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
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.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/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
95 ktime_t soft
, hard
, now
;
98 if (hrtimer_active(period_timer
))
101 now
= hrtimer_cb_get_time(period_timer
);
102 hrtimer_forward(period_timer
, now
, period
);
104 soft
= hrtimer_get_softexpires(period_timer
);
105 hard
= hrtimer_get_expires(period_timer
);
106 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
107 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
108 HRTIMER_MODE_ABS_PINNED
, 0);
112 DEFINE_MUTEX(sched_domains_mutex
);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
115 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
117 void update_rq_clock(struct rq
*rq
)
121 if (rq
->skip_clock_update
> 0)
124 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
126 update_rq_clock_task(rq
, delta
);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug
unsigned int sysctl_sched_features
=
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names
[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp
)
201 if (strncmp(cmp
, "NO_", 3) == 0) {
206 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
207 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
209 sysctl_sched_features
&= ~(1UL << i
);
210 sched_feat_disable(i
);
212 sysctl_sched_features
|= (1UL << i
);
213 sched_feat_enable(i
);
223 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
224 size_t cnt
, loff_t
*ppos
)
233 if (copy_from_user(&buf
, ubuf
, cnt
))
239 i
= sched_feat_set(cmp
);
240 if (i
== __SCHED_FEAT_NR
)
248 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
250 return single_open(filp
, sched_feat_show
, NULL
);
253 static const struct file_operations sched_feat_fops
= {
254 .open
= sched_feat_open
,
255 .write
= sched_feat_write
,
258 .release
= single_release
,
261 static __init
int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
268 late_initcall(sched_init_debug
);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
278 * period over which we average the RT time consumption, measured
283 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period
= 1000000;
291 __read_mostly
int scheduler_running
;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime
= 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
309 lockdep_assert_held(&p
->pi_lock
);
313 raw_spin_lock(&rq
->lock
);
314 if (likely(rq
== task_rq(p
)))
316 raw_spin_unlock(&rq
->lock
);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
324 __acquires(p
->pi_lock
)
330 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
332 raw_spin_lock(&rq
->lock
);
333 if (likely(rq
== task_rq(p
)))
335 raw_spin_unlock(&rq
->lock
);
336 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
340 static void __task_rq_unlock(struct rq
*rq
)
343 raw_spin_unlock(&rq
->lock
);
347 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
349 __releases(p
->pi_lock
)
351 raw_spin_unlock(&rq
->lock
);
352 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq
*this_rq_lock(void)
365 raw_spin_lock(&rq
->lock
);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq
*rq
)
384 if (hrtimer_active(&rq
->hrtick_timer
))
385 hrtimer_cancel(&rq
->hrtick_timer
);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
394 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
396 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
398 raw_spin_lock(&rq
->lock
);
400 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
401 raw_spin_unlock(&rq
->lock
);
403 return HRTIMER_NORESTART
;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg
)
414 raw_spin_lock(&rq
->lock
);
415 hrtimer_restart(&rq
->hrtick_timer
);
416 rq
->hrtick_csd_pending
= 0;
417 raw_spin_unlock(&rq
->lock
);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq
*rq
, u64 delay
)
427 struct hrtimer
*timer
= &rq
->hrtick_timer
;
428 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
430 hrtimer_set_expires(timer
, time
);
432 if (rq
== this_rq()) {
433 hrtimer_restart(timer
);
434 } else if (!rq
->hrtick_csd_pending
) {
435 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
436 rq
->hrtick_csd_pending
= 1;
441 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
443 int cpu
= (int)(long)hcpu
;
446 case CPU_UP_CANCELED
:
447 case CPU_UP_CANCELED_FROZEN
:
448 case CPU_DOWN_PREPARE
:
449 case CPU_DOWN_PREPARE_FROZEN
:
451 case CPU_DEAD_FROZEN
:
452 hrtick_clear(cpu_rq(cpu
));
459 static __init
void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick
, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq
*rq
, u64 delay
)
471 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
472 HRTIMER_MODE_REL_PINNED
, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq
*rq
)
483 rq
->hrtick_csd_pending
= 0;
485 rq
->hrtick_csd
.flags
= 0;
486 rq
->hrtick_csd
.func
= __hrtick_start
;
487 rq
->hrtick_csd
.info
= rq
;
490 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
491 rq
->hrtick_timer
.function
= hrtick
;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq
*rq
)
498 static inline void init_rq_hrtick(struct rq
*rq
)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
516 #ifndef tsk_is_polling
517 #define tsk_is_polling(t) 0
520 void resched_task(struct task_struct
*p
)
524 assert_raw_spin_locked(&task_rq(p
)->lock
);
526 if (test_tsk_need_resched(p
))
529 set_tsk_need_resched(p
);
532 if (cpu
== smp_processor_id())
535 /* NEED_RESCHED must be visible before we test polling */
537 if (!tsk_is_polling(p
))
538 smp_send_reschedule(cpu
);
541 void resched_cpu(int cpu
)
543 struct rq
*rq
= cpu_rq(cpu
);
546 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
548 resched_task(cpu_curr(cpu
));
549 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
554 * In the semi idle case, use the nearest busy cpu for migrating timers
555 * from an idle cpu. This is good for power-savings.
557 * We don't do similar optimization for completely idle system, as
558 * selecting an idle cpu will add more delays to the timers than intended
559 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
561 int get_nohz_timer_target(void)
563 int cpu
= smp_processor_id();
565 struct sched_domain
*sd
;
568 for_each_domain(cpu
, sd
) {
569 for_each_cpu(i
, sched_domain_span(sd
)) {
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 static void wake_up_idle_cpu(int cpu
)
592 struct rq
*rq
= cpu_rq(cpu
);
594 if (cpu
== smp_processor_id())
598 * This is safe, as this function is called with the timer
599 * wheel base lock of (cpu) held. When the CPU is on the way
600 * to idle and has not yet set rq->curr to idle then it will
601 * be serialized on the timer wheel base lock and take the new
602 * timer into account automatically.
604 if (rq
->curr
!= rq
->idle
)
608 * We can set TIF_RESCHED on the idle task of the other CPU
609 * lockless. The worst case is that the other CPU runs the
610 * idle task through an additional NOOP schedule()
612 set_tsk_need_resched(rq
->idle
);
614 /* NEED_RESCHED must be visible before we test polling */
616 if (!tsk_is_polling(rq
->idle
))
617 smp_send_reschedule(cpu
);
620 static bool wake_up_extended_nohz_cpu(int cpu
)
622 if (tick_nohz_extended_cpu(cpu
)) {
623 if (cpu
!= smp_processor_id() ||
624 tick_nohz_tick_stopped())
625 smp_send_reschedule(cpu
);
632 void wake_up_nohz_cpu(int cpu
)
634 if (!wake_up_extended_nohz_cpu(cpu
))
635 wake_up_idle_cpu(cpu
);
638 static inline bool got_nohz_idle_kick(void)
640 int cpu
= smp_processor_id();
641 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
644 #else /* CONFIG_NO_HZ */
646 static inline bool got_nohz_idle_kick(void)
651 #endif /* CONFIG_NO_HZ */
653 void sched_avg_update(struct rq
*rq
)
655 s64 period
= sched_avg_period();
657 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
659 * Inline assembly required to prevent the compiler
660 * optimising this loop into a divmod call.
661 * See __iter_div_u64_rem() for another example of this.
663 asm("" : "+rm" (rq
->age_stamp
));
664 rq
->age_stamp
+= period
;
669 #else /* !CONFIG_SMP */
670 void resched_task(struct task_struct
*p
)
672 assert_raw_spin_locked(&task_rq(p
)->lock
);
673 set_tsk_need_resched(p
);
675 #endif /* CONFIG_SMP */
677 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
678 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
680 * Iterate task_group tree rooted at *from, calling @down when first entering a
681 * node and @up when leaving it for the final time.
683 * Caller must hold rcu_lock or sufficient equivalent.
685 int walk_tg_tree_from(struct task_group
*from
,
686 tg_visitor down
, tg_visitor up
, void *data
)
688 struct task_group
*parent
, *child
;
694 ret
= (*down
)(parent
, data
);
697 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
704 ret
= (*up
)(parent
, data
);
705 if (ret
|| parent
== from
)
709 parent
= parent
->parent
;
716 int tg_nop(struct task_group
*tg
, void *data
)
722 static void set_load_weight(struct task_struct
*p
)
724 int prio
= p
->static_prio
- MAX_RT_PRIO
;
725 struct load_weight
*load
= &p
->se
.load
;
728 * SCHED_IDLE tasks get minimal weight:
730 if (p
->policy
== SCHED_IDLE
) {
731 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
732 load
->inv_weight
= WMULT_IDLEPRIO
;
736 load
->weight
= scale_load(prio_to_weight
[prio
]);
737 load
->inv_weight
= prio_to_wmult
[prio
];
740 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
743 sched_info_queued(p
);
744 p
->sched_class
->enqueue_task(rq
, p
, flags
);
747 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
750 sched_info_dequeued(p
);
751 p
->sched_class
->dequeue_task(rq
, p
, flags
);
754 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
756 if (task_contributes_to_load(p
))
757 rq
->nr_uninterruptible
--;
759 enqueue_task(rq
, p
, flags
);
762 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
764 if (task_contributes_to_load(p
))
765 rq
->nr_uninterruptible
++;
767 dequeue_task(rq
, p
, flags
);
770 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
773 * In theory, the compile should just see 0 here, and optimize out the call
774 * to sched_rt_avg_update. But I don't trust it...
776 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
777 s64 steal
= 0, irq_delta
= 0;
779 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
780 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
783 * Since irq_time is only updated on {soft,}irq_exit, we might run into
784 * this case when a previous update_rq_clock() happened inside a
787 * When this happens, we stop ->clock_task and only update the
788 * prev_irq_time stamp to account for the part that fit, so that a next
789 * update will consume the rest. This ensures ->clock_task is
792 * It does however cause some slight miss-attribution of {soft,}irq
793 * time, a more accurate solution would be to update the irq_time using
794 * the current rq->clock timestamp, except that would require using
797 if (irq_delta
> delta
)
800 rq
->prev_irq_time
+= irq_delta
;
803 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
804 if (static_key_false((¶virt_steal_rq_enabled
))) {
807 steal
= paravirt_steal_clock(cpu_of(rq
));
808 steal
-= rq
->prev_steal_time_rq
;
810 if (unlikely(steal
> delta
))
813 st
= steal_ticks(steal
);
814 steal
= st
* TICK_NSEC
;
816 rq
->prev_steal_time_rq
+= steal
;
822 rq
->clock_task
+= delta
;
824 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
825 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
826 sched_rt_avg_update(rq
, irq_delta
+ steal
);
830 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
832 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
833 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
837 * Make it appear like a SCHED_FIFO task, its something
838 * userspace knows about and won't get confused about.
840 * Also, it will make PI more or less work without too
841 * much confusion -- but then, stop work should not
842 * rely on PI working anyway.
844 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
846 stop
->sched_class
= &stop_sched_class
;
849 cpu_rq(cpu
)->stop
= stop
;
853 * Reset it back to a normal scheduling class so that
854 * it can die in pieces.
856 old_stop
->sched_class
= &rt_sched_class
;
861 * __normal_prio - return the priority that is based on the static prio
863 static inline int __normal_prio(struct task_struct
*p
)
865 return p
->static_prio
;
869 * Calculate the expected normal priority: i.e. priority
870 * without taking RT-inheritance into account. Might be
871 * boosted by interactivity modifiers. Changes upon fork,
872 * setprio syscalls, and whenever the interactivity
873 * estimator recalculates.
875 static inline int normal_prio(struct task_struct
*p
)
879 if (task_has_rt_policy(p
))
880 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
882 prio
= __normal_prio(p
);
887 * Calculate the current priority, i.e. the priority
888 * taken into account by the scheduler. This value might
889 * be boosted by RT tasks, or might be boosted by
890 * interactivity modifiers. Will be RT if the task got
891 * RT-boosted. If not then it returns p->normal_prio.
893 static int effective_prio(struct task_struct
*p
)
895 p
->normal_prio
= normal_prio(p
);
897 * If we are RT tasks or we were boosted to RT priority,
898 * keep the priority unchanged. Otherwise, update priority
899 * to the normal priority:
901 if (!rt_prio(p
->prio
))
902 return p
->normal_prio
;
907 * task_curr - is this task currently executing on a CPU?
908 * @p: the task in question.
910 inline int task_curr(const struct task_struct
*p
)
912 return cpu_curr(task_cpu(p
)) == p
;
915 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
916 const struct sched_class
*prev_class
,
919 if (prev_class
!= p
->sched_class
) {
920 if (prev_class
->switched_from
)
921 prev_class
->switched_from(rq
, p
);
922 p
->sched_class
->switched_to(rq
, p
);
923 } else if (oldprio
!= p
->prio
)
924 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
927 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
929 const struct sched_class
*class;
931 if (p
->sched_class
== rq
->curr
->sched_class
) {
932 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
934 for_each_class(class) {
935 if (class == rq
->curr
->sched_class
)
937 if (class == p
->sched_class
) {
938 resched_task(rq
->curr
);
945 * A queue event has occurred, and we're going to schedule. In
946 * this case, we can save a useless back to back clock update.
948 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
949 rq
->skip_clock_update
= 1;
952 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
954 void register_task_migration_notifier(struct notifier_block
*n
)
956 atomic_notifier_chain_register(&task_migration_notifier
, n
);
960 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
962 #ifdef CONFIG_SCHED_DEBUG
964 * We should never call set_task_cpu() on a blocked task,
965 * ttwu() will sort out the placement.
967 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
968 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
970 #ifdef CONFIG_LOCKDEP
972 * The caller should hold either p->pi_lock or rq->lock, when changing
973 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
975 * sched_move_task() holds both and thus holding either pins the cgroup,
978 * Furthermore, all task_rq users should acquire both locks, see
981 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
982 lockdep_is_held(&task_rq(p
)->lock
)));
986 trace_sched_migrate_task(p
, new_cpu
);
988 if (task_cpu(p
) != new_cpu
) {
989 struct task_migration_notifier tmn
;
991 if (p
->sched_class
->migrate_task_rq
)
992 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
993 p
->se
.nr_migrations
++;
994 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
997 tmn
.from_cpu
= task_cpu(p
);
998 tmn
.to_cpu
= new_cpu
;
1000 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
1003 __set_task_cpu(p
, new_cpu
);
1006 struct migration_arg
{
1007 struct task_struct
*task
;
1011 static int migration_cpu_stop(void *data
);
1014 * wait_task_inactive - wait for a thread to unschedule.
1016 * If @match_state is nonzero, it's the @p->state value just checked and
1017 * not expected to change. If it changes, i.e. @p might have woken up,
1018 * then return zero. When we succeed in waiting for @p to be off its CPU,
1019 * we return a positive number (its total switch count). If a second call
1020 * a short while later returns the same number, the caller can be sure that
1021 * @p has remained unscheduled the whole time.
1023 * The caller must ensure that the task *will* unschedule sometime soon,
1024 * else this function might spin for a *long* time. This function can't
1025 * be called with interrupts off, or it may introduce deadlock with
1026 * smp_call_function() if an IPI is sent by the same process we are
1027 * waiting to become inactive.
1029 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1031 unsigned long flags
;
1038 * We do the initial early heuristics without holding
1039 * any task-queue locks at all. We'll only try to get
1040 * the runqueue lock when things look like they will
1046 * If the task is actively running on another CPU
1047 * still, just relax and busy-wait without holding
1050 * NOTE! Since we don't hold any locks, it's not
1051 * even sure that "rq" stays as the right runqueue!
1052 * But we don't care, since "task_running()" will
1053 * return false if the runqueue has changed and p
1054 * is actually now running somewhere else!
1056 while (task_running(rq
, p
)) {
1057 if (match_state
&& unlikely(p
->state
!= match_state
))
1063 * Ok, time to look more closely! We need the rq
1064 * lock now, to be *sure*. If we're wrong, we'll
1065 * just go back and repeat.
1067 rq
= task_rq_lock(p
, &flags
);
1068 trace_sched_wait_task(p
);
1069 running
= task_running(rq
, p
);
1072 if (!match_state
|| p
->state
== match_state
)
1073 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1074 task_rq_unlock(rq
, p
, &flags
);
1077 * If it changed from the expected state, bail out now.
1079 if (unlikely(!ncsw
))
1083 * Was it really running after all now that we
1084 * checked with the proper locks actually held?
1086 * Oops. Go back and try again..
1088 if (unlikely(running
)) {
1094 * It's not enough that it's not actively running,
1095 * it must be off the runqueue _entirely_, and not
1098 * So if it was still runnable (but just not actively
1099 * running right now), it's preempted, and we should
1100 * yield - it could be a while.
1102 if (unlikely(on_rq
)) {
1103 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1105 set_current_state(TASK_UNINTERRUPTIBLE
);
1106 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1111 * Ahh, all good. It wasn't running, and it wasn't
1112 * runnable, which means that it will never become
1113 * running in the future either. We're all done!
1122 * kick_process - kick a running thread to enter/exit the kernel
1123 * @p: the to-be-kicked thread
1125 * Cause a process which is running on another CPU to enter
1126 * kernel-mode, without any delay. (to get signals handled.)
1128 * NOTE: this function doesn't have to take the runqueue lock,
1129 * because all it wants to ensure is that the remote task enters
1130 * the kernel. If the IPI races and the task has been migrated
1131 * to another CPU then no harm is done and the purpose has been
1134 void kick_process(struct task_struct
*p
)
1140 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1141 smp_send_reschedule(cpu
);
1144 EXPORT_SYMBOL_GPL(kick_process
);
1145 #endif /* CONFIG_SMP */
1149 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1151 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1153 int nid
= cpu_to_node(cpu
);
1154 const struct cpumask
*nodemask
= NULL
;
1155 enum { cpuset
, possible
, fail
} state
= cpuset
;
1159 * If the node that the cpu is on has been offlined, cpu_to_node()
1160 * will return -1. There is no cpu on the node, and we should
1161 * select the cpu on the other node.
1164 nodemask
= cpumask_of_node(nid
);
1166 /* Look for allowed, online CPU in same node. */
1167 for_each_cpu(dest_cpu
, nodemask
) {
1168 if (!cpu_online(dest_cpu
))
1170 if (!cpu_active(dest_cpu
))
1172 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1178 /* Any allowed, online CPU? */
1179 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1180 if (!cpu_online(dest_cpu
))
1182 if (!cpu_active(dest_cpu
))
1189 /* No more Mr. Nice Guy. */
1190 cpuset_cpus_allowed_fallback(p
);
1195 do_set_cpus_allowed(p
, cpu_possible_mask
);
1206 if (state
!= cpuset
) {
1208 * Don't tell them about moving exiting tasks or
1209 * kernel threads (both mm NULL), since they never
1212 if (p
->mm
&& printk_ratelimit()) {
1213 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1214 task_pid_nr(p
), p
->comm
, cpu
);
1222 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1225 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1227 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1230 * In order not to call set_task_cpu() on a blocking task we need
1231 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1234 * Since this is common to all placement strategies, this lives here.
1236 * [ this allows ->select_task() to simply return task_cpu(p) and
1237 * not worry about this generic constraint ]
1239 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1241 cpu
= select_fallback_rq(task_cpu(p
), p
);
1246 static void update_avg(u64
*avg
, u64 sample
)
1248 s64 diff
= sample
- *avg
;
1254 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1256 #ifdef CONFIG_SCHEDSTATS
1257 struct rq
*rq
= this_rq();
1260 int this_cpu
= smp_processor_id();
1262 if (cpu
== this_cpu
) {
1263 schedstat_inc(rq
, ttwu_local
);
1264 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1266 struct sched_domain
*sd
;
1268 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1270 for_each_domain(this_cpu
, sd
) {
1271 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1272 schedstat_inc(sd
, ttwu_wake_remote
);
1279 if (wake_flags
& WF_MIGRATED
)
1280 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1282 #endif /* CONFIG_SMP */
1284 schedstat_inc(rq
, ttwu_count
);
1285 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1287 if (wake_flags
& WF_SYNC
)
1288 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1290 #endif /* CONFIG_SCHEDSTATS */
1293 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1295 activate_task(rq
, p
, en_flags
);
1298 /* if a worker is waking up, notify workqueue */
1299 if (p
->flags
& PF_WQ_WORKER
)
1300 wq_worker_waking_up(p
, cpu_of(rq
));
1304 * Mark the task runnable and perform wakeup-preemption.
1307 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1309 check_preempt_curr(rq
, p
, wake_flags
);
1310 trace_sched_wakeup(p
, true);
1312 p
->state
= TASK_RUNNING
;
1314 if (p
->sched_class
->task_woken
)
1315 p
->sched_class
->task_woken(rq
, p
);
1317 if (rq
->idle_stamp
) {
1318 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1319 u64 max
= 2*sysctl_sched_migration_cost
;
1324 update_avg(&rq
->avg_idle
, delta
);
1331 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1334 if (p
->sched_contributes_to_load
)
1335 rq
->nr_uninterruptible
--;
1338 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1339 ttwu_do_wakeup(rq
, p
, wake_flags
);
1343 * Called in case the task @p isn't fully descheduled from its runqueue,
1344 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1345 * since all we need to do is flip p->state to TASK_RUNNING, since
1346 * the task is still ->on_rq.
1348 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1353 rq
= __task_rq_lock(p
);
1355 ttwu_do_wakeup(rq
, p
, wake_flags
);
1358 __task_rq_unlock(rq
);
1364 static void sched_ttwu_pending(void)
1366 struct rq
*rq
= this_rq();
1367 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1368 struct task_struct
*p
;
1370 raw_spin_lock(&rq
->lock
);
1373 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1374 llist
= llist_next(llist
);
1375 ttwu_do_activate(rq
, p
, 0);
1378 raw_spin_unlock(&rq
->lock
);
1381 void scheduler_ipi(void)
1383 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1387 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1388 * traditionally all their work was done from the interrupt return
1389 * path. Now that we actually do some work, we need to make sure
1392 * Some archs already do call them, luckily irq_enter/exit nest
1395 * Arguably we should visit all archs and update all handlers,
1396 * however a fair share of IPIs are still resched only so this would
1397 * somewhat pessimize the simple resched case.
1400 sched_ttwu_pending();
1403 * Check if someone kicked us for doing the nohz idle load balance.
1405 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1406 this_rq()->idle_balance
= 1;
1407 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1412 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1414 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1415 smp_send_reschedule(cpu
);
1418 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1420 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1422 #endif /* CONFIG_SMP */
1424 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1426 struct rq
*rq
= cpu_rq(cpu
);
1428 #if defined(CONFIG_SMP)
1429 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1430 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1431 ttwu_queue_remote(p
, cpu
);
1436 raw_spin_lock(&rq
->lock
);
1437 ttwu_do_activate(rq
, p
, 0);
1438 raw_spin_unlock(&rq
->lock
);
1442 * try_to_wake_up - wake up a thread
1443 * @p: the thread to be awakened
1444 * @state: the mask of task states that can be woken
1445 * @wake_flags: wake modifier flags (WF_*)
1447 * Put it on the run-queue if it's not already there. The "current"
1448 * thread is always on the run-queue (except when the actual
1449 * re-schedule is in progress), and as such you're allowed to do
1450 * the simpler "current->state = TASK_RUNNING" to mark yourself
1451 * runnable without the overhead of this.
1453 * Returns %true if @p was woken up, %false if it was already running
1454 * or @state didn't match @p's state.
1457 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1459 unsigned long flags
;
1460 int cpu
, success
= 0;
1463 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1464 if (!(p
->state
& state
))
1467 success
= 1; /* we're going to change ->state */
1470 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1475 * If the owning (remote) cpu is still in the middle of schedule() with
1476 * this task as prev, wait until its done referencing the task.
1481 * Pairs with the smp_wmb() in finish_lock_switch().
1485 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1486 p
->state
= TASK_WAKING
;
1488 if (p
->sched_class
->task_waking
)
1489 p
->sched_class
->task_waking(p
);
1491 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1492 if (task_cpu(p
) != cpu
) {
1493 wake_flags
|= WF_MIGRATED
;
1494 set_task_cpu(p
, cpu
);
1496 #endif /* CONFIG_SMP */
1500 ttwu_stat(p
, cpu
, wake_flags
);
1502 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1508 * try_to_wake_up_local - try to wake up a local task with rq lock held
1509 * @p: the thread to be awakened
1511 * Put @p on the run-queue if it's not already there. The caller must
1512 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1515 static void try_to_wake_up_local(struct task_struct
*p
)
1517 struct rq
*rq
= task_rq(p
);
1519 BUG_ON(rq
!= this_rq());
1520 BUG_ON(p
== current
);
1521 lockdep_assert_held(&rq
->lock
);
1523 if (!raw_spin_trylock(&p
->pi_lock
)) {
1524 raw_spin_unlock(&rq
->lock
);
1525 raw_spin_lock(&p
->pi_lock
);
1526 raw_spin_lock(&rq
->lock
);
1529 if (!(p
->state
& TASK_NORMAL
))
1533 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1535 ttwu_do_wakeup(rq
, p
, 0);
1536 ttwu_stat(p
, smp_processor_id(), 0);
1538 raw_spin_unlock(&p
->pi_lock
);
1542 * wake_up_process - Wake up a specific process
1543 * @p: The process to be woken up.
1545 * Attempt to wake up the nominated process and move it to the set of runnable
1546 * processes. Returns 1 if the process was woken up, 0 if it was already
1549 * It may be assumed that this function implies a write memory barrier before
1550 * changing the task state if and only if any tasks are woken up.
1552 int wake_up_process(struct task_struct
*p
)
1554 WARN_ON(task_is_stopped_or_traced(p
));
1555 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1557 EXPORT_SYMBOL(wake_up_process
);
1559 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1561 return try_to_wake_up(p
, state
, 0);
1565 * Perform scheduler related setup for a newly forked process p.
1566 * p is forked by current.
1568 * __sched_fork() is basic setup used by init_idle() too:
1570 static void __sched_fork(struct task_struct
*p
)
1575 p
->se
.exec_start
= 0;
1576 p
->se
.sum_exec_runtime
= 0;
1577 p
->se
.prev_sum_exec_runtime
= 0;
1578 p
->se
.nr_migrations
= 0;
1580 INIT_LIST_HEAD(&p
->se
.group_node
);
1583 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1584 * removed when useful for applications beyond shares distribution (e.g.
1587 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1588 p
->se
.avg
.runnable_avg_period
= 0;
1589 p
->se
.avg
.runnable_avg_sum
= 0;
1591 #ifdef CONFIG_SCHEDSTATS
1592 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1595 INIT_LIST_HEAD(&p
->rt
.run_list
);
1597 #ifdef CONFIG_PREEMPT_NOTIFIERS
1598 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1601 #ifdef CONFIG_NUMA_BALANCING
1602 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1603 p
->mm
->numa_next_scan
= jiffies
;
1604 p
->mm
->numa_next_reset
= jiffies
;
1605 p
->mm
->numa_scan_seq
= 0;
1608 p
->node_stamp
= 0ULL;
1609 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1610 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1611 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1612 p
->numa_work
.next
= &p
->numa_work
;
1613 #endif /* CONFIG_NUMA_BALANCING */
1616 #ifdef CONFIG_NUMA_BALANCING
1617 #ifdef CONFIG_SCHED_DEBUG
1618 void set_numabalancing_state(bool enabled
)
1621 sched_feat_set("NUMA");
1623 sched_feat_set("NO_NUMA");
1626 __read_mostly
bool numabalancing_enabled
;
1628 void set_numabalancing_state(bool enabled
)
1630 numabalancing_enabled
= enabled
;
1632 #endif /* CONFIG_SCHED_DEBUG */
1633 #endif /* CONFIG_NUMA_BALANCING */
1636 * fork()/clone()-time setup:
1638 void sched_fork(struct task_struct
*p
)
1640 unsigned long flags
;
1641 int cpu
= get_cpu();
1645 * We mark the process as running here. This guarantees that
1646 * nobody will actually run it, and a signal or other external
1647 * event cannot wake it up and insert it on the runqueue either.
1649 p
->state
= TASK_RUNNING
;
1652 * Make sure we do not leak PI boosting priority to the child.
1654 p
->prio
= current
->normal_prio
;
1657 * Revert to default priority/policy on fork if requested.
1659 if (unlikely(p
->sched_reset_on_fork
)) {
1660 if (task_has_rt_policy(p
)) {
1661 p
->policy
= SCHED_NORMAL
;
1662 p
->static_prio
= NICE_TO_PRIO(0);
1664 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1665 p
->static_prio
= NICE_TO_PRIO(0);
1667 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1671 * We don't need the reset flag anymore after the fork. It has
1672 * fulfilled its duty:
1674 p
->sched_reset_on_fork
= 0;
1677 if (!rt_prio(p
->prio
))
1678 p
->sched_class
= &fair_sched_class
;
1680 if (p
->sched_class
->task_fork
)
1681 p
->sched_class
->task_fork(p
);
1684 * The child is not yet in the pid-hash so no cgroup attach races,
1685 * and the cgroup is pinned to this child due to cgroup_fork()
1686 * is ran before sched_fork().
1688 * Silence PROVE_RCU.
1690 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1691 set_task_cpu(p
, cpu
);
1692 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1694 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1695 if (likely(sched_info_on()))
1696 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1698 #if defined(CONFIG_SMP)
1701 #ifdef CONFIG_PREEMPT_COUNT
1702 /* Want to start with kernel preemption disabled. */
1703 task_thread_info(p
)->preempt_count
= 1;
1706 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1713 * wake_up_new_task - wake up a newly created task for the first time.
1715 * This function will do some initial scheduler statistics housekeeping
1716 * that must be done for every newly created context, then puts the task
1717 * on the runqueue and wakes it.
1719 void wake_up_new_task(struct task_struct
*p
)
1721 unsigned long flags
;
1724 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1727 * Fork balancing, do it here and not earlier because:
1728 * - cpus_allowed can change in the fork path
1729 * - any previously selected cpu might disappear through hotplug
1731 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1734 rq
= __task_rq_lock(p
);
1735 activate_task(rq
, p
, 0);
1737 trace_sched_wakeup_new(p
, true);
1738 check_preempt_curr(rq
, p
, WF_FORK
);
1740 if (p
->sched_class
->task_woken
)
1741 p
->sched_class
->task_woken(rq
, p
);
1743 task_rq_unlock(rq
, p
, &flags
);
1746 #ifdef CONFIG_PREEMPT_NOTIFIERS
1749 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1750 * @notifier: notifier struct to register
1752 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1754 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1756 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1759 * preempt_notifier_unregister - no longer interested in preemption notifications
1760 * @notifier: notifier struct to unregister
1762 * This is safe to call from within a preemption notifier.
1764 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1766 hlist_del(¬ifier
->link
);
1768 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1770 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1772 struct preempt_notifier
*notifier
;
1774 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1775 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1779 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1780 struct task_struct
*next
)
1782 struct preempt_notifier
*notifier
;
1784 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1785 notifier
->ops
->sched_out(notifier
, next
);
1788 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1790 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1795 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1796 struct task_struct
*next
)
1800 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1803 * prepare_task_switch - prepare to switch tasks
1804 * @rq: the runqueue preparing to switch
1805 * @prev: the current task that is being switched out
1806 * @next: the task we are going to switch to.
1808 * This is called with the rq lock held and interrupts off. It must
1809 * be paired with a subsequent finish_task_switch after the context
1812 * prepare_task_switch sets up locking and calls architecture specific
1816 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1817 struct task_struct
*next
)
1819 trace_sched_switch(prev
, next
);
1820 sched_info_switch(prev
, next
);
1821 perf_event_task_sched_out(prev
, next
);
1822 fire_sched_out_preempt_notifiers(prev
, next
);
1823 prepare_lock_switch(rq
, next
);
1824 prepare_arch_switch(next
);
1828 * finish_task_switch - clean up after a task-switch
1829 * @rq: runqueue associated with task-switch
1830 * @prev: the thread we just switched away from.
1832 * finish_task_switch must be called after the context switch, paired
1833 * with a prepare_task_switch call before the context switch.
1834 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1835 * and do any other architecture-specific cleanup actions.
1837 * Note that we may have delayed dropping an mm in context_switch(). If
1838 * so, we finish that here outside of the runqueue lock. (Doing it
1839 * with the lock held can cause deadlocks; see schedule() for
1842 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1843 __releases(rq
->lock
)
1845 struct mm_struct
*mm
= rq
->prev_mm
;
1851 * A task struct has one reference for the use as "current".
1852 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1853 * schedule one last time. The schedule call will never return, and
1854 * the scheduled task must drop that reference.
1855 * The test for TASK_DEAD must occur while the runqueue locks are
1856 * still held, otherwise prev could be scheduled on another cpu, die
1857 * there before we look at prev->state, and then the reference would
1859 * Manfred Spraul <manfred@colorfullife.com>
1861 prev_state
= prev
->state
;
1862 vtime_task_switch(prev
);
1863 finish_arch_switch(prev
);
1864 perf_event_task_sched_in(prev
, current
);
1865 finish_lock_switch(rq
, prev
);
1866 finish_arch_post_lock_switch();
1868 fire_sched_in_preempt_notifiers(current
);
1871 if (unlikely(prev_state
== TASK_DEAD
)) {
1873 * Remove function-return probe instances associated with this
1874 * task and put them back on the free list.
1876 kprobe_flush_task(prev
);
1877 put_task_struct(prev
);
1883 /* assumes rq->lock is held */
1884 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1886 if (prev
->sched_class
->pre_schedule
)
1887 prev
->sched_class
->pre_schedule(rq
, prev
);
1890 /* rq->lock is NOT held, but preemption is disabled */
1891 static inline void post_schedule(struct rq
*rq
)
1893 if (rq
->post_schedule
) {
1894 unsigned long flags
;
1896 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1897 if (rq
->curr
->sched_class
->post_schedule
)
1898 rq
->curr
->sched_class
->post_schedule(rq
);
1899 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1901 rq
->post_schedule
= 0;
1907 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1911 static inline void post_schedule(struct rq
*rq
)
1918 * schedule_tail - first thing a freshly forked thread must call.
1919 * @prev: the thread we just switched away from.
1921 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1922 __releases(rq
->lock
)
1924 struct rq
*rq
= this_rq();
1926 finish_task_switch(rq
, prev
);
1929 * FIXME: do we need to worry about rq being invalidated by the
1934 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1935 /* In this case, finish_task_switch does not reenable preemption */
1938 if (current
->set_child_tid
)
1939 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1943 * context_switch - switch to the new MM and the new
1944 * thread's register state.
1947 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1948 struct task_struct
*next
)
1950 struct mm_struct
*mm
, *oldmm
;
1952 prepare_task_switch(rq
, prev
, next
);
1955 oldmm
= prev
->active_mm
;
1957 * For paravirt, this is coupled with an exit in switch_to to
1958 * combine the page table reload and the switch backend into
1961 arch_start_context_switch(prev
);
1964 next
->active_mm
= oldmm
;
1965 atomic_inc(&oldmm
->mm_count
);
1966 enter_lazy_tlb(oldmm
, next
);
1968 switch_mm(oldmm
, mm
, next
);
1971 prev
->active_mm
= NULL
;
1972 rq
->prev_mm
= oldmm
;
1975 * Since the runqueue lock will be released by the next
1976 * task (which is an invalid locking op but in the case
1977 * of the scheduler it's an obvious special-case), so we
1978 * do an early lockdep release here:
1980 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1981 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1984 context_tracking_task_switch(prev
, next
);
1985 /* Here we just switch the register state and the stack. */
1986 switch_to(prev
, next
, prev
);
1990 * this_rq must be evaluated again because prev may have moved
1991 * CPUs since it called schedule(), thus the 'rq' on its stack
1992 * frame will be invalid.
1994 finish_task_switch(this_rq(), prev
);
1998 * nr_running and nr_context_switches:
2000 * externally visible scheduler statistics: current number of runnable
2001 * threads, total number of context switches performed since bootup.
2003 unsigned long nr_running(void)
2005 unsigned long i
, sum
= 0;
2007 for_each_online_cpu(i
)
2008 sum
+= cpu_rq(i
)->nr_running
;
2013 unsigned long long nr_context_switches(void)
2016 unsigned long long sum
= 0;
2018 for_each_possible_cpu(i
)
2019 sum
+= cpu_rq(i
)->nr_switches
;
2024 unsigned long nr_iowait(void)
2026 unsigned long i
, sum
= 0;
2028 for_each_possible_cpu(i
)
2029 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2034 unsigned long nr_iowait_cpu(int cpu
)
2036 struct rq
*this = cpu_rq(cpu
);
2037 return atomic_read(&this->nr_iowait
);
2040 unsigned long this_cpu_load(void)
2042 struct rq
*this = this_rq();
2043 return this->cpu_load
[0];
2048 * Global load-average calculations
2050 * We take a distributed and async approach to calculating the global load-avg
2051 * in order to minimize overhead.
2053 * The global load average is an exponentially decaying average of nr_running +
2054 * nr_uninterruptible.
2056 * Once every LOAD_FREQ:
2059 * for_each_possible_cpu(cpu)
2060 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2062 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2064 * Due to a number of reasons the above turns in the mess below:
2066 * - for_each_possible_cpu() is prohibitively expensive on machines with
2067 * serious number of cpus, therefore we need to take a distributed approach
2068 * to calculating nr_active.
2070 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2071 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2073 * So assuming nr_active := 0 when we start out -- true per definition, we
2074 * can simply take per-cpu deltas and fold those into a global accumulate
2075 * to obtain the same result. See calc_load_fold_active().
2077 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2078 * across the machine, we assume 10 ticks is sufficient time for every
2079 * cpu to have completed this task.
2081 * This places an upper-bound on the IRQ-off latency of the machine. Then
2082 * again, being late doesn't loose the delta, just wrecks the sample.
2084 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2085 * this would add another cross-cpu cacheline miss and atomic operation
2086 * to the wakeup path. Instead we increment on whatever cpu the task ran
2087 * when it went into uninterruptible state and decrement on whatever cpu
2088 * did the wakeup. This means that only the sum of nr_uninterruptible over
2089 * all cpus yields the correct result.
2091 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2094 /* Variables and functions for calc_load */
2095 static atomic_long_t calc_load_tasks
;
2096 static unsigned long calc_load_update
;
2097 unsigned long avenrun
[3];
2098 EXPORT_SYMBOL(avenrun
); /* should be removed */
2101 * get_avenrun - get the load average array
2102 * @loads: pointer to dest load array
2103 * @offset: offset to add
2104 * @shift: shift count to shift the result left
2106 * These values are estimates at best, so no need for locking.
2108 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2110 loads
[0] = (avenrun
[0] + offset
) << shift
;
2111 loads
[1] = (avenrun
[1] + offset
) << shift
;
2112 loads
[2] = (avenrun
[2] + offset
) << shift
;
2115 static long calc_load_fold_active(struct rq
*this_rq
)
2117 long nr_active
, delta
= 0;
2119 nr_active
= this_rq
->nr_running
;
2120 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2122 if (nr_active
!= this_rq
->calc_load_active
) {
2123 delta
= nr_active
- this_rq
->calc_load_active
;
2124 this_rq
->calc_load_active
= nr_active
;
2131 * a1 = a0 * e + a * (1 - e)
2133 static unsigned long
2134 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2137 load
+= active
* (FIXED_1
- exp
);
2138 load
+= 1UL << (FSHIFT
- 1);
2139 return load
>> FSHIFT
;
2144 * Handle NO_HZ for the global load-average.
2146 * Since the above described distributed algorithm to compute the global
2147 * load-average relies on per-cpu sampling from the tick, it is affected by
2150 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2151 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2152 * when we read the global state.
2154 * Obviously reality has to ruin such a delightfully simple scheme:
2156 * - When we go NO_HZ idle during the window, we can negate our sample
2157 * contribution, causing under-accounting.
2159 * We avoid this by keeping two idle-delta counters and flipping them
2160 * when the window starts, thus separating old and new NO_HZ load.
2162 * The only trick is the slight shift in index flip for read vs write.
2166 * |-|-----------|-|-----------|-|-----------|-|
2167 * r:0 0 1 1 0 0 1 1 0
2168 * w:0 1 1 0 0 1 1 0 0
2170 * This ensures we'll fold the old idle contribution in this window while
2171 * accumlating the new one.
2173 * - When we wake up from NO_HZ idle during the window, we push up our
2174 * contribution, since we effectively move our sample point to a known
2177 * This is solved by pushing the window forward, and thus skipping the
2178 * sample, for this cpu (effectively using the idle-delta for this cpu which
2179 * was in effect at the time the window opened). This also solves the issue
2180 * of having to deal with a cpu having been in NOHZ idle for multiple
2181 * LOAD_FREQ intervals.
2183 * When making the ILB scale, we should try to pull this in as well.
2185 static atomic_long_t calc_load_idle
[2];
2186 static int calc_load_idx
;
2188 static inline int calc_load_write_idx(void)
2190 int idx
= calc_load_idx
;
2193 * See calc_global_nohz(), if we observe the new index, we also
2194 * need to observe the new update time.
2199 * If the folding window started, make sure we start writing in the
2202 if (!time_before(jiffies
, calc_load_update
))
2208 static inline int calc_load_read_idx(void)
2210 return calc_load_idx
& 1;
2213 void calc_load_enter_idle(void)
2215 struct rq
*this_rq
= this_rq();
2219 * We're going into NOHZ mode, if there's any pending delta, fold it
2220 * into the pending idle delta.
2222 delta
= calc_load_fold_active(this_rq
);
2224 int idx
= calc_load_write_idx();
2225 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2229 void calc_load_exit_idle(void)
2231 struct rq
*this_rq
= this_rq();
2234 * If we're still before the sample window, we're done.
2236 if (time_before(jiffies
, this_rq
->calc_load_update
))
2240 * We woke inside or after the sample window, this means we're already
2241 * accounted through the nohz accounting, so skip the entire deal and
2242 * sync up for the next window.
2244 this_rq
->calc_load_update
= calc_load_update
;
2245 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2246 this_rq
->calc_load_update
+= LOAD_FREQ
;
2249 static long calc_load_fold_idle(void)
2251 int idx
= calc_load_read_idx();
2254 if (atomic_long_read(&calc_load_idle
[idx
]))
2255 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2261 * fixed_power_int - compute: x^n, in O(log n) time
2263 * @x: base of the power
2264 * @frac_bits: fractional bits of @x
2265 * @n: power to raise @x to.
2267 * By exploiting the relation between the definition of the natural power
2268 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2269 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2270 * (where: n_i \elem {0, 1}, the binary vector representing n),
2271 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2272 * of course trivially computable in O(log_2 n), the length of our binary
2275 static unsigned long
2276 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2278 unsigned long result
= 1UL << frac_bits
;
2283 result
+= 1UL << (frac_bits
- 1);
2284 result
>>= frac_bits
;
2290 x
+= 1UL << (frac_bits
- 1);
2298 * a1 = a0 * e + a * (1 - e)
2300 * a2 = a1 * e + a * (1 - e)
2301 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2302 * = a0 * e^2 + a * (1 - e) * (1 + e)
2304 * a3 = a2 * e + a * (1 - e)
2305 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2306 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2310 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2311 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2312 * = a0 * e^n + a * (1 - e^n)
2314 * [1] application of the geometric series:
2317 * S_n := \Sum x^i = -------------
2320 static unsigned long
2321 calc_load_n(unsigned long load
, unsigned long exp
,
2322 unsigned long active
, unsigned int n
)
2325 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2329 * NO_HZ can leave us missing all per-cpu ticks calling
2330 * calc_load_account_active(), but since an idle CPU folds its delta into
2331 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2332 * in the pending idle delta if our idle period crossed a load cycle boundary.
2334 * Once we've updated the global active value, we need to apply the exponential
2335 * weights adjusted to the number of cycles missed.
2337 static void calc_global_nohz(void)
2339 long delta
, active
, n
;
2341 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2343 * Catch-up, fold however many we are behind still
2345 delta
= jiffies
- calc_load_update
- 10;
2346 n
= 1 + (delta
/ LOAD_FREQ
);
2348 active
= atomic_long_read(&calc_load_tasks
);
2349 active
= active
> 0 ? active
* FIXED_1
: 0;
2351 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2352 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2353 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2355 calc_load_update
+= n
* LOAD_FREQ
;
2359 * Flip the idle index...
2361 * Make sure we first write the new time then flip the index, so that
2362 * calc_load_write_idx() will see the new time when it reads the new
2363 * index, this avoids a double flip messing things up.
2368 #else /* !CONFIG_NO_HZ */
2370 static inline long calc_load_fold_idle(void) { return 0; }
2371 static inline void calc_global_nohz(void) { }
2373 #endif /* CONFIG_NO_HZ */
2376 * calc_load - update the avenrun load estimates 10 ticks after the
2377 * CPUs have updated calc_load_tasks.
2379 void calc_global_load(unsigned long ticks
)
2383 if (time_before(jiffies
, calc_load_update
+ 10))
2387 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2389 delta
= calc_load_fold_idle();
2391 atomic_long_add(delta
, &calc_load_tasks
);
2393 active
= atomic_long_read(&calc_load_tasks
);
2394 active
= active
> 0 ? active
* FIXED_1
: 0;
2396 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2397 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2398 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2400 calc_load_update
+= LOAD_FREQ
;
2403 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2409 * Called from update_cpu_load() to periodically update this CPU's
2412 static void calc_load_account_active(struct rq
*this_rq
)
2416 if (time_before(jiffies
, this_rq
->calc_load_update
))
2419 delta
= calc_load_fold_active(this_rq
);
2421 atomic_long_add(delta
, &calc_load_tasks
);
2423 this_rq
->calc_load_update
+= LOAD_FREQ
;
2427 * End of global load-average stuff
2431 * The exact cpuload at various idx values, calculated at every tick would be
2432 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2434 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2435 * on nth tick when cpu may be busy, then we have:
2436 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2437 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2439 * decay_load_missed() below does efficient calculation of
2440 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2441 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2443 * The calculation is approximated on a 128 point scale.
2444 * degrade_zero_ticks is the number of ticks after which load at any
2445 * particular idx is approximated to be zero.
2446 * degrade_factor is a precomputed table, a row for each load idx.
2447 * Each column corresponds to degradation factor for a power of two ticks,
2448 * based on 128 point scale.
2450 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2451 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2453 * With this power of 2 load factors, we can degrade the load n times
2454 * by looking at 1 bits in n and doing as many mult/shift instead of
2455 * n mult/shifts needed by the exact degradation.
2457 #define DEGRADE_SHIFT 7
2458 static const unsigned char
2459 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2460 static const unsigned char
2461 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2462 {0, 0, 0, 0, 0, 0, 0, 0},
2463 {64, 32, 8, 0, 0, 0, 0, 0},
2464 {96, 72, 40, 12, 1, 0, 0},
2465 {112, 98, 75, 43, 15, 1, 0},
2466 {120, 112, 98, 76, 45, 16, 2} };
2469 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2470 * would be when CPU is idle and so we just decay the old load without
2471 * adding any new load.
2473 static unsigned long
2474 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2478 if (!missed_updates
)
2481 if (missed_updates
>= degrade_zero_ticks
[idx
])
2485 return load
>> missed_updates
;
2487 while (missed_updates
) {
2488 if (missed_updates
% 2)
2489 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2491 missed_updates
>>= 1;
2498 * Update rq->cpu_load[] statistics. This function is usually called every
2499 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2500 * every tick. We fix it up based on jiffies.
2502 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2503 unsigned long pending_updates
)
2507 this_rq
->nr_load_updates
++;
2509 /* Update our load: */
2510 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2511 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2512 unsigned long old_load
, new_load
;
2514 /* scale is effectively 1 << i now, and >> i divides by scale */
2516 old_load
= this_rq
->cpu_load
[i
];
2517 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2518 new_load
= this_load
;
2520 * Round up the averaging division if load is increasing. This
2521 * prevents us from getting stuck on 9 if the load is 10, for
2524 if (new_load
> old_load
)
2525 new_load
+= scale
- 1;
2527 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2530 sched_avg_update(this_rq
);
2535 * There is no sane way to deal with nohz on smp when using jiffies because the
2536 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2537 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2539 * Therefore we cannot use the delta approach from the regular tick since that
2540 * would seriously skew the load calculation. However we'll make do for those
2541 * updates happening while idle (nohz_idle_balance) or coming out of idle
2542 * (tick_nohz_idle_exit).
2544 * This means we might still be one tick off for nohz periods.
2548 * Called from nohz_idle_balance() to update the load ratings before doing the
2551 void update_idle_cpu_load(struct rq
*this_rq
)
2553 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2554 unsigned long load
= this_rq
->load
.weight
;
2555 unsigned long pending_updates
;
2558 * bail if there's load or we're actually up-to-date.
2560 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2563 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2564 this_rq
->last_load_update_tick
= curr_jiffies
;
2566 __update_cpu_load(this_rq
, load
, pending_updates
);
2570 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2572 void update_cpu_load_nohz(void)
2574 struct rq
*this_rq
= this_rq();
2575 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2576 unsigned long pending_updates
;
2578 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2581 raw_spin_lock(&this_rq
->lock
);
2582 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2583 if (pending_updates
) {
2584 this_rq
->last_load_update_tick
= curr_jiffies
;
2586 * We were idle, this means load 0, the current load might be
2587 * !0 due to remote wakeups and the sort.
2589 __update_cpu_load(this_rq
, 0, pending_updates
);
2591 raw_spin_unlock(&this_rq
->lock
);
2593 #endif /* CONFIG_NO_HZ */
2596 * Called from scheduler_tick()
2598 static void update_cpu_load_active(struct rq
*this_rq
)
2601 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2603 this_rq
->last_load_update_tick
= jiffies
;
2604 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2606 calc_load_account_active(this_rq
);
2612 * sched_exec - execve() is a valuable balancing opportunity, because at
2613 * this point the task has the smallest effective memory and cache footprint.
2615 void sched_exec(void)
2617 struct task_struct
*p
= current
;
2618 unsigned long flags
;
2621 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2622 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2623 if (dest_cpu
== smp_processor_id())
2626 if (likely(cpu_active(dest_cpu
))) {
2627 struct migration_arg arg
= { p
, dest_cpu
};
2629 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2630 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2634 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2639 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2640 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2642 EXPORT_PER_CPU_SYMBOL(kstat
);
2643 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2646 * Return any ns on the sched_clock that have not yet been accounted in
2647 * @p in case that task is currently running.
2649 * Called with task_rq_lock() held on @rq.
2651 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2655 if (task_current(rq
, p
)) {
2656 update_rq_clock(rq
);
2657 ns
= rq
->clock_task
- p
->se
.exec_start
;
2665 unsigned long long task_delta_exec(struct task_struct
*p
)
2667 unsigned long flags
;
2671 rq
= task_rq_lock(p
, &flags
);
2672 ns
= do_task_delta_exec(p
, rq
);
2673 task_rq_unlock(rq
, p
, &flags
);
2679 * Return accounted runtime for the task.
2680 * In case the task is currently running, return the runtime plus current's
2681 * pending runtime that have not been accounted yet.
2683 unsigned long long task_sched_runtime(struct task_struct
*p
)
2685 unsigned long flags
;
2689 rq
= task_rq_lock(p
, &flags
);
2690 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2691 task_rq_unlock(rq
, p
, &flags
);
2697 * This function gets called by the timer code, with HZ frequency.
2698 * We call it with interrupts disabled.
2700 void scheduler_tick(void)
2702 int cpu
= smp_processor_id();
2703 struct rq
*rq
= cpu_rq(cpu
);
2704 struct task_struct
*curr
= rq
->curr
;
2708 raw_spin_lock(&rq
->lock
);
2709 update_rq_clock(rq
);
2710 update_cpu_load_active(rq
);
2711 curr
->sched_class
->task_tick(rq
, curr
, 0);
2712 raw_spin_unlock(&rq
->lock
);
2714 perf_event_task_tick();
2717 rq
->idle_balance
= idle_cpu(cpu
);
2718 trigger_load_balance(rq
, cpu
);
2722 notrace
unsigned long get_parent_ip(unsigned long addr
)
2724 if (in_lock_functions(addr
)) {
2725 addr
= CALLER_ADDR2
;
2726 if (in_lock_functions(addr
))
2727 addr
= CALLER_ADDR3
;
2732 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2733 defined(CONFIG_PREEMPT_TRACER))
2735 void __kprobes
add_preempt_count(int val
)
2737 #ifdef CONFIG_DEBUG_PREEMPT
2741 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2744 preempt_count() += val
;
2745 #ifdef CONFIG_DEBUG_PREEMPT
2747 * Spinlock count overflowing soon?
2749 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2752 if (preempt_count() == val
)
2753 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2755 EXPORT_SYMBOL(add_preempt_count
);
2757 void __kprobes
sub_preempt_count(int val
)
2759 #ifdef CONFIG_DEBUG_PREEMPT
2763 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2766 * Is the spinlock portion underflowing?
2768 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2769 !(preempt_count() & PREEMPT_MASK
)))
2773 if (preempt_count() == val
)
2774 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2775 preempt_count() -= val
;
2777 EXPORT_SYMBOL(sub_preempt_count
);
2782 * Print scheduling while atomic bug:
2784 static noinline
void __schedule_bug(struct task_struct
*prev
)
2786 if (oops_in_progress
)
2789 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2790 prev
->comm
, prev
->pid
, preempt_count());
2792 debug_show_held_locks(prev
);
2794 if (irqs_disabled())
2795 print_irqtrace_events(prev
);
2797 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2801 * Various schedule()-time debugging checks and statistics:
2803 static inline void schedule_debug(struct task_struct
*prev
)
2806 * Test if we are atomic. Since do_exit() needs to call into
2807 * schedule() atomically, we ignore that path for now.
2808 * Otherwise, whine if we are scheduling when we should not be.
2810 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2811 __schedule_bug(prev
);
2814 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2816 schedstat_inc(this_rq(), sched_count
);
2819 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2821 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2822 update_rq_clock(rq
);
2823 prev
->sched_class
->put_prev_task(rq
, prev
);
2827 * Pick up the highest-prio task:
2829 static inline struct task_struct
*
2830 pick_next_task(struct rq
*rq
)
2832 const struct sched_class
*class;
2833 struct task_struct
*p
;
2836 * Optimization: we know that if all tasks are in
2837 * the fair class we can call that function directly:
2839 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2840 p
= fair_sched_class
.pick_next_task(rq
);
2845 for_each_class(class) {
2846 p
= class->pick_next_task(rq
);
2851 BUG(); /* the idle class will always have a runnable task */
2855 * __schedule() is the main scheduler function.
2857 * The main means of driving the scheduler and thus entering this function are:
2859 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2861 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2862 * paths. For example, see arch/x86/entry_64.S.
2864 * To drive preemption between tasks, the scheduler sets the flag in timer
2865 * interrupt handler scheduler_tick().
2867 * 3. Wakeups don't really cause entry into schedule(). They add a
2868 * task to the run-queue and that's it.
2870 * Now, if the new task added to the run-queue preempts the current
2871 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2872 * called on the nearest possible occasion:
2874 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2876 * - in syscall or exception context, at the next outmost
2877 * preempt_enable(). (this might be as soon as the wake_up()'s
2880 * - in IRQ context, return from interrupt-handler to
2881 * preemptible context
2883 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2886 * - cond_resched() call
2887 * - explicit schedule() call
2888 * - return from syscall or exception to user-space
2889 * - return from interrupt-handler to user-space
2891 static void __sched
__schedule(void)
2893 struct task_struct
*prev
, *next
;
2894 unsigned long *switch_count
;
2900 cpu
= smp_processor_id();
2902 rcu_note_context_switch(cpu
);
2905 schedule_debug(prev
);
2907 if (sched_feat(HRTICK
))
2910 raw_spin_lock_irq(&rq
->lock
);
2912 switch_count
= &prev
->nivcsw
;
2913 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2914 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2915 prev
->state
= TASK_RUNNING
;
2917 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2921 * If a worker went to sleep, notify and ask workqueue
2922 * whether it wants to wake up a task to maintain
2925 if (prev
->flags
& PF_WQ_WORKER
) {
2926 struct task_struct
*to_wakeup
;
2928 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2930 try_to_wake_up_local(to_wakeup
);
2933 switch_count
= &prev
->nvcsw
;
2936 pre_schedule(rq
, prev
);
2938 if (unlikely(!rq
->nr_running
))
2939 idle_balance(cpu
, rq
);
2941 put_prev_task(rq
, prev
);
2942 next
= pick_next_task(rq
);
2943 clear_tsk_need_resched(prev
);
2944 rq
->skip_clock_update
= 0;
2946 if (likely(prev
!= next
)) {
2951 context_switch(rq
, prev
, next
); /* unlocks the rq */
2953 * The context switch have flipped the stack from under us
2954 * and restored the local variables which were saved when
2955 * this task called schedule() in the past. prev == current
2956 * is still correct, but it can be moved to another cpu/rq.
2958 cpu
= smp_processor_id();
2961 raw_spin_unlock_irq(&rq
->lock
);
2965 sched_preempt_enable_no_resched();
2970 static inline void sched_submit_work(struct task_struct
*tsk
)
2972 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2975 * If we are going to sleep and we have plugged IO queued,
2976 * make sure to submit it to avoid deadlocks.
2978 if (blk_needs_flush_plug(tsk
))
2979 blk_schedule_flush_plug(tsk
);
2982 asmlinkage
void __sched
schedule(void)
2984 struct task_struct
*tsk
= current
;
2986 sched_submit_work(tsk
);
2989 EXPORT_SYMBOL(schedule
);
2991 #ifdef CONFIG_CONTEXT_TRACKING
2992 asmlinkage
void __sched
schedule_user(void)
2995 * If we come here after a random call to set_need_resched(),
2996 * or we have been woken up remotely but the IPI has not yet arrived,
2997 * we haven't yet exited the RCU idle mode. Do it here manually until
2998 * we find a better solution.
3007 * schedule_preempt_disabled - called with preemption disabled
3009 * Returns with preemption disabled. Note: preempt_count must be 1
3011 void __sched
schedule_preempt_disabled(void)
3013 sched_preempt_enable_no_resched();
3018 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3020 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3022 if (lock
->owner
!= owner
)
3026 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3027 * lock->owner still matches owner, if that fails, owner might
3028 * point to free()d memory, if it still matches, the rcu_read_lock()
3029 * ensures the memory stays valid.
3033 return owner
->on_cpu
;
3037 * Look out! "owner" is an entirely speculative pointer
3038 * access and not reliable.
3040 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3042 if (!sched_feat(OWNER_SPIN
))
3046 while (owner_running(lock
, owner
)) {
3050 arch_mutex_cpu_relax();
3055 * We break out the loop above on need_resched() and when the
3056 * owner changed, which is a sign for heavy contention. Return
3057 * success only when lock->owner is NULL.
3059 return lock
->owner
== NULL
;
3063 #ifdef CONFIG_PREEMPT
3065 * this is the entry point to schedule() from in-kernel preemption
3066 * off of preempt_enable. Kernel preemptions off return from interrupt
3067 * occur there and call schedule directly.
3069 asmlinkage
void __sched notrace
preempt_schedule(void)
3071 struct thread_info
*ti
= current_thread_info();
3074 * If there is a non-zero preempt_count or interrupts are disabled,
3075 * we do not want to preempt the current task. Just return..
3077 if (likely(ti
->preempt_count
|| irqs_disabled()))
3081 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3083 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3086 * Check again in case we missed a preemption opportunity
3087 * between schedule and now.
3090 } while (need_resched());
3092 EXPORT_SYMBOL(preempt_schedule
);
3095 * this is the entry point to schedule() from kernel preemption
3096 * off of irq context.
3097 * Note, that this is called and return with irqs disabled. This will
3098 * protect us against recursive calling from irq.
3100 asmlinkage
void __sched
preempt_schedule_irq(void)
3102 struct thread_info
*ti
= current_thread_info();
3103 enum ctx_state prev_state
;
3105 /* Catch callers which need to be fixed */
3106 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3108 prev_state
= exception_enter();
3111 add_preempt_count(PREEMPT_ACTIVE
);
3114 local_irq_disable();
3115 sub_preempt_count(PREEMPT_ACTIVE
);
3118 * Check again in case we missed a preemption opportunity
3119 * between schedule and now.
3122 } while (need_resched());
3124 exception_exit(prev_state
);
3127 #endif /* CONFIG_PREEMPT */
3129 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3132 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3134 EXPORT_SYMBOL(default_wake_function
);
3137 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3138 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3139 * number) then we wake all the non-exclusive tasks and one exclusive task.
3141 * There are circumstances in which we can try to wake a task which has already
3142 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3143 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3145 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3146 int nr_exclusive
, int wake_flags
, void *key
)
3148 wait_queue_t
*curr
, *next
;
3150 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3151 unsigned flags
= curr
->flags
;
3153 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3154 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3160 * __wake_up - wake up threads blocked on a waitqueue.
3162 * @mode: which threads
3163 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3164 * @key: is directly passed to the wakeup function
3166 * It may be assumed that this function implies a write memory barrier before
3167 * changing the task state if and only if any tasks are woken up.
3169 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3170 int nr_exclusive
, void *key
)
3172 unsigned long flags
;
3174 spin_lock_irqsave(&q
->lock
, flags
);
3175 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3176 spin_unlock_irqrestore(&q
->lock
, flags
);
3178 EXPORT_SYMBOL(__wake_up
);
3181 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3183 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3185 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3187 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3189 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3191 __wake_up_common(q
, mode
, 1, 0, key
);
3193 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3196 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3198 * @mode: which threads
3199 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3200 * @key: opaque value to be passed to wakeup targets
3202 * The sync wakeup differs that the waker knows that it will schedule
3203 * away soon, so while the target thread will be woken up, it will not
3204 * be migrated to another CPU - ie. the two threads are 'synchronized'
3205 * with each other. This can prevent needless bouncing between CPUs.
3207 * On UP it can prevent extra preemption.
3209 * It may be assumed that this function implies a write memory barrier before
3210 * changing the task state if and only if any tasks are woken up.
3212 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3213 int nr_exclusive
, void *key
)
3215 unsigned long flags
;
3216 int wake_flags
= WF_SYNC
;
3221 if (unlikely(!nr_exclusive
))
3224 spin_lock_irqsave(&q
->lock
, flags
);
3225 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3226 spin_unlock_irqrestore(&q
->lock
, flags
);
3228 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3231 * __wake_up_sync - see __wake_up_sync_key()
3233 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3235 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3237 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3240 * complete: - signals a single thread waiting on this completion
3241 * @x: holds the state of this particular completion
3243 * This will wake up a single thread waiting on this completion. Threads will be
3244 * awakened in the same order in which they were queued.
3246 * See also complete_all(), wait_for_completion() and related routines.
3248 * It may be assumed that this function implies a write memory barrier before
3249 * changing the task state if and only if any tasks are woken up.
3251 void complete(struct completion
*x
)
3253 unsigned long flags
;
3255 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3257 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3258 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3260 EXPORT_SYMBOL(complete
);
3263 * complete_all: - signals all threads waiting on this completion
3264 * @x: holds the state of this particular completion
3266 * This will wake up all threads waiting on this particular completion event.
3268 * It may be assumed that this function implies a write memory barrier before
3269 * changing the task state if and only if any tasks are woken up.
3271 void complete_all(struct completion
*x
)
3273 unsigned long flags
;
3275 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3276 x
->done
+= UINT_MAX
/2;
3277 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3278 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3280 EXPORT_SYMBOL(complete_all
);
3282 static inline long __sched
3283 do_wait_for_common(struct completion
*x
,
3284 long (*action
)(long), long timeout
, int state
)
3287 DECLARE_WAITQUEUE(wait
, current
);
3289 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3291 if (signal_pending_state(state
, current
)) {
3292 timeout
= -ERESTARTSYS
;
3295 __set_current_state(state
);
3296 spin_unlock_irq(&x
->wait
.lock
);
3297 timeout
= action(timeout
);
3298 spin_lock_irq(&x
->wait
.lock
);
3299 } while (!x
->done
&& timeout
);
3300 __remove_wait_queue(&x
->wait
, &wait
);
3305 return timeout
?: 1;
3308 static inline long __sched
3309 __wait_for_common(struct completion
*x
,
3310 long (*action
)(long), long timeout
, int state
)
3314 spin_lock_irq(&x
->wait
.lock
);
3315 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
3316 spin_unlock_irq(&x
->wait
.lock
);
3321 wait_for_common(struct completion
*x
, long timeout
, int state
)
3323 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
3327 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
3329 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
3333 * wait_for_completion: - waits for completion of a task
3334 * @x: holds the state of this particular completion
3336 * This waits to be signaled for completion of a specific task. It is NOT
3337 * interruptible and there is no timeout.
3339 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3340 * and interrupt capability. Also see complete().
3342 void __sched
wait_for_completion(struct completion
*x
)
3344 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3346 EXPORT_SYMBOL(wait_for_completion
);
3349 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3350 * @x: holds the state of this particular completion
3351 * @timeout: timeout value in jiffies
3353 * This waits for either a completion of a specific task to be signaled or for a
3354 * specified timeout to expire. The timeout is in jiffies. It is not
3357 * The return value is 0 if timed out, and positive (at least 1, or number of
3358 * jiffies left till timeout) if completed.
3360 unsigned long __sched
3361 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3363 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3365 EXPORT_SYMBOL(wait_for_completion_timeout
);
3368 * wait_for_completion_io: - waits for completion of a task
3369 * @x: holds the state of this particular completion
3371 * This waits to be signaled for completion of a specific task. It is NOT
3372 * interruptible and there is no timeout. The caller is accounted as waiting
3375 void __sched
wait_for_completion_io(struct completion
*x
)
3377 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3379 EXPORT_SYMBOL(wait_for_completion_io
);
3382 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3383 * @x: holds the state of this particular completion
3384 * @timeout: timeout value in jiffies
3386 * This waits for either a completion of a specific task to be signaled or for a
3387 * specified timeout to expire. The timeout is in jiffies. It is not
3388 * interruptible. The caller is accounted as waiting for IO.
3390 * The return value is 0 if timed out, and positive (at least 1, or number of
3391 * jiffies left till timeout) if completed.
3393 unsigned long __sched
3394 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
3396 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3398 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
3401 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3402 * @x: holds the state of this particular completion
3404 * This waits for completion of a specific task to be signaled. It is
3407 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3409 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3411 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3412 if (t
== -ERESTARTSYS
)
3416 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3419 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3420 * @x: holds the state of this particular completion
3421 * @timeout: timeout value in jiffies
3423 * This waits for either a completion of a specific task to be signaled or for a
3424 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3426 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3427 * positive (at least 1, or number of jiffies left till timeout) if completed.
3430 wait_for_completion_interruptible_timeout(struct completion
*x
,
3431 unsigned long timeout
)
3433 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3435 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3438 * wait_for_completion_killable: - waits for completion of a task (killable)
3439 * @x: holds the state of this particular completion
3441 * This waits to be signaled for completion of a specific task. It can be
3442 * interrupted by a kill signal.
3444 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3446 int __sched
wait_for_completion_killable(struct completion
*x
)
3448 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3449 if (t
== -ERESTARTSYS
)
3453 EXPORT_SYMBOL(wait_for_completion_killable
);
3456 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3457 * @x: holds the state of this particular completion
3458 * @timeout: timeout value in jiffies
3460 * This waits for either a completion of a specific task to be
3461 * signaled or for a specified timeout to expire. It can be
3462 * interrupted by a kill signal. The timeout is in jiffies.
3464 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3465 * positive (at least 1, or number of jiffies left till timeout) if completed.
3468 wait_for_completion_killable_timeout(struct completion
*x
,
3469 unsigned long timeout
)
3471 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3473 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3476 * try_wait_for_completion - try to decrement a completion without blocking
3477 * @x: completion structure
3479 * Returns: 0 if a decrement cannot be done without blocking
3480 * 1 if a decrement succeeded.
3482 * If a completion is being used as a counting completion,
3483 * attempt to decrement the counter without blocking. This
3484 * enables us to avoid waiting if the resource the completion
3485 * is protecting is not available.
3487 bool try_wait_for_completion(struct completion
*x
)
3489 unsigned long flags
;
3492 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3497 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3500 EXPORT_SYMBOL(try_wait_for_completion
);
3503 * completion_done - Test to see if a completion has any waiters
3504 * @x: completion structure
3506 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3507 * 1 if there are no waiters.
3510 bool completion_done(struct completion
*x
)
3512 unsigned long flags
;
3515 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3518 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3521 EXPORT_SYMBOL(completion_done
);
3524 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3526 unsigned long flags
;
3529 init_waitqueue_entry(&wait
, current
);
3531 __set_current_state(state
);
3533 spin_lock_irqsave(&q
->lock
, flags
);
3534 __add_wait_queue(q
, &wait
);
3535 spin_unlock(&q
->lock
);
3536 timeout
= schedule_timeout(timeout
);
3537 spin_lock_irq(&q
->lock
);
3538 __remove_wait_queue(q
, &wait
);
3539 spin_unlock_irqrestore(&q
->lock
, flags
);
3544 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3546 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3548 EXPORT_SYMBOL(interruptible_sleep_on
);
3551 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3553 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3555 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3557 void __sched
sleep_on(wait_queue_head_t
*q
)
3559 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3561 EXPORT_SYMBOL(sleep_on
);
3563 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3565 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3567 EXPORT_SYMBOL(sleep_on_timeout
);
3569 #ifdef CONFIG_RT_MUTEXES
3572 * rt_mutex_setprio - set the current priority of a task
3574 * @prio: prio value (kernel-internal form)
3576 * This function changes the 'effective' priority of a task. It does
3577 * not touch ->normal_prio like __setscheduler().
3579 * Used by the rt_mutex code to implement priority inheritance logic.
3581 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3583 int oldprio
, on_rq
, running
;
3585 const struct sched_class
*prev_class
;
3587 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3589 rq
= __task_rq_lock(p
);
3592 * Idle task boosting is a nono in general. There is one
3593 * exception, when PREEMPT_RT and NOHZ is active:
3595 * The idle task calls get_next_timer_interrupt() and holds
3596 * the timer wheel base->lock on the CPU and another CPU wants
3597 * to access the timer (probably to cancel it). We can safely
3598 * ignore the boosting request, as the idle CPU runs this code
3599 * with interrupts disabled and will complete the lock
3600 * protected section without being interrupted. So there is no
3601 * real need to boost.
3603 if (unlikely(p
== rq
->idle
)) {
3604 WARN_ON(p
!= rq
->curr
);
3605 WARN_ON(p
->pi_blocked_on
);
3609 trace_sched_pi_setprio(p
, prio
);
3611 prev_class
= p
->sched_class
;
3613 running
= task_current(rq
, p
);
3615 dequeue_task(rq
, p
, 0);
3617 p
->sched_class
->put_prev_task(rq
, p
);
3620 p
->sched_class
= &rt_sched_class
;
3622 p
->sched_class
= &fair_sched_class
;
3627 p
->sched_class
->set_curr_task(rq
);
3629 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3631 check_class_changed(rq
, p
, prev_class
, oldprio
);
3633 __task_rq_unlock(rq
);
3636 void set_user_nice(struct task_struct
*p
, long nice
)
3638 int old_prio
, delta
, on_rq
;
3639 unsigned long flags
;
3642 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3645 * We have to be careful, if called from sys_setpriority(),
3646 * the task might be in the middle of scheduling on another CPU.
3648 rq
= task_rq_lock(p
, &flags
);
3650 * The RT priorities are set via sched_setscheduler(), but we still
3651 * allow the 'normal' nice value to be set - but as expected
3652 * it wont have any effect on scheduling until the task is
3653 * SCHED_FIFO/SCHED_RR:
3655 if (task_has_rt_policy(p
)) {
3656 p
->static_prio
= NICE_TO_PRIO(nice
);
3661 dequeue_task(rq
, p
, 0);
3663 p
->static_prio
= NICE_TO_PRIO(nice
);
3666 p
->prio
= effective_prio(p
);
3667 delta
= p
->prio
- old_prio
;
3670 enqueue_task(rq
, p
, 0);
3672 * If the task increased its priority or is running and
3673 * lowered its priority, then reschedule its CPU:
3675 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3676 resched_task(rq
->curr
);
3679 task_rq_unlock(rq
, p
, &flags
);
3681 EXPORT_SYMBOL(set_user_nice
);
3684 * can_nice - check if a task can reduce its nice value
3688 int can_nice(const struct task_struct
*p
, const int nice
)
3690 /* convert nice value [19,-20] to rlimit style value [1,40] */
3691 int nice_rlim
= 20 - nice
;
3693 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3694 capable(CAP_SYS_NICE
));
3697 #ifdef __ARCH_WANT_SYS_NICE
3700 * sys_nice - change the priority of the current process.
3701 * @increment: priority increment
3703 * sys_setpriority is a more generic, but much slower function that
3704 * does similar things.
3706 SYSCALL_DEFINE1(nice
, int, increment
)
3711 * Setpriority might change our priority at the same moment.
3712 * We don't have to worry. Conceptually one call occurs first
3713 * and we have a single winner.
3715 if (increment
< -40)
3720 nice
= TASK_NICE(current
) + increment
;
3726 if (increment
< 0 && !can_nice(current
, nice
))
3729 retval
= security_task_setnice(current
, nice
);
3733 set_user_nice(current
, nice
);
3740 * task_prio - return the priority value of a given task.
3741 * @p: the task in question.
3743 * This is the priority value as seen by users in /proc.
3744 * RT tasks are offset by -200. Normal tasks are centered
3745 * around 0, value goes from -16 to +15.
3747 int task_prio(const struct task_struct
*p
)
3749 return p
->prio
- MAX_RT_PRIO
;
3753 * task_nice - return the nice value of a given task.
3754 * @p: the task in question.
3756 int task_nice(const struct task_struct
*p
)
3758 return TASK_NICE(p
);
3760 EXPORT_SYMBOL(task_nice
);
3763 * idle_cpu - is a given cpu idle currently?
3764 * @cpu: the processor in question.
3766 int idle_cpu(int cpu
)
3768 struct rq
*rq
= cpu_rq(cpu
);
3770 if (rq
->curr
!= rq
->idle
)
3777 if (!llist_empty(&rq
->wake_list
))
3785 * idle_task - return the idle task for a given cpu.
3786 * @cpu: the processor in question.
3788 struct task_struct
*idle_task(int cpu
)
3790 return cpu_rq(cpu
)->idle
;
3794 * find_process_by_pid - find a process with a matching PID value.
3795 * @pid: the pid in question.
3797 static struct task_struct
*find_process_by_pid(pid_t pid
)
3799 return pid
? find_task_by_vpid(pid
) : current
;
3802 /* Actually do priority change: must hold rq lock. */
3804 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3807 p
->rt_priority
= prio
;
3808 p
->normal_prio
= normal_prio(p
);
3809 /* we are holding p->pi_lock already */
3810 p
->prio
= rt_mutex_getprio(p
);
3811 if (rt_prio(p
->prio
))
3812 p
->sched_class
= &rt_sched_class
;
3814 p
->sched_class
= &fair_sched_class
;
3819 * check the target process has a UID that matches the current process's
3821 static bool check_same_owner(struct task_struct
*p
)
3823 const struct cred
*cred
= current_cred(), *pcred
;
3827 pcred
= __task_cred(p
);
3828 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3829 uid_eq(cred
->euid
, pcred
->uid
));
3834 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3835 const struct sched_param
*param
, bool user
)
3837 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3838 unsigned long flags
;
3839 const struct sched_class
*prev_class
;
3843 /* may grab non-irq protected spin_locks */
3844 BUG_ON(in_interrupt());
3846 /* double check policy once rq lock held */
3848 reset_on_fork
= p
->sched_reset_on_fork
;
3849 policy
= oldpolicy
= p
->policy
;
3851 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3852 policy
&= ~SCHED_RESET_ON_FORK
;
3854 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3855 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3856 policy
!= SCHED_IDLE
)
3861 * Valid priorities for SCHED_FIFO and SCHED_RR are
3862 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3863 * SCHED_BATCH and SCHED_IDLE is 0.
3865 if (param
->sched_priority
< 0 ||
3866 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3867 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3869 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3873 * Allow unprivileged RT tasks to decrease priority:
3875 if (user
&& !capable(CAP_SYS_NICE
)) {
3876 if (rt_policy(policy
)) {
3877 unsigned long rlim_rtprio
=
3878 task_rlimit(p
, RLIMIT_RTPRIO
);
3880 /* can't set/change the rt policy */
3881 if (policy
!= p
->policy
&& !rlim_rtprio
)
3884 /* can't increase priority */
3885 if (param
->sched_priority
> p
->rt_priority
&&
3886 param
->sched_priority
> rlim_rtprio
)
3891 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3892 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3894 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3895 if (!can_nice(p
, TASK_NICE(p
)))
3899 /* can't change other user's priorities */
3900 if (!check_same_owner(p
))
3903 /* Normal users shall not reset the sched_reset_on_fork flag */
3904 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3909 retval
= security_task_setscheduler(p
);
3915 * make sure no PI-waiters arrive (or leave) while we are
3916 * changing the priority of the task:
3918 * To be able to change p->policy safely, the appropriate
3919 * runqueue lock must be held.
3921 rq
= task_rq_lock(p
, &flags
);
3924 * Changing the policy of the stop threads its a very bad idea
3926 if (p
== rq
->stop
) {
3927 task_rq_unlock(rq
, p
, &flags
);
3932 * If not changing anything there's no need to proceed further:
3934 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3935 param
->sched_priority
== p
->rt_priority
))) {
3936 task_rq_unlock(rq
, p
, &flags
);
3940 #ifdef CONFIG_RT_GROUP_SCHED
3943 * Do not allow realtime tasks into groups that have no runtime
3946 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3947 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3948 !task_group_is_autogroup(task_group(p
))) {
3949 task_rq_unlock(rq
, p
, &flags
);
3955 /* recheck policy now with rq lock held */
3956 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3957 policy
= oldpolicy
= -1;
3958 task_rq_unlock(rq
, p
, &flags
);
3962 running
= task_current(rq
, p
);
3964 dequeue_task(rq
, p
, 0);
3966 p
->sched_class
->put_prev_task(rq
, p
);
3968 p
->sched_reset_on_fork
= reset_on_fork
;
3971 prev_class
= p
->sched_class
;
3972 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3975 p
->sched_class
->set_curr_task(rq
);
3977 enqueue_task(rq
, p
, 0);
3979 check_class_changed(rq
, p
, prev_class
, oldprio
);
3980 task_rq_unlock(rq
, p
, &flags
);
3982 rt_mutex_adjust_pi(p
);
3988 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3989 * @p: the task in question.
3990 * @policy: new policy.
3991 * @param: structure containing the new RT priority.
3993 * NOTE that the task may be already dead.
3995 int sched_setscheduler(struct task_struct
*p
, int policy
,
3996 const struct sched_param
*param
)
3998 return __sched_setscheduler(p
, policy
, param
, true);
4000 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4003 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4004 * @p: the task in question.
4005 * @policy: new policy.
4006 * @param: structure containing the new RT priority.
4008 * Just like sched_setscheduler, only don't bother checking if the
4009 * current context has permission. For example, this is needed in
4010 * stop_machine(): we create temporary high priority worker threads,
4011 * but our caller might not have that capability.
4013 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4014 const struct sched_param
*param
)
4016 return __sched_setscheduler(p
, policy
, param
, false);
4020 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4022 struct sched_param lparam
;
4023 struct task_struct
*p
;
4026 if (!param
|| pid
< 0)
4028 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4033 p
= find_process_by_pid(pid
);
4035 retval
= sched_setscheduler(p
, policy
, &lparam
);
4042 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4043 * @pid: the pid in question.
4044 * @policy: new policy.
4045 * @param: structure containing the new RT priority.
4047 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4048 struct sched_param __user
*, param
)
4050 /* negative values for policy are not valid */
4054 return do_sched_setscheduler(pid
, policy
, param
);
4058 * sys_sched_setparam - set/change the RT priority of a thread
4059 * @pid: the pid in question.
4060 * @param: structure containing the new RT priority.
4062 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4064 return do_sched_setscheduler(pid
, -1, param
);
4068 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4069 * @pid: the pid in question.
4071 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4073 struct task_struct
*p
;
4081 p
= find_process_by_pid(pid
);
4083 retval
= security_task_getscheduler(p
);
4086 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4093 * sys_sched_getparam - get the RT priority of a thread
4094 * @pid: the pid in question.
4095 * @param: structure containing the RT priority.
4097 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4099 struct sched_param lp
;
4100 struct task_struct
*p
;
4103 if (!param
|| pid
< 0)
4107 p
= find_process_by_pid(pid
);
4112 retval
= security_task_getscheduler(p
);
4116 lp
.sched_priority
= p
->rt_priority
;
4120 * This one might sleep, we cannot do it with a spinlock held ...
4122 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4131 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4133 cpumask_var_t cpus_allowed
, new_mask
;
4134 struct task_struct
*p
;
4140 p
= find_process_by_pid(pid
);
4147 /* Prevent p going away */
4151 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4155 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4157 goto out_free_cpus_allowed
;
4160 if (!check_same_owner(p
)) {
4162 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4169 retval
= security_task_setscheduler(p
);
4173 cpuset_cpus_allowed(p
, cpus_allowed
);
4174 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4176 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4179 cpuset_cpus_allowed(p
, cpus_allowed
);
4180 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4182 * We must have raced with a concurrent cpuset
4183 * update. Just reset the cpus_allowed to the
4184 * cpuset's cpus_allowed
4186 cpumask_copy(new_mask
, cpus_allowed
);
4191 free_cpumask_var(new_mask
);
4192 out_free_cpus_allowed
:
4193 free_cpumask_var(cpus_allowed
);
4200 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4201 struct cpumask
*new_mask
)
4203 if (len
< cpumask_size())
4204 cpumask_clear(new_mask
);
4205 else if (len
> cpumask_size())
4206 len
= cpumask_size();
4208 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4212 * sys_sched_setaffinity - set the cpu affinity of a process
4213 * @pid: pid of the process
4214 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4215 * @user_mask_ptr: user-space pointer to the new cpu mask
4217 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4218 unsigned long __user
*, user_mask_ptr
)
4220 cpumask_var_t new_mask
;
4223 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4226 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4228 retval
= sched_setaffinity(pid
, new_mask
);
4229 free_cpumask_var(new_mask
);
4233 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4235 struct task_struct
*p
;
4236 unsigned long flags
;
4243 p
= find_process_by_pid(pid
);
4247 retval
= security_task_getscheduler(p
);
4251 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4252 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4253 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4263 * sys_sched_getaffinity - get the cpu affinity of a process
4264 * @pid: pid of the process
4265 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4266 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4268 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4269 unsigned long __user
*, user_mask_ptr
)
4274 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4276 if (len
& (sizeof(unsigned long)-1))
4279 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4282 ret
= sched_getaffinity(pid
, mask
);
4284 size_t retlen
= min_t(size_t, len
, cpumask_size());
4286 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4291 free_cpumask_var(mask
);
4297 * sys_sched_yield - yield the current processor to other threads.
4299 * This function yields the current CPU to other tasks. If there are no
4300 * other threads running on this CPU then this function will return.
4302 SYSCALL_DEFINE0(sched_yield
)
4304 struct rq
*rq
= this_rq_lock();
4306 schedstat_inc(rq
, yld_count
);
4307 current
->sched_class
->yield_task(rq
);
4310 * Since we are going to call schedule() anyway, there's
4311 * no need to preempt or enable interrupts:
4313 __release(rq
->lock
);
4314 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4315 do_raw_spin_unlock(&rq
->lock
);
4316 sched_preempt_enable_no_resched();
4323 static inline int should_resched(void)
4325 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4328 static void __cond_resched(void)
4330 add_preempt_count(PREEMPT_ACTIVE
);
4332 sub_preempt_count(PREEMPT_ACTIVE
);
4335 int __sched
_cond_resched(void)
4337 if (should_resched()) {
4343 EXPORT_SYMBOL(_cond_resched
);
4346 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4347 * call schedule, and on return reacquire the lock.
4349 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4350 * operations here to prevent schedule() from being called twice (once via
4351 * spin_unlock(), once by hand).
4353 int __cond_resched_lock(spinlock_t
*lock
)
4355 int resched
= should_resched();
4358 lockdep_assert_held(lock
);
4360 if (spin_needbreak(lock
) || resched
) {
4371 EXPORT_SYMBOL(__cond_resched_lock
);
4373 int __sched
__cond_resched_softirq(void)
4375 BUG_ON(!in_softirq());
4377 if (should_resched()) {
4385 EXPORT_SYMBOL(__cond_resched_softirq
);
4388 * yield - yield the current processor to other threads.
4390 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4392 * The scheduler is at all times free to pick the calling task as the most
4393 * eligible task to run, if removing the yield() call from your code breaks
4394 * it, its already broken.
4396 * Typical broken usage is:
4401 * where one assumes that yield() will let 'the other' process run that will
4402 * make event true. If the current task is a SCHED_FIFO task that will never
4403 * happen. Never use yield() as a progress guarantee!!
4405 * If you want to use yield() to wait for something, use wait_event().
4406 * If you want to use yield() to be 'nice' for others, use cond_resched().
4407 * If you still want to use yield(), do not!
4409 void __sched
yield(void)
4411 set_current_state(TASK_RUNNING
);
4414 EXPORT_SYMBOL(yield
);
4417 * yield_to - yield the current processor to another thread in
4418 * your thread group, or accelerate that thread toward the
4419 * processor it's on.
4421 * @preempt: whether task preemption is allowed or not
4423 * It's the caller's job to ensure that the target task struct
4424 * can't go away on us before we can do any checks.
4427 * true (>0) if we indeed boosted the target task.
4428 * false (0) if we failed to boost the target.
4429 * -ESRCH if there's no task to yield to.
4431 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4433 struct task_struct
*curr
= current
;
4434 struct rq
*rq
, *p_rq
;
4435 unsigned long flags
;
4438 local_irq_save(flags
);
4444 * If we're the only runnable task on the rq and target rq also
4445 * has only one task, there's absolutely no point in yielding.
4447 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4452 double_rq_lock(rq
, p_rq
);
4453 while (task_rq(p
) != p_rq
) {
4454 double_rq_unlock(rq
, p_rq
);
4458 if (!curr
->sched_class
->yield_to_task
)
4461 if (curr
->sched_class
!= p
->sched_class
)
4464 if (task_running(p_rq
, p
) || p
->state
)
4467 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4469 schedstat_inc(rq
, yld_count
);
4471 * Make p's CPU reschedule; pick_next_entity takes care of
4474 if (preempt
&& rq
!= p_rq
)
4475 resched_task(p_rq
->curr
);
4479 double_rq_unlock(rq
, p_rq
);
4481 local_irq_restore(flags
);
4488 EXPORT_SYMBOL_GPL(yield_to
);
4491 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4492 * that process accounting knows that this is a task in IO wait state.
4494 void __sched
io_schedule(void)
4496 struct rq
*rq
= raw_rq();
4498 delayacct_blkio_start();
4499 atomic_inc(&rq
->nr_iowait
);
4500 blk_flush_plug(current
);
4501 current
->in_iowait
= 1;
4503 current
->in_iowait
= 0;
4504 atomic_dec(&rq
->nr_iowait
);
4505 delayacct_blkio_end();
4507 EXPORT_SYMBOL(io_schedule
);
4509 long __sched
io_schedule_timeout(long timeout
)
4511 struct rq
*rq
= raw_rq();
4514 delayacct_blkio_start();
4515 atomic_inc(&rq
->nr_iowait
);
4516 blk_flush_plug(current
);
4517 current
->in_iowait
= 1;
4518 ret
= schedule_timeout(timeout
);
4519 current
->in_iowait
= 0;
4520 atomic_dec(&rq
->nr_iowait
);
4521 delayacct_blkio_end();
4526 * sys_sched_get_priority_max - return maximum RT priority.
4527 * @policy: scheduling class.
4529 * this syscall returns the maximum rt_priority that can be used
4530 * by a given scheduling class.
4532 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4539 ret
= MAX_USER_RT_PRIO
-1;
4551 * sys_sched_get_priority_min - return minimum RT priority.
4552 * @policy: scheduling class.
4554 * this syscall returns the minimum rt_priority that can be used
4555 * by a given scheduling class.
4557 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4575 * sys_sched_rr_get_interval - return the default timeslice of a process.
4576 * @pid: pid of the process.
4577 * @interval: userspace pointer to the timeslice value.
4579 * this syscall writes the default timeslice value of a given process
4580 * into the user-space timespec buffer. A value of '0' means infinity.
4582 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4583 struct timespec __user
*, interval
)
4585 struct task_struct
*p
;
4586 unsigned int time_slice
;
4587 unsigned long flags
;
4597 p
= find_process_by_pid(pid
);
4601 retval
= security_task_getscheduler(p
);
4605 rq
= task_rq_lock(p
, &flags
);
4606 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4607 task_rq_unlock(rq
, p
, &flags
);
4610 jiffies_to_timespec(time_slice
, &t
);
4611 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4619 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4621 void sched_show_task(struct task_struct
*p
)
4623 unsigned long free
= 0;
4627 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4628 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4629 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4630 #if BITS_PER_LONG == 32
4631 if (state
== TASK_RUNNING
)
4632 printk(KERN_CONT
" running ");
4634 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4636 if (state
== TASK_RUNNING
)
4637 printk(KERN_CONT
" running task ");
4639 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4641 #ifdef CONFIG_DEBUG_STACK_USAGE
4642 free
= stack_not_used(p
);
4645 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4647 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4648 task_pid_nr(p
), ppid
,
4649 (unsigned long)task_thread_info(p
)->flags
);
4651 show_stack(p
, NULL
);
4654 void show_state_filter(unsigned long state_filter
)
4656 struct task_struct
*g
, *p
;
4658 #if BITS_PER_LONG == 32
4660 " task PC stack pid father\n");
4663 " task PC stack pid father\n");
4666 do_each_thread(g
, p
) {
4668 * reset the NMI-timeout, listing all files on a slow
4669 * console might take a lot of time:
4671 touch_nmi_watchdog();
4672 if (!state_filter
|| (p
->state
& state_filter
))
4674 } while_each_thread(g
, p
);
4676 touch_all_softlockup_watchdogs();
4678 #ifdef CONFIG_SCHED_DEBUG
4679 sysrq_sched_debug_show();
4683 * Only show locks if all tasks are dumped:
4686 debug_show_all_locks();
4689 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4691 idle
->sched_class
= &idle_sched_class
;
4695 * init_idle - set up an idle thread for a given CPU
4696 * @idle: task in question
4697 * @cpu: cpu the idle task belongs to
4699 * NOTE: this function does not set the idle thread's NEED_RESCHED
4700 * flag, to make booting more robust.
4702 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4704 struct rq
*rq
= cpu_rq(cpu
);
4705 unsigned long flags
;
4707 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4710 idle
->state
= TASK_RUNNING
;
4711 idle
->se
.exec_start
= sched_clock();
4713 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4715 * We're having a chicken and egg problem, even though we are
4716 * holding rq->lock, the cpu isn't yet set to this cpu so the
4717 * lockdep check in task_group() will fail.
4719 * Similar case to sched_fork(). / Alternatively we could
4720 * use task_rq_lock() here and obtain the other rq->lock.
4725 __set_task_cpu(idle
, cpu
);
4728 rq
->curr
= rq
->idle
= idle
;
4729 #if defined(CONFIG_SMP)
4732 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4734 /* Set the preempt count _outside_ the spinlocks! */
4735 task_thread_info(idle
)->preempt_count
= 0;
4738 * The idle tasks have their own, simple scheduling class:
4740 idle
->sched_class
= &idle_sched_class
;
4741 ftrace_graph_init_idle_task(idle
, cpu
);
4742 vtime_init_idle(idle
);
4743 #if defined(CONFIG_SMP)
4744 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4749 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4751 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4752 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4754 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4755 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4759 * This is how migration works:
4761 * 1) we invoke migration_cpu_stop() on the target CPU using
4763 * 2) stopper starts to run (implicitly forcing the migrated thread
4765 * 3) it checks whether the migrated task is still in the wrong runqueue.
4766 * 4) if it's in the wrong runqueue then the migration thread removes
4767 * it and puts it into the right queue.
4768 * 5) stopper completes and stop_one_cpu() returns and the migration
4773 * Change a given task's CPU affinity. Migrate the thread to a
4774 * proper CPU and schedule it away if the CPU it's executing on
4775 * is removed from the allowed bitmask.
4777 * NOTE: the caller must have a valid reference to the task, the
4778 * task must not exit() & deallocate itself prematurely. The
4779 * call is not atomic; no spinlocks may be held.
4781 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4783 unsigned long flags
;
4785 unsigned int dest_cpu
;
4788 rq
= task_rq_lock(p
, &flags
);
4790 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4793 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4798 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4803 do_set_cpus_allowed(p
, new_mask
);
4805 /* Can the task run on the task's current CPU? If so, we're done */
4806 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4809 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4811 struct migration_arg arg
= { p
, dest_cpu
};
4812 /* Need help from migration thread: drop lock and wait. */
4813 task_rq_unlock(rq
, p
, &flags
);
4814 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4815 tlb_migrate_finish(p
->mm
);
4819 task_rq_unlock(rq
, p
, &flags
);
4823 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4826 * Move (not current) task off this cpu, onto dest cpu. We're doing
4827 * this because either it can't run here any more (set_cpus_allowed()
4828 * away from this CPU, or CPU going down), or because we're
4829 * attempting to rebalance this task on exec (sched_exec).
4831 * So we race with normal scheduler movements, but that's OK, as long
4832 * as the task is no longer on this CPU.
4834 * Returns non-zero if task was successfully migrated.
4836 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4838 struct rq
*rq_dest
, *rq_src
;
4841 if (unlikely(!cpu_active(dest_cpu
)))
4844 rq_src
= cpu_rq(src_cpu
);
4845 rq_dest
= cpu_rq(dest_cpu
);
4847 raw_spin_lock(&p
->pi_lock
);
4848 double_rq_lock(rq_src
, rq_dest
);
4849 /* Already moved. */
4850 if (task_cpu(p
) != src_cpu
)
4852 /* Affinity changed (again). */
4853 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4857 * If we're not on a rq, the next wake-up will ensure we're
4861 dequeue_task(rq_src
, p
, 0);
4862 set_task_cpu(p
, dest_cpu
);
4863 enqueue_task(rq_dest
, p
, 0);
4864 check_preempt_curr(rq_dest
, p
, 0);
4869 double_rq_unlock(rq_src
, rq_dest
);
4870 raw_spin_unlock(&p
->pi_lock
);
4875 * migration_cpu_stop - this will be executed by a highprio stopper thread
4876 * and performs thread migration by bumping thread off CPU then
4877 * 'pushing' onto another runqueue.
4879 static int migration_cpu_stop(void *data
)
4881 struct migration_arg
*arg
= data
;
4884 * The original target cpu might have gone down and we might
4885 * be on another cpu but it doesn't matter.
4887 local_irq_disable();
4888 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4893 #ifdef CONFIG_HOTPLUG_CPU
4896 * Ensures that the idle task is using init_mm right before its cpu goes
4899 void idle_task_exit(void)
4901 struct mm_struct
*mm
= current
->active_mm
;
4903 BUG_ON(cpu_online(smp_processor_id()));
4906 switch_mm(mm
, &init_mm
, current
);
4911 * Since this CPU is going 'away' for a while, fold any nr_active delta
4912 * we might have. Assumes we're called after migrate_tasks() so that the
4913 * nr_active count is stable.
4915 * Also see the comment "Global load-average calculations".
4917 static void calc_load_migrate(struct rq
*rq
)
4919 long delta
= calc_load_fold_active(rq
);
4921 atomic_long_add(delta
, &calc_load_tasks
);
4925 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4926 * try_to_wake_up()->select_task_rq().
4928 * Called with rq->lock held even though we'er in stop_machine() and
4929 * there's no concurrency possible, we hold the required locks anyway
4930 * because of lock validation efforts.
4932 static void migrate_tasks(unsigned int dead_cpu
)
4934 struct rq
*rq
= cpu_rq(dead_cpu
);
4935 struct task_struct
*next
, *stop
= rq
->stop
;
4939 * Fudge the rq selection such that the below task selection loop
4940 * doesn't get stuck on the currently eligible stop task.
4942 * We're currently inside stop_machine() and the rq is either stuck
4943 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4944 * either way we should never end up calling schedule() until we're
4951 * There's this thread running, bail when that's the only
4954 if (rq
->nr_running
== 1)
4957 next
= pick_next_task(rq
);
4959 next
->sched_class
->put_prev_task(rq
, next
);
4961 /* Find suitable destination for @next, with force if needed. */
4962 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4963 raw_spin_unlock(&rq
->lock
);
4965 __migrate_task(next
, dead_cpu
, dest_cpu
);
4967 raw_spin_lock(&rq
->lock
);
4973 #endif /* CONFIG_HOTPLUG_CPU */
4975 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4977 static struct ctl_table sd_ctl_dir
[] = {
4979 .procname
= "sched_domain",
4985 static struct ctl_table sd_ctl_root
[] = {
4987 .procname
= "kernel",
4989 .child
= sd_ctl_dir
,
4994 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4996 struct ctl_table
*entry
=
4997 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5002 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5004 struct ctl_table
*entry
;
5007 * In the intermediate directories, both the child directory and
5008 * procname are dynamically allocated and could fail but the mode
5009 * will always be set. In the lowest directory the names are
5010 * static strings and all have proc handlers.
5012 for (entry
= *tablep
; entry
->mode
; entry
++) {
5014 sd_free_ctl_entry(&entry
->child
);
5015 if (entry
->proc_handler
== NULL
)
5016 kfree(entry
->procname
);
5023 static int min_load_idx
= 0;
5024 static int max_load_idx
= CPU_LOAD_IDX_MAX
;
5027 set_table_entry(struct ctl_table
*entry
,
5028 const char *procname
, void *data
, int maxlen
,
5029 umode_t mode
, proc_handler
*proc_handler
,
5032 entry
->procname
= procname
;
5034 entry
->maxlen
= maxlen
;
5036 entry
->proc_handler
= proc_handler
;
5039 entry
->extra1
= &min_load_idx
;
5040 entry
->extra2
= &max_load_idx
;
5044 static struct ctl_table
*
5045 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5047 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5052 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5053 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5054 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5055 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5056 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5057 sizeof(int), 0644, proc_dointvec_minmax
, true);
5058 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5059 sizeof(int), 0644, proc_dointvec_minmax
, true);
5060 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5061 sizeof(int), 0644, proc_dointvec_minmax
, true);
5062 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5063 sizeof(int), 0644, proc_dointvec_minmax
, true);
5064 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5065 sizeof(int), 0644, proc_dointvec_minmax
, true);
5066 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5067 sizeof(int), 0644, proc_dointvec_minmax
, false);
5068 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5069 sizeof(int), 0644, proc_dointvec_minmax
, false);
5070 set_table_entry(&table
[9], "cache_nice_tries",
5071 &sd
->cache_nice_tries
,
5072 sizeof(int), 0644, proc_dointvec_minmax
, false);
5073 set_table_entry(&table
[10], "flags", &sd
->flags
,
5074 sizeof(int), 0644, proc_dointvec_minmax
, false);
5075 set_table_entry(&table
[11], "name", sd
->name
,
5076 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5077 /* &table[12] is terminator */
5082 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5084 struct ctl_table
*entry
, *table
;
5085 struct sched_domain
*sd
;
5086 int domain_num
= 0, i
;
5089 for_each_domain(cpu
, sd
)
5091 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5096 for_each_domain(cpu
, sd
) {
5097 snprintf(buf
, 32, "domain%d", i
);
5098 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5100 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5107 static struct ctl_table_header
*sd_sysctl_header
;
5108 static void register_sched_domain_sysctl(void)
5110 int i
, cpu_num
= num_possible_cpus();
5111 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5114 WARN_ON(sd_ctl_dir
[0].child
);
5115 sd_ctl_dir
[0].child
= entry
;
5120 for_each_possible_cpu(i
) {
5121 snprintf(buf
, 32, "cpu%d", i
);
5122 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5124 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5128 WARN_ON(sd_sysctl_header
);
5129 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5132 /* may be called multiple times per register */
5133 static void unregister_sched_domain_sysctl(void)
5135 if (sd_sysctl_header
)
5136 unregister_sysctl_table(sd_sysctl_header
);
5137 sd_sysctl_header
= NULL
;
5138 if (sd_ctl_dir
[0].child
)
5139 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5142 static void register_sched_domain_sysctl(void)
5145 static void unregister_sched_domain_sysctl(void)
5150 static void set_rq_online(struct rq
*rq
)
5153 const struct sched_class
*class;
5155 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5158 for_each_class(class) {
5159 if (class->rq_online
)
5160 class->rq_online(rq
);
5165 static void set_rq_offline(struct rq
*rq
)
5168 const struct sched_class
*class;
5170 for_each_class(class) {
5171 if (class->rq_offline
)
5172 class->rq_offline(rq
);
5175 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5181 * migration_call - callback that gets triggered when a CPU is added.
5182 * Here we can start up the necessary migration thread for the new CPU.
5184 static int __cpuinit
5185 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5187 int cpu
= (long)hcpu
;
5188 unsigned long flags
;
5189 struct rq
*rq
= cpu_rq(cpu
);
5191 switch (action
& ~CPU_TASKS_FROZEN
) {
5193 case CPU_UP_PREPARE
:
5194 rq
->calc_load_update
= calc_load_update
;
5198 /* Update our root-domain */
5199 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5201 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5205 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5208 #ifdef CONFIG_HOTPLUG_CPU
5210 sched_ttwu_pending();
5211 /* Update our root-domain */
5212 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5214 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5218 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5219 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5223 calc_load_migrate(rq
);
5228 update_max_interval();
5234 * Register at high priority so that task migration (migrate_all_tasks)
5235 * happens before everything else. This has to be lower priority than
5236 * the notifier in the perf_event subsystem, though.
5238 static struct notifier_block __cpuinitdata migration_notifier
= {
5239 .notifier_call
= migration_call
,
5240 .priority
= CPU_PRI_MIGRATION
,
5243 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5244 unsigned long action
, void *hcpu
)
5246 switch (action
& ~CPU_TASKS_FROZEN
) {
5248 case CPU_DOWN_FAILED
:
5249 set_cpu_active((long)hcpu
, true);
5256 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5257 unsigned long action
, void *hcpu
)
5259 switch (action
& ~CPU_TASKS_FROZEN
) {
5260 case CPU_DOWN_PREPARE
:
5261 set_cpu_active((long)hcpu
, false);
5268 static int __init
migration_init(void)
5270 void *cpu
= (void *)(long)smp_processor_id();
5273 /* Initialize migration for the boot CPU */
5274 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5275 BUG_ON(err
== NOTIFY_BAD
);
5276 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5277 register_cpu_notifier(&migration_notifier
);
5279 /* Register cpu active notifiers */
5280 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5281 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5285 early_initcall(migration_init
);
5290 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5292 #ifdef CONFIG_SCHED_DEBUG
5294 static __read_mostly
int sched_debug_enabled
;
5296 static int __init
sched_debug_setup(char *str
)
5298 sched_debug_enabled
= 1;
5302 early_param("sched_debug", sched_debug_setup
);
5304 static inline bool sched_debug(void)
5306 return sched_debug_enabled
;
5309 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5310 struct cpumask
*groupmask
)
5312 struct sched_group
*group
= sd
->groups
;
5315 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5316 cpumask_clear(groupmask
);
5318 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5320 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5321 printk("does not load-balance\n");
5323 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5328 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5330 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5331 printk(KERN_ERR
"ERROR: domain->span does not contain "
5334 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5335 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5339 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5343 printk(KERN_ERR
"ERROR: group is NULL\n");
5348 * Even though we initialize ->power to something semi-sane,
5349 * we leave power_orig unset. This allows us to detect if
5350 * domain iteration is still funny without causing /0 traps.
5352 if (!group
->sgp
->power_orig
) {
5353 printk(KERN_CONT
"\n");
5354 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5359 if (!cpumask_weight(sched_group_cpus(group
))) {
5360 printk(KERN_CONT
"\n");
5361 printk(KERN_ERR
"ERROR: empty group\n");
5365 if (!(sd
->flags
& SD_OVERLAP
) &&
5366 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5367 printk(KERN_CONT
"\n");
5368 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5372 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5374 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5376 printk(KERN_CONT
" %s", str
);
5377 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5378 printk(KERN_CONT
" (cpu_power = %d)",
5382 group
= group
->next
;
5383 } while (group
!= sd
->groups
);
5384 printk(KERN_CONT
"\n");
5386 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5387 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5390 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5391 printk(KERN_ERR
"ERROR: parent span is not a superset "
5392 "of domain->span\n");
5396 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5400 if (!sched_debug_enabled
)
5404 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5408 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5411 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5419 #else /* !CONFIG_SCHED_DEBUG */
5420 # define sched_domain_debug(sd, cpu) do { } while (0)
5421 static inline bool sched_debug(void)
5425 #endif /* CONFIG_SCHED_DEBUG */
5427 static int sd_degenerate(struct sched_domain
*sd
)
5429 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5432 /* Following flags need at least 2 groups */
5433 if (sd
->flags
& (SD_LOAD_BALANCE
|
5434 SD_BALANCE_NEWIDLE
|
5438 SD_SHARE_PKG_RESOURCES
)) {
5439 if (sd
->groups
!= sd
->groups
->next
)
5443 /* Following flags don't use groups */
5444 if (sd
->flags
& (SD_WAKE_AFFINE
))
5451 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5453 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5455 if (sd_degenerate(parent
))
5458 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5461 /* Flags needing groups don't count if only 1 group in parent */
5462 if (parent
->groups
== parent
->groups
->next
) {
5463 pflags
&= ~(SD_LOAD_BALANCE
|
5464 SD_BALANCE_NEWIDLE
|
5468 SD_SHARE_PKG_RESOURCES
);
5469 if (nr_node_ids
== 1)
5470 pflags
&= ~SD_SERIALIZE
;
5472 if (~cflags
& pflags
)
5478 static void free_rootdomain(struct rcu_head
*rcu
)
5480 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5482 cpupri_cleanup(&rd
->cpupri
);
5483 free_cpumask_var(rd
->rto_mask
);
5484 free_cpumask_var(rd
->online
);
5485 free_cpumask_var(rd
->span
);
5489 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5491 struct root_domain
*old_rd
= NULL
;
5492 unsigned long flags
;
5494 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5499 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5502 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5505 * If we dont want to free the old_rt yet then
5506 * set old_rd to NULL to skip the freeing later
5509 if (!atomic_dec_and_test(&old_rd
->refcount
))
5513 atomic_inc(&rd
->refcount
);
5516 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5517 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5520 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5523 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5526 static int init_rootdomain(struct root_domain
*rd
)
5528 memset(rd
, 0, sizeof(*rd
));
5530 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5532 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5534 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5537 if (cpupri_init(&rd
->cpupri
) != 0)
5542 free_cpumask_var(rd
->rto_mask
);
5544 free_cpumask_var(rd
->online
);
5546 free_cpumask_var(rd
->span
);
5552 * By default the system creates a single root-domain with all cpus as
5553 * members (mimicking the global state we have today).
5555 struct root_domain def_root_domain
;
5557 static void init_defrootdomain(void)
5559 init_rootdomain(&def_root_domain
);
5561 atomic_set(&def_root_domain
.refcount
, 1);
5564 static struct root_domain
*alloc_rootdomain(void)
5566 struct root_domain
*rd
;
5568 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5572 if (init_rootdomain(rd
) != 0) {
5580 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5582 struct sched_group
*tmp
, *first
;
5591 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5596 } while (sg
!= first
);
5599 static void free_sched_domain(struct rcu_head
*rcu
)
5601 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5604 * If its an overlapping domain it has private groups, iterate and
5607 if (sd
->flags
& SD_OVERLAP
) {
5608 free_sched_groups(sd
->groups
, 1);
5609 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5610 kfree(sd
->groups
->sgp
);
5616 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5618 call_rcu(&sd
->rcu
, free_sched_domain
);
5621 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5623 for (; sd
; sd
= sd
->parent
)
5624 destroy_sched_domain(sd
, cpu
);
5628 * Keep a special pointer to the highest sched_domain that has
5629 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5630 * allows us to avoid some pointer chasing select_idle_sibling().
5632 * Also keep a unique ID per domain (we use the first cpu number in
5633 * the cpumask of the domain), this allows us to quickly tell if
5634 * two cpus are in the same cache domain, see cpus_share_cache().
5636 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5637 DEFINE_PER_CPU(int, sd_llc_id
);
5639 static void update_top_cache_domain(int cpu
)
5641 struct sched_domain
*sd
;
5644 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5646 id
= cpumask_first(sched_domain_span(sd
));
5648 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5649 per_cpu(sd_llc_id
, cpu
) = id
;
5653 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5654 * hold the hotplug lock.
5657 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5659 struct rq
*rq
= cpu_rq(cpu
);
5660 struct sched_domain
*tmp
;
5662 /* Remove the sched domains which do not contribute to scheduling. */
5663 for (tmp
= sd
; tmp
; ) {
5664 struct sched_domain
*parent
= tmp
->parent
;
5668 if (sd_parent_degenerate(tmp
, parent
)) {
5669 tmp
->parent
= parent
->parent
;
5671 parent
->parent
->child
= tmp
;
5672 destroy_sched_domain(parent
, cpu
);
5677 if (sd
&& sd_degenerate(sd
)) {
5680 destroy_sched_domain(tmp
, cpu
);
5685 sched_domain_debug(sd
, cpu
);
5687 rq_attach_root(rq
, rd
);
5689 rcu_assign_pointer(rq
->sd
, sd
);
5690 destroy_sched_domains(tmp
, cpu
);
5692 update_top_cache_domain(cpu
);
5695 /* cpus with isolated domains */
5696 static cpumask_var_t cpu_isolated_map
;
5698 /* Setup the mask of cpus configured for isolated domains */
5699 static int __init
isolated_cpu_setup(char *str
)
5701 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5702 cpulist_parse(str
, cpu_isolated_map
);
5706 __setup("isolcpus=", isolated_cpu_setup
);
5708 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5710 return cpumask_of_node(cpu_to_node(cpu
));
5714 struct sched_domain
**__percpu sd
;
5715 struct sched_group
**__percpu sg
;
5716 struct sched_group_power
**__percpu sgp
;
5720 struct sched_domain
** __percpu sd
;
5721 struct root_domain
*rd
;
5731 struct sched_domain_topology_level
;
5733 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5734 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5736 #define SDTL_OVERLAP 0x01
5738 struct sched_domain_topology_level
{
5739 sched_domain_init_f init
;
5740 sched_domain_mask_f mask
;
5743 struct sd_data data
;
5747 * Build an iteration mask that can exclude certain CPUs from the upwards
5750 * Asymmetric node setups can result in situations where the domain tree is of
5751 * unequal depth, make sure to skip domains that already cover the entire
5754 * In that case build_sched_domains() will have terminated the iteration early
5755 * and our sibling sd spans will be empty. Domains should always include the
5756 * cpu they're built on, so check that.
5759 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5761 const struct cpumask
*span
= sched_domain_span(sd
);
5762 struct sd_data
*sdd
= sd
->private;
5763 struct sched_domain
*sibling
;
5766 for_each_cpu(i
, span
) {
5767 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5768 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5771 cpumask_set_cpu(i
, sched_group_mask(sg
));
5776 * Return the canonical balance cpu for this group, this is the first cpu
5777 * of this group that's also in the iteration mask.
5779 int group_balance_cpu(struct sched_group
*sg
)
5781 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5785 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5787 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5788 const struct cpumask
*span
= sched_domain_span(sd
);
5789 struct cpumask
*covered
= sched_domains_tmpmask
;
5790 struct sd_data
*sdd
= sd
->private;
5791 struct sched_domain
*child
;
5794 cpumask_clear(covered
);
5796 for_each_cpu(i
, span
) {
5797 struct cpumask
*sg_span
;
5799 if (cpumask_test_cpu(i
, covered
))
5802 child
= *per_cpu_ptr(sdd
->sd
, i
);
5804 /* See the comment near build_group_mask(). */
5805 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5808 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5809 GFP_KERNEL
, cpu_to_node(cpu
));
5814 sg_span
= sched_group_cpus(sg
);
5816 child
= child
->child
;
5817 cpumask_copy(sg_span
, sched_domain_span(child
));
5819 cpumask_set_cpu(i
, sg_span
);
5821 cpumask_or(covered
, covered
, sg_span
);
5823 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5824 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5825 build_group_mask(sd
, sg
);
5828 * Initialize sgp->power such that even if we mess up the
5829 * domains and no possible iteration will get us here, we won't
5832 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5835 * Make sure the first group of this domain contains the
5836 * canonical balance cpu. Otherwise the sched_domain iteration
5837 * breaks. See update_sg_lb_stats().
5839 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5840 group_balance_cpu(sg
) == cpu
)
5850 sd
->groups
= groups
;
5855 free_sched_groups(first
, 0);
5860 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5862 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5863 struct sched_domain
*child
= sd
->child
;
5866 cpu
= cpumask_first(sched_domain_span(child
));
5869 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5870 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5871 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5878 * build_sched_groups will build a circular linked list of the groups
5879 * covered by the given span, and will set each group's ->cpumask correctly,
5880 * and ->cpu_power to 0.
5882 * Assumes the sched_domain tree is fully constructed
5885 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5887 struct sched_group
*first
= NULL
, *last
= NULL
;
5888 struct sd_data
*sdd
= sd
->private;
5889 const struct cpumask
*span
= sched_domain_span(sd
);
5890 struct cpumask
*covered
;
5893 get_group(cpu
, sdd
, &sd
->groups
);
5894 atomic_inc(&sd
->groups
->ref
);
5896 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5899 lockdep_assert_held(&sched_domains_mutex
);
5900 covered
= sched_domains_tmpmask
;
5902 cpumask_clear(covered
);
5904 for_each_cpu(i
, span
) {
5905 struct sched_group
*sg
;
5906 int group
= get_group(i
, sdd
, &sg
);
5909 if (cpumask_test_cpu(i
, covered
))
5912 cpumask_clear(sched_group_cpus(sg
));
5914 cpumask_setall(sched_group_mask(sg
));
5916 for_each_cpu(j
, span
) {
5917 if (get_group(j
, sdd
, NULL
) != group
)
5920 cpumask_set_cpu(j
, covered
);
5921 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5936 * Initialize sched groups cpu_power.
5938 * cpu_power indicates the capacity of sched group, which is used while
5939 * distributing the load between different sched groups in a sched domain.
5940 * Typically cpu_power for all the groups in a sched domain will be same unless
5941 * there are asymmetries in the topology. If there are asymmetries, group
5942 * having more cpu_power will pickup more load compared to the group having
5945 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5947 struct sched_group
*sg
= sd
->groups
;
5949 WARN_ON(!sd
|| !sg
);
5952 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5954 } while (sg
!= sd
->groups
);
5956 if (cpu
!= group_balance_cpu(sg
))
5959 update_group_power(sd
, cpu
);
5960 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5963 int __weak
arch_sd_sibling_asym_packing(void)
5965 return 0*SD_ASYM_PACKING
;
5969 * Initializers for schedule domains
5970 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5973 #ifdef CONFIG_SCHED_DEBUG
5974 # define SD_INIT_NAME(sd, type) sd->name = #type
5976 # define SD_INIT_NAME(sd, type) do { } while (0)
5979 #define SD_INIT_FUNC(type) \
5980 static noinline struct sched_domain * \
5981 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5983 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5984 *sd = SD_##type##_INIT; \
5985 SD_INIT_NAME(sd, type); \
5986 sd->private = &tl->data; \
5991 #ifdef CONFIG_SCHED_SMT
5992 SD_INIT_FUNC(SIBLING
)
5994 #ifdef CONFIG_SCHED_MC
5997 #ifdef CONFIG_SCHED_BOOK
6001 static int default_relax_domain_level
= -1;
6002 int sched_domain_level_max
;
6004 static int __init
setup_relax_domain_level(char *str
)
6006 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6007 pr_warn("Unable to set relax_domain_level\n");
6011 __setup("relax_domain_level=", setup_relax_domain_level
);
6013 static void set_domain_attribute(struct sched_domain
*sd
,
6014 struct sched_domain_attr
*attr
)
6018 if (!attr
|| attr
->relax_domain_level
< 0) {
6019 if (default_relax_domain_level
< 0)
6022 request
= default_relax_domain_level
;
6024 request
= attr
->relax_domain_level
;
6025 if (request
< sd
->level
) {
6026 /* turn off idle balance on this domain */
6027 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6029 /* turn on idle balance on this domain */
6030 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6034 static void __sdt_free(const struct cpumask
*cpu_map
);
6035 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6037 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6038 const struct cpumask
*cpu_map
)
6042 if (!atomic_read(&d
->rd
->refcount
))
6043 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6045 free_percpu(d
->sd
); /* fall through */
6047 __sdt_free(cpu_map
); /* fall through */
6053 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6054 const struct cpumask
*cpu_map
)
6056 memset(d
, 0, sizeof(*d
));
6058 if (__sdt_alloc(cpu_map
))
6059 return sa_sd_storage
;
6060 d
->sd
= alloc_percpu(struct sched_domain
*);
6062 return sa_sd_storage
;
6063 d
->rd
= alloc_rootdomain();
6066 return sa_rootdomain
;
6070 * NULL the sd_data elements we've used to build the sched_domain and
6071 * sched_group structure so that the subsequent __free_domain_allocs()
6072 * will not free the data we're using.
6074 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6076 struct sd_data
*sdd
= sd
->private;
6078 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6079 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6081 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6082 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6084 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6085 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6088 #ifdef CONFIG_SCHED_SMT
6089 static const struct cpumask
*cpu_smt_mask(int cpu
)
6091 return topology_thread_cpumask(cpu
);
6096 * Topology list, bottom-up.
6098 static struct sched_domain_topology_level default_topology
[] = {
6099 #ifdef CONFIG_SCHED_SMT
6100 { sd_init_SIBLING
, cpu_smt_mask
, },
6102 #ifdef CONFIG_SCHED_MC
6103 { sd_init_MC
, cpu_coregroup_mask
, },
6105 #ifdef CONFIG_SCHED_BOOK
6106 { sd_init_BOOK
, cpu_book_mask
, },
6108 { sd_init_CPU
, cpu_cpu_mask
, },
6112 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6116 static int sched_domains_numa_levels
;
6117 static int *sched_domains_numa_distance
;
6118 static struct cpumask
***sched_domains_numa_masks
;
6119 static int sched_domains_curr_level
;
6121 static inline int sd_local_flags(int level
)
6123 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6126 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6129 static struct sched_domain
*
6130 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6132 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6133 int level
= tl
->numa_level
;
6134 int sd_weight
= cpumask_weight(
6135 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6137 *sd
= (struct sched_domain
){
6138 .min_interval
= sd_weight
,
6139 .max_interval
= 2*sd_weight
,
6141 .imbalance_pct
= 125,
6142 .cache_nice_tries
= 2,
6149 .flags
= 1*SD_LOAD_BALANCE
6150 | 1*SD_BALANCE_NEWIDLE
6155 | 0*SD_SHARE_CPUPOWER
6156 | 0*SD_SHARE_PKG_RESOURCES
6158 | 0*SD_PREFER_SIBLING
6159 | sd_local_flags(level
)
6161 .last_balance
= jiffies
,
6162 .balance_interval
= sd_weight
,
6164 SD_INIT_NAME(sd
, NUMA
);
6165 sd
->private = &tl
->data
;
6168 * Ugly hack to pass state to sd_numa_mask()...
6170 sched_domains_curr_level
= tl
->numa_level
;
6175 static const struct cpumask
*sd_numa_mask(int cpu
)
6177 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6180 static void sched_numa_warn(const char *str
)
6182 static int done
= false;
6190 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6192 for (i
= 0; i
< nr_node_ids
; i
++) {
6193 printk(KERN_WARNING
" ");
6194 for (j
= 0; j
< nr_node_ids
; j
++)
6195 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6196 printk(KERN_CONT
"\n");
6198 printk(KERN_WARNING
"\n");
6201 static bool find_numa_distance(int distance
)
6205 if (distance
== node_distance(0, 0))
6208 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6209 if (sched_domains_numa_distance
[i
] == distance
)
6216 static void sched_init_numa(void)
6218 int next_distance
, curr_distance
= node_distance(0, 0);
6219 struct sched_domain_topology_level
*tl
;
6223 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6224 if (!sched_domains_numa_distance
)
6228 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6229 * unique distances in the node_distance() table.
6231 * Assumes node_distance(0,j) includes all distances in
6232 * node_distance(i,j) in order to avoid cubic time.
6234 next_distance
= curr_distance
;
6235 for (i
= 0; i
< nr_node_ids
; i
++) {
6236 for (j
= 0; j
< nr_node_ids
; j
++) {
6237 for (k
= 0; k
< nr_node_ids
; k
++) {
6238 int distance
= node_distance(i
, k
);
6240 if (distance
> curr_distance
&&
6241 (distance
< next_distance
||
6242 next_distance
== curr_distance
))
6243 next_distance
= distance
;
6246 * While not a strong assumption it would be nice to know
6247 * about cases where if node A is connected to B, B is not
6248 * equally connected to A.
6250 if (sched_debug() && node_distance(k
, i
) != distance
)
6251 sched_numa_warn("Node-distance not symmetric");
6253 if (sched_debug() && i
&& !find_numa_distance(distance
))
6254 sched_numa_warn("Node-0 not representative");
6256 if (next_distance
!= curr_distance
) {
6257 sched_domains_numa_distance
[level
++] = next_distance
;
6258 sched_domains_numa_levels
= level
;
6259 curr_distance
= next_distance
;
6264 * In case of sched_debug() we verify the above assumption.
6270 * 'level' contains the number of unique distances, excluding the
6271 * identity distance node_distance(i,i).
6273 * The sched_domains_nume_distance[] array includes the actual distance
6278 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6279 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6280 * the array will contain less then 'level' members. This could be
6281 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6282 * in other functions.
6284 * We reset it to 'level' at the end of this function.
6286 sched_domains_numa_levels
= 0;
6288 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6289 if (!sched_domains_numa_masks
)
6293 * Now for each level, construct a mask per node which contains all
6294 * cpus of nodes that are that many hops away from us.
6296 for (i
= 0; i
< level
; i
++) {
6297 sched_domains_numa_masks
[i
] =
6298 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6299 if (!sched_domains_numa_masks
[i
])
6302 for (j
= 0; j
< nr_node_ids
; j
++) {
6303 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6307 sched_domains_numa_masks
[i
][j
] = mask
;
6309 for (k
= 0; k
< nr_node_ids
; k
++) {
6310 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6313 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6318 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6319 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6324 * Copy the default topology bits..
6326 for (i
= 0; default_topology
[i
].init
; i
++)
6327 tl
[i
] = default_topology
[i
];
6330 * .. and append 'j' levels of NUMA goodness.
6332 for (j
= 0; j
< level
; i
++, j
++) {
6333 tl
[i
] = (struct sched_domain_topology_level
){
6334 .init
= sd_numa_init
,
6335 .mask
= sd_numa_mask
,
6336 .flags
= SDTL_OVERLAP
,
6341 sched_domain_topology
= tl
;
6343 sched_domains_numa_levels
= level
;
6346 static void sched_domains_numa_masks_set(int cpu
)
6349 int node
= cpu_to_node(cpu
);
6351 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6352 for (j
= 0; j
< nr_node_ids
; j
++) {
6353 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6354 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6359 static void sched_domains_numa_masks_clear(int cpu
)
6362 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6363 for (j
= 0; j
< nr_node_ids
; j
++)
6364 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6369 * Update sched_domains_numa_masks[level][node] array when new cpus
6372 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6373 unsigned long action
,
6376 int cpu
= (long)hcpu
;
6378 switch (action
& ~CPU_TASKS_FROZEN
) {
6380 sched_domains_numa_masks_set(cpu
);
6384 sched_domains_numa_masks_clear(cpu
);
6394 static inline void sched_init_numa(void)
6398 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6399 unsigned long action
,
6404 #endif /* CONFIG_NUMA */
6406 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6408 struct sched_domain_topology_level
*tl
;
6411 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6412 struct sd_data
*sdd
= &tl
->data
;
6414 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6418 sdd
->sg
= alloc_percpu(struct sched_group
*);
6422 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6426 for_each_cpu(j
, cpu_map
) {
6427 struct sched_domain
*sd
;
6428 struct sched_group
*sg
;
6429 struct sched_group_power
*sgp
;
6431 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6432 GFP_KERNEL
, cpu_to_node(j
));
6436 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6438 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6439 GFP_KERNEL
, cpu_to_node(j
));
6445 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6447 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6448 GFP_KERNEL
, cpu_to_node(j
));
6452 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6459 static void __sdt_free(const struct cpumask
*cpu_map
)
6461 struct sched_domain_topology_level
*tl
;
6464 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6465 struct sd_data
*sdd
= &tl
->data
;
6467 for_each_cpu(j
, cpu_map
) {
6468 struct sched_domain
*sd
;
6471 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6472 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6473 free_sched_groups(sd
->groups
, 0);
6474 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6478 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6480 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6482 free_percpu(sdd
->sd
);
6484 free_percpu(sdd
->sg
);
6486 free_percpu(sdd
->sgp
);
6491 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6492 struct s_data
*d
, const struct cpumask
*cpu_map
,
6493 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6496 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6500 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6502 sd
->level
= child
->level
+ 1;
6503 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6507 set_domain_attribute(sd
, attr
);
6513 * Build sched domains for a given set of cpus and attach the sched domains
6514 * to the individual cpus
6516 static int build_sched_domains(const struct cpumask
*cpu_map
,
6517 struct sched_domain_attr
*attr
)
6519 enum s_alloc alloc_state
= sa_none
;
6520 struct sched_domain
*sd
;
6522 int i
, ret
= -ENOMEM
;
6524 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6525 if (alloc_state
!= sa_rootdomain
)
6528 /* Set up domains for cpus specified by the cpu_map. */
6529 for_each_cpu(i
, cpu_map
) {
6530 struct sched_domain_topology_level
*tl
;
6533 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6534 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6535 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6536 sd
->flags
|= SD_OVERLAP
;
6537 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6544 *per_cpu_ptr(d
.sd
, i
) = sd
;
6547 /* Build the groups for the domains */
6548 for_each_cpu(i
, cpu_map
) {
6549 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6550 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6551 if (sd
->flags
& SD_OVERLAP
) {
6552 if (build_overlap_sched_groups(sd
, i
))
6555 if (build_sched_groups(sd
, i
))
6561 /* Calculate CPU power for physical packages and nodes */
6562 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6563 if (!cpumask_test_cpu(i
, cpu_map
))
6566 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6567 claim_allocations(i
, sd
);
6568 init_sched_groups_power(i
, sd
);
6572 /* Attach the domains */
6574 for_each_cpu(i
, cpu_map
) {
6575 sd
= *per_cpu_ptr(d
.sd
, i
);
6576 cpu_attach_domain(sd
, d
.rd
, i
);
6582 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6586 static cpumask_var_t
*doms_cur
; /* current sched domains */
6587 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6588 static struct sched_domain_attr
*dattr_cur
;
6589 /* attribues of custom domains in 'doms_cur' */
6592 * Special case: If a kmalloc of a doms_cur partition (array of
6593 * cpumask) fails, then fallback to a single sched domain,
6594 * as determined by the single cpumask fallback_doms.
6596 static cpumask_var_t fallback_doms
;
6599 * arch_update_cpu_topology lets virtualized architectures update the
6600 * cpu core maps. It is supposed to return 1 if the topology changed
6601 * or 0 if it stayed the same.
6603 int __attribute__((weak
)) arch_update_cpu_topology(void)
6608 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6611 cpumask_var_t
*doms
;
6613 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6616 for (i
= 0; i
< ndoms
; i
++) {
6617 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6618 free_sched_domains(doms
, i
);
6625 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6628 for (i
= 0; i
< ndoms
; i
++)
6629 free_cpumask_var(doms
[i
]);
6634 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6635 * For now this just excludes isolated cpus, but could be used to
6636 * exclude other special cases in the future.
6638 static int init_sched_domains(const struct cpumask
*cpu_map
)
6642 arch_update_cpu_topology();
6644 doms_cur
= alloc_sched_domains(ndoms_cur
);
6646 doms_cur
= &fallback_doms
;
6647 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6648 err
= build_sched_domains(doms_cur
[0], NULL
);
6649 register_sched_domain_sysctl();
6655 * Detach sched domains from a group of cpus specified in cpu_map
6656 * These cpus will now be attached to the NULL domain
6658 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6663 for_each_cpu(i
, cpu_map
)
6664 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6668 /* handle null as "default" */
6669 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6670 struct sched_domain_attr
*new, int idx_new
)
6672 struct sched_domain_attr tmp
;
6679 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6680 new ? (new + idx_new
) : &tmp
,
6681 sizeof(struct sched_domain_attr
));
6685 * Partition sched domains as specified by the 'ndoms_new'
6686 * cpumasks in the array doms_new[] of cpumasks. This compares
6687 * doms_new[] to the current sched domain partitioning, doms_cur[].
6688 * It destroys each deleted domain and builds each new domain.
6690 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6691 * The masks don't intersect (don't overlap.) We should setup one
6692 * sched domain for each mask. CPUs not in any of the cpumasks will
6693 * not be load balanced. If the same cpumask appears both in the
6694 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6697 * The passed in 'doms_new' should be allocated using
6698 * alloc_sched_domains. This routine takes ownership of it and will
6699 * free_sched_domains it when done with it. If the caller failed the
6700 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6701 * and partition_sched_domains() will fallback to the single partition
6702 * 'fallback_doms', it also forces the domains to be rebuilt.
6704 * If doms_new == NULL it will be replaced with cpu_online_mask.
6705 * ndoms_new == 0 is a special case for destroying existing domains,
6706 * and it will not create the default domain.
6708 * Call with hotplug lock held
6710 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6711 struct sched_domain_attr
*dattr_new
)
6716 mutex_lock(&sched_domains_mutex
);
6718 /* always unregister in case we don't destroy any domains */
6719 unregister_sched_domain_sysctl();
6721 /* Let architecture update cpu core mappings. */
6722 new_topology
= arch_update_cpu_topology();
6724 n
= doms_new
? ndoms_new
: 0;
6726 /* Destroy deleted domains */
6727 for (i
= 0; i
< ndoms_cur
; i
++) {
6728 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6729 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6730 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6733 /* no match - a current sched domain not in new doms_new[] */
6734 detach_destroy_domains(doms_cur
[i
]);
6739 if (doms_new
== NULL
) {
6741 doms_new
= &fallback_doms
;
6742 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6743 WARN_ON_ONCE(dattr_new
);
6746 /* Build new domains */
6747 for (i
= 0; i
< ndoms_new
; i
++) {
6748 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6749 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6750 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6753 /* no match - add a new doms_new */
6754 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6759 /* Remember the new sched domains */
6760 if (doms_cur
!= &fallback_doms
)
6761 free_sched_domains(doms_cur
, ndoms_cur
);
6762 kfree(dattr_cur
); /* kfree(NULL) is safe */
6763 doms_cur
= doms_new
;
6764 dattr_cur
= dattr_new
;
6765 ndoms_cur
= ndoms_new
;
6767 register_sched_domain_sysctl();
6769 mutex_unlock(&sched_domains_mutex
);
6772 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6775 * Update cpusets according to cpu_active mask. If cpusets are
6776 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6777 * around partition_sched_domains().
6779 * If we come here as part of a suspend/resume, don't touch cpusets because we
6780 * want to restore it back to its original state upon resume anyway.
6782 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6786 case CPU_ONLINE_FROZEN
:
6787 case CPU_DOWN_FAILED_FROZEN
:
6790 * num_cpus_frozen tracks how many CPUs are involved in suspend
6791 * resume sequence. As long as this is not the last online
6792 * operation in the resume sequence, just build a single sched
6793 * domain, ignoring cpusets.
6796 if (likely(num_cpus_frozen
)) {
6797 partition_sched_domains(1, NULL
, NULL
);
6802 * This is the last CPU online operation. So fall through and
6803 * restore the original sched domains by considering the
6804 * cpuset configurations.
6808 case CPU_DOWN_FAILED
:
6809 cpuset_update_active_cpus(true);
6817 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6821 case CPU_DOWN_PREPARE
:
6822 cpuset_update_active_cpus(false);
6824 case CPU_DOWN_PREPARE_FROZEN
:
6826 partition_sched_domains(1, NULL
, NULL
);
6834 void __init
sched_init_smp(void)
6836 cpumask_var_t non_isolated_cpus
;
6838 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6839 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6844 mutex_lock(&sched_domains_mutex
);
6845 init_sched_domains(cpu_active_mask
);
6846 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6847 if (cpumask_empty(non_isolated_cpus
))
6848 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6849 mutex_unlock(&sched_domains_mutex
);
6852 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6853 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6854 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6856 /* RT runtime code needs to handle some hotplug events */
6857 hotcpu_notifier(update_runtime
, 0);
6861 /* Move init over to a non-isolated CPU */
6862 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6864 sched_init_granularity();
6865 free_cpumask_var(non_isolated_cpus
);
6867 init_sched_rt_class();
6870 void __init
sched_init_smp(void)
6872 sched_init_granularity();
6874 #endif /* CONFIG_SMP */
6876 const_debug
unsigned int sysctl_timer_migration
= 1;
6878 int in_sched_functions(unsigned long addr
)
6880 return in_lock_functions(addr
) ||
6881 (addr
>= (unsigned long)__sched_text_start
6882 && addr
< (unsigned long)__sched_text_end
);
6885 #ifdef CONFIG_CGROUP_SCHED
6887 * Default task group.
6888 * Every task in system belongs to this group at bootup.
6890 struct task_group root_task_group
;
6891 LIST_HEAD(task_groups
);
6894 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6896 void __init
sched_init(void)
6899 unsigned long alloc_size
= 0, ptr
;
6901 #ifdef CONFIG_FAIR_GROUP_SCHED
6902 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6904 #ifdef CONFIG_RT_GROUP_SCHED
6905 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6907 #ifdef CONFIG_CPUMASK_OFFSTACK
6908 alloc_size
+= num_possible_cpus() * cpumask_size();
6911 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6913 #ifdef CONFIG_FAIR_GROUP_SCHED
6914 root_task_group
.se
= (struct sched_entity
**)ptr
;
6915 ptr
+= nr_cpu_ids
* sizeof(void **);
6917 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6918 ptr
+= nr_cpu_ids
* sizeof(void **);
6920 #endif /* CONFIG_FAIR_GROUP_SCHED */
6921 #ifdef CONFIG_RT_GROUP_SCHED
6922 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6923 ptr
+= nr_cpu_ids
* sizeof(void **);
6925 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6926 ptr
+= nr_cpu_ids
* sizeof(void **);
6928 #endif /* CONFIG_RT_GROUP_SCHED */
6929 #ifdef CONFIG_CPUMASK_OFFSTACK
6930 for_each_possible_cpu(i
) {
6931 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6932 ptr
+= cpumask_size();
6934 #endif /* CONFIG_CPUMASK_OFFSTACK */
6938 init_defrootdomain();
6941 init_rt_bandwidth(&def_rt_bandwidth
,
6942 global_rt_period(), global_rt_runtime());
6944 #ifdef CONFIG_RT_GROUP_SCHED
6945 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6946 global_rt_period(), global_rt_runtime());
6947 #endif /* CONFIG_RT_GROUP_SCHED */
6949 #ifdef CONFIG_CGROUP_SCHED
6950 list_add(&root_task_group
.list
, &task_groups
);
6951 INIT_LIST_HEAD(&root_task_group
.children
);
6952 INIT_LIST_HEAD(&root_task_group
.siblings
);
6953 autogroup_init(&init_task
);
6955 #endif /* CONFIG_CGROUP_SCHED */
6957 #ifdef CONFIG_CGROUP_CPUACCT
6958 root_cpuacct
.cpustat
= &kernel_cpustat
;
6959 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6960 /* Too early, not expected to fail */
6961 BUG_ON(!root_cpuacct
.cpuusage
);
6963 for_each_possible_cpu(i
) {
6967 raw_spin_lock_init(&rq
->lock
);
6969 rq
->calc_load_active
= 0;
6970 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6971 init_cfs_rq(&rq
->cfs
);
6972 init_rt_rq(&rq
->rt
, rq
);
6973 #ifdef CONFIG_FAIR_GROUP_SCHED
6974 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6975 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6977 * How much cpu bandwidth does root_task_group get?
6979 * In case of task-groups formed thr' the cgroup filesystem, it
6980 * gets 100% of the cpu resources in the system. This overall
6981 * system cpu resource is divided among the tasks of
6982 * root_task_group and its child task-groups in a fair manner,
6983 * based on each entity's (task or task-group's) weight
6984 * (se->load.weight).
6986 * In other words, if root_task_group has 10 tasks of weight
6987 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6988 * then A0's share of the cpu resource is:
6990 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6992 * We achieve this by letting root_task_group's tasks sit
6993 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6995 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6996 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6997 #endif /* CONFIG_FAIR_GROUP_SCHED */
6999 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7000 #ifdef CONFIG_RT_GROUP_SCHED
7001 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7002 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7005 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7006 rq
->cpu_load
[j
] = 0;
7008 rq
->last_load_update_tick
= jiffies
;
7013 rq
->cpu_power
= SCHED_POWER_SCALE
;
7014 rq
->post_schedule
= 0;
7015 rq
->active_balance
= 0;
7016 rq
->next_balance
= jiffies
;
7021 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7023 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7025 rq_attach_root(rq
, &def_root_domain
);
7031 atomic_set(&rq
->nr_iowait
, 0);
7034 set_load_weight(&init_task
);
7036 #ifdef CONFIG_PREEMPT_NOTIFIERS
7037 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7040 #ifdef CONFIG_RT_MUTEXES
7041 plist_head_init(&init_task
.pi_waiters
);
7045 * The boot idle thread does lazy MMU switching as well:
7047 atomic_inc(&init_mm
.mm_count
);
7048 enter_lazy_tlb(&init_mm
, current
);
7051 * Make us the idle thread. Technically, schedule() should not be
7052 * called from this thread, however somewhere below it might be,
7053 * but because we are the idle thread, we just pick up running again
7054 * when this runqueue becomes "idle".
7056 init_idle(current
, smp_processor_id());
7058 calc_load_update
= jiffies
+ LOAD_FREQ
;
7061 * During early bootup we pretend to be a normal task:
7063 current
->sched_class
= &fair_sched_class
;
7066 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7067 /* May be allocated at isolcpus cmdline parse time */
7068 if (cpu_isolated_map
== NULL
)
7069 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7070 idle_thread_set_boot_cpu();
7072 init_sched_fair_class();
7074 scheduler_running
= 1;
7077 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7078 static inline int preempt_count_equals(int preempt_offset
)
7080 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7082 return (nested
== preempt_offset
);
7085 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7087 static unsigned long prev_jiffy
; /* ratelimiting */
7089 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7090 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7091 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7093 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7095 prev_jiffy
= jiffies
;
7098 "BUG: sleeping function called from invalid context at %s:%d\n",
7101 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7102 in_atomic(), irqs_disabled(),
7103 current
->pid
, current
->comm
);
7105 debug_show_held_locks(current
);
7106 if (irqs_disabled())
7107 print_irqtrace_events(current
);
7110 EXPORT_SYMBOL(__might_sleep
);
7113 #ifdef CONFIG_MAGIC_SYSRQ
7114 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7116 const struct sched_class
*prev_class
= p
->sched_class
;
7117 int old_prio
= p
->prio
;
7122 dequeue_task(rq
, p
, 0);
7123 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7125 enqueue_task(rq
, p
, 0);
7126 resched_task(rq
->curr
);
7129 check_class_changed(rq
, p
, prev_class
, old_prio
);
7132 void normalize_rt_tasks(void)
7134 struct task_struct
*g
, *p
;
7135 unsigned long flags
;
7138 read_lock_irqsave(&tasklist_lock
, flags
);
7139 do_each_thread(g
, p
) {
7141 * Only normalize user tasks:
7146 p
->se
.exec_start
= 0;
7147 #ifdef CONFIG_SCHEDSTATS
7148 p
->se
.statistics
.wait_start
= 0;
7149 p
->se
.statistics
.sleep_start
= 0;
7150 p
->se
.statistics
.block_start
= 0;
7155 * Renice negative nice level userspace
7158 if (TASK_NICE(p
) < 0 && p
->mm
)
7159 set_user_nice(p
, 0);
7163 raw_spin_lock(&p
->pi_lock
);
7164 rq
= __task_rq_lock(p
);
7166 normalize_task(rq
, p
);
7168 __task_rq_unlock(rq
);
7169 raw_spin_unlock(&p
->pi_lock
);
7170 } while_each_thread(g
, p
);
7172 read_unlock_irqrestore(&tasklist_lock
, flags
);
7175 #endif /* CONFIG_MAGIC_SYSRQ */
7177 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7179 * These functions are only useful for the IA64 MCA handling, or kdb.
7181 * They can only be called when the whole system has been
7182 * stopped - every CPU needs to be quiescent, and no scheduling
7183 * activity can take place. Using them for anything else would
7184 * be a serious bug, and as a result, they aren't even visible
7185 * under any other configuration.
7189 * curr_task - return the current task for a given cpu.
7190 * @cpu: the processor in question.
7192 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7194 struct task_struct
*curr_task(int cpu
)
7196 return cpu_curr(cpu
);
7199 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7203 * set_curr_task - set the current task for a given cpu.
7204 * @cpu: the processor in question.
7205 * @p: the task pointer to set.
7207 * Description: This function must only be used when non-maskable interrupts
7208 * are serviced on a separate stack. It allows the architecture to switch the
7209 * notion of the current task on a cpu in a non-blocking manner. This function
7210 * must be called with all CPU's synchronized, and interrupts disabled, the
7211 * and caller must save the original value of the current task (see
7212 * curr_task() above) and restore that value before reenabling interrupts and
7213 * re-starting the system.
7215 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7217 void set_curr_task(int cpu
, struct task_struct
*p
)
7224 #ifdef CONFIG_CGROUP_SCHED
7225 /* task_group_lock serializes the addition/removal of task groups */
7226 static DEFINE_SPINLOCK(task_group_lock
);
7228 static void free_sched_group(struct task_group
*tg
)
7230 free_fair_sched_group(tg
);
7231 free_rt_sched_group(tg
);
7236 /* allocate runqueue etc for a new task group */
7237 struct task_group
*sched_create_group(struct task_group
*parent
)
7239 struct task_group
*tg
;
7241 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7243 return ERR_PTR(-ENOMEM
);
7245 if (!alloc_fair_sched_group(tg
, parent
))
7248 if (!alloc_rt_sched_group(tg
, parent
))
7254 free_sched_group(tg
);
7255 return ERR_PTR(-ENOMEM
);
7258 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7260 unsigned long flags
;
7262 spin_lock_irqsave(&task_group_lock
, flags
);
7263 list_add_rcu(&tg
->list
, &task_groups
);
7265 WARN_ON(!parent
); /* root should already exist */
7267 tg
->parent
= parent
;
7268 INIT_LIST_HEAD(&tg
->children
);
7269 list_add_rcu(&tg
->siblings
, &parent
->children
);
7270 spin_unlock_irqrestore(&task_group_lock
, flags
);
7273 /* rcu callback to free various structures associated with a task group */
7274 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7276 /* now it should be safe to free those cfs_rqs */
7277 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7280 /* Destroy runqueue etc associated with a task group */
7281 void sched_destroy_group(struct task_group
*tg
)
7283 /* wait for possible concurrent references to cfs_rqs complete */
7284 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7287 void sched_offline_group(struct task_group
*tg
)
7289 unsigned long flags
;
7292 /* end participation in shares distribution */
7293 for_each_possible_cpu(i
)
7294 unregister_fair_sched_group(tg
, i
);
7296 spin_lock_irqsave(&task_group_lock
, flags
);
7297 list_del_rcu(&tg
->list
);
7298 list_del_rcu(&tg
->siblings
);
7299 spin_unlock_irqrestore(&task_group_lock
, flags
);
7302 /* change task's runqueue when it moves between groups.
7303 * The caller of this function should have put the task in its new group
7304 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7305 * reflect its new group.
7307 void sched_move_task(struct task_struct
*tsk
)
7309 struct task_group
*tg
;
7311 unsigned long flags
;
7314 rq
= task_rq_lock(tsk
, &flags
);
7316 running
= task_current(rq
, tsk
);
7320 dequeue_task(rq
, tsk
, 0);
7321 if (unlikely(running
))
7322 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7324 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7325 lockdep_is_held(&tsk
->sighand
->siglock
)),
7326 struct task_group
, css
);
7327 tg
= autogroup_task_group(tsk
, tg
);
7328 tsk
->sched_task_group
= tg
;
7330 #ifdef CONFIG_FAIR_GROUP_SCHED
7331 if (tsk
->sched_class
->task_move_group
)
7332 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7335 set_task_rq(tsk
, task_cpu(tsk
));
7337 if (unlikely(running
))
7338 tsk
->sched_class
->set_curr_task(rq
);
7340 enqueue_task(rq
, tsk
, 0);
7342 task_rq_unlock(rq
, tsk
, &flags
);
7344 #endif /* CONFIG_CGROUP_SCHED */
7346 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7347 static unsigned long to_ratio(u64 period
, u64 runtime
)
7349 if (runtime
== RUNTIME_INF
)
7352 return div64_u64(runtime
<< 20, period
);
7356 #ifdef CONFIG_RT_GROUP_SCHED
7358 * Ensure that the real time constraints are schedulable.
7360 static DEFINE_MUTEX(rt_constraints_mutex
);
7362 /* Must be called with tasklist_lock held */
7363 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7365 struct task_struct
*g
, *p
;
7367 do_each_thread(g
, p
) {
7368 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7370 } while_each_thread(g
, p
);
7375 struct rt_schedulable_data
{
7376 struct task_group
*tg
;
7381 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7383 struct rt_schedulable_data
*d
= data
;
7384 struct task_group
*child
;
7385 unsigned long total
, sum
= 0;
7386 u64 period
, runtime
;
7388 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7389 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7392 period
= d
->rt_period
;
7393 runtime
= d
->rt_runtime
;
7397 * Cannot have more runtime than the period.
7399 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7403 * Ensure we don't starve existing RT tasks.
7405 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7408 total
= to_ratio(period
, runtime
);
7411 * Nobody can have more than the global setting allows.
7413 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7417 * The sum of our children's runtime should not exceed our own.
7419 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7420 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7421 runtime
= child
->rt_bandwidth
.rt_runtime
;
7423 if (child
== d
->tg
) {
7424 period
= d
->rt_period
;
7425 runtime
= d
->rt_runtime
;
7428 sum
+= to_ratio(period
, runtime
);
7437 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7441 struct rt_schedulable_data data
= {
7443 .rt_period
= period
,
7444 .rt_runtime
= runtime
,
7448 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7454 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7455 u64 rt_period
, u64 rt_runtime
)
7459 mutex_lock(&rt_constraints_mutex
);
7460 read_lock(&tasklist_lock
);
7461 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7465 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7466 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7467 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7469 for_each_possible_cpu(i
) {
7470 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7472 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7473 rt_rq
->rt_runtime
= rt_runtime
;
7474 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7476 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7478 read_unlock(&tasklist_lock
);
7479 mutex_unlock(&rt_constraints_mutex
);
7484 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7486 u64 rt_runtime
, rt_period
;
7488 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7489 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7490 if (rt_runtime_us
< 0)
7491 rt_runtime
= RUNTIME_INF
;
7493 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7496 static long sched_group_rt_runtime(struct task_group
*tg
)
7500 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7503 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7504 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7505 return rt_runtime_us
;
7508 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7510 u64 rt_runtime
, rt_period
;
7512 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7513 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7518 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7521 static long sched_group_rt_period(struct task_group
*tg
)
7525 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7526 do_div(rt_period_us
, NSEC_PER_USEC
);
7527 return rt_period_us
;
7530 static int sched_rt_global_constraints(void)
7532 u64 runtime
, period
;
7535 if (sysctl_sched_rt_period
<= 0)
7538 runtime
= global_rt_runtime();
7539 period
= global_rt_period();
7542 * Sanity check on the sysctl variables.
7544 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7547 mutex_lock(&rt_constraints_mutex
);
7548 read_lock(&tasklist_lock
);
7549 ret
= __rt_schedulable(NULL
, 0, 0);
7550 read_unlock(&tasklist_lock
);
7551 mutex_unlock(&rt_constraints_mutex
);
7556 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7558 /* Don't accept realtime tasks when there is no way for them to run */
7559 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7565 #else /* !CONFIG_RT_GROUP_SCHED */
7566 static int sched_rt_global_constraints(void)
7568 unsigned long flags
;
7571 if (sysctl_sched_rt_period
<= 0)
7575 * There's always some RT tasks in the root group
7576 * -- migration, kstopmachine etc..
7578 if (sysctl_sched_rt_runtime
== 0)
7581 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7582 for_each_possible_cpu(i
) {
7583 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7585 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7586 rt_rq
->rt_runtime
= global_rt_runtime();
7587 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7589 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7593 #endif /* CONFIG_RT_GROUP_SCHED */
7595 int sched_rr_handler(struct ctl_table
*table
, int write
,
7596 void __user
*buffer
, size_t *lenp
,
7600 static DEFINE_MUTEX(mutex
);
7603 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7604 /* make sure that internally we keep jiffies */
7605 /* also, writing zero resets timeslice to default */
7606 if (!ret
&& write
) {
7607 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7608 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7610 mutex_unlock(&mutex
);
7614 int sched_rt_handler(struct ctl_table
*table
, int write
,
7615 void __user
*buffer
, size_t *lenp
,
7619 int old_period
, old_runtime
;
7620 static DEFINE_MUTEX(mutex
);
7623 old_period
= sysctl_sched_rt_period
;
7624 old_runtime
= sysctl_sched_rt_runtime
;
7626 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7628 if (!ret
&& write
) {
7629 ret
= sched_rt_global_constraints();
7631 sysctl_sched_rt_period
= old_period
;
7632 sysctl_sched_rt_runtime
= old_runtime
;
7634 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7635 def_rt_bandwidth
.rt_period
=
7636 ns_to_ktime(global_rt_period());
7639 mutex_unlock(&mutex
);
7644 #ifdef CONFIG_CGROUP_SCHED
7646 /* return corresponding task_group object of a cgroup */
7647 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7649 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7650 struct task_group
, css
);
7653 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7655 struct task_group
*tg
, *parent
;
7657 if (!cgrp
->parent
) {
7658 /* This is early initialization for the top cgroup */
7659 return &root_task_group
.css
;
7662 parent
= cgroup_tg(cgrp
->parent
);
7663 tg
= sched_create_group(parent
);
7665 return ERR_PTR(-ENOMEM
);
7670 static int cpu_cgroup_css_online(struct cgroup
*cgrp
)
7672 struct task_group
*tg
= cgroup_tg(cgrp
);
7673 struct task_group
*parent
;
7678 parent
= cgroup_tg(cgrp
->parent
);
7679 sched_online_group(tg
, parent
);
7683 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7685 struct task_group
*tg
= cgroup_tg(cgrp
);
7687 sched_destroy_group(tg
);
7690 static void cpu_cgroup_css_offline(struct cgroup
*cgrp
)
7692 struct task_group
*tg
= cgroup_tg(cgrp
);
7694 sched_offline_group(tg
);
7697 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7698 struct cgroup_taskset
*tset
)
7700 struct task_struct
*task
;
7702 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7703 #ifdef CONFIG_RT_GROUP_SCHED
7704 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7707 /* We don't support RT-tasks being in separate groups */
7708 if (task
->sched_class
!= &fair_sched_class
)
7715 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7716 struct cgroup_taskset
*tset
)
7718 struct task_struct
*task
;
7720 cgroup_taskset_for_each(task
, cgrp
, tset
)
7721 sched_move_task(task
);
7725 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7726 struct task_struct
*task
)
7729 * cgroup_exit() is called in the copy_process() failure path.
7730 * Ignore this case since the task hasn't ran yet, this avoids
7731 * trying to poke a half freed task state from generic code.
7733 if (!(task
->flags
& PF_EXITING
))
7736 sched_move_task(task
);
7739 #ifdef CONFIG_FAIR_GROUP_SCHED
7740 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7743 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7746 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7748 struct task_group
*tg
= cgroup_tg(cgrp
);
7750 return (u64
) scale_load_down(tg
->shares
);
7753 #ifdef CONFIG_CFS_BANDWIDTH
7754 static DEFINE_MUTEX(cfs_constraints_mutex
);
7756 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7757 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7759 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7761 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7763 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7764 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7766 if (tg
== &root_task_group
)
7770 * Ensure we have at some amount of bandwidth every period. This is
7771 * to prevent reaching a state of large arrears when throttled via
7772 * entity_tick() resulting in prolonged exit starvation.
7774 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7778 * Likewise, bound things on the otherside by preventing insane quota
7779 * periods. This also allows us to normalize in computing quota
7782 if (period
> max_cfs_quota_period
)
7785 mutex_lock(&cfs_constraints_mutex
);
7786 ret
= __cfs_schedulable(tg
, period
, quota
);
7790 runtime_enabled
= quota
!= RUNTIME_INF
;
7791 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7792 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7793 raw_spin_lock_irq(&cfs_b
->lock
);
7794 cfs_b
->period
= ns_to_ktime(period
);
7795 cfs_b
->quota
= quota
;
7797 __refill_cfs_bandwidth_runtime(cfs_b
);
7798 /* restart the period timer (if active) to handle new period expiry */
7799 if (runtime_enabled
&& cfs_b
->timer_active
) {
7800 /* force a reprogram */
7801 cfs_b
->timer_active
= 0;
7802 __start_cfs_bandwidth(cfs_b
);
7804 raw_spin_unlock_irq(&cfs_b
->lock
);
7806 for_each_possible_cpu(i
) {
7807 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7808 struct rq
*rq
= cfs_rq
->rq
;
7810 raw_spin_lock_irq(&rq
->lock
);
7811 cfs_rq
->runtime_enabled
= runtime_enabled
;
7812 cfs_rq
->runtime_remaining
= 0;
7814 if (cfs_rq
->throttled
)
7815 unthrottle_cfs_rq(cfs_rq
);
7816 raw_spin_unlock_irq(&rq
->lock
);
7819 mutex_unlock(&cfs_constraints_mutex
);
7824 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7828 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7829 if (cfs_quota_us
< 0)
7830 quota
= RUNTIME_INF
;
7832 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7834 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7837 long tg_get_cfs_quota(struct task_group
*tg
)
7841 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7844 quota_us
= tg
->cfs_bandwidth
.quota
;
7845 do_div(quota_us
, NSEC_PER_USEC
);
7850 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7854 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7855 quota
= tg
->cfs_bandwidth
.quota
;
7857 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7860 long tg_get_cfs_period(struct task_group
*tg
)
7864 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7865 do_div(cfs_period_us
, NSEC_PER_USEC
);
7867 return cfs_period_us
;
7870 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7872 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7875 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7878 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7881 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7883 return tg_get_cfs_period(cgroup_tg(cgrp
));
7886 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7889 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7892 struct cfs_schedulable_data
{
7893 struct task_group
*tg
;
7898 * normalize group quota/period to be quota/max_period
7899 * note: units are usecs
7901 static u64
normalize_cfs_quota(struct task_group
*tg
,
7902 struct cfs_schedulable_data
*d
)
7910 period
= tg_get_cfs_period(tg
);
7911 quota
= tg_get_cfs_quota(tg
);
7914 /* note: these should typically be equivalent */
7915 if (quota
== RUNTIME_INF
|| quota
== -1)
7918 return to_ratio(period
, quota
);
7921 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7923 struct cfs_schedulable_data
*d
= data
;
7924 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7925 s64 quota
= 0, parent_quota
= -1;
7928 quota
= RUNTIME_INF
;
7930 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7932 quota
= normalize_cfs_quota(tg
, d
);
7933 parent_quota
= parent_b
->hierarchal_quota
;
7936 * ensure max(child_quota) <= parent_quota, inherit when no
7939 if (quota
== RUNTIME_INF
)
7940 quota
= parent_quota
;
7941 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7944 cfs_b
->hierarchal_quota
= quota
;
7949 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7952 struct cfs_schedulable_data data
= {
7958 if (quota
!= RUNTIME_INF
) {
7959 do_div(data
.period
, NSEC_PER_USEC
);
7960 do_div(data
.quota
, NSEC_PER_USEC
);
7964 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7970 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7971 struct cgroup_map_cb
*cb
)
7973 struct task_group
*tg
= cgroup_tg(cgrp
);
7974 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7976 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7977 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7978 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7982 #endif /* CONFIG_CFS_BANDWIDTH */
7983 #endif /* CONFIG_FAIR_GROUP_SCHED */
7985 #ifdef CONFIG_RT_GROUP_SCHED
7986 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7989 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7992 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7994 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7997 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8000 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8003 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8005 return sched_group_rt_period(cgroup_tg(cgrp
));
8007 #endif /* CONFIG_RT_GROUP_SCHED */
8009 static struct cftype cpu_files
[] = {
8010 #ifdef CONFIG_FAIR_GROUP_SCHED
8013 .read_u64
= cpu_shares_read_u64
,
8014 .write_u64
= cpu_shares_write_u64
,
8017 #ifdef CONFIG_CFS_BANDWIDTH
8019 .name
= "cfs_quota_us",
8020 .read_s64
= cpu_cfs_quota_read_s64
,
8021 .write_s64
= cpu_cfs_quota_write_s64
,
8024 .name
= "cfs_period_us",
8025 .read_u64
= cpu_cfs_period_read_u64
,
8026 .write_u64
= cpu_cfs_period_write_u64
,
8030 .read_map
= cpu_stats_show
,
8033 #ifdef CONFIG_RT_GROUP_SCHED
8035 .name
= "rt_runtime_us",
8036 .read_s64
= cpu_rt_runtime_read
,
8037 .write_s64
= cpu_rt_runtime_write
,
8040 .name
= "rt_period_us",
8041 .read_u64
= cpu_rt_period_read_uint
,
8042 .write_u64
= cpu_rt_period_write_uint
,
8048 struct cgroup_subsys cpu_cgroup_subsys
= {
8050 .css_alloc
= cpu_cgroup_css_alloc
,
8051 .css_free
= cpu_cgroup_css_free
,
8052 .css_online
= cpu_cgroup_css_online
,
8053 .css_offline
= cpu_cgroup_css_offline
,
8054 .can_attach
= cpu_cgroup_can_attach
,
8055 .attach
= cpu_cgroup_attach
,
8056 .exit
= cpu_cgroup_exit
,
8057 .subsys_id
= cpu_cgroup_subsys_id
,
8058 .base_cftypes
= cpu_files
,
8062 #endif /* CONFIG_CGROUP_SCHED */
8064 #ifdef CONFIG_CGROUP_CPUACCT
8067 * CPU accounting code for task groups.
8069 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8070 * (balbir@in.ibm.com).
8073 struct cpuacct root_cpuacct
;
8075 /* create a new cpu accounting group */
8076 static struct cgroup_subsys_state
*cpuacct_css_alloc(struct cgroup
*cgrp
)
8081 return &root_cpuacct
.css
;
8083 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8087 ca
->cpuusage
= alloc_percpu(u64
);
8091 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
8093 goto out_free_cpuusage
;
8098 free_percpu(ca
->cpuusage
);
8102 return ERR_PTR(-ENOMEM
);
8105 /* destroy an existing cpu accounting group */
8106 static void cpuacct_css_free(struct cgroup
*cgrp
)
8108 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8110 free_percpu(ca
->cpustat
);
8111 free_percpu(ca
->cpuusage
);
8115 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8117 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8120 #ifndef CONFIG_64BIT
8122 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8124 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8126 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8134 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8136 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8138 #ifndef CONFIG_64BIT
8140 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8142 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8144 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8150 /* return total cpu usage (in nanoseconds) of a group */
8151 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8153 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8154 u64 totalcpuusage
= 0;
8157 for_each_present_cpu(i
)
8158 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8160 return totalcpuusage
;
8163 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8166 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8175 for_each_present_cpu(i
)
8176 cpuacct_cpuusage_write(ca
, i
, 0);
8182 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8185 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8189 for_each_present_cpu(i
) {
8190 percpu
= cpuacct_cpuusage_read(ca
, i
);
8191 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8193 seq_printf(m
, "\n");
8197 static const char *cpuacct_stat_desc
[] = {
8198 [CPUACCT_STAT_USER
] = "user",
8199 [CPUACCT_STAT_SYSTEM
] = "system",
8202 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8203 struct cgroup_map_cb
*cb
)
8205 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8209 for_each_online_cpu(cpu
) {
8210 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8211 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8212 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8214 val
= cputime64_to_clock_t(val
);
8215 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8218 for_each_online_cpu(cpu
) {
8219 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8220 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8221 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8222 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8225 val
= cputime64_to_clock_t(val
);
8226 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8231 static struct cftype files
[] = {
8234 .read_u64
= cpuusage_read
,
8235 .write_u64
= cpuusage_write
,
8238 .name
= "usage_percpu",
8239 .read_seq_string
= cpuacct_percpu_seq_read
,
8243 .read_map
= cpuacct_stats_show
,
8249 * charge this task's execution time to its accounting group.
8251 * called with rq->lock held.
8253 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8258 if (unlikely(!cpuacct_subsys
.active
))
8261 cpu
= task_cpu(tsk
);
8267 for (; ca
; ca
= parent_ca(ca
)) {
8268 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8269 *cpuusage
+= cputime
;
8275 struct cgroup_subsys cpuacct_subsys
= {
8277 .css_alloc
= cpuacct_css_alloc
,
8278 .css_free
= cpuacct_css_free
,
8279 .subsys_id
= cpuacct_subsys_id
,
8280 .base_cftypes
= files
,
8282 #endif /* CONFIG_CGROUP_CPUACCT */
8284 void dump_cpu_task(int cpu
)
8286 pr_info("Task dump for CPU %d:\n", cpu
);
8287 sched_show_task(cpu_curr(cpu
));