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>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
94 ktime_t soft
, hard
, now
;
97 if (hrtimer_active(period_timer
))
100 now
= hrtimer_cb_get_time(period_timer
);
101 hrtimer_forward(period_timer
, now
, period
);
103 soft
= hrtimer_get_softexpires(period_timer
);
104 hard
= hrtimer_get_expires(period_timer
);
105 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
106 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
107 HRTIMER_MODE_ABS_PINNED
, 0);
111 DEFINE_MUTEX(sched_domains_mutex
);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
114 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
116 void update_rq_clock(struct rq
*rq
)
120 if (rq
->skip_clock_update
> 0)
123 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
125 update_rq_clock_task(rq
, delta
);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug
unsigned int sysctl_sched_features
=
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static const char * const sched_feat_names
[] = {
146 #include "features.h"
151 static int sched_feat_show(struct seq_file
*m
, void *v
)
155 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
156 if (!(sysctl_sched_features
& (1UL << i
)))
158 seq_printf(m
, "%s ", sched_feat_names
[i
]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
174 #include "features.h"
179 static void sched_feat_disable(int i
)
181 if (static_key_enabled(&sched_feat_keys
[i
]))
182 static_key_slow_dec(&sched_feat_keys
[i
]);
185 static void sched_feat_enable(int i
)
187 if (!static_key_enabled(&sched_feat_keys
[i
]))
188 static_key_slow_inc(&sched_feat_keys
[i
]);
191 static void sched_feat_disable(int i
) { };
192 static void sched_feat_enable(int i
) { };
193 #endif /* HAVE_JUMP_LABEL */
195 static int sched_feat_set(char *cmp
)
200 if (strncmp(cmp
, "NO_", 3) == 0) {
205 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
206 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
208 sysctl_sched_features
&= ~(1UL << i
);
209 sched_feat_disable(i
);
211 sysctl_sched_features
|= (1UL << i
);
212 sched_feat_enable(i
);
222 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
223 size_t cnt
, loff_t
*ppos
)
232 if (copy_from_user(&buf
, ubuf
, cnt
))
238 i
= sched_feat_set(cmp
);
239 if (i
== __SCHED_FEAT_NR
)
247 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
249 return single_open(filp
, sched_feat_show
, NULL
);
252 static const struct file_operations sched_feat_fops
= {
253 .open
= sched_feat_open
,
254 .write
= sched_feat_write
,
257 .release
= single_release
,
260 static __init
int sched_init_debug(void)
262 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
267 late_initcall(sched_init_debug
);
268 #endif /* CONFIG_SCHED_DEBUG */
271 * Number of tasks to iterate in a single balance run.
272 * Limited because this is done with IRQs disabled.
274 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
277 * period over which we average the RT time consumption, measured
282 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
285 * period over which we measure -rt task cpu usage in us.
288 unsigned int sysctl_sched_rt_period
= 1000000;
290 __read_mostly
int scheduler_running
;
293 * part of the period that we allow rt tasks to run in us.
296 int sysctl_sched_rt_runtime
= 950000;
301 * __task_rq_lock - lock the rq @p resides on.
303 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
308 lockdep_assert_held(&p
->pi_lock
);
312 raw_spin_lock(&rq
->lock
);
313 if (likely(rq
== task_rq(p
)))
315 raw_spin_unlock(&rq
->lock
);
320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
322 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
323 __acquires(p
->pi_lock
)
329 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
331 raw_spin_lock(&rq
->lock
);
332 if (likely(rq
== task_rq(p
)))
334 raw_spin_unlock(&rq
->lock
);
335 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
339 static void __task_rq_unlock(struct rq
*rq
)
342 raw_spin_unlock(&rq
->lock
);
346 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
348 __releases(p
->pi_lock
)
350 raw_spin_unlock(&rq
->lock
);
351 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
355 * this_rq_lock - lock this runqueue and disable interrupts.
357 static struct rq
*this_rq_lock(void)
364 raw_spin_lock(&rq
->lock
);
369 #ifdef CONFIG_SCHED_HRTICK
371 * Use HR-timers to deliver accurate preemption points.
373 * Its all a bit involved since we cannot program an hrt while holding the
374 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
377 * When we get rescheduled we reprogram the hrtick_timer outside of the
381 static void hrtick_clear(struct rq
*rq
)
383 if (hrtimer_active(&rq
->hrtick_timer
))
384 hrtimer_cancel(&rq
->hrtick_timer
);
388 * High-resolution timer tick.
389 * Runs from hardirq context with interrupts disabled.
391 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
393 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
395 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
397 raw_spin_lock(&rq
->lock
);
399 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
400 raw_spin_unlock(&rq
->lock
);
402 return HRTIMER_NORESTART
;
407 * called from hardirq (IPI) context
409 static void __hrtick_start(void *arg
)
413 raw_spin_lock(&rq
->lock
);
414 hrtimer_restart(&rq
->hrtick_timer
);
415 rq
->hrtick_csd_pending
= 0;
416 raw_spin_unlock(&rq
->lock
);
420 * Called to set the hrtick timer state.
422 * called with rq->lock held and irqs disabled
424 void hrtick_start(struct rq
*rq
, u64 delay
)
426 struct hrtimer
*timer
= &rq
->hrtick_timer
;
427 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
429 hrtimer_set_expires(timer
, time
);
431 if (rq
== this_rq()) {
432 hrtimer_restart(timer
);
433 } else if (!rq
->hrtick_csd_pending
) {
434 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
435 rq
->hrtick_csd_pending
= 1;
440 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
442 int cpu
= (int)(long)hcpu
;
445 case CPU_UP_CANCELED
:
446 case CPU_UP_CANCELED_FROZEN
:
447 case CPU_DOWN_PREPARE
:
448 case CPU_DOWN_PREPARE_FROZEN
:
450 case CPU_DEAD_FROZEN
:
451 hrtick_clear(cpu_rq(cpu
));
458 static __init
void init_hrtick(void)
460 hotcpu_notifier(hotplug_hrtick
, 0);
464 * Called to set the hrtick timer state.
466 * called with rq->lock held and irqs disabled
468 void hrtick_start(struct rq
*rq
, u64 delay
)
470 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
471 HRTIMER_MODE_REL_PINNED
, 0);
474 static inline void init_hrtick(void)
477 #endif /* CONFIG_SMP */
479 static void init_rq_hrtick(struct rq
*rq
)
482 rq
->hrtick_csd_pending
= 0;
484 rq
->hrtick_csd
.flags
= 0;
485 rq
->hrtick_csd
.func
= __hrtick_start
;
486 rq
->hrtick_csd
.info
= rq
;
489 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
490 rq
->hrtick_timer
.function
= hrtick
;
492 #else /* CONFIG_SCHED_HRTICK */
493 static inline void hrtick_clear(struct rq
*rq
)
497 static inline void init_rq_hrtick(struct rq
*rq
)
501 static inline void init_hrtick(void)
504 #endif /* CONFIG_SCHED_HRTICK */
507 * resched_task - mark a task 'to be rescheduled now'.
509 * On UP this means the setting of the need_resched flag, on SMP it
510 * might also involve a cross-CPU call to trigger the scheduler on
515 #ifndef tsk_is_polling
516 #define tsk_is_polling(t) 0
519 void resched_task(struct task_struct
*p
)
523 assert_raw_spin_locked(&task_rq(p
)->lock
);
525 if (test_tsk_need_resched(p
))
528 set_tsk_need_resched(p
);
531 if (cpu
== smp_processor_id())
534 /* NEED_RESCHED must be visible before we test polling */
536 if (!tsk_is_polling(p
))
537 smp_send_reschedule(cpu
);
540 void resched_cpu(int cpu
)
542 struct rq
*rq
= cpu_rq(cpu
);
545 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
547 resched_task(cpu_curr(cpu
));
548 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
553 * In the semi idle case, use the nearest busy cpu for migrating timers
554 * from an idle cpu. This is good for power-savings.
556 * We don't do similar optimization for completely idle system, as
557 * selecting an idle cpu will add more delays to the timers than intended
558 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 int get_nohz_timer_target(void)
562 int cpu
= smp_processor_id();
564 struct sched_domain
*sd
;
567 for_each_domain(cpu
, sd
) {
568 for_each_cpu(i
, sched_domain_span(sd
)) {
580 * When add_timer_on() enqueues a timer into the timer wheel of an
581 * idle CPU then this timer might expire before the next timer event
582 * which is scheduled to wake up that CPU. In case of a completely
583 * idle system the next event might even be infinite time into the
584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
585 * leaves the inner idle loop so the newly added timer is taken into
586 * account when the CPU goes back to idle and evaluates the timer
587 * wheel for the next timer event.
589 void wake_up_idle_cpu(int cpu
)
591 struct rq
*rq
= cpu_rq(cpu
);
593 if (cpu
== smp_processor_id())
597 * This is safe, as this function is called with the timer
598 * wheel base lock of (cpu) held. When the CPU is on the way
599 * to idle and has not yet set rq->curr to idle then it will
600 * be serialized on the timer wheel base lock and take the new
601 * timer into account automatically.
603 if (rq
->curr
!= rq
->idle
)
607 * We can set TIF_RESCHED on the idle task of the other CPU
608 * lockless. The worst case is that the other CPU runs the
609 * idle task through an additional NOOP schedule()
611 set_tsk_need_resched(rq
->idle
);
613 /* NEED_RESCHED must be visible before we test polling */
615 if (!tsk_is_polling(rq
->idle
))
616 smp_send_reschedule(cpu
);
619 static inline bool got_nohz_idle_kick(void)
621 int cpu
= smp_processor_id();
622 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
625 #else /* CONFIG_NO_HZ */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ */
634 void sched_avg_update(struct rq
*rq
)
636 s64 period
= sched_avg_period();
638 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
640 * Inline assembly required to prevent the compiler
641 * optimising this loop into a divmod call.
642 * See __iter_div_u64_rem() for another example of this.
644 asm("" : "+rm" (rq
->age_stamp
));
645 rq
->age_stamp
+= period
;
650 #else /* !CONFIG_SMP */
651 void resched_task(struct task_struct
*p
)
653 assert_raw_spin_locked(&task_rq(p
)->lock
);
654 set_tsk_need_resched(p
);
656 #endif /* CONFIG_SMP */
658 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
659 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
661 * Iterate task_group tree rooted at *from, calling @down when first entering a
662 * node and @up when leaving it for the final time.
664 * Caller must hold rcu_lock or sufficient equivalent.
666 int walk_tg_tree_from(struct task_group
*from
,
667 tg_visitor down
, tg_visitor up
, void *data
)
669 struct task_group
*parent
, *child
;
675 ret
= (*down
)(parent
, data
);
678 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
685 ret
= (*up
)(parent
, data
);
686 if (ret
|| parent
== from
)
690 parent
= parent
->parent
;
697 int tg_nop(struct task_group
*tg
, void *data
)
703 static void set_load_weight(struct task_struct
*p
)
705 int prio
= p
->static_prio
- MAX_RT_PRIO
;
706 struct load_weight
*load
= &p
->se
.load
;
709 * SCHED_IDLE tasks get minimal weight:
711 if (p
->policy
== SCHED_IDLE
) {
712 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
713 load
->inv_weight
= WMULT_IDLEPRIO
;
717 load
->weight
= scale_load(prio_to_weight
[prio
]);
718 load
->inv_weight
= prio_to_wmult
[prio
];
721 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
724 sched_info_queued(p
);
725 p
->sched_class
->enqueue_task(rq
, p
, flags
);
728 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
731 sched_info_dequeued(p
);
732 p
->sched_class
->dequeue_task(rq
, p
, flags
);
735 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
737 if (task_contributes_to_load(p
))
738 rq
->nr_uninterruptible
--;
740 enqueue_task(rq
, p
, flags
);
743 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
745 if (task_contributes_to_load(p
))
746 rq
->nr_uninterruptible
++;
748 dequeue_task(rq
, p
, flags
);
751 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
754 * In theory, the compile should just see 0 here, and optimize out the call
755 * to sched_rt_avg_update. But I don't trust it...
757 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
758 s64 steal
= 0, irq_delta
= 0;
760 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
761 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
764 * Since irq_time is only updated on {soft,}irq_exit, we might run into
765 * this case when a previous update_rq_clock() happened inside a
768 * When this happens, we stop ->clock_task and only update the
769 * prev_irq_time stamp to account for the part that fit, so that a next
770 * update will consume the rest. This ensures ->clock_task is
773 * It does however cause some slight miss-attribution of {soft,}irq
774 * time, a more accurate solution would be to update the irq_time using
775 * the current rq->clock timestamp, except that would require using
778 if (irq_delta
> delta
)
781 rq
->prev_irq_time
+= irq_delta
;
784 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
785 if (static_key_false((¶virt_steal_rq_enabled
))) {
788 steal
= paravirt_steal_clock(cpu_of(rq
));
789 steal
-= rq
->prev_steal_time_rq
;
791 if (unlikely(steal
> delta
))
794 st
= steal_ticks(steal
);
795 steal
= st
* TICK_NSEC
;
797 rq
->prev_steal_time_rq
+= steal
;
803 rq
->clock_task
+= delta
;
805 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
806 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
807 sched_rt_avg_update(rq
, irq_delta
+ steal
);
811 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
813 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
814 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
818 * Make it appear like a SCHED_FIFO task, its something
819 * userspace knows about and won't get confused about.
821 * Also, it will make PI more or less work without too
822 * much confusion -- but then, stop work should not
823 * rely on PI working anyway.
825 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
827 stop
->sched_class
= &stop_sched_class
;
830 cpu_rq(cpu
)->stop
= stop
;
834 * Reset it back to a normal scheduling class so that
835 * it can die in pieces.
837 old_stop
->sched_class
= &rt_sched_class
;
842 * __normal_prio - return the priority that is based on the static prio
844 static inline int __normal_prio(struct task_struct
*p
)
846 return p
->static_prio
;
850 * Calculate the expected normal priority: i.e. priority
851 * without taking RT-inheritance into account. Might be
852 * boosted by interactivity modifiers. Changes upon fork,
853 * setprio syscalls, and whenever the interactivity
854 * estimator recalculates.
856 static inline int normal_prio(struct task_struct
*p
)
860 if (task_has_rt_policy(p
))
861 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
863 prio
= __normal_prio(p
);
868 * Calculate the current priority, i.e. the priority
869 * taken into account by the scheduler. This value might
870 * be boosted by RT tasks, or might be boosted by
871 * interactivity modifiers. Will be RT if the task got
872 * RT-boosted. If not then it returns p->normal_prio.
874 static int effective_prio(struct task_struct
*p
)
876 p
->normal_prio
= normal_prio(p
);
878 * If we are RT tasks or we were boosted to RT priority,
879 * keep the priority unchanged. Otherwise, update priority
880 * to the normal priority:
882 if (!rt_prio(p
->prio
))
883 return p
->normal_prio
;
888 * task_curr - is this task currently executing on a CPU?
889 * @p: the task in question.
891 inline int task_curr(const struct task_struct
*p
)
893 return cpu_curr(task_cpu(p
)) == p
;
896 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
897 const struct sched_class
*prev_class
,
900 if (prev_class
!= p
->sched_class
) {
901 if (prev_class
->switched_from
)
902 prev_class
->switched_from(rq
, p
);
903 p
->sched_class
->switched_to(rq
, p
);
904 } else if (oldprio
!= p
->prio
)
905 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
908 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
910 const struct sched_class
*class;
912 if (p
->sched_class
== rq
->curr
->sched_class
) {
913 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
915 for_each_class(class) {
916 if (class == rq
->curr
->sched_class
)
918 if (class == p
->sched_class
) {
919 resched_task(rq
->curr
);
926 * A queue event has occurred, and we're going to schedule. In
927 * this case, we can save a useless back to back clock update.
929 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
930 rq
->skip_clock_update
= 1;
934 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
936 #ifdef CONFIG_SCHED_DEBUG
938 * We should never call set_task_cpu() on a blocked task,
939 * ttwu() will sort out the placement.
941 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
942 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
944 #ifdef CONFIG_LOCKDEP
946 * The caller should hold either p->pi_lock or rq->lock, when changing
947 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
949 * sched_move_task() holds both and thus holding either pins the cgroup,
952 * Furthermore, all task_rq users should acquire both locks, see
955 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
956 lockdep_is_held(&task_rq(p
)->lock
)));
960 trace_sched_migrate_task(p
, new_cpu
);
962 if (task_cpu(p
) != new_cpu
) {
963 p
->se
.nr_migrations
++;
964 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
967 __set_task_cpu(p
, new_cpu
);
970 struct migration_arg
{
971 struct task_struct
*task
;
975 static int migration_cpu_stop(void *data
);
978 * wait_task_inactive - wait for a thread to unschedule.
980 * If @match_state is nonzero, it's the @p->state value just checked and
981 * not expected to change. If it changes, i.e. @p might have woken up,
982 * then return zero. When we succeed in waiting for @p to be off its CPU,
983 * we return a positive number (its total switch count). If a second call
984 * a short while later returns the same number, the caller can be sure that
985 * @p has remained unscheduled the whole time.
987 * The caller must ensure that the task *will* unschedule sometime soon,
988 * else this function might spin for a *long* time. This function can't
989 * be called with interrupts off, or it may introduce deadlock with
990 * smp_call_function() if an IPI is sent by the same process we are
991 * waiting to become inactive.
993 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1002 * We do the initial early heuristics without holding
1003 * any task-queue locks at all. We'll only try to get
1004 * the runqueue lock when things look like they will
1010 * If the task is actively running on another CPU
1011 * still, just relax and busy-wait without holding
1014 * NOTE! Since we don't hold any locks, it's not
1015 * even sure that "rq" stays as the right runqueue!
1016 * But we don't care, since "task_running()" will
1017 * return false if the runqueue has changed and p
1018 * is actually now running somewhere else!
1020 while (task_running(rq
, p
)) {
1021 if (match_state
&& unlikely(p
->state
!= match_state
))
1027 * Ok, time to look more closely! We need the rq
1028 * lock now, to be *sure*. If we're wrong, we'll
1029 * just go back and repeat.
1031 rq
= task_rq_lock(p
, &flags
);
1032 trace_sched_wait_task(p
);
1033 running
= task_running(rq
, p
);
1036 if (!match_state
|| p
->state
== match_state
)
1037 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1038 task_rq_unlock(rq
, p
, &flags
);
1041 * If it changed from the expected state, bail out now.
1043 if (unlikely(!ncsw
))
1047 * Was it really running after all now that we
1048 * checked with the proper locks actually held?
1050 * Oops. Go back and try again..
1052 if (unlikely(running
)) {
1058 * It's not enough that it's not actively running,
1059 * it must be off the runqueue _entirely_, and not
1062 * So if it was still runnable (but just not actively
1063 * running right now), it's preempted, and we should
1064 * yield - it could be a while.
1066 if (unlikely(on_rq
)) {
1067 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1069 set_current_state(TASK_UNINTERRUPTIBLE
);
1070 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1075 * Ahh, all good. It wasn't running, and it wasn't
1076 * runnable, which means that it will never become
1077 * running in the future either. We're all done!
1086 * kick_process - kick a running thread to enter/exit the kernel
1087 * @p: the to-be-kicked thread
1089 * Cause a process which is running on another CPU to enter
1090 * kernel-mode, without any delay. (to get signals handled.)
1092 * NOTE: this function doesn't have to take the runqueue lock,
1093 * because all it wants to ensure is that the remote task enters
1094 * the kernel. If the IPI races and the task has been migrated
1095 * to another CPU then no harm is done and the purpose has been
1098 void kick_process(struct task_struct
*p
)
1104 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1105 smp_send_reschedule(cpu
);
1108 EXPORT_SYMBOL_GPL(kick_process
);
1109 #endif /* CONFIG_SMP */
1113 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1115 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1117 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1118 enum { cpuset
, possible
, fail
} state
= cpuset
;
1121 /* Look for allowed, online CPU in same node. */
1122 for_each_cpu(dest_cpu
, nodemask
) {
1123 if (!cpu_online(dest_cpu
))
1125 if (!cpu_active(dest_cpu
))
1127 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1132 /* Any allowed, online CPU? */
1133 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1134 if (!cpu_online(dest_cpu
))
1136 if (!cpu_active(dest_cpu
))
1143 /* No more Mr. Nice Guy. */
1144 cpuset_cpus_allowed_fallback(p
);
1149 do_set_cpus_allowed(p
, cpu_possible_mask
);
1160 if (state
!= cpuset
) {
1162 * Don't tell them about moving exiting tasks or
1163 * kernel threads (both mm NULL), since they never
1166 if (p
->mm
&& printk_ratelimit()) {
1167 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1168 task_pid_nr(p
), p
->comm
, cpu
);
1176 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1179 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1181 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1184 * In order not to call set_task_cpu() on a blocking task we need
1185 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1188 * Since this is common to all placement strategies, this lives here.
1190 * [ this allows ->select_task() to simply return task_cpu(p) and
1191 * not worry about this generic constraint ]
1193 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1195 cpu
= select_fallback_rq(task_cpu(p
), p
);
1200 static void update_avg(u64
*avg
, u64 sample
)
1202 s64 diff
= sample
- *avg
;
1208 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1210 #ifdef CONFIG_SCHEDSTATS
1211 struct rq
*rq
= this_rq();
1214 int this_cpu
= smp_processor_id();
1216 if (cpu
== this_cpu
) {
1217 schedstat_inc(rq
, ttwu_local
);
1218 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1220 struct sched_domain
*sd
;
1222 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1224 for_each_domain(this_cpu
, sd
) {
1225 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1226 schedstat_inc(sd
, ttwu_wake_remote
);
1233 if (wake_flags
& WF_MIGRATED
)
1234 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1236 #endif /* CONFIG_SMP */
1238 schedstat_inc(rq
, ttwu_count
);
1239 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1241 if (wake_flags
& WF_SYNC
)
1242 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1244 #endif /* CONFIG_SCHEDSTATS */
1247 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1249 activate_task(rq
, p
, en_flags
);
1252 /* if a worker is waking up, notify workqueue */
1253 if (p
->flags
& PF_WQ_WORKER
)
1254 wq_worker_waking_up(p
, cpu_of(rq
));
1258 * Mark the task runnable and perform wakeup-preemption.
1261 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1263 trace_sched_wakeup(p
, true);
1264 check_preempt_curr(rq
, p
, wake_flags
);
1266 p
->state
= TASK_RUNNING
;
1268 if (p
->sched_class
->task_woken
)
1269 p
->sched_class
->task_woken(rq
, p
);
1271 if (rq
->idle_stamp
) {
1272 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1273 u64 max
= 2*sysctl_sched_migration_cost
;
1278 update_avg(&rq
->avg_idle
, delta
);
1285 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1288 if (p
->sched_contributes_to_load
)
1289 rq
->nr_uninterruptible
--;
1292 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1293 ttwu_do_wakeup(rq
, p
, wake_flags
);
1297 * Called in case the task @p isn't fully descheduled from its runqueue,
1298 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1299 * since all we need to do is flip p->state to TASK_RUNNING, since
1300 * the task is still ->on_rq.
1302 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1307 rq
= __task_rq_lock(p
);
1309 ttwu_do_wakeup(rq
, p
, wake_flags
);
1312 __task_rq_unlock(rq
);
1318 static void sched_ttwu_pending(void)
1320 struct rq
*rq
= this_rq();
1321 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1322 struct task_struct
*p
;
1324 raw_spin_lock(&rq
->lock
);
1327 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1328 llist
= llist_next(llist
);
1329 ttwu_do_activate(rq
, p
, 0);
1332 raw_spin_unlock(&rq
->lock
);
1335 void scheduler_ipi(void)
1337 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1341 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1342 * traditionally all their work was done from the interrupt return
1343 * path. Now that we actually do some work, we need to make sure
1346 * Some archs already do call them, luckily irq_enter/exit nest
1349 * Arguably we should visit all archs and update all handlers,
1350 * however a fair share of IPIs are still resched only so this would
1351 * somewhat pessimize the simple resched case.
1354 sched_ttwu_pending();
1357 * Check if someone kicked us for doing the nohz idle load balance.
1359 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1360 this_rq()->idle_balance
= 1;
1361 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1366 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1368 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1369 smp_send_reschedule(cpu
);
1372 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1374 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1376 #endif /* CONFIG_SMP */
1378 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1380 struct rq
*rq
= cpu_rq(cpu
);
1382 #if defined(CONFIG_SMP)
1383 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1384 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1385 ttwu_queue_remote(p
, cpu
);
1390 raw_spin_lock(&rq
->lock
);
1391 ttwu_do_activate(rq
, p
, 0);
1392 raw_spin_unlock(&rq
->lock
);
1396 * try_to_wake_up - wake up a thread
1397 * @p: the thread to be awakened
1398 * @state: the mask of task states that can be woken
1399 * @wake_flags: wake modifier flags (WF_*)
1401 * Put it on the run-queue if it's not already there. The "current"
1402 * thread is always on the run-queue (except when the actual
1403 * re-schedule is in progress), and as such you're allowed to do
1404 * the simpler "current->state = TASK_RUNNING" to mark yourself
1405 * runnable without the overhead of this.
1407 * Returns %true if @p was woken up, %false if it was already running
1408 * or @state didn't match @p's state.
1411 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1413 unsigned long flags
;
1414 int cpu
, success
= 0;
1417 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1418 if (!(p
->state
& state
))
1421 success
= 1; /* we're going to change ->state */
1424 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1429 * If the owning (remote) cpu is still in the middle of schedule() with
1430 * this task as prev, wait until its done referencing the task.
1435 * Pairs with the smp_wmb() in finish_lock_switch().
1439 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1440 p
->state
= TASK_WAKING
;
1442 if (p
->sched_class
->task_waking
)
1443 p
->sched_class
->task_waking(p
);
1445 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1446 if (task_cpu(p
) != cpu
) {
1447 wake_flags
|= WF_MIGRATED
;
1448 set_task_cpu(p
, cpu
);
1450 #endif /* CONFIG_SMP */
1454 ttwu_stat(p
, cpu
, wake_flags
);
1456 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1462 * try_to_wake_up_local - try to wake up a local task with rq lock held
1463 * @p: the thread to be awakened
1465 * Put @p on the run-queue if it's not already there. The caller must
1466 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1469 static void try_to_wake_up_local(struct task_struct
*p
)
1471 struct rq
*rq
= task_rq(p
);
1473 BUG_ON(rq
!= this_rq());
1474 BUG_ON(p
== current
);
1475 lockdep_assert_held(&rq
->lock
);
1477 if (!raw_spin_trylock(&p
->pi_lock
)) {
1478 raw_spin_unlock(&rq
->lock
);
1479 raw_spin_lock(&p
->pi_lock
);
1480 raw_spin_lock(&rq
->lock
);
1483 if (!(p
->state
& TASK_NORMAL
))
1487 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1489 ttwu_do_wakeup(rq
, p
, 0);
1490 ttwu_stat(p
, smp_processor_id(), 0);
1492 raw_spin_unlock(&p
->pi_lock
);
1496 * wake_up_process - Wake up a specific process
1497 * @p: The process to be woken up.
1499 * Attempt to wake up the nominated process and move it to the set of runnable
1500 * processes. Returns 1 if the process was woken up, 0 if it was already
1503 * It may be assumed that this function implies a write memory barrier before
1504 * changing the task state if and only if any tasks are woken up.
1506 int wake_up_process(struct task_struct
*p
)
1508 return try_to_wake_up(p
, TASK_ALL
, 0);
1510 EXPORT_SYMBOL(wake_up_process
);
1512 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1514 return try_to_wake_up(p
, state
, 0);
1518 * Perform scheduler related setup for a newly forked process p.
1519 * p is forked by current.
1521 * __sched_fork() is basic setup used by init_idle() too:
1523 static void __sched_fork(struct task_struct
*p
)
1528 p
->se
.exec_start
= 0;
1529 p
->se
.sum_exec_runtime
= 0;
1530 p
->se
.prev_sum_exec_runtime
= 0;
1531 p
->se
.nr_migrations
= 0;
1533 INIT_LIST_HEAD(&p
->se
.group_node
);
1535 #ifdef CONFIG_SCHEDSTATS
1536 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1539 INIT_LIST_HEAD(&p
->rt
.run_list
);
1541 #ifdef CONFIG_PREEMPT_NOTIFIERS
1542 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1545 #ifdef CONFIG_NUMA_BALANCING
1546 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1547 p
->mm
->numa_next_scan
= jiffies
;
1548 p
->mm
->numa_next_reset
= jiffies
;
1549 p
->mm
->numa_scan_seq
= 0;
1552 p
->node_stamp
= 0ULL;
1553 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1554 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1555 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1556 p
->numa_work
.next
= &p
->numa_work
;
1557 #endif /* CONFIG_NUMA_BALANCING */
1560 #ifdef CONFIG_NUMA_BALANCING
1561 void set_numabalancing_state(bool enabled
)
1564 sched_feat_set("NUMA");
1566 sched_feat_set("NO_NUMA");
1568 #endif /* CONFIG_NUMA_BALANCING */
1571 * fork()/clone()-time setup:
1573 void sched_fork(struct task_struct
*p
)
1575 unsigned long flags
;
1576 int cpu
= get_cpu();
1580 * We mark the process as running here. This guarantees that
1581 * nobody will actually run it, and a signal or other external
1582 * event cannot wake it up and insert it on the runqueue either.
1584 p
->state
= TASK_RUNNING
;
1587 * Make sure we do not leak PI boosting priority to the child.
1589 p
->prio
= current
->normal_prio
;
1592 * Revert to default priority/policy on fork if requested.
1594 if (unlikely(p
->sched_reset_on_fork
)) {
1595 if (task_has_rt_policy(p
)) {
1596 p
->policy
= SCHED_NORMAL
;
1597 p
->static_prio
= NICE_TO_PRIO(0);
1599 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1600 p
->static_prio
= NICE_TO_PRIO(0);
1602 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1606 * We don't need the reset flag anymore after the fork. It has
1607 * fulfilled its duty:
1609 p
->sched_reset_on_fork
= 0;
1612 if (!rt_prio(p
->prio
))
1613 p
->sched_class
= &fair_sched_class
;
1615 if (p
->sched_class
->task_fork
)
1616 p
->sched_class
->task_fork(p
);
1619 * The child is not yet in the pid-hash so no cgroup attach races,
1620 * and the cgroup is pinned to this child due to cgroup_fork()
1621 * is ran before sched_fork().
1623 * Silence PROVE_RCU.
1625 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1626 set_task_cpu(p
, cpu
);
1627 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1629 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1630 if (likely(sched_info_on()))
1631 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1633 #if defined(CONFIG_SMP)
1636 #ifdef CONFIG_PREEMPT_COUNT
1637 /* Want to start with kernel preemption disabled. */
1638 task_thread_info(p
)->preempt_count
= 1;
1641 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1648 * wake_up_new_task - wake up a newly created task for the first time.
1650 * This function will do some initial scheduler statistics housekeeping
1651 * that must be done for every newly created context, then puts the task
1652 * on the runqueue and wakes it.
1654 void wake_up_new_task(struct task_struct
*p
)
1656 unsigned long flags
;
1659 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1662 * Fork balancing, do it here and not earlier because:
1663 * - cpus_allowed can change in the fork path
1664 * - any previously selected cpu might disappear through hotplug
1666 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1669 rq
= __task_rq_lock(p
);
1670 activate_task(rq
, p
, 0);
1672 trace_sched_wakeup_new(p
, true);
1673 check_preempt_curr(rq
, p
, WF_FORK
);
1675 if (p
->sched_class
->task_woken
)
1676 p
->sched_class
->task_woken(rq
, p
);
1678 task_rq_unlock(rq
, p
, &flags
);
1681 #ifdef CONFIG_PREEMPT_NOTIFIERS
1684 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1685 * @notifier: notifier struct to register
1687 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1689 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1691 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1694 * preempt_notifier_unregister - no longer interested in preemption notifications
1695 * @notifier: notifier struct to unregister
1697 * This is safe to call from within a preemption notifier.
1699 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1701 hlist_del(¬ifier
->link
);
1703 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1705 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1707 struct preempt_notifier
*notifier
;
1708 struct hlist_node
*node
;
1710 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1711 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1715 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1716 struct task_struct
*next
)
1718 struct preempt_notifier
*notifier
;
1719 struct hlist_node
*node
;
1721 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1722 notifier
->ops
->sched_out(notifier
, next
);
1725 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1727 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1732 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1733 struct task_struct
*next
)
1737 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1740 * prepare_task_switch - prepare to switch tasks
1741 * @rq: the runqueue preparing to switch
1742 * @prev: the current task that is being switched out
1743 * @next: the task we are going to switch to.
1745 * This is called with the rq lock held and interrupts off. It must
1746 * be paired with a subsequent finish_task_switch after the context
1749 * prepare_task_switch sets up locking and calls architecture specific
1753 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1754 struct task_struct
*next
)
1756 trace_sched_switch(prev
, next
);
1757 sched_info_switch(prev
, next
);
1758 perf_event_task_sched_out(prev
, next
);
1759 fire_sched_out_preempt_notifiers(prev
, next
);
1760 prepare_lock_switch(rq
, next
);
1761 prepare_arch_switch(next
);
1765 * finish_task_switch - clean up after a task-switch
1766 * @rq: runqueue associated with task-switch
1767 * @prev: the thread we just switched away from.
1769 * finish_task_switch must be called after the context switch, paired
1770 * with a prepare_task_switch call before the context switch.
1771 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1772 * and do any other architecture-specific cleanup actions.
1774 * Note that we may have delayed dropping an mm in context_switch(). If
1775 * so, we finish that here outside of the runqueue lock. (Doing it
1776 * with the lock held can cause deadlocks; see schedule() for
1779 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1780 __releases(rq
->lock
)
1782 struct mm_struct
*mm
= rq
->prev_mm
;
1788 * A task struct has one reference for the use as "current".
1789 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1790 * schedule one last time. The schedule call will never return, and
1791 * the scheduled task must drop that reference.
1792 * The test for TASK_DEAD must occur while the runqueue locks are
1793 * still held, otherwise prev could be scheduled on another cpu, die
1794 * there before we look at prev->state, and then the reference would
1796 * Manfred Spraul <manfred@colorfullife.com>
1798 prev_state
= prev
->state
;
1799 vtime_task_switch(prev
);
1800 finish_arch_switch(prev
);
1801 perf_event_task_sched_in(prev
, current
);
1802 finish_lock_switch(rq
, prev
);
1803 finish_arch_post_lock_switch();
1805 fire_sched_in_preempt_notifiers(current
);
1808 if (unlikely(prev_state
== TASK_DEAD
)) {
1810 * Remove function-return probe instances associated with this
1811 * task and put them back on the free list.
1813 kprobe_flush_task(prev
);
1814 put_task_struct(prev
);
1820 /* assumes rq->lock is held */
1821 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1823 if (prev
->sched_class
->pre_schedule
)
1824 prev
->sched_class
->pre_schedule(rq
, prev
);
1827 /* rq->lock is NOT held, but preemption is disabled */
1828 static inline void post_schedule(struct rq
*rq
)
1830 if (rq
->post_schedule
) {
1831 unsigned long flags
;
1833 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1834 if (rq
->curr
->sched_class
->post_schedule
)
1835 rq
->curr
->sched_class
->post_schedule(rq
);
1836 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1838 rq
->post_schedule
= 0;
1844 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1848 static inline void post_schedule(struct rq
*rq
)
1855 * schedule_tail - first thing a freshly forked thread must call.
1856 * @prev: the thread we just switched away from.
1858 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1859 __releases(rq
->lock
)
1861 struct rq
*rq
= this_rq();
1863 finish_task_switch(rq
, prev
);
1866 * FIXME: do we need to worry about rq being invalidated by the
1871 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1872 /* In this case, finish_task_switch does not reenable preemption */
1875 if (current
->set_child_tid
)
1876 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1880 * context_switch - switch to the new MM and the new
1881 * thread's register state.
1884 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1885 struct task_struct
*next
)
1887 struct mm_struct
*mm
, *oldmm
;
1889 prepare_task_switch(rq
, prev
, next
);
1892 oldmm
= prev
->active_mm
;
1894 * For paravirt, this is coupled with an exit in switch_to to
1895 * combine the page table reload and the switch backend into
1898 arch_start_context_switch(prev
);
1901 next
->active_mm
= oldmm
;
1902 atomic_inc(&oldmm
->mm_count
);
1903 enter_lazy_tlb(oldmm
, next
);
1905 switch_mm(oldmm
, mm
, next
);
1908 prev
->active_mm
= NULL
;
1909 rq
->prev_mm
= oldmm
;
1912 * Since the runqueue lock will be released by the next
1913 * task (which is an invalid locking op but in the case
1914 * of the scheduler it's an obvious special-case), so we
1915 * do an early lockdep release here:
1917 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1918 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1921 /* Here we just switch the register state and the stack. */
1922 rcu_switch(prev
, next
);
1923 switch_to(prev
, next
, prev
);
1927 * this_rq must be evaluated again because prev may have moved
1928 * CPUs since it called schedule(), thus the 'rq' on its stack
1929 * frame will be invalid.
1931 finish_task_switch(this_rq(), prev
);
1935 * nr_running, nr_uninterruptible and nr_context_switches:
1937 * externally visible scheduler statistics: current number of runnable
1938 * threads, current number of uninterruptible-sleeping threads, total
1939 * number of context switches performed since bootup.
1941 unsigned long nr_running(void)
1943 unsigned long i
, sum
= 0;
1945 for_each_online_cpu(i
)
1946 sum
+= cpu_rq(i
)->nr_running
;
1951 unsigned long nr_uninterruptible(void)
1953 unsigned long i
, sum
= 0;
1955 for_each_possible_cpu(i
)
1956 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1959 * Since we read the counters lockless, it might be slightly
1960 * inaccurate. Do not allow it to go below zero though:
1962 if (unlikely((long)sum
< 0))
1968 unsigned long long nr_context_switches(void)
1971 unsigned long long sum
= 0;
1973 for_each_possible_cpu(i
)
1974 sum
+= cpu_rq(i
)->nr_switches
;
1979 unsigned long nr_iowait(void)
1981 unsigned long i
, sum
= 0;
1983 for_each_possible_cpu(i
)
1984 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1989 unsigned long nr_iowait_cpu(int cpu
)
1991 struct rq
*this = cpu_rq(cpu
);
1992 return atomic_read(&this->nr_iowait
);
1995 unsigned long this_cpu_load(void)
1997 struct rq
*this = this_rq();
1998 return this->cpu_load
[0];
2003 * Global load-average calculations
2005 * We take a distributed and async approach to calculating the global load-avg
2006 * in order to minimize overhead.
2008 * The global load average is an exponentially decaying average of nr_running +
2009 * nr_uninterruptible.
2011 * Once every LOAD_FREQ:
2014 * for_each_possible_cpu(cpu)
2015 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2017 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2019 * Due to a number of reasons the above turns in the mess below:
2021 * - for_each_possible_cpu() is prohibitively expensive on machines with
2022 * serious number of cpus, therefore we need to take a distributed approach
2023 * to calculating nr_active.
2025 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2026 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2028 * So assuming nr_active := 0 when we start out -- true per definition, we
2029 * can simply take per-cpu deltas and fold those into a global accumulate
2030 * to obtain the same result. See calc_load_fold_active().
2032 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2033 * across the machine, we assume 10 ticks is sufficient time for every
2034 * cpu to have completed this task.
2036 * This places an upper-bound on the IRQ-off latency of the machine. Then
2037 * again, being late doesn't loose the delta, just wrecks the sample.
2039 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2040 * this would add another cross-cpu cacheline miss and atomic operation
2041 * to the wakeup path. Instead we increment on whatever cpu the task ran
2042 * when it went into uninterruptible state and decrement on whatever cpu
2043 * did the wakeup. This means that only the sum of nr_uninterruptible over
2044 * all cpus yields the correct result.
2046 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2049 /* Variables and functions for calc_load */
2050 static atomic_long_t calc_load_tasks
;
2051 static unsigned long calc_load_update
;
2052 unsigned long avenrun
[3];
2053 EXPORT_SYMBOL(avenrun
); /* should be removed */
2056 * get_avenrun - get the load average array
2057 * @loads: pointer to dest load array
2058 * @offset: offset to add
2059 * @shift: shift count to shift the result left
2061 * These values are estimates at best, so no need for locking.
2063 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2065 loads
[0] = (avenrun
[0] + offset
) << shift
;
2066 loads
[1] = (avenrun
[1] + offset
) << shift
;
2067 loads
[2] = (avenrun
[2] + offset
) << shift
;
2070 static long calc_load_fold_active(struct rq
*this_rq
)
2072 long nr_active
, delta
= 0;
2074 nr_active
= this_rq
->nr_running
;
2075 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2077 if (nr_active
!= this_rq
->calc_load_active
) {
2078 delta
= nr_active
- this_rq
->calc_load_active
;
2079 this_rq
->calc_load_active
= nr_active
;
2086 * a1 = a0 * e + a * (1 - e)
2088 static unsigned long
2089 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2092 load
+= active
* (FIXED_1
- exp
);
2093 load
+= 1UL << (FSHIFT
- 1);
2094 return load
>> FSHIFT
;
2099 * Handle NO_HZ for the global load-average.
2101 * Since the above described distributed algorithm to compute the global
2102 * load-average relies on per-cpu sampling from the tick, it is affected by
2105 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2106 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2107 * when we read the global state.
2109 * Obviously reality has to ruin such a delightfully simple scheme:
2111 * - When we go NO_HZ idle during the window, we can negate our sample
2112 * contribution, causing under-accounting.
2114 * We avoid this by keeping two idle-delta counters and flipping them
2115 * when the window starts, thus separating old and new NO_HZ load.
2117 * The only trick is the slight shift in index flip for read vs write.
2121 * |-|-----------|-|-----------|-|-----------|-|
2122 * r:0 0 1 1 0 0 1 1 0
2123 * w:0 1 1 0 0 1 1 0 0
2125 * This ensures we'll fold the old idle contribution in this window while
2126 * accumlating the new one.
2128 * - When we wake up from NO_HZ idle during the window, we push up our
2129 * contribution, since we effectively move our sample point to a known
2132 * This is solved by pushing the window forward, and thus skipping the
2133 * sample, for this cpu (effectively using the idle-delta for this cpu which
2134 * was in effect at the time the window opened). This also solves the issue
2135 * of having to deal with a cpu having been in NOHZ idle for multiple
2136 * LOAD_FREQ intervals.
2138 * When making the ILB scale, we should try to pull this in as well.
2140 static atomic_long_t calc_load_idle
[2];
2141 static int calc_load_idx
;
2143 static inline int calc_load_write_idx(void)
2145 int idx
= calc_load_idx
;
2148 * See calc_global_nohz(), if we observe the new index, we also
2149 * need to observe the new update time.
2154 * If the folding window started, make sure we start writing in the
2157 if (!time_before(jiffies
, calc_load_update
))
2163 static inline int calc_load_read_idx(void)
2165 return calc_load_idx
& 1;
2168 void calc_load_enter_idle(void)
2170 struct rq
*this_rq
= this_rq();
2174 * We're going into NOHZ mode, if there's any pending delta, fold it
2175 * into the pending idle delta.
2177 delta
= calc_load_fold_active(this_rq
);
2179 int idx
= calc_load_write_idx();
2180 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2184 void calc_load_exit_idle(void)
2186 struct rq
*this_rq
= this_rq();
2189 * If we're still before the sample window, we're done.
2191 if (time_before(jiffies
, this_rq
->calc_load_update
))
2195 * We woke inside or after the sample window, this means we're already
2196 * accounted through the nohz accounting, so skip the entire deal and
2197 * sync up for the next window.
2199 this_rq
->calc_load_update
= calc_load_update
;
2200 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2201 this_rq
->calc_load_update
+= LOAD_FREQ
;
2204 static long calc_load_fold_idle(void)
2206 int idx
= calc_load_read_idx();
2209 if (atomic_long_read(&calc_load_idle
[idx
]))
2210 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2216 * fixed_power_int - compute: x^n, in O(log n) time
2218 * @x: base of the power
2219 * @frac_bits: fractional bits of @x
2220 * @n: power to raise @x to.
2222 * By exploiting the relation between the definition of the natural power
2223 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2224 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2225 * (where: n_i \elem {0, 1}, the binary vector representing n),
2226 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2227 * of course trivially computable in O(log_2 n), the length of our binary
2230 static unsigned long
2231 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2233 unsigned long result
= 1UL << frac_bits
;
2238 result
+= 1UL << (frac_bits
- 1);
2239 result
>>= frac_bits
;
2245 x
+= 1UL << (frac_bits
- 1);
2253 * a1 = a0 * e + a * (1 - e)
2255 * a2 = a1 * e + a * (1 - e)
2256 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2257 * = a0 * e^2 + a * (1 - e) * (1 + e)
2259 * a3 = a2 * e + a * (1 - e)
2260 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2261 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2265 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2266 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2267 * = a0 * e^n + a * (1 - e^n)
2269 * [1] application of the geometric series:
2272 * S_n := \Sum x^i = -------------
2275 static unsigned long
2276 calc_load_n(unsigned long load
, unsigned long exp
,
2277 unsigned long active
, unsigned int n
)
2280 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2284 * NO_HZ can leave us missing all per-cpu ticks calling
2285 * calc_load_account_active(), but since an idle CPU folds its delta into
2286 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2287 * in the pending idle delta if our idle period crossed a load cycle boundary.
2289 * Once we've updated the global active value, we need to apply the exponential
2290 * weights adjusted to the number of cycles missed.
2292 static void calc_global_nohz(void)
2294 long delta
, active
, n
;
2296 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2298 * Catch-up, fold however many we are behind still
2300 delta
= jiffies
- calc_load_update
- 10;
2301 n
= 1 + (delta
/ LOAD_FREQ
);
2303 active
= atomic_long_read(&calc_load_tasks
);
2304 active
= active
> 0 ? active
* FIXED_1
: 0;
2306 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2307 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2308 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2310 calc_load_update
+= n
* LOAD_FREQ
;
2314 * Flip the idle index...
2316 * Make sure we first write the new time then flip the index, so that
2317 * calc_load_write_idx() will see the new time when it reads the new
2318 * index, this avoids a double flip messing things up.
2323 #else /* !CONFIG_NO_HZ */
2325 static inline long calc_load_fold_idle(void) { return 0; }
2326 static inline void calc_global_nohz(void) { }
2328 #endif /* CONFIG_NO_HZ */
2331 * calc_load - update the avenrun load estimates 10 ticks after the
2332 * CPUs have updated calc_load_tasks.
2334 void calc_global_load(unsigned long ticks
)
2338 if (time_before(jiffies
, calc_load_update
+ 10))
2342 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2344 delta
= calc_load_fold_idle();
2346 atomic_long_add(delta
, &calc_load_tasks
);
2348 active
= atomic_long_read(&calc_load_tasks
);
2349 active
= active
> 0 ? active
* FIXED_1
: 0;
2351 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2352 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2353 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2355 calc_load_update
+= LOAD_FREQ
;
2358 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2364 * Called from update_cpu_load() to periodically update this CPU's
2367 static void calc_load_account_active(struct rq
*this_rq
)
2371 if (time_before(jiffies
, this_rq
->calc_load_update
))
2374 delta
= calc_load_fold_active(this_rq
);
2376 atomic_long_add(delta
, &calc_load_tasks
);
2378 this_rq
->calc_load_update
+= LOAD_FREQ
;
2382 * End of global load-average stuff
2386 * The exact cpuload at various idx values, calculated at every tick would be
2387 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2389 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2390 * on nth tick when cpu may be busy, then we have:
2391 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2392 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2394 * decay_load_missed() below does efficient calculation of
2395 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2396 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2398 * The calculation is approximated on a 128 point scale.
2399 * degrade_zero_ticks is the number of ticks after which load at any
2400 * particular idx is approximated to be zero.
2401 * degrade_factor is a precomputed table, a row for each load idx.
2402 * Each column corresponds to degradation factor for a power of two ticks,
2403 * based on 128 point scale.
2405 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2406 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2408 * With this power of 2 load factors, we can degrade the load n times
2409 * by looking at 1 bits in n and doing as many mult/shift instead of
2410 * n mult/shifts needed by the exact degradation.
2412 #define DEGRADE_SHIFT 7
2413 static const unsigned char
2414 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2415 static const unsigned char
2416 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2417 {0, 0, 0, 0, 0, 0, 0, 0},
2418 {64, 32, 8, 0, 0, 0, 0, 0},
2419 {96, 72, 40, 12, 1, 0, 0},
2420 {112, 98, 75, 43, 15, 1, 0},
2421 {120, 112, 98, 76, 45, 16, 2} };
2424 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2425 * would be when CPU is idle and so we just decay the old load without
2426 * adding any new load.
2428 static unsigned long
2429 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2433 if (!missed_updates
)
2436 if (missed_updates
>= degrade_zero_ticks
[idx
])
2440 return load
>> missed_updates
;
2442 while (missed_updates
) {
2443 if (missed_updates
% 2)
2444 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2446 missed_updates
>>= 1;
2453 * Update rq->cpu_load[] statistics. This function is usually called every
2454 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2455 * every tick. We fix it up based on jiffies.
2457 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2458 unsigned long pending_updates
)
2462 this_rq
->nr_load_updates
++;
2464 /* Update our load: */
2465 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2466 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2467 unsigned long old_load
, new_load
;
2469 /* scale is effectively 1 << i now, and >> i divides by scale */
2471 old_load
= this_rq
->cpu_load
[i
];
2472 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2473 new_load
= this_load
;
2475 * Round up the averaging division if load is increasing. This
2476 * prevents us from getting stuck on 9 if the load is 10, for
2479 if (new_load
> old_load
)
2480 new_load
+= scale
- 1;
2482 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2485 sched_avg_update(this_rq
);
2490 * There is no sane way to deal with nohz on smp when using jiffies because the
2491 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2492 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2494 * Therefore we cannot use the delta approach from the regular tick since that
2495 * would seriously skew the load calculation. However we'll make do for those
2496 * updates happening while idle (nohz_idle_balance) or coming out of idle
2497 * (tick_nohz_idle_exit).
2499 * This means we might still be one tick off for nohz periods.
2503 * Called from nohz_idle_balance() to update the load ratings before doing the
2506 void update_idle_cpu_load(struct rq
*this_rq
)
2508 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2509 unsigned long load
= this_rq
->load
.weight
;
2510 unsigned long pending_updates
;
2513 * bail if there's load or we're actually up-to-date.
2515 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2518 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2519 this_rq
->last_load_update_tick
= curr_jiffies
;
2521 __update_cpu_load(this_rq
, load
, pending_updates
);
2525 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2527 void update_cpu_load_nohz(void)
2529 struct rq
*this_rq
= this_rq();
2530 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2531 unsigned long pending_updates
;
2533 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2536 raw_spin_lock(&this_rq
->lock
);
2537 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2538 if (pending_updates
) {
2539 this_rq
->last_load_update_tick
= curr_jiffies
;
2541 * We were idle, this means load 0, the current load might be
2542 * !0 due to remote wakeups and the sort.
2544 __update_cpu_load(this_rq
, 0, pending_updates
);
2546 raw_spin_unlock(&this_rq
->lock
);
2548 #endif /* CONFIG_NO_HZ */
2551 * Called from scheduler_tick()
2553 static void update_cpu_load_active(struct rq
*this_rq
)
2556 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2558 this_rq
->last_load_update_tick
= jiffies
;
2559 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2561 calc_load_account_active(this_rq
);
2567 * sched_exec - execve() is a valuable balancing opportunity, because at
2568 * this point the task has the smallest effective memory and cache footprint.
2570 void sched_exec(void)
2572 struct task_struct
*p
= current
;
2573 unsigned long flags
;
2576 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2577 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2578 if (dest_cpu
== smp_processor_id())
2581 if (likely(cpu_active(dest_cpu
))) {
2582 struct migration_arg arg
= { p
, dest_cpu
};
2584 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2585 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2589 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2594 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2595 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2597 EXPORT_PER_CPU_SYMBOL(kstat
);
2598 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2601 * Return any ns on the sched_clock that have not yet been accounted in
2602 * @p in case that task is currently running.
2604 * Called with task_rq_lock() held on @rq.
2606 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2610 if (task_current(rq
, p
)) {
2611 update_rq_clock(rq
);
2612 ns
= rq
->clock_task
- p
->se
.exec_start
;
2620 unsigned long long task_delta_exec(struct task_struct
*p
)
2622 unsigned long flags
;
2626 rq
= task_rq_lock(p
, &flags
);
2627 ns
= do_task_delta_exec(p
, rq
);
2628 task_rq_unlock(rq
, p
, &flags
);
2634 * Return accounted runtime for the task.
2635 * In case the task is currently running, return the runtime plus current's
2636 * pending runtime that have not been accounted yet.
2638 unsigned long long task_sched_runtime(struct task_struct
*p
)
2640 unsigned long flags
;
2644 rq
= task_rq_lock(p
, &flags
);
2645 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2646 task_rq_unlock(rq
, p
, &flags
);
2652 * This function gets called by the timer code, with HZ frequency.
2653 * We call it with interrupts disabled.
2655 void scheduler_tick(void)
2657 int cpu
= smp_processor_id();
2658 struct rq
*rq
= cpu_rq(cpu
);
2659 struct task_struct
*curr
= rq
->curr
;
2663 raw_spin_lock(&rq
->lock
);
2664 update_rq_clock(rq
);
2665 update_cpu_load_active(rq
);
2666 curr
->sched_class
->task_tick(rq
, curr
, 0);
2667 raw_spin_unlock(&rq
->lock
);
2669 perf_event_task_tick();
2672 rq
->idle_balance
= idle_cpu(cpu
);
2673 trigger_load_balance(rq
, cpu
);
2677 notrace
unsigned long get_parent_ip(unsigned long addr
)
2679 if (in_lock_functions(addr
)) {
2680 addr
= CALLER_ADDR2
;
2681 if (in_lock_functions(addr
))
2682 addr
= CALLER_ADDR3
;
2687 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2688 defined(CONFIG_PREEMPT_TRACER))
2690 void __kprobes
add_preempt_count(int val
)
2692 #ifdef CONFIG_DEBUG_PREEMPT
2696 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2699 preempt_count() += val
;
2700 #ifdef CONFIG_DEBUG_PREEMPT
2702 * Spinlock count overflowing soon?
2704 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2707 if (preempt_count() == val
)
2708 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2710 EXPORT_SYMBOL(add_preempt_count
);
2712 void __kprobes
sub_preempt_count(int val
)
2714 #ifdef CONFIG_DEBUG_PREEMPT
2718 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2721 * Is the spinlock portion underflowing?
2723 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2724 !(preempt_count() & PREEMPT_MASK
)))
2728 if (preempt_count() == val
)
2729 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2730 preempt_count() -= val
;
2732 EXPORT_SYMBOL(sub_preempt_count
);
2737 * Print scheduling while atomic bug:
2739 static noinline
void __schedule_bug(struct task_struct
*prev
)
2741 if (oops_in_progress
)
2744 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2745 prev
->comm
, prev
->pid
, preempt_count());
2747 debug_show_held_locks(prev
);
2749 if (irqs_disabled())
2750 print_irqtrace_events(prev
);
2752 add_taint(TAINT_WARN
);
2756 * Various schedule()-time debugging checks and statistics:
2758 static inline void schedule_debug(struct task_struct
*prev
)
2761 * Test if we are atomic. Since do_exit() needs to call into
2762 * schedule() atomically, we ignore that path for now.
2763 * Otherwise, whine if we are scheduling when we should not be.
2765 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2766 __schedule_bug(prev
);
2769 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2771 schedstat_inc(this_rq(), sched_count
);
2774 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2776 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2777 update_rq_clock(rq
);
2778 prev
->sched_class
->put_prev_task(rq
, prev
);
2782 * Pick up the highest-prio task:
2784 static inline struct task_struct
*
2785 pick_next_task(struct rq
*rq
)
2787 const struct sched_class
*class;
2788 struct task_struct
*p
;
2791 * Optimization: we know that if all tasks are in
2792 * the fair class we can call that function directly:
2794 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2795 p
= fair_sched_class
.pick_next_task(rq
);
2800 for_each_class(class) {
2801 p
= class->pick_next_task(rq
);
2806 BUG(); /* the idle class will always have a runnable task */
2810 * __schedule() is the main scheduler function.
2812 * The main means of driving the scheduler and thus entering this function are:
2814 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2816 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2817 * paths. For example, see arch/x86/entry_64.S.
2819 * To drive preemption between tasks, the scheduler sets the flag in timer
2820 * interrupt handler scheduler_tick().
2822 * 3. Wakeups don't really cause entry into schedule(). They add a
2823 * task to the run-queue and that's it.
2825 * Now, if the new task added to the run-queue preempts the current
2826 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2827 * called on the nearest possible occasion:
2829 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2831 * - in syscall or exception context, at the next outmost
2832 * preempt_enable(). (this might be as soon as the wake_up()'s
2835 * - in IRQ context, return from interrupt-handler to
2836 * preemptible context
2838 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2841 * - cond_resched() call
2842 * - explicit schedule() call
2843 * - return from syscall or exception to user-space
2844 * - return from interrupt-handler to user-space
2846 static void __sched
__schedule(void)
2848 struct task_struct
*prev
, *next
;
2849 unsigned long *switch_count
;
2855 cpu
= smp_processor_id();
2857 rcu_note_context_switch(cpu
);
2860 schedule_debug(prev
);
2862 if (sched_feat(HRTICK
))
2865 raw_spin_lock_irq(&rq
->lock
);
2867 switch_count
= &prev
->nivcsw
;
2868 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2869 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2870 prev
->state
= TASK_RUNNING
;
2872 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2876 * If a worker went to sleep, notify and ask workqueue
2877 * whether it wants to wake up a task to maintain
2880 if (prev
->flags
& PF_WQ_WORKER
) {
2881 struct task_struct
*to_wakeup
;
2883 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2885 try_to_wake_up_local(to_wakeup
);
2888 switch_count
= &prev
->nvcsw
;
2891 pre_schedule(rq
, prev
);
2893 if (unlikely(!rq
->nr_running
))
2894 idle_balance(cpu
, rq
);
2896 put_prev_task(rq
, prev
);
2897 next
= pick_next_task(rq
);
2898 clear_tsk_need_resched(prev
);
2899 rq
->skip_clock_update
= 0;
2901 if (likely(prev
!= next
)) {
2906 context_switch(rq
, prev
, next
); /* unlocks the rq */
2908 * The context switch have flipped the stack from under us
2909 * and restored the local variables which were saved when
2910 * this task called schedule() in the past. prev == current
2911 * is still correct, but it can be moved to another cpu/rq.
2913 cpu
= smp_processor_id();
2916 raw_spin_unlock_irq(&rq
->lock
);
2920 sched_preempt_enable_no_resched();
2925 static inline void sched_submit_work(struct task_struct
*tsk
)
2927 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2930 * If we are going to sleep and we have plugged IO queued,
2931 * make sure to submit it to avoid deadlocks.
2933 if (blk_needs_flush_plug(tsk
))
2934 blk_schedule_flush_plug(tsk
);
2937 asmlinkage
void __sched
schedule(void)
2939 struct task_struct
*tsk
= current
;
2941 sched_submit_work(tsk
);
2944 EXPORT_SYMBOL(schedule
);
2946 #ifdef CONFIG_RCU_USER_QS
2947 asmlinkage
void __sched
schedule_user(void)
2950 * If we come here after a random call to set_need_resched(),
2951 * or we have been woken up remotely but the IPI has not yet arrived,
2952 * we haven't yet exited the RCU idle mode. Do it here manually until
2953 * we find a better solution.
2962 * schedule_preempt_disabled - called with preemption disabled
2964 * Returns with preemption disabled. Note: preempt_count must be 1
2966 void __sched
schedule_preempt_disabled(void)
2968 sched_preempt_enable_no_resched();
2973 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2975 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
2977 if (lock
->owner
!= owner
)
2981 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2982 * lock->owner still matches owner, if that fails, owner might
2983 * point to free()d memory, if it still matches, the rcu_read_lock()
2984 * ensures the memory stays valid.
2988 return owner
->on_cpu
;
2992 * Look out! "owner" is an entirely speculative pointer
2993 * access and not reliable.
2995 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
2997 if (!sched_feat(OWNER_SPIN
))
3001 while (owner_running(lock
, owner
)) {
3005 arch_mutex_cpu_relax();
3010 * We break out the loop above on need_resched() and when the
3011 * owner changed, which is a sign for heavy contention. Return
3012 * success only when lock->owner is NULL.
3014 return lock
->owner
== NULL
;
3018 #ifdef CONFIG_PREEMPT
3020 * this is the entry point to schedule() from in-kernel preemption
3021 * off of preempt_enable. Kernel preemptions off return from interrupt
3022 * occur there and call schedule directly.
3024 asmlinkage
void __sched notrace
preempt_schedule(void)
3026 struct thread_info
*ti
= current_thread_info();
3029 * If there is a non-zero preempt_count or interrupts are disabled,
3030 * we do not want to preempt the current task. Just return..
3032 if (likely(ti
->preempt_count
|| irqs_disabled()))
3036 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3038 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3041 * Check again in case we missed a preemption opportunity
3042 * between schedule and now.
3045 } while (need_resched());
3047 EXPORT_SYMBOL(preempt_schedule
);
3050 * this is the entry point to schedule() from kernel preemption
3051 * off of irq context.
3052 * Note, that this is called and return with irqs disabled. This will
3053 * protect us against recursive calling from irq.
3055 asmlinkage
void __sched
preempt_schedule_irq(void)
3057 struct thread_info
*ti
= current_thread_info();
3059 /* Catch callers which need to be fixed */
3060 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3064 add_preempt_count(PREEMPT_ACTIVE
);
3067 local_irq_disable();
3068 sub_preempt_count(PREEMPT_ACTIVE
);
3071 * Check again in case we missed a preemption opportunity
3072 * between schedule and now.
3075 } while (need_resched());
3078 #endif /* CONFIG_PREEMPT */
3080 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3083 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3085 EXPORT_SYMBOL(default_wake_function
);
3088 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3089 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3090 * number) then we wake all the non-exclusive tasks and one exclusive task.
3092 * There are circumstances in which we can try to wake a task which has already
3093 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3094 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3096 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3097 int nr_exclusive
, int wake_flags
, void *key
)
3099 wait_queue_t
*curr
, *next
;
3101 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3102 unsigned flags
= curr
->flags
;
3104 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3105 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3111 * __wake_up - wake up threads blocked on a waitqueue.
3113 * @mode: which threads
3114 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3115 * @key: is directly passed to the wakeup function
3117 * It may be assumed that this function implies a write memory barrier before
3118 * changing the task state if and only if any tasks are woken up.
3120 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3121 int nr_exclusive
, void *key
)
3123 unsigned long flags
;
3125 spin_lock_irqsave(&q
->lock
, flags
);
3126 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3127 spin_unlock_irqrestore(&q
->lock
, flags
);
3129 EXPORT_SYMBOL(__wake_up
);
3132 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3134 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3136 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3138 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3140 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3142 __wake_up_common(q
, mode
, 1, 0, key
);
3144 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3147 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3149 * @mode: which threads
3150 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3151 * @key: opaque value to be passed to wakeup targets
3153 * The sync wakeup differs that the waker knows that it will schedule
3154 * away soon, so while the target thread will be woken up, it will not
3155 * be migrated to another CPU - ie. the two threads are 'synchronized'
3156 * with each other. This can prevent needless bouncing between CPUs.
3158 * On UP it can prevent extra preemption.
3160 * It may be assumed that this function implies a write memory barrier before
3161 * changing the task state if and only if any tasks are woken up.
3163 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3164 int nr_exclusive
, void *key
)
3166 unsigned long flags
;
3167 int wake_flags
= WF_SYNC
;
3172 if (unlikely(!nr_exclusive
))
3175 spin_lock_irqsave(&q
->lock
, flags
);
3176 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3177 spin_unlock_irqrestore(&q
->lock
, flags
);
3179 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3182 * __wake_up_sync - see __wake_up_sync_key()
3184 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3186 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3188 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3191 * complete: - signals a single thread waiting on this completion
3192 * @x: holds the state of this particular completion
3194 * This will wake up a single thread waiting on this completion. Threads will be
3195 * awakened in the same order in which they were queued.
3197 * See also complete_all(), wait_for_completion() and related routines.
3199 * It may be assumed that this function implies a write memory barrier before
3200 * changing the task state if and only if any tasks are woken up.
3202 void complete(struct completion
*x
)
3204 unsigned long flags
;
3206 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3208 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3209 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3211 EXPORT_SYMBOL(complete
);
3214 * complete_all: - signals all threads waiting on this completion
3215 * @x: holds the state of this particular completion
3217 * This will wake up all threads waiting on this particular completion event.
3219 * It may be assumed that this function implies a write memory barrier before
3220 * changing the task state if and only if any tasks are woken up.
3222 void complete_all(struct completion
*x
)
3224 unsigned long flags
;
3226 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3227 x
->done
+= UINT_MAX
/2;
3228 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3229 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3231 EXPORT_SYMBOL(complete_all
);
3233 static inline long __sched
3234 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3237 DECLARE_WAITQUEUE(wait
, current
);
3239 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3241 if (signal_pending_state(state
, current
)) {
3242 timeout
= -ERESTARTSYS
;
3245 __set_current_state(state
);
3246 spin_unlock_irq(&x
->wait
.lock
);
3247 timeout
= schedule_timeout(timeout
);
3248 spin_lock_irq(&x
->wait
.lock
);
3249 } while (!x
->done
&& timeout
);
3250 __remove_wait_queue(&x
->wait
, &wait
);
3255 return timeout
?: 1;
3259 wait_for_common(struct completion
*x
, long timeout
, int state
)
3263 spin_lock_irq(&x
->wait
.lock
);
3264 timeout
= do_wait_for_common(x
, timeout
, state
);
3265 spin_unlock_irq(&x
->wait
.lock
);
3270 * wait_for_completion: - waits for completion of a task
3271 * @x: holds the state of this particular completion
3273 * This waits to be signaled for completion of a specific task. It is NOT
3274 * interruptible and there is no timeout.
3276 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3277 * and interrupt capability. Also see complete().
3279 void __sched
wait_for_completion(struct completion
*x
)
3281 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3283 EXPORT_SYMBOL(wait_for_completion
);
3286 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3287 * @x: holds the state of this particular completion
3288 * @timeout: timeout value in jiffies
3290 * This waits for either a completion of a specific task to be signaled or for a
3291 * specified timeout to expire. The timeout is in jiffies. It is not
3294 * The return value is 0 if timed out, and positive (at least 1, or number of
3295 * jiffies left till timeout) if completed.
3297 unsigned long __sched
3298 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3300 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3302 EXPORT_SYMBOL(wait_for_completion_timeout
);
3305 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3306 * @x: holds the state of this particular completion
3308 * This waits for completion of a specific task to be signaled. It is
3311 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3313 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3315 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3316 if (t
== -ERESTARTSYS
)
3320 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3323 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3324 * @x: holds the state of this particular completion
3325 * @timeout: timeout value in jiffies
3327 * This waits for either a completion of a specific task to be signaled or for a
3328 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3330 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3331 * positive (at least 1, or number of jiffies left till timeout) if completed.
3334 wait_for_completion_interruptible_timeout(struct completion
*x
,
3335 unsigned long timeout
)
3337 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3339 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3342 * wait_for_completion_killable: - waits for completion of a task (killable)
3343 * @x: holds the state of this particular completion
3345 * This waits to be signaled for completion of a specific task. It can be
3346 * interrupted by a kill signal.
3348 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3350 int __sched
wait_for_completion_killable(struct completion
*x
)
3352 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3353 if (t
== -ERESTARTSYS
)
3357 EXPORT_SYMBOL(wait_for_completion_killable
);
3360 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3361 * @x: holds the state of this particular completion
3362 * @timeout: timeout value in jiffies
3364 * This waits for either a completion of a specific task to be
3365 * signaled or for a specified timeout to expire. It can be
3366 * interrupted by a kill signal. The timeout is in jiffies.
3368 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3369 * positive (at least 1, or number of jiffies left till timeout) if completed.
3372 wait_for_completion_killable_timeout(struct completion
*x
,
3373 unsigned long timeout
)
3375 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3377 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3380 * try_wait_for_completion - try to decrement a completion without blocking
3381 * @x: completion structure
3383 * Returns: 0 if a decrement cannot be done without blocking
3384 * 1 if a decrement succeeded.
3386 * If a completion is being used as a counting completion,
3387 * attempt to decrement the counter without blocking. This
3388 * enables us to avoid waiting if the resource the completion
3389 * is protecting is not available.
3391 bool try_wait_for_completion(struct completion
*x
)
3393 unsigned long flags
;
3396 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3401 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3404 EXPORT_SYMBOL(try_wait_for_completion
);
3407 * completion_done - Test to see if a completion has any waiters
3408 * @x: completion structure
3410 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3411 * 1 if there are no waiters.
3414 bool completion_done(struct completion
*x
)
3416 unsigned long flags
;
3419 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3422 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3425 EXPORT_SYMBOL(completion_done
);
3428 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3430 unsigned long flags
;
3433 init_waitqueue_entry(&wait
, current
);
3435 __set_current_state(state
);
3437 spin_lock_irqsave(&q
->lock
, flags
);
3438 __add_wait_queue(q
, &wait
);
3439 spin_unlock(&q
->lock
);
3440 timeout
= schedule_timeout(timeout
);
3441 spin_lock_irq(&q
->lock
);
3442 __remove_wait_queue(q
, &wait
);
3443 spin_unlock_irqrestore(&q
->lock
, flags
);
3448 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3450 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3452 EXPORT_SYMBOL(interruptible_sleep_on
);
3455 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3457 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3459 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3461 void __sched
sleep_on(wait_queue_head_t
*q
)
3463 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3465 EXPORT_SYMBOL(sleep_on
);
3467 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3469 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3471 EXPORT_SYMBOL(sleep_on_timeout
);
3473 #ifdef CONFIG_RT_MUTEXES
3476 * rt_mutex_setprio - set the current priority of a task
3478 * @prio: prio value (kernel-internal form)
3480 * This function changes the 'effective' priority of a task. It does
3481 * not touch ->normal_prio like __setscheduler().
3483 * Used by the rt_mutex code to implement priority inheritance logic.
3485 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3487 int oldprio
, on_rq
, running
;
3489 const struct sched_class
*prev_class
;
3491 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3493 rq
= __task_rq_lock(p
);
3496 * Idle task boosting is a nono in general. There is one
3497 * exception, when PREEMPT_RT and NOHZ is active:
3499 * The idle task calls get_next_timer_interrupt() and holds
3500 * the timer wheel base->lock on the CPU and another CPU wants
3501 * to access the timer (probably to cancel it). We can safely
3502 * ignore the boosting request, as the idle CPU runs this code
3503 * with interrupts disabled and will complete the lock
3504 * protected section without being interrupted. So there is no
3505 * real need to boost.
3507 if (unlikely(p
== rq
->idle
)) {
3508 WARN_ON(p
!= rq
->curr
);
3509 WARN_ON(p
->pi_blocked_on
);
3513 trace_sched_pi_setprio(p
, prio
);
3515 prev_class
= p
->sched_class
;
3517 running
= task_current(rq
, p
);
3519 dequeue_task(rq
, p
, 0);
3521 p
->sched_class
->put_prev_task(rq
, p
);
3524 p
->sched_class
= &rt_sched_class
;
3526 p
->sched_class
= &fair_sched_class
;
3531 p
->sched_class
->set_curr_task(rq
);
3533 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3535 check_class_changed(rq
, p
, prev_class
, oldprio
);
3537 __task_rq_unlock(rq
);
3540 void set_user_nice(struct task_struct
*p
, long nice
)
3542 int old_prio
, delta
, on_rq
;
3543 unsigned long flags
;
3546 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3549 * We have to be careful, if called from sys_setpriority(),
3550 * the task might be in the middle of scheduling on another CPU.
3552 rq
= task_rq_lock(p
, &flags
);
3554 * The RT priorities are set via sched_setscheduler(), but we still
3555 * allow the 'normal' nice value to be set - but as expected
3556 * it wont have any effect on scheduling until the task is
3557 * SCHED_FIFO/SCHED_RR:
3559 if (task_has_rt_policy(p
)) {
3560 p
->static_prio
= NICE_TO_PRIO(nice
);
3565 dequeue_task(rq
, p
, 0);
3567 p
->static_prio
= NICE_TO_PRIO(nice
);
3570 p
->prio
= effective_prio(p
);
3571 delta
= p
->prio
- old_prio
;
3574 enqueue_task(rq
, p
, 0);
3576 * If the task increased its priority or is running and
3577 * lowered its priority, then reschedule its CPU:
3579 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3580 resched_task(rq
->curr
);
3583 task_rq_unlock(rq
, p
, &flags
);
3585 EXPORT_SYMBOL(set_user_nice
);
3588 * can_nice - check if a task can reduce its nice value
3592 int can_nice(const struct task_struct
*p
, const int nice
)
3594 /* convert nice value [19,-20] to rlimit style value [1,40] */
3595 int nice_rlim
= 20 - nice
;
3597 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3598 capable(CAP_SYS_NICE
));
3601 #ifdef __ARCH_WANT_SYS_NICE
3604 * sys_nice - change the priority of the current process.
3605 * @increment: priority increment
3607 * sys_setpriority is a more generic, but much slower function that
3608 * does similar things.
3610 SYSCALL_DEFINE1(nice
, int, increment
)
3615 * Setpriority might change our priority at the same moment.
3616 * We don't have to worry. Conceptually one call occurs first
3617 * and we have a single winner.
3619 if (increment
< -40)
3624 nice
= TASK_NICE(current
) + increment
;
3630 if (increment
< 0 && !can_nice(current
, nice
))
3633 retval
= security_task_setnice(current
, nice
);
3637 set_user_nice(current
, nice
);
3644 * task_prio - return the priority value of a given task.
3645 * @p: the task in question.
3647 * This is the priority value as seen by users in /proc.
3648 * RT tasks are offset by -200. Normal tasks are centered
3649 * around 0, value goes from -16 to +15.
3651 int task_prio(const struct task_struct
*p
)
3653 return p
->prio
- MAX_RT_PRIO
;
3657 * task_nice - return the nice value of a given task.
3658 * @p: the task in question.
3660 int task_nice(const struct task_struct
*p
)
3662 return TASK_NICE(p
);
3664 EXPORT_SYMBOL(task_nice
);
3667 * idle_cpu - is a given cpu idle currently?
3668 * @cpu: the processor in question.
3670 int idle_cpu(int cpu
)
3672 struct rq
*rq
= cpu_rq(cpu
);
3674 if (rq
->curr
!= rq
->idle
)
3681 if (!llist_empty(&rq
->wake_list
))
3689 * idle_task - return the idle task for a given cpu.
3690 * @cpu: the processor in question.
3692 struct task_struct
*idle_task(int cpu
)
3694 return cpu_rq(cpu
)->idle
;
3698 * find_process_by_pid - find a process with a matching PID value.
3699 * @pid: the pid in question.
3701 static struct task_struct
*find_process_by_pid(pid_t pid
)
3703 return pid
? find_task_by_vpid(pid
) : current
;
3706 /* Actually do priority change: must hold rq lock. */
3708 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3711 p
->rt_priority
= prio
;
3712 p
->normal_prio
= normal_prio(p
);
3713 /* we are holding p->pi_lock already */
3714 p
->prio
= rt_mutex_getprio(p
);
3715 if (rt_prio(p
->prio
))
3716 p
->sched_class
= &rt_sched_class
;
3718 p
->sched_class
= &fair_sched_class
;
3723 * check the target process has a UID that matches the current process's
3725 static bool check_same_owner(struct task_struct
*p
)
3727 const struct cred
*cred
= current_cred(), *pcred
;
3731 pcred
= __task_cred(p
);
3732 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3733 uid_eq(cred
->euid
, pcred
->uid
));
3738 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3739 const struct sched_param
*param
, bool user
)
3741 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3742 unsigned long flags
;
3743 const struct sched_class
*prev_class
;
3747 /* may grab non-irq protected spin_locks */
3748 BUG_ON(in_interrupt());
3750 /* double check policy once rq lock held */
3752 reset_on_fork
= p
->sched_reset_on_fork
;
3753 policy
= oldpolicy
= p
->policy
;
3755 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3756 policy
&= ~SCHED_RESET_ON_FORK
;
3758 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3759 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3760 policy
!= SCHED_IDLE
)
3765 * Valid priorities for SCHED_FIFO and SCHED_RR are
3766 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3767 * SCHED_BATCH and SCHED_IDLE is 0.
3769 if (param
->sched_priority
< 0 ||
3770 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3771 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3773 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3777 * Allow unprivileged RT tasks to decrease priority:
3779 if (user
&& !capable(CAP_SYS_NICE
)) {
3780 if (rt_policy(policy
)) {
3781 unsigned long rlim_rtprio
=
3782 task_rlimit(p
, RLIMIT_RTPRIO
);
3784 /* can't set/change the rt policy */
3785 if (policy
!= p
->policy
&& !rlim_rtprio
)
3788 /* can't increase priority */
3789 if (param
->sched_priority
> p
->rt_priority
&&
3790 param
->sched_priority
> rlim_rtprio
)
3795 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3796 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3798 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3799 if (!can_nice(p
, TASK_NICE(p
)))
3803 /* can't change other user's priorities */
3804 if (!check_same_owner(p
))
3807 /* Normal users shall not reset the sched_reset_on_fork flag */
3808 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3813 retval
= security_task_setscheduler(p
);
3819 * make sure no PI-waiters arrive (or leave) while we are
3820 * changing the priority of the task:
3822 * To be able to change p->policy safely, the appropriate
3823 * runqueue lock must be held.
3825 rq
= task_rq_lock(p
, &flags
);
3828 * Changing the policy of the stop threads its a very bad idea
3830 if (p
== rq
->stop
) {
3831 task_rq_unlock(rq
, p
, &flags
);
3836 * If not changing anything there's no need to proceed further:
3838 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3839 param
->sched_priority
== p
->rt_priority
))) {
3840 task_rq_unlock(rq
, p
, &flags
);
3844 #ifdef CONFIG_RT_GROUP_SCHED
3847 * Do not allow realtime tasks into groups that have no runtime
3850 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3851 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3852 !task_group_is_autogroup(task_group(p
))) {
3853 task_rq_unlock(rq
, p
, &flags
);
3859 /* recheck policy now with rq lock held */
3860 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3861 policy
= oldpolicy
= -1;
3862 task_rq_unlock(rq
, p
, &flags
);
3866 running
= task_current(rq
, p
);
3868 dequeue_task(rq
, p
, 0);
3870 p
->sched_class
->put_prev_task(rq
, p
);
3872 p
->sched_reset_on_fork
= reset_on_fork
;
3875 prev_class
= p
->sched_class
;
3876 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3879 p
->sched_class
->set_curr_task(rq
);
3881 enqueue_task(rq
, p
, 0);
3883 check_class_changed(rq
, p
, prev_class
, oldprio
);
3884 task_rq_unlock(rq
, p
, &flags
);
3886 rt_mutex_adjust_pi(p
);
3892 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3893 * @p: the task in question.
3894 * @policy: new policy.
3895 * @param: structure containing the new RT priority.
3897 * NOTE that the task may be already dead.
3899 int sched_setscheduler(struct task_struct
*p
, int policy
,
3900 const struct sched_param
*param
)
3902 return __sched_setscheduler(p
, policy
, param
, true);
3904 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3907 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3908 * @p: the task in question.
3909 * @policy: new policy.
3910 * @param: structure containing the new RT priority.
3912 * Just like sched_setscheduler, only don't bother checking if the
3913 * current context has permission. For example, this is needed in
3914 * stop_machine(): we create temporary high priority worker threads,
3915 * but our caller might not have that capability.
3917 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3918 const struct sched_param
*param
)
3920 return __sched_setscheduler(p
, policy
, param
, false);
3924 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3926 struct sched_param lparam
;
3927 struct task_struct
*p
;
3930 if (!param
|| pid
< 0)
3932 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3937 p
= find_process_by_pid(pid
);
3939 retval
= sched_setscheduler(p
, policy
, &lparam
);
3946 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3947 * @pid: the pid in question.
3948 * @policy: new policy.
3949 * @param: structure containing the new RT priority.
3951 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3952 struct sched_param __user
*, param
)
3954 /* negative values for policy are not valid */
3958 return do_sched_setscheduler(pid
, policy
, param
);
3962 * sys_sched_setparam - set/change the RT priority of a thread
3963 * @pid: the pid in question.
3964 * @param: structure containing the new RT priority.
3966 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3968 return do_sched_setscheduler(pid
, -1, param
);
3972 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3973 * @pid: the pid in question.
3975 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3977 struct task_struct
*p
;
3985 p
= find_process_by_pid(pid
);
3987 retval
= security_task_getscheduler(p
);
3990 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3997 * sys_sched_getparam - get the RT priority of a thread
3998 * @pid: the pid in question.
3999 * @param: structure containing the RT priority.
4001 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4003 struct sched_param lp
;
4004 struct task_struct
*p
;
4007 if (!param
|| pid
< 0)
4011 p
= find_process_by_pid(pid
);
4016 retval
= security_task_getscheduler(p
);
4020 lp
.sched_priority
= p
->rt_priority
;
4024 * This one might sleep, we cannot do it with a spinlock held ...
4026 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4035 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4037 cpumask_var_t cpus_allowed
, new_mask
;
4038 struct task_struct
*p
;
4044 p
= find_process_by_pid(pid
);
4051 /* Prevent p going away */
4055 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4059 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4061 goto out_free_cpus_allowed
;
4064 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4067 retval
= security_task_setscheduler(p
);
4071 cpuset_cpus_allowed(p
, cpus_allowed
);
4072 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4074 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4077 cpuset_cpus_allowed(p
, cpus_allowed
);
4078 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4080 * We must have raced with a concurrent cpuset
4081 * update. Just reset the cpus_allowed to the
4082 * cpuset's cpus_allowed
4084 cpumask_copy(new_mask
, cpus_allowed
);
4089 free_cpumask_var(new_mask
);
4090 out_free_cpus_allowed
:
4091 free_cpumask_var(cpus_allowed
);
4098 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4099 struct cpumask
*new_mask
)
4101 if (len
< cpumask_size())
4102 cpumask_clear(new_mask
);
4103 else if (len
> cpumask_size())
4104 len
= cpumask_size();
4106 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4110 * sys_sched_setaffinity - set the cpu affinity of a process
4111 * @pid: pid of the process
4112 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4113 * @user_mask_ptr: user-space pointer to the new cpu mask
4115 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4116 unsigned long __user
*, user_mask_ptr
)
4118 cpumask_var_t new_mask
;
4121 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4124 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4126 retval
= sched_setaffinity(pid
, new_mask
);
4127 free_cpumask_var(new_mask
);
4131 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4133 struct task_struct
*p
;
4134 unsigned long flags
;
4141 p
= find_process_by_pid(pid
);
4145 retval
= security_task_getscheduler(p
);
4149 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4150 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4151 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4161 * sys_sched_getaffinity - get the cpu affinity of a process
4162 * @pid: pid of the process
4163 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4164 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4166 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4167 unsigned long __user
*, user_mask_ptr
)
4172 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4174 if (len
& (sizeof(unsigned long)-1))
4177 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4180 ret
= sched_getaffinity(pid
, mask
);
4182 size_t retlen
= min_t(size_t, len
, cpumask_size());
4184 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4189 free_cpumask_var(mask
);
4195 * sys_sched_yield - yield the current processor to other threads.
4197 * This function yields the current CPU to other tasks. If there are no
4198 * other threads running on this CPU then this function will return.
4200 SYSCALL_DEFINE0(sched_yield
)
4202 struct rq
*rq
= this_rq_lock();
4204 schedstat_inc(rq
, yld_count
);
4205 current
->sched_class
->yield_task(rq
);
4208 * Since we are going to call schedule() anyway, there's
4209 * no need to preempt or enable interrupts:
4211 __release(rq
->lock
);
4212 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4213 do_raw_spin_unlock(&rq
->lock
);
4214 sched_preempt_enable_no_resched();
4221 static inline int should_resched(void)
4223 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4226 static void __cond_resched(void)
4228 add_preempt_count(PREEMPT_ACTIVE
);
4230 sub_preempt_count(PREEMPT_ACTIVE
);
4233 int __sched
_cond_resched(void)
4235 if (should_resched()) {
4241 EXPORT_SYMBOL(_cond_resched
);
4244 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4245 * call schedule, and on return reacquire the lock.
4247 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4248 * operations here to prevent schedule() from being called twice (once via
4249 * spin_unlock(), once by hand).
4251 int __cond_resched_lock(spinlock_t
*lock
)
4253 int resched
= should_resched();
4256 lockdep_assert_held(lock
);
4258 if (spin_needbreak(lock
) || resched
) {
4269 EXPORT_SYMBOL(__cond_resched_lock
);
4271 int __sched
__cond_resched_softirq(void)
4273 BUG_ON(!in_softirq());
4275 if (should_resched()) {
4283 EXPORT_SYMBOL(__cond_resched_softirq
);
4286 * yield - yield the current processor to other threads.
4288 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4290 * The scheduler is at all times free to pick the calling task as the most
4291 * eligible task to run, if removing the yield() call from your code breaks
4292 * it, its already broken.
4294 * Typical broken usage is:
4299 * where one assumes that yield() will let 'the other' process run that will
4300 * make event true. If the current task is a SCHED_FIFO task that will never
4301 * happen. Never use yield() as a progress guarantee!!
4303 * If you want to use yield() to wait for something, use wait_event().
4304 * If you want to use yield() to be 'nice' for others, use cond_resched().
4305 * If you still want to use yield(), do not!
4307 void __sched
yield(void)
4309 set_current_state(TASK_RUNNING
);
4312 EXPORT_SYMBOL(yield
);
4315 * yield_to - yield the current processor to another thread in
4316 * your thread group, or accelerate that thread toward the
4317 * processor it's on.
4319 * @preempt: whether task preemption is allowed or not
4321 * It's the caller's job to ensure that the target task struct
4322 * can't go away on us before we can do any checks.
4324 * Returns true if we indeed boosted the target task.
4326 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4328 struct task_struct
*curr
= current
;
4329 struct rq
*rq
, *p_rq
;
4330 unsigned long flags
;
4333 local_irq_save(flags
);
4338 double_rq_lock(rq
, p_rq
);
4339 while (task_rq(p
) != p_rq
) {
4340 double_rq_unlock(rq
, p_rq
);
4344 if (!curr
->sched_class
->yield_to_task
)
4347 if (curr
->sched_class
!= p
->sched_class
)
4350 if (task_running(p_rq
, p
) || p
->state
)
4353 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4355 schedstat_inc(rq
, yld_count
);
4357 * Make p's CPU reschedule; pick_next_entity takes care of
4360 if (preempt
&& rq
!= p_rq
)
4361 resched_task(p_rq
->curr
);
4365 double_rq_unlock(rq
, p_rq
);
4366 local_irq_restore(flags
);
4373 EXPORT_SYMBOL_GPL(yield_to
);
4376 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4377 * that process accounting knows that this is a task in IO wait state.
4379 void __sched
io_schedule(void)
4381 struct rq
*rq
= raw_rq();
4383 delayacct_blkio_start();
4384 atomic_inc(&rq
->nr_iowait
);
4385 blk_flush_plug(current
);
4386 current
->in_iowait
= 1;
4388 current
->in_iowait
= 0;
4389 atomic_dec(&rq
->nr_iowait
);
4390 delayacct_blkio_end();
4392 EXPORT_SYMBOL(io_schedule
);
4394 long __sched
io_schedule_timeout(long timeout
)
4396 struct rq
*rq
= raw_rq();
4399 delayacct_blkio_start();
4400 atomic_inc(&rq
->nr_iowait
);
4401 blk_flush_plug(current
);
4402 current
->in_iowait
= 1;
4403 ret
= schedule_timeout(timeout
);
4404 current
->in_iowait
= 0;
4405 atomic_dec(&rq
->nr_iowait
);
4406 delayacct_blkio_end();
4411 * sys_sched_get_priority_max - return maximum RT priority.
4412 * @policy: scheduling class.
4414 * this syscall returns the maximum rt_priority that can be used
4415 * by a given scheduling class.
4417 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4424 ret
= MAX_USER_RT_PRIO
-1;
4436 * sys_sched_get_priority_min - return minimum RT priority.
4437 * @policy: scheduling class.
4439 * this syscall returns the minimum rt_priority that can be used
4440 * by a given scheduling class.
4442 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4460 * sys_sched_rr_get_interval - return the default timeslice of a process.
4461 * @pid: pid of the process.
4462 * @interval: userspace pointer to the timeslice value.
4464 * this syscall writes the default timeslice value of a given process
4465 * into the user-space timespec buffer. A value of '0' means infinity.
4467 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4468 struct timespec __user
*, interval
)
4470 struct task_struct
*p
;
4471 unsigned int time_slice
;
4472 unsigned long flags
;
4482 p
= find_process_by_pid(pid
);
4486 retval
= security_task_getscheduler(p
);
4490 rq
= task_rq_lock(p
, &flags
);
4491 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4492 task_rq_unlock(rq
, p
, &flags
);
4495 jiffies_to_timespec(time_slice
, &t
);
4496 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4504 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4506 void sched_show_task(struct task_struct
*p
)
4508 unsigned long free
= 0;
4511 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4512 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4513 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4514 #if BITS_PER_LONG == 32
4515 if (state
== TASK_RUNNING
)
4516 printk(KERN_CONT
" running ");
4518 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4520 if (state
== TASK_RUNNING
)
4521 printk(KERN_CONT
" running task ");
4523 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4525 #ifdef CONFIG_DEBUG_STACK_USAGE
4526 free
= stack_not_used(p
);
4528 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4529 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4530 (unsigned long)task_thread_info(p
)->flags
);
4532 show_stack(p
, NULL
);
4535 void show_state_filter(unsigned long state_filter
)
4537 struct task_struct
*g
, *p
;
4539 #if BITS_PER_LONG == 32
4541 " task PC stack pid father\n");
4544 " task PC stack pid father\n");
4547 do_each_thread(g
, p
) {
4549 * reset the NMI-timeout, listing all files on a slow
4550 * console might take a lot of time:
4552 touch_nmi_watchdog();
4553 if (!state_filter
|| (p
->state
& state_filter
))
4555 } while_each_thread(g
, p
);
4557 touch_all_softlockup_watchdogs();
4559 #ifdef CONFIG_SCHED_DEBUG
4560 sysrq_sched_debug_show();
4564 * Only show locks if all tasks are dumped:
4567 debug_show_all_locks();
4570 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4572 idle
->sched_class
= &idle_sched_class
;
4576 * init_idle - set up an idle thread for a given CPU
4577 * @idle: task in question
4578 * @cpu: cpu the idle task belongs to
4580 * NOTE: this function does not set the idle thread's NEED_RESCHED
4581 * flag, to make booting more robust.
4583 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4585 struct rq
*rq
= cpu_rq(cpu
);
4586 unsigned long flags
;
4588 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4591 idle
->state
= TASK_RUNNING
;
4592 idle
->se
.exec_start
= sched_clock();
4594 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4596 * We're having a chicken and egg problem, even though we are
4597 * holding rq->lock, the cpu isn't yet set to this cpu so the
4598 * lockdep check in task_group() will fail.
4600 * Similar case to sched_fork(). / Alternatively we could
4601 * use task_rq_lock() here and obtain the other rq->lock.
4606 __set_task_cpu(idle
, cpu
);
4609 rq
->curr
= rq
->idle
= idle
;
4610 #if defined(CONFIG_SMP)
4613 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4615 /* Set the preempt count _outside_ the spinlocks! */
4616 task_thread_info(idle
)->preempt_count
= 0;
4619 * The idle tasks have their own, simple scheduling class:
4621 idle
->sched_class
= &idle_sched_class
;
4622 ftrace_graph_init_idle_task(idle
, cpu
);
4623 #if defined(CONFIG_SMP)
4624 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4629 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4631 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4632 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4634 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4635 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4639 * This is how migration works:
4641 * 1) we invoke migration_cpu_stop() on the target CPU using
4643 * 2) stopper starts to run (implicitly forcing the migrated thread
4645 * 3) it checks whether the migrated task is still in the wrong runqueue.
4646 * 4) if it's in the wrong runqueue then the migration thread removes
4647 * it and puts it into the right queue.
4648 * 5) stopper completes and stop_one_cpu() returns and the migration
4653 * Change a given task's CPU affinity. Migrate the thread to a
4654 * proper CPU and schedule it away if the CPU it's executing on
4655 * is removed from the allowed bitmask.
4657 * NOTE: the caller must have a valid reference to the task, the
4658 * task must not exit() & deallocate itself prematurely. The
4659 * call is not atomic; no spinlocks may be held.
4661 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4663 unsigned long flags
;
4665 unsigned int dest_cpu
;
4668 rq
= task_rq_lock(p
, &flags
);
4670 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4673 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4678 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4683 do_set_cpus_allowed(p
, new_mask
);
4685 /* Can the task run on the task's current CPU? If so, we're done */
4686 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4689 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4691 struct migration_arg arg
= { p
, dest_cpu
};
4692 /* Need help from migration thread: drop lock and wait. */
4693 task_rq_unlock(rq
, p
, &flags
);
4694 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4695 tlb_migrate_finish(p
->mm
);
4699 task_rq_unlock(rq
, p
, &flags
);
4703 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4706 * Move (not current) task off this cpu, onto dest cpu. We're doing
4707 * this because either it can't run here any more (set_cpus_allowed()
4708 * away from this CPU, or CPU going down), or because we're
4709 * attempting to rebalance this task on exec (sched_exec).
4711 * So we race with normal scheduler movements, but that's OK, as long
4712 * as the task is no longer on this CPU.
4714 * Returns non-zero if task was successfully migrated.
4716 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4718 struct rq
*rq_dest
, *rq_src
;
4721 if (unlikely(!cpu_active(dest_cpu
)))
4724 rq_src
= cpu_rq(src_cpu
);
4725 rq_dest
= cpu_rq(dest_cpu
);
4727 raw_spin_lock(&p
->pi_lock
);
4728 double_rq_lock(rq_src
, rq_dest
);
4729 /* Already moved. */
4730 if (task_cpu(p
) != src_cpu
)
4732 /* Affinity changed (again). */
4733 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4737 * If we're not on a rq, the next wake-up will ensure we're
4741 dequeue_task(rq_src
, p
, 0);
4742 set_task_cpu(p
, dest_cpu
);
4743 enqueue_task(rq_dest
, p
, 0);
4744 check_preempt_curr(rq_dest
, p
, 0);
4749 double_rq_unlock(rq_src
, rq_dest
);
4750 raw_spin_unlock(&p
->pi_lock
);
4755 * migration_cpu_stop - this will be executed by a highprio stopper thread
4756 * and performs thread migration by bumping thread off CPU then
4757 * 'pushing' onto another runqueue.
4759 static int migration_cpu_stop(void *data
)
4761 struct migration_arg
*arg
= data
;
4764 * The original target cpu might have gone down and we might
4765 * be on another cpu but it doesn't matter.
4767 local_irq_disable();
4768 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4773 #ifdef CONFIG_HOTPLUG_CPU
4776 * Ensures that the idle task is using init_mm right before its cpu goes
4779 void idle_task_exit(void)
4781 struct mm_struct
*mm
= current
->active_mm
;
4783 BUG_ON(cpu_online(smp_processor_id()));
4786 switch_mm(mm
, &init_mm
, current
);
4791 * Since this CPU is going 'away' for a while, fold any nr_active delta
4792 * we might have. Assumes we're called after migrate_tasks() so that the
4793 * nr_active count is stable.
4795 * Also see the comment "Global load-average calculations".
4797 static void calc_load_migrate(struct rq
*rq
)
4799 long delta
= calc_load_fold_active(rq
);
4801 atomic_long_add(delta
, &calc_load_tasks
);
4805 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4806 * try_to_wake_up()->select_task_rq().
4808 * Called with rq->lock held even though we'er in stop_machine() and
4809 * there's no concurrency possible, we hold the required locks anyway
4810 * because of lock validation efforts.
4812 static void migrate_tasks(unsigned int dead_cpu
)
4814 struct rq
*rq
= cpu_rq(dead_cpu
);
4815 struct task_struct
*next
, *stop
= rq
->stop
;
4819 * Fudge the rq selection such that the below task selection loop
4820 * doesn't get stuck on the currently eligible stop task.
4822 * We're currently inside stop_machine() and the rq is either stuck
4823 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4824 * either way we should never end up calling schedule() until we're
4831 * There's this thread running, bail when that's the only
4834 if (rq
->nr_running
== 1)
4837 next
= pick_next_task(rq
);
4839 next
->sched_class
->put_prev_task(rq
, next
);
4841 /* Find suitable destination for @next, with force if needed. */
4842 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4843 raw_spin_unlock(&rq
->lock
);
4845 __migrate_task(next
, dead_cpu
, dest_cpu
);
4847 raw_spin_lock(&rq
->lock
);
4853 #endif /* CONFIG_HOTPLUG_CPU */
4855 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4857 static struct ctl_table sd_ctl_dir
[] = {
4859 .procname
= "sched_domain",
4865 static struct ctl_table sd_ctl_root
[] = {
4867 .procname
= "kernel",
4869 .child
= sd_ctl_dir
,
4874 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4876 struct ctl_table
*entry
=
4877 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4882 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4884 struct ctl_table
*entry
;
4887 * In the intermediate directories, both the child directory and
4888 * procname are dynamically allocated and could fail but the mode
4889 * will always be set. In the lowest directory the names are
4890 * static strings and all have proc handlers.
4892 for (entry
= *tablep
; entry
->mode
; entry
++) {
4894 sd_free_ctl_entry(&entry
->child
);
4895 if (entry
->proc_handler
== NULL
)
4896 kfree(entry
->procname
);
4903 static int min_load_idx
= 0;
4904 static int max_load_idx
= CPU_LOAD_IDX_MAX
;
4907 set_table_entry(struct ctl_table
*entry
,
4908 const char *procname
, void *data
, int maxlen
,
4909 umode_t mode
, proc_handler
*proc_handler
,
4912 entry
->procname
= procname
;
4914 entry
->maxlen
= maxlen
;
4916 entry
->proc_handler
= proc_handler
;
4919 entry
->extra1
= &min_load_idx
;
4920 entry
->extra2
= &max_load_idx
;
4924 static struct ctl_table
*
4925 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4927 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4932 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4933 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4934 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4935 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4936 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4937 sizeof(int), 0644, proc_dointvec_minmax
, true);
4938 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4939 sizeof(int), 0644, proc_dointvec_minmax
, true);
4940 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4941 sizeof(int), 0644, proc_dointvec_minmax
, true);
4942 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4943 sizeof(int), 0644, proc_dointvec_minmax
, true);
4944 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4945 sizeof(int), 0644, proc_dointvec_minmax
, true);
4946 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4947 sizeof(int), 0644, proc_dointvec_minmax
, false);
4948 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4949 sizeof(int), 0644, proc_dointvec_minmax
, false);
4950 set_table_entry(&table
[9], "cache_nice_tries",
4951 &sd
->cache_nice_tries
,
4952 sizeof(int), 0644, proc_dointvec_minmax
, false);
4953 set_table_entry(&table
[10], "flags", &sd
->flags
,
4954 sizeof(int), 0644, proc_dointvec_minmax
, false);
4955 set_table_entry(&table
[11], "name", sd
->name
,
4956 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4957 /* &table[12] is terminator */
4962 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4964 struct ctl_table
*entry
, *table
;
4965 struct sched_domain
*sd
;
4966 int domain_num
= 0, i
;
4969 for_each_domain(cpu
, sd
)
4971 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4976 for_each_domain(cpu
, sd
) {
4977 snprintf(buf
, 32, "domain%d", i
);
4978 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4980 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4987 static struct ctl_table_header
*sd_sysctl_header
;
4988 static void register_sched_domain_sysctl(void)
4990 int i
, cpu_num
= num_possible_cpus();
4991 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4994 WARN_ON(sd_ctl_dir
[0].child
);
4995 sd_ctl_dir
[0].child
= entry
;
5000 for_each_possible_cpu(i
) {
5001 snprintf(buf
, 32, "cpu%d", i
);
5002 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5004 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5008 WARN_ON(sd_sysctl_header
);
5009 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5012 /* may be called multiple times per register */
5013 static void unregister_sched_domain_sysctl(void)
5015 if (sd_sysctl_header
)
5016 unregister_sysctl_table(sd_sysctl_header
);
5017 sd_sysctl_header
= NULL
;
5018 if (sd_ctl_dir
[0].child
)
5019 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5022 static void register_sched_domain_sysctl(void)
5025 static void unregister_sched_domain_sysctl(void)
5030 static void set_rq_online(struct rq
*rq
)
5033 const struct sched_class
*class;
5035 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5038 for_each_class(class) {
5039 if (class->rq_online
)
5040 class->rq_online(rq
);
5045 static void set_rq_offline(struct rq
*rq
)
5048 const struct sched_class
*class;
5050 for_each_class(class) {
5051 if (class->rq_offline
)
5052 class->rq_offline(rq
);
5055 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5061 * migration_call - callback that gets triggered when a CPU is added.
5062 * Here we can start up the necessary migration thread for the new CPU.
5064 static int __cpuinit
5065 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5067 int cpu
= (long)hcpu
;
5068 unsigned long flags
;
5069 struct rq
*rq
= cpu_rq(cpu
);
5071 switch (action
& ~CPU_TASKS_FROZEN
) {
5073 case CPU_UP_PREPARE
:
5074 rq
->calc_load_update
= calc_load_update
;
5078 /* Update our root-domain */
5079 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5081 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5085 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5088 #ifdef CONFIG_HOTPLUG_CPU
5090 sched_ttwu_pending();
5091 /* Update our root-domain */
5092 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5094 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5098 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5099 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5103 calc_load_migrate(rq
);
5108 update_max_interval();
5114 * Register at high priority so that task migration (migrate_all_tasks)
5115 * happens before everything else. This has to be lower priority than
5116 * the notifier in the perf_event subsystem, though.
5118 static struct notifier_block __cpuinitdata migration_notifier
= {
5119 .notifier_call
= migration_call
,
5120 .priority
= CPU_PRI_MIGRATION
,
5123 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5124 unsigned long action
, void *hcpu
)
5126 switch (action
& ~CPU_TASKS_FROZEN
) {
5128 case CPU_DOWN_FAILED
:
5129 set_cpu_active((long)hcpu
, true);
5136 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5137 unsigned long action
, void *hcpu
)
5139 switch (action
& ~CPU_TASKS_FROZEN
) {
5140 case CPU_DOWN_PREPARE
:
5141 set_cpu_active((long)hcpu
, false);
5148 static int __init
migration_init(void)
5150 void *cpu
= (void *)(long)smp_processor_id();
5153 /* Initialize migration for the boot CPU */
5154 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5155 BUG_ON(err
== NOTIFY_BAD
);
5156 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5157 register_cpu_notifier(&migration_notifier
);
5159 /* Register cpu active notifiers */
5160 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5161 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5165 early_initcall(migration_init
);
5170 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5172 #ifdef CONFIG_SCHED_DEBUG
5174 static __read_mostly
int sched_debug_enabled
;
5176 static int __init
sched_debug_setup(char *str
)
5178 sched_debug_enabled
= 1;
5182 early_param("sched_debug", sched_debug_setup
);
5184 static inline bool sched_debug(void)
5186 return sched_debug_enabled
;
5189 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5190 struct cpumask
*groupmask
)
5192 struct sched_group
*group
= sd
->groups
;
5195 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5196 cpumask_clear(groupmask
);
5198 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5200 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5201 printk("does not load-balance\n");
5203 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5208 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5210 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5211 printk(KERN_ERR
"ERROR: domain->span does not contain "
5214 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5215 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5219 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5223 printk(KERN_ERR
"ERROR: group is NULL\n");
5228 * Even though we initialize ->power to something semi-sane,
5229 * we leave power_orig unset. This allows us to detect if
5230 * domain iteration is still funny without causing /0 traps.
5232 if (!group
->sgp
->power_orig
) {
5233 printk(KERN_CONT
"\n");
5234 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5239 if (!cpumask_weight(sched_group_cpus(group
))) {
5240 printk(KERN_CONT
"\n");
5241 printk(KERN_ERR
"ERROR: empty group\n");
5245 if (!(sd
->flags
& SD_OVERLAP
) &&
5246 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5247 printk(KERN_CONT
"\n");
5248 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5252 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5254 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5256 printk(KERN_CONT
" %s", str
);
5257 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5258 printk(KERN_CONT
" (cpu_power = %d)",
5262 group
= group
->next
;
5263 } while (group
!= sd
->groups
);
5264 printk(KERN_CONT
"\n");
5266 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5267 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5270 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5271 printk(KERN_ERR
"ERROR: parent span is not a superset "
5272 "of domain->span\n");
5276 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5280 if (!sched_debug_enabled
)
5284 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5288 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5291 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5299 #else /* !CONFIG_SCHED_DEBUG */
5300 # define sched_domain_debug(sd, cpu) do { } while (0)
5301 static inline bool sched_debug(void)
5305 #endif /* CONFIG_SCHED_DEBUG */
5307 static int sd_degenerate(struct sched_domain
*sd
)
5309 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5312 /* Following flags need at least 2 groups */
5313 if (sd
->flags
& (SD_LOAD_BALANCE
|
5314 SD_BALANCE_NEWIDLE
|
5318 SD_SHARE_PKG_RESOURCES
)) {
5319 if (sd
->groups
!= sd
->groups
->next
)
5323 /* Following flags don't use groups */
5324 if (sd
->flags
& (SD_WAKE_AFFINE
))
5331 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5333 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5335 if (sd_degenerate(parent
))
5338 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5341 /* Flags needing groups don't count if only 1 group in parent */
5342 if (parent
->groups
== parent
->groups
->next
) {
5343 pflags
&= ~(SD_LOAD_BALANCE
|
5344 SD_BALANCE_NEWIDLE
|
5348 SD_SHARE_PKG_RESOURCES
);
5349 if (nr_node_ids
== 1)
5350 pflags
&= ~SD_SERIALIZE
;
5352 if (~cflags
& pflags
)
5358 static void free_rootdomain(struct rcu_head
*rcu
)
5360 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5362 cpupri_cleanup(&rd
->cpupri
);
5363 free_cpumask_var(rd
->rto_mask
);
5364 free_cpumask_var(rd
->online
);
5365 free_cpumask_var(rd
->span
);
5369 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5371 struct root_domain
*old_rd
= NULL
;
5372 unsigned long flags
;
5374 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5379 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5382 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5385 * If we dont want to free the old_rt yet then
5386 * set old_rd to NULL to skip the freeing later
5389 if (!atomic_dec_and_test(&old_rd
->refcount
))
5393 atomic_inc(&rd
->refcount
);
5396 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5397 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5400 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5403 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5406 static int init_rootdomain(struct root_domain
*rd
)
5408 memset(rd
, 0, sizeof(*rd
));
5410 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5412 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5414 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5417 if (cpupri_init(&rd
->cpupri
) != 0)
5422 free_cpumask_var(rd
->rto_mask
);
5424 free_cpumask_var(rd
->online
);
5426 free_cpumask_var(rd
->span
);
5432 * By default the system creates a single root-domain with all cpus as
5433 * members (mimicking the global state we have today).
5435 struct root_domain def_root_domain
;
5437 static void init_defrootdomain(void)
5439 init_rootdomain(&def_root_domain
);
5441 atomic_set(&def_root_domain
.refcount
, 1);
5444 static struct root_domain
*alloc_rootdomain(void)
5446 struct root_domain
*rd
;
5448 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5452 if (init_rootdomain(rd
) != 0) {
5460 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5462 struct sched_group
*tmp
, *first
;
5471 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5476 } while (sg
!= first
);
5479 static void free_sched_domain(struct rcu_head
*rcu
)
5481 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5484 * If its an overlapping domain it has private groups, iterate and
5487 if (sd
->flags
& SD_OVERLAP
) {
5488 free_sched_groups(sd
->groups
, 1);
5489 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5490 kfree(sd
->groups
->sgp
);
5496 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5498 call_rcu(&sd
->rcu
, free_sched_domain
);
5501 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5503 for (; sd
; sd
= sd
->parent
)
5504 destroy_sched_domain(sd
, cpu
);
5508 * Keep a special pointer to the highest sched_domain that has
5509 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5510 * allows us to avoid some pointer chasing select_idle_sibling().
5512 * Also keep a unique ID per domain (we use the first cpu number in
5513 * the cpumask of the domain), this allows us to quickly tell if
5514 * two cpus are in the same cache domain, see cpus_share_cache().
5516 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5517 DEFINE_PER_CPU(int, sd_llc_id
);
5519 static void update_top_cache_domain(int cpu
)
5521 struct sched_domain
*sd
;
5524 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5526 id
= cpumask_first(sched_domain_span(sd
));
5528 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5529 per_cpu(sd_llc_id
, cpu
) = id
;
5533 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5534 * hold the hotplug lock.
5537 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5539 struct rq
*rq
= cpu_rq(cpu
);
5540 struct sched_domain
*tmp
;
5542 /* Remove the sched domains which do not contribute to scheduling. */
5543 for (tmp
= sd
; tmp
; ) {
5544 struct sched_domain
*parent
= tmp
->parent
;
5548 if (sd_parent_degenerate(tmp
, parent
)) {
5549 tmp
->parent
= parent
->parent
;
5551 parent
->parent
->child
= tmp
;
5552 destroy_sched_domain(parent
, cpu
);
5557 if (sd
&& sd_degenerate(sd
)) {
5560 destroy_sched_domain(tmp
, cpu
);
5565 sched_domain_debug(sd
, cpu
);
5567 rq_attach_root(rq
, rd
);
5569 rcu_assign_pointer(rq
->sd
, sd
);
5570 destroy_sched_domains(tmp
, cpu
);
5572 update_top_cache_domain(cpu
);
5575 /* cpus with isolated domains */
5576 static cpumask_var_t cpu_isolated_map
;
5578 /* Setup the mask of cpus configured for isolated domains */
5579 static int __init
isolated_cpu_setup(char *str
)
5581 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5582 cpulist_parse(str
, cpu_isolated_map
);
5586 __setup("isolcpus=", isolated_cpu_setup
);
5588 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5590 return cpumask_of_node(cpu_to_node(cpu
));
5594 struct sched_domain
**__percpu sd
;
5595 struct sched_group
**__percpu sg
;
5596 struct sched_group_power
**__percpu sgp
;
5600 struct sched_domain
** __percpu sd
;
5601 struct root_domain
*rd
;
5611 struct sched_domain_topology_level
;
5613 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5614 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5616 #define SDTL_OVERLAP 0x01
5618 struct sched_domain_topology_level
{
5619 sched_domain_init_f init
;
5620 sched_domain_mask_f mask
;
5623 struct sd_data data
;
5627 * Build an iteration mask that can exclude certain CPUs from the upwards
5630 * Asymmetric node setups can result in situations where the domain tree is of
5631 * unequal depth, make sure to skip domains that already cover the entire
5634 * In that case build_sched_domains() will have terminated the iteration early
5635 * and our sibling sd spans will be empty. Domains should always include the
5636 * cpu they're built on, so check that.
5639 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5641 const struct cpumask
*span
= sched_domain_span(sd
);
5642 struct sd_data
*sdd
= sd
->private;
5643 struct sched_domain
*sibling
;
5646 for_each_cpu(i
, span
) {
5647 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5648 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5651 cpumask_set_cpu(i
, sched_group_mask(sg
));
5656 * Return the canonical balance cpu for this group, this is the first cpu
5657 * of this group that's also in the iteration mask.
5659 int group_balance_cpu(struct sched_group
*sg
)
5661 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5665 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5667 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5668 const struct cpumask
*span
= sched_domain_span(sd
);
5669 struct cpumask
*covered
= sched_domains_tmpmask
;
5670 struct sd_data
*sdd
= sd
->private;
5671 struct sched_domain
*child
;
5674 cpumask_clear(covered
);
5676 for_each_cpu(i
, span
) {
5677 struct cpumask
*sg_span
;
5679 if (cpumask_test_cpu(i
, covered
))
5682 child
= *per_cpu_ptr(sdd
->sd
, i
);
5684 /* See the comment near build_group_mask(). */
5685 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5688 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5689 GFP_KERNEL
, cpu_to_node(cpu
));
5694 sg_span
= sched_group_cpus(sg
);
5696 child
= child
->child
;
5697 cpumask_copy(sg_span
, sched_domain_span(child
));
5699 cpumask_set_cpu(i
, sg_span
);
5701 cpumask_or(covered
, covered
, sg_span
);
5703 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5704 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5705 build_group_mask(sd
, sg
);
5708 * Initialize sgp->power such that even if we mess up the
5709 * domains and no possible iteration will get us here, we won't
5712 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5715 * Make sure the first group of this domain contains the
5716 * canonical balance cpu. Otherwise the sched_domain iteration
5717 * breaks. See update_sg_lb_stats().
5719 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5720 group_balance_cpu(sg
) == cpu
)
5730 sd
->groups
= groups
;
5735 free_sched_groups(first
, 0);
5740 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5742 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5743 struct sched_domain
*child
= sd
->child
;
5746 cpu
= cpumask_first(sched_domain_span(child
));
5749 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5750 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5751 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5758 * build_sched_groups will build a circular linked list of the groups
5759 * covered by the given span, and will set each group's ->cpumask correctly,
5760 * and ->cpu_power to 0.
5762 * Assumes the sched_domain tree is fully constructed
5765 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5767 struct sched_group
*first
= NULL
, *last
= NULL
;
5768 struct sd_data
*sdd
= sd
->private;
5769 const struct cpumask
*span
= sched_domain_span(sd
);
5770 struct cpumask
*covered
;
5773 get_group(cpu
, sdd
, &sd
->groups
);
5774 atomic_inc(&sd
->groups
->ref
);
5776 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5779 lockdep_assert_held(&sched_domains_mutex
);
5780 covered
= sched_domains_tmpmask
;
5782 cpumask_clear(covered
);
5784 for_each_cpu(i
, span
) {
5785 struct sched_group
*sg
;
5786 int group
= get_group(i
, sdd
, &sg
);
5789 if (cpumask_test_cpu(i
, covered
))
5792 cpumask_clear(sched_group_cpus(sg
));
5794 cpumask_setall(sched_group_mask(sg
));
5796 for_each_cpu(j
, span
) {
5797 if (get_group(j
, sdd
, NULL
) != group
)
5800 cpumask_set_cpu(j
, covered
);
5801 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5816 * Initialize sched groups cpu_power.
5818 * cpu_power indicates the capacity of sched group, which is used while
5819 * distributing the load between different sched groups in a sched domain.
5820 * Typically cpu_power for all the groups in a sched domain will be same unless
5821 * there are asymmetries in the topology. If there are asymmetries, group
5822 * having more cpu_power will pickup more load compared to the group having
5825 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5827 struct sched_group
*sg
= sd
->groups
;
5829 WARN_ON(!sd
|| !sg
);
5832 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5834 } while (sg
!= sd
->groups
);
5836 if (cpu
!= group_balance_cpu(sg
))
5839 update_group_power(sd
, cpu
);
5840 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5843 int __weak
arch_sd_sibling_asym_packing(void)
5845 return 0*SD_ASYM_PACKING
;
5849 * Initializers for schedule domains
5850 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5853 #ifdef CONFIG_SCHED_DEBUG
5854 # define SD_INIT_NAME(sd, type) sd->name = #type
5856 # define SD_INIT_NAME(sd, type) do { } while (0)
5859 #define SD_INIT_FUNC(type) \
5860 static noinline struct sched_domain * \
5861 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5863 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5864 *sd = SD_##type##_INIT; \
5865 SD_INIT_NAME(sd, type); \
5866 sd->private = &tl->data; \
5871 #ifdef CONFIG_SCHED_SMT
5872 SD_INIT_FUNC(SIBLING
)
5874 #ifdef CONFIG_SCHED_MC
5877 #ifdef CONFIG_SCHED_BOOK
5881 static int default_relax_domain_level
= -1;
5882 int sched_domain_level_max
;
5884 static int __init
setup_relax_domain_level(char *str
)
5886 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5887 pr_warn("Unable to set relax_domain_level\n");
5891 __setup("relax_domain_level=", setup_relax_domain_level
);
5893 static void set_domain_attribute(struct sched_domain
*sd
,
5894 struct sched_domain_attr
*attr
)
5898 if (!attr
|| attr
->relax_domain_level
< 0) {
5899 if (default_relax_domain_level
< 0)
5902 request
= default_relax_domain_level
;
5904 request
= attr
->relax_domain_level
;
5905 if (request
< sd
->level
) {
5906 /* turn off idle balance on this domain */
5907 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5909 /* turn on idle balance on this domain */
5910 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5914 static void __sdt_free(const struct cpumask
*cpu_map
);
5915 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5917 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5918 const struct cpumask
*cpu_map
)
5922 if (!atomic_read(&d
->rd
->refcount
))
5923 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5925 free_percpu(d
->sd
); /* fall through */
5927 __sdt_free(cpu_map
); /* fall through */
5933 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5934 const struct cpumask
*cpu_map
)
5936 memset(d
, 0, sizeof(*d
));
5938 if (__sdt_alloc(cpu_map
))
5939 return sa_sd_storage
;
5940 d
->sd
= alloc_percpu(struct sched_domain
*);
5942 return sa_sd_storage
;
5943 d
->rd
= alloc_rootdomain();
5946 return sa_rootdomain
;
5950 * NULL the sd_data elements we've used to build the sched_domain and
5951 * sched_group structure so that the subsequent __free_domain_allocs()
5952 * will not free the data we're using.
5954 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5956 struct sd_data
*sdd
= sd
->private;
5958 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5959 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5961 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5962 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5964 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5965 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5968 #ifdef CONFIG_SCHED_SMT
5969 static const struct cpumask
*cpu_smt_mask(int cpu
)
5971 return topology_thread_cpumask(cpu
);
5976 * Topology list, bottom-up.
5978 static struct sched_domain_topology_level default_topology
[] = {
5979 #ifdef CONFIG_SCHED_SMT
5980 { sd_init_SIBLING
, cpu_smt_mask
, },
5982 #ifdef CONFIG_SCHED_MC
5983 { sd_init_MC
, cpu_coregroup_mask
, },
5985 #ifdef CONFIG_SCHED_BOOK
5986 { sd_init_BOOK
, cpu_book_mask
, },
5988 { sd_init_CPU
, cpu_cpu_mask
, },
5992 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5996 static int sched_domains_numa_levels
;
5997 static int *sched_domains_numa_distance
;
5998 static struct cpumask
***sched_domains_numa_masks
;
5999 static int sched_domains_curr_level
;
6001 static inline int sd_local_flags(int level
)
6003 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6006 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6009 static struct sched_domain
*
6010 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6012 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6013 int level
= tl
->numa_level
;
6014 int sd_weight
= cpumask_weight(
6015 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6017 *sd
= (struct sched_domain
){
6018 .min_interval
= sd_weight
,
6019 .max_interval
= 2*sd_weight
,
6021 .imbalance_pct
= 125,
6022 .cache_nice_tries
= 2,
6029 .flags
= 1*SD_LOAD_BALANCE
6030 | 1*SD_BALANCE_NEWIDLE
6035 | 0*SD_SHARE_CPUPOWER
6036 | 0*SD_SHARE_PKG_RESOURCES
6038 | 0*SD_PREFER_SIBLING
6039 | sd_local_flags(level
)
6041 .last_balance
= jiffies
,
6042 .balance_interval
= sd_weight
,
6044 SD_INIT_NAME(sd
, NUMA
);
6045 sd
->private = &tl
->data
;
6048 * Ugly hack to pass state to sd_numa_mask()...
6050 sched_domains_curr_level
= tl
->numa_level
;
6055 static const struct cpumask
*sd_numa_mask(int cpu
)
6057 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6060 static void sched_numa_warn(const char *str
)
6062 static int done
= false;
6070 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6072 for (i
= 0; i
< nr_node_ids
; i
++) {
6073 printk(KERN_WARNING
" ");
6074 for (j
= 0; j
< nr_node_ids
; j
++)
6075 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6076 printk(KERN_CONT
"\n");
6078 printk(KERN_WARNING
"\n");
6081 static bool find_numa_distance(int distance
)
6085 if (distance
== node_distance(0, 0))
6088 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6089 if (sched_domains_numa_distance
[i
] == distance
)
6096 static void sched_init_numa(void)
6098 int next_distance
, curr_distance
= node_distance(0, 0);
6099 struct sched_domain_topology_level
*tl
;
6103 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6104 if (!sched_domains_numa_distance
)
6108 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6109 * unique distances in the node_distance() table.
6111 * Assumes node_distance(0,j) includes all distances in
6112 * node_distance(i,j) in order to avoid cubic time.
6114 next_distance
= curr_distance
;
6115 for (i
= 0; i
< nr_node_ids
; i
++) {
6116 for (j
= 0; j
< nr_node_ids
; j
++) {
6117 for (k
= 0; k
< nr_node_ids
; k
++) {
6118 int distance
= node_distance(i
, k
);
6120 if (distance
> curr_distance
&&
6121 (distance
< next_distance
||
6122 next_distance
== curr_distance
))
6123 next_distance
= distance
;
6126 * While not a strong assumption it would be nice to know
6127 * about cases where if node A is connected to B, B is not
6128 * equally connected to A.
6130 if (sched_debug() && node_distance(k
, i
) != distance
)
6131 sched_numa_warn("Node-distance not symmetric");
6133 if (sched_debug() && i
&& !find_numa_distance(distance
))
6134 sched_numa_warn("Node-0 not representative");
6136 if (next_distance
!= curr_distance
) {
6137 sched_domains_numa_distance
[level
++] = next_distance
;
6138 sched_domains_numa_levels
= level
;
6139 curr_distance
= next_distance
;
6144 * In case of sched_debug() we verify the above assumption.
6150 * 'level' contains the number of unique distances, excluding the
6151 * identity distance node_distance(i,i).
6153 * The sched_domains_nume_distance[] array includes the actual distance
6158 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6159 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6160 * the array will contain less then 'level' members. This could be
6161 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6162 * in other functions.
6164 * We reset it to 'level' at the end of this function.
6166 sched_domains_numa_levels
= 0;
6168 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6169 if (!sched_domains_numa_masks
)
6173 * Now for each level, construct a mask per node which contains all
6174 * cpus of nodes that are that many hops away from us.
6176 for (i
= 0; i
< level
; i
++) {
6177 sched_domains_numa_masks
[i
] =
6178 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6179 if (!sched_domains_numa_masks
[i
])
6182 for (j
= 0; j
< nr_node_ids
; j
++) {
6183 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6187 sched_domains_numa_masks
[i
][j
] = mask
;
6189 for (k
= 0; k
< nr_node_ids
; k
++) {
6190 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6193 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6198 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6199 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6204 * Copy the default topology bits..
6206 for (i
= 0; default_topology
[i
].init
; i
++)
6207 tl
[i
] = default_topology
[i
];
6210 * .. and append 'j' levels of NUMA goodness.
6212 for (j
= 0; j
< level
; i
++, j
++) {
6213 tl
[i
] = (struct sched_domain_topology_level
){
6214 .init
= sd_numa_init
,
6215 .mask
= sd_numa_mask
,
6216 .flags
= SDTL_OVERLAP
,
6221 sched_domain_topology
= tl
;
6223 sched_domains_numa_levels
= level
;
6226 static void sched_domains_numa_masks_set(int cpu
)
6229 int node
= cpu_to_node(cpu
);
6231 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6232 for (j
= 0; j
< nr_node_ids
; j
++) {
6233 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6234 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6239 static void sched_domains_numa_masks_clear(int cpu
)
6242 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6243 for (j
= 0; j
< nr_node_ids
; j
++)
6244 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6249 * Update sched_domains_numa_masks[level][node] array when new cpus
6252 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6253 unsigned long action
,
6256 int cpu
= (long)hcpu
;
6258 switch (action
& ~CPU_TASKS_FROZEN
) {
6260 sched_domains_numa_masks_set(cpu
);
6264 sched_domains_numa_masks_clear(cpu
);
6274 static inline void sched_init_numa(void)
6278 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6279 unsigned long action
,
6284 #endif /* CONFIG_NUMA */
6286 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6288 struct sched_domain_topology_level
*tl
;
6291 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6292 struct sd_data
*sdd
= &tl
->data
;
6294 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6298 sdd
->sg
= alloc_percpu(struct sched_group
*);
6302 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6306 for_each_cpu(j
, cpu_map
) {
6307 struct sched_domain
*sd
;
6308 struct sched_group
*sg
;
6309 struct sched_group_power
*sgp
;
6311 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6312 GFP_KERNEL
, cpu_to_node(j
));
6316 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6318 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6319 GFP_KERNEL
, cpu_to_node(j
));
6325 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6327 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6328 GFP_KERNEL
, cpu_to_node(j
));
6332 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6339 static void __sdt_free(const struct cpumask
*cpu_map
)
6341 struct sched_domain_topology_level
*tl
;
6344 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6345 struct sd_data
*sdd
= &tl
->data
;
6347 for_each_cpu(j
, cpu_map
) {
6348 struct sched_domain
*sd
;
6351 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6352 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6353 free_sched_groups(sd
->groups
, 0);
6354 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6358 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6360 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6362 free_percpu(sdd
->sd
);
6364 free_percpu(sdd
->sg
);
6366 free_percpu(sdd
->sgp
);
6371 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6372 struct s_data
*d
, const struct cpumask
*cpu_map
,
6373 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6376 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6380 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6382 sd
->level
= child
->level
+ 1;
6383 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6387 set_domain_attribute(sd
, attr
);
6393 * Build sched domains for a given set of cpus and attach the sched domains
6394 * to the individual cpus
6396 static int build_sched_domains(const struct cpumask
*cpu_map
,
6397 struct sched_domain_attr
*attr
)
6399 enum s_alloc alloc_state
= sa_none
;
6400 struct sched_domain
*sd
;
6402 int i
, ret
= -ENOMEM
;
6404 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6405 if (alloc_state
!= sa_rootdomain
)
6408 /* Set up domains for cpus specified by the cpu_map. */
6409 for_each_cpu(i
, cpu_map
) {
6410 struct sched_domain_topology_level
*tl
;
6413 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6414 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6415 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6416 sd
->flags
|= SD_OVERLAP
;
6417 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6424 *per_cpu_ptr(d
.sd
, i
) = sd
;
6427 /* Build the groups for the domains */
6428 for_each_cpu(i
, cpu_map
) {
6429 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6430 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6431 if (sd
->flags
& SD_OVERLAP
) {
6432 if (build_overlap_sched_groups(sd
, i
))
6435 if (build_sched_groups(sd
, i
))
6441 /* Calculate CPU power for physical packages and nodes */
6442 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6443 if (!cpumask_test_cpu(i
, cpu_map
))
6446 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6447 claim_allocations(i
, sd
);
6448 init_sched_groups_power(i
, sd
);
6452 /* Attach the domains */
6454 for_each_cpu(i
, cpu_map
) {
6455 sd
= *per_cpu_ptr(d
.sd
, i
);
6456 cpu_attach_domain(sd
, d
.rd
, i
);
6462 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6466 static cpumask_var_t
*doms_cur
; /* current sched domains */
6467 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6468 static struct sched_domain_attr
*dattr_cur
;
6469 /* attribues of custom domains in 'doms_cur' */
6472 * Special case: If a kmalloc of a doms_cur partition (array of
6473 * cpumask) fails, then fallback to a single sched domain,
6474 * as determined by the single cpumask fallback_doms.
6476 static cpumask_var_t fallback_doms
;
6479 * arch_update_cpu_topology lets virtualized architectures update the
6480 * cpu core maps. It is supposed to return 1 if the topology changed
6481 * or 0 if it stayed the same.
6483 int __attribute__((weak
)) arch_update_cpu_topology(void)
6488 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6491 cpumask_var_t
*doms
;
6493 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6496 for (i
= 0; i
< ndoms
; i
++) {
6497 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6498 free_sched_domains(doms
, i
);
6505 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6508 for (i
= 0; i
< ndoms
; i
++)
6509 free_cpumask_var(doms
[i
]);
6514 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6515 * For now this just excludes isolated cpus, but could be used to
6516 * exclude other special cases in the future.
6518 static int init_sched_domains(const struct cpumask
*cpu_map
)
6522 arch_update_cpu_topology();
6524 doms_cur
= alloc_sched_domains(ndoms_cur
);
6526 doms_cur
= &fallback_doms
;
6527 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6528 err
= build_sched_domains(doms_cur
[0], NULL
);
6529 register_sched_domain_sysctl();
6535 * Detach sched domains from a group of cpus specified in cpu_map
6536 * These cpus will now be attached to the NULL domain
6538 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6543 for_each_cpu(i
, cpu_map
)
6544 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6548 /* handle null as "default" */
6549 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6550 struct sched_domain_attr
*new, int idx_new
)
6552 struct sched_domain_attr tmp
;
6559 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6560 new ? (new + idx_new
) : &tmp
,
6561 sizeof(struct sched_domain_attr
));
6565 * Partition sched domains as specified by the 'ndoms_new'
6566 * cpumasks in the array doms_new[] of cpumasks. This compares
6567 * doms_new[] to the current sched domain partitioning, doms_cur[].
6568 * It destroys each deleted domain and builds each new domain.
6570 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6571 * The masks don't intersect (don't overlap.) We should setup one
6572 * sched domain for each mask. CPUs not in any of the cpumasks will
6573 * not be load balanced. If the same cpumask appears both in the
6574 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6577 * The passed in 'doms_new' should be allocated using
6578 * alloc_sched_domains. This routine takes ownership of it and will
6579 * free_sched_domains it when done with it. If the caller failed the
6580 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6581 * and partition_sched_domains() will fallback to the single partition
6582 * 'fallback_doms', it also forces the domains to be rebuilt.
6584 * If doms_new == NULL it will be replaced with cpu_online_mask.
6585 * ndoms_new == 0 is a special case for destroying existing domains,
6586 * and it will not create the default domain.
6588 * Call with hotplug lock held
6590 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6591 struct sched_domain_attr
*dattr_new
)
6596 mutex_lock(&sched_domains_mutex
);
6598 /* always unregister in case we don't destroy any domains */
6599 unregister_sched_domain_sysctl();
6601 /* Let architecture update cpu core mappings. */
6602 new_topology
= arch_update_cpu_topology();
6604 n
= doms_new
? ndoms_new
: 0;
6606 /* Destroy deleted domains */
6607 for (i
= 0; i
< ndoms_cur
; i
++) {
6608 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6609 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6610 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6613 /* no match - a current sched domain not in new doms_new[] */
6614 detach_destroy_domains(doms_cur
[i
]);
6619 if (doms_new
== NULL
) {
6621 doms_new
= &fallback_doms
;
6622 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6623 WARN_ON_ONCE(dattr_new
);
6626 /* Build new domains */
6627 for (i
= 0; i
< ndoms_new
; i
++) {
6628 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6629 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6630 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6633 /* no match - add a new doms_new */
6634 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6639 /* Remember the new sched domains */
6640 if (doms_cur
!= &fallback_doms
)
6641 free_sched_domains(doms_cur
, ndoms_cur
);
6642 kfree(dattr_cur
); /* kfree(NULL) is safe */
6643 doms_cur
= doms_new
;
6644 dattr_cur
= dattr_new
;
6645 ndoms_cur
= ndoms_new
;
6647 register_sched_domain_sysctl();
6649 mutex_unlock(&sched_domains_mutex
);
6652 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6655 * Update cpusets according to cpu_active mask. If cpusets are
6656 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6657 * around partition_sched_domains().
6659 * If we come here as part of a suspend/resume, don't touch cpusets because we
6660 * want to restore it back to its original state upon resume anyway.
6662 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6666 case CPU_ONLINE_FROZEN
:
6667 case CPU_DOWN_FAILED_FROZEN
:
6670 * num_cpus_frozen tracks how many CPUs are involved in suspend
6671 * resume sequence. As long as this is not the last online
6672 * operation in the resume sequence, just build a single sched
6673 * domain, ignoring cpusets.
6676 if (likely(num_cpus_frozen
)) {
6677 partition_sched_domains(1, NULL
, NULL
);
6682 * This is the last CPU online operation. So fall through and
6683 * restore the original sched domains by considering the
6684 * cpuset configurations.
6688 case CPU_DOWN_FAILED
:
6689 cpuset_update_active_cpus(true);
6697 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6701 case CPU_DOWN_PREPARE
:
6702 cpuset_update_active_cpus(false);
6704 case CPU_DOWN_PREPARE_FROZEN
:
6706 partition_sched_domains(1, NULL
, NULL
);
6714 void __init
sched_init_smp(void)
6716 cpumask_var_t non_isolated_cpus
;
6718 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6719 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6724 mutex_lock(&sched_domains_mutex
);
6725 init_sched_domains(cpu_active_mask
);
6726 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6727 if (cpumask_empty(non_isolated_cpus
))
6728 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6729 mutex_unlock(&sched_domains_mutex
);
6732 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6733 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6734 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6736 /* RT runtime code needs to handle some hotplug events */
6737 hotcpu_notifier(update_runtime
, 0);
6741 /* Move init over to a non-isolated CPU */
6742 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6744 sched_init_granularity();
6745 free_cpumask_var(non_isolated_cpus
);
6747 init_sched_rt_class();
6750 void __init
sched_init_smp(void)
6752 sched_init_granularity();
6754 #endif /* CONFIG_SMP */
6756 const_debug
unsigned int sysctl_timer_migration
= 1;
6758 int in_sched_functions(unsigned long addr
)
6760 return in_lock_functions(addr
) ||
6761 (addr
>= (unsigned long)__sched_text_start
6762 && addr
< (unsigned long)__sched_text_end
);
6765 #ifdef CONFIG_CGROUP_SCHED
6766 struct task_group root_task_group
;
6767 LIST_HEAD(task_groups
);
6770 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6772 void __init
sched_init(void)
6775 unsigned long alloc_size
= 0, ptr
;
6777 #ifdef CONFIG_FAIR_GROUP_SCHED
6778 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6780 #ifdef CONFIG_RT_GROUP_SCHED
6781 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6783 #ifdef CONFIG_CPUMASK_OFFSTACK
6784 alloc_size
+= num_possible_cpus() * cpumask_size();
6787 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6789 #ifdef CONFIG_FAIR_GROUP_SCHED
6790 root_task_group
.se
= (struct sched_entity
**)ptr
;
6791 ptr
+= nr_cpu_ids
* sizeof(void **);
6793 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6794 ptr
+= nr_cpu_ids
* sizeof(void **);
6796 #endif /* CONFIG_FAIR_GROUP_SCHED */
6797 #ifdef CONFIG_RT_GROUP_SCHED
6798 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6799 ptr
+= nr_cpu_ids
* sizeof(void **);
6801 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6802 ptr
+= nr_cpu_ids
* sizeof(void **);
6804 #endif /* CONFIG_RT_GROUP_SCHED */
6805 #ifdef CONFIG_CPUMASK_OFFSTACK
6806 for_each_possible_cpu(i
) {
6807 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6808 ptr
+= cpumask_size();
6810 #endif /* CONFIG_CPUMASK_OFFSTACK */
6814 init_defrootdomain();
6817 init_rt_bandwidth(&def_rt_bandwidth
,
6818 global_rt_period(), global_rt_runtime());
6820 #ifdef CONFIG_RT_GROUP_SCHED
6821 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6822 global_rt_period(), global_rt_runtime());
6823 #endif /* CONFIG_RT_GROUP_SCHED */
6825 #ifdef CONFIG_CGROUP_SCHED
6826 list_add(&root_task_group
.list
, &task_groups
);
6827 INIT_LIST_HEAD(&root_task_group
.children
);
6828 INIT_LIST_HEAD(&root_task_group
.siblings
);
6829 autogroup_init(&init_task
);
6831 #endif /* CONFIG_CGROUP_SCHED */
6833 #ifdef CONFIG_CGROUP_CPUACCT
6834 root_cpuacct
.cpustat
= &kernel_cpustat
;
6835 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6836 /* Too early, not expected to fail */
6837 BUG_ON(!root_cpuacct
.cpuusage
);
6839 for_each_possible_cpu(i
) {
6843 raw_spin_lock_init(&rq
->lock
);
6845 rq
->calc_load_active
= 0;
6846 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6847 init_cfs_rq(&rq
->cfs
);
6848 init_rt_rq(&rq
->rt
, rq
);
6849 #ifdef CONFIG_FAIR_GROUP_SCHED
6850 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6851 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6853 * How much cpu bandwidth does root_task_group get?
6855 * In case of task-groups formed thr' the cgroup filesystem, it
6856 * gets 100% of the cpu resources in the system. This overall
6857 * system cpu resource is divided among the tasks of
6858 * root_task_group and its child task-groups in a fair manner,
6859 * based on each entity's (task or task-group's) weight
6860 * (se->load.weight).
6862 * In other words, if root_task_group has 10 tasks of weight
6863 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6864 * then A0's share of the cpu resource is:
6866 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6868 * We achieve this by letting root_task_group's tasks sit
6869 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6871 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6872 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6873 #endif /* CONFIG_FAIR_GROUP_SCHED */
6875 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6876 #ifdef CONFIG_RT_GROUP_SCHED
6877 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6878 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6881 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6882 rq
->cpu_load
[j
] = 0;
6884 rq
->last_load_update_tick
= jiffies
;
6889 rq
->cpu_power
= SCHED_POWER_SCALE
;
6890 rq
->post_schedule
= 0;
6891 rq
->active_balance
= 0;
6892 rq
->next_balance
= jiffies
;
6897 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6899 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6901 rq_attach_root(rq
, &def_root_domain
);
6907 atomic_set(&rq
->nr_iowait
, 0);
6910 set_load_weight(&init_task
);
6912 #ifdef CONFIG_PREEMPT_NOTIFIERS
6913 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6916 #ifdef CONFIG_RT_MUTEXES
6917 plist_head_init(&init_task
.pi_waiters
);
6921 * The boot idle thread does lazy MMU switching as well:
6923 atomic_inc(&init_mm
.mm_count
);
6924 enter_lazy_tlb(&init_mm
, current
);
6927 * Make us the idle thread. Technically, schedule() should not be
6928 * called from this thread, however somewhere below it might be,
6929 * but because we are the idle thread, we just pick up running again
6930 * when this runqueue becomes "idle".
6932 init_idle(current
, smp_processor_id());
6934 calc_load_update
= jiffies
+ LOAD_FREQ
;
6937 * During early bootup we pretend to be a normal task:
6939 current
->sched_class
= &fair_sched_class
;
6942 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6943 /* May be allocated at isolcpus cmdline parse time */
6944 if (cpu_isolated_map
== NULL
)
6945 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6946 idle_thread_set_boot_cpu();
6948 init_sched_fair_class();
6950 scheduler_running
= 1;
6953 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6954 static inline int preempt_count_equals(int preempt_offset
)
6956 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6958 return (nested
== preempt_offset
);
6961 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6963 static unsigned long prev_jiffy
; /* ratelimiting */
6965 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6966 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6967 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6969 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6971 prev_jiffy
= jiffies
;
6974 "BUG: sleeping function called from invalid context at %s:%d\n",
6977 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6978 in_atomic(), irqs_disabled(),
6979 current
->pid
, current
->comm
);
6981 debug_show_held_locks(current
);
6982 if (irqs_disabled())
6983 print_irqtrace_events(current
);
6986 EXPORT_SYMBOL(__might_sleep
);
6989 #ifdef CONFIG_MAGIC_SYSRQ
6990 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6992 const struct sched_class
*prev_class
= p
->sched_class
;
6993 int old_prio
= p
->prio
;
6998 dequeue_task(rq
, p
, 0);
6999 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7001 enqueue_task(rq
, p
, 0);
7002 resched_task(rq
->curr
);
7005 check_class_changed(rq
, p
, prev_class
, old_prio
);
7008 void normalize_rt_tasks(void)
7010 struct task_struct
*g
, *p
;
7011 unsigned long flags
;
7014 read_lock_irqsave(&tasklist_lock
, flags
);
7015 do_each_thread(g
, p
) {
7017 * Only normalize user tasks:
7022 p
->se
.exec_start
= 0;
7023 #ifdef CONFIG_SCHEDSTATS
7024 p
->se
.statistics
.wait_start
= 0;
7025 p
->se
.statistics
.sleep_start
= 0;
7026 p
->se
.statistics
.block_start
= 0;
7031 * Renice negative nice level userspace
7034 if (TASK_NICE(p
) < 0 && p
->mm
)
7035 set_user_nice(p
, 0);
7039 raw_spin_lock(&p
->pi_lock
);
7040 rq
= __task_rq_lock(p
);
7042 normalize_task(rq
, p
);
7044 __task_rq_unlock(rq
);
7045 raw_spin_unlock(&p
->pi_lock
);
7046 } while_each_thread(g
, p
);
7048 read_unlock_irqrestore(&tasklist_lock
, flags
);
7051 #endif /* CONFIG_MAGIC_SYSRQ */
7053 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7055 * These functions are only useful for the IA64 MCA handling, or kdb.
7057 * They can only be called when the whole system has been
7058 * stopped - every CPU needs to be quiescent, and no scheduling
7059 * activity can take place. Using them for anything else would
7060 * be a serious bug, and as a result, they aren't even visible
7061 * under any other configuration.
7065 * curr_task - return the current task for a given cpu.
7066 * @cpu: the processor in question.
7068 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7070 struct task_struct
*curr_task(int cpu
)
7072 return cpu_curr(cpu
);
7075 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7079 * set_curr_task - set the current task for a given cpu.
7080 * @cpu: the processor in question.
7081 * @p: the task pointer to set.
7083 * Description: This function must only be used when non-maskable interrupts
7084 * are serviced on a separate stack. It allows the architecture to switch the
7085 * notion of the current task on a cpu in a non-blocking manner. This function
7086 * must be called with all CPU's synchronized, and interrupts disabled, the
7087 * and caller must save the original value of the current task (see
7088 * curr_task() above) and restore that value before reenabling interrupts and
7089 * re-starting the system.
7091 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7093 void set_curr_task(int cpu
, struct task_struct
*p
)
7100 #ifdef CONFIG_CGROUP_SCHED
7101 /* task_group_lock serializes the addition/removal of task groups */
7102 static DEFINE_SPINLOCK(task_group_lock
);
7104 static void free_sched_group(struct task_group
*tg
)
7106 free_fair_sched_group(tg
);
7107 free_rt_sched_group(tg
);
7112 /* allocate runqueue etc for a new task group */
7113 struct task_group
*sched_create_group(struct task_group
*parent
)
7115 struct task_group
*tg
;
7116 unsigned long flags
;
7118 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7120 return ERR_PTR(-ENOMEM
);
7122 if (!alloc_fair_sched_group(tg
, parent
))
7125 if (!alloc_rt_sched_group(tg
, parent
))
7128 spin_lock_irqsave(&task_group_lock
, flags
);
7129 list_add_rcu(&tg
->list
, &task_groups
);
7131 WARN_ON(!parent
); /* root should already exist */
7133 tg
->parent
= parent
;
7134 INIT_LIST_HEAD(&tg
->children
);
7135 list_add_rcu(&tg
->siblings
, &parent
->children
);
7136 spin_unlock_irqrestore(&task_group_lock
, flags
);
7141 free_sched_group(tg
);
7142 return ERR_PTR(-ENOMEM
);
7145 /* rcu callback to free various structures associated with a task group */
7146 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7148 /* now it should be safe to free those cfs_rqs */
7149 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7152 /* Destroy runqueue etc associated with a task group */
7153 void sched_destroy_group(struct task_group
*tg
)
7155 unsigned long flags
;
7158 /* end participation in shares distribution */
7159 for_each_possible_cpu(i
)
7160 unregister_fair_sched_group(tg
, i
);
7162 spin_lock_irqsave(&task_group_lock
, flags
);
7163 list_del_rcu(&tg
->list
);
7164 list_del_rcu(&tg
->siblings
);
7165 spin_unlock_irqrestore(&task_group_lock
, flags
);
7167 /* wait for possible concurrent references to cfs_rqs complete */
7168 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7171 /* change task's runqueue when it moves between groups.
7172 * The caller of this function should have put the task in its new group
7173 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7174 * reflect its new group.
7176 void sched_move_task(struct task_struct
*tsk
)
7178 struct task_group
*tg
;
7180 unsigned long flags
;
7183 rq
= task_rq_lock(tsk
, &flags
);
7185 running
= task_current(rq
, tsk
);
7189 dequeue_task(rq
, tsk
, 0);
7190 if (unlikely(running
))
7191 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7193 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7194 lockdep_is_held(&tsk
->sighand
->siglock
)),
7195 struct task_group
, css
);
7196 tg
= autogroup_task_group(tsk
, tg
);
7197 tsk
->sched_task_group
= tg
;
7199 #ifdef CONFIG_FAIR_GROUP_SCHED
7200 if (tsk
->sched_class
->task_move_group
)
7201 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7204 set_task_rq(tsk
, task_cpu(tsk
));
7206 if (unlikely(running
))
7207 tsk
->sched_class
->set_curr_task(rq
);
7209 enqueue_task(rq
, tsk
, 0);
7211 task_rq_unlock(rq
, tsk
, &flags
);
7213 #endif /* CONFIG_CGROUP_SCHED */
7215 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7216 static unsigned long to_ratio(u64 period
, u64 runtime
)
7218 if (runtime
== RUNTIME_INF
)
7221 return div64_u64(runtime
<< 20, period
);
7225 #ifdef CONFIG_RT_GROUP_SCHED
7227 * Ensure that the real time constraints are schedulable.
7229 static DEFINE_MUTEX(rt_constraints_mutex
);
7231 /* Must be called with tasklist_lock held */
7232 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7234 struct task_struct
*g
, *p
;
7236 do_each_thread(g
, p
) {
7237 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7239 } while_each_thread(g
, p
);
7244 struct rt_schedulable_data
{
7245 struct task_group
*tg
;
7250 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7252 struct rt_schedulable_data
*d
= data
;
7253 struct task_group
*child
;
7254 unsigned long total
, sum
= 0;
7255 u64 period
, runtime
;
7257 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7258 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7261 period
= d
->rt_period
;
7262 runtime
= d
->rt_runtime
;
7266 * Cannot have more runtime than the period.
7268 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7272 * Ensure we don't starve existing RT tasks.
7274 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7277 total
= to_ratio(period
, runtime
);
7280 * Nobody can have more than the global setting allows.
7282 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7286 * The sum of our children's runtime should not exceed our own.
7288 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7289 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7290 runtime
= child
->rt_bandwidth
.rt_runtime
;
7292 if (child
== d
->tg
) {
7293 period
= d
->rt_period
;
7294 runtime
= d
->rt_runtime
;
7297 sum
+= to_ratio(period
, runtime
);
7306 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7310 struct rt_schedulable_data data
= {
7312 .rt_period
= period
,
7313 .rt_runtime
= runtime
,
7317 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7323 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7324 u64 rt_period
, u64 rt_runtime
)
7328 mutex_lock(&rt_constraints_mutex
);
7329 read_lock(&tasklist_lock
);
7330 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7334 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7335 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7336 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7338 for_each_possible_cpu(i
) {
7339 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7341 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7342 rt_rq
->rt_runtime
= rt_runtime
;
7343 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7345 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7347 read_unlock(&tasklist_lock
);
7348 mutex_unlock(&rt_constraints_mutex
);
7353 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7355 u64 rt_runtime
, rt_period
;
7357 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7358 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7359 if (rt_runtime_us
< 0)
7360 rt_runtime
= RUNTIME_INF
;
7362 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7365 long sched_group_rt_runtime(struct task_group
*tg
)
7369 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7372 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7373 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7374 return rt_runtime_us
;
7377 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7379 u64 rt_runtime
, rt_period
;
7381 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7382 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7387 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7390 long sched_group_rt_period(struct task_group
*tg
)
7394 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7395 do_div(rt_period_us
, NSEC_PER_USEC
);
7396 return rt_period_us
;
7399 static int sched_rt_global_constraints(void)
7401 u64 runtime
, period
;
7404 if (sysctl_sched_rt_period
<= 0)
7407 runtime
= global_rt_runtime();
7408 period
= global_rt_period();
7411 * Sanity check on the sysctl variables.
7413 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7416 mutex_lock(&rt_constraints_mutex
);
7417 read_lock(&tasklist_lock
);
7418 ret
= __rt_schedulable(NULL
, 0, 0);
7419 read_unlock(&tasklist_lock
);
7420 mutex_unlock(&rt_constraints_mutex
);
7425 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7427 /* Don't accept realtime tasks when there is no way for them to run */
7428 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7434 #else /* !CONFIG_RT_GROUP_SCHED */
7435 static int sched_rt_global_constraints(void)
7437 unsigned long flags
;
7440 if (sysctl_sched_rt_period
<= 0)
7444 * There's always some RT tasks in the root group
7445 * -- migration, kstopmachine etc..
7447 if (sysctl_sched_rt_runtime
== 0)
7450 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7451 for_each_possible_cpu(i
) {
7452 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7454 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7455 rt_rq
->rt_runtime
= global_rt_runtime();
7456 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7458 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7462 #endif /* CONFIG_RT_GROUP_SCHED */
7464 int sched_rt_handler(struct ctl_table
*table
, int write
,
7465 void __user
*buffer
, size_t *lenp
,
7469 int old_period
, old_runtime
;
7470 static DEFINE_MUTEX(mutex
);
7473 old_period
= sysctl_sched_rt_period
;
7474 old_runtime
= sysctl_sched_rt_runtime
;
7476 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7478 if (!ret
&& write
) {
7479 ret
= sched_rt_global_constraints();
7481 sysctl_sched_rt_period
= old_period
;
7482 sysctl_sched_rt_runtime
= old_runtime
;
7484 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7485 def_rt_bandwidth
.rt_period
=
7486 ns_to_ktime(global_rt_period());
7489 mutex_unlock(&mutex
);
7494 #ifdef CONFIG_CGROUP_SCHED
7496 /* return corresponding task_group object of a cgroup */
7497 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7499 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7500 struct task_group
, css
);
7503 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7505 struct task_group
*tg
, *parent
;
7507 if (!cgrp
->parent
) {
7508 /* This is early initialization for the top cgroup */
7509 return &root_task_group
.css
;
7512 parent
= cgroup_tg(cgrp
->parent
);
7513 tg
= sched_create_group(parent
);
7515 return ERR_PTR(-ENOMEM
);
7520 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7522 struct task_group
*tg
= cgroup_tg(cgrp
);
7524 sched_destroy_group(tg
);
7527 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7528 struct cgroup_taskset
*tset
)
7530 struct task_struct
*task
;
7532 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7533 #ifdef CONFIG_RT_GROUP_SCHED
7534 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7537 /* We don't support RT-tasks being in separate groups */
7538 if (task
->sched_class
!= &fair_sched_class
)
7545 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7546 struct cgroup_taskset
*tset
)
7548 struct task_struct
*task
;
7550 cgroup_taskset_for_each(task
, cgrp
, tset
)
7551 sched_move_task(task
);
7555 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7556 struct task_struct
*task
)
7559 * cgroup_exit() is called in the copy_process() failure path.
7560 * Ignore this case since the task hasn't ran yet, this avoids
7561 * trying to poke a half freed task state from generic code.
7563 if (!(task
->flags
& PF_EXITING
))
7566 sched_move_task(task
);
7569 #ifdef CONFIG_FAIR_GROUP_SCHED
7570 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7573 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7576 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7578 struct task_group
*tg
= cgroup_tg(cgrp
);
7580 return (u64
) scale_load_down(tg
->shares
);
7583 #ifdef CONFIG_CFS_BANDWIDTH
7584 static DEFINE_MUTEX(cfs_constraints_mutex
);
7586 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7587 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7589 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7591 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7593 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7594 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7596 if (tg
== &root_task_group
)
7600 * Ensure we have at some amount of bandwidth every period. This is
7601 * to prevent reaching a state of large arrears when throttled via
7602 * entity_tick() resulting in prolonged exit starvation.
7604 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7608 * Likewise, bound things on the otherside by preventing insane quota
7609 * periods. This also allows us to normalize in computing quota
7612 if (period
> max_cfs_quota_period
)
7615 mutex_lock(&cfs_constraints_mutex
);
7616 ret
= __cfs_schedulable(tg
, period
, quota
);
7620 runtime_enabled
= quota
!= RUNTIME_INF
;
7621 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7622 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7623 raw_spin_lock_irq(&cfs_b
->lock
);
7624 cfs_b
->period
= ns_to_ktime(period
);
7625 cfs_b
->quota
= quota
;
7627 __refill_cfs_bandwidth_runtime(cfs_b
);
7628 /* restart the period timer (if active) to handle new period expiry */
7629 if (runtime_enabled
&& cfs_b
->timer_active
) {
7630 /* force a reprogram */
7631 cfs_b
->timer_active
= 0;
7632 __start_cfs_bandwidth(cfs_b
);
7634 raw_spin_unlock_irq(&cfs_b
->lock
);
7636 for_each_possible_cpu(i
) {
7637 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7638 struct rq
*rq
= cfs_rq
->rq
;
7640 raw_spin_lock_irq(&rq
->lock
);
7641 cfs_rq
->runtime_enabled
= runtime_enabled
;
7642 cfs_rq
->runtime_remaining
= 0;
7644 if (cfs_rq
->throttled
)
7645 unthrottle_cfs_rq(cfs_rq
);
7646 raw_spin_unlock_irq(&rq
->lock
);
7649 mutex_unlock(&cfs_constraints_mutex
);
7654 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7658 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7659 if (cfs_quota_us
< 0)
7660 quota
= RUNTIME_INF
;
7662 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7664 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7667 long tg_get_cfs_quota(struct task_group
*tg
)
7671 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7674 quota_us
= tg
->cfs_bandwidth
.quota
;
7675 do_div(quota_us
, NSEC_PER_USEC
);
7680 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7684 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7685 quota
= tg
->cfs_bandwidth
.quota
;
7687 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7690 long tg_get_cfs_period(struct task_group
*tg
)
7694 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7695 do_div(cfs_period_us
, NSEC_PER_USEC
);
7697 return cfs_period_us
;
7700 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7702 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7705 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7708 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7711 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7713 return tg_get_cfs_period(cgroup_tg(cgrp
));
7716 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7719 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7722 struct cfs_schedulable_data
{
7723 struct task_group
*tg
;
7728 * normalize group quota/period to be quota/max_period
7729 * note: units are usecs
7731 static u64
normalize_cfs_quota(struct task_group
*tg
,
7732 struct cfs_schedulable_data
*d
)
7740 period
= tg_get_cfs_period(tg
);
7741 quota
= tg_get_cfs_quota(tg
);
7744 /* note: these should typically be equivalent */
7745 if (quota
== RUNTIME_INF
|| quota
== -1)
7748 return to_ratio(period
, quota
);
7751 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7753 struct cfs_schedulable_data
*d
= data
;
7754 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7755 s64 quota
= 0, parent_quota
= -1;
7758 quota
= RUNTIME_INF
;
7760 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7762 quota
= normalize_cfs_quota(tg
, d
);
7763 parent_quota
= parent_b
->hierarchal_quota
;
7766 * ensure max(child_quota) <= parent_quota, inherit when no
7769 if (quota
== RUNTIME_INF
)
7770 quota
= parent_quota
;
7771 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7774 cfs_b
->hierarchal_quota
= quota
;
7779 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7782 struct cfs_schedulable_data data
= {
7788 if (quota
!= RUNTIME_INF
) {
7789 do_div(data
.period
, NSEC_PER_USEC
);
7790 do_div(data
.quota
, NSEC_PER_USEC
);
7794 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7800 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7801 struct cgroup_map_cb
*cb
)
7803 struct task_group
*tg
= cgroup_tg(cgrp
);
7804 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7806 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7807 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7808 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7812 #endif /* CONFIG_CFS_BANDWIDTH */
7813 #endif /* CONFIG_FAIR_GROUP_SCHED */
7815 #ifdef CONFIG_RT_GROUP_SCHED
7816 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7819 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7822 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7824 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7827 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7830 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7833 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7835 return sched_group_rt_period(cgroup_tg(cgrp
));
7837 #endif /* CONFIG_RT_GROUP_SCHED */
7839 static struct cftype cpu_files
[] = {
7840 #ifdef CONFIG_FAIR_GROUP_SCHED
7843 .read_u64
= cpu_shares_read_u64
,
7844 .write_u64
= cpu_shares_write_u64
,
7847 #ifdef CONFIG_CFS_BANDWIDTH
7849 .name
= "cfs_quota_us",
7850 .read_s64
= cpu_cfs_quota_read_s64
,
7851 .write_s64
= cpu_cfs_quota_write_s64
,
7854 .name
= "cfs_period_us",
7855 .read_u64
= cpu_cfs_period_read_u64
,
7856 .write_u64
= cpu_cfs_period_write_u64
,
7860 .read_map
= cpu_stats_show
,
7863 #ifdef CONFIG_RT_GROUP_SCHED
7865 .name
= "rt_runtime_us",
7866 .read_s64
= cpu_rt_runtime_read
,
7867 .write_s64
= cpu_rt_runtime_write
,
7870 .name
= "rt_period_us",
7871 .read_u64
= cpu_rt_period_read_uint
,
7872 .write_u64
= cpu_rt_period_write_uint
,
7878 struct cgroup_subsys cpu_cgroup_subsys
= {
7880 .create
= cpu_cgroup_create
,
7881 .destroy
= cpu_cgroup_destroy
,
7882 .can_attach
= cpu_cgroup_can_attach
,
7883 .attach
= cpu_cgroup_attach
,
7884 .exit
= cpu_cgroup_exit
,
7885 .subsys_id
= cpu_cgroup_subsys_id
,
7886 .base_cftypes
= cpu_files
,
7890 #endif /* CONFIG_CGROUP_SCHED */
7892 #ifdef CONFIG_CGROUP_CPUACCT
7895 * CPU accounting code for task groups.
7897 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7898 * (balbir@in.ibm.com).
7901 struct cpuacct root_cpuacct
;
7903 /* create a new cpu accounting group */
7904 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
7909 return &root_cpuacct
.css
;
7911 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7915 ca
->cpuusage
= alloc_percpu(u64
);
7919 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7921 goto out_free_cpuusage
;
7926 free_percpu(ca
->cpuusage
);
7930 return ERR_PTR(-ENOMEM
);
7933 /* destroy an existing cpu accounting group */
7934 static void cpuacct_destroy(struct cgroup
*cgrp
)
7936 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7938 free_percpu(ca
->cpustat
);
7939 free_percpu(ca
->cpuusage
);
7943 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
7945 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7948 #ifndef CONFIG_64BIT
7950 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7952 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7954 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7962 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
7964 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7966 #ifndef CONFIG_64BIT
7968 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7970 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7972 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7978 /* return total cpu usage (in nanoseconds) of a group */
7979 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7981 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7982 u64 totalcpuusage
= 0;
7985 for_each_present_cpu(i
)
7986 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
7988 return totalcpuusage
;
7991 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
7994 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8003 for_each_present_cpu(i
)
8004 cpuacct_cpuusage_write(ca
, i
, 0);
8010 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8013 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8017 for_each_present_cpu(i
) {
8018 percpu
= cpuacct_cpuusage_read(ca
, i
);
8019 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8021 seq_printf(m
, "\n");
8025 static const char *cpuacct_stat_desc
[] = {
8026 [CPUACCT_STAT_USER
] = "user",
8027 [CPUACCT_STAT_SYSTEM
] = "system",
8030 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8031 struct cgroup_map_cb
*cb
)
8033 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8037 for_each_online_cpu(cpu
) {
8038 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8039 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8040 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8042 val
= cputime64_to_clock_t(val
);
8043 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8046 for_each_online_cpu(cpu
) {
8047 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8048 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8049 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8050 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8053 val
= cputime64_to_clock_t(val
);
8054 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8059 static struct cftype files
[] = {
8062 .read_u64
= cpuusage_read
,
8063 .write_u64
= cpuusage_write
,
8066 .name
= "usage_percpu",
8067 .read_seq_string
= cpuacct_percpu_seq_read
,
8071 .read_map
= cpuacct_stats_show
,
8077 * charge this task's execution time to its accounting group.
8079 * called with rq->lock held.
8081 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8086 if (unlikely(!cpuacct_subsys
.active
))
8089 cpu
= task_cpu(tsk
);
8095 for (; ca
; ca
= parent_ca(ca
)) {
8096 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8097 *cpuusage
+= cputime
;
8103 struct cgroup_subsys cpuacct_subsys
= {
8105 .create
= cpuacct_create
,
8106 .destroy
= cpuacct_destroy
,
8107 .subsys_id
= cpuacct_subsys_id
,
8108 .base_cftypes
= files
,
8110 #endif /* CONFIG_CGROUP_CPUACCT */