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
);
552 #ifdef CONFIG_NO_HZ_COMMON
554 * In the semi idle case, use the nearest busy cpu for migrating timers
555 * from an idle cpu. This is good for power-savings.
557 * We don't do similar optimization for completely idle system, as
558 * selecting an idle cpu will add more delays to the timers than intended
559 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
561 int get_nohz_timer_target(void)
563 int cpu
= smp_processor_id();
565 struct sched_domain
*sd
;
568 for_each_domain(cpu
, sd
) {
569 for_each_cpu(i
, sched_domain_span(sd
)) {
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 static void wake_up_idle_cpu(int cpu
)
592 struct rq
*rq
= cpu_rq(cpu
);
594 if (cpu
== smp_processor_id())
598 * This is safe, as this function is called with the timer
599 * wheel base lock of (cpu) held. When the CPU is on the way
600 * to idle and has not yet set rq->curr to idle then it will
601 * be serialized on the timer wheel base lock and take the new
602 * timer into account automatically.
604 if (rq
->curr
!= rq
->idle
)
608 * We can set TIF_RESCHED on the idle task of the other CPU
609 * lockless. The worst case is that the other CPU runs the
610 * idle task through an additional NOOP schedule()
612 set_tsk_need_resched(rq
->idle
);
614 /* NEED_RESCHED must be visible before we test polling */
616 if (!tsk_is_polling(rq
->idle
))
617 smp_send_reschedule(cpu
);
620 static bool wake_up_full_nohz_cpu(int cpu
)
622 if (tick_nohz_full_cpu(cpu
)) {
623 if (cpu
!= smp_processor_id() ||
624 tick_nohz_tick_stopped())
625 smp_send_reschedule(cpu
);
632 void wake_up_nohz_cpu(int cpu
)
634 if (!wake_up_full_nohz_cpu(cpu
))
635 wake_up_idle_cpu(cpu
);
638 static inline bool got_nohz_idle_kick(void)
640 int cpu
= smp_processor_id();
641 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
644 #else /* CONFIG_NO_HZ_COMMON */
646 static inline bool got_nohz_idle_kick(void)
651 #endif /* CONFIG_NO_HZ_COMMON */
653 #ifdef CONFIG_NO_HZ_FULL
654 bool sched_can_stop_tick(void)
660 /* Make sure rq->nr_running update is visible after the IPI */
663 /* More than one running task need preemption */
664 if (rq
->nr_running
> 1)
669 #endif /* CONFIG_NO_HZ_FULL */
671 void sched_avg_update(struct rq
*rq
)
673 s64 period
= sched_avg_period();
675 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
677 * Inline assembly required to prevent the compiler
678 * optimising this loop into a divmod call.
679 * See __iter_div_u64_rem() for another example of this.
681 asm("" : "+rm" (rq
->age_stamp
));
682 rq
->age_stamp
+= period
;
687 #else /* !CONFIG_SMP */
688 void resched_task(struct task_struct
*p
)
690 assert_raw_spin_locked(&task_rq(p
)->lock
);
691 set_tsk_need_resched(p
);
693 #endif /* CONFIG_SMP */
695 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
696 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
698 * Iterate task_group tree rooted at *from, calling @down when first entering a
699 * node and @up when leaving it for the final time.
701 * Caller must hold rcu_lock or sufficient equivalent.
703 int walk_tg_tree_from(struct task_group
*from
,
704 tg_visitor down
, tg_visitor up
, void *data
)
706 struct task_group
*parent
, *child
;
712 ret
= (*down
)(parent
, data
);
715 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
722 ret
= (*up
)(parent
, data
);
723 if (ret
|| parent
== from
)
727 parent
= parent
->parent
;
734 int tg_nop(struct task_group
*tg
, void *data
)
740 static void set_load_weight(struct task_struct
*p
)
742 int prio
= p
->static_prio
- MAX_RT_PRIO
;
743 struct load_weight
*load
= &p
->se
.load
;
746 * SCHED_IDLE tasks get minimal weight:
748 if (p
->policy
== SCHED_IDLE
) {
749 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
750 load
->inv_weight
= WMULT_IDLEPRIO
;
754 load
->weight
= scale_load(prio_to_weight
[prio
]);
755 load
->inv_weight
= prio_to_wmult
[prio
];
758 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
761 sched_info_queued(p
);
762 p
->sched_class
->enqueue_task(rq
, p
, flags
);
765 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
768 sched_info_dequeued(p
);
769 p
->sched_class
->dequeue_task(rq
, p
, flags
);
772 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
774 if (task_contributes_to_load(p
))
775 rq
->nr_uninterruptible
--;
777 enqueue_task(rq
, p
, flags
);
780 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
782 if (task_contributes_to_load(p
))
783 rq
->nr_uninterruptible
++;
785 dequeue_task(rq
, p
, flags
);
788 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
791 * In theory, the compile should just see 0 here, and optimize out the call
792 * to sched_rt_avg_update. But I don't trust it...
794 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
795 s64 steal
= 0, irq_delta
= 0;
797 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
798 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
801 * Since irq_time is only updated on {soft,}irq_exit, we might run into
802 * this case when a previous update_rq_clock() happened inside a
805 * When this happens, we stop ->clock_task and only update the
806 * prev_irq_time stamp to account for the part that fit, so that a next
807 * update will consume the rest. This ensures ->clock_task is
810 * It does however cause some slight miss-attribution of {soft,}irq
811 * time, a more accurate solution would be to update the irq_time using
812 * the current rq->clock timestamp, except that would require using
815 if (irq_delta
> delta
)
818 rq
->prev_irq_time
+= irq_delta
;
821 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
822 if (static_key_false((¶virt_steal_rq_enabled
))) {
825 steal
= paravirt_steal_clock(cpu_of(rq
));
826 steal
-= rq
->prev_steal_time_rq
;
828 if (unlikely(steal
> delta
))
831 st
= steal_ticks(steal
);
832 steal
= st
* TICK_NSEC
;
834 rq
->prev_steal_time_rq
+= steal
;
840 rq
->clock_task
+= delta
;
842 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
843 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
844 sched_rt_avg_update(rq
, irq_delta
+ steal
);
848 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
850 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
851 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
855 * Make it appear like a SCHED_FIFO task, its something
856 * userspace knows about and won't get confused about.
858 * Also, it will make PI more or less work without too
859 * much confusion -- but then, stop work should not
860 * rely on PI working anyway.
862 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
864 stop
->sched_class
= &stop_sched_class
;
867 cpu_rq(cpu
)->stop
= stop
;
871 * Reset it back to a normal scheduling class so that
872 * it can die in pieces.
874 old_stop
->sched_class
= &rt_sched_class
;
879 * __normal_prio - return the priority that is based on the static prio
881 static inline int __normal_prio(struct task_struct
*p
)
883 return p
->static_prio
;
887 * Calculate the expected normal priority: i.e. priority
888 * without taking RT-inheritance into account. Might be
889 * boosted by interactivity modifiers. Changes upon fork,
890 * setprio syscalls, and whenever the interactivity
891 * estimator recalculates.
893 static inline int normal_prio(struct task_struct
*p
)
897 if (task_has_rt_policy(p
))
898 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
900 prio
= __normal_prio(p
);
905 * Calculate the current priority, i.e. the priority
906 * taken into account by the scheduler. This value might
907 * be boosted by RT tasks, or might be boosted by
908 * interactivity modifiers. Will be RT if the task got
909 * RT-boosted. If not then it returns p->normal_prio.
911 static int effective_prio(struct task_struct
*p
)
913 p
->normal_prio
= normal_prio(p
);
915 * If we are RT tasks or we were boosted to RT priority,
916 * keep the priority unchanged. Otherwise, update priority
917 * to the normal priority:
919 if (!rt_prio(p
->prio
))
920 return p
->normal_prio
;
925 * task_curr - is this task currently executing on a CPU?
926 * @p: the task in question.
928 inline int task_curr(const struct task_struct
*p
)
930 return cpu_curr(task_cpu(p
)) == p
;
933 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
934 const struct sched_class
*prev_class
,
937 if (prev_class
!= p
->sched_class
) {
938 if (prev_class
->switched_from
)
939 prev_class
->switched_from(rq
, p
);
940 p
->sched_class
->switched_to(rq
, p
);
941 } else if (oldprio
!= p
->prio
)
942 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
945 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
947 const struct sched_class
*class;
949 if (p
->sched_class
== rq
->curr
->sched_class
) {
950 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
952 for_each_class(class) {
953 if (class == rq
->curr
->sched_class
)
955 if (class == p
->sched_class
) {
956 resched_task(rq
->curr
);
963 * A queue event has occurred, and we're going to schedule. In
964 * this case, we can save a useless back to back clock update.
966 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
967 rq
->skip_clock_update
= 1;
970 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
972 void register_task_migration_notifier(struct notifier_block
*n
)
974 atomic_notifier_chain_register(&task_migration_notifier
, n
);
978 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
986 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1000 lockdep_is_held(&task_rq(p
)->lock
)));
1004 trace_sched_migrate_task(p
, new_cpu
);
1006 if (task_cpu(p
) != new_cpu
) {
1007 struct task_migration_notifier tmn
;
1009 if (p
->sched_class
->migrate_task_rq
)
1010 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1011 p
->se
.nr_migrations
++;
1012 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1015 tmn
.from_cpu
= task_cpu(p
);
1016 tmn
.to_cpu
= new_cpu
;
1018 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
1021 __set_task_cpu(p
, new_cpu
);
1024 struct migration_arg
{
1025 struct task_struct
*task
;
1029 static int migration_cpu_stop(void *data
);
1032 * wait_task_inactive - wait for a thread to unschedule.
1034 * If @match_state is nonzero, it's the @p->state value just checked and
1035 * not expected to change. If it changes, i.e. @p might have woken up,
1036 * then return zero. When we succeed in waiting for @p to be off its CPU,
1037 * we return a positive number (its total switch count). If a second call
1038 * a short while later returns the same number, the caller can be sure that
1039 * @p has remained unscheduled the whole time.
1041 * The caller must ensure that the task *will* unschedule sometime soon,
1042 * else this function might spin for a *long* time. This function can't
1043 * be called with interrupts off, or it may introduce deadlock with
1044 * smp_call_function() if an IPI is sent by the same process we are
1045 * waiting to become inactive.
1047 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1049 unsigned long flags
;
1056 * We do the initial early heuristics without holding
1057 * any task-queue locks at all. We'll only try to get
1058 * the runqueue lock when things look like they will
1064 * If the task is actively running on another CPU
1065 * still, just relax and busy-wait without holding
1068 * NOTE! Since we don't hold any locks, it's not
1069 * even sure that "rq" stays as the right runqueue!
1070 * But we don't care, since "task_running()" will
1071 * return false if the runqueue has changed and p
1072 * is actually now running somewhere else!
1074 while (task_running(rq
, p
)) {
1075 if (match_state
&& unlikely(p
->state
!= match_state
))
1081 * Ok, time to look more closely! We need the rq
1082 * lock now, to be *sure*. If we're wrong, we'll
1083 * just go back and repeat.
1085 rq
= task_rq_lock(p
, &flags
);
1086 trace_sched_wait_task(p
);
1087 running
= task_running(rq
, p
);
1090 if (!match_state
|| p
->state
== match_state
)
1091 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1092 task_rq_unlock(rq
, p
, &flags
);
1095 * If it changed from the expected state, bail out now.
1097 if (unlikely(!ncsw
))
1101 * Was it really running after all now that we
1102 * checked with the proper locks actually held?
1104 * Oops. Go back and try again..
1106 if (unlikely(running
)) {
1112 * It's not enough that it's not actively running,
1113 * it must be off the runqueue _entirely_, and not
1116 * So if it was still runnable (but just not actively
1117 * running right now), it's preempted, and we should
1118 * yield - it could be a while.
1120 if (unlikely(on_rq
)) {
1121 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1123 set_current_state(TASK_UNINTERRUPTIBLE
);
1124 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1129 * Ahh, all good. It wasn't running, and it wasn't
1130 * runnable, which means that it will never become
1131 * running in the future either. We're all done!
1140 * kick_process - kick a running thread to enter/exit the kernel
1141 * @p: the to-be-kicked thread
1143 * Cause a process which is running on another CPU to enter
1144 * kernel-mode, without any delay. (to get signals handled.)
1146 * NOTE: this function doesn't have to take the runqueue lock,
1147 * because all it wants to ensure is that the remote task enters
1148 * the kernel. If the IPI races and the task has been migrated
1149 * to another CPU then no harm is done and the purpose has been
1152 void kick_process(struct task_struct
*p
)
1158 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1159 smp_send_reschedule(cpu
);
1162 EXPORT_SYMBOL_GPL(kick_process
);
1163 #endif /* CONFIG_SMP */
1167 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1169 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1171 int nid
= cpu_to_node(cpu
);
1172 const struct cpumask
*nodemask
= NULL
;
1173 enum { cpuset
, possible
, fail
} state
= cpuset
;
1177 * If the node that the cpu is on has been offlined, cpu_to_node()
1178 * will return -1. There is no cpu on the node, and we should
1179 * select the cpu on the other node.
1182 nodemask
= cpumask_of_node(nid
);
1184 /* Look for allowed, online CPU in same node. */
1185 for_each_cpu(dest_cpu
, nodemask
) {
1186 if (!cpu_online(dest_cpu
))
1188 if (!cpu_active(dest_cpu
))
1190 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1196 /* Any allowed, online CPU? */
1197 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1198 if (!cpu_online(dest_cpu
))
1200 if (!cpu_active(dest_cpu
))
1207 /* No more Mr. Nice Guy. */
1208 cpuset_cpus_allowed_fallback(p
);
1213 do_set_cpus_allowed(p
, cpu_possible_mask
);
1224 if (state
!= cpuset
) {
1226 * Don't tell them about moving exiting tasks or
1227 * kernel threads (both mm NULL), since they never
1230 if (p
->mm
&& printk_ratelimit()) {
1231 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1232 task_pid_nr(p
), p
->comm
, cpu
);
1240 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1243 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1245 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1248 * In order not to call set_task_cpu() on a blocking task we need
1249 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1252 * Since this is common to all placement strategies, this lives here.
1254 * [ this allows ->select_task() to simply return task_cpu(p) and
1255 * not worry about this generic constraint ]
1257 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1259 cpu
= select_fallback_rq(task_cpu(p
), p
);
1264 static void update_avg(u64
*avg
, u64 sample
)
1266 s64 diff
= sample
- *avg
;
1272 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1274 #ifdef CONFIG_SCHEDSTATS
1275 struct rq
*rq
= this_rq();
1278 int this_cpu
= smp_processor_id();
1280 if (cpu
== this_cpu
) {
1281 schedstat_inc(rq
, ttwu_local
);
1282 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1284 struct sched_domain
*sd
;
1286 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1288 for_each_domain(this_cpu
, sd
) {
1289 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1290 schedstat_inc(sd
, ttwu_wake_remote
);
1297 if (wake_flags
& WF_MIGRATED
)
1298 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1300 #endif /* CONFIG_SMP */
1302 schedstat_inc(rq
, ttwu_count
);
1303 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1305 if (wake_flags
& WF_SYNC
)
1306 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1308 #endif /* CONFIG_SCHEDSTATS */
1311 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1313 activate_task(rq
, p
, en_flags
);
1316 /* if a worker is waking up, notify workqueue */
1317 if (p
->flags
& PF_WQ_WORKER
)
1318 wq_worker_waking_up(p
, cpu_of(rq
));
1322 * Mark the task runnable and perform wakeup-preemption.
1325 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1327 check_preempt_curr(rq
, p
, wake_flags
);
1328 trace_sched_wakeup(p
, true);
1330 p
->state
= TASK_RUNNING
;
1332 if (p
->sched_class
->task_woken
)
1333 p
->sched_class
->task_woken(rq
, p
);
1335 if (rq
->idle_stamp
) {
1336 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1337 u64 max
= 2*sysctl_sched_migration_cost
;
1342 update_avg(&rq
->avg_idle
, delta
);
1349 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1352 if (p
->sched_contributes_to_load
)
1353 rq
->nr_uninterruptible
--;
1356 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1357 ttwu_do_wakeup(rq
, p
, wake_flags
);
1361 * Called in case the task @p isn't fully descheduled from its runqueue,
1362 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1363 * since all we need to do is flip p->state to TASK_RUNNING, since
1364 * the task is still ->on_rq.
1366 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1371 rq
= __task_rq_lock(p
);
1373 ttwu_do_wakeup(rq
, p
, wake_flags
);
1376 __task_rq_unlock(rq
);
1382 static void sched_ttwu_pending(void)
1384 struct rq
*rq
= this_rq();
1385 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1386 struct task_struct
*p
;
1388 raw_spin_lock(&rq
->lock
);
1391 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1392 llist
= llist_next(llist
);
1393 ttwu_do_activate(rq
, p
, 0);
1396 raw_spin_unlock(&rq
->lock
);
1399 void scheduler_ipi(void)
1401 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1405 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1406 * traditionally all their work was done from the interrupt return
1407 * path. Now that we actually do some work, we need to make sure
1410 * Some archs already do call them, luckily irq_enter/exit nest
1413 * Arguably we should visit all archs and update all handlers,
1414 * however a fair share of IPIs are still resched only so this would
1415 * somewhat pessimize the simple resched case.
1418 sched_ttwu_pending();
1421 * Check if someone kicked us for doing the nohz idle load balance.
1423 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1424 this_rq()->idle_balance
= 1;
1425 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1430 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1432 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1433 smp_send_reschedule(cpu
);
1436 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1438 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1440 #endif /* CONFIG_SMP */
1442 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1444 struct rq
*rq
= cpu_rq(cpu
);
1446 #if defined(CONFIG_SMP)
1447 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1448 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1449 ttwu_queue_remote(p
, cpu
);
1454 raw_spin_lock(&rq
->lock
);
1455 ttwu_do_activate(rq
, p
, 0);
1456 raw_spin_unlock(&rq
->lock
);
1460 * try_to_wake_up - wake up a thread
1461 * @p: the thread to be awakened
1462 * @state: the mask of task states that can be woken
1463 * @wake_flags: wake modifier flags (WF_*)
1465 * Put it on the run-queue if it's not already there. The "current"
1466 * thread is always on the run-queue (except when the actual
1467 * re-schedule is in progress), and as such you're allowed to do
1468 * the simpler "current->state = TASK_RUNNING" to mark yourself
1469 * runnable without the overhead of this.
1471 * Returns %true if @p was woken up, %false if it was already running
1472 * or @state didn't match @p's state.
1475 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1477 unsigned long flags
;
1478 int cpu
, success
= 0;
1481 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1482 if (!(p
->state
& state
))
1485 success
= 1; /* we're going to change ->state */
1488 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1493 * If the owning (remote) cpu is still in the middle of schedule() with
1494 * this task as prev, wait until its done referencing the task.
1499 * Pairs with the smp_wmb() in finish_lock_switch().
1503 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1504 p
->state
= TASK_WAKING
;
1506 if (p
->sched_class
->task_waking
)
1507 p
->sched_class
->task_waking(p
);
1509 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1510 if (task_cpu(p
) != cpu
) {
1511 wake_flags
|= WF_MIGRATED
;
1512 set_task_cpu(p
, cpu
);
1514 #endif /* CONFIG_SMP */
1518 ttwu_stat(p
, cpu
, wake_flags
);
1520 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1526 * try_to_wake_up_local - try to wake up a local task with rq lock held
1527 * @p: the thread to be awakened
1529 * Put @p on the run-queue if it's not already there. The caller must
1530 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1533 static void try_to_wake_up_local(struct task_struct
*p
)
1535 struct rq
*rq
= task_rq(p
);
1537 BUG_ON(rq
!= this_rq());
1538 BUG_ON(p
== current
);
1539 lockdep_assert_held(&rq
->lock
);
1541 if (!raw_spin_trylock(&p
->pi_lock
)) {
1542 raw_spin_unlock(&rq
->lock
);
1543 raw_spin_lock(&p
->pi_lock
);
1544 raw_spin_lock(&rq
->lock
);
1547 if (!(p
->state
& TASK_NORMAL
))
1551 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1553 ttwu_do_wakeup(rq
, p
, 0);
1554 ttwu_stat(p
, smp_processor_id(), 0);
1556 raw_spin_unlock(&p
->pi_lock
);
1560 * wake_up_process - Wake up a specific process
1561 * @p: The process to be woken up.
1563 * Attempt to wake up the nominated process and move it to the set of runnable
1564 * processes. Returns 1 if the process was woken up, 0 if it was already
1567 * It may be assumed that this function implies a write memory barrier before
1568 * changing the task state if and only if any tasks are woken up.
1570 int wake_up_process(struct task_struct
*p
)
1572 WARN_ON(task_is_stopped_or_traced(p
));
1573 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1575 EXPORT_SYMBOL(wake_up_process
);
1577 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1579 return try_to_wake_up(p
, state
, 0);
1583 * Perform scheduler related setup for a newly forked process p.
1584 * p is forked by current.
1586 * __sched_fork() is basic setup used by init_idle() too:
1588 static void __sched_fork(struct task_struct
*p
)
1593 p
->se
.exec_start
= 0;
1594 p
->se
.sum_exec_runtime
= 0;
1595 p
->se
.prev_sum_exec_runtime
= 0;
1596 p
->se
.nr_migrations
= 0;
1598 INIT_LIST_HEAD(&p
->se
.group_node
);
1601 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1602 * removed when useful for applications beyond shares distribution (e.g.
1605 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1606 p
->se
.avg
.runnable_avg_period
= 0;
1607 p
->se
.avg
.runnable_avg_sum
= 0;
1609 #ifdef CONFIG_SCHEDSTATS
1610 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1613 INIT_LIST_HEAD(&p
->rt
.run_list
);
1615 #ifdef CONFIG_PREEMPT_NOTIFIERS
1616 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1619 #ifdef CONFIG_NUMA_BALANCING
1620 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1621 p
->mm
->numa_next_scan
= jiffies
;
1622 p
->mm
->numa_next_reset
= jiffies
;
1623 p
->mm
->numa_scan_seq
= 0;
1626 p
->node_stamp
= 0ULL;
1627 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1628 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1629 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1630 p
->numa_work
.next
= &p
->numa_work
;
1631 #endif /* CONFIG_NUMA_BALANCING */
1634 #ifdef CONFIG_NUMA_BALANCING
1635 #ifdef CONFIG_SCHED_DEBUG
1636 void set_numabalancing_state(bool enabled
)
1639 sched_feat_set("NUMA");
1641 sched_feat_set("NO_NUMA");
1644 __read_mostly
bool numabalancing_enabled
;
1646 void set_numabalancing_state(bool enabled
)
1648 numabalancing_enabled
= enabled
;
1650 #endif /* CONFIG_SCHED_DEBUG */
1651 #endif /* CONFIG_NUMA_BALANCING */
1654 * fork()/clone()-time setup:
1656 void sched_fork(struct task_struct
*p
)
1658 unsigned long flags
;
1659 int cpu
= get_cpu();
1663 * We mark the process as running here. This guarantees that
1664 * nobody will actually run it, and a signal or other external
1665 * event cannot wake it up and insert it on the runqueue either.
1667 p
->state
= TASK_RUNNING
;
1670 * Make sure we do not leak PI boosting priority to the child.
1672 p
->prio
= current
->normal_prio
;
1675 * Revert to default priority/policy on fork if requested.
1677 if (unlikely(p
->sched_reset_on_fork
)) {
1678 if (task_has_rt_policy(p
)) {
1679 p
->policy
= SCHED_NORMAL
;
1680 p
->static_prio
= NICE_TO_PRIO(0);
1682 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1683 p
->static_prio
= NICE_TO_PRIO(0);
1685 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1689 * We don't need the reset flag anymore after the fork. It has
1690 * fulfilled its duty:
1692 p
->sched_reset_on_fork
= 0;
1695 if (!rt_prio(p
->prio
))
1696 p
->sched_class
= &fair_sched_class
;
1698 if (p
->sched_class
->task_fork
)
1699 p
->sched_class
->task_fork(p
);
1702 * The child is not yet in the pid-hash so no cgroup attach races,
1703 * and the cgroup is pinned to this child due to cgroup_fork()
1704 * is ran before sched_fork().
1706 * Silence PROVE_RCU.
1708 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1709 set_task_cpu(p
, cpu
);
1710 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1712 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1713 if (likely(sched_info_on()))
1714 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1716 #if defined(CONFIG_SMP)
1719 #ifdef CONFIG_PREEMPT_COUNT
1720 /* Want to start with kernel preemption disabled. */
1721 task_thread_info(p
)->preempt_count
= 1;
1724 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1731 * wake_up_new_task - wake up a newly created task for the first time.
1733 * This function will do some initial scheduler statistics housekeeping
1734 * that must be done for every newly created context, then puts the task
1735 * on the runqueue and wakes it.
1737 void wake_up_new_task(struct task_struct
*p
)
1739 unsigned long flags
;
1742 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1745 * Fork balancing, do it here and not earlier because:
1746 * - cpus_allowed can change in the fork path
1747 * - any previously selected cpu might disappear through hotplug
1749 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1752 rq
= __task_rq_lock(p
);
1753 activate_task(rq
, p
, 0);
1755 trace_sched_wakeup_new(p
, true);
1756 check_preempt_curr(rq
, p
, WF_FORK
);
1758 if (p
->sched_class
->task_woken
)
1759 p
->sched_class
->task_woken(rq
, p
);
1761 task_rq_unlock(rq
, p
, &flags
);
1764 #ifdef CONFIG_PREEMPT_NOTIFIERS
1767 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1768 * @notifier: notifier struct to register
1770 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1772 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1774 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1777 * preempt_notifier_unregister - no longer interested in preemption notifications
1778 * @notifier: notifier struct to unregister
1780 * This is safe to call from within a preemption notifier.
1782 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1784 hlist_del(¬ifier
->link
);
1786 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1788 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1790 struct preempt_notifier
*notifier
;
1792 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1793 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1797 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1798 struct task_struct
*next
)
1800 struct preempt_notifier
*notifier
;
1802 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1803 notifier
->ops
->sched_out(notifier
, next
);
1806 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1808 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1813 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1814 struct task_struct
*next
)
1818 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1821 * prepare_task_switch - prepare to switch tasks
1822 * @rq: the runqueue preparing to switch
1823 * @prev: the current task that is being switched out
1824 * @next: the task we are going to switch to.
1826 * This is called with the rq lock held and interrupts off. It must
1827 * be paired with a subsequent finish_task_switch after the context
1830 * prepare_task_switch sets up locking and calls architecture specific
1834 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1835 struct task_struct
*next
)
1837 trace_sched_switch(prev
, next
);
1838 sched_info_switch(prev
, next
);
1839 perf_event_task_sched_out(prev
, next
);
1840 fire_sched_out_preempt_notifiers(prev
, next
);
1841 prepare_lock_switch(rq
, next
);
1842 prepare_arch_switch(next
);
1846 * finish_task_switch - clean up after a task-switch
1847 * @rq: runqueue associated with task-switch
1848 * @prev: the thread we just switched away from.
1850 * finish_task_switch must be called after the context switch, paired
1851 * with a prepare_task_switch call before the context switch.
1852 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1853 * and do any other architecture-specific cleanup actions.
1855 * Note that we may have delayed dropping an mm in context_switch(). If
1856 * so, we finish that here outside of the runqueue lock. (Doing it
1857 * with the lock held can cause deadlocks; see schedule() for
1860 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1861 __releases(rq
->lock
)
1863 struct mm_struct
*mm
= rq
->prev_mm
;
1869 * A task struct has one reference for the use as "current".
1870 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1871 * schedule one last time. The schedule call will never return, and
1872 * the scheduled task must drop that reference.
1873 * The test for TASK_DEAD must occur while the runqueue locks are
1874 * still held, otherwise prev could be scheduled on another cpu, die
1875 * there before we look at prev->state, and then the reference would
1877 * Manfred Spraul <manfred@colorfullife.com>
1879 prev_state
= prev
->state
;
1880 vtime_task_switch(prev
);
1881 finish_arch_switch(prev
);
1882 perf_event_task_sched_in(prev
, current
);
1883 finish_lock_switch(rq
, prev
);
1884 finish_arch_post_lock_switch();
1886 fire_sched_in_preempt_notifiers(current
);
1889 if (unlikely(prev_state
== TASK_DEAD
)) {
1891 * Remove function-return probe instances associated with this
1892 * task and put them back on the free list.
1894 kprobe_flush_task(prev
);
1895 put_task_struct(prev
);
1901 /* assumes rq->lock is held */
1902 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1904 if (prev
->sched_class
->pre_schedule
)
1905 prev
->sched_class
->pre_schedule(rq
, prev
);
1908 /* rq->lock is NOT held, but preemption is disabled */
1909 static inline void post_schedule(struct rq
*rq
)
1911 if (rq
->post_schedule
) {
1912 unsigned long flags
;
1914 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1915 if (rq
->curr
->sched_class
->post_schedule
)
1916 rq
->curr
->sched_class
->post_schedule(rq
);
1917 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1919 rq
->post_schedule
= 0;
1925 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1929 static inline void post_schedule(struct rq
*rq
)
1936 * schedule_tail - first thing a freshly forked thread must call.
1937 * @prev: the thread we just switched away from.
1939 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1940 __releases(rq
->lock
)
1942 struct rq
*rq
= this_rq();
1944 finish_task_switch(rq
, prev
);
1947 * FIXME: do we need to worry about rq being invalidated by the
1952 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1953 /* In this case, finish_task_switch does not reenable preemption */
1956 if (current
->set_child_tid
)
1957 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1961 * context_switch - switch to the new MM and the new
1962 * thread's register state.
1965 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1966 struct task_struct
*next
)
1968 struct mm_struct
*mm
, *oldmm
;
1970 prepare_task_switch(rq
, prev
, next
);
1973 oldmm
= prev
->active_mm
;
1975 * For paravirt, this is coupled with an exit in switch_to to
1976 * combine the page table reload and the switch backend into
1979 arch_start_context_switch(prev
);
1982 next
->active_mm
= oldmm
;
1983 atomic_inc(&oldmm
->mm_count
);
1984 enter_lazy_tlb(oldmm
, next
);
1986 switch_mm(oldmm
, mm
, next
);
1989 prev
->active_mm
= NULL
;
1990 rq
->prev_mm
= oldmm
;
1993 * Since the runqueue lock will be released by the next
1994 * task (which is an invalid locking op but in the case
1995 * of the scheduler it's an obvious special-case), so we
1996 * do an early lockdep release here:
1998 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1999 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2002 context_tracking_task_switch(prev
, next
);
2003 /* Here we just switch the register state and the stack. */
2004 switch_to(prev
, next
, prev
);
2008 * this_rq must be evaluated again because prev may have moved
2009 * CPUs since it called schedule(), thus the 'rq' on its stack
2010 * frame will be invalid.
2012 finish_task_switch(this_rq(), prev
);
2016 * nr_running and nr_context_switches:
2018 * externally visible scheduler statistics: current number of runnable
2019 * threads, total number of context switches performed since bootup.
2021 unsigned long nr_running(void)
2023 unsigned long i
, sum
= 0;
2025 for_each_online_cpu(i
)
2026 sum
+= cpu_rq(i
)->nr_running
;
2031 unsigned long long nr_context_switches(void)
2034 unsigned long long sum
= 0;
2036 for_each_possible_cpu(i
)
2037 sum
+= cpu_rq(i
)->nr_switches
;
2042 unsigned long nr_iowait(void)
2044 unsigned long i
, sum
= 0;
2046 for_each_possible_cpu(i
)
2047 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2052 unsigned long nr_iowait_cpu(int cpu
)
2054 struct rq
*this = cpu_rq(cpu
);
2055 return atomic_read(&this->nr_iowait
);
2058 unsigned long this_cpu_load(void)
2060 struct rq
*this = this_rq();
2061 return this->cpu_load
[0];
2066 * Global load-average calculations
2068 * We take a distributed and async approach to calculating the global load-avg
2069 * in order to minimize overhead.
2071 * The global load average is an exponentially decaying average of nr_running +
2072 * nr_uninterruptible.
2074 * Once every LOAD_FREQ:
2077 * for_each_possible_cpu(cpu)
2078 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2080 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2082 * Due to a number of reasons the above turns in the mess below:
2084 * - for_each_possible_cpu() is prohibitively expensive on machines with
2085 * serious number of cpus, therefore we need to take a distributed approach
2086 * to calculating nr_active.
2088 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2089 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2091 * So assuming nr_active := 0 when we start out -- true per definition, we
2092 * can simply take per-cpu deltas and fold those into a global accumulate
2093 * to obtain the same result. See calc_load_fold_active().
2095 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2096 * across the machine, we assume 10 ticks is sufficient time for every
2097 * cpu to have completed this task.
2099 * This places an upper-bound on the IRQ-off latency of the machine. Then
2100 * again, being late doesn't loose the delta, just wrecks the sample.
2102 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2103 * this would add another cross-cpu cacheline miss and atomic operation
2104 * to the wakeup path. Instead we increment on whatever cpu the task ran
2105 * when it went into uninterruptible state and decrement on whatever cpu
2106 * did the wakeup. This means that only the sum of nr_uninterruptible over
2107 * all cpus yields the correct result.
2109 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2112 /* Variables and functions for calc_load */
2113 static atomic_long_t calc_load_tasks
;
2114 static unsigned long calc_load_update
;
2115 unsigned long avenrun
[3];
2116 EXPORT_SYMBOL(avenrun
); /* should be removed */
2119 * get_avenrun - get the load average array
2120 * @loads: pointer to dest load array
2121 * @offset: offset to add
2122 * @shift: shift count to shift the result left
2124 * These values are estimates at best, so no need for locking.
2126 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2128 loads
[0] = (avenrun
[0] + offset
) << shift
;
2129 loads
[1] = (avenrun
[1] + offset
) << shift
;
2130 loads
[2] = (avenrun
[2] + offset
) << shift
;
2133 static long calc_load_fold_active(struct rq
*this_rq
)
2135 long nr_active
, delta
= 0;
2137 nr_active
= this_rq
->nr_running
;
2138 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2140 if (nr_active
!= this_rq
->calc_load_active
) {
2141 delta
= nr_active
- this_rq
->calc_load_active
;
2142 this_rq
->calc_load_active
= nr_active
;
2149 * a1 = a0 * e + a * (1 - e)
2151 static unsigned long
2152 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2155 load
+= active
* (FIXED_1
- exp
);
2156 load
+= 1UL << (FSHIFT
- 1);
2157 return load
>> FSHIFT
;
2160 #ifdef CONFIG_NO_HZ_COMMON
2162 * Handle NO_HZ for the global load-average.
2164 * Since the above described distributed algorithm to compute the global
2165 * load-average relies on per-cpu sampling from the tick, it is affected by
2168 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2169 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2170 * when we read the global state.
2172 * Obviously reality has to ruin such a delightfully simple scheme:
2174 * - When we go NO_HZ idle during the window, we can negate our sample
2175 * contribution, causing under-accounting.
2177 * We avoid this by keeping two idle-delta counters and flipping them
2178 * when the window starts, thus separating old and new NO_HZ load.
2180 * The only trick is the slight shift in index flip for read vs write.
2184 * |-|-----------|-|-----------|-|-----------|-|
2185 * r:0 0 1 1 0 0 1 1 0
2186 * w:0 1 1 0 0 1 1 0 0
2188 * This ensures we'll fold the old idle contribution in this window while
2189 * accumlating the new one.
2191 * - When we wake up from NO_HZ idle during the window, we push up our
2192 * contribution, since we effectively move our sample point to a known
2195 * This is solved by pushing the window forward, and thus skipping the
2196 * sample, for this cpu (effectively using the idle-delta for this cpu which
2197 * was in effect at the time the window opened). This also solves the issue
2198 * of having to deal with a cpu having been in NOHZ idle for multiple
2199 * LOAD_FREQ intervals.
2201 * When making the ILB scale, we should try to pull this in as well.
2203 static atomic_long_t calc_load_idle
[2];
2204 static int calc_load_idx
;
2206 static inline int calc_load_write_idx(void)
2208 int idx
= calc_load_idx
;
2211 * See calc_global_nohz(), if we observe the new index, we also
2212 * need to observe the new update time.
2217 * If the folding window started, make sure we start writing in the
2220 if (!time_before(jiffies
, calc_load_update
))
2226 static inline int calc_load_read_idx(void)
2228 return calc_load_idx
& 1;
2231 void calc_load_enter_idle(void)
2233 struct rq
*this_rq
= this_rq();
2237 * We're going into NOHZ mode, if there's any pending delta, fold it
2238 * into the pending idle delta.
2240 delta
= calc_load_fold_active(this_rq
);
2242 int idx
= calc_load_write_idx();
2243 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2247 void calc_load_exit_idle(void)
2249 struct rq
*this_rq
= this_rq();
2252 * If we're still before the sample window, we're done.
2254 if (time_before(jiffies
, this_rq
->calc_load_update
))
2258 * We woke inside or after the sample window, this means we're already
2259 * accounted through the nohz accounting, so skip the entire deal and
2260 * sync up for the next window.
2262 this_rq
->calc_load_update
= calc_load_update
;
2263 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2264 this_rq
->calc_load_update
+= LOAD_FREQ
;
2267 static long calc_load_fold_idle(void)
2269 int idx
= calc_load_read_idx();
2272 if (atomic_long_read(&calc_load_idle
[idx
]))
2273 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2279 * fixed_power_int - compute: x^n, in O(log n) time
2281 * @x: base of the power
2282 * @frac_bits: fractional bits of @x
2283 * @n: power to raise @x to.
2285 * By exploiting the relation between the definition of the natural power
2286 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2287 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2288 * (where: n_i \elem {0, 1}, the binary vector representing n),
2289 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2290 * of course trivially computable in O(log_2 n), the length of our binary
2293 static unsigned long
2294 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2296 unsigned long result
= 1UL << frac_bits
;
2301 result
+= 1UL << (frac_bits
- 1);
2302 result
>>= frac_bits
;
2308 x
+= 1UL << (frac_bits
- 1);
2316 * a1 = a0 * e + a * (1 - e)
2318 * a2 = a1 * e + a * (1 - e)
2319 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2320 * = a0 * e^2 + a * (1 - e) * (1 + e)
2322 * a3 = a2 * e + a * (1 - e)
2323 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2324 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2328 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2329 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2330 * = a0 * e^n + a * (1 - e^n)
2332 * [1] application of the geometric series:
2335 * S_n := \Sum x^i = -------------
2338 static unsigned long
2339 calc_load_n(unsigned long load
, unsigned long exp
,
2340 unsigned long active
, unsigned int n
)
2343 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2347 * NO_HZ can leave us missing all per-cpu ticks calling
2348 * calc_load_account_active(), but since an idle CPU folds its delta into
2349 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2350 * in the pending idle delta if our idle period crossed a load cycle boundary.
2352 * Once we've updated the global active value, we need to apply the exponential
2353 * weights adjusted to the number of cycles missed.
2355 static void calc_global_nohz(void)
2357 long delta
, active
, n
;
2359 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2361 * Catch-up, fold however many we are behind still
2363 delta
= jiffies
- calc_load_update
- 10;
2364 n
= 1 + (delta
/ LOAD_FREQ
);
2366 active
= atomic_long_read(&calc_load_tasks
);
2367 active
= active
> 0 ? active
* FIXED_1
: 0;
2369 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2370 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2371 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2373 calc_load_update
+= n
* LOAD_FREQ
;
2377 * Flip the idle index...
2379 * Make sure we first write the new time then flip the index, so that
2380 * calc_load_write_idx() will see the new time when it reads the new
2381 * index, this avoids a double flip messing things up.
2386 #else /* !CONFIG_NO_HZ_COMMON */
2388 static inline long calc_load_fold_idle(void) { return 0; }
2389 static inline void calc_global_nohz(void) { }
2391 #endif /* CONFIG_NO_HZ_COMMON */
2394 * calc_load - update the avenrun load estimates 10 ticks after the
2395 * CPUs have updated calc_load_tasks.
2397 void calc_global_load(unsigned long ticks
)
2401 if (time_before(jiffies
, calc_load_update
+ 10))
2405 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2407 delta
= calc_load_fold_idle();
2409 atomic_long_add(delta
, &calc_load_tasks
);
2411 active
= atomic_long_read(&calc_load_tasks
);
2412 active
= active
> 0 ? active
* FIXED_1
: 0;
2414 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2415 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2416 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2418 calc_load_update
+= LOAD_FREQ
;
2421 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2427 * Called from update_cpu_load() to periodically update this CPU's
2430 static void calc_load_account_active(struct rq
*this_rq
)
2434 if (time_before(jiffies
, this_rq
->calc_load_update
))
2437 delta
= calc_load_fold_active(this_rq
);
2439 atomic_long_add(delta
, &calc_load_tasks
);
2441 this_rq
->calc_load_update
+= LOAD_FREQ
;
2445 * End of global load-average stuff
2449 * The exact cpuload at various idx values, calculated at every tick would be
2450 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2452 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2453 * on nth tick when cpu may be busy, then we have:
2454 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2455 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2457 * decay_load_missed() below does efficient calculation of
2458 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2459 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2461 * The calculation is approximated on a 128 point scale.
2462 * degrade_zero_ticks is the number of ticks after which load at any
2463 * particular idx is approximated to be zero.
2464 * degrade_factor is a precomputed table, a row for each load idx.
2465 * Each column corresponds to degradation factor for a power of two ticks,
2466 * based on 128 point scale.
2468 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2469 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2471 * With this power of 2 load factors, we can degrade the load n times
2472 * by looking at 1 bits in n and doing as many mult/shift instead of
2473 * n mult/shifts needed by the exact degradation.
2475 #define DEGRADE_SHIFT 7
2476 static const unsigned char
2477 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2478 static const unsigned char
2479 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2480 {0, 0, 0, 0, 0, 0, 0, 0},
2481 {64, 32, 8, 0, 0, 0, 0, 0},
2482 {96, 72, 40, 12, 1, 0, 0},
2483 {112, 98, 75, 43, 15, 1, 0},
2484 {120, 112, 98, 76, 45, 16, 2} };
2487 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2488 * would be when CPU is idle and so we just decay the old load without
2489 * adding any new load.
2491 static unsigned long
2492 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2496 if (!missed_updates
)
2499 if (missed_updates
>= degrade_zero_ticks
[idx
])
2503 return load
>> missed_updates
;
2505 while (missed_updates
) {
2506 if (missed_updates
% 2)
2507 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2509 missed_updates
>>= 1;
2516 * Update rq->cpu_load[] statistics. This function is usually called every
2517 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2518 * every tick. We fix it up based on jiffies.
2520 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2521 unsigned long pending_updates
)
2525 this_rq
->nr_load_updates
++;
2527 /* Update our load: */
2528 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2529 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2530 unsigned long old_load
, new_load
;
2532 /* scale is effectively 1 << i now, and >> i divides by scale */
2534 old_load
= this_rq
->cpu_load
[i
];
2535 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2536 new_load
= this_load
;
2538 * Round up the averaging division if load is increasing. This
2539 * prevents us from getting stuck on 9 if the load is 10, for
2542 if (new_load
> old_load
)
2543 new_load
+= scale
- 1;
2545 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2548 sched_avg_update(this_rq
);
2551 #ifdef CONFIG_NO_HZ_COMMON
2553 * There is no sane way to deal with nohz on smp when using jiffies because the
2554 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2555 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2557 * Therefore we cannot use the delta approach from the regular tick since that
2558 * would seriously skew the load calculation. However we'll make do for those
2559 * updates happening while idle (nohz_idle_balance) or coming out of idle
2560 * (tick_nohz_idle_exit).
2562 * This means we might still be one tick off for nohz periods.
2566 * Called from nohz_idle_balance() to update the load ratings before doing the
2569 void update_idle_cpu_load(struct rq
*this_rq
)
2571 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2572 unsigned long load
= this_rq
->load
.weight
;
2573 unsigned long pending_updates
;
2576 * bail if there's load or we're actually up-to-date.
2578 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2581 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2582 this_rq
->last_load_update_tick
= curr_jiffies
;
2584 __update_cpu_load(this_rq
, load
, pending_updates
);
2588 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2590 void update_cpu_load_nohz(void)
2592 struct rq
*this_rq
= this_rq();
2593 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2594 unsigned long pending_updates
;
2596 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2599 raw_spin_lock(&this_rq
->lock
);
2600 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2601 if (pending_updates
) {
2602 this_rq
->last_load_update_tick
= curr_jiffies
;
2604 * We were idle, this means load 0, the current load might be
2605 * !0 due to remote wakeups and the sort.
2607 __update_cpu_load(this_rq
, 0, pending_updates
);
2609 raw_spin_unlock(&this_rq
->lock
);
2611 #endif /* CONFIG_NO_HZ_COMMON */
2614 * Called from scheduler_tick()
2616 static void update_cpu_load_active(struct rq
*this_rq
)
2619 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2621 this_rq
->last_load_update_tick
= jiffies
;
2622 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2624 calc_load_account_active(this_rq
);
2630 * sched_exec - execve() is a valuable balancing opportunity, because at
2631 * this point the task has the smallest effective memory and cache footprint.
2633 void sched_exec(void)
2635 struct task_struct
*p
= current
;
2636 unsigned long flags
;
2639 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2640 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2641 if (dest_cpu
== smp_processor_id())
2644 if (likely(cpu_active(dest_cpu
))) {
2645 struct migration_arg arg
= { p
, dest_cpu
};
2647 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2648 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2652 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2657 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2658 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2660 EXPORT_PER_CPU_SYMBOL(kstat
);
2661 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2664 * Return any ns on the sched_clock that have not yet been accounted in
2665 * @p in case that task is currently running.
2667 * Called with task_rq_lock() held on @rq.
2669 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2673 if (task_current(rq
, p
)) {
2674 update_rq_clock(rq
);
2675 ns
= rq
->clock_task
- p
->se
.exec_start
;
2683 unsigned long long task_delta_exec(struct task_struct
*p
)
2685 unsigned long flags
;
2689 rq
= task_rq_lock(p
, &flags
);
2690 ns
= do_task_delta_exec(p
, rq
);
2691 task_rq_unlock(rq
, p
, &flags
);
2697 * Return accounted runtime for the task.
2698 * In case the task is currently running, return the runtime plus current's
2699 * pending runtime that have not been accounted yet.
2701 unsigned long long task_sched_runtime(struct task_struct
*p
)
2703 unsigned long flags
;
2707 rq
= task_rq_lock(p
, &flags
);
2708 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2709 task_rq_unlock(rq
, p
, &flags
);
2715 * This function gets called by the timer code, with HZ frequency.
2716 * We call it with interrupts disabled.
2718 void scheduler_tick(void)
2720 int cpu
= smp_processor_id();
2721 struct rq
*rq
= cpu_rq(cpu
);
2722 struct task_struct
*curr
= rq
->curr
;
2726 raw_spin_lock(&rq
->lock
);
2727 update_rq_clock(rq
);
2728 update_cpu_load_active(rq
);
2729 curr
->sched_class
->task_tick(rq
, curr
, 0);
2730 raw_spin_unlock(&rq
->lock
);
2732 perf_event_task_tick();
2735 rq
->idle_balance
= idle_cpu(cpu
);
2736 trigger_load_balance(rq
, cpu
);
2740 notrace
unsigned long get_parent_ip(unsigned long addr
)
2742 if (in_lock_functions(addr
)) {
2743 addr
= CALLER_ADDR2
;
2744 if (in_lock_functions(addr
))
2745 addr
= CALLER_ADDR3
;
2750 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2751 defined(CONFIG_PREEMPT_TRACER))
2753 void __kprobes
add_preempt_count(int val
)
2755 #ifdef CONFIG_DEBUG_PREEMPT
2759 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2762 preempt_count() += val
;
2763 #ifdef CONFIG_DEBUG_PREEMPT
2765 * Spinlock count overflowing soon?
2767 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2770 if (preempt_count() == val
)
2771 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2773 EXPORT_SYMBOL(add_preempt_count
);
2775 void __kprobes
sub_preempt_count(int val
)
2777 #ifdef CONFIG_DEBUG_PREEMPT
2781 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2784 * Is the spinlock portion underflowing?
2786 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2787 !(preempt_count() & PREEMPT_MASK
)))
2791 if (preempt_count() == val
)
2792 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2793 preempt_count() -= val
;
2795 EXPORT_SYMBOL(sub_preempt_count
);
2800 * Print scheduling while atomic bug:
2802 static noinline
void __schedule_bug(struct task_struct
*prev
)
2804 if (oops_in_progress
)
2807 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2808 prev
->comm
, prev
->pid
, preempt_count());
2810 debug_show_held_locks(prev
);
2812 if (irqs_disabled())
2813 print_irqtrace_events(prev
);
2815 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2819 * Various schedule()-time debugging checks and statistics:
2821 static inline void schedule_debug(struct task_struct
*prev
)
2824 * Test if we are atomic. Since do_exit() needs to call into
2825 * schedule() atomically, we ignore that path for now.
2826 * Otherwise, whine if we are scheduling when we should not be.
2828 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2829 __schedule_bug(prev
);
2832 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2834 schedstat_inc(this_rq(), sched_count
);
2837 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2839 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2840 update_rq_clock(rq
);
2841 prev
->sched_class
->put_prev_task(rq
, prev
);
2845 * Pick up the highest-prio task:
2847 static inline struct task_struct
*
2848 pick_next_task(struct rq
*rq
)
2850 const struct sched_class
*class;
2851 struct task_struct
*p
;
2854 * Optimization: we know that if all tasks are in
2855 * the fair class we can call that function directly:
2857 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2858 p
= fair_sched_class
.pick_next_task(rq
);
2863 for_each_class(class) {
2864 p
= class->pick_next_task(rq
);
2869 BUG(); /* the idle class will always have a runnable task */
2873 * __schedule() is the main scheduler function.
2875 * The main means of driving the scheduler and thus entering this function are:
2877 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2879 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2880 * paths. For example, see arch/x86/entry_64.S.
2882 * To drive preemption between tasks, the scheduler sets the flag in timer
2883 * interrupt handler scheduler_tick().
2885 * 3. Wakeups don't really cause entry into schedule(). They add a
2886 * task to the run-queue and that's it.
2888 * Now, if the new task added to the run-queue preempts the current
2889 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2890 * called on the nearest possible occasion:
2892 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2894 * - in syscall or exception context, at the next outmost
2895 * preempt_enable(). (this might be as soon as the wake_up()'s
2898 * - in IRQ context, return from interrupt-handler to
2899 * preemptible context
2901 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2904 * - cond_resched() call
2905 * - explicit schedule() call
2906 * - return from syscall or exception to user-space
2907 * - return from interrupt-handler to user-space
2909 static void __sched
__schedule(void)
2911 struct task_struct
*prev
, *next
;
2912 unsigned long *switch_count
;
2918 cpu
= smp_processor_id();
2920 rcu_note_context_switch(cpu
);
2923 schedule_debug(prev
);
2925 if (sched_feat(HRTICK
))
2928 raw_spin_lock_irq(&rq
->lock
);
2930 switch_count
= &prev
->nivcsw
;
2931 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2932 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2933 prev
->state
= TASK_RUNNING
;
2935 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2939 * If a worker went to sleep, notify and ask workqueue
2940 * whether it wants to wake up a task to maintain
2943 if (prev
->flags
& PF_WQ_WORKER
) {
2944 struct task_struct
*to_wakeup
;
2946 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2948 try_to_wake_up_local(to_wakeup
);
2951 switch_count
= &prev
->nvcsw
;
2954 pre_schedule(rq
, prev
);
2956 if (unlikely(!rq
->nr_running
))
2957 idle_balance(cpu
, rq
);
2959 put_prev_task(rq
, prev
);
2960 next
= pick_next_task(rq
);
2961 clear_tsk_need_resched(prev
);
2962 rq
->skip_clock_update
= 0;
2964 if (likely(prev
!= next
)) {
2969 context_switch(rq
, prev
, next
); /* unlocks the rq */
2971 * The context switch have flipped the stack from under us
2972 * and restored the local variables which were saved when
2973 * this task called schedule() in the past. prev == current
2974 * is still correct, but it can be moved to another cpu/rq.
2976 cpu
= smp_processor_id();
2979 raw_spin_unlock_irq(&rq
->lock
);
2983 sched_preempt_enable_no_resched();
2988 static inline void sched_submit_work(struct task_struct
*tsk
)
2990 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2993 * If we are going to sleep and we have plugged IO queued,
2994 * make sure to submit it to avoid deadlocks.
2996 if (blk_needs_flush_plug(tsk
))
2997 blk_schedule_flush_plug(tsk
);
3000 asmlinkage
void __sched
schedule(void)
3002 struct task_struct
*tsk
= current
;
3004 sched_submit_work(tsk
);
3007 EXPORT_SYMBOL(schedule
);
3009 #ifdef CONFIG_CONTEXT_TRACKING
3010 asmlinkage
void __sched
schedule_user(void)
3013 * If we come here after a random call to set_need_resched(),
3014 * or we have been woken up remotely but the IPI has not yet arrived,
3015 * we haven't yet exited the RCU idle mode. Do it here manually until
3016 * we find a better solution.
3025 * schedule_preempt_disabled - called with preemption disabled
3027 * Returns with preemption disabled. Note: preempt_count must be 1
3029 void __sched
schedule_preempt_disabled(void)
3031 sched_preempt_enable_no_resched();
3036 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3038 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3040 if (lock
->owner
!= owner
)
3044 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3045 * lock->owner still matches owner, if that fails, owner might
3046 * point to free()d memory, if it still matches, the rcu_read_lock()
3047 * ensures the memory stays valid.
3051 return owner
->on_cpu
;
3055 * Look out! "owner" is an entirely speculative pointer
3056 * access and not reliable.
3058 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3060 if (!sched_feat(OWNER_SPIN
))
3064 while (owner_running(lock
, owner
)) {
3068 arch_mutex_cpu_relax();
3073 * We break out the loop above on need_resched() and when the
3074 * owner changed, which is a sign for heavy contention. Return
3075 * success only when lock->owner is NULL.
3077 return lock
->owner
== NULL
;
3081 #ifdef CONFIG_PREEMPT
3083 * this is the entry point to schedule() from in-kernel preemption
3084 * off of preempt_enable. Kernel preemptions off return from interrupt
3085 * occur there and call schedule directly.
3087 asmlinkage
void __sched notrace
preempt_schedule(void)
3089 struct thread_info
*ti
= current_thread_info();
3092 * If there is a non-zero preempt_count or interrupts are disabled,
3093 * we do not want to preempt the current task. Just return..
3095 if (likely(ti
->preempt_count
|| irqs_disabled()))
3099 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3101 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3104 * Check again in case we missed a preemption opportunity
3105 * between schedule and now.
3108 } while (need_resched());
3110 EXPORT_SYMBOL(preempt_schedule
);
3113 * this is the entry point to schedule() from kernel preemption
3114 * off of irq context.
3115 * Note, that this is called and return with irqs disabled. This will
3116 * protect us against recursive calling from irq.
3118 asmlinkage
void __sched
preempt_schedule_irq(void)
3120 struct thread_info
*ti
= current_thread_info();
3121 enum ctx_state prev_state
;
3123 /* Catch callers which need to be fixed */
3124 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3126 prev_state
= exception_enter();
3129 add_preempt_count(PREEMPT_ACTIVE
);
3132 local_irq_disable();
3133 sub_preempt_count(PREEMPT_ACTIVE
);
3136 * Check again in case we missed a preemption opportunity
3137 * between schedule and now.
3140 } while (need_resched());
3142 exception_exit(prev_state
);
3145 #endif /* CONFIG_PREEMPT */
3147 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3150 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3152 EXPORT_SYMBOL(default_wake_function
);
3155 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3156 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3157 * number) then we wake all the non-exclusive tasks and one exclusive task.
3159 * There are circumstances in which we can try to wake a task which has already
3160 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3161 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3163 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3164 int nr_exclusive
, int wake_flags
, void *key
)
3166 wait_queue_t
*curr
, *next
;
3168 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3169 unsigned flags
= curr
->flags
;
3171 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3172 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3178 * __wake_up - wake up threads blocked on a waitqueue.
3180 * @mode: which threads
3181 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3182 * @key: is directly passed to the wakeup function
3184 * It may be assumed that this function implies a write memory barrier before
3185 * changing the task state if and only if any tasks are woken up.
3187 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3188 int nr_exclusive
, void *key
)
3190 unsigned long flags
;
3192 spin_lock_irqsave(&q
->lock
, flags
);
3193 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3194 spin_unlock_irqrestore(&q
->lock
, flags
);
3196 EXPORT_SYMBOL(__wake_up
);
3199 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3201 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3203 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3205 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3207 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3209 __wake_up_common(q
, mode
, 1, 0, key
);
3211 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3214 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3216 * @mode: which threads
3217 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3218 * @key: opaque value to be passed to wakeup targets
3220 * The sync wakeup differs that the waker knows that it will schedule
3221 * away soon, so while the target thread will be woken up, it will not
3222 * be migrated to another CPU - ie. the two threads are 'synchronized'
3223 * with each other. This can prevent needless bouncing between CPUs.
3225 * On UP it can prevent extra preemption.
3227 * It may be assumed that this function implies a write memory barrier before
3228 * changing the task state if and only if any tasks are woken up.
3230 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3231 int nr_exclusive
, void *key
)
3233 unsigned long flags
;
3234 int wake_flags
= WF_SYNC
;
3239 if (unlikely(!nr_exclusive
))
3242 spin_lock_irqsave(&q
->lock
, flags
);
3243 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3244 spin_unlock_irqrestore(&q
->lock
, flags
);
3246 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3249 * __wake_up_sync - see __wake_up_sync_key()
3251 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3253 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3255 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3258 * complete: - signals a single thread waiting on this completion
3259 * @x: holds the state of this particular completion
3261 * This will wake up a single thread waiting on this completion. Threads will be
3262 * awakened in the same order in which they were queued.
3264 * See also complete_all(), wait_for_completion() and related routines.
3266 * It may be assumed that this function implies a write memory barrier before
3267 * changing the task state if and only if any tasks are woken up.
3269 void complete(struct completion
*x
)
3271 unsigned long flags
;
3273 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3275 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3276 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3278 EXPORT_SYMBOL(complete
);
3281 * complete_all: - signals all threads waiting on this completion
3282 * @x: holds the state of this particular completion
3284 * This will wake up all threads waiting on this particular completion event.
3286 * It may be assumed that this function implies a write memory barrier before
3287 * changing the task state if and only if any tasks are woken up.
3289 void complete_all(struct completion
*x
)
3291 unsigned long flags
;
3293 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3294 x
->done
+= UINT_MAX
/2;
3295 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3296 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3298 EXPORT_SYMBOL(complete_all
);
3300 static inline long __sched
3301 do_wait_for_common(struct completion
*x
,
3302 long (*action
)(long), long timeout
, int state
)
3305 DECLARE_WAITQUEUE(wait
, current
);
3307 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3309 if (signal_pending_state(state
, current
)) {
3310 timeout
= -ERESTARTSYS
;
3313 __set_current_state(state
);
3314 spin_unlock_irq(&x
->wait
.lock
);
3315 timeout
= action(timeout
);
3316 spin_lock_irq(&x
->wait
.lock
);
3317 } while (!x
->done
&& timeout
);
3318 __remove_wait_queue(&x
->wait
, &wait
);
3323 return timeout
?: 1;
3326 static inline long __sched
3327 __wait_for_common(struct completion
*x
,
3328 long (*action
)(long), long timeout
, int state
)
3332 spin_lock_irq(&x
->wait
.lock
);
3333 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
3334 spin_unlock_irq(&x
->wait
.lock
);
3339 wait_for_common(struct completion
*x
, long timeout
, int state
)
3341 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
3345 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
3347 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
3351 * wait_for_completion: - waits for completion of a task
3352 * @x: holds the state of this particular completion
3354 * This waits to be signaled for completion of a specific task. It is NOT
3355 * interruptible and there is no timeout.
3357 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3358 * and interrupt capability. Also see complete().
3360 void __sched
wait_for_completion(struct completion
*x
)
3362 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3364 EXPORT_SYMBOL(wait_for_completion
);
3367 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3368 * @x: holds the state of this particular completion
3369 * @timeout: timeout value in jiffies
3371 * This waits for either a completion of a specific task to be signaled or for a
3372 * specified timeout to expire. The timeout is in jiffies. It is not
3375 * The return value is 0 if timed out, and positive (at least 1, or number of
3376 * jiffies left till timeout) if completed.
3378 unsigned long __sched
3379 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3381 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3383 EXPORT_SYMBOL(wait_for_completion_timeout
);
3386 * wait_for_completion_io: - waits for completion of a task
3387 * @x: holds the state of this particular completion
3389 * This waits to be signaled for completion of a specific task. It is NOT
3390 * interruptible and there is no timeout. The caller is accounted as waiting
3393 void __sched
wait_for_completion_io(struct completion
*x
)
3395 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3397 EXPORT_SYMBOL(wait_for_completion_io
);
3400 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3401 * @x: holds the state of this particular completion
3402 * @timeout: timeout value in jiffies
3404 * This waits for either a completion of a specific task to be signaled or for a
3405 * specified timeout to expire. The timeout is in jiffies. It is not
3406 * interruptible. The caller is accounted as waiting for IO.
3408 * The return value is 0 if timed out, and positive (at least 1, or number of
3409 * jiffies left till timeout) if completed.
3411 unsigned long __sched
3412 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
3414 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3416 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
3419 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3420 * @x: holds the state of this particular completion
3422 * This waits for completion of a specific task to be signaled. It is
3425 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3427 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3429 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3430 if (t
== -ERESTARTSYS
)
3434 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3437 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3438 * @x: holds the state of this particular completion
3439 * @timeout: timeout value in jiffies
3441 * This waits for either a completion of a specific task to be signaled or for a
3442 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3444 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3445 * positive (at least 1, or number of jiffies left till timeout) if completed.
3448 wait_for_completion_interruptible_timeout(struct completion
*x
,
3449 unsigned long timeout
)
3451 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3453 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3456 * wait_for_completion_killable: - waits for completion of a task (killable)
3457 * @x: holds the state of this particular completion
3459 * This waits to be signaled for completion of a specific task. It can be
3460 * interrupted by a kill signal.
3462 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3464 int __sched
wait_for_completion_killable(struct completion
*x
)
3466 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3467 if (t
== -ERESTARTSYS
)
3471 EXPORT_SYMBOL(wait_for_completion_killable
);
3474 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3475 * @x: holds the state of this particular completion
3476 * @timeout: timeout value in jiffies
3478 * This waits for either a completion of a specific task to be
3479 * signaled or for a specified timeout to expire. It can be
3480 * interrupted by a kill signal. The timeout is in jiffies.
3482 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3483 * positive (at least 1, or number of jiffies left till timeout) if completed.
3486 wait_for_completion_killable_timeout(struct completion
*x
,
3487 unsigned long timeout
)
3489 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3491 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3494 * try_wait_for_completion - try to decrement a completion without blocking
3495 * @x: completion structure
3497 * Returns: 0 if a decrement cannot be done without blocking
3498 * 1 if a decrement succeeded.
3500 * If a completion is being used as a counting completion,
3501 * attempt to decrement the counter without blocking. This
3502 * enables us to avoid waiting if the resource the completion
3503 * is protecting is not available.
3505 bool try_wait_for_completion(struct completion
*x
)
3507 unsigned long flags
;
3510 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3515 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3518 EXPORT_SYMBOL(try_wait_for_completion
);
3521 * completion_done - Test to see if a completion has any waiters
3522 * @x: completion structure
3524 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3525 * 1 if there are no waiters.
3528 bool completion_done(struct completion
*x
)
3530 unsigned long flags
;
3533 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3536 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3539 EXPORT_SYMBOL(completion_done
);
3542 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3544 unsigned long flags
;
3547 init_waitqueue_entry(&wait
, current
);
3549 __set_current_state(state
);
3551 spin_lock_irqsave(&q
->lock
, flags
);
3552 __add_wait_queue(q
, &wait
);
3553 spin_unlock(&q
->lock
);
3554 timeout
= schedule_timeout(timeout
);
3555 spin_lock_irq(&q
->lock
);
3556 __remove_wait_queue(q
, &wait
);
3557 spin_unlock_irqrestore(&q
->lock
, flags
);
3562 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3564 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3566 EXPORT_SYMBOL(interruptible_sleep_on
);
3569 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3571 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3573 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3575 void __sched
sleep_on(wait_queue_head_t
*q
)
3577 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3579 EXPORT_SYMBOL(sleep_on
);
3581 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3583 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3585 EXPORT_SYMBOL(sleep_on_timeout
);
3587 #ifdef CONFIG_RT_MUTEXES
3590 * rt_mutex_setprio - set the current priority of a task
3592 * @prio: prio value (kernel-internal form)
3594 * This function changes the 'effective' priority of a task. It does
3595 * not touch ->normal_prio like __setscheduler().
3597 * Used by the rt_mutex code to implement priority inheritance logic.
3599 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3601 int oldprio
, on_rq
, running
;
3603 const struct sched_class
*prev_class
;
3605 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3607 rq
= __task_rq_lock(p
);
3610 * Idle task boosting is a nono in general. There is one
3611 * exception, when PREEMPT_RT and NOHZ is active:
3613 * The idle task calls get_next_timer_interrupt() and holds
3614 * the timer wheel base->lock on the CPU and another CPU wants
3615 * to access the timer (probably to cancel it). We can safely
3616 * ignore the boosting request, as the idle CPU runs this code
3617 * with interrupts disabled and will complete the lock
3618 * protected section without being interrupted. So there is no
3619 * real need to boost.
3621 if (unlikely(p
== rq
->idle
)) {
3622 WARN_ON(p
!= rq
->curr
);
3623 WARN_ON(p
->pi_blocked_on
);
3627 trace_sched_pi_setprio(p
, prio
);
3629 prev_class
= p
->sched_class
;
3631 running
= task_current(rq
, p
);
3633 dequeue_task(rq
, p
, 0);
3635 p
->sched_class
->put_prev_task(rq
, p
);
3638 p
->sched_class
= &rt_sched_class
;
3640 p
->sched_class
= &fair_sched_class
;
3645 p
->sched_class
->set_curr_task(rq
);
3647 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3649 check_class_changed(rq
, p
, prev_class
, oldprio
);
3651 __task_rq_unlock(rq
);
3654 void set_user_nice(struct task_struct
*p
, long nice
)
3656 int old_prio
, delta
, on_rq
;
3657 unsigned long flags
;
3660 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3663 * We have to be careful, if called from sys_setpriority(),
3664 * the task might be in the middle of scheduling on another CPU.
3666 rq
= task_rq_lock(p
, &flags
);
3668 * The RT priorities are set via sched_setscheduler(), but we still
3669 * allow the 'normal' nice value to be set - but as expected
3670 * it wont have any effect on scheduling until the task is
3671 * SCHED_FIFO/SCHED_RR:
3673 if (task_has_rt_policy(p
)) {
3674 p
->static_prio
= NICE_TO_PRIO(nice
);
3679 dequeue_task(rq
, p
, 0);
3681 p
->static_prio
= NICE_TO_PRIO(nice
);
3684 p
->prio
= effective_prio(p
);
3685 delta
= p
->prio
- old_prio
;
3688 enqueue_task(rq
, p
, 0);
3690 * If the task increased its priority or is running and
3691 * lowered its priority, then reschedule its CPU:
3693 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3694 resched_task(rq
->curr
);
3697 task_rq_unlock(rq
, p
, &flags
);
3699 EXPORT_SYMBOL(set_user_nice
);
3702 * can_nice - check if a task can reduce its nice value
3706 int can_nice(const struct task_struct
*p
, const int nice
)
3708 /* convert nice value [19,-20] to rlimit style value [1,40] */
3709 int nice_rlim
= 20 - nice
;
3711 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3712 capable(CAP_SYS_NICE
));
3715 #ifdef __ARCH_WANT_SYS_NICE
3718 * sys_nice - change the priority of the current process.
3719 * @increment: priority increment
3721 * sys_setpriority is a more generic, but much slower function that
3722 * does similar things.
3724 SYSCALL_DEFINE1(nice
, int, increment
)
3729 * Setpriority might change our priority at the same moment.
3730 * We don't have to worry. Conceptually one call occurs first
3731 * and we have a single winner.
3733 if (increment
< -40)
3738 nice
= TASK_NICE(current
) + increment
;
3744 if (increment
< 0 && !can_nice(current
, nice
))
3747 retval
= security_task_setnice(current
, nice
);
3751 set_user_nice(current
, nice
);
3758 * task_prio - return the priority value of a given task.
3759 * @p: the task in question.
3761 * This is the priority value as seen by users in /proc.
3762 * RT tasks are offset by -200. Normal tasks are centered
3763 * around 0, value goes from -16 to +15.
3765 int task_prio(const struct task_struct
*p
)
3767 return p
->prio
- MAX_RT_PRIO
;
3771 * task_nice - return the nice value of a given task.
3772 * @p: the task in question.
3774 int task_nice(const struct task_struct
*p
)
3776 return TASK_NICE(p
);
3778 EXPORT_SYMBOL(task_nice
);
3781 * idle_cpu - is a given cpu idle currently?
3782 * @cpu: the processor in question.
3784 int idle_cpu(int cpu
)
3786 struct rq
*rq
= cpu_rq(cpu
);
3788 if (rq
->curr
!= rq
->idle
)
3795 if (!llist_empty(&rq
->wake_list
))
3803 * idle_task - return the idle task for a given cpu.
3804 * @cpu: the processor in question.
3806 struct task_struct
*idle_task(int cpu
)
3808 return cpu_rq(cpu
)->idle
;
3812 * find_process_by_pid - find a process with a matching PID value.
3813 * @pid: the pid in question.
3815 static struct task_struct
*find_process_by_pid(pid_t pid
)
3817 return pid
? find_task_by_vpid(pid
) : current
;
3820 /* Actually do priority change: must hold rq lock. */
3822 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3825 p
->rt_priority
= prio
;
3826 p
->normal_prio
= normal_prio(p
);
3827 /* we are holding p->pi_lock already */
3828 p
->prio
= rt_mutex_getprio(p
);
3829 if (rt_prio(p
->prio
))
3830 p
->sched_class
= &rt_sched_class
;
3832 p
->sched_class
= &fair_sched_class
;
3837 * check the target process has a UID that matches the current process's
3839 static bool check_same_owner(struct task_struct
*p
)
3841 const struct cred
*cred
= current_cred(), *pcred
;
3845 pcred
= __task_cred(p
);
3846 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3847 uid_eq(cred
->euid
, pcred
->uid
));
3852 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3853 const struct sched_param
*param
, bool user
)
3855 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3856 unsigned long flags
;
3857 const struct sched_class
*prev_class
;
3861 /* may grab non-irq protected spin_locks */
3862 BUG_ON(in_interrupt());
3864 /* double check policy once rq lock held */
3866 reset_on_fork
= p
->sched_reset_on_fork
;
3867 policy
= oldpolicy
= p
->policy
;
3869 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3870 policy
&= ~SCHED_RESET_ON_FORK
;
3872 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3873 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3874 policy
!= SCHED_IDLE
)
3879 * Valid priorities for SCHED_FIFO and SCHED_RR are
3880 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3881 * SCHED_BATCH and SCHED_IDLE is 0.
3883 if (param
->sched_priority
< 0 ||
3884 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3885 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3887 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3891 * Allow unprivileged RT tasks to decrease priority:
3893 if (user
&& !capable(CAP_SYS_NICE
)) {
3894 if (rt_policy(policy
)) {
3895 unsigned long rlim_rtprio
=
3896 task_rlimit(p
, RLIMIT_RTPRIO
);
3898 /* can't set/change the rt policy */
3899 if (policy
!= p
->policy
&& !rlim_rtprio
)
3902 /* can't increase priority */
3903 if (param
->sched_priority
> p
->rt_priority
&&
3904 param
->sched_priority
> rlim_rtprio
)
3909 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3910 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3912 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3913 if (!can_nice(p
, TASK_NICE(p
)))
3917 /* can't change other user's priorities */
3918 if (!check_same_owner(p
))
3921 /* Normal users shall not reset the sched_reset_on_fork flag */
3922 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3927 retval
= security_task_setscheduler(p
);
3933 * make sure no PI-waiters arrive (or leave) while we are
3934 * changing the priority of the task:
3936 * To be able to change p->policy safely, the appropriate
3937 * runqueue lock must be held.
3939 rq
= task_rq_lock(p
, &flags
);
3942 * Changing the policy of the stop threads its a very bad idea
3944 if (p
== rq
->stop
) {
3945 task_rq_unlock(rq
, p
, &flags
);
3950 * If not changing anything there's no need to proceed further:
3952 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3953 param
->sched_priority
== p
->rt_priority
))) {
3954 task_rq_unlock(rq
, p
, &flags
);
3958 #ifdef CONFIG_RT_GROUP_SCHED
3961 * Do not allow realtime tasks into groups that have no runtime
3964 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3965 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3966 !task_group_is_autogroup(task_group(p
))) {
3967 task_rq_unlock(rq
, p
, &flags
);
3973 /* recheck policy now with rq lock held */
3974 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3975 policy
= oldpolicy
= -1;
3976 task_rq_unlock(rq
, p
, &flags
);
3980 running
= task_current(rq
, p
);
3982 dequeue_task(rq
, p
, 0);
3984 p
->sched_class
->put_prev_task(rq
, p
);
3986 p
->sched_reset_on_fork
= reset_on_fork
;
3989 prev_class
= p
->sched_class
;
3990 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3993 p
->sched_class
->set_curr_task(rq
);
3995 enqueue_task(rq
, p
, 0);
3997 check_class_changed(rq
, p
, prev_class
, oldprio
);
3998 task_rq_unlock(rq
, p
, &flags
);
4000 rt_mutex_adjust_pi(p
);
4006 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4007 * @p: the task in question.
4008 * @policy: new policy.
4009 * @param: structure containing the new RT priority.
4011 * NOTE that the task may be already dead.
4013 int sched_setscheduler(struct task_struct
*p
, int policy
,
4014 const struct sched_param
*param
)
4016 return __sched_setscheduler(p
, policy
, param
, true);
4018 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4021 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4022 * @p: the task in question.
4023 * @policy: new policy.
4024 * @param: structure containing the new RT priority.
4026 * Just like sched_setscheduler, only don't bother checking if the
4027 * current context has permission. For example, this is needed in
4028 * stop_machine(): we create temporary high priority worker threads,
4029 * but our caller might not have that capability.
4031 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4032 const struct sched_param
*param
)
4034 return __sched_setscheduler(p
, policy
, param
, false);
4038 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4040 struct sched_param lparam
;
4041 struct task_struct
*p
;
4044 if (!param
|| pid
< 0)
4046 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4051 p
= find_process_by_pid(pid
);
4053 retval
= sched_setscheduler(p
, policy
, &lparam
);
4060 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4061 * @pid: the pid in question.
4062 * @policy: new policy.
4063 * @param: structure containing the new RT priority.
4065 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4066 struct sched_param __user
*, param
)
4068 /* negative values for policy are not valid */
4072 return do_sched_setscheduler(pid
, policy
, param
);
4076 * sys_sched_setparam - set/change the RT priority of a thread
4077 * @pid: the pid in question.
4078 * @param: structure containing the new RT priority.
4080 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4082 return do_sched_setscheduler(pid
, -1, param
);
4086 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4087 * @pid: the pid in question.
4089 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4091 struct task_struct
*p
;
4099 p
= find_process_by_pid(pid
);
4101 retval
= security_task_getscheduler(p
);
4104 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4111 * sys_sched_getparam - get the RT priority of a thread
4112 * @pid: the pid in question.
4113 * @param: structure containing the RT priority.
4115 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4117 struct sched_param lp
;
4118 struct task_struct
*p
;
4121 if (!param
|| pid
< 0)
4125 p
= find_process_by_pid(pid
);
4130 retval
= security_task_getscheduler(p
);
4134 lp
.sched_priority
= p
->rt_priority
;
4138 * This one might sleep, we cannot do it with a spinlock held ...
4140 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4149 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4151 cpumask_var_t cpus_allowed
, new_mask
;
4152 struct task_struct
*p
;
4158 p
= find_process_by_pid(pid
);
4165 /* Prevent p going away */
4169 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4173 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4175 goto out_free_cpus_allowed
;
4178 if (!check_same_owner(p
)) {
4180 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4187 retval
= security_task_setscheduler(p
);
4191 cpuset_cpus_allowed(p
, cpus_allowed
);
4192 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4194 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4197 cpuset_cpus_allowed(p
, cpus_allowed
);
4198 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4200 * We must have raced with a concurrent cpuset
4201 * update. Just reset the cpus_allowed to the
4202 * cpuset's cpus_allowed
4204 cpumask_copy(new_mask
, cpus_allowed
);
4209 free_cpumask_var(new_mask
);
4210 out_free_cpus_allowed
:
4211 free_cpumask_var(cpus_allowed
);
4218 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4219 struct cpumask
*new_mask
)
4221 if (len
< cpumask_size())
4222 cpumask_clear(new_mask
);
4223 else if (len
> cpumask_size())
4224 len
= cpumask_size();
4226 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4230 * sys_sched_setaffinity - set the cpu affinity of a process
4231 * @pid: pid of the process
4232 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4233 * @user_mask_ptr: user-space pointer to the new cpu mask
4235 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4236 unsigned long __user
*, user_mask_ptr
)
4238 cpumask_var_t new_mask
;
4241 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4244 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4246 retval
= sched_setaffinity(pid
, new_mask
);
4247 free_cpumask_var(new_mask
);
4251 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4253 struct task_struct
*p
;
4254 unsigned long flags
;
4261 p
= find_process_by_pid(pid
);
4265 retval
= security_task_getscheduler(p
);
4269 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4270 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4271 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4281 * sys_sched_getaffinity - get the cpu affinity of a process
4282 * @pid: pid of the process
4283 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4284 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4286 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4287 unsigned long __user
*, user_mask_ptr
)
4292 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4294 if (len
& (sizeof(unsigned long)-1))
4297 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4300 ret
= sched_getaffinity(pid
, mask
);
4302 size_t retlen
= min_t(size_t, len
, cpumask_size());
4304 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4309 free_cpumask_var(mask
);
4315 * sys_sched_yield - yield the current processor to other threads.
4317 * This function yields the current CPU to other tasks. If there are no
4318 * other threads running on this CPU then this function will return.
4320 SYSCALL_DEFINE0(sched_yield
)
4322 struct rq
*rq
= this_rq_lock();
4324 schedstat_inc(rq
, yld_count
);
4325 current
->sched_class
->yield_task(rq
);
4328 * Since we are going to call schedule() anyway, there's
4329 * no need to preempt or enable interrupts:
4331 __release(rq
->lock
);
4332 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4333 do_raw_spin_unlock(&rq
->lock
);
4334 sched_preempt_enable_no_resched();
4341 static inline int should_resched(void)
4343 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4346 static void __cond_resched(void)
4348 add_preempt_count(PREEMPT_ACTIVE
);
4350 sub_preempt_count(PREEMPT_ACTIVE
);
4353 int __sched
_cond_resched(void)
4355 if (should_resched()) {
4361 EXPORT_SYMBOL(_cond_resched
);
4364 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4365 * call schedule, and on return reacquire the lock.
4367 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4368 * operations here to prevent schedule() from being called twice (once via
4369 * spin_unlock(), once by hand).
4371 int __cond_resched_lock(spinlock_t
*lock
)
4373 int resched
= should_resched();
4376 lockdep_assert_held(lock
);
4378 if (spin_needbreak(lock
) || resched
) {
4389 EXPORT_SYMBOL(__cond_resched_lock
);
4391 int __sched
__cond_resched_softirq(void)
4393 BUG_ON(!in_softirq());
4395 if (should_resched()) {
4403 EXPORT_SYMBOL(__cond_resched_softirq
);
4406 * yield - yield the current processor to other threads.
4408 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4410 * The scheduler is at all times free to pick the calling task as the most
4411 * eligible task to run, if removing the yield() call from your code breaks
4412 * it, its already broken.
4414 * Typical broken usage is:
4419 * where one assumes that yield() will let 'the other' process run that will
4420 * make event true. If the current task is a SCHED_FIFO task that will never
4421 * happen. Never use yield() as a progress guarantee!!
4423 * If you want to use yield() to wait for something, use wait_event().
4424 * If you want to use yield() to be 'nice' for others, use cond_resched().
4425 * If you still want to use yield(), do not!
4427 void __sched
yield(void)
4429 set_current_state(TASK_RUNNING
);
4432 EXPORT_SYMBOL(yield
);
4435 * yield_to - yield the current processor to another thread in
4436 * your thread group, or accelerate that thread toward the
4437 * processor it's on.
4439 * @preempt: whether task preemption is allowed or not
4441 * It's the caller's job to ensure that the target task struct
4442 * can't go away on us before we can do any checks.
4445 * true (>0) if we indeed boosted the target task.
4446 * false (0) if we failed to boost the target.
4447 * -ESRCH if there's no task to yield to.
4449 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4451 struct task_struct
*curr
= current
;
4452 struct rq
*rq
, *p_rq
;
4453 unsigned long flags
;
4456 local_irq_save(flags
);
4462 * If we're the only runnable task on the rq and target rq also
4463 * has only one task, there's absolutely no point in yielding.
4465 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4470 double_rq_lock(rq
, p_rq
);
4471 while (task_rq(p
) != p_rq
) {
4472 double_rq_unlock(rq
, p_rq
);
4476 if (!curr
->sched_class
->yield_to_task
)
4479 if (curr
->sched_class
!= p
->sched_class
)
4482 if (task_running(p_rq
, p
) || p
->state
)
4485 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4487 schedstat_inc(rq
, yld_count
);
4489 * Make p's CPU reschedule; pick_next_entity takes care of
4492 if (preempt
&& rq
!= p_rq
)
4493 resched_task(p_rq
->curr
);
4497 double_rq_unlock(rq
, p_rq
);
4499 local_irq_restore(flags
);
4506 EXPORT_SYMBOL_GPL(yield_to
);
4509 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4510 * that process accounting knows that this is a task in IO wait state.
4512 void __sched
io_schedule(void)
4514 struct rq
*rq
= raw_rq();
4516 delayacct_blkio_start();
4517 atomic_inc(&rq
->nr_iowait
);
4518 blk_flush_plug(current
);
4519 current
->in_iowait
= 1;
4521 current
->in_iowait
= 0;
4522 atomic_dec(&rq
->nr_iowait
);
4523 delayacct_blkio_end();
4525 EXPORT_SYMBOL(io_schedule
);
4527 long __sched
io_schedule_timeout(long timeout
)
4529 struct rq
*rq
= raw_rq();
4532 delayacct_blkio_start();
4533 atomic_inc(&rq
->nr_iowait
);
4534 blk_flush_plug(current
);
4535 current
->in_iowait
= 1;
4536 ret
= schedule_timeout(timeout
);
4537 current
->in_iowait
= 0;
4538 atomic_dec(&rq
->nr_iowait
);
4539 delayacct_blkio_end();
4544 * sys_sched_get_priority_max - return maximum RT priority.
4545 * @policy: scheduling class.
4547 * this syscall returns the maximum rt_priority that can be used
4548 * by a given scheduling class.
4550 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4557 ret
= MAX_USER_RT_PRIO
-1;
4569 * sys_sched_get_priority_min - return minimum RT priority.
4570 * @policy: scheduling class.
4572 * this syscall returns the minimum rt_priority that can be used
4573 * by a given scheduling class.
4575 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4593 * sys_sched_rr_get_interval - return the default timeslice of a process.
4594 * @pid: pid of the process.
4595 * @interval: userspace pointer to the timeslice value.
4597 * this syscall writes the default timeslice value of a given process
4598 * into the user-space timespec buffer. A value of '0' means infinity.
4600 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4601 struct timespec __user
*, interval
)
4603 struct task_struct
*p
;
4604 unsigned int time_slice
;
4605 unsigned long flags
;
4615 p
= find_process_by_pid(pid
);
4619 retval
= security_task_getscheduler(p
);
4623 rq
= task_rq_lock(p
, &flags
);
4624 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4625 task_rq_unlock(rq
, p
, &flags
);
4628 jiffies_to_timespec(time_slice
, &t
);
4629 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4637 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4639 void sched_show_task(struct task_struct
*p
)
4641 unsigned long free
= 0;
4645 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4646 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4647 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4648 #if BITS_PER_LONG == 32
4649 if (state
== TASK_RUNNING
)
4650 printk(KERN_CONT
" running ");
4652 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4654 if (state
== TASK_RUNNING
)
4655 printk(KERN_CONT
" running task ");
4657 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4659 #ifdef CONFIG_DEBUG_STACK_USAGE
4660 free
= stack_not_used(p
);
4663 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4665 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4666 task_pid_nr(p
), ppid
,
4667 (unsigned long)task_thread_info(p
)->flags
);
4669 show_stack(p
, NULL
);
4672 void show_state_filter(unsigned long state_filter
)
4674 struct task_struct
*g
, *p
;
4676 #if BITS_PER_LONG == 32
4678 " task PC stack pid father\n");
4681 " task PC stack pid father\n");
4684 do_each_thread(g
, p
) {
4686 * reset the NMI-timeout, listing all files on a slow
4687 * console might take a lot of time:
4689 touch_nmi_watchdog();
4690 if (!state_filter
|| (p
->state
& state_filter
))
4692 } while_each_thread(g
, p
);
4694 touch_all_softlockup_watchdogs();
4696 #ifdef CONFIG_SCHED_DEBUG
4697 sysrq_sched_debug_show();
4701 * Only show locks if all tasks are dumped:
4704 debug_show_all_locks();
4707 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4709 idle
->sched_class
= &idle_sched_class
;
4713 * init_idle - set up an idle thread for a given CPU
4714 * @idle: task in question
4715 * @cpu: cpu the idle task belongs to
4717 * NOTE: this function does not set the idle thread's NEED_RESCHED
4718 * flag, to make booting more robust.
4720 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4722 struct rq
*rq
= cpu_rq(cpu
);
4723 unsigned long flags
;
4725 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4728 idle
->state
= TASK_RUNNING
;
4729 idle
->se
.exec_start
= sched_clock();
4731 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4733 * We're having a chicken and egg problem, even though we are
4734 * holding rq->lock, the cpu isn't yet set to this cpu so the
4735 * lockdep check in task_group() will fail.
4737 * Similar case to sched_fork(). / Alternatively we could
4738 * use task_rq_lock() here and obtain the other rq->lock.
4743 __set_task_cpu(idle
, cpu
);
4746 rq
->curr
= rq
->idle
= idle
;
4747 #if defined(CONFIG_SMP)
4750 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4752 /* Set the preempt count _outside_ the spinlocks! */
4753 task_thread_info(idle
)->preempt_count
= 0;
4756 * The idle tasks have their own, simple scheduling class:
4758 idle
->sched_class
= &idle_sched_class
;
4759 ftrace_graph_init_idle_task(idle
, cpu
);
4760 vtime_init_idle(idle
);
4761 #if defined(CONFIG_SMP)
4762 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4767 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4769 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4770 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4772 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4773 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4777 * This is how migration works:
4779 * 1) we invoke migration_cpu_stop() on the target CPU using
4781 * 2) stopper starts to run (implicitly forcing the migrated thread
4783 * 3) it checks whether the migrated task is still in the wrong runqueue.
4784 * 4) if it's in the wrong runqueue then the migration thread removes
4785 * it and puts it into the right queue.
4786 * 5) stopper completes and stop_one_cpu() returns and the migration
4791 * Change a given task's CPU affinity. Migrate the thread to a
4792 * proper CPU and schedule it away if the CPU it's executing on
4793 * is removed from the allowed bitmask.
4795 * NOTE: the caller must have a valid reference to the task, the
4796 * task must not exit() & deallocate itself prematurely. The
4797 * call is not atomic; no spinlocks may be held.
4799 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4801 unsigned long flags
;
4803 unsigned int dest_cpu
;
4806 rq
= task_rq_lock(p
, &flags
);
4808 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4811 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4816 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4821 do_set_cpus_allowed(p
, new_mask
);
4823 /* Can the task run on the task's current CPU? If so, we're done */
4824 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4827 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4829 struct migration_arg arg
= { p
, dest_cpu
};
4830 /* Need help from migration thread: drop lock and wait. */
4831 task_rq_unlock(rq
, p
, &flags
);
4832 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4833 tlb_migrate_finish(p
->mm
);
4837 task_rq_unlock(rq
, p
, &flags
);
4841 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4844 * Move (not current) task off this cpu, onto dest cpu. We're doing
4845 * this because either it can't run here any more (set_cpus_allowed()
4846 * away from this CPU, or CPU going down), or because we're
4847 * attempting to rebalance this task on exec (sched_exec).
4849 * So we race with normal scheduler movements, but that's OK, as long
4850 * as the task is no longer on this CPU.
4852 * Returns non-zero if task was successfully migrated.
4854 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4856 struct rq
*rq_dest
, *rq_src
;
4859 if (unlikely(!cpu_active(dest_cpu
)))
4862 rq_src
= cpu_rq(src_cpu
);
4863 rq_dest
= cpu_rq(dest_cpu
);
4865 raw_spin_lock(&p
->pi_lock
);
4866 double_rq_lock(rq_src
, rq_dest
);
4867 /* Already moved. */
4868 if (task_cpu(p
) != src_cpu
)
4870 /* Affinity changed (again). */
4871 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4875 * If we're not on a rq, the next wake-up will ensure we're
4879 dequeue_task(rq_src
, p
, 0);
4880 set_task_cpu(p
, dest_cpu
);
4881 enqueue_task(rq_dest
, p
, 0);
4882 check_preempt_curr(rq_dest
, p
, 0);
4887 double_rq_unlock(rq_src
, rq_dest
);
4888 raw_spin_unlock(&p
->pi_lock
);
4893 * migration_cpu_stop - this will be executed by a highprio stopper thread
4894 * and performs thread migration by bumping thread off CPU then
4895 * 'pushing' onto another runqueue.
4897 static int migration_cpu_stop(void *data
)
4899 struct migration_arg
*arg
= data
;
4902 * The original target cpu might have gone down and we might
4903 * be on another cpu but it doesn't matter.
4905 local_irq_disable();
4906 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4911 #ifdef CONFIG_HOTPLUG_CPU
4914 * Ensures that the idle task is using init_mm right before its cpu goes
4917 void idle_task_exit(void)
4919 struct mm_struct
*mm
= current
->active_mm
;
4921 BUG_ON(cpu_online(smp_processor_id()));
4924 switch_mm(mm
, &init_mm
, current
);
4929 * Since this CPU is going 'away' for a while, fold any nr_active delta
4930 * we might have. Assumes we're called after migrate_tasks() so that the
4931 * nr_active count is stable.
4933 * Also see the comment "Global load-average calculations".
4935 static void calc_load_migrate(struct rq
*rq
)
4937 long delta
= calc_load_fold_active(rq
);
4939 atomic_long_add(delta
, &calc_load_tasks
);
4943 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4944 * try_to_wake_up()->select_task_rq().
4946 * Called with rq->lock held even though we'er in stop_machine() and
4947 * there's no concurrency possible, we hold the required locks anyway
4948 * because of lock validation efforts.
4950 static void migrate_tasks(unsigned int dead_cpu
)
4952 struct rq
*rq
= cpu_rq(dead_cpu
);
4953 struct task_struct
*next
, *stop
= rq
->stop
;
4957 * Fudge the rq selection such that the below task selection loop
4958 * doesn't get stuck on the currently eligible stop task.
4960 * We're currently inside stop_machine() and the rq is either stuck
4961 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4962 * either way we should never end up calling schedule() until we're
4969 * There's this thread running, bail when that's the only
4972 if (rq
->nr_running
== 1)
4975 next
= pick_next_task(rq
);
4977 next
->sched_class
->put_prev_task(rq
, next
);
4979 /* Find suitable destination for @next, with force if needed. */
4980 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4981 raw_spin_unlock(&rq
->lock
);
4983 __migrate_task(next
, dead_cpu
, dest_cpu
);
4985 raw_spin_lock(&rq
->lock
);
4991 #endif /* CONFIG_HOTPLUG_CPU */
4993 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4995 static struct ctl_table sd_ctl_dir
[] = {
4997 .procname
= "sched_domain",
5003 static struct ctl_table sd_ctl_root
[] = {
5005 .procname
= "kernel",
5007 .child
= sd_ctl_dir
,
5012 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5014 struct ctl_table
*entry
=
5015 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5020 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5022 struct ctl_table
*entry
;
5025 * In the intermediate directories, both the child directory and
5026 * procname are dynamically allocated and could fail but the mode
5027 * will always be set. In the lowest directory the names are
5028 * static strings and all have proc handlers.
5030 for (entry
= *tablep
; entry
->mode
; entry
++) {
5032 sd_free_ctl_entry(&entry
->child
);
5033 if (entry
->proc_handler
== NULL
)
5034 kfree(entry
->procname
);
5041 static int min_load_idx
= 0;
5042 static int max_load_idx
= CPU_LOAD_IDX_MAX
;
5045 set_table_entry(struct ctl_table
*entry
,
5046 const char *procname
, void *data
, int maxlen
,
5047 umode_t mode
, proc_handler
*proc_handler
,
5050 entry
->procname
= procname
;
5052 entry
->maxlen
= maxlen
;
5054 entry
->proc_handler
= proc_handler
;
5057 entry
->extra1
= &min_load_idx
;
5058 entry
->extra2
= &max_load_idx
;
5062 static struct ctl_table
*
5063 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5065 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5070 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5071 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5072 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5073 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5074 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5075 sizeof(int), 0644, proc_dointvec_minmax
, true);
5076 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5077 sizeof(int), 0644, proc_dointvec_minmax
, true);
5078 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5079 sizeof(int), 0644, proc_dointvec_minmax
, true);
5080 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5081 sizeof(int), 0644, proc_dointvec_minmax
, true);
5082 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5083 sizeof(int), 0644, proc_dointvec_minmax
, true);
5084 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5085 sizeof(int), 0644, proc_dointvec_minmax
, false);
5086 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5087 sizeof(int), 0644, proc_dointvec_minmax
, false);
5088 set_table_entry(&table
[9], "cache_nice_tries",
5089 &sd
->cache_nice_tries
,
5090 sizeof(int), 0644, proc_dointvec_minmax
, false);
5091 set_table_entry(&table
[10], "flags", &sd
->flags
,
5092 sizeof(int), 0644, proc_dointvec_minmax
, false);
5093 set_table_entry(&table
[11], "name", sd
->name
,
5094 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5095 /* &table[12] is terminator */
5100 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5102 struct ctl_table
*entry
, *table
;
5103 struct sched_domain
*sd
;
5104 int domain_num
= 0, i
;
5107 for_each_domain(cpu
, sd
)
5109 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5114 for_each_domain(cpu
, sd
) {
5115 snprintf(buf
, 32, "domain%d", i
);
5116 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5118 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5125 static struct ctl_table_header
*sd_sysctl_header
;
5126 static void register_sched_domain_sysctl(void)
5128 int i
, cpu_num
= num_possible_cpus();
5129 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5132 WARN_ON(sd_ctl_dir
[0].child
);
5133 sd_ctl_dir
[0].child
= entry
;
5138 for_each_possible_cpu(i
) {
5139 snprintf(buf
, 32, "cpu%d", i
);
5140 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5142 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5146 WARN_ON(sd_sysctl_header
);
5147 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5150 /* may be called multiple times per register */
5151 static void unregister_sched_domain_sysctl(void)
5153 if (sd_sysctl_header
)
5154 unregister_sysctl_table(sd_sysctl_header
);
5155 sd_sysctl_header
= NULL
;
5156 if (sd_ctl_dir
[0].child
)
5157 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5160 static void register_sched_domain_sysctl(void)
5163 static void unregister_sched_domain_sysctl(void)
5168 static void set_rq_online(struct rq
*rq
)
5171 const struct sched_class
*class;
5173 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5176 for_each_class(class) {
5177 if (class->rq_online
)
5178 class->rq_online(rq
);
5183 static void set_rq_offline(struct rq
*rq
)
5186 const struct sched_class
*class;
5188 for_each_class(class) {
5189 if (class->rq_offline
)
5190 class->rq_offline(rq
);
5193 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5199 * migration_call - callback that gets triggered when a CPU is added.
5200 * Here we can start up the necessary migration thread for the new CPU.
5202 static int __cpuinit
5203 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5205 int cpu
= (long)hcpu
;
5206 unsigned long flags
;
5207 struct rq
*rq
= cpu_rq(cpu
);
5209 switch (action
& ~CPU_TASKS_FROZEN
) {
5211 case CPU_UP_PREPARE
:
5212 rq
->calc_load_update
= calc_load_update
;
5216 /* Update our root-domain */
5217 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5219 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5223 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5226 #ifdef CONFIG_HOTPLUG_CPU
5228 sched_ttwu_pending();
5229 /* Update our root-domain */
5230 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5232 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5236 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5237 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5241 calc_load_migrate(rq
);
5246 update_max_interval();
5252 * Register at high priority so that task migration (migrate_all_tasks)
5253 * happens before everything else. This has to be lower priority than
5254 * the notifier in the perf_event subsystem, though.
5256 static struct notifier_block __cpuinitdata migration_notifier
= {
5257 .notifier_call
= migration_call
,
5258 .priority
= CPU_PRI_MIGRATION
,
5261 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5262 unsigned long action
, void *hcpu
)
5264 switch (action
& ~CPU_TASKS_FROZEN
) {
5266 case CPU_DOWN_FAILED
:
5267 set_cpu_active((long)hcpu
, true);
5274 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5275 unsigned long action
, void *hcpu
)
5277 switch (action
& ~CPU_TASKS_FROZEN
) {
5278 case CPU_DOWN_PREPARE
:
5279 set_cpu_active((long)hcpu
, false);
5286 static int __init
migration_init(void)
5288 void *cpu
= (void *)(long)smp_processor_id();
5291 /* Initialize migration for the boot CPU */
5292 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5293 BUG_ON(err
== NOTIFY_BAD
);
5294 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5295 register_cpu_notifier(&migration_notifier
);
5297 /* Register cpu active notifiers */
5298 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5299 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5303 early_initcall(migration_init
);
5308 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5310 #ifdef CONFIG_SCHED_DEBUG
5312 static __read_mostly
int sched_debug_enabled
;
5314 static int __init
sched_debug_setup(char *str
)
5316 sched_debug_enabled
= 1;
5320 early_param("sched_debug", sched_debug_setup
);
5322 static inline bool sched_debug(void)
5324 return sched_debug_enabled
;
5327 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5328 struct cpumask
*groupmask
)
5330 struct sched_group
*group
= sd
->groups
;
5333 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5334 cpumask_clear(groupmask
);
5336 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5338 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5339 printk("does not load-balance\n");
5341 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5346 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5348 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5349 printk(KERN_ERR
"ERROR: domain->span does not contain "
5352 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5353 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5357 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5361 printk(KERN_ERR
"ERROR: group is NULL\n");
5366 * Even though we initialize ->power to something semi-sane,
5367 * we leave power_orig unset. This allows us to detect if
5368 * domain iteration is still funny without causing /0 traps.
5370 if (!group
->sgp
->power_orig
) {
5371 printk(KERN_CONT
"\n");
5372 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5377 if (!cpumask_weight(sched_group_cpus(group
))) {
5378 printk(KERN_CONT
"\n");
5379 printk(KERN_ERR
"ERROR: empty group\n");
5383 if (!(sd
->flags
& SD_OVERLAP
) &&
5384 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5385 printk(KERN_CONT
"\n");
5386 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5390 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5392 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5394 printk(KERN_CONT
" %s", str
);
5395 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5396 printk(KERN_CONT
" (cpu_power = %d)",
5400 group
= group
->next
;
5401 } while (group
!= sd
->groups
);
5402 printk(KERN_CONT
"\n");
5404 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5405 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5408 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5409 printk(KERN_ERR
"ERROR: parent span is not a superset "
5410 "of domain->span\n");
5414 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5418 if (!sched_debug_enabled
)
5422 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5426 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5429 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5437 #else /* !CONFIG_SCHED_DEBUG */
5438 # define sched_domain_debug(sd, cpu) do { } while (0)
5439 static inline bool sched_debug(void)
5443 #endif /* CONFIG_SCHED_DEBUG */
5445 static int sd_degenerate(struct sched_domain
*sd
)
5447 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5450 /* Following flags need at least 2 groups */
5451 if (sd
->flags
& (SD_LOAD_BALANCE
|
5452 SD_BALANCE_NEWIDLE
|
5456 SD_SHARE_PKG_RESOURCES
)) {
5457 if (sd
->groups
!= sd
->groups
->next
)
5461 /* Following flags don't use groups */
5462 if (sd
->flags
& (SD_WAKE_AFFINE
))
5469 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5471 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5473 if (sd_degenerate(parent
))
5476 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5479 /* Flags needing groups don't count if only 1 group in parent */
5480 if (parent
->groups
== parent
->groups
->next
) {
5481 pflags
&= ~(SD_LOAD_BALANCE
|
5482 SD_BALANCE_NEWIDLE
|
5486 SD_SHARE_PKG_RESOURCES
);
5487 if (nr_node_ids
== 1)
5488 pflags
&= ~SD_SERIALIZE
;
5490 if (~cflags
& pflags
)
5496 static void free_rootdomain(struct rcu_head
*rcu
)
5498 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5500 cpupri_cleanup(&rd
->cpupri
);
5501 free_cpumask_var(rd
->rto_mask
);
5502 free_cpumask_var(rd
->online
);
5503 free_cpumask_var(rd
->span
);
5507 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5509 struct root_domain
*old_rd
= NULL
;
5510 unsigned long flags
;
5512 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5517 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5520 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5523 * If we dont want to free the old_rt yet then
5524 * set old_rd to NULL to skip the freeing later
5527 if (!atomic_dec_and_test(&old_rd
->refcount
))
5531 atomic_inc(&rd
->refcount
);
5534 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5535 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5538 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5541 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5544 static int init_rootdomain(struct root_domain
*rd
)
5546 memset(rd
, 0, sizeof(*rd
));
5548 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5550 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5552 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5555 if (cpupri_init(&rd
->cpupri
) != 0)
5560 free_cpumask_var(rd
->rto_mask
);
5562 free_cpumask_var(rd
->online
);
5564 free_cpumask_var(rd
->span
);
5570 * By default the system creates a single root-domain with all cpus as
5571 * members (mimicking the global state we have today).
5573 struct root_domain def_root_domain
;
5575 static void init_defrootdomain(void)
5577 init_rootdomain(&def_root_domain
);
5579 atomic_set(&def_root_domain
.refcount
, 1);
5582 static struct root_domain
*alloc_rootdomain(void)
5584 struct root_domain
*rd
;
5586 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5590 if (init_rootdomain(rd
) != 0) {
5598 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5600 struct sched_group
*tmp
, *first
;
5609 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5614 } while (sg
!= first
);
5617 static void free_sched_domain(struct rcu_head
*rcu
)
5619 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5622 * If its an overlapping domain it has private groups, iterate and
5625 if (sd
->flags
& SD_OVERLAP
) {
5626 free_sched_groups(sd
->groups
, 1);
5627 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5628 kfree(sd
->groups
->sgp
);
5634 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5636 call_rcu(&sd
->rcu
, free_sched_domain
);
5639 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5641 for (; sd
; sd
= sd
->parent
)
5642 destroy_sched_domain(sd
, cpu
);
5646 * Keep a special pointer to the highest sched_domain that has
5647 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5648 * allows us to avoid some pointer chasing select_idle_sibling().
5650 * Also keep a unique ID per domain (we use the first cpu number in
5651 * the cpumask of the domain), this allows us to quickly tell if
5652 * two cpus are in the same cache domain, see cpus_share_cache().
5654 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5655 DEFINE_PER_CPU(int, sd_llc_id
);
5657 static void update_top_cache_domain(int cpu
)
5659 struct sched_domain
*sd
;
5662 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5664 id
= cpumask_first(sched_domain_span(sd
));
5666 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5667 per_cpu(sd_llc_id
, cpu
) = id
;
5671 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5672 * hold the hotplug lock.
5675 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5677 struct rq
*rq
= cpu_rq(cpu
);
5678 struct sched_domain
*tmp
;
5680 /* Remove the sched domains which do not contribute to scheduling. */
5681 for (tmp
= sd
; tmp
; ) {
5682 struct sched_domain
*parent
= tmp
->parent
;
5686 if (sd_parent_degenerate(tmp
, parent
)) {
5687 tmp
->parent
= parent
->parent
;
5689 parent
->parent
->child
= tmp
;
5690 destroy_sched_domain(parent
, cpu
);
5695 if (sd
&& sd_degenerate(sd
)) {
5698 destroy_sched_domain(tmp
, cpu
);
5703 sched_domain_debug(sd
, cpu
);
5705 rq_attach_root(rq
, rd
);
5707 rcu_assign_pointer(rq
->sd
, sd
);
5708 destroy_sched_domains(tmp
, cpu
);
5710 update_top_cache_domain(cpu
);
5713 /* cpus with isolated domains */
5714 static cpumask_var_t cpu_isolated_map
;
5716 /* Setup the mask of cpus configured for isolated domains */
5717 static int __init
isolated_cpu_setup(char *str
)
5719 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5720 cpulist_parse(str
, cpu_isolated_map
);
5724 __setup("isolcpus=", isolated_cpu_setup
);
5726 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5728 return cpumask_of_node(cpu_to_node(cpu
));
5732 struct sched_domain
**__percpu sd
;
5733 struct sched_group
**__percpu sg
;
5734 struct sched_group_power
**__percpu sgp
;
5738 struct sched_domain
** __percpu sd
;
5739 struct root_domain
*rd
;
5749 struct sched_domain_topology_level
;
5751 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5752 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5754 #define SDTL_OVERLAP 0x01
5756 struct sched_domain_topology_level
{
5757 sched_domain_init_f init
;
5758 sched_domain_mask_f mask
;
5761 struct sd_data data
;
5765 * Build an iteration mask that can exclude certain CPUs from the upwards
5768 * Asymmetric node setups can result in situations where the domain tree is of
5769 * unequal depth, make sure to skip domains that already cover the entire
5772 * In that case build_sched_domains() will have terminated the iteration early
5773 * and our sibling sd spans will be empty. Domains should always include the
5774 * cpu they're built on, so check that.
5777 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5779 const struct cpumask
*span
= sched_domain_span(sd
);
5780 struct sd_data
*sdd
= sd
->private;
5781 struct sched_domain
*sibling
;
5784 for_each_cpu(i
, span
) {
5785 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5786 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5789 cpumask_set_cpu(i
, sched_group_mask(sg
));
5794 * Return the canonical balance cpu for this group, this is the first cpu
5795 * of this group that's also in the iteration mask.
5797 int group_balance_cpu(struct sched_group
*sg
)
5799 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5803 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5805 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5806 const struct cpumask
*span
= sched_domain_span(sd
);
5807 struct cpumask
*covered
= sched_domains_tmpmask
;
5808 struct sd_data
*sdd
= sd
->private;
5809 struct sched_domain
*child
;
5812 cpumask_clear(covered
);
5814 for_each_cpu(i
, span
) {
5815 struct cpumask
*sg_span
;
5817 if (cpumask_test_cpu(i
, covered
))
5820 child
= *per_cpu_ptr(sdd
->sd
, i
);
5822 /* See the comment near build_group_mask(). */
5823 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5826 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5827 GFP_KERNEL
, cpu_to_node(cpu
));
5832 sg_span
= sched_group_cpus(sg
);
5834 child
= child
->child
;
5835 cpumask_copy(sg_span
, sched_domain_span(child
));
5837 cpumask_set_cpu(i
, sg_span
);
5839 cpumask_or(covered
, covered
, sg_span
);
5841 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5842 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5843 build_group_mask(sd
, sg
);
5846 * Initialize sgp->power such that even if we mess up the
5847 * domains and no possible iteration will get us here, we won't
5850 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5853 * Make sure the first group of this domain contains the
5854 * canonical balance cpu. Otherwise the sched_domain iteration
5855 * breaks. See update_sg_lb_stats().
5857 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5858 group_balance_cpu(sg
) == cpu
)
5868 sd
->groups
= groups
;
5873 free_sched_groups(first
, 0);
5878 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5880 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5881 struct sched_domain
*child
= sd
->child
;
5884 cpu
= cpumask_first(sched_domain_span(child
));
5887 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5888 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5889 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5896 * build_sched_groups will build a circular linked list of the groups
5897 * covered by the given span, and will set each group's ->cpumask correctly,
5898 * and ->cpu_power to 0.
5900 * Assumes the sched_domain tree is fully constructed
5903 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5905 struct sched_group
*first
= NULL
, *last
= NULL
;
5906 struct sd_data
*sdd
= sd
->private;
5907 const struct cpumask
*span
= sched_domain_span(sd
);
5908 struct cpumask
*covered
;
5911 get_group(cpu
, sdd
, &sd
->groups
);
5912 atomic_inc(&sd
->groups
->ref
);
5914 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5917 lockdep_assert_held(&sched_domains_mutex
);
5918 covered
= sched_domains_tmpmask
;
5920 cpumask_clear(covered
);
5922 for_each_cpu(i
, span
) {
5923 struct sched_group
*sg
;
5924 int group
= get_group(i
, sdd
, &sg
);
5927 if (cpumask_test_cpu(i
, covered
))
5930 cpumask_clear(sched_group_cpus(sg
));
5932 cpumask_setall(sched_group_mask(sg
));
5934 for_each_cpu(j
, span
) {
5935 if (get_group(j
, sdd
, NULL
) != group
)
5938 cpumask_set_cpu(j
, covered
);
5939 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5954 * Initialize sched groups cpu_power.
5956 * cpu_power indicates the capacity of sched group, which is used while
5957 * distributing the load between different sched groups in a sched domain.
5958 * Typically cpu_power for all the groups in a sched domain will be same unless
5959 * there are asymmetries in the topology. If there are asymmetries, group
5960 * having more cpu_power will pickup more load compared to the group having
5963 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5965 struct sched_group
*sg
= sd
->groups
;
5967 WARN_ON(!sd
|| !sg
);
5970 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5972 } while (sg
!= sd
->groups
);
5974 if (cpu
!= group_balance_cpu(sg
))
5977 update_group_power(sd
, cpu
);
5978 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5981 int __weak
arch_sd_sibling_asym_packing(void)
5983 return 0*SD_ASYM_PACKING
;
5987 * Initializers for schedule domains
5988 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5991 #ifdef CONFIG_SCHED_DEBUG
5992 # define SD_INIT_NAME(sd, type) sd->name = #type
5994 # define SD_INIT_NAME(sd, type) do { } while (0)
5997 #define SD_INIT_FUNC(type) \
5998 static noinline struct sched_domain * \
5999 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6001 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6002 *sd = SD_##type##_INIT; \
6003 SD_INIT_NAME(sd, type); \
6004 sd->private = &tl->data; \
6009 #ifdef CONFIG_SCHED_SMT
6010 SD_INIT_FUNC(SIBLING
)
6012 #ifdef CONFIG_SCHED_MC
6015 #ifdef CONFIG_SCHED_BOOK
6019 static int default_relax_domain_level
= -1;
6020 int sched_domain_level_max
;
6022 static int __init
setup_relax_domain_level(char *str
)
6024 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6025 pr_warn("Unable to set relax_domain_level\n");
6029 __setup("relax_domain_level=", setup_relax_domain_level
);
6031 static void set_domain_attribute(struct sched_domain
*sd
,
6032 struct sched_domain_attr
*attr
)
6036 if (!attr
|| attr
->relax_domain_level
< 0) {
6037 if (default_relax_domain_level
< 0)
6040 request
= default_relax_domain_level
;
6042 request
= attr
->relax_domain_level
;
6043 if (request
< sd
->level
) {
6044 /* turn off idle balance on this domain */
6045 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6047 /* turn on idle balance on this domain */
6048 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6052 static void __sdt_free(const struct cpumask
*cpu_map
);
6053 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6055 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6056 const struct cpumask
*cpu_map
)
6060 if (!atomic_read(&d
->rd
->refcount
))
6061 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6063 free_percpu(d
->sd
); /* fall through */
6065 __sdt_free(cpu_map
); /* fall through */
6071 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6072 const struct cpumask
*cpu_map
)
6074 memset(d
, 0, sizeof(*d
));
6076 if (__sdt_alloc(cpu_map
))
6077 return sa_sd_storage
;
6078 d
->sd
= alloc_percpu(struct sched_domain
*);
6080 return sa_sd_storage
;
6081 d
->rd
= alloc_rootdomain();
6084 return sa_rootdomain
;
6088 * NULL the sd_data elements we've used to build the sched_domain and
6089 * sched_group structure so that the subsequent __free_domain_allocs()
6090 * will not free the data we're using.
6092 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6094 struct sd_data
*sdd
= sd
->private;
6096 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6097 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6099 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6100 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6102 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6103 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6106 #ifdef CONFIG_SCHED_SMT
6107 static const struct cpumask
*cpu_smt_mask(int cpu
)
6109 return topology_thread_cpumask(cpu
);
6114 * Topology list, bottom-up.
6116 static struct sched_domain_topology_level default_topology
[] = {
6117 #ifdef CONFIG_SCHED_SMT
6118 { sd_init_SIBLING
, cpu_smt_mask
, },
6120 #ifdef CONFIG_SCHED_MC
6121 { sd_init_MC
, cpu_coregroup_mask
, },
6123 #ifdef CONFIG_SCHED_BOOK
6124 { sd_init_BOOK
, cpu_book_mask
, },
6126 { sd_init_CPU
, cpu_cpu_mask
, },
6130 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6134 static int sched_domains_numa_levels
;
6135 static int *sched_domains_numa_distance
;
6136 static struct cpumask
***sched_domains_numa_masks
;
6137 static int sched_domains_curr_level
;
6139 static inline int sd_local_flags(int level
)
6141 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6144 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6147 static struct sched_domain
*
6148 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6150 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6151 int level
= tl
->numa_level
;
6152 int sd_weight
= cpumask_weight(
6153 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6155 *sd
= (struct sched_domain
){
6156 .min_interval
= sd_weight
,
6157 .max_interval
= 2*sd_weight
,
6159 .imbalance_pct
= 125,
6160 .cache_nice_tries
= 2,
6167 .flags
= 1*SD_LOAD_BALANCE
6168 | 1*SD_BALANCE_NEWIDLE
6173 | 0*SD_SHARE_CPUPOWER
6174 | 0*SD_SHARE_PKG_RESOURCES
6176 | 0*SD_PREFER_SIBLING
6177 | sd_local_flags(level
)
6179 .last_balance
= jiffies
,
6180 .balance_interval
= sd_weight
,
6182 SD_INIT_NAME(sd
, NUMA
);
6183 sd
->private = &tl
->data
;
6186 * Ugly hack to pass state to sd_numa_mask()...
6188 sched_domains_curr_level
= tl
->numa_level
;
6193 static const struct cpumask
*sd_numa_mask(int cpu
)
6195 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6198 static void sched_numa_warn(const char *str
)
6200 static int done
= false;
6208 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6210 for (i
= 0; i
< nr_node_ids
; i
++) {
6211 printk(KERN_WARNING
" ");
6212 for (j
= 0; j
< nr_node_ids
; j
++)
6213 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6214 printk(KERN_CONT
"\n");
6216 printk(KERN_WARNING
"\n");
6219 static bool find_numa_distance(int distance
)
6223 if (distance
== node_distance(0, 0))
6226 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6227 if (sched_domains_numa_distance
[i
] == distance
)
6234 static void sched_init_numa(void)
6236 int next_distance
, curr_distance
= node_distance(0, 0);
6237 struct sched_domain_topology_level
*tl
;
6241 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6242 if (!sched_domains_numa_distance
)
6246 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6247 * unique distances in the node_distance() table.
6249 * Assumes node_distance(0,j) includes all distances in
6250 * node_distance(i,j) in order to avoid cubic time.
6252 next_distance
= curr_distance
;
6253 for (i
= 0; i
< nr_node_ids
; i
++) {
6254 for (j
= 0; j
< nr_node_ids
; j
++) {
6255 for (k
= 0; k
< nr_node_ids
; k
++) {
6256 int distance
= node_distance(i
, k
);
6258 if (distance
> curr_distance
&&
6259 (distance
< next_distance
||
6260 next_distance
== curr_distance
))
6261 next_distance
= distance
;
6264 * While not a strong assumption it would be nice to know
6265 * about cases where if node A is connected to B, B is not
6266 * equally connected to A.
6268 if (sched_debug() && node_distance(k
, i
) != distance
)
6269 sched_numa_warn("Node-distance not symmetric");
6271 if (sched_debug() && i
&& !find_numa_distance(distance
))
6272 sched_numa_warn("Node-0 not representative");
6274 if (next_distance
!= curr_distance
) {
6275 sched_domains_numa_distance
[level
++] = next_distance
;
6276 sched_domains_numa_levels
= level
;
6277 curr_distance
= next_distance
;
6282 * In case of sched_debug() we verify the above assumption.
6288 * 'level' contains the number of unique distances, excluding the
6289 * identity distance node_distance(i,i).
6291 * The sched_domains_nume_distance[] array includes the actual distance
6296 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6297 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6298 * the array will contain less then 'level' members. This could be
6299 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6300 * in other functions.
6302 * We reset it to 'level' at the end of this function.
6304 sched_domains_numa_levels
= 0;
6306 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6307 if (!sched_domains_numa_masks
)
6311 * Now for each level, construct a mask per node which contains all
6312 * cpus of nodes that are that many hops away from us.
6314 for (i
= 0; i
< level
; i
++) {
6315 sched_domains_numa_masks
[i
] =
6316 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6317 if (!sched_domains_numa_masks
[i
])
6320 for (j
= 0; j
< nr_node_ids
; j
++) {
6321 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6325 sched_domains_numa_masks
[i
][j
] = mask
;
6327 for (k
= 0; k
< nr_node_ids
; k
++) {
6328 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6331 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6336 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6337 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6342 * Copy the default topology bits..
6344 for (i
= 0; default_topology
[i
].init
; i
++)
6345 tl
[i
] = default_topology
[i
];
6348 * .. and append 'j' levels of NUMA goodness.
6350 for (j
= 0; j
< level
; i
++, j
++) {
6351 tl
[i
] = (struct sched_domain_topology_level
){
6352 .init
= sd_numa_init
,
6353 .mask
= sd_numa_mask
,
6354 .flags
= SDTL_OVERLAP
,
6359 sched_domain_topology
= tl
;
6361 sched_domains_numa_levels
= level
;
6364 static void sched_domains_numa_masks_set(int cpu
)
6367 int node
= cpu_to_node(cpu
);
6369 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6370 for (j
= 0; j
< nr_node_ids
; j
++) {
6371 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6372 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6377 static void sched_domains_numa_masks_clear(int cpu
)
6380 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6381 for (j
= 0; j
< nr_node_ids
; j
++)
6382 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6387 * Update sched_domains_numa_masks[level][node] array when new cpus
6390 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6391 unsigned long action
,
6394 int cpu
= (long)hcpu
;
6396 switch (action
& ~CPU_TASKS_FROZEN
) {
6398 sched_domains_numa_masks_set(cpu
);
6402 sched_domains_numa_masks_clear(cpu
);
6412 static inline void sched_init_numa(void)
6416 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6417 unsigned long action
,
6422 #endif /* CONFIG_NUMA */
6424 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6426 struct sched_domain_topology_level
*tl
;
6429 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6430 struct sd_data
*sdd
= &tl
->data
;
6432 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6436 sdd
->sg
= alloc_percpu(struct sched_group
*);
6440 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6444 for_each_cpu(j
, cpu_map
) {
6445 struct sched_domain
*sd
;
6446 struct sched_group
*sg
;
6447 struct sched_group_power
*sgp
;
6449 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6450 GFP_KERNEL
, cpu_to_node(j
));
6454 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6456 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6457 GFP_KERNEL
, cpu_to_node(j
));
6463 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6465 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6466 GFP_KERNEL
, cpu_to_node(j
));
6470 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6477 static void __sdt_free(const struct cpumask
*cpu_map
)
6479 struct sched_domain_topology_level
*tl
;
6482 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6483 struct sd_data
*sdd
= &tl
->data
;
6485 for_each_cpu(j
, cpu_map
) {
6486 struct sched_domain
*sd
;
6489 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6490 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6491 free_sched_groups(sd
->groups
, 0);
6492 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6496 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6498 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6500 free_percpu(sdd
->sd
);
6502 free_percpu(sdd
->sg
);
6504 free_percpu(sdd
->sgp
);
6509 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6510 struct s_data
*d
, const struct cpumask
*cpu_map
,
6511 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6514 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6518 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6520 sd
->level
= child
->level
+ 1;
6521 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6525 set_domain_attribute(sd
, attr
);
6531 * Build sched domains for a given set of cpus and attach the sched domains
6532 * to the individual cpus
6534 static int build_sched_domains(const struct cpumask
*cpu_map
,
6535 struct sched_domain_attr
*attr
)
6537 enum s_alloc alloc_state
= sa_none
;
6538 struct sched_domain
*sd
;
6540 int i
, ret
= -ENOMEM
;
6542 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6543 if (alloc_state
!= sa_rootdomain
)
6546 /* Set up domains for cpus specified by the cpu_map. */
6547 for_each_cpu(i
, cpu_map
) {
6548 struct sched_domain_topology_level
*tl
;
6551 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6552 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6553 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6554 sd
->flags
|= SD_OVERLAP
;
6555 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6562 *per_cpu_ptr(d
.sd
, i
) = sd
;
6565 /* Build the groups for the domains */
6566 for_each_cpu(i
, cpu_map
) {
6567 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6568 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6569 if (sd
->flags
& SD_OVERLAP
) {
6570 if (build_overlap_sched_groups(sd
, i
))
6573 if (build_sched_groups(sd
, i
))
6579 /* Calculate CPU power for physical packages and nodes */
6580 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6581 if (!cpumask_test_cpu(i
, cpu_map
))
6584 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6585 claim_allocations(i
, sd
);
6586 init_sched_groups_power(i
, sd
);
6590 /* Attach the domains */
6592 for_each_cpu(i
, cpu_map
) {
6593 sd
= *per_cpu_ptr(d
.sd
, i
);
6594 cpu_attach_domain(sd
, d
.rd
, i
);
6600 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6604 static cpumask_var_t
*doms_cur
; /* current sched domains */
6605 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6606 static struct sched_domain_attr
*dattr_cur
;
6607 /* attribues of custom domains in 'doms_cur' */
6610 * Special case: If a kmalloc of a doms_cur partition (array of
6611 * cpumask) fails, then fallback to a single sched domain,
6612 * as determined by the single cpumask fallback_doms.
6614 static cpumask_var_t fallback_doms
;
6617 * arch_update_cpu_topology lets virtualized architectures update the
6618 * cpu core maps. It is supposed to return 1 if the topology changed
6619 * or 0 if it stayed the same.
6621 int __attribute__((weak
)) arch_update_cpu_topology(void)
6626 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6629 cpumask_var_t
*doms
;
6631 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6634 for (i
= 0; i
< ndoms
; i
++) {
6635 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6636 free_sched_domains(doms
, i
);
6643 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6646 for (i
= 0; i
< ndoms
; i
++)
6647 free_cpumask_var(doms
[i
]);
6652 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6653 * For now this just excludes isolated cpus, but could be used to
6654 * exclude other special cases in the future.
6656 static int init_sched_domains(const struct cpumask
*cpu_map
)
6660 arch_update_cpu_topology();
6662 doms_cur
= alloc_sched_domains(ndoms_cur
);
6664 doms_cur
= &fallback_doms
;
6665 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6666 err
= build_sched_domains(doms_cur
[0], NULL
);
6667 register_sched_domain_sysctl();
6673 * Detach sched domains from a group of cpus specified in cpu_map
6674 * These cpus will now be attached to the NULL domain
6676 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6681 for_each_cpu(i
, cpu_map
)
6682 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6686 /* handle null as "default" */
6687 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6688 struct sched_domain_attr
*new, int idx_new
)
6690 struct sched_domain_attr tmp
;
6697 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6698 new ? (new + idx_new
) : &tmp
,
6699 sizeof(struct sched_domain_attr
));
6703 * Partition sched domains as specified by the 'ndoms_new'
6704 * cpumasks in the array doms_new[] of cpumasks. This compares
6705 * doms_new[] to the current sched domain partitioning, doms_cur[].
6706 * It destroys each deleted domain and builds each new domain.
6708 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6709 * The masks don't intersect (don't overlap.) We should setup one
6710 * sched domain for each mask. CPUs not in any of the cpumasks will
6711 * not be load balanced. If the same cpumask appears both in the
6712 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6715 * The passed in 'doms_new' should be allocated using
6716 * alloc_sched_domains. This routine takes ownership of it and will
6717 * free_sched_domains it when done with it. If the caller failed the
6718 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6719 * and partition_sched_domains() will fallback to the single partition
6720 * 'fallback_doms', it also forces the domains to be rebuilt.
6722 * If doms_new == NULL it will be replaced with cpu_online_mask.
6723 * ndoms_new == 0 is a special case for destroying existing domains,
6724 * and it will not create the default domain.
6726 * Call with hotplug lock held
6728 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6729 struct sched_domain_attr
*dattr_new
)
6734 mutex_lock(&sched_domains_mutex
);
6736 /* always unregister in case we don't destroy any domains */
6737 unregister_sched_domain_sysctl();
6739 /* Let architecture update cpu core mappings. */
6740 new_topology
= arch_update_cpu_topology();
6742 n
= doms_new
? ndoms_new
: 0;
6744 /* Destroy deleted domains */
6745 for (i
= 0; i
< ndoms_cur
; i
++) {
6746 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6747 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6748 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6751 /* no match - a current sched domain not in new doms_new[] */
6752 detach_destroy_domains(doms_cur
[i
]);
6757 if (doms_new
== NULL
) {
6759 doms_new
= &fallback_doms
;
6760 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6761 WARN_ON_ONCE(dattr_new
);
6764 /* Build new domains */
6765 for (i
= 0; i
< ndoms_new
; i
++) {
6766 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6767 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6768 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6771 /* no match - add a new doms_new */
6772 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6777 /* Remember the new sched domains */
6778 if (doms_cur
!= &fallback_doms
)
6779 free_sched_domains(doms_cur
, ndoms_cur
);
6780 kfree(dattr_cur
); /* kfree(NULL) is safe */
6781 doms_cur
= doms_new
;
6782 dattr_cur
= dattr_new
;
6783 ndoms_cur
= ndoms_new
;
6785 register_sched_domain_sysctl();
6787 mutex_unlock(&sched_domains_mutex
);
6790 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6793 * Update cpusets according to cpu_active mask. If cpusets are
6794 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6795 * around partition_sched_domains().
6797 * If we come here as part of a suspend/resume, don't touch cpusets because we
6798 * want to restore it back to its original state upon resume anyway.
6800 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6804 case CPU_ONLINE_FROZEN
:
6805 case CPU_DOWN_FAILED_FROZEN
:
6808 * num_cpus_frozen tracks how many CPUs are involved in suspend
6809 * resume sequence. As long as this is not the last online
6810 * operation in the resume sequence, just build a single sched
6811 * domain, ignoring cpusets.
6814 if (likely(num_cpus_frozen
)) {
6815 partition_sched_domains(1, NULL
, NULL
);
6820 * This is the last CPU online operation. So fall through and
6821 * restore the original sched domains by considering the
6822 * cpuset configurations.
6826 case CPU_DOWN_FAILED
:
6827 cpuset_update_active_cpus(true);
6835 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6839 case CPU_DOWN_PREPARE
:
6840 cpuset_update_active_cpus(false);
6842 case CPU_DOWN_PREPARE_FROZEN
:
6844 partition_sched_domains(1, NULL
, NULL
);
6852 void __init
sched_init_smp(void)
6854 cpumask_var_t non_isolated_cpus
;
6856 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6857 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6862 mutex_lock(&sched_domains_mutex
);
6863 init_sched_domains(cpu_active_mask
);
6864 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6865 if (cpumask_empty(non_isolated_cpus
))
6866 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6867 mutex_unlock(&sched_domains_mutex
);
6870 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6871 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6872 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6874 /* RT runtime code needs to handle some hotplug events */
6875 hotcpu_notifier(update_runtime
, 0);
6879 /* Move init over to a non-isolated CPU */
6880 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6882 sched_init_granularity();
6883 free_cpumask_var(non_isolated_cpus
);
6885 init_sched_rt_class();
6888 void __init
sched_init_smp(void)
6890 sched_init_granularity();
6892 #endif /* CONFIG_SMP */
6894 const_debug
unsigned int sysctl_timer_migration
= 1;
6896 int in_sched_functions(unsigned long addr
)
6898 return in_lock_functions(addr
) ||
6899 (addr
>= (unsigned long)__sched_text_start
6900 && addr
< (unsigned long)__sched_text_end
);
6903 #ifdef CONFIG_CGROUP_SCHED
6905 * Default task group.
6906 * Every task in system belongs to this group at bootup.
6908 struct task_group root_task_group
;
6909 LIST_HEAD(task_groups
);
6912 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6914 void __init
sched_init(void)
6917 unsigned long alloc_size
= 0, ptr
;
6919 #ifdef CONFIG_FAIR_GROUP_SCHED
6920 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6922 #ifdef CONFIG_RT_GROUP_SCHED
6923 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6925 #ifdef CONFIG_CPUMASK_OFFSTACK
6926 alloc_size
+= num_possible_cpus() * cpumask_size();
6929 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6931 #ifdef CONFIG_FAIR_GROUP_SCHED
6932 root_task_group
.se
= (struct sched_entity
**)ptr
;
6933 ptr
+= nr_cpu_ids
* sizeof(void **);
6935 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6936 ptr
+= nr_cpu_ids
* sizeof(void **);
6938 #endif /* CONFIG_FAIR_GROUP_SCHED */
6939 #ifdef CONFIG_RT_GROUP_SCHED
6940 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6941 ptr
+= nr_cpu_ids
* sizeof(void **);
6943 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6944 ptr
+= nr_cpu_ids
* sizeof(void **);
6946 #endif /* CONFIG_RT_GROUP_SCHED */
6947 #ifdef CONFIG_CPUMASK_OFFSTACK
6948 for_each_possible_cpu(i
) {
6949 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6950 ptr
+= cpumask_size();
6952 #endif /* CONFIG_CPUMASK_OFFSTACK */
6956 init_defrootdomain();
6959 init_rt_bandwidth(&def_rt_bandwidth
,
6960 global_rt_period(), global_rt_runtime());
6962 #ifdef CONFIG_RT_GROUP_SCHED
6963 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6964 global_rt_period(), global_rt_runtime());
6965 #endif /* CONFIG_RT_GROUP_SCHED */
6967 #ifdef CONFIG_CGROUP_SCHED
6968 list_add(&root_task_group
.list
, &task_groups
);
6969 INIT_LIST_HEAD(&root_task_group
.children
);
6970 INIT_LIST_HEAD(&root_task_group
.siblings
);
6971 autogroup_init(&init_task
);
6973 #endif /* CONFIG_CGROUP_SCHED */
6975 #ifdef CONFIG_CGROUP_CPUACCT
6976 root_cpuacct
.cpustat
= &kernel_cpustat
;
6977 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6978 /* Too early, not expected to fail */
6979 BUG_ON(!root_cpuacct
.cpuusage
);
6981 for_each_possible_cpu(i
) {
6985 raw_spin_lock_init(&rq
->lock
);
6987 rq
->calc_load_active
= 0;
6988 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6989 init_cfs_rq(&rq
->cfs
);
6990 init_rt_rq(&rq
->rt
, rq
);
6991 #ifdef CONFIG_FAIR_GROUP_SCHED
6992 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6993 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6995 * How much cpu bandwidth does root_task_group get?
6997 * In case of task-groups formed thr' the cgroup filesystem, it
6998 * gets 100% of the cpu resources in the system. This overall
6999 * system cpu resource is divided among the tasks of
7000 * root_task_group and its child task-groups in a fair manner,
7001 * based on each entity's (task or task-group's) weight
7002 * (se->load.weight).
7004 * In other words, if root_task_group has 10 tasks of weight
7005 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7006 * then A0's share of the cpu resource is:
7008 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7010 * We achieve this by letting root_task_group's tasks sit
7011 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7013 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7014 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7015 #endif /* CONFIG_FAIR_GROUP_SCHED */
7017 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7018 #ifdef CONFIG_RT_GROUP_SCHED
7019 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7020 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7023 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7024 rq
->cpu_load
[j
] = 0;
7026 rq
->last_load_update_tick
= jiffies
;
7031 rq
->cpu_power
= SCHED_POWER_SCALE
;
7032 rq
->post_schedule
= 0;
7033 rq
->active_balance
= 0;
7034 rq
->next_balance
= jiffies
;
7039 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7041 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7043 rq_attach_root(rq
, &def_root_domain
);
7044 #ifdef CONFIG_NO_HZ_COMMON
7049 atomic_set(&rq
->nr_iowait
, 0);
7052 set_load_weight(&init_task
);
7054 #ifdef CONFIG_PREEMPT_NOTIFIERS
7055 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7058 #ifdef CONFIG_RT_MUTEXES
7059 plist_head_init(&init_task
.pi_waiters
);
7063 * The boot idle thread does lazy MMU switching as well:
7065 atomic_inc(&init_mm
.mm_count
);
7066 enter_lazy_tlb(&init_mm
, current
);
7069 * Make us the idle thread. Technically, schedule() should not be
7070 * called from this thread, however somewhere below it might be,
7071 * but because we are the idle thread, we just pick up running again
7072 * when this runqueue becomes "idle".
7074 init_idle(current
, smp_processor_id());
7076 calc_load_update
= jiffies
+ LOAD_FREQ
;
7079 * During early bootup we pretend to be a normal task:
7081 current
->sched_class
= &fair_sched_class
;
7084 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7085 /* May be allocated at isolcpus cmdline parse time */
7086 if (cpu_isolated_map
== NULL
)
7087 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7088 idle_thread_set_boot_cpu();
7090 init_sched_fair_class();
7092 scheduler_running
= 1;
7095 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7096 static inline int preempt_count_equals(int preempt_offset
)
7098 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7100 return (nested
== preempt_offset
);
7103 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7105 static unsigned long prev_jiffy
; /* ratelimiting */
7107 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7108 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7109 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7111 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7113 prev_jiffy
= jiffies
;
7116 "BUG: sleeping function called from invalid context at %s:%d\n",
7119 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7120 in_atomic(), irqs_disabled(),
7121 current
->pid
, current
->comm
);
7123 debug_show_held_locks(current
);
7124 if (irqs_disabled())
7125 print_irqtrace_events(current
);
7128 EXPORT_SYMBOL(__might_sleep
);
7131 #ifdef CONFIG_MAGIC_SYSRQ
7132 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7134 const struct sched_class
*prev_class
= p
->sched_class
;
7135 int old_prio
= p
->prio
;
7140 dequeue_task(rq
, p
, 0);
7141 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7143 enqueue_task(rq
, p
, 0);
7144 resched_task(rq
->curr
);
7147 check_class_changed(rq
, p
, prev_class
, old_prio
);
7150 void normalize_rt_tasks(void)
7152 struct task_struct
*g
, *p
;
7153 unsigned long flags
;
7156 read_lock_irqsave(&tasklist_lock
, flags
);
7157 do_each_thread(g
, p
) {
7159 * Only normalize user tasks:
7164 p
->se
.exec_start
= 0;
7165 #ifdef CONFIG_SCHEDSTATS
7166 p
->se
.statistics
.wait_start
= 0;
7167 p
->se
.statistics
.sleep_start
= 0;
7168 p
->se
.statistics
.block_start
= 0;
7173 * Renice negative nice level userspace
7176 if (TASK_NICE(p
) < 0 && p
->mm
)
7177 set_user_nice(p
, 0);
7181 raw_spin_lock(&p
->pi_lock
);
7182 rq
= __task_rq_lock(p
);
7184 normalize_task(rq
, p
);
7186 __task_rq_unlock(rq
);
7187 raw_spin_unlock(&p
->pi_lock
);
7188 } while_each_thread(g
, p
);
7190 read_unlock_irqrestore(&tasklist_lock
, flags
);
7193 #endif /* CONFIG_MAGIC_SYSRQ */
7195 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7197 * These functions are only useful for the IA64 MCA handling, or kdb.
7199 * They can only be called when the whole system has been
7200 * stopped - every CPU needs to be quiescent, and no scheduling
7201 * activity can take place. Using them for anything else would
7202 * be a serious bug, and as a result, they aren't even visible
7203 * under any other configuration.
7207 * curr_task - return the current task for a given cpu.
7208 * @cpu: the processor in question.
7210 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7212 struct task_struct
*curr_task(int cpu
)
7214 return cpu_curr(cpu
);
7217 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7221 * set_curr_task - set the current task for a given cpu.
7222 * @cpu: the processor in question.
7223 * @p: the task pointer to set.
7225 * Description: This function must only be used when non-maskable interrupts
7226 * are serviced on a separate stack. It allows the architecture to switch the
7227 * notion of the current task on a cpu in a non-blocking manner. This function
7228 * must be called with all CPU's synchronized, and interrupts disabled, the
7229 * and caller must save the original value of the current task (see
7230 * curr_task() above) and restore that value before reenabling interrupts and
7231 * re-starting the system.
7233 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7235 void set_curr_task(int cpu
, struct task_struct
*p
)
7242 #ifdef CONFIG_CGROUP_SCHED
7243 /* task_group_lock serializes the addition/removal of task groups */
7244 static DEFINE_SPINLOCK(task_group_lock
);
7246 static void free_sched_group(struct task_group
*tg
)
7248 free_fair_sched_group(tg
);
7249 free_rt_sched_group(tg
);
7254 /* allocate runqueue etc for a new task group */
7255 struct task_group
*sched_create_group(struct task_group
*parent
)
7257 struct task_group
*tg
;
7259 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7261 return ERR_PTR(-ENOMEM
);
7263 if (!alloc_fair_sched_group(tg
, parent
))
7266 if (!alloc_rt_sched_group(tg
, parent
))
7272 free_sched_group(tg
);
7273 return ERR_PTR(-ENOMEM
);
7276 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7278 unsigned long flags
;
7280 spin_lock_irqsave(&task_group_lock
, flags
);
7281 list_add_rcu(&tg
->list
, &task_groups
);
7283 WARN_ON(!parent
); /* root should already exist */
7285 tg
->parent
= parent
;
7286 INIT_LIST_HEAD(&tg
->children
);
7287 list_add_rcu(&tg
->siblings
, &parent
->children
);
7288 spin_unlock_irqrestore(&task_group_lock
, flags
);
7291 /* rcu callback to free various structures associated with a task group */
7292 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7294 /* now it should be safe to free those cfs_rqs */
7295 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7298 /* Destroy runqueue etc associated with a task group */
7299 void sched_destroy_group(struct task_group
*tg
)
7301 /* wait for possible concurrent references to cfs_rqs complete */
7302 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7305 void sched_offline_group(struct task_group
*tg
)
7307 unsigned long flags
;
7310 /* end participation in shares distribution */
7311 for_each_possible_cpu(i
)
7312 unregister_fair_sched_group(tg
, i
);
7314 spin_lock_irqsave(&task_group_lock
, flags
);
7315 list_del_rcu(&tg
->list
);
7316 list_del_rcu(&tg
->siblings
);
7317 spin_unlock_irqrestore(&task_group_lock
, flags
);
7320 /* change task's runqueue when it moves between groups.
7321 * The caller of this function should have put the task in its new group
7322 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7323 * reflect its new group.
7325 void sched_move_task(struct task_struct
*tsk
)
7327 struct task_group
*tg
;
7329 unsigned long flags
;
7332 rq
= task_rq_lock(tsk
, &flags
);
7334 running
= task_current(rq
, tsk
);
7338 dequeue_task(rq
, tsk
, 0);
7339 if (unlikely(running
))
7340 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7342 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7343 lockdep_is_held(&tsk
->sighand
->siglock
)),
7344 struct task_group
, css
);
7345 tg
= autogroup_task_group(tsk
, tg
);
7346 tsk
->sched_task_group
= tg
;
7348 #ifdef CONFIG_FAIR_GROUP_SCHED
7349 if (tsk
->sched_class
->task_move_group
)
7350 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7353 set_task_rq(tsk
, task_cpu(tsk
));
7355 if (unlikely(running
))
7356 tsk
->sched_class
->set_curr_task(rq
);
7358 enqueue_task(rq
, tsk
, 0);
7360 task_rq_unlock(rq
, tsk
, &flags
);
7362 #endif /* CONFIG_CGROUP_SCHED */
7364 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7365 static unsigned long to_ratio(u64 period
, u64 runtime
)
7367 if (runtime
== RUNTIME_INF
)
7370 return div64_u64(runtime
<< 20, period
);
7374 #ifdef CONFIG_RT_GROUP_SCHED
7376 * Ensure that the real time constraints are schedulable.
7378 static DEFINE_MUTEX(rt_constraints_mutex
);
7380 /* Must be called with tasklist_lock held */
7381 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7383 struct task_struct
*g
, *p
;
7385 do_each_thread(g
, p
) {
7386 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7388 } while_each_thread(g
, p
);
7393 struct rt_schedulable_data
{
7394 struct task_group
*tg
;
7399 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7401 struct rt_schedulable_data
*d
= data
;
7402 struct task_group
*child
;
7403 unsigned long total
, sum
= 0;
7404 u64 period
, runtime
;
7406 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7407 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7410 period
= d
->rt_period
;
7411 runtime
= d
->rt_runtime
;
7415 * Cannot have more runtime than the period.
7417 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7421 * Ensure we don't starve existing RT tasks.
7423 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7426 total
= to_ratio(period
, runtime
);
7429 * Nobody can have more than the global setting allows.
7431 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7435 * The sum of our children's runtime should not exceed our own.
7437 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7438 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7439 runtime
= child
->rt_bandwidth
.rt_runtime
;
7441 if (child
== d
->tg
) {
7442 period
= d
->rt_period
;
7443 runtime
= d
->rt_runtime
;
7446 sum
+= to_ratio(period
, runtime
);
7455 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7459 struct rt_schedulable_data data
= {
7461 .rt_period
= period
,
7462 .rt_runtime
= runtime
,
7466 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7472 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7473 u64 rt_period
, u64 rt_runtime
)
7477 mutex_lock(&rt_constraints_mutex
);
7478 read_lock(&tasklist_lock
);
7479 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7483 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7484 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7485 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7487 for_each_possible_cpu(i
) {
7488 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7490 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7491 rt_rq
->rt_runtime
= rt_runtime
;
7492 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7494 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7496 read_unlock(&tasklist_lock
);
7497 mutex_unlock(&rt_constraints_mutex
);
7502 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7504 u64 rt_runtime
, rt_period
;
7506 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7507 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7508 if (rt_runtime_us
< 0)
7509 rt_runtime
= RUNTIME_INF
;
7511 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7514 static long sched_group_rt_runtime(struct task_group
*tg
)
7518 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7521 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7522 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7523 return rt_runtime_us
;
7526 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7528 u64 rt_runtime
, rt_period
;
7530 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7531 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7536 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7539 static long sched_group_rt_period(struct task_group
*tg
)
7543 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7544 do_div(rt_period_us
, NSEC_PER_USEC
);
7545 return rt_period_us
;
7548 static int sched_rt_global_constraints(void)
7550 u64 runtime
, period
;
7553 if (sysctl_sched_rt_period
<= 0)
7556 runtime
= global_rt_runtime();
7557 period
= global_rt_period();
7560 * Sanity check on the sysctl variables.
7562 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7565 mutex_lock(&rt_constraints_mutex
);
7566 read_lock(&tasklist_lock
);
7567 ret
= __rt_schedulable(NULL
, 0, 0);
7568 read_unlock(&tasklist_lock
);
7569 mutex_unlock(&rt_constraints_mutex
);
7574 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7576 /* Don't accept realtime tasks when there is no way for them to run */
7577 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7583 #else /* !CONFIG_RT_GROUP_SCHED */
7584 static int sched_rt_global_constraints(void)
7586 unsigned long flags
;
7589 if (sysctl_sched_rt_period
<= 0)
7593 * There's always some RT tasks in the root group
7594 * -- migration, kstopmachine etc..
7596 if (sysctl_sched_rt_runtime
== 0)
7599 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7600 for_each_possible_cpu(i
) {
7601 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7603 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7604 rt_rq
->rt_runtime
= global_rt_runtime();
7605 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7607 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7611 #endif /* CONFIG_RT_GROUP_SCHED */
7613 int sched_rr_handler(struct ctl_table
*table
, int write
,
7614 void __user
*buffer
, size_t *lenp
,
7618 static DEFINE_MUTEX(mutex
);
7621 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7622 /* make sure that internally we keep jiffies */
7623 /* also, writing zero resets timeslice to default */
7624 if (!ret
&& write
) {
7625 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7626 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7628 mutex_unlock(&mutex
);
7632 int sched_rt_handler(struct ctl_table
*table
, int write
,
7633 void __user
*buffer
, size_t *lenp
,
7637 int old_period
, old_runtime
;
7638 static DEFINE_MUTEX(mutex
);
7641 old_period
= sysctl_sched_rt_period
;
7642 old_runtime
= sysctl_sched_rt_runtime
;
7644 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7646 if (!ret
&& write
) {
7647 ret
= sched_rt_global_constraints();
7649 sysctl_sched_rt_period
= old_period
;
7650 sysctl_sched_rt_runtime
= old_runtime
;
7652 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7653 def_rt_bandwidth
.rt_period
=
7654 ns_to_ktime(global_rt_period());
7657 mutex_unlock(&mutex
);
7662 #ifdef CONFIG_CGROUP_SCHED
7664 /* return corresponding task_group object of a cgroup */
7665 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7667 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7668 struct task_group
, css
);
7671 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7673 struct task_group
*tg
, *parent
;
7675 if (!cgrp
->parent
) {
7676 /* This is early initialization for the top cgroup */
7677 return &root_task_group
.css
;
7680 parent
= cgroup_tg(cgrp
->parent
);
7681 tg
= sched_create_group(parent
);
7683 return ERR_PTR(-ENOMEM
);
7688 static int cpu_cgroup_css_online(struct cgroup
*cgrp
)
7690 struct task_group
*tg
= cgroup_tg(cgrp
);
7691 struct task_group
*parent
;
7696 parent
= cgroup_tg(cgrp
->parent
);
7697 sched_online_group(tg
, parent
);
7701 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7703 struct task_group
*tg
= cgroup_tg(cgrp
);
7705 sched_destroy_group(tg
);
7708 static void cpu_cgroup_css_offline(struct cgroup
*cgrp
)
7710 struct task_group
*tg
= cgroup_tg(cgrp
);
7712 sched_offline_group(tg
);
7715 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7716 struct cgroup_taskset
*tset
)
7718 struct task_struct
*task
;
7720 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7721 #ifdef CONFIG_RT_GROUP_SCHED
7722 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7725 /* We don't support RT-tasks being in separate groups */
7726 if (task
->sched_class
!= &fair_sched_class
)
7733 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7734 struct cgroup_taskset
*tset
)
7736 struct task_struct
*task
;
7738 cgroup_taskset_for_each(task
, cgrp
, tset
)
7739 sched_move_task(task
);
7743 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7744 struct task_struct
*task
)
7747 * cgroup_exit() is called in the copy_process() failure path.
7748 * Ignore this case since the task hasn't ran yet, this avoids
7749 * trying to poke a half freed task state from generic code.
7751 if (!(task
->flags
& PF_EXITING
))
7754 sched_move_task(task
);
7757 #ifdef CONFIG_FAIR_GROUP_SCHED
7758 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7761 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7764 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7766 struct task_group
*tg
= cgroup_tg(cgrp
);
7768 return (u64
) scale_load_down(tg
->shares
);
7771 #ifdef CONFIG_CFS_BANDWIDTH
7772 static DEFINE_MUTEX(cfs_constraints_mutex
);
7774 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7775 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7777 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7779 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7781 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7782 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7784 if (tg
== &root_task_group
)
7788 * Ensure we have at some amount of bandwidth every period. This is
7789 * to prevent reaching a state of large arrears when throttled via
7790 * entity_tick() resulting in prolonged exit starvation.
7792 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7796 * Likewise, bound things on the otherside by preventing insane quota
7797 * periods. This also allows us to normalize in computing quota
7800 if (period
> max_cfs_quota_period
)
7803 mutex_lock(&cfs_constraints_mutex
);
7804 ret
= __cfs_schedulable(tg
, period
, quota
);
7808 runtime_enabled
= quota
!= RUNTIME_INF
;
7809 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7810 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7811 raw_spin_lock_irq(&cfs_b
->lock
);
7812 cfs_b
->period
= ns_to_ktime(period
);
7813 cfs_b
->quota
= quota
;
7815 __refill_cfs_bandwidth_runtime(cfs_b
);
7816 /* restart the period timer (if active) to handle new period expiry */
7817 if (runtime_enabled
&& cfs_b
->timer_active
) {
7818 /* force a reprogram */
7819 cfs_b
->timer_active
= 0;
7820 __start_cfs_bandwidth(cfs_b
);
7822 raw_spin_unlock_irq(&cfs_b
->lock
);
7824 for_each_possible_cpu(i
) {
7825 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7826 struct rq
*rq
= cfs_rq
->rq
;
7828 raw_spin_lock_irq(&rq
->lock
);
7829 cfs_rq
->runtime_enabled
= runtime_enabled
;
7830 cfs_rq
->runtime_remaining
= 0;
7832 if (cfs_rq
->throttled
)
7833 unthrottle_cfs_rq(cfs_rq
);
7834 raw_spin_unlock_irq(&rq
->lock
);
7837 mutex_unlock(&cfs_constraints_mutex
);
7842 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7846 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7847 if (cfs_quota_us
< 0)
7848 quota
= RUNTIME_INF
;
7850 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7852 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7855 long tg_get_cfs_quota(struct task_group
*tg
)
7859 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7862 quota_us
= tg
->cfs_bandwidth
.quota
;
7863 do_div(quota_us
, NSEC_PER_USEC
);
7868 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7872 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7873 quota
= tg
->cfs_bandwidth
.quota
;
7875 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7878 long tg_get_cfs_period(struct task_group
*tg
)
7882 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7883 do_div(cfs_period_us
, NSEC_PER_USEC
);
7885 return cfs_period_us
;
7888 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7890 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7893 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7896 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7899 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7901 return tg_get_cfs_period(cgroup_tg(cgrp
));
7904 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7907 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7910 struct cfs_schedulable_data
{
7911 struct task_group
*tg
;
7916 * normalize group quota/period to be quota/max_period
7917 * note: units are usecs
7919 static u64
normalize_cfs_quota(struct task_group
*tg
,
7920 struct cfs_schedulable_data
*d
)
7928 period
= tg_get_cfs_period(tg
);
7929 quota
= tg_get_cfs_quota(tg
);
7932 /* note: these should typically be equivalent */
7933 if (quota
== RUNTIME_INF
|| quota
== -1)
7936 return to_ratio(period
, quota
);
7939 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7941 struct cfs_schedulable_data
*d
= data
;
7942 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7943 s64 quota
= 0, parent_quota
= -1;
7946 quota
= RUNTIME_INF
;
7948 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7950 quota
= normalize_cfs_quota(tg
, d
);
7951 parent_quota
= parent_b
->hierarchal_quota
;
7954 * ensure max(child_quota) <= parent_quota, inherit when no
7957 if (quota
== RUNTIME_INF
)
7958 quota
= parent_quota
;
7959 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7962 cfs_b
->hierarchal_quota
= quota
;
7967 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7970 struct cfs_schedulable_data data
= {
7976 if (quota
!= RUNTIME_INF
) {
7977 do_div(data
.period
, NSEC_PER_USEC
);
7978 do_div(data
.quota
, NSEC_PER_USEC
);
7982 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7988 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7989 struct cgroup_map_cb
*cb
)
7991 struct task_group
*tg
= cgroup_tg(cgrp
);
7992 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7994 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7995 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7996 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
8000 #endif /* CONFIG_CFS_BANDWIDTH */
8001 #endif /* CONFIG_FAIR_GROUP_SCHED */
8003 #ifdef CONFIG_RT_GROUP_SCHED
8004 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8007 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8010 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8012 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8015 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8018 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8021 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8023 return sched_group_rt_period(cgroup_tg(cgrp
));
8025 #endif /* CONFIG_RT_GROUP_SCHED */
8027 static struct cftype cpu_files
[] = {
8028 #ifdef CONFIG_FAIR_GROUP_SCHED
8031 .read_u64
= cpu_shares_read_u64
,
8032 .write_u64
= cpu_shares_write_u64
,
8035 #ifdef CONFIG_CFS_BANDWIDTH
8037 .name
= "cfs_quota_us",
8038 .read_s64
= cpu_cfs_quota_read_s64
,
8039 .write_s64
= cpu_cfs_quota_write_s64
,
8042 .name
= "cfs_period_us",
8043 .read_u64
= cpu_cfs_period_read_u64
,
8044 .write_u64
= cpu_cfs_period_write_u64
,
8048 .read_map
= cpu_stats_show
,
8051 #ifdef CONFIG_RT_GROUP_SCHED
8053 .name
= "rt_runtime_us",
8054 .read_s64
= cpu_rt_runtime_read
,
8055 .write_s64
= cpu_rt_runtime_write
,
8058 .name
= "rt_period_us",
8059 .read_u64
= cpu_rt_period_read_uint
,
8060 .write_u64
= cpu_rt_period_write_uint
,
8066 struct cgroup_subsys cpu_cgroup_subsys
= {
8068 .css_alloc
= cpu_cgroup_css_alloc
,
8069 .css_free
= cpu_cgroup_css_free
,
8070 .css_online
= cpu_cgroup_css_online
,
8071 .css_offline
= cpu_cgroup_css_offline
,
8072 .can_attach
= cpu_cgroup_can_attach
,
8073 .attach
= cpu_cgroup_attach
,
8074 .exit
= cpu_cgroup_exit
,
8075 .subsys_id
= cpu_cgroup_subsys_id
,
8076 .base_cftypes
= cpu_files
,
8080 #endif /* CONFIG_CGROUP_SCHED */
8082 #ifdef CONFIG_CGROUP_CPUACCT
8085 * CPU accounting code for task groups.
8087 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8088 * (balbir@in.ibm.com).
8091 struct cpuacct root_cpuacct
;
8093 /* create a new cpu accounting group */
8094 static struct cgroup_subsys_state
*cpuacct_css_alloc(struct cgroup
*cgrp
)
8099 return &root_cpuacct
.css
;
8101 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8105 ca
->cpuusage
= alloc_percpu(u64
);
8109 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
8111 goto out_free_cpuusage
;
8116 free_percpu(ca
->cpuusage
);
8120 return ERR_PTR(-ENOMEM
);
8123 /* destroy an existing cpu accounting group */
8124 static void cpuacct_css_free(struct cgroup
*cgrp
)
8126 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8128 free_percpu(ca
->cpustat
);
8129 free_percpu(ca
->cpuusage
);
8133 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8135 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8138 #ifndef CONFIG_64BIT
8140 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8142 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8144 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8152 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8154 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8156 #ifndef CONFIG_64BIT
8158 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8160 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8162 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8168 /* return total cpu usage (in nanoseconds) of a group */
8169 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8171 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8172 u64 totalcpuusage
= 0;
8175 for_each_present_cpu(i
)
8176 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8178 return totalcpuusage
;
8181 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8184 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8193 for_each_present_cpu(i
)
8194 cpuacct_cpuusage_write(ca
, i
, 0);
8200 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8203 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8207 for_each_present_cpu(i
) {
8208 percpu
= cpuacct_cpuusage_read(ca
, i
);
8209 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8211 seq_printf(m
, "\n");
8215 static const char *cpuacct_stat_desc
[] = {
8216 [CPUACCT_STAT_USER
] = "user",
8217 [CPUACCT_STAT_SYSTEM
] = "system",
8220 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8221 struct cgroup_map_cb
*cb
)
8223 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8227 for_each_online_cpu(cpu
) {
8228 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8229 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8230 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8232 val
= cputime64_to_clock_t(val
);
8233 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8236 for_each_online_cpu(cpu
) {
8237 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8238 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8239 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8240 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8243 val
= cputime64_to_clock_t(val
);
8244 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8249 static struct cftype files
[] = {
8252 .read_u64
= cpuusage_read
,
8253 .write_u64
= cpuusage_write
,
8256 .name
= "usage_percpu",
8257 .read_seq_string
= cpuacct_percpu_seq_read
,
8261 .read_map
= cpuacct_stats_show
,
8267 * charge this task's execution time to its accounting group.
8269 * called with rq->lock held.
8271 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8276 if (unlikely(!cpuacct_subsys
.active
))
8279 cpu
= task_cpu(tsk
);
8285 for (; ca
; ca
= parent_ca(ca
)) {
8286 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8287 *cpuusage
+= cputime
;
8293 struct cgroup_subsys cpuacct_subsys
= {
8295 .css_alloc
= cpuacct_css_alloc
,
8296 .css_free
= cpuacct_css_free
,
8297 .subsys_id
= cpuacct_subsys_id
,
8298 .base_cftypes
= files
,
8300 #endif /* CONFIG_CGROUP_CPUACCT */
8302 void dump_cpu_task(int cpu
)
8304 pr_info("Task dump for CPU %d:\n", cpu
);
8305 sched_show_task(cpu_curr(cpu
));