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 */
324 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
326 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
327 * this avoids any races wrt polling state changes and thereby avoids
330 static bool set_nr_and_not_polling(struct task_struct
*p
)
332 struct thread_info
*ti
= task_thread_info(p
);
333 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
337 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
339 * If this returns true, then the idle task promises to call
340 * sched_ttwu_pending() and reschedule soon.
342 static bool set_nr_if_polling(struct task_struct
*p
)
344 struct thread_info
*ti
= task_thread_info(p
);
345 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
348 if (!(val
& _TIF_POLLING_NRFLAG
))
350 if (val
& _TIF_NEED_RESCHED
)
352 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
361 static bool set_nr_and_not_polling(struct task_struct
*p
)
363 set_tsk_need_resched(p
);
368 static bool set_nr_if_polling(struct task_struct
*p
)
375 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
377 struct wake_q_node
*node
= &task
->wake_q
;
380 * Atomically grab the task, if ->wake_q is !nil already it means
381 * its already queued (either by us or someone else) and will get the
382 * wakeup due to that.
384 * This cmpxchg() implies a full barrier, which pairs with the write
385 * barrier implied by the wakeup in wake_up_list().
387 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
390 get_task_struct(task
);
393 * The head is context local, there can be no concurrency.
396 head
->lastp
= &node
->next
;
399 void wake_up_q(struct wake_q_head
*head
)
401 struct wake_q_node
*node
= head
->first
;
403 while (node
!= WAKE_Q_TAIL
) {
404 struct task_struct
*task
;
406 task
= container_of(node
, struct task_struct
, wake_q
);
408 /* task can safely be re-inserted now */
410 task
->wake_q
.next
= NULL
;
413 * wake_up_process() implies a wmb() to pair with the queueing
414 * in wake_q_add() so as not to miss wakeups.
416 wake_up_process(task
);
417 put_task_struct(task
);
422 * resched_curr - mark rq's current task 'to be rescheduled now'.
424 * On UP this means the setting of the need_resched flag, on SMP it
425 * might also involve a cross-CPU call to trigger the scheduler on
428 void resched_curr(struct rq
*rq
)
430 struct task_struct
*curr
= rq
->curr
;
433 lockdep_assert_held(&rq
->lock
);
435 if (test_tsk_need_resched(curr
))
440 if (cpu
== smp_processor_id()) {
441 set_tsk_need_resched(curr
);
442 set_preempt_need_resched();
446 if (set_nr_and_not_polling(curr
))
447 smp_send_reschedule(cpu
);
449 trace_sched_wake_idle_without_ipi(cpu
);
452 void resched_cpu(int cpu
)
454 struct rq
*rq
= cpu_rq(cpu
);
457 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
460 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
464 #ifdef CONFIG_NO_HZ_COMMON
466 * In the semi idle case, use the nearest busy cpu for migrating timers
467 * from an idle cpu. This is good for power-savings.
469 * We don't do similar optimization for completely idle system, as
470 * selecting an idle cpu will add more delays to the timers than intended
471 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
473 int get_nohz_timer_target(void)
475 int i
, cpu
= smp_processor_id();
476 struct sched_domain
*sd
;
478 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
482 for_each_domain(cpu
, sd
) {
483 for_each_cpu(i
, sched_domain_span(sd
)) {
484 if (!idle_cpu(i
) && is_housekeeping_cpu(cpu
)) {
491 if (!is_housekeeping_cpu(cpu
))
492 cpu
= housekeeping_any_cpu();
498 * When add_timer_on() enqueues a timer into the timer wheel of an
499 * idle CPU then this timer might expire before the next timer event
500 * which is scheduled to wake up that CPU. In case of a completely
501 * idle system the next event might even be infinite time into the
502 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
503 * leaves the inner idle loop so the newly added timer is taken into
504 * account when the CPU goes back to idle and evaluates the timer
505 * wheel for the next timer event.
507 static void wake_up_idle_cpu(int cpu
)
509 struct rq
*rq
= cpu_rq(cpu
);
511 if (cpu
== smp_processor_id())
514 if (set_nr_and_not_polling(rq
->idle
))
515 smp_send_reschedule(cpu
);
517 trace_sched_wake_idle_without_ipi(cpu
);
520 static bool wake_up_full_nohz_cpu(int cpu
)
523 * We just need the target to call irq_exit() and re-evaluate
524 * the next tick. The nohz full kick at least implies that.
525 * If needed we can still optimize that later with an
528 if (tick_nohz_full_cpu(cpu
)) {
529 if (cpu
!= smp_processor_id() ||
530 tick_nohz_tick_stopped())
531 tick_nohz_full_kick_cpu(cpu
);
538 void wake_up_nohz_cpu(int cpu
)
540 if (!wake_up_full_nohz_cpu(cpu
))
541 wake_up_idle_cpu(cpu
);
544 static inline bool got_nohz_idle_kick(void)
546 int cpu
= smp_processor_id();
548 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
551 if (idle_cpu(cpu
) && !need_resched())
555 * We can't run Idle Load Balance on this CPU for this time so we
556 * cancel it and clear NOHZ_BALANCE_KICK
558 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
562 #else /* CONFIG_NO_HZ_COMMON */
564 static inline bool got_nohz_idle_kick(void)
569 #endif /* CONFIG_NO_HZ_COMMON */
571 #ifdef CONFIG_NO_HZ_FULL
572 bool sched_can_stop_tick(struct rq
*rq
)
576 /* Deadline tasks, even if single, need the tick */
577 if (rq
->dl
.dl_nr_running
)
581 * FIFO realtime policy runs the highest priority task (after DEADLINE).
582 * Other runnable tasks are of a lower priority. The scheduler tick
585 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
590 * Round-robin realtime tasks time slice with other tasks at the same
593 if (rq
->rt
.rr_nr_running
) {
594 if (rq
->rt
.rr_nr_running
== 1)
600 /* Normal multitasking need periodic preemption checks */
601 if (rq
->cfs
.nr_running
> 1)
606 #endif /* CONFIG_NO_HZ_FULL */
608 void sched_avg_update(struct rq
*rq
)
610 s64 period
= sched_avg_period();
612 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
614 * Inline assembly required to prevent the compiler
615 * optimising this loop into a divmod call.
616 * See __iter_div_u64_rem() for another example of this.
618 asm("" : "+rm" (rq
->age_stamp
));
619 rq
->age_stamp
+= period
;
624 #endif /* CONFIG_SMP */
626 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
627 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
629 * Iterate task_group tree rooted at *from, calling @down when first entering a
630 * node and @up when leaving it for the final time.
632 * Caller must hold rcu_lock or sufficient equivalent.
634 int walk_tg_tree_from(struct task_group
*from
,
635 tg_visitor down
, tg_visitor up
, void *data
)
637 struct task_group
*parent
, *child
;
643 ret
= (*down
)(parent
, data
);
646 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
653 ret
= (*up
)(parent
, data
);
654 if (ret
|| parent
== from
)
658 parent
= parent
->parent
;
665 int tg_nop(struct task_group
*tg
, void *data
)
671 static void set_load_weight(struct task_struct
*p
)
673 int prio
= p
->static_prio
- MAX_RT_PRIO
;
674 struct load_weight
*load
= &p
->se
.load
;
677 * SCHED_IDLE tasks get minimal weight:
679 if (idle_policy(p
->policy
)) {
680 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
681 load
->inv_weight
= WMULT_IDLEPRIO
;
685 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
686 load
->inv_weight
= sched_prio_to_wmult
[prio
];
689 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
692 if (!(flags
& ENQUEUE_RESTORE
))
693 sched_info_queued(rq
, p
);
694 p
->sched_class
->enqueue_task(rq
, p
, flags
);
697 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
700 if (!(flags
& DEQUEUE_SAVE
))
701 sched_info_dequeued(rq
, p
);
702 p
->sched_class
->dequeue_task(rq
, p
, flags
);
705 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
707 if (task_contributes_to_load(p
))
708 rq
->nr_uninterruptible
--;
710 enqueue_task(rq
, p
, flags
);
713 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
715 if (task_contributes_to_load(p
))
716 rq
->nr_uninterruptible
++;
718 dequeue_task(rq
, p
, flags
);
721 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
724 * In theory, the compile should just see 0 here, and optimize out the call
725 * to sched_rt_avg_update. But I don't trust it...
727 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
728 s64 steal
= 0, irq_delta
= 0;
730 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
731 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
734 * Since irq_time is only updated on {soft,}irq_exit, we might run into
735 * this case when a previous update_rq_clock() happened inside a
738 * When this happens, we stop ->clock_task and only update the
739 * prev_irq_time stamp to account for the part that fit, so that a next
740 * update will consume the rest. This ensures ->clock_task is
743 * It does however cause some slight miss-attribution of {soft,}irq
744 * time, a more accurate solution would be to update the irq_time using
745 * the current rq->clock timestamp, except that would require using
748 if (irq_delta
> delta
)
751 rq
->prev_irq_time
+= irq_delta
;
754 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
755 if (static_key_false((¶virt_steal_rq_enabled
))) {
756 steal
= paravirt_steal_clock(cpu_of(rq
));
757 steal
-= rq
->prev_steal_time_rq
;
759 if (unlikely(steal
> delta
))
762 rq
->prev_steal_time_rq
+= steal
;
767 rq
->clock_task
+= delta
;
769 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
770 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
771 sched_rt_avg_update(rq
, irq_delta
+ steal
);
775 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
777 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
778 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
782 * Make it appear like a SCHED_FIFO task, its something
783 * userspace knows about and won't get confused about.
785 * Also, it will make PI more or less work without too
786 * much confusion -- but then, stop work should not
787 * rely on PI working anyway.
789 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
791 stop
->sched_class
= &stop_sched_class
;
794 cpu_rq(cpu
)->stop
= stop
;
798 * Reset it back to a normal scheduling class so that
799 * it can die in pieces.
801 old_stop
->sched_class
= &rt_sched_class
;
806 * __normal_prio - return the priority that is based on the static prio
808 static inline int __normal_prio(struct task_struct
*p
)
810 return p
->static_prio
;
814 * Calculate the expected normal priority: i.e. priority
815 * without taking RT-inheritance into account. Might be
816 * boosted by interactivity modifiers. Changes upon fork,
817 * setprio syscalls, and whenever the interactivity
818 * estimator recalculates.
820 static inline int normal_prio(struct task_struct
*p
)
824 if (task_has_dl_policy(p
))
825 prio
= MAX_DL_PRIO
-1;
826 else if (task_has_rt_policy(p
))
827 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
829 prio
= __normal_prio(p
);
834 * Calculate the current priority, i.e. the priority
835 * taken into account by the scheduler. This value might
836 * be boosted by RT tasks, or might be boosted by
837 * interactivity modifiers. Will be RT if the task got
838 * RT-boosted. If not then it returns p->normal_prio.
840 static int effective_prio(struct task_struct
*p
)
842 p
->normal_prio
= normal_prio(p
);
844 * If we are RT tasks or we were boosted to RT priority,
845 * keep the priority unchanged. Otherwise, update priority
846 * to the normal priority:
848 if (!rt_prio(p
->prio
))
849 return p
->normal_prio
;
854 * task_curr - is this task currently executing on a CPU?
855 * @p: the task in question.
857 * Return: 1 if the task is currently executing. 0 otherwise.
859 inline int task_curr(const struct task_struct
*p
)
861 return cpu_curr(task_cpu(p
)) == p
;
865 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
866 * use the balance_callback list if you want balancing.
868 * this means any call to check_class_changed() must be followed by a call to
869 * balance_callback().
871 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
872 const struct sched_class
*prev_class
,
875 if (prev_class
!= p
->sched_class
) {
876 if (prev_class
->switched_from
)
877 prev_class
->switched_from(rq
, p
);
879 p
->sched_class
->switched_to(rq
, p
);
880 } else if (oldprio
!= p
->prio
|| dl_task(p
))
881 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
884 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
886 const struct sched_class
*class;
888 if (p
->sched_class
== rq
->curr
->sched_class
) {
889 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
891 for_each_class(class) {
892 if (class == rq
->curr
->sched_class
)
894 if (class == p
->sched_class
) {
902 * A queue event has occurred, and we're going to schedule. In
903 * this case, we can save a useless back to back clock update.
905 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
906 rq_clock_skip_update(rq
, true);
911 * This is how migration works:
913 * 1) we invoke migration_cpu_stop() on the target CPU using
915 * 2) stopper starts to run (implicitly forcing the migrated thread
917 * 3) it checks whether the migrated task is still in the wrong runqueue.
918 * 4) if it's in the wrong runqueue then the migration thread removes
919 * it and puts it into the right queue.
920 * 5) stopper completes and stop_one_cpu() returns and the migration
925 * move_queued_task - move a queued task to new rq.
927 * Returns (locked) new rq. Old rq's lock is released.
929 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
931 lockdep_assert_held(&rq
->lock
);
933 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
934 dequeue_task(rq
, p
, 0);
935 set_task_cpu(p
, new_cpu
);
936 raw_spin_unlock(&rq
->lock
);
938 rq
= cpu_rq(new_cpu
);
940 raw_spin_lock(&rq
->lock
);
941 BUG_ON(task_cpu(p
) != new_cpu
);
942 enqueue_task(rq
, p
, 0);
943 p
->on_rq
= TASK_ON_RQ_QUEUED
;
944 check_preempt_curr(rq
, p
, 0);
949 struct migration_arg
{
950 struct task_struct
*task
;
955 * Move (not current) task off this cpu, onto dest cpu. We're doing
956 * this because either it can't run here any more (set_cpus_allowed()
957 * away from this CPU, or CPU going down), or because we're
958 * attempting to rebalance this task on exec (sched_exec).
960 * So we race with normal scheduler movements, but that's OK, as long
961 * as the task is no longer on this CPU.
963 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
965 if (unlikely(!cpu_active(dest_cpu
)))
968 /* Affinity changed (again). */
969 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
972 rq
= move_queued_task(rq
, p
, dest_cpu
);
978 * migration_cpu_stop - this will be executed by a highprio stopper thread
979 * and performs thread migration by bumping thread off CPU then
980 * 'pushing' onto another runqueue.
982 static int migration_cpu_stop(void *data
)
984 struct migration_arg
*arg
= data
;
985 struct task_struct
*p
= arg
->task
;
986 struct rq
*rq
= this_rq();
989 * The original target cpu might have gone down and we might
990 * be on another cpu but it doesn't matter.
994 * We need to explicitly wake pending tasks before running
995 * __migrate_task() such that we will not miss enforcing cpus_allowed
996 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
998 sched_ttwu_pending();
1000 raw_spin_lock(&p
->pi_lock
);
1001 raw_spin_lock(&rq
->lock
);
1003 * If task_rq(p) != rq, it cannot be migrated here, because we're
1004 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1005 * we're holding p->pi_lock.
1007 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1008 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1009 raw_spin_unlock(&rq
->lock
);
1010 raw_spin_unlock(&p
->pi_lock
);
1017 * sched_class::set_cpus_allowed must do the below, but is not required to
1018 * actually call this function.
1020 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1022 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1023 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1026 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1028 struct rq
*rq
= task_rq(p
);
1029 bool queued
, running
;
1031 lockdep_assert_held(&p
->pi_lock
);
1033 queued
= task_on_rq_queued(p
);
1034 running
= task_current(rq
, p
);
1038 * Because __kthread_bind() calls this on blocked tasks without
1041 lockdep_assert_held(&rq
->lock
);
1042 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1045 put_prev_task(rq
, p
);
1047 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1050 p
->sched_class
->set_curr_task(rq
);
1052 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1056 * Change a given task's CPU affinity. Migrate the thread to a
1057 * proper CPU and schedule it away if the CPU it's executing on
1058 * is removed from the allowed bitmask.
1060 * NOTE: the caller must have a valid reference to the task, the
1061 * task must not exit() & deallocate itself prematurely. The
1062 * call is not atomic; no spinlocks may be held.
1064 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1065 const struct cpumask
*new_mask
, bool check
)
1067 unsigned long flags
;
1069 unsigned int dest_cpu
;
1072 rq
= task_rq_lock(p
, &flags
);
1075 * Must re-check here, to close a race against __kthread_bind(),
1076 * sched_setaffinity() is not guaranteed to observe the flag.
1078 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1083 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1086 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1091 do_set_cpus_allowed(p
, new_mask
);
1093 /* Can the task run on the task's current CPU? If so, we're done */
1094 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1097 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1098 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1099 struct migration_arg arg
= { p
, dest_cpu
};
1100 /* Need help from migration thread: drop lock and wait. */
1101 task_rq_unlock(rq
, p
, &flags
);
1102 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1103 tlb_migrate_finish(p
->mm
);
1105 } else if (task_on_rq_queued(p
)) {
1107 * OK, since we're going to drop the lock immediately
1108 * afterwards anyway.
1110 lockdep_unpin_lock(&rq
->lock
);
1111 rq
= move_queued_task(rq
, p
, dest_cpu
);
1112 lockdep_pin_lock(&rq
->lock
);
1115 task_rq_unlock(rq
, p
, &flags
);
1120 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1122 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1124 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1126 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1128 #ifdef CONFIG_SCHED_DEBUG
1130 * We should never call set_task_cpu() on a blocked task,
1131 * ttwu() will sort out the placement.
1133 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1137 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1138 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1139 * time relying on p->on_rq.
1141 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1142 p
->sched_class
== &fair_sched_class
&&
1143 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1145 #ifdef CONFIG_LOCKDEP
1147 * The caller should hold either p->pi_lock or rq->lock, when changing
1148 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1150 * sched_move_task() holds both and thus holding either pins the cgroup,
1153 * Furthermore, all task_rq users should acquire both locks, see
1156 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1157 lockdep_is_held(&task_rq(p
)->lock
)));
1161 trace_sched_migrate_task(p
, new_cpu
);
1163 if (task_cpu(p
) != new_cpu
) {
1164 if (p
->sched_class
->migrate_task_rq
)
1165 p
->sched_class
->migrate_task_rq(p
);
1166 p
->se
.nr_migrations
++;
1167 perf_event_task_migrate(p
);
1170 __set_task_cpu(p
, new_cpu
);
1173 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1175 if (task_on_rq_queued(p
)) {
1176 struct rq
*src_rq
, *dst_rq
;
1178 src_rq
= task_rq(p
);
1179 dst_rq
= cpu_rq(cpu
);
1181 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1182 deactivate_task(src_rq
, p
, 0);
1183 set_task_cpu(p
, cpu
);
1184 activate_task(dst_rq
, p
, 0);
1185 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1186 check_preempt_curr(dst_rq
, p
, 0);
1189 * Task isn't running anymore; make it appear like we migrated
1190 * it before it went to sleep. This means on wakeup we make the
1191 * previous cpu our targer instead of where it really is.
1197 struct migration_swap_arg
{
1198 struct task_struct
*src_task
, *dst_task
;
1199 int src_cpu
, dst_cpu
;
1202 static int migrate_swap_stop(void *data
)
1204 struct migration_swap_arg
*arg
= data
;
1205 struct rq
*src_rq
, *dst_rq
;
1208 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1211 src_rq
= cpu_rq(arg
->src_cpu
);
1212 dst_rq
= cpu_rq(arg
->dst_cpu
);
1214 double_raw_lock(&arg
->src_task
->pi_lock
,
1215 &arg
->dst_task
->pi_lock
);
1216 double_rq_lock(src_rq
, dst_rq
);
1218 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1221 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1224 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1227 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1230 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1231 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1236 double_rq_unlock(src_rq
, dst_rq
);
1237 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1238 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1244 * Cross migrate two tasks
1246 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1248 struct migration_swap_arg arg
;
1251 arg
= (struct migration_swap_arg
){
1253 .src_cpu
= task_cpu(cur
),
1255 .dst_cpu
= task_cpu(p
),
1258 if (arg
.src_cpu
== arg
.dst_cpu
)
1262 * These three tests are all lockless; this is OK since all of them
1263 * will be re-checked with proper locks held further down the line.
1265 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1268 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1271 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1274 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1275 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1282 * wait_task_inactive - wait for a thread to unschedule.
1284 * If @match_state is nonzero, it's the @p->state value just checked and
1285 * not expected to change. If it changes, i.e. @p might have woken up,
1286 * then return zero. When we succeed in waiting for @p to be off its CPU,
1287 * we return a positive number (its total switch count). If a second call
1288 * a short while later returns the same number, the caller can be sure that
1289 * @p has remained unscheduled the whole time.
1291 * The caller must ensure that the task *will* unschedule sometime soon,
1292 * else this function might spin for a *long* time. This function can't
1293 * be called with interrupts off, or it may introduce deadlock with
1294 * smp_call_function() if an IPI is sent by the same process we are
1295 * waiting to become inactive.
1297 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1299 unsigned long flags
;
1300 int running
, queued
;
1306 * We do the initial early heuristics without holding
1307 * any task-queue locks at all. We'll only try to get
1308 * the runqueue lock when things look like they will
1314 * If the task is actively running on another CPU
1315 * still, just relax and busy-wait without holding
1318 * NOTE! Since we don't hold any locks, it's not
1319 * even sure that "rq" stays as the right runqueue!
1320 * But we don't care, since "task_running()" will
1321 * return false if the runqueue has changed and p
1322 * is actually now running somewhere else!
1324 while (task_running(rq
, p
)) {
1325 if (match_state
&& unlikely(p
->state
!= match_state
))
1331 * Ok, time to look more closely! We need the rq
1332 * lock now, to be *sure*. If we're wrong, we'll
1333 * just go back and repeat.
1335 rq
= task_rq_lock(p
, &flags
);
1336 trace_sched_wait_task(p
);
1337 running
= task_running(rq
, p
);
1338 queued
= task_on_rq_queued(p
);
1340 if (!match_state
|| p
->state
== match_state
)
1341 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1342 task_rq_unlock(rq
, p
, &flags
);
1345 * If it changed from the expected state, bail out now.
1347 if (unlikely(!ncsw
))
1351 * Was it really running after all now that we
1352 * checked with the proper locks actually held?
1354 * Oops. Go back and try again..
1356 if (unlikely(running
)) {
1362 * It's not enough that it's not actively running,
1363 * it must be off the runqueue _entirely_, and not
1366 * So if it was still runnable (but just not actively
1367 * running right now), it's preempted, and we should
1368 * yield - it could be a while.
1370 if (unlikely(queued
)) {
1371 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1373 set_current_state(TASK_UNINTERRUPTIBLE
);
1374 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1379 * Ahh, all good. It wasn't running, and it wasn't
1380 * runnable, which means that it will never become
1381 * running in the future either. We're all done!
1390 * kick_process - kick a running thread to enter/exit the kernel
1391 * @p: the to-be-kicked thread
1393 * Cause a process which is running on another CPU to enter
1394 * kernel-mode, without any delay. (to get signals handled.)
1396 * NOTE: this function doesn't have to take the runqueue lock,
1397 * because all it wants to ensure is that the remote task enters
1398 * the kernel. If the IPI races and the task has been migrated
1399 * to another CPU then no harm is done and the purpose has been
1402 void kick_process(struct task_struct
*p
)
1408 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1409 smp_send_reschedule(cpu
);
1412 EXPORT_SYMBOL_GPL(kick_process
);
1415 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1417 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1419 int nid
= cpu_to_node(cpu
);
1420 const struct cpumask
*nodemask
= NULL
;
1421 enum { cpuset
, possible
, fail
} state
= cpuset
;
1425 * If the node that the cpu is on has been offlined, cpu_to_node()
1426 * will return -1. There is no cpu on the node, and we should
1427 * select the cpu on the other node.
1430 nodemask
= cpumask_of_node(nid
);
1432 /* Look for allowed, online CPU in same node. */
1433 for_each_cpu(dest_cpu
, nodemask
) {
1434 if (!cpu_online(dest_cpu
))
1436 if (!cpu_active(dest_cpu
))
1438 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1444 /* Any allowed, online CPU? */
1445 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1446 if (!cpu_online(dest_cpu
))
1448 if (!cpu_active(dest_cpu
))
1453 /* No more Mr. Nice Guy. */
1456 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1457 cpuset_cpus_allowed_fallback(p
);
1463 do_set_cpus_allowed(p
, cpu_possible_mask
);
1474 if (state
!= cpuset
) {
1476 * Don't tell them about moving exiting tasks or
1477 * kernel threads (both mm NULL), since they never
1480 if (p
->mm
&& printk_ratelimit()) {
1481 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1482 task_pid_nr(p
), p
->comm
, cpu
);
1490 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1493 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1495 lockdep_assert_held(&p
->pi_lock
);
1497 if (p
->nr_cpus_allowed
> 1)
1498 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1501 * In order not to call set_task_cpu() on a blocking task we need
1502 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1505 * Since this is common to all placement strategies, this lives here.
1507 * [ this allows ->select_task() to simply return task_cpu(p) and
1508 * not worry about this generic constraint ]
1510 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1512 cpu
= select_fallback_rq(task_cpu(p
), p
);
1517 static void update_avg(u64
*avg
, u64 sample
)
1519 s64 diff
= sample
- *avg
;
1525 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1526 const struct cpumask
*new_mask
, bool check
)
1528 return set_cpus_allowed_ptr(p
, new_mask
);
1531 #endif /* CONFIG_SMP */
1534 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1536 #ifdef CONFIG_SCHEDSTATS
1537 struct rq
*rq
= this_rq();
1540 int this_cpu
= smp_processor_id();
1542 if (cpu
== this_cpu
) {
1543 schedstat_inc(rq
, ttwu_local
);
1544 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1546 struct sched_domain
*sd
;
1548 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1550 for_each_domain(this_cpu
, sd
) {
1551 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1552 schedstat_inc(sd
, ttwu_wake_remote
);
1559 if (wake_flags
& WF_MIGRATED
)
1560 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1562 #endif /* CONFIG_SMP */
1564 schedstat_inc(rq
, ttwu_count
);
1565 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1567 if (wake_flags
& WF_SYNC
)
1568 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1570 #endif /* CONFIG_SCHEDSTATS */
1573 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1575 activate_task(rq
, p
, en_flags
);
1576 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1578 /* if a worker is waking up, notify workqueue */
1579 if (p
->flags
& PF_WQ_WORKER
)
1580 wq_worker_waking_up(p
, cpu_of(rq
));
1584 * Mark the task runnable and perform wakeup-preemption.
1587 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1589 check_preempt_curr(rq
, p
, wake_flags
);
1590 p
->state
= TASK_RUNNING
;
1591 trace_sched_wakeup(p
);
1594 if (p
->sched_class
->task_woken
) {
1596 * Our task @p is fully woken up and running; so its safe to
1597 * drop the rq->lock, hereafter rq is only used for statistics.
1599 lockdep_unpin_lock(&rq
->lock
);
1600 p
->sched_class
->task_woken(rq
, p
);
1601 lockdep_pin_lock(&rq
->lock
);
1604 if (rq
->idle_stamp
) {
1605 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1606 u64 max
= 2*rq
->max_idle_balance_cost
;
1608 update_avg(&rq
->avg_idle
, delta
);
1610 if (rq
->avg_idle
> max
)
1619 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1621 lockdep_assert_held(&rq
->lock
);
1624 if (p
->sched_contributes_to_load
)
1625 rq
->nr_uninterruptible
--;
1628 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1629 ttwu_do_wakeup(rq
, p
, wake_flags
);
1633 * Called in case the task @p isn't fully descheduled from its runqueue,
1634 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1635 * since all we need to do is flip p->state to TASK_RUNNING, since
1636 * the task is still ->on_rq.
1638 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1643 rq
= __task_rq_lock(p
);
1644 if (task_on_rq_queued(p
)) {
1645 /* check_preempt_curr() may use rq clock */
1646 update_rq_clock(rq
);
1647 ttwu_do_wakeup(rq
, p
, wake_flags
);
1650 __task_rq_unlock(rq
);
1656 void sched_ttwu_pending(void)
1658 struct rq
*rq
= this_rq();
1659 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1660 struct task_struct
*p
;
1661 unsigned long flags
;
1666 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1667 lockdep_pin_lock(&rq
->lock
);
1670 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1671 llist
= llist_next(llist
);
1672 ttwu_do_activate(rq
, p
, 0);
1675 lockdep_unpin_lock(&rq
->lock
);
1676 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1679 void scheduler_ipi(void)
1682 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1683 * TIF_NEED_RESCHED remotely (for the first time) will also send
1686 preempt_fold_need_resched();
1688 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1692 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1693 * traditionally all their work was done from the interrupt return
1694 * path. Now that we actually do some work, we need to make sure
1697 * Some archs already do call them, luckily irq_enter/exit nest
1700 * Arguably we should visit all archs and update all handlers,
1701 * however a fair share of IPIs are still resched only so this would
1702 * somewhat pessimize the simple resched case.
1705 sched_ttwu_pending();
1708 * Check if someone kicked us for doing the nohz idle load balance.
1710 if (unlikely(got_nohz_idle_kick())) {
1711 this_rq()->idle_balance
= 1;
1712 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1717 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1719 struct rq
*rq
= cpu_rq(cpu
);
1721 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1722 if (!set_nr_if_polling(rq
->idle
))
1723 smp_send_reschedule(cpu
);
1725 trace_sched_wake_idle_without_ipi(cpu
);
1729 void wake_up_if_idle(int cpu
)
1731 struct rq
*rq
= cpu_rq(cpu
);
1732 unsigned long flags
;
1736 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1739 if (set_nr_if_polling(rq
->idle
)) {
1740 trace_sched_wake_idle_without_ipi(cpu
);
1742 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1743 if (is_idle_task(rq
->curr
))
1744 smp_send_reschedule(cpu
);
1745 /* Else cpu is not in idle, do nothing here */
1746 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1753 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1755 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1757 #endif /* CONFIG_SMP */
1759 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1761 struct rq
*rq
= cpu_rq(cpu
);
1763 #if defined(CONFIG_SMP)
1764 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1765 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1766 ttwu_queue_remote(p
, cpu
);
1771 raw_spin_lock(&rq
->lock
);
1772 lockdep_pin_lock(&rq
->lock
);
1773 ttwu_do_activate(rq
, p
, 0);
1774 lockdep_unpin_lock(&rq
->lock
);
1775 raw_spin_unlock(&rq
->lock
);
1779 * Notes on Program-Order guarantees on SMP systems.
1783 * The basic program-order guarantee on SMP systems is that when a task [t]
1784 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1785 * execution on its new cpu [c1].
1787 * For migration (of runnable tasks) this is provided by the following means:
1789 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1790 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1791 * rq(c1)->lock (if not at the same time, then in that order).
1792 * C) LOCK of the rq(c1)->lock scheduling in task
1794 * Transitivity guarantees that B happens after A and C after B.
1795 * Note: we only require RCpc transitivity.
1796 * Note: the cpu doing B need not be c0 or c1
1805 * UNLOCK rq(0)->lock
1807 * LOCK rq(0)->lock // orders against CPU0
1809 * UNLOCK rq(0)->lock
1813 * UNLOCK rq(1)->lock
1815 * LOCK rq(1)->lock // orders against CPU2
1818 * UNLOCK rq(1)->lock
1821 * BLOCKING -- aka. SLEEP + WAKEUP
1823 * For blocking we (obviously) need to provide the same guarantee as for
1824 * migration. However the means are completely different as there is no lock
1825 * chain to provide order. Instead we do:
1827 * 1) smp_store_release(X->on_cpu, 0)
1828 * 2) smp_cond_acquire(!X->on_cpu)
1832 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1834 * LOCK rq(0)->lock LOCK X->pi_lock
1837 * smp_store_release(X->on_cpu, 0);
1839 * smp_cond_acquire(!X->on_cpu);
1845 * X->state = RUNNING
1846 * UNLOCK rq(2)->lock
1848 * LOCK rq(2)->lock // orders against CPU1
1851 * UNLOCK rq(2)->lock
1854 * UNLOCK rq(0)->lock
1857 * However; for wakeups there is a second guarantee we must provide, namely we
1858 * must observe the state that lead to our wakeup. That is, not only must our
1859 * task observe its own prior state, it must also observe the stores prior to
1862 * This means that any means of doing remote wakeups must order the CPU doing
1863 * the wakeup against the CPU the task is going to end up running on. This,
1864 * however, is already required for the regular Program-Order guarantee above,
1865 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1870 * try_to_wake_up - wake up a thread
1871 * @p: the thread to be awakened
1872 * @state: the mask of task states that can be woken
1873 * @wake_flags: wake modifier flags (WF_*)
1875 * Put it on the run-queue if it's not already there. The "current"
1876 * thread is always on the run-queue (except when the actual
1877 * re-schedule is in progress), and as such you're allowed to do
1878 * the simpler "current->state = TASK_RUNNING" to mark yourself
1879 * runnable without the overhead of this.
1881 * Return: %true if @p was woken up, %false if it was already running.
1882 * or @state didn't match @p's state.
1885 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1887 unsigned long flags
;
1888 int cpu
, success
= 0;
1891 * If we are going to wake up a thread waiting for CONDITION we
1892 * need to ensure that CONDITION=1 done by the caller can not be
1893 * reordered with p->state check below. This pairs with mb() in
1894 * set_current_state() the waiting thread does.
1896 smp_mb__before_spinlock();
1897 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1898 if (!(p
->state
& state
))
1901 trace_sched_waking(p
);
1903 success
= 1; /* we're going to change ->state */
1906 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1911 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1912 * possible to, falsely, observe p->on_cpu == 0.
1914 * One must be running (->on_cpu == 1) in order to remove oneself
1915 * from the runqueue.
1917 * [S] ->on_cpu = 1; [L] ->on_rq
1921 * [S] ->on_rq = 0; [L] ->on_cpu
1923 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1924 * from the consecutive calls to schedule(); the first switching to our
1925 * task, the second putting it to sleep.
1930 * If the owning (remote) cpu is still in the middle of schedule() with
1931 * this task as prev, wait until its done referencing the task.
1933 * Pairs with the smp_store_release() in finish_lock_switch().
1935 * This ensures that tasks getting woken will be fully ordered against
1936 * their previous state and preserve Program Order.
1938 smp_cond_acquire(!p
->on_cpu
);
1940 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1941 p
->state
= TASK_WAKING
;
1943 if (p
->sched_class
->task_waking
)
1944 p
->sched_class
->task_waking(p
);
1946 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1947 if (task_cpu(p
) != cpu
) {
1948 wake_flags
|= WF_MIGRATED
;
1949 set_task_cpu(p
, cpu
);
1951 #endif /* CONFIG_SMP */
1955 if (schedstat_enabled())
1956 ttwu_stat(p
, cpu
, wake_flags
);
1958 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1964 * try_to_wake_up_local - try to wake up a local task with rq lock held
1965 * @p: the thread to be awakened
1967 * Put @p on the run-queue if it's not already there. The caller must
1968 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1971 static void try_to_wake_up_local(struct task_struct
*p
)
1973 struct rq
*rq
= task_rq(p
);
1975 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1976 WARN_ON_ONCE(p
== current
))
1979 lockdep_assert_held(&rq
->lock
);
1981 if (!raw_spin_trylock(&p
->pi_lock
)) {
1983 * This is OK, because current is on_cpu, which avoids it being
1984 * picked for load-balance and preemption/IRQs are still
1985 * disabled avoiding further scheduler activity on it and we've
1986 * not yet picked a replacement task.
1988 lockdep_unpin_lock(&rq
->lock
);
1989 raw_spin_unlock(&rq
->lock
);
1990 raw_spin_lock(&p
->pi_lock
);
1991 raw_spin_lock(&rq
->lock
);
1992 lockdep_pin_lock(&rq
->lock
);
1995 if (!(p
->state
& TASK_NORMAL
))
1998 trace_sched_waking(p
);
2000 if (!task_on_rq_queued(p
))
2001 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2003 ttwu_do_wakeup(rq
, p
, 0);
2004 if (schedstat_enabled())
2005 ttwu_stat(p
, smp_processor_id(), 0);
2007 raw_spin_unlock(&p
->pi_lock
);
2011 * wake_up_process - Wake up a specific process
2012 * @p: The process to be woken up.
2014 * Attempt to wake up the nominated process and move it to the set of runnable
2017 * Return: 1 if the process was woken up, 0 if it was already running.
2019 * It may be assumed that this function implies a write memory barrier before
2020 * changing the task state if and only if any tasks are woken up.
2022 int wake_up_process(struct task_struct
*p
)
2024 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2026 EXPORT_SYMBOL(wake_up_process
);
2028 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2030 return try_to_wake_up(p
, state
, 0);
2034 * This function clears the sched_dl_entity static params.
2036 void __dl_clear_params(struct task_struct
*p
)
2038 struct sched_dl_entity
*dl_se
= &p
->dl
;
2040 dl_se
->dl_runtime
= 0;
2041 dl_se
->dl_deadline
= 0;
2042 dl_se
->dl_period
= 0;
2046 dl_se
->dl_throttled
= 0;
2047 dl_se
->dl_yielded
= 0;
2051 * Perform scheduler related setup for a newly forked process p.
2052 * p is forked by current.
2054 * __sched_fork() is basic setup used by init_idle() too:
2056 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2061 p
->se
.exec_start
= 0;
2062 p
->se
.sum_exec_runtime
= 0;
2063 p
->se
.prev_sum_exec_runtime
= 0;
2064 p
->se
.nr_migrations
= 0;
2066 INIT_LIST_HEAD(&p
->se
.group_node
);
2068 #ifdef CONFIG_FAIR_GROUP_SCHED
2069 p
->se
.cfs_rq
= NULL
;
2072 #ifdef CONFIG_SCHEDSTATS
2073 /* Even if schedstat is disabled, there should not be garbage */
2074 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2077 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2078 init_dl_task_timer(&p
->dl
);
2079 __dl_clear_params(p
);
2081 INIT_LIST_HEAD(&p
->rt
.run_list
);
2083 p
->rt
.time_slice
= sched_rr_timeslice
;
2087 #ifdef CONFIG_PREEMPT_NOTIFIERS
2088 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2091 #ifdef CONFIG_NUMA_BALANCING
2092 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2093 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2094 p
->mm
->numa_scan_seq
= 0;
2097 if (clone_flags
& CLONE_VM
)
2098 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2100 p
->numa_preferred_nid
= -1;
2102 p
->node_stamp
= 0ULL;
2103 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2104 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2105 p
->numa_work
.next
= &p
->numa_work
;
2106 p
->numa_faults
= NULL
;
2107 p
->last_task_numa_placement
= 0;
2108 p
->last_sum_exec_runtime
= 0;
2110 p
->numa_group
= NULL
;
2111 #endif /* CONFIG_NUMA_BALANCING */
2114 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2116 #ifdef CONFIG_NUMA_BALANCING
2118 void set_numabalancing_state(bool enabled
)
2121 static_branch_enable(&sched_numa_balancing
);
2123 static_branch_disable(&sched_numa_balancing
);
2126 #ifdef CONFIG_PROC_SYSCTL
2127 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2128 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2132 int state
= static_branch_likely(&sched_numa_balancing
);
2134 if (write
&& !capable(CAP_SYS_ADMIN
))
2139 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2143 set_numabalancing_state(state
);
2149 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2151 #ifdef CONFIG_SCHEDSTATS
2152 static void set_schedstats(bool enabled
)
2155 static_branch_enable(&sched_schedstats
);
2157 static_branch_disable(&sched_schedstats
);
2160 void force_schedstat_enabled(void)
2162 if (!schedstat_enabled()) {
2163 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2164 static_branch_enable(&sched_schedstats
);
2168 static int __init
setup_schedstats(char *str
)
2174 if (!strcmp(str
, "enable")) {
2175 set_schedstats(true);
2177 } else if (!strcmp(str
, "disable")) {
2178 set_schedstats(false);
2183 pr_warn("Unable to parse schedstats=\n");
2187 __setup("schedstats=", setup_schedstats
);
2189 #ifdef CONFIG_PROC_SYSCTL
2190 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2191 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2195 int state
= static_branch_likely(&sched_schedstats
);
2197 if (write
&& !capable(CAP_SYS_ADMIN
))
2202 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2206 set_schedstats(state
);
2213 * fork()/clone()-time setup:
2215 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2217 unsigned long flags
;
2218 int cpu
= get_cpu();
2220 __sched_fork(clone_flags
, p
);
2222 * We mark the process as running here. This guarantees that
2223 * nobody will actually run it, and a signal or other external
2224 * event cannot wake it up and insert it on the runqueue either.
2226 p
->state
= TASK_RUNNING
;
2229 * Make sure we do not leak PI boosting priority to the child.
2231 p
->prio
= current
->normal_prio
;
2234 * Revert to default priority/policy on fork if requested.
2236 if (unlikely(p
->sched_reset_on_fork
)) {
2237 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2238 p
->policy
= SCHED_NORMAL
;
2239 p
->static_prio
= NICE_TO_PRIO(0);
2241 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2242 p
->static_prio
= NICE_TO_PRIO(0);
2244 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2248 * We don't need the reset flag anymore after the fork. It has
2249 * fulfilled its duty:
2251 p
->sched_reset_on_fork
= 0;
2254 if (dl_prio(p
->prio
)) {
2257 } else if (rt_prio(p
->prio
)) {
2258 p
->sched_class
= &rt_sched_class
;
2260 p
->sched_class
= &fair_sched_class
;
2263 if (p
->sched_class
->task_fork
)
2264 p
->sched_class
->task_fork(p
);
2267 * The child is not yet in the pid-hash so no cgroup attach races,
2268 * and the cgroup is pinned to this child due to cgroup_fork()
2269 * is ran before sched_fork().
2271 * Silence PROVE_RCU.
2273 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2274 set_task_cpu(p
, cpu
);
2275 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2277 #ifdef CONFIG_SCHED_INFO
2278 if (likely(sched_info_on()))
2279 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2281 #if defined(CONFIG_SMP)
2284 init_task_preempt_count(p
);
2286 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2287 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2294 unsigned long to_ratio(u64 period
, u64 runtime
)
2296 if (runtime
== RUNTIME_INF
)
2300 * Doing this here saves a lot of checks in all
2301 * the calling paths, and returning zero seems
2302 * safe for them anyway.
2307 return div64_u64(runtime
<< 20, period
);
2311 inline struct dl_bw
*dl_bw_of(int i
)
2313 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2314 "sched RCU must be held");
2315 return &cpu_rq(i
)->rd
->dl_bw
;
2318 static inline int dl_bw_cpus(int i
)
2320 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2323 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2324 "sched RCU must be held");
2325 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2331 inline struct dl_bw
*dl_bw_of(int i
)
2333 return &cpu_rq(i
)->dl
.dl_bw
;
2336 static inline int dl_bw_cpus(int i
)
2343 * We must be sure that accepting a new task (or allowing changing the
2344 * parameters of an existing one) is consistent with the bandwidth
2345 * constraints. If yes, this function also accordingly updates the currently
2346 * allocated bandwidth to reflect the new situation.
2348 * This function is called while holding p's rq->lock.
2350 * XXX we should delay bw change until the task's 0-lag point, see
2353 static int dl_overflow(struct task_struct
*p
, int policy
,
2354 const struct sched_attr
*attr
)
2357 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2358 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2359 u64 runtime
= attr
->sched_runtime
;
2360 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2363 if (new_bw
== p
->dl
.dl_bw
)
2367 * Either if a task, enters, leave, or stays -deadline but changes
2368 * its parameters, we may need to update accordingly the total
2369 * allocated bandwidth of the container.
2371 raw_spin_lock(&dl_b
->lock
);
2372 cpus
= dl_bw_cpus(task_cpu(p
));
2373 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2374 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2375 __dl_add(dl_b
, new_bw
);
2377 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2378 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2379 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2380 __dl_add(dl_b
, new_bw
);
2382 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2383 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2386 raw_spin_unlock(&dl_b
->lock
);
2391 extern void init_dl_bw(struct dl_bw
*dl_b
);
2394 * wake_up_new_task - wake up a newly created task for the first time.
2396 * This function will do some initial scheduler statistics housekeeping
2397 * that must be done for every newly created context, then puts the task
2398 * on the runqueue and wakes it.
2400 void wake_up_new_task(struct task_struct
*p
)
2402 unsigned long flags
;
2405 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2406 /* Initialize new task's runnable average */
2407 init_entity_runnable_average(&p
->se
);
2410 * Fork balancing, do it here and not earlier because:
2411 * - cpus_allowed can change in the fork path
2412 * - any previously selected cpu might disappear through hotplug
2414 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2417 rq
= __task_rq_lock(p
);
2418 activate_task(rq
, p
, 0);
2419 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2420 trace_sched_wakeup_new(p
);
2421 check_preempt_curr(rq
, p
, WF_FORK
);
2423 if (p
->sched_class
->task_woken
) {
2425 * Nothing relies on rq->lock after this, so its fine to
2428 lockdep_unpin_lock(&rq
->lock
);
2429 p
->sched_class
->task_woken(rq
, p
);
2430 lockdep_pin_lock(&rq
->lock
);
2433 task_rq_unlock(rq
, p
, &flags
);
2436 #ifdef CONFIG_PREEMPT_NOTIFIERS
2438 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2440 void preempt_notifier_inc(void)
2442 static_key_slow_inc(&preempt_notifier_key
);
2444 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2446 void preempt_notifier_dec(void)
2448 static_key_slow_dec(&preempt_notifier_key
);
2450 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2453 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2454 * @notifier: notifier struct to register
2456 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2458 if (!static_key_false(&preempt_notifier_key
))
2459 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2461 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2463 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2466 * preempt_notifier_unregister - no longer interested in preemption notifications
2467 * @notifier: notifier struct to unregister
2469 * This is *not* safe to call from within a preemption notifier.
2471 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2473 hlist_del(¬ifier
->link
);
2475 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2477 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2479 struct preempt_notifier
*notifier
;
2481 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2482 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2485 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2487 if (static_key_false(&preempt_notifier_key
))
2488 __fire_sched_in_preempt_notifiers(curr
);
2492 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2493 struct task_struct
*next
)
2495 struct preempt_notifier
*notifier
;
2497 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2498 notifier
->ops
->sched_out(notifier
, next
);
2501 static __always_inline
void
2502 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2503 struct task_struct
*next
)
2505 if (static_key_false(&preempt_notifier_key
))
2506 __fire_sched_out_preempt_notifiers(curr
, next
);
2509 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2511 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2516 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2517 struct task_struct
*next
)
2521 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2524 * prepare_task_switch - prepare to switch tasks
2525 * @rq: the runqueue preparing to switch
2526 * @prev: the current task that is being switched out
2527 * @next: the task we are going to switch to.
2529 * This is called with the rq lock held and interrupts off. It must
2530 * be paired with a subsequent finish_task_switch after the context
2533 * prepare_task_switch sets up locking and calls architecture specific
2537 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2538 struct task_struct
*next
)
2540 sched_info_switch(rq
, prev
, next
);
2541 perf_event_task_sched_out(prev
, next
);
2542 fire_sched_out_preempt_notifiers(prev
, next
);
2543 prepare_lock_switch(rq
, next
);
2544 prepare_arch_switch(next
);
2548 * finish_task_switch - clean up after a task-switch
2549 * @prev: the thread we just switched away from.
2551 * finish_task_switch must be called after the context switch, paired
2552 * with a prepare_task_switch call before the context switch.
2553 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2554 * and do any other architecture-specific cleanup actions.
2556 * Note that we may have delayed dropping an mm in context_switch(). If
2557 * so, we finish that here outside of the runqueue lock. (Doing it
2558 * with the lock held can cause deadlocks; see schedule() for
2561 * The context switch have flipped the stack from under us and restored the
2562 * local variables which were saved when this task called schedule() in the
2563 * past. prev == current is still correct but we need to recalculate this_rq
2564 * because prev may have moved to another CPU.
2566 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2567 __releases(rq
->lock
)
2569 struct rq
*rq
= this_rq();
2570 struct mm_struct
*mm
= rq
->prev_mm
;
2574 * The previous task will have left us with a preempt_count of 2
2575 * because it left us after:
2578 * preempt_disable(); // 1
2580 * raw_spin_lock_irq(&rq->lock) // 2
2582 * Also, see FORK_PREEMPT_COUNT.
2584 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2585 "corrupted preempt_count: %s/%d/0x%x\n",
2586 current
->comm
, current
->pid
, preempt_count()))
2587 preempt_count_set(FORK_PREEMPT_COUNT
);
2592 * A task struct has one reference for the use as "current".
2593 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2594 * schedule one last time. The schedule call will never return, and
2595 * the scheduled task must drop that reference.
2597 * We must observe prev->state before clearing prev->on_cpu (in
2598 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2599 * running on another CPU and we could rave with its RUNNING -> DEAD
2600 * transition, resulting in a double drop.
2602 prev_state
= prev
->state
;
2603 vtime_task_switch(prev
);
2604 perf_event_task_sched_in(prev
, current
);
2605 finish_lock_switch(rq
, prev
);
2606 finish_arch_post_lock_switch();
2608 fire_sched_in_preempt_notifiers(current
);
2611 if (unlikely(prev_state
== TASK_DEAD
)) {
2612 if (prev
->sched_class
->task_dead
)
2613 prev
->sched_class
->task_dead(prev
);
2616 * Remove function-return probe instances associated with this
2617 * task and put them back on the free list.
2619 kprobe_flush_task(prev
);
2620 put_task_struct(prev
);
2623 tick_nohz_task_switch();
2629 /* rq->lock is NOT held, but preemption is disabled */
2630 static void __balance_callback(struct rq
*rq
)
2632 struct callback_head
*head
, *next
;
2633 void (*func
)(struct rq
*rq
);
2634 unsigned long flags
;
2636 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2637 head
= rq
->balance_callback
;
2638 rq
->balance_callback
= NULL
;
2640 func
= (void (*)(struct rq
*))head
->func
;
2647 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2650 static inline void balance_callback(struct rq
*rq
)
2652 if (unlikely(rq
->balance_callback
))
2653 __balance_callback(rq
);
2658 static inline void balance_callback(struct rq
*rq
)
2665 * schedule_tail - first thing a freshly forked thread must call.
2666 * @prev: the thread we just switched away from.
2668 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2669 __releases(rq
->lock
)
2674 * New tasks start with FORK_PREEMPT_COUNT, see there and
2675 * finish_task_switch() for details.
2677 * finish_task_switch() will drop rq->lock() and lower preempt_count
2678 * and the preempt_enable() will end up enabling preemption (on
2679 * PREEMPT_COUNT kernels).
2682 rq
= finish_task_switch(prev
);
2683 balance_callback(rq
);
2686 if (current
->set_child_tid
)
2687 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2691 * context_switch - switch to the new MM and the new thread's register state.
2693 static __always_inline
struct rq
*
2694 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2695 struct task_struct
*next
)
2697 struct mm_struct
*mm
, *oldmm
;
2699 prepare_task_switch(rq
, prev
, next
);
2702 oldmm
= prev
->active_mm
;
2704 * For paravirt, this is coupled with an exit in switch_to to
2705 * combine the page table reload and the switch backend into
2708 arch_start_context_switch(prev
);
2711 next
->active_mm
= oldmm
;
2712 atomic_inc(&oldmm
->mm_count
);
2713 enter_lazy_tlb(oldmm
, next
);
2715 switch_mm(oldmm
, mm
, next
);
2718 prev
->active_mm
= NULL
;
2719 rq
->prev_mm
= oldmm
;
2722 * Since the runqueue lock will be released by the next
2723 * task (which is an invalid locking op but in the case
2724 * of the scheduler it's an obvious special-case), so we
2725 * do an early lockdep release here:
2727 lockdep_unpin_lock(&rq
->lock
);
2728 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2730 /* Here we just switch the register state and the stack. */
2731 switch_to(prev
, next
, prev
);
2734 return finish_task_switch(prev
);
2738 * nr_running and nr_context_switches:
2740 * externally visible scheduler statistics: current number of runnable
2741 * threads, total number of context switches performed since bootup.
2743 unsigned long nr_running(void)
2745 unsigned long i
, sum
= 0;
2747 for_each_online_cpu(i
)
2748 sum
+= cpu_rq(i
)->nr_running
;
2754 * Check if only the current task is running on the cpu.
2756 * Caution: this function does not check that the caller has disabled
2757 * preemption, thus the result might have a time-of-check-to-time-of-use
2758 * race. The caller is responsible to use it correctly, for example:
2760 * - from a non-preemptable section (of course)
2762 * - from a thread that is bound to a single CPU
2764 * - in a loop with very short iterations (e.g. a polling loop)
2766 bool single_task_running(void)
2768 return raw_rq()->nr_running
== 1;
2770 EXPORT_SYMBOL(single_task_running
);
2772 unsigned long long nr_context_switches(void)
2775 unsigned long long sum
= 0;
2777 for_each_possible_cpu(i
)
2778 sum
+= cpu_rq(i
)->nr_switches
;
2783 unsigned long nr_iowait(void)
2785 unsigned long i
, sum
= 0;
2787 for_each_possible_cpu(i
)
2788 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2793 unsigned long nr_iowait_cpu(int cpu
)
2795 struct rq
*this = cpu_rq(cpu
);
2796 return atomic_read(&this->nr_iowait
);
2799 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2801 struct rq
*rq
= this_rq();
2802 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2803 *load
= rq
->load
.weight
;
2809 * sched_exec - execve() is a valuable balancing opportunity, because at
2810 * this point the task has the smallest effective memory and cache footprint.
2812 void sched_exec(void)
2814 struct task_struct
*p
= current
;
2815 unsigned long flags
;
2818 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2819 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2820 if (dest_cpu
== smp_processor_id())
2823 if (likely(cpu_active(dest_cpu
))) {
2824 struct migration_arg arg
= { p
, dest_cpu
};
2826 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2827 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2831 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2836 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2837 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2839 EXPORT_PER_CPU_SYMBOL(kstat
);
2840 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2843 * Return accounted runtime for the task.
2844 * In case the task is currently running, return the runtime plus current's
2845 * pending runtime that have not been accounted yet.
2847 unsigned long long task_sched_runtime(struct task_struct
*p
)
2849 unsigned long flags
;
2853 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2855 * 64-bit doesn't need locks to atomically read a 64bit value.
2856 * So we have a optimization chance when the task's delta_exec is 0.
2857 * Reading ->on_cpu is racy, but this is ok.
2859 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2860 * If we race with it entering cpu, unaccounted time is 0. This is
2861 * indistinguishable from the read occurring a few cycles earlier.
2862 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2863 * been accounted, so we're correct here as well.
2865 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2866 return p
->se
.sum_exec_runtime
;
2869 rq
= task_rq_lock(p
, &flags
);
2871 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2872 * project cycles that may never be accounted to this
2873 * thread, breaking clock_gettime().
2875 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2876 update_rq_clock(rq
);
2877 p
->sched_class
->update_curr(rq
);
2879 ns
= p
->se
.sum_exec_runtime
;
2880 task_rq_unlock(rq
, p
, &flags
);
2886 * This function gets called by the timer code, with HZ frequency.
2887 * We call it with interrupts disabled.
2889 void scheduler_tick(void)
2891 int cpu
= smp_processor_id();
2892 struct rq
*rq
= cpu_rq(cpu
);
2893 struct task_struct
*curr
= rq
->curr
;
2897 raw_spin_lock(&rq
->lock
);
2898 update_rq_clock(rq
);
2899 curr
->sched_class
->task_tick(rq
, curr
, 0);
2900 update_cpu_load_active(rq
);
2901 calc_global_load_tick(rq
);
2902 raw_spin_unlock(&rq
->lock
);
2904 perf_event_task_tick();
2907 rq
->idle_balance
= idle_cpu(cpu
);
2908 trigger_load_balance(rq
);
2910 rq_last_tick_reset(rq
);
2913 #ifdef CONFIG_NO_HZ_FULL
2915 * scheduler_tick_max_deferment
2917 * Keep at least one tick per second when a single
2918 * active task is running because the scheduler doesn't
2919 * yet completely support full dynticks environment.
2921 * This makes sure that uptime, CFS vruntime, load
2922 * balancing, etc... continue to move forward, even
2923 * with a very low granularity.
2925 * Return: Maximum deferment in nanoseconds.
2927 u64
scheduler_tick_max_deferment(void)
2929 struct rq
*rq
= this_rq();
2930 unsigned long next
, now
= READ_ONCE(jiffies
);
2932 next
= rq
->last_sched_tick
+ HZ
;
2934 if (time_before_eq(next
, now
))
2937 return jiffies_to_nsecs(next
- now
);
2941 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2942 defined(CONFIG_PREEMPT_TRACER))
2944 void preempt_count_add(int val
)
2946 #ifdef CONFIG_DEBUG_PREEMPT
2950 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2953 __preempt_count_add(val
);
2954 #ifdef CONFIG_DEBUG_PREEMPT
2956 * Spinlock count overflowing soon?
2958 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2961 if (preempt_count() == val
) {
2962 unsigned long ip
= get_lock_parent_ip();
2963 #ifdef CONFIG_DEBUG_PREEMPT
2964 current
->preempt_disable_ip
= ip
;
2966 trace_preempt_off(CALLER_ADDR0
, ip
);
2969 EXPORT_SYMBOL(preempt_count_add
);
2970 NOKPROBE_SYMBOL(preempt_count_add
);
2972 void preempt_count_sub(int val
)
2974 #ifdef CONFIG_DEBUG_PREEMPT
2978 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2981 * Is the spinlock portion underflowing?
2983 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2984 !(preempt_count() & PREEMPT_MASK
)))
2988 if (preempt_count() == val
)
2989 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
2990 __preempt_count_sub(val
);
2992 EXPORT_SYMBOL(preempt_count_sub
);
2993 NOKPROBE_SYMBOL(preempt_count_sub
);
2998 * Print scheduling while atomic bug:
3000 static noinline
void __schedule_bug(struct task_struct
*prev
)
3002 if (oops_in_progress
)
3005 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3006 prev
->comm
, prev
->pid
, preempt_count());
3008 debug_show_held_locks(prev
);
3010 if (irqs_disabled())
3011 print_irqtrace_events(prev
);
3012 #ifdef CONFIG_DEBUG_PREEMPT
3013 if (in_atomic_preempt_off()) {
3014 pr_err("Preemption disabled at:");
3015 print_ip_sym(current
->preempt_disable_ip
);
3020 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3024 * Various schedule()-time debugging checks and statistics:
3026 static inline void schedule_debug(struct task_struct
*prev
)
3028 #ifdef CONFIG_SCHED_STACK_END_CHECK
3029 BUG_ON(task_stack_end_corrupted(prev
));
3032 if (unlikely(in_atomic_preempt_off())) {
3033 __schedule_bug(prev
);
3034 preempt_count_set(PREEMPT_DISABLED
);
3038 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3040 schedstat_inc(this_rq(), sched_count
);
3044 * Pick up the highest-prio task:
3046 static inline struct task_struct
*
3047 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3049 const struct sched_class
*class = &fair_sched_class
;
3050 struct task_struct
*p
;
3053 * Optimization: we know that if all tasks are in
3054 * the fair class we can call that function directly:
3056 if (likely(prev
->sched_class
== class &&
3057 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3058 p
= fair_sched_class
.pick_next_task(rq
, prev
);
3059 if (unlikely(p
== RETRY_TASK
))
3062 /* assumes fair_sched_class->next == idle_sched_class */
3064 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3070 for_each_class(class) {
3071 p
= class->pick_next_task(rq
, prev
);
3073 if (unlikely(p
== RETRY_TASK
))
3079 BUG(); /* the idle class will always have a runnable task */
3083 * __schedule() is the main scheduler function.
3085 * The main means of driving the scheduler and thus entering this function are:
3087 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3089 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3090 * paths. For example, see arch/x86/entry_64.S.
3092 * To drive preemption between tasks, the scheduler sets the flag in timer
3093 * interrupt handler scheduler_tick().
3095 * 3. Wakeups don't really cause entry into schedule(). They add a
3096 * task to the run-queue and that's it.
3098 * Now, if the new task added to the run-queue preempts the current
3099 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3100 * called on the nearest possible occasion:
3102 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3104 * - in syscall or exception context, at the next outmost
3105 * preempt_enable(). (this might be as soon as the wake_up()'s
3108 * - in IRQ context, return from interrupt-handler to
3109 * preemptible context
3111 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3114 * - cond_resched() call
3115 * - explicit schedule() call
3116 * - return from syscall or exception to user-space
3117 * - return from interrupt-handler to user-space
3119 * WARNING: must be called with preemption disabled!
3121 static void __sched notrace
__schedule(bool preempt
)
3123 struct task_struct
*prev
, *next
;
3124 unsigned long *switch_count
;
3128 cpu
= smp_processor_id();
3133 * do_exit() calls schedule() with preemption disabled as an exception;
3134 * however we must fix that up, otherwise the next task will see an
3135 * inconsistent (higher) preempt count.
3137 * It also avoids the below schedule_debug() test from complaining
3140 if (unlikely(prev
->state
== TASK_DEAD
))
3141 preempt_enable_no_resched_notrace();
3143 schedule_debug(prev
);
3145 if (sched_feat(HRTICK
))
3148 local_irq_disable();
3149 rcu_note_context_switch();
3152 * Make sure that signal_pending_state()->signal_pending() below
3153 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3154 * done by the caller to avoid the race with signal_wake_up().
3156 smp_mb__before_spinlock();
3157 raw_spin_lock(&rq
->lock
);
3158 lockdep_pin_lock(&rq
->lock
);
3160 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3162 switch_count
= &prev
->nivcsw
;
3163 if (!preempt
&& prev
->state
) {
3164 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3165 prev
->state
= TASK_RUNNING
;
3167 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3171 * If a worker went to sleep, notify and ask workqueue
3172 * whether it wants to wake up a task to maintain
3175 if (prev
->flags
& PF_WQ_WORKER
) {
3176 struct task_struct
*to_wakeup
;
3178 to_wakeup
= wq_worker_sleeping(prev
);
3180 try_to_wake_up_local(to_wakeup
);
3183 switch_count
= &prev
->nvcsw
;
3186 if (task_on_rq_queued(prev
))
3187 update_rq_clock(rq
);
3189 next
= pick_next_task(rq
, prev
);
3190 clear_tsk_need_resched(prev
);
3191 clear_preempt_need_resched();
3192 rq
->clock_skip_update
= 0;
3194 if (likely(prev
!= next
)) {
3199 trace_sched_switch(preempt
, prev
, next
);
3200 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3202 lockdep_unpin_lock(&rq
->lock
);
3203 raw_spin_unlock_irq(&rq
->lock
);
3206 balance_callback(rq
);
3208 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3210 static inline void sched_submit_work(struct task_struct
*tsk
)
3212 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3215 * If we are going to sleep and we have plugged IO queued,
3216 * make sure to submit it to avoid deadlocks.
3218 if (blk_needs_flush_plug(tsk
))
3219 blk_schedule_flush_plug(tsk
);
3222 asmlinkage __visible
void __sched
schedule(void)
3224 struct task_struct
*tsk
= current
;
3226 sched_submit_work(tsk
);
3230 sched_preempt_enable_no_resched();
3231 } while (need_resched());
3233 EXPORT_SYMBOL(schedule
);
3235 #ifdef CONFIG_CONTEXT_TRACKING
3236 asmlinkage __visible
void __sched
schedule_user(void)
3239 * If we come here after a random call to set_need_resched(),
3240 * or we have been woken up remotely but the IPI has not yet arrived,
3241 * we haven't yet exited the RCU idle mode. Do it here manually until
3242 * we find a better solution.
3244 * NB: There are buggy callers of this function. Ideally we
3245 * should warn if prev_state != CONTEXT_USER, but that will trigger
3246 * too frequently to make sense yet.
3248 enum ctx_state prev_state
= exception_enter();
3250 exception_exit(prev_state
);
3255 * schedule_preempt_disabled - called with preemption disabled
3257 * Returns with preemption disabled. Note: preempt_count must be 1
3259 void __sched
schedule_preempt_disabled(void)
3261 sched_preempt_enable_no_resched();
3266 static void __sched notrace
preempt_schedule_common(void)
3269 preempt_disable_notrace();
3271 preempt_enable_no_resched_notrace();
3274 * Check again in case we missed a preemption opportunity
3275 * between schedule and now.
3277 } while (need_resched());
3280 #ifdef CONFIG_PREEMPT
3282 * this is the entry point to schedule() from in-kernel preemption
3283 * off of preempt_enable. Kernel preemptions off return from interrupt
3284 * occur there and call schedule directly.
3286 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3289 * If there is a non-zero preempt_count or interrupts are disabled,
3290 * we do not want to preempt the current task. Just return..
3292 if (likely(!preemptible()))
3295 preempt_schedule_common();
3297 NOKPROBE_SYMBOL(preempt_schedule
);
3298 EXPORT_SYMBOL(preempt_schedule
);
3301 * preempt_schedule_notrace - preempt_schedule called by tracing
3303 * The tracing infrastructure uses preempt_enable_notrace to prevent
3304 * recursion and tracing preempt enabling caused by the tracing
3305 * infrastructure itself. But as tracing can happen in areas coming
3306 * from userspace or just about to enter userspace, a preempt enable
3307 * can occur before user_exit() is called. This will cause the scheduler
3308 * to be called when the system is still in usermode.
3310 * To prevent this, the preempt_enable_notrace will use this function
3311 * instead of preempt_schedule() to exit user context if needed before
3312 * calling the scheduler.
3314 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3316 enum ctx_state prev_ctx
;
3318 if (likely(!preemptible()))
3322 preempt_disable_notrace();
3324 * Needs preempt disabled in case user_exit() is traced
3325 * and the tracer calls preempt_enable_notrace() causing
3326 * an infinite recursion.
3328 prev_ctx
= exception_enter();
3330 exception_exit(prev_ctx
);
3332 preempt_enable_no_resched_notrace();
3333 } while (need_resched());
3335 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3337 #endif /* CONFIG_PREEMPT */
3340 * this is the entry point to schedule() from kernel preemption
3341 * off of irq context.
3342 * Note, that this is called and return with irqs disabled. This will
3343 * protect us against recursive calling from irq.
3345 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3347 enum ctx_state prev_state
;
3349 /* Catch callers which need to be fixed */
3350 BUG_ON(preempt_count() || !irqs_disabled());
3352 prev_state
= exception_enter();
3358 local_irq_disable();
3359 sched_preempt_enable_no_resched();
3360 } while (need_resched());
3362 exception_exit(prev_state
);
3365 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3368 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3370 EXPORT_SYMBOL(default_wake_function
);
3372 #ifdef CONFIG_RT_MUTEXES
3375 * rt_mutex_setprio - set the current priority of a task
3377 * @prio: prio value (kernel-internal form)
3379 * This function changes the 'effective' priority of a task. It does
3380 * not touch ->normal_prio like __setscheduler().
3382 * Used by the rt_mutex code to implement priority inheritance
3383 * logic. Call site only calls if the priority of the task changed.
3385 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3387 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3389 const struct sched_class
*prev_class
;
3391 BUG_ON(prio
> MAX_PRIO
);
3393 rq
= __task_rq_lock(p
);
3396 * Idle task boosting is a nono in general. There is one
3397 * exception, when PREEMPT_RT and NOHZ is active:
3399 * The idle task calls get_next_timer_interrupt() and holds
3400 * the timer wheel base->lock on the CPU and another CPU wants
3401 * to access the timer (probably to cancel it). We can safely
3402 * ignore the boosting request, as the idle CPU runs this code
3403 * with interrupts disabled and will complete the lock
3404 * protected section without being interrupted. So there is no
3405 * real need to boost.
3407 if (unlikely(p
== rq
->idle
)) {
3408 WARN_ON(p
!= rq
->curr
);
3409 WARN_ON(p
->pi_blocked_on
);
3413 trace_sched_pi_setprio(p
, prio
);
3416 if (oldprio
== prio
)
3417 queue_flag
&= ~DEQUEUE_MOVE
;
3419 prev_class
= p
->sched_class
;
3420 queued
= task_on_rq_queued(p
);
3421 running
= task_current(rq
, p
);
3423 dequeue_task(rq
, p
, queue_flag
);
3425 put_prev_task(rq
, p
);
3428 * Boosting condition are:
3429 * 1. -rt task is running and holds mutex A
3430 * --> -dl task blocks on mutex A
3432 * 2. -dl task is running and holds mutex A
3433 * --> -dl task blocks on mutex A and could preempt the
3436 if (dl_prio(prio
)) {
3437 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3438 if (!dl_prio(p
->normal_prio
) ||
3439 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3440 p
->dl
.dl_boosted
= 1;
3441 queue_flag
|= ENQUEUE_REPLENISH
;
3443 p
->dl
.dl_boosted
= 0;
3444 p
->sched_class
= &dl_sched_class
;
3445 } else if (rt_prio(prio
)) {
3446 if (dl_prio(oldprio
))
3447 p
->dl
.dl_boosted
= 0;
3449 queue_flag
|= ENQUEUE_HEAD
;
3450 p
->sched_class
= &rt_sched_class
;
3452 if (dl_prio(oldprio
))
3453 p
->dl
.dl_boosted
= 0;
3454 if (rt_prio(oldprio
))
3456 p
->sched_class
= &fair_sched_class
;
3462 p
->sched_class
->set_curr_task(rq
);
3464 enqueue_task(rq
, p
, queue_flag
);
3466 check_class_changed(rq
, p
, prev_class
, oldprio
);
3468 preempt_disable(); /* avoid rq from going away on us */
3469 __task_rq_unlock(rq
);
3471 balance_callback(rq
);
3476 void set_user_nice(struct task_struct
*p
, long nice
)
3478 int old_prio
, delta
, queued
;
3479 unsigned long flags
;
3482 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3485 * We have to be careful, if called from sys_setpriority(),
3486 * the task might be in the middle of scheduling on another CPU.
3488 rq
= task_rq_lock(p
, &flags
);
3490 * The RT priorities are set via sched_setscheduler(), but we still
3491 * allow the 'normal' nice value to be set - but as expected
3492 * it wont have any effect on scheduling until the task is
3493 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3495 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3496 p
->static_prio
= NICE_TO_PRIO(nice
);
3499 queued
= task_on_rq_queued(p
);
3501 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3503 p
->static_prio
= NICE_TO_PRIO(nice
);
3506 p
->prio
= effective_prio(p
);
3507 delta
= p
->prio
- old_prio
;
3510 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3512 * If the task increased its priority or is running and
3513 * lowered its priority, then reschedule its CPU:
3515 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3519 task_rq_unlock(rq
, p
, &flags
);
3521 EXPORT_SYMBOL(set_user_nice
);
3524 * can_nice - check if a task can reduce its nice value
3528 int can_nice(const struct task_struct
*p
, const int nice
)
3530 /* convert nice value [19,-20] to rlimit style value [1,40] */
3531 int nice_rlim
= nice_to_rlimit(nice
);
3533 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3534 capable(CAP_SYS_NICE
));
3537 #ifdef __ARCH_WANT_SYS_NICE
3540 * sys_nice - change the priority of the current process.
3541 * @increment: priority increment
3543 * sys_setpriority is a more generic, but much slower function that
3544 * does similar things.
3546 SYSCALL_DEFINE1(nice
, int, increment
)
3551 * Setpriority might change our priority at the same moment.
3552 * We don't have to worry. Conceptually one call occurs first
3553 * and we have a single winner.
3555 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3556 nice
= task_nice(current
) + increment
;
3558 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3559 if (increment
< 0 && !can_nice(current
, nice
))
3562 retval
= security_task_setnice(current
, nice
);
3566 set_user_nice(current
, nice
);
3573 * task_prio - return the priority value of a given task.
3574 * @p: the task in question.
3576 * Return: The priority value as seen by users in /proc.
3577 * RT tasks are offset by -200. Normal tasks are centered
3578 * around 0, value goes from -16 to +15.
3580 int task_prio(const struct task_struct
*p
)
3582 return p
->prio
- MAX_RT_PRIO
;
3586 * idle_cpu - is a given cpu idle currently?
3587 * @cpu: the processor in question.
3589 * Return: 1 if the CPU is currently idle. 0 otherwise.
3591 int idle_cpu(int cpu
)
3593 struct rq
*rq
= cpu_rq(cpu
);
3595 if (rq
->curr
!= rq
->idle
)
3602 if (!llist_empty(&rq
->wake_list
))
3610 * idle_task - return the idle task for a given cpu.
3611 * @cpu: the processor in question.
3613 * Return: The idle task for the cpu @cpu.
3615 struct task_struct
*idle_task(int cpu
)
3617 return cpu_rq(cpu
)->idle
;
3621 * find_process_by_pid - find a process with a matching PID value.
3622 * @pid: the pid in question.
3624 * The task of @pid, if found. %NULL otherwise.
3626 static struct task_struct
*find_process_by_pid(pid_t pid
)
3628 return pid
? find_task_by_vpid(pid
) : current
;
3632 * This function initializes the sched_dl_entity of a newly becoming
3633 * SCHED_DEADLINE task.
3635 * Only the static values are considered here, the actual runtime and the
3636 * absolute deadline will be properly calculated when the task is enqueued
3637 * for the first time with its new policy.
3640 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3642 struct sched_dl_entity
*dl_se
= &p
->dl
;
3644 dl_se
->dl_runtime
= attr
->sched_runtime
;
3645 dl_se
->dl_deadline
= attr
->sched_deadline
;
3646 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3647 dl_se
->flags
= attr
->sched_flags
;
3648 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3651 * Changing the parameters of a task is 'tricky' and we're not doing
3652 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3654 * What we SHOULD do is delay the bandwidth release until the 0-lag
3655 * point. This would include retaining the task_struct until that time
3656 * and change dl_overflow() to not immediately decrement the current
3659 * Instead we retain the current runtime/deadline and let the new
3660 * parameters take effect after the current reservation period lapses.
3661 * This is safe (albeit pessimistic) because the 0-lag point is always
3662 * before the current scheduling deadline.
3664 * We can still have temporary overloads because we do not delay the
3665 * change in bandwidth until that time; so admission control is
3666 * not on the safe side. It does however guarantee tasks will never
3667 * consume more than promised.
3672 * sched_setparam() passes in -1 for its policy, to let the functions
3673 * it calls know not to change it.
3675 #define SETPARAM_POLICY -1
3677 static void __setscheduler_params(struct task_struct
*p
,
3678 const struct sched_attr
*attr
)
3680 int policy
= attr
->sched_policy
;
3682 if (policy
== SETPARAM_POLICY
)
3687 if (dl_policy(policy
))
3688 __setparam_dl(p
, attr
);
3689 else if (fair_policy(policy
))
3690 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3693 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3694 * !rt_policy. Always setting this ensures that things like
3695 * getparam()/getattr() don't report silly values for !rt tasks.
3697 p
->rt_priority
= attr
->sched_priority
;
3698 p
->normal_prio
= normal_prio(p
);
3702 /* Actually do priority change: must hold pi & rq lock. */
3703 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3704 const struct sched_attr
*attr
, bool keep_boost
)
3706 __setscheduler_params(p
, attr
);
3709 * Keep a potential priority boosting if called from
3710 * sched_setscheduler().
3713 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3715 p
->prio
= normal_prio(p
);
3717 if (dl_prio(p
->prio
))
3718 p
->sched_class
= &dl_sched_class
;
3719 else if (rt_prio(p
->prio
))
3720 p
->sched_class
= &rt_sched_class
;
3722 p
->sched_class
= &fair_sched_class
;
3726 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3728 struct sched_dl_entity
*dl_se
= &p
->dl
;
3730 attr
->sched_priority
= p
->rt_priority
;
3731 attr
->sched_runtime
= dl_se
->dl_runtime
;
3732 attr
->sched_deadline
= dl_se
->dl_deadline
;
3733 attr
->sched_period
= dl_se
->dl_period
;
3734 attr
->sched_flags
= dl_se
->flags
;
3738 * This function validates the new parameters of a -deadline task.
3739 * We ask for the deadline not being zero, and greater or equal
3740 * than the runtime, as well as the period of being zero or
3741 * greater than deadline. Furthermore, we have to be sure that
3742 * user parameters are above the internal resolution of 1us (we
3743 * check sched_runtime only since it is always the smaller one) and
3744 * below 2^63 ns (we have to check both sched_deadline and
3745 * sched_period, as the latter can be zero).
3748 __checkparam_dl(const struct sched_attr
*attr
)
3751 if (attr
->sched_deadline
== 0)
3755 * Since we truncate DL_SCALE bits, make sure we're at least
3758 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3762 * Since we use the MSB for wrap-around and sign issues, make
3763 * sure it's not set (mind that period can be equal to zero).
3765 if (attr
->sched_deadline
& (1ULL << 63) ||
3766 attr
->sched_period
& (1ULL << 63))
3769 /* runtime <= deadline <= period (if period != 0) */
3770 if ((attr
->sched_period
!= 0 &&
3771 attr
->sched_period
< attr
->sched_deadline
) ||
3772 attr
->sched_deadline
< attr
->sched_runtime
)
3779 * check the target process has a UID that matches the current process's
3781 static bool check_same_owner(struct task_struct
*p
)
3783 const struct cred
*cred
= current_cred(), *pcred
;
3787 pcred
= __task_cred(p
);
3788 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3789 uid_eq(cred
->euid
, pcred
->uid
));
3794 static bool dl_param_changed(struct task_struct
*p
,
3795 const struct sched_attr
*attr
)
3797 struct sched_dl_entity
*dl_se
= &p
->dl
;
3799 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3800 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3801 dl_se
->dl_period
!= attr
->sched_period
||
3802 dl_se
->flags
!= attr
->sched_flags
)
3808 static int __sched_setscheduler(struct task_struct
*p
,
3809 const struct sched_attr
*attr
,
3812 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3813 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3814 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3815 int new_effective_prio
, policy
= attr
->sched_policy
;
3816 unsigned long flags
;
3817 const struct sched_class
*prev_class
;
3820 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3822 /* may grab non-irq protected spin_locks */
3823 BUG_ON(in_interrupt());
3825 /* double check policy once rq lock held */
3827 reset_on_fork
= p
->sched_reset_on_fork
;
3828 policy
= oldpolicy
= p
->policy
;
3830 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3832 if (!valid_policy(policy
))
3836 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3840 * Valid priorities for SCHED_FIFO and SCHED_RR are
3841 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3842 * SCHED_BATCH and SCHED_IDLE is 0.
3844 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3845 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3847 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3848 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3852 * Allow unprivileged RT tasks to decrease priority:
3854 if (user
&& !capable(CAP_SYS_NICE
)) {
3855 if (fair_policy(policy
)) {
3856 if (attr
->sched_nice
< task_nice(p
) &&
3857 !can_nice(p
, attr
->sched_nice
))
3861 if (rt_policy(policy
)) {
3862 unsigned long rlim_rtprio
=
3863 task_rlimit(p
, RLIMIT_RTPRIO
);
3865 /* can't set/change the rt policy */
3866 if (policy
!= p
->policy
&& !rlim_rtprio
)
3869 /* can't increase priority */
3870 if (attr
->sched_priority
> p
->rt_priority
&&
3871 attr
->sched_priority
> rlim_rtprio
)
3876 * Can't set/change SCHED_DEADLINE policy at all for now
3877 * (safest behavior); in the future we would like to allow
3878 * unprivileged DL tasks to increase their relative deadline
3879 * or reduce their runtime (both ways reducing utilization)
3881 if (dl_policy(policy
))
3885 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3886 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3888 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
3889 if (!can_nice(p
, task_nice(p
)))
3893 /* can't change other user's priorities */
3894 if (!check_same_owner(p
))
3897 /* Normal users shall not reset the sched_reset_on_fork flag */
3898 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3903 retval
= security_task_setscheduler(p
);
3909 * make sure no PI-waiters arrive (or leave) while we are
3910 * changing the priority of the task:
3912 * To be able to change p->policy safely, the appropriate
3913 * runqueue lock must be held.
3915 rq
= task_rq_lock(p
, &flags
);
3918 * Changing the policy of the stop threads its a very bad idea
3920 if (p
== rq
->stop
) {
3921 task_rq_unlock(rq
, p
, &flags
);
3926 * If not changing anything there's no need to proceed further,
3927 * but store a possible modification of reset_on_fork.
3929 if (unlikely(policy
== p
->policy
)) {
3930 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3932 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3934 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3937 p
->sched_reset_on_fork
= reset_on_fork
;
3938 task_rq_unlock(rq
, p
, &flags
);
3944 #ifdef CONFIG_RT_GROUP_SCHED
3946 * Do not allow realtime tasks into groups that have no runtime
3949 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3950 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3951 !task_group_is_autogroup(task_group(p
))) {
3952 task_rq_unlock(rq
, p
, &flags
);
3957 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3958 cpumask_t
*span
= rq
->rd
->span
;
3961 * Don't allow tasks with an affinity mask smaller than
3962 * the entire root_domain to become SCHED_DEADLINE. We
3963 * will also fail if there's no bandwidth available.
3965 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3966 rq
->rd
->dl_bw
.bw
== 0) {
3967 task_rq_unlock(rq
, p
, &flags
);
3974 /* recheck policy now with rq lock held */
3975 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3976 policy
= oldpolicy
= -1;
3977 task_rq_unlock(rq
, p
, &flags
);
3982 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3983 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3986 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3987 task_rq_unlock(rq
, p
, &flags
);
3991 p
->sched_reset_on_fork
= reset_on_fork
;
3996 * Take priority boosted tasks into account. If the new
3997 * effective priority is unchanged, we just store the new
3998 * normal parameters and do not touch the scheduler class and
3999 * the runqueue. This will be done when the task deboost
4002 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4003 if (new_effective_prio
== oldprio
)
4004 queue_flags
&= ~DEQUEUE_MOVE
;
4007 queued
= task_on_rq_queued(p
);
4008 running
= task_current(rq
, p
);
4010 dequeue_task(rq
, p
, queue_flags
);
4012 put_prev_task(rq
, p
);
4014 prev_class
= p
->sched_class
;
4015 __setscheduler(rq
, p
, attr
, pi
);
4018 p
->sched_class
->set_curr_task(rq
);
4021 * We enqueue to tail when the priority of a task is
4022 * increased (user space view).
4024 if (oldprio
< p
->prio
)
4025 queue_flags
|= ENQUEUE_HEAD
;
4027 enqueue_task(rq
, p
, queue_flags
);
4030 check_class_changed(rq
, p
, prev_class
, oldprio
);
4031 preempt_disable(); /* avoid rq from going away on us */
4032 task_rq_unlock(rq
, p
, &flags
);
4035 rt_mutex_adjust_pi(p
);
4038 * Run balance callbacks after we've adjusted the PI chain.
4040 balance_callback(rq
);
4046 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4047 const struct sched_param
*param
, bool check
)
4049 struct sched_attr attr
= {
4050 .sched_policy
= policy
,
4051 .sched_priority
= param
->sched_priority
,
4052 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4055 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4056 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4057 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4058 policy
&= ~SCHED_RESET_ON_FORK
;
4059 attr
.sched_policy
= policy
;
4062 return __sched_setscheduler(p
, &attr
, check
, true);
4065 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4066 * @p: the task in question.
4067 * @policy: new policy.
4068 * @param: structure containing the new RT priority.
4070 * Return: 0 on success. An error code otherwise.
4072 * NOTE that the task may be already dead.
4074 int sched_setscheduler(struct task_struct
*p
, int policy
,
4075 const struct sched_param
*param
)
4077 return _sched_setscheduler(p
, policy
, param
, true);
4079 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4081 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4083 return __sched_setscheduler(p
, attr
, true, true);
4085 EXPORT_SYMBOL_GPL(sched_setattr
);
4088 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4089 * @p: the task in question.
4090 * @policy: new policy.
4091 * @param: structure containing the new RT priority.
4093 * Just like sched_setscheduler, only don't bother checking if the
4094 * current context has permission. For example, this is needed in
4095 * stop_machine(): we create temporary high priority worker threads,
4096 * but our caller might not have that capability.
4098 * Return: 0 on success. An error code otherwise.
4100 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4101 const struct sched_param
*param
)
4103 return _sched_setscheduler(p
, policy
, param
, false);
4105 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4108 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4110 struct sched_param lparam
;
4111 struct task_struct
*p
;
4114 if (!param
|| pid
< 0)
4116 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4121 p
= find_process_by_pid(pid
);
4123 retval
= sched_setscheduler(p
, policy
, &lparam
);
4130 * Mimics kernel/events/core.c perf_copy_attr().
4132 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4133 struct sched_attr
*attr
)
4138 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4142 * zero the full structure, so that a short copy will be nice.
4144 memset(attr
, 0, sizeof(*attr
));
4146 ret
= get_user(size
, &uattr
->size
);
4150 if (size
> PAGE_SIZE
) /* silly large */
4153 if (!size
) /* abi compat */
4154 size
= SCHED_ATTR_SIZE_VER0
;
4156 if (size
< SCHED_ATTR_SIZE_VER0
)
4160 * If we're handed a bigger struct than we know of,
4161 * ensure all the unknown bits are 0 - i.e. new
4162 * user-space does not rely on any kernel feature
4163 * extensions we dont know about yet.
4165 if (size
> sizeof(*attr
)) {
4166 unsigned char __user
*addr
;
4167 unsigned char __user
*end
;
4170 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4171 end
= (void __user
*)uattr
+ size
;
4173 for (; addr
< end
; addr
++) {
4174 ret
= get_user(val
, addr
);
4180 size
= sizeof(*attr
);
4183 ret
= copy_from_user(attr
, uattr
, size
);
4188 * XXX: do we want to be lenient like existing syscalls; or do we want
4189 * to be strict and return an error on out-of-bounds values?
4191 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4196 put_user(sizeof(*attr
), &uattr
->size
);
4201 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4202 * @pid: the pid in question.
4203 * @policy: new policy.
4204 * @param: structure containing the new RT priority.
4206 * Return: 0 on success. An error code otherwise.
4208 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4209 struct sched_param __user
*, param
)
4211 /* negative values for policy are not valid */
4215 return do_sched_setscheduler(pid
, policy
, param
);
4219 * sys_sched_setparam - set/change the RT priority of a thread
4220 * @pid: the pid in question.
4221 * @param: structure containing the new RT priority.
4223 * Return: 0 on success. An error code otherwise.
4225 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4227 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4231 * sys_sched_setattr - same as above, but with extended sched_attr
4232 * @pid: the pid in question.
4233 * @uattr: structure containing the extended parameters.
4234 * @flags: for future extension.
4236 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4237 unsigned int, flags
)
4239 struct sched_attr attr
;
4240 struct task_struct
*p
;
4243 if (!uattr
|| pid
< 0 || flags
)
4246 retval
= sched_copy_attr(uattr
, &attr
);
4250 if ((int)attr
.sched_policy
< 0)
4255 p
= find_process_by_pid(pid
);
4257 retval
= sched_setattr(p
, &attr
);
4264 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4265 * @pid: the pid in question.
4267 * Return: On success, the policy of the thread. Otherwise, a negative error
4270 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4272 struct task_struct
*p
;
4280 p
= find_process_by_pid(pid
);
4282 retval
= security_task_getscheduler(p
);
4285 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4292 * sys_sched_getparam - get the RT priority of a thread
4293 * @pid: the pid in question.
4294 * @param: structure containing the RT priority.
4296 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4299 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4301 struct sched_param lp
= { .sched_priority
= 0 };
4302 struct task_struct
*p
;
4305 if (!param
|| pid
< 0)
4309 p
= find_process_by_pid(pid
);
4314 retval
= security_task_getscheduler(p
);
4318 if (task_has_rt_policy(p
))
4319 lp
.sched_priority
= p
->rt_priority
;
4323 * This one might sleep, we cannot do it with a spinlock held ...
4325 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4334 static int sched_read_attr(struct sched_attr __user
*uattr
,
4335 struct sched_attr
*attr
,
4340 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4344 * If we're handed a smaller struct than we know of,
4345 * ensure all the unknown bits are 0 - i.e. old
4346 * user-space does not get uncomplete information.
4348 if (usize
< sizeof(*attr
)) {
4349 unsigned char *addr
;
4352 addr
= (void *)attr
+ usize
;
4353 end
= (void *)attr
+ sizeof(*attr
);
4355 for (; addr
< end
; addr
++) {
4363 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4371 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4372 * @pid: the pid in question.
4373 * @uattr: structure containing the extended parameters.
4374 * @size: sizeof(attr) for fwd/bwd comp.
4375 * @flags: for future extension.
4377 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4378 unsigned int, size
, unsigned int, flags
)
4380 struct sched_attr attr
= {
4381 .size
= sizeof(struct sched_attr
),
4383 struct task_struct
*p
;
4386 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4387 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4391 p
= find_process_by_pid(pid
);
4396 retval
= security_task_getscheduler(p
);
4400 attr
.sched_policy
= p
->policy
;
4401 if (p
->sched_reset_on_fork
)
4402 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4403 if (task_has_dl_policy(p
))
4404 __getparam_dl(p
, &attr
);
4405 else if (task_has_rt_policy(p
))
4406 attr
.sched_priority
= p
->rt_priority
;
4408 attr
.sched_nice
= task_nice(p
);
4412 retval
= sched_read_attr(uattr
, &attr
, size
);
4420 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4422 cpumask_var_t cpus_allowed
, new_mask
;
4423 struct task_struct
*p
;
4428 p
= find_process_by_pid(pid
);
4434 /* Prevent p going away */
4438 if (p
->flags
& PF_NO_SETAFFINITY
) {
4442 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4446 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4448 goto out_free_cpus_allowed
;
4451 if (!check_same_owner(p
)) {
4453 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4455 goto out_free_new_mask
;
4460 retval
= security_task_setscheduler(p
);
4462 goto out_free_new_mask
;
4465 cpuset_cpus_allowed(p
, cpus_allowed
);
4466 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4469 * Since bandwidth control happens on root_domain basis,
4470 * if admission test is enabled, we only admit -deadline
4471 * tasks allowed to run on all the CPUs in the task's
4475 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4477 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4480 goto out_free_new_mask
;
4486 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4489 cpuset_cpus_allowed(p
, cpus_allowed
);
4490 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4492 * We must have raced with a concurrent cpuset
4493 * update. Just reset the cpus_allowed to the
4494 * cpuset's cpus_allowed
4496 cpumask_copy(new_mask
, cpus_allowed
);
4501 free_cpumask_var(new_mask
);
4502 out_free_cpus_allowed
:
4503 free_cpumask_var(cpus_allowed
);
4509 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4510 struct cpumask
*new_mask
)
4512 if (len
< cpumask_size())
4513 cpumask_clear(new_mask
);
4514 else if (len
> cpumask_size())
4515 len
= cpumask_size();
4517 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4521 * sys_sched_setaffinity - set the cpu affinity of a process
4522 * @pid: pid of the process
4523 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4524 * @user_mask_ptr: user-space pointer to the new cpu mask
4526 * Return: 0 on success. An error code otherwise.
4528 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4529 unsigned long __user
*, user_mask_ptr
)
4531 cpumask_var_t new_mask
;
4534 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4537 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4539 retval
= sched_setaffinity(pid
, new_mask
);
4540 free_cpumask_var(new_mask
);
4544 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4546 struct task_struct
*p
;
4547 unsigned long flags
;
4553 p
= find_process_by_pid(pid
);
4557 retval
= security_task_getscheduler(p
);
4561 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4562 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4563 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4572 * sys_sched_getaffinity - get the cpu affinity of a process
4573 * @pid: pid of the process
4574 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4575 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4577 * Return: 0 on success. An error code otherwise.
4579 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4580 unsigned long __user
*, user_mask_ptr
)
4585 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4587 if (len
& (sizeof(unsigned long)-1))
4590 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4593 ret
= sched_getaffinity(pid
, mask
);
4595 size_t retlen
= min_t(size_t, len
, cpumask_size());
4597 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4602 free_cpumask_var(mask
);
4608 * sys_sched_yield - yield the current processor to other threads.
4610 * This function yields the current CPU to other tasks. If there are no
4611 * other threads running on this CPU then this function will return.
4615 SYSCALL_DEFINE0(sched_yield
)
4617 struct rq
*rq
= this_rq_lock();
4619 schedstat_inc(rq
, yld_count
);
4620 current
->sched_class
->yield_task(rq
);
4623 * Since we are going to call schedule() anyway, there's
4624 * no need to preempt or enable interrupts:
4626 __release(rq
->lock
);
4627 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4628 do_raw_spin_unlock(&rq
->lock
);
4629 sched_preempt_enable_no_resched();
4636 int __sched
_cond_resched(void)
4638 if (should_resched(0)) {
4639 preempt_schedule_common();
4644 EXPORT_SYMBOL(_cond_resched
);
4647 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4648 * call schedule, and on return reacquire the lock.
4650 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4651 * operations here to prevent schedule() from being called twice (once via
4652 * spin_unlock(), once by hand).
4654 int __cond_resched_lock(spinlock_t
*lock
)
4656 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4659 lockdep_assert_held(lock
);
4661 if (spin_needbreak(lock
) || resched
) {
4664 preempt_schedule_common();
4672 EXPORT_SYMBOL(__cond_resched_lock
);
4674 int __sched
__cond_resched_softirq(void)
4676 BUG_ON(!in_softirq());
4678 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4680 preempt_schedule_common();
4686 EXPORT_SYMBOL(__cond_resched_softirq
);
4689 * yield - yield the current processor to other threads.
4691 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4693 * The scheduler is at all times free to pick the calling task as the most
4694 * eligible task to run, if removing the yield() call from your code breaks
4695 * it, its already broken.
4697 * Typical broken usage is:
4702 * where one assumes that yield() will let 'the other' process run that will
4703 * make event true. If the current task is a SCHED_FIFO task that will never
4704 * happen. Never use yield() as a progress guarantee!!
4706 * If you want to use yield() to wait for something, use wait_event().
4707 * If you want to use yield() to be 'nice' for others, use cond_resched().
4708 * If you still want to use yield(), do not!
4710 void __sched
yield(void)
4712 set_current_state(TASK_RUNNING
);
4715 EXPORT_SYMBOL(yield
);
4718 * yield_to - yield the current processor to another thread in
4719 * your thread group, or accelerate that thread toward the
4720 * processor it's on.
4722 * @preempt: whether task preemption is allowed or not
4724 * It's the caller's job to ensure that the target task struct
4725 * can't go away on us before we can do any checks.
4728 * true (>0) if we indeed boosted the target task.
4729 * false (0) if we failed to boost the target.
4730 * -ESRCH if there's no task to yield to.
4732 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4734 struct task_struct
*curr
= current
;
4735 struct rq
*rq
, *p_rq
;
4736 unsigned long flags
;
4739 local_irq_save(flags
);
4745 * If we're the only runnable task on the rq and target rq also
4746 * has only one task, there's absolutely no point in yielding.
4748 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4753 double_rq_lock(rq
, p_rq
);
4754 if (task_rq(p
) != p_rq
) {
4755 double_rq_unlock(rq
, p_rq
);
4759 if (!curr
->sched_class
->yield_to_task
)
4762 if (curr
->sched_class
!= p
->sched_class
)
4765 if (task_running(p_rq
, p
) || p
->state
)
4768 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4770 schedstat_inc(rq
, yld_count
);
4772 * Make p's CPU reschedule; pick_next_entity takes care of
4775 if (preempt
&& rq
!= p_rq
)
4780 double_rq_unlock(rq
, p_rq
);
4782 local_irq_restore(flags
);
4789 EXPORT_SYMBOL_GPL(yield_to
);
4792 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4793 * that process accounting knows that this is a task in IO wait state.
4795 long __sched
io_schedule_timeout(long timeout
)
4797 int old_iowait
= current
->in_iowait
;
4801 current
->in_iowait
= 1;
4802 blk_schedule_flush_plug(current
);
4804 delayacct_blkio_start();
4806 atomic_inc(&rq
->nr_iowait
);
4807 ret
= schedule_timeout(timeout
);
4808 current
->in_iowait
= old_iowait
;
4809 atomic_dec(&rq
->nr_iowait
);
4810 delayacct_blkio_end();
4814 EXPORT_SYMBOL(io_schedule_timeout
);
4817 * sys_sched_get_priority_max - return maximum RT priority.
4818 * @policy: scheduling class.
4820 * Return: On success, this syscall returns the maximum
4821 * rt_priority that can be used by a given scheduling class.
4822 * On failure, a negative error code is returned.
4824 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4831 ret
= MAX_USER_RT_PRIO
-1;
4833 case SCHED_DEADLINE
:
4844 * sys_sched_get_priority_min - return minimum RT priority.
4845 * @policy: scheduling class.
4847 * Return: On success, this syscall returns the minimum
4848 * rt_priority that can be used by a given scheduling class.
4849 * On failure, a negative error code is returned.
4851 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4860 case SCHED_DEADLINE
:
4870 * sys_sched_rr_get_interval - return the default timeslice of a process.
4871 * @pid: pid of the process.
4872 * @interval: userspace pointer to the timeslice value.
4874 * this syscall writes the default timeslice value of a given process
4875 * into the user-space timespec buffer. A value of '0' means infinity.
4877 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4880 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4881 struct timespec __user
*, interval
)
4883 struct task_struct
*p
;
4884 unsigned int time_slice
;
4885 unsigned long flags
;
4895 p
= find_process_by_pid(pid
);
4899 retval
= security_task_getscheduler(p
);
4903 rq
= task_rq_lock(p
, &flags
);
4905 if (p
->sched_class
->get_rr_interval
)
4906 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4907 task_rq_unlock(rq
, p
, &flags
);
4910 jiffies_to_timespec(time_slice
, &t
);
4911 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4919 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4921 void sched_show_task(struct task_struct
*p
)
4923 unsigned long free
= 0;
4925 unsigned long state
= p
->state
;
4928 state
= __ffs(state
) + 1;
4929 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4930 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4931 #if BITS_PER_LONG == 32
4932 if (state
== TASK_RUNNING
)
4933 printk(KERN_CONT
" running ");
4935 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4937 if (state
== TASK_RUNNING
)
4938 printk(KERN_CONT
" running task ");
4940 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4942 #ifdef CONFIG_DEBUG_STACK_USAGE
4943 free
= stack_not_used(p
);
4948 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4950 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4951 task_pid_nr(p
), ppid
,
4952 (unsigned long)task_thread_info(p
)->flags
);
4954 print_worker_info(KERN_INFO
, p
);
4955 show_stack(p
, NULL
);
4958 void show_state_filter(unsigned long state_filter
)
4960 struct task_struct
*g
, *p
;
4962 #if BITS_PER_LONG == 32
4964 " task PC stack pid father\n");
4967 " task PC stack pid father\n");
4970 for_each_process_thread(g
, p
) {
4972 * reset the NMI-timeout, listing all files on a slow
4973 * console might take a lot of time:
4975 touch_nmi_watchdog();
4976 if (!state_filter
|| (p
->state
& state_filter
))
4980 touch_all_softlockup_watchdogs();
4982 #ifdef CONFIG_SCHED_DEBUG
4983 sysrq_sched_debug_show();
4987 * Only show locks if all tasks are dumped:
4990 debug_show_all_locks();
4993 void init_idle_bootup_task(struct task_struct
*idle
)
4995 idle
->sched_class
= &idle_sched_class
;
4999 * init_idle - set up an idle thread for a given CPU
5000 * @idle: task in question
5001 * @cpu: cpu the idle task belongs to
5003 * NOTE: this function does not set the idle thread's NEED_RESCHED
5004 * flag, to make booting more robust.
5006 void init_idle(struct task_struct
*idle
, int cpu
)
5008 struct rq
*rq
= cpu_rq(cpu
);
5009 unsigned long flags
;
5011 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5012 raw_spin_lock(&rq
->lock
);
5014 __sched_fork(0, idle
);
5015 idle
->state
= TASK_RUNNING
;
5016 idle
->se
.exec_start
= sched_clock();
5018 kasan_unpoison_task_stack(idle
);
5022 * Its possible that init_idle() gets called multiple times on a task,
5023 * in that case do_set_cpus_allowed() will not do the right thing.
5025 * And since this is boot we can forgo the serialization.
5027 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5030 * We're having a chicken and egg problem, even though we are
5031 * holding rq->lock, the cpu isn't yet set to this cpu so the
5032 * lockdep check in task_group() will fail.
5034 * Similar case to sched_fork(). / Alternatively we could
5035 * use task_rq_lock() here and obtain the other rq->lock.
5040 __set_task_cpu(idle
, cpu
);
5043 rq
->curr
= rq
->idle
= idle
;
5044 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5048 raw_spin_unlock(&rq
->lock
);
5049 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5051 /* Set the preempt count _outside_ the spinlocks! */
5052 init_idle_preempt_count(idle
, cpu
);
5055 * The idle tasks have their own, simple scheduling class:
5057 idle
->sched_class
= &idle_sched_class
;
5058 ftrace_graph_init_idle_task(idle
, cpu
);
5059 vtime_init_idle(idle
, cpu
);
5061 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5065 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5066 const struct cpumask
*trial
)
5068 int ret
= 1, trial_cpus
;
5069 struct dl_bw
*cur_dl_b
;
5070 unsigned long flags
;
5072 if (!cpumask_weight(cur
))
5075 rcu_read_lock_sched();
5076 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5077 trial_cpus
= cpumask_weight(trial
);
5079 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5080 if (cur_dl_b
->bw
!= -1 &&
5081 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5083 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5084 rcu_read_unlock_sched();
5089 int task_can_attach(struct task_struct
*p
,
5090 const struct cpumask
*cs_cpus_allowed
)
5095 * Kthreads which disallow setaffinity shouldn't be moved
5096 * to a new cpuset; we don't want to change their cpu
5097 * affinity and isolating such threads by their set of
5098 * allowed nodes is unnecessary. Thus, cpusets are not
5099 * applicable for such threads. This prevents checking for
5100 * success of set_cpus_allowed_ptr() on all attached tasks
5101 * before cpus_allowed may be changed.
5103 if (p
->flags
& PF_NO_SETAFFINITY
) {
5109 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5111 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5116 unsigned long flags
;
5118 rcu_read_lock_sched();
5119 dl_b
= dl_bw_of(dest_cpu
);
5120 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5121 cpus
= dl_bw_cpus(dest_cpu
);
5122 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5127 * We reserve space for this task in the destination
5128 * root_domain, as we can't fail after this point.
5129 * We will free resources in the source root_domain
5130 * later on (see set_cpus_allowed_dl()).
5132 __dl_add(dl_b
, p
->dl
.dl_bw
);
5134 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5135 rcu_read_unlock_sched();
5145 #ifdef CONFIG_NUMA_BALANCING
5146 /* Migrate current task p to target_cpu */
5147 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5149 struct migration_arg arg
= { p
, target_cpu
};
5150 int curr_cpu
= task_cpu(p
);
5152 if (curr_cpu
== target_cpu
)
5155 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5158 /* TODO: This is not properly updating schedstats */
5160 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5161 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5165 * Requeue a task on a given node and accurately track the number of NUMA
5166 * tasks on the runqueues
5168 void sched_setnuma(struct task_struct
*p
, int nid
)
5171 unsigned long flags
;
5172 bool queued
, running
;
5174 rq
= task_rq_lock(p
, &flags
);
5175 queued
= task_on_rq_queued(p
);
5176 running
= task_current(rq
, p
);
5179 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5181 put_prev_task(rq
, p
);
5183 p
->numa_preferred_nid
= nid
;
5186 p
->sched_class
->set_curr_task(rq
);
5188 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5189 task_rq_unlock(rq
, p
, &flags
);
5191 #endif /* CONFIG_NUMA_BALANCING */
5193 #ifdef CONFIG_HOTPLUG_CPU
5195 * Ensures that the idle task is using init_mm right before its cpu goes
5198 void idle_task_exit(void)
5200 struct mm_struct
*mm
= current
->active_mm
;
5202 BUG_ON(cpu_online(smp_processor_id()));
5204 if (mm
!= &init_mm
) {
5205 switch_mm(mm
, &init_mm
, current
);
5206 finish_arch_post_lock_switch();
5212 * Since this CPU is going 'away' for a while, fold any nr_active delta
5213 * we might have. Assumes we're called after migrate_tasks() so that the
5214 * nr_active count is stable.
5216 * Also see the comment "Global load-average calculations".
5218 static void calc_load_migrate(struct rq
*rq
)
5220 long delta
= calc_load_fold_active(rq
);
5222 atomic_long_add(delta
, &calc_load_tasks
);
5225 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5229 static const struct sched_class fake_sched_class
= {
5230 .put_prev_task
= put_prev_task_fake
,
5233 static struct task_struct fake_task
= {
5235 * Avoid pull_{rt,dl}_task()
5237 .prio
= MAX_PRIO
+ 1,
5238 .sched_class
= &fake_sched_class
,
5242 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5243 * try_to_wake_up()->select_task_rq().
5245 * Called with rq->lock held even though we'er in stop_machine() and
5246 * there's no concurrency possible, we hold the required locks anyway
5247 * because of lock validation efforts.
5249 static void migrate_tasks(struct rq
*dead_rq
)
5251 struct rq
*rq
= dead_rq
;
5252 struct task_struct
*next
, *stop
= rq
->stop
;
5256 * Fudge the rq selection such that the below task selection loop
5257 * doesn't get stuck on the currently eligible stop task.
5259 * We're currently inside stop_machine() and the rq is either stuck
5260 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5261 * either way we should never end up calling schedule() until we're
5267 * put_prev_task() and pick_next_task() sched
5268 * class method both need to have an up-to-date
5269 * value of rq->clock[_task]
5271 update_rq_clock(rq
);
5275 * There's this thread running, bail when that's the only
5278 if (rq
->nr_running
== 1)
5282 * pick_next_task assumes pinned rq->lock.
5284 lockdep_pin_lock(&rq
->lock
);
5285 next
= pick_next_task(rq
, &fake_task
);
5287 next
->sched_class
->put_prev_task(rq
, next
);
5290 * Rules for changing task_struct::cpus_allowed are holding
5291 * both pi_lock and rq->lock, such that holding either
5292 * stabilizes the mask.
5294 * Drop rq->lock is not quite as disastrous as it usually is
5295 * because !cpu_active at this point, which means load-balance
5296 * will not interfere. Also, stop-machine.
5298 lockdep_unpin_lock(&rq
->lock
);
5299 raw_spin_unlock(&rq
->lock
);
5300 raw_spin_lock(&next
->pi_lock
);
5301 raw_spin_lock(&rq
->lock
);
5304 * Since we're inside stop-machine, _nothing_ should have
5305 * changed the task, WARN if weird stuff happened, because in
5306 * that case the above rq->lock drop is a fail too.
5308 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5309 raw_spin_unlock(&next
->pi_lock
);
5313 /* Find suitable destination for @next, with force if needed. */
5314 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5316 rq
= __migrate_task(rq
, next
, dest_cpu
);
5317 if (rq
!= dead_rq
) {
5318 raw_spin_unlock(&rq
->lock
);
5320 raw_spin_lock(&rq
->lock
);
5322 raw_spin_unlock(&next
->pi_lock
);
5327 #endif /* CONFIG_HOTPLUG_CPU */
5329 static void set_rq_online(struct rq
*rq
)
5332 const struct sched_class
*class;
5334 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5337 for_each_class(class) {
5338 if (class->rq_online
)
5339 class->rq_online(rq
);
5344 static void set_rq_offline(struct rq
*rq
)
5347 const struct sched_class
*class;
5349 for_each_class(class) {
5350 if (class->rq_offline
)
5351 class->rq_offline(rq
);
5354 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5360 * migration_call - callback that gets triggered when a CPU is added.
5361 * Here we can start up the necessary migration thread for the new CPU.
5364 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5366 int cpu
= (long)hcpu
;
5367 unsigned long flags
;
5368 struct rq
*rq
= cpu_rq(cpu
);
5370 switch (action
& ~CPU_TASKS_FROZEN
) {
5372 case CPU_UP_PREPARE
:
5373 rq
->calc_load_update
= calc_load_update
;
5374 account_reset_rq(rq
);
5378 /* Update our root-domain */
5379 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5381 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5385 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5388 #ifdef CONFIG_HOTPLUG_CPU
5390 sched_ttwu_pending();
5391 /* Update our root-domain */
5392 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5394 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5398 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5399 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5403 calc_load_migrate(rq
);
5408 update_max_interval();
5414 * Register at high priority so that task migration (migrate_all_tasks)
5415 * happens before everything else. This has to be lower priority than
5416 * the notifier in the perf_event subsystem, though.
5418 static struct notifier_block migration_notifier
= {
5419 .notifier_call
= migration_call
,
5420 .priority
= CPU_PRI_MIGRATION
,
5423 static void set_cpu_rq_start_time(void)
5425 int cpu
= smp_processor_id();
5426 struct rq
*rq
= cpu_rq(cpu
);
5427 rq
->age_stamp
= sched_clock_cpu(cpu
);
5430 static int sched_cpu_active(struct notifier_block
*nfb
,
5431 unsigned long action
, void *hcpu
)
5433 int cpu
= (long)hcpu
;
5435 switch (action
& ~CPU_TASKS_FROZEN
) {
5437 set_cpu_rq_start_time();
5440 case CPU_DOWN_FAILED
:
5441 set_cpu_active(cpu
, true);
5449 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5450 unsigned long action
, void *hcpu
)
5452 switch (action
& ~CPU_TASKS_FROZEN
) {
5453 case CPU_DOWN_PREPARE
:
5454 set_cpu_active((long)hcpu
, false);
5461 static int __init
migration_init(void)
5463 void *cpu
= (void *)(long)smp_processor_id();
5466 /* Initialize migration for the boot CPU */
5467 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5468 BUG_ON(err
== NOTIFY_BAD
);
5469 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5470 register_cpu_notifier(&migration_notifier
);
5472 /* Register cpu active notifiers */
5473 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5474 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5478 early_initcall(migration_init
);
5480 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5482 #ifdef CONFIG_SCHED_DEBUG
5484 static __read_mostly
int sched_debug_enabled
;
5486 static int __init
sched_debug_setup(char *str
)
5488 sched_debug_enabled
= 1;
5492 early_param("sched_debug", sched_debug_setup
);
5494 static inline bool sched_debug(void)
5496 return sched_debug_enabled
;
5499 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5500 struct cpumask
*groupmask
)
5502 struct sched_group
*group
= sd
->groups
;
5504 cpumask_clear(groupmask
);
5506 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5508 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5509 printk("does not load-balance\n");
5511 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5516 printk(KERN_CONT
"span %*pbl level %s\n",
5517 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5519 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5520 printk(KERN_ERR
"ERROR: domain->span does not contain "
5523 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5524 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5528 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5532 printk(KERN_ERR
"ERROR: group is NULL\n");
5536 if (!cpumask_weight(sched_group_cpus(group
))) {
5537 printk(KERN_CONT
"\n");
5538 printk(KERN_ERR
"ERROR: empty group\n");
5542 if (!(sd
->flags
& SD_OVERLAP
) &&
5543 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5544 printk(KERN_CONT
"\n");
5545 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5549 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5551 printk(KERN_CONT
" %*pbl",
5552 cpumask_pr_args(sched_group_cpus(group
)));
5553 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5554 printk(KERN_CONT
" (cpu_capacity = %d)",
5555 group
->sgc
->capacity
);
5558 group
= group
->next
;
5559 } while (group
!= sd
->groups
);
5560 printk(KERN_CONT
"\n");
5562 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5563 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5566 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5567 printk(KERN_ERR
"ERROR: parent span is not a superset "
5568 "of domain->span\n");
5572 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5576 if (!sched_debug_enabled
)
5580 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5584 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5587 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5595 #else /* !CONFIG_SCHED_DEBUG */
5596 # define sched_domain_debug(sd, cpu) do { } while (0)
5597 static inline bool sched_debug(void)
5601 #endif /* CONFIG_SCHED_DEBUG */
5603 static int sd_degenerate(struct sched_domain
*sd
)
5605 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5608 /* Following flags need at least 2 groups */
5609 if (sd
->flags
& (SD_LOAD_BALANCE
|
5610 SD_BALANCE_NEWIDLE
|
5613 SD_SHARE_CPUCAPACITY
|
5614 SD_SHARE_PKG_RESOURCES
|
5615 SD_SHARE_POWERDOMAIN
)) {
5616 if (sd
->groups
!= sd
->groups
->next
)
5620 /* Following flags don't use groups */
5621 if (sd
->flags
& (SD_WAKE_AFFINE
))
5628 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5630 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5632 if (sd_degenerate(parent
))
5635 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5638 /* Flags needing groups don't count if only 1 group in parent */
5639 if (parent
->groups
== parent
->groups
->next
) {
5640 pflags
&= ~(SD_LOAD_BALANCE
|
5641 SD_BALANCE_NEWIDLE
|
5644 SD_SHARE_CPUCAPACITY
|
5645 SD_SHARE_PKG_RESOURCES
|
5647 SD_SHARE_POWERDOMAIN
);
5648 if (nr_node_ids
== 1)
5649 pflags
&= ~SD_SERIALIZE
;
5651 if (~cflags
& pflags
)
5657 static void free_rootdomain(struct rcu_head
*rcu
)
5659 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5661 cpupri_cleanup(&rd
->cpupri
);
5662 cpudl_cleanup(&rd
->cpudl
);
5663 free_cpumask_var(rd
->dlo_mask
);
5664 free_cpumask_var(rd
->rto_mask
);
5665 free_cpumask_var(rd
->online
);
5666 free_cpumask_var(rd
->span
);
5670 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5672 struct root_domain
*old_rd
= NULL
;
5673 unsigned long flags
;
5675 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5680 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5683 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5686 * If we dont want to free the old_rd yet then
5687 * set old_rd to NULL to skip the freeing later
5690 if (!atomic_dec_and_test(&old_rd
->refcount
))
5694 atomic_inc(&rd
->refcount
);
5697 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5698 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5701 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5704 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5707 static int init_rootdomain(struct root_domain
*rd
)
5709 memset(rd
, 0, sizeof(*rd
));
5711 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5713 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5715 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5717 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5720 init_dl_bw(&rd
->dl_bw
);
5721 if (cpudl_init(&rd
->cpudl
) != 0)
5724 if (cpupri_init(&rd
->cpupri
) != 0)
5729 free_cpumask_var(rd
->rto_mask
);
5731 free_cpumask_var(rd
->dlo_mask
);
5733 free_cpumask_var(rd
->online
);
5735 free_cpumask_var(rd
->span
);
5741 * By default the system creates a single root-domain with all cpus as
5742 * members (mimicking the global state we have today).
5744 struct root_domain def_root_domain
;
5746 static void init_defrootdomain(void)
5748 init_rootdomain(&def_root_domain
);
5750 atomic_set(&def_root_domain
.refcount
, 1);
5753 static struct root_domain
*alloc_rootdomain(void)
5755 struct root_domain
*rd
;
5757 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5761 if (init_rootdomain(rd
) != 0) {
5769 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5771 struct sched_group
*tmp
, *first
;
5780 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5785 } while (sg
!= first
);
5788 static void free_sched_domain(struct rcu_head
*rcu
)
5790 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5793 * If its an overlapping domain it has private groups, iterate and
5796 if (sd
->flags
& SD_OVERLAP
) {
5797 free_sched_groups(sd
->groups
, 1);
5798 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5799 kfree(sd
->groups
->sgc
);
5805 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5807 call_rcu(&sd
->rcu
, free_sched_domain
);
5810 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5812 for (; sd
; sd
= sd
->parent
)
5813 destroy_sched_domain(sd
, cpu
);
5817 * Keep a special pointer to the highest sched_domain that has
5818 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5819 * allows us to avoid some pointer chasing select_idle_sibling().
5821 * Also keep a unique ID per domain (we use the first cpu number in
5822 * the cpumask of the domain), this allows us to quickly tell if
5823 * two cpus are in the same cache domain, see cpus_share_cache().
5825 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5826 DEFINE_PER_CPU(int, sd_llc_size
);
5827 DEFINE_PER_CPU(int, sd_llc_id
);
5828 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5829 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5830 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5832 static void update_top_cache_domain(int cpu
)
5834 struct sched_domain
*sd
;
5835 struct sched_domain
*busy_sd
= NULL
;
5839 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5841 id
= cpumask_first(sched_domain_span(sd
));
5842 size
= cpumask_weight(sched_domain_span(sd
));
5843 busy_sd
= sd
->parent
; /* sd_busy */
5845 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5847 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5848 per_cpu(sd_llc_size
, cpu
) = size
;
5849 per_cpu(sd_llc_id
, cpu
) = id
;
5851 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5852 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5854 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5855 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5859 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5860 * hold the hotplug lock.
5863 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5865 struct rq
*rq
= cpu_rq(cpu
);
5866 struct sched_domain
*tmp
;
5868 /* Remove the sched domains which do not contribute to scheduling. */
5869 for (tmp
= sd
; tmp
; ) {
5870 struct sched_domain
*parent
= tmp
->parent
;
5874 if (sd_parent_degenerate(tmp
, parent
)) {
5875 tmp
->parent
= parent
->parent
;
5877 parent
->parent
->child
= tmp
;
5879 * Transfer SD_PREFER_SIBLING down in case of a
5880 * degenerate parent; the spans match for this
5881 * so the property transfers.
5883 if (parent
->flags
& SD_PREFER_SIBLING
)
5884 tmp
->flags
|= SD_PREFER_SIBLING
;
5885 destroy_sched_domain(parent
, cpu
);
5890 if (sd
&& sd_degenerate(sd
)) {
5893 destroy_sched_domain(tmp
, cpu
);
5898 sched_domain_debug(sd
, cpu
);
5900 rq_attach_root(rq
, rd
);
5902 rcu_assign_pointer(rq
->sd
, sd
);
5903 destroy_sched_domains(tmp
, cpu
);
5905 update_top_cache_domain(cpu
);
5908 /* Setup the mask of cpus configured for isolated domains */
5909 static int __init
isolated_cpu_setup(char *str
)
5913 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5914 ret
= cpulist_parse(str
, cpu_isolated_map
);
5916 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5921 __setup("isolcpus=", isolated_cpu_setup
);
5924 struct sched_domain
** __percpu sd
;
5925 struct root_domain
*rd
;
5936 * Build an iteration mask that can exclude certain CPUs from the upwards
5939 * Asymmetric node setups can result in situations where the domain tree is of
5940 * unequal depth, make sure to skip domains that already cover the entire
5943 * In that case build_sched_domains() will have terminated the iteration early
5944 * and our sibling sd spans will be empty. Domains should always include the
5945 * cpu they're built on, so check that.
5948 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5950 const struct cpumask
*span
= sched_domain_span(sd
);
5951 struct sd_data
*sdd
= sd
->private;
5952 struct sched_domain
*sibling
;
5955 for_each_cpu(i
, span
) {
5956 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5957 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5960 cpumask_set_cpu(i
, sched_group_mask(sg
));
5965 * Return the canonical balance cpu for this group, this is the first cpu
5966 * of this group that's also in the iteration mask.
5968 int group_balance_cpu(struct sched_group
*sg
)
5970 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5974 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5976 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5977 const struct cpumask
*span
= sched_domain_span(sd
);
5978 struct cpumask
*covered
= sched_domains_tmpmask
;
5979 struct sd_data
*sdd
= sd
->private;
5980 struct sched_domain
*sibling
;
5983 cpumask_clear(covered
);
5985 for_each_cpu(i
, span
) {
5986 struct cpumask
*sg_span
;
5988 if (cpumask_test_cpu(i
, covered
))
5991 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5993 /* See the comment near build_group_mask(). */
5994 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5997 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5998 GFP_KERNEL
, cpu_to_node(cpu
));
6003 sg_span
= sched_group_cpus(sg
);
6005 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6007 cpumask_set_cpu(i
, sg_span
);
6009 cpumask_or(covered
, covered
, sg_span
);
6011 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6012 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6013 build_group_mask(sd
, sg
);
6016 * Initialize sgc->capacity such that even if we mess up the
6017 * domains and no possible iteration will get us here, we won't
6020 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6023 * Make sure the first group of this domain contains the
6024 * canonical balance cpu. Otherwise the sched_domain iteration
6025 * breaks. See update_sg_lb_stats().
6027 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6028 group_balance_cpu(sg
) == cpu
)
6038 sd
->groups
= groups
;
6043 free_sched_groups(first
, 0);
6048 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6050 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6051 struct sched_domain
*child
= sd
->child
;
6054 cpu
= cpumask_first(sched_domain_span(child
));
6057 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6058 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6059 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6066 * build_sched_groups will build a circular linked list of the groups
6067 * covered by the given span, and will set each group's ->cpumask correctly,
6068 * and ->cpu_capacity to 0.
6070 * Assumes the sched_domain tree is fully constructed
6073 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6075 struct sched_group
*first
= NULL
, *last
= NULL
;
6076 struct sd_data
*sdd
= sd
->private;
6077 const struct cpumask
*span
= sched_domain_span(sd
);
6078 struct cpumask
*covered
;
6081 get_group(cpu
, sdd
, &sd
->groups
);
6082 atomic_inc(&sd
->groups
->ref
);
6084 if (cpu
!= cpumask_first(span
))
6087 lockdep_assert_held(&sched_domains_mutex
);
6088 covered
= sched_domains_tmpmask
;
6090 cpumask_clear(covered
);
6092 for_each_cpu(i
, span
) {
6093 struct sched_group
*sg
;
6096 if (cpumask_test_cpu(i
, covered
))
6099 group
= get_group(i
, sdd
, &sg
);
6100 cpumask_setall(sched_group_mask(sg
));
6102 for_each_cpu(j
, span
) {
6103 if (get_group(j
, sdd
, NULL
) != group
)
6106 cpumask_set_cpu(j
, covered
);
6107 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6122 * Initialize sched groups cpu_capacity.
6124 * cpu_capacity indicates the capacity of sched group, which is used while
6125 * distributing the load between different sched groups in a sched domain.
6126 * Typically cpu_capacity for all the groups in a sched domain will be same
6127 * unless there are asymmetries in the topology. If there are asymmetries,
6128 * group having more cpu_capacity will pickup more load compared to the
6129 * group having less cpu_capacity.
6131 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6133 struct sched_group
*sg
= sd
->groups
;
6138 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6140 } while (sg
!= sd
->groups
);
6142 if (cpu
!= group_balance_cpu(sg
))
6145 update_group_capacity(sd
, cpu
);
6146 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6150 * Initializers for schedule domains
6151 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6154 static int default_relax_domain_level
= -1;
6155 int sched_domain_level_max
;
6157 static int __init
setup_relax_domain_level(char *str
)
6159 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6160 pr_warn("Unable to set relax_domain_level\n");
6164 __setup("relax_domain_level=", setup_relax_domain_level
);
6166 static void set_domain_attribute(struct sched_domain
*sd
,
6167 struct sched_domain_attr
*attr
)
6171 if (!attr
|| attr
->relax_domain_level
< 0) {
6172 if (default_relax_domain_level
< 0)
6175 request
= default_relax_domain_level
;
6177 request
= attr
->relax_domain_level
;
6178 if (request
< sd
->level
) {
6179 /* turn off idle balance on this domain */
6180 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6182 /* turn on idle balance on this domain */
6183 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6187 static void __sdt_free(const struct cpumask
*cpu_map
);
6188 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6190 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6191 const struct cpumask
*cpu_map
)
6195 if (!atomic_read(&d
->rd
->refcount
))
6196 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6198 free_percpu(d
->sd
); /* fall through */
6200 __sdt_free(cpu_map
); /* fall through */
6206 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6207 const struct cpumask
*cpu_map
)
6209 memset(d
, 0, sizeof(*d
));
6211 if (__sdt_alloc(cpu_map
))
6212 return sa_sd_storage
;
6213 d
->sd
= alloc_percpu(struct sched_domain
*);
6215 return sa_sd_storage
;
6216 d
->rd
= alloc_rootdomain();
6219 return sa_rootdomain
;
6223 * NULL the sd_data elements we've used to build the sched_domain and
6224 * sched_group structure so that the subsequent __free_domain_allocs()
6225 * will not free the data we're using.
6227 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6229 struct sd_data
*sdd
= sd
->private;
6231 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6232 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6234 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6235 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6237 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6238 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6242 static int sched_domains_numa_levels
;
6243 enum numa_topology_type sched_numa_topology_type
;
6244 static int *sched_domains_numa_distance
;
6245 int sched_max_numa_distance
;
6246 static struct cpumask
***sched_domains_numa_masks
;
6247 static int sched_domains_curr_level
;
6251 * SD_flags allowed in topology descriptions.
6253 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6254 * SD_SHARE_PKG_RESOURCES - describes shared caches
6255 * SD_NUMA - describes NUMA topologies
6256 * SD_SHARE_POWERDOMAIN - describes shared power domain
6259 * SD_ASYM_PACKING - describes SMT quirks
6261 #define TOPOLOGY_SD_FLAGS \
6262 (SD_SHARE_CPUCAPACITY | \
6263 SD_SHARE_PKG_RESOURCES | \
6266 SD_SHARE_POWERDOMAIN)
6268 static struct sched_domain
*
6269 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6271 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6272 int sd_weight
, sd_flags
= 0;
6276 * Ugly hack to pass state to sd_numa_mask()...
6278 sched_domains_curr_level
= tl
->numa_level
;
6281 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6284 sd_flags
= (*tl
->sd_flags
)();
6285 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6286 "wrong sd_flags in topology description\n"))
6287 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6289 *sd
= (struct sched_domain
){
6290 .min_interval
= sd_weight
,
6291 .max_interval
= 2*sd_weight
,
6293 .imbalance_pct
= 125,
6295 .cache_nice_tries
= 0,
6302 .flags
= 1*SD_LOAD_BALANCE
6303 | 1*SD_BALANCE_NEWIDLE
6308 | 0*SD_SHARE_CPUCAPACITY
6309 | 0*SD_SHARE_PKG_RESOURCES
6311 | 0*SD_PREFER_SIBLING
6316 .last_balance
= jiffies
,
6317 .balance_interval
= sd_weight
,
6319 .max_newidle_lb_cost
= 0,
6320 .next_decay_max_lb_cost
= jiffies
,
6321 #ifdef CONFIG_SCHED_DEBUG
6327 * Convert topological properties into behaviour.
6330 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6331 sd
->flags
|= SD_PREFER_SIBLING
;
6332 sd
->imbalance_pct
= 110;
6333 sd
->smt_gain
= 1178; /* ~15% */
6335 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6336 sd
->imbalance_pct
= 117;
6337 sd
->cache_nice_tries
= 1;
6341 } else if (sd
->flags
& SD_NUMA
) {
6342 sd
->cache_nice_tries
= 2;
6346 sd
->flags
|= SD_SERIALIZE
;
6347 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6348 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6355 sd
->flags
|= SD_PREFER_SIBLING
;
6356 sd
->cache_nice_tries
= 1;
6361 sd
->private = &tl
->data
;
6367 * Topology list, bottom-up.
6369 static struct sched_domain_topology_level default_topology
[] = {
6370 #ifdef CONFIG_SCHED_SMT
6371 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6373 #ifdef CONFIG_SCHED_MC
6374 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6376 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6380 static struct sched_domain_topology_level
*sched_domain_topology
=
6383 #define for_each_sd_topology(tl) \
6384 for (tl = sched_domain_topology; tl->mask; tl++)
6386 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6388 sched_domain_topology
= tl
;
6393 static const struct cpumask
*sd_numa_mask(int cpu
)
6395 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6398 static void sched_numa_warn(const char *str
)
6400 static int done
= false;
6408 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6410 for (i
= 0; i
< nr_node_ids
; i
++) {
6411 printk(KERN_WARNING
" ");
6412 for (j
= 0; j
< nr_node_ids
; j
++)
6413 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6414 printk(KERN_CONT
"\n");
6416 printk(KERN_WARNING
"\n");
6419 bool find_numa_distance(int distance
)
6423 if (distance
== node_distance(0, 0))
6426 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6427 if (sched_domains_numa_distance
[i
] == distance
)
6435 * A system can have three types of NUMA topology:
6436 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6437 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6438 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6440 * The difference between a glueless mesh topology and a backplane
6441 * topology lies in whether communication between not directly
6442 * connected nodes goes through intermediary nodes (where programs
6443 * could run), or through backplane controllers. This affects
6444 * placement of programs.
6446 * The type of topology can be discerned with the following tests:
6447 * - If the maximum distance between any nodes is 1 hop, the system
6448 * is directly connected.
6449 * - If for two nodes A and B, located N > 1 hops away from each other,
6450 * there is an intermediary node C, which is < N hops away from both
6451 * nodes A and B, the system is a glueless mesh.
6453 static void init_numa_topology_type(void)
6457 n
= sched_max_numa_distance
;
6459 if (sched_domains_numa_levels
<= 1) {
6460 sched_numa_topology_type
= NUMA_DIRECT
;
6464 for_each_online_node(a
) {
6465 for_each_online_node(b
) {
6466 /* Find two nodes furthest removed from each other. */
6467 if (node_distance(a
, b
) < n
)
6470 /* Is there an intermediary node between a and b? */
6471 for_each_online_node(c
) {
6472 if (node_distance(a
, c
) < n
&&
6473 node_distance(b
, c
) < n
) {
6474 sched_numa_topology_type
=
6480 sched_numa_topology_type
= NUMA_BACKPLANE
;
6486 static void sched_init_numa(void)
6488 int next_distance
, curr_distance
= node_distance(0, 0);
6489 struct sched_domain_topology_level
*tl
;
6493 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6494 if (!sched_domains_numa_distance
)
6498 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6499 * unique distances in the node_distance() table.
6501 * Assumes node_distance(0,j) includes all distances in
6502 * node_distance(i,j) in order to avoid cubic time.
6504 next_distance
= curr_distance
;
6505 for (i
= 0; i
< nr_node_ids
; i
++) {
6506 for (j
= 0; j
< nr_node_ids
; j
++) {
6507 for (k
= 0; k
< nr_node_ids
; k
++) {
6508 int distance
= node_distance(i
, k
);
6510 if (distance
> curr_distance
&&
6511 (distance
< next_distance
||
6512 next_distance
== curr_distance
))
6513 next_distance
= distance
;
6516 * While not a strong assumption it would be nice to know
6517 * about cases where if node A is connected to B, B is not
6518 * equally connected to A.
6520 if (sched_debug() && node_distance(k
, i
) != distance
)
6521 sched_numa_warn("Node-distance not symmetric");
6523 if (sched_debug() && i
&& !find_numa_distance(distance
))
6524 sched_numa_warn("Node-0 not representative");
6526 if (next_distance
!= curr_distance
) {
6527 sched_domains_numa_distance
[level
++] = next_distance
;
6528 sched_domains_numa_levels
= level
;
6529 curr_distance
= next_distance
;
6534 * In case of sched_debug() we verify the above assumption.
6544 * 'level' contains the number of unique distances, excluding the
6545 * identity distance node_distance(i,i).
6547 * The sched_domains_numa_distance[] array includes the actual distance
6552 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6553 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6554 * the array will contain less then 'level' members. This could be
6555 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6556 * in other functions.
6558 * We reset it to 'level' at the end of this function.
6560 sched_domains_numa_levels
= 0;
6562 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6563 if (!sched_domains_numa_masks
)
6567 * Now for each level, construct a mask per node which contains all
6568 * cpus of nodes that are that many hops away from us.
6570 for (i
= 0; i
< level
; i
++) {
6571 sched_domains_numa_masks
[i
] =
6572 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6573 if (!sched_domains_numa_masks
[i
])
6576 for (j
= 0; j
< nr_node_ids
; j
++) {
6577 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6581 sched_domains_numa_masks
[i
][j
] = mask
;
6584 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6587 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6592 /* Compute default topology size */
6593 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6595 tl
= kzalloc((i
+ level
+ 1) *
6596 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6601 * Copy the default topology bits..
6603 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6604 tl
[i
] = sched_domain_topology
[i
];
6607 * .. and append 'j' levels of NUMA goodness.
6609 for (j
= 0; j
< level
; i
++, j
++) {
6610 tl
[i
] = (struct sched_domain_topology_level
){
6611 .mask
= sd_numa_mask
,
6612 .sd_flags
= cpu_numa_flags
,
6613 .flags
= SDTL_OVERLAP
,
6619 sched_domain_topology
= tl
;
6621 sched_domains_numa_levels
= level
;
6622 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6624 init_numa_topology_type();
6627 static void sched_domains_numa_masks_set(int cpu
)
6630 int node
= cpu_to_node(cpu
);
6632 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6633 for (j
= 0; j
< nr_node_ids
; j
++) {
6634 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6635 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6640 static void sched_domains_numa_masks_clear(int cpu
)
6643 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6644 for (j
= 0; j
< nr_node_ids
; j
++)
6645 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6650 * Update sched_domains_numa_masks[level][node] array when new cpus
6653 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6654 unsigned long action
,
6657 int cpu
= (long)hcpu
;
6659 switch (action
& ~CPU_TASKS_FROZEN
) {
6661 sched_domains_numa_masks_set(cpu
);
6665 sched_domains_numa_masks_clear(cpu
);
6675 static inline void sched_init_numa(void)
6679 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6680 unsigned long action
,
6685 #endif /* CONFIG_NUMA */
6687 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6689 struct sched_domain_topology_level
*tl
;
6692 for_each_sd_topology(tl
) {
6693 struct sd_data
*sdd
= &tl
->data
;
6695 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6699 sdd
->sg
= alloc_percpu(struct sched_group
*);
6703 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6707 for_each_cpu(j
, cpu_map
) {
6708 struct sched_domain
*sd
;
6709 struct sched_group
*sg
;
6710 struct sched_group_capacity
*sgc
;
6712 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6713 GFP_KERNEL
, cpu_to_node(j
));
6717 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6719 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6720 GFP_KERNEL
, cpu_to_node(j
));
6726 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6728 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6729 GFP_KERNEL
, cpu_to_node(j
));
6733 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6740 static void __sdt_free(const struct cpumask
*cpu_map
)
6742 struct sched_domain_topology_level
*tl
;
6745 for_each_sd_topology(tl
) {
6746 struct sd_data
*sdd
= &tl
->data
;
6748 for_each_cpu(j
, cpu_map
) {
6749 struct sched_domain
*sd
;
6752 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6753 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6754 free_sched_groups(sd
->groups
, 0);
6755 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6759 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6761 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6763 free_percpu(sdd
->sd
);
6765 free_percpu(sdd
->sg
);
6767 free_percpu(sdd
->sgc
);
6772 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6773 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6774 struct sched_domain
*child
, int cpu
)
6776 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6780 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6782 sd
->level
= child
->level
+ 1;
6783 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6787 if (!cpumask_subset(sched_domain_span(child
),
6788 sched_domain_span(sd
))) {
6789 pr_err("BUG: arch topology borken\n");
6790 #ifdef CONFIG_SCHED_DEBUG
6791 pr_err(" the %s domain not a subset of the %s domain\n",
6792 child
->name
, sd
->name
);
6794 /* Fixup, ensure @sd has at least @child cpus. */
6795 cpumask_or(sched_domain_span(sd
),
6796 sched_domain_span(sd
),
6797 sched_domain_span(child
));
6801 set_domain_attribute(sd
, attr
);
6807 * Build sched domains for a given set of cpus and attach the sched domains
6808 * to the individual cpus
6810 static int build_sched_domains(const struct cpumask
*cpu_map
,
6811 struct sched_domain_attr
*attr
)
6813 enum s_alloc alloc_state
;
6814 struct sched_domain
*sd
;
6816 int i
, ret
= -ENOMEM
;
6818 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6819 if (alloc_state
!= sa_rootdomain
)
6822 /* Set up domains for cpus specified by the cpu_map. */
6823 for_each_cpu(i
, cpu_map
) {
6824 struct sched_domain_topology_level
*tl
;
6827 for_each_sd_topology(tl
) {
6828 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6829 if (tl
== sched_domain_topology
)
6830 *per_cpu_ptr(d
.sd
, i
) = sd
;
6831 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6832 sd
->flags
|= SD_OVERLAP
;
6833 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6838 /* Build the groups for the domains */
6839 for_each_cpu(i
, cpu_map
) {
6840 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6841 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6842 if (sd
->flags
& SD_OVERLAP
) {
6843 if (build_overlap_sched_groups(sd
, i
))
6846 if (build_sched_groups(sd
, i
))
6852 /* Calculate CPU capacity for physical packages and nodes */
6853 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6854 if (!cpumask_test_cpu(i
, cpu_map
))
6857 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6858 claim_allocations(i
, sd
);
6859 init_sched_groups_capacity(i
, sd
);
6863 /* Attach the domains */
6865 for_each_cpu(i
, cpu_map
) {
6866 sd
= *per_cpu_ptr(d
.sd
, i
);
6867 cpu_attach_domain(sd
, d
.rd
, i
);
6873 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6877 static cpumask_var_t
*doms_cur
; /* current sched domains */
6878 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6879 static struct sched_domain_attr
*dattr_cur
;
6880 /* attribues of custom domains in 'doms_cur' */
6883 * Special case: If a kmalloc of a doms_cur partition (array of
6884 * cpumask) fails, then fallback to a single sched domain,
6885 * as determined by the single cpumask fallback_doms.
6887 static cpumask_var_t fallback_doms
;
6890 * arch_update_cpu_topology lets virtualized architectures update the
6891 * cpu core maps. It is supposed to return 1 if the topology changed
6892 * or 0 if it stayed the same.
6894 int __weak
arch_update_cpu_topology(void)
6899 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6902 cpumask_var_t
*doms
;
6904 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6907 for (i
= 0; i
< ndoms
; i
++) {
6908 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6909 free_sched_domains(doms
, i
);
6916 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6919 for (i
= 0; i
< ndoms
; i
++)
6920 free_cpumask_var(doms
[i
]);
6925 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6926 * For now this just excludes isolated cpus, but could be used to
6927 * exclude other special cases in the future.
6929 static int init_sched_domains(const struct cpumask
*cpu_map
)
6933 arch_update_cpu_topology();
6935 doms_cur
= alloc_sched_domains(ndoms_cur
);
6937 doms_cur
= &fallback_doms
;
6938 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6939 err
= build_sched_domains(doms_cur
[0], NULL
);
6940 register_sched_domain_sysctl();
6946 * Detach sched domains from a group of cpus specified in cpu_map
6947 * These cpus will now be attached to the NULL domain
6949 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6954 for_each_cpu(i
, cpu_map
)
6955 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6959 /* handle null as "default" */
6960 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6961 struct sched_domain_attr
*new, int idx_new
)
6963 struct sched_domain_attr tmp
;
6970 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6971 new ? (new + idx_new
) : &tmp
,
6972 sizeof(struct sched_domain_attr
));
6976 * Partition sched domains as specified by the 'ndoms_new'
6977 * cpumasks in the array doms_new[] of cpumasks. This compares
6978 * doms_new[] to the current sched domain partitioning, doms_cur[].
6979 * It destroys each deleted domain and builds each new domain.
6981 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6982 * The masks don't intersect (don't overlap.) We should setup one
6983 * sched domain for each mask. CPUs not in any of the cpumasks will
6984 * not be load balanced. If the same cpumask appears both in the
6985 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6988 * The passed in 'doms_new' should be allocated using
6989 * alloc_sched_domains. This routine takes ownership of it and will
6990 * free_sched_domains it when done with it. If the caller failed the
6991 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6992 * and partition_sched_domains() will fallback to the single partition
6993 * 'fallback_doms', it also forces the domains to be rebuilt.
6995 * If doms_new == NULL it will be replaced with cpu_online_mask.
6996 * ndoms_new == 0 is a special case for destroying existing domains,
6997 * and it will not create the default domain.
6999 * Call with hotplug lock held
7001 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7002 struct sched_domain_attr
*dattr_new
)
7007 mutex_lock(&sched_domains_mutex
);
7009 /* always unregister in case we don't destroy any domains */
7010 unregister_sched_domain_sysctl();
7012 /* Let architecture update cpu core mappings. */
7013 new_topology
= arch_update_cpu_topology();
7015 n
= doms_new
? ndoms_new
: 0;
7017 /* Destroy deleted domains */
7018 for (i
= 0; i
< ndoms_cur
; i
++) {
7019 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7020 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7021 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7024 /* no match - a current sched domain not in new doms_new[] */
7025 detach_destroy_domains(doms_cur
[i
]);
7031 if (doms_new
== NULL
) {
7033 doms_new
= &fallback_doms
;
7034 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7035 WARN_ON_ONCE(dattr_new
);
7038 /* Build new domains */
7039 for (i
= 0; i
< ndoms_new
; i
++) {
7040 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7041 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7042 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7045 /* no match - add a new doms_new */
7046 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7051 /* Remember the new sched domains */
7052 if (doms_cur
!= &fallback_doms
)
7053 free_sched_domains(doms_cur
, ndoms_cur
);
7054 kfree(dattr_cur
); /* kfree(NULL) is safe */
7055 doms_cur
= doms_new
;
7056 dattr_cur
= dattr_new
;
7057 ndoms_cur
= ndoms_new
;
7059 register_sched_domain_sysctl();
7061 mutex_unlock(&sched_domains_mutex
);
7064 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7067 * Update cpusets according to cpu_active mask. If cpusets are
7068 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7069 * around partition_sched_domains().
7071 * If we come here as part of a suspend/resume, don't touch cpusets because we
7072 * want to restore it back to its original state upon resume anyway.
7074 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7078 case CPU_ONLINE_FROZEN
:
7079 case CPU_DOWN_FAILED_FROZEN
:
7082 * num_cpus_frozen tracks how many CPUs are involved in suspend
7083 * resume sequence. As long as this is not the last online
7084 * operation in the resume sequence, just build a single sched
7085 * domain, ignoring cpusets.
7088 if (likely(num_cpus_frozen
)) {
7089 partition_sched_domains(1, NULL
, NULL
);
7094 * This is the last CPU online operation. So fall through and
7095 * restore the original sched domains by considering the
7096 * cpuset configurations.
7100 cpuset_update_active_cpus(true);
7108 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7111 unsigned long flags
;
7112 long cpu
= (long)hcpu
;
7118 case CPU_DOWN_PREPARE
:
7119 rcu_read_lock_sched();
7120 dl_b
= dl_bw_of(cpu
);
7122 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7123 cpus
= dl_bw_cpus(cpu
);
7124 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7125 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7127 rcu_read_unlock_sched();
7130 return notifier_from_errno(-EBUSY
);
7131 cpuset_update_active_cpus(false);
7133 case CPU_DOWN_PREPARE_FROZEN
:
7135 partition_sched_domains(1, NULL
, NULL
);
7143 void __init
sched_init_smp(void)
7145 cpumask_var_t non_isolated_cpus
;
7147 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7148 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7153 * There's no userspace yet to cause hotplug operations; hence all the
7154 * cpu masks are stable and all blatant races in the below code cannot
7157 mutex_lock(&sched_domains_mutex
);
7158 init_sched_domains(cpu_active_mask
);
7159 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7160 if (cpumask_empty(non_isolated_cpus
))
7161 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7162 mutex_unlock(&sched_domains_mutex
);
7164 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7165 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7166 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7170 /* Move init over to a non-isolated CPU */
7171 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7173 sched_init_granularity();
7174 free_cpumask_var(non_isolated_cpus
);
7176 init_sched_rt_class();
7177 init_sched_dl_class();
7180 void __init
sched_init_smp(void)
7182 sched_init_granularity();
7184 #endif /* CONFIG_SMP */
7186 int in_sched_functions(unsigned long addr
)
7188 return in_lock_functions(addr
) ||
7189 (addr
>= (unsigned long)__sched_text_start
7190 && addr
< (unsigned long)__sched_text_end
);
7193 #ifdef CONFIG_CGROUP_SCHED
7195 * Default task group.
7196 * Every task in system belongs to this group at bootup.
7198 struct task_group root_task_group
;
7199 LIST_HEAD(task_groups
);
7201 /* Cacheline aligned slab cache for task_group */
7202 static struct kmem_cache
*task_group_cache __read_mostly
;
7205 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7207 void __init
sched_init(void)
7210 unsigned long alloc_size
= 0, ptr
;
7212 #ifdef CONFIG_FAIR_GROUP_SCHED
7213 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7215 #ifdef CONFIG_RT_GROUP_SCHED
7216 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7219 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7221 #ifdef CONFIG_FAIR_GROUP_SCHED
7222 root_task_group
.se
= (struct sched_entity
**)ptr
;
7223 ptr
+= nr_cpu_ids
* sizeof(void **);
7225 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7226 ptr
+= nr_cpu_ids
* sizeof(void **);
7228 #endif /* CONFIG_FAIR_GROUP_SCHED */
7229 #ifdef CONFIG_RT_GROUP_SCHED
7230 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7231 ptr
+= nr_cpu_ids
* sizeof(void **);
7233 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7234 ptr
+= nr_cpu_ids
* sizeof(void **);
7236 #endif /* CONFIG_RT_GROUP_SCHED */
7238 #ifdef CONFIG_CPUMASK_OFFSTACK
7239 for_each_possible_cpu(i
) {
7240 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7241 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7243 #endif /* CONFIG_CPUMASK_OFFSTACK */
7245 init_rt_bandwidth(&def_rt_bandwidth
,
7246 global_rt_period(), global_rt_runtime());
7247 init_dl_bandwidth(&def_dl_bandwidth
,
7248 global_rt_period(), global_rt_runtime());
7251 init_defrootdomain();
7254 #ifdef CONFIG_RT_GROUP_SCHED
7255 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7256 global_rt_period(), global_rt_runtime());
7257 #endif /* CONFIG_RT_GROUP_SCHED */
7259 #ifdef CONFIG_CGROUP_SCHED
7260 task_group_cache
= KMEM_CACHE(task_group
, 0);
7262 list_add(&root_task_group
.list
, &task_groups
);
7263 INIT_LIST_HEAD(&root_task_group
.children
);
7264 INIT_LIST_HEAD(&root_task_group
.siblings
);
7265 autogroup_init(&init_task
);
7266 #endif /* CONFIG_CGROUP_SCHED */
7268 for_each_possible_cpu(i
) {
7272 raw_spin_lock_init(&rq
->lock
);
7274 rq
->calc_load_active
= 0;
7275 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7276 init_cfs_rq(&rq
->cfs
);
7277 init_rt_rq(&rq
->rt
);
7278 init_dl_rq(&rq
->dl
);
7279 #ifdef CONFIG_FAIR_GROUP_SCHED
7280 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7281 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7283 * How much cpu bandwidth does root_task_group get?
7285 * In case of task-groups formed thr' the cgroup filesystem, it
7286 * gets 100% of the cpu resources in the system. This overall
7287 * system cpu resource is divided among the tasks of
7288 * root_task_group and its child task-groups in a fair manner,
7289 * based on each entity's (task or task-group's) weight
7290 * (se->load.weight).
7292 * In other words, if root_task_group has 10 tasks of weight
7293 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7294 * then A0's share of the cpu resource is:
7296 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7298 * We achieve this by letting root_task_group's tasks sit
7299 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7301 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7302 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7303 #endif /* CONFIG_FAIR_GROUP_SCHED */
7305 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7306 #ifdef CONFIG_RT_GROUP_SCHED
7307 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7310 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7311 rq
->cpu_load
[j
] = 0;
7313 rq
->last_load_update_tick
= jiffies
;
7318 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7319 rq
->balance_callback
= NULL
;
7320 rq
->active_balance
= 0;
7321 rq
->next_balance
= jiffies
;
7326 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7327 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7329 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7331 rq_attach_root(rq
, &def_root_domain
);
7332 #ifdef CONFIG_NO_HZ_COMMON
7335 #ifdef CONFIG_NO_HZ_FULL
7336 rq
->last_sched_tick
= 0;
7340 atomic_set(&rq
->nr_iowait
, 0);
7343 set_load_weight(&init_task
);
7345 #ifdef CONFIG_PREEMPT_NOTIFIERS
7346 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7350 * The boot idle thread does lazy MMU switching as well:
7352 atomic_inc(&init_mm
.mm_count
);
7353 enter_lazy_tlb(&init_mm
, current
);
7356 * During early bootup we pretend to be a normal task:
7358 current
->sched_class
= &fair_sched_class
;
7361 * Make us the idle thread. Technically, schedule() should not be
7362 * called from this thread, however somewhere below it might be,
7363 * but because we are the idle thread, we just pick up running again
7364 * when this runqueue becomes "idle".
7366 init_idle(current
, smp_processor_id());
7368 calc_load_update
= jiffies
+ LOAD_FREQ
;
7371 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7372 /* May be allocated at isolcpus cmdline parse time */
7373 if (cpu_isolated_map
== NULL
)
7374 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7375 idle_thread_set_boot_cpu();
7376 set_cpu_rq_start_time();
7378 init_sched_fair_class();
7380 scheduler_running
= 1;
7383 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7384 static inline int preempt_count_equals(int preempt_offset
)
7386 int nested
= preempt_count() + rcu_preempt_depth();
7388 return (nested
== preempt_offset
);
7391 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7394 * Blocking primitives will set (and therefore destroy) current->state,
7395 * since we will exit with TASK_RUNNING make sure we enter with it,
7396 * otherwise we will destroy state.
7398 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7399 "do not call blocking ops when !TASK_RUNNING; "
7400 "state=%lx set at [<%p>] %pS\n",
7402 (void *)current
->task_state_change
,
7403 (void *)current
->task_state_change
);
7405 ___might_sleep(file
, line
, preempt_offset
);
7407 EXPORT_SYMBOL(__might_sleep
);
7409 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7411 static unsigned long prev_jiffy
; /* ratelimiting */
7413 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7414 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7415 !is_idle_task(current
)) ||
7416 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7418 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7420 prev_jiffy
= jiffies
;
7423 "BUG: sleeping function called from invalid context at %s:%d\n",
7426 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7427 in_atomic(), irqs_disabled(),
7428 current
->pid
, current
->comm
);
7430 if (task_stack_end_corrupted(current
))
7431 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7433 debug_show_held_locks(current
);
7434 if (irqs_disabled())
7435 print_irqtrace_events(current
);
7436 #ifdef CONFIG_DEBUG_PREEMPT
7437 if (!preempt_count_equals(preempt_offset
)) {
7438 pr_err("Preemption disabled at:");
7439 print_ip_sym(current
->preempt_disable_ip
);
7445 EXPORT_SYMBOL(___might_sleep
);
7448 #ifdef CONFIG_MAGIC_SYSRQ
7449 void normalize_rt_tasks(void)
7451 struct task_struct
*g
, *p
;
7452 struct sched_attr attr
= {
7453 .sched_policy
= SCHED_NORMAL
,
7456 read_lock(&tasklist_lock
);
7457 for_each_process_thread(g
, p
) {
7459 * Only normalize user tasks:
7461 if (p
->flags
& PF_KTHREAD
)
7464 p
->se
.exec_start
= 0;
7465 #ifdef CONFIG_SCHEDSTATS
7466 p
->se
.statistics
.wait_start
= 0;
7467 p
->se
.statistics
.sleep_start
= 0;
7468 p
->se
.statistics
.block_start
= 0;
7471 if (!dl_task(p
) && !rt_task(p
)) {
7473 * Renice negative nice level userspace
7476 if (task_nice(p
) < 0)
7477 set_user_nice(p
, 0);
7481 __sched_setscheduler(p
, &attr
, false, false);
7483 read_unlock(&tasklist_lock
);
7486 #endif /* CONFIG_MAGIC_SYSRQ */
7488 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7490 * These functions are only useful for the IA64 MCA handling, or kdb.
7492 * They can only be called when the whole system has been
7493 * stopped - every CPU needs to be quiescent, and no scheduling
7494 * activity can take place. Using them for anything else would
7495 * be a serious bug, and as a result, they aren't even visible
7496 * under any other configuration.
7500 * curr_task - return the current task for a given cpu.
7501 * @cpu: the processor in question.
7503 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7505 * Return: The current task for @cpu.
7507 struct task_struct
*curr_task(int cpu
)
7509 return cpu_curr(cpu
);
7512 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7516 * set_curr_task - set the current task for a given cpu.
7517 * @cpu: the processor in question.
7518 * @p: the task pointer to set.
7520 * Description: This function must only be used when non-maskable interrupts
7521 * are serviced on a separate stack. It allows the architecture to switch the
7522 * notion of the current task on a cpu in a non-blocking manner. This function
7523 * must be called with all CPU's synchronized, and interrupts disabled, the
7524 * and caller must save the original value of the current task (see
7525 * curr_task() above) and restore that value before reenabling interrupts and
7526 * re-starting the system.
7528 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7530 void set_curr_task(int cpu
, struct task_struct
*p
)
7537 #ifdef CONFIG_CGROUP_SCHED
7538 /* task_group_lock serializes the addition/removal of task groups */
7539 static DEFINE_SPINLOCK(task_group_lock
);
7541 static void sched_free_group(struct task_group
*tg
)
7543 free_fair_sched_group(tg
);
7544 free_rt_sched_group(tg
);
7546 kmem_cache_free(task_group_cache
, tg
);
7549 /* allocate runqueue etc for a new task group */
7550 struct task_group
*sched_create_group(struct task_group
*parent
)
7552 struct task_group
*tg
;
7554 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7556 return ERR_PTR(-ENOMEM
);
7558 if (!alloc_fair_sched_group(tg
, parent
))
7561 if (!alloc_rt_sched_group(tg
, parent
))
7567 sched_free_group(tg
);
7568 return ERR_PTR(-ENOMEM
);
7571 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7573 unsigned long flags
;
7575 spin_lock_irqsave(&task_group_lock
, flags
);
7576 list_add_rcu(&tg
->list
, &task_groups
);
7578 WARN_ON(!parent
); /* root should already exist */
7580 tg
->parent
= parent
;
7581 INIT_LIST_HEAD(&tg
->children
);
7582 list_add_rcu(&tg
->siblings
, &parent
->children
);
7583 spin_unlock_irqrestore(&task_group_lock
, flags
);
7586 /* rcu callback to free various structures associated with a task group */
7587 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7589 /* now it should be safe to free those cfs_rqs */
7590 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7593 void sched_destroy_group(struct task_group
*tg
)
7595 /* wait for possible concurrent references to cfs_rqs complete */
7596 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7599 void sched_offline_group(struct task_group
*tg
)
7601 unsigned long flags
;
7603 /* end participation in shares distribution */
7604 unregister_fair_sched_group(tg
);
7606 spin_lock_irqsave(&task_group_lock
, flags
);
7607 list_del_rcu(&tg
->list
);
7608 list_del_rcu(&tg
->siblings
);
7609 spin_unlock_irqrestore(&task_group_lock
, flags
);
7612 /* change task's runqueue when it moves between groups.
7613 * The caller of this function should have put the task in its new group
7614 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7615 * reflect its new group.
7617 void sched_move_task(struct task_struct
*tsk
)
7619 struct task_group
*tg
;
7620 int queued
, running
;
7621 unsigned long flags
;
7624 rq
= task_rq_lock(tsk
, &flags
);
7626 running
= task_current(rq
, tsk
);
7627 queued
= task_on_rq_queued(tsk
);
7630 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7631 if (unlikely(running
))
7632 put_prev_task(rq
, tsk
);
7635 * All callers are synchronized by task_rq_lock(); we do not use RCU
7636 * which is pointless here. Thus, we pass "true" to task_css_check()
7637 * to prevent lockdep warnings.
7639 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7640 struct task_group
, css
);
7641 tg
= autogroup_task_group(tsk
, tg
);
7642 tsk
->sched_task_group
= tg
;
7644 #ifdef CONFIG_FAIR_GROUP_SCHED
7645 if (tsk
->sched_class
->task_move_group
)
7646 tsk
->sched_class
->task_move_group(tsk
);
7649 set_task_rq(tsk
, task_cpu(tsk
));
7651 if (unlikely(running
))
7652 tsk
->sched_class
->set_curr_task(rq
);
7654 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7656 task_rq_unlock(rq
, tsk
, &flags
);
7658 #endif /* CONFIG_CGROUP_SCHED */
7660 #ifdef CONFIG_RT_GROUP_SCHED
7662 * Ensure that the real time constraints are schedulable.
7664 static DEFINE_MUTEX(rt_constraints_mutex
);
7666 /* Must be called with tasklist_lock held */
7667 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7669 struct task_struct
*g
, *p
;
7672 * Autogroups do not have RT tasks; see autogroup_create().
7674 if (task_group_is_autogroup(tg
))
7677 for_each_process_thread(g
, p
) {
7678 if (rt_task(p
) && task_group(p
) == tg
)
7685 struct rt_schedulable_data
{
7686 struct task_group
*tg
;
7691 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7693 struct rt_schedulable_data
*d
= data
;
7694 struct task_group
*child
;
7695 unsigned long total
, sum
= 0;
7696 u64 period
, runtime
;
7698 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7699 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7702 period
= d
->rt_period
;
7703 runtime
= d
->rt_runtime
;
7707 * Cannot have more runtime than the period.
7709 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7713 * Ensure we don't starve existing RT tasks.
7715 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7718 total
= to_ratio(period
, runtime
);
7721 * Nobody can have more than the global setting allows.
7723 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7727 * The sum of our children's runtime should not exceed our own.
7729 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7730 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7731 runtime
= child
->rt_bandwidth
.rt_runtime
;
7733 if (child
== d
->tg
) {
7734 period
= d
->rt_period
;
7735 runtime
= d
->rt_runtime
;
7738 sum
+= to_ratio(period
, runtime
);
7747 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7751 struct rt_schedulable_data data
= {
7753 .rt_period
= period
,
7754 .rt_runtime
= runtime
,
7758 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7764 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7765 u64 rt_period
, u64 rt_runtime
)
7770 * Disallowing the root group RT runtime is BAD, it would disallow the
7771 * kernel creating (and or operating) RT threads.
7773 if (tg
== &root_task_group
&& rt_runtime
== 0)
7776 /* No period doesn't make any sense. */
7780 mutex_lock(&rt_constraints_mutex
);
7781 read_lock(&tasklist_lock
);
7782 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7786 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7787 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7788 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7790 for_each_possible_cpu(i
) {
7791 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7793 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7794 rt_rq
->rt_runtime
= rt_runtime
;
7795 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7797 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7799 read_unlock(&tasklist_lock
);
7800 mutex_unlock(&rt_constraints_mutex
);
7805 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7807 u64 rt_runtime
, rt_period
;
7809 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7810 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7811 if (rt_runtime_us
< 0)
7812 rt_runtime
= RUNTIME_INF
;
7814 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7817 static long sched_group_rt_runtime(struct task_group
*tg
)
7821 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7824 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7825 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7826 return rt_runtime_us
;
7829 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7831 u64 rt_runtime
, rt_period
;
7833 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7834 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7836 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7839 static long sched_group_rt_period(struct task_group
*tg
)
7843 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7844 do_div(rt_period_us
, NSEC_PER_USEC
);
7845 return rt_period_us
;
7847 #endif /* CONFIG_RT_GROUP_SCHED */
7849 #ifdef CONFIG_RT_GROUP_SCHED
7850 static int sched_rt_global_constraints(void)
7854 mutex_lock(&rt_constraints_mutex
);
7855 read_lock(&tasklist_lock
);
7856 ret
= __rt_schedulable(NULL
, 0, 0);
7857 read_unlock(&tasklist_lock
);
7858 mutex_unlock(&rt_constraints_mutex
);
7863 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7865 /* Don't accept realtime tasks when there is no way for them to run */
7866 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7872 #else /* !CONFIG_RT_GROUP_SCHED */
7873 static int sched_rt_global_constraints(void)
7875 unsigned long flags
;
7878 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7879 for_each_possible_cpu(i
) {
7880 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7882 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7883 rt_rq
->rt_runtime
= global_rt_runtime();
7884 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7886 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7890 #endif /* CONFIG_RT_GROUP_SCHED */
7892 static int sched_dl_global_validate(void)
7894 u64 runtime
= global_rt_runtime();
7895 u64 period
= global_rt_period();
7896 u64 new_bw
= to_ratio(period
, runtime
);
7899 unsigned long flags
;
7902 * Here we want to check the bandwidth not being set to some
7903 * value smaller than the currently allocated bandwidth in
7904 * any of the root_domains.
7906 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7907 * cycling on root_domains... Discussion on different/better
7908 * solutions is welcome!
7910 for_each_possible_cpu(cpu
) {
7911 rcu_read_lock_sched();
7912 dl_b
= dl_bw_of(cpu
);
7914 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7915 if (new_bw
< dl_b
->total_bw
)
7917 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7919 rcu_read_unlock_sched();
7928 static void sched_dl_do_global(void)
7933 unsigned long flags
;
7935 def_dl_bandwidth
.dl_period
= global_rt_period();
7936 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7938 if (global_rt_runtime() != RUNTIME_INF
)
7939 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7942 * FIXME: As above...
7944 for_each_possible_cpu(cpu
) {
7945 rcu_read_lock_sched();
7946 dl_b
= dl_bw_of(cpu
);
7948 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7950 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7952 rcu_read_unlock_sched();
7956 static int sched_rt_global_validate(void)
7958 if (sysctl_sched_rt_period
<= 0)
7961 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7962 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7968 static void sched_rt_do_global(void)
7970 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7971 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7974 int sched_rt_handler(struct ctl_table
*table
, int write
,
7975 void __user
*buffer
, size_t *lenp
,
7978 int old_period
, old_runtime
;
7979 static DEFINE_MUTEX(mutex
);
7983 old_period
= sysctl_sched_rt_period
;
7984 old_runtime
= sysctl_sched_rt_runtime
;
7986 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7988 if (!ret
&& write
) {
7989 ret
= sched_rt_global_validate();
7993 ret
= sched_dl_global_validate();
7997 ret
= sched_rt_global_constraints();
8001 sched_rt_do_global();
8002 sched_dl_do_global();
8006 sysctl_sched_rt_period
= old_period
;
8007 sysctl_sched_rt_runtime
= old_runtime
;
8009 mutex_unlock(&mutex
);
8014 int sched_rr_handler(struct ctl_table
*table
, int write
,
8015 void __user
*buffer
, size_t *lenp
,
8019 static DEFINE_MUTEX(mutex
);
8022 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8023 /* make sure that internally we keep jiffies */
8024 /* also, writing zero resets timeslice to default */
8025 if (!ret
&& write
) {
8026 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8027 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8029 mutex_unlock(&mutex
);
8033 #ifdef CONFIG_CGROUP_SCHED
8035 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8037 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8040 static struct cgroup_subsys_state
*
8041 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8043 struct task_group
*parent
= css_tg(parent_css
);
8044 struct task_group
*tg
;
8047 /* This is early initialization for the top cgroup */
8048 return &root_task_group
.css
;
8051 tg
= sched_create_group(parent
);
8053 return ERR_PTR(-ENOMEM
);
8055 sched_online_group(tg
, parent
);
8060 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8062 struct task_group
*tg
= css_tg(css
);
8064 sched_offline_group(tg
);
8067 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8069 struct task_group
*tg
= css_tg(css
);
8072 * Relies on the RCU grace period between css_released() and this.
8074 sched_free_group(tg
);
8077 static void cpu_cgroup_fork(struct task_struct
*task
)
8079 sched_move_task(task
);
8082 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8084 struct task_struct
*task
;
8085 struct cgroup_subsys_state
*css
;
8087 cgroup_taskset_for_each(task
, css
, tset
) {
8088 #ifdef CONFIG_RT_GROUP_SCHED
8089 if (!sched_rt_can_attach(css_tg(css
), task
))
8092 /* We don't support RT-tasks being in separate groups */
8093 if (task
->sched_class
!= &fair_sched_class
)
8100 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8102 struct task_struct
*task
;
8103 struct cgroup_subsys_state
*css
;
8105 cgroup_taskset_for_each(task
, css
, tset
)
8106 sched_move_task(task
);
8109 #ifdef CONFIG_FAIR_GROUP_SCHED
8110 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8111 struct cftype
*cftype
, u64 shareval
)
8113 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8116 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8119 struct task_group
*tg
= css_tg(css
);
8121 return (u64
) scale_load_down(tg
->shares
);
8124 #ifdef CONFIG_CFS_BANDWIDTH
8125 static DEFINE_MUTEX(cfs_constraints_mutex
);
8127 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8128 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8130 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8132 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8134 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8135 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8137 if (tg
== &root_task_group
)
8141 * Ensure we have at some amount of bandwidth every period. This is
8142 * to prevent reaching a state of large arrears when throttled via
8143 * entity_tick() resulting in prolonged exit starvation.
8145 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8149 * Likewise, bound things on the otherside by preventing insane quota
8150 * periods. This also allows us to normalize in computing quota
8153 if (period
> max_cfs_quota_period
)
8157 * Prevent race between setting of cfs_rq->runtime_enabled and
8158 * unthrottle_offline_cfs_rqs().
8161 mutex_lock(&cfs_constraints_mutex
);
8162 ret
= __cfs_schedulable(tg
, period
, quota
);
8166 runtime_enabled
= quota
!= RUNTIME_INF
;
8167 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8169 * If we need to toggle cfs_bandwidth_used, off->on must occur
8170 * before making related changes, and on->off must occur afterwards
8172 if (runtime_enabled
&& !runtime_was_enabled
)
8173 cfs_bandwidth_usage_inc();
8174 raw_spin_lock_irq(&cfs_b
->lock
);
8175 cfs_b
->period
= ns_to_ktime(period
);
8176 cfs_b
->quota
= quota
;
8178 __refill_cfs_bandwidth_runtime(cfs_b
);
8179 /* restart the period timer (if active) to handle new period expiry */
8180 if (runtime_enabled
)
8181 start_cfs_bandwidth(cfs_b
);
8182 raw_spin_unlock_irq(&cfs_b
->lock
);
8184 for_each_online_cpu(i
) {
8185 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8186 struct rq
*rq
= cfs_rq
->rq
;
8188 raw_spin_lock_irq(&rq
->lock
);
8189 cfs_rq
->runtime_enabled
= runtime_enabled
;
8190 cfs_rq
->runtime_remaining
= 0;
8192 if (cfs_rq
->throttled
)
8193 unthrottle_cfs_rq(cfs_rq
);
8194 raw_spin_unlock_irq(&rq
->lock
);
8196 if (runtime_was_enabled
&& !runtime_enabled
)
8197 cfs_bandwidth_usage_dec();
8199 mutex_unlock(&cfs_constraints_mutex
);
8205 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8209 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8210 if (cfs_quota_us
< 0)
8211 quota
= RUNTIME_INF
;
8213 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8215 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8218 long tg_get_cfs_quota(struct task_group
*tg
)
8222 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8225 quota_us
= tg
->cfs_bandwidth
.quota
;
8226 do_div(quota_us
, NSEC_PER_USEC
);
8231 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8235 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8236 quota
= tg
->cfs_bandwidth
.quota
;
8238 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8241 long tg_get_cfs_period(struct task_group
*tg
)
8245 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8246 do_div(cfs_period_us
, NSEC_PER_USEC
);
8248 return cfs_period_us
;
8251 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8254 return tg_get_cfs_quota(css_tg(css
));
8257 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8258 struct cftype
*cftype
, s64 cfs_quota_us
)
8260 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8263 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8266 return tg_get_cfs_period(css_tg(css
));
8269 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8270 struct cftype
*cftype
, u64 cfs_period_us
)
8272 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8275 struct cfs_schedulable_data
{
8276 struct task_group
*tg
;
8281 * normalize group quota/period to be quota/max_period
8282 * note: units are usecs
8284 static u64
normalize_cfs_quota(struct task_group
*tg
,
8285 struct cfs_schedulable_data
*d
)
8293 period
= tg_get_cfs_period(tg
);
8294 quota
= tg_get_cfs_quota(tg
);
8297 /* note: these should typically be equivalent */
8298 if (quota
== RUNTIME_INF
|| quota
== -1)
8301 return to_ratio(period
, quota
);
8304 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8306 struct cfs_schedulable_data
*d
= data
;
8307 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8308 s64 quota
= 0, parent_quota
= -1;
8311 quota
= RUNTIME_INF
;
8313 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8315 quota
= normalize_cfs_quota(tg
, d
);
8316 parent_quota
= parent_b
->hierarchical_quota
;
8319 * ensure max(child_quota) <= parent_quota, inherit when no
8322 if (quota
== RUNTIME_INF
)
8323 quota
= parent_quota
;
8324 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8327 cfs_b
->hierarchical_quota
= quota
;
8332 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8335 struct cfs_schedulable_data data
= {
8341 if (quota
!= RUNTIME_INF
) {
8342 do_div(data
.period
, NSEC_PER_USEC
);
8343 do_div(data
.quota
, NSEC_PER_USEC
);
8347 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8353 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8355 struct task_group
*tg
= css_tg(seq_css(sf
));
8356 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8358 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8359 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8360 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8364 #endif /* CONFIG_CFS_BANDWIDTH */
8365 #endif /* CONFIG_FAIR_GROUP_SCHED */
8367 #ifdef CONFIG_RT_GROUP_SCHED
8368 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8369 struct cftype
*cft
, s64 val
)
8371 return sched_group_set_rt_runtime(css_tg(css
), val
);
8374 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8377 return sched_group_rt_runtime(css_tg(css
));
8380 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8381 struct cftype
*cftype
, u64 rt_period_us
)
8383 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8386 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8389 return sched_group_rt_period(css_tg(css
));
8391 #endif /* CONFIG_RT_GROUP_SCHED */
8393 static struct cftype cpu_files
[] = {
8394 #ifdef CONFIG_FAIR_GROUP_SCHED
8397 .read_u64
= cpu_shares_read_u64
,
8398 .write_u64
= cpu_shares_write_u64
,
8401 #ifdef CONFIG_CFS_BANDWIDTH
8403 .name
= "cfs_quota_us",
8404 .read_s64
= cpu_cfs_quota_read_s64
,
8405 .write_s64
= cpu_cfs_quota_write_s64
,
8408 .name
= "cfs_period_us",
8409 .read_u64
= cpu_cfs_period_read_u64
,
8410 .write_u64
= cpu_cfs_period_write_u64
,
8414 .seq_show
= cpu_stats_show
,
8417 #ifdef CONFIG_RT_GROUP_SCHED
8419 .name
= "rt_runtime_us",
8420 .read_s64
= cpu_rt_runtime_read
,
8421 .write_s64
= cpu_rt_runtime_write
,
8424 .name
= "rt_period_us",
8425 .read_u64
= cpu_rt_period_read_uint
,
8426 .write_u64
= cpu_rt_period_write_uint
,
8432 struct cgroup_subsys cpu_cgrp_subsys
= {
8433 .css_alloc
= cpu_cgroup_css_alloc
,
8434 .css_released
= cpu_cgroup_css_released
,
8435 .css_free
= cpu_cgroup_css_free
,
8436 .fork
= cpu_cgroup_fork
,
8437 .can_attach
= cpu_cgroup_can_attach
,
8438 .attach
= cpu_cgroup_attach
,
8439 .legacy_cftypes
= cpu_files
,
8443 #endif /* CONFIG_CGROUP_SCHED */
8445 void dump_cpu_task(int cpu
)
8447 pr_info("Task dump for CPU %d:\n", cpu
);
8448 sched_show_task(cpu_curr(cpu
));
8452 * Nice levels are multiplicative, with a gentle 10% change for every
8453 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8454 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8455 * that remained on nice 0.
8457 * The "10% effect" is relative and cumulative: from _any_ nice level,
8458 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8459 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8460 * If a task goes up by ~10% and another task goes down by ~10% then
8461 * the relative distance between them is ~25%.)
8463 const int sched_prio_to_weight
[40] = {
8464 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8465 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8466 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8467 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8468 /* 0 */ 1024, 820, 655, 526, 423,
8469 /* 5 */ 335, 272, 215, 172, 137,
8470 /* 10 */ 110, 87, 70, 56, 45,
8471 /* 15 */ 36, 29, 23, 18, 15,
8475 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8477 * In cases where the weight does not change often, we can use the
8478 * precalculated inverse to speed up arithmetics by turning divisions
8479 * into multiplications:
8481 const u32 sched_prio_to_wmult
[40] = {
8482 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8483 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8484 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8485 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8486 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8487 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8488 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8489 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,