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 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 inline bool got_nohz_idle_kick(void)
622 int cpu
= smp_processor_id();
623 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
626 #else /* CONFIG_NO_HZ */
628 static inline bool got_nohz_idle_kick(void)
633 #endif /* CONFIG_NO_HZ */
635 void sched_avg_update(struct rq
*rq
)
637 s64 period
= sched_avg_period();
639 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
641 * Inline assembly required to prevent the compiler
642 * optimising this loop into a divmod call.
643 * See __iter_div_u64_rem() for another example of this.
645 asm("" : "+rm" (rq
->age_stamp
));
646 rq
->age_stamp
+= period
;
651 #else /* !CONFIG_SMP */
652 void resched_task(struct task_struct
*p
)
654 assert_raw_spin_locked(&task_rq(p
)->lock
);
655 set_tsk_need_resched(p
);
657 #endif /* CONFIG_SMP */
659 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
662 * Iterate task_group tree rooted at *from, calling @down when first entering a
663 * node and @up when leaving it for the final time.
665 * Caller must hold rcu_lock or sufficient equivalent.
667 int walk_tg_tree_from(struct task_group
*from
,
668 tg_visitor down
, tg_visitor up
, void *data
)
670 struct task_group
*parent
, *child
;
676 ret
= (*down
)(parent
, data
);
679 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
686 ret
= (*up
)(parent
, data
);
687 if (ret
|| parent
== from
)
691 parent
= parent
->parent
;
698 int tg_nop(struct task_group
*tg
, void *data
)
704 static void set_load_weight(struct task_struct
*p
)
706 int prio
= p
->static_prio
- MAX_RT_PRIO
;
707 struct load_weight
*load
= &p
->se
.load
;
710 * SCHED_IDLE tasks get minimal weight:
712 if (p
->policy
== SCHED_IDLE
) {
713 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
714 load
->inv_weight
= WMULT_IDLEPRIO
;
718 load
->weight
= scale_load(prio_to_weight
[prio
]);
719 load
->inv_weight
= prio_to_wmult
[prio
];
722 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
725 sched_info_queued(p
);
726 p
->sched_class
->enqueue_task(rq
, p
, flags
);
729 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
732 sched_info_dequeued(p
);
733 p
->sched_class
->dequeue_task(rq
, p
, flags
);
736 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
738 if (task_contributes_to_load(p
))
739 rq
->nr_uninterruptible
--;
741 enqueue_task(rq
, p
, flags
);
744 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
746 if (task_contributes_to_load(p
))
747 rq
->nr_uninterruptible
++;
749 dequeue_task(rq
, p
, flags
);
752 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
755 * In theory, the compile should just see 0 here, and optimize out the call
756 * to sched_rt_avg_update. But I don't trust it...
758 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
759 s64 steal
= 0, irq_delta
= 0;
761 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
762 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
765 * Since irq_time is only updated on {soft,}irq_exit, we might run into
766 * this case when a previous update_rq_clock() happened inside a
769 * When this happens, we stop ->clock_task and only update the
770 * prev_irq_time stamp to account for the part that fit, so that a next
771 * update will consume the rest. This ensures ->clock_task is
774 * It does however cause some slight miss-attribution of {soft,}irq
775 * time, a more accurate solution would be to update the irq_time using
776 * the current rq->clock timestamp, except that would require using
779 if (irq_delta
> delta
)
782 rq
->prev_irq_time
+= irq_delta
;
785 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
786 if (static_key_false((¶virt_steal_rq_enabled
))) {
789 steal
= paravirt_steal_clock(cpu_of(rq
));
790 steal
-= rq
->prev_steal_time_rq
;
792 if (unlikely(steal
> delta
))
795 st
= steal_ticks(steal
);
796 steal
= st
* TICK_NSEC
;
798 rq
->prev_steal_time_rq
+= steal
;
804 rq
->clock_task
+= delta
;
806 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
807 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
808 sched_rt_avg_update(rq
, irq_delta
+ steal
);
812 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
814 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
815 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
819 * Make it appear like a SCHED_FIFO task, its something
820 * userspace knows about and won't get confused about.
822 * Also, it will make PI more or less work without too
823 * much confusion -- but then, stop work should not
824 * rely on PI working anyway.
826 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
828 stop
->sched_class
= &stop_sched_class
;
831 cpu_rq(cpu
)->stop
= stop
;
835 * Reset it back to a normal scheduling class so that
836 * it can die in pieces.
838 old_stop
->sched_class
= &rt_sched_class
;
843 * __normal_prio - return the priority that is based on the static prio
845 static inline int __normal_prio(struct task_struct
*p
)
847 return p
->static_prio
;
851 * Calculate the expected normal priority: i.e. priority
852 * without taking RT-inheritance into account. Might be
853 * boosted by interactivity modifiers. Changes upon fork,
854 * setprio syscalls, and whenever the interactivity
855 * estimator recalculates.
857 static inline int normal_prio(struct task_struct
*p
)
861 if (task_has_rt_policy(p
))
862 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
864 prio
= __normal_prio(p
);
869 * Calculate the current priority, i.e. the priority
870 * taken into account by the scheduler. This value might
871 * be boosted by RT tasks, or might be boosted by
872 * interactivity modifiers. Will be RT if the task got
873 * RT-boosted. If not then it returns p->normal_prio.
875 static int effective_prio(struct task_struct
*p
)
877 p
->normal_prio
= normal_prio(p
);
879 * If we are RT tasks or we were boosted to RT priority,
880 * keep the priority unchanged. Otherwise, update priority
881 * to the normal priority:
883 if (!rt_prio(p
->prio
))
884 return p
->normal_prio
;
889 * task_curr - is this task currently executing on a CPU?
890 * @p: the task in question.
892 inline int task_curr(const struct task_struct
*p
)
894 return cpu_curr(task_cpu(p
)) == p
;
897 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
898 const struct sched_class
*prev_class
,
901 if (prev_class
!= p
->sched_class
) {
902 if (prev_class
->switched_from
)
903 prev_class
->switched_from(rq
, p
);
904 p
->sched_class
->switched_to(rq
, p
);
905 } else if (oldprio
!= p
->prio
)
906 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
909 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
911 const struct sched_class
*class;
913 if (p
->sched_class
== rq
->curr
->sched_class
) {
914 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
916 for_each_class(class) {
917 if (class == rq
->curr
->sched_class
)
919 if (class == p
->sched_class
) {
920 resched_task(rq
->curr
);
927 * A queue event has occurred, and we're going to schedule. In
928 * this case, we can save a useless back to back clock update.
930 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
931 rq
->skip_clock_update
= 1;
934 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
936 void register_task_migration_notifier(struct notifier_block
*n
)
938 atomic_notifier_chain_register(&task_migration_notifier
, n
);
942 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
944 #ifdef CONFIG_SCHED_DEBUG
946 * We should never call set_task_cpu() on a blocked task,
947 * ttwu() will sort out the placement.
949 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
950 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
952 #ifdef CONFIG_LOCKDEP
954 * The caller should hold either p->pi_lock or rq->lock, when changing
955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
957 * sched_move_task() holds both and thus holding either pins the cgroup,
960 * Furthermore, all task_rq users should acquire both locks, see
963 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
964 lockdep_is_held(&task_rq(p
)->lock
)));
968 trace_sched_migrate_task(p
, new_cpu
);
970 if (task_cpu(p
) != new_cpu
) {
971 struct task_migration_notifier tmn
;
973 if (p
->sched_class
->migrate_task_rq
)
974 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
975 p
->se
.nr_migrations
++;
976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
979 tmn
.from_cpu
= task_cpu(p
);
980 tmn
.to_cpu
= new_cpu
;
982 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
985 __set_task_cpu(p
, new_cpu
);
988 struct migration_arg
{
989 struct task_struct
*task
;
993 static int migration_cpu_stop(void *data
);
996 * wait_task_inactive - wait for a thread to unschedule.
998 * If @match_state is nonzero, it's the @p->state value just checked and
999 * not expected to change. If it changes, i.e. @p might have woken up,
1000 * then return zero. When we succeed in waiting for @p to be off its CPU,
1001 * we return a positive number (its total switch count). If a second call
1002 * a short while later returns the same number, the caller can be sure that
1003 * @p has remained unscheduled the whole time.
1005 * The caller must ensure that the task *will* unschedule sometime soon,
1006 * else this function might spin for a *long* time. This function can't
1007 * be called with interrupts off, or it may introduce deadlock with
1008 * smp_call_function() if an IPI is sent by the same process we are
1009 * waiting to become inactive.
1011 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1013 unsigned long flags
;
1020 * We do the initial early heuristics without holding
1021 * any task-queue locks at all. We'll only try to get
1022 * the runqueue lock when things look like they will
1028 * If the task is actively running on another CPU
1029 * still, just relax and busy-wait without holding
1032 * NOTE! Since we don't hold any locks, it's not
1033 * even sure that "rq" stays as the right runqueue!
1034 * But we don't care, since "task_running()" will
1035 * return false if the runqueue has changed and p
1036 * is actually now running somewhere else!
1038 while (task_running(rq
, p
)) {
1039 if (match_state
&& unlikely(p
->state
!= match_state
))
1045 * Ok, time to look more closely! We need the rq
1046 * lock now, to be *sure*. If we're wrong, we'll
1047 * just go back and repeat.
1049 rq
= task_rq_lock(p
, &flags
);
1050 trace_sched_wait_task(p
);
1051 running
= task_running(rq
, p
);
1054 if (!match_state
|| p
->state
== match_state
)
1055 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1056 task_rq_unlock(rq
, p
, &flags
);
1059 * If it changed from the expected state, bail out now.
1061 if (unlikely(!ncsw
))
1065 * Was it really running after all now that we
1066 * checked with the proper locks actually held?
1068 * Oops. Go back and try again..
1070 if (unlikely(running
)) {
1076 * It's not enough that it's not actively running,
1077 * it must be off the runqueue _entirely_, and not
1080 * So if it was still runnable (but just not actively
1081 * running right now), it's preempted, and we should
1082 * yield - it could be a while.
1084 if (unlikely(on_rq
)) {
1085 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1087 set_current_state(TASK_UNINTERRUPTIBLE
);
1088 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1093 * Ahh, all good. It wasn't running, and it wasn't
1094 * runnable, which means that it will never become
1095 * running in the future either. We're all done!
1104 * kick_process - kick a running thread to enter/exit the kernel
1105 * @p: the to-be-kicked thread
1107 * Cause a process which is running on another CPU to enter
1108 * kernel-mode, without any delay. (to get signals handled.)
1110 * NOTE: this function doesn't have to take the runqueue lock,
1111 * because all it wants to ensure is that the remote task enters
1112 * the kernel. If the IPI races and the task has been migrated
1113 * to another CPU then no harm is done and the purpose has been
1116 void kick_process(struct task_struct
*p
)
1122 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1123 smp_send_reschedule(cpu
);
1126 EXPORT_SYMBOL_GPL(kick_process
);
1127 #endif /* CONFIG_SMP */
1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1133 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1135 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1136 enum { cpuset
, possible
, fail
} state
= cpuset
;
1139 /* Look for allowed, online CPU in same node. */
1140 for_each_cpu(dest_cpu
, nodemask
) {
1141 if (!cpu_online(dest_cpu
))
1143 if (!cpu_active(dest_cpu
))
1145 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1150 /* Any allowed, online CPU? */
1151 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1152 if (!cpu_online(dest_cpu
))
1154 if (!cpu_active(dest_cpu
))
1161 /* No more Mr. Nice Guy. */
1162 cpuset_cpus_allowed_fallback(p
);
1167 do_set_cpus_allowed(p
, cpu_possible_mask
);
1178 if (state
!= cpuset
) {
1180 * Don't tell them about moving exiting tasks or
1181 * kernel threads (both mm NULL), since they never
1184 if (p
->mm
&& printk_ratelimit()) {
1185 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1186 task_pid_nr(p
), p
->comm
, cpu
);
1194 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1197 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1199 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1202 * In order not to call set_task_cpu() on a blocking task we need
1203 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1206 * Since this is common to all placement strategies, this lives here.
1208 * [ this allows ->select_task() to simply return task_cpu(p) and
1209 * not worry about this generic constraint ]
1211 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1213 cpu
= select_fallback_rq(task_cpu(p
), p
);
1218 static void update_avg(u64
*avg
, u64 sample
)
1220 s64 diff
= sample
- *avg
;
1226 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1228 #ifdef CONFIG_SCHEDSTATS
1229 struct rq
*rq
= this_rq();
1232 int this_cpu
= smp_processor_id();
1234 if (cpu
== this_cpu
) {
1235 schedstat_inc(rq
, ttwu_local
);
1236 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1238 struct sched_domain
*sd
;
1240 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1242 for_each_domain(this_cpu
, sd
) {
1243 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1244 schedstat_inc(sd
, ttwu_wake_remote
);
1251 if (wake_flags
& WF_MIGRATED
)
1252 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1254 #endif /* CONFIG_SMP */
1256 schedstat_inc(rq
, ttwu_count
);
1257 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1259 if (wake_flags
& WF_SYNC
)
1260 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1262 #endif /* CONFIG_SCHEDSTATS */
1265 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1267 activate_task(rq
, p
, en_flags
);
1270 /* if a worker is waking up, notify workqueue */
1271 if (p
->flags
& PF_WQ_WORKER
)
1272 wq_worker_waking_up(p
, cpu_of(rq
));
1276 * Mark the task runnable and perform wakeup-preemption.
1279 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1281 trace_sched_wakeup(p
, true);
1282 check_preempt_curr(rq
, p
, wake_flags
);
1284 p
->state
= TASK_RUNNING
;
1286 if (p
->sched_class
->task_woken
)
1287 p
->sched_class
->task_woken(rq
, p
);
1289 if (rq
->idle_stamp
) {
1290 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1291 u64 max
= 2*sysctl_sched_migration_cost
;
1296 update_avg(&rq
->avg_idle
, delta
);
1303 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1306 if (p
->sched_contributes_to_load
)
1307 rq
->nr_uninterruptible
--;
1310 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1311 ttwu_do_wakeup(rq
, p
, wake_flags
);
1315 * Called in case the task @p isn't fully descheduled from its runqueue,
1316 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1317 * since all we need to do is flip p->state to TASK_RUNNING, since
1318 * the task is still ->on_rq.
1320 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1325 rq
= __task_rq_lock(p
);
1327 ttwu_do_wakeup(rq
, p
, wake_flags
);
1330 __task_rq_unlock(rq
);
1336 static void sched_ttwu_pending(void)
1338 struct rq
*rq
= this_rq();
1339 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1340 struct task_struct
*p
;
1342 raw_spin_lock(&rq
->lock
);
1345 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1346 llist
= llist_next(llist
);
1347 ttwu_do_activate(rq
, p
, 0);
1350 raw_spin_unlock(&rq
->lock
);
1353 void scheduler_ipi(void)
1355 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1359 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1360 * traditionally all their work was done from the interrupt return
1361 * path. Now that we actually do some work, we need to make sure
1364 * Some archs already do call them, luckily irq_enter/exit nest
1367 * Arguably we should visit all archs and update all handlers,
1368 * however a fair share of IPIs are still resched only so this would
1369 * somewhat pessimize the simple resched case.
1372 sched_ttwu_pending();
1375 * Check if someone kicked us for doing the nohz idle load balance.
1377 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1378 this_rq()->idle_balance
= 1;
1379 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1384 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1386 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1387 smp_send_reschedule(cpu
);
1390 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1392 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1394 #endif /* CONFIG_SMP */
1396 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1398 struct rq
*rq
= cpu_rq(cpu
);
1400 #if defined(CONFIG_SMP)
1401 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1402 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1403 ttwu_queue_remote(p
, cpu
);
1408 raw_spin_lock(&rq
->lock
);
1409 ttwu_do_activate(rq
, p
, 0);
1410 raw_spin_unlock(&rq
->lock
);
1414 * try_to_wake_up - wake up a thread
1415 * @p: the thread to be awakened
1416 * @state: the mask of task states that can be woken
1417 * @wake_flags: wake modifier flags (WF_*)
1419 * Put it on the run-queue if it's not already there. The "current"
1420 * thread is always on the run-queue (except when the actual
1421 * re-schedule is in progress), and as such you're allowed to do
1422 * the simpler "current->state = TASK_RUNNING" to mark yourself
1423 * runnable without the overhead of this.
1425 * Returns %true if @p was woken up, %false if it was already running
1426 * or @state didn't match @p's state.
1429 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1431 unsigned long flags
;
1432 int cpu
, success
= 0;
1435 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1436 if (!(p
->state
& state
))
1439 success
= 1; /* we're going to change ->state */
1442 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1447 * If the owning (remote) cpu is still in the middle of schedule() with
1448 * this task as prev, wait until its done referencing the task.
1453 * Pairs with the smp_wmb() in finish_lock_switch().
1457 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1458 p
->state
= TASK_WAKING
;
1460 if (p
->sched_class
->task_waking
)
1461 p
->sched_class
->task_waking(p
);
1463 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1464 if (task_cpu(p
) != cpu
) {
1465 wake_flags
|= WF_MIGRATED
;
1466 set_task_cpu(p
, cpu
);
1468 #endif /* CONFIG_SMP */
1472 ttwu_stat(p
, cpu
, wake_flags
);
1474 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1480 * try_to_wake_up_local - try to wake up a local task with rq lock held
1481 * @p: the thread to be awakened
1483 * Put @p on the run-queue if it's not already there. The caller must
1484 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1487 static void try_to_wake_up_local(struct task_struct
*p
)
1489 struct rq
*rq
= task_rq(p
);
1491 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1492 WARN_ON_ONCE(p
== current
))
1495 lockdep_assert_held(&rq
->lock
);
1497 if (!raw_spin_trylock(&p
->pi_lock
)) {
1498 raw_spin_unlock(&rq
->lock
);
1499 raw_spin_lock(&p
->pi_lock
);
1500 raw_spin_lock(&rq
->lock
);
1503 if (!(p
->state
& TASK_NORMAL
))
1507 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1509 ttwu_do_wakeup(rq
, p
, 0);
1510 ttwu_stat(p
, smp_processor_id(), 0);
1512 raw_spin_unlock(&p
->pi_lock
);
1516 * wake_up_process - Wake up a specific process
1517 * @p: The process to be woken up.
1519 * Attempt to wake up the nominated process and move it to the set of runnable
1520 * processes. Returns 1 if the process was woken up, 0 if it was already
1523 * It may be assumed that this function implies a write memory barrier before
1524 * changing the task state if and only if any tasks are woken up.
1526 int wake_up_process(struct task_struct
*p
)
1528 WARN_ON(task_is_stopped_or_traced(p
));
1529 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1531 EXPORT_SYMBOL(wake_up_process
);
1533 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1535 return try_to_wake_up(p
, state
, 0);
1539 * Perform scheduler related setup for a newly forked process p.
1540 * p is forked by current.
1542 * __sched_fork() is basic setup used by init_idle() too:
1544 static void __sched_fork(struct task_struct
*p
)
1549 p
->se
.exec_start
= 0;
1550 p
->se
.sum_exec_runtime
= 0;
1551 p
->se
.prev_sum_exec_runtime
= 0;
1552 p
->se
.nr_migrations
= 0;
1554 INIT_LIST_HEAD(&p
->se
.group_node
);
1557 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1558 * removed when useful for applications beyond shares distribution (e.g.
1561 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1562 p
->se
.avg
.runnable_avg_period
= 0;
1563 p
->se
.avg
.runnable_avg_sum
= 0;
1565 #ifdef CONFIG_SCHEDSTATS
1566 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1569 INIT_LIST_HEAD(&p
->rt
.run_list
);
1571 #ifdef CONFIG_PREEMPT_NOTIFIERS
1572 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1575 #ifdef CONFIG_NUMA_BALANCING
1576 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1577 p
->mm
->numa_next_scan
= jiffies
;
1578 p
->mm
->numa_next_reset
= jiffies
;
1579 p
->mm
->numa_scan_seq
= 0;
1582 p
->node_stamp
= 0ULL;
1583 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1584 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1585 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1586 p
->numa_work
.next
= &p
->numa_work
;
1587 #endif /* CONFIG_NUMA_BALANCING */
1590 #ifdef CONFIG_NUMA_BALANCING
1591 #ifdef CONFIG_SCHED_DEBUG
1592 void set_numabalancing_state(bool enabled
)
1595 sched_feat_set("NUMA");
1597 sched_feat_set("NO_NUMA");
1600 __read_mostly
bool numabalancing_enabled
;
1602 void set_numabalancing_state(bool enabled
)
1604 numabalancing_enabled
= enabled
;
1606 #endif /* CONFIG_SCHED_DEBUG */
1607 #endif /* CONFIG_NUMA_BALANCING */
1610 * fork()/clone()-time setup:
1612 void sched_fork(struct task_struct
*p
)
1614 unsigned long flags
;
1615 int cpu
= get_cpu();
1619 * We mark the process as running here. This guarantees that
1620 * nobody will actually run it, and a signal or other external
1621 * event cannot wake it up and insert it on the runqueue either.
1623 p
->state
= TASK_RUNNING
;
1626 * Make sure we do not leak PI boosting priority to the child.
1628 p
->prio
= current
->normal_prio
;
1631 * Revert to default priority/policy on fork if requested.
1633 if (unlikely(p
->sched_reset_on_fork
)) {
1634 if (task_has_rt_policy(p
)) {
1635 p
->policy
= SCHED_NORMAL
;
1636 p
->static_prio
= NICE_TO_PRIO(0);
1638 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1639 p
->static_prio
= NICE_TO_PRIO(0);
1641 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1645 * We don't need the reset flag anymore after the fork. It has
1646 * fulfilled its duty:
1648 p
->sched_reset_on_fork
= 0;
1651 if (!rt_prio(p
->prio
))
1652 p
->sched_class
= &fair_sched_class
;
1654 if (p
->sched_class
->task_fork
)
1655 p
->sched_class
->task_fork(p
);
1658 * The child is not yet in the pid-hash so no cgroup attach races,
1659 * and the cgroup is pinned to this child due to cgroup_fork()
1660 * is ran before sched_fork().
1662 * Silence PROVE_RCU.
1664 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1665 set_task_cpu(p
, cpu
);
1666 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1668 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1669 if (likely(sched_info_on()))
1670 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1672 #if defined(CONFIG_SMP)
1675 #ifdef CONFIG_PREEMPT_COUNT
1676 /* Want to start with kernel preemption disabled. */
1677 task_thread_info(p
)->preempt_count
= 1;
1680 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1687 * wake_up_new_task - wake up a newly created task for the first time.
1689 * This function will do some initial scheduler statistics housekeeping
1690 * that must be done for every newly created context, then puts the task
1691 * on the runqueue and wakes it.
1693 void wake_up_new_task(struct task_struct
*p
)
1695 unsigned long flags
;
1698 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1701 * Fork balancing, do it here and not earlier because:
1702 * - cpus_allowed can change in the fork path
1703 * - any previously selected cpu might disappear through hotplug
1705 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1708 rq
= __task_rq_lock(p
);
1709 activate_task(rq
, p
, 0);
1711 trace_sched_wakeup_new(p
, true);
1712 check_preempt_curr(rq
, p
, WF_FORK
);
1714 if (p
->sched_class
->task_woken
)
1715 p
->sched_class
->task_woken(rq
, p
);
1717 task_rq_unlock(rq
, p
, &flags
);
1720 #ifdef CONFIG_PREEMPT_NOTIFIERS
1723 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1724 * @notifier: notifier struct to register
1726 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1728 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1730 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1733 * preempt_notifier_unregister - no longer interested in preemption notifications
1734 * @notifier: notifier struct to unregister
1736 * This is safe to call from within a preemption notifier.
1738 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1740 hlist_del(¬ifier
->link
);
1742 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1744 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1746 struct preempt_notifier
*notifier
;
1747 struct hlist_node
*node
;
1749 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1750 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1754 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1755 struct task_struct
*next
)
1757 struct preempt_notifier
*notifier
;
1758 struct hlist_node
*node
;
1760 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1761 notifier
->ops
->sched_out(notifier
, next
);
1764 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1766 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1771 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1772 struct task_struct
*next
)
1776 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1779 * prepare_task_switch - prepare to switch tasks
1780 * @rq: the runqueue preparing to switch
1781 * @prev: the current task that is being switched out
1782 * @next: the task we are going to switch to.
1784 * This is called with the rq lock held and interrupts off. It must
1785 * be paired with a subsequent finish_task_switch after the context
1788 * prepare_task_switch sets up locking and calls architecture specific
1792 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1793 struct task_struct
*next
)
1795 trace_sched_switch(prev
, next
);
1796 sched_info_switch(prev
, next
);
1797 perf_event_task_sched_out(prev
, next
);
1798 fire_sched_out_preempt_notifiers(prev
, next
);
1799 prepare_lock_switch(rq
, next
);
1800 prepare_arch_switch(next
);
1804 * finish_task_switch - clean up after a task-switch
1805 * @rq: runqueue associated with task-switch
1806 * @prev: the thread we just switched away from.
1808 * finish_task_switch must be called after the context switch, paired
1809 * with a prepare_task_switch call before the context switch.
1810 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1811 * and do any other architecture-specific cleanup actions.
1813 * Note that we may have delayed dropping an mm in context_switch(). If
1814 * so, we finish that here outside of the runqueue lock. (Doing it
1815 * with the lock held can cause deadlocks; see schedule() for
1818 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1819 __releases(rq
->lock
)
1821 struct mm_struct
*mm
= rq
->prev_mm
;
1827 * A task struct has one reference for the use as "current".
1828 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1829 * schedule one last time. The schedule call will never return, and
1830 * the scheduled task must drop that reference.
1831 * The test for TASK_DEAD must occur while the runqueue locks are
1832 * still held, otherwise prev could be scheduled on another cpu, die
1833 * there before we look at prev->state, and then the reference would
1835 * Manfred Spraul <manfred@colorfullife.com>
1837 prev_state
= prev
->state
;
1838 vtime_task_switch(prev
);
1839 finish_arch_switch(prev
);
1840 perf_event_task_sched_in(prev
, current
);
1841 finish_lock_switch(rq
, prev
);
1842 finish_arch_post_lock_switch();
1844 fire_sched_in_preempt_notifiers(current
);
1847 if (unlikely(prev_state
== TASK_DEAD
)) {
1849 * Remove function-return probe instances associated with this
1850 * task and put them back on the free list.
1852 kprobe_flush_task(prev
);
1853 put_task_struct(prev
);
1859 /* assumes rq->lock is held */
1860 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1862 if (prev
->sched_class
->pre_schedule
)
1863 prev
->sched_class
->pre_schedule(rq
, prev
);
1866 /* rq->lock is NOT held, but preemption is disabled */
1867 static inline void post_schedule(struct rq
*rq
)
1869 if (rq
->post_schedule
) {
1870 unsigned long flags
;
1872 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1873 if (rq
->curr
->sched_class
->post_schedule
)
1874 rq
->curr
->sched_class
->post_schedule(rq
);
1875 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1877 rq
->post_schedule
= 0;
1883 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1887 static inline void post_schedule(struct rq
*rq
)
1894 * schedule_tail - first thing a freshly forked thread must call.
1895 * @prev: the thread we just switched away from.
1897 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1898 __releases(rq
->lock
)
1900 struct rq
*rq
= this_rq();
1902 finish_task_switch(rq
, prev
);
1905 * FIXME: do we need to worry about rq being invalidated by the
1910 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1911 /* In this case, finish_task_switch does not reenable preemption */
1914 if (current
->set_child_tid
)
1915 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1919 * context_switch - switch to the new MM and the new
1920 * thread's register state.
1923 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1924 struct task_struct
*next
)
1926 struct mm_struct
*mm
, *oldmm
;
1928 prepare_task_switch(rq
, prev
, next
);
1931 oldmm
= prev
->active_mm
;
1933 * For paravirt, this is coupled with an exit in switch_to to
1934 * combine the page table reload and the switch backend into
1937 arch_start_context_switch(prev
);
1940 next
->active_mm
= oldmm
;
1941 atomic_inc(&oldmm
->mm_count
);
1942 enter_lazy_tlb(oldmm
, next
);
1944 switch_mm(oldmm
, mm
, next
);
1947 prev
->active_mm
= NULL
;
1948 rq
->prev_mm
= oldmm
;
1951 * Since the runqueue lock will be released by the next
1952 * task (which is an invalid locking op but in the case
1953 * of the scheduler it's an obvious special-case), so we
1954 * do an early lockdep release here:
1956 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1957 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1960 context_tracking_task_switch(prev
, next
);
1961 /* Here we just switch the register state and the stack. */
1962 switch_to(prev
, next
, prev
);
1966 * this_rq must be evaluated again because prev may have moved
1967 * CPUs since it called schedule(), thus the 'rq' on its stack
1968 * frame will be invalid.
1970 finish_task_switch(this_rq(), prev
);
1974 * nr_running and nr_context_switches:
1976 * externally visible scheduler statistics: current number of runnable
1977 * threads, total number of context switches performed since bootup.
1979 unsigned long nr_running(void)
1981 unsigned long i
, sum
= 0;
1983 for_each_online_cpu(i
)
1984 sum
+= cpu_rq(i
)->nr_running
;
1989 unsigned long long nr_context_switches(void)
1992 unsigned long long sum
= 0;
1994 for_each_possible_cpu(i
)
1995 sum
+= cpu_rq(i
)->nr_switches
;
2000 unsigned long nr_iowait(void)
2002 unsigned long i
, sum
= 0;
2004 for_each_possible_cpu(i
)
2005 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2010 unsigned long nr_iowait_cpu(int cpu
)
2012 struct rq
*this = cpu_rq(cpu
);
2013 return atomic_read(&this->nr_iowait
);
2016 unsigned long this_cpu_load(void)
2018 struct rq
*this = this_rq();
2019 return this->cpu_load
[0];
2024 * Global load-average calculations
2026 * We take a distributed and async approach to calculating the global load-avg
2027 * in order to minimize overhead.
2029 * The global load average is an exponentially decaying average of nr_running +
2030 * nr_uninterruptible.
2032 * Once every LOAD_FREQ:
2035 * for_each_possible_cpu(cpu)
2036 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2038 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2040 * Due to a number of reasons the above turns in the mess below:
2042 * - for_each_possible_cpu() is prohibitively expensive on machines with
2043 * serious number of cpus, therefore we need to take a distributed approach
2044 * to calculating nr_active.
2046 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2047 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2049 * So assuming nr_active := 0 when we start out -- true per definition, we
2050 * can simply take per-cpu deltas and fold those into a global accumulate
2051 * to obtain the same result. See calc_load_fold_active().
2053 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2054 * across the machine, we assume 10 ticks is sufficient time for every
2055 * cpu to have completed this task.
2057 * This places an upper-bound on the IRQ-off latency of the machine. Then
2058 * again, being late doesn't loose the delta, just wrecks the sample.
2060 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2061 * this would add another cross-cpu cacheline miss and atomic operation
2062 * to the wakeup path. Instead we increment on whatever cpu the task ran
2063 * when it went into uninterruptible state and decrement on whatever cpu
2064 * did the wakeup. This means that only the sum of nr_uninterruptible over
2065 * all cpus yields the correct result.
2067 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2070 /* Variables and functions for calc_load */
2071 static atomic_long_t calc_load_tasks
;
2072 static unsigned long calc_load_update
;
2073 unsigned long avenrun
[3];
2074 EXPORT_SYMBOL(avenrun
); /* should be removed */
2077 * get_avenrun - get the load average array
2078 * @loads: pointer to dest load array
2079 * @offset: offset to add
2080 * @shift: shift count to shift the result left
2082 * These values are estimates at best, so no need for locking.
2084 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2086 loads
[0] = (avenrun
[0] + offset
) << shift
;
2087 loads
[1] = (avenrun
[1] + offset
) << shift
;
2088 loads
[2] = (avenrun
[2] + offset
) << shift
;
2091 static long calc_load_fold_active(struct rq
*this_rq
)
2093 long nr_active
, delta
= 0;
2095 nr_active
= this_rq
->nr_running
;
2096 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2098 if (nr_active
!= this_rq
->calc_load_active
) {
2099 delta
= nr_active
- this_rq
->calc_load_active
;
2100 this_rq
->calc_load_active
= nr_active
;
2107 * a1 = a0 * e + a * (1 - e)
2109 static unsigned long
2110 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2113 load
+= active
* (FIXED_1
- exp
);
2114 load
+= 1UL << (FSHIFT
- 1);
2115 return load
>> FSHIFT
;
2120 * Handle NO_HZ for the global load-average.
2122 * Since the above described distributed algorithm to compute the global
2123 * load-average relies on per-cpu sampling from the tick, it is affected by
2126 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2127 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2128 * when we read the global state.
2130 * Obviously reality has to ruin such a delightfully simple scheme:
2132 * - When we go NO_HZ idle during the window, we can negate our sample
2133 * contribution, causing under-accounting.
2135 * We avoid this by keeping two idle-delta counters and flipping them
2136 * when the window starts, thus separating old and new NO_HZ load.
2138 * The only trick is the slight shift in index flip for read vs write.
2142 * |-|-----------|-|-----------|-|-----------|-|
2143 * r:0 0 1 1 0 0 1 1 0
2144 * w:0 1 1 0 0 1 1 0 0
2146 * This ensures we'll fold the old idle contribution in this window while
2147 * accumlating the new one.
2149 * - When we wake up from NO_HZ idle during the window, we push up our
2150 * contribution, since we effectively move our sample point to a known
2153 * This is solved by pushing the window forward, and thus skipping the
2154 * sample, for this cpu (effectively using the idle-delta for this cpu which
2155 * was in effect at the time the window opened). This also solves the issue
2156 * of having to deal with a cpu having been in NOHZ idle for multiple
2157 * LOAD_FREQ intervals.
2159 * When making the ILB scale, we should try to pull this in as well.
2161 static atomic_long_t calc_load_idle
[2];
2162 static int calc_load_idx
;
2164 static inline int calc_load_write_idx(void)
2166 int idx
= calc_load_idx
;
2169 * See calc_global_nohz(), if we observe the new index, we also
2170 * need to observe the new update time.
2175 * If the folding window started, make sure we start writing in the
2178 if (!time_before(jiffies
, calc_load_update
))
2184 static inline int calc_load_read_idx(void)
2186 return calc_load_idx
& 1;
2189 void calc_load_enter_idle(void)
2191 struct rq
*this_rq
= this_rq();
2195 * We're going into NOHZ mode, if there's any pending delta, fold it
2196 * into the pending idle delta.
2198 delta
= calc_load_fold_active(this_rq
);
2200 int idx
= calc_load_write_idx();
2201 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2205 void calc_load_exit_idle(void)
2207 struct rq
*this_rq
= this_rq();
2210 * If we're still before the sample window, we're done.
2212 if (time_before(jiffies
, this_rq
->calc_load_update
))
2216 * We woke inside or after the sample window, this means we're already
2217 * accounted through the nohz accounting, so skip the entire deal and
2218 * sync up for the next window.
2220 this_rq
->calc_load_update
= calc_load_update
;
2221 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2222 this_rq
->calc_load_update
+= LOAD_FREQ
;
2225 static long calc_load_fold_idle(void)
2227 int idx
= calc_load_read_idx();
2230 if (atomic_long_read(&calc_load_idle
[idx
]))
2231 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2237 * fixed_power_int - compute: x^n, in O(log n) time
2239 * @x: base of the power
2240 * @frac_bits: fractional bits of @x
2241 * @n: power to raise @x to.
2243 * By exploiting the relation between the definition of the natural power
2244 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2245 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2246 * (where: n_i \elem {0, 1}, the binary vector representing n),
2247 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2248 * of course trivially computable in O(log_2 n), the length of our binary
2251 static unsigned long
2252 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2254 unsigned long result
= 1UL << frac_bits
;
2259 result
+= 1UL << (frac_bits
- 1);
2260 result
>>= frac_bits
;
2266 x
+= 1UL << (frac_bits
- 1);
2274 * a1 = a0 * e + a * (1 - e)
2276 * a2 = a1 * e + a * (1 - e)
2277 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2278 * = a0 * e^2 + a * (1 - e) * (1 + e)
2280 * a3 = a2 * e + a * (1 - e)
2281 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2282 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2286 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2287 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2288 * = a0 * e^n + a * (1 - e^n)
2290 * [1] application of the geometric series:
2293 * S_n := \Sum x^i = -------------
2296 static unsigned long
2297 calc_load_n(unsigned long load
, unsigned long exp
,
2298 unsigned long active
, unsigned int n
)
2301 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2305 * NO_HZ can leave us missing all per-cpu ticks calling
2306 * calc_load_account_active(), but since an idle CPU folds its delta into
2307 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2308 * in the pending idle delta if our idle period crossed a load cycle boundary.
2310 * Once we've updated the global active value, we need to apply the exponential
2311 * weights adjusted to the number of cycles missed.
2313 static void calc_global_nohz(void)
2315 long delta
, active
, n
;
2317 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2319 * Catch-up, fold however many we are behind still
2321 delta
= jiffies
- calc_load_update
- 10;
2322 n
= 1 + (delta
/ LOAD_FREQ
);
2324 active
= atomic_long_read(&calc_load_tasks
);
2325 active
= active
> 0 ? active
* FIXED_1
: 0;
2327 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2328 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2329 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2331 calc_load_update
+= n
* LOAD_FREQ
;
2335 * Flip the idle index...
2337 * Make sure we first write the new time then flip the index, so that
2338 * calc_load_write_idx() will see the new time when it reads the new
2339 * index, this avoids a double flip messing things up.
2344 #else /* !CONFIG_NO_HZ */
2346 static inline long calc_load_fold_idle(void) { return 0; }
2347 static inline void calc_global_nohz(void) { }
2349 #endif /* CONFIG_NO_HZ */
2352 * calc_load - update the avenrun load estimates 10 ticks after the
2353 * CPUs have updated calc_load_tasks.
2355 void calc_global_load(unsigned long ticks
)
2359 if (time_before(jiffies
, calc_load_update
+ 10))
2363 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2365 delta
= calc_load_fold_idle();
2367 atomic_long_add(delta
, &calc_load_tasks
);
2369 active
= atomic_long_read(&calc_load_tasks
);
2370 active
= active
> 0 ? active
* FIXED_1
: 0;
2372 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2373 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2374 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2376 calc_load_update
+= LOAD_FREQ
;
2379 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2385 * Called from update_cpu_load() to periodically update this CPU's
2388 static void calc_load_account_active(struct rq
*this_rq
)
2392 if (time_before(jiffies
, this_rq
->calc_load_update
))
2395 delta
= calc_load_fold_active(this_rq
);
2397 atomic_long_add(delta
, &calc_load_tasks
);
2399 this_rq
->calc_load_update
+= LOAD_FREQ
;
2403 * End of global load-average stuff
2407 * The exact cpuload at various idx values, calculated at every tick would be
2408 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2410 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2411 * on nth tick when cpu may be busy, then we have:
2412 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2413 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2415 * decay_load_missed() below does efficient calculation of
2416 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2417 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2419 * The calculation is approximated on a 128 point scale.
2420 * degrade_zero_ticks is the number of ticks after which load at any
2421 * particular idx is approximated to be zero.
2422 * degrade_factor is a precomputed table, a row for each load idx.
2423 * Each column corresponds to degradation factor for a power of two ticks,
2424 * based on 128 point scale.
2426 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2427 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2429 * With this power of 2 load factors, we can degrade the load n times
2430 * by looking at 1 bits in n and doing as many mult/shift instead of
2431 * n mult/shifts needed by the exact degradation.
2433 #define DEGRADE_SHIFT 7
2434 static const unsigned char
2435 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2436 static const unsigned char
2437 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2438 {0, 0, 0, 0, 0, 0, 0, 0},
2439 {64, 32, 8, 0, 0, 0, 0, 0},
2440 {96, 72, 40, 12, 1, 0, 0},
2441 {112, 98, 75, 43, 15, 1, 0},
2442 {120, 112, 98, 76, 45, 16, 2} };
2445 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2446 * would be when CPU is idle and so we just decay the old load without
2447 * adding any new load.
2449 static unsigned long
2450 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2454 if (!missed_updates
)
2457 if (missed_updates
>= degrade_zero_ticks
[idx
])
2461 return load
>> missed_updates
;
2463 while (missed_updates
) {
2464 if (missed_updates
% 2)
2465 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2467 missed_updates
>>= 1;
2474 * Update rq->cpu_load[] statistics. This function is usually called every
2475 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2476 * every tick. We fix it up based on jiffies.
2478 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2479 unsigned long pending_updates
)
2483 this_rq
->nr_load_updates
++;
2485 /* Update our load: */
2486 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2487 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2488 unsigned long old_load
, new_load
;
2490 /* scale is effectively 1 << i now, and >> i divides by scale */
2492 old_load
= this_rq
->cpu_load
[i
];
2493 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2494 new_load
= this_load
;
2496 * Round up the averaging division if load is increasing. This
2497 * prevents us from getting stuck on 9 if the load is 10, for
2500 if (new_load
> old_load
)
2501 new_load
+= scale
- 1;
2503 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2506 sched_avg_update(this_rq
);
2511 * There is no sane way to deal with nohz on smp when using jiffies because the
2512 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2513 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2515 * Therefore we cannot use the delta approach from the regular tick since that
2516 * would seriously skew the load calculation. However we'll make do for those
2517 * updates happening while idle (nohz_idle_balance) or coming out of idle
2518 * (tick_nohz_idle_exit).
2520 * This means we might still be one tick off for nohz periods.
2524 * Called from nohz_idle_balance() to update the load ratings before doing the
2527 void update_idle_cpu_load(struct rq
*this_rq
)
2529 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2530 unsigned long load
= this_rq
->load
.weight
;
2531 unsigned long pending_updates
;
2534 * bail if there's load or we're actually up-to-date.
2536 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2539 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2540 this_rq
->last_load_update_tick
= curr_jiffies
;
2542 __update_cpu_load(this_rq
, load
, pending_updates
);
2546 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2548 void update_cpu_load_nohz(void)
2550 struct rq
*this_rq
= this_rq();
2551 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2552 unsigned long pending_updates
;
2554 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2557 raw_spin_lock(&this_rq
->lock
);
2558 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2559 if (pending_updates
) {
2560 this_rq
->last_load_update_tick
= curr_jiffies
;
2562 * We were idle, this means load 0, the current load might be
2563 * !0 due to remote wakeups and the sort.
2565 __update_cpu_load(this_rq
, 0, pending_updates
);
2567 raw_spin_unlock(&this_rq
->lock
);
2569 #endif /* CONFIG_NO_HZ */
2572 * Called from scheduler_tick()
2574 static void update_cpu_load_active(struct rq
*this_rq
)
2577 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2579 this_rq
->last_load_update_tick
= jiffies
;
2580 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2582 calc_load_account_active(this_rq
);
2588 * sched_exec - execve() is a valuable balancing opportunity, because at
2589 * this point the task has the smallest effective memory and cache footprint.
2591 void sched_exec(void)
2593 struct task_struct
*p
= current
;
2594 unsigned long flags
;
2597 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2598 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2599 if (dest_cpu
== smp_processor_id())
2602 if (likely(cpu_active(dest_cpu
))) {
2603 struct migration_arg arg
= { p
, dest_cpu
};
2605 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2606 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2610 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2615 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2616 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2618 EXPORT_PER_CPU_SYMBOL(kstat
);
2619 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2622 * Return any ns on the sched_clock that have not yet been accounted in
2623 * @p in case that task is currently running.
2625 * Called with task_rq_lock() held on @rq.
2627 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2631 if (task_current(rq
, p
)) {
2632 update_rq_clock(rq
);
2633 ns
= rq
->clock_task
- p
->se
.exec_start
;
2641 unsigned long long task_delta_exec(struct task_struct
*p
)
2643 unsigned long flags
;
2647 rq
= task_rq_lock(p
, &flags
);
2648 ns
= do_task_delta_exec(p
, rq
);
2649 task_rq_unlock(rq
, p
, &flags
);
2655 * Return accounted runtime for the task.
2656 * In case the task is currently running, return the runtime plus current's
2657 * pending runtime that have not been accounted yet.
2659 unsigned long long task_sched_runtime(struct task_struct
*p
)
2661 unsigned long flags
;
2665 rq
= task_rq_lock(p
, &flags
);
2666 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2667 task_rq_unlock(rq
, p
, &flags
);
2673 * This function gets called by the timer code, with HZ frequency.
2674 * We call it with interrupts disabled.
2676 void scheduler_tick(void)
2678 int cpu
= smp_processor_id();
2679 struct rq
*rq
= cpu_rq(cpu
);
2680 struct task_struct
*curr
= rq
->curr
;
2684 raw_spin_lock(&rq
->lock
);
2685 update_rq_clock(rq
);
2686 update_cpu_load_active(rq
);
2687 curr
->sched_class
->task_tick(rq
, curr
, 0);
2688 raw_spin_unlock(&rq
->lock
);
2690 perf_event_task_tick();
2693 rq
->idle_balance
= idle_cpu(cpu
);
2694 trigger_load_balance(rq
, cpu
);
2698 notrace
unsigned long get_parent_ip(unsigned long addr
)
2700 if (in_lock_functions(addr
)) {
2701 addr
= CALLER_ADDR2
;
2702 if (in_lock_functions(addr
))
2703 addr
= CALLER_ADDR3
;
2708 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2709 defined(CONFIG_PREEMPT_TRACER))
2711 void __kprobes
add_preempt_count(int val
)
2713 #ifdef CONFIG_DEBUG_PREEMPT
2717 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2720 preempt_count() += val
;
2721 #ifdef CONFIG_DEBUG_PREEMPT
2723 * Spinlock count overflowing soon?
2725 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2728 if (preempt_count() == val
)
2729 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2731 EXPORT_SYMBOL(add_preempt_count
);
2733 void __kprobes
sub_preempt_count(int val
)
2735 #ifdef CONFIG_DEBUG_PREEMPT
2739 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2742 * Is the spinlock portion underflowing?
2744 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2745 !(preempt_count() & PREEMPT_MASK
)))
2749 if (preempt_count() == val
)
2750 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2751 preempt_count() -= val
;
2753 EXPORT_SYMBOL(sub_preempt_count
);
2758 * Print scheduling while atomic bug:
2760 static noinline
void __schedule_bug(struct task_struct
*prev
)
2762 if (oops_in_progress
)
2765 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2766 prev
->comm
, prev
->pid
, preempt_count());
2768 debug_show_held_locks(prev
);
2770 if (irqs_disabled())
2771 print_irqtrace_events(prev
);
2773 add_taint(TAINT_WARN
);
2777 * Various schedule()-time debugging checks and statistics:
2779 static inline void schedule_debug(struct task_struct
*prev
)
2782 * Test if we are atomic. Since do_exit() needs to call into
2783 * schedule() atomically, we ignore that path for now.
2784 * Otherwise, whine if we are scheduling when we should not be.
2786 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2787 __schedule_bug(prev
);
2790 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2792 schedstat_inc(this_rq(), sched_count
);
2795 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2797 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2798 update_rq_clock(rq
);
2799 prev
->sched_class
->put_prev_task(rq
, prev
);
2803 * Pick up the highest-prio task:
2805 static inline struct task_struct
*
2806 pick_next_task(struct rq
*rq
)
2808 const struct sched_class
*class;
2809 struct task_struct
*p
;
2812 * Optimization: we know that if all tasks are in
2813 * the fair class we can call that function directly:
2815 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2816 p
= fair_sched_class
.pick_next_task(rq
);
2821 for_each_class(class) {
2822 p
= class->pick_next_task(rq
);
2827 BUG(); /* the idle class will always have a runnable task */
2831 * __schedule() is the main scheduler function.
2833 * The main means of driving the scheduler and thus entering this function are:
2835 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2837 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2838 * paths. For example, see arch/x86/entry_64.S.
2840 * To drive preemption between tasks, the scheduler sets the flag in timer
2841 * interrupt handler scheduler_tick().
2843 * 3. Wakeups don't really cause entry into schedule(). They add a
2844 * task to the run-queue and that's it.
2846 * Now, if the new task added to the run-queue preempts the current
2847 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2848 * called on the nearest possible occasion:
2850 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2852 * - in syscall or exception context, at the next outmost
2853 * preempt_enable(). (this might be as soon as the wake_up()'s
2856 * - in IRQ context, return from interrupt-handler to
2857 * preemptible context
2859 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2862 * - cond_resched() call
2863 * - explicit schedule() call
2864 * - return from syscall or exception to user-space
2865 * - return from interrupt-handler to user-space
2867 static void __sched
__schedule(void)
2869 struct task_struct
*prev
, *next
;
2870 unsigned long *switch_count
;
2876 cpu
= smp_processor_id();
2878 rcu_note_context_switch(cpu
);
2881 schedule_debug(prev
);
2883 if (sched_feat(HRTICK
))
2886 raw_spin_lock_irq(&rq
->lock
);
2888 switch_count
= &prev
->nivcsw
;
2889 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2890 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2891 prev
->state
= TASK_RUNNING
;
2893 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2897 * If a worker went to sleep, notify and ask workqueue
2898 * whether it wants to wake up a task to maintain
2901 if (prev
->flags
& PF_WQ_WORKER
) {
2902 struct task_struct
*to_wakeup
;
2904 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2906 try_to_wake_up_local(to_wakeup
);
2909 switch_count
= &prev
->nvcsw
;
2912 pre_schedule(rq
, prev
);
2914 if (unlikely(!rq
->nr_running
))
2915 idle_balance(cpu
, rq
);
2917 put_prev_task(rq
, prev
);
2918 next
= pick_next_task(rq
);
2919 clear_tsk_need_resched(prev
);
2920 rq
->skip_clock_update
= 0;
2922 if (likely(prev
!= next
)) {
2927 context_switch(rq
, prev
, next
); /* unlocks the rq */
2929 * The context switch have flipped the stack from under us
2930 * and restored the local variables which were saved when
2931 * this task called schedule() in the past. prev == current
2932 * is still correct, but it can be moved to another cpu/rq.
2934 cpu
= smp_processor_id();
2937 raw_spin_unlock_irq(&rq
->lock
);
2941 sched_preempt_enable_no_resched();
2946 static inline void sched_submit_work(struct task_struct
*tsk
)
2948 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2951 * If we are going to sleep and we have plugged IO queued,
2952 * make sure to submit it to avoid deadlocks.
2954 if (blk_needs_flush_plug(tsk
))
2955 blk_schedule_flush_plug(tsk
);
2958 asmlinkage
void __sched
schedule(void)
2960 struct task_struct
*tsk
= current
;
2962 sched_submit_work(tsk
);
2965 EXPORT_SYMBOL(schedule
);
2967 #ifdef CONFIG_CONTEXT_TRACKING
2968 asmlinkage
void __sched
schedule_user(void)
2971 * If we come here after a random call to set_need_resched(),
2972 * or we have been woken up remotely but the IPI has not yet arrived,
2973 * we haven't yet exited the RCU idle mode. Do it here manually until
2974 * we find a better solution.
2983 * schedule_preempt_disabled - called with preemption disabled
2985 * Returns with preemption disabled. Note: preempt_count must be 1
2987 void __sched
schedule_preempt_disabled(void)
2989 sched_preempt_enable_no_resched();
2994 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2996 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
2998 if (lock
->owner
!= owner
)
3002 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3003 * lock->owner still matches owner, if that fails, owner might
3004 * point to free()d memory, if it still matches, the rcu_read_lock()
3005 * ensures the memory stays valid.
3009 return owner
->on_cpu
;
3013 * Look out! "owner" is an entirely speculative pointer
3014 * access and not reliable.
3016 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3018 if (!sched_feat(OWNER_SPIN
))
3022 while (owner_running(lock
, owner
)) {
3026 arch_mutex_cpu_relax();
3031 * We break out the loop above on need_resched() and when the
3032 * owner changed, which is a sign for heavy contention. Return
3033 * success only when lock->owner is NULL.
3035 return lock
->owner
== NULL
;
3039 #ifdef CONFIG_PREEMPT
3041 * this is the entry point to schedule() from in-kernel preemption
3042 * off of preempt_enable. Kernel preemptions off return from interrupt
3043 * occur there and call schedule directly.
3045 asmlinkage
void __sched notrace
preempt_schedule(void)
3047 struct thread_info
*ti
= current_thread_info();
3050 * If there is a non-zero preempt_count or interrupts are disabled,
3051 * we do not want to preempt the current task. Just return..
3053 if (likely(ti
->preempt_count
|| irqs_disabled()))
3057 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3059 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3062 * Check again in case we missed a preemption opportunity
3063 * between schedule and now.
3066 } while (need_resched());
3068 EXPORT_SYMBOL(preempt_schedule
);
3071 * this is the entry point to schedule() from kernel preemption
3072 * off of irq context.
3073 * Note, that this is called and return with irqs disabled. This will
3074 * protect us against recursive calling from irq.
3076 asmlinkage
void __sched
preempt_schedule_irq(void)
3078 struct thread_info
*ti
= current_thread_info();
3080 /* Catch callers which need to be fixed */
3081 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3085 add_preempt_count(PREEMPT_ACTIVE
);
3088 local_irq_disable();
3089 sub_preempt_count(PREEMPT_ACTIVE
);
3092 * Check again in case we missed a preemption opportunity
3093 * between schedule and now.
3096 } while (need_resched());
3099 #endif /* CONFIG_PREEMPT */
3101 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3104 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3106 EXPORT_SYMBOL(default_wake_function
);
3109 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3110 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3111 * number) then we wake all the non-exclusive tasks and one exclusive task.
3113 * There are circumstances in which we can try to wake a task which has already
3114 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3115 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3117 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3118 int nr_exclusive
, int wake_flags
, void *key
)
3120 wait_queue_t
*curr
, *next
;
3122 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3123 unsigned flags
= curr
->flags
;
3125 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3126 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3132 * __wake_up - wake up threads blocked on a waitqueue.
3134 * @mode: which threads
3135 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3136 * @key: is directly passed to the wakeup function
3138 * It may be assumed that this function implies a write memory barrier before
3139 * changing the task state if and only if any tasks are woken up.
3141 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3142 int nr_exclusive
, void *key
)
3144 unsigned long flags
;
3146 spin_lock_irqsave(&q
->lock
, flags
);
3147 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3148 spin_unlock_irqrestore(&q
->lock
, flags
);
3150 EXPORT_SYMBOL(__wake_up
);
3153 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3155 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3157 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3159 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3161 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3163 __wake_up_common(q
, mode
, 1, 0, key
);
3165 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3168 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3170 * @mode: which threads
3171 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3172 * @key: opaque value to be passed to wakeup targets
3174 * The sync wakeup differs that the waker knows that it will schedule
3175 * away soon, so while the target thread will be woken up, it will not
3176 * be migrated to another CPU - ie. the two threads are 'synchronized'
3177 * with each other. This can prevent needless bouncing between CPUs.
3179 * On UP it can prevent extra preemption.
3181 * It may be assumed that this function implies a write memory barrier before
3182 * changing the task state if and only if any tasks are woken up.
3184 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3185 int nr_exclusive
, void *key
)
3187 unsigned long flags
;
3188 int wake_flags
= WF_SYNC
;
3193 if (unlikely(!nr_exclusive
))
3196 spin_lock_irqsave(&q
->lock
, flags
);
3197 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3198 spin_unlock_irqrestore(&q
->lock
, flags
);
3200 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3203 * __wake_up_sync - see __wake_up_sync_key()
3205 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3207 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3209 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3212 * complete: - signals a single thread waiting on this completion
3213 * @x: holds the state of this particular completion
3215 * This will wake up a single thread waiting on this completion. Threads will be
3216 * awakened in the same order in which they were queued.
3218 * See also complete_all(), wait_for_completion() and related routines.
3220 * It may be assumed that this function implies a write memory barrier before
3221 * changing the task state if and only if any tasks are woken up.
3223 void complete(struct completion
*x
)
3225 unsigned long flags
;
3227 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3229 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3230 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3232 EXPORT_SYMBOL(complete
);
3235 * complete_all: - signals all threads waiting on this completion
3236 * @x: holds the state of this particular completion
3238 * This will wake up all threads waiting on this particular completion event.
3240 * It may be assumed that this function implies a write memory barrier before
3241 * changing the task state if and only if any tasks are woken up.
3243 void complete_all(struct completion
*x
)
3245 unsigned long flags
;
3247 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3248 x
->done
+= UINT_MAX
/2;
3249 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3250 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3252 EXPORT_SYMBOL(complete_all
);
3254 static inline long __sched
3255 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3258 DECLARE_WAITQUEUE(wait
, current
);
3260 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3262 if (signal_pending_state(state
, current
)) {
3263 timeout
= -ERESTARTSYS
;
3266 __set_current_state(state
);
3267 spin_unlock_irq(&x
->wait
.lock
);
3268 timeout
= schedule_timeout(timeout
);
3269 spin_lock_irq(&x
->wait
.lock
);
3270 } while (!x
->done
&& timeout
);
3271 __remove_wait_queue(&x
->wait
, &wait
);
3276 return timeout
?: 1;
3280 wait_for_common(struct completion
*x
, long timeout
, int state
)
3284 spin_lock_irq(&x
->wait
.lock
);
3285 timeout
= do_wait_for_common(x
, timeout
, state
);
3286 spin_unlock_irq(&x
->wait
.lock
);
3291 * wait_for_completion: - waits for completion of a task
3292 * @x: holds the state of this particular completion
3294 * This waits to be signaled for completion of a specific task. It is NOT
3295 * interruptible and there is no timeout.
3297 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3298 * and interrupt capability. Also see complete().
3300 void __sched
wait_for_completion(struct completion
*x
)
3302 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3304 EXPORT_SYMBOL(wait_for_completion
);
3307 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3308 * @x: holds the state of this particular completion
3309 * @timeout: timeout value in jiffies
3311 * This waits for either a completion of a specific task to be signaled or for a
3312 * specified timeout to expire. The timeout is in jiffies. It is not
3315 * The return value is 0 if timed out, and positive (at least 1, or number of
3316 * jiffies left till timeout) if completed.
3318 unsigned long __sched
3319 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3321 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3323 EXPORT_SYMBOL(wait_for_completion_timeout
);
3326 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3327 * @x: holds the state of this particular completion
3329 * This waits for completion of a specific task to be signaled. It is
3332 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3334 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3336 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3337 if (t
== -ERESTARTSYS
)
3341 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3344 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3345 * @x: holds the state of this particular completion
3346 * @timeout: timeout value in jiffies
3348 * This waits for either a completion of a specific task to be signaled or for a
3349 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3351 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3352 * positive (at least 1, or number of jiffies left till timeout) if completed.
3355 wait_for_completion_interruptible_timeout(struct completion
*x
,
3356 unsigned long timeout
)
3358 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3360 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3363 * wait_for_completion_killable: - waits for completion of a task (killable)
3364 * @x: holds the state of this particular completion
3366 * This waits to be signaled for completion of a specific task. It can be
3367 * interrupted by a kill signal.
3369 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3371 int __sched
wait_for_completion_killable(struct completion
*x
)
3373 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3374 if (t
== -ERESTARTSYS
)
3378 EXPORT_SYMBOL(wait_for_completion_killable
);
3381 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3382 * @x: holds the state of this particular completion
3383 * @timeout: timeout value in jiffies
3385 * This waits for either a completion of a specific task to be
3386 * signaled or for a specified timeout to expire. It can be
3387 * interrupted by a kill signal. The timeout is in jiffies.
3389 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3390 * positive (at least 1, or number of jiffies left till timeout) if completed.
3393 wait_for_completion_killable_timeout(struct completion
*x
,
3394 unsigned long timeout
)
3396 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3398 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3401 * try_wait_for_completion - try to decrement a completion without blocking
3402 * @x: completion structure
3404 * Returns: 0 if a decrement cannot be done without blocking
3405 * 1 if a decrement succeeded.
3407 * If a completion is being used as a counting completion,
3408 * attempt to decrement the counter without blocking. This
3409 * enables us to avoid waiting if the resource the completion
3410 * is protecting is not available.
3412 bool try_wait_for_completion(struct completion
*x
)
3414 unsigned long flags
;
3417 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3422 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3425 EXPORT_SYMBOL(try_wait_for_completion
);
3428 * completion_done - Test to see if a completion has any waiters
3429 * @x: completion structure
3431 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3432 * 1 if there are no waiters.
3435 bool completion_done(struct completion
*x
)
3437 unsigned long flags
;
3440 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3443 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3446 EXPORT_SYMBOL(completion_done
);
3449 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3451 unsigned long flags
;
3454 init_waitqueue_entry(&wait
, current
);
3456 __set_current_state(state
);
3458 spin_lock_irqsave(&q
->lock
, flags
);
3459 __add_wait_queue(q
, &wait
);
3460 spin_unlock(&q
->lock
);
3461 timeout
= schedule_timeout(timeout
);
3462 spin_lock_irq(&q
->lock
);
3463 __remove_wait_queue(q
, &wait
);
3464 spin_unlock_irqrestore(&q
->lock
, flags
);
3469 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3471 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3473 EXPORT_SYMBOL(interruptible_sleep_on
);
3476 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3478 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3480 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3482 void __sched
sleep_on(wait_queue_head_t
*q
)
3484 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3486 EXPORT_SYMBOL(sleep_on
);
3488 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3490 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3492 EXPORT_SYMBOL(sleep_on_timeout
);
3494 #ifdef CONFIG_RT_MUTEXES
3497 * rt_mutex_setprio - set the current priority of a task
3499 * @prio: prio value (kernel-internal form)
3501 * This function changes the 'effective' priority of a task. It does
3502 * not touch ->normal_prio like __setscheduler().
3504 * Used by the rt_mutex code to implement priority inheritance logic.
3506 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3508 int oldprio
, on_rq
, running
;
3510 const struct sched_class
*prev_class
;
3512 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3514 rq
= __task_rq_lock(p
);
3517 * Idle task boosting is a nono in general. There is one
3518 * exception, when PREEMPT_RT and NOHZ is active:
3520 * The idle task calls get_next_timer_interrupt() and holds
3521 * the timer wheel base->lock on the CPU and another CPU wants
3522 * to access the timer (probably to cancel it). We can safely
3523 * ignore the boosting request, as the idle CPU runs this code
3524 * with interrupts disabled and will complete the lock
3525 * protected section without being interrupted. So there is no
3526 * real need to boost.
3528 if (unlikely(p
== rq
->idle
)) {
3529 WARN_ON(p
!= rq
->curr
);
3530 WARN_ON(p
->pi_blocked_on
);
3534 trace_sched_pi_setprio(p
, prio
);
3536 prev_class
= p
->sched_class
;
3538 running
= task_current(rq
, p
);
3540 dequeue_task(rq
, p
, 0);
3542 p
->sched_class
->put_prev_task(rq
, p
);
3545 p
->sched_class
= &rt_sched_class
;
3547 p
->sched_class
= &fair_sched_class
;
3552 p
->sched_class
->set_curr_task(rq
);
3554 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3556 check_class_changed(rq
, p
, prev_class
, oldprio
);
3558 __task_rq_unlock(rq
);
3561 void set_user_nice(struct task_struct
*p
, long nice
)
3563 int old_prio
, delta
, on_rq
;
3564 unsigned long flags
;
3567 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3570 * We have to be careful, if called from sys_setpriority(),
3571 * the task might be in the middle of scheduling on another CPU.
3573 rq
= task_rq_lock(p
, &flags
);
3575 * The RT priorities are set via sched_setscheduler(), but we still
3576 * allow the 'normal' nice value to be set - but as expected
3577 * it wont have any effect on scheduling until the task is
3578 * SCHED_FIFO/SCHED_RR:
3580 if (task_has_rt_policy(p
)) {
3581 p
->static_prio
= NICE_TO_PRIO(nice
);
3586 dequeue_task(rq
, p
, 0);
3588 p
->static_prio
= NICE_TO_PRIO(nice
);
3591 p
->prio
= effective_prio(p
);
3592 delta
= p
->prio
- old_prio
;
3595 enqueue_task(rq
, p
, 0);
3597 * If the task increased its priority or is running and
3598 * lowered its priority, then reschedule its CPU:
3600 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3601 resched_task(rq
->curr
);
3604 task_rq_unlock(rq
, p
, &flags
);
3606 EXPORT_SYMBOL(set_user_nice
);
3609 * can_nice - check if a task can reduce its nice value
3613 int can_nice(const struct task_struct
*p
, const int nice
)
3615 /* convert nice value [19,-20] to rlimit style value [1,40] */
3616 int nice_rlim
= 20 - nice
;
3618 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3619 capable(CAP_SYS_NICE
));
3622 #ifdef __ARCH_WANT_SYS_NICE
3625 * sys_nice - change the priority of the current process.
3626 * @increment: priority increment
3628 * sys_setpriority is a more generic, but much slower function that
3629 * does similar things.
3631 SYSCALL_DEFINE1(nice
, int, increment
)
3636 * Setpriority might change our priority at the same moment.
3637 * We don't have to worry. Conceptually one call occurs first
3638 * and we have a single winner.
3640 if (increment
< -40)
3645 nice
= TASK_NICE(current
) + increment
;
3651 if (increment
< 0 && !can_nice(current
, nice
))
3654 retval
= security_task_setnice(current
, nice
);
3658 set_user_nice(current
, nice
);
3665 * task_prio - return the priority value of a given task.
3666 * @p: the task in question.
3668 * This is the priority value as seen by users in /proc.
3669 * RT tasks are offset by -200. Normal tasks are centered
3670 * around 0, value goes from -16 to +15.
3672 int task_prio(const struct task_struct
*p
)
3674 return p
->prio
- MAX_RT_PRIO
;
3678 * task_nice - return the nice value of a given task.
3679 * @p: the task in question.
3681 int task_nice(const struct task_struct
*p
)
3683 return TASK_NICE(p
);
3685 EXPORT_SYMBOL(task_nice
);
3688 * idle_cpu - is a given cpu idle currently?
3689 * @cpu: the processor in question.
3691 int idle_cpu(int cpu
)
3693 struct rq
*rq
= cpu_rq(cpu
);
3695 if (rq
->curr
!= rq
->idle
)
3702 if (!llist_empty(&rq
->wake_list
))
3710 * idle_task - return the idle task for a given cpu.
3711 * @cpu: the processor in question.
3713 struct task_struct
*idle_task(int cpu
)
3715 return cpu_rq(cpu
)->idle
;
3719 * find_process_by_pid - find a process with a matching PID value.
3720 * @pid: the pid in question.
3722 static struct task_struct
*find_process_by_pid(pid_t pid
)
3724 return pid
? find_task_by_vpid(pid
) : current
;
3727 /* Actually do priority change: must hold rq lock. */
3729 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3732 p
->rt_priority
= prio
;
3733 p
->normal_prio
= normal_prio(p
);
3734 /* we are holding p->pi_lock already */
3735 p
->prio
= rt_mutex_getprio(p
);
3736 if (rt_prio(p
->prio
))
3737 p
->sched_class
= &rt_sched_class
;
3739 p
->sched_class
= &fair_sched_class
;
3744 * check the target process has a UID that matches the current process's
3746 static bool check_same_owner(struct task_struct
*p
)
3748 const struct cred
*cred
= current_cred(), *pcred
;
3752 pcred
= __task_cred(p
);
3753 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3754 uid_eq(cred
->euid
, pcred
->uid
));
3759 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3760 const struct sched_param
*param
, bool user
)
3762 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3763 unsigned long flags
;
3764 const struct sched_class
*prev_class
;
3768 /* may grab non-irq protected spin_locks */
3769 BUG_ON(in_interrupt());
3771 /* double check policy once rq lock held */
3773 reset_on_fork
= p
->sched_reset_on_fork
;
3774 policy
= oldpolicy
= p
->policy
;
3776 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3777 policy
&= ~SCHED_RESET_ON_FORK
;
3779 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3780 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3781 policy
!= SCHED_IDLE
)
3786 * Valid priorities for SCHED_FIFO and SCHED_RR are
3787 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3788 * SCHED_BATCH and SCHED_IDLE is 0.
3790 if (param
->sched_priority
< 0 ||
3791 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3792 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3794 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3798 * Allow unprivileged RT tasks to decrease priority:
3800 if (user
&& !capable(CAP_SYS_NICE
)) {
3801 if (rt_policy(policy
)) {
3802 unsigned long rlim_rtprio
=
3803 task_rlimit(p
, RLIMIT_RTPRIO
);
3805 /* can't set/change the rt policy */
3806 if (policy
!= p
->policy
&& !rlim_rtprio
)
3809 /* can't increase priority */
3810 if (param
->sched_priority
> p
->rt_priority
&&
3811 param
->sched_priority
> rlim_rtprio
)
3816 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3817 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3819 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3820 if (!can_nice(p
, TASK_NICE(p
)))
3824 /* can't change other user's priorities */
3825 if (!check_same_owner(p
))
3828 /* Normal users shall not reset the sched_reset_on_fork flag */
3829 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3834 retval
= security_task_setscheduler(p
);
3840 * make sure no PI-waiters arrive (or leave) while we are
3841 * changing the priority of the task:
3843 * To be able to change p->policy safely, the appropriate
3844 * runqueue lock must be held.
3846 rq
= task_rq_lock(p
, &flags
);
3849 * Changing the policy of the stop threads its a very bad idea
3851 if (p
== rq
->stop
) {
3852 task_rq_unlock(rq
, p
, &flags
);
3857 * If not changing anything there's no need to proceed further:
3859 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3860 param
->sched_priority
== p
->rt_priority
))) {
3861 task_rq_unlock(rq
, p
, &flags
);
3865 #ifdef CONFIG_RT_GROUP_SCHED
3868 * Do not allow realtime tasks into groups that have no runtime
3871 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3872 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3873 !task_group_is_autogroup(task_group(p
))) {
3874 task_rq_unlock(rq
, p
, &flags
);
3880 /* recheck policy now with rq lock held */
3881 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3882 policy
= oldpolicy
= -1;
3883 task_rq_unlock(rq
, p
, &flags
);
3887 running
= task_current(rq
, p
);
3889 dequeue_task(rq
, p
, 0);
3891 p
->sched_class
->put_prev_task(rq
, p
);
3893 p
->sched_reset_on_fork
= reset_on_fork
;
3896 prev_class
= p
->sched_class
;
3897 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3900 p
->sched_class
->set_curr_task(rq
);
3902 enqueue_task(rq
, p
, 0);
3904 check_class_changed(rq
, p
, prev_class
, oldprio
);
3905 task_rq_unlock(rq
, p
, &flags
);
3907 rt_mutex_adjust_pi(p
);
3913 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3914 * @p: the task in question.
3915 * @policy: new policy.
3916 * @param: structure containing the new RT priority.
3918 * NOTE that the task may be already dead.
3920 int sched_setscheduler(struct task_struct
*p
, int policy
,
3921 const struct sched_param
*param
)
3923 return __sched_setscheduler(p
, policy
, param
, true);
3925 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3928 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3929 * @p: the task in question.
3930 * @policy: new policy.
3931 * @param: structure containing the new RT priority.
3933 * Just like sched_setscheduler, only don't bother checking if the
3934 * current context has permission. For example, this is needed in
3935 * stop_machine(): we create temporary high priority worker threads,
3936 * but our caller might not have that capability.
3938 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3939 const struct sched_param
*param
)
3941 return __sched_setscheduler(p
, policy
, param
, false);
3945 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3947 struct sched_param lparam
;
3948 struct task_struct
*p
;
3951 if (!param
|| pid
< 0)
3953 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3958 p
= find_process_by_pid(pid
);
3960 retval
= sched_setscheduler(p
, policy
, &lparam
);
3967 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3968 * @pid: the pid in question.
3969 * @policy: new policy.
3970 * @param: structure containing the new RT priority.
3972 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3973 struct sched_param __user
*, param
)
3975 /* negative values for policy are not valid */
3979 return do_sched_setscheduler(pid
, policy
, param
);
3983 * sys_sched_setparam - set/change the RT priority of a thread
3984 * @pid: the pid in question.
3985 * @param: structure containing the new RT priority.
3987 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3989 return do_sched_setscheduler(pid
, -1, param
);
3993 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3994 * @pid: the pid in question.
3996 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3998 struct task_struct
*p
;
4006 p
= find_process_by_pid(pid
);
4008 retval
= security_task_getscheduler(p
);
4011 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4018 * sys_sched_getparam - get the RT priority of a thread
4019 * @pid: the pid in question.
4020 * @param: structure containing the RT priority.
4022 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4024 struct sched_param lp
;
4025 struct task_struct
*p
;
4028 if (!param
|| pid
< 0)
4032 p
= find_process_by_pid(pid
);
4037 retval
= security_task_getscheduler(p
);
4041 lp
.sched_priority
= p
->rt_priority
;
4045 * This one might sleep, we cannot do it with a spinlock held ...
4047 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4056 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4058 cpumask_var_t cpus_allowed
, new_mask
;
4059 struct task_struct
*p
;
4065 p
= find_process_by_pid(pid
);
4072 /* Prevent p going away */
4076 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4080 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4082 goto out_free_cpus_allowed
;
4085 if (!check_same_owner(p
)) {
4087 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4094 retval
= security_task_setscheduler(p
);
4098 cpuset_cpus_allowed(p
, cpus_allowed
);
4099 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4101 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4104 cpuset_cpus_allowed(p
, cpus_allowed
);
4105 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4107 * We must have raced with a concurrent cpuset
4108 * update. Just reset the cpus_allowed to the
4109 * cpuset's cpus_allowed
4111 cpumask_copy(new_mask
, cpus_allowed
);
4116 free_cpumask_var(new_mask
);
4117 out_free_cpus_allowed
:
4118 free_cpumask_var(cpus_allowed
);
4125 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4126 struct cpumask
*new_mask
)
4128 if (len
< cpumask_size())
4129 cpumask_clear(new_mask
);
4130 else if (len
> cpumask_size())
4131 len
= cpumask_size();
4133 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4137 * sys_sched_setaffinity - set the cpu affinity of a process
4138 * @pid: pid of the process
4139 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4140 * @user_mask_ptr: user-space pointer to the new cpu mask
4142 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4143 unsigned long __user
*, user_mask_ptr
)
4145 cpumask_var_t new_mask
;
4148 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4151 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4153 retval
= sched_setaffinity(pid
, new_mask
);
4154 free_cpumask_var(new_mask
);
4158 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4160 struct task_struct
*p
;
4161 unsigned long flags
;
4168 p
= find_process_by_pid(pid
);
4172 retval
= security_task_getscheduler(p
);
4176 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4177 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4178 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4188 * sys_sched_getaffinity - get the cpu affinity of a process
4189 * @pid: pid of the process
4190 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4191 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4193 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4194 unsigned long __user
*, user_mask_ptr
)
4199 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4201 if (len
& (sizeof(unsigned long)-1))
4204 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4207 ret
= sched_getaffinity(pid
, mask
);
4209 size_t retlen
= min_t(size_t, len
, cpumask_size());
4211 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4216 free_cpumask_var(mask
);
4222 * sys_sched_yield - yield the current processor to other threads.
4224 * This function yields the current CPU to other tasks. If there are no
4225 * other threads running on this CPU then this function will return.
4227 SYSCALL_DEFINE0(sched_yield
)
4229 struct rq
*rq
= this_rq_lock();
4231 schedstat_inc(rq
, yld_count
);
4232 current
->sched_class
->yield_task(rq
);
4235 * Since we are going to call schedule() anyway, there's
4236 * no need to preempt or enable interrupts:
4238 __release(rq
->lock
);
4239 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4240 do_raw_spin_unlock(&rq
->lock
);
4241 sched_preempt_enable_no_resched();
4248 static inline int should_resched(void)
4250 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4253 static void __cond_resched(void)
4255 add_preempt_count(PREEMPT_ACTIVE
);
4257 sub_preempt_count(PREEMPT_ACTIVE
);
4260 int __sched
_cond_resched(void)
4262 if (should_resched()) {
4268 EXPORT_SYMBOL(_cond_resched
);
4271 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4272 * call schedule, and on return reacquire the lock.
4274 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4275 * operations here to prevent schedule() from being called twice (once via
4276 * spin_unlock(), once by hand).
4278 int __cond_resched_lock(spinlock_t
*lock
)
4280 int resched
= should_resched();
4283 lockdep_assert_held(lock
);
4285 if (spin_needbreak(lock
) || resched
) {
4296 EXPORT_SYMBOL(__cond_resched_lock
);
4298 int __sched
__cond_resched_softirq(void)
4300 BUG_ON(!in_softirq());
4302 if (should_resched()) {
4310 EXPORT_SYMBOL(__cond_resched_softirq
);
4313 * yield - yield the current processor to other threads.
4315 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4317 * The scheduler is at all times free to pick the calling task as the most
4318 * eligible task to run, if removing the yield() call from your code breaks
4319 * it, its already broken.
4321 * Typical broken usage is:
4326 * where one assumes that yield() will let 'the other' process run that will
4327 * make event true. If the current task is a SCHED_FIFO task that will never
4328 * happen. Never use yield() as a progress guarantee!!
4330 * If you want to use yield() to wait for something, use wait_event().
4331 * If you want to use yield() to be 'nice' for others, use cond_resched().
4332 * If you still want to use yield(), do not!
4334 void __sched
yield(void)
4336 set_current_state(TASK_RUNNING
);
4339 EXPORT_SYMBOL(yield
);
4342 * yield_to - yield the current processor to another thread in
4343 * your thread group, or accelerate that thread toward the
4344 * processor it's on.
4346 * @preempt: whether task preemption is allowed or not
4348 * It's the caller's job to ensure that the target task struct
4349 * can't go away on us before we can do any checks.
4351 * Returns true if we indeed boosted the target task.
4353 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4355 struct task_struct
*curr
= current
;
4356 struct rq
*rq
, *p_rq
;
4357 unsigned long flags
;
4360 local_irq_save(flags
);
4365 double_rq_lock(rq
, p_rq
);
4366 while (task_rq(p
) != p_rq
) {
4367 double_rq_unlock(rq
, p_rq
);
4371 if (!curr
->sched_class
->yield_to_task
)
4374 if (curr
->sched_class
!= p
->sched_class
)
4377 if (task_running(p_rq
, p
) || p
->state
)
4380 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4382 schedstat_inc(rq
, yld_count
);
4384 * Make p's CPU reschedule; pick_next_entity takes care of
4387 if (preempt
&& rq
!= p_rq
)
4388 resched_task(p_rq
->curr
);
4392 double_rq_unlock(rq
, p_rq
);
4393 local_irq_restore(flags
);
4400 EXPORT_SYMBOL_GPL(yield_to
);
4403 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4404 * that process accounting knows that this is a task in IO wait state.
4406 void __sched
io_schedule(void)
4408 struct rq
*rq
= raw_rq();
4410 delayacct_blkio_start();
4411 atomic_inc(&rq
->nr_iowait
);
4412 blk_flush_plug(current
);
4413 current
->in_iowait
= 1;
4415 current
->in_iowait
= 0;
4416 atomic_dec(&rq
->nr_iowait
);
4417 delayacct_blkio_end();
4419 EXPORT_SYMBOL(io_schedule
);
4421 long __sched
io_schedule_timeout(long timeout
)
4423 struct rq
*rq
= raw_rq();
4426 delayacct_blkio_start();
4427 atomic_inc(&rq
->nr_iowait
);
4428 blk_flush_plug(current
);
4429 current
->in_iowait
= 1;
4430 ret
= schedule_timeout(timeout
);
4431 current
->in_iowait
= 0;
4432 atomic_dec(&rq
->nr_iowait
);
4433 delayacct_blkio_end();
4438 * sys_sched_get_priority_max - return maximum RT priority.
4439 * @policy: scheduling class.
4441 * this syscall returns the maximum rt_priority that can be used
4442 * by a given scheduling class.
4444 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4451 ret
= MAX_USER_RT_PRIO
-1;
4463 * sys_sched_get_priority_min - return minimum RT priority.
4464 * @policy: scheduling class.
4466 * this syscall returns the minimum rt_priority that can be used
4467 * by a given scheduling class.
4469 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4487 * sys_sched_rr_get_interval - return the default timeslice of a process.
4488 * @pid: pid of the process.
4489 * @interval: userspace pointer to the timeslice value.
4491 * this syscall writes the default timeslice value of a given process
4492 * into the user-space timespec buffer. A value of '0' means infinity.
4494 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4495 struct timespec __user
*, interval
)
4497 struct task_struct
*p
;
4498 unsigned int time_slice
;
4499 unsigned long flags
;
4509 p
= find_process_by_pid(pid
);
4513 retval
= security_task_getscheduler(p
);
4517 rq
= task_rq_lock(p
, &flags
);
4518 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4519 task_rq_unlock(rq
, p
, &flags
);
4522 jiffies_to_timespec(time_slice
, &t
);
4523 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4531 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4533 void sched_show_task(struct task_struct
*p
)
4535 unsigned long free
= 0;
4539 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4540 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4541 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4542 #if BITS_PER_LONG == 32
4543 if (state
== TASK_RUNNING
)
4544 printk(KERN_CONT
" running ");
4546 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4548 if (state
== TASK_RUNNING
)
4549 printk(KERN_CONT
" running task ");
4551 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4553 #ifdef CONFIG_DEBUG_STACK_USAGE
4554 free
= stack_not_used(p
);
4557 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4559 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4560 task_pid_nr(p
), ppid
,
4561 (unsigned long)task_thread_info(p
)->flags
);
4563 show_stack(p
, NULL
);
4566 void show_state_filter(unsigned long state_filter
)
4568 struct task_struct
*g
, *p
;
4570 #if BITS_PER_LONG == 32
4572 " task PC stack pid father\n");
4575 " task PC stack pid father\n");
4578 do_each_thread(g
, p
) {
4580 * reset the NMI-timeout, listing all files on a slow
4581 * console might take a lot of time:
4583 touch_nmi_watchdog();
4584 if (!state_filter
|| (p
->state
& state_filter
))
4586 } while_each_thread(g
, p
);
4588 touch_all_softlockup_watchdogs();
4590 #ifdef CONFIG_SCHED_DEBUG
4591 sysrq_sched_debug_show();
4595 * Only show locks if all tasks are dumped:
4598 debug_show_all_locks();
4601 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4603 idle
->sched_class
= &idle_sched_class
;
4607 * init_idle - set up an idle thread for a given CPU
4608 * @idle: task in question
4609 * @cpu: cpu the idle task belongs to
4611 * NOTE: this function does not set the idle thread's NEED_RESCHED
4612 * flag, to make booting more robust.
4614 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4616 struct rq
*rq
= cpu_rq(cpu
);
4617 unsigned long flags
;
4619 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4622 idle
->state
= TASK_RUNNING
;
4623 idle
->se
.exec_start
= sched_clock();
4625 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4627 * We're having a chicken and egg problem, even though we are
4628 * holding rq->lock, the cpu isn't yet set to this cpu so the
4629 * lockdep check in task_group() will fail.
4631 * Similar case to sched_fork(). / Alternatively we could
4632 * use task_rq_lock() here and obtain the other rq->lock.
4637 __set_task_cpu(idle
, cpu
);
4640 rq
->curr
= rq
->idle
= idle
;
4641 #if defined(CONFIG_SMP)
4644 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4646 /* Set the preempt count _outside_ the spinlocks! */
4647 task_thread_info(idle
)->preempt_count
= 0;
4650 * The idle tasks have their own, simple scheduling class:
4652 idle
->sched_class
= &idle_sched_class
;
4653 ftrace_graph_init_idle_task(idle
, cpu
);
4654 vtime_init_idle(idle
);
4655 #if defined(CONFIG_SMP)
4656 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4661 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4663 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4664 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4666 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4667 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4671 * This is how migration works:
4673 * 1) we invoke migration_cpu_stop() on the target CPU using
4675 * 2) stopper starts to run (implicitly forcing the migrated thread
4677 * 3) it checks whether the migrated task is still in the wrong runqueue.
4678 * 4) if it's in the wrong runqueue then the migration thread removes
4679 * it and puts it into the right queue.
4680 * 5) stopper completes and stop_one_cpu() returns and the migration
4685 * Change a given task's CPU affinity. Migrate the thread to a
4686 * proper CPU and schedule it away if the CPU it's executing on
4687 * is removed from the allowed bitmask.
4689 * NOTE: the caller must have a valid reference to the task, the
4690 * task must not exit() & deallocate itself prematurely. The
4691 * call is not atomic; no spinlocks may be held.
4693 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4695 unsigned long flags
;
4697 unsigned int dest_cpu
;
4700 rq
= task_rq_lock(p
, &flags
);
4702 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4705 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4710 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4715 do_set_cpus_allowed(p
, new_mask
);
4717 /* Can the task run on the task's current CPU? If so, we're done */
4718 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4721 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4723 struct migration_arg arg
= { p
, dest_cpu
};
4724 /* Need help from migration thread: drop lock and wait. */
4725 task_rq_unlock(rq
, p
, &flags
);
4726 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4727 tlb_migrate_finish(p
->mm
);
4731 task_rq_unlock(rq
, p
, &flags
);
4735 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4738 * Move (not current) task off this cpu, onto dest cpu. We're doing
4739 * this because either it can't run here any more (set_cpus_allowed()
4740 * away from this CPU, or CPU going down), or because we're
4741 * attempting to rebalance this task on exec (sched_exec).
4743 * So we race with normal scheduler movements, but that's OK, as long
4744 * as the task is no longer on this CPU.
4746 * Returns non-zero if task was successfully migrated.
4748 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4750 struct rq
*rq_dest
, *rq_src
;
4753 if (unlikely(!cpu_active(dest_cpu
)))
4756 rq_src
= cpu_rq(src_cpu
);
4757 rq_dest
= cpu_rq(dest_cpu
);
4759 raw_spin_lock(&p
->pi_lock
);
4760 double_rq_lock(rq_src
, rq_dest
);
4761 /* Already moved. */
4762 if (task_cpu(p
) != src_cpu
)
4764 /* Affinity changed (again). */
4765 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4769 * If we're not on a rq, the next wake-up will ensure we're
4773 dequeue_task(rq_src
, p
, 0);
4774 set_task_cpu(p
, dest_cpu
);
4775 enqueue_task(rq_dest
, p
, 0);
4776 check_preempt_curr(rq_dest
, p
, 0);
4781 double_rq_unlock(rq_src
, rq_dest
);
4782 raw_spin_unlock(&p
->pi_lock
);
4787 * migration_cpu_stop - this will be executed by a highprio stopper thread
4788 * and performs thread migration by bumping thread off CPU then
4789 * 'pushing' onto another runqueue.
4791 static int migration_cpu_stop(void *data
)
4793 struct migration_arg
*arg
= data
;
4796 * The original target cpu might have gone down and we might
4797 * be on another cpu but it doesn't matter.
4799 local_irq_disable();
4800 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4805 #ifdef CONFIG_HOTPLUG_CPU
4808 * Ensures that the idle task is using init_mm right before its cpu goes
4811 void idle_task_exit(void)
4813 struct mm_struct
*mm
= current
->active_mm
;
4815 BUG_ON(cpu_online(smp_processor_id()));
4818 switch_mm(mm
, &init_mm
, current
);
4823 * Since this CPU is going 'away' for a while, fold any nr_active delta
4824 * we might have. Assumes we're called after migrate_tasks() so that the
4825 * nr_active count is stable.
4827 * Also see the comment "Global load-average calculations".
4829 static void calc_load_migrate(struct rq
*rq
)
4831 long delta
= calc_load_fold_active(rq
);
4833 atomic_long_add(delta
, &calc_load_tasks
);
4837 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4838 * try_to_wake_up()->select_task_rq().
4840 * Called with rq->lock held even though we'er in stop_machine() and
4841 * there's no concurrency possible, we hold the required locks anyway
4842 * because of lock validation efforts.
4844 static void migrate_tasks(unsigned int dead_cpu
)
4846 struct rq
*rq
= cpu_rq(dead_cpu
);
4847 struct task_struct
*next
, *stop
= rq
->stop
;
4851 * Fudge the rq selection such that the below task selection loop
4852 * doesn't get stuck on the currently eligible stop task.
4854 * We're currently inside stop_machine() and the rq is either stuck
4855 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4856 * either way we should never end up calling schedule() until we're
4863 * There's this thread running, bail when that's the only
4866 if (rq
->nr_running
== 1)
4869 next
= pick_next_task(rq
);
4871 next
->sched_class
->put_prev_task(rq
, next
);
4873 /* Find suitable destination for @next, with force if needed. */
4874 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4875 raw_spin_unlock(&rq
->lock
);
4877 __migrate_task(next
, dead_cpu
, dest_cpu
);
4879 raw_spin_lock(&rq
->lock
);
4885 #endif /* CONFIG_HOTPLUG_CPU */
4887 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4889 static struct ctl_table sd_ctl_dir
[] = {
4891 .procname
= "sched_domain",
4897 static struct ctl_table sd_ctl_root
[] = {
4899 .procname
= "kernel",
4901 .child
= sd_ctl_dir
,
4906 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4908 struct ctl_table
*entry
=
4909 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4914 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4916 struct ctl_table
*entry
;
4919 * In the intermediate directories, both the child directory and
4920 * procname are dynamically allocated and could fail but the mode
4921 * will always be set. In the lowest directory the names are
4922 * static strings and all have proc handlers.
4924 for (entry
= *tablep
; entry
->mode
; entry
++) {
4926 sd_free_ctl_entry(&entry
->child
);
4927 if (entry
->proc_handler
== NULL
)
4928 kfree(entry
->procname
);
4935 static int min_load_idx
= 0;
4936 static int max_load_idx
= CPU_LOAD_IDX_MAX
;
4939 set_table_entry(struct ctl_table
*entry
,
4940 const char *procname
, void *data
, int maxlen
,
4941 umode_t mode
, proc_handler
*proc_handler
,
4944 entry
->procname
= procname
;
4946 entry
->maxlen
= maxlen
;
4948 entry
->proc_handler
= proc_handler
;
4951 entry
->extra1
= &min_load_idx
;
4952 entry
->extra2
= &max_load_idx
;
4956 static struct ctl_table
*
4957 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4959 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4964 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4965 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4966 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4967 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4968 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4969 sizeof(int), 0644, proc_dointvec_minmax
, true);
4970 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4971 sizeof(int), 0644, proc_dointvec_minmax
, true);
4972 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4973 sizeof(int), 0644, proc_dointvec_minmax
, true);
4974 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4975 sizeof(int), 0644, proc_dointvec_minmax
, true);
4976 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4977 sizeof(int), 0644, proc_dointvec_minmax
, true);
4978 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4979 sizeof(int), 0644, proc_dointvec_minmax
, false);
4980 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4981 sizeof(int), 0644, proc_dointvec_minmax
, false);
4982 set_table_entry(&table
[9], "cache_nice_tries",
4983 &sd
->cache_nice_tries
,
4984 sizeof(int), 0644, proc_dointvec_minmax
, false);
4985 set_table_entry(&table
[10], "flags", &sd
->flags
,
4986 sizeof(int), 0644, proc_dointvec_minmax
, false);
4987 set_table_entry(&table
[11], "name", sd
->name
,
4988 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4989 /* &table[12] is terminator */
4994 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4996 struct ctl_table
*entry
, *table
;
4997 struct sched_domain
*sd
;
4998 int domain_num
= 0, i
;
5001 for_each_domain(cpu
, sd
)
5003 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5008 for_each_domain(cpu
, sd
) {
5009 snprintf(buf
, 32, "domain%d", i
);
5010 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5012 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5019 static struct ctl_table_header
*sd_sysctl_header
;
5020 static void register_sched_domain_sysctl(void)
5022 int i
, cpu_num
= num_possible_cpus();
5023 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5026 WARN_ON(sd_ctl_dir
[0].child
);
5027 sd_ctl_dir
[0].child
= entry
;
5032 for_each_possible_cpu(i
) {
5033 snprintf(buf
, 32, "cpu%d", i
);
5034 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5036 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5040 WARN_ON(sd_sysctl_header
);
5041 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5044 /* may be called multiple times per register */
5045 static void unregister_sched_domain_sysctl(void)
5047 if (sd_sysctl_header
)
5048 unregister_sysctl_table(sd_sysctl_header
);
5049 sd_sysctl_header
= NULL
;
5050 if (sd_ctl_dir
[0].child
)
5051 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5054 static void register_sched_domain_sysctl(void)
5057 static void unregister_sched_domain_sysctl(void)
5062 static void set_rq_online(struct rq
*rq
)
5065 const struct sched_class
*class;
5067 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5070 for_each_class(class) {
5071 if (class->rq_online
)
5072 class->rq_online(rq
);
5077 static void set_rq_offline(struct rq
*rq
)
5080 const struct sched_class
*class;
5082 for_each_class(class) {
5083 if (class->rq_offline
)
5084 class->rq_offline(rq
);
5087 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5093 * migration_call - callback that gets triggered when a CPU is added.
5094 * Here we can start up the necessary migration thread for the new CPU.
5096 static int __cpuinit
5097 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5099 int cpu
= (long)hcpu
;
5100 unsigned long flags
;
5101 struct rq
*rq
= cpu_rq(cpu
);
5103 switch (action
& ~CPU_TASKS_FROZEN
) {
5105 case CPU_UP_PREPARE
:
5106 rq
->calc_load_update
= calc_load_update
;
5110 /* Update our root-domain */
5111 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5113 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5117 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5120 #ifdef CONFIG_HOTPLUG_CPU
5122 sched_ttwu_pending();
5123 /* Update our root-domain */
5124 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5126 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5130 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5131 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5135 calc_load_migrate(rq
);
5140 update_max_interval();
5146 * Register at high priority so that task migration (migrate_all_tasks)
5147 * happens before everything else. This has to be lower priority than
5148 * the notifier in the perf_event subsystem, though.
5150 static struct notifier_block __cpuinitdata migration_notifier
= {
5151 .notifier_call
= migration_call
,
5152 .priority
= CPU_PRI_MIGRATION
,
5155 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5156 unsigned long action
, void *hcpu
)
5158 switch (action
& ~CPU_TASKS_FROZEN
) {
5160 case CPU_DOWN_FAILED
:
5161 set_cpu_active((long)hcpu
, true);
5168 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5169 unsigned long action
, void *hcpu
)
5171 switch (action
& ~CPU_TASKS_FROZEN
) {
5172 case CPU_DOWN_PREPARE
:
5173 set_cpu_active((long)hcpu
, false);
5180 static int __init
migration_init(void)
5182 void *cpu
= (void *)(long)smp_processor_id();
5185 /* Initialize migration for the boot CPU */
5186 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5187 BUG_ON(err
== NOTIFY_BAD
);
5188 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5189 register_cpu_notifier(&migration_notifier
);
5191 /* Register cpu active notifiers */
5192 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5193 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5197 early_initcall(migration_init
);
5202 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5204 #ifdef CONFIG_SCHED_DEBUG
5206 static __read_mostly
int sched_debug_enabled
;
5208 static int __init
sched_debug_setup(char *str
)
5210 sched_debug_enabled
= 1;
5214 early_param("sched_debug", sched_debug_setup
);
5216 static inline bool sched_debug(void)
5218 return sched_debug_enabled
;
5221 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5222 struct cpumask
*groupmask
)
5224 struct sched_group
*group
= sd
->groups
;
5227 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5228 cpumask_clear(groupmask
);
5230 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5232 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5233 printk("does not load-balance\n");
5235 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5240 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5242 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5243 printk(KERN_ERR
"ERROR: domain->span does not contain "
5246 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5247 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5251 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5255 printk(KERN_ERR
"ERROR: group is NULL\n");
5260 * Even though we initialize ->power to something semi-sane,
5261 * we leave power_orig unset. This allows us to detect if
5262 * domain iteration is still funny without causing /0 traps.
5264 if (!group
->sgp
->power_orig
) {
5265 printk(KERN_CONT
"\n");
5266 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5271 if (!cpumask_weight(sched_group_cpus(group
))) {
5272 printk(KERN_CONT
"\n");
5273 printk(KERN_ERR
"ERROR: empty group\n");
5277 if (!(sd
->flags
& SD_OVERLAP
) &&
5278 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5279 printk(KERN_CONT
"\n");
5280 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5284 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5286 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5288 printk(KERN_CONT
" %s", str
);
5289 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5290 printk(KERN_CONT
" (cpu_power = %d)",
5294 group
= group
->next
;
5295 } while (group
!= sd
->groups
);
5296 printk(KERN_CONT
"\n");
5298 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5299 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5302 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5303 printk(KERN_ERR
"ERROR: parent span is not a superset "
5304 "of domain->span\n");
5308 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5312 if (!sched_debug_enabled
)
5316 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5320 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5323 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5331 #else /* !CONFIG_SCHED_DEBUG */
5332 # define sched_domain_debug(sd, cpu) do { } while (0)
5333 static inline bool sched_debug(void)
5337 #endif /* CONFIG_SCHED_DEBUG */
5339 static int sd_degenerate(struct sched_domain
*sd
)
5341 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5344 /* Following flags need at least 2 groups */
5345 if (sd
->flags
& (SD_LOAD_BALANCE
|
5346 SD_BALANCE_NEWIDLE
|
5350 SD_SHARE_PKG_RESOURCES
)) {
5351 if (sd
->groups
!= sd
->groups
->next
)
5355 /* Following flags don't use groups */
5356 if (sd
->flags
& (SD_WAKE_AFFINE
))
5363 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5365 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5367 if (sd_degenerate(parent
))
5370 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5373 /* Flags needing groups don't count if only 1 group in parent */
5374 if (parent
->groups
== parent
->groups
->next
) {
5375 pflags
&= ~(SD_LOAD_BALANCE
|
5376 SD_BALANCE_NEWIDLE
|
5380 SD_SHARE_PKG_RESOURCES
);
5381 if (nr_node_ids
== 1)
5382 pflags
&= ~SD_SERIALIZE
;
5384 if (~cflags
& pflags
)
5390 static void free_rootdomain(struct rcu_head
*rcu
)
5392 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5394 cpupri_cleanup(&rd
->cpupri
);
5395 free_cpumask_var(rd
->rto_mask
);
5396 free_cpumask_var(rd
->online
);
5397 free_cpumask_var(rd
->span
);
5401 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5403 struct root_domain
*old_rd
= NULL
;
5404 unsigned long flags
;
5406 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5411 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5414 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5417 * If we dont want to free the old_rt yet then
5418 * set old_rd to NULL to skip the freeing later
5421 if (!atomic_dec_and_test(&old_rd
->refcount
))
5425 atomic_inc(&rd
->refcount
);
5428 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5429 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5432 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5435 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5438 static int init_rootdomain(struct root_domain
*rd
)
5440 memset(rd
, 0, sizeof(*rd
));
5442 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5444 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5446 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5449 if (cpupri_init(&rd
->cpupri
) != 0)
5454 free_cpumask_var(rd
->rto_mask
);
5456 free_cpumask_var(rd
->online
);
5458 free_cpumask_var(rd
->span
);
5464 * By default the system creates a single root-domain with all cpus as
5465 * members (mimicking the global state we have today).
5467 struct root_domain def_root_domain
;
5469 static void init_defrootdomain(void)
5471 init_rootdomain(&def_root_domain
);
5473 atomic_set(&def_root_domain
.refcount
, 1);
5476 static struct root_domain
*alloc_rootdomain(void)
5478 struct root_domain
*rd
;
5480 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5484 if (init_rootdomain(rd
) != 0) {
5492 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5494 struct sched_group
*tmp
, *first
;
5503 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5508 } while (sg
!= first
);
5511 static void free_sched_domain(struct rcu_head
*rcu
)
5513 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5516 * If its an overlapping domain it has private groups, iterate and
5519 if (sd
->flags
& SD_OVERLAP
) {
5520 free_sched_groups(sd
->groups
, 1);
5521 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5522 kfree(sd
->groups
->sgp
);
5528 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5530 call_rcu(&sd
->rcu
, free_sched_domain
);
5533 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5535 for (; sd
; sd
= sd
->parent
)
5536 destroy_sched_domain(sd
, cpu
);
5540 * Keep a special pointer to the highest sched_domain that has
5541 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5542 * allows us to avoid some pointer chasing select_idle_sibling().
5544 * Also keep a unique ID per domain (we use the first cpu number in
5545 * the cpumask of the domain), this allows us to quickly tell if
5546 * two cpus are in the same cache domain, see cpus_share_cache().
5548 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5549 DEFINE_PER_CPU(int, sd_llc_id
);
5551 static void update_top_cache_domain(int cpu
)
5553 struct sched_domain
*sd
;
5556 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5558 id
= cpumask_first(sched_domain_span(sd
));
5560 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5561 per_cpu(sd_llc_id
, cpu
) = id
;
5565 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5566 * hold the hotplug lock.
5569 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5571 struct rq
*rq
= cpu_rq(cpu
);
5572 struct sched_domain
*tmp
;
5574 /* Remove the sched domains which do not contribute to scheduling. */
5575 for (tmp
= sd
; tmp
; ) {
5576 struct sched_domain
*parent
= tmp
->parent
;
5580 if (sd_parent_degenerate(tmp
, parent
)) {
5581 tmp
->parent
= parent
->parent
;
5583 parent
->parent
->child
= tmp
;
5584 destroy_sched_domain(parent
, cpu
);
5589 if (sd
&& sd_degenerate(sd
)) {
5592 destroy_sched_domain(tmp
, cpu
);
5597 sched_domain_debug(sd
, cpu
);
5599 rq_attach_root(rq
, rd
);
5601 rcu_assign_pointer(rq
->sd
, sd
);
5602 destroy_sched_domains(tmp
, cpu
);
5604 update_top_cache_domain(cpu
);
5607 /* cpus with isolated domains */
5608 static cpumask_var_t cpu_isolated_map
;
5610 /* Setup the mask of cpus configured for isolated domains */
5611 static int __init
isolated_cpu_setup(char *str
)
5613 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5614 cpulist_parse(str
, cpu_isolated_map
);
5618 __setup("isolcpus=", isolated_cpu_setup
);
5620 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5622 return cpumask_of_node(cpu_to_node(cpu
));
5626 struct sched_domain
**__percpu sd
;
5627 struct sched_group
**__percpu sg
;
5628 struct sched_group_power
**__percpu sgp
;
5632 struct sched_domain
** __percpu sd
;
5633 struct root_domain
*rd
;
5643 struct sched_domain_topology_level
;
5645 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5646 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5648 #define SDTL_OVERLAP 0x01
5650 struct sched_domain_topology_level
{
5651 sched_domain_init_f init
;
5652 sched_domain_mask_f mask
;
5655 struct sd_data data
;
5659 * Build an iteration mask that can exclude certain CPUs from the upwards
5662 * Asymmetric node setups can result in situations where the domain tree is of
5663 * unequal depth, make sure to skip domains that already cover the entire
5666 * In that case build_sched_domains() will have terminated the iteration early
5667 * and our sibling sd spans will be empty. Domains should always include the
5668 * cpu they're built on, so check that.
5671 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5673 const struct cpumask
*span
= sched_domain_span(sd
);
5674 struct sd_data
*sdd
= sd
->private;
5675 struct sched_domain
*sibling
;
5678 for_each_cpu(i
, span
) {
5679 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5680 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5683 cpumask_set_cpu(i
, sched_group_mask(sg
));
5688 * Return the canonical balance cpu for this group, this is the first cpu
5689 * of this group that's also in the iteration mask.
5691 int group_balance_cpu(struct sched_group
*sg
)
5693 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5697 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5699 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5700 const struct cpumask
*span
= sched_domain_span(sd
);
5701 struct cpumask
*covered
= sched_domains_tmpmask
;
5702 struct sd_data
*sdd
= sd
->private;
5703 struct sched_domain
*child
;
5706 cpumask_clear(covered
);
5708 for_each_cpu(i
, span
) {
5709 struct cpumask
*sg_span
;
5711 if (cpumask_test_cpu(i
, covered
))
5714 child
= *per_cpu_ptr(sdd
->sd
, i
);
5716 /* See the comment near build_group_mask(). */
5717 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5720 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5721 GFP_KERNEL
, cpu_to_node(cpu
));
5726 sg_span
= sched_group_cpus(sg
);
5728 child
= child
->child
;
5729 cpumask_copy(sg_span
, sched_domain_span(child
));
5731 cpumask_set_cpu(i
, sg_span
);
5733 cpumask_or(covered
, covered
, sg_span
);
5735 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5736 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5737 build_group_mask(sd
, sg
);
5740 * Initialize sgp->power such that even if we mess up the
5741 * domains and no possible iteration will get us here, we won't
5744 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5747 * Make sure the first group of this domain contains the
5748 * canonical balance cpu. Otherwise the sched_domain iteration
5749 * breaks. See update_sg_lb_stats().
5751 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5752 group_balance_cpu(sg
) == cpu
)
5762 sd
->groups
= groups
;
5767 free_sched_groups(first
, 0);
5772 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5774 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5775 struct sched_domain
*child
= sd
->child
;
5778 cpu
= cpumask_first(sched_domain_span(child
));
5781 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5782 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5783 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5790 * build_sched_groups will build a circular linked list of the groups
5791 * covered by the given span, and will set each group's ->cpumask correctly,
5792 * and ->cpu_power to 0.
5794 * Assumes the sched_domain tree is fully constructed
5797 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5799 struct sched_group
*first
= NULL
, *last
= NULL
;
5800 struct sd_data
*sdd
= sd
->private;
5801 const struct cpumask
*span
= sched_domain_span(sd
);
5802 struct cpumask
*covered
;
5805 get_group(cpu
, sdd
, &sd
->groups
);
5806 atomic_inc(&sd
->groups
->ref
);
5808 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5811 lockdep_assert_held(&sched_domains_mutex
);
5812 covered
= sched_domains_tmpmask
;
5814 cpumask_clear(covered
);
5816 for_each_cpu(i
, span
) {
5817 struct sched_group
*sg
;
5818 int group
= get_group(i
, sdd
, &sg
);
5821 if (cpumask_test_cpu(i
, covered
))
5824 cpumask_clear(sched_group_cpus(sg
));
5826 cpumask_setall(sched_group_mask(sg
));
5828 for_each_cpu(j
, span
) {
5829 if (get_group(j
, sdd
, NULL
) != group
)
5832 cpumask_set_cpu(j
, covered
);
5833 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5848 * Initialize sched groups cpu_power.
5850 * cpu_power indicates the capacity of sched group, which is used while
5851 * distributing the load between different sched groups in a sched domain.
5852 * Typically cpu_power for all the groups in a sched domain will be same unless
5853 * there are asymmetries in the topology. If there are asymmetries, group
5854 * having more cpu_power will pickup more load compared to the group having
5857 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5859 struct sched_group
*sg
= sd
->groups
;
5861 WARN_ON(!sd
|| !sg
);
5864 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5866 } while (sg
!= sd
->groups
);
5868 if (cpu
!= group_balance_cpu(sg
))
5871 update_group_power(sd
, cpu
);
5872 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5875 int __weak
arch_sd_sibling_asym_packing(void)
5877 return 0*SD_ASYM_PACKING
;
5881 * Initializers for schedule domains
5882 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5885 #ifdef CONFIG_SCHED_DEBUG
5886 # define SD_INIT_NAME(sd, type) sd->name = #type
5888 # define SD_INIT_NAME(sd, type) do { } while (0)
5891 #define SD_INIT_FUNC(type) \
5892 static noinline struct sched_domain * \
5893 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5895 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5896 *sd = SD_##type##_INIT; \
5897 SD_INIT_NAME(sd, type); \
5898 sd->private = &tl->data; \
5903 #ifdef CONFIG_SCHED_SMT
5904 SD_INIT_FUNC(SIBLING
)
5906 #ifdef CONFIG_SCHED_MC
5909 #ifdef CONFIG_SCHED_BOOK
5913 static int default_relax_domain_level
= -1;
5914 int sched_domain_level_max
;
5916 static int __init
setup_relax_domain_level(char *str
)
5918 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5919 pr_warn("Unable to set relax_domain_level\n");
5923 __setup("relax_domain_level=", setup_relax_domain_level
);
5925 static void set_domain_attribute(struct sched_domain
*sd
,
5926 struct sched_domain_attr
*attr
)
5930 if (!attr
|| attr
->relax_domain_level
< 0) {
5931 if (default_relax_domain_level
< 0)
5934 request
= default_relax_domain_level
;
5936 request
= attr
->relax_domain_level
;
5937 if (request
< sd
->level
) {
5938 /* turn off idle balance on this domain */
5939 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5941 /* turn on idle balance on this domain */
5942 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5946 static void __sdt_free(const struct cpumask
*cpu_map
);
5947 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5949 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5950 const struct cpumask
*cpu_map
)
5954 if (!atomic_read(&d
->rd
->refcount
))
5955 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5957 free_percpu(d
->sd
); /* fall through */
5959 __sdt_free(cpu_map
); /* fall through */
5965 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5966 const struct cpumask
*cpu_map
)
5968 memset(d
, 0, sizeof(*d
));
5970 if (__sdt_alloc(cpu_map
))
5971 return sa_sd_storage
;
5972 d
->sd
= alloc_percpu(struct sched_domain
*);
5974 return sa_sd_storage
;
5975 d
->rd
= alloc_rootdomain();
5978 return sa_rootdomain
;
5982 * NULL the sd_data elements we've used to build the sched_domain and
5983 * sched_group structure so that the subsequent __free_domain_allocs()
5984 * will not free the data we're using.
5986 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5988 struct sd_data
*sdd
= sd
->private;
5990 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5991 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5993 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5994 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5996 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5997 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6000 #ifdef CONFIG_SCHED_SMT
6001 static const struct cpumask
*cpu_smt_mask(int cpu
)
6003 return topology_thread_cpumask(cpu
);
6008 * Topology list, bottom-up.
6010 static struct sched_domain_topology_level default_topology
[] = {
6011 #ifdef CONFIG_SCHED_SMT
6012 { sd_init_SIBLING
, cpu_smt_mask
, },
6014 #ifdef CONFIG_SCHED_MC
6015 { sd_init_MC
, cpu_coregroup_mask
, },
6017 #ifdef CONFIG_SCHED_BOOK
6018 { sd_init_BOOK
, cpu_book_mask
, },
6020 { sd_init_CPU
, cpu_cpu_mask
, },
6024 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6028 static int sched_domains_numa_levels
;
6029 static int *sched_domains_numa_distance
;
6030 static struct cpumask
***sched_domains_numa_masks
;
6031 static int sched_domains_curr_level
;
6033 static inline int sd_local_flags(int level
)
6035 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6038 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6041 static struct sched_domain
*
6042 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6044 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6045 int level
= tl
->numa_level
;
6046 int sd_weight
= cpumask_weight(
6047 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6049 *sd
= (struct sched_domain
){
6050 .min_interval
= sd_weight
,
6051 .max_interval
= 2*sd_weight
,
6053 .imbalance_pct
= 125,
6054 .cache_nice_tries
= 2,
6061 .flags
= 1*SD_LOAD_BALANCE
6062 | 1*SD_BALANCE_NEWIDLE
6067 | 0*SD_SHARE_CPUPOWER
6068 | 0*SD_SHARE_PKG_RESOURCES
6070 | 0*SD_PREFER_SIBLING
6071 | sd_local_flags(level
)
6073 .last_balance
= jiffies
,
6074 .balance_interval
= sd_weight
,
6076 SD_INIT_NAME(sd
, NUMA
);
6077 sd
->private = &tl
->data
;
6080 * Ugly hack to pass state to sd_numa_mask()...
6082 sched_domains_curr_level
= tl
->numa_level
;
6087 static const struct cpumask
*sd_numa_mask(int cpu
)
6089 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6092 static void sched_numa_warn(const char *str
)
6094 static int done
= false;
6102 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6104 for (i
= 0; i
< nr_node_ids
; i
++) {
6105 printk(KERN_WARNING
" ");
6106 for (j
= 0; j
< nr_node_ids
; j
++)
6107 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6108 printk(KERN_CONT
"\n");
6110 printk(KERN_WARNING
"\n");
6113 static bool find_numa_distance(int distance
)
6117 if (distance
== node_distance(0, 0))
6120 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6121 if (sched_domains_numa_distance
[i
] == distance
)
6128 static void sched_init_numa(void)
6130 int next_distance
, curr_distance
= node_distance(0, 0);
6131 struct sched_domain_topology_level
*tl
;
6135 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6136 if (!sched_domains_numa_distance
)
6140 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6141 * unique distances in the node_distance() table.
6143 * Assumes node_distance(0,j) includes all distances in
6144 * node_distance(i,j) in order to avoid cubic time.
6146 next_distance
= curr_distance
;
6147 for (i
= 0; i
< nr_node_ids
; i
++) {
6148 for (j
= 0; j
< nr_node_ids
; j
++) {
6149 for (k
= 0; k
< nr_node_ids
; k
++) {
6150 int distance
= node_distance(i
, k
);
6152 if (distance
> curr_distance
&&
6153 (distance
< next_distance
||
6154 next_distance
== curr_distance
))
6155 next_distance
= distance
;
6158 * While not a strong assumption it would be nice to know
6159 * about cases where if node A is connected to B, B is not
6160 * equally connected to A.
6162 if (sched_debug() && node_distance(k
, i
) != distance
)
6163 sched_numa_warn("Node-distance not symmetric");
6165 if (sched_debug() && i
&& !find_numa_distance(distance
))
6166 sched_numa_warn("Node-0 not representative");
6168 if (next_distance
!= curr_distance
) {
6169 sched_domains_numa_distance
[level
++] = next_distance
;
6170 sched_domains_numa_levels
= level
;
6171 curr_distance
= next_distance
;
6176 * In case of sched_debug() we verify the above assumption.
6182 * 'level' contains the number of unique distances, excluding the
6183 * identity distance node_distance(i,i).
6185 * The sched_domains_nume_distance[] array includes the actual distance
6190 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6191 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6192 * the array will contain less then 'level' members. This could be
6193 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6194 * in other functions.
6196 * We reset it to 'level' at the end of this function.
6198 sched_domains_numa_levels
= 0;
6200 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6201 if (!sched_domains_numa_masks
)
6205 * Now for each level, construct a mask per node which contains all
6206 * cpus of nodes that are that many hops away from us.
6208 for (i
= 0; i
< level
; i
++) {
6209 sched_domains_numa_masks
[i
] =
6210 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6211 if (!sched_domains_numa_masks
[i
])
6214 for (j
= 0; j
< nr_node_ids
; j
++) {
6215 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6219 sched_domains_numa_masks
[i
][j
] = mask
;
6221 for (k
= 0; k
< nr_node_ids
; k
++) {
6222 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6225 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6230 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6231 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6236 * Copy the default topology bits..
6238 for (i
= 0; default_topology
[i
].init
; i
++)
6239 tl
[i
] = default_topology
[i
];
6242 * .. and append 'j' levels of NUMA goodness.
6244 for (j
= 0; j
< level
; i
++, j
++) {
6245 tl
[i
] = (struct sched_domain_topology_level
){
6246 .init
= sd_numa_init
,
6247 .mask
= sd_numa_mask
,
6248 .flags
= SDTL_OVERLAP
,
6253 sched_domain_topology
= tl
;
6255 sched_domains_numa_levels
= level
;
6258 static void sched_domains_numa_masks_set(int cpu
)
6261 int node
= cpu_to_node(cpu
);
6263 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6264 for (j
= 0; j
< nr_node_ids
; j
++) {
6265 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6266 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6271 static void sched_domains_numa_masks_clear(int cpu
)
6274 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6275 for (j
= 0; j
< nr_node_ids
; j
++)
6276 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6281 * Update sched_domains_numa_masks[level][node] array when new cpus
6284 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6285 unsigned long action
,
6288 int cpu
= (long)hcpu
;
6290 switch (action
& ~CPU_TASKS_FROZEN
) {
6292 sched_domains_numa_masks_set(cpu
);
6296 sched_domains_numa_masks_clear(cpu
);
6306 static inline void sched_init_numa(void)
6310 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6311 unsigned long action
,
6316 #endif /* CONFIG_NUMA */
6318 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6320 struct sched_domain_topology_level
*tl
;
6323 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6324 struct sd_data
*sdd
= &tl
->data
;
6326 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6330 sdd
->sg
= alloc_percpu(struct sched_group
*);
6334 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6338 for_each_cpu(j
, cpu_map
) {
6339 struct sched_domain
*sd
;
6340 struct sched_group
*sg
;
6341 struct sched_group_power
*sgp
;
6343 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6344 GFP_KERNEL
, cpu_to_node(j
));
6348 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6350 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6351 GFP_KERNEL
, cpu_to_node(j
));
6357 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6359 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6360 GFP_KERNEL
, cpu_to_node(j
));
6364 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6371 static void __sdt_free(const struct cpumask
*cpu_map
)
6373 struct sched_domain_topology_level
*tl
;
6376 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6377 struct sd_data
*sdd
= &tl
->data
;
6379 for_each_cpu(j
, cpu_map
) {
6380 struct sched_domain
*sd
;
6383 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6384 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6385 free_sched_groups(sd
->groups
, 0);
6386 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6390 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6392 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6394 free_percpu(sdd
->sd
);
6396 free_percpu(sdd
->sg
);
6398 free_percpu(sdd
->sgp
);
6403 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6404 struct s_data
*d
, const struct cpumask
*cpu_map
,
6405 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6408 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6412 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6414 sd
->level
= child
->level
+ 1;
6415 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6419 set_domain_attribute(sd
, attr
);
6425 * Build sched domains for a given set of cpus and attach the sched domains
6426 * to the individual cpus
6428 static int build_sched_domains(const struct cpumask
*cpu_map
,
6429 struct sched_domain_attr
*attr
)
6431 enum s_alloc alloc_state
= sa_none
;
6432 struct sched_domain
*sd
;
6434 int i
, ret
= -ENOMEM
;
6436 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6437 if (alloc_state
!= sa_rootdomain
)
6440 /* Set up domains for cpus specified by the cpu_map. */
6441 for_each_cpu(i
, cpu_map
) {
6442 struct sched_domain_topology_level
*tl
;
6445 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6446 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6447 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6448 sd
->flags
|= SD_OVERLAP
;
6449 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6456 *per_cpu_ptr(d
.sd
, i
) = sd
;
6459 /* Build the groups for the domains */
6460 for_each_cpu(i
, cpu_map
) {
6461 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6462 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6463 if (sd
->flags
& SD_OVERLAP
) {
6464 if (build_overlap_sched_groups(sd
, i
))
6467 if (build_sched_groups(sd
, i
))
6473 /* Calculate CPU power for physical packages and nodes */
6474 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6475 if (!cpumask_test_cpu(i
, cpu_map
))
6478 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6479 claim_allocations(i
, sd
);
6480 init_sched_groups_power(i
, sd
);
6484 /* Attach the domains */
6486 for_each_cpu(i
, cpu_map
) {
6487 sd
= *per_cpu_ptr(d
.sd
, i
);
6488 cpu_attach_domain(sd
, d
.rd
, i
);
6494 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6498 static cpumask_var_t
*doms_cur
; /* current sched domains */
6499 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6500 static struct sched_domain_attr
*dattr_cur
;
6501 /* attribues of custom domains in 'doms_cur' */
6504 * Special case: If a kmalloc of a doms_cur partition (array of
6505 * cpumask) fails, then fallback to a single sched domain,
6506 * as determined by the single cpumask fallback_doms.
6508 static cpumask_var_t fallback_doms
;
6511 * arch_update_cpu_topology lets virtualized architectures update the
6512 * cpu core maps. It is supposed to return 1 if the topology changed
6513 * or 0 if it stayed the same.
6515 int __attribute__((weak
)) arch_update_cpu_topology(void)
6520 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6523 cpumask_var_t
*doms
;
6525 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6528 for (i
= 0; i
< ndoms
; i
++) {
6529 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6530 free_sched_domains(doms
, i
);
6537 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6540 for (i
= 0; i
< ndoms
; i
++)
6541 free_cpumask_var(doms
[i
]);
6546 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6547 * For now this just excludes isolated cpus, but could be used to
6548 * exclude other special cases in the future.
6550 static int init_sched_domains(const struct cpumask
*cpu_map
)
6554 arch_update_cpu_topology();
6556 doms_cur
= alloc_sched_domains(ndoms_cur
);
6558 doms_cur
= &fallback_doms
;
6559 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6560 err
= build_sched_domains(doms_cur
[0], NULL
);
6561 register_sched_domain_sysctl();
6567 * Detach sched domains from a group of cpus specified in cpu_map
6568 * These cpus will now be attached to the NULL domain
6570 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6575 for_each_cpu(i
, cpu_map
)
6576 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6580 /* handle null as "default" */
6581 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6582 struct sched_domain_attr
*new, int idx_new
)
6584 struct sched_domain_attr tmp
;
6591 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6592 new ? (new + idx_new
) : &tmp
,
6593 sizeof(struct sched_domain_attr
));
6597 * Partition sched domains as specified by the 'ndoms_new'
6598 * cpumasks in the array doms_new[] of cpumasks. This compares
6599 * doms_new[] to the current sched domain partitioning, doms_cur[].
6600 * It destroys each deleted domain and builds each new domain.
6602 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6603 * The masks don't intersect (don't overlap.) We should setup one
6604 * sched domain for each mask. CPUs not in any of the cpumasks will
6605 * not be load balanced. If the same cpumask appears both in the
6606 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6609 * The passed in 'doms_new' should be allocated using
6610 * alloc_sched_domains. This routine takes ownership of it and will
6611 * free_sched_domains it when done with it. If the caller failed the
6612 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6613 * and partition_sched_domains() will fallback to the single partition
6614 * 'fallback_doms', it also forces the domains to be rebuilt.
6616 * If doms_new == NULL it will be replaced with cpu_online_mask.
6617 * ndoms_new == 0 is a special case for destroying existing domains,
6618 * and it will not create the default domain.
6620 * Call with hotplug lock held
6622 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6623 struct sched_domain_attr
*dattr_new
)
6628 mutex_lock(&sched_domains_mutex
);
6630 /* always unregister in case we don't destroy any domains */
6631 unregister_sched_domain_sysctl();
6633 /* Let architecture update cpu core mappings. */
6634 new_topology
= arch_update_cpu_topology();
6636 n
= doms_new
? ndoms_new
: 0;
6638 /* Destroy deleted domains */
6639 for (i
= 0; i
< ndoms_cur
; i
++) {
6640 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6641 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6642 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6645 /* no match - a current sched domain not in new doms_new[] */
6646 detach_destroy_domains(doms_cur
[i
]);
6651 if (doms_new
== NULL
) {
6653 doms_new
= &fallback_doms
;
6654 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6655 WARN_ON_ONCE(dattr_new
);
6658 /* Build new domains */
6659 for (i
= 0; i
< ndoms_new
; i
++) {
6660 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6661 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6662 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6665 /* no match - add a new doms_new */
6666 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6671 /* Remember the new sched domains */
6672 if (doms_cur
!= &fallback_doms
)
6673 free_sched_domains(doms_cur
, ndoms_cur
);
6674 kfree(dattr_cur
); /* kfree(NULL) is safe */
6675 doms_cur
= doms_new
;
6676 dattr_cur
= dattr_new
;
6677 ndoms_cur
= ndoms_new
;
6679 register_sched_domain_sysctl();
6681 mutex_unlock(&sched_domains_mutex
);
6684 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6687 * Update cpusets according to cpu_active mask. If cpusets are
6688 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6689 * around partition_sched_domains().
6691 * If we come here as part of a suspend/resume, don't touch cpusets because we
6692 * want to restore it back to its original state upon resume anyway.
6694 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6698 case CPU_ONLINE_FROZEN
:
6699 case CPU_DOWN_FAILED_FROZEN
:
6702 * num_cpus_frozen tracks how many CPUs are involved in suspend
6703 * resume sequence. As long as this is not the last online
6704 * operation in the resume sequence, just build a single sched
6705 * domain, ignoring cpusets.
6708 if (likely(num_cpus_frozen
)) {
6709 partition_sched_domains(1, NULL
, NULL
);
6714 * This is the last CPU online operation. So fall through and
6715 * restore the original sched domains by considering the
6716 * cpuset configurations.
6720 case CPU_DOWN_FAILED
:
6721 cpuset_update_active_cpus(true);
6729 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6733 case CPU_DOWN_PREPARE
:
6734 cpuset_update_active_cpus(false);
6736 case CPU_DOWN_PREPARE_FROZEN
:
6738 partition_sched_domains(1, NULL
, NULL
);
6746 void __init
sched_init_smp(void)
6748 cpumask_var_t non_isolated_cpus
;
6750 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6751 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6756 mutex_lock(&sched_domains_mutex
);
6757 init_sched_domains(cpu_active_mask
);
6758 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6759 if (cpumask_empty(non_isolated_cpus
))
6760 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6761 mutex_unlock(&sched_domains_mutex
);
6764 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6765 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6766 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6768 /* RT runtime code needs to handle some hotplug events */
6769 hotcpu_notifier(update_runtime
, 0);
6773 /* Move init over to a non-isolated CPU */
6774 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6776 sched_init_granularity();
6777 free_cpumask_var(non_isolated_cpus
);
6779 init_sched_rt_class();
6782 void __init
sched_init_smp(void)
6784 sched_init_granularity();
6786 #endif /* CONFIG_SMP */
6788 const_debug
unsigned int sysctl_timer_migration
= 1;
6790 int in_sched_functions(unsigned long addr
)
6792 return in_lock_functions(addr
) ||
6793 (addr
>= (unsigned long)__sched_text_start
6794 && addr
< (unsigned long)__sched_text_end
);
6797 #ifdef CONFIG_CGROUP_SCHED
6798 struct task_group root_task_group
;
6799 LIST_HEAD(task_groups
);
6802 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6804 void __init
sched_init(void)
6807 unsigned long alloc_size
= 0, ptr
;
6809 #ifdef CONFIG_FAIR_GROUP_SCHED
6810 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6812 #ifdef CONFIG_RT_GROUP_SCHED
6813 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6815 #ifdef CONFIG_CPUMASK_OFFSTACK
6816 alloc_size
+= num_possible_cpus() * cpumask_size();
6819 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6821 #ifdef CONFIG_FAIR_GROUP_SCHED
6822 root_task_group
.se
= (struct sched_entity
**)ptr
;
6823 ptr
+= nr_cpu_ids
* sizeof(void **);
6825 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6826 ptr
+= nr_cpu_ids
* sizeof(void **);
6828 #endif /* CONFIG_FAIR_GROUP_SCHED */
6829 #ifdef CONFIG_RT_GROUP_SCHED
6830 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6831 ptr
+= nr_cpu_ids
* sizeof(void **);
6833 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6834 ptr
+= nr_cpu_ids
* sizeof(void **);
6836 #endif /* CONFIG_RT_GROUP_SCHED */
6837 #ifdef CONFIG_CPUMASK_OFFSTACK
6838 for_each_possible_cpu(i
) {
6839 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6840 ptr
+= cpumask_size();
6842 #endif /* CONFIG_CPUMASK_OFFSTACK */
6846 init_defrootdomain();
6849 init_rt_bandwidth(&def_rt_bandwidth
,
6850 global_rt_period(), global_rt_runtime());
6852 #ifdef CONFIG_RT_GROUP_SCHED
6853 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6854 global_rt_period(), global_rt_runtime());
6855 #endif /* CONFIG_RT_GROUP_SCHED */
6857 #ifdef CONFIG_CGROUP_SCHED
6858 list_add(&root_task_group
.list
, &task_groups
);
6859 INIT_LIST_HEAD(&root_task_group
.children
);
6860 INIT_LIST_HEAD(&root_task_group
.siblings
);
6861 autogroup_init(&init_task
);
6863 #endif /* CONFIG_CGROUP_SCHED */
6865 #ifdef CONFIG_CGROUP_CPUACCT
6866 root_cpuacct
.cpustat
= &kernel_cpustat
;
6867 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6868 /* Too early, not expected to fail */
6869 BUG_ON(!root_cpuacct
.cpuusage
);
6871 for_each_possible_cpu(i
) {
6875 raw_spin_lock_init(&rq
->lock
);
6877 rq
->calc_load_active
= 0;
6878 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6879 init_cfs_rq(&rq
->cfs
);
6880 init_rt_rq(&rq
->rt
, rq
);
6881 #ifdef CONFIG_FAIR_GROUP_SCHED
6882 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6883 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6885 * How much cpu bandwidth does root_task_group get?
6887 * In case of task-groups formed thr' the cgroup filesystem, it
6888 * gets 100% of the cpu resources in the system. This overall
6889 * system cpu resource is divided among the tasks of
6890 * root_task_group and its child task-groups in a fair manner,
6891 * based on each entity's (task or task-group's) weight
6892 * (se->load.weight).
6894 * In other words, if root_task_group has 10 tasks of weight
6895 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6896 * then A0's share of the cpu resource is:
6898 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6900 * We achieve this by letting root_task_group's tasks sit
6901 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6903 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6904 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6905 #endif /* CONFIG_FAIR_GROUP_SCHED */
6907 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6908 #ifdef CONFIG_RT_GROUP_SCHED
6909 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6910 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6913 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6914 rq
->cpu_load
[j
] = 0;
6916 rq
->last_load_update_tick
= jiffies
;
6921 rq
->cpu_power
= SCHED_POWER_SCALE
;
6922 rq
->post_schedule
= 0;
6923 rq
->active_balance
= 0;
6924 rq
->next_balance
= jiffies
;
6929 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6931 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6933 rq_attach_root(rq
, &def_root_domain
);
6939 atomic_set(&rq
->nr_iowait
, 0);
6942 set_load_weight(&init_task
);
6944 #ifdef CONFIG_PREEMPT_NOTIFIERS
6945 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6948 #ifdef CONFIG_RT_MUTEXES
6949 plist_head_init(&init_task
.pi_waiters
);
6953 * The boot idle thread does lazy MMU switching as well:
6955 atomic_inc(&init_mm
.mm_count
);
6956 enter_lazy_tlb(&init_mm
, current
);
6959 * Make us the idle thread. Technically, schedule() should not be
6960 * called from this thread, however somewhere below it might be,
6961 * but because we are the idle thread, we just pick up running again
6962 * when this runqueue becomes "idle".
6964 init_idle(current
, smp_processor_id());
6966 calc_load_update
= jiffies
+ LOAD_FREQ
;
6969 * During early bootup we pretend to be a normal task:
6971 current
->sched_class
= &fair_sched_class
;
6974 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6975 /* May be allocated at isolcpus cmdline parse time */
6976 if (cpu_isolated_map
== NULL
)
6977 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6978 idle_thread_set_boot_cpu();
6980 init_sched_fair_class();
6982 scheduler_running
= 1;
6985 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6986 static inline int preempt_count_equals(int preempt_offset
)
6988 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6990 return (nested
== preempt_offset
);
6993 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6995 static unsigned long prev_jiffy
; /* ratelimiting */
6997 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6998 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6999 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7001 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7003 prev_jiffy
= jiffies
;
7006 "BUG: sleeping function called from invalid context at %s:%d\n",
7009 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7010 in_atomic(), irqs_disabled(),
7011 current
->pid
, current
->comm
);
7013 debug_show_held_locks(current
);
7014 if (irqs_disabled())
7015 print_irqtrace_events(current
);
7018 EXPORT_SYMBOL(__might_sleep
);
7021 #ifdef CONFIG_MAGIC_SYSRQ
7022 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7024 const struct sched_class
*prev_class
= p
->sched_class
;
7025 int old_prio
= p
->prio
;
7030 dequeue_task(rq
, p
, 0);
7031 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7033 enqueue_task(rq
, p
, 0);
7034 resched_task(rq
->curr
);
7037 check_class_changed(rq
, p
, prev_class
, old_prio
);
7040 void normalize_rt_tasks(void)
7042 struct task_struct
*g
, *p
;
7043 unsigned long flags
;
7046 read_lock_irqsave(&tasklist_lock
, flags
);
7047 do_each_thread(g
, p
) {
7049 * Only normalize user tasks:
7054 p
->se
.exec_start
= 0;
7055 #ifdef CONFIG_SCHEDSTATS
7056 p
->se
.statistics
.wait_start
= 0;
7057 p
->se
.statistics
.sleep_start
= 0;
7058 p
->se
.statistics
.block_start
= 0;
7063 * Renice negative nice level userspace
7066 if (TASK_NICE(p
) < 0 && p
->mm
)
7067 set_user_nice(p
, 0);
7071 raw_spin_lock(&p
->pi_lock
);
7072 rq
= __task_rq_lock(p
);
7074 normalize_task(rq
, p
);
7076 __task_rq_unlock(rq
);
7077 raw_spin_unlock(&p
->pi_lock
);
7078 } while_each_thread(g
, p
);
7080 read_unlock_irqrestore(&tasklist_lock
, flags
);
7083 #endif /* CONFIG_MAGIC_SYSRQ */
7085 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7087 * These functions are only useful for the IA64 MCA handling, or kdb.
7089 * They can only be called when the whole system has been
7090 * stopped - every CPU needs to be quiescent, and no scheduling
7091 * activity can take place. Using them for anything else would
7092 * be a serious bug, and as a result, they aren't even visible
7093 * under any other configuration.
7097 * curr_task - return the current task for a given cpu.
7098 * @cpu: the processor in question.
7100 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7102 struct task_struct
*curr_task(int cpu
)
7104 return cpu_curr(cpu
);
7107 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7111 * set_curr_task - set the current task for a given cpu.
7112 * @cpu: the processor in question.
7113 * @p: the task pointer to set.
7115 * Description: This function must only be used when non-maskable interrupts
7116 * are serviced on a separate stack. It allows the architecture to switch the
7117 * notion of the current task on a cpu in a non-blocking manner. This function
7118 * must be called with all CPU's synchronized, and interrupts disabled, the
7119 * and caller must save the original value of the current task (see
7120 * curr_task() above) and restore that value before reenabling interrupts and
7121 * re-starting the system.
7123 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7125 void set_curr_task(int cpu
, struct task_struct
*p
)
7132 #ifdef CONFIG_CGROUP_SCHED
7133 /* task_group_lock serializes the addition/removal of task groups */
7134 static DEFINE_SPINLOCK(task_group_lock
);
7136 static void free_sched_group(struct task_group
*tg
)
7138 free_fair_sched_group(tg
);
7139 free_rt_sched_group(tg
);
7144 /* allocate runqueue etc for a new task group */
7145 struct task_group
*sched_create_group(struct task_group
*parent
)
7147 struct task_group
*tg
;
7148 unsigned long flags
;
7150 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7152 return ERR_PTR(-ENOMEM
);
7154 if (!alloc_fair_sched_group(tg
, parent
))
7157 if (!alloc_rt_sched_group(tg
, parent
))
7160 spin_lock_irqsave(&task_group_lock
, flags
);
7161 list_add_rcu(&tg
->list
, &task_groups
);
7163 WARN_ON(!parent
); /* root should already exist */
7165 tg
->parent
= parent
;
7166 INIT_LIST_HEAD(&tg
->children
);
7167 list_add_rcu(&tg
->siblings
, &parent
->children
);
7168 spin_unlock_irqrestore(&task_group_lock
, flags
);
7173 free_sched_group(tg
);
7174 return ERR_PTR(-ENOMEM
);
7177 /* rcu callback to free various structures associated with a task group */
7178 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7180 /* now it should be safe to free those cfs_rqs */
7181 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7184 /* Destroy runqueue etc associated with a task group */
7185 void sched_destroy_group(struct task_group
*tg
)
7187 unsigned long flags
;
7190 /* end participation in shares distribution */
7191 for_each_possible_cpu(i
)
7192 unregister_fair_sched_group(tg
, i
);
7194 spin_lock_irqsave(&task_group_lock
, flags
);
7195 list_del_rcu(&tg
->list
);
7196 list_del_rcu(&tg
->siblings
);
7197 spin_unlock_irqrestore(&task_group_lock
, flags
);
7199 /* wait for possible concurrent references to cfs_rqs complete */
7200 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7203 /* change task's runqueue when it moves between groups.
7204 * The caller of this function should have put the task in its new group
7205 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7206 * reflect its new group.
7208 void sched_move_task(struct task_struct
*tsk
)
7210 struct task_group
*tg
;
7212 unsigned long flags
;
7215 rq
= task_rq_lock(tsk
, &flags
);
7217 running
= task_current(rq
, tsk
);
7221 dequeue_task(rq
, tsk
, 0);
7222 if (unlikely(running
))
7223 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7225 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7226 lockdep_is_held(&tsk
->sighand
->siglock
)),
7227 struct task_group
, css
);
7228 tg
= autogroup_task_group(tsk
, tg
);
7229 tsk
->sched_task_group
= tg
;
7231 #ifdef CONFIG_FAIR_GROUP_SCHED
7232 if (tsk
->sched_class
->task_move_group
)
7233 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7236 set_task_rq(tsk
, task_cpu(tsk
));
7238 if (unlikely(running
))
7239 tsk
->sched_class
->set_curr_task(rq
);
7241 enqueue_task(rq
, tsk
, 0);
7243 task_rq_unlock(rq
, tsk
, &flags
);
7245 #endif /* CONFIG_CGROUP_SCHED */
7247 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7248 static unsigned long to_ratio(u64 period
, u64 runtime
)
7250 if (runtime
== RUNTIME_INF
)
7253 return div64_u64(runtime
<< 20, period
);
7257 #ifdef CONFIG_RT_GROUP_SCHED
7259 * Ensure that the real time constraints are schedulable.
7261 static DEFINE_MUTEX(rt_constraints_mutex
);
7263 /* Must be called with tasklist_lock held */
7264 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7266 struct task_struct
*g
, *p
;
7268 do_each_thread(g
, p
) {
7269 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7271 } while_each_thread(g
, p
);
7276 struct rt_schedulable_data
{
7277 struct task_group
*tg
;
7282 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7284 struct rt_schedulable_data
*d
= data
;
7285 struct task_group
*child
;
7286 unsigned long total
, sum
= 0;
7287 u64 period
, runtime
;
7289 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7290 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7293 period
= d
->rt_period
;
7294 runtime
= d
->rt_runtime
;
7298 * Cannot have more runtime than the period.
7300 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7304 * Ensure we don't starve existing RT tasks.
7306 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7309 total
= to_ratio(period
, runtime
);
7312 * Nobody can have more than the global setting allows.
7314 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7318 * The sum of our children's runtime should not exceed our own.
7320 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7321 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7322 runtime
= child
->rt_bandwidth
.rt_runtime
;
7324 if (child
== d
->tg
) {
7325 period
= d
->rt_period
;
7326 runtime
= d
->rt_runtime
;
7329 sum
+= to_ratio(period
, runtime
);
7338 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7342 struct rt_schedulable_data data
= {
7344 .rt_period
= period
,
7345 .rt_runtime
= runtime
,
7349 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7355 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7356 u64 rt_period
, u64 rt_runtime
)
7360 mutex_lock(&rt_constraints_mutex
);
7361 read_lock(&tasklist_lock
);
7362 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7366 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7367 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7368 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7370 for_each_possible_cpu(i
) {
7371 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7373 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7374 rt_rq
->rt_runtime
= rt_runtime
;
7375 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7377 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7379 read_unlock(&tasklist_lock
);
7380 mutex_unlock(&rt_constraints_mutex
);
7385 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7387 u64 rt_runtime
, rt_period
;
7389 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7390 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7391 if (rt_runtime_us
< 0)
7392 rt_runtime
= RUNTIME_INF
;
7394 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7397 long sched_group_rt_runtime(struct task_group
*tg
)
7401 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7404 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7405 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7406 return rt_runtime_us
;
7409 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7411 u64 rt_runtime
, rt_period
;
7413 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7414 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7419 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7422 long sched_group_rt_period(struct task_group
*tg
)
7426 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7427 do_div(rt_period_us
, NSEC_PER_USEC
);
7428 return rt_period_us
;
7431 static int sched_rt_global_constraints(void)
7433 u64 runtime
, period
;
7436 if (sysctl_sched_rt_period
<= 0)
7439 runtime
= global_rt_runtime();
7440 period
= global_rt_period();
7443 * Sanity check on the sysctl variables.
7445 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7448 mutex_lock(&rt_constraints_mutex
);
7449 read_lock(&tasklist_lock
);
7450 ret
= __rt_schedulable(NULL
, 0, 0);
7451 read_unlock(&tasklist_lock
);
7452 mutex_unlock(&rt_constraints_mutex
);
7457 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7459 /* Don't accept realtime tasks when there is no way for them to run */
7460 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7466 #else /* !CONFIG_RT_GROUP_SCHED */
7467 static int sched_rt_global_constraints(void)
7469 unsigned long flags
;
7472 if (sysctl_sched_rt_period
<= 0)
7476 * There's always some RT tasks in the root group
7477 * -- migration, kstopmachine etc..
7479 if (sysctl_sched_rt_runtime
== 0)
7482 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7483 for_each_possible_cpu(i
) {
7484 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7486 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7487 rt_rq
->rt_runtime
= global_rt_runtime();
7488 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7490 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7494 #endif /* CONFIG_RT_GROUP_SCHED */
7496 int sched_rr_handler(struct ctl_table
*table
, int write
,
7497 void __user
*buffer
, size_t *lenp
,
7501 static DEFINE_MUTEX(mutex
);
7504 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7505 /* make sure that internally we keep jiffies */
7506 /* also, writing zero resets timeslice to default */
7507 if (!ret
&& write
) {
7508 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7509 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7511 mutex_unlock(&mutex
);
7515 int sched_rt_handler(struct ctl_table
*table
, int write
,
7516 void __user
*buffer
, size_t *lenp
,
7520 int old_period
, old_runtime
;
7521 static DEFINE_MUTEX(mutex
);
7524 old_period
= sysctl_sched_rt_period
;
7525 old_runtime
= sysctl_sched_rt_runtime
;
7527 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7529 if (!ret
&& write
) {
7530 ret
= sched_rt_global_constraints();
7532 sysctl_sched_rt_period
= old_period
;
7533 sysctl_sched_rt_runtime
= old_runtime
;
7535 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7536 def_rt_bandwidth
.rt_period
=
7537 ns_to_ktime(global_rt_period());
7540 mutex_unlock(&mutex
);
7545 #ifdef CONFIG_CGROUP_SCHED
7547 /* return corresponding task_group object of a cgroup */
7548 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7550 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7551 struct task_group
, css
);
7554 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7556 struct task_group
*tg
, *parent
;
7558 if (!cgrp
->parent
) {
7559 /* This is early initialization for the top cgroup */
7560 return &root_task_group
.css
;
7563 parent
= cgroup_tg(cgrp
->parent
);
7564 tg
= sched_create_group(parent
);
7566 return ERR_PTR(-ENOMEM
);
7571 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7573 struct task_group
*tg
= cgroup_tg(cgrp
);
7575 sched_destroy_group(tg
);
7578 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7579 struct cgroup_taskset
*tset
)
7581 struct task_struct
*task
;
7583 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7584 #ifdef CONFIG_RT_GROUP_SCHED
7585 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7588 /* We don't support RT-tasks being in separate groups */
7589 if (task
->sched_class
!= &fair_sched_class
)
7596 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7597 struct cgroup_taskset
*tset
)
7599 struct task_struct
*task
;
7601 cgroup_taskset_for_each(task
, cgrp
, tset
)
7602 sched_move_task(task
);
7606 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7607 struct task_struct
*task
)
7610 * cgroup_exit() is called in the copy_process() failure path.
7611 * Ignore this case since the task hasn't ran yet, this avoids
7612 * trying to poke a half freed task state from generic code.
7614 if (!(task
->flags
& PF_EXITING
))
7617 sched_move_task(task
);
7620 #ifdef CONFIG_FAIR_GROUP_SCHED
7621 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7624 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7627 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7629 struct task_group
*tg
= cgroup_tg(cgrp
);
7631 return (u64
) scale_load_down(tg
->shares
);
7634 #ifdef CONFIG_CFS_BANDWIDTH
7635 static DEFINE_MUTEX(cfs_constraints_mutex
);
7637 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7638 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7640 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7642 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7644 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7645 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7647 if (tg
== &root_task_group
)
7651 * Ensure we have at some amount of bandwidth every period. This is
7652 * to prevent reaching a state of large arrears when throttled via
7653 * entity_tick() resulting in prolonged exit starvation.
7655 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7659 * Likewise, bound things on the otherside by preventing insane quota
7660 * periods. This also allows us to normalize in computing quota
7663 if (period
> max_cfs_quota_period
)
7666 mutex_lock(&cfs_constraints_mutex
);
7667 ret
= __cfs_schedulable(tg
, period
, quota
);
7671 runtime_enabled
= quota
!= RUNTIME_INF
;
7672 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7673 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7674 raw_spin_lock_irq(&cfs_b
->lock
);
7675 cfs_b
->period
= ns_to_ktime(period
);
7676 cfs_b
->quota
= quota
;
7678 __refill_cfs_bandwidth_runtime(cfs_b
);
7679 /* restart the period timer (if active) to handle new period expiry */
7680 if (runtime_enabled
&& cfs_b
->timer_active
) {
7681 /* force a reprogram */
7682 cfs_b
->timer_active
= 0;
7683 __start_cfs_bandwidth(cfs_b
);
7685 raw_spin_unlock_irq(&cfs_b
->lock
);
7687 for_each_possible_cpu(i
) {
7688 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7689 struct rq
*rq
= cfs_rq
->rq
;
7691 raw_spin_lock_irq(&rq
->lock
);
7692 cfs_rq
->runtime_enabled
= runtime_enabled
;
7693 cfs_rq
->runtime_remaining
= 0;
7695 if (cfs_rq
->throttled
)
7696 unthrottle_cfs_rq(cfs_rq
);
7697 raw_spin_unlock_irq(&rq
->lock
);
7700 mutex_unlock(&cfs_constraints_mutex
);
7705 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7709 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7710 if (cfs_quota_us
< 0)
7711 quota
= RUNTIME_INF
;
7713 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7715 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7718 long tg_get_cfs_quota(struct task_group
*tg
)
7722 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7725 quota_us
= tg
->cfs_bandwidth
.quota
;
7726 do_div(quota_us
, NSEC_PER_USEC
);
7731 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7735 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7736 quota
= tg
->cfs_bandwidth
.quota
;
7738 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7741 long tg_get_cfs_period(struct task_group
*tg
)
7745 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7746 do_div(cfs_period_us
, NSEC_PER_USEC
);
7748 return cfs_period_us
;
7751 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7753 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7756 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7759 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7762 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7764 return tg_get_cfs_period(cgroup_tg(cgrp
));
7767 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7770 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7773 struct cfs_schedulable_data
{
7774 struct task_group
*tg
;
7779 * normalize group quota/period to be quota/max_period
7780 * note: units are usecs
7782 static u64
normalize_cfs_quota(struct task_group
*tg
,
7783 struct cfs_schedulable_data
*d
)
7791 period
= tg_get_cfs_period(tg
);
7792 quota
= tg_get_cfs_quota(tg
);
7795 /* note: these should typically be equivalent */
7796 if (quota
== RUNTIME_INF
|| quota
== -1)
7799 return to_ratio(period
, quota
);
7802 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7804 struct cfs_schedulable_data
*d
= data
;
7805 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7806 s64 quota
= 0, parent_quota
= -1;
7809 quota
= RUNTIME_INF
;
7811 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7813 quota
= normalize_cfs_quota(tg
, d
);
7814 parent_quota
= parent_b
->hierarchal_quota
;
7817 * ensure max(child_quota) <= parent_quota, inherit when no
7820 if (quota
== RUNTIME_INF
)
7821 quota
= parent_quota
;
7822 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7825 cfs_b
->hierarchal_quota
= quota
;
7830 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7833 struct cfs_schedulable_data data
= {
7839 if (quota
!= RUNTIME_INF
) {
7840 do_div(data
.period
, NSEC_PER_USEC
);
7841 do_div(data
.quota
, NSEC_PER_USEC
);
7845 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7851 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7852 struct cgroup_map_cb
*cb
)
7854 struct task_group
*tg
= cgroup_tg(cgrp
);
7855 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7857 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7858 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7859 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7863 #endif /* CONFIG_CFS_BANDWIDTH */
7864 #endif /* CONFIG_FAIR_GROUP_SCHED */
7866 #ifdef CONFIG_RT_GROUP_SCHED
7867 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7870 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7873 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7875 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7878 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7881 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7884 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7886 return sched_group_rt_period(cgroup_tg(cgrp
));
7888 #endif /* CONFIG_RT_GROUP_SCHED */
7890 static struct cftype cpu_files
[] = {
7891 #ifdef CONFIG_FAIR_GROUP_SCHED
7894 .read_u64
= cpu_shares_read_u64
,
7895 .write_u64
= cpu_shares_write_u64
,
7898 #ifdef CONFIG_CFS_BANDWIDTH
7900 .name
= "cfs_quota_us",
7901 .read_s64
= cpu_cfs_quota_read_s64
,
7902 .write_s64
= cpu_cfs_quota_write_s64
,
7905 .name
= "cfs_period_us",
7906 .read_u64
= cpu_cfs_period_read_u64
,
7907 .write_u64
= cpu_cfs_period_write_u64
,
7911 .read_map
= cpu_stats_show
,
7914 #ifdef CONFIG_RT_GROUP_SCHED
7916 .name
= "rt_runtime_us",
7917 .read_s64
= cpu_rt_runtime_read
,
7918 .write_s64
= cpu_rt_runtime_write
,
7921 .name
= "rt_period_us",
7922 .read_u64
= cpu_rt_period_read_uint
,
7923 .write_u64
= cpu_rt_period_write_uint
,
7929 struct cgroup_subsys cpu_cgroup_subsys
= {
7931 .css_alloc
= cpu_cgroup_css_alloc
,
7932 .css_free
= cpu_cgroup_css_free
,
7933 .can_attach
= cpu_cgroup_can_attach
,
7934 .attach
= cpu_cgroup_attach
,
7935 .exit
= cpu_cgroup_exit
,
7936 .subsys_id
= cpu_cgroup_subsys_id
,
7937 .base_cftypes
= cpu_files
,
7941 #endif /* CONFIG_CGROUP_SCHED */
7943 #ifdef CONFIG_CGROUP_CPUACCT
7946 * CPU accounting code for task groups.
7948 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7949 * (balbir@in.ibm.com).
7952 struct cpuacct root_cpuacct
;
7954 /* create a new cpu accounting group */
7955 static struct cgroup_subsys_state
*cpuacct_css_alloc(struct cgroup
*cgrp
)
7960 return &root_cpuacct
.css
;
7962 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7966 ca
->cpuusage
= alloc_percpu(u64
);
7970 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7972 goto out_free_cpuusage
;
7977 free_percpu(ca
->cpuusage
);
7981 return ERR_PTR(-ENOMEM
);
7984 /* destroy an existing cpu accounting group */
7985 static void cpuacct_css_free(struct cgroup
*cgrp
)
7987 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7989 free_percpu(ca
->cpustat
);
7990 free_percpu(ca
->cpuusage
);
7994 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
7996 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7999 #ifndef CONFIG_64BIT
8001 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8003 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8005 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8013 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8015 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8017 #ifndef CONFIG_64BIT
8019 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8021 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8023 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8029 /* return total cpu usage (in nanoseconds) of a group */
8030 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8032 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8033 u64 totalcpuusage
= 0;
8036 for_each_present_cpu(i
)
8037 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8039 return totalcpuusage
;
8042 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8045 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8054 for_each_present_cpu(i
)
8055 cpuacct_cpuusage_write(ca
, i
, 0);
8061 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8064 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8068 for_each_present_cpu(i
) {
8069 percpu
= cpuacct_cpuusage_read(ca
, i
);
8070 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8072 seq_printf(m
, "\n");
8076 static const char *cpuacct_stat_desc
[] = {
8077 [CPUACCT_STAT_USER
] = "user",
8078 [CPUACCT_STAT_SYSTEM
] = "system",
8081 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8082 struct cgroup_map_cb
*cb
)
8084 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8088 for_each_online_cpu(cpu
) {
8089 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8090 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8091 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8093 val
= cputime64_to_clock_t(val
);
8094 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8097 for_each_online_cpu(cpu
) {
8098 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8099 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8100 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8101 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8104 val
= cputime64_to_clock_t(val
);
8105 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8110 static struct cftype files
[] = {
8113 .read_u64
= cpuusage_read
,
8114 .write_u64
= cpuusage_write
,
8117 .name
= "usage_percpu",
8118 .read_seq_string
= cpuacct_percpu_seq_read
,
8122 .read_map
= cpuacct_stats_show
,
8128 * charge this task's execution time to its accounting group.
8130 * called with rq->lock held.
8132 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8137 if (unlikely(!cpuacct_subsys
.active
))
8140 cpu
= task_cpu(tsk
);
8146 for (; ca
; ca
= parent_ca(ca
)) {
8147 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8148 *cpuusage
+= cputime
;
8154 struct cgroup_subsys cpuacct_subsys
= {
8156 .css_alloc
= cpuacct_css_alloc
,
8157 .css_free
= cpuacct_css_free
,
8158 .subsys_id
= cpuacct_subsys_id
,
8159 .base_cftypes
= files
,
8161 #endif /* CONFIG_CGROUP_CPUACCT */
8163 void dump_cpu_task(int cpu
)
8165 pr_info("Task dump for CPU %d:\n", cpu
);
8166 sched_show_task(cpu_curr(cpu
));