4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex
);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
96 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
98 void update_rq_clock(struct rq
*rq
)
102 lockdep_assert_held(&rq
->lock
);
104 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
107 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
111 update_rq_clock_task(rq
, delta
);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug
unsigned int sysctl_sched_features
=
122 #include "features.h"
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
131 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
134 * period over which we average the RT time consumption, measured
139 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
142 * period over which we measure -rt task cpu usage in us.
145 unsigned int sysctl_sched_rt_period
= 1000000;
147 __read_mostly
int scheduler_running
;
150 * part of the period that we allow rt tasks to run in us.
153 int sysctl_sched_rt_runtime
= 950000;
155 /* cpus with isolated domains */
156 cpumask_var_t cpu_isolated_map
;
159 * this_rq_lock - lock this runqueue and disable interrupts.
161 static struct rq
*this_rq_lock(void)
168 raw_spin_lock(&rq
->lock
);
173 #ifdef CONFIG_SCHED_HRTICK
175 * Use HR-timers to deliver accurate preemption points.
178 static void hrtick_clear(struct rq
*rq
)
180 if (hrtimer_active(&rq
->hrtick_timer
))
181 hrtimer_cancel(&rq
->hrtick_timer
);
185 * High-resolution timer tick.
186 * Runs from hardirq context with interrupts disabled.
188 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
190 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
192 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
194 raw_spin_lock(&rq
->lock
);
196 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
197 raw_spin_unlock(&rq
->lock
);
199 return HRTIMER_NORESTART
;
204 static void __hrtick_restart(struct rq
*rq
)
206 struct hrtimer
*timer
= &rq
->hrtick_timer
;
208 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
212 * called from hardirq (IPI) context
214 static void __hrtick_start(void *arg
)
218 raw_spin_lock(&rq
->lock
);
219 __hrtick_restart(rq
);
220 rq
->hrtick_csd_pending
= 0;
221 raw_spin_unlock(&rq
->lock
);
225 * Called to set the hrtick timer state.
227 * called with rq->lock held and irqs disabled
229 void hrtick_start(struct rq
*rq
, u64 delay
)
231 struct hrtimer
*timer
= &rq
->hrtick_timer
;
236 * Don't schedule slices shorter than 10000ns, that just
237 * doesn't make sense and can cause timer DoS.
239 delta
= max_t(s64
, delay
, 10000LL);
240 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
242 hrtimer_set_expires(timer
, time
);
244 if (rq
== this_rq()) {
245 __hrtick_restart(rq
);
246 } else if (!rq
->hrtick_csd_pending
) {
247 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
248 rq
->hrtick_csd_pending
= 1;
253 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
255 int cpu
= (int)(long)hcpu
;
258 case CPU_UP_CANCELED
:
259 case CPU_UP_CANCELED_FROZEN
:
260 case CPU_DOWN_PREPARE
:
261 case CPU_DOWN_PREPARE_FROZEN
:
263 case CPU_DEAD_FROZEN
:
264 hrtick_clear(cpu_rq(cpu
));
271 static __init
void init_hrtick(void)
273 hotcpu_notifier(hotplug_hrtick
, 0);
277 * Called to set the hrtick timer state.
279 * called with rq->lock held and irqs disabled
281 void hrtick_start(struct rq
*rq
, u64 delay
)
284 * Don't schedule slices shorter than 10000ns, that just
285 * doesn't make sense. Rely on vruntime for fairness.
287 delay
= max_t(u64
, delay
, 10000LL);
288 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
289 HRTIMER_MODE_REL_PINNED
);
292 static inline void init_hrtick(void)
295 #endif /* CONFIG_SMP */
297 static void init_rq_hrtick(struct rq
*rq
)
300 rq
->hrtick_csd_pending
= 0;
302 rq
->hrtick_csd
.flags
= 0;
303 rq
->hrtick_csd
.func
= __hrtick_start
;
304 rq
->hrtick_csd
.info
= rq
;
307 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
308 rq
->hrtick_timer
.function
= hrtick
;
310 #else /* CONFIG_SCHED_HRTICK */
311 static inline void hrtick_clear(struct rq
*rq
)
315 static inline void init_rq_hrtick(struct rq
*rq
)
319 static inline void init_hrtick(void)
322 #endif /* CONFIG_SCHED_HRTICK */
325 * cmpxchg based fetch_or, macro so it works for different integer types
327 #define fetch_or(ptr, mask) \
329 typeof(ptr) _ptr = (ptr); \
330 typeof(mask) _mask = (mask); \
331 typeof(*_ptr) _old, _val = *_ptr; \
334 _old = cmpxchg(_ptr, _val, _val | _mask); \
342 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
344 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
345 * this avoids any races wrt polling state changes and thereby avoids
348 static bool set_nr_and_not_polling(struct task_struct
*p
)
350 struct thread_info
*ti
= task_thread_info(p
);
351 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
355 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
357 * If this returns true, then the idle task promises to call
358 * sched_ttwu_pending() and reschedule soon.
360 static bool set_nr_if_polling(struct task_struct
*p
)
362 struct thread_info
*ti
= task_thread_info(p
);
363 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
366 if (!(val
& _TIF_POLLING_NRFLAG
))
368 if (val
& _TIF_NEED_RESCHED
)
370 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
379 static bool set_nr_and_not_polling(struct task_struct
*p
)
381 set_tsk_need_resched(p
);
386 static bool set_nr_if_polling(struct task_struct
*p
)
393 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
395 struct wake_q_node
*node
= &task
->wake_q
;
398 * Atomically grab the task, if ->wake_q is !nil already it means
399 * its already queued (either by us or someone else) and will get the
400 * wakeup due to that.
402 * This cmpxchg() implies a full barrier, which pairs with the write
403 * barrier implied by the wakeup in wake_up_list().
405 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
408 get_task_struct(task
);
411 * The head is context local, there can be no concurrency.
414 head
->lastp
= &node
->next
;
417 void wake_up_q(struct wake_q_head
*head
)
419 struct wake_q_node
*node
= head
->first
;
421 while (node
!= WAKE_Q_TAIL
) {
422 struct task_struct
*task
;
424 task
= container_of(node
, struct task_struct
, wake_q
);
426 /* task can safely be re-inserted now */
428 task
->wake_q
.next
= NULL
;
431 * wake_up_process() implies a wmb() to pair with the queueing
432 * in wake_q_add() so as not to miss wakeups.
434 wake_up_process(task
);
435 put_task_struct(task
);
440 * resched_curr - mark rq's current task 'to be rescheduled now'.
442 * On UP this means the setting of the need_resched flag, on SMP it
443 * might also involve a cross-CPU call to trigger the scheduler on
446 void resched_curr(struct rq
*rq
)
448 struct task_struct
*curr
= rq
->curr
;
451 lockdep_assert_held(&rq
->lock
);
453 if (test_tsk_need_resched(curr
))
458 if (cpu
== smp_processor_id()) {
459 set_tsk_need_resched(curr
);
460 set_preempt_need_resched();
464 if (set_nr_and_not_polling(curr
))
465 smp_send_reschedule(cpu
);
467 trace_sched_wake_idle_without_ipi(cpu
);
470 void resched_cpu(int cpu
)
472 struct rq
*rq
= cpu_rq(cpu
);
475 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
478 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
482 #ifdef CONFIG_NO_HZ_COMMON
484 * In the semi idle case, use the nearest busy cpu for migrating timers
485 * from an idle cpu. This is good for power-savings.
487 * We don't do similar optimization for completely idle system, as
488 * selecting an idle cpu will add more delays to the timers than intended
489 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
491 int get_nohz_timer_target(void)
493 int i
, cpu
= smp_processor_id();
494 struct sched_domain
*sd
;
496 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
500 for_each_domain(cpu
, sd
) {
501 for_each_cpu(i
, sched_domain_span(sd
)) {
502 if (!idle_cpu(i
) && is_housekeeping_cpu(cpu
)) {
509 if (!is_housekeeping_cpu(cpu
))
510 cpu
= housekeeping_any_cpu();
516 * When add_timer_on() enqueues a timer into the timer wheel of an
517 * idle CPU then this timer might expire before the next timer event
518 * which is scheduled to wake up that CPU. In case of a completely
519 * idle system the next event might even be infinite time into the
520 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
521 * leaves the inner idle loop so the newly added timer is taken into
522 * account when the CPU goes back to idle and evaluates the timer
523 * wheel for the next timer event.
525 static void wake_up_idle_cpu(int cpu
)
527 struct rq
*rq
= cpu_rq(cpu
);
529 if (cpu
== smp_processor_id())
532 if (set_nr_and_not_polling(rq
->idle
))
533 smp_send_reschedule(cpu
);
535 trace_sched_wake_idle_without_ipi(cpu
);
538 static bool wake_up_full_nohz_cpu(int cpu
)
541 * We just need the target to call irq_exit() and re-evaluate
542 * the next tick. The nohz full kick at least implies that.
543 * If needed we can still optimize that later with an
546 if (tick_nohz_full_cpu(cpu
)) {
547 if (cpu
!= smp_processor_id() ||
548 tick_nohz_tick_stopped())
549 tick_nohz_full_kick_cpu(cpu
);
556 void wake_up_nohz_cpu(int cpu
)
558 if (!wake_up_full_nohz_cpu(cpu
))
559 wake_up_idle_cpu(cpu
);
562 static inline bool got_nohz_idle_kick(void)
564 int cpu
= smp_processor_id();
566 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
569 if (idle_cpu(cpu
) && !need_resched())
573 * We can't run Idle Load Balance on this CPU for this time so we
574 * cancel it and clear NOHZ_BALANCE_KICK
576 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
580 #else /* CONFIG_NO_HZ_COMMON */
582 static inline bool got_nohz_idle_kick(void)
587 #endif /* CONFIG_NO_HZ_COMMON */
589 #ifdef CONFIG_NO_HZ_FULL
590 bool sched_can_stop_tick(struct rq
*rq
)
594 /* Deadline tasks, even if single, need the tick */
595 if (rq
->dl
.dl_nr_running
)
599 * FIFO realtime policy runs the highest priority task (after DEADLINE).
600 * Other runnable tasks are of a lower priority. The scheduler tick
603 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
608 * Round-robin realtime tasks time slice with other tasks at the same
611 if (rq
->rt
.rr_nr_running
) {
612 if (rq
->rt
.rr_nr_running
== 1)
618 /* Normal multitasking need periodic preemption checks */
619 if (rq
->cfs
.nr_running
> 1)
624 #endif /* CONFIG_NO_HZ_FULL */
626 void sched_avg_update(struct rq
*rq
)
628 s64 period
= sched_avg_period();
630 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq
->age_stamp
));
637 rq
->age_stamp
+= period
;
642 #endif /* CONFIG_SMP */
644 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
645 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
647 * Iterate task_group tree rooted at *from, calling @down when first entering a
648 * node and @up when leaving it for the final time.
650 * Caller must hold rcu_lock or sufficient equivalent.
652 int walk_tg_tree_from(struct task_group
*from
,
653 tg_visitor down
, tg_visitor up
, void *data
)
655 struct task_group
*parent
, *child
;
661 ret
= (*down
)(parent
, data
);
664 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
671 ret
= (*up
)(parent
, data
);
672 if (ret
|| parent
== from
)
676 parent
= parent
->parent
;
683 int tg_nop(struct task_group
*tg
, void *data
)
689 static void set_load_weight(struct task_struct
*p
)
691 int prio
= p
->static_prio
- MAX_RT_PRIO
;
692 struct load_weight
*load
= &p
->se
.load
;
695 * SCHED_IDLE tasks get minimal weight:
697 if (idle_policy(p
->policy
)) {
698 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
699 load
->inv_weight
= WMULT_IDLEPRIO
;
703 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
704 load
->inv_weight
= sched_prio_to_wmult
[prio
];
707 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
710 if (!(flags
& ENQUEUE_RESTORE
))
711 sched_info_queued(rq
, p
);
712 p
->sched_class
->enqueue_task(rq
, p
, flags
);
715 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
718 if (!(flags
& DEQUEUE_SAVE
))
719 sched_info_dequeued(rq
, p
);
720 p
->sched_class
->dequeue_task(rq
, p
, flags
);
723 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
725 if (task_contributes_to_load(p
))
726 rq
->nr_uninterruptible
--;
728 enqueue_task(rq
, p
, flags
);
731 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
733 if (task_contributes_to_load(p
))
734 rq
->nr_uninterruptible
++;
736 dequeue_task(rq
, p
, flags
);
739 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
742 * In theory, the compile should just see 0 here, and optimize out the call
743 * to sched_rt_avg_update. But I don't trust it...
745 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
746 s64 steal
= 0, irq_delta
= 0;
748 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
749 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
752 * Since irq_time is only updated on {soft,}irq_exit, we might run into
753 * this case when a previous update_rq_clock() happened inside a
756 * When this happens, we stop ->clock_task and only update the
757 * prev_irq_time stamp to account for the part that fit, so that a next
758 * update will consume the rest. This ensures ->clock_task is
761 * It does however cause some slight miss-attribution of {soft,}irq
762 * time, a more accurate solution would be to update the irq_time using
763 * the current rq->clock timestamp, except that would require using
766 if (irq_delta
> delta
)
769 rq
->prev_irq_time
+= irq_delta
;
772 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
773 if (static_key_false((¶virt_steal_rq_enabled
))) {
774 steal
= paravirt_steal_clock(cpu_of(rq
));
775 steal
-= rq
->prev_steal_time_rq
;
777 if (unlikely(steal
> delta
))
780 rq
->prev_steal_time_rq
+= steal
;
785 rq
->clock_task
+= delta
;
787 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
788 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
789 sched_rt_avg_update(rq
, irq_delta
+ steal
);
793 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
795 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
796 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
800 * Make it appear like a SCHED_FIFO task, its something
801 * userspace knows about and won't get confused about.
803 * Also, it will make PI more or less work without too
804 * much confusion -- but then, stop work should not
805 * rely on PI working anyway.
807 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
809 stop
->sched_class
= &stop_sched_class
;
812 cpu_rq(cpu
)->stop
= stop
;
816 * Reset it back to a normal scheduling class so that
817 * it can die in pieces.
819 old_stop
->sched_class
= &rt_sched_class
;
824 * __normal_prio - return the priority that is based on the static prio
826 static inline int __normal_prio(struct task_struct
*p
)
828 return p
->static_prio
;
832 * Calculate the expected normal priority: i.e. priority
833 * without taking RT-inheritance into account. Might be
834 * boosted by interactivity modifiers. Changes upon fork,
835 * setprio syscalls, and whenever the interactivity
836 * estimator recalculates.
838 static inline int normal_prio(struct task_struct
*p
)
842 if (task_has_dl_policy(p
))
843 prio
= MAX_DL_PRIO
-1;
844 else if (task_has_rt_policy(p
))
845 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
847 prio
= __normal_prio(p
);
852 * Calculate the current priority, i.e. the priority
853 * taken into account by the scheduler. This value might
854 * be boosted by RT tasks, or might be boosted by
855 * interactivity modifiers. Will be RT if the task got
856 * RT-boosted. If not then it returns p->normal_prio.
858 static int effective_prio(struct task_struct
*p
)
860 p
->normal_prio
= normal_prio(p
);
862 * If we are RT tasks or we were boosted to RT priority,
863 * keep the priority unchanged. Otherwise, update priority
864 * to the normal priority:
866 if (!rt_prio(p
->prio
))
867 return p
->normal_prio
;
872 * task_curr - is this task currently executing on a CPU?
873 * @p: the task in question.
875 * Return: 1 if the task is currently executing. 0 otherwise.
877 inline int task_curr(const struct task_struct
*p
)
879 return cpu_curr(task_cpu(p
)) == p
;
883 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
884 * use the balance_callback list if you want balancing.
886 * this means any call to check_class_changed() must be followed by a call to
887 * balance_callback().
889 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
890 const struct sched_class
*prev_class
,
893 if (prev_class
!= p
->sched_class
) {
894 if (prev_class
->switched_from
)
895 prev_class
->switched_from(rq
, p
);
897 p
->sched_class
->switched_to(rq
, p
);
898 } else if (oldprio
!= p
->prio
|| dl_task(p
))
899 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
902 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
904 const struct sched_class
*class;
906 if (p
->sched_class
== rq
->curr
->sched_class
) {
907 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
909 for_each_class(class) {
910 if (class == rq
->curr
->sched_class
)
912 if (class == p
->sched_class
) {
920 * A queue event has occurred, and we're going to schedule. In
921 * this case, we can save a useless back to back clock update.
923 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
924 rq_clock_skip_update(rq
, true);
929 * This is how migration works:
931 * 1) we invoke migration_cpu_stop() on the target CPU using
933 * 2) stopper starts to run (implicitly forcing the migrated thread
935 * 3) it checks whether the migrated task is still in the wrong runqueue.
936 * 4) if it's in the wrong runqueue then the migration thread removes
937 * it and puts it into the right queue.
938 * 5) stopper completes and stop_one_cpu() returns and the migration
943 * move_queued_task - move a queued task to new rq.
945 * Returns (locked) new rq. Old rq's lock is released.
947 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
949 lockdep_assert_held(&rq
->lock
);
951 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
952 dequeue_task(rq
, p
, 0);
953 set_task_cpu(p
, new_cpu
);
954 raw_spin_unlock(&rq
->lock
);
956 rq
= cpu_rq(new_cpu
);
958 raw_spin_lock(&rq
->lock
);
959 BUG_ON(task_cpu(p
) != new_cpu
);
960 enqueue_task(rq
, p
, 0);
961 p
->on_rq
= TASK_ON_RQ_QUEUED
;
962 check_preempt_curr(rq
, p
, 0);
967 struct migration_arg
{
968 struct task_struct
*task
;
973 * Move (not current) task off this cpu, onto dest cpu. We're doing
974 * this because either it can't run here any more (set_cpus_allowed()
975 * away from this CPU, or CPU going down), or because we're
976 * attempting to rebalance this task on exec (sched_exec).
978 * So we race with normal scheduler movements, but that's OK, as long
979 * as the task is no longer on this CPU.
981 static struct rq
*__migrate_task(struct rq
*rq
, 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
, tsk_cpus_allowed(p
)))
990 rq
= move_queued_task(rq
, p
, dest_cpu
);
996 * migration_cpu_stop - this will be executed by a highprio stopper thread
997 * and performs thread migration by bumping thread off CPU then
998 * 'pushing' onto another runqueue.
1000 static int migration_cpu_stop(void *data
)
1002 struct migration_arg
*arg
= data
;
1003 struct task_struct
*p
= arg
->task
;
1004 struct rq
*rq
= this_rq();
1007 * The original target cpu might have gone down and we might
1008 * be on another cpu but it doesn't matter.
1010 local_irq_disable();
1012 * We need to explicitly wake pending tasks before running
1013 * __migrate_task() such that we will not miss enforcing cpus_allowed
1014 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1016 sched_ttwu_pending();
1018 raw_spin_lock(&p
->pi_lock
);
1019 raw_spin_lock(&rq
->lock
);
1021 * If task_rq(p) != rq, it cannot be migrated here, because we're
1022 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1023 * we're holding p->pi_lock.
1025 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1026 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1027 raw_spin_unlock(&rq
->lock
);
1028 raw_spin_unlock(&p
->pi_lock
);
1035 * sched_class::set_cpus_allowed must do the below, but is not required to
1036 * actually call this function.
1038 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1040 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1041 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1044 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1046 struct rq
*rq
= task_rq(p
);
1047 bool queued
, running
;
1049 lockdep_assert_held(&p
->pi_lock
);
1051 queued
= task_on_rq_queued(p
);
1052 running
= task_current(rq
, p
);
1056 * Because __kthread_bind() calls this on blocked tasks without
1059 lockdep_assert_held(&rq
->lock
);
1060 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1063 put_prev_task(rq
, p
);
1065 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1068 p
->sched_class
->set_curr_task(rq
);
1070 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1074 * Change a given task's CPU affinity. Migrate the thread to a
1075 * proper CPU and schedule it away if the CPU it's executing on
1076 * is removed from the allowed bitmask.
1078 * NOTE: the caller must have a valid reference to the task, the
1079 * task must not exit() & deallocate itself prematurely. The
1080 * call is not atomic; no spinlocks may be held.
1082 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1083 const struct cpumask
*new_mask
, bool check
)
1085 unsigned long flags
;
1087 unsigned int dest_cpu
;
1090 rq
= task_rq_lock(p
, &flags
);
1093 * Must re-check here, to close a race against __kthread_bind(),
1094 * sched_setaffinity() is not guaranteed to observe the flag.
1096 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1101 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1104 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1109 do_set_cpus_allowed(p
, new_mask
);
1111 /* Can the task run on the task's current CPU? If so, we're done */
1112 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1115 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1116 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1117 struct migration_arg arg
= { p
, dest_cpu
};
1118 /* Need help from migration thread: drop lock and wait. */
1119 task_rq_unlock(rq
, p
, &flags
);
1120 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1121 tlb_migrate_finish(p
->mm
);
1123 } else if (task_on_rq_queued(p
)) {
1125 * OK, since we're going to drop the lock immediately
1126 * afterwards anyway.
1128 lockdep_unpin_lock(&rq
->lock
);
1129 rq
= move_queued_task(rq
, p
, dest_cpu
);
1130 lockdep_pin_lock(&rq
->lock
);
1133 task_rq_unlock(rq
, p
, &flags
);
1138 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1140 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1142 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1144 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1146 #ifdef CONFIG_SCHED_DEBUG
1148 * We should never call set_task_cpu() on a blocked task,
1149 * ttwu() will sort out the placement.
1151 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1155 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1156 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1157 * time relying on p->on_rq.
1159 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1160 p
->sched_class
== &fair_sched_class
&&
1161 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1163 #ifdef CONFIG_LOCKDEP
1165 * The caller should hold either p->pi_lock or rq->lock, when changing
1166 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1168 * sched_move_task() holds both and thus holding either pins the cgroup,
1171 * Furthermore, all task_rq users should acquire both locks, see
1174 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1175 lockdep_is_held(&task_rq(p
)->lock
)));
1179 trace_sched_migrate_task(p
, new_cpu
);
1181 if (task_cpu(p
) != new_cpu
) {
1182 if (p
->sched_class
->migrate_task_rq
)
1183 p
->sched_class
->migrate_task_rq(p
);
1184 p
->se
.nr_migrations
++;
1185 perf_event_task_migrate(p
);
1188 __set_task_cpu(p
, new_cpu
);
1191 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1193 if (task_on_rq_queued(p
)) {
1194 struct rq
*src_rq
, *dst_rq
;
1196 src_rq
= task_rq(p
);
1197 dst_rq
= cpu_rq(cpu
);
1199 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1200 deactivate_task(src_rq
, p
, 0);
1201 set_task_cpu(p
, cpu
);
1202 activate_task(dst_rq
, p
, 0);
1203 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1204 check_preempt_curr(dst_rq
, p
, 0);
1207 * Task isn't running anymore; make it appear like we migrated
1208 * it before it went to sleep. This means on wakeup we make the
1209 * previous cpu our targer instead of where it really is.
1215 struct migration_swap_arg
{
1216 struct task_struct
*src_task
, *dst_task
;
1217 int src_cpu
, dst_cpu
;
1220 static int migrate_swap_stop(void *data
)
1222 struct migration_swap_arg
*arg
= data
;
1223 struct rq
*src_rq
, *dst_rq
;
1226 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1229 src_rq
= cpu_rq(arg
->src_cpu
);
1230 dst_rq
= cpu_rq(arg
->dst_cpu
);
1232 double_raw_lock(&arg
->src_task
->pi_lock
,
1233 &arg
->dst_task
->pi_lock
);
1234 double_rq_lock(src_rq
, dst_rq
);
1236 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1239 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1242 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1245 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1248 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1249 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1254 double_rq_unlock(src_rq
, dst_rq
);
1255 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1256 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1262 * Cross migrate two tasks
1264 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1266 struct migration_swap_arg arg
;
1269 arg
= (struct migration_swap_arg
){
1271 .src_cpu
= task_cpu(cur
),
1273 .dst_cpu
= task_cpu(p
),
1276 if (arg
.src_cpu
== arg
.dst_cpu
)
1280 * These three tests are all lockless; this is OK since all of them
1281 * will be re-checked with proper locks held further down the line.
1283 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1286 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1289 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1292 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1293 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1300 * wait_task_inactive - wait for a thread to unschedule.
1302 * If @match_state is nonzero, it's the @p->state value just checked and
1303 * not expected to change. If it changes, i.e. @p might have woken up,
1304 * then return zero. When we succeed in waiting for @p to be off its CPU,
1305 * we return a positive number (its total switch count). If a second call
1306 * a short while later returns the same number, the caller can be sure that
1307 * @p has remained unscheduled the whole time.
1309 * The caller must ensure that the task *will* unschedule sometime soon,
1310 * else this function might spin for a *long* time. This function can't
1311 * be called with interrupts off, or it may introduce deadlock with
1312 * smp_call_function() if an IPI is sent by the same process we are
1313 * waiting to become inactive.
1315 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1317 unsigned long flags
;
1318 int running
, queued
;
1324 * We do the initial early heuristics without holding
1325 * any task-queue locks at all. We'll only try to get
1326 * the runqueue lock when things look like they will
1332 * If the task is actively running on another CPU
1333 * still, just relax and busy-wait without holding
1336 * NOTE! Since we don't hold any locks, it's not
1337 * even sure that "rq" stays as the right runqueue!
1338 * But we don't care, since "task_running()" will
1339 * return false if the runqueue has changed and p
1340 * is actually now running somewhere else!
1342 while (task_running(rq
, p
)) {
1343 if (match_state
&& unlikely(p
->state
!= match_state
))
1349 * Ok, time to look more closely! We need the rq
1350 * lock now, to be *sure*. If we're wrong, we'll
1351 * just go back and repeat.
1353 rq
= task_rq_lock(p
, &flags
);
1354 trace_sched_wait_task(p
);
1355 running
= task_running(rq
, p
);
1356 queued
= task_on_rq_queued(p
);
1358 if (!match_state
|| p
->state
== match_state
)
1359 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1360 task_rq_unlock(rq
, p
, &flags
);
1363 * If it changed from the expected state, bail out now.
1365 if (unlikely(!ncsw
))
1369 * Was it really running after all now that we
1370 * checked with the proper locks actually held?
1372 * Oops. Go back and try again..
1374 if (unlikely(running
)) {
1380 * It's not enough that it's not actively running,
1381 * it must be off the runqueue _entirely_, and not
1384 * So if it was still runnable (but just not actively
1385 * running right now), it's preempted, and we should
1386 * yield - it could be a while.
1388 if (unlikely(queued
)) {
1389 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1391 set_current_state(TASK_UNINTERRUPTIBLE
);
1392 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1397 * Ahh, all good. It wasn't running, and it wasn't
1398 * runnable, which means that it will never become
1399 * running in the future either. We're all done!
1408 * kick_process - kick a running thread to enter/exit the kernel
1409 * @p: the to-be-kicked thread
1411 * Cause a process which is running on another CPU to enter
1412 * kernel-mode, without any delay. (to get signals handled.)
1414 * NOTE: this function doesn't have to take the runqueue lock,
1415 * because all it wants to ensure is that the remote task enters
1416 * the kernel. If the IPI races and the task has been migrated
1417 * to another CPU then no harm is done and the purpose has been
1420 void kick_process(struct task_struct
*p
)
1426 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1427 smp_send_reschedule(cpu
);
1430 EXPORT_SYMBOL_GPL(kick_process
);
1433 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1435 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1437 int nid
= cpu_to_node(cpu
);
1438 const struct cpumask
*nodemask
= NULL
;
1439 enum { cpuset
, possible
, fail
} state
= cpuset
;
1443 * If the node that the cpu is on has been offlined, cpu_to_node()
1444 * will return -1. There is no cpu on the node, and we should
1445 * select the cpu on the other node.
1448 nodemask
= cpumask_of_node(nid
);
1450 /* Look for allowed, online CPU in same node. */
1451 for_each_cpu(dest_cpu
, nodemask
) {
1452 if (!cpu_online(dest_cpu
))
1454 if (!cpu_active(dest_cpu
))
1456 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1462 /* Any allowed, online CPU? */
1463 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1464 if (!cpu_online(dest_cpu
))
1466 if (!cpu_active(dest_cpu
))
1471 /* No more Mr. Nice Guy. */
1474 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1475 cpuset_cpus_allowed_fallback(p
);
1481 do_set_cpus_allowed(p
, cpu_possible_mask
);
1492 if (state
!= cpuset
) {
1494 * Don't tell them about moving exiting tasks or
1495 * kernel threads (both mm NULL), since they never
1498 if (p
->mm
&& printk_ratelimit()) {
1499 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1500 task_pid_nr(p
), p
->comm
, cpu
);
1508 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1511 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1513 lockdep_assert_held(&p
->pi_lock
);
1515 if (p
->nr_cpus_allowed
> 1)
1516 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1519 * In order not to call set_task_cpu() on a blocking task we need
1520 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1523 * Since this is common to all placement strategies, this lives here.
1525 * [ this allows ->select_task() to simply return task_cpu(p) and
1526 * not worry about this generic constraint ]
1528 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1530 cpu
= select_fallback_rq(task_cpu(p
), p
);
1535 static void update_avg(u64
*avg
, u64 sample
)
1537 s64 diff
= sample
- *avg
;
1543 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1544 const struct cpumask
*new_mask
, bool check
)
1546 return set_cpus_allowed_ptr(p
, new_mask
);
1549 #endif /* CONFIG_SMP */
1552 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1554 #ifdef CONFIG_SCHEDSTATS
1555 struct rq
*rq
= this_rq();
1558 int this_cpu
= smp_processor_id();
1560 if (cpu
== this_cpu
) {
1561 schedstat_inc(rq
, ttwu_local
);
1562 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1564 struct sched_domain
*sd
;
1566 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1568 for_each_domain(this_cpu
, sd
) {
1569 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1570 schedstat_inc(sd
, ttwu_wake_remote
);
1577 if (wake_flags
& WF_MIGRATED
)
1578 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1580 #endif /* CONFIG_SMP */
1582 schedstat_inc(rq
, ttwu_count
);
1583 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1585 if (wake_flags
& WF_SYNC
)
1586 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1588 #endif /* CONFIG_SCHEDSTATS */
1591 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1593 activate_task(rq
, p
, en_flags
);
1594 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1596 /* if a worker is waking up, notify workqueue */
1597 if (p
->flags
& PF_WQ_WORKER
)
1598 wq_worker_waking_up(p
, cpu_of(rq
));
1602 * Mark the task runnable and perform wakeup-preemption.
1605 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1607 check_preempt_curr(rq
, p
, wake_flags
);
1608 p
->state
= TASK_RUNNING
;
1609 trace_sched_wakeup(p
);
1612 if (p
->sched_class
->task_woken
) {
1614 * Our task @p is fully woken up and running; so its safe to
1615 * drop the rq->lock, hereafter rq is only used for statistics.
1617 lockdep_unpin_lock(&rq
->lock
);
1618 p
->sched_class
->task_woken(rq
, p
);
1619 lockdep_pin_lock(&rq
->lock
);
1622 if (rq
->idle_stamp
) {
1623 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1624 u64 max
= 2*rq
->max_idle_balance_cost
;
1626 update_avg(&rq
->avg_idle
, delta
);
1628 if (rq
->avg_idle
> max
)
1637 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1639 lockdep_assert_held(&rq
->lock
);
1642 if (p
->sched_contributes_to_load
)
1643 rq
->nr_uninterruptible
--;
1646 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1647 ttwu_do_wakeup(rq
, p
, wake_flags
);
1651 * Called in case the task @p isn't fully descheduled from its runqueue,
1652 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1653 * since all we need to do is flip p->state to TASK_RUNNING, since
1654 * the task is still ->on_rq.
1656 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1661 rq
= __task_rq_lock(p
);
1662 if (task_on_rq_queued(p
)) {
1663 /* check_preempt_curr() may use rq clock */
1664 update_rq_clock(rq
);
1665 ttwu_do_wakeup(rq
, p
, wake_flags
);
1668 __task_rq_unlock(rq
);
1674 void sched_ttwu_pending(void)
1676 struct rq
*rq
= this_rq();
1677 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1678 struct task_struct
*p
;
1679 unsigned long flags
;
1684 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1685 lockdep_pin_lock(&rq
->lock
);
1688 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1689 llist
= llist_next(llist
);
1690 ttwu_do_activate(rq
, p
, 0);
1693 lockdep_unpin_lock(&rq
->lock
);
1694 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1697 void scheduler_ipi(void)
1700 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1701 * TIF_NEED_RESCHED remotely (for the first time) will also send
1704 preempt_fold_need_resched();
1706 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1710 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1711 * traditionally all their work was done from the interrupt return
1712 * path. Now that we actually do some work, we need to make sure
1715 * Some archs already do call them, luckily irq_enter/exit nest
1718 * Arguably we should visit all archs and update all handlers,
1719 * however a fair share of IPIs are still resched only so this would
1720 * somewhat pessimize the simple resched case.
1723 sched_ttwu_pending();
1726 * Check if someone kicked us for doing the nohz idle load balance.
1728 if (unlikely(got_nohz_idle_kick())) {
1729 this_rq()->idle_balance
= 1;
1730 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1735 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1737 struct rq
*rq
= cpu_rq(cpu
);
1739 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1740 if (!set_nr_if_polling(rq
->idle
))
1741 smp_send_reschedule(cpu
);
1743 trace_sched_wake_idle_without_ipi(cpu
);
1747 void wake_up_if_idle(int cpu
)
1749 struct rq
*rq
= cpu_rq(cpu
);
1750 unsigned long flags
;
1754 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1757 if (set_nr_if_polling(rq
->idle
)) {
1758 trace_sched_wake_idle_without_ipi(cpu
);
1760 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1761 if (is_idle_task(rq
->curr
))
1762 smp_send_reschedule(cpu
);
1763 /* Else cpu is not in idle, do nothing here */
1764 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1771 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1773 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1775 #endif /* CONFIG_SMP */
1777 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1779 struct rq
*rq
= cpu_rq(cpu
);
1781 #if defined(CONFIG_SMP)
1782 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1783 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1784 ttwu_queue_remote(p
, cpu
);
1789 raw_spin_lock(&rq
->lock
);
1790 lockdep_pin_lock(&rq
->lock
);
1791 ttwu_do_activate(rq
, p
, 0);
1792 lockdep_unpin_lock(&rq
->lock
);
1793 raw_spin_unlock(&rq
->lock
);
1797 * Notes on Program-Order guarantees on SMP systems.
1801 * The basic program-order guarantee on SMP systems is that when a task [t]
1802 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1803 * execution on its new cpu [c1].
1805 * For migration (of runnable tasks) this is provided by the following means:
1807 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1808 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1809 * rq(c1)->lock (if not at the same time, then in that order).
1810 * C) LOCK of the rq(c1)->lock scheduling in task
1812 * Transitivity guarantees that B happens after A and C after B.
1813 * Note: we only require RCpc transitivity.
1814 * Note: the cpu doing B need not be c0 or c1
1823 * UNLOCK rq(0)->lock
1825 * LOCK rq(0)->lock // orders against CPU0
1827 * UNLOCK rq(0)->lock
1831 * UNLOCK rq(1)->lock
1833 * LOCK rq(1)->lock // orders against CPU2
1836 * UNLOCK rq(1)->lock
1839 * BLOCKING -- aka. SLEEP + WAKEUP
1841 * For blocking we (obviously) need to provide the same guarantee as for
1842 * migration. However the means are completely different as there is no lock
1843 * chain to provide order. Instead we do:
1845 * 1) smp_store_release(X->on_cpu, 0)
1846 * 2) smp_cond_acquire(!X->on_cpu)
1850 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1852 * LOCK rq(0)->lock LOCK X->pi_lock
1855 * smp_store_release(X->on_cpu, 0);
1857 * smp_cond_acquire(!X->on_cpu);
1863 * X->state = RUNNING
1864 * UNLOCK rq(2)->lock
1866 * LOCK rq(2)->lock // orders against CPU1
1869 * UNLOCK rq(2)->lock
1872 * UNLOCK rq(0)->lock
1875 * However; for wakeups there is a second guarantee we must provide, namely we
1876 * must observe the state that lead to our wakeup. That is, not only must our
1877 * task observe its own prior state, it must also observe the stores prior to
1880 * This means that any means of doing remote wakeups must order the CPU doing
1881 * the wakeup against the CPU the task is going to end up running on. This,
1882 * however, is already required for the regular Program-Order guarantee above,
1883 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1888 * try_to_wake_up - wake up a thread
1889 * @p: the thread to be awakened
1890 * @state: the mask of task states that can be woken
1891 * @wake_flags: wake modifier flags (WF_*)
1893 * Put it on the run-queue if it's not already there. The "current"
1894 * thread is always on the run-queue (except when the actual
1895 * re-schedule is in progress), and as such you're allowed to do
1896 * the simpler "current->state = TASK_RUNNING" to mark yourself
1897 * runnable without the overhead of this.
1899 * Return: %true if @p was woken up, %false if it was already running.
1900 * or @state didn't match @p's state.
1903 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1905 unsigned long flags
;
1906 int cpu
, success
= 0;
1909 * If we are going to wake up a thread waiting for CONDITION we
1910 * need to ensure that CONDITION=1 done by the caller can not be
1911 * reordered with p->state check below. This pairs with mb() in
1912 * set_current_state() the waiting thread does.
1914 smp_mb__before_spinlock();
1915 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1916 if (!(p
->state
& state
))
1919 trace_sched_waking(p
);
1921 success
= 1; /* we're going to change ->state */
1924 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1929 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1930 * possible to, falsely, observe p->on_cpu == 0.
1932 * One must be running (->on_cpu == 1) in order to remove oneself
1933 * from the runqueue.
1935 * [S] ->on_cpu = 1; [L] ->on_rq
1939 * [S] ->on_rq = 0; [L] ->on_cpu
1941 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1942 * from the consecutive calls to schedule(); the first switching to our
1943 * task, the second putting it to sleep.
1948 * If the owning (remote) cpu is still in the middle of schedule() with
1949 * this task as prev, wait until its done referencing the task.
1951 * Pairs with the smp_store_release() in finish_lock_switch().
1953 * This ensures that tasks getting woken will be fully ordered against
1954 * their previous state and preserve Program Order.
1956 smp_cond_acquire(!p
->on_cpu
);
1958 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1959 p
->state
= TASK_WAKING
;
1961 if (p
->sched_class
->task_waking
)
1962 p
->sched_class
->task_waking(p
);
1964 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1965 if (task_cpu(p
) != cpu
) {
1966 wake_flags
|= WF_MIGRATED
;
1967 set_task_cpu(p
, cpu
);
1969 #endif /* CONFIG_SMP */
1973 if (schedstat_enabled())
1974 ttwu_stat(p
, cpu
, wake_flags
);
1976 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1982 * try_to_wake_up_local - try to wake up a local task with rq lock held
1983 * @p: the thread to be awakened
1985 * Put @p on the run-queue if it's not already there. The caller must
1986 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1989 static void try_to_wake_up_local(struct task_struct
*p
)
1991 struct rq
*rq
= task_rq(p
);
1993 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1994 WARN_ON_ONCE(p
== current
))
1997 lockdep_assert_held(&rq
->lock
);
1999 if (!raw_spin_trylock(&p
->pi_lock
)) {
2001 * This is OK, because current is on_cpu, which avoids it being
2002 * picked for load-balance and preemption/IRQs are still
2003 * disabled avoiding further scheduler activity on it and we've
2004 * not yet picked a replacement task.
2006 lockdep_unpin_lock(&rq
->lock
);
2007 raw_spin_unlock(&rq
->lock
);
2008 raw_spin_lock(&p
->pi_lock
);
2009 raw_spin_lock(&rq
->lock
);
2010 lockdep_pin_lock(&rq
->lock
);
2013 if (!(p
->state
& TASK_NORMAL
))
2016 trace_sched_waking(p
);
2018 if (!task_on_rq_queued(p
))
2019 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2021 ttwu_do_wakeup(rq
, p
, 0);
2022 if (schedstat_enabled())
2023 ttwu_stat(p
, smp_processor_id(), 0);
2025 raw_spin_unlock(&p
->pi_lock
);
2029 * wake_up_process - Wake up a specific process
2030 * @p: The process to be woken up.
2032 * Attempt to wake up the nominated process and move it to the set of runnable
2035 * Return: 1 if the process was woken up, 0 if it was already running.
2037 * It may be assumed that this function implies a write memory barrier before
2038 * changing the task state if and only if any tasks are woken up.
2040 int wake_up_process(struct task_struct
*p
)
2042 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2044 EXPORT_SYMBOL(wake_up_process
);
2046 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2048 return try_to_wake_up(p
, state
, 0);
2052 * This function clears the sched_dl_entity static params.
2054 void __dl_clear_params(struct task_struct
*p
)
2056 struct sched_dl_entity
*dl_se
= &p
->dl
;
2058 dl_se
->dl_runtime
= 0;
2059 dl_se
->dl_deadline
= 0;
2060 dl_se
->dl_period
= 0;
2064 dl_se
->dl_throttled
= 0;
2065 dl_se
->dl_yielded
= 0;
2069 * Perform scheduler related setup for a newly forked process p.
2070 * p is forked by current.
2072 * __sched_fork() is basic setup used by init_idle() too:
2074 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2079 p
->se
.exec_start
= 0;
2080 p
->se
.sum_exec_runtime
= 0;
2081 p
->se
.prev_sum_exec_runtime
= 0;
2082 p
->se
.nr_migrations
= 0;
2084 INIT_LIST_HEAD(&p
->se
.group_node
);
2086 #ifdef CONFIG_FAIR_GROUP_SCHED
2087 p
->se
.cfs_rq
= NULL
;
2090 #ifdef CONFIG_SCHEDSTATS
2091 /* Even if schedstat is disabled, there should not be garbage */
2092 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2095 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2096 init_dl_task_timer(&p
->dl
);
2097 __dl_clear_params(p
);
2099 INIT_LIST_HEAD(&p
->rt
.run_list
);
2101 p
->rt
.time_slice
= sched_rr_timeslice
;
2105 #ifdef CONFIG_PREEMPT_NOTIFIERS
2106 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2109 #ifdef CONFIG_NUMA_BALANCING
2110 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2111 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2112 p
->mm
->numa_scan_seq
= 0;
2115 if (clone_flags
& CLONE_VM
)
2116 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2118 p
->numa_preferred_nid
= -1;
2120 p
->node_stamp
= 0ULL;
2121 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2122 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2123 p
->numa_work
.next
= &p
->numa_work
;
2124 p
->numa_faults
= NULL
;
2125 p
->last_task_numa_placement
= 0;
2126 p
->last_sum_exec_runtime
= 0;
2128 p
->numa_group
= NULL
;
2129 #endif /* CONFIG_NUMA_BALANCING */
2132 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2134 #ifdef CONFIG_NUMA_BALANCING
2136 void set_numabalancing_state(bool enabled
)
2139 static_branch_enable(&sched_numa_balancing
);
2141 static_branch_disable(&sched_numa_balancing
);
2144 #ifdef CONFIG_PROC_SYSCTL
2145 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2146 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2150 int state
= static_branch_likely(&sched_numa_balancing
);
2152 if (write
&& !capable(CAP_SYS_ADMIN
))
2157 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2161 set_numabalancing_state(state
);
2167 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2169 #ifdef CONFIG_SCHEDSTATS
2170 static void set_schedstats(bool enabled
)
2173 static_branch_enable(&sched_schedstats
);
2175 static_branch_disable(&sched_schedstats
);
2178 void force_schedstat_enabled(void)
2180 if (!schedstat_enabled()) {
2181 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2182 static_branch_enable(&sched_schedstats
);
2186 static int __init
setup_schedstats(char *str
)
2192 if (!strcmp(str
, "enable")) {
2193 set_schedstats(true);
2195 } else if (!strcmp(str
, "disable")) {
2196 set_schedstats(false);
2201 pr_warn("Unable to parse schedstats=\n");
2205 __setup("schedstats=", setup_schedstats
);
2207 #ifdef CONFIG_PROC_SYSCTL
2208 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2209 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2213 int state
= static_branch_likely(&sched_schedstats
);
2215 if (write
&& !capable(CAP_SYS_ADMIN
))
2220 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2224 set_schedstats(state
);
2231 * fork()/clone()-time setup:
2233 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2235 unsigned long flags
;
2236 int cpu
= get_cpu();
2238 __sched_fork(clone_flags
, p
);
2240 * We mark the process as running here. This guarantees that
2241 * nobody will actually run it, and a signal or other external
2242 * event cannot wake it up and insert it on the runqueue either.
2244 p
->state
= TASK_RUNNING
;
2247 * Make sure we do not leak PI boosting priority to the child.
2249 p
->prio
= current
->normal_prio
;
2252 * Revert to default priority/policy on fork if requested.
2254 if (unlikely(p
->sched_reset_on_fork
)) {
2255 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2256 p
->policy
= SCHED_NORMAL
;
2257 p
->static_prio
= NICE_TO_PRIO(0);
2259 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2260 p
->static_prio
= NICE_TO_PRIO(0);
2262 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2266 * We don't need the reset flag anymore after the fork. It has
2267 * fulfilled its duty:
2269 p
->sched_reset_on_fork
= 0;
2272 if (dl_prio(p
->prio
)) {
2275 } else if (rt_prio(p
->prio
)) {
2276 p
->sched_class
= &rt_sched_class
;
2278 p
->sched_class
= &fair_sched_class
;
2281 if (p
->sched_class
->task_fork
)
2282 p
->sched_class
->task_fork(p
);
2285 * The child is not yet in the pid-hash so no cgroup attach races,
2286 * and the cgroup is pinned to this child due to cgroup_fork()
2287 * is ran before sched_fork().
2289 * Silence PROVE_RCU.
2291 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2292 set_task_cpu(p
, cpu
);
2293 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2295 #ifdef CONFIG_SCHED_INFO
2296 if (likely(sched_info_on()))
2297 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2299 #if defined(CONFIG_SMP)
2302 init_task_preempt_count(p
);
2304 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2305 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2312 unsigned long to_ratio(u64 period
, u64 runtime
)
2314 if (runtime
== RUNTIME_INF
)
2318 * Doing this here saves a lot of checks in all
2319 * the calling paths, and returning zero seems
2320 * safe for them anyway.
2325 return div64_u64(runtime
<< 20, period
);
2329 inline struct dl_bw
*dl_bw_of(int i
)
2331 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2332 "sched RCU must be held");
2333 return &cpu_rq(i
)->rd
->dl_bw
;
2336 static inline int dl_bw_cpus(int i
)
2338 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2341 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2342 "sched RCU must be held");
2343 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2349 inline struct dl_bw
*dl_bw_of(int i
)
2351 return &cpu_rq(i
)->dl
.dl_bw
;
2354 static inline int dl_bw_cpus(int i
)
2361 * We must be sure that accepting a new task (or allowing changing the
2362 * parameters of an existing one) is consistent with the bandwidth
2363 * constraints. If yes, this function also accordingly updates the currently
2364 * allocated bandwidth to reflect the new situation.
2366 * This function is called while holding p's rq->lock.
2368 * XXX we should delay bw change until the task's 0-lag point, see
2371 static int dl_overflow(struct task_struct
*p
, int policy
,
2372 const struct sched_attr
*attr
)
2375 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2376 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2377 u64 runtime
= attr
->sched_runtime
;
2378 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2381 if (new_bw
== p
->dl
.dl_bw
)
2385 * Either if a task, enters, leave, or stays -deadline but changes
2386 * its parameters, we may need to update accordingly the total
2387 * allocated bandwidth of the container.
2389 raw_spin_lock(&dl_b
->lock
);
2390 cpus
= dl_bw_cpus(task_cpu(p
));
2391 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2392 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2393 __dl_add(dl_b
, new_bw
);
2395 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2396 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2397 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2398 __dl_add(dl_b
, new_bw
);
2400 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2401 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2404 raw_spin_unlock(&dl_b
->lock
);
2409 extern void init_dl_bw(struct dl_bw
*dl_b
);
2412 * wake_up_new_task - wake up a newly created task for the first time.
2414 * This function will do some initial scheduler statistics housekeeping
2415 * that must be done for every newly created context, then puts the task
2416 * on the runqueue and wakes it.
2418 void wake_up_new_task(struct task_struct
*p
)
2420 unsigned long flags
;
2423 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2424 /* Initialize new task's runnable average */
2425 init_entity_runnable_average(&p
->se
);
2428 * Fork balancing, do it here and not earlier because:
2429 * - cpus_allowed can change in the fork path
2430 * - any previously selected cpu might disappear through hotplug
2432 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2434 /* Post initialize new task's util average when its cfs_rq is set */
2435 post_init_entity_util_avg(&p
->se
);
2437 rq
= __task_rq_lock(p
);
2438 activate_task(rq
, p
, 0);
2439 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2440 trace_sched_wakeup_new(p
);
2441 check_preempt_curr(rq
, p
, WF_FORK
);
2443 if (p
->sched_class
->task_woken
) {
2445 * Nothing relies on rq->lock after this, so its fine to
2448 lockdep_unpin_lock(&rq
->lock
);
2449 p
->sched_class
->task_woken(rq
, p
);
2450 lockdep_pin_lock(&rq
->lock
);
2453 task_rq_unlock(rq
, p
, &flags
);
2456 #ifdef CONFIG_PREEMPT_NOTIFIERS
2458 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2460 void preempt_notifier_inc(void)
2462 static_key_slow_inc(&preempt_notifier_key
);
2464 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2466 void preempt_notifier_dec(void)
2468 static_key_slow_dec(&preempt_notifier_key
);
2470 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2473 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2474 * @notifier: notifier struct to register
2476 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2478 if (!static_key_false(&preempt_notifier_key
))
2479 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2481 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2483 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2486 * preempt_notifier_unregister - no longer interested in preemption notifications
2487 * @notifier: notifier struct to unregister
2489 * This is *not* safe to call from within a preemption notifier.
2491 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2493 hlist_del(¬ifier
->link
);
2495 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2497 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2499 struct preempt_notifier
*notifier
;
2501 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2502 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2505 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2507 if (static_key_false(&preempt_notifier_key
))
2508 __fire_sched_in_preempt_notifiers(curr
);
2512 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2513 struct task_struct
*next
)
2515 struct preempt_notifier
*notifier
;
2517 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2518 notifier
->ops
->sched_out(notifier
, next
);
2521 static __always_inline
void
2522 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2523 struct task_struct
*next
)
2525 if (static_key_false(&preempt_notifier_key
))
2526 __fire_sched_out_preempt_notifiers(curr
, next
);
2529 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2531 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2536 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2537 struct task_struct
*next
)
2541 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2544 * prepare_task_switch - prepare to switch tasks
2545 * @rq: the runqueue preparing to switch
2546 * @prev: the current task that is being switched out
2547 * @next: the task we are going to switch to.
2549 * This is called with the rq lock held and interrupts off. It must
2550 * be paired with a subsequent finish_task_switch after the context
2553 * prepare_task_switch sets up locking and calls architecture specific
2557 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2558 struct task_struct
*next
)
2560 sched_info_switch(rq
, prev
, next
);
2561 perf_event_task_sched_out(prev
, next
);
2562 fire_sched_out_preempt_notifiers(prev
, next
);
2563 prepare_lock_switch(rq
, next
);
2564 prepare_arch_switch(next
);
2568 * finish_task_switch - clean up after a task-switch
2569 * @prev: the thread we just switched away from.
2571 * finish_task_switch must be called after the context switch, paired
2572 * with a prepare_task_switch call before the context switch.
2573 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2574 * and do any other architecture-specific cleanup actions.
2576 * Note that we may have delayed dropping an mm in context_switch(). If
2577 * so, we finish that here outside of the runqueue lock. (Doing it
2578 * with the lock held can cause deadlocks; see schedule() for
2581 * The context switch have flipped the stack from under us and restored the
2582 * local variables which were saved when this task called schedule() in the
2583 * past. prev == current is still correct but we need to recalculate this_rq
2584 * because prev may have moved to another CPU.
2586 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2587 __releases(rq
->lock
)
2589 struct rq
*rq
= this_rq();
2590 struct mm_struct
*mm
= rq
->prev_mm
;
2594 * The previous task will have left us with a preempt_count of 2
2595 * because it left us after:
2598 * preempt_disable(); // 1
2600 * raw_spin_lock_irq(&rq->lock) // 2
2602 * Also, see FORK_PREEMPT_COUNT.
2604 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2605 "corrupted preempt_count: %s/%d/0x%x\n",
2606 current
->comm
, current
->pid
, preempt_count()))
2607 preempt_count_set(FORK_PREEMPT_COUNT
);
2612 * A task struct has one reference for the use as "current".
2613 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2614 * schedule one last time. The schedule call will never return, and
2615 * the scheduled task must drop that reference.
2617 * We must observe prev->state before clearing prev->on_cpu (in
2618 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2619 * running on another CPU and we could rave with its RUNNING -> DEAD
2620 * transition, resulting in a double drop.
2622 prev_state
= prev
->state
;
2623 vtime_task_switch(prev
);
2624 perf_event_task_sched_in(prev
, current
);
2625 finish_lock_switch(rq
, prev
);
2626 finish_arch_post_lock_switch();
2628 fire_sched_in_preempt_notifiers(current
);
2631 if (unlikely(prev_state
== TASK_DEAD
)) {
2632 if (prev
->sched_class
->task_dead
)
2633 prev
->sched_class
->task_dead(prev
);
2636 * Remove function-return probe instances associated with this
2637 * task and put them back on the free list.
2639 kprobe_flush_task(prev
);
2640 put_task_struct(prev
);
2643 tick_nohz_task_switch();
2649 /* rq->lock is NOT held, but preemption is disabled */
2650 static void __balance_callback(struct rq
*rq
)
2652 struct callback_head
*head
, *next
;
2653 void (*func
)(struct rq
*rq
);
2654 unsigned long flags
;
2656 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2657 head
= rq
->balance_callback
;
2658 rq
->balance_callback
= NULL
;
2660 func
= (void (*)(struct rq
*))head
->func
;
2667 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2670 static inline void balance_callback(struct rq
*rq
)
2672 if (unlikely(rq
->balance_callback
))
2673 __balance_callback(rq
);
2678 static inline void balance_callback(struct rq
*rq
)
2685 * schedule_tail - first thing a freshly forked thread must call.
2686 * @prev: the thread we just switched away from.
2688 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2689 __releases(rq
->lock
)
2694 * New tasks start with FORK_PREEMPT_COUNT, see there and
2695 * finish_task_switch() for details.
2697 * finish_task_switch() will drop rq->lock() and lower preempt_count
2698 * and the preempt_enable() will end up enabling preemption (on
2699 * PREEMPT_COUNT kernels).
2702 rq
= finish_task_switch(prev
);
2703 balance_callback(rq
);
2706 if (current
->set_child_tid
)
2707 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2711 * context_switch - switch to the new MM and the new thread's register state.
2713 static __always_inline
struct rq
*
2714 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2715 struct task_struct
*next
)
2717 struct mm_struct
*mm
, *oldmm
;
2719 prepare_task_switch(rq
, prev
, next
);
2722 oldmm
= prev
->active_mm
;
2724 * For paravirt, this is coupled with an exit in switch_to to
2725 * combine the page table reload and the switch backend into
2728 arch_start_context_switch(prev
);
2731 next
->active_mm
= oldmm
;
2732 atomic_inc(&oldmm
->mm_count
);
2733 enter_lazy_tlb(oldmm
, next
);
2735 switch_mm(oldmm
, mm
, next
);
2738 prev
->active_mm
= NULL
;
2739 rq
->prev_mm
= oldmm
;
2742 * Since the runqueue lock will be released by the next
2743 * task (which is an invalid locking op but in the case
2744 * of the scheduler it's an obvious special-case), so we
2745 * do an early lockdep release here:
2747 lockdep_unpin_lock(&rq
->lock
);
2748 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2750 /* Here we just switch the register state and the stack. */
2751 switch_to(prev
, next
, prev
);
2754 return finish_task_switch(prev
);
2758 * nr_running and nr_context_switches:
2760 * externally visible scheduler statistics: current number of runnable
2761 * threads, total number of context switches performed since bootup.
2763 unsigned long nr_running(void)
2765 unsigned long i
, sum
= 0;
2767 for_each_online_cpu(i
)
2768 sum
+= cpu_rq(i
)->nr_running
;
2774 * Check if only the current task is running on the cpu.
2776 * Caution: this function does not check that the caller has disabled
2777 * preemption, thus the result might have a time-of-check-to-time-of-use
2778 * race. The caller is responsible to use it correctly, for example:
2780 * - from a non-preemptable section (of course)
2782 * - from a thread that is bound to a single CPU
2784 * - in a loop with very short iterations (e.g. a polling loop)
2786 bool single_task_running(void)
2788 return raw_rq()->nr_running
== 1;
2790 EXPORT_SYMBOL(single_task_running
);
2792 unsigned long long nr_context_switches(void)
2795 unsigned long long sum
= 0;
2797 for_each_possible_cpu(i
)
2798 sum
+= cpu_rq(i
)->nr_switches
;
2803 unsigned long nr_iowait(void)
2805 unsigned long i
, sum
= 0;
2807 for_each_possible_cpu(i
)
2808 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2813 unsigned long nr_iowait_cpu(int cpu
)
2815 struct rq
*this = cpu_rq(cpu
);
2816 return atomic_read(&this->nr_iowait
);
2819 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2821 struct rq
*rq
= this_rq();
2822 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2823 *load
= rq
->load
.weight
;
2829 * sched_exec - execve() is a valuable balancing opportunity, because at
2830 * this point the task has the smallest effective memory and cache footprint.
2832 void sched_exec(void)
2834 struct task_struct
*p
= current
;
2835 unsigned long flags
;
2838 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2839 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2840 if (dest_cpu
== smp_processor_id())
2843 if (likely(cpu_active(dest_cpu
))) {
2844 struct migration_arg arg
= { p
, dest_cpu
};
2846 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2847 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2851 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2856 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2857 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2859 EXPORT_PER_CPU_SYMBOL(kstat
);
2860 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2863 * Return accounted runtime for the task.
2864 * In case the task is currently running, return the runtime plus current's
2865 * pending runtime that have not been accounted yet.
2867 unsigned long long task_sched_runtime(struct task_struct
*p
)
2869 unsigned long flags
;
2873 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2875 * 64-bit doesn't need locks to atomically read a 64bit value.
2876 * So we have a optimization chance when the task's delta_exec is 0.
2877 * Reading ->on_cpu is racy, but this is ok.
2879 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2880 * If we race with it entering cpu, unaccounted time is 0. This is
2881 * indistinguishable from the read occurring a few cycles earlier.
2882 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2883 * been accounted, so we're correct here as well.
2885 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2886 return p
->se
.sum_exec_runtime
;
2889 rq
= task_rq_lock(p
, &flags
);
2891 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2892 * project cycles that may never be accounted to this
2893 * thread, breaking clock_gettime().
2895 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2896 update_rq_clock(rq
);
2897 p
->sched_class
->update_curr(rq
);
2899 ns
= p
->se
.sum_exec_runtime
;
2900 task_rq_unlock(rq
, p
, &flags
);
2906 * This function gets called by the timer code, with HZ frequency.
2907 * We call it with interrupts disabled.
2909 void scheduler_tick(void)
2911 int cpu
= smp_processor_id();
2912 struct rq
*rq
= cpu_rq(cpu
);
2913 struct task_struct
*curr
= rq
->curr
;
2917 raw_spin_lock(&rq
->lock
);
2918 update_rq_clock(rq
);
2919 curr
->sched_class
->task_tick(rq
, curr
, 0);
2920 cpu_load_update_active(rq
);
2921 calc_global_load_tick(rq
);
2922 raw_spin_unlock(&rq
->lock
);
2924 perf_event_task_tick();
2927 rq
->idle_balance
= idle_cpu(cpu
);
2928 trigger_load_balance(rq
);
2930 rq_last_tick_reset(rq
);
2933 #ifdef CONFIG_NO_HZ_FULL
2935 * scheduler_tick_max_deferment
2937 * Keep at least one tick per second when a single
2938 * active task is running because the scheduler doesn't
2939 * yet completely support full dynticks environment.
2941 * This makes sure that uptime, CFS vruntime, load
2942 * balancing, etc... continue to move forward, even
2943 * with a very low granularity.
2945 * Return: Maximum deferment in nanoseconds.
2947 u64
scheduler_tick_max_deferment(void)
2949 struct rq
*rq
= this_rq();
2950 unsigned long next
, now
= READ_ONCE(jiffies
);
2952 next
= rq
->last_sched_tick
+ HZ
;
2954 if (time_before_eq(next
, now
))
2957 return jiffies_to_nsecs(next
- now
);
2961 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2962 defined(CONFIG_PREEMPT_TRACER))
2964 * If the value passed in is equal to the current preempt count
2965 * then we just disabled preemption. Start timing the latency.
2967 static inline void preempt_latency_start(int val
)
2969 if (preempt_count() == val
) {
2970 unsigned long ip
= get_lock_parent_ip();
2971 #ifdef CONFIG_DEBUG_PREEMPT
2972 current
->preempt_disable_ip
= ip
;
2974 trace_preempt_off(CALLER_ADDR0
, ip
);
2978 void preempt_count_add(int val
)
2980 #ifdef CONFIG_DEBUG_PREEMPT
2984 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2987 __preempt_count_add(val
);
2988 #ifdef CONFIG_DEBUG_PREEMPT
2990 * Spinlock count overflowing soon?
2992 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2995 preempt_latency_start(val
);
2997 EXPORT_SYMBOL(preempt_count_add
);
2998 NOKPROBE_SYMBOL(preempt_count_add
);
3001 * If the value passed in equals to the current preempt count
3002 * then we just enabled preemption. Stop timing the latency.
3004 static inline void preempt_latency_stop(int val
)
3006 if (preempt_count() == val
)
3007 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3010 void preempt_count_sub(int val
)
3012 #ifdef CONFIG_DEBUG_PREEMPT
3016 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3019 * Is the spinlock portion underflowing?
3021 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3022 !(preempt_count() & PREEMPT_MASK
)))
3026 preempt_latency_stop(val
);
3027 __preempt_count_sub(val
);
3029 EXPORT_SYMBOL(preempt_count_sub
);
3030 NOKPROBE_SYMBOL(preempt_count_sub
);
3033 static inline void preempt_latency_start(int val
) { }
3034 static inline void preempt_latency_stop(int val
) { }
3038 * Print scheduling while atomic bug:
3040 static noinline
void __schedule_bug(struct task_struct
*prev
)
3042 if (oops_in_progress
)
3045 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3046 prev
->comm
, prev
->pid
, preempt_count());
3048 debug_show_held_locks(prev
);
3050 if (irqs_disabled())
3051 print_irqtrace_events(prev
);
3052 #ifdef CONFIG_DEBUG_PREEMPT
3053 if (in_atomic_preempt_off()) {
3054 pr_err("Preemption disabled at:");
3055 print_ip_sym(current
->preempt_disable_ip
);
3060 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3064 * Various schedule()-time debugging checks and statistics:
3066 static inline void schedule_debug(struct task_struct
*prev
)
3068 #ifdef CONFIG_SCHED_STACK_END_CHECK
3069 BUG_ON(task_stack_end_corrupted(prev
));
3072 if (unlikely(in_atomic_preempt_off())) {
3073 __schedule_bug(prev
);
3074 preempt_count_set(PREEMPT_DISABLED
);
3078 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3080 schedstat_inc(this_rq(), sched_count
);
3084 * Pick up the highest-prio task:
3086 static inline struct task_struct
*
3087 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3089 const struct sched_class
*class = &fair_sched_class
;
3090 struct task_struct
*p
;
3093 * Optimization: we know that if all tasks are in
3094 * the fair class we can call that function directly:
3096 if (likely(prev
->sched_class
== class &&
3097 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3098 p
= fair_sched_class
.pick_next_task(rq
, prev
);
3099 if (unlikely(p
== RETRY_TASK
))
3102 /* assumes fair_sched_class->next == idle_sched_class */
3104 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3110 for_each_class(class) {
3111 p
= class->pick_next_task(rq
, prev
);
3113 if (unlikely(p
== RETRY_TASK
))
3119 BUG(); /* the idle class will always have a runnable task */
3123 * __schedule() is the main scheduler function.
3125 * The main means of driving the scheduler and thus entering this function are:
3127 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3129 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3130 * paths. For example, see arch/x86/entry_64.S.
3132 * To drive preemption between tasks, the scheduler sets the flag in timer
3133 * interrupt handler scheduler_tick().
3135 * 3. Wakeups don't really cause entry into schedule(). They add a
3136 * task to the run-queue and that's it.
3138 * Now, if the new task added to the run-queue preempts the current
3139 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3140 * called on the nearest possible occasion:
3142 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3144 * - in syscall or exception context, at the next outmost
3145 * preempt_enable(). (this might be as soon as the wake_up()'s
3148 * - in IRQ context, return from interrupt-handler to
3149 * preemptible context
3151 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3154 * - cond_resched() call
3155 * - explicit schedule() call
3156 * - return from syscall or exception to user-space
3157 * - return from interrupt-handler to user-space
3159 * WARNING: must be called with preemption disabled!
3161 static void __sched notrace
__schedule(bool preempt
)
3163 struct task_struct
*prev
, *next
;
3164 unsigned long *switch_count
;
3168 cpu
= smp_processor_id();
3173 * do_exit() calls schedule() with preemption disabled as an exception;
3174 * however we must fix that up, otherwise the next task will see an
3175 * inconsistent (higher) preempt count.
3177 * It also avoids the below schedule_debug() test from complaining
3180 if (unlikely(prev
->state
== TASK_DEAD
))
3181 preempt_enable_no_resched_notrace();
3183 schedule_debug(prev
);
3185 if (sched_feat(HRTICK
))
3188 local_irq_disable();
3189 rcu_note_context_switch();
3192 * Make sure that signal_pending_state()->signal_pending() below
3193 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3194 * done by the caller to avoid the race with signal_wake_up().
3196 smp_mb__before_spinlock();
3197 raw_spin_lock(&rq
->lock
);
3198 lockdep_pin_lock(&rq
->lock
);
3200 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3202 switch_count
= &prev
->nivcsw
;
3203 if (!preempt
&& prev
->state
) {
3204 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3205 prev
->state
= TASK_RUNNING
;
3207 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3211 * If a worker went to sleep, notify and ask workqueue
3212 * whether it wants to wake up a task to maintain
3215 if (prev
->flags
& PF_WQ_WORKER
) {
3216 struct task_struct
*to_wakeup
;
3218 to_wakeup
= wq_worker_sleeping(prev
);
3220 try_to_wake_up_local(to_wakeup
);
3223 switch_count
= &prev
->nvcsw
;
3226 if (task_on_rq_queued(prev
))
3227 update_rq_clock(rq
);
3229 next
= pick_next_task(rq
, prev
);
3230 clear_tsk_need_resched(prev
);
3231 clear_preempt_need_resched();
3232 rq
->clock_skip_update
= 0;
3234 if (likely(prev
!= next
)) {
3239 trace_sched_switch(preempt
, prev
, next
);
3240 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3242 lockdep_unpin_lock(&rq
->lock
);
3243 raw_spin_unlock_irq(&rq
->lock
);
3246 balance_callback(rq
);
3248 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3250 static inline void sched_submit_work(struct task_struct
*tsk
)
3252 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3255 * If we are going to sleep and we have plugged IO queued,
3256 * make sure to submit it to avoid deadlocks.
3258 if (blk_needs_flush_plug(tsk
))
3259 blk_schedule_flush_plug(tsk
);
3262 asmlinkage __visible
void __sched
schedule(void)
3264 struct task_struct
*tsk
= current
;
3266 sched_submit_work(tsk
);
3270 sched_preempt_enable_no_resched();
3271 } while (need_resched());
3273 EXPORT_SYMBOL(schedule
);
3275 #ifdef CONFIG_CONTEXT_TRACKING
3276 asmlinkage __visible
void __sched
schedule_user(void)
3279 * If we come here after a random call to set_need_resched(),
3280 * or we have been woken up remotely but the IPI has not yet arrived,
3281 * we haven't yet exited the RCU idle mode. Do it here manually until
3282 * we find a better solution.
3284 * NB: There are buggy callers of this function. Ideally we
3285 * should warn if prev_state != CONTEXT_USER, but that will trigger
3286 * too frequently to make sense yet.
3288 enum ctx_state prev_state
= exception_enter();
3290 exception_exit(prev_state
);
3295 * schedule_preempt_disabled - called with preemption disabled
3297 * Returns with preemption disabled. Note: preempt_count must be 1
3299 void __sched
schedule_preempt_disabled(void)
3301 sched_preempt_enable_no_resched();
3306 static void __sched notrace
preempt_schedule_common(void)
3310 * Because the function tracer can trace preempt_count_sub()
3311 * and it also uses preempt_enable/disable_notrace(), if
3312 * NEED_RESCHED is set, the preempt_enable_notrace() called
3313 * by the function tracer will call this function again and
3314 * cause infinite recursion.
3316 * Preemption must be disabled here before the function
3317 * tracer can trace. Break up preempt_disable() into two
3318 * calls. One to disable preemption without fear of being
3319 * traced. The other to still record the preemption latency,
3320 * which can also be traced by the function tracer.
3322 preempt_disable_notrace();
3323 preempt_latency_start(1);
3325 preempt_latency_stop(1);
3326 preempt_enable_no_resched_notrace();
3329 * Check again in case we missed a preemption opportunity
3330 * between schedule and now.
3332 } while (need_resched());
3335 #ifdef CONFIG_PREEMPT
3337 * this is the entry point to schedule() from in-kernel preemption
3338 * off of preempt_enable. Kernel preemptions off return from interrupt
3339 * occur there and call schedule directly.
3341 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3344 * If there is a non-zero preempt_count or interrupts are disabled,
3345 * we do not want to preempt the current task. Just return..
3347 if (likely(!preemptible()))
3350 preempt_schedule_common();
3352 NOKPROBE_SYMBOL(preempt_schedule
);
3353 EXPORT_SYMBOL(preempt_schedule
);
3356 * preempt_schedule_notrace - preempt_schedule called by tracing
3358 * The tracing infrastructure uses preempt_enable_notrace to prevent
3359 * recursion and tracing preempt enabling caused by the tracing
3360 * infrastructure itself. But as tracing can happen in areas coming
3361 * from userspace or just about to enter userspace, a preempt enable
3362 * can occur before user_exit() is called. This will cause the scheduler
3363 * to be called when the system is still in usermode.
3365 * To prevent this, the preempt_enable_notrace will use this function
3366 * instead of preempt_schedule() to exit user context if needed before
3367 * calling the scheduler.
3369 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3371 enum ctx_state prev_ctx
;
3373 if (likely(!preemptible()))
3378 * Because the function tracer can trace preempt_count_sub()
3379 * and it also uses preempt_enable/disable_notrace(), if
3380 * NEED_RESCHED is set, the preempt_enable_notrace() called
3381 * by the function tracer will call this function again and
3382 * cause infinite recursion.
3384 * Preemption must be disabled here before the function
3385 * tracer can trace. Break up preempt_disable() into two
3386 * calls. One to disable preemption without fear of being
3387 * traced. The other to still record the preemption latency,
3388 * which can also be traced by the function tracer.
3390 preempt_disable_notrace();
3391 preempt_latency_start(1);
3393 * Needs preempt disabled in case user_exit() is traced
3394 * and the tracer calls preempt_enable_notrace() causing
3395 * an infinite recursion.
3397 prev_ctx
= exception_enter();
3399 exception_exit(prev_ctx
);
3401 preempt_latency_stop(1);
3402 preempt_enable_no_resched_notrace();
3403 } while (need_resched());
3405 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3407 #endif /* CONFIG_PREEMPT */
3410 * this is the entry point to schedule() from kernel preemption
3411 * off of irq context.
3412 * Note, that this is called and return with irqs disabled. This will
3413 * protect us against recursive calling from irq.
3415 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3417 enum ctx_state prev_state
;
3419 /* Catch callers which need to be fixed */
3420 BUG_ON(preempt_count() || !irqs_disabled());
3422 prev_state
= exception_enter();
3428 local_irq_disable();
3429 sched_preempt_enable_no_resched();
3430 } while (need_resched());
3432 exception_exit(prev_state
);
3435 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3438 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3440 EXPORT_SYMBOL(default_wake_function
);
3442 #ifdef CONFIG_RT_MUTEXES
3445 * rt_mutex_setprio - set the current priority of a task
3447 * @prio: prio value (kernel-internal form)
3449 * This function changes the 'effective' priority of a task. It does
3450 * not touch ->normal_prio like __setscheduler().
3452 * Used by the rt_mutex code to implement priority inheritance
3453 * logic. Call site only calls if the priority of the task changed.
3455 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3457 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3459 const struct sched_class
*prev_class
;
3461 BUG_ON(prio
> MAX_PRIO
);
3463 rq
= __task_rq_lock(p
);
3466 * Idle task boosting is a nono in general. There is one
3467 * exception, when PREEMPT_RT and NOHZ is active:
3469 * The idle task calls get_next_timer_interrupt() and holds
3470 * the timer wheel base->lock on the CPU and another CPU wants
3471 * to access the timer (probably to cancel it). We can safely
3472 * ignore the boosting request, as the idle CPU runs this code
3473 * with interrupts disabled and will complete the lock
3474 * protected section without being interrupted. So there is no
3475 * real need to boost.
3477 if (unlikely(p
== rq
->idle
)) {
3478 WARN_ON(p
!= rq
->curr
);
3479 WARN_ON(p
->pi_blocked_on
);
3483 trace_sched_pi_setprio(p
, prio
);
3486 if (oldprio
== prio
)
3487 queue_flag
&= ~DEQUEUE_MOVE
;
3489 prev_class
= p
->sched_class
;
3490 queued
= task_on_rq_queued(p
);
3491 running
= task_current(rq
, p
);
3493 dequeue_task(rq
, p
, queue_flag
);
3495 put_prev_task(rq
, p
);
3498 * Boosting condition are:
3499 * 1. -rt task is running and holds mutex A
3500 * --> -dl task blocks on mutex A
3502 * 2. -dl task is running and holds mutex A
3503 * --> -dl task blocks on mutex A and could preempt the
3506 if (dl_prio(prio
)) {
3507 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3508 if (!dl_prio(p
->normal_prio
) ||
3509 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3510 p
->dl
.dl_boosted
= 1;
3511 queue_flag
|= ENQUEUE_REPLENISH
;
3513 p
->dl
.dl_boosted
= 0;
3514 p
->sched_class
= &dl_sched_class
;
3515 } else if (rt_prio(prio
)) {
3516 if (dl_prio(oldprio
))
3517 p
->dl
.dl_boosted
= 0;
3519 queue_flag
|= ENQUEUE_HEAD
;
3520 p
->sched_class
= &rt_sched_class
;
3522 if (dl_prio(oldprio
))
3523 p
->dl
.dl_boosted
= 0;
3524 if (rt_prio(oldprio
))
3526 p
->sched_class
= &fair_sched_class
;
3532 p
->sched_class
->set_curr_task(rq
);
3534 enqueue_task(rq
, p
, queue_flag
);
3536 check_class_changed(rq
, p
, prev_class
, oldprio
);
3538 preempt_disable(); /* avoid rq from going away on us */
3539 __task_rq_unlock(rq
);
3541 balance_callback(rq
);
3546 void set_user_nice(struct task_struct
*p
, long nice
)
3548 int old_prio
, delta
, queued
;
3549 unsigned long flags
;
3552 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3555 * We have to be careful, if called from sys_setpriority(),
3556 * the task might be in the middle of scheduling on another CPU.
3558 rq
= task_rq_lock(p
, &flags
);
3560 * The RT priorities are set via sched_setscheduler(), but we still
3561 * allow the 'normal' nice value to be set - but as expected
3562 * it wont have any effect on scheduling until the task is
3563 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3565 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3566 p
->static_prio
= NICE_TO_PRIO(nice
);
3569 queued
= task_on_rq_queued(p
);
3571 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3573 p
->static_prio
= NICE_TO_PRIO(nice
);
3576 p
->prio
= effective_prio(p
);
3577 delta
= p
->prio
- old_prio
;
3580 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3582 * If the task increased its priority or is running and
3583 * lowered its priority, then reschedule its CPU:
3585 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3589 task_rq_unlock(rq
, p
, &flags
);
3591 EXPORT_SYMBOL(set_user_nice
);
3594 * can_nice - check if a task can reduce its nice value
3598 int can_nice(const struct task_struct
*p
, const int nice
)
3600 /* convert nice value [19,-20] to rlimit style value [1,40] */
3601 int nice_rlim
= nice_to_rlimit(nice
);
3603 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3604 capable(CAP_SYS_NICE
));
3607 #ifdef __ARCH_WANT_SYS_NICE
3610 * sys_nice - change the priority of the current process.
3611 * @increment: priority increment
3613 * sys_setpriority is a more generic, but much slower function that
3614 * does similar things.
3616 SYSCALL_DEFINE1(nice
, int, increment
)
3621 * Setpriority might change our priority at the same moment.
3622 * We don't have to worry. Conceptually one call occurs first
3623 * and we have a single winner.
3625 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3626 nice
= task_nice(current
) + increment
;
3628 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3629 if (increment
< 0 && !can_nice(current
, nice
))
3632 retval
= security_task_setnice(current
, nice
);
3636 set_user_nice(current
, nice
);
3643 * task_prio - return the priority value of a given task.
3644 * @p: the task in question.
3646 * Return: The priority value as seen by users in /proc.
3647 * RT tasks are offset by -200. Normal tasks are centered
3648 * around 0, value goes from -16 to +15.
3650 int task_prio(const struct task_struct
*p
)
3652 return p
->prio
- MAX_RT_PRIO
;
3656 * idle_cpu - is a given cpu idle currently?
3657 * @cpu: the processor in question.
3659 * Return: 1 if the CPU is currently idle. 0 otherwise.
3661 int idle_cpu(int cpu
)
3663 struct rq
*rq
= cpu_rq(cpu
);
3665 if (rq
->curr
!= rq
->idle
)
3672 if (!llist_empty(&rq
->wake_list
))
3680 * idle_task - return the idle task for a given cpu.
3681 * @cpu: the processor in question.
3683 * Return: The idle task for the cpu @cpu.
3685 struct task_struct
*idle_task(int cpu
)
3687 return cpu_rq(cpu
)->idle
;
3691 * find_process_by_pid - find a process with a matching PID value.
3692 * @pid: the pid in question.
3694 * The task of @pid, if found. %NULL otherwise.
3696 static struct task_struct
*find_process_by_pid(pid_t pid
)
3698 return pid
? find_task_by_vpid(pid
) : current
;
3702 * This function initializes the sched_dl_entity of a newly becoming
3703 * SCHED_DEADLINE task.
3705 * Only the static values are considered here, the actual runtime and the
3706 * absolute deadline will be properly calculated when the task is enqueued
3707 * for the first time with its new policy.
3710 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3712 struct sched_dl_entity
*dl_se
= &p
->dl
;
3714 dl_se
->dl_runtime
= attr
->sched_runtime
;
3715 dl_se
->dl_deadline
= attr
->sched_deadline
;
3716 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3717 dl_se
->flags
= attr
->sched_flags
;
3718 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3721 * Changing the parameters of a task is 'tricky' and we're not doing
3722 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3724 * What we SHOULD do is delay the bandwidth release until the 0-lag
3725 * point. This would include retaining the task_struct until that time
3726 * and change dl_overflow() to not immediately decrement the current
3729 * Instead we retain the current runtime/deadline and let the new
3730 * parameters take effect after the current reservation period lapses.
3731 * This is safe (albeit pessimistic) because the 0-lag point is always
3732 * before the current scheduling deadline.
3734 * We can still have temporary overloads because we do not delay the
3735 * change in bandwidth until that time; so admission control is
3736 * not on the safe side. It does however guarantee tasks will never
3737 * consume more than promised.
3742 * sched_setparam() passes in -1 for its policy, to let the functions
3743 * it calls know not to change it.
3745 #define SETPARAM_POLICY -1
3747 static void __setscheduler_params(struct task_struct
*p
,
3748 const struct sched_attr
*attr
)
3750 int policy
= attr
->sched_policy
;
3752 if (policy
== SETPARAM_POLICY
)
3757 if (dl_policy(policy
))
3758 __setparam_dl(p
, attr
);
3759 else if (fair_policy(policy
))
3760 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3763 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3764 * !rt_policy. Always setting this ensures that things like
3765 * getparam()/getattr() don't report silly values for !rt tasks.
3767 p
->rt_priority
= attr
->sched_priority
;
3768 p
->normal_prio
= normal_prio(p
);
3772 /* Actually do priority change: must hold pi & rq lock. */
3773 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3774 const struct sched_attr
*attr
, bool keep_boost
)
3776 __setscheduler_params(p
, attr
);
3779 * Keep a potential priority boosting if called from
3780 * sched_setscheduler().
3783 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3785 p
->prio
= normal_prio(p
);
3787 if (dl_prio(p
->prio
))
3788 p
->sched_class
= &dl_sched_class
;
3789 else if (rt_prio(p
->prio
))
3790 p
->sched_class
= &rt_sched_class
;
3792 p
->sched_class
= &fair_sched_class
;
3796 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3798 struct sched_dl_entity
*dl_se
= &p
->dl
;
3800 attr
->sched_priority
= p
->rt_priority
;
3801 attr
->sched_runtime
= dl_se
->dl_runtime
;
3802 attr
->sched_deadline
= dl_se
->dl_deadline
;
3803 attr
->sched_period
= dl_se
->dl_period
;
3804 attr
->sched_flags
= dl_se
->flags
;
3808 * This function validates the new parameters of a -deadline task.
3809 * We ask for the deadline not being zero, and greater or equal
3810 * than the runtime, as well as the period of being zero or
3811 * greater than deadline. Furthermore, we have to be sure that
3812 * user parameters are above the internal resolution of 1us (we
3813 * check sched_runtime only since it is always the smaller one) and
3814 * below 2^63 ns (we have to check both sched_deadline and
3815 * sched_period, as the latter can be zero).
3818 __checkparam_dl(const struct sched_attr
*attr
)
3821 if (attr
->sched_deadline
== 0)
3825 * Since we truncate DL_SCALE bits, make sure we're at least
3828 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3832 * Since we use the MSB for wrap-around and sign issues, make
3833 * sure it's not set (mind that period can be equal to zero).
3835 if (attr
->sched_deadline
& (1ULL << 63) ||
3836 attr
->sched_period
& (1ULL << 63))
3839 /* runtime <= deadline <= period (if period != 0) */
3840 if ((attr
->sched_period
!= 0 &&
3841 attr
->sched_period
< attr
->sched_deadline
) ||
3842 attr
->sched_deadline
< attr
->sched_runtime
)
3849 * check the target process has a UID that matches the current process's
3851 static bool check_same_owner(struct task_struct
*p
)
3853 const struct cred
*cred
= current_cred(), *pcred
;
3857 pcred
= __task_cred(p
);
3858 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3859 uid_eq(cred
->euid
, pcred
->uid
));
3864 static bool dl_param_changed(struct task_struct
*p
,
3865 const struct sched_attr
*attr
)
3867 struct sched_dl_entity
*dl_se
= &p
->dl
;
3869 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3870 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3871 dl_se
->dl_period
!= attr
->sched_period
||
3872 dl_se
->flags
!= attr
->sched_flags
)
3878 static int __sched_setscheduler(struct task_struct
*p
,
3879 const struct sched_attr
*attr
,
3882 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3883 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3884 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3885 int new_effective_prio
, policy
= attr
->sched_policy
;
3886 unsigned long flags
;
3887 const struct sched_class
*prev_class
;
3890 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3892 /* may grab non-irq protected spin_locks */
3893 BUG_ON(in_interrupt());
3895 /* double check policy once rq lock held */
3897 reset_on_fork
= p
->sched_reset_on_fork
;
3898 policy
= oldpolicy
= p
->policy
;
3900 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3902 if (!valid_policy(policy
))
3906 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3910 * Valid priorities for SCHED_FIFO and SCHED_RR are
3911 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3912 * SCHED_BATCH and SCHED_IDLE is 0.
3914 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3915 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3917 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3918 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3922 * Allow unprivileged RT tasks to decrease priority:
3924 if (user
&& !capable(CAP_SYS_NICE
)) {
3925 if (fair_policy(policy
)) {
3926 if (attr
->sched_nice
< task_nice(p
) &&
3927 !can_nice(p
, attr
->sched_nice
))
3931 if (rt_policy(policy
)) {
3932 unsigned long rlim_rtprio
=
3933 task_rlimit(p
, RLIMIT_RTPRIO
);
3935 /* can't set/change the rt policy */
3936 if (policy
!= p
->policy
&& !rlim_rtprio
)
3939 /* can't increase priority */
3940 if (attr
->sched_priority
> p
->rt_priority
&&
3941 attr
->sched_priority
> rlim_rtprio
)
3946 * Can't set/change SCHED_DEADLINE policy at all for now
3947 * (safest behavior); in the future we would like to allow
3948 * unprivileged DL tasks to increase their relative deadline
3949 * or reduce their runtime (both ways reducing utilization)
3951 if (dl_policy(policy
))
3955 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3956 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3958 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
3959 if (!can_nice(p
, task_nice(p
)))
3963 /* can't change other user's priorities */
3964 if (!check_same_owner(p
))
3967 /* Normal users shall not reset the sched_reset_on_fork flag */
3968 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3973 retval
= security_task_setscheduler(p
);
3979 * make sure no PI-waiters arrive (or leave) while we are
3980 * changing the priority of the task:
3982 * To be able to change p->policy safely, the appropriate
3983 * runqueue lock must be held.
3985 rq
= task_rq_lock(p
, &flags
);
3988 * Changing the policy of the stop threads its a very bad idea
3990 if (p
== rq
->stop
) {
3991 task_rq_unlock(rq
, p
, &flags
);
3996 * If not changing anything there's no need to proceed further,
3997 * but store a possible modification of reset_on_fork.
3999 if (unlikely(policy
== p
->policy
)) {
4000 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4002 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4004 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4007 p
->sched_reset_on_fork
= reset_on_fork
;
4008 task_rq_unlock(rq
, p
, &flags
);
4014 #ifdef CONFIG_RT_GROUP_SCHED
4016 * Do not allow realtime tasks into groups that have no runtime
4019 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4020 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4021 !task_group_is_autogroup(task_group(p
))) {
4022 task_rq_unlock(rq
, p
, &flags
);
4027 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4028 cpumask_t
*span
= rq
->rd
->span
;
4031 * Don't allow tasks with an affinity mask smaller than
4032 * the entire root_domain to become SCHED_DEADLINE. We
4033 * will also fail if there's no bandwidth available.
4035 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4036 rq
->rd
->dl_bw
.bw
== 0) {
4037 task_rq_unlock(rq
, p
, &flags
);
4044 /* recheck policy now with rq lock held */
4045 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4046 policy
= oldpolicy
= -1;
4047 task_rq_unlock(rq
, p
, &flags
);
4052 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4053 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4056 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4057 task_rq_unlock(rq
, p
, &flags
);
4061 p
->sched_reset_on_fork
= reset_on_fork
;
4066 * Take priority boosted tasks into account. If the new
4067 * effective priority is unchanged, we just store the new
4068 * normal parameters and do not touch the scheduler class and
4069 * the runqueue. This will be done when the task deboost
4072 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4073 if (new_effective_prio
== oldprio
)
4074 queue_flags
&= ~DEQUEUE_MOVE
;
4077 queued
= task_on_rq_queued(p
);
4078 running
= task_current(rq
, p
);
4080 dequeue_task(rq
, p
, queue_flags
);
4082 put_prev_task(rq
, p
);
4084 prev_class
= p
->sched_class
;
4085 __setscheduler(rq
, p
, attr
, pi
);
4088 p
->sched_class
->set_curr_task(rq
);
4091 * We enqueue to tail when the priority of a task is
4092 * increased (user space view).
4094 if (oldprio
< p
->prio
)
4095 queue_flags
|= ENQUEUE_HEAD
;
4097 enqueue_task(rq
, p
, queue_flags
);
4100 check_class_changed(rq
, p
, prev_class
, oldprio
);
4101 preempt_disable(); /* avoid rq from going away on us */
4102 task_rq_unlock(rq
, p
, &flags
);
4105 rt_mutex_adjust_pi(p
);
4108 * Run balance callbacks after we've adjusted the PI chain.
4110 balance_callback(rq
);
4116 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4117 const struct sched_param
*param
, bool check
)
4119 struct sched_attr attr
= {
4120 .sched_policy
= policy
,
4121 .sched_priority
= param
->sched_priority
,
4122 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4125 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4126 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4127 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4128 policy
&= ~SCHED_RESET_ON_FORK
;
4129 attr
.sched_policy
= policy
;
4132 return __sched_setscheduler(p
, &attr
, check
, true);
4135 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4136 * @p: the task in question.
4137 * @policy: new policy.
4138 * @param: structure containing the new RT priority.
4140 * Return: 0 on success. An error code otherwise.
4142 * NOTE that the task may be already dead.
4144 int sched_setscheduler(struct task_struct
*p
, int policy
,
4145 const struct sched_param
*param
)
4147 return _sched_setscheduler(p
, policy
, param
, true);
4149 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4151 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4153 return __sched_setscheduler(p
, attr
, true, true);
4155 EXPORT_SYMBOL_GPL(sched_setattr
);
4158 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4159 * @p: the task in question.
4160 * @policy: new policy.
4161 * @param: structure containing the new RT priority.
4163 * Just like sched_setscheduler, only don't bother checking if the
4164 * current context has permission. For example, this is needed in
4165 * stop_machine(): we create temporary high priority worker threads,
4166 * but our caller might not have that capability.
4168 * Return: 0 on success. An error code otherwise.
4170 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4171 const struct sched_param
*param
)
4173 return _sched_setscheduler(p
, policy
, param
, false);
4175 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4178 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4180 struct sched_param lparam
;
4181 struct task_struct
*p
;
4184 if (!param
|| pid
< 0)
4186 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4191 p
= find_process_by_pid(pid
);
4193 retval
= sched_setscheduler(p
, policy
, &lparam
);
4200 * Mimics kernel/events/core.c perf_copy_attr().
4202 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4203 struct sched_attr
*attr
)
4208 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4212 * zero the full structure, so that a short copy will be nice.
4214 memset(attr
, 0, sizeof(*attr
));
4216 ret
= get_user(size
, &uattr
->size
);
4220 if (size
> PAGE_SIZE
) /* silly large */
4223 if (!size
) /* abi compat */
4224 size
= SCHED_ATTR_SIZE_VER0
;
4226 if (size
< SCHED_ATTR_SIZE_VER0
)
4230 * If we're handed a bigger struct than we know of,
4231 * ensure all the unknown bits are 0 - i.e. new
4232 * user-space does not rely on any kernel feature
4233 * extensions we dont know about yet.
4235 if (size
> sizeof(*attr
)) {
4236 unsigned char __user
*addr
;
4237 unsigned char __user
*end
;
4240 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4241 end
= (void __user
*)uattr
+ size
;
4243 for (; addr
< end
; addr
++) {
4244 ret
= get_user(val
, addr
);
4250 size
= sizeof(*attr
);
4253 ret
= copy_from_user(attr
, uattr
, size
);
4258 * XXX: do we want to be lenient like existing syscalls; or do we want
4259 * to be strict and return an error on out-of-bounds values?
4261 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4266 put_user(sizeof(*attr
), &uattr
->size
);
4271 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4272 * @pid: the pid in question.
4273 * @policy: new policy.
4274 * @param: structure containing the new RT priority.
4276 * Return: 0 on success. An error code otherwise.
4278 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4279 struct sched_param __user
*, param
)
4281 /* negative values for policy are not valid */
4285 return do_sched_setscheduler(pid
, policy
, param
);
4289 * sys_sched_setparam - set/change the RT priority of a thread
4290 * @pid: the pid in question.
4291 * @param: structure containing the new RT priority.
4293 * Return: 0 on success. An error code otherwise.
4295 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4297 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4301 * sys_sched_setattr - same as above, but with extended sched_attr
4302 * @pid: the pid in question.
4303 * @uattr: structure containing the extended parameters.
4304 * @flags: for future extension.
4306 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4307 unsigned int, flags
)
4309 struct sched_attr attr
;
4310 struct task_struct
*p
;
4313 if (!uattr
|| pid
< 0 || flags
)
4316 retval
= sched_copy_attr(uattr
, &attr
);
4320 if ((int)attr
.sched_policy
< 0)
4325 p
= find_process_by_pid(pid
);
4327 retval
= sched_setattr(p
, &attr
);
4334 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4335 * @pid: the pid in question.
4337 * Return: On success, the policy of the thread. Otherwise, a negative error
4340 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4342 struct task_struct
*p
;
4350 p
= find_process_by_pid(pid
);
4352 retval
= security_task_getscheduler(p
);
4355 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4362 * sys_sched_getparam - get the RT priority of a thread
4363 * @pid: the pid in question.
4364 * @param: structure containing the RT priority.
4366 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4369 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4371 struct sched_param lp
= { .sched_priority
= 0 };
4372 struct task_struct
*p
;
4375 if (!param
|| pid
< 0)
4379 p
= find_process_by_pid(pid
);
4384 retval
= security_task_getscheduler(p
);
4388 if (task_has_rt_policy(p
))
4389 lp
.sched_priority
= p
->rt_priority
;
4393 * This one might sleep, we cannot do it with a spinlock held ...
4395 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4404 static int sched_read_attr(struct sched_attr __user
*uattr
,
4405 struct sched_attr
*attr
,
4410 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4414 * If we're handed a smaller struct than we know of,
4415 * ensure all the unknown bits are 0 - i.e. old
4416 * user-space does not get uncomplete information.
4418 if (usize
< sizeof(*attr
)) {
4419 unsigned char *addr
;
4422 addr
= (void *)attr
+ usize
;
4423 end
= (void *)attr
+ sizeof(*attr
);
4425 for (; addr
< end
; addr
++) {
4433 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4441 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4442 * @pid: the pid in question.
4443 * @uattr: structure containing the extended parameters.
4444 * @size: sizeof(attr) for fwd/bwd comp.
4445 * @flags: for future extension.
4447 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4448 unsigned int, size
, unsigned int, flags
)
4450 struct sched_attr attr
= {
4451 .size
= sizeof(struct sched_attr
),
4453 struct task_struct
*p
;
4456 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4457 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4461 p
= find_process_by_pid(pid
);
4466 retval
= security_task_getscheduler(p
);
4470 attr
.sched_policy
= p
->policy
;
4471 if (p
->sched_reset_on_fork
)
4472 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4473 if (task_has_dl_policy(p
))
4474 __getparam_dl(p
, &attr
);
4475 else if (task_has_rt_policy(p
))
4476 attr
.sched_priority
= p
->rt_priority
;
4478 attr
.sched_nice
= task_nice(p
);
4482 retval
= sched_read_attr(uattr
, &attr
, size
);
4490 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4492 cpumask_var_t cpus_allowed
, new_mask
;
4493 struct task_struct
*p
;
4498 p
= find_process_by_pid(pid
);
4504 /* Prevent p going away */
4508 if (p
->flags
& PF_NO_SETAFFINITY
) {
4512 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4516 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4518 goto out_free_cpus_allowed
;
4521 if (!check_same_owner(p
)) {
4523 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4525 goto out_free_new_mask
;
4530 retval
= security_task_setscheduler(p
);
4532 goto out_free_new_mask
;
4535 cpuset_cpus_allowed(p
, cpus_allowed
);
4536 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4539 * Since bandwidth control happens on root_domain basis,
4540 * if admission test is enabled, we only admit -deadline
4541 * tasks allowed to run on all the CPUs in the task's
4545 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4547 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4550 goto out_free_new_mask
;
4556 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4559 cpuset_cpus_allowed(p
, cpus_allowed
);
4560 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4562 * We must have raced with a concurrent cpuset
4563 * update. Just reset the cpus_allowed to the
4564 * cpuset's cpus_allowed
4566 cpumask_copy(new_mask
, cpus_allowed
);
4571 free_cpumask_var(new_mask
);
4572 out_free_cpus_allowed
:
4573 free_cpumask_var(cpus_allowed
);
4579 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4580 struct cpumask
*new_mask
)
4582 if (len
< cpumask_size())
4583 cpumask_clear(new_mask
);
4584 else if (len
> cpumask_size())
4585 len
= cpumask_size();
4587 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4591 * sys_sched_setaffinity - set the cpu affinity of a process
4592 * @pid: pid of the process
4593 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4594 * @user_mask_ptr: user-space pointer to the new cpu mask
4596 * Return: 0 on success. An error code otherwise.
4598 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4599 unsigned long __user
*, user_mask_ptr
)
4601 cpumask_var_t new_mask
;
4604 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4607 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4609 retval
= sched_setaffinity(pid
, new_mask
);
4610 free_cpumask_var(new_mask
);
4614 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4616 struct task_struct
*p
;
4617 unsigned long flags
;
4623 p
= find_process_by_pid(pid
);
4627 retval
= security_task_getscheduler(p
);
4631 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4632 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4633 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4642 * sys_sched_getaffinity - get the cpu affinity of a process
4643 * @pid: pid of the process
4644 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4645 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4647 * Return: 0 on success. An error code otherwise.
4649 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4650 unsigned long __user
*, user_mask_ptr
)
4655 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4657 if (len
& (sizeof(unsigned long)-1))
4660 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4663 ret
= sched_getaffinity(pid
, mask
);
4665 size_t retlen
= min_t(size_t, len
, cpumask_size());
4667 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4672 free_cpumask_var(mask
);
4678 * sys_sched_yield - yield the current processor to other threads.
4680 * This function yields the current CPU to other tasks. If there are no
4681 * other threads running on this CPU then this function will return.
4685 SYSCALL_DEFINE0(sched_yield
)
4687 struct rq
*rq
= this_rq_lock();
4689 schedstat_inc(rq
, yld_count
);
4690 current
->sched_class
->yield_task(rq
);
4693 * Since we are going to call schedule() anyway, there's
4694 * no need to preempt or enable interrupts:
4696 __release(rq
->lock
);
4697 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4698 do_raw_spin_unlock(&rq
->lock
);
4699 sched_preempt_enable_no_resched();
4706 int __sched
_cond_resched(void)
4708 if (should_resched(0)) {
4709 preempt_schedule_common();
4714 EXPORT_SYMBOL(_cond_resched
);
4717 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4718 * call schedule, and on return reacquire the lock.
4720 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4721 * operations here to prevent schedule() from being called twice (once via
4722 * spin_unlock(), once by hand).
4724 int __cond_resched_lock(spinlock_t
*lock
)
4726 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4729 lockdep_assert_held(lock
);
4731 if (spin_needbreak(lock
) || resched
) {
4734 preempt_schedule_common();
4742 EXPORT_SYMBOL(__cond_resched_lock
);
4744 int __sched
__cond_resched_softirq(void)
4746 BUG_ON(!in_softirq());
4748 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4750 preempt_schedule_common();
4756 EXPORT_SYMBOL(__cond_resched_softirq
);
4759 * yield - yield the current processor to other threads.
4761 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4763 * The scheduler is at all times free to pick the calling task as the most
4764 * eligible task to run, if removing the yield() call from your code breaks
4765 * it, its already broken.
4767 * Typical broken usage is:
4772 * where one assumes that yield() will let 'the other' process run that will
4773 * make event true. If the current task is a SCHED_FIFO task that will never
4774 * happen. Never use yield() as a progress guarantee!!
4776 * If you want to use yield() to wait for something, use wait_event().
4777 * If you want to use yield() to be 'nice' for others, use cond_resched().
4778 * If you still want to use yield(), do not!
4780 void __sched
yield(void)
4782 set_current_state(TASK_RUNNING
);
4785 EXPORT_SYMBOL(yield
);
4788 * yield_to - yield the current processor to another thread in
4789 * your thread group, or accelerate that thread toward the
4790 * processor it's on.
4792 * @preempt: whether task preemption is allowed or not
4794 * It's the caller's job to ensure that the target task struct
4795 * can't go away on us before we can do any checks.
4798 * true (>0) if we indeed boosted the target task.
4799 * false (0) if we failed to boost the target.
4800 * -ESRCH if there's no task to yield to.
4802 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4804 struct task_struct
*curr
= current
;
4805 struct rq
*rq
, *p_rq
;
4806 unsigned long flags
;
4809 local_irq_save(flags
);
4815 * If we're the only runnable task on the rq and target rq also
4816 * has only one task, there's absolutely no point in yielding.
4818 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4823 double_rq_lock(rq
, p_rq
);
4824 if (task_rq(p
) != p_rq
) {
4825 double_rq_unlock(rq
, p_rq
);
4829 if (!curr
->sched_class
->yield_to_task
)
4832 if (curr
->sched_class
!= p
->sched_class
)
4835 if (task_running(p_rq
, p
) || p
->state
)
4838 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4840 schedstat_inc(rq
, yld_count
);
4842 * Make p's CPU reschedule; pick_next_entity takes care of
4845 if (preempt
&& rq
!= p_rq
)
4850 double_rq_unlock(rq
, p_rq
);
4852 local_irq_restore(flags
);
4859 EXPORT_SYMBOL_GPL(yield_to
);
4862 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4863 * that process accounting knows that this is a task in IO wait state.
4865 long __sched
io_schedule_timeout(long timeout
)
4867 int old_iowait
= current
->in_iowait
;
4871 current
->in_iowait
= 1;
4872 blk_schedule_flush_plug(current
);
4874 delayacct_blkio_start();
4876 atomic_inc(&rq
->nr_iowait
);
4877 ret
= schedule_timeout(timeout
);
4878 current
->in_iowait
= old_iowait
;
4879 atomic_dec(&rq
->nr_iowait
);
4880 delayacct_blkio_end();
4884 EXPORT_SYMBOL(io_schedule_timeout
);
4887 * sys_sched_get_priority_max - return maximum RT priority.
4888 * @policy: scheduling class.
4890 * Return: On success, this syscall returns the maximum
4891 * rt_priority that can be used by a given scheduling class.
4892 * On failure, a negative error code is returned.
4894 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4901 ret
= MAX_USER_RT_PRIO
-1;
4903 case SCHED_DEADLINE
:
4914 * sys_sched_get_priority_min - return minimum RT priority.
4915 * @policy: scheduling class.
4917 * Return: On success, this syscall returns the minimum
4918 * rt_priority that can be used by a given scheduling class.
4919 * On failure, a negative error code is returned.
4921 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4930 case SCHED_DEADLINE
:
4940 * sys_sched_rr_get_interval - return the default timeslice of a process.
4941 * @pid: pid of the process.
4942 * @interval: userspace pointer to the timeslice value.
4944 * this syscall writes the default timeslice value of a given process
4945 * into the user-space timespec buffer. A value of '0' means infinity.
4947 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4950 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4951 struct timespec __user
*, interval
)
4953 struct task_struct
*p
;
4954 unsigned int time_slice
;
4955 unsigned long flags
;
4965 p
= find_process_by_pid(pid
);
4969 retval
= security_task_getscheduler(p
);
4973 rq
= task_rq_lock(p
, &flags
);
4975 if (p
->sched_class
->get_rr_interval
)
4976 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4977 task_rq_unlock(rq
, p
, &flags
);
4980 jiffies_to_timespec(time_slice
, &t
);
4981 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4989 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4991 void sched_show_task(struct task_struct
*p
)
4993 unsigned long free
= 0;
4995 unsigned long state
= p
->state
;
4998 state
= __ffs(state
) + 1;
4999 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5000 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5001 #if BITS_PER_LONG == 32
5002 if (state
== TASK_RUNNING
)
5003 printk(KERN_CONT
" running ");
5005 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5007 if (state
== TASK_RUNNING
)
5008 printk(KERN_CONT
" running task ");
5010 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5012 #ifdef CONFIG_DEBUG_STACK_USAGE
5013 free
= stack_not_used(p
);
5018 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5020 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5021 task_pid_nr(p
), ppid
,
5022 (unsigned long)task_thread_info(p
)->flags
);
5024 print_worker_info(KERN_INFO
, p
);
5025 show_stack(p
, NULL
);
5028 void show_state_filter(unsigned long state_filter
)
5030 struct task_struct
*g
, *p
;
5032 #if BITS_PER_LONG == 32
5034 " task PC stack pid father\n");
5037 " task PC stack pid father\n");
5040 for_each_process_thread(g
, p
) {
5042 * reset the NMI-timeout, listing all files on a slow
5043 * console might take a lot of time:
5045 touch_nmi_watchdog();
5046 if (!state_filter
|| (p
->state
& state_filter
))
5050 touch_all_softlockup_watchdogs();
5052 #ifdef CONFIG_SCHED_DEBUG
5054 sysrq_sched_debug_show();
5058 * Only show locks if all tasks are dumped:
5061 debug_show_all_locks();
5064 void init_idle_bootup_task(struct task_struct
*idle
)
5066 idle
->sched_class
= &idle_sched_class
;
5070 * init_idle - set up an idle thread for a given CPU
5071 * @idle: task in question
5072 * @cpu: cpu the idle task belongs to
5074 * NOTE: this function does not set the idle thread's NEED_RESCHED
5075 * flag, to make booting more robust.
5077 void init_idle(struct task_struct
*idle
, int cpu
)
5079 struct rq
*rq
= cpu_rq(cpu
);
5080 unsigned long flags
;
5082 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5083 raw_spin_lock(&rq
->lock
);
5085 __sched_fork(0, idle
);
5086 idle
->state
= TASK_RUNNING
;
5087 idle
->se
.exec_start
= sched_clock();
5089 kasan_unpoison_task_stack(idle
);
5093 * Its possible that init_idle() gets called multiple times on a task,
5094 * in that case do_set_cpus_allowed() will not do the right thing.
5096 * And since this is boot we can forgo the serialization.
5098 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5101 * We're having a chicken and egg problem, even though we are
5102 * holding rq->lock, the cpu isn't yet set to this cpu so the
5103 * lockdep check in task_group() will fail.
5105 * Similar case to sched_fork(). / Alternatively we could
5106 * use task_rq_lock() here and obtain the other rq->lock.
5111 __set_task_cpu(idle
, cpu
);
5114 rq
->curr
= rq
->idle
= idle
;
5115 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5119 raw_spin_unlock(&rq
->lock
);
5120 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5122 /* Set the preempt count _outside_ the spinlocks! */
5123 init_idle_preempt_count(idle
, cpu
);
5126 * The idle tasks have their own, simple scheduling class:
5128 idle
->sched_class
= &idle_sched_class
;
5129 ftrace_graph_init_idle_task(idle
, cpu
);
5130 vtime_init_idle(idle
, cpu
);
5132 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5136 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5137 const struct cpumask
*trial
)
5139 int ret
= 1, trial_cpus
;
5140 struct dl_bw
*cur_dl_b
;
5141 unsigned long flags
;
5143 if (!cpumask_weight(cur
))
5146 rcu_read_lock_sched();
5147 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5148 trial_cpus
= cpumask_weight(trial
);
5150 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5151 if (cur_dl_b
->bw
!= -1 &&
5152 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5154 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5155 rcu_read_unlock_sched();
5160 int task_can_attach(struct task_struct
*p
,
5161 const struct cpumask
*cs_cpus_allowed
)
5166 * Kthreads which disallow setaffinity shouldn't be moved
5167 * to a new cpuset; we don't want to change their cpu
5168 * affinity and isolating such threads by their set of
5169 * allowed nodes is unnecessary. Thus, cpusets are not
5170 * applicable for such threads. This prevents checking for
5171 * success of set_cpus_allowed_ptr() on all attached tasks
5172 * before cpus_allowed may be changed.
5174 if (p
->flags
& PF_NO_SETAFFINITY
) {
5180 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5182 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5187 unsigned long flags
;
5189 rcu_read_lock_sched();
5190 dl_b
= dl_bw_of(dest_cpu
);
5191 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5192 cpus
= dl_bw_cpus(dest_cpu
);
5193 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5198 * We reserve space for this task in the destination
5199 * root_domain, as we can't fail after this point.
5200 * We will free resources in the source root_domain
5201 * later on (see set_cpus_allowed_dl()).
5203 __dl_add(dl_b
, p
->dl
.dl_bw
);
5205 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5206 rcu_read_unlock_sched();
5216 #ifdef CONFIG_NUMA_BALANCING
5217 /* Migrate current task p to target_cpu */
5218 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5220 struct migration_arg arg
= { p
, target_cpu
};
5221 int curr_cpu
= task_cpu(p
);
5223 if (curr_cpu
== target_cpu
)
5226 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5229 /* TODO: This is not properly updating schedstats */
5231 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5232 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5236 * Requeue a task on a given node and accurately track the number of NUMA
5237 * tasks on the runqueues
5239 void sched_setnuma(struct task_struct
*p
, int nid
)
5242 unsigned long flags
;
5243 bool queued
, running
;
5245 rq
= task_rq_lock(p
, &flags
);
5246 queued
= task_on_rq_queued(p
);
5247 running
= task_current(rq
, p
);
5250 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5252 put_prev_task(rq
, p
);
5254 p
->numa_preferred_nid
= nid
;
5257 p
->sched_class
->set_curr_task(rq
);
5259 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5260 task_rq_unlock(rq
, p
, &flags
);
5262 #endif /* CONFIG_NUMA_BALANCING */
5264 #ifdef CONFIG_HOTPLUG_CPU
5266 * Ensures that the idle task is using init_mm right before its cpu goes
5269 void idle_task_exit(void)
5271 struct mm_struct
*mm
= current
->active_mm
;
5273 BUG_ON(cpu_online(smp_processor_id()));
5275 if (mm
!= &init_mm
) {
5276 switch_mm(mm
, &init_mm
, current
);
5277 finish_arch_post_lock_switch();
5283 * Since this CPU is going 'away' for a while, fold any nr_active delta
5284 * we might have. Assumes we're called after migrate_tasks() so that the
5285 * nr_active count is stable.
5287 * Also see the comment "Global load-average calculations".
5289 static void calc_load_migrate(struct rq
*rq
)
5291 long delta
= calc_load_fold_active(rq
);
5293 atomic_long_add(delta
, &calc_load_tasks
);
5296 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5300 static const struct sched_class fake_sched_class
= {
5301 .put_prev_task
= put_prev_task_fake
,
5304 static struct task_struct fake_task
= {
5306 * Avoid pull_{rt,dl}_task()
5308 .prio
= MAX_PRIO
+ 1,
5309 .sched_class
= &fake_sched_class
,
5313 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5314 * try_to_wake_up()->select_task_rq().
5316 * Called with rq->lock held even though we'er in stop_machine() and
5317 * there's no concurrency possible, we hold the required locks anyway
5318 * because of lock validation efforts.
5320 static void migrate_tasks(struct rq
*dead_rq
)
5322 struct rq
*rq
= dead_rq
;
5323 struct task_struct
*next
, *stop
= rq
->stop
;
5327 * Fudge the rq selection such that the below task selection loop
5328 * doesn't get stuck on the currently eligible stop task.
5330 * We're currently inside stop_machine() and the rq is either stuck
5331 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5332 * either way we should never end up calling schedule() until we're
5338 * put_prev_task() and pick_next_task() sched
5339 * class method both need to have an up-to-date
5340 * value of rq->clock[_task]
5342 update_rq_clock(rq
);
5346 * There's this thread running, bail when that's the only
5349 if (rq
->nr_running
== 1)
5353 * pick_next_task assumes pinned rq->lock.
5355 lockdep_pin_lock(&rq
->lock
);
5356 next
= pick_next_task(rq
, &fake_task
);
5358 next
->sched_class
->put_prev_task(rq
, next
);
5361 * Rules for changing task_struct::cpus_allowed are holding
5362 * both pi_lock and rq->lock, such that holding either
5363 * stabilizes the mask.
5365 * Drop rq->lock is not quite as disastrous as it usually is
5366 * because !cpu_active at this point, which means load-balance
5367 * will not interfere. Also, stop-machine.
5369 lockdep_unpin_lock(&rq
->lock
);
5370 raw_spin_unlock(&rq
->lock
);
5371 raw_spin_lock(&next
->pi_lock
);
5372 raw_spin_lock(&rq
->lock
);
5375 * Since we're inside stop-machine, _nothing_ should have
5376 * changed the task, WARN if weird stuff happened, because in
5377 * that case the above rq->lock drop is a fail too.
5379 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5380 raw_spin_unlock(&next
->pi_lock
);
5384 /* Find suitable destination for @next, with force if needed. */
5385 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5387 rq
= __migrate_task(rq
, next
, dest_cpu
);
5388 if (rq
!= dead_rq
) {
5389 raw_spin_unlock(&rq
->lock
);
5391 raw_spin_lock(&rq
->lock
);
5393 raw_spin_unlock(&next
->pi_lock
);
5398 #endif /* CONFIG_HOTPLUG_CPU */
5400 static void set_rq_online(struct rq
*rq
)
5403 const struct sched_class
*class;
5405 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5408 for_each_class(class) {
5409 if (class->rq_online
)
5410 class->rq_online(rq
);
5415 static void set_rq_offline(struct rq
*rq
)
5418 const struct sched_class
*class;
5420 for_each_class(class) {
5421 if (class->rq_offline
)
5422 class->rq_offline(rq
);
5425 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5431 * migration_call - callback that gets triggered when a CPU is added.
5432 * Here we can start up the necessary migration thread for the new CPU.
5435 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5437 int cpu
= (long)hcpu
;
5438 unsigned long flags
;
5439 struct rq
*rq
= cpu_rq(cpu
);
5441 switch (action
& ~CPU_TASKS_FROZEN
) {
5443 case CPU_UP_PREPARE
:
5444 rq
->calc_load_update
= calc_load_update
;
5445 account_reset_rq(rq
);
5449 /* Update our root-domain */
5450 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5452 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5456 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5459 #ifdef CONFIG_HOTPLUG_CPU
5461 sched_ttwu_pending();
5462 /* Update our root-domain */
5463 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5465 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5469 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5470 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5474 calc_load_migrate(rq
);
5479 update_max_interval();
5485 * Register at high priority so that task migration (migrate_all_tasks)
5486 * happens before everything else. This has to be lower priority than
5487 * the notifier in the perf_event subsystem, though.
5489 static struct notifier_block migration_notifier
= {
5490 .notifier_call
= migration_call
,
5491 .priority
= CPU_PRI_MIGRATION
,
5494 static void set_cpu_rq_start_time(void)
5496 int cpu
= smp_processor_id();
5497 struct rq
*rq
= cpu_rq(cpu
);
5498 rq
->age_stamp
= sched_clock_cpu(cpu
);
5501 static int sched_cpu_active(struct notifier_block
*nfb
,
5502 unsigned long action
, void *hcpu
)
5504 int cpu
= (long)hcpu
;
5506 switch (action
& ~CPU_TASKS_FROZEN
) {
5508 set_cpu_rq_start_time();
5511 case CPU_DOWN_FAILED
:
5512 set_cpu_active(cpu
, true);
5520 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5521 unsigned long action
, void *hcpu
)
5523 switch (action
& ~CPU_TASKS_FROZEN
) {
5524 case CPU_DOWN_PREPARE
:
5525 set_cpu_active((long)hcpu
, false);
5532 static int __init
migration_init(void)
5534 void *cpu
= (void *)(long)smp_processor_id();
5537 /* Initialize migration for the boot CPU */
5538 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5539 BUG_ON(err
== NOTIFY_BAD
);
5540 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5541 register_cpu_notifier(&migration_notifier
);
5543 /* Register cpu active notifiers */
5544 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5545 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5549 early_initcall(migration_init
);
5551 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5553 #ifdef CONFIG_SCHED_DEBUG
5555 static __read_mostly
int sched_debug_enabled
;
5557 static int __init
sched_debug_setup(char *str
)
5559 sched_debug_enabled
= 1;
5563 early_param("sched_debug", sched_debug_setup
);
5565 static inline bool sched_debug(void)
5567 return sched_debug_enabled
;
5570 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5571 struct cpumask
*groupmask
)
5573 struct sched_group
*group
= sd
->groups
;
5575 cpumask_clear(groupmask
);
5577 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5579 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5580 printk("does not load-balance\n");
5582 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5587 printk(KERN_CONT
"span %*pbl level %s\n",
5588 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5590 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5591 printk(KERN_ERR
"ERROR: domain->span does not contain "
5594 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5595 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5599 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5603 printk(KERN_ERR
"ERROR: group is NULL\n");
5607 if (!cpumask_weight(sched_group_cpus(group
))) {
5608 printk(KERN_CONT
"\n");
5609 printk(KERN_ERR
"ERROR: empty group\n");
5613 if (!(sd
->flags
& SD_OVERLAP
) &&
5614 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5615 printk(KERN_CONT
"\n");
5616 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5620 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5622 printk(KERN_CONT
" %*pbl",
5623 cpumask_pr_args(sched_group_cpus(group
)));
5624 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5625 printk(KERN_CONT
" (cpu_capacity = %d)",
5626 group
->sgc
->capacity
);
5629 group
= group
->next
;
5630 } while (group
!= sd
->groups
);
5631 printk(KERN_CONT
"\n");
5633 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5634 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5637 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5638 printk(KERN_ERR
"ERROR: parent span is not a superset "
5639 "of domain->span\n");
5643 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5647 if (!sched_debug_enabled
)
5651 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5655 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5658 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5666 #else /* !CONFIG_SCHED_DEBUG */
5667 # define sched_domain_debug(sd, cpu) do { } while (0)
5668 static inline bool sched_debug(void)
5672 #endif /* CONFIG_SCHED_DEBUG */
5674 static int sd_degenerate(struct sched_domain
*sd
)
5676 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5679 /* Following flags need at least 2 groups */
5680 if (sd
->flags
& (SD_LOAD_BALANCE
|
5681 SD_BALANCE_NEWIDLE
|
5684 SD_SHARE_CPUCAPACITY
|
5685 SD_SHARE_PKG_RESOURCES
|
5686 SD_SHARE_POWERDOMAIN
)) {
5687 if (sd
->groups
!= sd
->groups
->next
)
5691 /* Following flags don't use groups */
5692 if (sd
->flags
& (SD_WAKE_AFFINE
))
5699 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5701 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5703 if (sd_degenerate(parent
))
5706 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5709 /* Flags needing groups don't count if only 1 group in parent */
5710 if (parent
->groups
== parent
->groups
->next
) {
5711 pflags
&= ~(SD_LOAD_BALANCE
|
5712 SD_BALANCE_NEWIDLE
|
5715 SD_SHARE_CPUCAPACITY
|
5716 SD_SHARE_PKG_RESOURCES
|
5718 SD_SHARE_POWERDOMAIN
);
5719 if (nr_node_ids
== 1)
5720 pflags
&= ~SD_SERIALIZE
;
5722 if (~cflags
& pflags
)
5728 static void free_rootdomain(struct rcu_head
*rcu
)
5730 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5732 cpupri_cleanup(&rd
->cpupri
);
5733 cpudl_cleanup(&rd
->cpudl
);
5734 free_cpumask_var(rd
->dlo_mask
);
5735 free_cpumask_var(rd
->rto_mask
);
5736 free_cpumask_var(rd
->online
);
5737 free_cpumask_var(rd
->span
);
5741 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5743 struct root_domain
*old_rd
= NULL
;
5744 unsigned long flags
;
5746 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5751 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5754 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5757 * If we dont want to free the old_rd yet then
5758 * set old_rd to NULL to skip the freeing later
5761 if (!atomic_dec_and_test(&old_rd
->refcount
))
5765 atomic_inc(&rd
->refcount
);
5768 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5769 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5772 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5775 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5778 static int init_rootdomain(struct root_domain
*rd
)
5780 memset(rd
, 0, sizeof(*rd
));
5782 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5784 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5786 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5788 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5791 init_dl_bw(&rd
->dl_bw
);
5792 if (cpudl_init(&rd
->cpudl
) != 0)
5795 if (cpupri_init(&rd
->cpupri
) != 0)
5800 free_cpumask_var(rd
->rto_mask
);
5802 free_cpumask_var(rd
->dlo_mask
);
5804 free_cpumask_var(rd
->online
);
5806 free_cpumask_var(rd
->span
);
5812 * By default the system creates a single root-domain with all cpus as
5813 * members (mimicking the global state we have today).
5815 struct root_domain def_root_domain
;
5817 static void init_defrootdomain(void)
5819 init_rootdomain(&def_root_domain
);
5821 atomic_set(&def_root_domain
.refcount
, 1);
5824 static struct root_domain
*alloc_rootdomain(void)
5826 struct root_domain
*rd
;
5828 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5832 if (init_rootdomain(rd
) != 0) {
5840 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5842 struct sched_group
*tmp
, *first
;
5851 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5856 } while (sg
!= first
);
5859 static void free_sched_domain(struct rcu_head
*rcu
)
5861 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5864 * If its an overlapping domain it has private groups, iterate and
5867 if (sd
->flags
& SD_OVERLAP
) {
5868 free_sched_groups(sd
->groups
, 1);
5869 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5870 kfree(sd
->groups
->sgc
);
5876 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5878 call_rcu(&sd
->rcu
, free_sched_domain
);
5881 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5883 for (; sd
; sd
= sd
->parent
)
5884 destroy_sched_domain(sd
, cpu
);
5888 * Keep a special pointer to the highest sched_domain that has
5889 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5890 * allows us to avoid some pointer chasing select_idle_sibling().
5892 * Also keep a unique ID per domain (we use the first cpu number in
5893 * the cpumask of the domain), this allows us to quickly tell if
5894 * two cpus are in the same cache domain, see cpus_share_cache().
5896 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5897 DEFINE_PER_CPU(int, sd_llc_size
);
5898 DEFINE_PER_CPU(int, sd_llc_id
);
5899 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5900 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5901 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5903 static void update_top_cache_domain(int cpu
)
5905 struct sched_domain
*sd
;
5906 struct sched_domain
*busy_sd
= NULL
;
5910 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5912 id
= cpumask_first(sched_domain_span(sd
));
5913 size
= cpumask_weight(sched_domain_span(sd
));
5914 busy_sd
= sd
->parent
; /* sd_busy */
5916 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5918 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5919 per_cpu(sd_llc_size
, cpu
) = size
;
5920 per_cpu(sd_llc_id
, cpu
) = id
;
5922 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5923 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5925 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5926 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5930 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5931 * hold the hotplug lock.
5934 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5936 struct rq
*rq
= cpu_rq(cpu
);
5937 struct sched_domain
*tmp
;
5939 /* Remove the sched domains which do not contribute to scheduling. */
5940 for (tmp
= sd
; tmp
; ) {
5941 struct sched_domain
*parent
= tmp
->parent
;
5945 if (sd_parent_degenerate(tmp
, parent
)) {
5946 tmp
->parent
= parent
->parent
;
5948 parent
->parent
->child
= tmp
;
5950 * Transfer SD_PREFER_SIBLING down in case of a
5951 * degenerate parent; the spans match for this
5952 * so the property transfers.
5954 if (parent
->flags
& SD_PREFER_SIBLING
)
5955 tmp
->flags
|= SD_PREFER_SIBLING
;
5956 destroy_sched_domain(parent
, cpu
);
5961 if (sd
&& sd_degenerate(sd
)) {
5964 destroy_sched_domain(tmp
, cpu
);
5969 sched_domain_debug(sd
, cpu
);
5971 rq_attach_root(rq
, rd
);
5973 rcu_assign_pointer(rq
->sd
, sd
);
5974 destroy_sched_domains(tmp
, cpu
);
5976 update_top_cache_domain(cpu
);
5979 /* Setup the mask of cpus configured for isolated domains */
5980 static int __init
isolated_cpu_setup(char *str
)
5984 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5985 ret
= cpulist_parse(str
, cpu_isolated_map
);
5987 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5992 __setup("isolcpus=", isolated_cpu_setup
);
5995 struct sched_domain
** __percpu sd
;
5996 struct root_domain
*rd
;
6007 * Build an iteration mask that can exclude certain CPUs from the upwards
6010 * Asymmetric node setups can result in situations where the domain tree is of
6011 * unequal depth, make sure to skip domains that already cover the entire
6014 * In that case build_sched_domains() will have terminated the iteration early
6015 * and our sibling sd spans will be empty. Domains should always include the
6016 * cpu they're built on, so check that.
6019 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6021 const struct cpumask
*span
= sched_domain_span(sd
);
6022 struct sd_data
*sdd
= sd
->private;
6023 struct sched_domain
*sibling
;
6026 for_each_cpu(i
, span
) {
6027 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6028 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6031 cpumask_set_cpu(i
, sched_group_mask(sg
));
6036 * Return the canonical balance cpu for this group, this is the first cpu
6037 * of this group that's also in the iteration mask.
6039 int group_balance_cpu(struct sched_group
*sg
)
6041 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6045 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6047 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6048 const struct cpumask
*span
= sched_domain_span(sd
);
6049 struct cpumask
*covered
= sched_domains_tmpmask
;
6050 struct sd_data
*sdd
= sd
->private;
6051 struct sched_domain
*sibling
;
6054 cpumask_clear(covered
);
6056 for_each_cpu(i
, span
) {
6057 struct cpumask
*sg_span
;
6059 if (cpumask_test_cpu(i
, covered
))
6062 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6064 /* See the comment near build_group_mask(). */
6065 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6068 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6069 GFP_KERNEL
, cpu_to_node(cpu
));
6074 sg_span
= sched_group_cpus(sg
);
6076 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6078 cpumask_set_cpu(i
, sg_span
);
6080 cpumask_or(covered
, covered
, sg_span
);
6082 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6083 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6084 build_group_mask(sd
, sg
);
6087 * Initialize sgc->capacity such that even if we mess up the
6088 * domains and no possible iteration will get us here, we won't
6091 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6094 * Make sure the first group of this domain contains the
6095 * canonical balance cpu. Otherwise the sched_domain iteration
6096 * breaks. See update_sg_lb_stats().
6098 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6099 group_balance_cpu(sg
) == cpu
)
6109 sd
->groups
= groups
;
6114 free_sched_groups(first
, 0);
6119 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6121 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6122 struct sched_domain
*child
= sd
->child
;
6125 cpu
= cpumask_first(sched_domain_span(child
));
6128 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6129 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6130 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6137 * build_sched_groups will build a circular linked list of the groups
6138 * covered by the given span, and will set each group's ->cpumask correctly,
6139 * and ->cpu_capacity to 0.
6141 * Assumes the sched_domain tree is fully constructed
6144 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6146 struct sched_group
*first
= NULL
, *last
= NULL
;
6147 struct sd_data
*sdd
= sd
->private;
6148 const struct cpumask
*span
= sched_domain_span(sd
);
6149 struct cpumask
*covered
;
6152 get_group(cpu
, sdd
, &sd
->groups
);
6153 atomic_inc(&sd
->groups
->ref
);
6155 if (cpu
!= cpumask_first(span
))
6158 lockdep_assert_held(&sched_domains_mutex
);
6159 covered
= sched_domains_tmpmask
;
6161 cpumask_clear(covered
);
6163 for_each_cpu(i
, span
) {
6164 struct sched_group
*sg
;
6167 if (cpumask_test_cpu(i
, covered
))
6170 group
= get_group(i
, sdd
, &sg
);
6171 cpumask_setall(sched_group_mask(sg
));
6173 for_each_cpu(j
, span
) {
6174 if (get_group(j
, sdd
, NULL
) != group
)
6177 cpumask_set_cpu(j
, covered
);
6178 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6193 * Initialize sched groups cpu_capacity.
6195 * cpu_capacity indicates the capacity of sched group, which is used while
6196 * distributing the load between different sched groups in a sched domain.
6197 * Typically cpu_capacity for all the groups in a sched domain will be same
6198 * unless there are asymmetries in the topology. If there are asymmetries,
6199 * group having more cpu_capacity will pickup more load compared to the
6200 * group having less cpu_capacity.
6202 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6204 struct sched_group
*sg
= sd
->groups
;
6209 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6211 } while (sg
!= sd
->groups
);
6213 if (cpu
!= group_balance_cpu(sg
))
6216 update_group_capacity(sd
, cpu
);
6217 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6221 * Initializers for schedule domains
6222 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6225 static int default_relax_domain_level
= -1;
6226 int sched_domain_level_max
;
6228 static int __init
setup_relax_domain_level(char *str
)
6230 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6231 pr_warn("Unable to set relax_domain_level\n");
6235 __setup("relax_domain_level=", setup_relax_domain_level
);
6237 static void set_domain_attribute(struct sched_domain
*sd
,
6238 struct sched_domain_attr
*attr
)
6242 if (!attr
|| attr
->relax_domain_level
< 0) {
6243 if (default_relax_domain_level
< 0)
6246 request
= default_relax_domain_level
;
6248 request
= attr
->relax_domain_level
;
6249 if (request
< sd
->level
) {
6250 /* turn off idle balance on this domain */
6251 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6253 /* turn on idle balance on this domain */
6254 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6258 static void __sdt_free(const struct cpumask
*cpu_map
);
6259 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6261 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6262 const struct cpumask
*cpu_map
)
6266 if (!atomic_read(&d
->rd
->refcount
))
6267 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6269 free_percpu(d
->sd
); /* fall through */
6271 __sdt_free(cpu_map
); /* fall through */
6277 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6278 const struct cpumask
*cpu_map
)
6280 memset(d
, 0, sizeof(*d
));
6282 if (__sdt_alloc(cpu_map
))
6283 return sa_sd_storage
;
6284 d
->sd
= alloc_percpu(struct sched_domain
*);
6286 return sa_sd_storage
;
6287 d
->rd
= alloc_rootdomain();
6290 return sa_rootdomain
;
6294 * NULL the sd_data elements we've used to build the sched_domain and
6295 * sched_group structure so that the subsequent __free_domain_allocs()
6296 * will not free the data we're using.
6298 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6300 struct sd_data
*sdd
= sd
->private;
6302 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6303 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6305 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6306 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6308 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6309 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6313 static int sched_domains_numa_levels
;
6314 enum numa_topology_type sched_numa_topology_type
;
6315 static int *sched_domains_numa_distance
;
6316 int sched_max_numa_distance
;
6317 static struct cpumask
***sched_domains_numa_masks
;
6318 static int sched_domains_curr_level
;
6322 * SD_flags allowed in topology descriptions.
6324 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6325 * SD_SHARE_PKG_RESOURCES - describes shared caches
6326 * SD_NUMA - describes NUMA topologies
6327 * SD_SHARE_POWERDOMAIN - describes shared power domain
6330 * SD_ASYM_PACKING - describes SMT quirks
6332 #define TOPOLOGY_SD_FLAGS \
6333 (SD_SHARE_CPUCAPACITY | \
6334 SD_SHARE_PKG_RESOURCES | \
6337 SD_SHARE_POWERDOMAIN)
6339 static struct sched_domain
*
6340 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6342 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6343 int sd_weight
, sd_flags
= 0;
6347 * Ugly hack to pass state to sd_numa_mask()...
6349 sched_domains_curr_level
= tl
->numa_level
;
6352 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6355 sd_flags
= (*tl
->sd_flags
)();
6356 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6357 "wrong sd_flags in topology description\n"))
6358 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6360 *sd
= (struct sched_domain
){
6361 .min_interval
= sd_weight
,
6362 .max_interval
= 2*sd_weight
,
6364 .imbalance_pct
= 125,
6366 .cache_nice_tries
= 0,
6373 .flags
= 1*SD_LOAD_BALANCE
6374 | 1*SD_BALANCE_NEWIDLE
6379 | 0*SD_SHARE_CPUCAPACITY
6380 | 0*SD_SHARE_PKG_RESOURCES
6382 | 0*SD_PREFER_SIBLING
6387 .last_balance
= jiffies
,
6388 .balance_interval
= sd_weight
,
6390 .max_newidle_lb_cost
= 0,
6391 .next_decay_max_lb_cost
= jiffies
,
6392 #ifdef CONFIG_SCHED_DEBUG
6398 * Convert topological properties into behaviour.
6401 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6402 sd
->flags
|= SD_PREFER_SIBLING
;
6403 sd
->imbalance_pct
= 110;
6404 sd
->smt_gain
= 1178; /* ~15% */
6406 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6407 sd
->imbalance_pct
= 117;
6408 sd
->cache_nice_tries
= 1;
6412 } else if (sd
->flags
& SD_NUMA
) {
6413 sd
->cache_nice_tries
= 2;
6417 sd
->flags
|= SD_SERIALIZE
;
6418 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6419 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6426 sd
->flags
|= SD_PREFER_SIBLING
;
6427 sd
->cache_nice_tries
= 1;
6432 sd
->private = &tl
->data
;
6438 * Topology list, bottom-up.
6440 static struct sched_domain_topology_level default_topology
[] = {
6441 #ifdef CONFIG_SCHED_SMT
6442 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6444 #ifdef CONFIG_SCHED_MC
6445 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6447 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6451 static struct sched_domain_topology_level
*sched_domain_topology
=
6454 #define for_each_sd_topology(tl) \
6455 for (tl = sched_domain_topology; tl->mask; tl++)
6457 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6459 sched_domain_topology
= tl
;
6464 static const struct cpumask
*sd_numa_mask(int cpu
)
6466 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6469 static void sched_numa_warn(const char *str
)
6471 static int done
= false;
6479 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6481 for (i
= 0; i
< nr_node_ids
; i
++) {
6482 printk(KERN_WARNING
" ");
6483 for (j
= 0; j
< nr_node_ids
; j
++)
6484 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6485 printk(KERN_CONT
"\n");
6487 printk(KERN_WARNING
"\n");
6490 bool find_numa_distance(int distance
)
6494 if (distance
== node_distance(0, 0))
6497 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6498 if (sched_domains_numa_distance
[i
] == distance
)
6506 * A system can have three types of NUMA topology:
6507 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6508 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6509 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6511 * The difference between a glueless mesh topology and a backplane
6512 * topology lies in whether communication between not directly
6513 * connected nodes goes through intermediary nodes (where programs
6514 * could run), or through backplane controllers. This affects
6515 * placement of programs.
6517 * The type of topology can be discerned with the following tests:
6518 * - If the maximum distance between any nodes is 1 hop, the system
6519 * is directly connected.
6520 * - If for two nodes A and B, located N > 1 hops away from each other,
6521 * there is an intermediary node C, which is < N hops away from both
6522 * nodes A and B, the system is a glueless mesh.
6524 static void init_numa_topology_type(void)
6528 n
= sched_max_numa_distance
;
6530 if (sched_domains_numa_levels
<= 1) {
6531 sched_numa_topology_type
= NUMA_DIRECT
;
6535 for_each_online_node(a
) {
6536 for_each_online_node(b
) {
6537 /* Find two nodes furthest removed from each other. */
6538 if (node_distance(a
, b
) < n
)
6541 /* Is there an intermediary node between a and b? */
6542 for_each_online_node(c
) {
6543 if (node_distance(a
, c
) < n
&&
6544 node_distance(b
, c
) < n
) {
6545 sched_numa_topology_type
=
6551 sched_numa_topology_type
= NUMA_BACKPLANE
;
6557 static void sched_init_numa(void)
6559 int next_distance
, curr_distance
= node_distance(0, 0);
6560 struct sched_domain_topology_level
*tl
;
6564 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6565 if (!sched_domains_numa_distance
)
6569 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6570 * unique distances in the node_distance() table.
6572 * Assumes node_distance(0,j) includes all distances in
6573 * node_distance(i,j) in order to avoid cubic time.
6575 next_distance
= curr_distance
;
6576 for (i
= 0; i
< nr_node_ids
; i
++) {
6577 for (j
= 0; j
< nr_node_ids
; j
++) {
6578 for (k
= 0; k
< nr_node_ids
; k
++) {
6579 int distance
= node_distance(i
, k
);
6581 if (distance
> curr_distance
&&
6582 (distance
< next_distance
||
6583 next_distance
== curr_distance
))
6584 next_distance
= distance
;
6587 * While not a strong assumption it would be nice to know
6588 * about cases where if node A is connected to B, B is not
6589 * equally connected to A.
6591 if (sched_debug() && node_distance(k
, i
) != distance
)
6592 sched_numa_warn("Node-distance not symmetric");
6594 if (sched_debug() && i
&& !find_numa_distance(distance
))
6595 sched_numa_warn("Node-0 not representative");
6597 if (next_distance
!= curr_distance
) {
6598 sched_domains_numa_distance
[level
++] = next_distance
;
6599 sched_domains_numa_levels
= level
;
6600 curr_distance
= next_distance
;
6605 * In case of sched_debug() we verify the above assumption.
6615 * 'level' contains the number of unique distances, excluding the
6616 * identity distance node_distance(i,i).
6618 * The sched_domains_numa_distance[] array includes the actual distance
6623 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6624 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6625 * the array will contain less then 'level' members. This could be
6626 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6627 * in other functions.
6629 * We reset it to 'level' at the end of this function.
6631 sched_domains_numa_levels
= 0;
6633 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6634 if (!sched_domains_numa_masks
)
6638 * Now for each level, construct a mask per node which contains all
6639 * cpus of nodes that are that many hops away from us.
6641 for (i
= 0; i
< level
; i
++) {
6642 sched_domains_numa_masks
[i
] =
6643 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6644 if (!sched_domains_numa_masks
[i
])
6647 for (j
= 0; j
< nr_node_ids
; j
++) {
6648 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6652 sched_domains_numa_masks
[i
][j
] = mask
;
6655 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6658 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6663 /* Compute default topology size */
6664 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6666 tl
= kzalloc((i
+ level
+ 1) *
6667 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6672 * Copy the default topology bits..
6674 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6675 tl
[i
] = sched_domain_topology
[i
];
6678 * .. and append 'j' levels of NUMA goodness.
6680 for (j
= 0; j
< level
; i
++, j
++) {
6681 tl
[i
] = (struct sched_domain_topology_level
){
6682 .mask
= sd_numa_mask
,
6683 .sd_flags
= cpu_numa_flags
,
6684 .flags
= SDTL_OVERLAP
,
6690 sched_domain_topology
= tl
;
6692 sched_domains_numa_levels
= level
;
6693 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6695 init_numa_topology_type();
6698 static void sched_domains_numa_masks_set(int cpu
)
6701 int node
= cpu_to_node(cpu
);
6703 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6704 for (j
= 0; j
< nr_node_ids
; j
++) {
6705 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6706 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6711 static void sched_domains_numa_masks_clear(int cpu
)
6714 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6715 for (j
= 0; j
< nr_node_ids
; j
++)
6716 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6721 * Update sched_domains_numa_masks[level][node] array when new cpus
6724 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6725 unsigned long action
,
6728 int cpu
= (long)hcpu
;
6730 switch (action
& ~CPU_TASKS_FROZEN
) {
6732 sched_domains_numa_masks_set(cpu
);
6736 sched_domains_numa_masks_clear(cpu
);
6746 static inline void sched_init_numa(void)
6750 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6751 unsigned long action
,
6756 #endif /* CONFIG_NUMA */
6758 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6760 struct sched_domain_topology_level
*tl
;
6763 for_each_sd_topology(tl
) {
6764 struct sd_data
*sdd
= &tl
->data
;
6766 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6770 sdd
->sg
= alloc_percpu(struct sched_group
*);
6774 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6778 for_each_cpu(j
, cpu_map
) {
6779 struct sched_domain
*sd
;
6780 struct sched_group
*sg
;
6781 struct sched_group_capacity
*sgc
;
6783 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6784 GFP_KERNEL
, cpu_to_node(j
));
6788 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6790 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6791 GFP_KERNEL
, cpu_to_node(j
));
6797 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6799 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6800 GFP_KERNEL
, cpu_to_node(j
));
6804 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6811 static void __sdt_free(const struct cpumask
*cpu_map
)
6813 struct sched_domain_topology_level
*tl
;
6816 for_each_sd_topology(tl
) {
6817 struct sd_data
*sdd
= &tl
->data
;
6819 for_each_cpu(j
, cpu_map
) {
6820 struct sched_domain
*sd
;
6823 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6824 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6825 free_sched_groups(sd
->groups
, 0);
6826 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6830 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6832 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6834 free_percpu(sdd
->sd
);
6836 free_percpu(sdd
->sg
);
6838 free_percpu(sdd
->sgc
);
6843 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6844 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6845 struct sched_domain
*child
, int cpu
)
6847 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6851 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6853 sd
->level
= child
->level
+ 1;
6854 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6858 if (!cpumask_subset(sched_domain_span(child
),
6859 sched_domain_span(sd
))) {
6860 pr_err("BUG: arch topology borken\n");
6861 #ifdef CONFIG_SCHED_DEBUG
6862 pr_err(" the %s domain not a subset of the %s domain\n",
6863 child
->name
, sd
->name
);
6865 /* Fixup, ensure @sd has at least @child cpus. */
6866 cpumask_or(sched_domain_span(sd
),
6867 sched_domain_span(sd
),
6868 sched_domain_span(child
));
6872 set_domain_attribute(sd
, attr
);
6878 * Build sched domains for a given set of cpus and attach the sched domains
6879 * to the individual cpus
6881 static int build_sched_domains(const struct cpumask
*cpu_map
,
6882 struct sched_domain_attr
*attr
)
6884 enum s_alloc alloc_state
;
6885 struct sched_domain
*sd
;
6887 int i
, ret
= -ENOMEM
;
6889 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6890 if (alloc_state
!= sa_rootdomain
)
6893 /* Set up domains for cpus specified by the cpu_map. */
6894 for_each_cpu(i
, cpu_map
) {
6895 struct sched_domain_topology_level
*tl
;
6898 for_each_sd_topology(tl
) {
6899 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6900 if (tl
== sched_domain_topology
)
6901 *per_cpu_ptr(d
.sd
, i
) = sd
;
6902 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6903 sd
->flags
|= SD_OVERLAP
;
6904 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6909 /* Build the groups for the domains */
6910 for_each_cpu(i
, cpu_map
) {
6911 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6912 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6913 if (sd
->flags
& SD_OVERLAP
) {
6914 if (build_overlap_sched_groups(sd
, i
))
6917 if (build_sched_groups(sd
, i
))
6923 /* Calculate CPU capacity for physical packages and nodes */
6924 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6925 if (!cpumask_test_cpu(i
, cpu_map
))
6928 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6929 claim_allocations(i
, sd
);
6930 init_sched_groups_capacity(i
, sd
);
6934 /* Attach the domains */
6936 for_each_cpu(i
, cpu_map
) {
6937 sd
= *per_cpu_ptr(d
.sd
, i
);
6938 cpu_attach_domain(sd
, d
.rd
, i
);
6944 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6948 static cpumask_var_t
*doms_cur
; /* current sched domains */
6949 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6950 static struct sched_domain_attr
*dattr_cur
;
6951 /* attribues of custom domains in 'doms_cur' */
6954 * Special case: If a kmalloc of a doms_cur partition (array of
6955 * cpumask) fails, then fallback to a single sched domain,
6956 * as determined by the single cpumask fallback_doms.
6958 static cpumask_var_t fallback_doms
;
6961 * arch_update_cpu_topology lets virtualized architectures update the
6962 * cpu core maps. It is supposed to return 1 if the topology changed
6963 * or 0 if it stayed the same.
6965 int __weak
arch_update_cpu_topology(void)
6970 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6973 cpumask_var_t
*doms
;
6975 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6978 for (i
= 0; i
< ndoms
; i
++) {
6979 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6980 free_sched_domains(doms
, i
);
6987 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6990 for (i
= 0; i
< ndoms
; i
++)
6991 free_cpumask_var(doms
[i
]);
6996 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6997 * For now this just excludes isolated cpus, but could be used to
6998 * exclude other special cases in the future.
7000 static int init_sched_domains(const struct cpumask
*cpu_map
)
7004 arch_update_cpu_topology();
7006 doms_cur
= alloc_sched_domains(ndoms_cur
);
7008 doms_cur
= &fallback_doms
;
7009 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7010 err
= build_sched_domains(doms_cur
[0], NULL
);
7011 register_sched_domain_sysctl();
7017 * Detach sched domains from a group of cpus specified in cpu_map
7018 * These cpus will now be attached to the NULL domain
7020 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7025 for_each_cpu(i
, cpu_map
)
7026 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7030 /* handle null as "default" */
7031 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7032 struct sched_domain_attr
*new, int idx_new
)
7034 struct sched_domain_attr tmp
;
7041 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7042 new ? (new + idx_new
) : &tmp
,
7043 sizeof(struct sched_domain_attr
));
7047 * Partition sched domains as specified by the 'ndoms_new'
7048 * cpumasks in the array doms_new[] of cpumasks. This compares
7049 * doms_new[] to the current sched domain partitioning, doms_cur[].
7050 * It destroys each deleted domain and builds each new domain.
7052 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7053 * The masks don't intersect (don't overlap.) We should setup one
7054 * sched domain for each mask. CPUs not in any of the cpumasks will
7055 * not be load balanced. If the same cpumask appears both in the
7056 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7059 * The passed in 'doms_new' should be allocated using
7060 * alloc_sched_domains. This routine takes ownership of it and will
7061 * free_sched_domains it when done with it. If the caller failed the
7062 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7063 * and partition_sched_domains() will fallback to the single partition
7064 * 'fallback_doms', it also forces the domains to be rebuilt.
7066 * If doms_new == NULL it will be replaced with cpu_online_mask.
7067 * ndoms_new == 0 is a special case for destroying existing domains,
7068 * and it will not create the default domain.
7070 * Call with hotplug lock held
7072 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7073 struct sched_domain_attr
*dattr_new
)
7078 mutex_lock(&sched_domains_mutex
);
7080 /* always unregister in case we don't destroy any domains */
7081 unregister_sched_domain_sysctl();
7083 /* Let architecture update cpu core mappings. */
7084 new_topology
= arch_update_cpu_topology();
7086 n
= doms_new
? ndoms_new
: 0;
7088 /* Destroy deleted domains */
7089 for (i
= 0; i
< ndoms_cur
; i
++) {
7090 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7091 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7092 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7095 /* no match - a current sched domain not in new doms_new[] */
7096 detach_destroy_domains(doms_cur
[i
]);
7102 if (doms_new
== NULL
) {
7104 doms_new
= &fallback_doms
;
7105 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7106 WARN_ON_ONCE(dattr_new
);
7109 /* Build new domains */
7110 for (i
= 0; i
< ndoms_new
; i
++) {
7111 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7112 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7113 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7116 /* no match - add a new doms_new */
7117 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7122 /* Remember the new sched domains */
7123 if (doms_cur
!= &fallback_doms
)
7124 free_sched_domains(doms_cur
, ndoms_cur
);
7125 kfree(dattr_cur
); /* kfree(NULL) is safe */
7126 doms_cur
= doms_new
;
7127 dattr_cur
= dattr_new
;
7128 ndoms_cur
= ndoms_new
;
7130 register_sched_domain_sysctl();
7132 mutex_unlock(&sched_domains_mutex
);
7135 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7138 * Update cpusets according to cpu_active mask. If cpusets are
7139 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7140 * around partition_sched_domains().
7142 * If we come here as part of a suspend/resume, don't touch cpusets because we
7143 * want to restore it back to its original state upon resume anyway.
7145 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7149 case CPU_ONLINE_FROZEN
:
7150 case CPU_DOWN_FAILED_FROZEN
:
7153 * num_cpus_frozen tracks how many CPUs are involved in suspend
7154 * resume sequence. As long as this is not the last online
7155 * operation in the resume sequence, just build a single sched
7156 * domain, ignoring cpusets.
7159 if (likely(num_cpus_frozen
)) {
7160 partition_sched_domains(1, NULL
, NULL
);
7165 * This is the last CPU online operation. So fall through and
7166 * restore the original sched domains by considering the
7167 * cpuset configurations.
7171 cpuset_update_active_cpus(true);
7179 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7182 unsigned long flags
;
7183 long cpu
= (long)hcpu
;
7189 case CPU_DOWN_PREPARE
:
7190 rcu_read_lock_sched();
7191 dl_b
= dl_bw_of(cpu
);
7193 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7194 cpus
= dl_bw_cpus(cpu
);
7195 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7196 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7198 rcu_read_unlock_sched();
7201 return notifier_from_errno(-EBUSY
);
7202 cpuset_update_active_cpus(false);
7204 case CPU_DOWN_PREPARE_FROZEN
:
7206 partition_sched_domains(1, NULL
, NULL
);
7214 void __init
sched_init_smp(void)
7216 cpumask_var_t non_isolated_cpus
;
7218 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7219 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7224 * There's no userspace yet to cause hotplug operations; hence all the
7225 * cpu masks are stable and all blatant races in the below code cannot
7228 mutex_lock(&sched_domains_mutex
);
7229 init_sched_domains(cpu_active_mask
);
7230 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7231 if (cpumask_empty(non_isolated_cpus
))
7232 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7233 mutex_unlock(&sched_domains_mutex
);
7235 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7236 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7237 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7241 /* Move init over to a non-isolated CPU */
7242 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7244 sched_init_granularity();
7245 free_cpumask_var(non_isolated_cpus
);
7247 init_sched_rt_class();
7248 init_sched_dl_class();
7251 void __init
sched_init_smp(void)
7253 sched_init_granularity();
7255 #endif /* CONFIG_SMP */
7257 int in_sched_functions(unsigned long addr
)
7259 return in_lock_functions(addr
) ||
7260 (addr
>= (unsigned long)__sched_text_start
7261 && addr
< (unsigned long)__sched_text_end
);
7264 #ifdef CONFIG_CGROUP_SCHED
7266 * Default task group.
7267 * Every task in system belongs to this group at bootup.
7269 struct task_group root_task_group
;
7270 LIST_HEAD(task_groups
);
7272 /* Cacheline aligned slab cache for task_group */
7273 static struct kmem_cache
*task_group_cache __read_mostly
;
7276 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7278 void __init
sched_init(void)
7281 unsigned long alloc_size
= 0, ptr
;
7283 #ifdef CONFIG_FAIR_GROUP_SCHED
7284 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7286 #ifdef CONFIG_RT_GROUP_SCHED
7287 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7290 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7292 #ifdef CONFIG_FAIR_GROUP_SCHED
7293 root_task_group
.se
= (struct sched_entity
**)ptr
;
7294 ptr
+= nr_cpu_ids
* sizeof(void **);
7296 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7297 ptr
+= nr_cpu_ids
* sizeof(void **);
7299 #endif /* CONFIG_FAIR_GROUP_SCHED */
7300 #ifdef CONFIG_RT_GROUP_SCHED
7301 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7302 ptr
+= nr_cpu_ids
* sizeof(void **);
7304 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7305 ptr
+= nr_cpu_ids
* sizeof(void **);
7307 #endif /* CONFIG_RT_GROUP_SCHED */
7309 #ifdef CONFIG_CPUMASK_OFFSTACK
7310 for_each_possible_cpu(i
) {
7311 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7312 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7314 #endif /* CONFIG_CPUMASK_OFFSTACK */
7316 init_rt_bandwidth(&def_rt_bandwidth
,
7317 global_rt_period(), global_rt_runtime());
7318 init_dl_bandwidth(&def_dl_bandwidth
,
7319 global_rt_period(), global_rt_runtime());
7322 init_defrootdomain();
7325 #ifdef CONFIG_RT_GROUP_SCHED
7326 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7327 global_rt_period(), global_rt_runtime());
7328 #endif /* CONFIG_RT_GROUP_SCHED */
7330 #ifdef CONFIG_CGROUP_SCHED
7331 task_group_cache
= KMEM_CACHE(task_group
, 0);
7333 list_add(&root_task_group
.list
, &task_groups
);
7334 INIT_LIST_HEAD(&root_task_group
.children
);
7335 INIT_LIST_HEAD(&root_task_group
.siblings
);
7336 autogroup_init(&init_task
);
7337 #endif /* CONFIG_CGROUP_SCHED */
7339 for_each_possible_cpu(i
) {
7343 raw_spin_lock_init(&rq
->lock
);
7345 rq
->calc_load_active
= 0;
7346 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7347 init_cfs_rq(&rq
->cfs
);
7348 init_rt_rq(&rq
->rt
);
7349 init_dl_rq(&rq
->dl
);
7350 #ifdef CONFIG_FAIR_GROUP_SCHED
7351 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7352 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7354 * How much cpu bandwidth does root_task_group get?
7356 * In case of task-groups formed thr' the cgroup filesystem, it
7357 * gets 100% of the cpu resources in the system. This overall
7358 * system cpu resource is divided among the tasks of
7359 * root_task_group and its child task-groups in a fair manner,
7360 * based on each entity's (task or task-group's) weight
7361 * (se->load.weight).
7363 * In other words, if root_task_group has 10 tasks of weight
7364 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7365 * then A0's share of the cpu resource is:
7367 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7369 * We achieve this by letting root_task_group's tasks sit
7370 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7372 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7373 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7374 #endif /* CONFIG_FAIR_GROUP_SCHED */
7376 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7377 #ifdef CONFIG_RT_GROUP_SCHED
7378 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7381 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7382 rq
->cpu_load
[j
] = 0;
7387 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7388 rq
->balance_callback
= NULL
;
7389 rq
->active_balance
= 0;
7390 rq
->next_balance
= jiffies
;
7395 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7396 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7398 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7400 rq_attach_root(rq
, &def_root_domain
);
7401 #ifdef CONFIG_NO_HZ_COMMON
7402 rq
->last_load_update_tick
= jiffies
;
7405 #ifdef CONFIG_NO_HZ_FULL
7406 rq
->last_sched_tick
= 0;
7408 #endif /* CONFIG_SMP */
7410 atomic_set(&rq
->nr_iowait
, 0);
7413 set_load_weight(&init_task
);
7415 #ifdef CONFIG_PREEMPT_NOTIFIERS
7416 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7420 * The boot idle thread does lazy MMU switching as well:
7422 atomic_inc(&init_mm
.mm_count
);
7423 enter_lazy_tlb(&init_mm
, current
);
7426 * During early bootup we pretend to be a normal task:
7428 current
->sched_class
= &fair_sched_class
;
7431 * Make us the idle thread. Technically, schedule() should not be
7432 * called from this thread, however somewhere below it might be,
7433 * but because we are the idle thread, we just pick up running again
7434 * when this runqueue becomes "idle".
7436 init_idle(current
, smp_processor_id());
7438 calc_load_update
= jiffies
+ LOAD_FREQ
;
7441 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7442 /* May be allocated at isolcpus cmdline parse time */
7443 if (cpu_isolated_map
== NULL
)
7444 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7445 idle_thread_set_boot_cpu();
7446 set_cpu_rq_start_time();
7448 init_sched_fair_class();
7450 scheduler_running
= 1;
7453 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7454 static inline int preempt_count_equals(int preempt_offset
)
7456 int nested
= preempt_count() + rcu_preempt_depth();
7458 return (nested
== preempt_offset
);
7461 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7464 * Blocking primitives will set (and therefore destroy) current->state,
7465 * since we will exit with TASK_RUNNING make sure we enter with it,
7466 * otherwise we will destroy state.
7468 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7469 "do not call blocking ops when !TASK_RUNNING; "
7470 "state=%lx set at [<%p>] %pS\n",
7472 (void *)current
->task_state_change
,
7473 (void *)current
->task_state_change
);
7475 ___might_sleep(file
, line
, preempt_offset
);
7477 EXPORT_SYMBOL(__might_sleep
);
7479 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7481 static unsigned long prev_jiffy
; /* ratelimiting */
7483 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7484 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7485 !is_idle_task(current
)) ||
7486 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7488 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7490 prev_jiffy
= jiffies
;
7493 "BUG: sleeping function called from invalid context at %s:%d\n",
7496 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7497 in_atomic(), irqs_disabled(),
7498 current
->pid
, current
->comm
);
7500 if (task_stack_end_corrupted(current
))
7501 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7503 debug_show_held_locks(current
);
7504 if (irqs_disabled())
7505 print_irqtrace_events(current
);
7506 #ifdef CONFIG_DEBUG_PREEMPT
7507 if (!preempt_count_equals(preempt_offset
)) {
7508 pr_err("Preemption disabled at:");
7509 print_ip_sym(current
->preempt_disable_ip
);
7515 EXPORT_SYMBOL(___might_sleep
);
7518 #ifdef CONFIG_MAGIC_SYSRQ
7519 void normalize_rt_tasks(void)
7521 struct task_struct
*g
, *p
;
7522 struct sched_attr attr
= {
7523 .sched_policy
= SCHED_NORMAL
,
7526 read_lock(&tasklist_lock
);
7527 for_each_process_thread(g
, p
) {
7529 * Only normalize user tasks:
7531 if (p
->flags
& PF_KTHREAD
)
7534 p
->se
.exec_start
= 0;
7535 #ifdef CONFIG_SCHEDSTATS
7536 p
->se
.statistics
.wait_start
= 0;
7537 p
->se
.statistics
.sleep_start
= 0;
7538 p
->se
.statistics
.block_start
= 0;
7541 if (!dl_task(p
) && !rt_task(p
)) {
7543 * Renice negative nice level userspace
7546 if (task_nice(p
) < 0)
7547 set_user_nice(p
, 0);
7551 __sched_setscheduler(p
, &attr
, false, false);
7553 read_unlock(&tasklist_lock
);
7556 #endif /* CONFIG_MAGIC_SYSRQ */
7558 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7560 * These functions are only useful for the IA64 MCA handling, or kdb.
7562 * They can only be called when the whole system has been
7563 * stopped - every CPU needs to be quiescent, and no scheduling
7564 * activity can take place. Using them for anything else would
7565 * be a serious bug, and as a result, they aren't even visible
7566 * under any other configuration.
7570 * curr_task - return the current task for a given cpu.
7571 * @cpu: the processor in question.
7573 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7575 * Return: The current task for @cpu.
7577 struct task_struct
*curr_task(int cpu
)
7579 return cpu_curr(cpu
);
7582 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7586 * set_curr_task - set the current task for a given cpu.
7587 * @cpu: the processor in question.
7588 * @p: the task pointer to set.
7590 * Description: This function must only be used when non-maskable interrupts
7591 * are serviced on a separate stack. It allows the architecture to switch the
7592 * notion of the current task on a cpu in a non-blocking manner. This function
7593 * must be called with all CPU's synchronized, and interrupts disabled, the
7594 * and caller must save the original value of the current task (see
7595 * curr_task() above) and restore that value before reenabling interrupts and
7596 * re-starting the system.
7598 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7600 void set_curr_task(int cpu
, struct task_struct
*p
)
7607 #ifdef CONFIG_CGROUP_SCHED
7608 /* task_group_lock serializes the addition/removal of task groups */
7609 static DEFINE_SPINLOCK(task_group_lock
);
7611 static void sched_free_group(struct task_group
*tg
)
7613 free_fair_sched_group(tg
);
7614 free_rt_sched_group(tg
);
7616 kmem_cache_free(task_group_cache
, tg
);
7619 /* allocate runqueue etc for a new task group */
7620 struct task_group
*sched_create_group(struct task_group
*parent
)
7622 struct task_group
*tg
;
7624 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7626 return ERR_PTR(-ENOMEM
);
7628 if (!alloc_fair_sched_group(tg
, parent
))
7631 if (!alloc_rt_sched_group(tg
, parent
))
7637 sched_free_group(tg
);
7638 return ERR_PTR(-ENOMEM
);
7641 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7643 unsigned long flags
;
7645 spin_lock_irqsave(&task_group_lock
, flags
);
7646 list_add_rcu(&tg
->list
, &task_groups
);
7648 WARN_ON(!parent
); /* root should already exist */
7650 tg
->parent
= parent
;
7651 INIT_LIST_HEAD(&tg
->children
);
7652 list_add_rcu(&tg
->siblings
, &parent
->children
);
7653 spin_unlock_irqrestore(&task_group_lock
, flags
);
7656 /* rcu callback to free various structures associated with a task group */
7657 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7659 /* now it should be safe to free those cfs_rqs */
7660 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7663 void sched_destroy_group(struct task_group
*tg
)
7665 /* wait for possible concurrent references to cfs_rqs complete */
7666 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7669 void sched_offline_group(struct task_group
*tg
)
7671 unsigned long flags
;
7673 /* end participation in shares distribution */
7674 unregister_fair_sched_group(tg
);
7676 spin_lock_irqsave(&task_group_lock
, flags
);
7677 list_del_rcu(&tg
->list
);
7678 list_del_rcu(&tg
->siblings
);
7679 spin_unlock_irqrestore(&task_group_lock
, flags
);
7682 /* change task's runqueue when it moves between groups.
7683 * The caller of this function should have put the task in its new group
7684 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7685 * reflect its new group.
7687 void sched_move_task(struct task_struct
*tsk
)
7689 struct task_group
*tg
;
7690 int queued
, running
;
7691 unsigned long flags
;
7694 rq
= task_rq_lock(tsk
, &flags
);
7696 running
= task_current(rq
, tsk
);
7697 queued
= task_on_rq_queued(tsk
);
7700 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7701 if (unlikely(running
))
7702 put_prev_task(rq
, tsk
);
7705 * All callers are synchronized by task_rq_lock(); we do not use RCU
7706 * which is pointless here. Thus, we pass "true" to task_css_check()
7707 * to prevent lockdep warnings.
7709 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7710 struct task_group
, css
);
7711 tg
= autogroup_task_group(tsk
, tg
);
7712 tsk
->sched_task_group
= tg
;
7714 #ifdef CONFIG_FAIR_GROUP_SCHED
7715 if (tsk
->sched_class
->task_move_group
)
7716 tsk
->sched_class
->task_move_group(tsk
);
7719 set_task_rq(tsk
, task_cpu(tsk
));
7721 if (unlikely(running
))
7722 tsk
->sched_class
->set_curr_task(rq
);
7724 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7726 task_rq_unlock(rq
, tsk
, &flags
);
7728 #endif /* CONFIG_CGROUP_SCHED */
7730 #ifdef CONFIG_RT_GROUP_SCHED
7732 * Ensure that the real time constraints are schedulable.
7734 static DEFINE_MUTEX(rt_constraints_mutex
);
7736 /* Must be called with tasklist_lock held */
7737 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7739 struct task_struct
*g
, *p
;
7742 * Autogroups do not have RT tasks; see autogroup_create().
7744 if (task_group_is_autogroup(tg
))
7747 for_each_process_thread(g
, p
) {
7748 if (rt_task(p
) && task_group(p
) == tg
)
7755 struct rt_schedulable_data
{
7756 struct task_group
*tg
;
7761 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7763 struct rt_schedulable_data
*d
= data
;
7764 struct task_group
*child
;
7765 unsigned long total
, sum
= 0;
7766 u64 period
, runtime
;
7768 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7769 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7772 period
= d
->rt_period
;
7773 runtime
= d
->rt_runtime
;
7777 * Cannot have more runtime than the period.
7779 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7783 * Ensure we don't starve existing RT tasks.
7785 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7788 total
= to_ratio(period
, runtime
);
7791 * Nobody can have more than the global setting allows.
7793 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7797 * The sum of our children's runtime should not exceed our own.
7799 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7800 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7801 runtime
= child
->rt_bandwidth
.rt_runtime
;
7803 if (child
== d
->tg
) {
7804 period
= d
->rt_period
;
7805 runtime
= d
->rt_runtime
;
7808 sum
+= to_ratio(period
, runtime
);
7817 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7821 struct rt_schedulable_data data
= {
7823 .rt_period
= period
,
7824 .rt_runtime
= runtime
,
7828 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7834 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7835 u64 rt_period
, u64 rt_runtime
)
7840 * Disallowing the root group RT runtime is BAD, it would disallow the
7841 * kernel creating (and or operating) RT threads.
7843 if (tg
== &root_task_group
&& rt_runtime
== 0)
7846 /* No period doesn't make any sense. */
7850 mutex_lock(&rt_constraints_mutex
);
7851 read_lock(&tasklist_lock
);
7852 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7856 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7857 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7858 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7860 for_each_possible_cpu(i
) {
7861 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7863 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7864 rt_rq
->rt_runtime
= rt_runtime
;
7865 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7867 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7869 read_unlock(&tasklist_lock
);
7870 mutex_unlock(&rt_constraints_mutex
);
7875 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7877 u64 rt_runtime
, rt_period
;
7879 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7880 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7881 if (rt_runtime_us
< 0)
7882 rt_runtime
= RUNTIME_INF
;
7884 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7887 static long sched_group_rt_runtime(struct task_group
*tg
)
7891 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7894 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7895 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7896 return rt_runtime_us
;
7899 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7901 u64 rt_runtime
, rt_period
;
7903 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7904 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7906 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7909 static long sched_group_rt_period(struct task_group
*tg
)
7913 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7914 do_div(rt_period_us
, NSEC_PER_USEC
);
7915 return rt_period_us
;
7917 #endif /* CONFIG_RT_GROUP_SCHED */
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 static int sched_rt_global_constraints(void)
7924 mutex_lock(&rt_constraints_mutex
);
7925 read_lock(&tasklist_lock
);
7926 ret
= __rt_schedulable(NULL
, 0, 0);
7927 read_unlock(&tasklist_lock
);
7928 mutex_unlock(&rt_constraints_mutex
);
7933 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7935 /* Don't accept realtime tasks when there is no way for them to run */
7936 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7942 #else /* !CONFIG_RT_GROUP_SCHED */
7943 static int sched_rt_global_constraints(void)
7945 unsigned long flags
;
7948 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7949 for_each_possible_cpu(i
) {
7950 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7952 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7953 rt_rq
->rt_runtime
= global_rt_runtime();
7954 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7956 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7960 #endif /* CONFIG_RT_GROUP_SCHED */
7962 static int sched_dl_global_validate(void)
7964 u64 runtime
= global_rt_runtime();
7965 u64 period
= global_rt_period();
7966 u64 new_bw
= to_ratio(period
, runtime
);
7969 unsigned long flags
;
7972 * Here we want to check the bandwidth not being set to some
7973 * value smaller than the currently allocated bandwidth in
7974 * any of the root_domains.
7976 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7977 * cycling on root_domains... Discussion on different/better
7978 * solutions is welcome!
7980 for_each_possible_cpu(cpu
) {
7981 rcu_read_lock_sched();
7982 dl_b
= dl_bw_of(cpu
);
7984 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7985 if (new_bw
< dl_b
->total_bw
)
7987 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7989 rcu_read_unlock_sched();
7998 static void sched_dl_do_global(void)
8003 unsigned long flags
;
8005 def_dl_bandwidth
.dl_period
= global_rt_period();
8006 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8008 if (global_rt_runtime() != RUNTIME_INF
)
8009 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8012 * FIXME: As above...
8014 for_each_possible_cpu(cpu
) {
8015 rcu_read_lock_sched();
8016 dl_b
= dl_bw_of(cpu
);
8018 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8020 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8022 rcu_read_unlock_sched();
8026 static int sched_rt_global_validate(void)
8028 if (sysctl_sched_rt_period
<= 0)
8031 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8032 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8038 static void sched_rt_do_global(void)
8040 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8041 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8044 int sched_rt_handler(struct ctl_table
*table
, int write
,
8045 void __user
*buffer
, size_t *lenp
,
8048 int old_period
, old_runtime
;
8049 static DEFINE_MUTEX(mutex
);
8053 old_period
= sysctl_sched_rt_period
;
8054 old_runtime
= sysctl_sched_rt_runtime
;
8056 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8058 if (!ret
&& write
) {
8059 ret
= sched_rt_global_validate();
8063 ret
= sched_dl_global_validate();
8067 ret
= sched_rt_global_constraints();
8071 sched_rt_do_global();
8072 sched_dl_do_global();
8076 sysctl_sched_rt_period
= old_period
;
8077 sysctl_sched_rt_runtime
= old_runtime
;
8079 mutex_unlock(&mutex
);
8084 int sched_rr_handler(struct ctl_table
*table
, int write
,
8085 void __user
*buffer
, size_t *lenp
,
8089 static DEFINE_MUTEX(mutex
);
8092 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8093 /* make sure that internally we keep jiffies */
8094 /* also, writing zero resets timeslice to default */
8095 if (!ret
&& write
) {
8096 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8097 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8099 mutex_unlock(&mutex
);
8103 #ifdef CONFIG_CGROUP_SCHED
8105 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8107 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8110 static struct cgroup_subsys_state
*
8111 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8113 struct task_group
*parent
= css_tg(parent_css
);
8114 struct task_group
*tg
;
8117 /* This is early initialization for the top cgroup */
8118 return &root_task_group
.css
;
8121 tg
= sched_create_group(parent
);
8123 return ERR_PTR(-ENOMEM
);
8125 sched_online_group(tg
, parent
);
8130 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8132 struct task_group
*tg
= css_tg(css
);
8134 sched_offline_group(tg
);
8137 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8139 struct task_group
*tg
= css_tg(css
);
8142 * Relies on the RCU grace period between css_released() and this.
8144 sched_free_group(tg
);
8147 static void cpu_cgroup_fork(struct task_struct
*task
)
8149 sched_move_task(task
);
8152 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8154 struct task_struct
*task
;
8155 struct cgroup_subsys_state
*css
;
8157 cgroup_taskset_for_each(task
, css
, tset
) {
8158 #ifdef CONFIG_RT_GROUP_SCHED
8159 if (!sched_rt_can_attach(css_tg(css
), task
))
8162 /* We don't support RT-tasks being in separate groups */
8163 if (task
->sched_class
!= &fair_sched_class
)
8170 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8172 struct task_struct
*task
;
8173 struct cgroup_subsys_state
*css
;
8175 cgroup_taskset_for_each(task
, css
, tset
)
8176 sched_move_task(task
);
8179 #ifdef CONFIG_FAIR_GROUP_SCHED
8180 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8181 struct cftype
*cftype
, u64 shareval
)
8183 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8186 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8189 struct task_group
*tg
= css_tg(css
);
8191 return (u64
) scale_load_down(tg
->shares
);
8194 #ifdef CONFIG_CFS_BANDWIDTH
8195 static DEFINE_MUTEX(cfs_constraints_mutex
);
8197 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8198 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8200 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8202 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8204 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8205 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8207 if (tg
== &root_task_group
)
8211 * Ensure we have at some amount of bandwidth every period. This is
8212 * to prevent reaching a state of large arrears when throttled via
8213 * entity_tick() resulting in prolonged exit starvation.
8215 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8219 * Likewise, bound things on the otherside by preventing insane quota
8220 * periods. This also allows us to normalize in computing quota
8223 if (period
> max_cfs_quota_period
)
8227 * Prevent race between setting of cfs_rq->runtime_enabled and
8228 * unthrottle_offline_cfs_rqs().
8231 mutex_lock(&cfs_constraints_mutex
);
8232 ret
= __cfs_schedulable(tg
, period
, quota
);
8236 runtime_enabled
= quota
!= RUNTIME_INF
;
8237 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8239 * If we need to toggle cfs_bandwidth_used, off->on must occur
8240 * before making related changes, and on->off must occur afterwards
8242 if (runtime_enabled
&& !runtime_was_enabled
)
8243 cfs_bandwidth_usage_inc();
8244 raw_spin_lock_irq(&cfs_b
->lock
);
8245 cfs_b
->period
= ns_to_ktime(period
);
8246 cfs_b
->quota
= quota
;
8248 __refill_cfs_bandwidth_runtime(cfs_b
);
8249 /* restart the period timer (if active) to handle new period expiry */
8250 if (runtime_enabled
)
8251 start_cfs_bandwidth(cfs_b
);
8252 raw_spin_unlock_irq(&cfs_b
->lock
);
8254 for_each_online_cpu(i
) {
8255 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8256 struct rq
*rq
= cfs_rq
->rq
;
8258 raw_spin_lock_irq(&rq
->lock
);
8259 cfs_rq
->runtime_enabled
= runtime_enabled
;
8260 cfs_rq
->runtime_remaining
= 0;
8262 if (cfs_rq
->throttled
)
8263 unthrottle_cfs_rq(cfs_rq
);
8264 raw_spin_unlock_irq(&rq
->lock
);
8266 if (runtime_was_enabled
&& !runtime_enabled
)
8267 cfs_bandwidth_usage_dec();
8269 mutex_unlock(&cfs_constraints_mutex
);
8275 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8279 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8280 if (cfs_quota_us
< 0)
8281 quota
= RUNTIME_INF
;
8283 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8285 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8288 long tg_get_cfs_quota(struct task_group
*tg
)
8292 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8295 quota_us
= tg
->cfs_bandwidth
.quota
;
8296 do_div(quota_us
, NSEC_PER_USEC
);
8301 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8305 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8306 quota
= tg
->cfs_bandwidth
.quota
;
8308 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8311 long tg_get_cfs_period(struct task_group
*tg
)
8315 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8316 do_div(cfs_period_us
, NSEC_PER_USEC
);
8318 return cfs_period_us
;
8321 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8324 return tg_get_cfs_quota(css_tg(css
));
8327 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8328 struct cftype
*cftype
, s64 cfs_quota_us
)
8330 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8333 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8336 return tg_get_cfs_period(css_tg(css
));
8339 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8340 struct cftype
*cftype
, u64 cfs_period_us
)
8342 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8345 struct cfs_schedulable_data
{
8346 struct task_group
*tg
;
8351 * normalize group quota/period to be quota/max_period
8352 * note: units are usecs
8354 static u64
normalize_cfs_quota(struct task_group
*tg
,
8355 struct cfs_schedulable_data
*d
)
8363 period
= tg_get_cfs_period(tg
);
8364 quota
= tg_get_cfs_quota(tg
);
8367 /* note: these should typically be equivalent */
8368 if (quota
== RUNTIME_INF
|| quota
== -1)
8371 return to_ratio(period
, quota
);
8374 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8376 struct cfs_schedulable_data
*d
= data
;
8377 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8378 s64 quota
= 0, parent_quota
= -1;
8381 quota
= RUNTIME_INF
;
8383 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8385 quota
= normalize_cfs_quota(tg
, d
);
8386 parent_quota
= parent_b
->hierarchical_quota
;
8389 * ensure max(child_quota) <= parent_quota, inherit when no
8392 if (quota
== RUNTIME_INF
)
8393 quota
= parent_quota
;
8394 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8397 cfs_b
->hierarchical_quota
= quota
;
8402 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8405 struct cfs_schedulable_data data
= {
8411 if (quota
!= RUNTIME_INF
) {
8412 do_div(data
.period
, NSEC_PER_USEC
);
8413 do_div(data
.quota
, NSEC_PER_USEC
);
8417 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8423 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8425 struct task_group
*tg
= css_tg(seq_css(sf
));
8426 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8428 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8429 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8430 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8434 #endif /* CONFIG_CFS_BANDWIDTH */
8435 #endif /* CONFIG_FAIR_GROUP_SCHED */
8437 #ifdef CONFIG_RT_GROUP_SCHED
8438 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8439 struct cftype
*cft
, s64 val
)
8441 return sched_group_set_rt_runtime(css_tg(css
), val
);
8444 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8447 return sched_group_rt_runtime(css_tg(css
));
8450 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8451 struct cftype
*cftype
, u64 rt_period_us
)
8453 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8456 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8459 return sched_group_rt_period(css_tg(css
));
8461 #endif /* CONFIG_RT_GROUP_SCHED */
8463 static struct cftype cpu_files
[] = {
8464 #ifdef CONFIG_FAIR_GROUP_SCHED
8467 .read_u64
= cpu_shares_read_u64
,
8468 .write_u64
= cpu_shares_write_u64
,
8471 #ifdef CONFIG_CFS_BANDWIDTH
8473 .name
= "cfs_quota_us",
8474 .read_s64
= cpu_cfs_quota_read_s64
,
8475 .write_s64
= cpu_cfs_quota_write_s64
,
8478 .name
= "cfs_period_us",
8479 .read_u64
= cpu_cfs_period_read_u64
,
8480 .write_u64
= cpu_cfs_period_write_u64
,
8484 .seq_show
= cpu_stats_show
,
8487 #ifdef CONFIG_RT_GROUP_SCHED
8489 .name
= "rt_runtime_us",
8490 .read_s64
= cpu_rt_runtime_read
,
8491 .write_s64
= cpu_rt_runtime_write
,
8494 .name
= "rt_period_us",
8495 .read_u64
= cpu_rt_period_read_uint
,
8496 .write_u64
= cpu_rt_period_write_uint
,
8502 struct cgroup_subsys cpu_cgrp_subsys
= {
8503 .css_alloc
= cpu_cgroup_css_alloc
,
8504 .css_released
= cpu_cgroup_css_released
,
8505 .css_free
= cpu_cgroup_css_free
,
8506 .fork
= cpu_cgroup_fork
,
8507 .can_attach
= cpu_cgroup_can_attach
,
8508 .attach
= cpu_cgroup_attach
,
8509 .legacy_cftypes
= cpu_files
,
8513 #endif /* CONFIG_CGROUP_SCHED */
8515 void dump_cpu_task(int cpu
)
8517 pr_info("Task dump for CPU %d:\n", cpu
);
8518 sched_show_task(cpu_curr(cpu
));
8522 * Nice levels are multiplicative, with a gentle 10% change for every
8523 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8524 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8525 * that remained on nice 0.
8527 * The "10% effect" is relative and cumulative: from _any_ nice level,
8528 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8529 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8530 * If a task goes up by ~10% and another task goes down by ~10% then
8531 * the relative distance between them is ~25%.)
8533 const int sched_prio_to_weight
[40] = {
8534 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8535 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8536 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8537 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8538 /* 0 */ 1024, 820, 655, 526, 423,
8539 /* 5 */ 335, 272, 215, 172, 137,
8540 /* 10 */ 110, 87, 70, 56, 45,
8541 /* 15 */ 36, 29, 23, 18, 15,
8545 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8547 * In cases where the weight does not change often, we can use the
8548 * precalculated inverse to speed up arithmetics by turning divisions
8549 * into multiplications:
8551 const u32 sched_prio_to_wmult
[40] = {
8552 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8553 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8554 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8555 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8556 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8557 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8558 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8559 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,