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
)
2235 trace_sched_set_prio(p
, prio
);
2240 * fork()/clone()-time setup:
2242 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2244 unsigned long flags
;
2245 int cpu
= get_cpu();
2247 __sched_fork(clone_flags
, p
);
2249 * We mark the process as running here. This guarantees that
2250 * nobody will actually run it, and a signal or other external
2251 * event cannot wake it up and insert it on the runqueue either.
2253 p
->state
= TASK_RUNNING
;
2256 * Make sure we do not leak PI boosting priority to the child.
2258 sched_set_prio(p
, current
->normal_prio
);
2261 * Revert to default priority/policy on fork if requested.
2263 if (unlikely(p
->sched_reset_on_fork
)) {
2264 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2265 p
->policy
= SCHED_NORMAL
;
2266 p
->static_prio
= NICE_TO_PRIO(0);
2268 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2269 p
->static_prio
= NICE_TO_PRIO(0);
2271 p
->normal_prio
= __normal_prio(p
);
2272 sched_set_prio(p
, p
->normal_prio
);
2276 * We don't need the reset flag anymore after the fork. It has
2277 * fulfilled its duty:
2279 p
->sched_reset_on_fork
= 0;
2282 if (dl_prio(p
->prio
)) {
2285 } else if (rt_prio(p
->prio
)) {
2286 p
->sched_class
= &rt_sched_class
;
2288 p
->sched_class
= &fair_sched_class
;
2291 if (p
->sched_class
->task_fork
)
2292 p
->sched_class
->task_fork(p
);
2295 * The child is not yet in the pid-hash so no cgroup attach races,
2296 * and the cgroup is pinned to this child due to cgroup_fork()
2297 * is ran before sched_fork().
2299 * Silence PROVE_RCU.
2301 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2302 set_task_cpu(p
, cpu
);
2303 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2305 #ifdef CONFIG_SCHED_INFO
2306 if (likely(sched_info_on()))
2307 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2309 #if defined(CONFIG_SMP)
2312 init_task_preempt_count(p
);
2314 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2315 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2322 unsigned long to_ratio(u64 period
, u64 runtime
)
2324 if (runtime
== RUNTIME_INF
)
2328 * Doing this here saves a lot of checks in all
2329 * the calling paths, and returning zero seems
2330 * safe for them anyway.
2335 return div64_u64(runtime
<< 20, period
);
2339 inline struct dl_bw
*dl_bw_of(int i
)
2341 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2342 "sched RCU must be held");
2343 return &cpu_rq(i
)->rd
->dl_bw
;
2346 static inline int dl_bw_cpus(int i
)
2348 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2351 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2352 "sched RCU must be held");
2353 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2359 inline struct dl_bw
*dl_bw_of(int i
)
2361 return &cpu_rq(i
)->dl
.dl_bw
;
2364 static inline int dl_bw_cpus(int i
)
2371 * We must be sure that accepting a new task (or allowing changing the
2372 * parameters of an existing one) is consistent with the bandwidth
2373 * constraints. If yes, this function also accordingly updates the currently
2374 * allocated bandwidth to reflect the new situation.
2376 * This function is called while holding p's rq->lock.
2378 * XXX we should delay bw change until the task's 0-lag point, see
2381 static int dl_overflow(struct task_struct
*p
, int policy
,
2382 const struct sched_attr
*attr
)
2385 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2386 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2387 u64 runtime
= attr
->sched_runtime
;
2388 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2391 if (new_bw
== p
->dl
.dl_bw
)
2395 * Either if a task, enters, leave, or stays -deadline but changes
2396 * its parameters, we may need to update accordingly the total
2397 * allocated bandwidth of the container.
2399 raw_spin_lock(&dl_b
->lock
);
2400 cpus
= dl_bw_cpus(task_cpu(p
));
2401 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2402 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2403 __dl_add(dl_b
, new_bw
);
2405 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2406 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2407 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2408 __dl_add(dl_b
, new_bw
);
2410 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2411 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2414 raw_spin_unlock(&dl_b
->lock
);
2419 extern void init_dl_bw(struct dl_bw
*dl_b
);
2422 * wake_up_new_task - wake up a newly created task for the first time.
2424 * This function will do some initial scheduler statistics housekeeping
2425 * that must be done for every newly created context, then puts the task
2426 * on the runqueue and wakes it.
2428 void wake_up_new_task(struct task_struct
*p
)
2430 unsigned long flags
;
2433 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2434 /* Initialize new task's runnable average */
2435 init_entity_runnable_average(&p
->se
);
2438 * Fork balancing, do it here and not earlier because:
2439 * - cpus_allowed can change in the fork path
2440 * - any previously selected cpu might disappear through hotplug
2442 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2445 rq
= __task_rq_lock(p
);
2446 activate_task(rq
, p
, 0);
2447 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2448 trace_sched_wakeup_new(p
);
2449 check_preempt_curr(rq
, p
, WF_FORK
);
2451 if (p
->sched_class
->task_woken
) {
2453 * Nothing relies on rq->lock after this, so its fine to
2456 lockdep_unpin_lock(&rq
->lock
);
2457 p
->sched_class
->task_woken(rq
, p
);
2458 lockdep_pin_lock(&rq
->lock
);
2461 task_rq_unlock(rq
, p
, &flags
);
2464 #ifdef CONFIG_PREEMPT_NOTIFIERS
2466 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2468 void preempt_notifier_inc(void)
2470 static_key_slow_inc(&preempt_notifier_key
);
2472 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2474 void preempt_notifier_dec(void)
2476 static_key_slow_dec(&preempt_notifier_key
);
2478 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2481 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2482 * @notifier: notifier struct to register
2484 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2486 if (!static_key_false(&preempt_notifier_key
))
2487 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2489 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2491 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2494 * preempt_notifier_unregister - no longer interested in preemption notifications
2495 * @notifier: notifier struct to unregister
2497 * This is *not* safe to call from within a preemption notifier.
2499 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2501 hlist_del(¬ifier
->link
);
2503 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2505 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2507 struct preempt_notifier
*notifier
;
2509 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2510 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2513 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2515 if (static_key_false(&preempt_notifier_key
))
2516 __fire_sched_in_preempt_notifiers(curr
);
2520 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2521 struct task_struct
*next
)
2523 struct preempt_notifier
*notifier
;
2525 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2526 notifier
->ops
->sched_out(notifier
, next
);
2529 static __always_inline
void
2530 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2531 struct task_struct
*next
)
2533 if (static_key_false(&preempt_notifier_key
))
2534 __fire_sched_out_preempt_notifiers(curr
, next
);
2537 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2539 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2544 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2545 struct task_struct
*next
)
2549 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2552 * prepare_task_switch - prepare to switch tasks
2553 * @rq: the runqueue preparing to switch
2554 * @prev: the current task that is being switched out
2555 * @next: the task we are going to switch to.
2557 * This is called with the rq lock held and interrupts off. It must
2558 * be paired with a subsequent finish_task_switch after the context
2561 * prepare_task_switch sets up locking and calls architecture specific
2565 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2566 struct task_struct
*next
)
2568 sched_info_switch(rq
, prev
, next
);
2569 perf_event_task_sched_out(prev
, next
);
2570 fire_sched_out_preempt_notifiers(prev
, next
);
2571 prepare_lock_switch(rq
, next
);
2572 prepare_arch_switch(next
);
2576 * finish_task_switch - clean up after a task-switch
2577 * @prev: the thread we just switched away from.
2579 * finish_task_switch must be called after the context switch, paired
2580 * with a prepare_task_switch call before the context switch.
2581 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2582 * and do any other architecture-specific cleanup actions.
2584 * Note that we may have delayed dropping an mm in context_switch(). If
2585 * so, we finish that here outside of the runqueue lock. (Doing it
2586 * with the lock held can cause deadlocks; see schedule() for
2589 * The context switch have flipped the stack from under us and restored the
2590 * local variables which were saved when this task called schedule() in the
2591 * past. prev == current is still correct but we need to recalculate this_rq
2592 * because prev may have moved to another CPU.
2594 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2595 __releases(rq
->lock
)
2597 struct rq
*rq
= this_rq();
2598 struct mm_struct
*mm
= rq
->prev_mm
;
2602 * The previous task will have left us with a preempt_count of 2
2603 * because it left us after:
2606 * preempt_disable(); // 1
2608 * raw_spin_lock_irq(&rq->lock) // 2
2610 * Also, see FORK_PREEMPT_COUNT.
2612 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2613 "corrupted preempt_count: %s/%d/0x%x\n",
2614 current
->comm
, current
->pid
, preempt_count()))
2615 preempt_count_set(FORK_PREEMPT_COUNT
);
2620 * A task struct has one reference for the use as "current".
2621 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2622 * schedule one last time. The schedule call will never return, and
2623 * the scheduled task must drop that reference.
2625 * We must observe prev->state before clearing prev->on_cpu (in
2626 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2627 * running on another CPU and we could rave with its RUNNING -> DEAD
2628 * transition, resulting in a double drop.
2630 prev_state
= prev
->state
;
2631 vtime_task_switch(prev
);
2632 perf_event_task_sched_in(prev
, current
);
2633 finish_lock_switch(rq
, prev
);
2634 finish_arch_post_lock_switch();
2636 fire_sched_in_preempt_notifiers(current
);
2639 if (unlikely(prev_state
== TASK_DEAD
)) {
2640 if (prev
->sched_class
->task_dead
)
2641 prev
->sched_class
->task_dead(prev
);
2644 * Remove function-return probe instances associated with this
2645 * task and put them back on the free list.
2647 kprobe_flush_task(prev
);
2648 put_task_struct(prev
);
2651 tick_nohz_task_switch();
2657 /* rq->lock is NOT held, but preemption is disabled */
2658 static void __balance_callback(struct rq
*rq
)
2660 struct callback_head
*head
, *next
;
2661 void (*func
)(struct rq
*rq
);
2662 unsigned long flags
;
2664 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2665 head
= rq
->balance_callback
;
2666 rq
->balance_callback
= NULL
;
2668 func
= (void (*)(struct rq
*))head
->func
;
2675 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2678 static inline void balance_callback(struct rq
*rq
)
2680 if (unlikely(rq
->balance_callback
))
2681 __balance_callback(rq
);
2686 static inline void balance_callback(struct rq
*rq
)
2693 * schedule_tail - first thing a freshly forked thread must call.
2694 * @prev: the thread we just switched away from.
2696 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2697 __releases(rq
->lock
)
2702 * New tasks start with FORK_PREEMPT_COUNT, see there and
2703 * finish_task_switch() for details.
2705 * finish_task_switch() will drop rq->lock() and lower preempt_count
2706 * and the preempt_enable() will end up enabling preemption (on
2707 * PREEMPT_COUNT kernels).
2710 rq
= finish_task_switch(prev
);
2711 balance_callback(rq
);
2714 if (current
->set_child_tid
)
2715 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2719 * context_switch - switch to the new MM and the new thread's register state.
2721 static __always_inline
struct rq
*
2722 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2723 struct task_struct
*next
)
2725 struct mm_struct
*mm
, *oldmm
;
2727 prepare_task_switch(rq
, prev
, next
);
2730 oldmm
= prev
->active_mm
;
2732 * For paravirt, this is coupled with an exit in switch_to to
2733 * combine the page table reload and the switch backend into
2736 arch_start_context_switch(prev
);
2739 next
->active_mm
= oldmm
;
2740 atomic_inc(&oldmm
->mm_count
);
2741 enter_lazy_tlb(oldmm
, next
);
2743 switch_mm(oldmm
, mm
, next
);
2746 prev
->active_mm
= NULL
;
2747 rq
->prev_mm
= oldmm
;
2750 * Since the runqueue lock will be released by the next
2751 * task (which is an invalid locking op but in the case
2752 * of the scheduler it's an obvious special-case), so we
2753 * do an early lockdep release here:
2755 lockdep_unpin_lock(&rq
->lock
);
2756 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2758 /* Here we just switch the register state and the stack. */
2759 switch_to(prev
, next
, prev
);
2762 return finish_task_switch(prev
);
2766 * nr_running and nr_context_switches:
2768 * externally visible scheduler statistics: current number of runnable
2769 * threads, total number of context switches performed since bootup.
2771 unsigned long nr_running(void)
2773 unsigned long i
, sum
= 0;
2775 for_each_online_cpu(i
)
2776 sum
+= cpu_rq(i
)->nr_running
;
2782 * Check if only the current task is running on the cpu.
2784 * Caution: this function does not check that the caller has disabled
2785 * preemption, thus the result might have a time-of-check-to-time-of-use
2786 * race. The caller is responsible to use it correctly, for example:
2788 * - from a non-preemptable section (of course)
2790 * - from a thread that is bound to a single CPU
2792 * - in a loop with very short iterations (e.g. a polling loop)
2794 bool single_task_running(void)
2796 return raw_rq()->nr_running
== 1;
2798 EXPORT_SYMBOL(single_task_running
);
2800 unsigned long long nr_context_switches(void)
2803 unsigned long long sum
= 0;
2805 for_each_possible_cpu(i
)
2806 sum
+= cpu_rq(i
)->nr_switches
;
2811 unsigned long nr_iowait(void)
2813 unsigned long i
, sum
= 0;
2815 for_each_possible_cpu(i
)
2816 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2821 unsigned long nr_iowait_cpu(int cpu
)
2823 struct rq
*this = cpu_rq(cpu
);
2824 return atomic_read(&this->nr_iowait
);
2827 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2829 struct rq
*rq
= this_rq();
2830 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2831 *load
= rq
->load
.weight
;
2837 * sched_exec - execve() is a valuable balancing opportunity, because at
2838 * this point the task has the smallest effective memory and cache footprint.
2840 void sched_exec(void)
2842 struct task_struct
*p
= current
;
2843 unsigned long flags
;
2846 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2847 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2848 if (dest_cpu
== smp_processor_id())
2851 if (likely(cpu_active(dest_cpu
))) {
2852 struct migration_arg arg
= { p
, dest_cpu
};
2854 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2855 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2859 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2864 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2865 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2867 EXPORT_PER_CPU_SYMBOL(kstat
);
2868 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2871 * Return accounted runtime for the task.
2872 * In case the task is currently running, return the runtime plus current's
2873 * pending runtime that have not been accounted yet.
2875 unsigned long long task_sched_runtime(struct task_struct
*p
)
2877 unsigned long flags
;
2881 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2883 * 64-bit doesn't need locks to atomically read a 64bit value.
2884 * So we have a optimization chance when the task's delta_exec is 0.
2885 * Reading ->on_cpu is racy, but this is ok.
2887 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2888 * If we race with it entering cpu, unaccounted time is 0. This is
2889 * indistinguishable from the read occurring a few cycles earlier.
2890 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2891 * been accounted, so we're correct here as well.
2893 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2894 return p
->se
.sum_exec_runtime
;
2897 rq
= task_rq_lock(p
, &flags
);
2899 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2900 * project cycles that may never be accounted to this
2901 * thread, breaking clock_gettime().
2903 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2904 update_rq_clock(rq
);
2905 p
->sched_class
->update_curr(rq
);
2907 ns
= p
->se
.sum_exec_runtime
;
2908 task_rq_unlock(rq
, p
, &flags
);
2914 * This function gets called by the timer code, with HZ frequency.
2915 * We call it with interrupts disabled.
2917 void scheduler_tick(void)
2919 int cpu
= smp_processor_id();
2920 struct rq
*rq
= cpu_rq(cpu
);
2921 struct task_struct
*curr
= rq
->curr
;
2925 raw_spin_lock(&rq
->lock
);
2926 update_rq_clock(rq
);
2927 curr
->sched_class
->task_tick(rq
, curr
, 0);
2928 update_cpu_load_active(rq
);
2929 calc_global_load_tick(rq
);
2930 raw_spin_unlock(&rq
->lock
);
2932 perf_event_task_tick();
2935 rq
->idle_balance
= idle_cpu(cpu
);
2936 trigger_load_balance(rq
);
2938 rq_last_tick_reset(rq
);
2941 #ifdef CONFIG_NO_HZ_FULL
2943 * scheduler_tick_max_deferment
2945 * Keep at least one tick per second when a single
2946 * active task is running because the scheduler doesn't
2947 * yet completely support full dynticks environment.
2949 * This makes sure that uptime, CFS vruntime, load
2950 * balancing, etc... continue to move forward, even
2951 * with a very low granularity.
2953 * Return: Maximum deferment in nanoseconds.
2955 u64
scheduler_tick_max_deferment(void)
2957 struct rq
*rq
= this_rq();
2958 unsigned long next
, now
= READ_ONCE(jiffies
);
2960 next
= rq
->last_sched_tick
+ HZ
;
2962 if (time_before_eq(next
, now
))
2965 return jiffies_to_nsecs(next
- now
);
2969 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2970 defined(CONFIG_PREEMPT_TRACER))
2972 void preempt_count_add(int val
)
2974 #ifdef CONFIG_DEBUG_PREEMPT
2978 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2981 __preempt_count_add(val
);
2982 #ifdef CONFIG_DEBUG_PREEMPT
2984 * Spinlock count overflowing soon?
2986 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2989 if (preempt_count() == val
) {
2990 unsigned long ip
= get_lock_parent_ip();
2991 #ifdef CONFIG_DEBUG_PREEMPT
2992 current
->preempt_disable_ip
= ip
;
2994 trace_preempt_off(CALLER_ADDR0
, ip
);
2997 EXPORT_SYMBOL(preempt_count_add
);
2998 NOKPROBE_SYMBOL(preempt_count_add
);
3000 void preempt_count_sub(int val
)
3002 #ifdef CONFIG_DEBUG_PREEMPT
3006 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3009 * Is the spinlock portion underflowing?
3011 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3012 !(preempt_count() & PREEMPT_MASK
)))
3016 if (preempt_count() == val
)
3017 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3018 __preempt_count_sub(val
);
3020 EXPORT_SYMBOL(preempt_count_sub
);
3021 NOKPROBE_SYMBOL(preempt_count_sub
);
3026 * Print scheduling while atomic bug:
3028 static noinline
void __schedule_bug(struct task_struct
*prev
)
3030 if (oops_in_progress
)
3033 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3034 prev
->comm
, prev
->pid
, preempt_count());
3036 debug_show_held_locks(prev
);
3038 if (irqs_disabled())
3039 print_irqtrace_events(prev
);
3040 #ifdef CONFIG_DEBUG_PREEMPT
3041 if (in_atomic_preempt_off()) {
3042 pr_err("Preemption disabled at:");
3043 print_ip_sym(current
->preempt_disable_ip
);
3048 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3052 * Various schedule()-time debugging checks and statistics:
3054 static inline void schedule_debug(struct task_struct
*prev
)
3056 #ifdef CONFIG_SCHED_STACK_END_CHECK
3057 BUG_ON(task_stack_end_corrupted(prev
));
3060 if (unlikely(in_atomic_preempt_off())) {
3061 __schedule_bug(prev
);
3062 preempt_count_set(PREEMPT_DISABLED
);
3066 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3068 schedstat_inc(this_rq(), sched_count
);
3072 * Pick up the highest-prio task:
3074 static inline struct task_struct
*
3075 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3077 const struct sched_class
*class = &fair_sched_class
;
3078 struct task_struct
*p
;
3081 * Optimization: we know that if all tasks are in
3082 * the fair class we can call that function directly:
3084 if (likely(prev
->sched_class
== class &&
3085 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3086 p
= fair_sched_class
.pick_next_task(rq
, prev
);
3087 if (unlikely(p
== RETRY_TASK
))
3090 /* assumes fair_sched_class->next == idle_sched_class */
3092 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3098 for_each_class(class) {
3099 p
= class->pick_next_task(rq
, prev
);
3101 if (unlikely(p
== RETRY_TASK
))
3107 BUG(); /* the idle class will always have a runnable task */
3111 * __schedule() is the main scheduler function.
3113 * The main means of driving the scheduler and thus entering this function are:
3115 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3117 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3118 * paths. For example, see arch/x86/entry_64.S.
3120 * To drive preemption between tasks, the scheduler sets the flag in timer
3121 * interrupt handler scheduler_tick().
3123 * 3. Wakeups don't really cause entry into schedule(). They add a
3124 * task to the run-queue and that's it.
3126 * Now, if the new task added to the run-queue preempts the current
3127 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3128 * called on the nearest possible occasion:
3130 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3132 * - in syscall or exception context, at the next outmost
3133 * preempt_enable(). (this might be as soon as the wake_up()'s
3136 * - in IRQ context, return from interrupt-handler to
3137 * preemptible context
3139 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3142 * - cond_resched() call
3143 * - explicit schedule() call
3144 * - return from syscall or exception to user-space
3145 * - return from interrupt-handler to user-space
3147 * WARNING: must be called with preemption disabled!
3149 static void __sched notrace
__schedule(bool preempt
)
3151 struct task_struct
*prev
, *next
;
3152 unsigned long *switch_count
;
3156 cpu
= smp_processor_id();
3161 * do_exit() calls schedule() with preemption disabled as an exception;
3162 * however we must fix that up, otherwise the next task will see an
3163 * inconsistent (higher) preempt count.
3165 * It also avoids the below schedule_debug() test from complaining
3168 if (unlikely(prev
->state
== TASK_DEAD
))
3169 preempt_enable_no_resched_notrace();
3171 schedule_debug(prev
);
3173 if (sched_feat(HRTICK
))
3176 local_irq_disable();
3177 rcu_note_context_switch();
3180 * Make sure that signal_pending_state()->signal_pending() below
3181 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3182 * done by the caller to avoid the race with signal_wake_up().
3184 smp_mb__before_spinlock();
3185 raw_spin_lock(&rq
->lock
);
3186 lockdep_pin_lock(&rq
->lock
);
3188 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3190 switch_count
= &prev
->nivcsw
;
3191 if (!preempt
&& prev
->state
) {
3192 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3193 prev
->state
= TASK_RUNNING
;
3195 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3199 * If a worker went to sleep, notify and ask workqueue
3200 * whether it wants to wake up a task to maintain
3203 if (prev
->flags
& PF_WQ_WORKER
) {
3204 struct task_struct
*to_wakeup
;
3206 to_wakeup
= wq_worker_sleeping(prev
);
3208 try_to_wake_up_local(to_wakeup
);
3211 switch_count
= &prev
->nvcsw
;
3214 if (task_on_rq_queued(prev
))
3215 update_rq_clock(rq
);
3217 next
= pick_next_task(rq
, prev
);
3218 clear_tsk_need_resched(prev
);
3219 clear_preempt_need_resched();
3220 rq
->clock_skip_update
= 0;
3222 if (likely(prev
!= next
)) {
3227 trace_sched_switch(preempt
, prev
, next
);
3228 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3230 lockdep_unpin_lock(&rq
->lock
);
3231 raw_spin_unlock_irq(&rq
->lock
);
3234 balance_callback(rq
);
3236 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3238 static inline void sched_submit_work(struct task_struct
*tsk
)
3240 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3243 * If we are going to sleep and we have plugged IO queued,
3244 * make sure to submit it to avoid deadlocks.
3246 if (blk_needs_flush_plug(tsk
))
3247 blk_schedule_flush_plug(tsk
);
3250 asmlinkage __visible
void __sched
schedule(void)
3252 struct task_struct
*tsk
= current
;
3254 sched_submit_work(tsk
);
3258 sched_preempt_enable_no_resched();
3259 } while (need_resched());
3261 EXPORT_SYMBOL(schedule
);
3263 #ifdef CONFIG_CONTEXT_TRACKING
3264 asmlinkage __visible
void __sched
schedule_user(void)
3267 * If we come here after a random call to set_need_resched(),
3268 * or we have been woken up remotely but the IPI has not yet arrived,
3269 * we haven't yet exited the RCU idle mode. Do it here manually until
3270 * we find a better solution.
3272 * NB: There are buggy callers of this function. Ideally we
3273 * should warn if prev_state != CONTEXT_USER, but that will trigger
3274 * too frequently to make sense yet.
3276 enum ctx_state prev_state
= exception_enter();
3278 exception_exit(prev_state
);
3283 * schedule_preempt_disabled - called with preemption disabled
3285 * Returns with preemption disabled. Note: preempt_count must be 1
3287 void __sched
schedule_preempt_disabled(void)
3289 sched_preempt_enable_no_resched();
3294 static void __sched notrace
preempt_schedule_common(void)
3297 preempt_disable_notrace();
3299 preempt_enable_no_resched_notrace();
3302 * Check again in case we missed a preemption opportunity
3303 * between schedule and now.
3305 } while (need_resched());
3308 #ifdef CONFIG_PREEMPT
3310 * this is the entry point to schedule() from in-kernel preemption
3311 * off of preempt_enable. Kernel preemptions off return from interrupt
3312 * occur there and call schedule directly.
3314 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3317 * If there is a non-zero preempt_count or interrupts are disabled,
3318 * we do not want to preempt the current task. Just return..
3320 if (likely(!preemptible()))
3323 preempt_schedule_common();
3325 NOKPROBE_SYMBOL(preempt_schedule
);
3326 EXPORT_SYMBOL(preempt_schedule
);
3329 * preempt_schedule_notrace - preempt_schedule called by tracing
3331 * The tracing infrastructure uses preempt_enable_notrace to prevent
3332 * recursion and tracing preempt enabling caused by the tracing
3333 * infrastructure itself. But as tracing can happen in areas coming
3334 * from userspace or just about to enter userspace, a preempt enable
3335 * can occur before user_exit() is called. This will cause the scheduler
3336 * to be called when the system is still in usermode.
3338 * To prevent this, the preempt_enable_notrace will use this function
3339 * instead of preempt_schedule() to exit user context if needed before
3340 * calling the scheduler.
3342 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3344 enum ctx_state prev_ctx
;
3346 if (likely(!preemptible()))
3350 preempt_disable_notrace();
3352 * Needs preempt disabled in case user_exit() is traced
3353 * and the tracer calls preempt_enable_notrace() causing
3354 * an infinite recursion.
3356 prev_ctx
= exception_enter();
3358 exception_exit(prev_ctx
);
3360 preempt_enable_no_resched_notrace();
3361 } while (need_resched());
3363 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3365 #endif /* CONFIG_PREEMPT */
3368 * this is the entry point to schedule() from kernel preemption
3369 * off of irq context.
3370 * Note, that this is called and return with irqs disabled. This will
3371 * protect us against recursive calling from irq.
3373 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3375 enum ctx_state prev_state
;
3377 /* Catch callers which need to be fixed */
3378 BUG_ON(preempt_count() || !irqs_disabled());
3380 prev_state
= exception_enter();
3386 local_irq_disable();
3387 sched_preempt_enable_no_resched();
3388 } while (need_resched());
3390 exception_exit(prev_state
);
3393 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3396 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3398 EXPORT_SYMBOL(default_wake_function
);
3400 #ifdef CONFIG_RT_MUTEXES
3403 * rt_mutex_setprio - set the current priority of a task
3405 * @prio: prio value (kernel-internal form)
3407 * This function changes the 'effective' priority of a task. It does
3408 * not touch ->normal_prio like __setscheduler().
3410 * Used by the rt_mutex code to implement priority inheritance
3411 * logic. Call site only calls if the priority of the task changed.
3413 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3415 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3417 const struct sched_class
*prev_class
;
3419 BUG_ON(prio
> MAX_PRIO
);
3421 rq
= __task_rq_lock(p
);
3424 * Idle task boosting is a nono in general. There is one
3425 * exception, when PREEMPT_RT and NOHZ is active:
3427 * The idle task calls get_next_timer_interrupt() and holds
3428 * the timer wheel base->lock on the CPU and another CPU wants
3429 * to access the timer (probably to cancel it). We can safely
3430 * ignore the boosting request, as the idle CPU runs this code
3431 * with interrupts disabled and will complete the lock
3432 * protected section without being interrupted. So there is no
3433 * real need to boost.
3435 if (unlikely(p
== rq
->idle
)) {
3436 WARN_ON(p
!= rq
->curr
);
3437 WARN_ON(p
->pi_blocked_on
);
3441 trace_sched_pi_setprio(p
, prio
);
3444 if (oldprio
== prio
)
3445 queue_flag
&= ~DEQUEUE_MOVE
;
3447 prev_class
= p
->sched_class
;
3448 queued
= task_on_rq_queued(p
);
3449 running
= task_current(rq
, p
);
3451 dequeue_task(rq
, p
, queue_flag
);
3453 put_prev_task(rq
, p
);
3456 * Boosting condition are:
3457 * 1. -rt task is running and holds mutex A
3458 * --> -dl task blocks on mutex A
3460 * 2. -dl task is running and holds mutex A
3461 * --> -dl task blocks on mutex A and could preempt the
3464 if (dl_prio(prio
)) {
3465 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3466 if (!dl_prio(p
->normal_prio
) ||
3467 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3468 p
->dl
.dl_boosted
= 1;
3469 queue_flag
|= ENQUEUE_REPLENISH
;
3471 p
->dl
.dl_boosted
= 0;
3472 p
->sched_class
= &dl_sched_class
;
3473 } else if (rt_prio(prio
)) {
3474 if (dl_prio(oldprio
))
3475 p
->dl
.dl_boosted
= 0;
3477 queue_flag
|= ENQUEUE_HEAD
;
3478 p
->sched_class
= &rt_sched_class
;
3480 if (dl_prio(oldprio
))
3481 p
->dl
.dl_boosted
= 0;
3482 if (rt_prio(oldprio
))
3484 p
->sched_class
= &fair_sched_class
;
3487 sched_set_prio(p
, prio
);
3490 p
->sched_class
->set_curr_task(rq
);
3492 enqueue_task(rq
, p
, queue_flag
);
3494 check_class_changed(rq
, p
, prev_class
, oldprio
);
3496 preempt_disable(); /* avoid rq from going away on us */
3497 __task_rq_unlock(rq
);
3499 balance_callback(rq
);
3504 void set_user_nice(struct task_struct
*p
, long nice
)
3506 int old_prio
, delta
, queued
;
3507 unsigned long flags
;
3510 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3513 * We have to be careful, if called from sys_setpriority(),
3514 * the task might be in the middle of scheduling on another CPU.
3516 rq
= task_rq_lock(p
, &flags
);
3518 * The RT priorities are set via sched_setscheduler(), but we still
3519 * allow the 'normal' nice value to be set - but as expected
3520 * it wont have any effect on scheduling until the task is
3521 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3523 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3524 p
->static_prio
= NICE_TO_PRIO(nice
);
3527 queued
= task_on_rq_queued(p
);
3529 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3531 p
->static_prio
= NICE_TO_PRIO(nice
);
3534 sched_set_prio(p
, effective_prio(p
));
3535 delta
= p
->prio
- old_prio
;
3538 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3540 * If the task increased its priority or is running and
3541 * lowered its priority, then reschedule its CPU:
3543 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3547 task_rq_unlock(rq
, p
, &flags
);
3549 EXPORT_SYMBOL(set_user_nice
);
3552 * can_nice - check if a task can reduce its nice value
3556 int can_nice(const struct task_struct
*p
, const int nice
)
3558 /* convert nice value [19,-20] to rlimit style value [1,40] */
3559 int nice_rlim
= nice_to_rlimit(nice
);
3561 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3562 capable(CAP_SYS_NICE
));
3565 #ifdef __ARCH_WANT_SYS_NICE
3568 * sys_nice - change the priority of the current process.
3569 * @increment: priority increment
3571 * sys_setpriority is a more generic, but much slower function that
3572 * does similar things.
3574 SYSCALL_DEFINE1(nice
, int, increment
)
3579 * Setpriority might change our priority at the same moment.
3580 * We don't have to worry. Conceptually one call occurs first
3581 * and we have a single winner.
3583 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3584 nice
= task_nice(current
) + increment
;
3586 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3587 if (increment
< 0 && !can_nice(current
, nice
))
3590 retval
= security_task_setnice(current
, nice
);
3594 set_user_nice(current
, nice
);
3601 * task_prio - return the priority value of a given task.
3602 * @p: the task in question.
3604 * Return: The priority value as seen by users in /proc.
3605 * RT tasks are offset by -200. Normal tasks are centered
3606 * around 0, value goes from -16 to +15.
3608 int task_prio(const struct task_struct
*p
)
3610 return p
->prio
- MAX_RT_PRIO
;
3614 * idle_cpu - is a given cpu idle currently?
3615 * @cpu: the processor in question.
3617 * Return: 1 if the CPU is currently idle. 0 otherwise.
3619 int idle_cpu(int cpu
)
3621 struct rq
*rq
= cpu_rq(cpu
);
3623 if (rq
->curr
!= rq
->idle
)
3630 if (!llist_empty(&rq
->wake_list
))
3638 * idle_task - return the idle task for a given cpu.
3639 * @cpu: the processor in question.
3641 * Return: The idle task for the cpu @cpu.
3643 struct task_struct
*idle_task(int cpu
)
3645 return cpu_rq(cpu
)->idle
;
3649 * find_process_by_pid - find a process with a matching PID value.
3650 * @pid: the pid in question.
3652 * The task of @pid, if found. %NULL otherwise.
3654 static struct task_struct
*find_process_by_pid(pid_t pid
)
3656 return pid
? find_task_by_vpid(pid
) : current
;
3660 * This function initializes the sched_dl_entity of a newly becoming
3661 * SCHED_DEADLINE task.
3663 * Only the static values are considered here, the actual runtime and the
3664 * absolute deadline will be properly calculated when the task is enqueued
3665 * for the first time with its new policy.
3668 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3670 struct sched_dl_entity
*dl_se
= &p
->dl
;
3672 dl_se
->dl_runtime
= attr
->sched_runtime
;
3673 dl_se
->dl_deadline
= attr
->sched_deadline
;
3674 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3675 dl_se
->flags
= attr
->sched_flags
;
3676 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3679 * Changing the parameters of a task is 'tricky' and we're not doing
3680 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3682 * What we SHOULD do is delay the bandwidth release until the 0-lag
3683 * point. This would include retaining the task_struct until that time
3684 * and change dl_overflow() to not immediately decrement the current
3687 * Instead we retain the current runtime/deadline and let the new
3688 * parameters take effect after the current reservation period lapses.
3689 * This is safe (albeit pessimistic) because the 0-lag point is always
3690 * before the current scheduling deadline.
3692 * We can still have temporary overloads because we do not delay the
3693 * change in bandwidth until that time; so admission control is
3694 * not on the safe side. It does however guarantee tasks will never
3695 * consume more than promised.
3700 * sched_setparam() passes in -1 for its policy, to let the functions
3701 * it calls know not to change it.
3703 #define SETPARAM_POLICY -1
3705 static void __setscheduler_params(struct task_struct
*p
,
3706 const struct sched_attr
*attr
)
3708 int policy
= attr
->sched_policy
;
3710 if (policy
== SETPARAM_POLICY
)
3715 if (dl_policy(policy
))
3716 __setparam_dl(p
, attr
);
3717 else if (fair_policy(policy
))
3718 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3721 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3722 * !rt_policy. Always setting this ensures that things like
3723 * getparam()/getattr() don't report silly values for !rt tasks.
3725 p
->rt_priority
= attr
->sched_priority
;
3726 p
->normal_prio
= normal_prio(p
);
3730 /* Actually do priority change: must hold pi & rq lock. */
3731 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3732 const struct sched_attr
*attr
, bool keep_boost
)
3734 __setscheduler_params(p
, attr
);
3737 * Keep a potential priority boosting if called from
3738 * sched_setscheduler().
3741 sched_set_prio(p
, rt_mutex_get_effective_prio(p
,
3744 sched_set_prio(p
, normal_prio(p
));
3746 if (dl_prio(p
->prio
))
3747 p
->sched_class
= &dl_sched_class
;
3748 else if (rt_prio(p
->prio
))
3749 p
->sched_class
= &rt_sched_class
;
3751 p
->sched_class
= &fair_sched_class
;
3755 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3757 struct sched_dl_entity
*dl_se
= &p
->dl
;
3759 attr
->sched_priority
= p
->rt_priority
;
3760 attr
->sched_runtime
= dl_se
->dl_runtime
;
3761 attr
->sched_deadline
= dl_se
->dl_deadline
;
3762 attr
->sched_period
= dl_se
->dl_period
;
3763 attr
->sched_flags
= dl_se
->flags
;
3767 * This function validates the new parameters of a -deadline task.
3768 * We ask for the deadline not being zero, and greater or equal
3769 * than the runtime, as well as the period of being zero or
3770 * greater than deadline. Furthermore, we have to be sure that
3771 * user parameters are above the internal resolution of 1us (we
3772 * check sched_runtime only since it is always the smaller one) and
3773 * below 2^63 ns (we have to check both sched_deadline and
3774 * sched_period, as the latter can be zero).
3777 __checkparam_dl(const struct sched_attr
*attr
)
3780 if (attr
->sched_deadline
== 0)
3784 * Since we truncate DL_SCALE bits, make sure we're at least
3787 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3791 * Since we use the MSB for wrap-around and sign issues, make
3792 * sure it's not set (mind that period can be equal to zero).
3794 if (attr
->sched_deadline
& (1ULL << 63) ||
3795 attr
->sched_period
& (1ULL << 63))
3798 /* runtime <= deadline <= period (if period != 0) */
3799 if ((attr
->sched_period
!= 0 &&
3800 attr
->sched_period
< attr
->sched_deadline
) ||
3801 attr
->sched_deadline
< attr
->sched_runtime
)
3808 * check the target process has a UID that matches the current process's
3810 static bool check_same_owner(struct task_struct
*p
)
3812 const struct cred
*cred
= current_cred(), *pcred
;
3816 pcred
= __task_cred(p
);
3817 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3818 uid_eq(cred
->euid
, pcred
->uid
));
3823 static bool dl_param_changed(struct task_struct
*p
,
3824 const struct sched_attr
*attr
)
3826 struct sched_dl_entity
*dl_se
= &p
->dl
;
3828 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3829 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3830 dl_se
->dl_period
!= attr
->sched_period
||
3831 dl_se
->flags
!= attr
->sched_flags
)
3837 static int __sched_setscheduler(struct task_struct
*p
,
3838 const struct sched_attr
*attr
,
3841 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3842 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3843 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3844 int new_effective_prio
, policy
= attr
->sched_policy
;
3845 unsigned long flags
;
3846 const struct sched_class
*prev_class
;
3849 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3851 /* may grab non-irq protected spin_locks */
3852 BUG_ON(in_interrupt());
3854 /* double check policy once rq lock held */
3856 reset_on_fork
= p
->sched_reset_on_fork
;
3857 policy
= oldpolicy
= p
->policy
;
3859 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3861 if (!valid_policy(policy
))
3865 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3869 * Valid priorities for SCHED_FIFO and SCHED_RR are
3870 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3871 * SCHED_BATCH and SCHED_IDLE is 0.
3873 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3874 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3876 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3877 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3881 * Allow unprivileged RT tasks to decrease priority:
3883 if (user
&& !capable(CAP_SYS_NICE
)) {
3884 if (fair_policy(policy
)) {
3885 if (attr
->sched_nice
< task_nice(p
) &&
3886 !can_nice(p
, attr
->sched_nice
))
3890 if (rt_policy(policy
)) {
3891 unsigned long rlim_rtprio
=
3892 task_rlimit(p
, RLIMIT_RTPRIO
);
3894 /* can't set/change the rt policy */
3895 if (policy
!= p
->policy
&& !rlim_rtprio
)
3898 /* can't increase priority */
3899 if (attr
->sched_priority
> p
->rt_priority
&&
3900 attr
->sched_priority
> rlim_rtprio
)
3905 * Can't set/change SCHED_DEADLINE policy at all for now
3906 * (safest behavior); in the future we would like to allow
3907 * unprivileged DL tasks to increase their relative deadline
3908 * or reduce their runtime (both ways reducing utilization)
3910 if (dl_policy(policy
))
3914 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3915 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3917 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
3918 if (!can_nice(p
, task_nice(p
)))
3922 /* can't change other user's priorities */
3923 if (!check_same_owner(p
))
3926 /* Normal users shall not reset the sched_reset_on_fork flag */
3927 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3932 retval
= security_task_setscheduler(p
);
3938 * make sure no PI-waiters arrive (or leave) while we are
3939 * changing the priority of the task:
3941 * To be able to change p->policy safely, the appropriate
3942 * runqueue lock must be held.
3944 rq
= task_rq_lock(p
, &flags
);
3947 * Changing the policy of the stop threads its a very bad idea
3949 if (p
== rq
->stop
) {
3950 task_rq_unlock(rq
, p
, &flags
);
3955 * If not changing anything there's no need to proceed further,
3956 * but store a possible modification of reset_on_fork.
3958 if (unlikely(policy
== p
->policy
)) {
3959 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3961 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3963 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3966 p
->sched_reset_on_fork
= reset_on_fork
;
3967 task_rq_unlock(rq
, p
, &flags
);
3973 #ifdef CONFIG_RT_GROUP_SCHED
3975 * Do not allow realtime tasks into groups that have no runtime
3978 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3979 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3980 !task_group_is_autogroup(task_group(p
))) {
3981 task_rq_unlock(rq
, p
, &flags
);
3986 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3987 cpumask_t
*span
= rq
->rd
->span
;
3990 * Don't allow tasks with an affinity mask smaller than
3991 * the entire root_domain to become SCHED_DEADLINE. We
3992 * will also fail if there's no bandwidth available.
3994 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3995 rq
->rd
->dl_bw
.bw
== 0) {
3996 task_rq_unlock(rq
, p
, &flags
);
4003 /* recheck policy now with rq lock held */
4004 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4005 policy
= oldpolicy
= -1;
4006 task_rq_unlock(rq
, p
, &flags
);
4011 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4012 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4015 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4016 task_rq_unlock(rq
, p
, &flags
);
4020 p
->sched_reset_on_fork
= reset_on_fork
;
4025 * Take priority boosted tasks into account. If the new
4026 * effective priority is unchanged, we just store the new
4027 * normal parameters and do not touch the scheduler class and
4028 * the runqueue. This will be done when the task deboost
4031 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4032 if (new_effective_prio
== oldprio
)
4033 queue_flags
&= ~DEQUEUE_MOVE
;
4036 queued
= task_on_rq_queued(p
);
4037 running
= task_current(rq
, p
);
4039 dequeue_task(rq
, p
, queue_flags
);
4041 put_prev_task(rq
, p
);
4043 prev_class
= p
->sched_class
;
4044 __setscheduler(rq
, p
, attr
, pi
);
4047 p
->sched_class
->set_curr_task(rq
);
4050 * We enqueue to tail when the priority of a task is
4051 * increased (user space view).
4053 if (oldprio
< p
->prio
)
4054 queue_flags
|= ENQUEUE_HEAD
;
4056 enqueue_task(rq
, p
, queue_flags
);
4059 check_class_changed(rq
, p
, prev_class
, oldprio
);
4060 preempt_disable(); /* avoid rq from going away on us */
4061 task_rq_unlock(rq
, p
, &flags
);
4064 rt_mutex_adjust_pi(p
);
4067 * Run balance callbacks after we've adjusted the PI chain.
4069 balance_callback(rq
);
4075 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4076 const struct sched_param
*param
, bool check
)
4078 struct sched_attr attr
= {
4079 .sched_policy
= policy
,
4080 .sched_priority
= param
->sched_priority
,
4081 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4084 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4085 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4086 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4087 policy
&= ~SCHED_RESET_ON_FORK
;
4088 attr
.sched_policy
= policy
;
4091 return __sched_setscheduler(p
, &attr
, check
, true);
4094 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4095 * @p: the task in question.
4096 * @policy: new policy.
4097 * @param: structure containing the new RT priority.
4099 * Return: 0 on success. An error code otherwise.
4101 * NOTE that the task may be already dead.
4103 int sched_setscheduler(struct task_struct
*p
, int policy
,
4104 const struct sched_param
*param
)
4106 return _sched_setscheduler(p
, policy
, param
, true);
4108 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4110 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4112 return __sched_setscheduler(p
, attr
, true, true);
4114 EXPORT_SYMBOL_GPL(sched_setattr
);
4117 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4118 * @p: the task in question.
4119 * @policy: new policy.
4120 * @param: structure containing the new RT priority.
4122 * Just like sched_setscheduler, only don't bother checking if the
4123 * current context has permission. For example, this is needed in
4124 * stop_machine(): we create temporary high priority worker threads,
4125 * but our caller might not have that capability.
4127 * Return: 0 on success. An error code otherwise.
4129 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4130 const struct sched_param
*param
)
4132 return _sched_setscheduler(p
, policy
, param
, false);
4134 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4137 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4139 struct sched_param lparam
;
4140 struct task_struct
*p
;
4143 if (!param
|| pid
< 0)
4145 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4150 p
= find_process_by_pid(pid
);
4152 retval
= sched_setscheduler(p
, policy
, &lparam
);
4159 * Mimics kernel/events/core.c perf_copy_attr().
4161 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4162 struct sched_attr
*attr
)
4167 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4171 * zero the full structure, so that a short copy will be nice.
4173 memset(attr
, 0, sizeof(*attr
));
4175 ret
= get_user(size
, &uattr
->size
);
4179 if (size
> PAGE_SIZE
) /* silly large */
4182 if (!size
) /* abi compat */
4183 size
= SCHED_ATTR_SIZE_VER0
;
4185 if (size
< SCHED_ATTR_SIZE_VER0
)
4189 * If we're handed a bigger struct than we know of,
4190 * ensure all the unknown bits are 0 - i.e. new
4191 * user-space does not rely on any kernel feature
4192 * extensions we dont know about yet.
4194 if (size
> sizeof(*attr
)) {
4195 unsigned char __user
*addr
;
4196 unsigned char __user
*end
;
4199 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4200 end
= (void __user
*)uattr
+ size
;
4202 for (; addr
< end
; addr
++) {
4203 ret
= get_user(val
, addr
);
4209 size
= sizeof(*attr
);
4212 ret
= copy_from_user(attr
, uattr
, size
);
4217 * XXX: do we want to be lenient like existing syscalls; or do we want
4218 * to be strict and return an error on out-of-bounds values?
4220 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4225 put_user(sizeof(*attr
), &uattr
->size
);
4230 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4231 * @pid: the pid in question.
4232 * @policy: new policy.
4233 * @param: structure containing the new RT priority.
4235 * Return: 0 on success. An error code otherwise.
4237 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4238 struct sched_param __user
*, param
)
4240 /* negative values for policy are not valid */
4244 return do_sched_setscheduler(pid
, policy
, param
);
4248 * sys_sched_setparam - set/change the RT priority of a thread
4249 * @pid: the pid in question.
4250 * @param: structure containing the new RT priority.
4252 * Return: 0 on success. An error code otherwise.
4254 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4256 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4260 * sys_sched_setattr - same as above, but with extended sched_attr
4261 * @pid: the pid in question.
4262 * @uattr: structure containing the extended parameters.
4263 * @flags: for future extension.
4265 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4266 unsigned int, flags
)
4268 struct sched_attr attr
;
4269 struct task_struct
*p
;
4272 if (!uattr
|| pid
< 0 || flags
)
4275 retval
= sched_copy_attr(uattr
, &attr
);
4279 if ((int)attr
.sched_policy
< 0)
4284 p
= find_process_by_pid(pid
);
4286 retval
= sched_setattr(p
, &attr
);
4293 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4294 * @pid: the pid in question.
4296 * Return: On success, the policy of the thread. Otherwise, a negative error
4299 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4301 struct task_struct
*p
;
4309 p
= find_process_by_pid(pid
);
4311 retval
= security_task_getscheduler(p
);
4314 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4321 * sys_sched_getparam - get the RT priority of a thread
4322 * @pid: the pid in question.
4323 * @param: structure containing the RT priority.
4325 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4328 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4330 struct sched_param lp
= { .sched_priority
= 0 };
4331 struct task_struct
*p
;
4334 if (!param
|| pid
< 0)
4338 p
= find_process_by_pid(pid
);
4343 retval
= security_task_getscheduler(p
);
4347 if (task_has_rt_policy(p
))
4348 lp
.sched_priority
= p
->rt_priority
;
4352 * This one might sleep, we cannot do it with a spinlock held ...
4354 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4363 static int sched_read_attr(struct sched_attr __user
*uattr
,
4364 struct sched_attr
*attr
,
4369 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4373 * If we're handed a smaller struct than we know of,
4374 * ensure all the unknown bits are 0 - i.e. old
4375 * user-space does not get uncomplete information.
4377 if (usize
< sizeof(*attr
)) {
4378 unsigned char *addr
;
4381 addr
= (void *)attr
+ usize
;
4382 end
= (void *)attr
+ sizeof(*attr
);
4384 for (; addr
< end
; addr
++) {
4392 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4400 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4401 * @pid: the pid in question.
4402 * @uattr: structure containing the extended parameters.
4403 * @size: sizeof(attr) for fwd/bwd comp.
4404 * @flags: for future extension.
4406 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4407 unsigned int, size
, unsigned int, flags
)
4409 struct sched_attr attr
= {
4410 .size
= sizeof(struct sched_attr
),
4412 struct task_struct
*p
;
4415 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4416 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4420 p
= find_process_by_pid(pid
);
4425 retval
= security_task_getscheduler(p
);
4429 attr
.sched_policy
= p
->policy
;
4430 if (p
->sched_reset_on_fork
)
4431 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4432 if (task_has_dl_policy(p
))
4433 __getparam_dl(p
, &attr
);
4434 else if (task_has_rt_policy(p
))
4435 attr
.sched_priority
= p
->rt_priority
;
4437 attr
.sched_nice
= task_nice(p
);
4441 retval
= sched_read_attr(uattr
, &attr
, size
);
4449 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4451 cpumask_var_t cpus_allowed
, new_mask
;
4452 struct task_struct
*p
;
4457 p
= find_process_by_pid(pid
);
4463 /* Prevent p going away */
4467 if (p
->flags
& PF_NO_SETAFFINITY
) {
4471 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4475 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4477 goto out_free_cpus_allowed
;
4480 if (!check_same_owner(p
)) {
4482 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4484 goto out_free_new_mask
;
4489 retval
= security_task_setscheduler(p
);
4491 goto out_free_new_mask
;
4494 cpuset_cpus_allowed(p
, cpus_allowed
);
4495 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4498 * Since bandwidth control happens on root_domain basis,
4499 * if admission test is enabled, we only admit -deadline
4500 * tasks allowed to run on all the CPUs in the task's
4504 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4506 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4509 goto out_free_new_mask
;
4515 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4518 cpuset_cpus_allowed(p
, cpus_allowed
);
4519 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4521 * We must have raced with a concurrent cpuset
4522 * update. Just reset the cpus_allowed to the
4523 * cpuset's cpus_allowed
4525 cpumask_copy(new_mask
, cpus_allowed
);
4530 free_cpumask_var(new_mask
);
4531 out_free_cpus_allowed
:
4532 free_cpumask_var(cpus_allowed
);
4538 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4539 struct cpumask
*new_mask
)
4541 if (len
< cpumask_size())
4542 cpumask_clear(new_mask
);
4543 else if (len
> cpumask_size())
4544 len
= cpumask_size();
4546 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4550 * sys_sched_setaffinity - set the cpu affinity of a process
4551 * @pid: pid of the process
4552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4553 * @user_mask_ptr: user-space pointer to the new cpu mask
4555 * Return: 0 on success. An error code otherwise.
4557 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4558 unsigned long __user
*, user_mask_ptr
)
4560 cpumask_var_t new_mask
;
4563 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4566 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4568 retval
= sched_setaffinity(pid
, new_mask
);
4569 free_cpumask_var(new_mask
);
4573 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4575 struct task_struct
*p
;
4576 unsigned long flags
;
4582 p
= find_process_by_pid(pid
);
4586 retval
= security_task_getscheduler(p
);
4590 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4591 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4592 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4601 * sys_sched_getaffinity - get the cpu affinity of a process
4602 * @pid: pid of the process
4603 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4604 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4606 * Return: 0 on success. An error code otherwise.
4608 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4609 unsigned long __user
*, user_mask_ptr
)
4614 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4616 if (len
& (sizeof(unsigned long)-1))
4619 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4622 ret
= sched_getaffinity(pid
, mask
);
4624 size_t retlen
= min_t(size_t, len
, cpumask_size());
4626 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4631 free_cpumask_var(mask
);
4637 * sys_sched_yield - yield the current processor to other threads.
4639 * This function yields the current CPU to other tasks. If there are no
4640 * other threads running on this CPU then this function will return.
4644 SYSCALL_DEFINE0(sched_yield
)
4646 struct rq
*rq
= this_rq_lock();
4648 schedstat_inc(rq
, yld_count
);
4649 current
->sched_class
->yield_task(rq
);
4652 * Since we are going to call schedule() anyway, there's
4653 * no need to preempt or enable interrupts:
4655 __release(rq
->lock
);
4656 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4657 do_raw_spin_unlock(&rq
->lock
);
4658 sched_preempt_enable_no_resched();
4665 int __sched
_cond_resched(void)
4667 if (should_resched(0)) {
4668 preempt_schedule_common();
4673 EXPORT_SYMBOL(_cond_resched
);
4676 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4677 * call schedule, and on return reacquire the lock.
4679 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4680 * operations here to prevent schedule() from being called twice (once via
4681 * spin_unlock(), once by hand).
4683 int __cond_resched_lock(spinlock_t
*lock
)
4685 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4688 lockdep_assert_held(lock
);
4690 if (spin_needbreak(lock
) || resched
) {
4693 preempt_schedule_common();
4701 EXPORT_SYMBOL(__cond_resched_lock
);
4703 int __sched
__cond_resched_softirq(void)
4705 BUG_ON(!in_softirq());
4707 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4709 preempt_schedule_common();
4715 EXPORT_SYMBOL(__cond_resched_softirq
);
4718 * yield - yield the current processor to other threads.
4720 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4722 * The scheduler is at all times free to pick the calling task as the most
4723 * eligible task to run, if removing the yield() call from your code breaks
4724 * it, its already broken.
4726 * Typical broken usage is:
4731 * where one assumes that yield() will let 'the other' process run that will
4732 * make event true. If the current task is a SCHED_FIFO task that will never
4733 * happen. Never use yield() as a progress guarantee!!
4735 * If you want to use yield() to wait for something, use wait_event().
4736 * If you want to use yield() to be 'nice' for others, use cond_resched().
4737 * If you still want to use yield(), do not!
4739 void __sched
yield(void)
4741 set_current_state(TASK_RUNNING
);
4744 EXPORT_SYMBOL(yield
);
4747 * yield_to - yield the current processor to another thread in
4748 * your thread group, or accelerate that thread toward the
4749 * processor it's on.
4751 * @preempt: whether task preemption is allowed or not
4753 * It's the caller's job to ensure that the target task struct
4754 * can't go away on us before we can do any checks.
4757 * true (>0) if we indeed boosted the target task.
4758 * false (0) if we failed to boost the target.
4759 * -ESRCH if there's no task to yield to.
4761 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4763 struct task_struct
*curr
= current
;
4764 struct rq
*rq
, *p_rq
;
4765 unsigned long flags
;
4768 local_irq_save(flags
);
4774 * If we're the only runnable task on the rq and target rq also
4775 * has only one task, there's absolutely no point in yielding.
4777 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4782 double_rq_lock(rq
, p_rq
);
4783 if (task_rq(p
) != p_rq
) {
4784 double_rq_unlock(rq
, p_rq
);
4788 if (!curr
->sched_class
->yield_to_task
)
4791 if (curr
->sched_class
!= p
->sched_class
)
4794 if (task_running(p_rq
, p
) || p
->state
)
4797 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4799 schedstat_inc(rq
, yld_count
);
4801 * Make p's CPU reschedule; pick_next_entity takes care of
4804 if (preempt
&& rq
!= p_rq
)
4809 double_rq_unlock(rq
, p_rq
);
4811 local_irq_restore(flags
);
4818 EXPORT_SYMBOL_GPL(yield_to
);
4821 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4822 * that process accounting knows that this is a task in IO wait state.
4824 long __sched
io_schedule_timeout(long timeout
)
4826 int old_iowait
= current
->in_iowait
;
4830 current
->in_iowait
= 1;
4831 blk_schedule_flush_plug(current
);
4833 delayacct_blkio_start();
4835 atomic_inc(&rq
->nr_iowait
);
4836 ret
= schedule_timeout(timeout
);
4837 current
->in_iowait
= old_iowait
;
4838 atomic_dec(&rq
->nr_iowait
);
4839 delayacct_blkio_end();
4843 EXPORT_SYMBOL(io_schedule_timeout
);
4846 * sys_sched_get_priority_max - return maximum RT priority.
4847 * @policy: scheduling class.
4849 * Return: On success, this syscall returns the maximum
4850 * rt_priority that can be used by a given scheduling class.
4851 * On failure, a negative error code is returned.
4853 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4860 ret
= MAX_USER_RT_PRIO
-1;
4862 case SCHED_DEADLINE
:
4873 * sys_sched_get_priority_min - return minimum RT priority.
4874 * @policy: scheduling class.
4876 * Return: On success, this syscall returns the minimum
4877 * rt_priority that can be used by a given scheduling class.
4878 * On failure, a negative error code is returned.
4880 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4889 case SCHED_DEADLINE
:
4899 * sys_sched_rr_get_interval - return the default timeslice of a process.
4900 * @pid: pid of the process.
4901 * @interval: userspace pointer to the timeslice value.
4903 * this syscall writes the default timeslice value of a given process
4904 * into the user-space timespec buffer. A value of '0' means infinity.
4906 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4909 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4910 struct timespec __user
*, interval
)
4912 struct task_struct
*p
;
4913 unsigned int time_slice
;
4914 unsigned long flags
;
4924 p
= find_process_by_pid(pid
);
4928 retval
= security_task_getscheduler(p
);
4932 rq
= task_rq_lock(p
, &flags
);
4934 if (p
->sched_class
->get_rr_interval
)
4935 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4936 task_rq_unlock(rq
, p
, &flags
);
4939 jiffies_to_timespec(time_slice
, &t
);
4940 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4948 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4950 void sched_show_task(struct task_struct
*p
)
4952 unsigned long free
= 0;
4954 unsigned long state
= p
->state
;
4957 state
= __ffs(state
) + 1;
4958 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4959 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4960 #if BITS_PER_LONG == 32
4961 if (state
== TASK_RUNNING
)
4962 printk(KERN_CONT
" running ");
4964 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4966 if (state
== TASK_RUNNING
)
4967 printk(KERN_CONT
" running task ");
4969 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4971 #ifdef CONFIG_DEBUG_STACK_USAGE
4972 free
= stack_not_used(p
);
4977 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4979 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4980 task_pid_nr(p
), ppid
,
4981 (unsigned long)task_thread_info(p
)->flags
);
4983 print_worker_info(KERN_INFO
, p
);
4984 show_stack(p
, NULL
);
4987 void show_state_filter(unsigned long state_filter
)
4989 struct task_struct
*g
, *p
;
4991 #if BITS_PER_LONG == 32
4993 " task PC stack pid father\n");
4996 " task PC stack pid father\n");
4999 for_each_process_thread(g
, p
) {
5001 * reset the NMI-timeout, listing all files on a slow
5002 * console might take a lot of time:
5004 touch_nmi_watchdog();
5005 if (!state_filter
|| (p
->state
& state_filter
))
5009 touch_all_softlockup_watchdogs();
5011 #ifdef CONFIG_SCHED_DEBUG
5012 sysrq_sched_debug_show();
5016 * Only show locks if all tasks are dumped:
5019 debug_show_all_locks();
5022 void init_idle_bootup_task(struct task_struct
*idle
)
5024 idle
->sched_class
= &idle_sched_class
;
5028 * init_idle - set up an idle thread for a given CPU
5029 * @idle: task in question
5030 * @cpu: cpu the idle task belongs to
5032 * NOTE: this function does not set the idle thread's NEED_RESCHED
5033 * flag, to make booting more robust.
5035 void init_idle(struct task_struct
*idle
, int cpu
)
5037 struct rq
*rq
= cpu_rq(cpu
);
5038 unsigned long flags
;
5040 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5041 raw_spin_lock(&rq
->lock
);
5043 __sched_fork(0, idle
);
5044 idle
->state
= TASK_RUNNING
;
5045 idle
->se
.exec_start
= sched_clock();
5047 kasan_unpoison_task_stack(idle
);
5051 * Its possible that init_idle() gets called multiple times on a task,
5052 * in that case do_set_cpus_allowed() will not do the right thing.
5054 * And since this is boot we can forgo the serialization.
5056 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5059 * We're having a chicken and egg problem, even though we are
5060 * holding rq->lock, the cpu isn't yet set to this cpu so the
5061 * lockdep check in task_group() will fail.
5063 * Similar case to sched_fork(). / Alternatively we could
5064 * use task_rq_lock() here and obtain the other rq->lock.
5069 __set_task_cpu(idle
, cpu
);
5072 rq
->curr
= rq
->idle
= idle
;
5073 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5077 raw_spin_unlock(&rq
->lock
);
5078 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5080 /* Set the preempt count _outside_ the spinlocks! */
5081 init_idle_preempt_count(idle
, cpu
);
5084 * The idle tasks have their own, simple scheduling class:
5086 idle
->sched_class
= &idle_sched_class
;
5087 ftrace_graph_init_idle_task(idle
, cpu
);
5088 vtime_init_idle(idle
, cpu
);
5090 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5094 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5095 const struct cpumask
*trial
)
5097 int ret
= 1, trial_cpus
;
5098 struct dl_bw
*cur_dl_b
;
5099 unsigned long flags
;
5101 if (!cpumask_weight(cur
))
5104 rcu_read_lock_sched();
5105 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5106 trial_cpus
= cpumask_weight(trial
);
5108 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5109 if (cur_dl_b
->bw
!= -1 &&
5110 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5112 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5113 rcu_read_unlock_sched();
5118 int task_can_attach(struct task_struct
*p
,
5119 const struct cpumask
*cs_cpus_allowed
)
5124 * Kthreads which disallow setaffinity shouldn't be moved
5125 * to a new cpuset; we don't want to change their cpu
5126 * affinity and isolating such threads by their set of
5127 * allowed nodes is unnecessary. Thus, cpusets are not
5128 * applicable for such threads. This prevents checking for
5129 * success of set_cpus_allowed_ptr() on all attached tasks
5130 * before cpus_allowed may be changed.
5132 if (p
->flags
& PF_NO_SETAFFINITY
) {
5138 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5140 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5145 unsigned long flags
;
5147 rcu_read_lock_sched();
5148 dl_b
= dl_bw_of(dest_cpu
);
5149 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5150 cpus
= dl_bw_cpus(dest_cpu
);
5151 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5156 * We reserve space for this task in the destination
5157 * root_domain, as we can't fail after this point.
5158 * We will free resources in the source root_domain
5159 * later on (see set_cpus_allowed_dl()).
5161 __dl_add(dl_b
, p
->dl
.dl_bw
);
5163 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5164 rcu_read_unlock_sched();
5174 #ifdef CONFIG_NUMA_BALANCING
5175 /* Migrate current task p to target_cpu */
5176 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5178 struct migration_arg arg
= { p
, target_cpu
};
5179 int curr_cpu
= task_cpu(p
);
5181 if (curr_cpu
== target_cpu
)
5184 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5187 /* TODO: This is not properly updating schedstats */
5189 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5190 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5194 * Requeue a task on a given node and accurately track the number of NUMA
5195 * tasks on the runqueues
5197 void sched_setnuma(struct task_struct
*p
, int nid
)
5200 unsigned long flags
;
5201 bool queued
, running
;
5203 rq
= task_rq_lock(p
, &flags
);
5204 queued
= task_on_rq_queued(p
);
5205 running
= task_current(rq
, p
);
5208 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5210 put_prev_task(rq
, p
);
5212 p
->numa_preferred_nid
= nid
;
5215 p
->sched_class
->set_curr_task(rq
);
5217 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5218 task_rq_unlock(rq
, p
, &flags
);
5220 #endif /* CONFIG_NUMA_BALANCING */
5222 #ifdef CONFIG_HOTPLUG_CPU
5224 * Ensures that the idle task is using init_mm right before its cpu goes
5227 void idle_task_exit(void)
5229 struct mm_struct
*mm
= current
->active_mm
;
5231 BUG_ON(cpu_online(smp_processor_id()));
5233 if (mm
!= &init_mm
) {
5234 switch_mm(mm
, &init_mm
, current
);
5235 finish_arch_post_lock_switch();
5241 * Since this CPU is going 'away' for a while, fold any nr_active delta
5242 * we might have. Assumes we're called after migrate_tasks() so that the
5243 * nr_active count is stable.
5245 * Also see the comment "Global load-average calculations".
5247 static void calc_load_migrate(struct rq
*rq
)
5249 long delta
= calc_load_fold_active(rq
);
5251 atomic_long_add(delta
, &calc_load_tasks
);
5254 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5258 static const struct sched_class fake_sched_class
= {
5259 .put_prev_task
= put_prev_task_fake
,
5262 static struct task_struct fake_task
= {
5264 * Avoid pull_{rt,dl}_task()
5266 .prio
= MAX_PRIO
+ 1,
5267 .sched_class
= &fake_sched_class
,
5271 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5272 * try_to_wake_up()->select_task_rq().
5274 * Called with rq->lock held even though we'er in stop_machine() and
5275 * there's no concurrency possible, we hold the required locks anyway
5276 * because of lock validation efforts.
5278 static void migrate_tasks(struct rq
*dead_rq
)
5280 struct rq
*rq
= dead_rq
;
5281 struct task_struct
*next
, *stop
= rq
->stop
;
5285 * Fudge the rq selection such that the below task selection loop
5286 * doesn't get stuck on the currently eligible stop task.
5288 * We're currently inside stop_machine() and the rq is either stuck
5289 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5290 * either way we should never end up calling schedule() until we're
5296 * put_prev_task() and pick_next_task() sched
5297 * class method both need to have an up-to-date
5298 * value of rq->clock[_task]
5300 update_rq_clock(rq
);
5304 * There's this thread running, bail when that's the only
5307 if (rq
->nr_running
== 1)
5311 * pick_next_task assumes pinned rq->lock.
5313 lockdep_pin_lock(&rq
->lock
);
5314 next
= pick_next_task(rq
, &fake_task
);
5316 next
->sched_class
->put_prev_task(rq
, next
);
5319 * Rules for changing task_struct::cpus_allowed are holding
5320 * both pi_lock and rq->lock, such that holding either
5321 * stabilizes the mask.
5323 * Drop rq->lock is not quite as disastrous as it usually is
5324 * because !cpu_active at this point, which means load-balance
5325 * will not interfere. Also, stop-machine.
5327 lockdep_unpin_lock(&rq
->lock
);
5328 raw_spin_unlock(&rq
->lock
);
5329 raw_spin_lock(&next
->pi_lock
);
5330 raw_spin_lock(&rq
->lock
);
5333 * Since we're inside stop-machine, _nothing_ should have
5334 * changed the task, WARN if weird stuff happened, because in
5335 * that case the above rq->lock drop is a fail too.
5337 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5338 raw_spin_unlock(&next
->pi_lock
);
5342 /* Find suitable destination for @next, with force if needed. */
5343 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5345 rq
= __migrate_task(rq
, next
, dest_cpu
);
5346 if (rq
!= dead_rq
) {
5347 raw_spin_unlock(&rq
->lock
);
5349 raw_spin_lock(&rq
->lock
);
5351 raw_spin_unlock(&next
->pi_lock
);
5356 #endif /* CONFIG_HOTPLUG_CPU */
5358 static void set_rq_online(struct rq
*rq
)
5361 const struct sched_class
*class;
5363 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5366 for_each_class(class) {
5367 if (class->rq_online
)
5368 class->rq_online(rq
);
5373 static void set_rq_offline(struct rq
*rq
)
5376 const struct sched_class
*class;
5378 for_each_class(class) {
5379 if (class->rq_offline
)
5380 class->rq_offline(rq
);
5383 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5389 * migration_call - callback that gets triggered when a CPU is added.
5390 * Here we can start up the necessary migration thread for the new CPU.
5393 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5395 int cpu
= (long)hcpu
;
5396 unsigned long flags
;
5397 struct rq
*rq
= cpu_rq(cpu
);
5399 switch (action
& ~CPU_TASKS_FROZEN
) {
5401 case CPU_UP_PREPARE
:
5402 rq
->calc_load_update
= calc_load_update
;
5403 account_reset_rq(rq
);
5407 /* Update our root-domain */
5408 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5410 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5414 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5417 #ifdef CONFIG_HOTPLUG_CPU
5419 sched_ttwu_pending();
5420 /* Update our root-domain */
5421 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5423 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5427 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5428 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5432 calc_load_migrate(rq
);
5437 update_max_interval();
5443 * Register at high priority so that task migration (migrate_all_tasks)
5444 * happens before everything else. This has to be lower priority than
5445 * the notifier in the perf_event subsystem, though.
5447 static struct notifier_block migration_notifier
= {
5448 .notifier_call
= migration_call
,
5449 .priority
= CPU_PRI_MIGRATION
,
5452 static void set_cpu_rq_start_time(void)
5454 int cpu
= smp_processor_id();
5455 struct rq
*rq
= cpu_rq(cpu
);
5456 rq
->age_stamp
= sched_clock_cpu(cpu
);
5459 static int sched_cpu_active(struct notifier_block
*nfb
,
5460 unsigned long action
, void *hcpu
)
5462 int cpu
= (long)hcpu
;
5464 switch (action
& ~CPU_TASKS_FROZEN
) {
5466 set_cpu_rq_start_time();
5469 case CPU_DOWN_FAILED
:
5470 set_cpu_active(cpu
, true);
5478 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5479 unsigned long action
, void *hcpu
)
5481 switch (action
& ~CPU_TASKS_FROZEN
) {
5482 case CPU_DOWN_PREPARE
:
5483 set_cpu_active((long)hcpu
, false);
5490 static int __init
migration_init(void)
5492 void *cpu
= (void *)(long)smp_processor_id();
5495 /* Initialize migration for the boot CPU */
5496 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5497 BUG_ON(err
== NOTIFY_BAD
);
5498 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5499 register_cpu_notifier(&migration_notifier
);
5501 /* Register cpu active notifiers */
5502 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5503 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5507 early_initcall(migration_init
);
5509 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5511 #ifdef CONFIG_SCHED_DEBUG
5513 static __read_mostly
int sched_debug_enabled
;
5515 static int __init
sched_debug_setup(char *str
)
5517 sched_debug_enabled
= 1;
5521 early_param("sched_debug", sched_debug_setup
);
5523 static inline bool sched_debug(void)
5525 return sched_debug_enabled
;
5528 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5529 struct cpumask
*groupmask
)
5531 struct sched_group
*group
= sd
->groups
;
5533 cpumask_clear(groupmask
);
5535 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5537 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5538 printk("does not load-balance\n");
5540 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5545 printk(KERN_CONT
"span %*pbl level %s\n",
5546 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5548 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5549 printk(KERN_ERR
"ERROR: domain->span does not contain "
5552 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5553 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5557 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5561 printk(KERN_ERR
"ERROR: group is NULL\n");
5565 if (!cpumask_weight(sched_group_cpus(group
))) {
5566 printk(KERN_CONT
"\n");
5567 printk(KERN_ERR
"ERROR: empty group\n");
5571 if (!(sd
->flags
& SD_OVERLAP
) &&
5572 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5573 printk(KERN_CONT
"\n");
5574 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5578 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5580 printk(KERN_CONT
" %*pbl",
5581 cpumask_pr_args(sched_group_cpus(group
)));
5582 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5583 printk(KERN_CONT
" (cpu_capacity = %d)",
5584 group
->sgc
->capacity
);
5587 group
= group
->next
;
5588 } while (group
!= sd
->groups
);
5589 printk(KERN_CONT
"\n");
5591 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5592 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5595 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5596 printk(KERN_ERR
"ERROR: parent span is not a superset "
5597 "of domain->span\n");
5601 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5605 if (!sched_debug_enabled
)
5609 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5613 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5616 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5624 #else /* !CONFIG_SCHED_DEBUG */
5625 # define sched_domain_debug(sd, cpu) do { } while (0)
5626 static inline bool sched_debug(void)
5630 #endif /* CONFIG_SCHED_DEBUG */
5632 static int sd_degenerate(struct sched_domain
*sd
)
5634 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5637 /* Following flags need at least 2 groups */
5638 if (sd
->flags
& (SD_LOAD_BALANCE
|
5639 SD_BALANCE_NEWIDLE
|
5642 SD_SHARE_CPUCAPACITY
|
5643 SD_SHARE_PKG_RESOURCES
|
5644 SD_SHARE_POWERDOMAIN
)) {
5645 if (sd
->groups
!= sd
->groups
->next
)
5649 /* Following flags don't use groups */
5650 if (sd
->flags
& (SD_WAKE_AFFINE
))
5657 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5659 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5661 if (sd_degenerate(parent
))
5664 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5667 /* Flags needing groups don't count if only 1 group in parent */
5668 if (parent
->groups
== parent
->groups
->next
) {
5669 pflags
&= ~(SD_LOAD_BALANCE
|
5670 SD_BALANCE_NEWIDLE
|
5673 SD_SHARE_CPUCAPACITY
|
5674 SD_SHARE_PKG_RESOURCES
|
5676 SD_SHARE_POWERDOMAIN
);
5677 if (nr_node_ids
== 1)
5678 pflags
&= ~SD_SERIALIZE
;
5680 if (~cflags
& pflags
)
5686 static void free_rootdomain(struct rcu_head
*rcu
)
5688 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5690 cpupri_cleanup(&rd
->cpupri
);
5691 cpudl_cleanup(&rd
->cpudl
);
5692 free_cpumask_var(rd
->dlo_mask
);
5693 free_cpumask_var(rd
->rto_mask
);
5694 free_cpumask_var(rd
->online
);
5695 free_cpumask_var(rd
->span
);
5699 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5701 struct root_domain
*old_rd
= NULL
;
5702 unsigned long flags
;
5704 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5709 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5712 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5715 * If we dont want to free the old_rd yet then
5716 * set old_rd to NULL to skip the freeing later
5719 if (!atomic_dec_and_test(&old_rd
->refcount
))
5723 atomic_inc(&rd
->refcount
);
5726 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5727 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5730 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5733 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5736 static int init_rootdomain(struct root_domain
*rd
)
5738 memset(rd
, 0, sizeof(*rd
));
5740 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5742 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5744 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5746 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5749 init_dl_bw(&rd
->dl_bw
);
5750 if (cpudl_init(&rd
->cpudl
) != 0)
5753 if (cpupri_init(&rd
->cpupri
) != 0)
5758 free_cpumask_var(rd
->rto_mask
);
5760 free_cpumask_var(rd
->dlo_mask
);
5762 free_cpumask_var(rd
->online
);
5764 free_cpumask_var(rd
->span
);
5770 * By default the system creates a single root-domain with all cpus as
5771 * members (mimicking the global state we have today).
5773 struct root_domain def_root_domain
;
5775 static void init_defrootdomain(void)
5777 init_rootdomain(&def_root_domain
);
5779 atomic_set(&def_root_domain
.refcount
, 1);
5782 static struct root_domain
*alloc_rootdomain(void)
5784 struct root_domain
*rd
;
5786 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5790 if (init_rootdomain(rd
) != 0) {
5798 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5800 struct sched_group
*tmp
, *first
;
5809 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5814 } while (sg
!= first
);
5817 static void free_sched_domain(struct rcu_head
*rcu
)
5819 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5822 * If its an overlapping domain it has private groups, iterate and
5825 if (sd
->flags
& SD_OVERLAP
) {
5826 free_sched_groups(sd
->groups
, 1);
5827 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5828 kfree(sd
->groups
->sgc
);
5834 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5836 call_rcu(&sd
->rcu
, free_sched_domain
);
5839 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5841 for (; sd
; sd
= sd
->parent
)
5842 destroy_sched_domain(sd
, cpu
);
5846 * Keep a special pointer to the highest sched_domain that has
5847 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5848 * allows us to avoid some pointer chasing select_idle_sibling().
5850 * Also keep a unique ID per domain (we use the first cpu number in
5851 * the cpumask of the domain), this allows us to quickly tell if
5852 * two cpus are in the same cache domain, see cpus_share_cache().
5854 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5855 DEFINE_PER_CPU(int, sd_llc_size
);
5856 DEFINE_PER_CPU(int, sd_llc_id
);
5857 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5858 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5859 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5861 static void update_top_cache_domain(int cpu
)
5863 struct sched_domain
*sd
;
5864 struct sched_domain
*busy_sd
= NULL
;
5868 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5870 id
= cpumask_first(sched_domain_span(sd
));
5871 size
= cpumask_weight(sched_domain_span(sd
));
5872 busy_sd
= sd
->parent
; /* sd_busy */
5874 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5876 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5877 per_cpu(sd_llc_size
, cpu
) = size
;
5878 per_cpu(sd_llc_id
, cpu
) = id
;
5880 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5881 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5883 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5884 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5888 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5889 * hold the hotplug lock.
5892 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5894 struct rq
*rq
= cpu_rq(cpu
);
5895 struct sched_domain
*tmp
;
5897 /* Remove the sched domains which do not contribute to scheduling. */
5898 for (tmp
= sd
; tmp
; ) {
5899 struct sched_domain
*parent
= tmp
->parent
;
5903 if (sd_parent_degenerate(tmp
, parent
)) {
5904 tmp
->parent
= parent
->parent
;
5906 parent
->parent
->child
= tmp
;
5908 * Transfer SD_PREFER_SIBLING down in case of a
5909 * degenerate parent; the spans match for this
5910 * so the property transfers.
5912 if (parent
->flags
& SD_PREFER_SIBLING
)
5913 tmp
->flags
|= SD_PREFER_SIBLING
;
5914 destroy_sched_domain(parent
, cpu
);
5919 if (sd
&& sd_degenerate(sd
)) {
5922 destroy_sched_domain(tmp
, cpu
);
5927 sched_domain_debug(sd
, cpu
);
5929 rq_attach_root(rq
, rd
);
5931 rcu_assign_pointer(rq
->sd
, sd
);
5932 destroy_sched_domains(tmp
, cpu
);
5934 update_top_cache_domain(cpu
);
5937 /* Setup the mask of cpus configured for isolated domains */
5938 static int __init
isolated_cpu_setup(char *str
)
5942 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5943 ret
= cpulist_parse(str
, cpu_isolated_map
);
5945 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5950 __setup("isolcpus=", isolated_cpu_setup
);
5953 struct sched_domain
** __percpu sd
;
5954 struct root_domain
*rd
;
5965 * Build an iteration mask that can exclude certain CPUs from the upwards
5968 * Asymmetric node setups can result in situations where the domain tree is of
5969 * unequal depth, make sure to skip domains that already cover the entire
5972 * In that case build_sched_domains() will have terminated the iteration early
5973 * and our sibling sd spans will be empty. Domains should always include the
5974 * cpu they're built on, so check that.
5977 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5979 const struct cpumask
*span
= sched_domain_span(sd
);
5980 struct sd_data
*sdd
= sd
->private;
5981 struct sched_domain
*sibling
;
5984 for_each_cpu(i
, span
) {
5985 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5986 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5989 cpumask_set_cpu(i
, sched_group_mask(sg
));
5994 * Return the canonical balance cpu for this group, this is the first cpu
5995 * of this group that's also in the iteration mask.
5997 int group_balance_cpu(struct sched_group
*sg
)
5999 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6003 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6005 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6006 const struct cpumask
*span
= sched_domain_span(sd
);
6007 struct cpumask
*covered
= sched_domains_tmpmask
;
6008 struct sd_data
*sdd
= sd
->private;
6009 struct sched_domain
*sibling
;
6012 cpumask_clear(covered
);
6014 for_each_cpu(i
, span
) {
6015 struct cpumask
*sg_span
;
6017 if (cpumask_test_cpu(i
, covered
))
6020 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6022 /* See the comment near build_group_mask(). */
6023 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6026 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6027 GFP_KERNEL
, cpu_to_node(cpu
));
6032 sg_span
= sched_group_cpus(sg
);
6034 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6036 cpumask_set_cpu(i
, sg_span
);
6038 cpumask_or(covered
, covered
, sg_span
);
6040 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6041 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6042 build_group_mask(sd
, sg
);
6045 * Initialize sgc->capacity such that even if we mess up the
6046 * domains and no possible iteration will get us here, we won't
6049 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6052 * Make sure the first group of this domain contains the
6053 * canonical balance cpu. Otherwise the sched_domain iteration
6054 * breaks. See update_sg_lb_stats().
6056 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6057 group_balance_cpu(sg
) == cpu
)
6067 sd
->groups
= groups
;
6072 free_sched_groups(first
, 0);
6077 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6079 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6080 struct sched_domain
*child
= sd
->child
;
6083 cpu
= cpumask_first(sched_domain_span(child
));
6086 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6087 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6088 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6095 * build_sched_groups will build a circular linked list of the groups
6096 * covered by the given span, and will set each group's ->cpumask correctly,
6097 * and ->cpu_capacity to 0.
6099 * Assumes the sched_domain tree is fully constructed
6102 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6104 struct sched_group
*first
= NULL
, *last
= NULL
;
6105 struct sd_data
*sdd
= sd
->private;
6106 const struct cpumask
*span
= sched_domain_span(sd
);
6107 struct cpumask
*covered
;
6110 get_group(cpu
, sdd
, &sd
->groups
);
6111 atomic_inc(&sd
->groups
->ref
);
6113 if (cpu
!= cpumask_first(span
))
6116 lockdep_assert_held(&sched_domains_mutex
);
6117 covered
= sched_domains_tmpmask
;
6119 cpumask_clear(covered
);
6121 for_each_cpu(i
, span
) {
6122 struct sched_group
*sg
;
6125 if (cpumask_test_cpu(i
, covered
))
6128 group
= get_group(i
, sdd
, &sg
);
6129 cpumask_setall(sched_group_mask(sg
));
6131 for_each_cpu(j
, span
) {
6132 if (get_group(j
, sdd
, NULL
) != group
)
6135 cpumask_set_cpu(j
, covered
);
6136 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6151 * Initialize sched groups cpu_capacity.
6153 * cpu_capacity indicates the capacity of sched group, which is used while
6154 * distributing the load between different sched groups in a sched domain.
6155 * Typically cpu_capacity for all the groups in a sched domain will be same
6156 * unless there are asymmetries in the topology. If there are asymmetries,
6157 * group having more cpu_capacity will pickup more load compared to the
6158 * group having less cpu_capacity.
6160 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6162 struct sched_group
*sg
= sd
->groups
;
6167 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6169 } while (sg
!= sd
->groups
);
6171 if (cpu
!= group_balance_cpu(sg
))
6174 update_group_capacity(sd
, cpu
);
6175 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6179 * Initializers for schedule domains
6180 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6183 static int default_relax_domain_level
= -1;
6184 int sched_domain_level_max
;
6186 static int __init
setup_relax_domain_level(char *str
)
6188 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6189 pr_warn("Unable to set relax_domain_level\n");
6193 __setup("relax_domain_level=", setup_relax_domain_level
);
6195 static void set_domain_attribute(struct sched_domain
*sd
,
6196 struct sched_domain_attr
*attr
)
6200 if (!attr
|| attr
->relax_domain_level
< 0) {
6201 if (default_relax_domain_level
< 0)
6204 request
= default_relax_domain_level
;
6206 request
= attr
->relax_domain_level
;
6207 if (request
< sd
->level
) {
6208 /* turn off idle balance on this domain */
6209 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6211 /* turn on idle balance on this domain */
6212 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6216 static void __sdt_free(const struct cpumask
*cpu_map
);
6217 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6219 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6220 const struct cpumask
*cpu_map
)
6224 if (!atomic_read(&d
->rd
->refcount
))
6225 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6227 free_percpu(d
->sd
); /* fall through */
6229 __sdt_free(cpu_map
); /* fall through */
6235 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6236 const struct cpumask
*cpu_map
)
6238 memset(d
, 0, sizeof(*d
));
6240 if (__sdt_alloc(cpu_map
))
6241 return sa_sd_storage
;
6242 d
->sd
= alloc_percpu(struct sched_domain
*);
6244 return sa_sd_storage
;
6245 d
->rd
= alloc_rootdomain();
6248 return sa_rootdomain
;
6252 * NULL the sd_data elements we've used to build the sched_domain and
6253 * sched_group structure so that the subsequent __free_domain_allocs()
6254 * will not free the data we're using.
6256 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6258 struct sd_data
*sdd
= sd
->private;
6260 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6261 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6263 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6264 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6266 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6267 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6271 static int sched_domains_numa_levels
;
6272 enum numa_topology_type sched_numa_topology_type
;
6273 static int *sched_domains_numa_distance
;
6274 int sched_max_numa_distance
;
6275 static struct cpumask
***sched_domains_numa_masks
;
6276 static int sched_domains_curr_level
;
6280 * SD_flags allowed in topology descriptions.
6282 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6283 * SD_SHARE_PKG_RESOURCES - describes shared caches
6284 * SD_NUMA - describes NUMA topologies
6285 * SD_SHARE_POWERDOMAIN - describes shared power domain
6288 * SD_ASYM_PACKING - describes SMT quirks
6290 #define TOPOLOGY_SD_FLAGS \
6291 (SD_SHARE_CPUCAPACITY | \
6292 SD_SHARE_PKG_RESOURCES | \
6295 SD_SHARE_POWERDOMAIN)
6297 static struct sched_domain
*
6298 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6300 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6301 int sd_weight
, sd_flags
= 0;
6305 * Ugly hack to pass state to sd_numa_mask()...
6307 sched_domains_curr_level
= tl
->numa_level
;
6310 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6313 sd_flags
= (*tl
->sd_flags
)();
6314 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6315 "wrong sd_flags in topology description\n"))
6316 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6318 *sd
= (struct sched_domain
){
6319 .min_interval
= sd_weight
,
6320 .max_interval
= 2*sd_weight
,
6322 .imbalance_pct
= 125,
6324 .cache_nice_tries
= 0,
6331 .flags
= 1*SD_LOAD_BALANCE
6332 | 1*SD_BALANCE_NEWIDLE
6337 | 0*SD_SHARE_CPUCAPACITY
6338 | 0*SD_SHARE_PKG_RESOURCES
6340 | 0*SD_PREFER_SIBLING
6345 .last_balance
= jiffies
,
6346 .balance_interval
= sd_weight
,
6348 .max_newidle_lb_cost
= 0,
6349 .next_decay_max_lb_cost
= jiffies
,
6350 #ifdef CONFIG_SCHED_DEBUG
6356 * Convert topological properties into behaviour.
6359 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6360 sd
->flags
|= SD_PREFER_SIBLING
;
6361 sd
->imbalance_pct
= 110;
6362 sd
->smt_gain
= 1178; /* ~15% */
6364 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6365 sd
->imbalance_pct
= 117;
6366 sd
->cache_nice_tries
= 1;
6370 } else if (sd
->flags
& SD_NUMA
) {
6371 sd
->cache_nice_tries
= 2;
6375 sd
->flags
|= SD_SERIALIZE
;
6376 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6377 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6384 sd
->flags
|= SD_PREFER_SIBLING
;
6385 sd
->cache_nice_tries
= 1;
6390 sd
->private = &tl
->data
;
6396 * Topology list, bottom-up.
6398 static struct sched_domain_topology_level default_topology
[] = {
6399 #ifdef CONFIG_SCHED_SMT
6400 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6402 #ifdef CONFIG_SCHED_MC
6403 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6405 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6409 static struct sched_domain_topology_level
*sched_domain_topology
=
6412 #define for_each_sd_topology(tl) \
6413 for (tl = sched_domain_topology; tl->mask; tl++)
6415 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6417 sched_domain_topology
= tl
;
6422 static const struct cpumask
*sd_numa_mask(int cpu
)
6424 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6427 static void sched_numa_warn(const char *str
)
6429 static int done
= false;
6437 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6439 for (i
= 0; i
< nr_node_ids
; i
++) {
6440 printk(KERN_WARNING
" ");
6441 for (j
= 0; j
< nr_node_ids
; j
++)
6442 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6443 printk(KERN_CONT
"\n");
6445 printk(KERN_WARNING
"\n");
6448 bool find_numa_distance(int distance
)
6452 if (distance
== node_distance(0, 0))
6455 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6456 if (sched_domains_numa_distance
[i
] == distance
)
6464 * A system can have three types of NUMA topology:
6465 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6466 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6467 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6469 * The difference between a glueless mesh topology and a backplane
6470 * topology lies in whether communication between not directly
6471 * connected nodes goes through intermediary nodes (where programs
6472 * could run), or through backplane controllers. This affects
6473 * placement of programs.
6475 * The type of topology can be discerned with the following tests:
6476 * - If the maximum distance between any nodes is 1 hop, the system
6477 * is directly connected.
6478 * - If for two nodes A and B, located N > 1 hops away from each other,
6479 * there is an intermediary node C, which is < N hops away from both
6480 * nodes A and B, the system is a glueless mesh.
6482 static void init_numa_topology_type(void)
6486 n
= sched_max_numa_distance
;
6488 if (sched_domains_numa_levels
<= 1) {
6489 sched_numa_topology_type
= NUMA_DIRECT
;
6493 for_each_online_node(a
) {
6494 for_each_online_node(b
) {
6495 /* Find two nodes furthest removed from each other. */
6496 if (node_distance(a
, b
) < n
)
6499 /* Is there an intermediary node between a and b? */
6500 for_each_online_node(c
) {
6501 if (node_distance(a
, c
) < n
&&
6502 node_distance(b
, c
) < n
) {
6503 sched_numa_topology_type
=
6509 sched_numa_topology_type
= NUMA_BACKPLANE
;
6515 static void sched_init_numa(void)
6517 int next_distance
, curr_distance
= node_distance(0, 0);
6518 struct sched_domain_topology_level
*tl
;
6522 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6523 if (!sched_domains_numa_distance
)
6527 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6528 * unique distances in the node_distance() table.
6530 * Assumes node_distance(0,j) includes all distances in
6531 * node_distance(i,j) in order to avoid cubic time.
6533 next_distance
= curr_distance
;
6534 for (i
= 0; i
< nr_node_ids
; i
++) {
6535 for (j
= 0; j
< nr_node_ids
; j
++) {
6536 for (k
= 0; k
< nr_node_ids
; k
++) {
6537 int distance
= node_distance(i
, k
);
6539 if (distance
> curr_distance
&&
6540 (distance
< next_distance
||
6541 next_distance
== curr_distance
))
6542 next_distance
= distance
;
6545 * While not a strong assumption it would be nice to know
6546 * about cases where if node A is connected to B, B is not
6547 * equally connected to A.
6549 if (sched_debug() && node_distance(k
, i
) != distance
)
6550 sched_numa_warn("Node-distance not symmetric");
6552 if (sched_debug() && i
&& !find_numa_distance(distance
))
6553 sched_numa_warn("Node-0 not representative");
6555 if (next_distance
!= curr_distance
) {
6556 sched_domains_numa_distance
[level
++] = next_distance
;
6557 sched_domains_numa_levels
= level
;
6558 curr_distance
= next_distance
;
6563 * In case of sched_debug() we verify the above assumption.
6573 * 'level' contains the number of unique distances, excluding the
6574 * identity distance node_distance(i,i).
6576 * The sched_domains_numa_distance[] array includes the actual distance
6581 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6582 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6583 * the array will contain less then 'level' members. This could be
6584 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6585 * in other functions.
6587 * We reset it to 'level' at the end of this function.
6589 sched_domains_numa_levels
= 0;
6591 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6592 if (!sched_domains_numa_masks
)
6596 * Now for each level, construct a mask per node which contains all
6597 * cpus of nodes that are that many hops away from us.
6599 for (i
= 0; i
< level
; i
++) {
6600 sched_domains_numa_masks
[i
] =
6601 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6602 if (!sched_domains_numa_masks
[i
])
6605 for (j
= 0; j
< nr_node_ids
; j
++) {
6606 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6610 sched_domains_numa_masks
[i
][j
] = mask
;
6613 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6616 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6621 /* Compute default topology size */
6622 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6624 tl
= kzalloc((i
+ level
+ 1) *
6625 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6630 * Copy the default topology bits..
6632 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6633 tl
[i
] = sched_domain_topology
[i
];
6636 * .. and append 'j' levels of NUMA goodness.
6638 for (j
= 0; j
< level
; i
++, j
++) {
6639 tl
[i
] = (struct sched_domain_topology_level
){
6640 .mask
= sd_numa_mask
,
6641 .sd_flags
= cpu_numa_flags
,
6642 .flags
= SDTL_OVERLAP
,
6648 sched_domain_topology
= tl
;
6650 sched_domains_numa_levels
= level
;
6651 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6653 init_numa_topology_type();
6656 static void sched_domains_numa_masks_set(int cpu
)
6659 int node
= cpu_to_node(cpu
);
6661 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6662 for (j
= 0; j
< nr_node_ids
; j
++) {
6663 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6664 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6669 static void sched_domains_numa_masks_clear(int cpu
)
6672 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6673 for (j
= 0; j
< nr_node_ids
; j
++)
6674 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6679 * Update sched_domains_numa_masks[level][node] array when new cpus
6682 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6683 unsigned long action
,
6686 int cpu
= (long)hcpu
;
6688 switch (action
& ~CPU_TASKS_FROZEN
) {
6690 sched_domains_numa_masks_set(cpu
);
6694 sched_domains_numa_masks_clear(cpu
);
6704 static inline void sched_init_numa(void)
6708 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6709 unsigned long action
,
6714 #endif /* CONFIG_NUMA */
6716 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6718 struct sched_domain_topology_level
*tl
;
6721 for_each_sd_topology(tl
) {
6722 struct sd_data
*sdd
= &tl
->data
;
6724 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6728 sdd
->sg
= alloc_percpu(struct sched_group
*);
6732 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6736 for_each_cpu(j
, cpu_map
) {
6737 struct sched_domain
*sd
;
6738 struct sched_group
*sg
;
6739 struct sched_group_capacity
*sgc
;
6741 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6742 GFP_KERNEL
, cpu_to_node(j
));
6746 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6748 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6749 GFP_KERNEL
, cpu_to_node(j
));
6755 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6757 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6758 GFP_KERNEL
, cpu_to_node(j
));
6762 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6769 static void __sdt_free(const struct cpumask
*cpu_map
)
6771 struct sched_domain_topology_level
*tl
;
6774 for_each_sd_topology(tl
) {
6775 struct sd_data
*sdd
= &tl
->data
;
6777 for_each_cpu(j
, cpu_map
) {
6778 struct sched_domain
*sd
;
6781 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6782 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6783 free_sched_groups(sd
->groups
, 0);
6784 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6788 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6790 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6792 free_percpu(sdd
->sd
);
6794 free_percpu(sdd
->sg
);
6796 free_percpu(sdd
->sgc
);
6801 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6802 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6803 struct sched_domain
*child
, int cpu
)
6805 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6809 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6811 sd
->level
= child
->level
+ 1;
6812 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6816 if (!cpumask_subset(sched_domain_span(child
),
6817 sched_domain_span(sd
))) {
6818 pr_err("BUG: arch topology borken\n");
6819 #ifdef CONFIG_SCHED_DEBUG
6820 pr_err(" the %s domain not a subset of the %s domain\n",
6821 child
->name
, sd
->name
);
6823 /* Fixup, ensure @sd has at least @child cpus. */
6824 cpumask_or(sched_domain_span(sd
),
6825 sched_domain_span(sd
),
6826 sched_domain_span(child
));
6830 set_domain_attribute(sd
, attr
);
6836 * Build sched domains for a given set of cpus and attach the sched domains
6837 * to the individual cpus
6839 static int build_sched_domains(const struct cpumask
*cpu_map
,
6840 struct sched_domain_attr
*attr
)
6842 enum s_alloc alloc_state
;
6843 struct sched_domain
*sd
;
6845 int i
, ret
= -ENOMEM
;
6847 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6848 if (alloc_state
!= sa_rootdomain
)
6851 /* Set up domains for cpus specified by the cpu_map. */
6852 for_each_cpu(i
, cpu_map
) {
6853 struct sched_domain_topology_level
*tl
;
6856 for_each_sd_topology(tl
) {
6857 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6858 if (tl
== sched_domain_topology
)
6859 *per_cpu_ptr(d
.sd
, i
) = sd
;
6860 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6861 sd
->flags
|= SD_OVERLAP
;
6862 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6867 /* Build the groups for the domains */
6868 for_each_cpu(i
, cpu_map
) {
6869 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6870 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6871 if (sd
->flags
& SD_OVERLAP
) {
6872 if (build_overlap_sched_groups(sd
, i
))
6875 if (build_sched_groups(sd
, i
))
6881 /* Calculate CPU capacity for physical packages and nodes */
6882 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6883 if (!cpumask_test_cpu(i
, cpu_map
))
6886 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6887 claim_allocations(i
, sd
);
6888 init_sched_groups_capacity(i
, sd
);
6892 /* Attach the domains */
6894 for_each_cpu(i
, cpu_map
) {
6895 sd
= *per_cpu_ptr(d
.sd
, i
);
6896 cpu_attach_domain(sd
, d
.rd
, i
);
6902 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6906 static cpumask_var_t
*doms_cur
; /* current sched domains */
6907 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6908 static struct sched_domain_attr
*dattr_cur
;
6909 /* attribues of custom domains in 'doms_cur' */
6912 * Special case: If a kmalloc of a doms_cur partition (array of
6913 * cpumask) fails, then fallback to a single sched domain,
6914 * as determined by the single cpumask fallback_doms.
6916 static cpumask_var_t fallback_doms
;
6919 * arch_update_cpu_topology lets virtualized architectures update the
6920 * cpu core maps. It is supposed to return 1 if the topology changed
6921 * or 0 if it stayed the same.
6923 int __weak
arch_update_cpu_topology(void)
6928 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6931 cpumask_var_t
*doms
;
6933 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6936 for (i
= 0; i
< ndoms
; i
++) {
6937 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6938 free_sched_domains(doms
, i
);
6945 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6948 for (i
= 0; i
< ndoms
; i
++)
6949 free_cpumask_var(doms
[i
]);
6954 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6955 * For now this just excludes isolated cpus, but could be used to
6956 * exclude other special cases in the future.
6958 static int init_sched_domains(const struct cpumask
*cpu_map
)
6962 arch_update_cpu_topology();
6964 doms_cur
= alloc_sched_domains(ndoms_cur
);
6966 doms_cur
= &fallback_doms
;
6967 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6968 err
= build_sched_domains(doms_cur
[0], NULL
);
6969 register_sched_domain_sysctl();
6975 * Detach sched domains from a group of cpus specified in cpu_map
6976 * These cpus will now be attached to the NULL domain
6978 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6983 for_each_cpu(i
, cpu_map
)
6984 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6988 /* handle null as "default" */
6989 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6990 struct sched_domain_attr
*new, int idx_new
)
6992 struct sched_domain_attr tmp
;
6999 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7000 new ? (new + idx_new
) : &tmp
,
7001 sizeof(struct sched_domain_attr
));
7005 * Partition sched domains as specified by the 'ndoms_new'
7006 * cpumasks in the array doms_new[] of cpumasks. This compares
7007 * doms_new[] to the current sched domain partitioning, doms_cur[].
7008 * It destroys each deleted domain and builds each new domain.
7010 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7011 * The masks don't intersect (don't overlap.) We should setup one
7012 * sched domain for each mask. CPUs not in any of the cpumasks will
7013 * not be load balanced. If the same cpumask appears both in the
7014 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7017 * The passed in 'doms_new' should be allocated using
7018 * alloc_sched_domains. This routine takes ownership of it and will
7019 * free_sched_domains it when done with it. If the caller failed the
7020 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7021 * and partition_sched_domains() will fallback to the single partition
7022 * 'fallback_doms', it also forces the domains to be rebuilt.
7024 * If doms_new == NULL it will be replaced with cpu_online_mask.
7025 * ndoms_new == 0 is a special case for destroying existing domains,
7026 * and it will not create the default domain.
7028 * Call with hotplug lock held
7030 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7031 struct sched_domain_attr
*dattr_new
)
7036 mutex_lock(&sched_domains_mutex
);
7038 /* always unregister in case we don't destroy any domains */
7039 unregister_sched_domain_sysctl();
7041 /* Let architecture update cpu core mappings. */
7042 new_topology
= arch_update_cpu_topology();
7044 n
= doms_new
? ndoms_new
: 0;
7046 /* Destroy deleted domains */
7047 for (i
= 0; i
< ndoms_cur
; i
++) {
7048 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7049 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7050 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7053 /* no match - a current sched domain not in new doms_new[] */
7054 detach_destroy_domains(doms_cur
[i
]);
7060 if (doms_new
== NULL
) {
7062 doms_new
= &fallback_doms
;
7063 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7064 WARN_ON_ONCE(dattr_new
);
7067 /* Build new domains */
7068 for (i
= 0; i
< ndoms_new
; i
++) {
7069 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7070 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7071 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7074 /* no match - add a new doms_new */
7075 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7080 /* Remember the new sched domains */
7081 if (doms_cur
!= &fallback_doms
)
7082 free_sched_domains(doms_cur
, ndoms_cur
);
7083 kfree(dattr_cur
); /* kfree(NULL) is safe */
7084 doms_cur
= doms_new
;
7085 dattr_cur
= dattr_new
;
7086 ndoms_cur
= ndoms_new
;
7088 register_sched_domain_sysctl();
7090 mutex_unlock(&sched_domains_mutex
);
7093 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7096 * Update cpusets according to cpu_active mask. If cpusets are
7097 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7098 * around partition_sched_domains().
7100 * If we come here as part of a suspend/resume, don't touch cpusets because we
7101 * want to restore it back to its original state upon resume anyway.
7103 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7107 case CPU_ONLINE_FROZEN
:
7108 case CPU_DOWN_FAILED_FROZEN
:
7111 * num_cpus_frozen tracks how many CPUs are involved in suspend
7112 * resume sequence. As long as this is not the last online
7113 * operation in the resume sequence, just build a single sched
7114 * domain, ignoring cpusets.
7117 if (likely(num_cpus_frozen
)) {
7118 partition_sched_domains(1, NULL
, NULL
);
7123 * This is the last CPU online operation. So fall through and
7124 * restore the original sched domains by considering the
7125 * cpuset configurations.
7129 cpuset_update_active_cpus(true);
7137 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7140 unsigned long flags
;
7141 long cpu
= (long)hcpu
;
7147 case CPU_DOWN_PREPARE
:
7148 rcu_read_lock_sched();
7149 dl_b
= dl_bw_of(cpu
);
7151 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7152 cpus
= dl_bw_cpus(cpu
);
7153 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7154 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7156 rcu_read_unlock_sched();
7159 return notifier_from_errno(-EBUSY
);
7160 cpuset_update_active_cpus(false);
7162 case CPU_DOWN_PREPARE_FROZEN
:
7164 partition_sched_domains(1, NULL
, NULL
);
7172 void __init
sched_init_smp(void)
7174 cpumask_var_t non_isolated_cpus
;
7176 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7177 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7182 * There's no userspace yet to cause hotplug operations; hence all the
7183 * cpu masks are stable and all blatant races in the below code cannot
7186 mutex_lock(&sched_domains_mutex
);
7187 init_sched_domains(cpu_active_mask
);
7188 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7189 if (cpumask_empty(non_isolated_cpus
))
7190 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7191 mutex_unlock(&sched_domains_mutex
);
7193 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7194 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7195 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7199 /* Move init over to a non-isolated CPU */
7200 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7202 sched_init_granularity();
7203 free_cpumask_var(non_isolated_cpus
);
7205 init_sched_rt_class();
7206 init_sched_dl_class();
7209 void __init
sched_init_smp(void)
7211 sched_init_granularity();
7213 #endif /* CONFIG_SMP */
7215 int in_sched_functions(unsigned long addr
)
7217 return in_lock_functions(addr
) ||
7218 (addr
>= (unsigned long)__sched_text_start
7219 && addr
< (unsigned long)__sched_text_end
);
7222 #ifdef CONFIG_CGROUP_SCHED
7224 * Default task group.
7225 * Every task in system belongs to this group at bootup.
7227 struct task_group root_task_group
;
7228 LIST_HEAD(task_groups
);
7230 /* Cacheline aligned slab cache for task_group */
7231 static struct kmem_cache
*task_group_cache __read_mostly
;
7234 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7236 void __init
sched_init(void)
7239 unsigned long alloc_size
= 0, ptr
;
7241 #ifdef CONFIG_FAIR_GROUP_SCHED
7242 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7244 #ifdef CONFIG_RT_GROUP_SCHED
7245 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7248 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7250 #ifdef CONFIG_FAIR_GROUP_SCHED
7251 root_task_group
.se
= (struct sched_entity
**)ptr
;
7252 ptr
+= nr_cpu_ids
* sizeof(void **);
7254 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7255 ptr
+= nr_cpu_ids
* sizeof(void **);
7257 #endif /* CONFIG_FAIR_GROUP_SCHED */
7258 #ifdef CONFIG_RT_GROUP_SCHED
7259 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7260 ptr
+= nr_cpu_ids
* sizeof(void **);
7262 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7263 ptr
+= nr_cpu_ids
* sizeof(void **);
7265 #endif /* CONFIG_RT_GROUP_SCHED */
7267 #ifdef CONFIG_CPUMASK_OFFSTACK
7268 for_each_possible_cpu(i
) {
7269 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7270 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7272 #endif /* CONFIG_CPUMASK_OFFSTACK */
7274 init_rt_bandwidth(&def_rt_bandwidth
,
7275 global_rt_period(), global_rt_runtime());
7276 init_dl_bandwidth(&def_dl_bandwidth
,
7277 global_rt_period(), global_rt_runtime());
7280 init_defrootdomain();
7283 #ifdef CONFIG_RT_GROUP_SCHED
7284 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7285 global_rt_period(), global_rt_runtime());
7286 #endif /* CONFIG_RT_GROUP_SCHED */
7288 #ifdef CONFIG_CGROUP_SCHED
7289 task_group_cache
= KMEM_CACHE(task_group
, 0);
7291 list_add(&root_task_group
.list
, &task_groups
);
7292 INIT_LIST_HEAD(&root_task_group
.children
);
7293 INIT_LIST_HEAD(&root_task_group
.siblings
);
7294 autogroup_init(&init_task
);
7295 #endif /* CONFIG_CGROUP_SCHED */
7297 for_each_possible_cpu(i
) {
7301 raw_spin_lock_init(&rq
->lock
);
7303 rq
->calc_load_active
= 0;
7304 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7305 init_cfs_rq(&rq
->cfs
);
7306 init_rt_rq(&rq
->rt
);
7307 init_dl_rq(&rq
->dl
);
7308 #ifdef CONFIG_FAIR_GROUP_SCHED
7309 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7310 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7312 * How much cpu bandwidth does root_task_group get?
7314 * In case of task-groups formed thr' the cgroup filesystem, it
7315 * gets 100% of the cpu resources in the system. This overall
7316 * system cpu resource is divided among the tasks of
7317 * root_task_group and its child task-groups in a fair manner,
7318 * based on each entity's (task or task-group's) weight
7319 * (se->load.weight).
7321 * In other words, if root_task_group has 10 tasks of weight
7322 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7323 * then A0's share of the cpu resource is:
7325 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7327 * We achieve this by letting root_task_group's tasks sit
7328 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7330 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7331 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7332 #endif /* CONFIG_FAIR_GROUP_SCHED */
7334 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7335 #ifdef CONFIG_RT_GROUP_SCHED
7336 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7339 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7340 rq
->cpu_load
[j
] = 0;
7342 rq
->last_load_update_tick
= jiffies
;
7347 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7348 rq
->balance_callback
= NULL
;
7349 rq
->active_balance
= 0;
7350 rq
->next_balance
= jiffies
;
7355 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7356 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7358 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7360 rq_attach_root(rq
, &def_root_domain
);
7361 #ifdef CONFIG_NO_HZ_COMMON
7364 #ifdef CONFIG_NO_HZ_FULL
7365 rq
->last_sched_tick
= 0;
7369 atomic_set(&rq
->nr_iowait
, 0);
7372 set_load_weight(&init_task
);
7374 #ifdef CONFIG_PREEMPT_NOTIFIERS
7375 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7379 * The boot idle thread does lazy MMU switching as well:
7381 atomic_inc(&init_mm
.mm_count
);
7382 enter_lazy_tlb(&init_mm
, current
);
7385 * During early bootup we pretend to be a normal task:
7387 current
->sched_class
= &fair_sched_class
;
7390 * Make us the idle thread. Technically, schedule() should not be
7391 * called from this thread, however somewhere below it might be,
7392 * but because we are the idle thread, we just pick up running again
7393 * when this runqueue becomes "idle".
7395 init_idle(current
, smp_processor_id());
7397 calc_load_update
= jiffies
+ LOAD_FREQ
;
7400 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7401 /* May be allocated at isolcpus cmdline parse time */
7402 if (cpu_isolated_map
== NULL
)
7403 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7404 idle_thread_set_boot_cpu();
7405 set_cpu_rq_start_time();
7407 init_sched_fair_class();
7409 scheduler_running
= 1;
7412 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7413 static inline int preempt_count_equals(int preempt_offset
)
7415 int nested
= preempt_count() + rcu_preempt_depth();
7417 return (nested
== preempt_offset
);
7420 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7423 * Blocking primitives will set (and therefore destroy) current->state,
7424 * since we will exit with TASK_RUNNING make sure we enter with it,
7425 * otherwise we will destroy state.
7427 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7428 "do not call blocking ops when !TASK_RUNNING; "
7429 "state=%lx set at [<%p>] %pS\n",
7431 (void *)current
->task_state_change
,
7432 (void *)current
->task_state_change
);
7434 ___might_sleep(file
, line
, preempt_offset
);
7436 EXPORT_SYMBOL(__might_sleep
);
7438 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7440 static unsigned long prev_jiffy
; /* ratelimiting */
7442 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7443 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7444 !is_idle_task(current
)) ||
7445 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7447 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7449 prev_jiffy
= jiffies
;
7452 "BUG: sleeping function called from invalid context at %s:%d\n",
7455 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7456 in_atomic(), irqs_disabled(),
7457 current
->pid
, current
->comm
);
7459 if (task_stack_end_corrupted(current
))
7460 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7462 debug_show_held_locks(current
);
7463 if (irqs_disabled())
7464 print_irqtrace_events(current
);
7465 #ifdef CONFIG_DEBUG_PREEMPT
7466 if (!preempt_count_equals(preempt_offset
)) {
7467 pr_err("Preemption disabled at:");
7468 print_ip_sym(current
->preempt_disable_ip
);
7474 EXPORT_SYMBOL(___might_sleep
);
7477 #ifdef CONFIG_MAGIC_SYSRQ
7478 void normalize_rt_tasks(void)
7480 struct task_struct
*g
, *p
;
7481 struct sched_attr attr
= {
7482 .sched_policy
= SCHED_NORMAL
,
7485 read_lock(&tasklist_lock
);
7486 for_each_process_thread(g
, p
) {
7488 * Only normalize user tasks:
7490 if (p
->flags
& PF_KTHREAD
)
7493 p
->se
.exec_start
= 0;
7494 #ifdef CONFIG_SCHEDSTATS
7495 p
->se
.statistics
.wait_start
= 0;
7496 p
->se
.statistics
.sleep_start
= 0;
7497 p
->se
.statistics
.block_start
= 0;
7500 if (!dl_task(p
) && !rt_task(p
)) {
7502 * Renice negative nice level userspace
7505 if (task_nice(p
) < 0)
7506 set_user_nice(p
, 0);
7510 __sched_setscheduler(p
, &attr
, false, false);
7512 read_unlock(&tasklist_lock
);
7515 #endif /* CONFIG_MAGIC_SYSRQ */
7517 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7519 * These functions are only useful for the IA64 MCA handling, or kdb.
7521 * They can only be called when the whole system has been
7522 * stopped - every CPU needs to be quiescent, and no scheduling
7523 * activity can take place. Using them for anything else would
7524 * be a serious bug, and as a result, they aren't even visible
7525 * under any other configuration.
7529 * curr_task - return the current task for a given cpu.
7530 * @cpu: the processor in question.
7532 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7534 * Return: The current task for @cpu.
7536 struct task_struct
*curr_task(int cpu
)
7538 return cpu_curr(cpu
);
7541 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7545 * set_curr_task - set the current task for a given cpu.
7546 * @cpu: the processor in question.
7547 * @p: the task pointer to set.
7549 * Description: This function must only be used when non-maskable interrupts
7550 * are serviced on a separate stack. It allows the architecture to switch the
7551 * notion of the current task on a cpu in a non-blocking manner. This function
7552 * must be called with all CPU's synchronized, and interrupts disabled, the
7553 * and caller must save the original value of the current task (see
7554 * curr_task() above) and restore that value before reenabling interrupts and
7555 * re-starting the system.
7557 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7559 void set_curr_task(int cpu
, struct task_struct
*p
)
7566 #ifdef CONFIG_CGROUP_SCHED
7567 /* task_group_lock serializes the addition/removal of task groups */
7568 static DEFINE_SPINLOCK(task_group_lock
);
7570 static void sched_free_group(struct task_group
*tg
)
7572 free_fair_sched_group(tg
);
7573 free_rt_sched_group(tg
);
7575 kmem_cache_free(task_group_cache
, tg
);
7578 /* allocate runqueue etc for a new task group */
7579 struct task_group
*sched_create_group(struct task_group
*parent
)
7581 struct task_group
*tg
;
7583 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7585 return ERR_PTR(-ENOMEM
);
7587 if (!alloc_fair_sched_group(tg
, parent
))
7590 if (!alloc_rt_sched_group(tg
, parent
))
7596 sched_free_group(tg
);
7597 return ERR_PTR(-ENOMEM
);
7600 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7602 unsigned long flags
;
7604 spin_lock_irqsave(&task_group_lock
, flags
);
7605 list_add_rcu(&tg
->list
, &task_groups
);
7607 WARN_ON(!parent
); /* root should already exist */
7609 tg
->parent
= parent
;
7610 INIT_LIST_HEAD(&tg
->children
);
7611 list_add_rcu(&tg
->siblings
, &parent
->children
);
7612 spin_unlock_irqrestore(&task_group_lock
, flags
);
7615 /* rcu callback to free various structures associated with a task group */
7616 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7618 /* now it should be safe to free those cfs_rqs */
7619 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7622 void sched_destroy_group(struct task_group
*tg
)
7624 /* wait for possible concurrent references to cfs_rqs complete */
7625 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7628 void sched_offline_group(struct task_group
*tg
)
7630 unsigned long flags
;
7632 /* end participation in shares distribution */
7633 unregister_fair_sched_group(tg
);
7635 spin_lock_irqsave(&task_group_lock
, flags
);
7636 list_del_rcu(&tg
->list
);
7637 list_del_rcu(&tg
->siblings
);
7638 spin_unlock_irqrestore(&task_group_lock
, flags
);
7641 /* change task's runqueue when it moves between groups.
7642 * The caller of this function should have put the task in its new group
7643 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7644 * reflect its new group.
7646 void sched_move_task(struct task_struct
*tsk
)
7648 struct task_group
*tg
;
7649 int queued
, running
;
7650 unsigned long flags
;
7653 rq
= task_rq_lock(tsk
, &flags
);
7655 running
= task_current(rq
, tsk
);
7656 queued
= task_on_rq_queued(tsk
);
7659 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7660 if (unlikely(running
))
7661 put_prev_task(rq
, tsk
);
7664 * All callers are synchronized by task_rq_lock(); we do not use RCU
7665 * which is pointless here. Thus, we pass "true" to task_css_check()
7666 * to prevent lockdep warnings.
7668 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7669 struct task_group
, css
);
7670 tg
= autogroup_task_group(tsk
, tg
);
7671 tsk
->sched_task_group
= tg
;
7673 #ifdef CONFIG_FAIR_GROUP_SCHED
7674 if (tsk
->sched_class
->task_move_group
)
7675 tsk
->sched_class
->task_move_group(tsk
);
7678 set_task_rq(tsk
, task_cpu(tsk
));
7680 if (unlikely(running
))
7681 tsk
->sched_class
->set_curr_task(rq
);
7683 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7685 task_rq_unlock(rq
, tsk
, &flags
);
7687 #endif /* CONFIG_CGROUP_SCHED */
7689 #ifdef CONFIG_RT_GROUP_SCHED
7691 * Ensure that the real time constraints are schedulable.
7693 static DEFINE_MUTEX(rt_constraints_mutex
);
7695 /* Must be called with tasklist_lock held */
7696 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7698 struct task_struct
*g
, *p
;
7701 * Autogroups do not have RT tasks; see autogroup_create().
7703 if (task_group_is_autogroup(tg
))
7706 for_each_process_thread(g
, p
) {
7707 if (rt_task(p
) && task_group(p
) == tg
)
7714 struct rt_schedulable_data
{
7715 struct task_group
*tg
;
7720 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7722 struct rt_schedulable_data
*d
= data
;
7723 struct task_group
*child
;
7724 unsigned long total
, sum
= 0;
7725 u64 period
, runtime
;
7727 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7728 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7731 period
= d
->rt_period
;
7732 runtime
= d
->rt_runtime
;
7736 * Cannot have more runtime than the period.
7738 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7742 * Ensure we don't starve existing RT tasks.
7744 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7747 total
= to_ratio(period
, runtime
);
7750 * Nobody can have more than the global setting allows.
7752 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7756 * The sum of our children's runtime should not exceed our own.
7758 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7759 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7760 runtime
= child
->rt_bandwidth
.rt_runtime
;
7762 if (child
== d
->tg
) {
7763 period
= d
->rt_period
;
7764 runtime
= d
->rt_runtime
;
7767 sum
+= to_ratio(period
, runtime
);
7776 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7780 struct rt_schedulable_data data
= {
7782 .rt_period
= period
,
7783 .rt_runtime
= runtime
,
7787 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7793 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7794 u64 rt_period
, u64 rt_runtime
)
7799 * Disallowing the root group RT runtime is BAD, it would disallow the
7800 * kernel creating (and or operating) RT threads.
7802 if (tg
== &root_task_group
&& rt_runtime
== 0)
7805 /* No period doesn't make any sense. */
7809 mutex_lock(&rt_constraints_mutex
);
7810 read_lock(&tasklist_lock
);
7811 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7815 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7816 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7817 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7819 for_each_possible_cpu(i
) {
7820 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7822 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7823 rt_rq
->rt_runtime
= rt_runtime
;
7824 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7826 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7828 read_unlock(&tasklist_lock
);
7829 mutex_unlock(&rt_constraints_mutex
);
7834 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7836 u64 rt_runtime
, rt_period
;
7838 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7839 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7840 if (rt_runtime_us
< 0)
7841 rt_runtime
= RUNTIME_INF
;
7843 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7846 static long sched_group_rt_runtime(struct task_group
*tg
)
7850 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7853 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7854 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7855 return rt_runtime_us
;
7858 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7860 u64 rt_runtime
, rt_period
;
7862 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7863 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7865 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7868 static long sched_group_rt_period(struct task_group
*tg
)
7872 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7873 do_div(rt_period_us
, NSEC_PER_USEC
);
7874 return rt_period_us
;
7876 #endif /* CONFIG_RT_GROUP_SCHED */
7878 #ifdef CONFIG_RT_GROUP_SCHED
7879 static int sched_rt_global_constraints(void)
7883 mutex_lock(&rt_constraints_mutex
);
7884 read_lock(&tasklist_lock
);
7885 ret
= __rt_schedulable(NULL
, 0, 0);
7886 read_unlock(&tasklist_lock
);
7887 mutex_unlock(&rt_constraints_mutex
);
7892 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7894 /* Don't accept realtime tasks when there is no way for them to run */
7895 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7901 #else /* !CONFIG_RT_GROUP_SCHED */
7902 static int sched_rt_global_constraints(void)
7904 unsigned long flags
;
7907 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7908 for_each_possible_cpu(i
) {
7909 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7911 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7912 rt_rq
->rt_runtime
= global_rt_runtime();
7913 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7915 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7919 #endif /* CONFIG_RT_GROUP_SCHED */
7921 static int sched_dl_global_validate(void)
7923 u64 runtime
= global_rt_runtime();
7924 u64 period
= global_rt_period();
7925 u64 new_bw
= to_ratio(period
, runtime
);
7928 unsigned long flags
;
7931 * Here we want to check the bandwidth not being set to some
7932 * value smaller than the currently allocated bandwidth in
7933 * any of the root_domains.
7935 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7936 * cycling on root_domains... Discussion on different/better
7937 * solutions is welcome!
7939 for_each_possible_cpu(cpu
) {
7940 rcu_read_lock_sched();
7941 dl_b
= dl_bw_of(cpu
);
7943 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7944 if (new_bw
< dl_b
->total_bw
)
7946 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7948 rcu_read_unlock_sched();
7957 static void sched_dl_do_global(void)
7962 unsigned long flags
;
7964 def_dl_bandwidth
.dl_period
= global_rt_period();
7965 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7967 if (global_rt_runtime() != RUNTIME_INF
)
7968 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7971 * FIXME: As above...
7973 for_each_possible_cpu(cpu
) {
7974 rcu_read_lock_sched();
7975 dl_b
= dl_bw_of(cpu
);
7977 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7979 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7981 rcu_read_unlock_sched();
7985 static int sched_rt_global_validate(void)
7987 if (sysctl_sched_rt_period
<= 0)
7990 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7991 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7997 static void sched_rt_do_global(void)
7999 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8000 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8003 int sched_rt_handler(struct ctl_table
*table
, int write
,
8004 void __user
*buffer
, size_t *lenp
,
8007 int old_period
, old_runtime
;
8008 static DEFINE_MUTEX(mutex
);
8012 old_period
= sysctl_sched_rt_period
;
8013 old_runtime
= sysctl_sched_rt_runtime
;
8015 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8017 if (!ret
&& write
) {
8018 ret
= sched_rt_global_validate();
8022 ret
= sched_dl_global_validate();
8026 ret
= sched_rt_global_constraints();
8030 sched_rt_do_global();
8031 sched_dl_do_global();
8035 sysctl_sched_rt_period
= old_period
;
8036 sysctl_sched_rt_runtime
= old_runtime
;
8038 mutex_unlock(&mutex
);
8043 int sched_rr_handler(struct ctl_table
*table
, int write
,
8044 void __user
*buffer
, size_t *lenp
,
8048 static DEFINE_MUTEX(mutex
);
8051 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8052 /* make sure that internally we keep jiffies */
8053 /* also, writing zero resets timeslice to default */
8054 if (!ret
&& write
) {
8055 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8056 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8058 mutex_unlock(&mutex
);
8062 #ifdef CONFIG_CGROUP_SCHED
8064 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8066 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8069 static struct cgroup_subsys_state
*
8070 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8072 struct task_group
*parent
= css_tg(parent_css
);
8073 struct task_group
*tg
;
8076 /* This is early initialization for the top cgroup */
8077 return &root_task_group
.css
;
8080 tg
= sched_create_group(parent
);
8082 return ERR_PTR(-ENOMEM
);
8084 sched_online_group(tg
, parent
);
8089 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8091 struct task_group
*tg
= css_tg(css
);
8093 sched_offline_group(tg
);
8096 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8098 struct task_group
*tg
= css_tg(css
);
8101 * Relies on the RCU grace period between css_released() and this.
8103 sched_free_group(tg
);
8106 static void cpu_cgroup_fork(struct task_struct
*task
)
8108 sched_move_task(task
);
8111 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8113 struct task_struct
*task
;
8114 struct cgroup_subsys_state
*css
;
8116 cgroup_taskset_for_each(task
, css
, tset
) {
8117 #ifdef CONFIG_RT_GROUP_SCHED
8118 if (!sched_rt_can_attach(css_tg(css
), task
))
8121 /* We don't support RT-tasks being in separate groups */
8122 if (task
->sched_class
!= &fair_sched_class
)
8129 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8131 struct task_struct
*task
;
8132 struct cgroup_subsys_state
*css
;
8134 cgroup_taskset_for_each(task
, css
, tset
)
8135 sched_move_task(task
);
8138 #ifdef CONFIG_FAIR_GROUP_SCHED
8139 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8140 struct cftype
*cftype
, u64 shareval
)
8142 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8145 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8148 struct task_group
*tg
= css_tg(css
);
8150 return (u64
) scale_load_down(tg
->shares
);
8153 #ifdef CONFIG_CFS_BANDWIDTH
8154 static DEFINE_MUTEX(cfs_constraints_mutex
);
8156 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8157 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8159 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8161 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8163 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8164 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8166 if (tg
== &root_task_group
)
8170 * Ensure we have at some amount of bandwidth every period. This is
8171 * to prevent reaching a state of large arrears when throttled via
8172 * entity_tick() resulting in prolonged exit starvation.
8174 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8178 * Likewise, bound things on the otherside by preventing insane quota
8179 * periods. This also allows us to normalize in computing quota
8182 if (period
> max_cfs_quota_period
)
8186 * Prevent race between setting of cfs_rq->runtime_enabled and
8187 * unthrottle_offline_cfs_rqs().
8190 mutex_lock(&cfs_constraints_mutex
);
8191 ret
= __cfs_schedulable(tg
, period
, quota
);
8195 runtime_enabled
= quota
!= RUNTIME_INF
;
8196 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8198 * If we need to toggle cfs_bandwidth_used, off->on must occur
8199 * before making related changes, and on->off must occur afterwards
8201 if (runtime_enabled
&& !runtime_was_enabled
)
8202 cfs_bandwidth_usage_inc();
8203 raw_spin_lock_irq(&cfs_b
->lock
);
8204 cfs_b
->period
= ns_to_ktime(period
);
8205 cfs_b
->quota
= quota
;
8207 __refill_cfs_bandwidth_runtime(cfs_b
);
8208 /* restart the period timer (if active) to handle new period expiry */
8209 if (runtime_enabled
)
8210 start_cfs_bandwidth(cfs_b
);
8211 raw_spin_unlock_irq(&cfs_b
->lock
);
8213 for_each_online_cpu(i
) {
8214 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8215 struct rq
*rq
= cfs_rq
->rq
;
8217 raw_spin_lock_irq(&rq
->lock
);
8218 cfs_rq
->runtime_enabled
= runtime_enabled
;
8219 cfs_rq
->runtime_remaining
= 0;
8221 if (cfs_rq
->throttled
)
8222 unthrottle_cfs_rq(cfs_rq
);
8223 raw_spin_unlock_irq(&rq
->lock
);
8225 if (runtime_was_enabled
&& !runtime_enabled
)
8226 cfs_bandwidth_usage_dec();
8228 mutex_unlock(&cfs_constraints_mutex
);
8234 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8238 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8239 if (cfs_quota_us
< 0)
8240 quota
= RUNTIME_INF
;
8242 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8244 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8247 long tg_get_cfs_quota(struct task_group
*tg
)
8251 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8254 quota_us
= tg
->cfs_bandwidth
.quota
;
8255 do_div(quota_us
, NSEC_PER_USEC
);
8260 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8264 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8265 quota
= tg
->cfs_bandwidth
.quota
;
8267 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8270 long tg_get_cfs_period(struct task_group
*tg
)
8274 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8275 do_div(cfs_period_us
, NSEC_PER_USEC
);
8277 return cfs_period_us
;
8280 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8283 return tg_get_cfs_quota(css_tg(css
));
8286 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8287 struct cftype
*cftype
, s64 cfs_quota_us
)
8289 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8292 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8295 return tg_get_cfs_period(css_tg(css
));
8298 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8299 struct cftype
*cftype
, u64 cfs_period_us
)
8301 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8304 struct cfs_schedulable_data
{
8305 struct task_group
*tg
;
8310 * normalize group quota/period to be quota/max_period
8311 * note: units are usecs
8313 static u64
normalize_cfs_quota(struct task_group
*tg
,
8314 struct cfs_schedulable_data
*d
)
8322 period
= tg_get_cfs_period(tg
);
8323 quota
= tg_get_cfs_quota(tg
);
8326 /* note: these should typically be equivalent */
8327 if (quota
== RUNTIME_INF
|| quota
== -1)
8330 return to_ratio(period
, quota
);
8333 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8335 struct cfs_schedulable_data
*d
= data
;
8336 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8337 s64 quota
= 0, parent_quota
= -1;
8340 quota
= RUNTIME_INF
;
8342 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8344 quota
= normalize_cfs_quota(tg
, d
);
8345 parent_quota
= parent_b
->hierarchical_quota
;
8348 * ensure max(child_quota) <= parent_quota, inherit when no
8351 if (quota
== RUNTIME_INF
)
8352 quota
= parent_quota
;
8353 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8356 cfs_b
->hierarchical_quota
= quota
;
8361 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8364 struct cfs_schedulable_data data
= {
8370 if (quota
!= RUNTIME_INF
) {
8371 do_div(data
.period
, NSEC_PER_USEC
);
8372 do_div(data
.quota
, NSEC_PER_USEC
);
8376 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8382 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8384 struct task_group
*tg
= css_tg(seq_css(sf
));
8385 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8387 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8388 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8389 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8393 #endif /* CONFIG_CFS_BANDWIDTH */
8394 #endif /* CONFIG_FAIR_GROUP_SCHED */
8396 #ifdef CONFIG_RT_GROUP_SCHED
8397 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8398 struct cftype
*cft
, s64 val
)
8400 return sched_group_set_rt_runtime(css_tg(css
), val
);
8403 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8406 return sched_group_rt_runtime(css_tg(css
));
8409 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8410 struct cftype
*cftype
, u64 rt_period_us
)
8412 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8415 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8418 return sched_group_rt_period(css_tg(css
));
8420 #endif /* CONFIG_RT_GROUP_SCHED */
8422 static struct cftype cpu_files
[] = {
8423 #ifdef CONFIG_FAIR_GROUP_SCHED
8426 .read_u64
= cpu_shares_read_u64
,
8427 .write_u64
= cpu_shares_write_u64
,
8430 #ifdef CONFIG_CFS_BANDWIDTH
8432 .name
= "cfs_quota_us",
8433 .read_s64
= cpu_cfs_quota_read_s64
,
8434 .write_s64
= cpu_cfs_quota_write_s64
,
8437 .name
= "cfs_period_us",
8438 .read_u64
= cpu_cfs_period_read_u64
,
8439 .write_u64
= cpu_cfs_period_write_u64
,
8443 .seq_show
= cpu_stats_show
,
8446 #ifdef CONFIG_RT_GROUP_SCHED
8448 .name
= "rt_runtime_us",
8449 .read_s64
= cpu_rt_runtime_read
,
8450 .write_s64
= cpu_rt_runtime_write
,
8453 .name
= "rt_period_us",
8454 .read_u64
= cpu_rt_period_read_uint
,
8455 .write_u64
= cpu_rt_period_write_uint
,
8461 struct cgroup_subsys cpu_cgrp_subsys
= {
8462 .css_alloc
= cpu_cgroup_css_alloc
,
8463 .css_released
= cpu_cgroup_css_released
,
8464 .css_free
= cpu_cgroup_css_free
,
8465 .fork
= cpu_cgroup_fork
,
8466 .can_attach
= cpu_cgroup_can_attach
,
8467 .attach
= cpu_cgroup_attach
,
8468 .legacy_cftypes
= cpu_files
,
8472 #endif /* CONFIG_CGROUP_SCHED */
8474 void dump_cpu_task(int cpu
)
8476 pr_info("Task dump for CPU %d:\n", cpu
);
8477 sched_show_task(cpu_curr(cpu
));
8481 * Nice levels are multiplicative, with a gentle 10% change for every
8482 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8483 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8484 * that remained on nice 0.
8486 * The "10% effect" is relative and cumulative: from _any_ nice level,
8487 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8488 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8489 * If a task goes up by ~10% and another task goes down by ~10% then
8490 * the relative distance between them is ~25%.)
8492 const int sched_prio_to_weight
[40] = {
8493 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8494 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8495 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8496 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8497 /* 0 */ 1024, 820, 655, 526, 423,
8498 /* 5 */ 335, 272, 215, 172, 137,
8499 /* 10 */ 110, 87, 70, 56, 45,
8500 /* 15 */ 36, 29, 23, 18, 15,
8504 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8506 * In cases where the weight does not change often, we can use the
8507 * precalculated inverse to speed up arithmetics by turning divisions
8508 * into multiplications:
8510 const u32 sched_prio_to_wmult
[40] = {
8511 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8512 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8513 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8514 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8515 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8516 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8517 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8518 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,