,
6802 void __user
*buffer
, size_t *lenp
,
6805 int old_period
, old_runtime
;
6806 static DEFINE_MUTEX(mutex
);
6810 old_period
= sysctl_sched_rt_period
;
6811 old_runtime
= sysctl_sched_rt_runtime
;
6813 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
6815 if (!ret
&& write
) {
6816 ret
= sched_rt_global_validate();
6820 ret
= sched_dl_global_validate();
6824 ret
= sched_rt_global_constraints();
6828 sched_rt_do_global();
6829 sched_dl_do_global();
6833 sysctl_sched_rt_period
= old_period
;
6834 sysctl_sched_rt_runtime
= old_runtime
;
6836 mutex_unlock(&mutex
);
6841 int sched_rr_handler(struct ctl_table
*table
, int write
,
6842 void __user
*buffer
, size_t *lenp
,
6846 static DEFINE_MUTEX(mutex
);
6849 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
6851 * Make sure that internally we keep jiffies.
6852 * Also, writing zero resets the timeslice to default:
6854 if (!ret
&& write
) {
6855 sched_rr_timeslice
=
6856 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
6857 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
6859 mutex_unlock(&mutex
);
6863 #ifdef CONFIG_CGROUP_SCHED
6865 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
6867 return css
? container_of(css
, struct task_group
, css
) : NULL
;
6870 static struct cgroup_subsys_state
*
6871 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6873 struct task_group
*parent
= css_tg(parent_css
);
6874 struct task_group
*tg
;
6877 /* This is early initialization for the top cgroup */
6878 return &root_task_group
.css
;
6881 tg
= sched_create_group(parent
);
6883 return ERR_PTR(-ENOMEM
);
6888 /* Expose task group only after completing cgroup initialization */
6889 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
6891 struct task_group
*tg
= css_tg(css
);
6892 struct task_group
*parent
= css_tg(css
->parent
);
6895 sched_online_group(tg
, parent
);
6899 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
6901 struct task_group
*tg
= css_tg(css
);
6903 sched_offline_group(tg
);
6906 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
6908 struct task_group
*tg
= css_tg(css
);
6911 * Relies on the RCU grace period between css_released() and this.
6913 sched_free_group(tg
);
6917 * This is called before wake_up_new_task(), therefore we really only
6918 * have to set its group bits, all the other stuff does not apply.
6920 static void cpu_cgroup_fork(struct task_struct
*task
)
6925 rq
= task_rq_lock(task
, &rf
);
6927 update_rq_clock(rq
);
6928 sched_change_group(task
, TASK_SET_GROUP
);
6930 task_rq_unlock(rq
, task
, &rf
);
6933 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
6935 struct task_struct
*task
;
6936 struct cgroup_subsys_state
*css
;
6939 cgroup_taskset_for_each(task
, css
, tset
) {
6940 #ifdef CONFIG_RT_GROUP_SCHED
6941 if (!sched_rt_can_attach(css_tg(css
), task
))
6944 /* We don't support RT-tasks being in separate groups */
6945 if (task
->sched_class
!= &fair_sched_class
)
6949 * Serialize against wake_up_new_task() such that if its
6950 * running, we're sure to observe its full state.
6952 raw_spin_lock_irq(&task
->pi_lock
);
6954 * Avoid calling sched_move_task() before wake_up_new_task()
6955 * has happened. This would lead to problems with PELT, due to
6956 * move wanting to detach+attach while we're not attached yet.
6958 if (task
->state
== TASK_NEW
)
6960 raw_spin_unlock_irq(&task
->pi_lock
);
6968 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
6970 struct task_struct
*task
;
6971 struct cgroup_subsys_state
*css
;
6973 cgroup_taskset_for_each(task
, css
, tset
)
6974 sched_move_task(task
);
6977 #ifdef CONFIG_FAIR_GROUP_SCHED
6978 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
6979 struct cftype
*cftype
, u64 shareval
)
6981 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
6984 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
6987 struct task_group
*tg
= css_tg(css
);
6989 return (u64
) scale_load_down(tg
->shares
);
6992 #ifdef CONFIG_CFS_BANDWIDTH
6993 static DEFINE_MUTEX(cfs_constraints_mutex
);
6995 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
6996 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
6998 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7000 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7002 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7003 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7005 if (tg
== &root_task_group
)
7009 * Ensure we have at some amount of bandwidth every period. This is
7010 * to prevent reaching a state of large arrears when throttled via
7011 * entity_tick() resulting in prolonged exit starvation.
7013 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7017 * Likewise, bound things on the otherside by preventing insane quota
7018 * periods. This also allows us to normalize in computing quota
7021 if (period
> max_cfs_quota_period
)
7025 * Prevent race between setting of cfs_rq->runtime_enabled and
7026 * unthrottle_offline_cfs_rqs().
7029 mutex_lock(&cfs_constraints_mutex
);
7030 ret
= __cfs_schedulable(tg
, period
, quota
);
7034 runtime_enabled
= quota
!= RUNTIME_INF
;
7035 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7037 * If we need to toggle cfs_bandwidth_used, off->on must occur
7038 * before making related changes, and on->off must occur afterwards
7040 if (runtime_enabled
&& !runtime_was_enabled
)
7041 cfs_bandwidth_usage_inc();
7042 raw_spin_lock_irq(&cfs_b
->lock
);
7043 cfs_b
->period
= ns_to_ktime(period
);
7044 cfs_b
->quota
= quota
;
7046 __refill_cfs_bandwidth_runtime(cfs_b
);
7048 /* Restart the period timer (if active) to handle new period expiry: */
7049 if (runtime_enabled
)
7050 start_cfs_bandwidth(cfs_b
);
7052 raw_spin_unlock_irq(&cfs_b
->lock
);
7054 for_each_online_cpu(i
) {
7055 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7056 struct rq
*rq
= cfs_rq
->rq
;
7059 rq_lock_irq(rq
, &rf
);
7060 cfs_rq
->runtime_enabled
= runtime_enabled
;
7061 cfs_rq
->runtime_remaining
= 0;
7063 if (cfs_rq
->throttled
)
7064 unthrottle_cfs_rq(cfs_rq
);
7065 rq_unlock_irq(rq
, &rf
);
7067 if (runtime_was_enabled
&& !runtime_enabled
)
7068 cfs_bandwidth_usage_dec();
7070 mutex_unlock(&cfs_constraints_mutex
);
7076 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7080 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7081 if (cfs_quota_us
< 0)
7082 quota
= RUNTIME_INF
;
7084 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7086 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7089 long tg_get_cfs_quota(struct task_group
*tg
)
7093 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7096 quota_us
= tg
->cfs_bandwidth
.quota
;
7097 do_div(quota_us
, NSEC_PER_USEC
);
7102 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7106 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7107 quota
= tg
->cfs_bandwidth
.quota
;
7109 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7112 long tg_get_cfs_period(struct task_group
*tg
)
7116 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7117 do_div(cfs_period_us
, NSEC_PER_USEC
);
7119 return cfs_period_us
;
7122 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7125 return tg_get_cfs_quota(css_tg(css
));
7128 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7129 struct cftype
*cftype
, s64 cfs_quota_us
)
7131 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7134 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7137 return tg_get_cfs_period(css_tg(css
));
7140 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7141 struct cftype
*cftype
, u64 cfs_period_us
)
7143 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7146 struct cfs_schedulable_data
{
7147 struct task_group
*tg
;
7152 * normalize group quota/period to be quota/max_period
7153 * note: units are usecs
7155 static u64
normalize_cfs_quota(struct task_group
*tg
,
7156 struct cfs_schedulable_data
*d
)
7164 period
= tg_get_cfs_period(tg
);
7165 quota
= tg_get_cfs_quota(tg
);
7168 /* note: these should typically be equivalent */
7169 if (quota
== RUNTIME_INF
|| quota
== -1)
7172 return to_ratio(period
, quota
);
7175 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7177 struct cfs_schedulable_data
*d
= data
;
7178 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7179 s64 quota
= 0, parent_quota
= -1;
7182 quota
= RUNTIME_INF
;
7184 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7186 quota
= normalize_cfs_quota(tg
, d
);
7187 parent_quota
= parent_b
->hierarchical_quota
;
7190 * Ensure max(child_quota) <= parent_quota, inherit when no
7193 if (quota
== RUNTIME_INF
)
7194 quota
= parent_quota
;
7195 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7198 cfs_b
->hierarchical_quota
= quota
;
7203 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7206 struct cfs_schedulable_data data
= {
7212 if (quota
!= RUNTIME_INF
) {
7213 do_div(data
.period
, NSEC_PER_USEC
);
7214 do_div(data
.quota
, NSEC_PER_USEC
);
7218 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7224 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7226 struct task_group
*tg
= css_tg(seq_css(sf
));
7227 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7229 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7230 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7231 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7235 #endif /* CONFIG_CFS_BANDWIDTH */
7236 #endif /* CONFIG_FAIR_GROUP_SCHED */
7238 #ifdef CONFIG_RT_GROUP_SCHED
7239 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7240 struct cftype
*cft
, s64 val
)
7242 return sched_group_set_rt_runtime(css_tg(css
), val
);
7245 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7248 return sched_group_rt_runtime(css_tg(css
));
7251 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7252 struct cftype
*cftype
, u64 rt_period_us
)
7254 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7257 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7260 return sched_group_rt_period(css_tg(css
));
7262 #endif /* CONFIG_RT_GROUP_SCHED */
7264 static struct cftype cpu_files
[] = {
7265 #ifdef CONFIG_FAIR_GROUP_SCHED
7268 .read_u64
= cpu_shares_read_u64
,
7269 .write_u64
= cpu_shares_write_u64
,
7272 #ifdef CONFIG_CFS_BANDWIDTH
7274 .name
= "cfs_quota_us",
7275 .read_s64
= cpu_cfs_quota_read_s64
,
7276 .write_s64
= cpu_cfs_quota_write_s64
,
7279 .name
= "cfs_period_us",
7280 .read_u64
= cpu_cfs_period_read_u64
,
7281 .write_u64
= cpu_cfs_period_write_u64
,
7285 .seq_show
= cpu_stats_show
,
7288 #ifdef CONFIG_RT_GROUP_SCHED
7290 .name
= "rt_runtime_us",
7291 .read_s64
= cpu_rt_runtime_read
,
7292 .write_s64
= cpu_rt_runtime_write
,
7295 .name
= "rt_period_us",
7296 .read_u64
= cpu_rt_period_read_uint
,
7297 .write_u64
= cpu_rt_period_write_uint
,
7303 struct cgroup_subsys cpu_cgrp_subsys
= {
7304 .css_alloc
= cpu_cgroup_css_alloc
,
7305 .css_online
= cpu_cgroup_css_online
,
7306 .css_released
= cpu_cgroup_css_released
,
7307 .css_free
= cpu_cgroup_css_free
,
7308 .fork
= cpu_cgroup_fork
,
7309 .can_attach
= cpu_cgroup_can_attach
,
7310 .attach
= cpu_cgroup_attach
,
7311 .legacy_cftypes
= cpu_files
,
7315 #endif /* CONFIG_CGROUP_SCHED */
7317 void dump_cpu_task(int cpu
)
7319 pr_info("Task dump for CPU %d:\n", cpu
);
7320 sched_show_task(cpu_curr(cpu
));
7324 * Nice levels are multiplicative, with a gentle 10% change for every
7325 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7326 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7327 * that remained on nice 0.
7329 * The "10% effect" is relative and cumulative: from _any_ nice level,
7330 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7331 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7332 * If a task goes up by ~10% and another task goes down by ~10% then
7333 * the relative distance between them is ~25%.)
7335 const int sched_prio_to_weight
[40] = {
7336 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7337 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7338 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7339 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7340 /* 0 */ 1024, 820, 655, 526, 423,
7341 /* 5 */ 335, 272, 215, 172, 137,
7342 /* 10 */ 110, 87, 70, 56, 45,
7343 /* 15 */ 36, 29, 23, 18, 15,
7347 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7349 * In cases where the weight does not change often, we can use the
7350 * precalculated inverse to speed up arithmetics by turning divisions
7351 * into multiplications:
7353 const u32 sched_prio_to_wmult
[40] = {
7354 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7355 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7356 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7357 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7358 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7359 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7360 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7361 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,