4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
95 ktime_t soft
, hard
, now
;
98 if (hrtimer_active(period_timer
))
101 now
= hrtimer_cb_get_time(period_timer
);
102 hrtimer_forward(period_timer
, now
, period
);
104 soft
= hrtimer_get_softexpires(period_timer
);
105 hard
= hrtimer_get_expires(period_timer
);
106 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
107 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
108 HRTIMER_MODE_ABS_PINNED
, 0);
112 DEFINE_MUTEX(sched_domains_mutex
);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
115 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
117 void update_rq_clock(struct rq
*rq
)
121 if (rq
->skip_clock_update
> 0)
124 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
126 update_rq_clock_task(rq
, delta
);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug
unsigned int sysctl_sched_features
=
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names
[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp
)
201 if (strncmp(cmp
, "NO_", 3) == 0) {
206 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
207 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
209 sysctl_sched_features
&= ~(1UL << i
);
210 sched_feat_disable(i
);
212 sysctl_sched_features
|= (1UL << i
);
213 sched_feat_enable(i
);
223 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
224 size_t cnt
, loff_t
*ppos
)
233 if (copy_from_user(&buf
, ubuf
, cnt
))
239 i
= sched_feat_set(cmp
);
240 if (i
== __SCHED_FEAT_NR
)
248 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
250 return single_open(filp
, sched_feat_show
, NULL
);
253 static const struct file_operations sched_feat_fops
= {
254 .open
= sched_feat_open
,
255 .write
= sched_feat_write
,
258 .release
= single_release
,
261 static __init
int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
268 late_initcall(sched_init_debug
);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
278 * period over which we average the RT time consumption, measured
283 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period
= 1000000;
291 __read_mostly
int scheduler_running
;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime
= 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
309 lockdep_assert_held(&p
->pi_lock
);
313 raw_spin_lock(&rq
->lock
);
314 if (likely(rq
== task_rq(p
)))
316 raw_spin_unlock(&rq
->lock
);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
324 __acquires(p
->pi_lock
)
330 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
332 raw_spin_lock(&rq
->lock
);
333 if (likely(rq
== task_rq(p
)))
335 raw_spin_unlock(&rq
->lock
);
336 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
340 static void __task_rq_unlock(struct rq
*rq
)
343 raw_spin_unlock(&rq
->lock
);
347 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
349 __releases(p
->pi_lock
)
351 raw_spin_unlock(&rq
->lock
);
352 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq
*this_rq_lock(void)
365 raw_spin_lock(&rq
->lock
);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq
*rq
)
384 if (hrtimer_active(&rq
->hrtick_timer
))
385 hrtimer_cancel(&rq
->hrtick_timer
);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
394 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
396 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
398 raw_spin_lock(&rq
->lock
);
400 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
401 raw_spin_unlock(&rq
->lock
);
403 return HRTIMER_NORESTART
;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg
)
414 raw_spin_lock(&rq
->lock
);
415 hrtimer_restart(&rq
->hrtick_timer
);
416 rq
->hrtick_csd_pending
= 0;
417 raw_spin_unlock(&rq
->lock
);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq
*rq
, u64 delay
)
427 struct hrtimer
*timer
= &rq
->hrtick_timer
;
428 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
430 hrtimer_set_expires(timer
, time
);
432 if (rq
== this_rq()) {
433 hrtimer_restart(timer
);
434 } else if (!rq
->hrtick_csd_pending
) {
435 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
436 rq
->hrtick_csd_pending
= 1;
441 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
443 int cpu
= (int)(long)hcpu
;
446 case CPU_UP_CANCELED
:
447 case CPU_UP_CANCELED_FROZEN
:
448 case CPU_DOWN_PREPARE
:
449 case CPU_DOWN_PREPARE_FROZEN
:
451 case CPU_DEAD_FROZEN
:
452 hrtick_clear(cpu_rq(cpu
));
459 static __init
void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick
, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq
*rq
, u64 delay
)
471 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
472 HRTIMER_MODE_REL_PINNED
, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq
*rq
)
483 rq
->hrtick_csd_pending
= 0;
485 rq
->hrtick_csd
.flags
= 0;
486 rq
->hrtick_csd
.func
= __hrtick_start
;
487 rq
->hrtick_csd
.info
= rq
;
490 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
491 rq
->hrtick_timer
.function
= hrtick
;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq
*rq
)
498 static inline void init_rq_hrtick(struct rq
*rq
)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
516 #ifndef tsk_is_polling
517 #define tsk_is_polling(t) 0
520 void resched_task(struct task_struct
*p
)
524 assert_raw_spin_locked(&task_rq(p
)->lock
);
526 if (test_tsk_need_resched(p
))
529 set_tsk_need_resched(p
);
532 if (cpu
== smp_processor_id())
535 /* NEED_RESCHED must be visible before we test polling */
537 if (!tsk_is_polling(p
))
538 smp_send_reschedule(cpu
);
541 void resched_cpu(int cpu
)
543 struct rq
*rq
= cpu_rq(cpu
);
546 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
548 resched_task(cpu_curr(cpu
));
549 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
554 * In the semi idle case, use the nearest busy cpu for migrating timers
555 * from an idle cpu. This is good for power-savings.
557 * We don't do similar optimization for completely idle system, as
558 * selecting an idle cpu will add more delays to the timers than intended
559 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
561 int get_nohz_timer_target(void)
563 int cpu
= smp_processor_id();
565 struct sched_domain
*sd
;
568 for_each_domain(cpu
, sd
) {
569 for_each_cpu(i
, sched_domain_span(sd
)) {
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 void wake_up_idle_cpu(int cpu
)
592 struct rq
*rq
= cpu_rq(cpu
);
594 if (cpu
== smp_processor_id())
598 * This is safe, as this function is called with the timer
599 * wheel base lock of (cpu) held. When the CPU is on the way
600 * to idle and has not yet set rq->curr to idle then it will
601 * be serialized on the timer wheel base lock and take the new
602 * timer into account automatically.
604 if (rq
->curr
!= rq
->idle
)
608 * We can set TIF_RESCHED on the idle task of the other CPU
609 * lockless. The worst case is that the other CPU runs the
610 * idle task through an additional NOOP schedule()
612 set_tsk_need_resched(rq
->idle
);
614 /* NEED_RESCHED must be visible before we test polling */
616 if (!tsk_is_polling(rq
->idle
))
617 smp_send_reschedule(cpu
);
620 static inline bool got_nohz_idle_kick(void)
622 int cpu
= smp_processor_id();
623 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
626 #else /* CONFIG_NO_HZ */
628 static inline bool got_nohz_idle_kick(void)
633 #endif /* CONFIG_NO_HZ */
635 void sched_avg_update(struct rq
*rq
)
637 s64 period
= sched_avg_period();
639 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
641 * Inline assembly required to prevent the compiler
642 * optimising this loop into a divmod call.
643 * See __iter_div_u64_rem() for another example of this.
645 asm("" : "+rm" (rq
->age_stamp
));
646 rq
->age_stamp
+= period
;
651 #else /* !CONFIG_SMP */
652 void resched_task(struct task_struct
*p
)
654 assert_raw_spin_locked(&task_rq(p
)->lock
);
655 set_tsk_need_resched(p
);
657 #endif /* CONFIG_SMP */
659 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
662 * Iterate task_group tree rooted at *from, calling @down when first entering a
663 * node and @up when leaving it for the final time.
665 * Caller must hold rcu_lock or sufficient equivalent.
667 int walk_tg_tree_from(struct task_group
*from
,
668 tg_visitor down
, tg_visitor up
, void *data
)
670 struct task_group
*parent
, *child
;
676 ret
= (*down
)(parent
, data
);
679 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
686 ret
= (*up
)(parent
, data
);
687 if (ret
|| parent
== from
)
691 parent
= parent
->parent
;
698 int tg_nop(struct task_group
*tg
, void *data
)
704 static void set_load_weight(struct task_struct
*p
)
706 int prio
= p
->static_prio
- MAX_RT_PRIO
;
707 struct load_weight
*load
= &p
->se
.load
;
710 * SCHED_IDLE tasks get minimal weight:
712 if (p
->policy
== SCHED_IDLE
) {
713 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
714 load
->inv_weight
= WMULT_IDLEPRIO
;
718 load
->weight
= scale_load(prio_to_weight
[prio
]);
719 load
->inv_weight
= prio_to_wmult
[prio
];
722 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
725 sched_info_queued(p
);
726 p
->sched_class
->enqueue_task(rq
, p
, flags
);
729 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
732 sched_info_dequeued(p
);
733 p
->sched_class
->dequeue_task(rq
, p
, flags
);
736 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
738 if (task_contributes_to_load(p
))
739 rq
->nr_uninterruptible
--;
741 enqueue_task(rq
, p
, flags
);
744 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
746 if (task_contributes_to_load(p
))
747 rq
->nr_uninterruptible
++;
749 dequeue_task(rq
, p
, flags
);
752 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
755 * In theory, the compile should just see 0 here, and optimize out the call
756 * to sched_rt_avg_update. But I don't trust it...
758 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
759 s64 steal
= 0, irq_delta
= 0;
761 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
762 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
765 * Since irq_time is only updated on {soft,}irq_exit, we might run into
766 * this case when a previous update_rq_clock() happened inside a
769 * When this happens, we stop ->clock_task and only update the
770 * prev_irq_time stamp to account for the part that fit, so that a next
771 * update will consume the rest. This ensures ->clock_task is
774 * It does however cause some slight miss-attribution of {soft,}irq
775 * time, a more accurate solution would be to update the irq_time using
776 * the current rq->clock timestamp, except that would require using
779 if (irq_delta
> delta
)
782 rq
->prev_irq_time
+= irq_delta
;
785 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
786 if (static_key_false((¶virt_steal_rq_enabled
))) {
789 steal
= paravirt_steal_clock(cpu_of(rq
));
790 steal
-= rq
->prev_steal_time_rq
;
792 if (unlikely(steal
> delta
))
795 st
= steal_ticks(steal
);
796 steal
= st
* TICK_NSEC
;
798 rq
->prev_steal_time_rq
+= steal
;
804 rq
->clock_task
+= delta
;
806 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
807 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
808 sched_rt_avg_update(rq
, irq_delta
+ steal
);
812 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
814 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
815 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
819 * Make it appear like a SCHED_FIFO task, its something
820 * userspace knows about and won't get confused about.
822 * Also, it will make PI more or less work without too
823 * much confusion -- but then, stop work should not
824 * rely on PI working anyway.
826 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
828 stop
->sched_class
= &stop_sched_class
;
831 cpu_rq(cpu
)->stop
= stop
;
835 * Reset it back to a normal scheduling class so that
836 * it can die in pieces.
838 old_stop
->sched_class
= &rt_sched_class
;
843 * __normal_prio - return the priority that is based on the static prio
845 static inline int __normal_prio(struct task_struct
*p
)
847 return p
->static_prio
;
851 * Calculate the expected normal priority: i.e. priority
852 * without taking RT-inheritance into account. Might be
853 * boosted by interactivity modifiers. Changes upon fork,
854 * setprio syscalls, and whenever the interactivity
855 * estimator recalculates.
857 static inline int normal_prio(struct task_struct
*p
)
861 if (task_has_rt_policy(p
))
862 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
864 prio
= __normal_prio(p
);
869 * Calculate the current priority, i.e. the priority
870 * taken into account by the scheduler. This value might
871 * be boosted by RT tasks, or might be boosted by
872 * interactivity modifiers. Will be RT if the task got
873 * RT-boosted. If not then it returns p->normal_prio.
875 static int effective_prio(struct task_struct
*p
)
877 p
->normal_prio
= normal_prio(p
);
879 * If we are RT tasks or we were boosted to RT priority,
880 * keep the priority unchanged. Otherwise, update priority
881 * to the normal priority:
883 if (!rt_prio(p
->prio
))
884 return p
->normal_prio
;
889 * task_curr - is this task currently executing on a CPU?
890 * @p: the task in question.
892 inline int task_curr(const struct task_struct
*p
)
894 return cpu_curr(task_cpu(p
)) == p
;
897 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
898 const struct sched_class
*prev_class
,
901 if (prev_class
!= p
->sched_class
) {
902 if (prev_class
->switched_from
)
903 prev_class
->switched_from(rq
, p
);
904 p
->sched_class
->switched_to(rq
, p
);
905 } else if (oldprio
!= p
->prio
)
906 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
909 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
911 const struct sched_class
*class;
913 if (p
->sched_class
== rq
->curr
->sched_class
) {
914 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
916 for_each_class(class) {
917 if (class == rq
->curr
->sched_class
)
919 if (class == p
->sched_class
) {
920 resched_task(rq
->curr
);
927 * A queue event has occurred, and we're going to schedule. In
928 * this case, we can save a useless back to back clock update.
930 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
931 rq
->skip_clock_update
= 1;
934 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
936 void register_task_migration_notifier(struct notifier_block
*n
)
938 atomic_notifier_chain_register(&task_migration_notifier
, n
);
942 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
944 #ifdef CONFIG_SCHED_DEBUG
946 * We should never call set_task_cpu() on a blocked task,
947 * ttwu() will sort out the placement.
949 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
950 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
952 #ifdef CONFIG_LOCKDEP
954 * The caller should hold either p->pi_lock or rq->lock, when changing
955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
957 * sched_move_task() holds both and thus holding either pins the cgroup,
960 * Furthermore, all task_rq users should acquire both locks, see
963 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
964 lockdep_is_held(&task_rq(p
)->lock
)));
968 trace_sched_migrate_task(p
, new_cpu
);
970 if (task_cpu(p
) != new_cpu
) {
971 struct task_migration_notifier tmn
;
973 if (p
->sched_class
->migrate_task_rq
)
974 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
975 p
->se
.nr_migrations
++;
976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
979 tmn
.from_cpu
= task_cpu(p
);
980 tmn
.to_cpu
= new_cpu
;
982 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
985 __set_task_cpu(p
, new_cpu
);
988 struct migration_arg
{
989 struct task_struct
*task
;
993 static int migration_cpu_stop(void *data
);
996 * wait_task_inactive - wait for a thread to unschedule.
998 * If @match_state is nonzero, it's the @p->state value just checked and
999 * not expected to change. If it changes, i.e. @p might have woken up,
1000 * then return zero. When we succeed in waiting for @p to be off its CPU,
1001 * we return a positive number (its total switch count). If a second call
1002 * a short while later returns the same number, the caller can be sure that
1003 * @p has remained unscheduled the whole time.
1005 * The caller must ensure that the task *will* unschedule sometime soon,
1006 * else this function might spin for a *long* time. This function can't
1007 * be called with interrupts off, or it may introduce deadlock with
1008 * smp_call_function() if an IPI is sent by the same process we are
1009 * waiting to become inactive.
1011 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1013 unsigned long flags
;
1020 * We do the initial early heuristics without holding
1021 * any task-queue locks at all. We'll only try to get
1022 * the runqueue lock when things look like they will
1028 * If the task is actively running on another CPU
1029 * still, just relax and busy-wait without holding
1032 * NOTE! Since we don't hold any locks, it's not
1033 * even sure that "rq" stays as the right runqueue!
1034 * But we don't care, since "task_running()" will
1035 * return false if the runqueue has changed and p
1036 * is actually now running somewhere else!
1038 while (task_running(rq
, p
)) {
1039 if (match_state
&& unlikely(p
->state
!= match_state
))
1045 * Ok, time to look more closely! We need the rq
1046 * lock now, to be *sure*. If we're wrong, we'll
1047 * just go back and repeat.
1049 rq
= task_rq_lock(p
, &flags
);
1050 trace_sched_wait_task(p
);
1051 running
= task_running(rq
, p
);
1054 if (!match_state
|| p
->state
== match_state
)
1055 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1056 task_rq_unlock(rq
, p
, &flags
);
1059 * If it changed from the expected state, bail out now.
1061 if (unlikely(!ncsw
))
1065 * Was it really running after all now that we
1066 * checked with the proper locks actually held?
1068 * Oops. Go back and try again..
1070 if (unlikely(running
)) {
1076 * It's not enough that it's not actively running,
1077 * it must be off the runqueue _entirely_, and not
1080 * So if it was still runnable (but just not actively
1081 * running right now), it's preempted, and we should
1082 * yield - it could be a while.
1084 if (unlikely(on_rq
)) {
1085 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1087 set_current_state(TASK_UNINTERRUPTIBLE
);
1088 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1093 * Ahh, all good. It wasn't running, and it wasn't
1094 * runnable, which means that it will never become
1095 * running in the future either. We're all done!
1104 * kick_process - kick a running thread to enter/exit the kernel
1105 * @p: the to-be-kicked thread
1107 * Cause a process which is running on another CPU to enter
1108 * kernel-mode, without any delay. (to get signals handled.)
1110 * NOTE: this function doesn't have to take the runqueue lock,
1111 * because all it wants to ensure is that the remote task enters
1112 * the kernel. If the IPI races and the task has been migrated
1113 * to another CPU then no harm is done and the purpose has been
1116 void kick_process(struct task_struct
*p
)
1122 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1123 smp_send_reschedule(cpu
);
1126 EXPORT_SYMBOL_GPL(kick_process
);
1127 #endif /* CONFIG_SMP */
1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1133 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1135 int nid
= cpu_to_node(cpu
);
1136 const struct cpumask
*nodemask
= NULL
;
1137 enum { cpuset
, possible
, fail
} state
= cpuset
;
1141 * If the node that the cpu is on has been offlined, cpu_to_node()
1142 * will return -1. There is no cpu on the node, and we should
1143 * select the cpu on the other node.
1146 nodemask
= cpumask_of_node(nid
);
1148 /* Look for allowed, online CPU in same node. */
1149 for_each_cpu(dest_cpu
, nodemask
) {
1150 if (!cpu_online(dest_cpu
))
1152 if (!cpu_active(dest_cpu
))
1154 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1160 /* Any allowed, online CPU? */
1161 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1162 if (!cpu_online(dest_cpu
))
1164 if (!cpu_active(dest_cpu
))
1171 /* No more Mr. Nice Guy. */
1172 cpuset_cpus_allowed_fallback(p
);
1177 do_set_cpus_allowed(p
, cpu_possible_mask
);
1188 if (state
!= cpuset
) {
1190 * Don't tell them about moving exiting tasks or
1191 * kernel threads (both mm NULL), since they never
1194 if (p
->mm
&& printk_ratelimit()) {
1195 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1196 task_pid_nr(p
), p
->comm
, cpu
);
1204 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1207 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1209 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1212 * In order not to call set_task_cpu() on a blocking task we need
1213 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1216 * Since this is common to all placement strategies, this lives here.
1218 * [ this allows ->select_task() to simply return task_cpu(p) and
1219 * not worry about this generic constraint ]
1221 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1223 cpu
= select_fallback_rq(task_cpu(p
), p
);
1228 static void update_avg(u64
*avg
, u64 sample
)
1230 s64 diff
= sample
- *avg
;
1236 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1238 #ifdef CONFIG_SCHEDSTATS
1239 struct rq
*rq
= this_rq();
1242 int this_cpu
= smp_processor_id();
1244 if (cpu
== this_cpu
) {
1245 schedstat_inc(rq
, ttwu_local
);
1246 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1248 struct sched_domain
*sd
;
1250 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1252 for_each_domain(this_cpu
, sd
) {
1253 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1254 schedstat_inc(sd
, ttwu_wake_remote
);
1261 if (wake_flags
& WF_MIGRATED
)
1262 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1264 #endif /* CONFIG_SMP */
1266 schedstat_inc(rq
, ttwu_count
);
1267 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1269 if (wake_flags
& WF_SYNC
)
1270 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1272 #endif /* CONFIG_SCHEDSTATS */
1275 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1277 activate_task(rq
, p
, en_flags
);
1280 /* if a worker is waking up, notify workqueue */
1281 if (p
->flags
& PF_WQ_WORKER
)
1282 wq_worker_waking_up(p
, cpu_of(rq
));
1286 * Mark the task runnable and perform wakeup-preemption.
1289 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1291 check_preempt_curr(rq
, p
, wake_flags
);
1292 trace_sched_wakeup(p
, true);
1294 p
->state
= TASK_RUNNING
;
1296 if (p
->sched_class
->task_woken
)
1297 p
->sched_class
->task_woken(rq
, p
);
1299 if (rq
->idle_stamp
) {
1300 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1301 u64 max
= 2*sysctl_sched_migration_cost
;
1306 update_avg(&rq
->avg_idle
, delta
);
1313 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1316 if (p
->sched_contributes_to_load
)
1317 rq
->nr_uninterruptible
--;
1320 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1321 ttwu_do_wakeup(rq
, p
, wake_flags
);
1325 * Called in case the task @p isn't fully descheduled from its runqueue,
1326 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1327 * since all we need to do is flip p->state to TASK_RUNNING, since
1328 * the task is still ->on_rq.
1330 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1335 rq
= __task_rq_lock(p
);
1337 ttwu_do_wakeup(rq
, p
, wake_flags
);
1340 __task_rq_unlock(rq
);
1346 static void sched_ttwu_pending(void)
1348 struct rq
*rq
= this_rq();
1349 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1350 struct task_struct
*p
;
1352 raw_spin_lock(&rq
->lock
);
1355 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1356 llist
= llist_next(llist
);
1357 ttwu_do_activate(rq
, p
, 0);
1360 raw_spin_unlock(&rq
->lock
);
1363 void scheduler_ipi(void)
1365 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1369 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1370 * traditionally all their work was done from the interrupt return
1371 * path. Now that we actually do some work, we need to make sure
1374 * Some archs already do call them, luckily irq_enter/exit nest
1377 * Arguably we should visit all archs and update all handlers,
1378 * however a fair share of IPIs are still resched only so this would
1379 * somewhat pessimize the simple resched case.
1382 sched_ttwu_pending();
1385 * Check if someone kicked us for doing the nohz idle load balance.
1387 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1388 this_rq()->idle_balance
= 1;
1389 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1394 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1396 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1397 smp_send_reschedule(cpu
);
1400 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1402 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1404 #endif /* CONFIG_SMP */
1406 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1408 struct rq
*rq
= cpu_rq(cpu
);
1410 #if defined(CONFIG_SMP)
1411 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1412 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1413 ttwu_queue_remote(p
, cpu
);
1418 raw_spin_lock(&rq
->lock
);
1419 ttwu_do_activate(rq
, p
, 0);
1420 raw_spin_unlock(&rq
->lock
);
1424 * try_to_wake_up - wake up a thread
1425 * @p: the thread to be awakened
1426 * @state: the mask of task states that can be woken
1427 * @wake_flags: wake modifier flags (WF_*)
1429 * Put it on the run-queue if it's not already there. The "current"
1430 * thread is always on the run-queue (except when the actual
1431 * re-schedule is in progress), and as such you're allowed to do
1432 * the simpler "current->state = TASK_RUNNING" to mark yourself
1433 * runnable without the overhead of this.
1435 * Returns %true if @p was woken up, %false if it was already running
1436 * or @state didn't match @p's state.
1439 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1441 unsigned long flags
;
1442 int cpu
, success
= 0;
1445 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1446 if (!(p
->state
& state
))
1449 success
= 1; /* we're going to change ->state */
1452 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1457 * If the owning (remote) cpu is still in the middle of schedule() with
1458 * this task as prev, wait until its done referencing the task.
1463 * Pairs with the smp_wmb() in finish_lock_switch().
1467 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1468 p
->state
= TASK_WAKING
;
1470 if (p
->sched_class
->task_waking
)
1471 p
->sched_class
->task_waking(p
);
1473 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1474 if (task_cpu(p
) != cpu
) {
1475 wake_flags
|= WF_MIGRATED
;
1476 set_task_cpu(p
, cpu
);
1478 #endif /* CONFIG_SMP */
1482 ttwu_stat(p
, cpu
, wake_flags
);
1484 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1490 * try_to_wake_up_local - try to wake up a local task with rq lock held
1491 * @p: the thread to be awakened
1493 * Put @p on the run-queue if it's not already there. The caller must
1494 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1497 static void try_to_wake_up_local(struct task_struct
*p
)
1499 struct rq
*rq
= task_rq(p
);
1501 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1502 WARN_ON_ONCE(p
== current
))
1505 lockdep_assert_held(&rq
->lock
);
1507 if (!raw_spin_trylock(&p
->pi_lock
)) {
1508 raw_spin_unlock(&rq
->lock
);
1509 raw_spin_lock(&p
->pi_lock
);
1510 raw_spin_lock(&rq
->lock
);
1513 if (!(p
->state
& TASK_NORMAL
))
1517 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1519 ttwu_do_wakeup(rq
, p
, 0);
1520 ttwu_stat(p
, smp_processor_id(), 0);
1522 raw_spin_unlock(&p
->pi_lock
);
1526 * wake_up_process - Wake up a specific process
1527 * @p: The process to be woken up.
1529 * Attempt to wake up the nominated process and move it to the set of runnable
1530 * processes. Returns 1 if the process was woken up, 0 if it was already
1533 * It may be assumed that this function implies a write memory barrier before
1534 * changing the task state if and only if any tasks are woken up.
1536 int wake_up_process(struct task_struct
*p
)
1538 WARN_ON(task_is_stopped_or_traced(p
));
1539 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1541 EXPORT_SYMBOL(wake_up_process
);
1543 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1545 return try_to_wake_up(p
, state
, 0);
1549 * Perform scheduler related setup for a newly forked process p.
1550 * p is forked by current.
1552 * __sched_fork() is basic setup used by init_idle() too:
1554 static void __sched_fork(struct task_struct
*p
)
1559 p
->se
.exec_start
= 0;
1560 p
->se
.sum_exec_runtime
= 0;
1561 p
->se
.prev_sum_exec_runtime
= 0;
1562 p
->se
.nr_migrations
= 0;
1564 INIT_LIST_HEAD(&p
->se
.group_node
);
1567 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1568 * removed when useful for applications beyond shares distribution (e.g.
1571 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1572 p
->se
.avg
.runnable_avg_period
= 0;
1573 p
->se
.avg
.runnable_avg_sum
= 0;
1575 #ifdef CONFIG_SCHEDSTATS
1576 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1579 INIT_LIST_HEAD(&p
->rt
.run_list
);
1581 #ifdef CONFIG_PREEMPT_NOTIFIERS
1582 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1585 #ifdef CONFIG_NUMA_BALANCING
1586 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1587 p
->mm
->numa_next_scan
= jiffies
;
1588 p
->mm
->numa_next_reset
= jiffies
;
1589 p
->mm
->numa_scan_seq
= 0;
1592 p
->node_stamp
= 0ULL;
1593 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1594 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1595 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1596 p
->numa_work
.next
= &p
->numa_work
;
1597 #endif /* CONFIG_NUMA_BALANCING */
1600 #ifdef CONFIG_NUMA_BALANCING
1601 #ifdef CONFIG_SCHED_DEBUG
1602 void set_numabalancing_state(bool enabled
)
1605 sched_feat_set("NUMA");
1607 sched_feat_set("NO_NUMA");
1610 __read_mostly
bool numabalancing_enabled
;
1612 void set_numabalancing_state(bool enabled
)
1614 numabalancing_enabled
= enabled
;
1616 #endif /* CONFIG_SCHED_DEBUG */
1617 #endif /* CONFIG_NUMA_BALANCING */
1620 * fork()/clone()-time setup:
1622 void sched_fork(struct task_struct
*p
)
1624 unsigned long flags
;
1625 int cpu
= get_cpu();
1629 * We mark the process as running here. This guarantees that
1630 * nobody will actually run it, and a signal or other external
1631 * event cannot wake it up and insert it on the runqueue either.
1633 p
->state
= TASK_RUNNING
;
1636 * Make sure we do not leak PI boosting priority to the child.
1638 p
->prio
= current
->normal_prio
;
1641 * Revert to default priority/policy on fork if requested.
1643 if (unlikely(p
->sched_reset_on_fork
)) {
1644 if (task_has_rt_policy(p
)) {
1645 p
->policy
= SCHED_NORMAL
;
1646 p
->static_prio
= NICE_TO_PRIO(0);
1648 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1649 p
->static_prio
= NICE_TO_PRIO(0);
1651 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1655 * We don't need the reset flag anymore after the fork. It has
1656 * fulfilled its duty:
1658 p
->sched_reset_on_fork
= 0;
1661 if (!rt_prio(p
->prio
))
1662 p
->sched_class
= &fair_sched_class
;
1664 if (p
->sched_class
->task_fork
)
1665 p
->sched_class
->task_fork(p
);
1668 * The child is not yet in the pid-hash so no cgroup attach races,
1669 * and the cgroup is pinned to this child due to cgroup_fork()
1670 * is ran before sched_fork().
1672 * Silence PROVE_RCU.
1674 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1675 set_task_cpu(p
, cpu
);
1676 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1678 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1679 if (likely(sched_info_on()))
1680 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1682 #if defined(CONFIG_SMP)
1685 #ifdef CONFIG_PREEMPT_COUNT
1686 /* Want to start with kernel preemption disabled. */
1687 task_thread_info(p
)->preempt_count
= 1;
1690 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1697 * wake_up_new_task - wake up a newly created task for the first time.
1699 * This function will do some initial scheduler statistics housekeeping
1700 * that must be done for every newly created context, then puts the task
1701 * on the runqueue and wakes it.
1703 void wake_up_new_task(struct task_struct
*p
)
1705 unsigned long flags
;
1708 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1711 * Fork balancing, do it here and not earlier because:
1712 * - cpus_allowed can change in the fork path
1713 * - any previously selected cpu might disappear through hotplug
1715 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1718 rq
= __task_rq_lock(p
);
1719 activate_task(rq
, p
, 0);
1721 trace_sched_wakeup_new(p
, true);
1722 check_preempt_curr(rq
, p
, WF_FORK
);
1724 if (p
->sched_class
->task_woken
)
1725 p
->sched_class
->task_woken(rq
, p
);
1727 task_rq_unlock(rq
, p
, &flags
);
1730 #ifdef CONFIG_PREEMPT_NOTIFIERS
1733 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1734 * @notifier: notifier struct to register
1736 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1738 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1740 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1743 * preempt_notifier_unregister - no longer interested in preemption notifications
1744 * @notifier: notifier struct to unregister
1746 * This is safe to call from within a preemption notifier.
1748 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1750 hlist_del(¬ifier
->link
);
1752 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1754 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1756 struct preempt_notifier
*notifier
;
1758 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1759 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1763 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1764 struct task_struct
*next
)
1766 struct preempt_notifier
*notifier
;
1768 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1769 notifier
->ops
->sched_out(notifier
, next
);
1772 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1774 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1779 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1780 struct task_struct
*next
)
1784 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1787 * prepare_task_switch - prepare to switch tasks
1788 * @rq: the runqueue preparing to switch
1789 * @prev: the current task that is being switched out
1790 * @next: the task we are going to switch to.
1792 * This is called with the rq lock held and interrupts off. It must
1793 * be paired with a subsequent finish_task_switch after the context
1796 * prepare_task_switch sets up locking and calls architecture specific
1800 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1801 struct task_struct
*next
)
1803 trace_sched_switch(prev
, next
);
1804 sched_info_switch(prev
, next
);
1805 perf_event_task_sched_out(prev
, next
);
1806 fire_sched_out_preempt_notifiers(prev
, next
);
1807 prepare_lock_switch(rq
, next
);
1808 prepare_arch_switch(next
);
1812 * finish_task_switch - clean up after a task-switch
1813 * @rq: runqueue associated with task-switch
1814 * @prev: the thread we just switched away from.
1816 * finish_task_switch must be called after the context switch, paired
1817 * with a prepare_task_switch call before the context switch.
1818 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1819 * and do any other architecture-specific cleanup actions.
1821 * Note that we may have delayed dropping an mm in context_switch(). If
1822 * so, we finish that here outside of the runqueue lock. (Doing it
1823 * with the lock held can cause deadlocks; see schedule() for
1826 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1827 __releases(rq
->lock
)
1829 struct mm_struct
*mm
= rq
->prev_mm
;
1835 * A task struct has one reference for the use as "current".
1836 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1837 * schedule one last time. The schedule call will never return, and
1838 * the scheduled task must drop that reference.
1839 * The test for TASK_DEAD must occur while the runqueue locks are
1840 * still held, otherwise prev could be scheduled on another cpu, die
1841 * there before we look at prev->state, and then the reference would
1843 * Manfred Spraul <manfred@colorfullife.com>
1845 prev_state
= prev
->state
;
1846 vtime_task_switch(prev
);
1847 finish_arch_switch(prev
);
1848 perf_event_task_sched_in(prev
, current
);
1849 finish_lock_switch(rq
, prev
);
1850 finish_arch_post_lock_switch();
1852 fire_sched_in_preempt_notifiers(current
);
1855 if (unlikely(prev_state
== TASK_DEAD
)) {
1857 * Remove function-return probe instances associated with this
1858 * task and put them back on the free list.
1860 kprobe_flush_task(prev
);
1861 put_task_struct(prev
);
1867 /* assumes rq->lock is held */
1868 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1870 if (prev
->sched_class
->pre_schedule
)
1871 prev
->sched_class
->pre_schedule(rq
, prev
);
1874 /* rq->lock is NOT held, but preemption is disabled */
1875 static inline void post_schedule(struct rq
*rq
)
1877 if (rq
->post_schedule
) {
1878 unsigned long flags
;
1880 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1881 if (rq
->curr
->sched_class
->post_schedule
)
1882 rq
->curr
->sched_class
->post_schedule(rq
);
1883 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1885 rq
->post_schedule
= 0;
1891 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1895 static inline void post_schedule(struct rq
*rq
)
1902 * schedule_tail - first thing a freshly forked thread must call.
1903 * @prev: the thread we just switched away from.
1905 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1906 __releases(rq
->lock
)
1908 struct rq
*rq
= this_rq();
1910 finish_task_switch(rq
, prev
);
1913 * FIXME: do we need to worry about rq being invalidated by the
1918 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1919 /* In this case, finish_task_switch does not reenable preemption */
1922 if (current
->set_child_tid
)
1923 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1927 * context_switch - switch to the new MM and the new
1928 * thread's register state.
1931 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1932 struct task_struct
*next
)
1934 struct mm_struct
*mm
, *oldmm
;
1936 prepare_task_switch(rq
, prev
, next
);
1939 oldmm
= prev
->active_mm
;
1941 * For paravirt, this is coupled with an exit in switch_to to
1942 * combine the page table reload and the switch backend into
1945 arch_start_context_switch(prev
);
1948 next
->active_mm
= oldmm
;
1949 atomic_inc(&oldmm
->mm_count
);
1950 enter_lazy_tlb(oldmm
, next
);
1952 switch_mm(oldmm
, mm
, next
);
1955 prev
->active_mm
= NULL
;
1956 rq
->prev_mm
= oldmm
;
1959 * Since the runqueue lock will be released by the next
1960 * task (which is an invalid locking op but in the case
1961 * of the scheduler it's an obvious special-case), so we
1962 * do an early lockdep release here:
1964 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1965 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1968 context_tracking_task_switch(prev
, next
);
1969 /* Here we just switch the register state and the stack. */
1970 switch_to(prev
, next
, prev
);
1974 * this_rq must be evaluated again because prev may have moved
1975 * CPUs since it called schedule(), thus the 'rq' on its stack
1976 * frame will be invalid.
1978 finish_task_switch(this_rq(), prev
);
1982 * nr_running and nr_context_switches:
1984 * externally visible scheduler statistics: current number of runnable
1985 * threads, total number of context switches performed since bootup.
1987 unsigned long nr_running(void)
1989 unsigned long i
, sum
= 0;
1991 for_each_online_cpu(i
)
1992 sum
+= cpu_rq(i
)->nr_running
;
1997 unsigned long long nr_context_switches(void)
2000 unsigned long long sum
= 0;
2002 for_each_possible_cpu(i
)
2003 sum
+= cpu_rq(i
)->nr_switches
;
2008 unsigned long nr_iowait(void)
2010 unsigned long i
, sum
= 0;
2012 for_each_possible_cpu(i
)
2013 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2018 unsigned long nr_iowait_cpu(int cpu
)
2020 struct rq
*this = cpu_rq(cpu
);
2021 return atomic_read(&this->nr_iowait
);
2024 unsigned long this_cpu_load(void)
2026 struct rq
*this = this_rq();
2027 return this->cpu_load
[0];
2032 * Global load-average calculations
2034 * We take a distributed and async approach to calculating the global load-avg
2035 * in order to minimize overhead.
2037 * The global load average is an exponentially decaying average of nr_running +
2038 * nr_uninterruptible.
2040 * Once every LOAD_FREQ:
2043 * for_each_possible_cpu(cpu)
2044 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2046 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2048 * Due to a number of reasons the above turns in the mess below:
2050 * - for_each_possible_cpu() is prohibitively expensive on machines with
2051 * serious number of cpus, therefore we need to take a distributed approach
2052 * to calculating nr_active.
2054 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2055 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2057 * So assuming nr_active := 0 when we start out -- true per definition, we
2058 * can simply take per-cpu deltas and fold those into a global accumulate
2059 * to obtain the same result. See calc_load_fold_active().
2061 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2062 * across the machine, we assume 10 ticks is sufficient time for every
2063 * cpu to have completed this task.
2065 * This places an upper-bound on the IRQ-off latency of the machine. Then
2066 * again, being late doesn't loose the delta, just wrecks the sample.
2068 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2069 * this would add another cross-cpu cacheline miss and atomic operation
2070 * to the wakeup path. Instead we increment on whatever cpu the task ran
2071 * when it went into uninterruptible state and decrement on whatever cpu
2072 * did the wakeup. This means that only the sum of nr_uninterruptible over
2073 * all cpus yields the correct result.
2075 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2078 /* Variables and functions for calc_load */
2079 static atomic_long_t calc_load_tasks
;
2080 static unsigned long calc_load_update
;
2081 unsigned long avenrun
[3];
2082 EXPORT_SYMBOL(avenrun
); /* should be removed */
2085 * get_avenrun - get the load average array
2086 * @loads: pointer to dest load array
2087 * @offset: offset to add
2088 * @shift: shift count to shift the result left
2090 * These values are estimates at best, so no need for locking.
2092 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2094 loads
[0] = (avenrun
[0] + offset
) << shift
;
2095 loads
[1] = (avenrun
[1] + offset
) << shift
;
2096 loads
[2] = (avenrun
[2] + offset
) << shift
;
2099 static long calc_load_fold_active(struct rq
*this_rq
)
2101 long nr_active
, delta
= 0;
2103 nr_active
= this_rq
->nr_running
;
2104 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2106 if (nr_active
!= this_rq
->calc_load_active
) {
2107 delta
= nr_active
- this_rq
->calc_load_active
;
2108 this_rq
->calc_load_active
= nr_active
;
2115 * a1 = a0 * e + a * (1 - e)
2117 static unsigned long
2118 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2121 load
+= active
* (FIXED_1
- exp
);
2122 load
+= 1UL << (FSHIFT
- 1);
2123 return load
>> FSHIFT
;
2128 * Handle NO_HZ for the global load-average.
2130 * Since the above described distributed algorithm to compute the global
2131 * load-average relies on per-cpu sampling from the tick, it is affected by
2134 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2135 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2136 * when we read the global state.
2138 * Obviously reality has to ruin such a delightfully simple scheme:
2140 * - When we go NO_HZ idle during the window, we can negate our sample
2141 * contribution, causing under-accounting.
2143 * We avoid this by keeping two idle-delta counters and flipping them
2144 * when the window starts, thus separating old and new NO_HZ load.
2146 * The only trick is the slight shift in index flip for read vs write.
2150 * |-|-----------|-|-----------|-|-----------|-|
2151 * r:0 0 1 1 0 0 1 1 0
2152 * w:0 1 1 0 0 1 1 0 0
2154 * This ensures we'll fold the old idle contribution in this window while
2155 * accumlating the new one.
2157 * - When we wake up from NO_HZ idle during the window, we push up our
2158 * contribution, since we effectively move our sample point to a known
2161 * This is solved by pushing the window forward, and thus skipping the
2162 * sample, for this cpu (effectively using the idle-delta for this cpu which
2163 * was in effect at the time the window opened). This also solves the issue
2164 * of having to deal with a cpu having been in NOHZ idle for multiple
2165 * LOAD_FREQ intervals.
2167 * When making the ILB scale, we should try to pull this in as well.
2169 static atomic_long_t calc_load_idle
[2];
2170 static int calc_load_idx
;
2172 static inline int calc_load_write_idx(void)
2174 int idx
= calc_load_idx
;
2177 * See calc_global_nohz(), if we observe the new index, we also
2178 * need to observe the new update time.
2183 * If the folding window started, make sure we start writing in the
2186 if (!time_before(jiffies
, calc_load_update
))
2192 static inline int calc_load_read_idx(void)
2194 return calc_load_idx
& 1;
2197 void calc_load_enter_idle(void)
2199 struct rq
*this_rq
= this_rq();
2203 * We're going into NOHZ mode, if there's any pending delta, fold it
2204 * into the pending idle delta.
2206 delta
= calc_load_fold_active(this_rq
);
2208 int idx
= calc_load_write_idx();
2209 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2213 void calc_load_exit_idle(void)
2215 struct rq
*this_rq
= this_rq();
2218 * If we're still before the sample window, we're done.
2220 if (time_before(jiffies
, this_rq
->calc_load_update
))
2224 * We woke inside or after the sample window, this means we're already
2225 * accounted through the nohz accounting, so skip the entire deal and
2226 * sync up for the next window.
2228 this_rq
->calc_load_update
= calc_load_update
;
2229 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2230 this_rq
->calc_load_update
+= LOAD_FREQ
;
2233 static long calc_load_fold_idle(void)
2235 int idx
= calc_load_read_idx();
2238 if (atomic_long_read(&calc_load_idle
[idx
]))
2239 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2245 * fixed_power_int - compute: x^n, in O(log n) time
2247 * @x: base of the power
2248 * @frac_bits: fractional bits of @x
2249 * @n: power to raise @x to.
2251 * By exploiting the relation between the definition of the natural power
2252 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2253 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2254 * (where: n_i \elem {0, 1}, the binary vector representing n),
2255 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2256 * of course trivially computable in O(log_2 n), the length of our binary
2259 static unsigned long
2260 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2262 unsigned long result
= 1UL << frac_bits
;
2267 result
+= 1UL << (frac_bits
- 1);
2268 result
>>= frac_bits
;
2274 x
+= 1UL << (frac_bits
- 1);
2282 * a1 = a0 * e + a * (1 - e)
2284 * a2 = a1 * e + a * (1 - e)
2285 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2286 * = a0 * e^2 + a * (1 - e) * (1 + e)
2288 * a3 = a2 * e + a * (1 - e)
2289 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2290 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2294 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2295 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2296 * = a0 * e^n + a * (1 - e^n)
2298 * [1] application of the geometric series:
2301 * S_n := \Sum x^i = -------------
2304 static unsigned long
2305 calc_load_n(unsigned long load
, unsigned long exp
,
2306 unsigned long active
, unsigned int n
)
2309 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2313 * NO_HZ can leave us missing all per-cpu ticks calling
2314 * calc_load_account_active(), but since an idle CPU folds its delta into
2315 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2316 * in the pending idle delta if our idle period crossed a load cycle boundary.
2318 * Once we've updated the global active value, we need to apply the exponential
2319 * weights adjusted to the number of cycles missed.
2321 static void calc_global_nohz(void)
2323 long delta
, active
, n
;
2325 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2327 * Catch-up, fold however many we are behind still
2329 delta
= jiffies
- calc_load_update
- 10;
2330 n
= 1 + (delta
/ LOAD_FREQ
);
2332 active
= atomic_long_read(&calc_load_tasks
);
2333 active
= active
> 0 ? active
* FIXED_1
: 0;
2335 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2336 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2337 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2339 calc_load_update
+= n
* LOAD_FREQ
;
2343 * Flip the idle index...
2345 * Make sure we first write the new time then flip the index, so that
2346 * calc_load_write_idx() will see the new time when it reads the new
2347 * index, this avoids a double flip messing things up.
2352 #else /* !CONFIG_NO_HZ */
2354 static inline long calc_load_fold_idle(void) { return 0; }
2355 static inline void calc_global_nohz(void) { }
2357 #endif /* CONFIG_NO_HZ */
2360 * calc_load - update the avenrun load estimates 10 ticks after the
2361 * CPUs have updated calc_load_tasks.
2363 void calc_global_load(unsigned long ticks
)
2367 if (time_before(jiffies
, calc_load_update
+ 10))
2371 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2373 delta
= calc_load_fold_idle();
2375 atomic_long_add(delta
, &calc_load_tasks
);
2377 active
= atomic_long_read(&calc_load_tasks
);
2378 active
= active
> 0 ? active
* FIXED_1
: 0;
2380 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2381 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2382 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2384 calc_load_update
+= LOAD_FREQ
;
2387 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2393 * Called from update_cpu_load() to periodically update this CPU's
2396 static void calc_load_account_active(struct rq
*this_rq
)
2400 if (time_before(jiffies
, this_rq
->calc_load_update
))
2403 delta
= calc_load_fold_active(this_rq
);
2405 atomic_long_add(delta
, &calc_load_tasks
);
2407 this_rq
->calc_load_update
+= LOAD_FREQ
;
2411 * End of global load-average stuff
2415 * The exact cpuload at various idx values, calculated at every tick would be
2416 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2418 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2419 * on nth tick when cpu may be busy, then we have:
2420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2421 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2423 * decay_load_missed() below does efficient calculation of
2424 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2425 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2427 * The calculation is approximated on a 128 point scale.
2428 * degrade_zero_ticks is the number of ticks after which load at any
2429 * particular idx is approximated to be zero.
2430 * degrade_factor is a precomputed table, a row for each load idx.
2431 * Each column corresponds to degradation factor for a power of two ticks,
2432 * based on 128 point scale.
2434 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2435 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2437 * With this power of 2 load factors, we can degrade the load n times
2438 * by looking at 1 bits in n and doing as many mult/shift instead of
2439 * n mult/shifts needed by the exact degradation.
2441 #define DEGRADE_SHIFT 7
2442 static const unsigned char
2443 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2444 static const unsigned char
2445 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2446 {0, 0, 0, 0, 0, 0, 0, 0},
2447 {64, 32, 8, 0, 0, 0, 0, 0},
2448 {96, 72, 40, 12, 1, 0, 0},
2449 {112, 98, 75, 43, 15, 1, 0},
2450 {120, 112, 98, 76, 45, 16, 2} };
2453 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2454 * would be when CPU is idle and so we just decay the old load without
2455 * adding any new load.
2457 static unsigned long
2458 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2462 if (!missed_updates
)
2465 if (missed_updates
>= degrade_zero_ticks
[idx
])
2469 return load
>> missed_updates
;
2471 while (missed_updates
) {
2472 if (missed_updates
% 2)
2473 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2475 missed_updates
>>= 1;
2482 * Update rq->cpu_load[] statistics. This function is usually called every
2483 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2484 * every tick. We fix it up based on jiffies.
2486 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2487 unsigned long pending_updates
)
2491 this_rq
->nr_load_updates
++;
2493 /* Update our load: */
2494 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2495 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2496 unsigned long old_load
, new_load
;
2498 /* scale is effectively 1 << i now, and >> i divides by scale */
2500 old_load
= this_rq
->cpu_load
[i
];
2501 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2502 new_load
= this_load
;
2504 * Round up the averaging division if load is increasing. This
2505 * prevents us from getting stuck on 9 if the load is 10, for
2508 if (new_load
> old_load
)
2509 new_load
+= scale
- 1;
2511 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2514 sched_avg_update(this_rq
);
2519 * There is no sane way to deal with nohz on smp when using jiffies because the
2520 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2521 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2523 * Therefore we cannot use the delta approach from the regular tick since that
2524 * would seriously skew the load calculation. However we'll make do for those
2525 * updates happening while idle (nohz_idle_balance) or coming out of idle
2526 * (tick_nohz_idle_exit).
2528 * This means we might still be one tick off for nohz periods.
2532 * Called from nohz_idle_balance() to update the load ratings before doing the
2535 void update_idle_cpu_load(struct rq
*this_rq
)
2537 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2538 unsigned long load
= this_rq
->load
.weight
;
2539 unsigned long pending_updates
;
2542 * bail if there's load or we're actually up-to-date.
2544 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2547 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2548 this_rq
->last_load_update_tick
= curr_jiffies
;
2550 __update_cpu_load(this_rq
, load
, pending_updates
);
2554 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2556 void update_cpu_load_nohz(void)
2558 struct rq
*this_rq
= this_rq();
2559 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2560 unsigned long pending_updates
;
2562 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2565 raw_spin_lock(&this_rq
->lock
);
2566 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2567 if (pending_updates
) {
2568 this_rq
->last_load_update_tick
= curr_jiffies
;
2570 * We were idle, this means load 0, the current load might be
2571 * !0 due to remote wakeups and the sort.
2573 __update_cpu_load(this_rq
, 0, pending_updates
);
2575 raw_spin_unlock(&this_rq
->lock
);
2577 #endif /* CONFIG_NO_HZ */
2580 * Called from scheduler_tick()
2582 static void update_cpu_load_active(struct rq
*this_rq
)
2585 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2587 this_rq
->last_load_update_tick
= jiffies
;
2588 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2590 calc_load_account_active(this_rq
);
2596 * sched_exec - execve() is a valuable balancing opportunity, because at
2597 * this point the task has the smallest effective memory and cache footprint.
2599 void sched_exec(void)
2601 struct task_struct
*p
= current
;
2602 unsigned long flags
;
2605 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2606 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2607 if (dest_cpu
== smp_processor_id())
2610 if (likely(cpu_active(dest_cpu
))) {
2611 struct migration_arg arg
= { p
, dest_cpu
};
2613 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2614 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2618 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2623 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2624 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2626 EXPORT_PER_CPU_SYMBOL(kstat
);
2627 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2630 * Return any ns on the sched_clock that have not yet been accounted in
2631 * @p in case that task is currently running.
2633 * Called with task_rq_lock() held on @rq.
2635 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2639 if (task_current(rq
, p
)) {
2640 update_rq_clock(rq
);
2641 ns
= rq
->clock_task
- p
->se
.exec_start
;
2649 unsigned long long task_delta_exec(struct task_struct
*p
)
2651 unsigned long flags
;
2655 rq
= task_rq_lock(p
, &flags
);
2656 ns
= do_task_delta_exec(p
, rq
);
2657 task_rq_unlock(rq
, p
, &flags
);
2663 * Return accounted runtime for the task.
2664 * In case the task is currently running, return the runtime plus current's
2665 * pending runtime that have not been accounted yet.
2667 unsigned long long task_sched_runtime(struct task_struct
*p
)
2669 unsigned long flags
;
2673 rq
= task_rq_lock(p
, &flags
);
2674 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2675 task_rq_unlock(rq
, p
, &flags
);
2681 * This function gets called by the timer code, with HZ frequency.
2682 * We call it with interrupts disabled.
2684 void scheduler_tick(void)
2686 int cpu
= smp_processor_id();
2687 struct rq
*rq
= cpu_rq(cpu
);
2688 struct task_struct
*curr
= rq
->curr
;
2692 raw_spin_lock(&rq
->lock
);
2693 update_rq_clock(rq
);
2694 update_cpu_load_active(rq
);
2695 curr
->sched_class
->task_tick(rq
, curr
, 0);
2696 raw_spin_unlock(&rq
->lock
);
2698 perf_event_task_tick();
2701 rq
->idle_balance
= idle_cpu(cpu
);
2702 trigger_load_balance(rq
, cpu
);
2706 notrace
unsigned long get_parent_ip(unsigned long addr
)
2708 if (in_lock_functions(addr
)) {
2709 addr
= CALLER_ADDR2
;
2710 if (in_lock_functions(addr
))
2711 addr
= CALLER_ADDR3
;
2716 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2717 defined(CONFIG_PREEMPT_TRACER))
2719 void __kprobes
add_preempt_count(int val
)
2721 #ifdef CONFIG_DEBUG_PREEMPT
2725 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2728 preempt_count() += val
;
2729 #ifdef CONFIG_DEBUG_PREEMPT
2731 * Spinlock count overflowing soon?
2733 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2736 if (preempt_count() == val
)
2737 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2739 EXPORT_SYMBOL(add_preempt_count
);
2741 void __kprobes
sub_preempt_count(int val
)
2743 #ifdef CONFIG_DEBUG_PREEMPT
2747 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2750 * Is the spinlock portion underflowing?
2752 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2753 !(preempt_count() & PREEMPT_MASK
)))
2757 if (preempt_count() == val
)
2758 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2759 preempt_count() -= val
;
2761 EXPORT_SYMBOL(sub_preempt_count
);
2766 * Print scheduling while atomic bug:
2768 static noinline
void __schedule_bug(struct task_struct
*prev
)
2770 if (oops_in_progress
)
2773 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2774 prev
->comm
, prev
->pid
, preempt_count());
2776 debug_show_held_locks(prev
);
2778 if (irqs_disabled())
2779 print_irqtrace_events(prev
);
2781 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2785 * Various schedule()-time debugging checks and statistics:
2787 static inline void schedule_debug(struct task_struct
*prev
)
2790 * Test if we are atomic. Since do_exit() needs to call into
2791 * schedule() atomically, we ignore that path for now.
2792 * Otherwise, whine if we are scheduling when we should not be.
2794 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2795 __schedule_bug(prev
);
2798 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2800 schedstat_inc(this_rq(), sched_count
);
2803 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2805 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2806 update_rq_clock(rq
);
2807 prev
->sched_class
->put_prev_task(rq
, prev
);
2811 * Pick up the highest-prio task:
2813 static inline struct task_struct
*
2814 pick_next_task(struct rq
*rq
)
2816 const struct sched_class
*class;
2817 struct task_struct
*p
;
2820 * Optimization: we know that if all tasks are in
2821 * the fair class we can call that function directly:
2823 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2824 p
= fair_sched_class
.pick_next_task(rq
);
2829 for_each_class(class) {
2830 p
= class->pick_next_task(rq
);
2835 BUG(); /* the idle class will always have a runnable task */
2839 * __schedule() is the main scheduler function.
2841 * The main means of driving the scheduler and thus entering this function are:
2843 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2845 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2846 * paths. For example, see arch/x86/entry_64.S.
2848 * To drive preemption between tasks, the scheduler sets the flag in timer
2849 * interrupt handler scheduler_tick().
2851 * 3. Wakeups don't really cause entry into schedule(). They add a
2852 * task to the run-queue and that's it.
2854 * Now, if the new task added to the run-queue preempts the current
2855 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2856 * called on the nearest possible occasion:
2858 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2860 * - in syscall or exception context, at the next outmost
2861 * preempt_enable(). (this might be as soon as the wake_up()'s
2864 * - in IRQ context, return from interrupt-handler to
2865 * preemptible context
2867 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2870 * - cond_resched() call
2871 * - explicit schedule() call
2872 * - return from syscall or exception to user-space
2873 * - return from interrupt-handler to user-space
2875 static void __sched
__schedule(void)
2877 struct task_struct
*prev
, *next
;
2878 unsigned long *switch_count
;
2884 cpu
= smp_processor_id();
2886 rcu_note_context_switch(cpu
);
2889 schedule_debug(prev
);
2891 if (sched_feat(HRTICK
))
2894 raw_spin_lock_irq(&rq
->lock
);
2896 switch_count
= &prev
->nivcsw
;
2897 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2898 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2899 prev
->state
= TASK_RUNNING
;
2901 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2905 * If a worker went to sleep, notify and ask workqueue
2906 * whether it wants to wake up a task to maintain
2909 if (prev
->flags
& PF_WQ_WORKER
) {
2910 struct task_struct
*to_wakeup
;
2912 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2914 try_to_wake_up_local(to_wakeup
);
2917 switch_count
= &prev
->nvcsw
;
2920 pre_schedule(rq
, prev
);
2922 if (unlikely(!rq
->nr_running
))
2923 idle_balance(cpu
, rq
);
2925 put_prev_task(rq
, prev
);
2926 next
= pick_next_task(rq
);
2927 clear_tsk_need_resched(prev
);
2928 rq
->skip_clock_update
= 0;
2930 if (likely(prev
!= next
)) {
2935 context_switch(rq
, prev
, next
); /* unlocks the rq */
2937 * The context switch have flipped the stack from under us
2938 * and restored the local variables which were saved when
2939 * this task called schedule() in the past. prev == current
2940 * is still correct, but it can be moved to another cpu/rq.
2942 cpu
= smp_processor_id();
2945 raw_spin_unlock_irq(&rq
->lock
);
2949 sched_preempt_enable_no_resched();
2954 static inline void sched_submit_work(struct task_struct
*tsk
)
2956 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2959 * If we are going to sleep and we have plugged IO queued,
2960 * make sure to submit it to avoid deadlocks.
2962 if (blk_needs_flush_plug(tsk
))
2963 blk_schedule_flush_plug(tsk
);
2966 asmlinkage
void __sched
schedule(void)
2968 struct task_struct
*tsk
= current
;
2970 sched_submit_work(tsk
);
2973 EXPORT_SYMBOL(schedule
);
2975 #ifdef CONFIG_CONTEXT_TRACKING
2976 asmlinkage
void __sched
schedule_user(void)
2979 * If we come here after a random call to set_need_resched(),
2980 * or we have been woken up remotely but the IPI has not yet arrived,
2981 * we haven't yet exited the RCU idle mode. Do it here manually until
2982 * we find a better solution.
2991 * schedule_preempt_disabled - called with preemption disabled
2993 * Returns with preemption disabled. Note: preempt_count must be 1
2995 void __sched
schedule_preempt_disabled(void)
2997 sched_preempt_enable_no_resched();
3002 #ifdef CONFIG_PREEMPT
3004 * this is the entry point to schedule() from in-kernel preemption
3005 * off of preempt_enable. Kernel preemptions off return from interrupt
3006 * occur there and call schedule directly.
3008 asmlinkage
void __sched notrace
preempt_schedule(void)
3010 struct thread_info
*ti
= current_thread_info();
3013 * If there is a non-zero preempt_count or interrupts are disabled,
3014 * we do not want to preempt the current task. Just return..
3016 if (likely(ti
->preempt_count
|| irqs_disabled()))
3020 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3022 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3025 * Check again in case we missed a preemption opportunity
3026 * between schedule and now.
3029 } while (need_resched());
3031 EXPORT_SYMBOL(preempt_schedule
);
3034 * this is the entry point to schedule() from kernel preemption
3035 * off of irq context.
3036 * Note, that this is called and return with irqs disabled. This will
3037 * protect us against recursive calling from irq.
3039 asmlinkage
void __sched
preempt_schedule_irq(void)
3041 struct thread_info
*ti
= current_thread_info();
3042 enum ctx_state prev_state
;
3044 /* Catch callers which need to be fixed */
3045 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3047 prev_state
= exception_enter();
3050 add_preempt_count(PREEMPT_ACTIVE
);
3053 local_irq_disable();
3054 sub_preempt_count(PREEMPT_ACTIVE
);
3057 * Check again in case we missed a preemption opportunity
3058 * between schedule and now.
3061 } while (need_resched());
3063 exception_exit(prev_state
);
3066 #endif /* CONFIG_PREEMPT */
3068 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3071 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3073 EXPORT_SYMBOL(default_wake_function
);
3076 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3077 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3078 * number) then we wake all the non-exclusive tasks and one exclusive task.
3080 * There are circumstances in which we can try to wake a task which has already
3081 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3082 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3084 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3085 int nr_exclusive
, int wake_flags
, void *key
)
3087 wait_queue_t
*curr
, *next
;
3089 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3090 unsigned flags
= curr
->flags
;
3092 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3093 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3099 * __wake_up - wake up threads blocked on a waitqueue.
3101 * @mode: which threads
3102 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3103 * @key: is directly passed to the wakeup function
3105 * It may be assumed that this function implies a write memory barrier before
3106 * changing the task state if and only if any tasks are woken up.
3108 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3109 int nr_exclusive
, void *key
)
3111 unsigned long flags
;
3113 spin_lock_irqsave(&q
->lock
, flags
);
3114 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3115 spin_unlock_irqrestore(&q
->lock
, flags
);
3117 EXPORT_SYMBOL(__wake_up
);
3120 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3122 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3124 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3126 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3128 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3130 __wake_up_common(q
, mode
, 1, 0, key
);
3132 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3135 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3137 * @mode: which threads
3138 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3139 * @key: opaque value to be passed to wakeup targets
3141 * The sync wakeup differs that the waker knows that it will schedule
3142 * away soon, so while the target thread will be woken up, it will not
3143 * be migrated to another CPU - ie. the two threads are 'synchronized'
3144 * with each other. This can prevent needless bouncing between CPUs.
3146 * On UP it can prevent extra preemption.
3148 * It may be assumed that this function implies a write memory barrier before
3149 * changing the task state if and only if any tasks are woken up.
3151 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3152 int nr_exclusive
, void *key
)
3154 unsigned long flags
;
3155 int wake_flags
= WF_SYNC
;
3160 if (unlikely(!nr_exclusive
))
3163 spin_lock_irqsave(&q
->lock
, flags
);
3164 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3165 spin_unlock_irqrestore(&q
->lock
, flags
);
3167 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3170 * __wake_up_sync - see __wake_up_sync_key()
3172 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3174 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3176 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3179 * complete: - signals a single thread waiting on this completion
3180 * @x: holds the state of this particular completion
3182 * This will wake up a single thread waiting on this completion. Threads will be
3183 * awakened in the same order in which they were queued.
3185 * See also complete_all(), wait_for_completion() and related routines.
3187 * It may be assumed that this function implies a write memory barrier before
3188 * changing the task state if and only if any tasks are woken up.
3190 void complete(struct completion
*x
)
3192 unsigned long flags
;
3194 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3196 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3197 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3199 EXPORT_SYMBOL(complete
);
3202 * complete_all: - signals all threads waiting on this completion
3203 * @x: holds the state of this particular completion
3205 * This will wake up all threads waiting on this particular completion event.
3207 * It may be assumed that this function implies a write memory barrier before
3208 * changing the task state if and only if any tasks are woken up.
3210 void complete_all(struct completion
*x
)
3212 unsigned long flags
;
3214 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3215 x
->done
+= UINT_MAX
/2;
3216 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3217 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3219 EXPORT_SYMBOL(complete_all
);
3221 static inline long __sched
3222 do_wait_for_common(struct completion
*x
,
3223 long (*action
)(long), long timeout
, int state
)
3226 DECLARE_WAITQUEUE(wait
, current
);
3228 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3230 if (signal_pending_state(state
, current
)) {
3231 timeout
= -ERESTARTSYS
;
3234 __set_current_state(state
);
3235 spin_unlock_irq(&x
->wait
.lock
);
3236 timeout
= action(timeout
);
3237 spin_lock_irq(&x
->wait
.lock
);
3238 } while (!x
->done
&& timeout
);
3239 __remove_wait_queue(&x
->wait
, &wait
);
3244 return timeout
?: 1;
3247 static inline long __sched
3248 __wait_for_common(struct completion
*x
,
3249 long (*action
)(long), long timeout
, int state
)
3253 spin_lock_irq(&x
->wait
.lock
);
3254 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
3255 spin_unlock_irq(&x
->wait
.lock
);
3260 wait_for_common(struct completion
*x
, long timeout
, int state
)
3262 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
3266 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
3268 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
3272 * wait_for_completion: - waits for completion of a task
3273 * @x: holds the state of this particular completion
3275 * This waits to be signaled for completion of a specific task. It is NOT
3276 * interruptible and there is no timeout.
3278 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3279 * and interrupt capability. Also see complete().
3281 void __sched
wait_for_completion(struct completion
*x
)
3283 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3285 EXPORT_SYMBOL(wait_for_completion
);
3288 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3289 * @x: holds the state of this particular completion
3290 * @timeout: timeout value in jiffies
3292 * This waits for either a completion of a specific task to be signaled or for a
3293 * specified timeout to expire. The timeout is in jiffies. It is not
3296 * The return value is 0 if timed out, and positive (at least 1, or number of
3297 * jiffies left till timeout) if completed.
3299 unsigned long __sched
3300 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3302 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3304 EXPORT_SYMBOL(wait_for_completion_timeout
);
3307 * wait_for_completion_io: - waits for completion of a task
3308 * @x: holds the state of this particular completion
3310 * This waits to be signaled for completion of a specific task. It is NOT
3311 * interruptible and there is no timeout. The caller is accounted as waiting
3314 void __sched
wait_for_completion_io(struct completion
*x
)
3316 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3318 EXPORT_SYMBOL(wait_for_completion_io
);
3321 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3322 * @x: holds the state of this particular completion
3323 * @timeout: timeout value in jiffies
3325 * This waits for either a completion of a specific task to be signaled or for a
3326 * specified timeout to expire. The timeout is in jiffies. It is not
3327 * interruptible. The caller is accounted as waiting for IO.
3329 * The return value is 0 if timed out, and positive (at least 1, or number of
3330 * jiffies left till timeout) if completed.
3332 unsigned long __sched
3333 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
3335 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3337 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
3340 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3341 * @x: holds the state of this particular completion
3343 * This waits for completion of a specific task to be signaled. It is
3346 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3348 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3350 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3351 if (t
== -ERESTARTSYS
)
3355 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3358 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3359 * @x: holds the state of this particular completion
3360 * @timeout: timeout value in jiffies
3362 * This waits for either a completion of a specific task to be signaled or for a
3363 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3365 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3366 * positive (at least 1, or number of jiffies left till timeout) if completed.
3369 wait_for_completion_interruptible_timeout(struct completion
*x
,
3370 unsigned long timeout
)
3372 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3374 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3377 * wait_for_completion_killable: - waits for completion of a task (killable)
3378 * @x: holds the state of this particular completion
3380 * This waits to be signaled for completion of a specific task. It can be
3381 * interrupted by a kill signal.
3383 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3385 int __sched
wait_for_completion_killable(struct completion
*x
)
3387 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3388 if (t
== -ERESTARTSYS
)
3392 EXPORT_SYMBOL(wait_for_completion_killable
);
3395 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3396 * @x: holds the state of this particular completion
3397 * @timeout: timeout value in jiffies
3399 * This waits for either a completion of a specific task to be
3400 * signaled or for a specified timeout to expire. It can be
3401 * interrupted by a kill signal. The timeout is in jiffies.
3403 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3404 * positive (at least 1, or number of jiffies left till timeout) if completed.
3407 wait_for_completion_killable_timeout(struct completion
*x
,
3408 unsigned long timeout
)
3410 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3412 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3415 * try_wait_for_completion - try to decrement a completion without blocking
3416 * @x: completion structure
3418 * Returns: 0 if a decrement cannot be done without blocking
3419 * 1 if a decrement succeeded.
3421 * If a completion is being used as a counting completion,
3422 * attempt to decrement the counter without blocking. This
3423 * enables us to avoid waiting if the resource the completion
3424 * is protecting is not available.
3426 bool try_wait_for_completion(struct completion
*x
)
3428 unsigned long flags
;
3431 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3436 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3439 EXPORT_SYMBOL(try_wait_for_completion
);
3442 * completion_done - Test to see if a completion has any waiters
3443 * @x: completion structure
3445 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3446 * 1 if there are no waiters.
3449 bool completion_done(struct completion
*x
)
3451 unsigned long flags
;
3454 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3457 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3460 EXPORT_SYMBOL(completion_done
);
3463 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3465 unsigned long flags
;
3468 init_waitqueue_entry(&wait
, current
);
3470 __set_current_state(state
);
3472 spin_lock_irqsave(&q
->lock
, flags
);
3473 __add_wait_queue(q
, &wait
);
3474 spin_unlock(&q
->lock
);
3475 timeout
= schedule_timeout(timeout
);
3476 spin_lock_irq(&q
->lock
);
3477 __remove_wait_queue(q
, &wait
);
3478 spin_unlock_irqrestore(&q
->lock
, flags
);
3483 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3485 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3487 EXPORT_SYMBOL(interruptible_sleep_on
);
3490 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3492 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3494 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3496 void __sched
sleep_on(wait_queue_head_t
*q
)
3498 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3500 EXPORT_SYMBOL(sleep_on
);
3502 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3504 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3506 EXPORT_SYMBOL(sleep_on_timeout
);
3508 #ifdef CONFIG_RT_MUTEXES
3511 * rt_mutex_setprio - set the current priority of a task
3513 * @prio: prio value (kernel-internal form)
3515 * This function changes the 'effective' priority of a task. It does
3516 * not touch ->normal_prio like __setscheduler().
3518 * Used by the rt_mutex code to implement priority inheritance logic.
3520 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3522 int oldprio
, on_rq
, running
;
3524 const struct sched_class
*prev_class
;
3526 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3528 rq
= __task_rq_lock(p
);
3531 * Idle task boosting is a nono in general. There is one
3532 * exception, when PREEMPT_RT and NOHZ is active:
3534 * The idle task calls get_next_timer_interrupt() and holds
3535 * the timer wheel base->lock on the CPU and another CPU wants
3536 * to access the timer (probably to cancel it). We can safely
3537 * ignore the boosting request, as the idle CPU runs this code
3538 * with interrupts disabled and will complete the lock
3539 * protected section without being interrupted. So there is no
3540 * real need to boost.
3542 if (unlikely(p
== rq
->idle
)) {
3543 WARN_ON(p
!= rq
->curr
);
3544 WARN_ON(p
->pi_blocked_on
);
3548 trace_sched_pi_setprio(p
, prio
);
3550 prev_class
= p
->sched_class
;
3552 running
= task_current(rq
, p
);
3554 dequeue_task(rq
, p
, 0);
3556 p
->sched_class
->put_prev_task(rq
, p
);
3559 p
->sched_class
= &rt_sched_class
;
3561 p
->sched_class
= &fair_sched_class
;
3566 p
->sched_class
->set_curr_task(rq
);
3568 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3570 check_class_changed(rq
, p
, prev_class
, oldprio
);
3572 __task_rq_unlock(rq
);
3575 void set_user_nice(struct task_struct
*p
, long nice
)
3577 int old_prio
, delta
, on_rq
;
3578 unsigned long flags
;
3581 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3584 * We have to be careful, if called from sys_setpriority(),
3585 * the task might be in the middle of scheduling on another CPU.
3587 rq
= task_rq_lock(p
, &flags
);
3589 * The RT priorities are set via sched_setscheduler(), but we still
3590 * allow the 'normal' nice value to be set - but as expected
3591 * it wont have any effect on scheduling until the task is
3592 * SCHED_FIFO/SCHED_RR:
3594 if (task_has_rt_policy(p
)) {
3595 p
->static_prio
= NICE_TO_PRIO(nice
);
3600 dequeue_task(rq
, p
, 0);
3602 p
->static_prio
= NICE_TO_PRIO(nice
);
3605 p
->prio
= effective_prio(p
);
3606 delta
= p
->prio
- old_prio
;
3609 enqueue_task(rq
, p
, 0);
3611 * If the task increased its priority or is running and
3612 * lowered its priority, then reschedule its CPU:
3614 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3615 resched_task(rq
->curr
);
3618 task_rq_unlock(rq
, p
, &flags
);
3620 EXPORT_SYMBOL(set_user_nice
);
3623 * can_nice - check if a task can reduce its nice value
3627 int can_nice(const struct task_struct
*p
, const int nice
)
3629 /* convert nice value [19,-20] to rlimit style value [1,40] */
3630 int nice_rlim
= 20 - nice
;
3632 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3633 capable(CAP_SYS_NICE
));
3636 #ifdef __ARCH_WANT_SYS_NICE
3639 * sys_nice - change the priority of the current process.
3640 * @increment: priority increment
3642 * sys_setpriority is a more generic, but much slower function that
3643 * does similar things.
3645 SYSCALL_DEFINE1(nice
, int, increment
)
3650 * Setpriority might change our priority at the same moment.
3651 * We don't have to worry. Conceptually one call occurs first
3652 * and we have a single winner.
3654 if (increment
< -40)
3659 nice
= TASK_NICE(current
) + increment
;
3665 if (increment
< 0 && !can_nice(current
, nice
))
3668 retval
= security_task_setnice(current
, nice
);
3672 set_user_nice(current
, nice
);
3679 * task_prio - return the priority value of a given task.
3680 * @p: the task in question.
3682 * This is the priority value as seen by users in /proc.
3683 * RT tasks are offset by -200. Normal tasks are centered
3684 * around 0, value goes from -16 to +15.
3686 int task_prio(const struct task_struct
*p
)
3688 return p
->prio
- MAX_RT_PRIO
;
3692 * task_nice - return the nice value of a given task.
3693 * @p: the task in question.
3695 int task_nice(const struct task_struct
*p
)
3697 return TASK_NICE(p
);
3699 EXPORT_SYMBOL(task_nice
);
3702 * idle_cpu - is a given cpu idle currently?
3703 * @cpu: the processor in question.
3705 int idle_cpu(int cpu
)
3707 struct rq
*rq
= cpu_rq(cpu
);
3709 if (rq
->curr
!= rq
->idle
)
3716 if (!llist_empty(&rq
->wake_list
))
3724 * idle_task - return the idle task for a given cpu.
3725 * @cpu: the processor in question.
3727 struct task_struct
*idle_task(int cpu
)
3729 return cpu_rq(cpu
)->idle
;
3733 * find_process_by_pid - find a process with a matching PID value.
3734 * @pid: the pid in question.
3736 static struct task_struct
*find_process_by_pid(pid_t pid
)
3738 return pid
? find_task_by_vpid(pid
) : current
;
3741 /* Actually do priority change: must hold rq lock. */
3743 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3746 p
->rt_priority
= prio
;
3747 p
->normal_prio
= normal_prio(p
);
3748 /* we are holding p->pi_lock already */
3749 p
->prio
= rt_mutex_getprio(p
);
3750 if (rt_prio(p
->prio
))
3751 p
->sched_class
= &rt_sched_class
;
3753 p
->sched_class
= &fair_sched_class
;
3758 * check the target process has a UID that matches the current process's
3760 static bool check_same_owner(struct task_struct
*p
)
3762 const struct cred
*cred
= current_cred(), *pcred
;
3766 pcred
= __task_cred(p
);
3767 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3768 uid_eq(cred
->euid
, pcred
->uid
));
3773 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3774 const struct sched_param
*param
, bool user
)
3776 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3777 unsigned long flags
;
3778 const struct sched_class
*prev_class
;
3782 /* may grab non-irq protected spin_locks */
3783 BUG_ON(in_interrupt());
3785 /* double check policy once rq lock held */
3787 reset_on_fork
= p
->sched_reset_on_fork
;
3788 policy
= oldpolicy
= p
->policy
;
3790 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3791 policy
&= ~SCHED_RESET_ON_FORK
;
3793 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3794 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3795 policy
!= SCHED_IDLE
)
3800 * Valid priorities for SCHED_FIFO and SCHED_RR are
3801 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3802 * SCHED_BATCH and SCHED_IDLE is 0.
3804 if (param
->sched_priority
< 0 ||
3805 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3806 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3808 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3812 * Allow unprivileged RT tasks to decrease priority:
3814 if (user
&& !capable(CAP_SYS_NICE
)) {
3815 if (rt_policy(policy
)) {
3816 unsigned long rlim_rtprio
=
3817 task_rlimit(p
, RLIMIT_RTPRIO
);
3819 /* can't set/change the rt policy */
3820 if (policy
!= p
->policy
&& !rlim_rtprio
)
3823 /* can't increase priority */
3824 if (param
->sched_priority
> p
->rt_priority
&&
3825 param
->sched_priority
> rlim_rtprio
)
3830 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3831 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3833 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3834 if (!can_nice(p
, TASK_NICE(p
)))
3838 /* can't change other user's priorities */
3839 if (!check_same_owner(p
))
3842 /* Normal users shall not reset the sched_reset_on_fork flag */
3843 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3848 retval
= security_task_setscheduler(p
);
3854 * make sure no PI-waiters arrive (or leave) while we are
3855 * changing the priority of the task:
3857 * To be able to change p->policy safely, the appropriate
3858 * runqueue lock must be held.
3860 rq
= task_rq_lock(p
, &flags
);
3863 * Changing the policy of the stop threads its a very bad idea
3865 if (p
== rq
->stop
) {
3866 task_rq_unlock(rq
, p
, &flags
);
3871 * If not changing anything there's no need to proceed further:
3873 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3874 param
->sched_priority
== p
->rt_priority
))) {
3875 task_rq_unlock(rq
, p
, &flags
);
3879 #ifdef CONFIG_RT_GROUP_SCHED
3882 * Do not allow realtime tasks into groups that have no runtime
3885 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3886 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3887 !task_group_is_autogroup(task_group(p
))) {
3888 task_rq_unlock(rq
, p
, &flags
);
3894 /* recheck policy now with rq lock held */
3895 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3896 policy
= oldpolicy
= -1;
3897 task_rq_unlock(rq
, p
, &flags
);
3901 running
= task_current(rq
, p
);
3903 dequeue_task(rq
, p
, 0);
3905 p
->sched_class
->put_prev_task(rq
, p
);
3907 p
->sched_reset_on_fork
= reset_on_fork
;
3910 prev_class
= p
->sched_class
;
3911 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3914 p
->sched_class
->set_curr_task(rq
);
3916 enqueue_task(rq
, p
, 0);
3918 check_class_changed(rq
, p
, prev_class
, oldprio
);
3919 task_rq_unlock(rq
, p
, &flags
);
3921 rt_mutex_adjust_pi(p
);
3927 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3928 * @p: the task in question.
3929 * @policy: new policy.
3930 * @param: structure containing the new RT priority.
3932 * NOTE that the task may be already dead.
3934 int sched_setscheduler(struct task_struct
*p
, int policy
,
3935 const struct sched_param
*param
)
3937 return __sched_setscheduler(p
, policy
, param
, true);
3939 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3942 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3943 * @p: the task in question.
3944 * @policy: new policy.
3945 * @param: structure containing the new RT priority.
3947 * Just like sched_setscheduler, only don't bother checking if the
3948 * current context has permission. For example, this is needed in
3949 * stop_machine(): we create temporary high priority worker threads,
3950 * but our caller might not have that capability.
3952 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3953 const struct sched_param
*param
)
3955 return __sched_setscheduler(p
, policy
, param
, false);
3959 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3961 struct sched_param lparam
;
3962 struct task_struct
*p
;
3965 if (!param
|| pid
< 0)
3967 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3972 p
= find_process_by_pid(pid
);
3974 retval
= sched_setscheduler(p
, policy
, &lparam
);
3981 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3982 * @pid: the pid in question.
3983 * @policy: new policy.
3984 * @param: structure containing the new RT priority.
3986 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3987 struct sched_param __user
*, param
)
3989 /* negative values for policy are not valid */
3993 return do_sched_setscheduler(pid
, policy
, param
);
3997 * sys_sched_setparam - set/change the RT priority of a thread
3998 * @pid: the pid in question.
3999 * @param: structure containing the new RT priority.
4001 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4003 return do_sched_setscheduler(pid
, -1, param
);
4007 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4008 * @pid: the pid in question.
4010 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4012 struct task_struct
*p
;
4020 p
= find_process_by_pid(pid
);
4022 retval
= security_task_getscheduler(p
);
4025 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4032 * sys_sched_getparam - get the RT priority of a thread
4033 * @pid: the pid in question.
4034 * @param: structure containing the RT priority.
4036 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4038 struct sched_param lp
;
4039 struct task_struct
*p
;
4042 if (!param
|| pid
< 0)
4046 p
= find_process_by_pid(pid
);
4051 retval
= security_task_getscheduler(p
);
4055 lp
.sched_priority
= p
->rt_priority
;
4059 * This one might sleep, we cannot do it with a spinlock held ...
4061 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4070 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4072 cpumask_var_t cpus_allowed
, new_mask
;
4073 struct task_struct
*p
;
4079 p
= find_process_by_pid(pid
);
4086 /* Prevent p going away */
4090 if (p
->flags
& PF_NO_SETAFFINITY
) {
4094 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4098 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4100 goto out_free_cpus_allowed
;
4103 if (!check_same_owner(p
)) {
4105 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4112 retval
= security_task_setscheduler(p
);
4116 cpuset_cpus_allowed(p
, cpus_allowed
);
4117 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4119 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4122 cpuset_cpus_allowed(p
, cpus_allowed
);
4123 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4125 * We must have raced with a concurrent cpuset
4126 * update. Just reset the cpus_allowed to the
4127 * cpuset's cpus_allowed
4129 cpumask_copy(new_mask
, cpus_allowed
);
4134 free_cpumask_var(new_mask
);
4135 out_free_cpus_allowed
:
4136 free_cpumask_var(cpus_allowed
);
4143 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4144 struct cpumask
*new_mask
)
4146 if (len
< cpumask_size())
4147 cpumask_clear(new_mask
);
4148 else if (len
> cpumask_size())
4149 len
= cpumask_size();
4151 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4155 * sys_sched_setaffinity - set the cpu affinity of a process
4156 * @pid: pid of the process
4157 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4158 * @user_mask_ptr: user-space pointer to the new cpu mask
4160 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4161 unsigned long __user
*, user_mask_ptr
)
4163 cpumask_var_t new_mask
;
4166 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4169 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4171 retval
= sched_setaffinity(pid
, new_mask
);
4172 free_cpumask_var(new_mask
);
4176 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4178 struct task_struct
*p
;
4179 unsigned long flags
;
4186 p
= find_process_by_pid(pid
);
4190 retval
= security_task_getscheduler(p
);
4194 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4195 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4196 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4206 * sys_sched_getaffinity - get the cpu affinity of a process
4207 * @pid: pid of the process
4208 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4209 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4211 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4212 unsigned long __user
*, user_mask_ptr
)
4217 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4219 if (len
& (sizeof(unsigned long)-1))
4222 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4225 ret
= sched_getaffinity(pid
, mask
);
4227 size_t retlen
= min_t(size_t, len
, cpumask_size());
4229 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4234 free_cpumask_var(mask
);
4240 * sys_sched_yield - yield the current processor to other threads.
4242 * This function yields the current CPU to other tasks. If there are no
4243 * other threads running on this CPU then this function will return.
4245 SYSCALL_DEFINE0(sched_yield
)
4247 struct rq
*rq
= this_rq_lock();
4249 schedstat_inc(rq
, yld_count
);
4250 current
->sched_class
->yield_task(rq
);
4253 * Since we are going to call schedule() anyway, there's
4254 * no need to preempt or enable interrupts:
4256 __release(rq
->lock
);
4257 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4258 do_raw_spin_unlock(&rq
->lock
);
4259 sched_preempt_enable_no_resched();
4266 static inline int should_resched(void)
4268 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4271 static void __cond_resched(void)
4273 add_preempt_count(PREEMPT_ACTIVE
);
4275 sub_preempt_count(PREEMPT_ACTIVE
);
4278 int __sched
_cond_resched(void)
4280 if (should_resched()) {
4286 EXPORT_SYMBOL(_cond_resched
);
4289 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4290 * call schedule, and on return reacquire the lock.
4292 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4293 * operations here to prevent schedule() from being called twice (once via
4294 * spin_unlock(), once by hand).
4296 int __cond_resched_lock(spinlock_t
*lock
)
4298 int resched
= should_resched();
4301 lockdep_assert_held(lock
);
4303 if (spin_needbreak(lock
) || resched
) {
4314 EXPORT_SYMBOL(__cond_resched_lock
);
4316 int __sched
__cond_resched_softirq(void)
4318 BUG_ON(!in_softirq());
4320 if (should_resched()) {
4328 EXPORT_SYMBOL(__cond_resched_softirq
);
4331 * yield - yield the current processor to other threads.
4333 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4335 * The scheduler is at all times free to pick the calling task as the most
4336 * eligible task to run, if removing the yield() call from your code breaks
4337 * it, its already broken.
4339 * Typical broken usage is:
4344 * where one assumes that yield() will let 'the other' process run that will
4345 * make event true. If the current task is a SCHED_FIFO task that will never
4346 * happen. Never use yield() as a progress guarantee!!
4348 * If you want to use yield() to wait for something, use wait_event().
4349 * If you want to use yield() to be 'nice' for others, use cond_resched().
4350 * If you still want to use yield(), do not!
4352 void __sched
yield(void)
4354 set_current_state(TASK_RUNNING
);
4357 EXPORT_SYMBOL(yield
);
4360 * yield_to - yield the current processor to another thread in
4361 * your thread group, or accelerate that thread toward the
4362 * processor it's on.
4364 * @preempt: whether task preemption is allowed or not
4366 * It's the caller's job to ensure that the target task struct
4367 * can't go away on us before we can do any checks.
4370 * true (>0) if we indeed boosted the target task.
4371 * false (0) if we failed to boost the target.
4372 * -ESRCH if there's no task to yield to.
4374 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4376 struct task_struct
*curr
= current
;
4377 struct rq
*rq
, *p_rq
;
4378 unsigned long flags
;
4381 local_irq_save(flags
);
4387 * If we're the only runnable task on the rq and target rq also
4388 * has only one task, there's absolutely no point in yielding.
4390 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4395 double_rq_lock(rq
, p_rq
);
4396 while (task_rq(p
) != p_rq
) {
4397 double_rq_unlock(rq
, p_rq
);
4401 if (!curr
->sched_class
->yield_to_task
)
4404 if (curr
->sched_class
!= p
->sched_class
)
4407 if (task_running(p_rq
, p
) || p
->state
)
4410 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4412 schedstat_inc(rq
, yld_count
);
4414 * Make p's CPU reschedule; pick_next_entity takes care of
4417 if (preempt
&& rq
!= p_rq
)
4418 resched_task(p_rq
->curr
);
4422 double_rq_unlock(rq
, p_rq
);
4424 local_irq_restore(flags
);
4431 EXPORT_SYMBOL_GPL(yield_to
);
4434 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4435 * that process accounting knows that this is a task in IO wait state.
4437 void __sched
io_schedule(void)
4439 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;
4446 current
->in_iowait
= 0;
4447 atomic_dec(&rq
->nr_iowait
);
4448 delayacct_blkio_end();
4450 EXPORT_SYMBOL(io_schedule
);
4452 long __sched
io_schedule_timeout(long timeout
)
4454 struct rq
*rq
= raw_rq();
4457 delayacct_blkio_start();
4458 atomic_inc(&rq
->nr_iowait
);
4459 blk_flush_plug(current
);
4460 current
->in_iowait
= 1;
4461 ret
= schedule_timeout(timeout
);
4462 current
->in_iowait
= 0;
4463 atomic_dec(&rq
->nr_iowait
);
4464 delayacct_blkio_end();
4469 * sys_sched_get_priority_max - return maximum RT priority.
4470 * @policy: scheduling class.
4472 * this syscall returns the maximum rt_priority that can be used
4473 * by a given scheduling class.
4475 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4482 ret
= MAX_USER_RT_PRIO
-1;
4494 * sys_sched_get_priority_min - return minimum RT priority.
4495 * @policy: scheduling class.
4497 * this syscall returns the minimum rt_priority that can be used
4498 * by a given scheduling class.
4500 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4518 * sys_sched_rr_get_interval - return the default timeslice of a process.
4519 * @pid: pid of the process.
4520 * @interval: userspace pointer to the timeslice value.
4522 * this syscall writes the default timeslice value of a given process
4523 * into the user-space timespec buffer. A value of '0' means infinity.
4525 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4526 struct timespec __user
*, interval
)
4528 struct task_struct
*p
;
4529 unsigned int time_slice
;
4530 unsigned long flags
;
4540 p
= find_process_by_pid(pid
);
4544 retval
= security_task_getscheduler(p
);
4548 rq
= task_rq_lock(p
, &flags
);
4549 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4550 task_rq_unlock(rq
, p
, &flags
);
4553 jiffies_to_timespec(time_slice
, &t
);
4554 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4562 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4564 void sched_show_task(struct task_struct
*p
)
4566 unsigned long free
= 0;
4570 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4571 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4572 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4573 #if BITS_PER_LONG == 32
4574 if (state
== TASK_RUNNING
)
4575 printk(KERN_CONT
" running ");
4577 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4579 if (state
== TASK_RUNNING
)
4580 printk(KERN_CONT
" running task ");
4582 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4584 #ifdef CONFIG_DEBUG_STACK_USAGE
4585 free
= stack_not_used(p
);
4588 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4590 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4591 task_pid_nr(p
), ppid
,
4592 (unsigned long)task_thread_info(p
)->flags
);
4594 show_stack(p
, NULL
);
4597 void show_state_filter(unsigned long state_filter
)
4599 struct task_struct
*g
, *p
;
4601 #if BITS_PER_LONG == 32
4603 " task PC stack pid father\n");
4606 " task PC stack pid father\n");
4609 do_each_thread(g
, p
) {
4611 * reset the NMI-timeout, listing all files on a slow
4612 * console might take a lot of time:
4614 touch_nmi_watchdog();
4615 if (!state_filter
|| (p
->state
& state_filter
))
4617 } while_each_thread(g
, p
);
4619 touch_all_softlockup_watchdogs();
4621 #ifdef CONFIG_SCHED_DEBUG
4622 sysrq_sched_debug_show();
4626 * Only show locks if all tasks are dumped:
4629 debug_show_all_locks();
4632 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4634 idle
->sched_class
= &idle_sched_class
;
4638 * init_idle - set up an idle thread for a given CPU
4639 * @idle: task in question
4640 * @cpu: cpu the idle task belongs to
4642 * NOTE: this function does not set the idle thread's NEED_RESCHED
4643 * flag, to make booting more robust.
4645 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4647 struct rq
*rq
= cpu_rq(cpu
);
4648 unsigned long flags
;
4650 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4653 idle
->state
= TASK_RUNNING
;
4654 idle
->se
.exec_start
= sched_clock();
4656 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4658 * We're having a chicken and egg problem, even though we are
4659 * holding rq->lock, the cpu isn't yet set to this cpu so the
4660 * lockdep check in task_group() will fail.
4662 * Similar case to sched_fork(). / Alternatively we could
4663 * use task_rq_lock() here and obtain the other rq->lock.
4668 __set_task_cpu(idle
, cpu
);
4671 rq
->curr
= rq
->idle
= idle
;
4672 #if defined(CONFIG_SMP)
4675 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4677 /* Set the preempt count _outside_ the spinlocks! */
4678 task_thread_info(idle
)->preempt_count
= 0;
4681 * The idle tasks have their own, simple scheduling class:
4683 idle
->sched_class
= &idle_sched_class
;
4684 ftrace_graph_init_idle_task(idle
, cpu
);
4685 vtime_init_idle(idle
);
4686 #if defined(CONFIG_SMP)
4687 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4692 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4694 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4695 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4697 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4698 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4702 * This is how migration works:
4704 * 1) we invoke migration_cpu_stop() on the target CPU using
4706 * 2) stopper starts to run (implicitly forcing the migrated thread
4708 * 3) it checks whether the migrated task is still in the wrong runqueue.
4709 * 4) if it's in the wrong runqueue then the migration thread removes
4710 * it and puts it into the right queue.
4711 * 5) stopper completes and stop_one_cpu() returns and the migration
4716 * Change a given task's CPU affinity. Migrate the thread to a
4717 * proper CPU and schedule it away if the CPU it's executing on
4718 * is removed from the allowed bitmask.
4720 * NOTE: the caller must have a valid reference to the task, the
4721 * task must not exit() & deallocate itself prematurely. The
4722 * call is not atomic; no spinlocks may be held.
4724 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4726 unsigned long flags
;
4728 unsigned int dest_cpu
;
4731 rq
= task_rq_lock(p
, &flags
);
4733 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4736 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4741 do_set_cpus_allowed(p
, new_mask
);
4743 /* Can the task run on the task's current CPU? If so, we're done */
4744 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4747 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4749 struct migration_arg arg
= { p
, dest_cpu
};
4750 /* Need help from migration thread: drop lock and wait. */
4751 task_rq_unlock(rq
, p
, &flags
);
4752 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4753 tlb_migrate_finish(p
->mm
);
4757 task_rq_unlock(rq
, p
, &flags
);
4761 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4764 * Move (not current) task off this cpu, onto dest cpu. We're doing
4765 * this because either it can't run here any more (set_cpus_allowed()
4766 * away from this CPU, or CPU going down), or because we're
4767 * attempting to rebalance this task on exec (sched_exec).
4769 * So we race with normal scheduler movements, but that's OK, as long
4770 * as the task is no longer on this CPU.
4772 * Returns non-zero if task was successfully migrated.
4774 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4776 struct rq
*rq_dest
, *rq_src
;
4779 if (unlikely(!cpu_active(dest_cpu
)))
4782 rq_src
= cpu_rq(src_cpu
);
4783 rq_dest
= cpu_rq(dest_cpu
);
4785 raw_spin_lock(&p
->pi_lock
);
4786 double_rq_lock(rq_src
, rq_dest
);
4787 /* Already moved. */
4788 if (task_cpu(p
) != src_cpu
)
4790 /* Affinity changed (again). */
4791 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4795 * If we're not on a rq, the next wake-up will ensure we're
4799 dequeue_task(rq_src
, p
, 0);
4800 set_task_cpu(p
, dest_cpu
);
4801 enqueue_task(rq_dest
, p
, 0);
4802 check_preempt_curr(rq_dest
, p
, 0);
4807 double_rq_unlock(rq_src
, rq_dest
);
4808 raw_spin_unlock(&p
->pi_lock
);
4813 * migration_cpu_stop - this will be executed by a highprio stopper thread
4814 * and performs thread migration by bumping thread off CPU then
4815 * 'pushing' onto another runqueue.
4817 static int migration_cpu_stop(void *data
)
4819 struct migration_arg
*arg
= data
;
4822 * The original target cpu might have gone down and we might
4823 * be on another cpu but it doesn't matter.
4825 local_irq_disable();
4826 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4831 #ifdef CONFIG_HOTPLUG_CPU
4834 * Ensures that the idle task is using init_mm right before its cpu goes
4837 void idle_task_exit(void)
4839 struct mm_struct
*mm
= current
->active_mm
;
4841 BUG_ON(cpu_online(smp_processor_id()));
4844 switch_mm(mm
, &init_mm
, current
);
4849 * Since this CPU is going 'away' for a while, fold any nr_active delta
4850 * we might have. Assumes we're called after migrate_tasks() so that the
4851 * nr_active count is stable.
4853 * Also see the comment "Global load-average calculations".
4855 static void calc_load_migrate(struct rq
*rq
)
4857 long delta
= calc_load_fold_active(rq
);
4859 atomic_long_add(delta
, &calc_load_tasks
);
4863 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4864 * try_to_wake_up()->select_task_rq().
4866 * Called with rq->lock held even though we'er in stop_machine() and
4867 * there's no concurrency possible, we hold the required locks anyway
4868 * because of lock validation efforts.
4870 static void migrate_tasks(unsigned int dead_cpu
)
4872 struct rq
*rq
= cpu_rq(dead_cpu
);
4873 struct task_struct
*next
, *stop
= rq
->stop
;
4877 * Fudge the rq selection such that the below task selection loop
4878 * doesn't get stuck on the currently eligible stop task.
4880 * We're currently inside stop_machine() and the rq is either stuck
4881 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4882 * either way we should never end up calling schedule() until we're
4889 * There's this thread running, bail when that's the only
4892 if (rq
->nr_running
== 1)
4895 next
= pick_next_task(rq
);
4897 next
->sched_class
->put_prev_task(rq
, next
);
4899 /* Find suitable destination for @next, with force if needed. */
4900 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4901 raw_spin_unlock(&rq
->lock
);
4903 __migrate_task(next
, dead_cpu
, dest_cpu
);
4905 raw_spin_lock(&rq
->lock
);
4911 #endif /* CONFIG_HOTPLUG_CPU */
4913 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4915 static struct ctl_table sd_ctl_dir
[] = {
4917 .procname
= "sched_domain",
4923 static struct ctl_table sd_ctl_root
[] = {
4925 .procname
= "kernel",
4927 .child
= sd_ctl_dir
,
4932 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4934 struct ctl_table
*entry
=
4935 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4940 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4942 struct ctl_table
*entry
;
4945 * In the intermediate directories, both the child directory and
4946 * procname are dynamically allocated and could fail but the mode
4947 * will always be set. In the lowest directory the names are
4948 * static strings and all have proc handlers.
4950 for (entry
= *tablep
; entry
->mode
; entry
++) {
4952 sd_free_ctl_entry(&entry
->child
);
4953 if (entry
->proc_handler
== NULL
)
4954 kfree(entry
->procname
);
4961 static int min_load_idx
= 0;
4962 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4965 set_table_entry(struct ctl_table
*entry
,
4966 const char *procname
, void *data
, int maxlen
,
4967 umode_t mode
, proc_handler
*proc_handler
,
4970 entry
->procname
= procname
;
4972 entry
->maxlen
= maxlen
;
4974 entry
->proc_handler
= proc_handler
;
4977 entry
->extra1
= &min_load_idx
;
4978 entry
->extra2
= &max_load_idx
;
4982 static struct ctl_table
*
4983 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4985 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4990 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4991 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4992 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4993 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4994 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4995 sizeof(int), 0644, proc_dointvec_minmax
, true);
4996 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4997 sizeof(int), 0644, proc_dointvec_minmax
, true);
4998 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4999 sizeof(int), 0644, proc_dointvec_minmax
, true);
5000 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5001 sizeof(int), 0644, proc_dointvec_minmax
, true);
5002 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5003 sizeof(int), 0644, proc_dointvec_minmax
, true);
5004 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5005 sizeof(int), 0644, proc_dointvec_minmax
, false);
5006 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5007 sizeof(int), 0644, proc_dointvec_minmax
, false);
5008 set_table_entry(&table
[9], "cache_nice_tries",
5009 &sd
->cache_nice_tries
,
5010 sizeof(int), 0644, proc_dointvec_minmax
, false);
5011 set_table_entry(&table
[10], "flags", &sd
->flags
,
5012 sizeof(int), 0644, proc_dointvec_minmax
, false);
5013 set_table_entry(&table
[11], "name", sd
->name
,
5014 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5015 /* &table[12] is terminator */
5020 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5022 struct ctl_table
*entry
, *table
;
5023 struct sched_domain
*sd
;
5024 int domain_num
= 0, i
;
5027 for_each_domain(cpu
, sd
)
5029 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5034 for_each_domain(cpu
, sd
) {
5035 snprintf(buf
, 32, "domain%d", i
);
5036 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5038 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5045 static struct ctl_table_header
*sd_sysctl_header
;
5046 static void register_sched_domain_sysctl(void)
5048 int i
, cpu_num
= num_possible_cpus();
5049 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5052 WARN_ON(sd_ctl_dir
[0].child
);
5053 sd_ctl_dir
[0].child
= entry
;
5058 for_each_possible_cpu(i
) {
5059 snprintf(buf
, 32, "cpu%d", i
);
5060 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5062 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5066 WARN_ON(sd_sysctl_header
);
5067 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5070 /* may be called multiple times per register */
5071 static void unregister_sched_domain_sysctl(void)
5073 if (sd_sysctl_header
)
5074 unregister_sysctl_table(sd_sysctl_header
);
5075 sd_sysctl_header
= NULL
;
5076 if (sd_ctl_dir
[0].child
)
5077 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5080 static void register_sched_domain_sysctl(void)
5083 static void unregister_sched_domain_sysctl(void)
5088 static void set_rq_online(struct rq
*rq
)
5091 const struct sched_class
*class;
5093 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5096 for_each_class(class) {
5097 if (class->rq_online
)
5098 class->rq_online(rq
);
5103 static void set_rq_offline(struct rq
*rq
)
5106 const struct sched_class
*class;
5108 for_each_class(class) {
5109 if (class->rq_offline
)
5110 class->rq_offline(rq
);
5113 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5119 * migration_call - callback that gets triggered when a CPU is added.
5120 * Here we can start up the necessary migration thread for the new CPU.
5122 static int __cpuinit
5123 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5125 int cpu
= (long)hcpu
;
5126 unsigned long flags
;
5127 struct rq
*rq
= cpu_rq(cpu
);
5129 switch (action
& ~CPU_TASKS_FROZEN
) {
5131 case CPU_UP_PREPARE
:
5132 rq
->calc_load_update
= calc_load_update
;
5136 /* Update our root-domain */
5137 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5139 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5143 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5146 #ifdef CONFIG_HOTPLUG_CPU
5148 sched_ttwu_pending();
5149 /* Update our root-domain */
5150 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5152 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5156 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5157 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5161 calc_load_migrate(rq
);
5166 update_max_interval();
5172 * Register at high priority so that task migration (migrate_all_tasks)
5173 * happens before everything else. This has to be lower priority than
5174 * the notifier in the perf_event subsystem, though.
5176 static struct notifier_block __cpuinitdata migration_notifier
= {
5177 .notifier_call
= migration_call
,
5178 .priority
= CPU_PRI_MIGRATION
,
5181 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5182 unsigned long action
, void *hcpu
)
5184 switch (action
& ~CPU_TASKS_FROZEN
) {
5186 case CPU_DOWN_FAILED
:
5187 set_cpu_active((long)hcpu
, true);
5194 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5195 unsigned long action
, void *hcpu
)
5197 switch (action
& ~CPU_TASKS_FROZEN
) {
5198 case CPU_DOWN_PREPARE
:
5199 set_cpu_active((long)hcpu
, false);
5206 static int __init
migration_init(void)
5208 void *cpu
= (void *)(long)smp_processor_id();
5211 /* Initialize migration for the boot CPU */
5212 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5213 BUG_ON(err
== NOTIFY_BAD
);
5214 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5215 register_cpu_notifier(&migration_notifier
);
5217 /* Register cpu active notifiers */
5218 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5219 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5223 early_initcall(migration_init
);
5228 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5230 #ifdef CONFIG_SCHED_DEBUG
5232 static __read_mostly
int sched_debug_enabled
;
5234 static int __init
sched_debug_setup(char *str
)
5236 sched_debug_enabled
= 1;
5240 early_param("sched_debug", sched_debug_setup
);
5242 static inline bool sched_debug(void)
5244 return sched_debug_enabled
;
5247 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5248 struct cpumask
*groupmask
)
5250 struct sched_group
*group
= sd
->groups
;
5253 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5254 cpumask_clear(groupmask
);
5256 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5258 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5259 printk("does not load-balance\n");
5261 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5266 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5268 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5269 printk(KERN_ERR
"ERROR: domain->span does not contain "
5272 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5273 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5277 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5281 printk(KERN_ERR
"ERROR: group is NULL\n");
5286 * Even though we initialize ->power to something semi-sane,
5287 * we leave power_orig unset. This allows us to detect if
5288 * domain iteration is still funny without causing /0 traps.
5290 if (!group
->sgp
->power_orig
) {
5291 printk(KERN_CONT
"\n");
5292 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5297 if (!cpumask_weight(sched_group_cpus(group
))) {
5298 printk(KERN_CONT
"\n");
5299 printk(KERN_ERR
"ERROR: empty group\n");
5303 if (!(sd
->flags
& SD_OVERLAP
) &&
5304 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5305 printk(KERN_CONT
"\n");
5306 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5310 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5312 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5314 printk(KERN_CONT
" %s", str
);
5315 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5316 printk(KERN_CONT
" (cpu_power = %d)",
5320 group
= group
->next
;
5321 } while (group
!= sd
->groups
);
5322 printk(KERN_CONT
"\n");
5324 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5325 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5328 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5329 printk(KERN_ERR
"ERROR: parent span is not a superset "
5330 "of domain->span\n");
5334 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5338 if (!sched_debug_enabled
)
5342 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5346 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5349 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5357 #else /* !CONFIG_SCHED_DEBUG */
5358 # define sched_domain_debug(sd, cpu) do { } while (0)
5359 static inline bool sched_debug(void)
5363 #endif /* CONFIG_SCHED_DEBUG */
5365 static int sd_degenerate(struct sched_domain
*sd
)
5367 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5370 /* Following flags need at least 2 groups */
5371 if (sd
->flags
& (SD_LOAD_BALANCE
|
5372 SD_BALANCE_NEWIDLE
|
5376 SD_SHARE_PKG_RESOURCES
)) {
5377 if (sd
->groups
!= sd
->groups
->next
)
5381 /* Following flags don't use groups */
5382 if (sd
->flags
& (SD_WAKE_AFFINE
))
5389 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5391 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5393 if (sd_degenerate(parent
))
5396 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5399 /* Flags needing groups don't count if only 1 group in parent */
5400 if (parent
->groups
== parent
->groups
->next
) {
5401 pflags
&= ~(SD_LOAD_BALANCE
|
5402 SD_BALANCE_NEWIDLE
|
5406 SD_SHARE_PKG_RESOURCES
);
5407 if (nr_node_ids
== 1)
5408 pflags
&= ~SD_SERIALIZE
;
5410 if (~cflags
& pflags
)
5416 static void free_rootdomain(struct rcu_head
*rcu
)
5418 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5420 cpupri_cleanup(&rd
->cpupri
);
5421 free_cpumask_var(rd
->rto_mask
);
5422 free_cpumask_var(rd
->online
);
5423 free_cpumask_var(rd
->span
);
5427 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5429 struct root_domain
*old_rd
= NULL
;
5430 unsigned long flags
;
5432 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5437 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5440 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5443 * If we dont want to free the old_rt yet then
5444 * set old_rd to NULL to skip the freeing later
5447 if (!atomic_dec_and_test(&old_rd
->refcount
))
5451 atomic_inc(&rd
->refcount
);
5454 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5455 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5458 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5461 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5464 static int init_rootdomain(struct root_domain
*rd
)
5466 memset(rd
, 0, sizeof(*rd
));
5468 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5470 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5472 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5475 if (cpupri_init(&rd
->cpupri
) != 0)
5480 free_cpumask_var(rd
->rto_mask
);
5482 free_cpumask_var(rd
->online
);
5484 free_cpumask_var(rd
->span
);
5490 * By default the system creates a single root-domain with all cpus as
5491 * members (mimicking the global state we have today).
5493 struct root_domain def_root_domain
;
5495 static void init_defrootdomain(void)
5497 init_rootdomain(&def_root_domain
);
5499 atomic_set(&def_root_domain
.refcount
, 1);
5502 static struct root_domain
*alloc_rootdomain(void)
5504 struct root_domain
*rd
;
5506 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5510 if (init_rootdomain(rd
) != 0) {
5518 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5520 struct sched_group
*tmp
, *first
;
5529 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5534 } while (sg
!= first
);
5537 static void free_sched_domain(struct rcu_head
*rcu
)
5539 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5542 * If its an overlapping domain it has private groups, iterate and
5545 if (sd
->flags
& SD_OVERLAP
) {
5546 free_sched_groups(sd
->groups
, 1);
5547 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5548 kfree(sd
->groups
->sgp
);
5554 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5556 call_rcu(&sd
->rcu
, free_sched_domain
);
5559 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5561 for (; sd
; sd
= sd
->parent
)
5562 destroy_sched_domain(sd
, cpu
);
5566 * Keep a special pointer to the highest sched_domain that has
5567 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5568 * allows us to avoid some pointer chasing select_idle_sibling().
5570 * Also keep a unique ID per domain (we use the first cpu number in
5571 * the cpumask of the domain), this allows us to quickly tell if
5572 * two cpus are in the same cache domain, see cpus_share_cache().
5574 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5575 DEFINE_PER_CPU(int, sd_llc_id
);
5577 static void update_top_cache_domain(int cpu
)
5579 struct sched_domain
*sd
;
5582 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5584 id
= cpumask_first(sched_domain_span(sd
));
5586 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5587 per_cpu(sd_llc_id
, cpu
) = id
;
5591 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5592 * hold the hotplug lock.
5595 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5597 struct rq
*rq
= cpu_rq(cpu
);
5598 struct sched_domain
*tmp
;
5600 /* Remove the sched domains which do not contribute to scheduling. */
5601 for (tmp
= sd
; tmp
; ) {
5602 struct sched_domain
*parent
= tmp
->parent
;
5606 if (sd_parent_degenerate(tmp
, parent
)) {
5607 tmp
->parent
= parent
->parent
;
5609 parent
->parent
->child
= tmp
;
5610 destroy_sched_domain(parent
, cpu
);
5615 if (sd
&& sd_degenerate(sd
)) {
5618 destroy_sched_domain(tmp
, cpu
);
5623 sched_domain_debug(sd
, cpu
);
5625 rq_attach_root(rq
, rd
);
5627 rcu_assign_pointer(rq
->sd
, sd
);
5628 destroy_sched_domains(tmp
, cpu
);
5630 update_top_cache_domain(cpu
);
5633 /* cpus with isolated domains */
5634 static cpumask_var_t cpu_isolated_map
;
5636 /* Setup the mask of cpus configured for isolated domains */
5637 static int __init
isolated_cpu_setup(char *str
)
5639 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5640 cpulist_parse(str
, cpu_isolated_map
);
5644 __setup("isolcpus=", isolated_cpu_setup
);
5646 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5648 return cpumask_of_node(cpu_to_node(cpu
));
5652 struct sched_domain
**__percpu sd
;
5653 struct sched_group
**__percpu sg
;
5654 struct sched_group_power
**__percpu sgp
;
5658 struct sched_domain
** __percpu sd
;
5659 struct root_domain
*rd
;
5669 struct sched_domain_topology_level
;
5671 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5672 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5674 #define SDTL_OVERLAP 0x01
5676 struct sched_domain_topology_level
{
5677 sched_domain_init_f init
;
5678 sched_domain_mask_f mask
;
5681 struct sd_data data
;
5685 * Build an iteration mask that can exclude certain CPUs from the upwards
5688 * Asymmetric node setups can result in situations where the domain tree is of
5689 * unequal depth, make sure to skip domains that already cover the entire
5692 * In that case build_sched_domains() will have terminated the iteration early
5693 * and our sibling sd spans will be empty. Domains should always include the
5694 * cpu they're built on, so check that.
5697 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5699 const struct cpumask
*span
= sched_domain_span(sd
);
5700 struct sd_data
*sdd
= sd
->private;
5701 struct sched_domain
*sibling
;
5704 for_each_cpu(i
, span
) {
5705 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5706 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5709 cpumask_set_cpu(i
, sched_group_mask(sg
));
5714 * Return the canonical balance cpu for this group, this is the first cpu
5715 * of this group that's also in the iteration mask.
5717 int group_balance_cpu(struct sched_group
*sg
)
5719 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5723 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5725 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5726 const struct cpumask
*span
= sched_domain_span(sd
);
5727 struct cpumask
*covered
= sched_domains_tmpmask
;
5728 struct sd_data
*sdd
= sd
->private;
5729 struct sched_domain
*child
;
5732 cpumask_clear(covered
);
5734 for_each_cpu(i
, span
) {
5735 struct cpumask
*sg_span
;
5737 if (cpumask_test_cpu(i
, covered
))
5740 child
= *per_cpu_ptr(sdd
->sd
, i
);
5742 /* See the comment near build_group_mask(). */
5743 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5746 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5747 GFP_KERNEL
, cpu_to_node(cpu
));
5752 sg_span
= sched_group_cpus(sg
);
5754 child
= child
->child
;
5755 cpumask_copy(sg_span
, sched_domain_span(child
));
5757 cpumask_set_cpu(i
, sg_span
);
5759 cpumask_or(covered
, covered
, sg_span
);
5761 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5762 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5763 build_group_mask(sd
, sg
);
5766 * Initialize sgp->power such that even if we mess up the
5767 * domains and no possible iteration will get us here, we won't
5770 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5773 * Make sure the first group of this domain contains the
5774 * canonical balance cpu. Otherwise the sched_domain iteration
5775 * breaks. See update_sg_lb_stats().
5777 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5778 group_balance_cpu(sg
) == cpu
)
5788 sd
->groups
= groups
;
5793 free_sched_groups(first
, 0);
5798 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5800 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5801 struct sched_domain
*child
= sd
->child
;
5804 cpu
= cpumask_first(sched_domain_span(child
));
5807 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5808 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5809 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5816 * build_sched_groups will build a circular linked list of the groups
5817 * covered by the given span, and will set each group's ->cpumask correctly,
5818 * and ->cpu_power to 0.
5820 * Assumes the sched_domain tree is fully constructed
5823 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5825 struct sched_group
*first
= NULL
, *last
= NULL
;
5826 struct sd_data
*sdd
= sd
->private;
5827 const struct cpumask
*span
= sched_domain_span(sd
);
5828 struct cpumask
*covered
;
5831 get_group(cpu
, sdd
, &sd
->groups
);
5832 atomic_inc(&sd
->groups
->ref
);
5834 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5837 lockdep_assert_held(&sched_domains_mutex
);
5838 covered
= sched_domains_tmpmask
;
5840 cpumask_clear(covered
);
5842 for_each_cpu(i
, span
) {
5843 struct sched_group
*sg
;
5844 int group
= get_group(i
, sdd
, &sg
);
5847 if (cpumask_test_cpu(i
, covered
))
5850 cpumask_clear(sched_group_cpus(sg
));
5852 cpumask_setall(sched_group_mask(sg
));
5854 for_each_cpu(j
, span
) {
5855 if (get_group(j
, sdd
, NULL
) != group
)
5858 cpumask_set_cpu(j
, covered
);
5859 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5874 * Initialize sched groups cpu_power.
5876 * cpu_power indicates the capacity of sched group, which is used while
5877 * distributing the load between different sched groups in a sched domain.
5878 * Typically cpu_power for all the groups in a sched domain will be same unless
5879 * there are asymmetries in the topology. If there are asymmetries, group
5880 * having more cpu_power will pickup more load compared to the group having
5883 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5885 struct sched_group
*sg
= sd
->groups
;
5887 WARN_ON(!sd
|| !sg
);
5890 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5892 } while (sg
!= sd
->groups
);
5894 if (cpu
!= group_balance_cpu(sg
))
5897 update_group_power(sd
, cpu
);
5898 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5901 int __weak
arch_sd_sibling_asym_packing(void)
5903 return 0*SD_ASYM_PACKING
;
5907 * Initializers for schedule domains
5908 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5911 #ifdef CONFIG_SCHED_DEBUG
5912 # define SD_INIT_NAME(sd, type) sd->name = #type
5914 # define SD_INIT_NAME(sd, type) do { } while (0)
5917 #define SD_INIT_FUNC(type) \
5918 static noinline struct sched_domain * \
5919 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5921 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5922 *sd = SD_##type##_INIT; \
5923 SD_INIT_NAME(sd, type); \
5924 sd->private = &tl->data; \
5929 #ifdef CONFIG_SCHED_SMT
5930 SD_INIT_FUNC(SIBLING
)
5932 #ifdef CONFIG_SCHED_MC
5935 #ifdef CONFIG_SCHED_BOOK
5939 static int default_relax_domain_level
= -1;
5940 int sched_domain_level_max
;
5942 static int __init
setup_relax_domain_level(char *str
)
5944 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5945 pr_warn("Unable to set relax_domain_level\n");
5949 __setup("relax_domain_level=", setup_relax_domain_level
);
5951 static void set_domain_attribute(struct sched_domain
*sd
,
5952 struct sched_domain_attr
*attr
)
5956 if (!attr
|| attr
->relax_domain_level
< 0) {
5957 if (default_relax_domain_level
< 0)
5960 request
= default_relax_domain_level
;
5962 request
= attr
->relax_domain_level
;
5963 if (request
< sd
->level
) {
5964 /* turn off idle balance on this domain */
5965 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5967 /* turn on idle balance on this domain */
5968 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5972 static void __sdt_free(const struct cpumask
*cpu_map
);
5973 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5975 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5976 const struct cpumask
*cpu_map
)
5980 if (!atomic_read(&d
->rd
->refcount
))
5981 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5983 free_percpu(d
->sd
); /* fall through */
5985 __sdt_free(cpu_map
); /* fall through */
5991 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5992 const struct cpumask
*cpu_map
)
5994 memset(d
, 0, sizeof(*d
));
5996 if (__sdt_alloc(cpu_map
))
5997 return sa_sd_storage
;
5998 d
->sd
= alloc_percpu(struct sched_domain
*);
6000 return sa_sd_storage
;
6001 d
->rd
= alloc_rootdomain();
6004 return sa_rootdomain
;
6008 * NULL the sd_data elements we've used to build the sched_domain and
6009 * sched_group structure so that the subsequent __free_domain_allocs()
6010 * will not free the data we're using.
6012 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6014 struct sd_data
*sdd
= sd
->private;
6016 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6017 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6019 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6020 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6022 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6023 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6026 #ifdef CONFIG_SCHED_SMT
6027 static const struct cpumask
*cpu_smt_mask(int cpu
)
6029 return topology_thread_cpumask(cpu
);
6034 * Topology list, bottom-up.
6036 static struct sched_domain_topology_level default_topology
[] = {
6037 #ifdef CONFIG_SCHED_SMT
6038 { sd_init_SIBLING
, cpu_smt_mask
, },
6040 #ifdef CONFIG_SCHED_MC
6041 { sd_init_MC
, cpu_coregroup_mask
, },
6043 #ifdef CONFIG_SCHED_BOOK
6044 { sd_init_BOOK
, cpu_book_mask
, },
6046 { sd_init_CPU
, cpu_cpu_mask
, },
6050 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6054 static int sched_domains_numa_levels
;
6055 static int *sched_domains_numa_distance
;
6056 static struct cpumask
***sched_domains_numa_masks
;
6057 static int sched_domains_curr_level
;
6059 static inline int sd_local_flags(int level
)
6061 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6064 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6067 static struct sched_domain
*
6068 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6070 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6071 int level
= tl
->numa_level
;
6072 int sd_weight
= cpumask_weight(
6073 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6075 *sd
= (struct sched_domain
){
6076 .min_interval
= sd_weight
,
6077 .max_interval
= 2*sd_weight
,
6079 .imbalance_pct
= 125,
6080 .cache_nice_tries
= 2,
6087 .flags
= 1*SD_LOAD_BALANCE
6088 | 1*SD_BALANCE_NEWIDLE
6093 | 0*SD_SHARE_CPUPOWER
6094 | 0*SD_SHARE_PKG_RESOURCES
6096 | 0*SD_PREFER_SIBLING
6097 | sd_local_flags(level
)
6099 .last_balance
= jiffies
,
6100 .balance_interval
= sd_weight
,
6102 SD_INIT_NAME(sd
, NUMA
);
6103 sd
->private = &tl
->data
;
6106 * Ugly hack to pass state to sd_numa_mask()...
6108 sched_domains_curr_level
= tl
->numa_level
;
6113 static const struct cpumask
*sd_numa_mask(int cpu
)
6115 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6118 static void sched_numa_warn(const char *str
)
6120 static int done
= false;
6128 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6130 for (i
= 0; i
< nr_node_ids
; i
++) {
6131 printk(KERN_WARNING
" ");
6132 for (j
= 0; j
< nr_node_ids
; j
++)
6133 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6134 printk(KERN_CONT
"\n");
6136 printk(KERN_WARNING
"\n");
6139 static bool find_numa_distance(int distance
)
6143 if (distance
== node_distance(0, 0))
6146 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6147 if (sched_domains_numa_distance
[i
] == distance
)
6154 static void sched_init_numa(void)
6156 int next_distance
, curr_distance
= node_distance(0, 0);
6157 struct sched_domain_topology_level
*tl
;
6161 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6162 if (!sched_domains_numa_distance
)
6166 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6167 * unique distances in the node_distance() table.
6169 * Assumes node_distance(0,j) includes all distances in
6170 * node_distance(i,j) in order to avoid cubic time.
6172 next_distance
= curr_distance
;
6173 for (i
= 0; i
< nr_node_ids
; i
++) {
6174 for (j
= 0; j
< nr_node_ids
; j
++) {
6175 for (k
= 0; k
< nr_node_ids
; k
++) {
6176 int distance
= node_distance(i
, k
);
6178 if (distance
> curr_distance
&&
6179 (distance
< next_distance
||
6180 next_distance
== curr_distance
))
6181 next_distance
= distance
;
6184 * While not a strong assumption it would be nice to know
6185 * about cases where if node A is connected to B, B is not
6186 * equally connected to A.
6188 if (sched_debug() && node_distance(k
, i
) != distance
)
6189 sched_numa_warn("Node-distance not symmetric");
6191 if (sched_debug() && i
&& !find_numa_distance(distance
))
6192 sched_numa_warn("Node-0 not representative");
6194 if (next_distance
!= curr_distance
) {
6195 sched_domains_numa_distance
[level
++] = next_distance
;
6196 sched_domains_numa_levels
= level
;
6197 curr_distance
= next_distance
;
6202 * In case of sched_debug() we verify the above assumption.
6208 * 'level' contains the number of unique distances, excluding the
6209 * identity distance node_distance(i,i).
6211 * The sched_domains_numa_distance[] array includes the actual distance
6216 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6217 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6218 * the array will contain less then 'level' members. This could be
6219 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6220 * in other functions.
6222 * We reset it to 'level' at the end of this function.
6224 sched_domains_numa_levels
= 0;
6226 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6227 if (!sched_domains_numa_masks
)
6231 * Now for each level, construct a mask per node which contains all
6232 * cpus of nodes that are that many hops away from us.
6234 for (i
= 0; i
< level
; i
++) {
6235 sched_domains_numa_masks
[i
] =
6236 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6237 if (!sched_domains_numa_masks
[i
])
6240 for (j
= 0; j
< nr_node_ids
; j
++) {
6241 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6245 sched_domains_numa_masks
[i
][j
] = mask
;
6247 for (k
= 0; k
< nr_node_ids
; k
++) {
6248 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6251 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6256 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6257 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6262 * Copy the default topology bits..
6264 for (i
= 0; default_topology
[i
].init
; i
++)
6265 tl
[i
] = default_topology
[i
];
6268 * .. and append 'j' levels of NUMA goodness.
6270 for (j
= 0; j
< level
; i
++, j
++) {
6271 tl
[i
] = (struct sched_domain_topology_level
){
6272 .init
= sd_numa_init
,
6273 .mask
= sd_numa_mask
,
6274 .flags
= SDTL_OVERLAP
,
6279 sched_domain_topology
= tl
;
6281 sched_domains_numa_levels
= level
;
6284 static void sched_domains_numa_masks_set(int cpu
)
6287 int node
= cpu_to_node(cpu
);
6289 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6290 for (j
= 0; j
< nr_node_ids
; j
++) {
6291 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6292 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6297 static void sched_domains_numa_masks_clear(int cpu
)
6300 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6301 for (j
= 0; j
< nr_node_ids
; j
++)
6302 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6307 * Update sched_domains_numa_masks[level][node] array when new cpus
6310 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6311 unsigned long action
,
6314 int cpu
= (long)hcpu
;
6316 switch (action
& ~CPU_TASKS_FROZEN
) {
6318 sched_domains_numa_masks_set(cpu
);
6322 sched_domains_numa_masks_clear(cpu
);
6332 static inline void sched_init_numa(void)
6336 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6337 unsigned long action
,
6342 #endif /* CONFIG_NUMA */
6344 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6346 struct sched_domain_topology_level
*tl
;
6349 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6350 struct sd_data
*sdd
= &tl
->data
;
6352 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6356 sdd
->sg
= alloc_percpu(struct sched_group
*);
6360 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6364 for_each_cpu(j
, cpu_map
) {
6365 struct sched_domain
*sd
;
6366 struct sched_group
*sg
;
6367 struct sched_group_power
*sgp
;
6369 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6370 GFP_KERNEL
, cpu_to_node(j
));
6374 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6376 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6377 GFP_KERNEL
, cpu_to_node(j
));
6383 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6385 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6386 GFP_KERNEL
, cpu_to_node(j
));
6390 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6397 static void __sdt_free(const struct cpumask
*cpu_map
)
6399 struct sched_domain_topology_level
*tl
;
6402 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6403 struct sd_data
*sdd
= &tl
->data
;
6405 for_each_cpu(j
, cpu_map
) {
6406 struct sched_domain
*sd
;
6409 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6410 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6411 free_sched_groups(sd
->groups
, 0);
6412 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6416 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6418 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6420 free_percpu(sdd
->sd
);
6422 free_percpu(sdd
->sg
);
6424 free_percpu(sdd
->sgp
);
6429 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6430 struct s_data
*d
, const struct cpumask
*cpu_map
,
6431 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6434 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6438 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6440 sd
->level
= child
->level
+ 1;
6441 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6445 set_domain_attribute(sd
, attr
);
6451 * Build sched domains for a given set of cpus and attach the sched domains
6452 * to the individual cpus
6454 static int build_sched_domains(const struct cpumask
*cpu_map
,
6455 struct sched_domain_attr
*attr
)
6457 enum s_alloc alloc_state
= sa_none
;
6458 struct sched_domain
*sd
;
6460 int i
, ret
= -ENOMEM
;
6462 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6463 if (alloc_state
!= sa_rootdomain
)
6466 /* Set up domains for cpus specified by the cpu_map. */
6467 for_each_cpu(i
, cpu_map
) {
6468 struct sched_domain_topology_level
*tl
;
6471 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6472 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6473 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6474 sd
->flags
|= SD_OVERLAP
;
6475 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6482 *per_cpu_ptr(d
.sd
, i
) = sd
;
6485 /* Build the groups for the domains */
6486 for_each_cpu(i
, cpu_map
) {
6487 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6488 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6489 if (sd
->flags
& SD_OVERLAP
) {
6490 if (build_overlap_sched_groups(sd
, i
))
6493 if (build_sched_groups(sd
, i
))
6499 /* Calculate CPU power for physical packages and nodes */
6500 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6501 if (!cpumask_test_cpu(i
, cpu_map
))
6504 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6505 claim_allocations(i
, sd
);
6506 init_sched_groups_power(i
, sd
);
6510 /* Attach the domains */
6512 for_each_cpu(i
, cpu_map
) {
6513 sd
= *per_cpu_ptr(d
.sd
, i
);
6514 cpu_attach_domain(sd
, d
.rd
, i
);
6520 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6524 static cpumask_var_t
*doms_cur
; /* current sched domains */
6525 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6526 static struct sched_domain_attr
*dattr_cur
;
6527 /* attribues of custom domains in 'doms_cur' */
6530 * Special case: If a kmalloc of a doms_cur partition (array of
6531 * cpumask) fails, then fallback to a single sched domain,
6532 * as determined by the single cpumask fallback_doms.
6534 static cpumask_var_t fallback_doms
;
6537 * arch_update_cpu_topology lets virtualized architectures update the
6538 * cpu core maps. It is supposed to return 1 if the topology changed
6539 * or 0 if it stayed the same.
6541 int __attribute__((weak
)) arch_update_cpu_topology(void)
6546 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6549 cpumask_var_t
*doms
;
6551 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6554 for (i
= 0; i
< ndoms
; i
++) {
6555 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6556 free_sched_domains(doms
, i
);
6563 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6566 for (i
= 0; i
< ndoms
; i
++)
6567 free_cpumask_var(doms
[i
]);
6572 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6573 * For now this just excludes isolated cpus, but could be used to
6574 * exclude other special cases in the future.
6576 static int init_sched_domains(const struct cpumask
*cpu_map
)
6580 arch_update_cpu_topology();
6582 doms_cur
= alloc_sched_domains(ndoms_cur
);
6584 doms_cur
= &fallback_doms
;
6585 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6586 err
= build_sched_domains(doms_cur
[0], NULL
);
6587 register_sched_domain_sysctl();
6593 * Detach sched domains from a group of cpus specified in cpu_map
6594 * These cpus will now be attached to the NULL domain
6596 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6601 for_each_cpu(i
, cpu_map
)
6602 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6606 /* handle null as "default" */
6607 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6608 struct sched_domain_attr
*new, int idx_new
)
6610 struct sched_domain_attr tmp
;
6617 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6618 new ? (new + idx_new
) : &tmp
,
6619 sizeof(struct sched_domain_attr
));
6623 * Partition sched domains as specified by the 'ndoms_new'
6624 * cpumasks in the array doms_new[] of cpumasks. This compares
6625 * doms_new[] to the current sched domain partitioning, doms_cur[].
6626 * It destroys each deleted domain and builds each new domain.
6628 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6629 * The masks don't intersect (don't overlap.) We should setup one
6630 * sched domain for each mask. CPUs not in any of the cpumasks will
6631 * not be load balanced. If the same cpumask appears both in the
6632 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6635 * The passed in 'doms_new' should be allocated using
6636 * alloc_sched_domains. This routine takes ownership of it and will
6637 * free_sched_domains it when done with it. If the caller failed the
6638 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6639 * and partition_sched_domains() will fallback to the single partition
6640 * 'fallback_doms', it also forces the domains to be rebuilt.
6642 * If doms_new == NULL it will be replaced with cpu_online_mask.
6643 * ndoms_new == 0 is a special case for destroying existing domains,
6644 * and it will not create the default domain.
6646 * Call with hotplug lock held
6648 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6649 struct sched_domain_attr
*dattr_new
)
6654 mutex_lock(&sched_domains_mutex
);
6656 /* always unregister in case we don't destroy any domains */
6657 unregister_sched_domain_sysctl();
6659 /* Let architecture update cpu core mappings. */
6660 new_topology
= arch_update_cpu_topology();
6662 n
= doms_new
? ndoms_new
: 0;
6664 /* Destroy deleted domains */
6665 for (i
= 0; i
< ndoms_cur
; i
++) {
6666 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6667 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6668 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6671 /* no match - a current sched domain not in new doms_new[] */
6672 detach_destroy_domains(doms_cur
[i
]);
6677 if (doms_new
== NULL
) {
6679 doms_new
= &fallback_doms
;
6680 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6681 WARN_ON_ONCE(dattr_new
);
6684 /* Build new domains */
6685 for (i
= 0; i
< ndoms_new
; i
++) {
6686 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6687 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6688 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6691 /* no match - add a new doms_new */
6692 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6697 /* Remember the new sched domains */
6698 if (doms_cur
!= &fallback_doms
)
6699 free_sched_domains(doms_cur
, ndoms_cur
);
6700 kfree(dattr_cur
); /* kfree(NULL) is safe */
6701 doms_cur
= doms_new
;
6702 dattr_cur
= dattr_new
;
6703 ndoms_cur
= ndoms_new
;
6705 register_sched_domain_sysctl();
6707 mutex_unlock(&sched_domains_mutex
);
6710 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6713 * Update cpusets according to cpu_active mask. If cpusets are
6714 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6715 * around partition_sched_domains().
6717 * If we come here as part of a suspend/resume, don't touch cpusets because we
6718 * want to restore it back to its original state upon resume anyway.
6720 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6724 case CPU_ONLINE_FROZEN
:
6725 case CPU_DOWN_FAILED_FROZEN
:
6728 * num_cpus_frozen tracks how many CPUs are involved in suspend
6729 * resume sequence. As long as this is not the last online
6730 * operation in the resume sequence, just build a single sched
6731 * domain, ignoring cpusets.
6734 if (likely(num_cpus_frozen
)) {
6735 partition_sched_domains(1, NULL
, NULL
);
6740 * This is the last CPU online operation. So fall through and
6741 * restore the original sched domains by considering the
6742 * cpuset configurations.
6746 case CPU_DOWN_FAILED
:
6747 cpuset_update_active_cpus(true);
6755 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6759 case CPU_DOWN_PREPARE
:
6760 cpuset_update_active_cpus(false);
6762 case CPU_DOWN_PREPARE_FROZEN
:
6764 partition_sched_domains(1, NULL
, NULL
);
6772 void __init
sched_init_smp(void)
6774 cpumask_var_t non_isolated_cpus
;
6776 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6777 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6782 mutex_lock(&sched_domains_mutex
);
6783 init_sched_domains(cpu_active_mask
);
6784 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6785 if (cpumask_empty(non_isolated_cpus
))
6786 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6787 mutex_unlock(&sched_domains_mutex
);
6790 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6791 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6792 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6794 /* RT runtime code needs to handle some hotplug events */
6795 hotcpu_notifier(update_runtime
, 0);
6799 /* Move init over to a non-isolated CPU */
6800 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6802 sched_init_granularity();
6803 free_cpumask_var(non_isolated_cpus
);
6805 init_sched_rt_class();
6808 void __init
sched_init_smp(void)
6810 sched_init_granularity();
6812 #endif /* CONFIG_SMP */
6814 const_debug
unsigned int sysctl_timer_migration
= 1;
6816 int in_sched_functions(unsigned long addr
)
6818 return in_lock_functions(addr
) ||
6819 (addr
>= (unsigned long)__sched_text_start
6820 && addr
< (unsigned long)__sched_text_end
);
6823 #ifdef CONFIG_CGROUP_SCHED
6825 * Default task group.
6826 * Every task in system belongs to this group at bootup.
6828 struct task_group root_task_group
;
6829 LIST_HEAD(task_groups
);
6832 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6834 void __init
sched_init(void)
6837 unsigned long alloc_size
= 0, ptr
;
6839 #ifdef CONFIG_FAIR_GROUP_SCHED
6840 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6842 #ifdef CONFIG_RT_GROUP_SCHED
6843 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6845 #ifdef CONFIG_CPUMASK_OFFSTACK
6846 alloc_size
+= num_possible_cpus() * cpumask_size();
6849 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6851 #ifdef CONFIG_FAIR_GROUP_SCHED
6852 root_task_group
.se
= (struct sched_entity
**)ptr
;
6853 ptr
+= nr_cpu_ids
* sizeof(void **);
6855 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6856 ptr
+= nr_cpu_ids
* sizeof(void **);
6858 #endif /* CONFIG_FAIR_GROUP_SCHED */
6859 #ifdef CONFIG_RT_GROUP_SCHED
6860 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6861 ptr
+= nr_cpu_ids
* sizeof(void **);
6863 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6864 ptr
+= nr_cpu_ids
* sizeof(void **);
6866 #endif /* CONFIG_RT_GROUP_SCHED */
6867 #ifdef CONFIG_CPUMASK_OFFSTACK
6868 for_each_possible_cpu(i
) {
6869 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6870 ptr
+= cpumask_size();
6872 #endif /* CONFIG_CPUMASK_OFFSTACK */
6876 init_defrootdomain();
6879 init_rt_bandwidth(&def_rt_bandwidth
,
6880 global_rt_period(), global_rt_runtime());
6882 #ifdef CONFIG_RT_GROUP_SCHED
6883 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6884 global_rt_period(), global_rt_runtime());
6885 #endif /* CONFIG_RT_GROUP_SCHED */
6887 #ifdef CONFIG_CGROUP_SCHED
6888 list_add(&root_task_group
.list
, &task_groups
);
6889 INIT_LIST_HEAD(&root_task_group
.children
);
6890 INIT_LIST_HEAD(&root_task_group
.siblings
);
6891 autogroup_init(&init_task
);
6893 #endif /* CONFIG_CGROUP_SCHED */
6895 for_each_possible_cpu(i
) {
6899 raw_spin_lock_init(&rq
->lock
);
6901 rq
->calc_load_active
= 0;
6902 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6903 init_cfs_rq(&rq
->cfs
);
6904 init_rt_rq(&rq
->rt
, rq
);
6905 #ifdef CONFIG_FAIR_GROUP_SCHED
6906 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6907 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6909 * How much cpu bandwidth does root_task_group get?
6911 * In case of task-groups formed thr' the cgroup filesystem, it
6912 * gets 100% of the cpu resources in the system. This overall
6913 * system cpu resource is divided among the tasks of
6914 * root_task_group and its child task-groups in a fair manner,
6915 * based on each entity's (task or task-group's) weight
6916 * (se->load.weight).
6918 * In other words, if root_task_group has 10 tasks of weight
6919 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6920 * then A0's share of the cpu resource is:
6922 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6924 * We achieve this by letting root_task_group's tasks sit
6925 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6927 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6928 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6929 #endif /* CONFIG_FAIR_GROUP_SCHED */
6931 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6932 #ifdef CONFIG_RT_GROUP_SCHED
6933 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6934 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6937 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6938 rq
->cpu_load
[j
] = 0;
6940 rq
->last_load_update_tick
= jiffies
;
6945 rq
->cpu_power
= SCHED_POWER_SCALE
;
6946 rq
->post_schedule
= 0;
6947 rq
->active_balance
= 0;
6948 rq
->next_balance
= jiffies
;
6953 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6955 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6957 rq_attach_root(rq
, &def_root_domain
);
6963 atomic_set(&rq
->nr_iowait
, 0);
6966 set_load_weight(&init_task
);
6968 #ifdef CONFIG_PREEMPT_NOTIFIERS
6969 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6972 #ifdef CONFIG_RT_MUTEXES
6973 plist_head_init(&init_task
.pi_waiters
);
6977 * The boot idle thread does lazy MMU switching as well:
6979 atomic_inc(&init_mm
.mm_count
);
6980 enter_lazy_tlb(&init_mm
, current
);
6983 * Make us the idle thread. Technically, schedule() should not be
6984 * called from this thread, however somewhere below it might be,
6985 * but because we are the idle thread, we just pick up running again
6986 * when this runqueue becomes "idle".
6988 init_idle(current
, smp_processor_id());
6990 calc_load_update
= jiffies
+ LOAD_FREQ
;
6993 * During early bootup we pretend to be a normal task:
6995 current
->sched_class
= &fair_sched_class
;
6998 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6999 /* May be allocated at isolcpus cmdline parse time */
7000 if (cpu_isolated_map
== NULL
)
7001 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7002 idle_thread_set_boot_cpu();
7004 init_sched_fair_class();
7006 scheduler_running
= 1;
7009 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7010 static inline int preempt_count_equals(int preempt_offset
)
7012 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7014 return (nested
== preempt_offset
);
7017 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7019 static unsigned long prev_jiffy
; /* ratelimiting */
7021 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7022 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7023 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7025 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7027 prev_jiffy
= jiffies
;
7030 "BUG: sleeping function called from invalid context at %s:%d\n",
7033 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7034 in_atomic(), irqs_disabled(),
7035 current
->pid
, current
->comm
);
7037 debug_show_held_locks(current
);
7038 if (irqs_disabled())
7039 print_irqtrace_events(current
);
7042 EXPORT_SYMBOL(__might_sleep
);
7045 #ifdef CONFIG_MAGIC_SYSRQ
7046 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7048 const struct sched_class
*prev_class
= p
->sched_class
;
7049 int old_prio
= p
->prio
;
7054 dequeue_task(rq
, p
, 0);
7055 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7057 enqueue_task(rq
, p
, 0);
7058 resched_task(rq
->curr
);
7061 check_class_changed(rq
, p
, prev_class
, old_prio
);
7064 void normalize_rt_tasks(void)
7066 struct task_struct
*g
, *p
;
7067 unsigned long flags
;
7070 read_lock_irqsave(&tasklist_lock
, flags
);
7071 do_each_thread(g
, p
) {
7073 * Only normalize user tasks:
7078 p
->se
.exec_start
= 0;
7079 #ifdef CONFIG_SCHEDSTATS
7080 p
->se
.statistics
.wait_start
= 0;
7081 p
->se
.statistics
.sleep_start
= 0;
7082 p
->se
.statistics
.block_start
= 0;
7087 * Renice negative nice level userspace
7090 if (TASK_NICE(p
) < 0 && p
->mm
)
7091 set_user_nice(p
, 0);
7095 raw_spin_lock(&p
->pi_lock
);
7096 rq
= __task_rq_lock(p
);
7098 normalize_task(rq
, p
);
7100 __task_rq_unlock(rq
);
7101 raw_spin_unlock(&p
->pi_lock
);
7102 } while_each_thread(g
, p
);
7104 read_unlock_irqrestore(&tasklist_lock
, flags
);
7107 #endif /* CONFIG_MAGIC_SYSRQ */
7109 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7111 * These functions are only useful for the IA64 MCA handling, or kdb.
7113 * They can only be called when the whole system has been
7114 * stopped - every CPU needs to be quiescent, and no scheduling
7115 * activity can take place. Using them for anything else would
7116 * be a serious bug, and as a result, they aren't even visible
7117 * under any other configuration.
7121 * curr_task - return the current task for a given cpu.
7122 * @cpu: the processor in question.
7124 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7126 struct task_struct
*curr_task(int cpu
)
7128 return cpu_curr(cpu
);
7131 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7135 * set_curr_task - set the current task for a given cpu.
7136 * @cpu: the processor in question.
7137 * @p: the task pointer to set.
7139 * Description: This function must only be used when non-maskable interrupts
7140 * are serviced on a separate stack. It allows the architecture to switch the
7141 * notion of the current task on a cpu in a non-blocking manner. This function
7142 * must be called with all CPU's synchronized, and interrupts disabled, the
7143 * and caller must save the original value of the current task (see
7144 * curr_task() above) and restore that value before reenabling interrupts and
7145 * re-starting the system.
7147 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7149 void set_curr_task(int cpu
, struct task_struct
*p
)
7156 #ifdef CONFIG_CGROUP_SCHED
7157 /* task_group_lock serializes the addition/removal of task groups */
7158 static DEFINE_SPINLOCK(task_group_lock
);
7160 static void free_sched_group(struct task_group
*tg
)
7162 free_fair_sched_group(tg
);
7163 free_rt_sched_group(tg
);
7168 /* allocate runqueue etc for a new task group */
7169 struct task_group
*sched_create_group(struct task_group
*parent
)
7171 struct task_group
*tg
;
7173 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7175 return ERR_PTR(-ENOMEM
);
7177 if (!alloc_fair_sched_group(tg
, parent
))
7180 if (!alloc_rt_sched_group(tg
, parent
))
7186 free_sched_group(tg
);
7187 return ERR_PTR(-ENOMEM
);
7190 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7192 unsigned long flags
;
7194 spin_lock_irqsave(&task_group_lock
, flags
);
7195 list_add_rcu(&tg
->list
, &task_groups
);
7197 WARN_ON(!parent
); /* root should already exist */
7199 tg
->parent
= parent
;
7200 INIT_LIST_HEAD(&tg
->children
);
7201 list_add_rcu(&tg
->siblings
, &parent
->children
);
7202 spin_unlock_irqrestore(&task_group_lock
, flags
);
7205 /* rcu callback to free various structures associated with a task group */
7206 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7208 /* now it should be safe to free those cfs_rqs */
7209 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7212 /* Destroy runqueue etc associated with a task group */
7213 void sched_destroy_group(struct task_group
*tg
)
7215 /* wait for possible concurrent references to cfs_rqs complete */
7216 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7219 void sched_offline_group(struct task_group
*tg
)
7221 unsigned long flags
;
7224 /* end participation in shares distribution */
7225 for_each_possible_cpu(i
)
7226 unregister_fair_sched_group(tg
, i
);
7228 spin_lock_irqsave(&task_group_lock
, flags
);
7229 list_del_rcu(&tg
->list
);
7230 list_del_rcu(&tg
->siblings
);
7231 spin_unlock_irqrestore(&task_group_lock
, flags
);
7234 /* change task's runqueue when it moves between groups.
7235 * The caller of this function should have put the task in its new group
7236 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7237 * reflect its new group.
7239 void sched_move_task(struct task_struct
*tsk
)
7241 struct task_group
*tg
;
7243 unsigned long flags
;
7246 rq
= task_rq_lock(tsk
, &flags
);
7248 running
= task_current(rq
, tsk
);
7252 dequeue_task(rq
, tsk
, 0);
7253 if (unlikely(running
))
7254 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7256 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7257 lockdep_is_held(&tsk
->sighand
->siglock
)),
7258 struct task_group
, css
);
7259 tg
= autogroup_task_group(tsk
, tg
);
7260 tsk
->sched_task_group
= tg
;
7262 #ifdef CONFIG_FAIR_GROUP_SCHED
7263 if (tsk
->sched_class
->task_move_group
)
7264 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7267 set_task_rq(tsk
, task_cpu(tsk
));
7269 if (unlikely(running
))
7270 tsk
->sched_class
->set_curr_task(rq
);
7272 enqueue_task(rq
, tsk
, 0);
7274 task_rq_unlock(rq
, tsk
, &flags
);
7276 #endif /* CONFIG_CGROUP_SCHED */
7278 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7279 static unsigned long to_ratio(u64 period
, u64 runtime
)
7281 if (runtime
== RUNTIME_INF
)
7284 return div64_u64(runtime
<< 20, period
);
7288 #ifdef CONFIG_RT_GROUP_SCHED
7290 * Ensure that the real time constraints are schedulable.
7292 static DEFINE_MUTEX(rt_constraints_mutex
);
7294 /* Must be called with tasklist_lock held */
7295 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7297 struct task_struct
*g
, *p
;
7299 do_each_thread(g
, p
) {
7300 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7302 } while_each_thread(g
, p
);
7307 struct rt_schedulable_data
{
7308 struct task_group
*tg
;
7313 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7315 struct rt_schedulable_data
*d
= data
;
7316 struct task_group
*child
;
7317 unsigned long total
, sum
= 0;
7318 u64 period
, runtime
;
7320 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7321 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7324 period
= d
->rt_period
;
7325 runtime
= d
->rt_runtime
;
7329 * Cannot have more runtime than the period.
7331 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7335 * Ensure we don't starve existing RT tasks.
7337 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7340 total
= to_ratio(period
, runtime
);
7343 * Nobody can have more than the global setting allows.
7345 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7349 * The sum of our children's runtime should not exceed our own.
7351 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7352 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7353 runtime
= child
->rt_bandwidth
.rt_runtime
;
7355 if (child
== d
->tg
) {
7356 period
= d
->rt_period
;
7357 runtime
= d
->rt_runtime
;
7360 sum
+= to_ratio(period
, runtime
);
7369 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7373 struct rt_schedulable_data data
= {
7375 .rt_period
= period
,
7376 .rt_runtime
= runtime
,
7380 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7386 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7387 u64 rt_period
, u64 rt_runtime
)
7391 mutex_lock(&rt_constraints_mutex
);
7392 read_lock(&tasklist_lock
);
7393 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7397 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7398 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7399 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7401 for_each_possible_cpu(i
) {
7402 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7404 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7405 rt_rq
->rt_runtime
= rt_runtime
;
7406 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7408 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7410 read_unlock(&tasklist_lock
);
7411 mutex_unlock(&rt_constraints_mutex
);
7416 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7418 u64 rt_runtime
, rt_period
;
7420 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7421 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7422 if (rt_runtime_us
< 0)
7423 rt_runtime
= RUNTIME_INF
;
7425 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7428 static long sched_group_rt_runtime(struct task_group
*tg
)
7432 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7435 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7436 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7437 return rt_runtime_us
;
7440 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7442 u64 rt_runtime
, rt_period
;
7444 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7445 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7450 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7453 static long sched_group_rt_period(struct task_group
*tg
)
7457 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7458 do_div(rt_period_us
, NSEC_PER_USEC
);
7459 return rt_period_us
;
7462 static int sched_rt_global_constraints(void)
7464 u64 runtime
, period
;
7467 if (sysctl_sched_rt_period
<= 0)
7470 runtime
= global_rt_runtime();
7471 period
= global_rt_period();
7474 * Sanity check on the sysctl variables.
7476 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7479 mutex_lock(&rt_constraints_mutex
);
7480 read_lock(&tasklist_lock
);
7481 ret
= __rt_schedulable(NULL
, 0, 0);
7482 read_unlock(&tasklist_lock
);
7483 mutex_unlock(&rt_constraints_mutex
);
7488 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7490 /* Don't accept realtime tasks when there is no way for them to run */
7491 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7497 #else /* !CONFIG_RT_GROUP_SCHED */
7498 static int sched_rt_global_constraints(void)
7500 unsigned long flags
;
7503 if (sysctl_sched_rt_period
<= 0)
7507 * There's always some RT tasks in the root group
7508 * -- migration, kstopmachine etc..
7510 if (sysctl_sched_rt_runtime
== 0)
7513 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7514 for_each_possible_cpu(i
) {
7515 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7517 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7518 rt_rq
->rt_runtime
= global_rt_runtime();
7519 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7521 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7525 #endif /* CONFIG_RT_GROUP_SCHED */
7527 int sched_rr_handler(struct ctl_table
*table
, int write
,
7528 void __user
*buffer
, size_t *lenp
,
7532 static DEFINE_MUTEX(mutex
);
7535 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7536 /* make sure that internally we keep jiffies */
7537 /* also, writing zero resets timeslice to default */
7538 if (!ret
&& write
) {
7539 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7540 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7542 mutex_unlock(&mutex
);
7546 int sched_rt_handler(struct ctl_table
*table
, int write
,
7547 void __user
*buffer
, size_t *lenp
,
7551 int old_period
, old_runtime
;
7552 static DEFINE_MUTEX(mutex
);
7555 old_period
= sysctl_sched_rt_period
;
7556 old_runtime
= sysctl_sched_rt_runtime
;
7558 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7560 if (!ret
&& write
) {
7561 ret
= sched_rt_global_constraints();
7563 sysctl_sched_rt_period
= old_period
;
7564 sysctl_sched_rt_runtime
= old_runtime
;
7566 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7567 def_rt_bandwidth
.rt_period
=
7568 ns_to_ktime(global_rt_period());
7571 mutex_unlock(&mutex
);
7576 #ifdef CONFIG_CGROUP_SCHED
7578 /* return corresponding task_group object of a cgroup */
7579 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7581 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7582 struct task_group
, css
);
7585 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7587 struct task_group
*tg
, *parent
;
7589 if (!cgrp
->parent
) {
7590 /* This is early initialization for the top cgroup */
7591 return &root_task_group
.css
;
7594 parent
= cgroup_tg(cgrp
->parent
);
7595 tg
= sched_create_group(parent
);
7597 return ERR_PTR(-ENOMEM
);
7602 static int cpu_cgroup_css_online(struct cgroup
*cgrp
)
7604 struct task_group
*tg
= cgroup_tg(cgrp
);
7605 struct task_group
*parent
;
7610 parent
= cgroup_tg(cgrp
->parent
);
7611 sched_online_group(tg
, parent
);
7615 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7617 struct task_group
*tg
= cgroup_tg(cgrp
);
7619 sched_destroy_group(tg
);
7622 static void cpu_cgroup_css_offline(struct cgroup
*cgrp
)
7624 struct task_group
*tg
= cgroup_tg(cgrp
);
7626 sched_offline_group(tg
);
7629 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7630 struct cgroup_taskset
*tset
)
7632 struct task_struct
*task
;
7634 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7635 #ifdef CONFIG_RT_GROUP_SCHED
7636 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7639 /* We don't support RT-tasks being in separate groups */
7640 if (task
->sched_class
!= &fair_sched_class
)
7647 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7648 struct cgroup_taskset
*tset
)
7650 struct task_struct
*task
;
7652 cgroup_taskset_for_each(task
, cgrp
, tset
)
7653 sched_move_task(task
);
7657 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7658 struct task_struct
*task
)
7661 * cgroup_exit() is called in the copy_process() failure path.
7662 * Ignore this case since the task hasn't ran yet, this avoids
7663 * trying to poke a half freed task state from generic code.
7665 if (!(task
->flags
& PF_EXITING
))
7668 sched_move_task(task
);
7671 #ifdef CONFIG_FAIR_GROUP_SCHED
7672 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7675 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7678 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7680 struct task_group
*tg
= cgroup_tg(cgrp
);
7682 return (u64
) scale_load_down(tg
->shares
);
7685 #ifdef CONFIG_CFS_BANDWIDTH
7686 static DEFINE_MUTEX(cfs_constraints_mutex
);
7688 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7689 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7691 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7693 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7695 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7696 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7698 if (tg
== &root_task_group
)
7702 * Ensure we have at some amount of bandwidth every period. This is
7703 * to prevent reaching a state of large arrears when throttled via
7704 * entity_tick() resulting in prolonged exit starvation.
7706 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7710 * Likewise, bound things on the otherside by preventing insane quota
7711 * periods. This also allows us to normalize in computing quota
7714 if (period
> max_cfs_quota_period
)
7717 mutex_lock(&cfs_constraints_mutex
);
7718 ret
= __cfs_schedulable(tg
, period
, quota
);
7722 runtime_enabled
= quota
!= RUNTIME_INF
;
7723 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7724 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7725 raw_spin_lock_irq(&cfs_b
->lock
);
7726 cfs_b
->period
= ns_to_ktime(period
);
7727 cfs_b
->quota
= quota
;
7729 __refill_cfs_bandwidth_runtime(cfs_b
);
7730 /* restart the period timer (if active) to handle new period expiry */
7731 if (runtime_enabled
&& cfs_b
->timer_active
) {
7732 /* force a reprogram */
7733 cfs_b
->timer_active
= 0;
7734 __start_cfs_bandwidth(cfs_b
);
7736 raw_spin_unlock_irq(&cfs_b
->lock
);
7738 for_each_possible_cpu(i
) {
7739 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7740 struct rq
*rq
= cfs_rq
->rq
;
7742 raw_spin_lock_irq(&rq
->lock
);
7743 cfs_rq
->runtime_enabled
= runtime_enabled
;
7744 cfs_rq
->runtime_remaining
= 0;
7746 if (cfs_rq
->throttled
)
7747 unthrottle_cfs_rq(cfs_rq
);
7748 raw_spin_unlock_irq(&rq
->lock
);
7751 mutex_unlock(&cfs_constraints_mutex
);
7756 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7760 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7761 if (cfs_quota_us
< 0)
7762 quota
= RUNTIME_INF
;
7764 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7766 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7769 long tg_get_cfs_quota(struct task_group
*tg
)
7773 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7776 quota_us
= tg
->cfs_bandwidth
.quota
;
7777 do_div(quota_us
, NSEC_PER_USEC
);
7782 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7786 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7787 quota
= tg
->cfs_bandwidth
.quota
;
7789 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7792 long tg_get_cfs_period(struct task_group
*tg
)
7796 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7797 do_div(cfs_period_us
, NSEC_PER_USEC
);
7799 return cfs_period_us
;
7802 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7804 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7807 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7810 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7813 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7815 return tg_get_cfs_period(cgroup_tg(cgrp
));
7818 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7821 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7824 struct cfs_schedulable_data
{
7825 struct task_group
*tg
;
7830 * normalize group quota/period to be quota/max_period
7831 * note: units are usecs
7833 static u64
normalize_cfs_quota(struct task_group
*tg
,
7834 struct cfs_schedulable_data
*d
)
7842 period
= tg_get_cfs_period(tg
);
7843 quota
= tg_get_cfs_quota(tg
);
7846 /* note: these should typically be equivalent */
7847 if (quota
== RUNTIME_INF
|| quota
== -1)
7850 return to_ratio(period
, quota
);
7853 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7855 struct cfs_schedulable_data
*d
= data
;
7856 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7857 s64 quota
= 0, parent_quota
= -1;
7860 quota
= RUNTIME_INF
;
7862 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7864 quota
= normalize_cfs_quota(tg
, d
);
7865 parent_quota
= parent_b
->hierarchal_quota
;
7868 * ensure max(child_quota) <= parent_quota, inherit when no
7871 if (quota
== RUNTIME_INF
)
7872 quota
= parent_quota
;
7873 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7876 cfs_b
->hierarchal_quota
= quota
;
7881 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7884 struct cfs_schedulable_data data
= {
7890 if (quota
!= RUNTIME_INF
) {
7891 do_div(data
.period
, NSEC_PER_USEC
);
7892 do_div(data
.quota
, NSEC_PER_USEC
);
7896 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7902 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7903 struct cgroup_map_cb
*cb
)
7905 struct task_group
*tg
= cgroup_tg(cgrp
);
7906 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7908 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7909 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7910 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7914 #endif /* CONFIG_CFS_BANDWIDTH */
7915 #endif /* CONFIG_FAIR_GROUP_SCHED */
7917 #ifdef CONFIG_RT_GROUP_SCHED
7918 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7921 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7924 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7926 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7929 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7932 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7935 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7937 return sched_group_rt_period(cgroup_tg(cgrp
));
7939 #endif /* CONFIG_RT_GROUP_SCHED */
7941 static struct cftype cpu_files
[] = {
7942 #ifdef CONFIG_FAIR_GROUP_SCHED
7945 .read_u64
= cpu_shares_read_u64
,
7946 .write_u64
= cpu_shares_write_u64
,
7949 #ifdef CONFIG_CFS_BANDWIDTH
7951 .name
= "cfs_quota_us",
7952 .read_s64
= cpu_cfs_quota_read_s64
,
7953 .write_s64
= cpu_cfs_quota_write_s64
,
7956 .name
= "cfs_period_us",
7957 .read_u64
= cpu_cfs_period_read_u64
,
7958 .write_u64
= cpu_cfs_period_write_u64
,
7962 .read_map
= cpu_stats_show
,
7965 #ifdef CONFIG_RT_GROUP_SCHED
7967 .name
= "rt_runtime_us",
7968 .read_s64
= cpu_rt_runtime_read
,
7969 .write_s64
= cpu_rt_runtime_write
,
7972 .name
= "rt_period_us",
7973 .read_u64
= cpu_rt_period_read_uint
,
7974 .write_u64
= cpu_rt_period_write_uint
,
7980 struct cgroup_subsys cpu_cgroup_subsys
= {
7982 .css_alloc
= cpu_cgroup_css_alloc
,
7983 .css_free
= cpu_cgroup_css_free
,
7984 .css_online
= cpu_cgroup_css_online
,
7985 .css_offline
= cpu_cgroup_css_offline
,
7986 .can_attach
= cpu_cgroup_can_attach
,
7987 .attach
= cpu_cgroup_attach
,
7988 .exit
= cpu_cgroup_exit
,
7989 .subsys_id
= cpu_cgroup_subsys_id
,
7990 .base_cftypes
= cpu_files
,
7994 #endif /* CONFIG_CGROUP_SCHED */
7996 void dump_cpu_task(int cpu
)
7998 pr_info("Task dump for CPU %d:\n", cpu
);
7999 sched_show_task(cpu_curr(cpu
));