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 * If there are more than one RR tasks, we need the tick to effect the
600 * actual RR behaviour.
602 if (rq
->rt
.rr_nr_running
) {
603 if (rq
->rt
.rr_nr_running
== 1)
610 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
611 * forced preemption between FIFO tasks.
613 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
618 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
619 * if there's more than one we need the tick for involuntary
622 if (rq
->nr_running
> 1)
627 #endif /* CONFIG_NO_HZ_FULL */
629 void sched_avg_update(struct rq
*rq
)
631 s64 period
= sched_avg_period();
633 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
635 * Inline assembly required to prevent the compiler
636 * optimising this loop into a divmod call.
637 * See __iter_div_u64_rem() for another example of this.
639 asm("" : "+rm" (rq
->age_stamp
));
640 rq
->age_stamp
+= period
;
645 #endif /* CONFIG_SMP */
647 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
653 * Caller must hold rcu_lock or sufficient equivalent.
655 int walk_tg_tree_from(struct task_group
*from
,
656 tg_visitor down
, tg_visitor up
, void *data
)
658 struct task_group
*parent
, *child
;
664 ret
= (*down
)(parent
, data
);
667 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
674 ret
= (*up
)(parent
, data
);
675 if (ret
|| parent
== from
)
679 parent
= parent
->parent
;
686 int tg_nop(struct task_group
*tg
, void *data
)
692 static void set_load_weight(struct task_struct
*p
)
694 int prio
= p
->static_prio
- MAX_RT_PRIO
;
695 struct load_weight
*load
= &p
->se
.load
;
698 * SCHED_IDLE tasks get minimal weight:
700 if (idle_policy(p
->policy
)) {
701 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
702 load
->inv_weight
= WMULT_IDLEPRIO
;
706 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
707 load
->inv_weight
= sched_prio_to_wmult
[prio
];
710 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
713 if (!(flags
& ENQUEUE_RESTORE
))
714 sched_info_queued(rq
, p
);
715 p
->sched_class
->enqueue_task(rq
, p
, flags
);
718 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
721 if (!(flags
& DEQUEUE_SAVE
))
722 sched_info_dequeued(rq
, p
);
723 p
->sched_class
->dequeue_task(rq
, p
, flags
);
726 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
728 if (task_contributes_to_load(p
))
729 rq
->nr_uninterruptible
--;
731 enqueue_task(rq
, p
, flags
);
734 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
736 if (task_contributes_to_load(p
))
737 rq
->nr_uninterruptible
++;
739 dequeue_task(rq
, p
, flags
);
742 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
745 * In theory, the compile should just see 0 here, and optimize out the call
746 * to sched_rt_avg_update. But I don't trust it...
748 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
749 s64 steal
= 0, irq_delta
= 0;
751 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
752 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
755 * Since irq_time is only updated on {soft,}irq_exit, we might run into
756 * this case when a previous update_rq_clock() happened inside a
759 * When this happens, we stop ->clock_task and only update the
760 * prev_irq_time stamp to account for the part that fit, so that a next
761 * update will consume the rest. This ensures ->clock_task is
764 * It does however cause some slight miss-attribution of {soft,}irq
765 * time, a more accurate solution would be to update the irq_time using
766 * the current rq->clock timestamp, except that would require using
769 if (irq_delta
> delta
)
772 rq
->prev_irq_time
+= irq_delta
;
775 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
776 if (static_key_false((¶virt_steal_rq_enabled
))) {
777 steal
= paravirt_steal_clock(cpu_of(rq
));
778 steal
-= rq
->prev_steal_time_rq
;
780 if (unlikely(steal
> delta
))
783 rq
->prev_steal_time_rq
+= steal
;
788 rq
->clock_task
+= delta
;
790 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
791 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
792 sched_rt_avg_update(rq
, irq_delta
+ steal
);
796 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
798 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
799 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
803 * Make it appear like a SCHED_FIFO task, its something
804 * userspace knows about and won't get confused about.
806 * Also, it will make PI more or less work without too
807 * much confusion -- but then, stop work should not
808 * rely on PI working anyway.
810 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
812 stop
->sched_class
= &stop_sched_class
;
815 cpu_rq(cpu
)->stop
= stop
;
819 * Reset it back to a normal scheduling class so that
820 * it can die in pieces.
822 old_stop
->sched_class
= &rt_sched_class
;
827 * __normal_prio - return the priority that is based on the static prio
829 static inline int __normal_prio(struct task_struct
*p
)
831 return p
->static_prio
;
835 * Calculate the expected normal priority: i.e. priority
836 * without taking RT-inheritance into account. Might be
837 * boosted by interactivity modifiers. Changes upon fork,
838 * setprio syscalls, and whenever the interactivity
839 * estimator recalculates.
841 static inline int normal_prio(struct task_struct
*p
)
845 if (task_has_dl_policy(p
))
846 prio
= MAX_DL_PRIO
-1;
847 else if (task_has_rt_policy(p
))
848 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
850 prio
= __normal_prio(p
);
855 * Calculate the current priority, i.e. the priority
856 * taken into account by the scheduler. This value might
857 * be boosted by RT tasks, or might be boosted by
858 * interactivity modifiers. Will be RT if the task got
859 * RT-boosted. If not then it returns p->normal_prio.
861 static int effective_prio(struct task_struct
*p
)
863 p
->normal_prio
= normal_prio(p
);
865 * If we are RT tasks or we were boosted to RT priority,
866 * keep the priority unchanged. Otherwise, update priority
867 * to the normal priority:
869 if (!rt_prio(p
->prio
))
870 return p
->normal_prio
;
875 * task_curr - is this task currently executing on a CPU?
876 * @p: the task in question.
878 * Return: 1 if the task is currently executing. 0 otherwise.
880 inline int task_curr(const struct task_struct
*p
)
882 return cpu_curr(task_cpu(p
)) == p
;
886 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
887 * use the balance_callback list if you want balancing.
889 * this means any call to check_class_changed() must be followed by a call to
890 * balance_callback().
892 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
893 const struct sched_class
*prev_class
,
896 if (prev_class
!= p
->sched_class
) {
897 if (prev_class
->switched_from
)
898 prev_class
->switched_from(rq
, p
);
900 p
->sched_class
->switched_to(rq
, p
);
901 } else if (oldprio
!= p
->prio
|| dl_task(p
))
902 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
905 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
907 const struct sched_class
*class;
909 if (p
->sched_class
== rq
->curr
->sched_class
) {
910 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
912 for_each_class(class) {
913 if (class == rq
->curr
->sched_class
)
915 if (class == p
->sched_class
) {
923 * A queue event has occurred, and we're going to schedule. In
924 * this case, we can save a useless back to back clock update.
926 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
927 rq_clock_skip_update(rq
, true);
932 * This is how migration works:
934 * 1) we invoke migration_cpu_stop() on the target CPU using
936 * 2) stopper starts to run (implicitly forcing the migrated thread
938 * 3) it checks whether the migrated task is still in the wrong runqueue.
939 * 4) if it's in the wrong runqueue then the migration thread removes
940 * it and puts it into the right queue.
941 * 5) stopper completes and stop_one_cpu() returns and the migration
946 * move_queued_task - move a queued task to new rq.
948 * Returns (locked) new rq. Old rq's lock is released.
950 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
952 lockdep_assert_held(&rq
->lock
);
954 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
955 dequeue_task(rq
, p
, 0);
956 set_task_cpu(p
, new_cpu
);
957 raw_spin_unlock(&rq
->lock
);
959 rq
= cpu_rq(new_cpu
);
961 raw_spin_lock(&rq
->lock
);
962 BUG_ON(task_cpu(p
) != new_cpu
);
963 enqueue_task(rq
, p
, 0);
964 p
->on_rq
= TASK_ON_RQ_QUEUED
;
965 check_preempt_curr(rq
, p
, 0);
970 struct migration_arg
{
971 struct task_struct
*task
;
976 * Move (not current) task off this cpu, onto dest cpu. We're doing
977 * this because either it can't run here any more (set_cpus_allowed()
978 * away from this CPU, or CPU going down), or because we're
979 * attempting to rebalance this task on exec (sched_exec).
981 * So we race with normal scheduler movements, but that's OK, as long
982 * as the task is no longer on this CPU.
984 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
986 if (unlikely(!cpu_active(dest_cpu
)))
989 /* Affinity changed (again). */
990 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
993 rq
= move_queued_task(rq
, p
, dest_cpu
);
999 * migration_cpu_stop - this will be executed by a highprio stopper thread
1000 * and performs thread migration by bumping thread off CPU then
1001 * 'pushing' onto another runqueue.
1003 static int migration_cpu_stop(void *data
)
1005 struct migration_arg
*arg
= data
;
1006 struct task_struct
*p
= arg
->task
;
1007 struct rq
*rq
= this_rq();
1010 * The original target cpu might have gone down and we might
1011 * be on another cpu but it doesn't matter.
1013 local_irq_disable();
1015 * We need to explicitly wake pending tasks before running
1016 * __migrate_task() such that we will not miss enforcing cpus_allowed
1017 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1019 sched_ttwu_pending();
1021 raw_spin_lock(&p
->pi_lock
);
1022 raw_spin_lock(&rq
->lock
);
1024 * If task_rq(p) != rq, it cannot be migrated here, because we're
1025 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1026 * we're holding p->pi_lock.
1028 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1029 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1030 raw_spin_unlock(&rq
->lock
);
1031 raw_spin_unlock(&p
->pi_lock
);
1038 * sched_class::set_cpus_allowed must do the below, but is not required to
1039 * actually call this function.
1041 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1043 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1044 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1047 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1049 struct rq
*rq
= task_rq(p
);
1050 bool queued
, running
;
1052 lockdep_assert_held(&p
->pi_lock
);
1054 queued
= task_on_rq_queued(p
);
1055 running
= task_current(rq
, p
);
1059 * Because __kthread_bind() calls this on blocked tasks without
1062 lockdep_assert_held(&rq
->lock
);
1063 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1066 put_prev_task(rq
, p
);
1068 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1071 p
->sched_class
->set_curr_task(rq
);
1073 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1077 * Change a given task's CPU affinity. Migrate the thread to a
1078 * proper CPU and schedule it away if the CPU it's executing on
1079 * is removed from the allowed bitmask.
1081 * NOTE: the caller must have a valid reference to the task, the
1082 * task must not exit() & deallocate itself prematurely. The
1083 * call is not atomic; no spinlocks may be held.
1085 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1086 const struct cpumask
*new_mask
, bool check
)
1088 unsigned long flags
;
1090 unsigned int dest_cpu
;
1093 rq
= task_rq_lock(p
, &flags
);
1096 * Must re-check here, to close a race against __kthread_bind(),
1097 * sched_setaffinity() is not guaranteed to observe the flag.
1099 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1104 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1107 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1112 do_set_cpus_allowed(p
, new_mask
);
1114 /* Can the task run on the task's current CPU? If so, we're done */
1115 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1118 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1119 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1120 struct migration_arg arg
= { p
, dest_cpu
};
1121 /* Need help from migration thread: drop lock and wait. */
1122 task_rq_unlock(rq
, p
, &flags
);
1123 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1124 tlb_migrate_finish(p
->mm
);
1126 } else if (task_on_rq_queued(p
)) {
1128 * OK, since we're going to drop the lock immediately
1129 * afterwards anyway.
1131 lockdep_unpin_lock(&rq
->lock
);
1132 rq
= move_queued_task(rq
, p
, dest_cpu
);
1133 lockdep_pin_lock(&rq
->lock
);
1136 task_rq_unlock(rq
, p
, &flags
);
1141 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1143 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1145 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1147 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1149 #ifdef CONFIG_SCHED_DEBUG
1151 * We should never call set_task_cpu() on a blocked task,
1152 * ttwu() will sort out the placement.
1154 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1158 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1159 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1160 * time relying on p->on_rq.
1162 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1163 p
->sched_class
== &fair_sched_class
&&
1164 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1166 #ifdef CONFIG_LOCKDEP
1168 * The caller should hold either p->pi_lock or rq->lock, when changing
1169 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1171 * sched_move_task() holds both and thus holding either pins the cgroup,
1174 * Furthermore, all task_rq users should acquire both locks, see
1177 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1178 lockdep_is_held(&task_rq(p
)->lock
)));
1182 trace_sched_migrate_task(p
, new_cpu
);
1184 if (task_cpu(p
) != new_cpu
) {
1185 if (p
->sched_class
->migrate_task_rq
)
1186 p
->sched_class
->migrate_task_rq(p
);
1187 p
->se
.nr_migrations
++;
1188 perf_event_task_migrate(p
);
1191 __set_task_cpu(p
, new_cpu
);
1194 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1196 if (task_on_rq_queued(p
)) {
1197 struct rq
*src_rq
, *dst_rq
;
1199 src_rq
= task_rq(p
);
1200 dst_rq
= cpu_rq(cpu
);
1202 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1203 deactivate_task(src_rq
, p
, 0);
1204 set_task_cpu(p
, cpu
);
1205 activate_task(dst_rq
, p
, 0);
1206 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1207 check_preempt_curr(dst_rq
, p
, 0);
1210 * Task isn't running anymore; make it appear like we migrated
1211 * it before it went to sleep. This means on wakeup we make the
1212 * previous cpu our targer instead of where it really is.
1218 struct migration_swap_arg
{
1219 struct task_struct
*src_task
, *dst_task
;
1220 int src_cpu
, dst_cpu
;
1223 static int migrate_swap_stop(void *data
)
1225 struct migration_swap_arg
*arg
= data
;
1226 struct rq
*src_rq
, *dst_rq
;
1229 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1232 src_rq
= cpu_rq(arg
->src_cpu
);
1233 dst_rq
= cpu_rq(arg
->dst_cpu
);
1235 double_raw_lock(&arg
->src_task
->pi_lock
,
1236 &arg
->dst_task
->pi_lock
);
1237 double_rq_lock(src_rq
, dst_rq
);
1239 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1242 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1245 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1248 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1251 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1252 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1257 double_rq_unlock(src_rq
, dst_rq
);
1258 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1259 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1265 * Cross migrate two tasks
1267 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1269 struct migration_swap_arg arg
;
1272 arg
= (struct migration_swap_arg
){
1274 .src_cpu
= task_cpu(cur
),
1276 .dst_cpu
= task_cpu(p
),
1279 if (arg
.src_cpu
== arg
.dst_cpu
)
1283 * These three tests are all lockless; this is OK since all of them
1284 * will be re-checked with proper locks held further down the line.
1286 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1289 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1292 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1295 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1296 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1303 * wait_task_inactive - wait for a thread to unschedule.
1305 * If @match_state is nonzero, it's the @p->state value just checked and
1306 * not expected to change. If it changes, i.e. @p might have woken up,
1307 * then return zero. When we succeed in waiting for @p to be off its CPU,
1308 * we return a positive number (its total switch count). If a second call
1309 * a short while later returns the same number, the caller can be sure that
1310 * @p has remained unscheduled the whole time.
1312 * The caller must ensure that the task *will* unschedule sometime soon,
1313 * else this function might spin for a *long* time. This function can't
1314 * be called with interrupts off, or it may introduce deadlock with
1315 * smp_call_function() if an IPI is sent by the same process we are
1316 * waiting to become inactive.
1318 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1320 unsigned long flags
;
1321 int running
, queued
;
1327 * We do the initial early heuristics without holding
1328 * any task-queue locks at all. We'll only try to get
1329 * the runqueue lock when things look like they will
1335 * If the task is actively running on another CPU
1336 * still, just relax and busy-wait without holding
1339 * NOTE! Since we don't hold any locks, it's not
1340 * even sure that "rq" stays as the right runqueue!
1341 * But we don't care, since "task_running()" will
1342 * return false if the runqueue has changed and p
1343 * is actually now running somewhere else!
1345 while (task_running(rq
, p
)) {
1346 if (match_state
&& unlikely(p
->state
!= match_state
))
1352 * Ok, time to look more closely! We need the rq
1353 * lock now, to be *sure*. If we're wrong, we'll
1354 * just go back and repeat.
1356 rq
= task_rq_lock(p
, &flags
);
1357 trace_sched_wait_task(p
);
1358 running
= task_running(rq
, p
);
1359 queued
= task_on_rq_queued(p
);
1361 if (!match_state
|| p
->state
== match_state
)
1362 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1363 task_rq_unlock(rq
, p
, &flags
);
1366 * If it changed from the expected state, bail out now.
1368 if (unlikely(!ncsw
))
1372 * Was it really running after all now that we
1373 * checked with the proper locks actually held?
1375 * Oops. Go back and try again..
1377 if (unlikely(running
)) {
1383 * It's not enough that it's not actively running,
1384 * it must be off the runqueue _entirely_, and not
1387 * So if it was still runnable (but just not actively
1388 * running right now), it's preempted, and we should
1389 * yield - it could be a while.
1391 if (unlikely(queued
)) {
1392 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1394 set_current_state(TASK_UNINTERRUPTIBLE
);
1395 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1400 * Ahh, all good. It wasn't running, and it wasn't
1401 * runnable, which means that it will never become
1402 * running in the future either. We're all done!
1411 * kick_process - kick a running thread to enter/exit the kernel
1412 * @p: the to-be-kicked thread
1414 * Cause a process which is running on another CPU to enter
1415 * kernel-mode, without any delay. (to get signals handled.)
1417 * NOTE: this function doesn't have to take the runqueue lock,
1418 * because all it wants to ensure is that the remote task enters
1419 * the kernel. If the IPI races and the task has been migrated
1420 * to another CPU then no harm is done and the purpose has been
1423 void kick_process(struct task_struct
*p
)
1429 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1430 smp_send_reschedule(cpu
);
1433 EXPORT_SYMBOL_GPL(kick_process
);
1436 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1438 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1440 int nid
= cpu_to_node(cpu
);
1441 const struct cpumask
*nodemask
= NULL
;
1442 enum { cpuset
, possible
, fail
} state
= cpuset
;
1446 * If the node that the cpu is on has been offlined, cpu_to_node()
1447 * will return -1. There is no cpu on the node, and we should
1448 * select the cpu on the other node.
1451 nodemask
= cpumask_of_node(nid
);
1453 /* Look for allowed, online CPU in same node. */
1454 for_each_cpu(dest_cpu
, nodemask
) {
1455 if (!cpu_online(dest_cpu
))
1457 if (!cpu_active(dest_cpu
))
1459 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1465 /* Any allowed, online CPU? */
1466 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1467 if (!cpu_online(dest_cpu
))
1469 if (!cpu_active(dest_cpu
))
1474 /* No more Mr. Nice Guy. */
1477 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1478 cpuset_cpus_allowed_fallback(p
);
1484 do_set_cpus_allowed(p
, cpu_possible_mask
);
1495 if (state
!= cpuset
) {
1497 * Don't tell them about moving exiting tasks or
1498 * kernel threads (both mm NULL), since they never
1501 if (p
->mm
&& printk_ratelimit()) {
1502 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1503 task_pid_nr(p
), p
->comm
, cpu
);
1511 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1514 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1516 lockdep_assert_held(&p
->pi_lock
);
1518 if (p
->nr_cpus_allowed
> 1)
1519 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1522 * In order not to call set_task_cpu() on a blocking task we need
1523 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1526 * Since this is common to all placement strategies, this lives here.
1528 * [ this allows ->select_task() to simply return task_cpu(p) and
1529 * not worry about this generic constraint ]
1531 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1533 cpu
= select_fallback_rq(task_cpu(p
), p
);
1538 static void update_avg(u64
*avg
, u64 sample
)
1540 s64 diff
= sample
- *avg
;
1546 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1547 const struct cpumask
*new_mask
, bool check
)
1549 return set_cpus_allowed_ptr(p
, new_mask
);
1552 #endif /* CONFIG_SMP */
1555 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1557 #ifdef CONFIG_SCHEDSTATS
1558 struct rq
*rq
= this_rq();
1561 int this_cpu
= smp_processor_id();
1563 if (cpu
== this_cpu
) {
1564 schedstat_inc(rq
, ttwu_local
);
1565 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1567 struct sched_domain
*sd
;
1569 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1571 for_each_domain(this_cpu
, sd
) {
1572 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1573 schedstat_inc(sd
, ttwu_wake_remote
);
1580 if (wake_flags
& WF_MIGRATED
)
1581 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1583 #endif /* CONFIG_SMP */
1585 schedstat_inc(rq
, ttwu_count
);
1586 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1588 if (wake_flags
& WF_SYNC
)
1589 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1591 #endif /* CONFIG_SCHEDSTATS */
1594 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1596 activate_task(rq
, p
, en_flags
);
1597 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1599 /* if a worker is waking up, notify workqueue */
1600 if (p
->flags
& PF_WQ_WORKER
)
1601 wq_worker_waking_up(p
, cpu_of(rq
));
1605 * Mark the task runnable and perform wakeup-preemption.
1608 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1610 check_preempt_curr(rq
, p
, wake_flags
);
1611 p
->state
= TASK_RUNNING
;
1612 trace_sched_wakeup(p
);
1615 if (p
->sched_class
->task_woken
) {
1617 * Our task @p is fully woken up and running; so its safe to
1618 * drop the rq->lock, hereafter rq is only used for statistics.
1620 lockdep_unpin_lock(&rq
->lock
);
1621 p
->sched_class
->task_woken(rq
, p
);
1622 lockdep_pin_lock(&rq
->lock
);
1625 if (rq
->idle_stamp
) {
1626 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1627 u64 max
= 2*rq
->max_idle_balance_cost
;
1629 update_avg(&rq
->avg_idle
, delta
);
1631 if (rq
->avg_idle
> max
)
1640 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1642 lockdep_assert_held(&rq
->lock
);
1645 if (p
->sched_contributes_to_load
)
1646 rq
->nr_uninterruptible
--;
1649 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1650 ttwu_do_wakeup(rq
, p
, wake_flags
);
1654 * Called in case the task @p isn't fully descheduled from its runqueue,
1655 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1656 * since all we need to do is flip p->state to TASK_RUNNING, since
1657 * the task is still ->on_rq.
1659 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1664 rq
= __task_rq_lock(p
);
1665 if (task_on_rq_queued(p
)) {
1666 /* check_preempt_curr() may use rq clock */
1667 update_rq_clock(rq
);
1668 ttwu_do_wakeup(rq
, p
, wake_flags
);
1671 __task_rq_unlock(rq
);
1677 void sched_ttwu_pending(void)
1679 struct rq
*rq
= this_rq();
1680 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1681 struct task_struct
*p
;
1682 unsigned long flags
;
1687 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1688 lockdep_pin_lock(&rq
->lock
);
1691 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1692 llist
= llist_next(llist
);
1693 ttwu_do_activate(rq
, p
, 0);
1696 lockdep_unpin_lock(&rq
->lock
);
1697 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1700 void scheduler_ipi(void)
1703 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1704 * TIF_NEED_RESCHED remotely (for the first time) will also send
1707 preempt_fold_need_resched();
1709 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1713 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1714 * traditionally all their work was done from the interrupt return
1715 * path. Now that we actually do some work, we need to make sure
1718 * Some archs already do call them, luckily irq_enter/exit nest
1721 * Arguably we should visit all archs and update all handlers,
1722 * however a fair share of IPIs are still resched only so this would
1723 * somewhat pessimize the simple resched case.
1726 sched_ttwu_pending();
1729 * Check if someone kicked us for doing the nohz idle load balance.
1731 if (unlikely(got_nohz_idle_kick())) {
1732 this_rq()->idle_balance
= 1;
1733 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1738 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1740 struct rq
*rq
= cpu_rq(cpu
);
1742 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1743 if (!set_nr_if_polling(rq
->idle
))
1744 smp_send_reschedule(cpu
);
1746 trace_sched_wake_idle_without_ipi(cpu
);
1750 void wake_up_if_idle(int cpu
)
1752 struct rq
*rq
= cpu_rq(cpu
);
1753 unsigned long flags
;
1757 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1760 if (set_nr_if_polling(rq
->idle
)) {
1761 trace_sched_wake_idle_without_ipi(cpu
);
1763 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1764 if (is_idle_task(rq
->curr
))
1765 smp_send_reschedule(cpu
);
1766 /* Else cpu is not in idle, do nothing here */
1767 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1774 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1776 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1778 #endif /* CONFIG_SMP */
1780 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1782 struct rq
*rq
= cpu_rq(cpu
);
1784 #if defined(CONFIG_SMP)
1785 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1786 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1787 ttwu_queue_remote(p
, cpu
);
1792 raw_spin_lock(&rq
->lock
);
1793 lockdep_pin_lock(&rq
->lock
);
1794 ttwu_do_activate(rq
, p
, 0);
1795 lockdep_unpin_lock(&rq
->lock
);
1796 raw_spin_unlock(&rq
->lock
);
1800 * Notes on Program-Order guarantees on SMP systems.
1804 * The basic program-order guarantee on SMP systems is that when a task [t]
1805 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1806 * execution on its new cpu [c1].
1808 * For migration (of runnable tasks) this is provided by the following means:
1810 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1811 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1812 * rq(c1)->lock (if not at the same time, then in that order).
1813 * C) LOCK of the rq(c1)->lock scheduling in task
1815 * Transitivity guarantees that B happens after A and C after B.
1816 * Note: we only require RCpc transitivity.
1817 * Note: the cpu doing B need not be c0 or c1
1826 * UNLOCK rq(0)->lock
1828 * LOCK rq(0)->lock // orders against CPU0
1830 * UNLOCK rq(0)->lock
1834 * UNLOCK rq(1)->lock
1836 * LOCK rq(1)->lock // orders against CPU2
1839 * UNLOCK rq(1)->lock
1842 * BLOCKING -- aka. SLEEP + WAKEUP
1844 * For blocking we (obviously) need to provide the same guarantee as for
1845 * migration. However the means are completely different as there is no lock
1846 * chain to provide order. Instead we do:
1848 * 1) smp_store_release(X->on_cpu, 0)
1849 * 2) smp_cond_acquire(!X->on_cpu)
1853 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1855 * LOCK rq(0)->lock LOCK X->pi_lock
1858 * smp_store_release(X->on_cpu, 0);
1860 * smp_cond_acquire(!X->on_cpu);
1866 * X->state = RUNNING
1867 * UNLOCK rq(2)->lock
1869 * LOCK rq(2)->lock // orders against CPU1
1872 * UNLOCK rq(2)->lock
1875 * UNLOCK rq(0)->lock
1878 * However; for wakeups there is a second guarantee we must provide, namely we
1879 * must observe the state that lead to our wakeup. That is, not only must our
1880 * task observe its own prior state, it must also observe the stores prior to
1883 * This means that any means of doing remote wakeups must order the CPU doing
1884 * the wakeup against the CPU the task is going to end up running on. This,
1885 * however, is already required for the regular Program-Order guarantee above,
1886 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1891 * try_to_wake_up - wake up a thread
1892 * @p: the thread to be awakened
1893 * @state: the mask of task states that can be woken
1894 * @wake_flags: wake modifier flags (WF_*)
1896 * Put it on the run-queue if it's not already there. The "current"
1897 * thread is always on the run-queue (except when the actual
1898 * re-schedule is in progress), and as such you're allowed to do
1899 * the simpler "current->state = TASK_RUNNING" to mark yourself
1900 * runnable without the overhead of this.
1902 * Return: %true if @p was woken up, %false if it was already running.
1903 * or @state didn't match @p's state.
1906 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1908 unsigned long flags
;
1909 int cpu
, success
= 0;
1912 * If we are going to wake up a thread waiting for CONDITION we
1913 * need to ensure that CONDITION=1 done by the caller can not be
1914 * reordered with p->state check below. This pairs with mb() in
1915 * set_current_state() the waiting thread does.
1917 smp_mb__before_spinlock();
1918 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1919 if (!(p
->state
& state
))
1922 trace_sched_waking(p
);
1924 success
= 1; /* we're going to change ->state */
1927 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1932 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1933 * possible to, falsely, observe p->on_cpu == 0.
1935 * One must be running (->on_cpu == 1) in order to remove oneself
1936 * from the runqueue.
1938 * [S] ->on_cpu = 1; [L] ->on_rq
1942 * [S] ->on_rq = 0; [L] ->on_cpu
1944 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1945 * from the consecutive calls to schedule(); the first switching to our
1946 * task, the second putting it to sleep.
1951 * If the owning (remote) cpu is still in the middle of schedule() with
1952 * this task as prev, wait until its done referencing the task.
1954 * Pairs with the smp_store_release() in finish_lock_switch().
1956 * This ensures that tasks getting woken will be fully ordered against
1957 * their previous state and preserve Program Order.
1959 smp_cond_acquire(!p
->on_cpu
);
1961 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1962 p
->state
= TASK_WAKING
;
1964 if (p
->sched_class
->task_waking
)
1965 p
->sched_class
->task_waking(p
);
1967 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1968 if (task_cpu(p
) != cpu
) {
1969 wake_flags
|= WF_MIGRATED
;
1970 set_task_cpu(p
, cpu
);
1972 #endif /* CONFIG_SMP */
1976 if (schedstat_enabled())
1977 ttwu_stat(p
, cpu
, wake_flags
);
1979 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1985 * try_to_wake_up_local - try to wake up a local task with rq lock held
1986 * @p: the thread to be awakened
1988 * Put @p on the run-queue if it's not already there. The caller must
1989 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1992 static void try_to_wake_up_local(struct task_struct
*p
)
1994 struct rq
*rq
= task_rq(p
);
1996 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1997 WARN_ON_ONCE(p
== current
))
2000 lockdep_assert_held(&rq
->lock
);
2002 if (!raw_spin_trylock(&p
->pi_lock
)) {
2004 * This is OK, because current is on_cpu, which avoids it being
2005 * picked for load-balance and preemption/IRQs are still
2006 * disabled avoiding further scheduler activity on it and we've
2007 * not yet picked a replacement task.
2009 lockdep_unpin_lock(&rq
->lock
);
2010 raw_spin_unlock(&rq
->lock
);
2011 raw_spin_lock(&p
->pi_lock
);
2012 raw_spin_lock(&rq
->lock
);
2013 lockdep_pin_lock(&rq
->lock
);
2016 if (!(p
->state
& TASK_NORMAL
))
2019 trace_sched_waking(p
);
2021 if (!task_on_rq_queued(p
))
2022 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2024 ttwu_do_wakeup(rq
, p
, 0);
2025 if (schedstat_enabled())
2026 ttwu_stat(p
, smp_processor_id(), 0);
2028 raw_spin_unlock(&p
->pi_lock
);
2032 * wake_up_process - Wake up a specific process
2033 * @p: The process to be woken up.
2035 * Attempt to wake up the nominated process and move it to the set of runnable
2038 * Return: 1 if the process was woken up, 0 if it was already running.
2040 * It may be assumed that this function implies a write memory barrier before
2041 * changing the task state if and only if any tasks are woken up.
2043 int wake_up_process(struct task_struct
*p
)
2045 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2047 EXPORT_SYMBOL(wake_up_process
);
2049 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2051 return try_to_wake_up(p
, state
, 0);
2055 * This function clears the sched_dl_entity static params.
2057 void __dl_clear_params(struct task_struct
*p
)
2059 struct sched_dl_entity
*dl_se
= &p
->dl
;
2061 dl_se
->dl_runtime
= 0;
2062 dl_se
->dl_deadline
= 0;
2063 dl_se
->dl_period
= 0;
2067 dl_se
->dl_throttled
= 0;
2068 dl_se
->dl_yielded
= 0;
2072 * Perform scheduler related setup for a newly forked process p.
2073 * p is forked by current.
2075 * __sched_fork() is basic setup used by init_idle() too:
2077 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2082 p
->se
.exec_start
= 0;
2083 p
->se
.sum_exec_runtime
= 0;
2084 p
->se
.prev_sum_exec_runtime
= 0;
2085 p
->se
.nr_migrations
= 0;
2087 INIT_LIST_HEAD(&p
->se
.group_node
);
2089 #ifdef CONFIG_FAIR_GROUP_SCHED
2090 p
->se
.cfs_rq
= NULL
;
2093 #ifdef CONFIG_SCHEDSTATS
2094 /* Even if schedstat is disabled, there should not be garbage */
2095 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2098 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2099 init_dl_task_timer(&p
->dl
);
2100 __dl_clear_params(p
);
2102 INIT_LIST_HEAD(&p
->rt
.run_list
);
2104 p
->rt
.time_slice
= sched_rr_timeslice
;
2108 #ifdef CONFIG_PREEMPT_NOTIFIERS
2109 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2112 #ifdef CONFIG_NUMA_BALANCING
2113 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2114 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2115 p
->mm
->numa_scan_seq
= 0;
2118 if (clone_flags
& CLONE_VM
)
2119 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2121 p
->numa_preferred_nid
= -1;
2123 p
->node_stamp
= 0ULL;
2124 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2125 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2126 p
->numa_work
.next
= &p
->numa_work
;
2127 p
->numa_faults
= NULL
;
2128 p
->last_task_numa_placement
= 0;
2129 p
->last_sum_exec_runtime
= 0;
2131 p
->numa_group
= NULL
;
2132 #endif /* CONFIG_NUMA_BALANCING */
2135 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2137 #ifdef CONFIG_NUMA_BALANCING
2139 void set_numabalancing_state(bool enabled
)
2142 static_branch_enable(&sched_numa_balancing
);
2144 static_branch_disable(&sched_numa_balancing
);
2147 #ifdef CONFIG_PROC_SYSCTL
2148 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2149 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2153 int state
= static_branch_likely(&sched_numa_balancing
);
2155 if (write
&& !capable(CAP_SYS_ADMIN
))
2160 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2164 set_numabalancing_state(state
);
2170 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2172 #ifdef CONFIG_SCHEDSTATS
2173 static void set_schedstats(bool enabled
)
2176 static_branch_enable(&sched_schedstats
);
2178 static_branch_disable(&sched_schedstats
);
2181 void force_schedstat_enabled(void)
2183 if (!schedstat_enabled()) {
2184 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2185 static_branch_enable(&sched_schedstats
);
2189 static int __init
setup_schedstats(char *str
)
2195 if (!strcmp(str
, "enable")) {
2196 set_schedstats(true);
2198 } else if (!strcmp(str
, "disable")) {
2199 set_schedstats(false);
2204 pr_warn("Unable to parse schedstats=\n");
2208 __setup("schedstats=", setup_schedstats
);
2210 #ifdef CONFIG_PROC_SYSCTL
2211 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2212 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2216 int state
= static_branch_likely(&sched_schedstats
);
2218 if (write
&& !capable(CAP_SYS_ADMIN
))
2223 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2227 set_schedstats(state
);
2233 static void sched_set_prio(struct task_struct
*p
, int prio
)
2239 * fork()/clone()-time setup:
2241 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2243 unsigned long flags
;
2244 int cpu
= get_cpu();
2246 __sched_fork(clone_flags
, p
);
2248 * We mark the process as running here. This guarantees that
2249 * nobody will actually run it, and a signal or other external
2250 * event cannot wake it up and insert it on the runqueue either.
2252 p
->state
= TASK_RUNNING
;
2255 * Make sure we do not leak PI boosting priority to the child.
2257 sched_set_prio(p
, current
->normal_prio
);
2260 * Revert to default priority/policy on fork if requested.
2262 if (unlikely(p
->sched_reset_on_fork
)) {
2263 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2264 p
->policy
= SCHED_NORMAL
;
2265 p
->static_prio
= NICE_TO_PRIO(0);
2267 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2268 p
->static_prio
= NICE_TO_PRIO(0);
2270 p
->normal_prio
= __normal_prio(p
);
2271 sched_set_prio(p
, p
->normal_prio
);
2275 * We don't need the reset flag anymore after the fork. It has
2276 * fulfilled its duty:
2278 p
->sched_reset_on_fork
= 0;
2281 if (dl_prio(p
->prio
)) {
2284 } else if (rt_prio(p
->prio
)) {
2285 p
->sched_class
= &rt_sched_class
;
2287 p
->sched_class
= &fair_sched_class
;
2290 if (p
->sched_class
->task_fork
)
2291 p
->sched_class
->task_fork(p
);
2294 * The child is not yet in the pid-hash so no cgroup attach races,
2295 * and the cgroup is pinned to this child due to cgroup_fork()
2296 * is ran before sched_fork().
2298 * Silence PROVE_RCU.
2300 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2301 set_task_cpu(p
, cpu
);
2302 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2304 #ifdef CONFIG_SCHED_INFO
2305 if (likely(sched_info_on()))
2306 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2308 #if defined(CONFIG_SMP)
2311 init_task_preempt_count(p
);
2313 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2314 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2321 unsigned long to_ratio(u64 period
, u64 runtime
)
2323 if (runtime
== RUNTIME_INF
)
2327 * Doing this here saves a lot of checks in all
2328 * the calling paths, and returning zero seems
2329 * safe for them anyway.
2334 return div64_u64(runtime
<< 20, period
);
2338 inline struct dl_bw
*dl_bw_of(int i
)
2340 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2341 "sched RCU must be held");
2342 return &cpu_rq(i
)->rd
->dl_bw
;
2345 static inline int dl_bw_cpus(int i
)
2347 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2350 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2351 "sched RCU must be held");
2352 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2358 inline struct dl_bw
*dl_bw_of(int i
)
2360 return &cpu_rq(i
)->dl
.dl_bw
;
2363 static inline int dl_bw_cpus(int i
)
2370 * We must be sure that accepting a new task (or allowing changing the
2371 * parameters of an existing one) is consistent with the bandwidth
2372 * constraints. If yes, this function also accordingly updates the currently
2373 * allocated bandwidth to reflect the new situation.
2375 * This function is called while holding p's rq->lock.
2377 * XXX we should delay bw change until the task's 0-lag point, see
2380 static int dl_overflow(struct task_struct
*p
, int policy
,
2381 const struct sched_attr
*attr
)
2384 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2385 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2386 u64 runtime
= attr
->sched_runtime
;
2387 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2390 if (new_bw
== p
->dl
.dl_bw
)
2394 * Either if a task, enters, leave, or stays -deadline but changes
2395 * its parameters, we may need to update accordingly the total
2396 * allocated bandwidth of the container.
2398 raw_spin_lock(&dl_b
->lock
);
2399 cpus
= dl_bw_cpus(task_cpu(p
));
2400 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2401 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2402 __dl_add(dl_b
, new_bw
);
2404 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2405 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2406 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2407 __dl_add(dl_b
, new_bw
);
2409 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2410 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2413 raw_spin_unlock(&dl_b
->lock
);
2418 extern void init_dl_bw(struct dl_bw
*dl_b
);
2421 * wake_up_new_task - wake up a newly created task for the first time.
2423 * This function will do some initial scheduler statistics housekeeping
2424 * that must be done for every newly created context, then puts the task
2425 * on the runqueue and wakes it.
2427 void wake_up_new_task(struct task_struct
*p
)
2429 unsigned long flags
;
2432 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2433 /* Initialize new task's runnable average */
2434 init_entity_runnable_average(&p
->se
);
2437 * Fork balancing, do it here and not earlier because:
2438 * - cpus_allowed can change in the fork path
2439 * - any previously selected cpu might disappear through hotplug
2441 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2444 rq
= __task_rq_lock(p
);
2445 activate_task(rq
, p
, 0);
2446 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2447 trace_sched_wakeup_new(p
);
2448 check_preempt_curr(rq
, p
, WF_FORK
);
2450 if (p
->sched_class
->task_woken
) {
2452 * Nothing relies on rq->lock after this, so its fine to
2455 lockdep_unpin_lock(&rq
->lock
);
2456 p
->sched_class
->task_woken(rq
, p
);
2457 lockdep_pin_lock(&rq
->lock
);
2460 task_rq_unlock(rq
, p
, &flags
);
2463 #ifdef CONFIG_PREEMPT_NOTIFIERS
2465 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2467 void preempt_notifier_inc(void)
2469 static_key_slow_inc(&preempt_notifier_key
);
2471 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2473 void preempt_notifier_dec(void)
2475 static_key_slow_dec(&preempt_notifier_key
);
2477 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2480 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2481 * @notifier: notifier struct to register
2483 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2485 if (!static_key_false(&preempt_notifier_key
))
2486 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2488 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2490 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2493 * preempt_notifier_unregister - no longer interested in preemption notifications
2494 * @notifier: notifier struct to unregister
2496 * This is *not* safe to call from within a preemption notifier.
2498 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2500 hlist_del(¬ifier
->link
);
2502 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2504 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2506 struct preempt_notifier
*notifier
;
2508 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2509 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2512 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2514 if (static_key_false(&preempt_notifier_key
))
2515 __fire_sched_in_preempt_notifiers(curr
);
2519 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2520 struct task_struct
*next
)
2522 struct preempt_notifier
*notifier
;
2524 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2525 notifier
->ops
->sched_out(notifier
, next
);
2528 static __always_inline
void
2529 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2530 struct task_struct
*next
)
2532 if (static_key_false(&preempt_notifier_key
))
2533 __fire_sched_out_preempt_notifiers(curr
, next
);
2536 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2538 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2543 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2544 struct task_struct
*next
)
2548 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2551 * prepare_task_switch - prepare to switch tasks
2552 * @rq: the runqueue preparing to switch
2553 * @prev: the current task that is being switched out
2554 * @next: the task we are going to switch to.
2556 * This is called with the rq lock held and interrupts off. It must
2557 * be paired with a subsequent finish_task_switch after the context
2560 * prepare_task_switch sets up locking and calls architecture specific
2564 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2565 struct task_struct
*next
)
2567 sched_info_switch(rq
, prev
, next
);
2568 perf_event_task_sched_out(prev
, next
);
2569 fire_sched_out_preempt_notifiers(prev
, next
);
2570 prepare_lock_switch(rq
, next
);
2571 prepare_arch_switch(next
);
2575 * finish_task_switch - clean up after a task-switch
2576 * @prev: the thread we just switched away from.
2578 * finish_task_switch must be called after the context switch, paired
2579 * with a prepare_task_switch call before the context switch.
2580 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2581 * and do any other architecture-specific cleanup actions.
2583 * Note that we may have delayed dropping an mm in context_switch(). If
2584 * so, we finish that here outside of the runqueue lock. (Doing it
2585 * with the lock held can cause deadlocks; see schedule() for
2588 * The context switch have flipped the stack from under us and restored the
2589 * local variables which were saved when this task called schedule() in the
2590 * past. prev == current is still correct but we need to recalculate this_rq
2591 * because prev may have moved to another CPU.
2593 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2594 __releases(rq
->lock
)
2596 struct rq
*rq
= this_rq();
2597 struct mm_struct
*mm
= rq
->prev_mm
;
2601 * The previous task will have left us with a preempt_count of 2
2602 * because it left us after:
2605 * preempt_disable(); // 1
2607 * raw_spin_lock_irq(&rq->lock) // 2
2609 * Also, see FORK_PREEMPT_COUNT.
2611 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2612 "corrupted preempt_count: %s/%d/0x%x\n",
2613 current
->comm
, current
->pid
, preempt_count()))
2614 preempt_count_set(FORK_PREEMPT_COUNT
);
2619 * A task struct has one reference for the use as "current".
2620 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2621 * schedule one last time. The schedule call will never return, and
2622 * the scheduled task must drop that reference.
2624 * We must observe prev->state before clearing prev->on_cpu (in
2625 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2626 * running on another CPU and we could rave with its RUNNING -> DEAD
2627 * transition, resulting in a double drop.
2629 prev_state
= prev
->state
;
2630 vtime_task_switch(prev
);
2631 perf_event_task_sched_in(prev
, current
);
2632 finish_lock_switch(rq
, prev
);
2633 finish_arch_post_lock_switch();
2635 fire_sched_in_preempt_notifiers(current
);
2638 if (unlikely(prev_state
== TASK_DEAD
)) {
2639 if (prev
->sched_class
->task_dead
)
2640 prev
->sched_class
->task_dead(prev
);
2643 * Remove function-return probe instances associated with this
2644 * task and put them back on the free list.
2646 kprobe_flush_task(prev
);
2647 put_task_struct(prev
);
2650 tick_nohz_task_switch();
2656 /* rq->lock is NOT held, but preemption is disabled */
2657 static void __balance_callback(struct rq
*rq
)
2659 struct callback_head
*head
, *next
;
2660 void (*func
)(struct rq
*rq
);
2661 unsigned long flags
;
2663 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2664 head
= rq
->balance_callback
;
2665 rq
->balance_callback
= NULL
;
2667 func
= (void (*)(struct rq
*))head
->func
;
2674 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2677 static inline void balance_callback(struct rq
*rq
)
2679 if (unlikely(rq
->balance_callback
))
2680 __balance_callback(rq
);
2685 static inline void balance_callback(struct rq
*rq
)
2692 * schedule_tail - first thing a freshly forked thread must call.
2693 * @prev: the thread we just switched away from.
2695 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2696 __releases(rq
->lock
)
2701 * New tasks start with FORK_PREEMPT_COUNT, see there and
2702 * finish_task_switch() for details.
2704 * finish_task_switch() will drop rq->lock() and lower preempt_count
2705 * and the preempt_enable() will end up enabling preemption (on
2706 * PREEMPT_COUNT kernels).
2709 rq
= finish_task_switch(prev
);
2710 balance_callback(rq
);
2713 if (current
->set_child_tid
)
2714 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2718 * context_switch - switch to the new MM and the new thread's register state.
2720 static __always_inline
struct rq
*
2721 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2722 struct task_struct
*next
)
2724 struct mm_struct
*mm
, *oldmm
;
2726 prepare_task_switch(rq
, prev
, next
);
2729 oldmm
= prev
->active_mm
;
2731 * For paravirt, this is coupled with an exit in switch_to to
2732 * combine the page table reload and the switch backend into
2735 arch_start_context_switch(prev
);
2738 next
->active_mm
= oldmm
;
2739 atomic_inc(&oldmm
->mm_count
);
2740 enter_lazy_tlb(oldmm
, next
);
2742 switch_mm(oldmm
, mm
, next
);
2745 prev
->active_mm
= NULL
;
2746 rq
->prev_mm
= oldmm
;
2749 * Since the runqueue lock will be released by the next
2750 * task (which is an invalid locking op but in the case
2751 * of the scheduler it's an obvious special-case), so we
2752 * do an early lockdep release here:
2754 lockdep_unpin_lock(&rq
->lock
);
2755 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2757 /* Here we just switch the register state and the stack. */
2758 switch_to(prev
, next
, prev
);
2761 return finish_task_switch(prev
);
2765 * nr_running and nr_context_switches:
2767 * externally visible scheduler statistics: current number of runnable
2768 * threads, total number of context switches performed since bootup.
2770 unsigned long nr_running(void)
2772 unsigned long i
, sum
= 0;
2774 for_each_online_cpu(i
)
2775 sum
+= cpu_rq(i
)->nr_running
;
2781 * Check if only the current task is running on the cpu.
2783 * Caution: this function does not check that the caller has disabled
2784 * preemption, thus the result might have a time-of-check-to-time-of-use
2785 * race. The caller is responsible to use it correctly, for example:
2787 * - from a non-preemptable section (of course)
2789 * - from a thread that is bound to a single CPU
2791 * - in a loop with very short iterations (e.g. a polling loop)
2793 bool single_task_running(void)
2795 return raw_rq()->nr_running
== 1;
2797 EXPORT_SYMBOL(single_task_running
);
2799 unsigned long long nr_context_switches(void)
2802 unsigned long long sum
= 0;
2804 for_each_possible_cpu(i
)
2805 sum
+= cpu_rq(i
)->nr_switches
;
2810 unsigned long nr_iowait(void)
2812 unsigned long i
, sum
= 0;
2814 for_each_possible_cpu(i
)
2815 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2820 unsigned long nr_iowait_cpu(int cpu
)
2822 struct rq
*this = cpu_rq(cpu
);
2823 return atomic_read(&this->nr_iowait
);
2826 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2828 struct rq
*rq
= this_rq();
2829 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2830 *load
= rq
->load
.weight
;
2836 * sched_exec - execve() is a valuable balancing opportunity, because at
2837 * this point the task has the smallest effective memory and cache footprint.
2839 void sched_exec(void)
2841 struct task_struct
*p
= current
;
2842 unsigned long flags
;
2845 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2846 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2847 if (dest_cpu
== smp_processor_id())
2850 if (likely(cpu_active(dest_cpu
))) {
2851 struct migration_arg arg
= { p
, dest_cpu
};
2853 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2854 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2858 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2863 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2864 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2866 EXPORT_PER_CPU_SYMBOL(kstat
);
2867 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2870 * Return accounted runtime for the task.
2871 * In case the task is currently running, return the runtime plus current's
2872 * pending runtime that have not been accounted yet.
2874 unsigned long long task_sched_runtime(struct task_struct
*p
)
2876 unsigned long flags
;
2880 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2882 * 64-bit doesn't need locks to atomically read a 64bit value.
2883 * So we have a optimization chance when the task's delta_exec is 0.
2884 * Reading ->on_cpu is racy, but this is ok.
2886 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2887 * If we race with it entering cpu, unaccounted time is 0. This is
2888 * indistinguishable from the read occurring a few cycles earlier.
2889 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2890 * been accounted, so we're correct here as well.
2892 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2893 return p
->se
.sum_exec_runtime
;
2896 rq
= task_rq_lock(p
, &flags
);
2898 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2899 * project cycles that may never be accounted to this
2900 * thread, breaking clock_gettime().
2902 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2903 update_rq_clock(rq
);
2904 p
->sched_class
->update_curr(rq
);
2906 ns
= p
->se
.sum_exec_runtime
;
2907 task_rq_unlock(rq
, p
, &flags
);
2913 * This function gets called by the timer code, with HZ frequency.
2914 * We call it with interrupts disabled.
2916 void scheduler_tick(void)
2918 int cpu
= smp_processor_id();
2919 struct rq
*rq
= cpu_rq(cpu
);
2920 struct task_struct
*curr
= rq
->curr
;
2924 raw_spin_lock(&rq
->lock
);
2925 update_rq_clock(rq
);
2926 curr
->sched_class
->task_tick(rq
, curr
, 0);
2927 update_cpu_load_active(rq
);
2928 calc_global_load_tick(rq
);
2929 raw_spin_unlock(&rq
->lock
);
2931 perf_event_task_tick();
2934 rq
->idle_balance
= idle_cpu(cpu
);
2935 trigger_load_balance(rq
);
2937 rq_last_tick_reset(rq
);
2940 #ifdef CONFIG_NO_HZ_FULL
2942 * scheduler_tick_max_deferment
2944 * Keep at least one tick per second when a single
2945 * active task is running because the scheduler doesn't
2946 * yet completely support full dynticks environment.
2948 * This makes sure that uptime, CFS vruntime, load
2949 * balancing, etc... continue to move forward, even
2950 * with a very low granularity.
2952 * Return: Maximum deferment in nanoseconds.
2954 u64
scheduler_tick_max_deferment(void)
2956 struct rq
*rq
= this_rq();
2957 unsigned long next
, now
= READ_ONCE(jiffies
);
2959 next
= rq
->last_sched_tick
+ HZ
;
2961 if (time_before_eq(next
, now
))
2964 return jiffies_to_nsecs(next
- now
);
2968 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2969 defined(CONFIG_PREEMPT_TRACER))
2971 void preempt_count_add(int val
)
2973 #ifdef CONFIG_DEBUG_PREEMPT
2977 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2980 __preempt_count_add(val
);
2981 #ifdef CONFIG_DEBUG_PREEMPT
2983 * Spinlock count overflowing soon?
2985 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2988 if (preempt_count() == val
) {
2989 unsigned long ip
= get_lock_parent_ip();
2990 #ifdef CONFIG_DEBUG_PREEMPT
2991 current
->preempt_disable_ip
= ip
;
2993 trace_preempt_off(CALLER_ADDR0
, ip
);
2996 EXPORT_SYMBOL(preempt_count_add
);
2997 NOKPROBE_SYMBOL(preempt_count_add
);
2999 void preempt_count_sub(int val
)
3001 #ifdef CONFIG_DEBUG_PREEMPT
3005 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3008 * Is the spinlock portion underflowing?
3010 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3011 !(preempt_count() & PREEMPT_MASK
)))
3015 if (preempt_count() == val
)
3016 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3017 __preempt_count_sub(val
);
3019 EXPORT_SYMBOL(preempt_count_sub
);
3020 NOKPROBE_SYMBOL(preempt_count_sub
);
3025 * Print scheduling while atomic bug:
3027 static noinline
void __schedule_bug(struct task_struct
*prev
)
3029 if (oops_in_progress
)
3032 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3033 prev
->comm
, prev
->pid
, preempt_count());
3035 debug_show_held_locks(prev
);
3037 if (irqs_disabled())
3038 print_irqtrace_events(prev
);
3039 #ifdef CONFIG_DEBUG_PREEMPT
3040 if (in_atomic_preempt_off()) {
3041 pr_err("Preemption disabled at:");
3042 print_ip_sym(current
->preempt_disable_ip
);
3047 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3051 * Various schedule()-time debugging checks and statistics:
3053 static inline void schedule_debug(struct task_struct
*prev
)
3055 #ifdef CONFIG_SCHED_STACK_END_CHECK
3056 BUG_ON(task_stack_end_corrupted(prev
));
3059 if (unlikely(in_atomic_preempt_off())) {
3060 __schedule_bug(prev
);
3061 preempt_count_set(PREEMPT_DISABLED
);
3065 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3067 schedstat_inc(this_rq(), sched_count
);
3071 * Pick up the highest-prio task:
3073 static inline struct task_struct
*
3074 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3076 const struct sched_class
*class = &fair_sched_class
;
3077 struct task_struct
*p
;
3080 * Optimization: we know that if all tasks are in
3081 * the fair class we can call that function directly:
3083 if (likely(prev
->sched_class
== class &&
3084 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3085 p
= fair_sched_class
.pick_next_task(rq
, prev
);
3086 if (unlikely(p
== RETRY_TASK
))
3089 /* assumes fair_sched_class->next == idle_sched_class */
3091 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3097 for_each_class(class) {
3098 p
= class->pick_next_task(rq
, prev
);
3100 if (unlikely(p
== RETRY_TASK
))
3106 BUG(); /* the idle class will always have a runnable task */
3110 * __schedule() is the main scheduler function.
3112 * The main means of driving the scheduler and thus entering this function are:
3114 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3116 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3117 * paths. For example, see arch/x86/entry_64.S.
3119 * To drive preemption between tasks, the scheduler sets the flag in timer
3120 * interrupt handler scheduler_tick().
3122 * 3. Wakeups don't really cause entry into schedule(). They add a
3123 * task to the run-queue and that's it.
3125 * Now, if the new task added to the run-queue preempts the current
3126 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3127 * called on the nearest possible occasion:
3129 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3131 * - in syscall or exception context, at the next outmost
3132 * preempt_enable(). (this might be as soon as the wake_up()'s
3135 * - in IRQ context, return from interrupt-handler to
3136 * preemptible context
3138 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3141 * - cond_resched() call
3142 * - explicit schedule() call
3143 * - return from syscall or exception to user-space
3144 * - return from interrupt-handler to user-space
3146 * WARNING: must be called with preemption disabled!
3148 static void __sched notrace
__schedule(bool preempt
)
3150 struct task_struct
*prev
, *next
;
3151 unsigned long *switch_count
;
3155 cpu
= smp_processor_id();
3160 * do_exit() calls schedule() with preemption disabled as an exception;
3161 * however we must fix that up, otherwise the next task will see an
3162 * inconsistent (higher) preempt count.
3164 * It also avoids the below schedule_debug() test from complaining
3167 if (unlikely(prev
->state
== TASK_DEAD
))
3168 preempt_enable_no_resched_notrace();
3170 schedule_debug(prev
);
3172 if (sched_feat(HRTICK
))
3175 local_irq_disable();
3176 rcu_note_context_switch();
3179 * Make sure that signal_pending_state()->signal_pending() below
3180 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3181 * done by the caller to avoid the race with signal_wake_up().
3183 smp_mb__before_spinlock();
3184 raw_spin_lock(&rq
->lock
);
3185 lockdep_pin_lock(&rq
->lock
);
3187 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3189 switch_count
= &prev
->nivcsw
;
3190 if (!preempt
&& prev
->state
) {
3191 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3192 prev
->state
= TASK_RUNNING
;
3194 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3198 * If a worker went to sleep, notify and ask workqueue
3199 * whether it wants to wake up a task to maintain
3202 if (prev
->flags
& PF_WQ_WORKER
) {
3203 struct task_struct
*to_wakeup
;
3205 to_wakeup
= wq_worker_sleeping(prev
);
3207 try_to_wake_up_local(to_wakeup
);
3210 switch_count
= &prev
->nvcsw
;
3213 if (task_on_rq_queued(prev
))
3214 update_rq_clock(rq
);
3216 next
= pick_next_task(rq
, prev
);
3217 clear_tsk_need_resched(prev
);
3218 clear_preempt_need_resched();
3219 rq
->clock_skip_update
= 0;
3221 if (likely(prev
!= next
)) {
3226 trace_sched_switch(preempt
, prev
, next
);
3227 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3229 lockdep_unpin_lock(&rq
->lock
);
3230 raw_spin_unlock_irq(&rq
->lock
);
3233 balance_callback(rq
);
3235 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3237 static inline void sched_submit_work(struct task_struct
*tsk
)
3239 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3242 * If we are going to sleep and we have plugged IO queued,
3243 * make sure to submit it to avoid deadlocks.
3245 if (blk_needs_flush_plug(tsk
))
3246 blk_schedule_flush_plug(tsk
);
3249 asmlinkage __visible
void __sched
schedule(void)
3251 struct task_struct
*tsk
= current
;
3253 sched_submit_work(tsk
);
3257 sched_preempt_enable_no_resched();
3258 } while (need_resched());
3260 EXPORT_SYMBOL(schedule
);
3262 #ifdef CONFIG_CONTEXT_TRACKING
3263 asmlinkage __visible
void __sched
schedule_user(void)
3266 * If we come here after a random call to set_need_resched(),
3267 * or we have been woken up remotely but the IPI has not yet arrived,
3268 * we haven't yet exited the RCU idle mode. Do it here manually until
3269 * we find a better solution.
3271 * NB: There are buggy callers of this function. Ideally we
3272 * should warn if prev_state != CONTEXT_USER, but that will trigger
3273 * too frequently to make sense yet.
3275 enum ctx_state prev_state
= exception_enter();
3277 exception_exit(prev_state
);
3282 * schedule_preempt_disabled - called with preemption disabled
3284 * Returns with preemption disabled. Note: preempt_count must be 1
3286 void __sched
schedule_preempt_disabled(void)
3288 sched_preempt_enable_no_resched();
3293 static void __sched notrace
preempt_schedule_common(void)
3296 preempt_disable_notrace();
3298 preempt_enable_no_resched_notrace();
3301 * Check again in case we missed a preemption opportunity
3302 * between schedule and now.
3304 } while (need_resched());
3307 #ifdef CONFIG_PREEMPT
3309 * this is the entry point to schedule() from in-kernel preemption
3310 * off of preempt_enable. Kernel preemptions off return from interrupt
3311 * occur there and call schedule directly.
3313 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3316 * If there is a non-zero preempt_count or interrupts are disabled,
3317 * we do not want to preempt the current task. Just return..
3319 if (likely(!preemptible()))
3322 preempt_schedule_common();
3324 NOKPROBE_SYMBOL(preempt_schedule
);
3325 EXPORT_SYMBOL(preempt_schedule
);
3328 * preempt_schedule_notrace - preempt_schedule called by tracing
3330 * The tracing infrastructure uses preempt_enable_notrace to prevent
3331 * recursion and tracing preempt enabling caused by the tracing
3332 * infrastructure itself. But as tracing can happen in areas coming
3333 * from userspace or just about to enter userspace, a preempt enable
3334 * can occur before user_exit() is called. This will cause the scheduler
3335 * to be called when the system is still in usermode.
3337 * To prevent this, the preempt_enable_notrace will use this function
3338 * instead of preempt_schedule() to exit user context if needed before
3339 * calling the scheduler.
3341 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3343 enum ctx_state prev_ctx
;
3345 if (likely(!preemptible()))
3349 preempt_disable_notrace();
3351 * Needs preempt disabled in case user_exit() is traced
3352 * and the tracer calls preempt_enable_notrace() causing
3353 * an infinite recursion.
3355 prev_ctx
= exception_enter();
3357 exception_exit(prev_ctx
);
3359 preempt_enable_no_resched_notrace();
3360 } while (need_resched());
3362 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3364 #endif /* CONFIG_PREEMPT */
3367 * this is the entry point to schedule() from kernel preemption
3368 * off of irq context.
3369 * Note, that this is called and return with irqs disabled. This will
3370 * protect us against recursive calling from irq.
3372 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3374 enum ctx_state prev_state
;
3376 /* Catch callers which need to be fixed */
3377 BUG_ON(preempt_count() || !irqs_disabled());
3379 prev_state
= exception_enter();
3385 local_irq_disable();
3386 sched_preempt_enable_no_resched();
3387 } while (need_resched());
3389 exception_exit(prev_state
);
3392 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3395 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3397 EXPORT_SYMBOL(default_wake_function
);
3399 #ifdef CONFIG_RT_MUTEXES
3402 * rt_mutex_setprio - set the current priority of a task
3404 * @prio: prio value (kernel-internal form)
3406 * This function changes the 'effective' priority of a task. It does
3407 * not touch ->normal_prio like __setscheduler().
3409 * Used by the rt_mutex code to implement priority inheritance
3410 * logic. Call site only calls if the priority of the task changed.
3412 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3414 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3416 const struct sched_class
*prev_class
;
3418 BUG_ON(prio
> MAX_PRIO
);
3420 rq
= __task_rq_lock(p
);
3423 * Idle task boosting is a nono in general. There is one
3424 * exception, when PREEMPT_RT and NOHZ is active:
3426 * The idle task calls get_next_timer_interrupt() and holds
3427 * the timer wheel base->lock on the CPU and another CPU wants
3428 * to access the timer (probably to cancel it). We can safely
3429 * ignore the boosting request, as the idle CPU runs this code
3430 * with interrupts disabled and will complete the lock
3431 * protected section without being interrupted. So there is no
3432 * real need to boost.
3434 if (unlikely(p
== rq
->idle
)) {
3435 WARN_ON(p
!= rq
->curr
);
3436 WARN_ON(p
->pi_blocked_on
);
3440 trace_sched_pi_setprio(p
, prio
);
3443 if (oldprio
== prio
)
3444 queue_flag
&= ~DEQUEUE_MOVE
;
3446 prev_class
= p
->sched_class
;
3447 queued
= task_on_rq_queued(p
);
3448 running
= task_current(rq
, p
);
3450 dequeue_task(rq
, p
, queue_flag
);
3452 put_prev_task(rq
, p
);
3455 * Boosting condition are:
3456 * 1. -rt task is running and holds mutex A
3457 * --> -dl task blocks on mutex A
3459 * 2. -dl task is running and holds mutex A
3460 * --> -dl task blocks on mutex A and could preempt the
3463 if (dl_prio(prio
)) {
3464 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3465 if (!dl_prio(p
->normal_prio
) ||
3466 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3467 p
->dl
.dl_boosted
= 1;
3468 queue_flag
|= ENQUEUE_REPLENISH
;
3470 p
->dl
.dl_boosted
= 0;
3471 p
->sched_class
= &dl_sched_class
;
3472 } else if (rt_prio(prio
)) {
3473 if (dl_prio(oldprio
))
3474 p
->dl
.dl_boosted
= 0;
3476 queue_flag
|= ENQUEUE_HEAD
;
3477 p
->sched_class
= &rt_sched_class
;
3479 if (dl_prio(oldprio
))
3480 p
->dl
.dl_boosted
= 0;
3481 if (rt_prio(oldprio
))
3483 p
->sched_class
= &fair_sched_class
;
3486 sched_set_prio(p
, prio
);
3489 p
->sched_class
->set_curr_task(rq
);
3491 enqueue_task(rq
, p
, queue_flag
);
3493 check_class_changed(rq
, p
, prev_class
, oldprio
);
3495 preempt_disable(); /* avoid rq from going away on us */
3496 __task_rq_unlock(rq
);
3498 balance_callback(rq
);
3503 void set_user_nice(struct task_struct
*p
, long nice
)
3505 int old_prio
, delta
, queued
;
3506 unsigned long flags
;
3509 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3512 * We have to be careful, if called from sys_setpriority(),
3513 * the task might be in the middle of scheduling on another CPU.
3515 rq
= task_rq_lock(p
, &flags
);
3517 * The RT priorities are set via sched_setscheduler(), but we still
3518 * allow the 'normal' nice value to be set - but as expected
3519 * it wont have any effect on scheduling until the task is
3520 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3522 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3523 p
->static_prio
= NICE_TO_PRIO(nice
);
3526 queued
= task_on_rq_queued(p
);
3528 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3530 p
->static_prio
= NICE_TO_PRIO(nice
);
3533 sched_set_prio(p
, effective_prio(p
));
3534 delta
= p
->prio
- old_prio
;
3537 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3539 * If the task increased its priority or is running and
3540 * lowered its priority, then reschedule its CPU:
3542 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3546 task_rq_unlock(rq
, p
, &flags
);
3548 EXPORT_SYMBOL(set_user_nice
);
3551 * can_nice - check if a task can reduce its nice value
3555 int can_nice(const struct task_struct
*p
, const int nice
)
3557 /* convert nice value [19,-20] to rlimit style value [1,40] */
3558 int nice_rlim
= nice_to_rlimit(nice
);
3560 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3561 capable(CAP_SYS_NICE
));
3564 #ifdef __ARCH_WANT_SYS_NICE
3567 * sys_nice - change the priority of the current process.
3568 * @increment: priority increment
3570 * sys_setpriority is a more generic, but much slower function that
3571 * does similar things.
3573 SYSCALL_DEFINE1(nice
, int, increment
)
3578 * Setpriority might change our priority at the same moment.
3579 * We don't have to worry. Conceptually one call occurs first
3580 * and we have a single winner.
3582 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3583 nice
= task_nice(current
) + increment
;
3585 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3586 if (increment
< 0 && !can_nice(current
, nice
))
3589 retval
= security_task_setnice(current
, nice
);
3593 set_user_nice(current
, nice
);
3600 * task_prio - return the priority value of a given task.
3601 * @p: the task in question.
3603 * Return: The priority value as seen by users in /proc.
3604 * RT tasks are offset by -200. Normal tasks are centered
3605 * around 0, value goes from -16 to +15.
3607 int task_prio(const struct task_struct
*p
)
3609 return p
->prio
- MAX_RT_PRIO
;
3613 * idle_cpu - is a given cpu idle currently?
3614 * @cpu: the processor in question.
3616 * Return: 1 if the CPU is currently idle. 0 otherwise.
3618 int idle_cpu(int cpu
)
3620 struct rq
*rq
= cpu_rq(cpu
);
3622 if (rq
->curr
!= rq
->idle
)
3629 if (!llist_empty(&rq
->wake_list
))
3637 * idle_task - return the idle task for a given cpu.
3638 * @cpu: the processor in question.
3640 * Return: The idle task for the cpu @cpu.
3642 struct task_struct
*idle_task(int cpu
)
3644 return cpu_rq(cpu
)->idle
;
3648 * find_process_by_pid - find a process with a matching PID value.
3649 * @pid: the pid in question.
3651 * The task of @pid, if found. %NULL otherwise.
3653 static struct task_struct
*find_process_by_pid(pid_t pid
)
3655 return pid
? find_task_by_vpid(pid
) : current
;
3659 * This function initializes the sched_dl_entity of a newly becoming
3660 * SCHED_DEADLINE task.
3662 * Only the static values are considered here, the actual runtime and the
3663 * absolute deadline will be properly calculated when the task is enqueued
3664 * for the first time with its new policy.
3667 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3669 struct sched_dl_entity
*dl_se
= &p
->dl
;
3671 dl_se
->dl_runtime
= attr
->sched_runtime
;
3672 dl_se
->dl_deadline
= attr
->sched_deadline
;
3673 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3674 dl_se
->flags
= attr
->sched_flags
;
3675 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3678 * Changing the parameters of a task is 'tricky' and we're not doing
3679 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3681 * What we SHOULD do is delay the bandwidth release until the 0-lag
3682 * point. This would include retaining the task_struct until that time
3683 * and change dl_overflow() to not immediately decrement the current
3686 * Instead we retain the current runtime/deadline and let the new
3687 * parameters take effect after the current reservation period lapses.
3688 * This is safe (albeit pessimistic) because the 0-lag point is always
3689 * before the current scheduling deadline.
3691 * We can still have temporary overloads because we do not delay the
3692 * change in bandwidth until that time; so admission control is
3693 * not on the safe side. It does however guarantee tasks will never
3694 * consume more than promised.
3699 * sched_setparam() passes in -1 for its policy, to let the functions
3700 * it calls know not to change it.
3702 #define SETPARAM_POLICY -1
3704 static void __setscheduler_params(struct task_struct
*p
,
3705 const struct sched_attr
*attr
)
3707 int policy
= attr
->sched_policy
;
3709 if (policy
== SETPARAM_POLICY
)
3714 if (dl_policy(policy
))
3715 __setparam_dl(p
, attr
);
3716 else if (fair_policy(policy
))
3717 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3720 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3721 * !rt_policy. Always setting this ensures that things like
3722 * getparam()/getattr() don't report silly values for !rt tasks.
3724 p
->rt_priority
= attr
->sched_priority
;
3725 p
->normal_prio
= normal_prio(p
);
3729 /* Actually do priority change: must hold pi & rq lock. */
3730 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3731 const struct sched_attr
*attr
, bool keep_boost
)
3733 __setscheduler_params(p
, attr
);
3736 * Keep a potential priority boosting if called from
3737 * sched_setscheduler().
3740 sched_set_prio(p
, rt_mutex_get_effective_prio(p
,
3743 sched_set_prio(p
, normal_prio(p
));
3745 if (dl_prio(p
->prio
))
3746 p
->sched_class
= &dl_sched_class
;
3747 else if (rt_prio(p
->prio
))
3748 p
->sched_class
= &rt_sched_class
;
3750 p
->sched_class
= &fair_sched_class
;
3754 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3756 struct sched_dl_entity
*dl_se
= &p
->dl
;
3758 attr
->sched_priority
= p
->rt_priority
;
3759 attr
->sched_runtime
= dl_se
->dl_runtime
;
3760 attr
->sched_deadline
= dl_se
->dl_deadline
;
3761 attr
->sched_period
= dl_se
->dl_period
;
3762 attr
->sched_flags
= dl_se
->flags
;
3766 * This function validates the new parameters of a -deadline task.
3767 * We ask for the deadline not being zero, and greater or equal
3768 * than the runtime, as well as the period of being zero or
3769 * greater than deadline. Furthermore, we have to be sure that
3770 * user parameters are above the internal resolution of 1us (we
3771 * check sched_runtime only since it is always the smaller one) and
3772 * below 2^63 ns (we have to check both sched_deadline and
3773 * sched_period, as the latter can be zero).
3776 __checkparam_dl(const struct sched_attr
*attr
)
3779 if (attr
->sched_deadline
== 0)
3783 * Since we truncate DL_SCALE bits, make sure we're at least
3786 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3790 * Since we use the MSB for wrap-around and sign issues, make
3791 * sure it's not set (mind that period can be equal to zero).
3793 if (attr
->sched_deadline
& (1ULL << 63) ||
3794 attr
->sched_period
& (1ULL << 63))
3797 /* runtime <= deadline <= period (if period != 0) */
3798 if ((attr
->sched_period
!= 0 &&
3799 attr
->sched_period
< attr
->sched_deadline
) ||
3800 attr
->sched_deadline
< attr
->sched_runtime
)
3807 * check the target process has a UID that matches the current process's
3809 static bool check_same_owner(struct task_struct
*p
)
3811 const struct cred
*cred
= current_cred(), *pcred
;
3815 pcred
= __task_cred(p
);
3816 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3817 uid_eq(cred
->euid
, pcred
->uid
));
3822 static bool dl_param_changed(struct task_struct
*p
,
3823 const struct sched_attr
*attr
)
3825 struct sched_dl_entity
*dl_se
= &p
->dl
;
3827 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3828 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3829 dl_se
->dl_period
!= attr
->sched_period
||
3830 dl_se
->flags
!= attr
->sched_flags
)
3836 static int __sched_setscheduler(struct task_struct
*p
,
3837 const struct sched_attr
*attr
,
3840 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3841 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3842 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3843 int new_effective_prio
, policy
= attr
->sched_policy
;
3844 unsigned long flags
;
3845 const struct sched_class
*prev_class
;
3848 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3850 /* may grab non-irq protected spin_locks */
3851 BUG_ON(in_interrupt());
3853 /* double check policy once rq lock held */
3855 reset_on_fork
= p
->sched_reset_on_fork
;
3856 policy
= oldpolicy
= p
->policy
;
3858 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3860 if (!valid_policy(policy
))
3864 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3868 * Valid priorities for SCHED_FIFO and SCHED_RR are
3869 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3870 * SCHED_BATCH and SCHED_IDLE is 0.
3872 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3873 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3875 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3876 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3880 * Allow unprivileged RT tasks to decrease priority:
3882 if (user
&& !capable(CAP_SYS_NICE
)) {
3883 if (fair_policy(policy
)) {
3884 if (attr
->sched_nice
< task_nice(p
) &&
3885 !can_nice(p
, attr
->sched_nice
))
3889 if (rt_policy(policy
)) {
3890 unsigned long rlim_rtprio
=
3891 task_rlimit(p
, RLIMIT_RTPRIO
);
3893 /* can't set/change the rt policy */
3894 if (policy
!= p
->policy
&& !rlim_rtprio
)
3897 /* can't increase priority */
3898 if (attr
->sched_priority
> p
->rt_priority
&&
3899 attr
->sched_priority
> rlim_rtprio
)
3904 * Can't set/change SCHED_DEADLINE policy at all for now
3905 * (safest behavior); in the future we would like to allow
3906 * unprivileged DL tasks to increase their relative deadline
3907 * or reduce their runtime (both ways reducing utilization)
3909 if (dl_policy(policy
))
3913 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3914 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3916 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
3917 if (!can_nice(p
, task_nice(p
)))
3921 /* can't change other user's priorities */
3922 if (!check_same_owner(p
))
3925 /* Normal users shall not reset the sched_reset_on_fork flag */
3926 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3931 retval
= security_task_setscheduler(p
);
3937 * make sure no PI-waiters arrive (or leave) while we are
3938 * changing the priority of the task:
3940 * To be able to change p->policy safely, the appropriate
3941 * runqueue lock must be held.
3943 rq
= task_rq_lock(p
, &flags
);
3946 * Changing the policy of the stop threads its a very bad idea
3948 if (p
== rq
->stop
) {
3949 task_rq_unlock(rq
, p
, &flags
);
3954 * If not changing anything there's no need to proceed further,
3955 * but store a possible modification of reset_on_fork.
3957 if (unlikely(policy
== p
->policy
)) {
3958 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3960 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3962 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3965 p
->sched_reset_on_fork
= reset_on_fork
;
3966 task_rq_unlock(rq
, p
, &flags
);
3972 #ifdef CONFIG_RT_GROUP_SCHED
3974 * Do not allow realtime tasks into groups that have no runtime
3977 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3978 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3979 !task_group_is_autogroup(task_group(p
))) {
3980 task_rq_unlock(rq
, p
, &flags
);
3985 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3986 cpumask_t
*span
= rq
->rd
->span
;
3989 * Don't allow tasks with an affinity mask smaller than
3990 * the entire root_domain to become SCHED_DEADLINE. We
3991 * will also fail if there's no bandwidth available.
3993 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3994 rq
->rd
->dl_bw
.bw
== 0) {
3995 task_rq_unlock(rq
, p
, &flags
);
4002 /* recheck policy now with rq lock held */
4003 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4004 policy
= oldpolicy
= -1;
4005 task_rq_unlock(rq
, p
, &flags
);
4010 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4011 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4014 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4015 task_rq_unlock(rq
, p
, &flags
);
4019 p
->sched_reset_on_fork
= reset_on_fork
;
4024 * Take priority boosted tasks into account. If the new
4025 * effective priority is unchanged, we just store the new
4026 * normal parameters and do not touch the scheduler class and
4027 * the runqueue. This will be done when the task deboost
4030 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4031 if (new_effective_prio
== oldprio
)
4032 queue_flags
&= ~DEQUEUE_MOVE
;
4035 queued
= task_on_rq_queued(p
);
4036 running
= task_current(rq
, p
);
4038 dequeue_task(rq
, p
, queue_flags
);
4040 put_prev_task(rq
, p
);
4042 prev_class
= p
->sched_class
;
4043 __setscheduler(rq
, p
, attr
, pi
);
4046 p
->sched_class
->set_curr_task(rq
);
4049 * We enqueue to tail when the priority of a task is
4050 * increased (user space view).
4052 if (oldprio
< p
->prio
)
4053 queue_flags
|= ENQUEUE_HEAD
;
4055 enqueue_task(rq
, p
, queue_flags
);
4058 check_class_changed(rq
, p
, prev_class
, oldprio
);
4059 preempt_disable(); /* avoid rq from going away on us */
4060 task_rq_unlock(rq
, p
, &flags
);
4063 rt_mutex_adjust_pi(p
);
4066 * Run balance callbacks after we've adjusted the PI chain.
4068 balance_callback(rq
);
4074 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4075 const struct sched_param
*param
, bool check
)
4077 struct sched_attr attr
= {
4078 .sched_policy
= policy
,
4079 .sched_priority
= param
->sched_priority
,
4080 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4083 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4084 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4085 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4086 policy
&= ~SCHED_RESET_ON_FORK
;
4087 attr
.sched_policy
= policy
;
4090 return __sched_setscheduler(p
, &attr
, check
, true);
4093 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4094 * @p: the task in question.
4095 * @policy: new policy.
4096 * @param: structure containing the new RT priority.
4098 * Return: 0 on success. An error code otherwise.
4100 * NOTE that the task may be already dead.
4102 int sched_setscheduler(struct task_struct
*p
, int policy
,
4103 const struct sched_param
*param
)
4105 return _sched_setscheduler(p
, policy
, param
, true);
4107 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4109 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4111 return __sched_setscheduler(p
, attr
, true, true);
4113 EXPORT_SYMBOL_GPL(sched_setattr
);
4116 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4117 * @p: the task in question.
4118 * @policy: new policy.
4119 * @param: structure containing the new RT priority.
4121 * Just like sched_setscheduler, only don't bother checking if the
4122 * current context has permission. For example, this is needed in
4123 * stop_machine(): we create temporary high priority worker threads,
4124 * but our caller might not have that capability.
4126 * Return: 0 on success. An error code otherwise.
4128 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4129 const struct sched_param
*param
)
4131 return _sched_setscheduler(p
, policy
, param
, false);
4133 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4136 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4138 struct sched_param lparam
;
4139 struct task_struct
*p
;
4142 if (!param
|| pid
< 0)
4144 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4149 p
= find_process_by_pid(pid
);
4151 retval
= sched_setscheduler(p
, policy
, &lparam
);
4158 * Mimics kernel/events/core.c perf_copy_attr().
4160 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4161 struct sched_attr
*attr
)
4166 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4170 * zero the full structure, so that a short copy will be nice.
4172 memset(attr
, 0, sizeof(*attr
));
4174 ret
= get_user(size
, &uattr
->size
);
4178 if (size
> PAGE_SIZE
) /* silly large */
4181 if (!size
) /* abi compat */
4182 size
= SCHED_ATTR_SIZE_VER0
;
4184 if (size
< SCHED_ATTR_SIZE_VER0
)
4188 * If we're handed a bigger struct than we know of,
4189 * ensure all the unknown bits are 0 - i.e. new
4190 * user-space does not rely on any kernel feature
4191 * extensions we dont know about yet.
4193 if (size
> sizeof(*attr
)) {
4194 unsigned char __user
*addr
;
4195 unsigned char __user
*end
;
4198 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4199 end
= (void __user
*)uattr
+ size
;
4201 for (; addr
< end
; addr
++) {
4202 ret
= get_user(val
, addr
);
4208 size
= sizeof(*attr
);
4211 ret
= copy_from_user(attr
, uattr
, size
);
4216 * XXX: do we want to be lenient like existing syscalls; or do we want
4217 * to be strict and return an error on out-of-bounds values?
4219 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4224 put_user(sizeof(*attr
), &uattr
->size
);
4229 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4230 * @pid: the pid in question.
4231 * @policy: new policy.
4232 * @param: structure containing the new RT priority.
4234 * Return: 0 on success. An error code otherwise.
4236 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4237 struct sched_param __user
*, param
)
4239 /* negative values for policy are not valid */
4243 return do_sched_setscheduler(pid
, policy
, param
);
4247 * sys_sched_setparam - set/change the RT priority of a thread
4248 * @pid: the pid in question.
4249 * @param: structure containing the new RT priority.
4251 * Return: 0 on success. An error code otherwise.
4253 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4255 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4259 * sys_sched_setattr - same as above, but with extended sched_attr
4260 * @pid: the pid in question.
4261 * @uattr: structure containing the extended parameters.
4262 * @flags: for future extension.
4264 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4265 unsigned int, flags
)
4267 struct sched_attr attr
;
4268 struct task_struct
*p
;
4271 if (!uattr
|| pid
< 0 || flags
)
4274 retval
= sched_copy_attr(uattr
, &attr
);
4278 if ((int)attr
.sched_policy
< 0)
4283 p
= find_process_by_pid(pid
);
4285 retval
= sched_setattr(p
, &attr
);
4292 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4293 * @pid: the pid in question.
4295 * Return: On success, the policy of the thread. Otherwise, a negative error
4298 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4300 struct task_struct
*p
;
4308 p
= find_process_by_pid(pid
);
4310 retval
= security_task_getscheduler(p
);
4313 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4320 * sys_sched_getparam - get the RT priority of a thread
4321 * @pid: the pid in question.
4322 * @param: structure containing the RT priority.
4324 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4327 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4329 struct sched_param lp
= { .sched_priority
= 0 };
4330 struct task_struct
*p
;
4333 if (!param
|| pid
< 0)
4337 p
= find_process_by_pid(pid
);
4342 retval
= security_task_getscheduler(p
);
4346 if (task_has_rt_policy(p
))
4347 lp
.sched_priority
= p
->rt_priority
;
4351 * This one might sleep, we cannot do it with a spinlock held ...
4353 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4362 static int sched_read_attr(struct sched_attr __user
*uattr
,
4363 struct sched_attr
*attr
,
4368 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4372 * If we're handed a smaller struct than we know of,
4373 * ensure all the unknown bits are 0 - i.e. old
4374 * user-space does not get uncomplete information.
4376 if (usize
< sizeof(*attr
)) {
4377 unsigned char *addr
;
4380 addr
= (void *)attr
+ usize
;
4381 end
= (void *)attr
+ sizeof(*attr
);
4383 for (; addr
< end
; addr
++) {
4391 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4399 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4400 * @pid: the pid in question.
4401 * @uattr: structure containing the extended parameters.
4402 * @size: sizeof(attr) for fwd/bwd comp.
4403 * @flags: for future extension.
4405 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4406 unsigned int, size
, unsigned int, flags
)
4408 struct sched_attr attr
= {
4409 .size
= sizeof(struct sched_attr
),
4411 struct task_struct
*p
;
4414 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4415 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4419 p
= find_process_by_pid(pid
);
4424 retval
= security_task_getscheduler(p
);
4428 attr
.sched_policy
= p
->policy
;
4429 if (p
->sched_reset_on_fork
)
4430 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4431 if (task_has_dl_policy(p
))
4432 __getparam_dl(p
, &attr
);
4433 else if (task_has_rt_policy(p
))
4434 attr
.sched_priority
= p
->rt_priority
;
4436 attr
.sched_nice
= task_nice(p
);
4440 retval
= sched_read_attr(uattr
, &attr
, size
);
4448 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4450 cpumask_var_t cpus_allowed
, new_mask
;
4451 struct task_struct
*p
;
4456 p
= find_process_by_pid(pid
);
4462 /* Prevent p going away */
4466 if (p
->flags
& PF_NO_SETAFFINITY
) {
4470 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4474 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4476 goto out_free_cpus_allowed
;
4479 if (!check_same_owner(p
)) {
4481 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4483 goto out_free_new_mask
;
4488 retval
= security_task_setscheduler(p
);
4490 goto out_free_new_mask
;
4493 cpuset_cpus_allowed(p
, cpus_allowed
);
4494 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4497 * Since bandwidth control happens on root_domain basis,
4498 * if admission test is enabled, we only admit -deadline
4499 * tasks allowed to run on all the CPUs in the task's
4503 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4505 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4508 goto out_free_new_mask
;
4514 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4517 cpuset_cpus_allowed(p
, cpus_allowed
);
4518 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4520 * We must have raced with a concurrent cpuset
4521 * update. Just reset the cpus_allowed to the
4522 * cpuset's cpus_allowed
4524 cpumask_copy(new_mask
, cpus_allowed
);
4529 free_cpumask_var(new_mask
);
4530 out_free_cpus_allowed
:
4531 free_cpumask_var(cpus_allowed
);
4537 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4538 struct cpumask
*new_mask
)
4540 if (len
< cpumask_size())
4541 cpumask_clear(new_mask
);
4542 else if (len
> cpumask_size())
4543 len
= cpumask_size();
4545 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4549 * sys_sched_setaffinity - set the cpu affinity of a process
4550 * @pid: pid of the process
4551 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4552 * @user_mask_ptr: user-space pointer to the new cpu mask
4554 * Return: 0 on success. An error code otherwise.
4556 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4557 unsigned long __user
*, user_mask_ptr
)
4559 cpumask_var_t new_mask
;
4562 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4565 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4567 retval
= sched_setaffinity(pid
, new_mask
);
4568 free_cpumask_var(new_mask
);
4572 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4574 struct task_struct
*p
;
4575 unsigned long flags
;
4581 p
= find_process_by_pid(pid
);
4585 retval
= security_task_getscheduler(p
);
4589 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4590 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4591 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4600 * sys_sched_getaffinity - get the cpu affinity of a process
4601 * @pid: pid of the process
4602 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4603 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4605 * Return: 0 on success. An error code otherwise.
4607 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4608 unsigned long __user
*, user_mask_ptr
)
4613 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4615 if (len
& (sizeof(unsigned long)-1))
4618 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4621 ret
= sched_getaffinity(pid
, mask
);
4623 size_t retlen
= min_t(size_t, len
, cpumask_size());
4625 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4630 free_cpumask_var(mask
);
4636 * sys_sched_yield - yield the current processor to other threads.
4638 * This function yields the current CPU to other tasks. If there are no
4639 * other threads running on this CPU then this function will return.
4643 SYSCALL_DEFINE0(sched_yield
)
4645 struct rq
*rq
= this_rq_lock();
4647 schedstat_inc(rq
, yld_count
);
4648 current
->sched_class
->yield_task(rq
);
4651 * Since we are going to call schedule() anyway, there's
4652 * no need to preempt or enable interrupts:
4654 __release(rq
->lock
);
4655 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4656 do_raw_spin_unlock(&rq
->lock
);
4657 sched_preempt_enable_no_resched();
4664 int __sched
_cond_resched(void)
4666 if (should_resched(0)) {
4667 preempt_schedule_common();
4672 EXPORT_SYMBOL(_cond_resched
);
4675 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4676 * call schedule, and on return reacquire the lock.
4678 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4679 * operations here to prevent schedule() from being called twice (once via
4680 * spin_unlock(), once by hand).
4682 int __cond_resched_lock(spinlock_t
*lock
)
4684 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4687 lockdep_assert_held(lock
);
4689 if (spin_needbreak(lock
) || resched
) {
4692 preempt_schedule_common();
4700 EXPORT_SYMBOL(__cond_resched_lock
);
4702 int __sched
__cond_resched_softirq(void)
4704 BUG_ON(!in_softirq());
4706 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4708 preempt_schedule_common();
4714 EXPORT_SYMBOL(__cond_resched_softirq
);
4717 * yield - yield the current processor to other threads.
4719 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4721 * The scheduler is at all times free to pick the calling task as the most
4722 * eligible task to run, if removing the yield() call from your code breaks
4723 * it, its already broken.
4725 * Typical broken usage is:
4730 * where one assumes that yield() will let 'the other' process run that will
4731 * make event true. If the current task is a SCHED_FIFO task that will never
4732 * happen. Never use yield() as a progress guarantee!!
4734 * If you want to use yield() to wait for something, use wait_event().
4735 * If you want to use yield() to be 'nice' for others, use cond_resched().
4736 * If you still want to use yield(), do not!
4738 void __sched
yield(void)
4740 set_current_state(TASK_RUNNING
);
4743 EXPORT_SYMBOL(yield
);
4746 * yield_to - yield the current processor to another thread in
4747 * your thread group, or accelerate that thread toward the
4748 * processor it's on.
4750 * @preempt: whether task preemption is allowed or not
4752 * It's the caller's job to ensure that the target task struct
4753 * can't go away on us before we can do any checks.
4756 * true (>0) if we indeed boosted the target task.
4757 * false (0) if we failed to boost the target.
4758 * -ESRCH if there's no task to yield to.
4760 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4762 struct task_struct
*curr
= current
;
4763 struct rq
*rq
, *p_rq
;
4764 unsigned long flags
;
4767 local_irq_save(flags
);
4773 * If we're the only runnable task on the rq and target rq also
4774 * has only one task, there's absolutely no point in yielding.
4776 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4781 double_rq_lock(rq
, p_rq
);
4782 if (task_rq(p
) != p_rq
) {
4783 double_rq_unlock(rq
, p_rq
);
4787 if (!curr
->sched_class
->yield_to_task
)
4790 if (curr
->sched_class
!= p
->sched_class
)
4793 if (task_running(p_rq
, p
) || p
->state
)
4796 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4798 schedstat_inc(rq
, yld_count
);
4800 * Make p's CPU reschedule; pick_next_entity takes care of
4803 if (preempt
&& rq
!= p_rq
)
4808 double_rq_unlock(rq
, p_rq
);
4810 local_irq_restore(flags
);
4817 EXPORT_SYMBOL_GPL(yield_to
);
4820 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4821 * that process accounting knows that this is a task in IO wait state.
4823 long __sched
io_schedule_timeout(long timeout
)
4825 int old_iowait
= current
->in_iowait
;
4829 current
->in_iowait
= 1;
4830 blk_schedule_flush_plug(current
);
4832 delayacct_blkio_start();
4834 atomic_inc(&rq
->nr_iowait
);
4835 ret
= schedule_timeout(timeout
);
4836 current
->in_iowait
= old_iowait
;
4837 atomic_dec(&rq
->nr_iowait
);
4838 delayacct_blkio_end();
4842 EXPORT_SYMBOL(io_schedule_timeout
);
4845 * sys_sched_get_priority_max - return maximum RT priority.
4846 * @policy: scheduling class.
4848 * Return: On success, this syscall returns the maximum
4849 * rt_priority that can be used by a given scheduling class.
4850 * On failure, a negative error code is returned.
4852 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4859 ret
= MAX_USER_RT_PRIO
-1;
4861 case SCHED_DEADLINE
:
4872 * sys_sched_get_priority_min - return minimum RT priority.
4873 * @policy: scheduling class.
4875 * Return: On success, this syscall returns the minimum
4876 * rt_priority that can be used by a given scheduling class.
4877 * On failure, a negative error code is returned.
4879 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4888 case SCHED_DEADLINE
:
4898 * sys_sched_rr_get_interval - return the default timeslice of a process.
4899 * @pid: pid of the process.
4900 * @interval: userspace pointer to the timeslice value.
4902 * this syscall writes the default timeslice value of a given process
4903 * into the user-space timespec buffer. A value of '0' means infinity.
4905 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4908 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4909 struct timespec __user
*, interval
)
4911 struct task_struct
*p
;
4912 unsigned int time_slice
;
4913 unsigned long flags
;
4923 p
= find_process_by_pid(pid
);
4927 retval
= security_task_getscheduler(p
);
4931 rq
= task_rq_lock(p
, &flags
);
4933 if (p
->sched_class
->get_rr_interval
)
4934 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4935 task_rq_unlock(rq
, p
, &flags
);
4938 jiffies_to_timespec(time_slice
, &t
);
4939 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4947 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4949 void sched_show_task(struct task_struct
*p
)
4951 unsigned long free
= 0;
4953 unsigned long state
= p
->state
;
4956 state
= __ffs(state
) + 1;
4957 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4958 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4959 #if BITS_PER_LONG == 32
4960 if (state
== TASK_RUNNING
)
4961 printk(KERN_CONT
" running ");
4963 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4965 if (state
== TASK_RUNNING
)
4966 printk(KERN_CONT
" running task ");
4968 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4970 #ifdef CONFIG_DEBUG_STACK_USAGE
4971 free
= stack_not_used(p
);
4976 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4978 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4979 task_pid_nr(p
), ppid
,
4980 (unsigned long)task_thread_info(p
)->flags
);
4982 print_worker_info(KERN_INFO
, p
);
4983 show_stack(p
, NULL
);
4986 void show_state_filter(unsigned long state_filter
)
4988 struct task_struct
*g
, *p
;
4990 #if BITS_PER_LONG == 32
4992 " task PC stack pid father\n");
4995 " task PC stack pid father\n");
4998 for_each_process_thread(g
, p
) {
5000 * reset the NMI-timeout, listing all files on a slow
5001 * console might take a lot of time:
5003 touch_nmi_watchdog();
5004 if (!state_filter
|| (p
->state
& state_filter
))
5008 touch_all_softlockup_watchdogs();
5010 #ifdef CONFIG_SCHED_DEBUG
5011 sysrq_sched_debug_show();
5015 * Only show locks if all tasks are dumped:
5018 debug_show_all_locks();
5021 void init_idle_bootup_task(struct task_struct
*idle
)
5023 idle
->sched_class
= &idle_sched_class
;
5027 * init_idle - set up an idle thread for a given CPU
5028 * @idle: task in question
5029 * @cpu: cpu the idle task belongs to
5031 * NOTE: this function does not set the idle thread's NEED_RESCHED
5032 * flag, to make booting more robust.
5034 void init_idle(struct task_struct
*idle
, int cpu
)
5036 struct rq
*rq
= cpu_rq(cpu
);
5037 unsigned long flags
;
5039 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5040 raw_spin_lock(&rq
->lock
);
5042 __sched_fork(0, idle
);
5043 idle
->state
= TASK_RUNNING
;
5044 idle
->se
.exec_start
= sched_clock();
5046 kasan_unpoison_task_stack(idle
);
5050 * Its possible that init_idle() gets called multiple times on a task,
5051 * in that case do_set_cpus_allowed() will not do the right thing.
5053 * And since this is boot we can forgo the serialization.
5055 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5058 * We're having a chicken and egg problem, even though we are
5059 * holding rq->lock, the cpu isn't yet set to this cpu so the
5060 * lockdep check in task_group() will fail.
5062 * Similar case to sched_fork(). / Alternatively we could
5063 * use task_rq_lock() here and obtain the other rq->lock.
5068 __set_task_cpu(idle
, cpu
);
5071 rq
->curr
= rq
->idle
= idle
;
5072 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5076 raw_spin_unlock(&rq
->lock
);
5077 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5079 /* Set the preempt count _outside_ the spinlocks! */
5080 init_idle_preempt_count(idle
, cpu
);
5083 * The idle tasks have their own, simple scheduling class:
5085 idle
->sched_class
= &idle_sched_class
;
5086 ftrace_graph_init_idle_task(idle
, cpu
);
5087 vtime_init_idle(idle
, cpu
);
5089 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5093 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5094 const struct cpumask
*trial
)
5096 int ret
= 1, trial_cpus
;
5097 struct dl_bw
*cur_dl_b
;
5098 unsigned long flags
;
5100 if (!cpumask_weight(cur
))
5103 rcu_read_lock_sched();
5104 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5105 trial_cpus
= cpumask_weight(trial
);
5107 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5108 if (cur_dl_b
->bw
!= -1 &&
5109 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5111 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5112 rcu_read_unlock_sched();
5117 int task_can_attach(struct task_struct
*p
,
5118 const struct cpumask
*cs_cpus_allowed
)
5123 * Kthreads which disallow setaffinity shouldn't be moved
5124 * to a new cpuset; we don't want to change their cpu
5125 * affinity and isolating such threads by their set of
5126 * allowed nodes is unnecessary. Thus, cpusets are not
5127 * applicable for such threads. This prevents checking for
5128 * success of set_cpus_allowed_ptr() on all attached tasks
5129 * before cpus_allowed may be changed.
5131 if (p
->flags
& PF_NO_SETAFFINITY
) {
5137 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5139 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5144 unsigned long flags
;
5146 rcu_read_lock_sched();
5147 dl_b
= dl_bw_of(dest_cpu
);
5148 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5149 cpus
= dl_bw_cpus(dest_cpu
);
5150 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5155 * We reserve space for this task in the destination
5156 * root_domain, as we can't fail after this point.
5157 * We will free resources in the source root_domain
5158 * later on (see set_cpus_allowed_dl()).
5160 __dl_add(dl_b
, p
->dl
.dl_bw
);
5162 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5163 rcu_read_unlock_sched();
5173 #ifdef CONFIG_NUMA_BALANCING
5174 /* Migrate current task p to target_cpu */
5175 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5177 struct migration_arg arg
= { p
, target_cpu
};
5178 int curr_cpu
= task_cpu(p
);
5180 if (curr_cpu
== target_cpu
)
5183 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5186 /* TODO: This is not properly updating schedstats */
5188 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5189 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5193 * Requeue a task on a given node and accurately track the number of NUMA
5194 * tasks on the runqueues
5196 void sched_setnuma(struct task_struct
*p
, int nid
)
5199 unsigned long flags
;
5200 bool queued
, running
;
5202 rq
= task_rq_lock(p
, &flags
);
5203 queued
= task_on_rq_queued(p
);
5204 running
= task_current(rq
, p
);
5207 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5209 put_prev_task(rq
, p
);
5211 p
->numa_preferred_nid
= nid
;
5214 p
->sched_class
->set_curr_task(rq
);
5216 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5217 task_rq_unlock(rq
, p
, &flags
);
5219 #endif /* CONFIG_NUMA_BALANCING */
5221 #ifdef CONFIG_HOTPLUG_CPU
5223 * Ensures that the idle task is using init_mm right before its cpu goes
5226 void idle_task_exit(void)
5228 struct mm_struct
*mm
= current
->active_mm
;
5230 BUG_ON(cpu_online(smp_processor_id()));
5232 if (mm
!= &init_mm
) {
5233 switch_mm(mm
, &init_mm
, current
);
5234 finish_arch_post_lock_switch();
5240 * Since this CPU is going 'away' for a while, fold any nr_active delta
5241 * we might have. Assumes we're called after migrate_tasks() so that the
5242 * nr_active count is stable.
5244 * Also see the comment "Global load-average calculations".
5246 static void calc_load_migrate(struct rq
*rq
)
5248 long delta
= calc_load_fold_active(rq
);
5250 atomic_long_add(delta
, &calc_load_tasks
);
5253 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5257 static const struct sched_class fake_sched_class
= {
5258 .put_prev_task
= put_prev_task_fake
,
5261 static struct task_struct fake_task
= {
5263 * Avoid pull_{rt,dl}_task()
5265 .prio
= MAX_PRIO
+ 1,
5266 .sched_class
= &fake_sched_class
,
5270 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5271 * try_to_wake_up()->select_task_rq().
5273 * Called with rq->lock held even though we'er in stop_machine() and
5274 * there's no concurrency possible, we hold the required locks anyway
5275 * because of lock validation efforts.
5277 static void migrate_tasks(struct rq
*dead_rq
)
5279 struct rq
*rq
= dead_rq
;
5280 struct task_struct
*next
, *stop
= rq
->stop
;
5284 * Fudge the rq selection such that the below task selection loop
5285 * doesn't get stuck on the currently eligible stop task.
5287 * We're currently inside stop_machine() and the rq is either stuck
5288 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5289 * either way we should never end up calling schedule() until we're
5295 * put_prev_task() and pick_next_task() sched
5296 * class method both need to have an up-to-date
5297 * value of rq->clock[_task]
5299 update_rq_clock(rq
);
5303 * There's this thread running, bail when that's the only
5306 if (rq
->nr_running
== 1)
5310 * pick_next_task assumes pinned rq->lock.
5312 lockdep_pin_lock(&rq
->lock
);
5313 next
= pick_next_task(rq
, &fake_task
);
5315 next
->sched_class
->put_prev_task(rq
, next
);
5318 * Rules for changing task_struct::cpus_allowed are holding
5319 * both pi_lock and rq->lock, such that holding either
5320 * stabilizes the mask.
5322 * Drop rq->lock is not quite as disastrous as it usually is
5323 * because !cpu_active at this point, which means load-balance
5324 * will not interfere. Also, stop-machine.
5326 lockdep_unpin_lock(&rq
->lock
);
5327 raw_spin_unlock(&rq
->lock
);
5328 raw_spin_lock(&next
->pi_lock
);
5329 raw_spin_lock(&rq
->lock
);
5332 * Since we're inside stop-machine, _nothing_ should have
5333 * changed the task, WARN if weird stuff happened, because in
5334 * that case the above rq->lock drop is a fail too.
5336 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5337 raw_spin_unlock(&next
->pi_lock
);
5341 /* Find suitable destination for @next, with force if needed. */
5342 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5344 rq
= __migrate_task(rq
, next
, dest_cpu
);
5345 if (rq
!= dead_rq
) {
5346 raw_spin_unlock(&rq
->lock
);
5348 raw_spin_lock(&rq
->lock
);
5350 raw_spin_unlock(&next
->pi_lock
);
5355 #endif /* CONFIG_HOTPLUG_CPU */
5357 static void set_rq_online(struct rq
*rq
)
5360 const struct sched_class
*class;
5362 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5365 for_each_class(class) {
5366 if (class->rq_online
)
5367 class->rq_online(rq
);
5372 static void set_rq_offline(struct rq
*rq
)
5375 const struct sched_class
*class;
5377 for_each_class(class) {
5378 if (class->rq_offline
)
5379 class->rq_offline(rq
);
5382 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5388 * migration_call - callback that gets triggered when a CPU is added.
5389 * Here we can start up the necessary migration thread for the new CPU.
5392 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5394 int cpu
= (long)hcpu
;
5395 unsigned long flags
;
5396 struct rq
*rq
= cpu_rq(cpu
);
5398 switch (action
& ~CPU_TASKS_FROZEN
) {
5400 case CPU_UP_PREPARE
:
5401 rq
->calc_load_update
= calc_load_update
;
5402 account_reset_rq(rq
);
5406 /* Update our root-domain */
5407 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5409 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5413 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5416 #ifdef CONFIG_HOTPLUG_CPU
5418 sched_ttwu_pending();
5419 /* Update our root-domain */
5420 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5422 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5426 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5427 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5431 calc_load_migrate(rq
);
5436 update_max_interval();
5442 * Register at high priority so that task migration (migrate_all_tasks)
5443 * happens before everything else. This has to be lower priority than
5444 * the notifier in the perf_event subsystem, though.
5446 static struct notifier_block migration_notifier
= {
5447 .notifier_call
= migration_call
,
5448 .priority
= CPU_PRI_MIGRATION
,
5451 static void set_cpu_rq_start_time(void)
5453 int cpu
= smp_processor_id();
5454 struct rq
*rq
= cpu_rq(cpu
);
5455 rq
->age_stamp
= sched_clock_cpu(cpu
);
5458 static int sched_cpu_active(struct notifier_block
*nfb
,
5459 unsigned long action
, void *hcpu
)
5461 int cpu
= (long)hcpu
;
5463 switch (action
& ~CPU_TASKS_FROZEN
) {
5465 set_cpu_rq_start_time();
5468 case CPU_DOWN_FAILED
:
5469 set_cpu_active(cpu
, true);
5477 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5478 unsigned long action
, void *hcpu
)
5480 switch (action
& ~CPU_TASKS_FROZEN
) {
5481 case CPU_DOWN_PREPARE
:
5482 set_cpu_active((long)hcpu
, false);
5489 static int __init
migration_init(void)
5491 void *cpu
= (void *)(long)smp_processor_id();
5494 /* Initialize migration for the boot CPU */
5495 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5496 BUG_ON(err
== NOTIFY_BAD
);
5497 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5498 register_cpu_notifier(&migration_notifier
);
5500 /* Register cpu active notifiers */
5501 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5502 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5506 early_initcall(migration_init
);
5508 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5510 #ifdef CONFIG_SCHED_DEBUG
5512 static __read_mostly
int sched_debug_enabled
;
5514 static int __init
sched_debug_setup(char *str
)
5516 sched_debug_enabled
= 1;
5520 early_param("sched_debug", sched_debug_setup
);
5522 static inline bool sched_debug(void)
5524 return sched_debug_enabled
;
5527 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5528 struct cpumask
*groupmask
)
5530 struct sched_group
*group
= sd
->groups
;
5532 cpumask_clear(groupmask
);
5534 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5536 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5537 printk("does not load-balance\n");
5539 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5544 printk(KERN_CONT
"span %*pbl level %s\n",
5545 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5547 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5548 printk(KERN_ERR
"ERROR: domain->span does not contain "
5551 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5552 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5556 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5560 printk(KERN_ERR
"ERROR: group is NULL\n");
5564 if (!cpumask_weight(sched_group_cpus(group
))) {
5565 printk(KERN_CONT
"\n");
5566 printk(KERN_ERR
"ERROR: empty group\n");
5570 if (!(sd
->flags
& SD_OVERLAP
) &&
5571 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5572 printk(KERN_CONT
"\n");
5573 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5577 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5579 printk(KERN_CONT
" %*pbl",
5580 cpumask_pr_args(sched_group_cpus(group
)));
5581 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5582 printk(KERN_CONT
" (cpu_capacity = %d)",
5583 group
->sgc
->capacity
);
5586 group
= group
->next
;
5587 } while (group
!= sd
->groups
);
5588 printk(KERN_CONT
"\n");
5590 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5591 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5594 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5595 printk(KERN_ERR
"ERROR: parent span is not a superset "
5596 "of domain->span\n");
5600 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5604 if (!sched_debug_enabled
)
5608 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5612 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5615 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5623 #else /* !CONFIG_SCHED_DEBUG */
5624 # define sched_domain_debug(sd, cpu) do { } while (0)
5625 static inline bool sched_debug(void)
5629 #endif /* CONFIG_SCHED_DEBUG */
5631 static int sd_degenerate(struct sched_domain
*sd
)
5633 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5636 /* Following flags need at least 2 groups */
5637 if (sd
->flags
& (SD_LOAD_BALANCE
|
5638 SD_BALANCE_NEWIDLE
|
5641 SD_SHARE_CPUCAPACITY
|
5642 SD_SHARE_PKG_RESOURCES
|
5643 SD_SHARE_POWERDOMAIN
)) {
5644 if (sd
->groups
!= sd
->groups
->next
)
5648 /* Following flags don't use groups */
5649 if (sd
->flags
& (SD_WAKE_AFFINE
))
5656 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5658 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5660 if (sd_degenerate(parent
))
5663 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5666 /* Flags needing groups don't count if only 1 group in parent */
5667 if (parent
->groups
== parent
->groups
->next
) {
5668 pflags
&= ~(SD_LOAD_BALANCE
|
5669 SD_BALANCE_NEWIDLE
|
5672 SD_SHARE_CPUCAPACITY
|
5673 SD_SHARE_PKG_RESOURCES
|
5675 SD_SHARE_POWERDOMAIN
);
5676 if (nr_node_ids
== 1)
5677 pflags
&= ~SD_SERIALIZE
;
5679 if (~cflags
& pflags
)
5685 static void free_rootdomain(struct rcu_head
*rcu
)
5687 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5689 cpupri_cleanup(&rd
->cpupri
);
5690 cpudl_cleanup(&rd
->cpudl
);
5691 free_cpumask_var(rd
->dlo_mask
);
5692 free_cpumask_var(rd
->rto_mask
);
5693 free_cpumask_var(rd
->online
);
5694 free_cpumask_var(rd
->span
);
5698 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5700 struct root_domain
*old_rd
= NULL
;
5701 unsigned long flags
;
5703 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5708 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5711 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5714 * If we dont want to free the old_rd yet then
5715 * set old_rd to NULL to skip the freeing later
5718 if (!atomic_dec_and_test(&old_rd
->refcount
))
5722 atomic_inc(&rd
->refcount
);
5725 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5726 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5729 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5732 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5735 static int init_rootdomain(struct root_domain
*rd
)
5737 memset(rd
, 0, sizeof(*rd
));
5739 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5741 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5743 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5745 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5748 init_dl_bw(&rd
->dl_bw
);
5749 if (cpudl_init(&rd
->cpudl
) != 0)
5752 if (cpupri_init(&rd
->cpupri
) != 0)
5757 free_cpumask_var(rd
->rto_mask
);
5759 free_cpumask_var(rd
->dlo_mask
);
5761 free_cpumask_var(rd
->online
);
5763 free_cpumask_var(rd
->span
);
5769 * By default the system creates a single root-domain with all cpus as
5770 * members (mimicking the global state we have today).
5772 struct root_domain def_root_domain
;
5774 static void init_defrootdomain(void)
5776 init_rootdomain(&def_root_domain
);
5778 atomic_set(&def_root_domain
.refcount
, 1);
5781 static struct root_domain
*alloc_rootdomain(void)
5783 struct root_domain
*rd
;
5785 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5789 if (init_rootdomain(rd
) != 0) {
5797 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5799 struct sched_group
*tmp
, *first
;
5808 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5813 } while (sg
!= first
);
5816 static void free_sched_domain(struct rcu_head
*rcu
)
5818 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5821 * If its an overlapping domain it has private groups, iterate and
5824 if (sd
->flags
& SD_OVERLAP
) {
5825 free_sched_groups(sd
->groups
, 1);
5826 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5827 kfree(sd
->groups
->sgc
);
5833 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5835 call_rcu(&sd
->rcu
, free_sched_domain
);
5838 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5840 for (; sd
; sd
= sd
->parent
)
5841 destroy_sched_domain(sd
, cpu
);
5845 * Keep a special pointer to the highest sched_domain that has
5846 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5847 * allows us to avoid some pointer chasing select_idle_sibling().
5849 * Also keep a unique ID per domain (we use the first cpu number in
5850 * the cpumask of the domain), this allows us to quickly tell if
5851 * two cpus are in the same cache domain, see cpus_share_cache().
5853 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5854 DEFINE_PER_CPU(int, sd_llc_size
);
5855 DEFINE_PER_CPU(int, sd_llc_id
);
5856 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5857 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5858 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5860 static void update_top_cache_domain(int cpu
)
5862 struct sched_domain
*sd
;
5863 struct sched_domain
*busy_sd
= NULL
;
5867 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5869 id
= cpumask_first(sched_domain_span(sd
));
5870 size
= cpumask_weight(sched_domain_span(sd
));
5871 busy_sd
= sd
->parent
; /* sd_busy */
5873 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5875 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5876 per_cpu(sd_llc_size
, cpu
) = size
;
5877 per_cpu(sd_llc_id
, cpu
) = id
;
5879 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5880 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5882 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5883 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5887 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5888 * hold the hotplug lock.
5891 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5893 struct rq
*rq
= cpu_rq(cpu
);
5894 struct sched_domain
*tmp
;
5896 /* Remove the sched domains which do not contribute to scheduling. */
5897 for (tmp
= sd
; tmp
; ) {
5898 struct sched_domain
*parent
= tmp
->parent
;
5902 if (sd_parent_degenerate(tmp
, parent
)) {
5903 tmp
->parent
= parent
->parent
;
5905 parent
->parent
->child
= tmp
;
5907 * Transfer SD_PREFER_SIBLING down in case of a
5908 * degenerate parent; the spans match for this
5909 * so the property transfers.
5911 if (parent
->flags
& SD_PREFER_SIBLING
)
5912 tmp
->flags
|= SD_PREFER_SIBLING
;
5913 destroy_sched_domain(parent
, cpu
);
5918 if (sd
&& sd_degenerate(sd
)) {
5921 destroy_sched_domain(tmp
, cpu
);
5926 sched_domain_debug(sd
, cpu
);
5928 rq_attach_root(rq
, rd
);
5930 rcu_assign_pointer(rq
->sd
, sd
);
5931 destroy_sched_domains(tmp
, cpu
);
5933 update_top_cache_domain(cpu
);
5936 /* Setup the mask of cpus configured for isolated domains */
5937 static int __init
isolated_cpu_setup(char *str
)
5941 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5942 ret
= cpulist_parse(str
, cpu_isolated_map
);
5944 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5949 __setup("isolcpus=", isolated_cpu_setup
);
5952 struct sched_domain
** __percpu sd
;
5953 struct root_domain
*rd
;
5964 * Build an iteration mask that can exclude certain CPUs from the upwards
5967 * Asymmetric node setups can result in situations where the domain tree is of
5968 * unequal depth, make sure to skip domains that already cover the entire
5971 * In that case build_sched_domains() will have terminated the iteration early
5972 * and our sibling sd spans will be empty. Domains should always include the
5973 * cpu they're built on, so check that.
5976 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5978 const struct cpumask
*span
= sched_domain_span(sd
);
5979 struct sd_data
*sdd
= sd
->private;
5980 struct sched_domain
*sibling
;
5983 for_each_cpu(i
, span
) {
5984 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5985 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5988 cpumask_set_cpu(i
, sched_group_mask(sg
));
5993 * Return the canonical balance cpu for this group, this is the first cpu
5994 * of this group that's also in the iteration mask.
5996 int group_balance_cpu(struct sched_group
*sg
)
5998 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6002 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6004 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6005 const struct cpumask
*span
= sched_domain_span(sd
);
6006 struct cpumask
*covered
= sched_domains_tmpmask
;
6007 struct sd_data
*sdd
= sd
->private;
6008 struct sched_domain
*sibling
;
6011 cpumask_clear(covered
);
6013 for_each_cpu(i
, span
) {
6014 struct cpumask
*sg_span
;
6016 if (cpumask_test_cpu(i
, covered
))
6019 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6021 /* See the comment near build_group_mask(). */
6022 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6025 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6026 GFP_KERNEL
, cpu_to_node(cpu
));
6031 sg_span
= sched_group_cpus(sg
);
6033 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6035 cpumask_set_cpu(i
, sg_span
);
6037 cpumask_or(covered
, covered
, sg_span
);
6039 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6040 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6041 build_group_mask(sd
, sg
);
6044 * Initialize sgc->capacity such that even if we mess up the
6045 * domains and no possible iteration will get us here, we won't
6048 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6051 * Make sure the first group of this domain contains the
6052 * canonical balance cpu. Otherwise the sched_domain iteration
6053 * breaks. See update_sg_lb_stats().
6055 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6056 group_balance_cpu(sg
) == cpu
)
6066 sd
->groups
= groups
;
6071 free_sched_groups(first
, 0);
6076 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6078 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6079 struct sched_domain
*child
= sd
->child
;
6082 cpu
= cpumask_first(sched_domain_span(child
));
6085 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6086 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6087 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6094 * build_sched_groups will build a circular linked list of the groups
6095 * covered by the given span, and will set each group's ->cpumask correctly,
6096 * and ->cpu_capacity to 0.
6098 * Assumes the sched_domain tree is fully constructed
6101 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6103 struct sched_group
*first
= NULL
, *last
= NULL
;
6104 struct sd_data
*sdd
= sd
->private;
6105 const struct cpumask
*span
= sched_domain_span(sd
);
6106 struct cpumask
*covered
;
6109 get_group(cpu
, sdd
, &sd
->groups
);
6110 atomic_inc(&sd
->groups
->ref
);
6112 if (cpu
!= cpumask_first(span
))
6115 lockdep_assert_held(&sched_domains_mutex
);
6116 covered
= sched_domains_tmpmask
;
6118 cpumask_clear(covered
);
6120 for_each_cpu(i
, span
) {
6121 struct sched_group
*sg
;
6124 if (cpumask_test_cpu(i
, covered
))
6127 group
= get_group(i
, sdd
, &sg
);
6128 cpumask_setall(sched_group_mask(sg
));
6130 for_each_cpu(j
, span
) {
6131 if (get_group(j
, sdd
, NULL
) != group
)
6134 cpumask_set_cpu(j
, covered
);
6135 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6150 * Initialize sched groups cpu_capacity.
6152 * cpu_capacity indicates the capacity of sched group, which is used while
6153 * distributing the load between different sched groups in a sched domain.
6154 * Typically cpu_capacity for all the groups in a sched domain will be same
6155 * unless there are asymmetries in the topology. If there are asymmetries,
6156 * group having more cpu_capacity will pickup more load compared to the
6157 * group having less cpu_capacity.
6159 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6161 struct sched_group
*sg
= sd
->groups
;
6166 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6168 } while (sg
!= sd
->groups
);
6170 if (cpu
!= group_balance_cpu(sg
))
6173 update_group_capacity(sd
, cpu
);
6174 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6178 * Initializers for schedule domains
6179 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6182 static int default_relax_domain_level
= -1;
6183 int sched_domain_level_max
;
6185 static int __init
setup_relax_domain_level(char *str
)
6187 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6188 pr_warn("Unable to set relax_domain_level\n");
6192 __setup("relax_domain_level=", setup_relax_domain_level
);
6194 static void set_domain_attribute(struct sched_domain
*sd
,
6195 struct sched_domain_attr
*attr
)
6199 if (!attr
|| attr
->relax_domain_level
< 0) {
6200 if (default_relax_domain_level
< 0)
6203 request
= default_relax_domain_level
;
6205 request
= attr
->relax_domain_level
;
6206 if (request
< sd
->level
) {
6207 /* turn off idle balance on this domain */
6208 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6210 /* turn on idle balance on this domain */
6211 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6215 static void __sdt_free(const struct cpumask
*cpu_map
);
6216 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6218 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6219 const struct cpumask
*cpu_map
)
6223 if (!atomic_read(&d
->rd
->refcount
))
6224 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6226 free_percpu(d
->sd
); /* fall through */
6228 __sdt_free(cpu_map
); /* fall through */
6234 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6235 const struct cpumask
*cpu_map
)
6237 memset(d
, 0, sizeof(*d
));
6239 if (__sdt_alloc(cpu_map
))
6240 return sa_sd_storage
;
6241 d
->sd
= alloc_percpu(struct sched_domain
*);
6243 return sa_sd_storage
;
6244 d
->rd
= alloc_rootdomain();
6247 return sa_rootdomain
;
6251 * NULL the sd_data elements we've used to build the sched_domain and
6252 * sched_group structure so that the subsequent __free_domain_allocs()
6253 * will not free the data we're using.
6255 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6257 struct sd_data
*sdd
= sd
->private;
6259 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6260 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6262 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6263 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6265 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6266 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6270 static int sched_domains_numa_levels
;
6271 enum numa_topology_type sched_numa_topology_type
;
6272 static int *sched_domains_numa_distance
;
6273 int sched_max_numa_distance
;
6274 static struct cpumask
***sched_domains_numa_masks
;
6275 static int sched_domains_curr_level
;
6279 * SD_flags allowed in topology descriptions.
6281 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6282 * SD_SHARE_PKG_RESOURCES - describes shared caches
6283 * SD_NUMA - describes NUMA topologies
6284 * SD_SHARE_POWERDOMAIN - describes shared power domain
6287 * SD_ASYM_PACKING - describes SMT quirks
6289 #define TOPOLOGY_SD_FLAGS \
6290 (SD_SHARE_CPUCAPACITY | \
6291 SD_SHARE_PKG_RESOURCES | \
6294 SD_SHARE_POWERDOMAIN)
6296 static struct sched_domain
*
6297 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6299 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6300 int sd_weight
, sd_flags
= 0;
6304 * Ugly hack to pass state to sd_numa_mask()...
6306 sched_domains_curr_level
= tl
->numa_level
;
6309 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6312 sd_flags
= (*tl
->sd_flags
)();
6313 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6314 "wrong sd_flags in topology description\n"))
6315 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6317 *sd
= (struct sched_domain
){
6318 .min_interval
= sd_weight
,
6319 .max_interval
= 2*sd_weight
,
6321 .imbalance_pct
= 125,
6323 .cache_nice_tries
= 0,
6330 .flags
= 1*SD_LOAD_BALANCE
6331 | 1*SD_BALANCE_NEWIDLE
6336 | 0*SD_SHARE_CPUCAPACITY
6337 | 0*SD_SHARE_PKG_RESOURCES
6339 | 0*SD_PREFER_SIBLING
6344 .last_balance
= jiffies
,
6345 .balance_interval
= sd_weight
,
6347 .max_newidle_lb_cost
= 0,
6348 .next_decay_max_lb_cost
= jiffies
,
6349 #ifdef CONFIG_SCHED_DEBUG
6355 * Convert topological properties into behaviour.
6358 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6359 sd
->flags
|= SD_PREFER_SIBLING
;
6360 sd
->imbalance_pct
= 110;
6361 sd
->smt_gain
= 1178; /* ~15% */
6363 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6364 sd
->imbalance_pct
= 117;
6365 sd
->cache_nice_tries
= 1;
6369 } else if (sd
->flags
& SD_NUMA
) {
6370 sd
->cache_nice_tries
= 2;
6374 sd
->flags
|= SD_SERIALIZE
;
6375 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6376 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6383 sd
->flags
|= SD_PREFER_SIBLING
;
6384 sd
->cache_nice_tries
= 1;
6389 sd
->private = &tl
->data
;
6395 * Topology list, bottom-up.
6397 static struct sched_domain_topology_level default_topology
[] = {
6398 #ifdef CONFIG_SCHED_SMT
6399 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6401 #ifdef CONFIG_SCHED_MC
6402 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6404 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6408 static struct sched_domain_topology_level
*sched_domain_topology
=
6411 #define for_each_sd_topology(tl) \
6412 for (tl = sched_domain_topology; tl->mask; tl++)
6414 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6416 sched_domain_topology
= tl
;
6421 static const struct cpumask
*sd_numa_mask(int cpu
)
6423 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6426 static void sched_numa_warn(const char *str
)
6428 static int done
= false;
6436 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6438 for (i
= 0; i
< nr_node_ids
; i
++) {
6439 printk(KERN_WARNING
" ");
6440 for (j
= 0; j
< nr_node_ids
; j
++)
6441 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6442 printk(KERN_CONT
"\n");
6444 printk(KERN_WARNING
"\n");
6447 bool find_numa_distance(int distance
)
6451 if (distance
== node_distance(0, 0))
6454 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6455 if (sched_domains_numa_distance
[i
] == distance
)
6463 * A system can have three types of NUMA topology:
6464 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6465 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6466 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6468 * The difference between a glueless mesh topology and a backplane
6469 * topology lies in whether communication between not directly
6470 * connected nodes goes through intermediary nodes (where programs
6471 * could run), or through backplane controllers. This affects
6472 * placement of programs.
6474 * The type of topology can be discerned with the following tests:
6475 * - If the maximum distance between any nodes is 1 hop, the system
6476 * is directly connected.
6477 * - If for two nodes A and B, located N > 1 hops away from each other,
6478 * there is an intermediary node C, which is < N hops away from both
6479 * nodes A and B, the system is a glueless mesh.
6481 static void init_numa_topology_type(void)
6485 n
= sched_max_numa_distance
;
6487 if (sched_domains_numa_levels
<= 1) {
6488 sched_numa_topology_type
= NUMA_DIRECT
;
6492 for_each_online_node(a
) {
6493 for_each_online_node(b
) {
6494 /* Find two nodes furthest removed from each other. */
6495 if (node_distance(a
, b
) < n
)
6498 /* Is there an intermediary node between a and b? */
6499 for_each_online_node(c
) {
6500 if (node_distance(a
, c
) < n
&&
6501 node_distance(b
, c
) < n
) {
6502 sched_numa_topology_type
=
6508 sched_numa_topology_type
= NUMA_BACKPLANE
;
6514 static void sched_init_numa(void)
6516 int next_distance
, curr_distance
= node_distance(0, 0);
6517 struct sched_domain_topology_level
*tl
;
6521 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6522 if (!sched_domains_numa_distance
)
6526 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6527 * unique distances in the node_distance() table.
6529 * Assumes node_distance(0,j) includes all distances in
6530 * node_distance(i,j) in order to avoid cubic time.
6532 next_distance
= curr_distance
;
6533 for (i
= 0; i
< nr_node_ids
; i
++) {
6534 for (j
= 0; j
< nr_node_ids
; j
++) {
6535 for (k
= 0; k
< nr_node_ids
; k
++) {
6536 int distance
= node_distance(i
, k
);
6538 if (distance
> curr_distance
&&
6539 (distance
< next_distance
||
6540 next_distance
== curr_distance
))
6541 next_distance
= distance
;
6544 * While not a strong assumption it would be nice to know
6545 * about cases where if node A is connected to B, B is not
6546 * equally connected to A.
6548 if (sched_debug() && node_distance(k
, i
) != distance
)
6549 sched_numa_warn("Node-distance not symmetric");
6551 if (sched_debug() && i
&& !find_numa_distance(distance
))
6552 sched_numa_warn("Node-0 not representative");
6554 if (next_distance
!= curr_distance
) {
6555 sched_domains_numa_distance
[level
++] = next_distance
;
6556 sched_domains_numa_levels
= level
;
6557 curr_distance
= next_distance
;
6562 * In case of sched_debug() we verify the above assumption.
6572 * 'level' contains the number of unique distances, excluding the
6573 * identity distance node_distance(i,i).
6575 * The sched_domains_numa_distance[] array includes the actual distance
6580 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6581 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6582 * the array will contain less then 'level' members. This could be
6583 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6584 * in other functions.
6586 * We reset it to 'level' at the end of this function.
6588 sched_domains_numa_levels
= 0;
6590 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6591 if (!sched_domains_numa_masks
)
6595 * Now for each level, construct a mask per node which contains all
6596 * cpus of nodes that are that many hops away from us.
6598 for (i
= 0; i
< level
; i
++) {
6599 sched_domains_numa_masks
[i
] =
6600 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6601 if (!sched_domains_numa_masks
[i
])
6604 for (j
= 0; j
< nr_node_ids
; j
++) {
6605 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6609 sched_domains_numa_masks
[i
][j
] = mask
;
6612 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6615 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6620 /* Compute default topology size */
6621 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6623 tl
= kzalloc((i
+ level
+ 1) *
6624 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6629 * Copy the default topology bits..
6631 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6632 tl
[i
] = sched_domain_topology
[i
];
6635 * .. and append 'j' levels of NUMA goodness.
6637 for (j
= 0; j
< level
; i
++, j
++) {
6638 tl
[i
] = (struct sched_domain_topology_level
){
6639 .mask
= sd_numa_mask
,
6640 .sd_flags
= cpu_numa_flags
,
6641 .flags
= SDTL_OVERLAP
,
6647 sched_domain_topology
= tl
;
6649 sched_domains_numa_levels
= level
;
6650 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6652 init_numa_topology_type();
6655 static void sched_domains_numa_masks_set(int cpu
)
6658 int node
= cpu_to_node(cpu
);
6660 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6661 for (j
= 0; j
< nr_node_ids
; j
++) {
6662 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6663 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6668 static void sched_domains_numa_masks_clear(int cpu
)
6671 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6672 for (j
= 0; j
< nr_node_ids
; j
++)
6673 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6678 * Update sched_domains_numa_masks[level][node] array when new cpus
6681 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6682 unsigned long action
,
6685 int cpu
= (long)hcpu
;
6687 switch (action
& ~CPU_TASKS_FROZEN
) {
6689 sched_domains_numa_masks_set(cpu
);
6693 sched_domains_numa_masks_clear(cpu
);
6703 static inline void sched_init_numa(void)
6707 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6708 unsigned long action
,
6713 #endif /* CONFIG_NUMA */
6715 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6717 struct sched_domain_topology_level
*tl
;
6720 for_each_sd_topology(tl
) {
6721 struct sd_data
*sdd
= &tl
->data
;
6723 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6727 sdd
->sg
= alloc_percpu(struct sched_group
*);
6731 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6735 for_each_cpu(j
, cpu_map
) {
6736 struct sched_domain
*sd
;
6737 struct sched_group
*sg
;
6738 struct sched_group_capacity
*sgc
;
6740 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6741 GFP_KERNEL
, cpu_to_node(j
));
6745 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6747 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6748 GFP_KERNEL
, cpu_to_node(j
));
6754 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6756 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6757 GFP_KERNEL
, cpu_to_node(j
));
6761 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6768 static void __sdt_free(const struct cpumask
*cpu_map
)
6770 struct sched_domain_topology_level
*tl
;
6773 for_each_sd_topology(tl
) {
6774 struct sd_data
*sdd
= &tl
->data
;
6776 for_each_cpu(j
, cpu_map
) {
6777 struct sched_domain
*sd
;
6780 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6781 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6782 free_sched_groups(sd
->groups
, 0);
6783 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6787 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6789 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6791 free_percpu(sdd
->sd
);
6793 free_percpu(sdd
->sg
);
6795 free_percpu(sdd
->sgc
);
6800 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6801 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6802 struct sched_domain
*child
, int cpu
)
6804 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6808 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6810 sd
->level
= child
->level
+ 1;
6811 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6815 if (!cpumask_subset(sched_domain_span(child
),
6816 sched_domain_span(sd
))) {
6817 pr_err("BUG: arch topology borken\n");
6818 #ifdef CONFIG_SCHED_DEBUG
6819 pr_err(" the %s domain not a subset of the %s domain\n",
6820 child
->name
, sd
->name
);
6822 /* Fixup, ensure @sd has at least @child cpus. */
6823 cpumask_or(sched_domain_span(sd
),
6824 sched_domain_span(sd
),
6825 sched_domain_span(child
));
6829 set_domain_attribute(sd
, attr
);
6835 * Build sched domains for a given set of cpus and attach the sched domains
6836 * to the individual cpus
6838 static int build_sched_domains(const struct cpumask
*cpu_map
,
6839 struct sched_domain_attr
*attr
)
6841 enum s_alloc alloc_state
;
6842 struct sched_domain
*sd
;
6844 int i
, ret
= -ENOMEM
;
6846 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6847 if (alloc_state
!= sa_rootdomain
)
6850 /* Set up domains for cpus specified by the cpu_map. */
6851 for_each_cpu(i
, cpu_map
) {
6852 struct sched_domain_topology_level
*tl
;
6855 for_each_sd_topology(tl
) {
6856 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6857 if (tl
== sched_domain_topology
)
6858 *per_cpu_ptr(d
.sd
, i
) = sd
;
6859 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6860 sd
->flags
|= SD_OVERLAP
;
6861 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6866 /* Build the groups for the domains */
6867 for_each_cpu(i
, cpu_map
) {
6868 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6869 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6870 if (sd
->flags
& SD_OVERLAP
) {
6871 if (build_overlap_sched_groups(sd
, i
))
6874 if (build_sched_groups(sd
, i
))
6880 /* Calculate CPU capacity for physical packages and nodes */
6881 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6882 if (!cpumask_test_cpu(i
, cpu_map
))
6885 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6886 claim_allocations(i
, sd
);
6887 init_sched_groups_capacity(i
, sd
);
6891 /* Attach the domains */
6893 for_each_cpu(i
, cpu_map
) {
6894 sd
= *per_cpu_ptr(d
.sd
, i
);
6895 cpu_attach_domain(sd
, d
.rd
, i
);
6901 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6905 static cpumask_var_t
*doms_cur
; /* current sched domains */
6906 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6907 static struct sched_domain_attr
*dattr_cur
;
6908 /* attribues of custom domains in 'doms_cur' */
6911 * Special case: If a kmalloc of a doms_cur partition (array of
6912 * cpumask) fails, then fallback to a single sched domain,
6913 * as determined by the single cpumask fallback_doms.
6915 static cpumask_var_t fallback_doms
;
6918 * arch_update_cpu_topology lets virtualized architectures update the
6919 * cpu core maps. It is supposed to return 1 if the topology changed
6920 * or 0 if it stayed the same.
6922 int __weak
arch_update_cpu_topology(void)
6927 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6930 cpumask_var_t
*doms
;
6932 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6935 for (i
= 0; i
< ndoms
; i
++) {
6936 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6937 free_sched_domains(doms
, i
);
6944 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6947 for (i
= 0; i
< ndoms
; i
++)
6948 free_cpumask_var(doms
[i
]);
6953 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6954 * For now this just excludes isolated cpus, but could be used to
6955 * exclude other special cases in the future.
6957 static int init_sched_domains(const struct cpumask
*cpu_map
)
6961 arch_update_cpu_topology();
6963 doms_cur
= alloc_sched_domains(ndoms_cur
);
6965 doms_cur
= &fallback_doms
;
6966 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6967 err
= build_sched_domains(doms_cur
[0], NULL
);
6968 register_sched_domain_sysctl();
6974 * Detach sched domains from a group of cpus specified in cpu_map
6975 * These cpus will now be attached to the NULL domain
6977 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6982 for_each_cpu(i
, cpu_map
)
6983 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6987 /* handle null as "default" */
6988 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6989 struct sched_domain_attr
*new, int idx_new
)
6991 struct sched_domain_attr tmp
;
6998 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6999 new ? (new + idx_new
) : &tmp
,
7000 sizeof(struct sched_domain_attr
));
7004 * Partition sched domains as specified by the 'ndoms_new'
7005 * cpumasks in the array doms_new[] of cpumasks. This compares
7006 * doms_new[] to the current sched domain partitioning, doms_cur[].
7007 * It destroys each deleted domain and builds each new domain.
7009 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7010 * The masks don't intersect (don't overlap.) We should setup one
7011 * sched domain for each mask. CPUs not in any of the cpumasks will
7012 * not be load balanced. If the same cpumask appears both in the
7013 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7016 * The passed in 'doms_new' should be allocated using
7017 * alloc_sched_domains. This routine takes ownership of it and will
7018 * free_sched_domains it when done with it. If the caller failed the
7019 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7020 * and partition_sched_domains() will fallback to the single partition
7021 * 'fallback_doms', it also forces the domains to be rebuilt.
7023 * If doms_new == NULL it will be replaced with cpu_online_mask.
7024 * ndoms_new == 0 is a special case for destroying existing domains,
7025 * and it will not create the default domain.
7027 * Call with hotplug lock held
7029 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7030 struct sched_domain_attr
*dattr_new
)
7035 mutex_lock(&sched_domains_mutex
);
7037 /* always unregister in case we don't destroy any domains */
7038 unregister_sched_domain_sysctl();
7040 /* Let architecture update cpu core mappings. */
7041 new_topology
= arch_update_cpu_topology();
7043 n
= doms_new
? ndoms_new
: 0;
7045 /* Destroy deleted domains */
7046 for (i
= 0; i
< ndoms_cur
; i
++) {
7047 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7048 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7049 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7052 /* no match - a current sched domain not in new doms_new[] */
7053 detach_destroy_domains(doms_cur
[i
]);
7059 if (doms_new
== NULL
) {
7061 doms_new
= &fallback_doms
;
7062 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7063 WARN_ON_ONCE(dattr_new
);
7066 /* Build new domains */
7067 for (i
= 0; i
< ndoms_new
; i
++) {
7068 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7069 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7070 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7073 /* no match - add a new doms_new */
7074 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7079 /* Remember the new sched domains */
7080 if (doms_cur
!= &fallback_doms
)
7081 free_sched_domains(doms_cur
, ndoms_cur
);
7082 kfree(dattr_cur
); /* kfree(NULL) is safe */
7083 doms_cur
= doms_new
;
7084 dattr_cur
= dattr_new
;
7085 ndoms_cur
= ndoms_new
;
7087 register_sched_domain_sysctl();
7089 mutex_unlock(&sched_domains_mutex
);
7092 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7095 * Update cpusets according to cpu_active mask. If cpusets are
7096 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7097 * around partition_sched_domains().
7099 * If we come here as part of a suspend/resume, don't touch cpusets because we
7100 * want to restore it back to its original state upon resume anyway.
7102 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7106 case CPU_ONLINE_FROZEN
:
7107 case CPU_DOWN_FAILED_FROZEN
:
7110 * num_cpus_frozen tracks how many CPUs are involved in suspend
7111 * resume sequence. As long as this is not the last online
7112 * operation in the resume sequence, just build a single sched
7113 * domain, ignoring cpusets.
7116 if (likely(num_cpus_frozen
)) {
7117 partition_sched_domains(1, NULL
, NULL
);
7122 * This is the last CPU online operation. So fall through and
7123 * restore the original sched domains by considering the
7124 * cpuset configurations.
7128 cpuset_update_active_cpus(true);
7136 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7139 unsigned long flags
;
7140 long cpu
= (long)hcpu
;
7146 case CPU_DOWN_PREPARE
:
7147 rcu_read_lock_sched();
7148 dl_b
= dl_bw_of(cpu
);
7150 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7151 cpus
= dl_bw_cpus(cpu
);
7152 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7153 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7155 rcu_read_unlock_sched();
7158 return notifier_from_errno(-EBUSY
);
7159 cpuset_update_active_cpus(false);
7161 case CPU_DOWN_PREPARE_FROZEN
:
7163 partition_sched_domains(1, NULL
, NULL
);
7171 void __init
sched_init_smp(void)
7173 cpumask_var_t non_isolated_cpus
;
7175 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7176 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7181 * There's no userspace yet to cause hotplug operations; hence all the
7182 * cpu masks are stable and all blatant races in the below code cannot
7185 mutex_lock(&sched_domains_mutex
);
7186 init_sched_domains(cpu_active_mask
);
7187 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7188 if (cpumask_empty(non_isolated_cpus
))
7189 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7190 mutex_unlock(&sched_domains_mutex
);
7192 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7193 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7194 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7198 /* Move init over to a non-isolated CPU */
7199 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7201 sched_init_granularity();
7202 free_cpumask_var(non_isolated_cpus
);
7204 init_sched_rt_class();
7205 init_sched_dl_class();
7208 void __init
sched_init_smp(void)
7210 sched_init_granularity();
7212 #endif /* CONFIG_SMP */
7214 int in_sched_functions(unsigned long addr
)
7216 return in_lock_functions(addr
) ||
7217 (addr
>= (unsigned long)__sched_text_start
7218 && addr
< (unsigned long)__sched_text_end
);
7221 #ifdef CONFIG_CGROUP_SCHED
7223 * Default task group.
7224 * Every task in system belongs to this group at bootup.
7226 struct task_group root_task_group
;
7227 LIST_HEAD(task_groups
);
7229 /* Cacheline aligned slab cache for task_group */
7230 static struct kmem_cache
*task_group_cache __read_mostly
;
7233 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7235 void __init
sched_init(void)
7238 unsigned long alloc_size
= 0, ptr
;
7240 #ifdef CONFIG_FAIR_GROUP_SCHED
7241 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7243 #ifdef CONFIG_RT_GROUP_SCHED
7244 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7247 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7249 #ifdef CONFIG_FAIR_GROUP_SCHED
7250 root_task_group
.se
= (struct sched_entity
**)ptr
;
7251 ptr
+= nr_cpu_ids
* sizeof(void **);
7253 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7254 ptr
+= nr_cpu_ids
* sizeof(void **);
7256 #endif /* CONFIG_FAIR_GROUP_SCHED */
7257 #ifdef CONFIG_RT_GROUP_SCHED
7258 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7259 ptr
+= nr_cpu_ids
* sizeof(void **);
7261 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7262 ptr
+= nr_cpu_ids
* sizeof(void **);
7264 #endif /* CONFIG_RT_GROUP_SCHED */
7266 #ifdef CONFIG_CPUMASK_OFFSTACK
7267 for_each_possible_cpu(i
) {
7268 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7269 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7271 #endif /* CONFIG_CPUMASK_OFFSTACK */
7273 init_rt_bandwidth(&def_rt_bandwidth
,
7274 global_rt_period(), global_rt_runtime());
7275 init_dl_bandwidth(&def_dl_bandwidth
,
7276 global_rt_period(), global_rt_runtime());
7279 init_defrootdomain();
7282 #ifdef CONFIG_RT_GROUP_SCHED
7283 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7284 global_rt_period(), global_rt_runtime());
7285 #endif /* CONFIG_RT_GROUP_SCHED */
7287 #ifdef CONFIG_CGROUP_SCHED
7288 task_group_cache
= KMEM_CACHE(task_group
, 0);
7290 list_add(&root_task_group
.list
, &task_groups
);
7291 INIT_LIST_HEAD(&root_task_group
.children
);
7292 INIT_LIST_HEAD(&root_task_group
.siblings
);
7293 autogroup_init(&init_task
);
7294 #endif /* CONFIG_CGROUP_SCHED */
7296 for_each_possible_cpu(i
) {
7300 raw_spin_lock_init(&rq
->lock
);
7302 rq
->calc_load_active
= 0;
7303 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7304 init_cfs_rq(&rq
->cfs
);
7305 init_rt_rq(&rq
->rt
);
7306 init_dl_rq(&rq
->dl
);
7307 #ifdef CONFIG_FAIR_GROUP_SCHED
7308 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7309 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7311 * How much cpu bandwidth does root_task_group get?
7313 * In case of task-groups formed thr' the cgroup filesystem, it
7314 * gets 100% of the cpu resources in the system. This overall
7315 * system cpu resource is divided among the tasks of
7316 * root_task_group and its child task-groups in a fair manner,
7317 * based on each entity's (task or task-group's) weight
7318 * (se->load.weight).
7320 * In other words, if root_task_group has 10 tasks of weight
7321 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7322 * then A0's share of the cpu resource is:
7324 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7326 * We achieve this by letting root_task_group's tasks sit
7327 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7329 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7330 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7331 #endif /* CONFIG_FAIR_GROUP_SCHED */
7333 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7334 #ifdef CONFIG_RT_GROUP_SCHED
7335 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7338 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7339 rq
->cpu_load
[j
] = 0;
7341 rq
->last_load_update_tick
= jiffies
;
7346 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7347 rq
->balance_callback
= NULL
;
7348 rq
->active_balance
= 0;
7349 rq
->next_balance
= jiffies
;
7354 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7355 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7357 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7359 rq_attach_root(rq
, &def_root_domain
);
7360 #ifdef CONFIG_NO_HZ_COMMON
7363 #ifdef CONFIG_NO_HZ_FULL
7364 rq
->last_sched_tick
= 0;
7368 atomic_set(&rq
->nr_iowait
, 0);
7371 set_load_weight(&init_task
);
7373 #ifdef CONFIG_PREEMPT_NOTIFIERS
7374 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7378 * The boot idle thread does lazy MMU switching as well:
7380 atomic_inc(&init_mm
.mm_count
);
7381 enter_lazy_tlb(&init_mm
, current
);
7384 * During early bootup we pretend to be a normal task:
7386 current
->sched_class
= &fair_sched_class
;
7389 * Make us the idle thread. Technically, schedule() should not be
7390 * called from this thread, however somewhere below it might be,
7391 * but because we are the idle thread, we just pick up running again
7392 * when this runqueue becomes "idle".
7394 init_idle(current
, smp_processor_id());
7396 calc_load_update
= jiffies
+ LOAD_FREQ
;
7399 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7400 /* May be allocated at isolcpus cmdline parse time */
7401 if (cpu_isolated_map
== NULL
)
7402 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7403 idle_thread_set_boot_cpu();
7404 set_cpu_rq_start_time();
7406 init_sched_fair_class();
7408 scheduler_running
= 1;
7411 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7412 static inline int preempt_count_equals(int preempt_offset
)
7414 int nested
= preempt_count() + rcu_preempt_depth();
7416 return (nested
== preempt_offset
);
7419 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7422 * Blocking primitives will set (and therefore destroy) current->state,
7423 * since we will exit with TASK_RUNNING make sure we enter with it,
7424 * otherwise we will destroy state.
7426 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7427 "do not call blocking ops when !TASK_RUNNING; "
7428 "state=%lx set at [<%p>] %pS\n",
7430 (void *)current
->task_state_change
,
7431 (void *)current
->task_state_change
);
7433 ___might_sleep(file
, line
, preempt_offset
);
7435 EXPORT_SYMBOL(__might_sleep
);
7437 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7439 static unsigned long prev_jiffy
; /* ratelimiting */
7441 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7442 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7443 !is_idle_task(current
)) ||
7444 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7446 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7448 prev_jiffy
= jiffies
;
7451 "BUG: sleeping function called from invalid context at %s:%d\n",
7454 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7455 in_atomic(), irqs_disabled(),
7456 current
->pid
, current
->comm
);
7458 if (task_stack_end_corrupted(current
))
7459 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7461 debug_show_held_locks(current
);
7462 if (irqs_disabled())
7463 print_irqtrace_events(current
);
7464 #ifdef CONFIG_DEBUG_PREEMPT
7465 if (!preempt_count_equals(preempt_offset
)) {
7466 pr_err("Preemption disabled at:");
7467 print_ip_sym(current
->preempt_disable_ip
);
7473 EXPORT_SYMBOL(___might_sleep
);
7476 #ifdef CONFIG_MAGIC_SYSRQ
7477 void normalize_rt_tasks(void)
7479 struct task_struct
*g
, *p
;
7480 struct sched_attr attr
= {
7481 .sched_policy
= SCHED_NORMAL
,
7484 read_lock(&tasklist_lock
);
7485 for_each_process_thread(g
, p
) {
7487 * Only normalize user tasks:
7489 if (p
->flags
& PF_KTHREAD
)
7492 p
->se
.exec_start
= 0;
7493 #ifdef CONFIG_SCHEDSTATS
7494 p
->se
.statistics
.wait_start
= 0;
7495 p
->se
.statistics
.sleep_start
= 0;
7496 p
->se
.statistics
.block_start
= 0;
7499 if (!dl_task(p
) && !rt_task(p
)) {
7501 * Renice negative nice level userspace
7504 if (task_nice(p
) < 0)
7505 set_user_nice(p
, 0);
7509 __sched_setscheduler(p
, &attr
, false, false);
7511 read_unlock(&tasklist_lock
);
7514 #endif /* CONFIG_MAGIC_SYSRQ */
7516 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7518 * These functions are only useful for the IA64 MCA handling, or kdb.
7520 * They can only be called when the whole system has been
7521 * stopped - every CPU needs to be quiescent, and no scheduling
7522 * activity can take place. Using them for anything else would
7523 * be a serious bug, and as a result, they aren't even visible
7524 * under any other configuration.
7528 * curr_task - return the current task for a given cpu.
7529 * @cpu: the processor in question.
7531 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7533 * Return: The current task for @cpu.
7535 struct task_struct
*curr_task(int cpu
)
7537 return cpu_curr(cpu
);
7540 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7544 * set_curr_task - set the current task for a given cpu.
7545 * @cpu: the processor in question.
7546 * @p: the task pointer to set.
7548 * Description: This function must only be used when non-maskable interrupts
7549 * are serviced on a separate stack. It allows the architecture to switch the
7550 * notion of the current task on a cpu in a non-blocking manner. This function
7551 * must be called with all CPU's synchronized, and interrupts disabled, the
7552 * and caller must save the original value of the current task (see
7553 * curr_task() above) and restore that value before reenabling interrupts and
7554 * re-starting the system.
7556 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7558 void set_curr_task(int cpu
, struct task_struct
*p
)
7565 #ifdef CONFIG_CGROUP_SCHED
7566 /* task_group_lock serializes the addition/removal of task groups */
7567 static DEFINE_SPINLOCK(task_group_lock
);
7569 static void sched_free_group(struct task_group
*tg
)
7571 free_fair_sched_group(tg
);
7572 free_rt_sched_group(tg
);
7574 kmem_cache_free(task_group_cache
, tg
);
7577 /* allocate runqueue etc for a new task group */
7578 struct task_group
*sched_create_group(struct task_group
*parent
)
7580 struct task_group
*tg
;
7582 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7584 return ERR_PTR(-ENOMEM
);
7586 if (!alloc_fair_sched_group(tg
, parent
))
7589 if (!alloc_rt_sched_group(tg
, parent
))
7595 sched_free_group(tg
);
7596 return ERR_PTR(-ENOMEM
);
7599 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7601 unsigned long flags
;
7603 spin_lock_irqsave(&task_group_lock
, flags
);
7604 list_add_rcu(&tg
->list
, &task_groups
);
7606 WARN_ON(!parent
); /* root should already exist */
7608 tg
->parent
= parent
;
7609 INIT_LIST_HEAD(&tg
->children
);
7610 list_add_rcu(&tg
->siblings
, &parent
->children
);
7611 spin_unlock_irqrestore(&task_group_lock
, flags
);
7614 /* rcu callback to free various structures associated with a task group */
7615 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7617 /* now it should be safe to free those cfs_rqs */
7618 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7621 void sched_destroy_group(struct task_group
*tg
)
7623 /* wait for possible concurrent references to cfs_rqs complete */
7624 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7627 void sched_offline_group(struct task_group
*tg
)
7629 unsigned long flags
;
7631 /* end participation in shares distribution */
7632 unregister_fair_sched_group(tg
);
7634 spin_lock_irqsave(&task_group_lock
, flags
);
7635 list_del_rcu(&tg
->list
);
7636 list_del_rcu(&tg
->siblings
);
7637 spin_unlock_irqrestore(&task_group_lock
, flags
);
7640 /* change task's runqueue when it moves between groups.
7641 * The caller of this function should have put the task in its new group
7642 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7643 * reflect its new group.
7645 void sched_move_task(struct task_struct
*tsk
)
7647 struct task_group
*tg
;
7648 int queued
, running
;
7649 unsigned long flags
;
7652 rq
= task_rq_lock(tsk
, &flags
);
7654 running
= task_current(rq
, tsk
);
7655 queued
= task_on_rq_queued(tsk
);
7658 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7659 if (unlikely(running
))
7660 put_prev_task(rq
, tsk
);
7663 * All callers are synchronized by task_rq_lock(); we do not use RCU
7664 * which is pointless here. Thus, we pass "true" to task_css_check()
7665 * to prevent lockdep warnings.
7667 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7668 struct task_group
, css
);
7669 tg
= autogroup_task_group(tsk
, tg
);
7670 tsk
->sched_task_group
= tg
;
7672 #ifdef CONFIG_FAIR_GROUP_SCHED
7673 if (tsk
->sched_class
->task_move_group
)
7674 tsk
->sched_class
->task_move_group(tsk
);
7677 set_task_rq(tsk
, task_cpu(tsk
));
7679 if (unlikely(running
))
7680 tsk
->sched_class
->set_curr_task(rq
);
7682 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7684 task_rq_unlock(rq
, tsk
, &flags
);
7686 #endif /* CONFIG_CGROUP_SCHED */
7688 #ifdef CONFIG_RT_GROUP_SCHED
7690 * Ensure that the real time constraints are schedulable.
7692 static DEFINE_MUTEX(rt_constraints_mutex
);
7694 /* Must be called with tasklist_lock held */
7695 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7697 struct task_struct
*g
, *p
;
7700 * Autogroups do not have RT tasks; see autogroup_create().
7702 if (task_group_is_autogroup(tg
))
7705 for_each_process_thread(g
, p
) {
7706 if (rt_task(p
) && task_group(p
) == tg
)
7713 struct rt_schedulable_data
{
7714 struct task_group
*tg
;
7719 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7721 struct rt_schedulable_data
*d
= data
;
7722 struct task_group
*child
;
7723 unsigned long total
, sum
= 0;
7724 u64 period
, runtime
;
7726 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7727 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7730 period
= d
->rt_period
;
7731 runtime
= d
->rt_runtime
;
7735 * Cannot have more runtime than the period.
7737 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7741 * Ensure we don't starve existing RT tasks.
7743 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7746 total
= to_ratio(period
, runtime
);
7749 * Nobody can have more than the global setting allows.
7751 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7755 * The sum of our children's runtime should not exceed our own.
7757 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7758 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7759 runtime
= child
->rt_bandwidth
.rt_runtime
;
7761 if (child
== d
->tg
) {
7762 period
= d
->rt_period
;
7763 runtime
= d
->rt_runtime
;
7766 sum
+= to_ratio(period
, runtime
);
7775 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7779 struct rt_schedulable_data data
= {
7781 .rt_period
= period
,
7782 .rt_runtime
= runtime
,
7786 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7792 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7793 u64 rt_period
, u64 rt_runtime
)
7798 * Disallowing the root group RT runtime is BAD, it would disallow the
7799 * kernel creating (and or operating) RT threads.
7801 if (tg
== &root_task_group
&& rt_runtime
== 0)
7804 /* No period doesn't make any sense. */
7808 mutex_lock(&rt_constraints_mutex
);
7809 read_lock(&tasklist_lock
);
7810 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7814 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7815 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7816 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7818 for_each_possible_cpu(i
) {
7819 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7821 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7822 rt_rq
->rt_runtime
= rt_runtime
;
7823 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7825 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7827 read_unlock(&tasklist_lock
);
7828 mutex_unlock(&rt_constraints_mutex
);
7833 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7835 u64 rt_runtime
, rt_period
;
7837 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7838 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7839 if (rt_runtime_us
< 0)
7840 rt_runtime
= RUNTIME_INF
;
7842 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7845 static long sched_group_rt_runtime(struct task_group
*tg
)
7849 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7852 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7853 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7854 return rt_runtime_us
;
7857 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7859 u64 rt_runtime
, rt_period
;
7861 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7862 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7864 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7867 static long sched_group_rt_period(struct task_group
*tg
)
7871 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7872 do_div(rt_period_us
, NSEC_PER_USEC
);
7873 return rt_period_us
;
7875 #endif /* CONFIG_RT_GROUP_SCHED */
7877 #ifdef CONFIG_RT_GROUP_SCHED
7878 static int sched_rt_global_constraints(void)
7882 mutex_lock(&rt_constraints_mutex
);
7883 read_lock(&tasklist_lock
);
7884 ret
= __rt_schedulable(NULL
, 0, 0);
7885 read_unlock(&tasklist_lock
);
7886 mutex_unlock(&rt_constraints_mutex
);
7891 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7893 /* Don't accept realtime tasks when there is no way for them to run */
7894 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7900 #else /* !CONFIG_RT_GROUP_SCHED */
7901 static int sched_rt_global_constraints(void)
7903 unsigned long flags
;
7906 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7907 for_each_possible_cpu(i
) {
7908 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7910 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7911 rt_rq
->rt_runtime
= global_rt_runtime();
7912 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7914 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7918 #endif /* CONFIG_RT_GROUP_SCHED */
7920 static int sched_dl_global_validate(void)
7922 u64 runtime
= global_rt_runtime();
7923 u64 period
= global_rt_period();
7924 u64 new_bw
= to_ratio(period
, runtime
);
7927 unsigned long flags
;
7930 * Here we want to check the bandwidth not being set to some
7931 * value smaller than the currently allocated bandwidth in
7932 * any of the root_domains.
7934 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7935 * cycling on root_domains... Discussion on different/better
7936 * solutions is welcome!
7938 for_each_possible_cpu(cpu
) {
7939 rcu_read_lock_sched();
7940 dl_b
= dl_bw_of(cpu
);
7942 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7943 if (new_bw
< dl_b
->total_bw
)
7945 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7947 rcu_read_unlock_sched();
7956 static void sched_dl_do_global(void)
7961 unsigned long flags
;
7963 def_dl_bandwidth
.dl_period
= global_rt_period();
7964 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7966 if (global_rt_runtime() != RUNTIME_INF
)
7967 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7970 * FIXME: As above...
7972 for_each_possible_cpu(cpu
) {
7973 rcu_read_lock_sched();
7974 dl_b
= dl_bw_of(cpu
);
7976 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7978 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7980 rcu_read_unlock_sched();
7984 static int sched_rt_global_validate(void)
7986 if (sysctl_sched_rt_period
<= 0)
7989 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7990 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7996 static void sched_rt_do_global(void)
7998 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7999 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8002 int sched_rt_handler(struct ctl_table
*table
, int write
,
8003 void __user
*buffer
, size_t *lenp
,
8006 int old_period
, old_runtime
;
8007 static DEFINE_MUTEX(mutex
);
8011 old_period
= sysctl_sched_rt_period
;
8012 old_runtime
= sysctl_sched_rt_runtime
;
8014 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8016 if (!ret
&& write
) {
8017 ret
= sched_rt_global_validate();
8021 ret
= sched_dl_global_validate();
8025 ret
= sched_rt_global_constraints();
8029 sched_rt_do_global();
8030 sched_dl_do_global();
8034 sysctl_sched_rt_period
= old_period
;
8035 sysctl_sched_rt_runtime
= old_runtime
;
8037 mutex_unlock(&mutex
);
8042 int sched_rr_handler(struct ctl_table
*table
, int write
,
8043 void __user
*buffer
, size_t *lenp
,
8047 static DEFINE_MUTEX(mutex
);
8050 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8051 /* make sure that internally we keep jiffies */
8052 /* also, writing zero resets timeslice to default */
8053 if (!ret
&& write
) {
8054 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8055 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8057 mutex_unlock(&mutex
);
8061 #ifdef CONFIG_CGROUP_SCHED
8063 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8065 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8068 static struct cgroup_subsys_state
*
8069 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8071 struct task_group
*parent
= css_tg(parent_css
);
8072 struct task_group
*tg
;
8075 /* This is early initialization for the top cgroup */
8076 return &root_task_group
.css
;
8079 tg
= sched_create_group(parent
);
8081 return ERR_PTR(-ENOMEM
);
8083 sched_online_group(tg
, parent
);
8088 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8090 struct task_group
*tg
= css_tg(css
);
8092 sched_offline_group(tg
);
8095 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8097 struct task_group
*tg
= css_tg(css
);
8100 * Relies on the RCU grace period between css_released() and this.
8102 sched_free_group(tg
);
8105 static void cpu_cgroup_fork(struct task_struct
*task
)
8107 sched_move_task(task
);
8110 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8112 struct task_struct
*task
;
8113 struct cgroup_subsys_state
*css
;
8115 cgroup_taskset_for_each(task
, css
, tset
) {
8116 #ifdef CONFIG_RT_GROUP_SCHED
8117 if (!sched_rt_can_attach(css_tg(css
), task
))
8120 /* We don't support RT-tasks being in separate groups */
8121 if (task
->sched_class
!= &fair_sched_class
)
8128 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8130 struct task_struct
*task
;
8131 struct cgroup_subsys_state
*css
;
8133 cgroup_taskset_for_each(task
, css
, tset
)
8134 sched_move_task(task
);
8137 #ifdef CONFIG_FAIR_GROUP_SCHED
8138 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8139 struct cftype
*cftype
, u64 shareval
)
8141 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8144 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8147 struct task_group
*tg
= css_tg(css
);
8149 return (u64
) scale_load_down(tg
->shares
);
8152 #ifdef CONFIG_CFS_BANDWIDTH
8153 static DEFINE_MUTEX(cfs_constraints_mutex
);
8155 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8156 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8158 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8160 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8162 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8163 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8165 if (tg
== &root_task_group
)
8169 * Ensure we have at some amount of bandwidth every period. This is
8170 * to prevent reaching a state of large arrears when throttled via
8171 * entity_tick() resulting in prolonged exit starvation.
8173 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8177 * Likewise, bound things on the otherside by preventing insane quota
8178 * periods. This also allows us to normalize in computing quota
8181 if (period
> max_cfs_quota_period
)
8185 * Prevent race between setting of cfs_rq->runtime_enabled and
8186 * unthrottle_offline_cfs_rqs().
8189 mutex_lock(&cfs_constraints_mutex
);
8190 ret
= __cfs_schedulable(tg
, period
, quota
);
8194 runtime_enabled
= quota
!= RUNTIME_INF
;
8195 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8197 * If we need to toggle cfs_bandwidth_used, off->on must occur
8198 * before making related changes, and on->off must occur afterwards
8200 if (runtime_enabled
&& !runtime_was_enabled
)
8201 cfs_bandwidth_usage_inc();
8202 raw_spin_lock_irq(&cfs_b
->lock
);
8203 cfs_b
->period
= ns_to_ktime(period
);
8204 cfs_b
->quota
= quota
;
8206 __refill_cfs_bandwidth_runtime(cfs_b
);
8207 /* restart the period timer (if active) to handle new period expiry */
8208 if (runtime_enabled
)
8209 start_cfs_bandwidth(cfs_b
);
8210 raw_spin_unlock_irq(&cfs_b
->lock
);
8212 for_each_online_cpu(i
) {
8213 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8214 struct rq
*rq
= cfs_rq
->rq
;
8216 raw_spin_lock_irq(&rq
->lock
);
8217 cfs_rq
->runtime_enabled
= runtime_enabled
;
8218 cfs_rq
->runtime_remaining
= 0;
8220 if (cfs_rq
->throttled
)
8221 unthrottle_cfs_rq(cfs_rq
);
8222 raw_spin_unlock_irq(&rq
->lock
);
8224 if (runtime_was_enabled
&& !runtime_enabled
)
8225 cfs_bandwidth_usage_dec();
8227 mutex_unlock(&cfs_constraints_mutex
);
8233 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8237 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8238 if (cfs_quota_us
< 0)
8239 quota
= RUNTIME_INF
;
8241 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8243 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8246 long tg_get_cfs_quota(struct task_group
*tg
)
8250 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8253 quota_us
= tg
->cfs_bandwidth
.quota
;
8254 do_div(quota_us
, NSEC_PER_USEC
);
8259 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8263 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8264 quota
= tg
->cfs_bandwidth
.quota
;
8266 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8269 long tg_get_cfs_period(struct task_group
*tg
)
8273 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8274 do_div(cfs_period_us
, NSEC_PER_USEC
);
8276 return cfs_period_us
;
8279 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8282 return tg_get_cfs_quota(css_tg(css
));
8285 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8286 struct cftype
*cftype
, s64 cfs_quota_us
)
8288 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8291 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8294 return tg_get_cfs_period(css_tg(css
));
8297 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8298 struct cftype
*cftype
, u64 cfs_period_us
)
8300 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8303 struct cfs_schedulable_data
{
8304 struct task_group
*tg
;
8309 * normalize group quota/period to be quota/max_period
8310 * note: units are usecs
8312 static u64
normalize_cfs_quota(struct task_group
*tg
,
8313 struct cfs_schedulable_data
*d
)
8321 period
= tg_get_cfs_period(tg
);
8322 quota
= tg_get_cfs_quota(tg
);
8325 /* note: these should typically be equivalent */
8326 if (quota
== RUNTIME_INF
|| quota
== -1)
8329 return to_ratio(period
, quota
);
8332 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8334 struct cfs_schedulable_data
*d
= data
;
8335 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8336 s64 quota
= 0, parent_quota
= -1;
8339 quota
= RUNTIME_INF
;
8341 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8343 quota
= normalize_cfs_quota(tg
, d
);
8344 parent_quota
= parent_b
->hierarchical_quota
;
8347 * ensure max(child_quota) <= parent_quota, inherit when no
8350 if (quota
== RUNTIME_INF
)
8351 quota
= parent_quota
;
8352 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8355 cfs_b
->hierarchical_quota
= quota
;
8360 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8363 struct cfs_schedulable_data data
= {
8369 if (quota
!= RUNTIME_INF
) {
8370 do_div(data
.period
, NSEC_PER_USEC
);
8371 do_div(data
.quota
, NSEC_PER_USEC
);
8375 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8381 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8383 struct task_group
*tg
= css_tg(seq_css(sf
));
8384 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8386 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8387 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8388 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8392 #endif /* CONFIG_CFS_BANDWIDTH */
8393 #endif /* CONFIG_FAIR_GROUP_SCHED */
8395 #ifdef CONFIG_RT_GROUP_SCHED
8396 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8397 struct cftype
*cft
, s64 val
)
8399 return sched_group_set_rt_runtime(css_tg(css
), val
);
8402 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8405 return sched_group_rt_runtime(css_tg(css
));
8408 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8409 struct cftype
*cftype
, u64 rt_period_us
)
8411 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8414 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8417 return sched_group_rt_period(css_tg(css
));
8419 #endif /* CONFIG_RT_GROUP_SCHED */
8421 static struct cftype cpu_files
[] = {
8422 #ifdef CONFIG_FAIR_GROUP_SCHED
8425 .read_u64
= cpu_shares_read_u64
,
8426 .write_u64
= cpu_shares_write_u64
,
8429 #ifdef CONFIG_CFS_BANDWIDTH
8431 .name
= "cfs_quota_us",
8432 .read_s64
= cpu_cfs_quota_read_s64
,
8433 .write_s64
= cpu_cfs_quota_write_s64
,
8436 .name
= "cfs_period_us",
8437 .read_u64
= cpu_cfs_period_read_u64
,
8438 .write_u64
= cpu_cfs_period_write_u64
,
8442 .seq_show
= cpu_stats_show
,
8445 #ifdef CONFIG_RT_GROUP_SCHED
8447 .name
= "rt_runtime_us",
8448 .read_s64
= cpu_rt_runtime_read
,
8449 .write_s64
= cpu_rt_runtime_write
,
8452 .name
= "rt_period_us",
8453 .read_u64
= cpu_rt_period_read_uint
,
8454 .write_u64
= cpu_rt_period_write_uint
,
8460 struct cgroup_subsys cpu_cgrp_subsys
= {
8461 .css_alloc
= cpu_cgroup_css_alloc
,
8462 .css_released
= cpu_cgroup_css_released
,
8463 .css_free
= cpu_cgroup_css_free
,
8464 .fork
= cpu_cgroup_fork
,
8465 .can_attach
= cpu_cgroup_can_attach
,
8466 .attach
= cpu_cgroup_attach
,
8467 .legacy_cftypes
= cpu_files
,
8471 #endif /* CONFIG_CGROUP_SCHED */
8473 void dump_cpu_task(int cpu
)
8475 pr_info("Task dump for CPU %d:\n", cpu
);
8476 sched_show_task(cpu_curr(cpu
));
8480 * Nice levels are multiplicative, with a gentle 10% change for every
8481 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8482 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8483 * that remained on nice 0.
8485 * The "10% effect" is relative and cumulative: from _any_ nice level,
8486 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8487 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8488 * If a task goes up by ~10% and another task goes down by ~10% then
8489 * the relative distance between them is ~25%.)
8491 const int sched_prio_to_weight
[40] = {
8492 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8493 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8494 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8495 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8496 /* 0 */ 1024, 820, 655, 526, 423,
8497 /* 5 */ 335, 272, 215, 172, 137,
8498 /* 10 */ 110, 87, 70, 56, 45,
8499 /* 15 */ 36, 29, 23, 18, 15,
8503 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8505 * In cases where the weight does not change often, we can use the
8506 * precalculated inverse to speed up arithmetics by turning divisions
8507 * into multiplications:
8509 const u32 sched_prio_to_wmult
[40] = {
8510 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8511 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8512 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8513 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8514 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8515 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8516 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8517 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,