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_sched.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 BUG_ON(rq
!= this_rq());
1492 BUG_ON(p
== current
);
1493 lockdep_assert_held(&rq
->lock
);
1495 if (!raw_spin_trylock(&p
->pi_lock
)) {
1496 raw_spin_unlock(&rq
->lock
);
1497 raw_spin_lock(&p
->pi_lock
);
1498 raw_spin_lock(&rq
->lock
);
1501 if (!(p
->state
& TASK_NORMAL
))
1505 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1507 ttwu_do_wakeup(rq
, p
, 0);
1508 ttwu_stat(p
, smp_processor_id(), 0);
1510 raw_spin_unlock(&p
->pi_lock
);
1514 * wake_up_process - Wake up a specific process
1515 * @p: The process to be woken up.
1517 * Attempt to wake up the nominated process and move it to the set of runnable
1518 * processes. Returns 1 if the process was woken up, 0 if it was already
1521 * It may be assumed that this function implies a write memory barrier before
1522 * changing the task state if and only if any tasks are woken up.
1524 int wake_up_process(struct task_struct
*p
)
1526 return try_to_wake_up(p
, TASK_ALL
, 0);
1528 EXPORT_SYMBOL(wake_up_process
);
1530 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1532 return try_to_wake_up(p
, state
, 0);
1536 * Perform scheduler related setup for a newly forked process p.
1537 * p is forked by current.
1539 * __sched_fork() is basic setup used by init_idle() too:
1541 static void __sched_fork(struct task_struct
*p
)
1546 p
->se
.exec_start
= 0;
1547 p
->se
.sum_exec_runtime
= 0;
1548 p
->se
.prev_sum_exec_runtime
= 0;
1549 p
->se
.nr_migrations
= 0;
1551 INIT_LIST_HEAD(&p
->se
.group_node
);
1554 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1555 * removed when useful for applications beyond shares distribution (e.g.
1558 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1559 p
->se
.avg
.runnable_avg_period
= 0;
1560 p
->se
.avg
.runnable_avg_sum
= 0;
1562 #ifdef CONFIG_SCHEDSTATS
1563 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1566 INIT_LIST_HEAD(&p
->rt
.run_list
);
1568 #ifdef CONFIG_PREEMPT_NOTIFIERS
1569 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1572 #ifdef CONFIG_NUMA_BALANCING
1573 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1574 p
->mm
->numa_next_scan
= jiffies
;
1575 p
->mm
->numa_next_reset
= jiffies
;
1576 p
->mm
->numa_scan_seq
= 0;
1579 p
->node_stamp
= 0ULL;
1580 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1581 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1582 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1583 p
->numa_work
.next
= &p
->numa_work
;
1584 #endif /* CONFIG_NUMA_BALANCING */
1587 #ifdef CONFIG_NUMA_BALANCING
1588 #ifdef CONFIG_SCHED_DEBUG
1589 void set_numabalancing_state(bool enabled
)
1592 sched_feat_set("NUMA");
1594 sched_feat_set("NO_NUMA");
1597 __read_mostly
bool numabalancing_enabled
;
1599 void set_numabalancing_state(bool enabled
)
1601 numabalancing_enabled
= enabled
;
1603 #endif /* CONFIG_SCHED_DEBUG */
1604 #endif /* CONFIG_NUMA_BALANCING */
1607 * fork()/clone()-time setup:
1609 void sched_fork(struct task_struct
*p
)
1611 unsigned long flags
;
1612 int cpu
= get_cpu();
1616 * We mark the process as running here. This guarantees that
1617 * nobody will actually run it, and a signal or other external
1618 * event cannot wake it up and insert it on the runqueue either.
1620 p
->state
= TASK_RUNNING
;
1623 * Make sure we do not leak PI boosting priority to the child.
1625 p
->prio
= current
->normal_prio
;
1628 * Revert to default priority/policy on fork if requested.
1630 if (unlikely(p
->sched_reset_on_fork
)) {
1631 if (task_has_rt_policy(p
)) {
1632 p
->policy
= SCHED_NORMAL
;
1633 p
->static_prio
= NICE_TO_PRIO(0);
1635 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1636 p
->static_prio
= NICE_TO_PRIO(0);
1638 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1642 * We don't need the reset flag anymore after the fork. It has
1643 * fulfilled its duty:
1645 p
->sched_reset_on_fork
= 0;
1648 if (!rt_prio(p
->prio
))
1649 p
->sched_class
= &fair_sched_class
;
1651 if (p
->sched_class
->task_fork
)
1652 p
->sched_class
->task_fork(p
);
1655 * The child is not yet in the pid-hash so no cgroup attach races,
1656 * and the cgroup is pinned to this child due to cgroup_fork()
1657 * is ran before sched_fork().
1659 * Silence PROVE_RCU.
1661 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1662 set_task_cpu(p
, cpu
);
1663 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1665 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1666 if (likely(sched_info_on()))
1667 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1669 #if defined(CONFIG_SMP)
1672 #ifdef CONFIG_PREEMPT_COUNT
1673 /* Want to start with kernel preemption disabled. */
1674 task_thread_info(p
)->preempt_count
= 1;
1677 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1684 * wake_up_new_task - wake up a newly created task for the first time.
1686 * This function will do some initial scheduler statistics housekeeping
1687 * that must be done for every newly created context, then puts the task
1688 * on the runqueue and wakes it.
1690 void wake_up_new_task(struct task_struct
*p
)
1692 unsigned long flags
;
1695 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1698 * Fork balancing, do it here and not earlier because:
1699 * - cpus_allowed can change in the fork path
1700 * - any previously selected cpu might disappear through hotplug
1702 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1705 rq
= __task_rq_lock(p
);
1706 activate_task(rq
, p
, 0);
1708 trace_sched_wakeup_new(p
, true);
1709 check_preempt_curr(rq
, p
, WF_FORK
);
1711 if (p
->sched_class
->task_woken
)
1712 p
->sched_class
->task_woken(rq
, p
);
1714 task_rq_unlock(rq
, p
, &flags
);
1717 #ifdef CONFIG_PREEMPT_NOTIFIERS
1720 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1721 * @notifier: notifier struct to register
1723 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1725 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1727 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1730 * preempt_notifier_unregister - no longer interested in preemption notifications
1731 * @notifier: notifier struct to unregister
1733 * This is safe to call from within a preemption notifier.
1735 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1737 hlist_del(¬ifier
->link
);
1739 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1741 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1743 struct preempt_notifier
*notifier
;
1744 struct hlist_node
*node
;
1746 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1747 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1751 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1752 struct task_struct
*next
)
1754 struct preempt_notifier
*notifier
;
1755 struct hlist_node
*node
;
1757 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1758 notifier
->ops
->sched_out(notifier
, next
);
1761 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1763 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1768 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1769 struct task_struct
*next
)
1773 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1776 * prepare_task_switch - prepare to switch tasks
1777 * @rq: the runqueue preparing to switch
1778 * @prev: the current task that is being switched out
1779 * @next: the task we are going to switch to.
1781 * This is called with the rq lock held and interrupts off. It must
1782 * be paired with a subsequent finish_task_switch after the context
1785 * prepare_task_switch sets up locking and calls architecture specific
1789 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1790 struct task_struct
*next
)
1792 trace_sched_switch(prev
, next
);
1793 sched_info_switch(prev
, next
);
1794 perf_event_task_sched_out(prev
, next
);
1795 fire_sched_out_preempt_notifiers(prev
, next
);
1796 prepare_lock_switch(rq
, next
);
1797 prepare_arch_switch(next
);
1801 * finish_task_switch - clean up after a task-switch
1802 * @rq: runqueue associated with task-switch
1803 * @prev: the thread we just switched away from.
1805 * finish_task_switch must be called after the context switch, paired
1806 * with a prepare_task_switch call before the context switch.
1807 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1808 * and do any other architecture-specific cleanup actions.
1810 * Note that we may have delayed dropping an mm in context_switch(). If
1811 * so, we finish that here outside of the runqueue lock. (Doing it
1812 * with the lock held can cause deadlocks; see schedule() for
1815 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1816 __releases(rq
->lock
)
1818 struct mm_struct
*mm
= rq
->prev_mm
;
1824 * A task struct has one reference for the use as "current".
1825 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1826 * schedule one last time. The schedule call will never return, and
1827 * the scheduled task must drop that reference.
1828 * The test for TASK_DEAD must occur while the runqueue locks are
1829 * still held, otherwise prev could be scheduled on another cpu, die
1830 * there before we look at prev->state, and then the reference would
1832 * Manfred Spraul <manfred@colorfullife.com>
1834 prev_state
= prev
->state
;
1835 vtime_task_switch(prev
);
1836 finish_arch_switch(prev
);
1837 perf_event_task_sched_in(prev
, current
);
1838 finish_lock_switch(rq
, prev
);
1839 finish_arch_post_lock_switch();
1841 fire_sched_in_preempt_notifiers(current
);
1844 if (unlikely(prev_state
== TASK_DEAD
)) {
1846 * Remove function-return probe instances associated with this
1847 * task and put them back on the free list.
1849 kprobe_flush_task(prev
);
1850 put_task_struct(prev
);
1856 /* assumes rq->lock is held */
1857 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1859 if (prev
->sched_class
->pre_schedule
)
1860 prev
->sched_class
->pre_schedule(rq
, prev
);
1863 /* rq->lock is NOT held, but preemption is disabled */
1864 static inline void post_schedule(struct rq
*rq
)
1866 if (rq
->post_schedule
) {
1867 unsigned long flags
;
1869 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1870 if (rq
->curr
->sched_class
->post_schedule
)
1871 rq
->curr
->sched_class
->post_schedule(rq
);
1872 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1874 rq
->post_schedule
= 0;
1880 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1884 static inline void post_schedule(struct rq
*rq
)
1891 * schedule_tail - first thing a freshly forked thread must call.
1892 * @prev: the thread we just switched away from.
1894 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1895 __releases(rq
->lock
)
1897 struct rq
*rq
= this_rq();
1899 finish_task_switch(rq
, prev
);
1902 * FIXME: do we need to worry about rq being invalidated by the
1907 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1908 /* In this case, finish_task_switch does not reenable preemption */
1911 if (current
->set_child_tid
)
1912 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1916 * context_switch - switch to the new MM and the new
1917 * thread's register state.
1920 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1921 struct task_struct
*next
)
1923 struct mm_struct
*mm
, *oldmm
;
1925 prepare_task_switch(rq
, prev
, next
);
1928 oldmm
= prev
->active_mm
;
1930 * For paravirt, this is coupled with an exit in switch_to to
1931 * combine the page table reload and the switch backend into
1934 arch_start_context_switch(prev
);
1937 next
->active_mm
= oldmm
;
1938 atomic_inc(&oldmm
->mm_count
);
1939 enter_lazy_tlb(oldmm
, next
);
1941 switch_mm(oldmm
, mm
, next
);
1944 prev
->active_mm
= NULL
;
1945 rq
->prev_mm
= oldmm
;
1948 * Since the runqueue lock will be released by the next
1949 * task (which is an invalid locking op but in the case
1950 * of the scheduler it's an obvious special-case), so we
1951 * do an early lockdep release here:
1953 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1954 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1957 context_tracking_task_switch(prev
, next
);
1958 /* Here we just switch the register state and the stack. */
1959 switch_to(prev
, next
, prev
);
1963 * this_rq must be evaluated again because prev may have moved
1964 * CPUs since it called schedule(), thus the 'rq' on its stack
1965 * frame will be invalid.
1967 finish_task_switch(this_rq(), prev
);
1971 * nr_running, nr_uninterruptible and nr_context_switches:
1973 * externally visible scheduler statistics: current number of runnable
1974 * threads, current number of uninterruptible-sleeping threads, total
1975 * number of context switches performed since bootup.
1977 unsigned long nr_running(void)
1979 unsigned long i
, sum
= 0;
1981 for_each_online_cpu(i
)
1982 sum
+= cpu_rq(i
)->nr_running
;
1987 unsigned long nr_uninterruptible(void)
1989 unsigned long i
, sum
= 0;
1991 for_each_possible_cpu(i
)
1992 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1995 * Since we read the counters lockless, it might be slightly
1996 * inaccurate. Do not allow it to go below zero though:
1998 if (unlikely((long)sum
< 0))
2004 unsigned long long nr_context_switches(void)
2007 unsigned long long sum
= 0;
2009 for_each_possible_cpu(i
)
2010 sum
+= cpu_rq(i
)->nr_switches
;
2015 unsigned long nr_iowait(void)
2017 unsigned long i
, sum
= 0;
2019 for_each_possible_cpu(i
)
2020 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2025 unsigned long nr_iowait_cpu(int cpu
)
2027 struct rq
*this = cpu_rq(cpu
);
2028 return atomic_read(&this->nr_iowait
);
2031 unsigned long this_cpu_load(void)
2033 struct rq
*this = this_rq();
2034 return this->cpu_load
[0];
2039 * Global load-average calculations
2041 * We take a distributed and async approach to calculating the global load-avg
2042 * in order to minimize overhead.
2044 * The global load average is an exponentially decaying average of nr_running +
2045 * nr_uninterruptible.
2047 * Once every LOAD_FREQ:
2050 * for_each_possible_cpu(cpu)
2051 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2053 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2055 * Due to a number of reasons the above turns in the mess below:
2057 * - for_each_possible_cpu() is prohibitively expensive on machines with
2058 * serious number of cpus, therefore we need to take a distributed approach
2059 * to calculating nr_active.
2061 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2062 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2064 * So assuming nr_active := 0 when we start out -- true per definition, we
2065 * can simply take per-cpu deltas and fold those into a global accumulate
2066 * to obtain the same result. See calc_load_fold_active().
2068 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2069 * across the machine, we assume 10 ticks is sufficient time for every
2070 * cpu to have completed this task.
2072 * This places an upper-bound on the IRQ-off latency of the machine. Then
2073 * again, being late doesn't loose the delta, just wrecks the sample.
2075 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2076 * this would add another cross-cpu cacheline miss and atomic operation
2077 * to the wakeup path. Instead we increment on whatever cpu the task ran
2078 * when it went into uninterruptible state and decrement on whatever cpu
2079 * did the wakeup. This means that only the sum of nr_uninterruptible over
2080 * all cpus yields the correct result.
2082 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2085 /* Variables and functions for calc_load */
2086 static atomic_long_t calc_load_tasks
;
2087 static unsigned long calc_load_update
;
2088 unsigned long avenrun
[3];
2089 EXPORT_SYMBOL(avenrun
); /* should be removed */
2092 * get_avenrun - get the load average array
2093 * @loads: pointer to dest load array
2094 * @offset: offset to add
2095 * @shift: shift count to shift the result left
2097 * These values are estimates at best, so no need for locking.
2099 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2101 loads
[0] = (avenrun
[0] + offset
) << shift
;
2102 loads
[1] = (avenrun
[1] + offset
) << shift
;
2103 loads
[2] = (avenrun
[2] + offset
) << shift
;
2106 static long calc_load_fold_active(struct rq
*this_rq
)
2108 long nr_active
, delta
= 0;
2110 nr_active
= this_rq
->nr_running
;
2111 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2113 if (nr_active
!= this_rq
->calc_load_active
) {
2114 delta
= nr_active
- this_rq
->calc_load_active
;
2115 this_rq
->calc_load_active
= nr_active
;
2122 * a1 = a0 * e + a * (1 - e)
2124 static unsigned long
2125 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2128 load
+= active
* (FIXED_1
- exp
);
2129 load
+= 1UL << (FSHIFT
- 1);
2130 return load
>> FSHIFT
;
2135 * Handle NO_HZ for the global load-average.
2137 * Since the above described distributed algorithm to compute the global
2138 * load-average relies on per-cpu sampling from the tick, it is affected by
2141 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2142 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2143 * when we read the global state.
2145 * Obviously reality has to ruin such a delightfully simple scheme:
2147 * - When we go NO_HZ idle during the window, we can negate our sample
2148 * contribution, causing under-accounting.
2150 * We avoid this by keeping two idle-delta counters and flipping them
2151 * when the window starts, thus separating old and new NO_HZ load.
2153 * The only trick is the slight shift in index flip for read vs write.
2157 * |-|-----------|-|-----------|-|-----------|-|
2158 * r:0 0 1 1 0 0 1 1 0
2159 * w:0 1 1 0 0 1 1 0 0
2161 * This ensures we'll fold the old idle contribution in this window while
2162 * accumlating the new one.
2164 * - When we wake up from NO_HZ idle during the window, we push up our
2165 * contribution, since we effectively move our sample point to a known
2168 * This is solved by pushing the window forward, and thus skipping the
2169 * sample, for this cpu (effectively using the idle-delta for this cpu which
2170 * was in effect at the time the window opened). This also solves the issue
2171 * of having to deal with a cpu having been in NOHZ idle for multiple
2172 * LOAD_FREQ intervals.
2174 * When making the ILB scale, we should try to pull this in as well.
2176 static atomic_long_t calc_load_idle
[2];
2177 static int calc_load_idx
;
2179 static inline int calc_load_write_idx(void)
2181 int idx
= calc_load_idx
;
2184 * See calc_global_nohz(), if we observe the new index, we also
2185 * need to observe the new update time.
2190 * If the folding window started, make sure we start writing in the
2193 if (!time_before(jiffies
, calc_load_update
))
2199 static inline int calc_load_read_idx(void)
2201 return calc_load_idx
& 1;
2204 void calc_load_enter_idle(void)
2206 struct rq
*this_rq
= this_rq();
2210 * We're going into NOHZ mode, if there's any pending delta, fold it
2211 * into the pending idle delta.
2213 delta
= calc_load_fold_active(this_rq
);
2215 int idx
= calc_load_write_idx();
2216 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2220 void calc_load_exit_idle(void)
2222 struct rq
*this_rq
= this_rq();
2225 * If we're still before the sample window, we're done.
2227 if (time_before(jiffies
, this_rq
->calc_load_update
))
2231 * We woke inside or after the sample window, this means we're already
2232 * accounted through the nohz accounting, so skip the entire deal and
2233 * sync up for the next window.
2235 this_rq
->calc_load_update
= calc_load_update
;
2236 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2237 this_rq
->calc_load_update
+= LOAD_FREQ
;
2240 static long calc_load_fold_idle(void)
2242 int idx
= calc_load_read_idx();
2245 if (atomic_long_read(&calc_load_idle
[idx
]))
2246 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2252 * fixed_power_int - compute: x^n, in O(log n) time
2254 * @x: base of the power
2255 * @frac_bits: fractional bits of @x
2256 * @n: power to raise @x to.
2258 * By exploiting the relation between the definition of the natural power
2259 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2260 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2261 * (where: n_i \elem {0, 1}, the binary vector representing n),
2262 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2263 * of course trivially computable in O(log_2 n), the length of our binary
2266 static unsigned long
2267 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2269 unsigned long result
= 1UL << frac_bits
;
2274 result
+= 1UL << (frac_bits
- 1);
2275 result
>>= frac_bits
;
2281 x
+= 1UL << (frac_bits
- 1);
2289 * a1 = a0 * e + a * (1 - e)
2291 * a2 = a1 * e + a * (1 - e)
2292 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2293 * = a0 * e^2 + a * (1 - e) * (1 + e)
2295 * a3 = a2 * e + a * (1 - e)
2296 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2297 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2301 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2302 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2303 * = a0 * e^n + a * (1 - e^n)
2305 * [1] application of the geometric series:
2308 * S_n := \Sum x^i = -------------
2311 static unsigned long
2312 calc_load_n(unsigned long load
, unsigned long exp
,
2313 unsigned long active
, unsigned int n
)
2316 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2320 * NO_HZ can leave us missing all per-cpu ticks calling
2321 * calc_load_account_active(), but since an idle CPU folds its delta into
2322 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2323 * in the pending idle delta if our idle period crossed a load cycle boundary.
2325 * Once we've updated the global active value, we need to apply the exponential
2326 * weights adjusted to the number of cycles missed.
2328 static void calc_global_nohz(void)
2330 long delta
, active
, n
;
2332 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2334 * Catch-up, fold however many we are behind still
2336 delta
= jiffies
- calc_load_update
- 10;
2337 n
= 1 + (delta
/ LOAD_FREQ
);
2339 active
= atomic_long_read(&calc_load_tasks
);
2340 active
= active
> 0 ? active
* FIXED_1
: 0;
2342 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2343 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2344 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2346 calc_load_update
+= n
* LOAD_FREQ
;
2350 * Flip the idle index...
2352 * Make sure we first write the new time then flip the index, so that
2353 * calc_load_write_idx() will see the new time when it reads the new
2354 * index, this avoids a double flip messing things up.
2359 #else /* !CONFIG_NO_HZ */
2361 static inline long calc_load_fold_idle(void) { return 0; }
2362 static inline void calc_global_nohz(void) { }
2364 #endif /* CONFIG_NO_HZ */
2367 * calc_load - update the avenrun load estimates 10 ticks after the
2368 * CPUs have updated calc_load_tasks.
2370 void calc_global_load(unsigned long ticks
)
2374 if (time_before(jiffies
, calc_load_update
+ 10))
2378 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2380 delta
= calc_load_fold_idle();
2382 atomic_long_add(delta
, &calc_load_tasks
);
2384 active
= atomic_long_read(&calc_load_tasks
);
2385 active
= active
> 0 ? active
* FIXED_1
: 0;
2387 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2388 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2389 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2391 calc_load_update
+= LOAD_FREQ
;
2394 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2400 * Called from update_cpu_load() to periodically update this CPU's
2403 static void calc_load_account_active(struct rq
*this_rq
)
2407 if (time_before(jiffies
, this_rq
->calc_load_update
))
2410 delta
= calc_load_fold_active(this_rq
);
2412 atomic_long_add(delta
, &calc_load_tasks
);
2414 this_rq
->calc_load_update
+= LOAD_FREQ
;
2418 * End of global load-average stuff
2422 * The exact cpuload at various idx values, calculated at every tick would be
2423 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2425 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2426 * on nth tick when cpu may be busy, then we have:
2427 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2428 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2430 * decay_load_missed() below does efficient calculation of
2431 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2432 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2434 * The calculation is approximated on a 128 point scale.
2435 * degrade_zero_ticks is the number of ticks after which load at any
2436 * particular idx is approximated to be zero.
2437 * degrade_factor is a precomputed table, a row for each load idx.
2438 * Each column corresponds to degradation factor for a power of two ticks,
2439 * based on 128 point scale.
2441 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2442 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2444 * With this power of 2 load factors, we can degrade the load n times
2445 * by looking at 1 bits in n and doing as many mult/shift instead of
2446 * n mult/shifts needed by the exact degradation.
2448 #define DEGRADE_SHIFT 7
2449 static const unsigned char
2450 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2451 static const unsigned char
2452 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2453 {0, 0, 0, 0, 0, 0, 0, 0},
2454 {64, 32, 8, 0, 0, 0, 0, 0},
2455 {96, 72, 40, 12, 1, 0, 0},
2456 {112, 98, 75, 43, 15, 1, 0},
2457 {120, 112, 98, 76, 45, 16, 2} };
2460 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2461 * would be when CPU is idle and so we just decay the old load without
2462 * adding any new load.
2464 static unsigned long
2465 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2469 if (!missed_updates
)
2472 if (missed_updates
>= degrade_zero_ticks
[idx
])
2476 return load
>> missed_updates
;
2478 while (missed_updates
) {
2479 if (missed_updates
% 2)
2480 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2482 missed_updates
>>= 1;
2489 * Update rq->cpu_load[] statistics. This function is usually called every
2490 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2491 * every tick. We fix it up based on jiffies.
2493 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2494 unsigned long pending_updates
)
2498 this_rq
->nr_load_updates
++;
2500 /* Update our load: */
2501 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2502 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2503 unsigned long old_load
, new_load
;
2505 /* scale is effectively 1 << i now, and >> i divides by scale */
2507 old_load
= this_rq
->cpu_load
[i
];
2508 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2509 new_load
= this_load
;
2511 * Round up the averaging division if load is increasing. This
2512 * prevents us from getting stuck on 9 if the load is 10, for
2515 if (new_load
> old_load
)
2516 new_load
+= scale
- 1;
2518 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2521 sched_avg_update(this_rq
);
2526 * There is no sane way to deal with nohz on smp when using jiffies because the
2527 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2528 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2530 * Therefore we cannot use the delta approach from the regular tick since that
2531 * would seriously skew the load calculation. However we'll make do for those
2532 * updates happening while idle (nohz_idle_balance) or coming out of idle
2533 * (tick_nohz_idle_exit).
2535 * This means we might still be one tick off for nohz periods.
2539 * Called from nohz_idle_balance() to update the load ratings before doing the
2542 void update_idle_cpu_load(struct rq
*this_rq
)
2544 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2545 unsigned long load
= this_rq
->load
.weight
;
2546 unsigned long pending_updates
;
2549 * bail if there's load or we're actually up-to-date.
2551 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2554 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2555 this_rq
->last_load_update_tick
= curr_jiffies
;
2557 __update_cpu_load(this_rq
, load
, pending_updates
);
2561 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2563 void update_cpu_load_nohz(void)
2565 struct rq
*this_rq
= this_rq();
2566 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2567 unsigned long pending_updates
;
2569 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2572 raw_spin_lock(&this_rq
->lock
);
2573 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2574 if (pending_updates
) {
2575 this_rq
->last_load_update_tick
= curr_jiffies
;
2577 * We were idle, this means load 0, the current load might be
2578 * !0 due to remote wakeups and the sort.
2580 __update_cpu_load(this_rq
, 0, pending_updates
);
2582 raw_spin_unlock(&this_rq
->lock
);
2584 #endif /* CONFIG_NO_HZ */
2587 * Called from scheduler_tick()
2589 static void update_cpu_load_active(struct rq
*this_rq
)
2592 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2594 this_rq
->last_load_update_tick
= jiffies
;
2595 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2597 calc_load_account_active(this_rq
);
2603 * sched_exec - execve() is a valuable balancing opportunity, because at
2604 * this point the task has the smallest effective memory and cache footprint.
2606 void sched_exec(void)
2608 struct task_struct
*p
= current
;
2609 unsigned long flags
;
2612 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2613 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2614 if (dest_cpu
== smp_processor_id())
2617 if (likely(cpu_active(dest_cpu
))) {
2618 struct migration_arg arg
= { p
, dest_cpu
};
2620 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2621 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2625 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2630 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2631 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2633 EXPORT_PER_CPU_SYMBOL(kstat
);
2634 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2637 * Return any ns on the sched_clock that have not yet been accounted in
2638 * @p in case that task is currently running.
2640 * Called with task_rq_lock() held on @rq.
2642 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2646 if (task_current(rq
, p
)) {
2647 update_rq_clock(rq
);
2648 ns
= rq
->clock_task
- p
->se
.exec_start
;
2656 unsigned long long task_delta_exec(struct task_struct
*p
)
2658 unsigned long flags
;
2662 rq
= task_rq_lock(p
, &flags
);
2663 ns
= do_task_delta_exec(p
, rq
);
2664 task_rq_unlock(rq
, p
, &flags
);
2670 * Return accounted runtime for the task.
2671 * In case the task is currently running, return the runtime plus current's
2672 * pending runtime that have not been accounted yet.
2674 unsigned long long task_sched_runtime(struct task_struct
*p
)
2676 unsigned long flags
;
2680 rq
= task_rq_lock(p
, &flags
);
2681 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2682 task_rq_unlock(rq
, p
, &flags
);
2688 * This function gets called by the timer code, with HZ frequency.
2689 * We call it with interrupts disabled.
2691 void scheduler_tick(void)
2693 int cpu
= smp_processor_id();
2694 struct rq
*rq
= cpu_rq(cpu
);
2695 struct task_struct
*curr
= rq
->curr
;
2699 raw_spin_lock(&rq
->lock
);
2700 update_rq_clock(rq
);
2701 update_cpu_load_active(rq
);
2702 curr
->sched_class
->task_tick(rq
, curr
, 0);
2703 raw_spin_unlock(&rq
->lock
);
2705 perf_event_task_tick();
2708 rq
->idle_balance
= idle_cpu(cpu
);
2709 trigger_load_balance(rq
, cpu
);
2713 notrace
unsigned long get_parent_ip(unsigned long addr
)
2715 if (in_lock_functions(addr
)) {
2716 addr
= CALLER_ADDR2
;
2717 if (in_lock_functions(addr
))
2718 addr
= CALLER_ADDR3
;
2723 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2724 defined(CONFIG_PREEMPT_TRACER))
2726 void __kprobes
add_preempt_count(int val
)
2728 #ifdef CONFIG_DEBUG_PREEMPT
2732 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2735 preempt_count() += val
;
2736 #ifdef CONFIG_DEBUG_PREEMPT
2738 * Spinlock count overflowing soon?
2740 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2743 if (preempt_count() == val
)
2744 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2746 EXPORT_SYMBOL(add_preempt_count
);
2748 void __kprobes
sub_preempt_count(int val
)
2750 #ifdef CONFIG_DEBUG_PREEMPT
2754 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2757 * Is the spinlock portion underflowing?
2759 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2760 !(preempt_count() & PREEMPT_MASK
)))
2764 if (preempt_count() == val
)
2765 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2766 preempt_count() -= val
;
2768 EXPORT_SYMBOL(sub_preempt_count
);
2773 * Print scheduling while atomic bug:
2775 static noinline
void __schedule_bug(struct task_struct
*prev
)
2777 if (oops_in_progress
)
2780 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2781 prev
->comm
, prev
->pid
, preempt_count());
2783 debug_show_held_locks(prev
);
2785 if (irqs_disabled())
2786 print_irqtrace_events(prev
);
2788 add_taint(TAINT_WARN
);
2792 * Various schedule()-time debugging checks and statistics:
2794 static inline void schedule_debug(struct task_struct
*prev
)
2797 * Test if we are atomic. Since do_exit() needs to call into
2798 * schedule() atomically, we ignore that path for now.
2799 * Otherwise, whine if we are scheduling when we should not be.
2801 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2802 __schedule_bug(prev
);
2805 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2807 schedstat_inc(this_rq(), sched_count
);
2810 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2812 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2813 update_rq_clock(rq
);
2814 prev
->sched_class
->put_prev_task(rq
, prev
);
2818 * Pick up the highest-prio task:
2820 static inline struct task_struct
*
2821 pick_next_task(struct rq
*rq
)
2823 const struct sched_class
*class;
2824 struct task_struct
*p
;
2827 * Optimization: we know that if all tasks are in
2828 * the fair class we can call that function directly:
2830 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2831 p
= fair_sched_class
.pick_next_task(rq
);
2836 for_each_class(class) {
2837 p
= class->pick_next_task(rq
);
2842 BUG(); /* the idle class will always have a runnable task */
2846 * __schedule() is the main scheduler function.
2848 * The main means of driving the scheduler and thus entering this function are:
2850 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2852 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2853 * paths. For example, see arch/x86/entry_64.S.
2855 * To drive preemption between tasks, the scheduler sets the flag in timer
2856 * interrupt handler scheduler_tick().
2858 * 3. Wakeups don't really cause entry into schedule(). They add a
2859 * task to the run-queue and that's it.
2861 * Now, if the new task added to the run-queue preempts the current
2862 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2863 * called on the nearest possible occasion:
2865 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2867 * - in syscall or exception context, at the next outmost
2868 * preempt_enable(). (this might be as soon as the wake_up()'s
2871 * - in IRQ context, return from interrupt-handler to
2872 * preemptible context
2874 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2877 * - cond_resched() call
2878 * - explicit schedule() call
2879 * - return from syscall or exception to user-space
2880 * - return from interrupt-handler to user-space
2882 static void __sched
__schedule(void)
2884 struct task_struct
*prev
, *next
;
2885 unsigned long *switch_count
;
2891 cpu
= smp_processor_id();
2893 rcu_note_context_switch(cpu
);
2896 schedule_debug(prev
);
2898 if (sched_feat(HRTICK
))
2901 raw_spin_lock_irq(&rq
->lock
);
2903 switch_count
= &prev
->nivcsw
;
2904 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2905 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2906 prev
->state
= TASK_RUNNING
;
2908 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2912 * If a worker went to sleep, notify and ask workqueue
2913 * whether it wants to wake up a task to maintain
2916 if (prev
->flags
& PF_WQ_WORKER
) {
2917 struct task_struct
*to_wakeup
;
2919 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2921 try_to_wake_up_local(to_wakeup
);
2924 switch_count
= &prev
->nvcsw
;
2927 pre_schedule(rq
, prev
);
2929 if (unlikely(!rq
->nr_running
))
2930 idle_balance(cpu
, rq
);
2932 put_prev_task(rq
, prev
);
2933 next
= pick_next_task(rq
);
2934 clear_tsk_need_resched(prev
);
2935 rq
->skip_clock_update
= 0;
2937 if (likely(prev
!= next
)) {
2942 context_switch(rq
, prev
, next
); /* unlocks the rq */
2944 * The context switch have flipped the stack from under us
2945 * and restored the local variables which were saved when
2946 * this task called schedule() in the past. prev == current
2947 * is still correct, but it can be moved to another cpu/rq.
2949 cpu
= smp_processor_id();
2952 raw_spin_unlock_irq(&rq
->lock
);
2956 sched_preempt_enable_no_resched();
2961 static inline void sched_submit_work(struct task_struct
*tsk
)
2963 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2966 * If we are going to sleep and we have plugged IO queued,
2967 * make sure to submit it to avoid deadlocks.
2969 if (blk_needs_flush_plug(tsk
))
2970 blk_schedule_flush_plug(tsk
);
2973 asmlinkage
void __sched
schedule(void)
2975 struct task_struct
*tsk
= current
;
2977 sched_submit_work(tsk
);
2980 EXPORT_SYMBOL(schedule
);
2982 #ifdef CONFIG_CONTEXT_TRACKING
2983 asmlinkage
void __sched
schedule_user(void)
2986 * If we come here after a random call to set_need_resched(),
2987 * or we have been woken up remotely but the IPI has not yet arrived,
2988 * we haven't yet exited the RCU idle mode. Do it here manually until
2989 * we find a better solution.
2998 * schedule_preempt_disabled - called with preemption disabled
3000 * Returns with preemption disabled. Note: preempt_count must be 1
3002 void __sched
schedule_preempt_disabled(void)
3004 sched_preempt_enable_no_resched();
3009 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3011 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3013 if (lock
->owner
!= owner
)
3017 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3018 * lock->owner still matches owner, if that fails, owner might
3019 * point to free()d memory, if it still matches, the rcu_read_lock()
3020 * ensures the memory stays valid.
3024 return owner
->on_cpu
;
3028 * Look out! "owner" is an entirely speculative pointer
3029 * access and not reliable.
3031 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3033 if (!sched_feat(OWNER_SPIN
))
3037 while (owner_running(lock
, owner
)) {
3041 arch_mutex_cpu_relax();
3046 * We break out the loop above on need_resched() and when the
3047 * owner changed, which is a sign for heavy contention. Return
3048 * success only when lock->owner is NULL.
3050 return lock
->owner
== NULL
;
3054 #ifdef CONFIG_PREEMPT
3056 * this is the entry point to schedule() from in-kernel preemption
3057 * off of preempt_enable. Kernel preemptions off return from interrupt
3058 * occur there and call schedule directly.
3060 asmlinkage
void __sched notrace
preempt_schedule(void)
3062 struct thread_info
*ti
= current_thread_info();
3065 * If there is a non-zero preempt_count or interrupts are disabled,
3066 * we do not want to preempt the current task. Just return..
3068 if (likely(ti
->preempt_count
|| irqs_disabled()))
3072 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3074 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3077 * Check again in case we missed a preemption opportunity
3078 * between schedule and now.
3081 } while (need_resched());
3083 EXPORT_SYMBOL(preempt_schedule
);
3086 * this is the entry point to schedule() from kernel preemption
3087 * off of irq context.
3088 * Note, that this is called and return with irqs disabled. This will
3089 * protect us against recursive calling from irq.
3091 asmlinkage
void __sched
preempt_schedule_irq(void)
3093 struct thread_info
*ti
= current_thread_info();
3095 /* Catch callers which need to be fixed */
3096 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3100 add_preempt_count(PREEMPT_ACTIVE
);
3103 local_irq_disable();
3104 sub_preempt_count(PREEMPT_ACTIVE
);
3107 * Check again in case we missed a preemption opportunity
3108 * between schedule and now.
3111 } while (need_resched());
3114 #endif /* CONFIG_PREEMPT */
3116 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3119 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3121 EXPORT_SYMBOL(default_wake_function
);
3124 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3125 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3126 * number) then we wake all the non-exclusive tasks and one exclusive task.
3128 * There are circumstances in which we can try to wake a task which has already
3129 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3130 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3132 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3133 int nr_exclusive
, int wake_flags
, void *key
)
3135 wait_queue_t
*curr
, *next
;
3137 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3138 unsigned flags
= curr
->flags
;
3140 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3141 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3147 * __wake_up - wake up threads blocked on a waitqueue.
3149 * @mode: which threads
3150 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3151 * @key: is directly passed to the wakeup function
3153 * It may be assumed that this function implies a write memory barrier before
3154 * changing the task state if and only if any tasks are woken up.
3156 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3157 int nr_exclusive
, void *key
)
3159 unsigned long flags
;
3161 spin_lock_irqsave(&q
->lock
, flags
);
3162 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3163 spin_unlock_irqrestore(&q
->lock
, flags
);
3165 EXPORT_SYMBOL(__wake_up
);
3168 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3170 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3172 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3174 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3176 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3178 __wake_up_common(q
, mode
, 1, 0, key
);
3180 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3183 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3185 * @mode: which threads
3186 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3187 * @key: opaque value to be passed to wakeup targets
3189 * The sync wakeup differs that the waker knows that it will schedule
3190 * away soon, so while the target thread will be woken up, it will not
3191 * be migrated to another CPU - ie. the two threads are 'synchronized'
3192 * with each other. This can prevent needless bouncing between CPUs.
3194 * On UP it can prevent extra preemption.
3196 * It may be assumed that this function implies a write memory barrier before
3197 * changing the task state if and only if any tasks are woken up.
3199 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3200 int nr_exclusive
, void *key
)
3202 unsigned long flags
;
3203 int wake_flags
= WF_SYNC
;
3208 if (unlikely(!nr_exclusive
))
3211 spin_lock_irqsave(&q
->lock
, flags
);
3212 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3213 spin_unlock_irqrestore(&q
->lock
, flags
);
3215 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3218 * __wake_up_sync - see __wake_up_sync_key()
3220 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3222 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3224 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3227 * complete: - signals a single thread waiting on this completion
3228 * @x: holds the state of this particular completion
3230 * This will wake up a single thread waiting on this completion. Threads will be
3231 * awakened in the same order in which they were queued.
3233 * See also complete_all(), wait_for_completion() and related routines.
3235 * It may be assumed that this function implies a write memory barrier before
3236 * changing the task state if and only if any tasks are woken up.
3238 void complete(struct completion
*x
)
3240 unsigned long flags
;
3242 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3244 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3245 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3247 EXPORT_SYMBOL(complete
);
3250 * complete_all: - signals all threads waiting on this completion
3251 * @x: holds the state of this particular completion
3253 * This will wake up all threads waiting on this particular completion event.
3255 * It may be assumed that this function implies a write memory barrier before
3256 * changing the task state if and only if any tasks are woken up.
3258 void complete_all(struct completion
*x
)
3260 unsigned long flags
;
3262 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3263 x
->done
+= UINT_MAX
/2;
3264 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3265 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3267 EXPORT_SYMBOL(complete_all
);
3269 static inline long __sched
3270 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3273 DECLARE_WAITQUEUE(wait
, current
);
3275 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3277 if (signal_pending_state(state
, current
)) {
3278 timeout
= -ERESTARTSYS
;
3281 __set_current_state(state
);
3282 spin_unlock_irq(&x
->wait
.lock
);
3283 timeout
= schedule_timeout(timeout
);
3284 spin_lock_irq(&x
->wait
.lock
);
3285 } while (!x
->done
&& timeout
);
3286 __remove_wait_queue(&x
->wait
, &wait
);
3291 return timeout
?: 1;
3295 wait_for_common(struct completion
*x
, long timeout
, int state
)
3299 spin_lock_irq(&x
->wait
.lock
);
3300 timeout
= do_wait_for_common(x
, timeout
, state
);
3301 spin_unlock_irq(&x
->wait
.lock
);
3306 * wait_for_completion: - waits for completion of a task
3307 * @x: holds the state of this particular completion
3309 * This waits to be signaled for completion of a specific task. It is NOT
3310 * interruptible and there is no timeout.
3312 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3313 * and interrupt capability. Also see complete().
3315 void __sched
wait_for_completion(struct completion
*x
)
3317 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3319 EXPORT_SYMBOL(wait_for_completion
);
3322 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3323 * @x: holds the state of this particular completion
3324 * @timeout: timeout value in jiffies
3326 * This waits for either a completion of a specific task to be signaled or for a
3327 * specified timeout to expire. The timeout is in jiffies. It is not
3330 * The return value is 0 if timed out, and positive (at least 1, or number of
3331 * jiffies left till timeout) if completed.
3333 unsigned long __sched
3334 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3336 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3338 EXPORT_SYMBOL(wait_for_completion_timeout
);
3341 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3342 * @x: holds the state of this particular completion
3344 * This waits for completion of a specific task to be signaled. It is
3347 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3349 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3351 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3352 if (t
== -ERESTARTSYS
)
3356 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3359 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3360 * @x: holds the state of this particular completion
3361 * @timeout: timeout value in jiffies
3363 * This waits for either a completion of a specific task to be signaled or for a
3364 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3366 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3367 * positive (at least 1, or number of jiffies left till timeout) if completed.
3370 wait_for_completion_interruptible_timeout(struct completion
*x
,
3371 unsigned long timeout
)
3373 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3375 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3378 * wait_for_completion_killable: - waits for completion of a task (killable)
3379 * @x: holds the state of this particular completion
3381 * This waits to be signaled for completion of a specific task. It can be
3382 * interrupted by a kill signal.
3384 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3386 int __sched
wait_for_completion_killable(struct completion
*x
)
3388 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3389 if (t
== -ERESTARTSYS
)
3393 EXPORT_SYMBOL(wait_for_completion_killable
);
3396 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3397 * @x: holds the state of this particular completion
3398 * @timeout: timeout value in jiffies
3400 * This waits for either a completion of a specific task to be
3401 * signaled or for a specified timeout to expire. It can be
3402 * interrupted by a kill signal. The timeout is in jiffies.
3404 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3405 * positive (at least 1, or number of jiffies left till timeout) if completed.
3408 wait_for_completion_killable_timeout(struct completion
*x
,
3409 unsigned long timeout
)
3411 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3413 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3416 * try_wait_for_completion - try to decrement a completion without blocking
3417 * @x: completion structure
3419 * Returns: 0 if a decrement cannot be done without blocking
3420 * 1 if a decrement succeeded.
3422 * If a completion is being used as a counting completion,
3423 * attempt to decrement the counter without blocking. This
3424 * enables us to avoid waiting if the resource the completion
3425 * is protecting is not available.
3427 bool try_wait_for_completion(struct completion
*x
)
3429 unsigned long flags
;
3432 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3437 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3440 EXPORT_SYMBOL(try_wait_for_completion
);
3443 * completion_done - Test to see if a completion has any waiters
3444 * @x: completion structure
3446 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3447 * 1 if there are no waiters.
3450 bool completion_done(struct completion
*x
)
3452 unsigned long flags
;
3455 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3458 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3461 EXPORT_SYMBOL(completion_done
);
3464 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3466 unsigned long flags
;
3469 init_waitqueue_entry(&wait
, current
);
3471 __set_current_state(state
);
3473 spin_lock_irqsave(&q
->lock
, flags
);
3474 __add_wait_queue(q
, &wait
);
3475 spin_unlock(&q
->lock
);
3476 timeout
= schedule_timeout(timeout
);
3477 spin_lock_irq(&q
->lock
);
3478 __remove_wait_queue(q
, &wait
);
3479 spin_unlock_irqrestore(&q
->lock
, flags
);
3484 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3486 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3488 EXPORT_SYMBOL(interruptible_sleep_on
);
3491 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3493 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3495 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3497 void __sched
sleep_on(wait_queue_head_t
*q
)
3499 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3501 EXPORT_SYMBOL(sleep_on
);
3503 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3505 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3507 EXPORT_SYMBOL(sleep_on_timeout
);
3509 #ifdef CONFIG_RT_MUTEXES
3512 * rt_mutex_setprio - set the current priority of a task
3514 * @prio: prio value (kernel-internal form)
3516 * This function changes the 'effective' priority of a task. It does
3517 * not touch ->normal_prio like __setscheduler().
3519 * Used by the rt_mutex code to implement priority inheritance logic.
3521 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3523 int oldprio
, on_rq
, running
;
3525 const struct sched_class
*prev_class
;
3527 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3529 rq
= __task_rq_lock(p
);
3532 * Idle task boosting is a nono in general. There is one
3533 * exception, when PREEMPT_RT and NOHZ is active:
3535 * The idle task calls get_next_timer_interrupt() and holds
3536 * the timer wheel base->lock on the CPU and another CPU wants
3537 * to access the timer (probably to cancel it). We can safely
3538 * ignore the boosting request, as the idle CPU runs this code
3539 * with interrupts disabled and will complete the lock
3540 * protected section without being interrupted. So there is no
3541 * real need to boost.
3543 if (unlikely(p
== rq
->idle
)) {
3544 WARN_ON(p
!= rq
->curr
);
3545 WARN_ON(p
->pi_blocked_on
);
3549 trace_sched_pi_setprio(p
, prio
);
3551 prev_class
= p
->sched_class
;
3553 running
= task_current(rq
, p
);
3555 dequeue_task(rq
, p
, 0);
3557 p
->sched_class
->put_prev_task(rq
, p
);
3560 p
->sched_class
= &rt_sched_class
;
3562 p
->sched_class
= &fair_sched_class
;
3567 p
->sched_class
->set_curr_task(rq
);
3569 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3571 check_class_changed(rq
, p
, prev_class
, oldprio
);
3573 __task_rq_unlock(rq
);
3576 void set_user_nice(struct task_struct
*p
, long nice
)
3578 int old_prio
, delta
, on_rq
;
3579 unsigned long flags
;
3582 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3585 * We have to be careful, if called from sys_setpriority(),
3586 * the task might be in the middle of scheduling on another CPU.
3588 rq
= task_rq_lock(p
, &flags
);
3590 * The RT priorities are set via sched_setscheduler(), but we still
3591 * allow the 'normal' nice value to be set - but as expected
3592 * it wont have any effect on scheduling until the task is
3593 * SCHED_FIFO/SCHED_RR:
3595 if (task_has_rt_policy(p
)) {
3596 p
->static_prio
= NICE_TO_PRIO(nice
);
3601 dequeue_task(rq
, p
, 0);
3603 p
->static_prio
= NICE_TO_PRIO(nice
);
3606 p
->prio
= effective_prio(p
);
3607 delta
= p
->prio
- old_prio
;
3610 enqueue_task(rq
, p
, 0);
3612 * If the task increased its priority or is running and
3613 * lowered its priority, then reschedule its CPU:
3615 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3616 resched_task(rq
->curr
);
3619 task_rq_unlock(rq
, p
, &flags
);
3621 EXPORT_SYMBOL(set_user_nice
);
3624 * can_nice - check if a task can reduce its nice value
3628 int can_nice(const struct task_struct
*p
, const int nice
)
3630 /* convert nice value [19,-20] to rlimit style value [1,40] */
3631 int nice_rlim
= 20 - nice
;
3633 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3634 capable(CAP_SYS_NICE
));
3637 #ifdef __ARCH_WANT_SYS_NICE
3640 * sys_nice - change the priority of the current process.
3641 * @increment: priority increment
3643 * sys_setpriority is a more generic, but much slower function that
3644 * does similar things.
3646 SYSCALL_DEFINE1(nice
, int, increment
)
3651 * Setpriority might change our priority at the same moment.
3652 * We don't have to worry. Conceptually one call occurs first
3653 * and we have a single winner.
3655 if (increment
< -40)
3660 nice
= TASK_NICE(current
) + increment
;
3666 if (increment
< 0 && !can_nice(current
, nice
))
3669 retval
= security_task_setnice(current
, nice
);
3673 set_user_nice(current
, nice
);
3680 * task_prio - return the priority value of a given task.
3681 * @p: the task in question.
3683 * This is the priority value as seen by users in /proc.
3684 * RT tasks are offset by -200. Normal tasks are centered
3685 * around 0, value goes from -16 to +15.
3687 int task_prio(const struct task_struct
*p
)
3689 return p
->prio
- MAX_RT_PRIO
;
3693 * task_nice - return the nice value of a given task.
3694 * @p: the task in question.
3696 int task_nice(const struct task_struct
*p
)
3698 return TASK_NICE(p
);
3700 EXPORT_SYMBOL(task_nice
);
3703 * idle_cpu - is a given cpu idle currently?
3704 * @cpu: the processor in question.
3706 int idle_cpu(int cpu
)
3708 struct rq
*rq
= cpu_rq(cpu
);
3710 if (rq
->curr
!= rq
->idle
)
3717 if (!llist_empty(&rq
->wake_list
))
3725 * idle_task - return the idle task for a given cpu.
3726 * @cpu: the processor in question.
3728 struct task_struct
*idle_task(int cpu
)
3730 return cpu_rq(cpu
)->idle
;
3734 * find_process_by_pid - find a process with a matching PID value.
3735 * @pid: the pid in question.
3737 static struct task_struct
*find_process_by_pid(pid_t pid
)
3739 return pid
? find_task_by_vpid(pid
) : current
;
3742 /* Actually do priority change: must hold rq lock. */
3744 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3747 p
->rt_priority
= prio
;
3748 p
->normal_prio
= normal_prio(p
);
3749 /* we are holding p->pi_lock already */
3750 p
->prio
= rt_mutex_getprio(p
);
3751 if (rt_prio(p
->prio
))
3752 p
->sched_class
= &rt_sched_class
;
3754 p
->sched_class
= &fair_sched_class
;
3759 * check the target process has a UID that matches the current process's
3761 static bool check_same_owner(struct task_struct
*p
)
3763 const struct cred
*cred
= current_cred(), *pcred
;
3767 pcred
= __task_cred(p
);
3768 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3769 uid_eq(cred
->euid
, pcred
->uid
));
3774 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3775 const struct sched_param
*param
, bool user
)
3777 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3778 unsigned long flags
;
3779 const struct sched_class
*prev_class
;
3783 /* may grab non-irq protected spin_locks */
3784 BUG_ON(in_interrupt());
3786 /* double check policy once rq lock held */
3788 reset_on_fork
= p
->sched_reset_on_fork
;
3789 policy
= oldpolicy
= p
->policy
;
3791 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3792 policy
&= ~SCHED_RESET_ON_FORK
;
3794 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3795 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3796 policy
!= SCHED_IDLE
)
3801 * Valid priorities for SCHED_FIFO and SCHED_RR are
3802 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3803 * SCHED_BATCH and SCHED_IDLE is 0.
3805 if (param
->sched_priority
< 0 ||
3806 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3807 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3809 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3813 * Allow unprivileged RT tasks to decrease priority:
3815 if (user
&& !capable(CAP_SYS_NICE
)) {
3816 if (rt_policy(policy
)) {
3817 unsigned long rlim_rtprio
=
3818 task_rlimit(p
, RLIMIT_RTPRIO
);
3820 /* can't set/change the rt policy */
3821 if (policy
!= p
->policy
&& !rlim_rtprio
)
3824 /* can't increase priority */
3825 if (param
->sched_priority
> p
->rt_priority
&&
3826 param
->sched_priority
> rlim_rtprio
)
3831 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3832 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3834 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3835 if (!can_nice(p
, TASK_NICE(p
)))
3839 /* can't change other user's priorities */
3840 if (!check_same_owner(p
))
3843 /* Normal users shall not reset the sched_reset_on_fork flag */
3844 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3849 retval
= security_task_setscheduler(p
);
3855 * make sure no PI-waiters arrive (or leave) while we are
3856 * changing the priority of the task:
3858 * To be able to change p->policy safely, the appropriate
3859 * runqueue lock must be held.
3861 rq
= task_rq_lock(p
, &flags
);
3864 * Changing the policy of the stop threads its a very bad idea
3866 if (p
== rq
->stop
) {
3867 task_rq_unlock(rq
, p
, &flags
);
3872 * If not changing anything there's no need to proceed further:
3874 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3875 param
->sched_priority
== p
->rt_priority
))) {
3876 task_rq_unlock(rq
, p
, &flags
);
3880 #ifdef CONFIG_RT_GROUP_SCHED
3883 * Do not allow realtime tasks into groups that have no runtime
3886 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3887 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3888 !task_group_is_autogroup(task_group(p
))) {
3889 task_rq_unlock(rq
, p
, &flags
);
3895 /* recheck policy now with rq lock held */
3896 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3897 policy
= oldpolicy
= -1;
3898 task_rq_unlock(rq
, p
, &flags
);
3902 running
= task_current(rq
, p
);
3904 dequeue_task(rq
, p
, 0);
3906 p
->sched_class
->put_prev_task(rq
, p
);
3908 p
->sched_reset_on_fork
= reset_on_fork
;
3911 prev_class
= p
->sched_class
;
3912 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3915 p
->sched_class
->set_curr_task(rq
);
3917 enqueue_task(rq
, p
, 0);
3919 check_class_changed(rq
, p
, prev_class
, oldprio
);
3920 task_rq_unlock(rq
, p
, &flags
);
3922 rt_mutex_adjust_pi(p
);
3928 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3929 * @p: the task in question.
3930 * @policy: new policy.
3931 * @param: structure containing the new RT priority.
3933 * NOTE that the task may be already dead.
3935 int sched_setscheduler(struct task_struct
*p
, int policy
,
3936 const struct sched_param
*param
)
3938 return __sched_setscheduler(p
, policy
, param
, true);
3940 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3943 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3944 * @p: the task in question.
3945 * @policy: new policy.
3946 * @param: structure containing the new RT priority.
3948 * Just like sched_setscheduler, only don't bother checking if the
3949 * current context has permission. For example, this is needed in
3950 * stop_machine(): we create temporary high priority worker threads,
3951 * but our caller might not have that capability.
3953 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3954 const struct sched_param
*param
)
3956 return __sched_setscheduler(p
, policy
, param
, false);
3960 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3962 struct sched_param lparam
;
3963 struct task_struct
*p
;
3966 if (!param
|| pid
< 0)
3968 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3973 p
= find_process_by_pid(pid
);
3975 retval
= sched_setscheduler(p
, policy
, &lparam
);
3982 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3983 * @pid: the pid in question.
3984 * @policy: new policy.
3985 * @param: structure containing the new RT priority.
3987 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3988 struct sched_param __user
*, param
)
3990 /* negative values for policy are not valid */
3994 return do_sched_setscheduler(pid
, policy
, param
);
3998 * sys_sched_setparam - set/change the RT priority of a thread
3999 * @pid: the pid in question.
4000 * @param: structure containing the new RT priority.
4002 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4004 return do_sched_setscheduler(pid
, -1, param
);
4008 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4009 * @pid: the pid in question.
4011 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4013 struct task_struct
*p
;
4021 p
= find_process_by_pid(pid
);
4023 retval
= security_task_getscheduler(p
);
4026 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4033 * sys_sched_getparam - get the RT priority of a thread
4034 * @pid: the pid in question.
4035 * @param: structure containing the RT priority.
4037 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4039 struct sched_param lp
;
4040 struct task_struct
*p
;
4043 if (!param
|| pid
< 0)
4047 p
= find_process_by_pid(pid
);
4052 retval
= security_task_getscheduler(p
);
4056 lp
.sched_priority
= p
->rt_priority
;
4060 * This one might sleep, we cannot do it with a spinlock held ...
4062 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4071 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4073 cpumask_var_t cpus_allowed
, new_mask
;
4074 struct task_struct
*p
;
4080 p
= find_process_by_pid(pid
);
4087 /* Prevent p going away */
4091 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4095 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4097 goto out_free_cpus_allowed
;
4100 if (!check_same_owner(p
)) {
4102 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4109 retval
= security_task_setscheduler(p
);
4113 cpuset_cpus_allowed(p
, cpus_allowed
);
4114 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4116 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4119 cpuset_cpus_allowed(p
, cpus_allowed
);
4120 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4122 * We must have raced with a concurrent cpuset
4123 * update. Just reset the cpus_allowed to the
4124 * cpuset's cpus_allowed
4126 cpumask_copy(new_mask
, cpus_allowed
);
4131 free_cpumask_var(new_mask
);
4132 out_free_cpus_allowed
:
4133 free_cpumask_var(cpus_allowed
);
4140 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4141 struct cpumask
*new_mask
)
4143 if (len
< cpumask_size())
4144 cpumask_clear(new_mask
);
4145 else if (len
> cpumask_size())
4146 len
= cpumask_size();
4148 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4152 * sys_sched_setaffinity - set the cpu affinity of a process
4153 * @pid: pid of the process
4154 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4155 * @user_mask_ptr: user-space pointer to the new cpu mask
4157 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4158 unsigned long __user
*, user_mask_ptr
)
4160 cpumask_var_t new_mask
;
4163 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4166 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4168 retval
= sched_setaffinity(pid
, new_mask
);
4169 free_cpumask_var(new_mask
);
4173 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4175 struct task_struct
*p
;
4176 unsigned long flags
;
4183 p
= find_process_by_pid(pid
);
4187 retval
= security_task_getscheduler(p
);
4191 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4192 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4193 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4203 * sys_sched_getaffinity - get the cpu affinity of a process
4204 * @pid: pid of the process
4205 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4206 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4208 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4209 unsigned long __user
*, user_mask_ptr
)
4214 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4216 if (len
& (sizeof(unsigned long)-1))
4219 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4222 ret
= sched_getaffinity(pid
, mask
);
4224 size_t retlen
= min_t(size_t, len
, cpumask_size());
4226 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4231 free_cpumask_var(mask
);
4237 * sys_sched_yield - yield the current processor to other threads.
4239 * This function yields the current CPU to other tasks. If there are no
4240 * other threads running on this CPU then this function will return.
4242 SYSCALL_DEFINE0(sched_yield
)
4244 struct rq
*rq
= this_rq_lock();
4246 schedstat_inc(rq
, yld_count
);
4247 current
->sched_class
->yield_task(rq
);
4250 * Since we are going to call schedule() anyway, there's
4251 * no need to preempt or enable interrupts:
4253 __release(rq
->lock
);
4254 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4255 do_raw_spin_unlock(&rq
->lock
);
4256 sched_preempt_enable_no_resched();
4263 static inline int should_resched(void)
4265 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4268 static void __cond_resched(void)
4270 add_preempt_count(PREEMPT_ACTIVE
);
4272 sub_preempt_count(PREEMPT_ACTIVE
);
4275 int __sched
_cond_resched(void)
4277 if (should_resched()) {
4283 EXPORT_SYMBOL(_cond_resched
);
4286 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4287 * call schedule, and on return reacquire the lock.
4289 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4290 * operations here to prevent schedule() from being called twice (once via
4291 * spin_unlock(), once by hand).
4293 int __cond_resched_lock(spinlock_t
*lock
)
4295 int resched
= should_resched();
4298 lockdep_assert_held(lock
);
4300 if (spin_needbreak(lock
) || resched
) {
4311 EXPORT_SYMBOL(__cond_resched_lock
);
4313 int __sched
__cond_resched_softirq(void)
4315 BUG_ON(!in_softirq());
4317 if (should_resched()) {
4325 EXPORT_SYMBOL(__cond_resched_softirq
);
4328 * yield - yield the current processor to other threads.
4330 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4332 * The scheduler is at all times free to pick the calling task as the most
4333 * eligible task to run, if removing the yield() call from your code breaks
4334 * it, its already broken.
4336 * Typical broken usage is:
4341 * where one assumes that yield() will let 'the other' process run that will
4342 * make event true. If the current task is a SCHED_FIFO task that will never
4343 * happen. Never use yield() as a progress guarantee!!
4345 * If you want to use yield() to wait for something, use wait_event().
4346 * If you want to use yield() to be 'nice' for others, use cond_resched().
4347 * If you still want to use yield(), do not!
4349 void __sched
yield(void)
4351 set_current_state(TASK_RUNNING
);
4354 EXPORT_SYMBOL(yield
);
4357 * yield_to - yield the current processor to another thread in
4358 * your thread group, or accelerate that thread toward the
4359 * processor it's on.
4361 * @preempt: whether task preemption is allowed or not
4363 * It's the caller's job to ensure that the target task struct
4364 * can't go away on us before we can do any checks.
4366 * Returns true if we indeed boosted the target task.
4368 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4370 struct task_struct
*curr
= current
;
4371 struct rq
*rq
, *p_rq
;
4372 unsigned long flags
;
4375 local_irq_save(flags
);
4380 double_rq_lock(rq
, p_rq
);
4381 while (task_rq(p
) != p_rq
) {
4382 double_rq_unlock(rq
, p_rq
);
4386 if (!curr
->sched_class
->yield_to_task
)
4389 if (curr
->sched_class
!= p
->sched_class
)
4392 if (task_running(p_rq
, p
) || p
->state
)
4395 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4397 schedstat_inc(rq
, yld_count
);
4399 * Make p's CPU reschedule; pick_next_entity takes care of
4402 if (preempt
&& rq
!= p_rq
)
4403 resched_task(p_rq
->curr
);
4407 double_rq_unlock(rq
, p_rq
);
4408 local_irq_restore(flags
);
4415 EXPORT_SYMBOL_GPL(yield_to
);
4418 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4419 * that process accounting knows that this is a task in IO wait state.
4421 void __sched
io_schedule(void)
4423 struct rq
*rq
= raw_rq();
4425 delayacct_blkio_start();
4426 atomic_inc(&rq
->nr_iowait
);
4427 blk_flush_plug(current
);
4428 current
->in_iowait
= 1;
4430 current
->in_iowait
= 0;
4431 atomic_dec(&rq
->nr_iowait
);
4432 delayacct_blkio_end();
4434 EXPORT_SYMBOL(io_schedule
);
4436 long __sched
io_schedule_timeout(long timeout
)
4438 struct rq
*rq
= raw_rq();
4441 delayacct_blkio_start();
4442 atomic_inc(&rq
->nr_iowait
);
4443 blk_flush_plug(current
);
4444 current
->in_iowait
= 1;
4445 ret
= schedule_timeout(timeout
);
4446 current
->in_iowait
= 0;
4447 atomic_dec(&rq
->nr_iowait
);
4448 delayacct_blkio_end();
4453 * sys_sched_get_priority_max - return maximum RT priority.
4454 * @policy: scheduling class.
4456 * this syscall returns the maximum rt_priority that can be used
4457 * by a given scheduling class.
4459 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4466 ret
= MAX_USER_RT_PRIO
-1;
4478 * sys_sched_get_priority_min - return minimum RT priority.
4479 * @policy: scheduling class.
4481 * this syscall returns the minimum rt_priority that can be used
4482 * by a given scheduling class.
4484 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4502 * sys_sched_rr_get_interval - return the default timeslice of a process.
4503 * @pid: pid of the process.
4504 * @interval: userspace pointer to the timeslice value.
4506 * this syscall writes the default timeslice value of a given process
4507 * into the user-space timespec buffer. A value of '0' means infinity.
4509 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4510 struct timespec __user
*, interval
)
4512 struct task_struct
*p
;
4513 unsigned int time_slice
;
4514 unsigned long flags
;
4524 p
= find_process_by_pid(pid
);
4528 retval
= security_task_getscheduler(p
);
4532 rq
= task_rq_lock(p
, &flags
);
4533 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4534 task_rq_unlock(rq
, p
, &flags
);
4537 jiffies_to_timespec(time_slice
, &t
);
4538 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4546 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4548 void sched_show_task(struct task_struct
*p
)
4550 unsigned long free
= 0;
4554 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4555 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4556 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4557 #if BITS_PER_LONG == 32
4558 if (state
== TASK_RUNNING
)
4559 printk(KERN_CONT
" running ");
4561 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4563 if (state
== TASK_RUNNING
)
4564 printk(KERN_CONT
" running task ");
4566 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4568 #ifdef CONFIG_DEBUG_STACK_USAGE
4569 free
= stack_not_used(p
);
4572 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4574 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4575 task_pid_nr(p
), ppid
,
4576 (unsigned long)task_thread_info(p
)->flags
);
4578 show_stack(p
, NULL
);
4581 void show_state_filter(unsigned long state_filter
)
4583 struct task_struct
*g
, *p
;
4585 #if BITS_PER_LONG == 32
4587 " task PC stack pid father\n");
4590 " task PC stack pid father\n");
4593 do_each_thread(g
, p
) {
4595 * reset the NMI-timeout, listing all files on a slow
4596 * console might take a lot of time:
4598 touch_nmi_watchdog();
4599 if (!state_filter
|| (p
->state
& state_filter
))
4601 } while_each_thread(g
, p
);
4603 touch_all_softlockup_watchdogs();
4605 #ifdef CONFIG_SCHED_DEBUG
4606 sysrq_sched_debug_show();
4610 * Only show locks if all tasks are dumped:
4613 debug_show_all_locks();
4616 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4618 idle
->sched_class
= &idle_sched_class
;
4622 * init_idle - set up an idle thread for a given CPU
4623 * @idle: task in question
4624 * @cpu: cpu the idle task belongs to
4626 * NOTE: this function does not set the idle thread's NEED_RESCHED
4627 * flag, to make booting more robust.
4629 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4631 struct rq
*rq
= cpu_rq(cpu
);
4632 unsigned long flags
;
4634 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4637 idle
->state
= TASK_RUNNING
;
4638 idle
->se
.exec_start
= sched_clock();
4640 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4642 * We're having a chicken and egg problem, even though we are
4643 * holding rq->lock, the cpu isn't yet set to this cpu so the
4644 * lockdep check in task_group() will fail.
4646 * Similar case to sched_fork(). / Alternatively we could
4647 * use task_rq_lock() here and obtain the other rq->lock.
4652 __set_task_cpu(idle
, cpu
);
4655 rq
->curr
= rq
->idle
= idle
;
4656 #if defined(CONFIG_SMP)
4659 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4661 /* Set the preempt count _outside_ the spinlocks! */
4662 task_thread_info(idle
)->preempt_count
= 0;
4665 * The idle tasks have their own, simple scheduling class:
4667 idle
->sched_class
= &idle_sched_class
;
4668 ftrace_graph_init_idle_task(idle
, cpu
);
4669 #if defined(CONFIG_SMP)
4670 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4675 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4677 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4678 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4680 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4681 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4685 * This is how migration works:
4687 * 1) we invoke migration_cpu_stop() on the target CPU using
4689 * 2) stopper starts to run (implicitly forcing the migrated thread
4691 * 3) it checks whether the migrated task is still in the wrong runqueue.
4692 * 4) if it's in the wrong runqueue then the migration thread removes
4693 * it and puts it into the right queue.
4694 * 5) stopper completes and stop_one_cpu() returns and the migration
4699 * Change a given task's CPU affinity. Migrate the thread to a
4700 * proper CPU and schedule it away if the CPU it's executing on
4701 * is removed from the allowed bitmask.
4703 * NOTE: the caller must have a valid reference to the task, the
4704 * task must not exit() & deallocate itself prematurely. The
4705 * call is not atomic; no spinlocks may be held.
4707 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4709 unsigned long flags
;
4711 unsigned int dest_cpu
;
4714 rq
= task_rq_lock(p
, &flags
);
4716 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4719 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4724 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4729 do_set_cpus_allowed(p
, new_mask
);
4731 /* Can the task run on the task's current CPU? If so, we're done */
4732 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4735 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4737 struct migration_arg arg
= { p
, dest_cpu
};
4738 /* Need help from migration thread: drop lock and wait. */
4739 task_rq_unlock(rq
, p
, &flags
);
4740 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4741 tlb_migrate_finish(p
->mm
);
4745 task_rq_unlock(rq
, p
, &flags
);
4749 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4752 * Move (not current) task off this cpu, onto dest cpu. We're doing
4753 * this because either it can't run here any more (set_cpus_allowed()
4754 * away from this CPU, or CPU going down), or because we're
4755 * attempting to rebalance this task on exec (sched_exec).
4757 * So we race with normal scheduler movements, but that's OK, as long
4758 * as the task is no longer on this CPU.
4760 * Returns non-zero if task was successfully migrated.
4762 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4764 struct rq
*rq_dest
, *rq_src
;
4767 if (unlikely(!cpu_active(dest_cpu
)))
4770 rq_src
= cpu_rq(src_cpu
);
4771 rq_dest
= cpu_rq(dest_cpu
);
4773 raw_spin_lock(&p
->pi_lock
);
4774 double_rq_lock(rq_src
, rq_dest
);
4775 /* Already moved. */
4776 if (task_cpu(p
) != src_cpu
)
4778 /* Affinity changed (again). */
4779 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4783 * If we're not on a rq, the next wake-up will ensure we're
4787 dequeue_task(rq_src
, p
, 0);
4788 set_task_cpu(p
, dest_cpu
);
4789 enqueue_task(rq_dest
, p
, 0);
4790 check_preempt_curr(rq_dest
, p
, 0);
4795 double_rq_unlock(rq_src
, rq_dest
);
4796 raw_spin_unlock(&p
->pi_lock
);
4801 * migration_cpu_stop - this will be executed by a highprio stopper thread
4802 * and performs thread migration by bumping thread off CPU then
4803 * 'pushing' onto another runqueue.
4805 static int migration_cpu_stop(void *data
)
4807 struct migration_arg
*arg
= data
;
4810 * The original target cpu might have gone down and we might
4811 * be on another cpu but it doesn't matter.
4813 local_irq_disable();
4814 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4819 #ifdef CONFIG_HOTPLUG_CPU
4822 * Ensures that the idle task is using init_mm right before its cpu goes
4825 void idle_task_exit(void)
4827 struct mm_struct
*mm
= current
->active_mm
;
4829 BUG_ON(cpu_online(smp_processor_id()));
4832 switch_mm(mm
, &init_mm
, current
);
4837 * Since this CPU is going 'away' for a while, fold any nr_active delta
4838 * we might have. Assumes we're called after migrate_tasks() so that the
4839 * nr_active count is stable.
4841 * Also see the comment "Global load-average calculations".
4843 static void calc_load_migrate(struct rq
*rq
)
4845 long delta
= calc_load_fold_active(rq
);
4847 atomic_long_add(delta
, &calc_load_tasks
);
4851 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4852 * try_to_wake_up()->select_task_rq().
4854 * Called with rq->lock held even though we'er in stop_machine() and
4855 * there's no concurrency possible, we hold the required locks anyway
4856 * because of lock validation efforts.
4858 static void migrate_tasks(unsigned int dead_cpu
)
4860 struct rq
*rq
= cpu_rq(dead_cpu
);
4861 struct task_struct
*next
, *stop
= rq
->stop
;
4865 * Fudge the rq selection such that the below task selection loop
4866 * doesn't get stuck on the currently eligible stop task.
4868 * We're currently inside stop_machine() and the rq is either stuck
4869 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4870 * either way we should never end up calling schedule() until we're
4877 * There's this thread running, bail when that's the only
4880 if (rq
->nr_running
== 1)
4883 next
= pick_next_task(rq
);
4885 next
->sched_class
->put_prev_task(rq
, next
);
4887 /* Find suitable destination for @next, with force if needed. */
4888 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4889 raw_spin_unlock(&rq
->lock
);
4891 __migrate_task(next
, dead_cpu
, dest_cpu
);
4893 raw_spin_lock(&rq
->lock
);
4899 #endif /* CONFIG_HOTPLUG_CPU */
4901 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4903 static struct ctl_table sd_ctl_dir
[] = {
4905 .procname
= "sched_domain",
4911 static struct ctl_table sd_ctl_root
[] = {
4913 .procname
= "kernel",
4915 .child
= sd_ctl_dir
,
4920 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4922 struct ctl_table
*entry
=
4923 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4928 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4930 struct ctl_table
*entry
;
4933 * In the intermediate directories, both the child directory and
4934 * procname are dynamically allocated and could fail but the mode
4935 * will always be set. In the lowest directory the names are
4936 * static strings and all have proc handlers.
4938 for (entry
= *tablep
; entry
->mode
; entry
++) {
4940 sd_free_ctl_entry(&entry
->child
);
4941 if (entry
->proc_handler
== NULL
)
4942 kfree(entry
->procname
);
4949 static int min_load_idx
= 0;
4950 static int max_load_idx
= CPU_LOAD_IDX_MAX
;
4953 set_table_entry(struct ctl_table
*entry
,
4954 const char *procname
, void *data
, int maxlen
,
4955 umode_t mode
, proc_handler
*proc_handler
,
4958 entry
->procname
= procname
;
4960 entry
->maxlen
= maxlen
;
4962 entry
->proc_handler
= proc_handler
;
4965 entry
->extra1
= &min_load_idx
;
4966 entry
->extra2
= &max_load_idx
;
4970 static struct ctl_table
*
4971 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4973 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4978 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4979 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4980 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4981 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4982 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4983 sizeof(int), 0644, proc_dointvec_minmax
, true);
4984 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4985 sizeof(int), 0644, proc_dointvec_minmax
, true);
4986 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4987 sizeof(int), 0644, proc_dointvec_minmax
, true);
4988 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4989 sizeof(int), 0644, proc_dointvec_minmax
, true);
4990 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4991 sizeof(int), 0644, proc_dointvec_minmax
, true);
4992 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4993 sizeof(int), 0644, proc_dointvec_minmax
, false);
4994 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4995 sizeof(int), 0644, proc_dointvec_minmax
, false);
4996 set_table_entry(&table
[9], "cache_nice_tries",
4997 &sd
->cache_nice_tries
,
4998 sizeof(int), 0644, proc_dointvec_minmax
, false);
4999 set_table_entry(&table
[10], "flags", &sd
->flags
,
5000 sizeof(int), 0644, proc_dointvec_minmax
, false);
5001 set_table_entry(&table
[11], "name", sd
->name
,
5002 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5003 /* &table[12] is terminator */
5008 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5010 struct ctl_table
*entry
, *table
;
5011 struct sched_domain
*sd
;
5012 int domain_num
= 0, i
;
5015 for_each_domain(cpu
, sd
)
5017 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5022 for_each_domain(cpu
, sd
) {
5023 snprintf(buf
, 32, "domain%d", i
);
5024 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5026 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5033 static struct ctl_table_header
*sd_sysctl_header
;
5034 static void register_sched_domain_sysctl(void)
5036 int i
, cpu_num
= num_possible_cpus();
5037 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5040 WARN_ON(sd_ctl_dir
[0].child
);
5041 sd_ctl_dir
[0].child
= entry
;
5046 for_each_possible_cpu(i
) {
5047 snprintf(buf
, 32, "cpu%d", i
);
5048 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5050 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5054 WARN_ON(sd_sysctl_header
);
5055 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5058 /* may be called multiple times per register */
5059 static void unregister_sched_domain_sysctl(void)
5061 if (sd_sysctl_header
)
5062 unregister_sysctl_table(sd_sysctl_header
);
5063 sd_sysctl_header
= NULL
;
5064 if (sd_ctl_dir
[0].child
)
5065 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5068 static void register_sched_domain_sysctl(void)
5071 static void unregister_sched_domain_sysctl(void)
5076 static void set_rq_online(struct rq
*rq
)
5079 const struct sched_class
*class;
5081 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5084 for_each_class(class) {
5085 if (class->rq_online
)
5086 class->rq_online(rq
);
5091 static void set_rq_offline(struct rq
*rq
)
5094 const struct sched_class
*class;
5096 for_each_class(class) {
5097 if (class->rq_offline
)
5098 class->rq_offline(rq
);
5101 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5107 * migration_call - callback that gets triggered when a CPU is added.
5108 * Here we can start up the necessary migration thread for the new CPU.
5110 static int __cpuinit
5111 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5113 int cpu
= (long)hcpu
;
5114 unsigned long flags
;
5115 struct rq
*rq
= cpu_rq(cpu
);
5117 switch (action
& ~CPU_TASKS_FROZEN
) {
5119 case CPU_UP_PREPARE
:
5120 rq
->calc_load_update
= calc_load_update
;
5124 /* Update our root-domain */
5125 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5127 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5131 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5134 #ifdef CONFIG_HOTPLUG_CPU
5136 sched_ttwu_pending();
5137 /* Update our root-domain */
5138 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5140 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5144 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5145 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5149 calc_load_migrate(rq
);
5154 update_max_interval();
5160 * Register at high priority so that task migration (migrate_all_tasks)
5161 * happens before everything else. This has to be lower priority than
5162 * the notifier in the perf_event subsystem, though.
5164 static struct notifier_block __cpuinitdata migration_notifier
= {
5165 .notifier_call
= migration_call
,
5166 .priority
= CPU_PRI_MIGRATION
,
5169 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5170 unsigned long action
, void *hcpu
)
5172 switch (action
& ~CPU_TASKS_FROZEN
) {
5174 case CPU_DOWN_FAILED
:
5175 set_cpu_active((long)hcpu
, true);
5182 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5183 unsigned long action
, void *hcpu
)
5185 switch (action
& ~CPU_TASKS_FROZEN
) {
5186 case CPU_DOWN_PREPARE
:
5187 set_cpu_active((long)hcpu
, false);
5194 static int __init
migration_init(void)
5196 void *cpu
= (void *)(long)smp_processor_id();
5199 /* Initialize migration for the boot CPU */
5200 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5201 BUG_ON(err
== NOTIFY_BAD
);
5202 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5203 register_cpu_notifier(&migration_notifier
);
5205 /* Register cpu active notifiers */
5206 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5207 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5211 early_initcall(migration_init
);
5216 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5218 #ifdef CONFIG_SCHED_DEBUG
5220 static __read_mostly
int sched_debug_enabled
;
5222 static int __init
sched_debug_setup(char *str
)
5224 sched_debug_enabled
= 1;
5228 early_param("sched_debug", sched_debug_setup
);
5230 static inline bool sched_debug(void)
5232 return sched_debug_enabled
;
5235 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5236 struct cpumask
*groupmask
)
5238 struct sched_group
*group
= sd
->groups
;
5241 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5242 cpumask_clear(groupmask
);
5244 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5246 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5247 printk("does not load-balance\n");
5249 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5254 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5256 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5257 printk(KERN_ERR
"ERROR: domain->span does not contain "
5260 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5261 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5265 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5269 printk(KERN_ERR
"ERROR: group is NULL\n");
5274 * Even though we initialize ->power to something semi-sane,
5275 * we leave power_orig unset. This allows us to detect if
5276 * domain iteration is still funny without causing /0 traps.
5278 if (!group
->sgp
->power_orig
) {
5279 printk(KERN_CONT
"\n");
5280 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5285 if (!cpumask_weight(sched_group_cpus(group
))) {
5286 printk(KERN_CONT
"\n");
5287 printk(KERN_ERR
"ERROR: empty group\n");
5291 if (!(sd
->flags
& SD_OVERLAP
) &&
5292 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5293 printk(KERN_CONT
"\n");
5294 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5298 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5300 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5302 printk(KERN_CONT
" %s", str
);
5303 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5304 printk(KERN_CONT
" (cpu_power = %d)",
5308 group
= group
->next
;
5309 } while (group
!= sd
->groups
);
5310 printk(KERN_CONT
"\n");
5312 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5313 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5316 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5317 printk(KERN_ERR
"ERROR: parent span is not a superset "
5318 "of domain->span\n");
5322 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5326 if (!sched_debug_enabled
)
5330 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5334 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5337 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5345 #else /* !CONFIG_SCHED_DEBUG */
5346 # define sched_domain_debug(sd, cpu) do { } while (0)
5347 static inline bool sched_debug(void)
5351 #endif /* CONFIG_SCHED_DEBUG */
5353 static int sd_degenerate(struct sched_domain
*sd
)
5355 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5358 /* Following flags need at least 2 groups */
5359 if (sd
->flags
& (SD_LOAD_BALANCE
|
5360 SD_BALANCE_NEWIDLE
|
5364 SD_SHARE_PKG_RESOURCES
)) {
5365 if (sd
->groups
!= sd
->groups
->next
)
5369 /* Following flags don't use groups */
5370 if (sd
->flags
& (SD_WAKE_AFFINE
))
5377 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5379 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5381 if (sd_degenerate(parent
))
5384 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5387 /* Flags needing groups don't count if only 1 group in parent */
5388 if (parent
->groups
== parent
->groups
->next
) {
5389 pflags
&= ~(SD_LOAD_BALANCE
|
5390 SD_BALANCE_NEWIDLE
|
5394 SD_SHARE_PKG_RESOURCES
);
5395 if (nr_node_ids
== 1)
5396 pflags
&= ~SD_SERIALIZE
;
5398 if (~cflags
& pflags
)
5404 static void free_rootdomain(struct rcu_head
*rcu
)
5406 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5408 cpupri_cleanup(&rd
->cpupri
);
5409 free_cpumask_var(rd
->rto_mask
);
5410 free_cpumask_var(rd
->online
);
5411 free_cpumask_var(rd
->span
);
5415 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5417 struct root_domain
*old_rd
= NULL
;
5418 unsigned long flags
;
5420 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5425 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5428 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5431 * If we dont want to free the old_rt yet then
5432 * set old_rd to NULL to skip the freeing later
5435 if (!atomic_dec_and_test(&old_rd
->refcount
))
5439 atomic_inc(&rd
->refcount
);
5442 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5443 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5446 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5449 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5452 static int init_rootdomain(struct root_domain
*rd
)
5454 memset(rd
, 0, sizeof(*rd
));
5456 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5458 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5460 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5463 if (cpupri_init(&rd
->cpupri
) != 0)
5468 free_cpumask_var(rd
->rto_mask
);
5470 free_cpumask_var(rd
->online
);
5472 free_cpumask_var(rd
->span
);
5478 * By default the system creates a single root-domain with all cpus as
5479 * members (mimicking the global state we have today).
5481 struct root_domain def_root_domain
;
5483 static void init_defrootdomain(void)
5485 init_rootdomain(&def_root_domain
);
5487 atomic_set(&def_root_domain
.refcount
, 1);
5490 static struct root_domain
*alloc_rootdomain(void)
5492 struct root_domain
*rd
;
5494 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5498 if (init_rootdomain(rd
) != 0) {
5506 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5508 struct sched_group
*tmp
, *first
;
5517 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5522 } while (sg
!= first
);
5525 static void free_sched_domain(struct rcu_head
*rcu
)
5527 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5530 * If its an overlapping domain it has private groups, iterate and
5533 if (sd
->flags
& SD_OVERLAP
) {
5534 free_sched_groups(sd
->groups
, 1);
5535 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5536 kfree(sd
->groups
->sgp
);
5542 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5544 call_rcu(&sd
->rcu
, free_sched_domain
);
5547 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5549 for (; sd
; sd
= sd
->parent
)
5550 destroy_sched_domain(sd
, cpu
);
5554 * Keep a special pointer to the highest sched_domain that has
5555 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5556 * allows us to avoid some pointer chasing select_idle_sibling().
5558 * Also keep a unique ID per domain (we use the first cpu number in
5559 * the cpumask of the domain), this allows us to quickly tell if
5560 * two cpus are in the same cache domain, see cpus_share_cache().
5562 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5563 DEFINE_PER_CPU(int, sd_llc_id
);
5565 static void update_top_cache_domain(int cpu
)
5567 struct sched_domain
*sd
;
5570 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5572 id
= cpumask_first(sched_domain_span(sd
));
5574 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5575 per_cpu(sd_llc_id
, cpu
) = id
;
5579 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5580 * hold the hotplug lock.
5583 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5585 struct rq
*rq
= cpu_rq(cpu
);
5586 struct sched_domain
*tmp
;
5588 /* Remove the sched domains which do not contribute to scheduling. */
5589 for (tmp
= sd
; tmp
; ) {
5590 struct sched_domain
*parent
= tmp
->parent
;
5594 if (sd_parent_degenerate(tmp
, parent
)) {
5595 tmp
->parent
= parent
->parent
;
5597 parent
->parent
->child
= tmp
;
5598 destroy_sched_domain(parent
, cpu
);
5603 if (sd
&& sd_degenerate(sd
)) {
5606 destroy_sched_domain(tmp
, cpu
);
5611 sched_domain_debug(sd
, cpu
);
5613 rq_attach_root(rq
, rd
);
5615 rcu_assign_pointer(rq
->sd
, sd
);
5616 destroy_sched_domains(tmp
, cpu
);
5618 update_top_cache_domain(cpu
);
5621 /* cpus with isolated domains */
5622 static cpumask_var_t cpu_isolated_map
;
5624 /* Setup the mask of cpus configured for isolated domains */
5625 static int __init
isolated_cpu_setup(char *str
)
5627 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5628 cpulist_parse(str
, cpu_isolated_map
);
5632 __setup("isolcpus=", isolated_cpu_setup
);
5634 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5636 return cpumask_of_node(cpu_to_node(cpu
));
5640 struct sched_domain
**__percpu sd
;
5641 struct sched_group
**__percpu sg
;
5642 struct sched_group_power
**__percpu sgp
;
5646 struct sched_domain
** __percpu sd
;
5647 struct root_domain
*rd
;
5657 struct sched_domain_topology_level
;
5659 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5660 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5662 #define SDTL_OVERLAP 0x01
5664 struct sched_domain_topology_level
{
5665 sched_domain_init_f init
;
5666 sched_domain_mask_f mask
;
5669 struct sd_data data
;
5673 * Build an iteration mask that can exclude certain CPUs from the upwards
5676 * Asymmetric node setups can result in situations where the domain tree is of
5677 * unequal depth, make sure to skip domains that already cover the entire
5680 * In that case build_sched_domains() will have terminated the iteration early
5681 * and our sibling sd spans will be empty. Domains should always include the
5682 * cpu they're built on, so check that.
5685 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5687 const struct cpumask
*span
= sched_domain_span(sd
);
5688 struct sd_data
*sdd
= sd
->private;
5689 struct sched_domain
*sibling
;
5692 for_each_cpu(i
, span
) {
5693 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5694 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5697 cpumask_set_cpu(i
, sched_group_mask(sg
));
5702 * Return the canonical balance cpu for this group, this is the first cpu
5703 * of this group that's also in the iteration mask.
5705 int group_balance_cpu(struct sched_group
*sg
)
5707 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5711 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5713 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5714 const struct cpumask
*span
= sched_domain_span(sd
);
5715 struct cpumask
*covered
= sched_domains_tmpmask
;
5716 struct sd_data
*sdd
= sd
->private;
5717 struct sched_domain
*child
;
5720 cpumask_clear(covered
);
5722 for_each_cpu(i
, span
) {
5723 struct cpumask
*sg_span
;
5725 if (cpumask_test_cpu(i
, covered
))
5728 child
= *per_cpu_ptr(sdd
->sd
, i
);
5730 /* See the comment near build_group_mask(). */
5731 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5734 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5735 GFP_KERNEL
, cpu_to_node(cpu
));
5740 sg_span
= sched_group_cpus(sg
);
5742 child
= child
->child
;
5743 cpumask_copy(sg_span
, sched_domain_span(child
));
5745 cpumask_set_cpu(i
, sg_span
);
5747 cpumask_or(covered
, covered
, sg_span
);
5749 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5750 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5751 build_group_mask(sd
, sg
);
5754 * Initialize sgp->power such that even if we mess up the
5755 * domains and no possible iteration will get us here, we won't
5758 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5761 * Make sure the first group of this domain contains the
5762 * canonical balance cpu. Otherwise the sched_domain iteration
5763 * breaks. See update_sg_lb_stats().
5765 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5766 group_balance_cpu(sg
) == cpu
)
5776 sd
->groups
= groups
;
5781 free_sched_groups(first
, 0);
5786 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5788 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5789 struct sched_domain
*child
= sd
->child
;
5792 cpu
= cpumask_first(sched_domain_span(child
));
5795 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5796 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5797 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5804 * build_sched_groups will build a circular linked list of the groups
5805 * covered by the given span, and will set each group's ->cpumask correctly,
5806 * and ->cpu_power to 0.
5808 * Assumes the sched_domain tree is fully constructed
5811 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5813 struct sched_group
*first
= NULL
, *last
= NULL
;
5814 struct sd_data
*sdd
= sd
->private;
5815 const struct cpumask
*span
= sched_domain_span(sd
);
5816 struct cpumask
*covered
;
5819 get_group(cpu
, sdd
, &sd
->groups
);
5820 atomic_inc(&sd
->groups
->ref
);
5822 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5825 lockdep_assert_held(&sched_domains_mutex
);
5826 covered
= sched_domains_tmpmask
;
5828 cpumask_clear(covered
);
5830 for_each_cpu(i
, span
) {
5831 struct sched_group
*sg
;
5832 int group
= get_group(i
, sdd
, &sg
);
5835 if (cpumask_test_cpu(i
, covered
))
5838 cpumask_clear(sched_group_cpus(sg
));
5840 cpumask_setall(sched_group_mask(sg
));
5842 for_each_cpu(j
, span
) {
5843 if (get_group(j
, sdd
, NULL
) != group
)
5846 cpumask_set_cpu(j
, covered
);
5847 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5862 * Initialize sched groups cpu_power.
5864 * cpu_power indicates the capacity of sched group, which is used while
5865 * distributing the load between different sched groups in a sched domain.
5866 * Typically cpu_power for all the groups in a sched domain will be same unless
5867 * there are asymmetries in the topology. If there are asymmetries, group
5868 * having more cpu_power will pickup more load compared to the group having
5871 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5873 struct sched_group
*sg
= sd
->groups
;
5875 WARN_ON(!sd
|| !sg
);
5878 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5880 } while (sg
!= sd
->groups
);
5882 if (cpu
!= group_balance_cpu(sg
))
5885 update_group_power(sd
, cpu
);
5886 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5889 int __weak
arch_sd_sibling_asym_packing(void)
5891 return 0*SD_ASYM_PACKING
;
5895 * Initializers for schedule domains
5896 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5899 #ifdef CONFIG_SCHED_DEBUG
5900 # define SD_INIT_NAME(sd, type) sd->name = #type
5902 # define SD_INIT_NAME(sd, type) do { } while (0)
5905 #define SD_INIT_FUNC(type) \
5906 static noinline struct sched_domain * \
5907 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5909 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5910 *sd = SD_##type##_INIT; \
5911 SD_INIT_NAME(sd, type); \
5912 sd->private = &tl->data; \
5917 #ifdef CONFIG_SCHED_SMT
5918 SD_INIT_FUNC(SIBLING
)
5920 #ifdef CONFIG_SCHED_MC
5923 #ifdef CONFIG_SCHED_BOOK
5927 static int default_relax_domain_level
= -1;
5928 int sched_domain_level_max
;
5930 static int __init
setup_relax_domain_level(char *str
)
5932 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5933 pr_warn("Unable to set relax_domain_level\n");
5937 __setup("relax_domain_level=", setup_relax_domain_level
);
5939 static void set_domain_attribute(struct sched_domain
*sd
,
5940 struct sched_domain_attr
*attr
)
5944 if (!attr
|| attr
->relax_domain_level
< 0) {
5945 if (default_relax_domain_level
< 0)
5948 request
= default_relax_domain_level
;
5950 request
= attr
->relax_domain_level
;
5951 if (request
< sd
->level
) {
5952 /* turn off idle balance on this domain */
5953 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5955 /* turn on idle balance on this domain */
5956 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5960 static void __sdt_free(const struct cpumask
*cpu_map
);
5961 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5963 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5964 const struct cpumask
*cpu_map
)
5968 if (!atomic_read(&d
->rd
->refcount
))
5969 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5971 free_percpu(d
->sd
); /* fall through */
5973 __sdt_free(cpu_map
); /* fall through */
5979 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5980 const struct cpumask
*cpu_map
)
5982 memset(d
, 0, sizeof(*d
));
5984 if (__sdt_alloc(cpu_map
))
5985 return sa_sd_storage
;
5986 d
->sd
= alloc_percpu(struct sched_domain
*);
5988 return sa_sd_storage
;
5989 d
->rd
= alloc_rootdomain();
5992 return sa_rootdomain
;
5996 * NULL the sd_data elements we've used to build the sched_domain and
5997 * sched_group structure so that the subsequent __free_domain_allocs()
5998 * will not free the data we're using.
6000 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6002 struct sd_data
*sdd
= sd
->private;
6004 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6005 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6007 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6008 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6010 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6011 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6014 #ifdef CONFIG_SCHED_SMT
6015 static const struct cpumask
*cpu_smt_mask(int cpu
)
6017 return topology_thread_cpumask(cpu
);
6022 * Topology list, bottom-up.
6024 static struct sched_domain_topology_level default_topology
[] = {
6025 #ifdef CONFIG_SCHED_SMT
6026 { sd_init_SIBLING
, cpu_smt_mask
, },
6028 #ifdef CONFIG_SCHED_MC
6029 { sd_init_MC
, cpu_coregroup_mask
, },
6031 #ifdef CONFIG_SCHED_BOOK
6032 { sd_init_BOOK
, cpu_book_mask
, },
6034 { sd_init_CPU
, cpu_cpu_mask
, },
6038 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6042 static int sched_domains_numa_levels
;
6043 static int *sched_domains_numa_distance
;
6044 static struct cpumask
***sched_domains_numa_masks
;
6045 static int sched_domains_curr_level
;
6047 static inline int sd_local_flags(int level
)
6049 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6052 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6055 static struct sched_domain
*
6056 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6058 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6059 int level
= tl
->numa_level
;
6060 int sd_weight
= cpumask_weight(
6061 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6063 *sd
= (struct sched_domain
){
6064 .min_interval
= sd_weight
,
6065 .max_interval
= 2*sd_weight
,
6067 .imbalance_pct
= 125,
6068 .cache_nice_tries
= 2,
6075 .flags
= 1*SD_LOAD_BALANCE
6076 | 1*SD_BALANCE_NEWIDLE
6081 | 0*SD_SHARE_CPUPOWER
6082 | 0*SD_SHARE_PKG_RESOURCES
6084 | 0*SD_PREFER_SIBLING
6085 | sd_local_flags(level
)
6087 .last_balance
= jiffies
,
6088 .balance_interval
= sd_weight
,
6090 SD_INIT_NAME(sd
, NUMA
);
6091 sd
->private = &tl
->data
;
6094 * Ugly hack to pass state to sd_numa_mask()...
6096 sched_domains_curr_level
= tl
->numa_level
;
6101 static const struct cpumask
*sd_numa_mask(int cpu
)
6103 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6106 static void sched_numa_warn(const char *str
)
6108 static int done
= false;
6116 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6118 for (i
= 0; i
< nr_node_ids
; i
++) {
6119 printk(KERN_WARNING
" ");
6120 for (j
= 0; j
< nr_node_ids
; j
++)
6121 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6122 printk(KERN_CONT
"\n");
6124 printk(KERN_WARNING
"\n");
6127 static bool find_numa_distance(int distance
)
6131 if (distance
== node_distance(0, 0))
6134 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6135 if (sched_domains_numa_distance
[i
] == distance
)
6142 static void sched_init_numa(void)
6144 int next_distance
, curr_distance
= node_distance(0, 0);
6145 struct sched_domain_topology_level
*tl
;
6149 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6150 if (!sched_domains_numa_distance
)
6154 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6155 * unique distances in the node_distance() table.
6157 * Assumes node_distance(0,j) includes all distances in
6158 * node_distance(i,j) in order to avoid cubic time.
6160 next_distance
= curr_distance
;
6161 for (i
= 0; i
< nr_node_ids
; i
++) {
6162 for (j
= 0; j
< nr_node_ids
; j
++) {
6163 for (k
= 0; k
< nr_node_ids
; k
++) {
6164 int distance
= node_distance(i
, k
);
6166 if (distance
> curr_distance
&&
6167 (distance
< next_distance
||
6168 next_distance
== curr_distance
))
6169 next_distance
= distance
;
6172 * While not a strong assumption it would be nice to know
6173 * about cases where if node A is connected to B, B is not
6174 * equally connected to A.
6176 if (sched_debug() && node_distance(k
, i
) != distance
)
6177 sched_numa_warn("Node-distance not symmetric");
6179 if (sched_debug() && i
&& !find_numa_distance(distance
))
6180 sched_numa_warn("Node-0 not representative");
6182 if (next_distance
!= curr_distance
) {
6183 sched_domains_numa_distance
[level
++] = next_distance
;
6184 sched_domains_numa_levels
= level
;
6185 curr_distance
= next_distance
;
6190 * In case of sched_debug() we verify the above assumption.
6196 * 'level' contains the number of unique distances, excluding the
6197 * identity distance node_distance(i,i).
6199 * The sched_domains_nume_distance[] array includes the actual distance
6204 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6205 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6206 * the array will contain less then 'level' members. This could be
6207 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6208 * in other functions.
6210 * We reset it to 'level' at the end of this function.
6212 sched_domains_numa_levels
= 0;
6214 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6215 if (!sched_domains_numa_masks
)
6219 * Now for each level, construct a mask per node which contains all
6220 * cpus of nodes that are that many hops away from us.
6222 for (i
= 0; i
< level
; i
++) {
6223 sched_domains_numa_masks
[i
] =
6224 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6225 if (!sched_domains_numa_masks
[i
])
6228 for (j
= 0; j
< nr_node_ids
; j
++) {
6229 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6233 sched_domains_numa_masks
[i
][j
] = mask
;
6235 for (k
= 0; k
< nr_node_ids
; k
++) {
6236 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6239 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6244 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6245 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6250 * Copy the default topology bits..
6252 for (i
= 0; default_topology
[i
].init
; i
++)
6253 tl
[i
] = default_topology
[i
];
6256 * .. and append 'j' levels of NUMA goodness.
6258 for (j
= 0; j
< level
; i
++, j
++) {
6259 tl
[i
] = (struct sched_domain_topology_level
){
6260 .init
= sd_numa_init
,
6261 .mask
= sd_numa_mask
,
6262 .flags
= SDTL_OVERLAP
,
6267 sched_domain_topology
= tl
;
6269 sched_domains_numa_levels
= level
;
6272 static void sched_domains_numa_masks_set(int cpu
)
6275 int node
= cpu_to_node(cpu
);
6277 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6278 for (j
= 0; j
< nr_node_ids
; j
++) {
6279 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6280 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6285 static void sched_domains_numa_masks_clear(int cpu
)
6288 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6289 for (j
= 0; j
< nr_node_ids
; j
++)
6290 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6295 * Update sched_domains_numa_masks[level][node] array when new cpus
6298 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6299 unsigned long action
,
6302 int cpu
= (long)hcpu
;
6304 switch (action
& ~CPU_TASKS_FROZEN
) {
6306 sched_domains_numa_masks_set(cpu
);
6310 sched_domains_numa_masks_clear(cpu
);
6320 static inline void sched_init_numa(void)
6324 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6325 unsigned long action
,
6330 #endif /* CONFIG_NUMA */
6332 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6334 struct sched_domain_topology_level
*tl
;
6337 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6338 struct sd_data
*sdd
= &tl
->data
;
6340 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6344 sdd
->sg
= alloc_percpu(struct sched_group
*);
6348 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6352 for_each_cpu(j
, cpu_map
) {
6353 struct sched_domain
*sd
;
6354 struct sched_group
*sg
;
6355 struct sched_group_power
*sgp
;
6357 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6358 GFP_KERNEL
, cpu_to_node(j
));
6362 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6364 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6365 GFP_KERNEL
, cpu_to_node(j
));
6371 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6373 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6374 GFP_KERNEL
, cpu_to_node(j
));
6378 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6385 static void __sdt_free(const struct cpumask
*cpu_map
)
6387 struct sched_domain_topology_level
*tl
;
6390 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6391 struct sd_data
*sdd
= &tl
->data
;
6393 for_each_cpu(j
, cpu_map
) {
6394 struct sched_domain
*sd
;
6397 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6398 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6399 free_sched_groups(sd
->groups
, 0);
6400 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6404 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6406 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6408 free_percpu(sdd
->sd
);
6410 free_percpu(sdd
->sg
);
6412 free_percpu(sdd
->sgp
);
6417 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6418 struct s_data
*d
, const struct cpumask
*cpu_map
,
6419 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6422 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6426 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6428 sd
->level
= child
->level
+ 1;
6429 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6433 set_domain_attribute(sd
, attr
);
6439 * Build sched domains for a given set of cpus and attach the sched domains
6440 * to the individual cpus
6442 static int build_sched_domains(const struct cpumask
*cpu_map
,
6443 struct sched_domain_attr
*attr
)
6445 enum s_alloc alloc_state
= sa_none
;
6446 struct sched_domain
*sd
;
6448 int i
, ret
= -ENOMEM
;
6450 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6451 if (alloc_state
!= sa_rootdomain
)
6454 /* Set up domains for cpus specified by the cpu_map. */
6455 for_each_cpu(i
, cpu_map
) {
6456 struct sched_domain_topology_level
*tl
;
6459 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6460 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6461 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6462 sd
->flags
|= SD_OVERLAP
;
6463 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6470 *per_cpu_ptr(d
.sd
, i
) = sd
;
6473 /* Build the groups for the domains */
6474 for_each_cpu(i
, cpu_map
) {
6475 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6476 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6477 if (sd
->flags
& SD_OVERLAP
) {
6478 if (build_overlap_sched_groups(sd
, i
))
6481 if (build_sched_groups(sd
, i
))
6487 /* Calculate CPU power for physical packages and nodes */
6488 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6489 if (!cpumask_test_cpu(i
, cpu_map
))
6492 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6493 claim_allocations(i
, sd
);
6494 init_sched_groups_power(i
, sd
);
6498 /* Attach the domains */
6500 for_each_cpu(i
, cpu_map
) {
6501 sd
= *per_cpu_ptr(d
.sd
, i
);
6502 cpu_attach_domain(sd
, d
.rd
, i
);
6508 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6512 static cpumask_var_t
*doms_cur
; /* current sched domains */
6513 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6514 static struct sched_domain_attr
*dattr_cur
;
6515 /* attribues of custom domains in 'doms_cur' */
6518 * Special case: If a kmalloc of a doms_cur partition (array of
6519 * cpumask) fails, then fallback to a single sched domain,
6520 * as determined by the single cpumask fallback_doms.
6522 static cpumask_var_t fallback_doms
;
6525 * arch_update_cpu_topology lets virtualized architectures update the
6526 * cpu core maps. It is supposed to return 1 if the topology changed
6527 * or 0 if it stayed the same.
6529 int __attribute__((weak
)) arch_update_cpu_topology(void)
6534 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6537 cpumask_var_t
*doms
;
6539 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6542 for (i
= 0; i
< ndoms
; i
++) {
6543 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6544 free_sched_domains(doms
, i
);
6551 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6554 for (i
= 0; i
< ndoms
; i
++)
6555 free_cpumask_var(doms
[i
]);
6560 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6561 * For now this just excludes isolated cpus, but could be used to
6562 * exclude other special cases in the future.
6564 static int init_sched_domains(const struct cpumask
*cpu_map
)
6568 arch_update_cpu_topology();
6570 doms_cur
= alloc_sched_domains(ndoms_cur
);
6572 doms_cur
= &fallback_doms
;
6573 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6574 err
= build_sched_domains(doms_cur
[0], NULL
);
6575 register_sched_domain_sysctl();
6581 * Detach sched domains from a group of cpus specified in cpu_map
6582 * These cpus will now be attached to the NULL domain
6584 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6589 for_each_cpu(i
, cpu_map
)
6590 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6594 /* handle null as "default" */
6595 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6596 struct sched_domain_attr
*new, int idx_new
)
6598 struct sched_domain_attr tmp
;
6605 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6606 new ? (new + idx_new
) : &tmp
,
6607 sizeof(struct sched_domain_attr
));
6611 * Partition sched domains as specified by the 'ndoms_new'
6612 * cpumasks in the array doms_new[] of cpumasks. This compares
6613 * doms_new[] to the current sched domain partitioning, doms_cur[].
6614 * It destroys each deleted domain and builds each new domain.
6616 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6617 * The masks don't intersect (don't overlap.) We should setup one
6618 * sched domain for each mask. CPUs not in any of the cpumasks will
6619 * not be load balanced. If the same cpumask appears both in the
6620 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6623 * The passed in 'doms_new' should be allocated using
6624 * alloc_sched_domains. This routine takes ownership of it and will
6625 * free_sched_domains it when done with it. If the caller failed the
6626 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6627 * and partition_sched_domains() will fallback to the single partition
6628 * 'fallback_doms', it also forces the domains to be rebuilt.
6630 * If doms_new == NULL it will be replaced with cpu_online_mask.
6631 * ndoms_new == 0 is a special case for destroying existing domains,
6632 * and it will not create the default domain.
6634 * Call with hotplug lock held
6636 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6637 struct sched_domain_attr
*dattr_new
)
6642 mutex_lock(&sched_domains_mutex
);
6644 /* always unregister in case we don't destroy any domains */
6645 unregister_sched_domain_sysctl();
6647 /* Let architecture update cpu core mappings. */
6648 new_topology
= arch_update_cpu_topology();
6650 n
= doms_new
? ndoms_new
: 0;
6652 /* Destroy deleted domains */
6653 for (i
= 0; i
< ndoms_cur
; i
++) {
6654 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6655 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6656 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6659 /* no match - a current sched domain not in new doms_new[] */
6660 detach_destroy_domains(doms_cur
[i
]);
6665 if (doms_new
== NULL
) {
6667 doms_new
= &fallback_doms
;
6668 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6669 WARN_ON_ONCE(dattr_new
);
6672 /* Build new domains */
6673 for (i
= 0; i
< ndoms_new
; i
++) {
6674 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6675 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6676 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6679 /* no match - add a new doms_new */
6680 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6685 /* Remember the new sched domains */
6686 if (doms_cur
!= &fallback_doms
)
6687 free_sched_domains(doms_cur
, ndoms_cur
);
6688 kfree(dattr_cur
); /* kfree(NULL) is safe */
6689 doms_cur
= doms_new
;
6690 dattr_cur
= dattr_new
;
6691 ndoms_cur
= ndoms_new
;
6693 register_sched_domain_sysctl();
6695 mutex_unlock(&sched_domains_mutex
);
6698 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6701 * Update cpusets according to cpu_active mask. If cpusets are
6702 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6703 * around partition_sched_domains().
6705 * If we come here as part of a suspend/resume, don't touch cpusets because we
6706 * want to restore it back to its original state upon resume anyway.
6708 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6712 case CPU_ONLINE_FROZEN
:
6713 case CPU_DOWN_FAILED_FROZEN
:
6716 * num_cpus_frozen tracks how many CPUs are involved in suspend
6717 * resume sequence. As long as this is not the last online
6718 * operation in the resume sequence, just build a single sched
6719 * domain, ignoring cpusets.
6722 if (likely(num_cpus_frozen
)) {
6723 partition_sched_domains(1, NULL
, NULL
);
6728 * This is the last CPU online operation. So fall through and
6729 * restore the original sched domains by considering the
6730 * cpuset configurations.
6734 case CPU_DOWN_FAILED
:
6735 cpuset_update_active_cpus(true);
6743 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6747 case CPU_DOWN_PREPARE
:
6748 cpuset_update_active_cpus(false);
6750 case CPU_DOWN_PREPARE_FROZEN
:
6752 partition_sched_domains(1, NULL
, NULL
);
6760 void __init
sched_init_smp(void)
6762 cpumask_var_t non_isolated_cpus
;
6764 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6765 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6770 mutex_lock(&sched_domains_mutex
);
6771 init_sched_domains(cpu_active_mask
);
6772 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6773 if (cpumask_empty(non_isolated_cpus
))
6774 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6775 mutex_unlock(&sched_domains_mutex
);
6778 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6779 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6780 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6782 /* RT runtime code needs to handle some hotplug events */
6783 hotcpu_notifier(update_runtime
, 0);
6787 /* Move init over to a non-isolated CPU */
6788 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6790 sched_init_granularity();
6791 free_cpumask_var(non_isolated_cpus
);
6793 init_sched_rt_class();
6796 void __init
sched_init_smp(void)
6798 sched_init_granularity();
6800 #endif /* CONFIG_SMP */
6802 const_debug
unsigned int sysctl_timer_migration
= 1;
6804 int in_sched_functions(unsigned long addr
)
6806 return in_lock_functions(addr
) ||
6807 (addr
>= (unsigned long)__sched_text_start
6808 && addr
< (unsigned long)__sched_text_end
);
6811 #ifdef CONFIG_CGROUP_SCHED
6812 struct task_group root_task_group
;
6813 LIST_HEAD(task_groups
);
6816 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6818 void __init
sched_init(void)
6821 unsigned long alloc_size
= 0, ptr
;
6823 #ifdef CONFIG_FAIR_GROUP_SCHED
6824 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6826 #ifdef CONFIG_RT_GROUP_SCHED
6827 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6829 #ifdef CONFIG_CPUMASK_OFFSTACK
6830 alloc_size
+= num_possible_cpus() * cpumask_size();
6833 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6835 #ifdef CONFIG_FAIR_GROUP_SCHED
6836 root_task_group
.se
= (struct sched_entity
**)ptr
;
6837 ptr
+= nr_cpu_ids
* sizeof(void **);
6839 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6840 ptr
+= nr_cpu_ids
* sizeof(void **);
6842 #endif /* CONFIG_FAIR_GROUP_SCHED */
6843 #ifdef CONFIG_RT_GROUP_SCHED
6844 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6845 ptr
+= nr_cpu_ids
* sizeof(void **);
6847 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6848 ptr
+= nr_cpu_ids
* sizeof(void **);
6850 #endif /* CONFIG_RT_GROUP_SCHED */
6851 #ifdef CONFIG_CPUMASK_OFFSTACK
6852 for_each_possible_cpu(i
) {
6853 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6854 ptr
+= cpumask_size();
6856 #endif /* CONFIG_CPUMASK_OFFSTACK */
6860 init_defrootdomain();
6863 init_rt_bandwidth(&def_rt_bandwidth
,
6864 global_rt_period(), global_rt_runtime());
6866 #ifdef CONFIG_RT_GROUP_SCHED
6867 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6868 global_rt_period(), global_rt_runtime());
6869 #endif /* CONFIG_RT_GROUP_SCHED */
6871 #ifdef CONFIG_CGROUP_SCHED
6872 list_add(&root_task_group
.list
, &task_groups
);
6873 INIT_LIST_HEAD(&root_task_group
.children
);
6874 INIT_LIST_HEAD(&root_task_group
.siblings
);
6875 autogroup_init(&init_task
);
6877 #endif /* CONFIG_CGROUP_SCHED */
6879 #ifdef CONFIG_CGROUP_CPUACCT
6880 root_cpuacct
.cpustat
= &kernel_cpustat
;
6881 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6882 /* Too early, not expected to fail */
6883 BUG_ON(!root_cpuacct
.cpuusage
);
6885 for_each_possible_cpu(i
) {
6889 raw_spin_lock_init(&rq
->lock
);
6891 rq
->calc_load_active
= 0;
6892 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6893 init_cfs_rq(&rq
->cfs
);
6894 init_rt_rq(&rq
->rt
, rq
);
6895 #ifdef CONFIG_FAIR_GROUP_SCHED
6896 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6897 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6899 * How much cpu bandwidth does root_task_group get?
6901 * In case of task-groups formed thr' the cgroup filesystem, it
6902 * gets 100% of the cpu resources in the system. This overall
6903 * system cpu resource is divided among the tasks of
6904 * root_task_group and its child task-groups in a fair manner,
6905 * based on each entity's (task or task-group's) weight
6906 * (se->load.weight).
6908 * In other words, if root_task_group has 10 tasks of weight
6909 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6910 * then A0's share of the cpu resource is:
6912 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6914 * We achieve this by letting root_task_group's tasks sit
6915 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6917 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6918 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6919 #endif /* CONFIG_FAIR_GROUP_SCHED */
6921 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6922 #ifdef CONFIG_RT_GROUP_SCHED
6923 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6924 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6927 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6928 rq
->cpu_load
[j
] = 0;
6930 rq
->last_load_update_tick
= jiffies
;
6935 rq
->cpu_power
= SCHED_POWER_SCALE
;
6936 rq
->post_schedule
= 0;
6937 rq
->active_balance
= 0;
6938 rq
->next_balance
= jiffies
;
6943 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6945 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6947 rq_attach_root(rq
, &def_root_domain
);
6953 atomic_set(&rq
->nr_iowait
, 0);
6956 set_load_weight(&init_task
);
6958 #ifdef CONFIG_PREEMPT_NOTIFIERS
6959 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6962 #ifdef CONFIG_RT_MUTEXES
6963 plist_head_init(&init_task
.pi_waiters
);
6967 * The boot idle thread does lazy MMU switching as well:
6969 atomic_inc(&init_mm
.mm_count
);
6970 enter_lazy_tlb(&init_mm
, current
);
6973 * Make us the idle thread. Technically, schedule() should not be
6974 * called from this thread, however somewhere below it might be,
6975 * but because we are the idle thread, we just pick up running again
6976 * when this runqueue becomes "idle".
6978 init_idle(current
, smp_processor_id());
6980 calc_load_update
= jiffies
+ LOAD_FREQ
;
6983 * During early bootup we pretend to be a normal task:
6985 current
->sched_class
= &fair_sched_class
;
6988 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6989 /* May be allocated at isolcpus cmdline parse time */
6990 if (cpu_isolated_map
== NULL
)
6991 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6992 idle_thread_set_boot_cpu();
6994 init_sched_fair_class();
6996 scheduler_running
= 1;
6999 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7000 static inline int preempt_count_equals(int preempt_offset
)
7002 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7004 return (nested
== preempt_offset
);
7007 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7009 static unsigned long prev_jiffy
; /* ratelimiting */
7011 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7012 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7013 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7015 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7017 prev_jiffy
= jiffies
;
7020 "BUG: sleeping function called from invalid context at %s:%d\n",
7023 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7024 in_atomic(), irqs_disabled(),
7025 current
->pid
, current
->comm
);
7027 debug_show_held_locks(current
);
7028 if (irqs_disabled())
7029 print_irqtrace_events(current
);
7032 EXPORT_SYMBOL(__might_sleep
);
7035 #ifdef CONFIG_MAGIC_SYSRQ
7036 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7038 const struct sched_class
*prev_class
= p
->sched_class
;
7039 int old_prio
= p
->prio
;
7044 dequeue_task(rq
, p
, 0);
7045 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7047 enqueue_task(rq
, p
, 0);
7048 resched_task(rq
->curr
);
7051 check_class_changed(rq
, p
, prev_class
, old_prio
);
7054 void normalize_rt_tasks(void)
7056 struct task_struct
*g
, *p
;
7057 unsigned long flags
;
7060 read_lock_irqsave(&tasklist_lock
, flags
);
7061 do_each_thread(g
, p
) {
7063 * Only normalize user tasks:
7068 p
->se
.exec_start
= 0;
7069 #ifdef CONFIG_SCHEDSTATS
7070 p
->se
.statistics
.wait_start
= 0;
7071 p
->se
.statistics
.sleep_start
= 0;
7072 p
->se
.statistics
.block_start
= 0;
7077 * Renice negative nice level userspace
7080 if (TASK_NICE(p
) < 0 && p
->mm
)
7081 set_user_nice(p
, 0);
7085 raw_spin_lock(&p
->pi_lock
);
7086 rq
= __task_rq_lock(p
);
7088 normalize_task(rq
, p
);
7090 __task_rq_unlock(rq
);
7091 raw_spin_unlock(&p
->pi_lock
);
7092 } while_each_thread(g
, p
);
7094 read_unlock_irqrestore(&tasklist_lock
, flags
);
7097 #endif /* CONFIG_MAGIC_SYSRQ */
7099 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7101 * These functions are only useful for the IA64 MCA handling, or kdb.
7103 * They can only be called when the whole system has been
7104 * stopped - every CPU needs to be quiescent, and no scheduling
7105 * activity can take place. Using them for anything else would
7106 * be a serious bug, and as a result, they aren't even visible
7107 * under any other configuration.
7111 * curr_task - return the current task for a given cpu.
7112 * @cpu: the processor in question.
7114 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7116 struct task_struct
*curr_task(int cpu
)
7118 return cpu_curr(cpu
);
7121 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7125 * set_curr_task - set the current task for a given cpu.
7126 * @cpu: the processor in question.
7127 * @p: the task pointer to set.
7129 * Description: This function must only be used when non-maskable interrupts
7130 * are serviced on a separate stack. It allows the architecture to switch the
7131 * notion of the current task on a cpu in a non-blocking manner. This function
7132 * must be called with all CPU's synchronized, and interrupts disabled, the
7133 * and caller must save the original value of the current task (see
7134 * curr_task() above) and restore that value before reenabling interrupts and
7135 * re-starting the system.
7137 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7139 void set_curr_task(int cpu
, struct task_struct
*p
)
7146 #ifdef CONFIG_CGROUP_SCHED
7147 /* task_group_lock serializes the addition/removal of task groups */
7148 static DEFINE_SPINLOCK(task_group_lock
);
7150 static void free_sched_group(struct task_group
*tg
)
7152 free_fair_sched_group(tg
);
7153 free_rt_sched_group(tg
);
7158 /* allocate runqueue etc for a new task group */
7159 struct task_group
*sched_create_group(struct task_group
*parent
)
7161 struct task_group
*tg
;
7162 unsigned long flags
;
7164 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7166 return ERR_PTR(-ENOMEM
);
7168 if (!alloc_fair_sched_group(tg
, parent
))
7171 if (!alloc_rt_sched_group(tg
, parent
))
7174 spin_lock_irqsave(&task_group_lock
, flags
);
7175 list_add_rcu(&tg
->list
, &task_groups
);
7177 WARN_ON(!parent
); /* root should already exist */
7179 tg
->parent
= parent
;
7180 INIT_LIST_HEAD(&tg
->children
);
7181 list_add_rcu(&tg
->siblings
, &parent
->children
);
7182 spin_unlock_irqrestore(&task_group_lock
, flags
);
7187 free_sched_group(tg
);
7188 return ERR_PTR(-ENOMEM
);
7191 /* rcu callback to free various structures associated with a task group */
7192 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7194 /* now it should be safe to free those cfs_rqs */
7195 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7198 /* Destroy runqueue etc associated with a task group */
7199 void sched_destroy_group(struct task_group
*tg
)
7201 unsigned long flags
;
7204 /* end participation in shares distribution */
7205 for_each_possible_cpu(i
)
7206 unregister_fair_sched_group(tg
, i
);
7208 spin_lock_irqsave(&task_group_lock
, flags
);
7209 list_del_rcu(&tg
->list
);
7210 list_del_rcu(&tg
->siblings
);
7211 spin_unlock_irqrestore(&task_group_lock
, flags
);
7213 /* wait for possible concurrent references to cfs_rqs complete */
7214 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7217 /* change task's runqueue when it moves between groups.
7218 * The caller of this function should have put the task in its new group
7219 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7220 * reflect its new group.
7222 void sched_move_task(struct task_struct
*tsk
)
7224 struct task_group
*tg
;
7226 unsigned long flags
;
7229 rq
= task_rq_lock(tsk
, &flags
);
7231 running
= task_current(rq
, tsk
);
7235 dequeue_task(rq
, tsk
, 0);
7236 if (unlikely(running
))
7237 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7239 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7240 lockdep_is_held(&tsk
->sighand
->siglock
)),
7241 struct task_group
, css
);
7242 tg
= autogroup_task_group(tsk
, tg
);
7243 tsk
->sched_task_group
= tg
;
7245 #ifdef CONFIG_FAIR_GROUP_SCHED
7246 if (tsk
->sched_class
->task_move_group
)
7247 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7250 set_task_rq(tsk
, task_cpu(tsk
));
7252 if (unlikely(running
))
7253 tsk
->sched_class
->set_curr_task(rq
);
7255 enqueue_task(rq
, tsk
, 0);
7257 task_rq_unlock(rq
, tsk
, &flags
);
7259 #endif /* CONFIG_CGROUP_SCHED */
7261 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7262 static unsigned long to_ratio(u64 period
, u64 runtime
)
7264 if (runtime
== RUNTIME_INF
)
7267 return div64_u64(runtime
<< 20, period
);
7271 #ifdef CONFIG_RT_GROUP_SCHED
7273 * Ensure that the real time constraints are schedulable.
7275 static DEFINE_MUTEX(rt_constraints_mutex
);
7277 /* Must be called with tasklist_lock held */
7278 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7280 struct task_struct
*g
, *p
;
7282 do_each_thread(g
, p
) {
7283 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7285 } while_each_thread(g
, p
);
7290 struct rt_schedulable_data
{
7291 struct task_group
*tg
;
7296 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7298 struct rt_schedulable_data
*d
= data
;
7299 struct task_group
*child
;
7300 unsigned long total
, sum
= 0;
7301 u64 period
, runtime
;
7303 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7304 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7307 period
= d
->rt_period
;
7308 runtime
= d
->rt_runtime
;
7312 * Cannot have more runtime than the period.
7314 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7318 * Ensure we don't starve existing RT tasks.
7320 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7323 total
= to_ratio(period
, runtime
);
7326 * Nobody can have more than the global setting allows.
7328 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7332 * The sum of our children's runtime should not exceed our own.
7334 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7335 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7336 runtime
= child
->rt_bandwidth
.rt_runtime
;
7338 if (child
== d
->tg
) {
7339 period
= d
->rt_period
;
7340 runtime
= d
->rt_runtime
;
7343 sum
+= to_ratio(period
, runtime
);
7352 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7356 struct rt_schedulable_data data
= {
7358 .rt_period
= period
,
7359 .rt_runtime
= runtime
,
7363 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7369 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7370 u64 rt_period
, u64 rt_runtime
)
7374 mutex_lock(&rt_constraints_mutex
);
7375 read_lock(&tasklist_lock
);
7376 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7380 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7381 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7382 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7384 for_each_possible_cpu(i
) {
7385 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7387 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7388 rt_rq
->rt_runtime
= rt_runtime
;
7389 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7391 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7393 read_unlock(&tasklist_lock
);
7394 mutex_unlock(&rt_constraints_mutex
);
7399 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7401 u64 rt_runtime
, rt_period
;
7403 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7404 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7405 if (rt_runtime_us
< 0)
7406 rt_runtime
= RUNTIME_INF
;
7408 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7411 long sched_group_rt_runtime(struct task_group
*tg
)
7415 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7418 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7419 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7420 return rt_runtime_us
;
7423 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7425 u64 rt_runtime
, rt_period
;
7427 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7428 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7433 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7436 long sched_group_rt_period(struct task_group
*tg
)
7440 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7441 do_div(rt_period_us
, NSEC_PER_USEC
);
7442 return rt_period_us
;
7445 static int sched_rt_global_constraints(void)
7447 u64 runtime
, period
;
7450 if (sysctl_sched_rt_period
<= 0)
7453 runtime
= global_rt_runtime();
7454 period
= global_rt_period();
7457 * Sanity check on the sysctl variables.
7459 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7462 mutex_lock(&rt_constraints_mutex
);
7463 read_lock(&tasklist_lock
);
7464 ret
= __rt_schedulable(NULL
, 0, 0);
7465 read_unlock(&tasklist_lock
);
7466 mutex_unlock(&rt_constraints_mutex
);
7471 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7473 /* Don't accept realtime tasks when there is no way for them to run */
7474 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7480 #else /* !CONFIG_RT_GROUP_SCHED */
7481 static int sched_rt_global_constraints(void)
7483 unsigned long flags
;
7486 if (sysctl_sched_rt_period
<= 0)
7490 * There's always some RT tasks in the root group
7491 * -- migration, kstopmachine etc..
7493 if (sysctl_sched_rt_runtime
== 0)
7496 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7497 for_each_possible_cpu(i
) {
7498 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7500 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7501 rt_rq
->rt_runtime
= global_rt_runtime();
7502 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7504 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7508 #endif /* CONFIG_RT_GROUP_SCHED */
7510 int sched_rt_handler(struct ctl_table
*table
, int write
,
7511 void __user
*buffer
, size_t *lenp
,
7515 int old_period
, old_runtime
;
7516 static DEFINE_MUTEX(mutex
);
7519 old_period
= sysctl_sched_rt_period
;
7520 old_runtime
= sysctl_sched_rt_runtime
;
7522 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7524 if (!ret
&& write
) {
7525 ret
= sched_rt_global_constraints();
7527 sysctl_sched_rt_period
= old_period
;
7528 sysctl_sched_rt_runtime
= old_runtime
;
7530 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7531 def_rt_bandwidth
.rt_period
=
7532 ns_to_ktime(global_rt_period());
7535 mutex_unlock(&mutex
);
7540 #ifdef CONFIG_CGROUP_SCHED
7542 /* return corresponding task_group object of a cgroup */
7543 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7545 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7546 struct task_group
, css
);
7549 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7551 struct task_group
*tg
, *parent
;
7553 if (!cgrp
->parent
) {
7554 /* This is early initialization for the top cgroup */
7555 return &root_task_group
.css
;
7558 parent
= cgroup_tg(cgrp
->parent
);
7559 tg
= sched_create_group(parent
);
7561 return ERR_PTR(-ENOMEM
);
7566 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7568 struct task_group
*tg
= cgroup_tg(cgrp
);
7570 sched_destroy_group(tg
);
7573 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7574 struct cgroup_taskset
*tset
)
7576 struct task_struct
*task
;
7578 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7579 #ifdef CONFIG_RT_GROUP_SCHED
7580 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7583 /* We don't support RT-tasks being in separate groups */
7584 if (task
->sched_class
!= &fair_sched_class
)
7591 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7592 struct cgroup_taskset
*tset
)
7594 struct task_struct
*task
;
7596 cgroup_taskset_for_each(task
, cgrp
, tset
)
7597 sched_move_task(task
);
7601 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7602 struct task_struct
*task
)
7605 * cgroup_exit() is called in the copy_process() failure path.
7606 * Ignore this case since the task hasn't ran yet, this avoids
7607 * trying to poke a half freed task state from generic code.
7609 if (!(task
->flags
& PF_EXITING
))
7612 sched_move_task(task
);
7615 #ifdef CONFIG_FAIR_GROUP_SCHED
7616 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7619 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7622 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7624 struct task_group
*tg
= cgroup_tg(cgrp
);
7626 return (u64
) scale_load_down(tg
->shares
);
7629 #ifdef CONFIG_CFS_BANDWIDTH
7630 static DEFINE_MUTEX(cfs_constraints_mutex
);
7632 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7633 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7635 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7637 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7639 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7640 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7642 if (tg
== &root_task_group
)
7646 * Ensure we have at some amount of bandwidth every period. This is
7647 * to prevent reaching a state of large arrears when throttled via
7648 * entity_tick() resulting in prolonged exit starvation.
7650 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7654 * Likewise, bound things on the otherside by preventing insane quota
7655 * periods. This also allows us to normalize in computing quota
7658 if (period
> max_cfs_quota_period
)
7661 mutex_lock(&cfs_constraints_mutex
);
7662 ret
= __cfs_schedulable(tg
, period
, quota
);
7666 runtime_enabled
= quota
!= RUNTIME_INF
;
7667 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7668 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7669 raw_spin_lock_irq(&cfs_b
->lock
);
7670 cfs_b
->period
= ns_to_ktime(period
);
7671 cfs_b
->quota
= quota
;
7673 __refill_cfs_bandwidth_runtime(cfs_b
);
7674 /* restart the period timer (if active) to handle new period expiry */
7675 if (runtime_enabled
&& cfs_b
->timer_active
) {
7676 /* force a reprogram */
7677 cfs_b
->timer_active
= 0;
7678 __start_cfs_bandwidth(cfs_b
);
7680 raw_spin_unlock_irq(&cfs_b
->lock
);
7682 for_each_possible_cpu(i
) {
7683 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7684 struct rq
*rq
= cfs_rq
->rq
;
7686 raw_spin_lock_irq(&rq
->lock
);
7687 cfs_rq
->runtime_enabled
= runtime_enabled
;
7688 cfs_rq
->runtime_remaining
= 0;
7690 if (cfs_rq
->throttled
)
7691 unthrottle_cfs_rq(cfs_rq
);
7692 raw_spin_unlock_irq(&rq
->lock
);
7695 mutex_unlock(&cfs_constraints_mutex
);
7700 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7704 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7705 if (cfs_quota_us
< 0)
7706 quota
= RUNTIME_INF
;
7708 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7710 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7713 long tg_get_cfs_quota(struct task_group
*tg
)
7717 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7720 quota_us
= tg
->cfs_bandwidth
.quota
;
7721 do_div(quota_us
, NSEC_PER_USEC
);
7726 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7730 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7731 quota
= tg
->cfs_bandwidth
.quota
;
7733 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7736 long tg_get_cfs_period(struct task_group
*tg
)
7740 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7741 do_div(cfs_period_us
, NSEC_PER_USEC
);
7743 return cfs_period_us
;
7746 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7748 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7751 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7754 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7757 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7759 return tg_get_cfs_period(cgroup_tg(cgrp
));
7762 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7765 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7768 struct cfs_schedulable_data
{
7769 struct task_group
*tg
;
7774 * normalize group quota/period to be quota/max_period
7775 * note: units are usecs
7777 static u64
normalize_cfs_quota(struct task_group
*tg
,
7778 struct cfs_schedulable_data
*d
)
7786 period
= tg_get_cfs_period(tg
);
7787 quota
= tg_get_cfs_quota(tg
);
7790 /* note: these should typically be equivalent */
7791 if (quota
== RUNTIME_INF
|| quota
== -1)
7794 return to_ratio(period
, quota
);
7797 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7799 struct cfs_schedulable_data
*d
= data
;
7800 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7801 s64 quota
= 0, parent_quota
= -1;
7804 quota
= RUNTIME_INF
;
7806 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7808 quota
= normalize_cfs_quota(tg
, d
);
7809 parent_quota
= parent_b
->hierarchal_quota
;
7812 * ensure max(child_quota) <= parent_quota, inherit when no
7815 if (quota
== RUNTIME_INF
)
7816 quota
= parent_quota
;
7817 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7820 cfs_b
->hierarchal_quota
= quota
;
7825 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7828 struct cfs_schedulable_data data
= {
7834 if (quota
!= RUNTIME_INF
) {
7835 do_div(data
.period
, NSEC_PER_USEC
);
7836 do_div(data
.quota
, NSEC_PER_USEC
);
7840 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7846 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7847 struct cgroup_map_cb
*cb
)
7849 struct task_group
*tg
= cgroup_tg(cgrp
);
7850 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7852 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7853 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7854 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7858 #endif /* CONFIG_CFS_BANDWIDTH */
7859 #endif /* CONFIG_FAIR_GROUP_SCHED */
7861 #ifdef CONFIG_RT_GROUP_SCHED
7862 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7865 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7868 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7870 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7873 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7876 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7879 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7881 return sched_group_rt_period(cgroup_tg(cgrp
));
7883 #endif /* CONFIG_RT_GROUP_SCHED */
7885 static struct cftype cpu_files
[] = {
7886 #ifdef CONFIG_FAIR_GROUP_SCHED
7889 .read_u64
= cpu_shares_read_u64
,
7890 .write_u64
= cpu_shares_write_u64
,
7893 #ifdef CONFIG_CFS_BANDWIDTH
7895 .name
= "cfs_quota_us",
7896 .read_s64
= cpu_cfs_quota_read_s64
,
7897 .write_s64
= cpu_cfs_quota_write_s64
,
7900 .name
= "cfs_period_us",
7901 .read_u64
= cpu_cfs_period_read_u64
,
7902 .write_u64
= cpu_cfs_period_write_u64
,
7906 .read_map
= cpu_stats_show
,
7909 #ifdef CONFIG_RT_GROUP_SCHED
7911 .name
= "rt_runtime_us",
7912 .read_s64
= cpu_rt_runtime_read
,
7913 .write_s64
= cpu_rt_runtime_write
,
7916 .name
= "rt_period_us",
7917 .read_u64
= cpu_rt_period_read_uint
,
7918 .write_u64
= cpu_rt_period_write_uint
,
7924 struct cgroup_subsys cpu_cgroup_subsys
= {
7926 .css_alloc
= cpu_cgroup_css_alloc
,
7927 .css_free
= cpu_cgroup_css_free
,
7928 .can_attach
= cpu_cgroup_can_attach
,
7929 .attach
= cpu_cgroup_attach
,
7930 .exit
= cpu_cgroup_exit
,
7931 .subsys_id
= cpu_cgroup_subsys_id
,
7932 .base_cftypes
= cpu_files
,
7936 #endif /* CONFIG_CGROUP_SCHED */
7938 #ifdef CONFIG_CGROUP_CPUACCT
7941 * CPU accounting code for task groups.
7943 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7944 * (balbir@in.ibm.com).
7947 struct cpuacct root_cpuacct
;
7949 /* create a new cpu accounting group */
7950 static struct cgroup_subsys_state
*cpuacct_css_alloc(struct cgroup
*cgrp
)
7955 return &root_cpuacct
.css
;
7957 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7961 ca
->cpuusage
= alloc_percpu(u64
);
7965 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7967 goto out_free_cpuusage
;
7972 free_percpu(ca
->cpuusage
);
7976 return ERR_PTR(-ENOMEM
);
7979 /* destroy an existing cpu accounting group */
7980 static void cpuacct_css_free(struct cgroup
*cgrp
)
7982 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7984 free_percpu(ca
->cpustat
);
7985 free_percpu(ca
->cpuusage
);
7989 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
7991 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7994 #ifndef CONFIG_64BIT
7996 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7998 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8000 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8008 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8010 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8012 #ifndef CONFIG_64BIT
8014 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8016 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8018 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8024 /* return total cpu usage (in nanoseconds) of a group */
8025 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8027 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8028 u64 totalcpuusage
= 0;
8031 for_each_present_cpu(i
)
8032 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8034 return totalcpuusage
;
8037 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8040 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8049 for_each_present_cpu(i
)
8050 cpuacct_cpuusage_write(ca
, i
, 0);
8056 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8059 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8063 for_each_present_cpu(i
) {
8064 percpu
= cpuacct_cpuusage_read(ca
, i
);
8065 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8067 seq_printf(m
, "\n");
8071 static const char *cpuacct_stat_desc
[] = {
8072 [CPUACCT_STAT_USER
] = "user",
8073 [CPUACCT_STAT_SYSTEM
] = "system",
8076 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8077 struct cgroup_map_cb
*cb
)
8079 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8083 for_each_online_cpu(cpu
) {
8084 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8085 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8086 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8088 val
= cputime64_to_clock_t(val
);
8089 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8092 for_each_online_cpu(cpu
) {
8093 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8094 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8095 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8096 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8099 val
= cputime64_to_clock_t(val
);
8100 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8105 static struct cftype files
[] = {
8108 .read_u64
= cpuusage_read
,
8109 .write_u64
= cpuusage_write
,
8112 .name
= "usage_percpu",
8113 .read_seq_string
= cpuacct_percpu_seq_read
,
8117 .read_map
= cpuacct_stats_show
,
8123 * charge this task's execution time to its accounting group.
8125 * called with rq->lock held.
8127 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8132 if (unlikely(!cpuacct_subsys
.active
))
8135 cpu
= task_cpu(tsk
);
8141 for (; ca
; ca
= parent_ca(ca
)) {
8142 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8143 *cpuusage
+= cputime
;
8149 struct cgroup_subsys cpuacct_subsys
= {
8151 .css_alloc
= cpuacct_css_alloc
,
8152 .css_free
= cpuacct_css_free
,
8153 .subsys_id
= cpuacct_subsys_id
,
8154 .base_cftypes
= files
,
8156 #endif /* CONFIG_CGROUP_CPUACCT */
8158 void dump_cpu_task(int cpu
)
8160 pr_info("Task dump for CPU %d:\n", cpu
);
8161 sched_show_task(cpu_curr(cpu
));