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>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
84 #include "../workqueue_sched.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
89 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
92 ktime_t soft
, hard
, now
;
95 if (hrtimer_active(period_timer
))
98 now
= hrtimer_cb_get_time(period_timer
);
99 hrtimer_forward(period_timer
, now
, period
);
101 soft
= hrtimer_get_softexpires(period_timer
);
102 hard
= hrtimer_get_expires(period_timer
);
103 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
104 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
105 HRTIMER_MODE_ABS_PINNED
, 0);
109 DEFINE_MUTEX(sched_domains_mutex
);
110 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
112 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
114 void update_rq_clock(struct rq
*rq
)
118 if (rq
->skip_clock_update
> 0)
121 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
123 update_rq_clock_task(rq
, delta
);
127 * Debugging: various feature bits
130 #define SCHED_FEAT(name, enabled) \
131 (1UL << __SCHED_FEAT_##name) * enabled |
133 const_debug
unsigned int sysctl_sched_features
=
134 #include "features.h"
139 #ifdef CONFIG_SCHED_DEBUG
140 #define SCHED_FEAT(name, enabled) \
143 static __read_mostly
char *sched_feat_names
[] = {
144 #include "features.h"
150 static int sched_feat_show(struct seq_file
*m
, void *v
)
154 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
155 if (!(sysctl_sched_features
& (1UL << i
)))
157 seq_printf(m
, "%s ", sched_feat_names
[i
]);
164 #ifdef HAVE_JUMP_LABEL
166 #define jump_label_key__true STATIC_KEY_INIT_TRUE
167 #define jump_label_key__false STATIC_KEY_INIT_FALSE
169 #define SCHED_FEAT(name, enabled) \
170 jump_label_key__##enabled ,
172 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
173 #include "features.h"
178 static void sched_feat_disable(int i
)
180 if (static_key_enabled(&sched_feat_keys
[i
]))
181 static_key_slow_dec(&sched_feat_keys
[i
]);
184 static void sched_feat_enable(int i
)
186 if (!static_key_enabled(&sched_feat_keys
[i
]))
187 static_key_slow_inc(&sched_feat_keys
[i
]);
190 static void sched_feat_disable(int i
) { };
191 static void sched_feat_enable(int i
) { };
192 #endif /* HAVE_JUMP_LABEL */
195 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
196 size_t cnt
, loff_t
*ppos
)
206 if (copy_from_user(&buf
, ubuf
, cnt
))
212 if (strncmp(cmp
, "NO_", 3) == 0) {
217 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
218 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
220 sysctl_sched_features
&= ~(1UL << i
);
221 sched_feat_disable(i
);
223 sysctl_sched_features
|= (1UL << i
);
224 sched_feat_enable(i
);
230 if (i
== __SCHED_FEAT_NR
)
238 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
240 return single_open(filp
, sched_feat_show
, NULL
);
243 static const struct file_operations sched_feat_fops
= {
244 .open
= sched_feat_open
,
245 .write
= sched_feat_write
,
248 .release
= single_release
,
251 static __init
int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
258 late_initcall(sched_init_debug
);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
268 * period over which we average the RT time consumption, measured
273 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period
= 1000000;
281 __read_mostly
int scheduler_running
;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime
= 950000;
292 * __task_rq_lock - lock the rq @p resides on.
294 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
299 lockdep_assert_held(&p
->pi_lock
);
303 raw_spin_lock(&rq
->lock
);
304 if (likely(rq
== task_rq(p
)))
306 raw_spin_unlock(&rq
->lock
);
311 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
314 __acquires(p
->pi_lock
)
320 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
322 raw_spin_lock(&rq
->lock
);
323 if (likely(rq
== task_rq(p
)))
325 raw_spin_unlock(&rq
->lock
);
326 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
330 static void __task_rq_unlock(struct rq
*rq
)
333 raw_spin_unlock(&rq
->lock
);
337 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
339 __releases(p
->pi_lock
)
341 raw_spin_unlock(&rq
->lock
);
342 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
346 * this_rq_lock - lock this runqueue and disable interrupts.
348 static struct rq
*this_rq_lock(void)
355 raw_spin_lock(&rq
->lock
);
360 #ifdef CONFIG_SCHED_HRTICK
362 * Use HR-timers to deliver accurate preemption points.
364 * Its all a bit involved since we cannot program an hrt while holding the
365 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
368 * When we get rescheduled we reprogram the hrtick_timer outside of the
372 static void hrtick_clear(struct rq
*rq
)
374 if (hrtimer_active(&rq
->hrtick_timer
))
375 hrtimer_cancel(&rq
->hrtick_timer
);
379 * High-resolution timer tick.
380 * Runs from hardirq context with interrupts disabled.
382 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
384 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
386 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
388 raw_spin_lock(&rq
->lock
);
390 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
391 raw_spin_unlock(&rq
->lock
);
393 return HRTIMER_NORESTART
;
398 * called from hardirq (IPI) context
400 static void __hrtick_start(void *arg
)
404 raw_spin_lock(&rq
->lock
);
405 hrtimer_restart(&rq
->hrtick_timer
);
406 rq
->hrtick_csd_pending
= 0;
407 raw_spin_unlock(&rq
->lock
);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq
*rq
, u64 delay
)
417 struct hrtimer
*timer
= &rq
->hrtick_timer
;
418 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
420 hrtimer_set_expires(timer
, time
);
422 if (rq
== this_rq()) {
423 hrtimer_restart(timer
);
424 } else if (!rq
->hrtick_csd_pending
) {
425 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
426 rq
->hrtick_csd_pending
= 1;
431 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
433 int cpu
= (int)(long)hcpu
;
436 case CPU_UP_CANCELED
:
437 case CPU_UP_CANCELED_FROZEN
:
438 case CPU_DOWN_PREPARE
:
439 case CPU_DOWN_PREPARE_FROZEN
:
441 case CPU_DEAD_FROZEN
:
442 hrtick_clear(cpu_rq(cpu
));
449 static __init
void init_hrtick(void)
451 hotcpu_notifier(hotplug_hrtick
, 0);
455 * Called to set the hrtick timer state.
457 * called with rq->lock held and irqs disabled
459 void hrtick_start(struct rq
*rq
, u64 delay
)
461 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
462 HRTIMER_MODE_REL_PINNED
, 0);
465 static inline void init_hrtick(void)
468 #endif /* CONFIG_SMP */
470 static void init_rq_hrtick(struct rq
*rq
)
473 rq
->hrtick_csd_pending
= 0;
475 rq
->hrtick_csd
.flags
= 0;
476 rq
->hrtick_csd
.func
= __hrtick_start
;
477 rq
->hrtick_csd
.info
= rq
;
480 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
481 rq
->hrtick_timer
.function
= hrtick
;
483 #else /* CONFIG_SCHED_HRTICK */
484 static inline void hrtick_clear(struct rq
*rq
)
488 static inline void init_rq_hrtick(struct rq
*rq
)
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SCHED_HRTICK */
498 * resched_task - mark a task 'to be rescheduled now'.
500 * On UP this means the setting of the need_resched flag, on SMP it
501 * might also involve a cross-CPU call to trigger the scheduler on
506 #ifndef tsk_is_polling
507 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
510 void resched_task(struct task_struct
*p
)
514 assert_raw_spin_locked(&task_rq(p
)->lock
);
516 if (test_tsk_need_resched(p
))
519 set_tsk_need_resched(p
);
522 if (cpu
== smp_processor_id())
525 /* NEED_RESCHED must be visible before we test polling */
527 if (!tsk_is_polling(p
))
528 smp_send_reschedule(cpu
);
531 void resched_cpu(int cpu
)
533 struct rq
*rq
= cpu_rq(cpu
);
536 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
538 resched_task(cpu_curr(cpu
));
539 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
544 * In the semi idle case, use the nearest busy cpu for migrating timers
545 * from an idle cpu. This is good for power-savings.
547 * We don't do similar optimization for completely idle system, as
548 * selecting an idle cpu will add more delays to the timers than intended
549 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 int get_nohz_timer_target(void)
553 int cpu
= smp_processor_id();
555 struct sched_domain
*sd
;
558 for_each_domain(cpu
, sd
) {
559 for_each_cpu(i
, sched_domain_span(sd
)) {
571 * When add_timer_on() enqueues a timer into the timer wheel of an
572 * idle CPU then this timer might expire before the next timer event
573 * which is scheduled to wake up that CPU. In case of a completely
574 * idle system the next event might even be infinite time into the
575 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
576 * leaves the inner idle loop so the newly added timer is taken into
577 * account when the CPU goes back to idle and evaluates the timer
578 * wheel for the next timer event.
580 void wake_up_idle_cpu(int cpu
)
582 struct rq
*rq
= cpu_rq(cpu
);
584 if (cpu
== smp_processor_id())
588 * This is safe, as this function is called with the timer
589 * wheel base lock of (cpu) held. When the CPU is on the way
590 * to idle and has not yet set rq->curr to idle then it will
591 * be serialized on the timer wheel base lock and take the new
592 * timer into account automatically.
594 if (rq
->curr
!= rq
->idle
)
598 * We can set TIF_RESCHED on the idle task of the other CPU
599 * lockless. The worst case is that the other CPU runs the
600 * idle task through an additional NOOP schedule()
602 set_tsk_need_resched(rq
->idle
);
604 /* NEED_RESCHED must be visible before we test polling */
606 if (!tsk_is_polling(rq
->idle
))
607 smp_send_reschedule(cpu
);
610 static inline bool got_nohz_idle_kick(void)
612 int cpu
= smp_processor_id();
613 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
616 #else /* CONFIG_NO_HZ */
618 static inline bool got_nohz_idle_kick(void)
623 #endif /* CONFIG_NO_HZ */
625 void sched_avg_update(struct rq
*rq
)
627 s64 period
= sched_avg_period();
629 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
631 * Inline assembly required to prevent the compiler
632 * optimising this loop into a divmod call.
633 * See __iter_div_u64_rem() for another example of this.
635 asm("" : "+rm" (rq
->age_stamp
));
636 rq
->age_stamp
+= period
;
641 #else /* !CONFIG_SMP */
642 void resched_task(struct task_struct
*p
)
644 assert_raw_spin_locked(&task_rq(p
)->lock
);
645 set_tsk_need_resched(p
);
647 #endif /* CONFIG_SMP */
649 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
650 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652 * Iterate task_group tree rooted at *from, calling @down when first entering a
653 * node and @up when leaving it for the final time.
655 * Caller must hold rcu_lock or sufficient equivalent.
657 int walk_tg_tree_from(struct task_group
*from
,
658 tg_visitor down
, tg_visitor up
, void *data
)
660 struct task_group
*parent
, *child
;
666 ret
= (*down
)(parent
, data
);
669 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
676 ret
= (*up
)(parent
, data
);
677 if (ret
|| parent
== from
)
681 parent
= parent
->parent
;
688 int tg_nop(struct task_group
*tg
, void *data
)
694 void update_cpu_load(struct rq
*this_rq
);
696 static void set_load_weight(struct task_struct
*p
)
698 int prio
= p
->static_prio
- MAX_RT_PRIO
;
699 struct load_weight
*load
= &p
->se
.load
;
702 * SCHED_IDLE tasks get minimal weight:
704 if (p
->policy
== SCHED_IDLE
) {
705 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
706 load
->inv_weight
= WMULT_IDLEPRIO
;
710 load
->weight
= scale_load(prio_to_weight
[prio
]);
711 load
->inv_weight
= prio_to_wmult
[prio
];
714 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
717 sched_info_queued(p
);
718 p
->sched_class
->enqueue_task(rq
, p
, flags
);
721 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
724 sched_info_dequeued(p
);
725 p
->sched_class
->dequeue_task(rq
, p
, flags
);
728 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
730 if (task_contributes_to_load(p
))
731 rq
->nr_uninterruptible
--;
733 enqueue_task(rq
, p
, flags
);
736 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
738 if (task_contributes_to_load(p
))
739 rq
->nr_uninterruptible
++;
741 dequeue_task(rq
, p
, flags
);
744 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
747 * There are no locks covering percpu hardirq/softirq time.
748 * They are only modified in account_system_vtime, on corresponding CPU
749 * with interrupts disabled. So, writes are safe.
750 * They are read and saved off onto struct rq in update_rq_clock().
751 * This may result in other CPU reading this CPU's irq time and can
752 * race with irq/account_system_vtime on this CPU. We would either get old
753 * or new value with a side effect of accounting a slice of irq time to wrong
754 * task when irq is in progress while we read rq->clock. That is a worthy
755 * compromise in place of having locks on each irq in account_system_time.
757 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
758 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
760 static DEFINE_PER_CPU(u64
, irq_start_time
);
761 static int sched_clock_irqtime
;
763 void enable_sched_clock_irqtime(void)
765 sched_clock_irqtime
= 1;
768 void disable_sched_clock_irqtime(void)
770 sched_clock_irqtime
= 0;
774 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
776 static inline void irq_time_write_begin(void)
778 __this_cpu_inc(irq_time_seq
.sequence
);
782 static inline void irq_time_write_end(void)
785 __this_cpu_inc(irq_time_seq
.sequence
);
788 static inline u64
irq_time_read(int cpu
)
794 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
795 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
796 per_cpu(cpu_hardirq_time
, cpu
);
797 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
801 #else /* CONFIG_64BIT */
802 static inline void irq_time_write_begin(void)
806 static inline void irq_time_write_end(void)
810 static inline u64
irq_time_read(int cpu
)
812 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
814 #endif /* CONFIG_64BIT */
817 * Called before incrementing preempt_count on {soft,}irq_enter
818 * and before decrementing preempt_count on {soft,}irq_exit.
820 void account_system_vtime(struct task_struct
*curr
)
826 if (!sched_clock_irqtime
)
829 local_irq_save(flags
);
831 cpu
= smp_processor_id();
832 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
833 __this_cpu_add(irq_start_time
, delta
);
835 irq_time_write_begin();
837 * We do not account for softirq time from ksoftirqd here.
838 * We want to continue accounting softirq time to ksoftirqd thread
839 * in that case, so as not to confuse scheduler with a special task
840 * that do not consume any time, but still wants to run.
843 __this_cpu_add(cpu_hardirq_time
, delta
);
844 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
845 __this_cpu_add(cpu_softirq_time
, delta
);
847 irq_time_write_end();
848 local_irq_restore(flags
);
850 EXPORT_SYMBOL_GPL(account_system_vtime
);
852 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
854 #ifdef CONFIG_PARAVIRT
855 static inline u64
steal_ticks(u64 steal
)
857 if (unlikely(steal
> NSEC_PER_SEC
))
858 return div_u64(steal
, TICK_NSEC
);
860 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
864 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
867 * In theory, the compile should just see 0 here, and optimize out the call
868 * to sched_rt_avg_update. But I don't trust it...
870 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
871 s64 steal
= 0, irq_delta
= 0;
873 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
874 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
877 * Since irq_time is only updated on {soft,}irq_exit, we might run into
878 * this case when a previous update_rq_clock() happened inside a
881 * When this happens, we stop ->clock_task and only update the
882 * prev_irq_time stamp to account for the part that fit, so that a next
883 * update will consume the rest. This ensures ->clock_task is
886 * It does however cause some slight miss-attribution of {soft,}irq
887 * time, a more accurate solution would be to update the irq_time using
888 * the current rq->clock timestamp, except that would require using
891 if (irq_delta
> delta
)
894 rq
->prev_irq_time
+= irq_delta
;
897 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
898 if (static_key_false((¶virt_steal_rq_enabled
))) {
901 steal
= paravirt_steal_clock(cpu_of(rq
));
902 steal
-= rq
->prev_steal_time_rq
;
904 if (unlikely(steal
> delta
))
907 st
= steal_ticks(steal
);
908 steal
= st
* TICK_NSEC
;
910 rq
->prev_steal_time_rq
+= steal
;
916 rq
->clock_task
+= delta
;
918 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
919 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
920 sched_rt_avg_update(rq
, irq_delta
+ steal
);
924 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
925 static int irqtime_account_hi_update(void)
927 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
932 local_irq_save(flags
);
933 latest_ns
= this_cpu_read(cpu_hardirq_time
);
934 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_IRQ
])
936 local_irq_restore(flags
);
940 static int irqtime_account_si_update(void)
942 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
947 local_irq_save(flags
);
948 latest_ns
= this_cpu_read(cpu_softirq_time
);
949 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_SOFTIRQ
])
951 local_irq_restore(flags
);
955 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
957 #define sched_clock_irqtime (0)
961 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
963 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
964 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
968 * Make it appear like a SCHED_FIFO task, its something
969 * userspace knows about and won't get confused about.
971 * Also, it will make PI more or less work without too
972 * much confusion -- but then, stop work should not
973 * rely on PI working anyway.
975 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
977 stop
->sched_class
= &stop_sched_class
;
980 cpu_rq(cpu
)->stop
= stop
;
984 * Reset it back to a normal scheduling class so that
985 * it can die in pieces.
987 old_stop
->sched_class
= &rt_sched_class
;
992 * __normal_prio - return the priority that is based on the static prio
994 static inline int __normal_prio(struct task_struct
*p
)
996 return p
->static_prio
;
1000 * Calculate the expected normal priority: i.e. priority
1001 * without taking RT-inheritance into account. Might be
1002 * boosted by interactivity modifiers. Changes upon fork,
1003 * setprio syscalls, and whenever the interactivity
1004 * estimator recalculates.
1006 static inline int normal_prio(struct task_struct
*p
)
1010 if (task_has_rt_policy(p
))
1011 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1013 prio
= __normal_prio(p
);
1018 * Calculate the current priority, i.e. the priority
1019 * taken into account by the scheduler. This value might
1020 * be boosted by RT tasks, or might be boosted by
1021 * interactivity modifiers. Will be RT if the task got
1022 * RT-boosted. If not then it returns p->normal_prio.
1024 static int effective_prio(struct task_struct
*p
)
1026 p
->normal_prio
= normal_prio(p
);
1028 * If we are RT tasks or we were boosted to RT priority,
1029 * keep the priority unchanged. Otherwise, update priority
1030 * to the normal priority:
1032 if (!rt_prio(p
->prio
))
1033 return p
->normal_prio
;
1038 * task_curr - is this task currently executing on a CPU?
1039 * @p: the task in question.
1041 inline int task_curr(const struct task_struct
*p
)
1043 return cpu_curr(task_cpu(p
)) == p
;
1046 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1047 const struct sched_class
*prev_class
,
1050 if (prev_class
!= p
->sched_class
) {
1051 if (prev_class
->switched_from
)
1052 prev_class
->switched_from(rq
, p
);
1053 p
->sched_class
->switched_to(rq
, p
);
1054 } else if (oldprio
!= p
->prio
)
1055 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1058 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1060 const struct sched_class
*class;
1062 if (p
->sched_class
== rq
->curr
->sched_class
) {
1063 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1065 for_each_class(class) {
1066 if (class == rq
->curr
->sched_class
)
1068 if (class == p
->sched_class
) {
1069 resched_task(rq
->curr
);
1076 * A queue event has occurred, and we're going to schedule. In
1077 * this case, we can save a useless back to back clock update.
1079 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1080 rq
->skip_clock_update
= 1;
1084 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1086 #ifdef CONFIG_SCHED_DEBUG
1088 * We should never call set_task_cpu() on a blocked task,
1089 * ttwu() will sort out the placement.
1091 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1092 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
1094 #ifdef CONFIG_LOCKDEP
1096 * The caller should hold either p->pi_lock or rq->lock, when changing
1097 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1099 * sched_move_task() holds both and thus holding either pins the cgroup,
1100 * see set_task_rq().
1102 * Furthermore, all task_rq users should acquire both locks, see
1105 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1106 lockdep_is_held(&task_rq(p
)->lock
)));
1110 trace_sched_migrate_task(p
, new_cpu
);
1112 if (task_cpu(p
) != new_cpu
) {
1113 p
->se
.nr_migrations
++;
1114 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1117 __set_task_cpu(p
, new_cpu
);
1120 struct migration_arg
{
1121 struct task_struct
*task
;
1125 static int migration_cpu_stop(void *data
);
1128 * wait_task_inactive - wait for a thread to unschedule.
1130 * If @match_state is nonzero, it's the @p->state value just checked and
1131 * not expected to change. If it changes, i.e. @p might have woken up,
1132 * then return zero. When we succeed in waiting for @p to be off its CPU,
1133 * we return a positive number (its total switch count). If a second call
1134 * a short while later returns the same number, the caller can be sure that
1135 * @p has remained unscheduled the whole time.
1137 * The caller must ensure that the task *will* unschedule sometime soon,
1138 * else this function might spin for a *long* time. This function can't
1139 * be called with interrupts off, or it may introduce deadlock with
1140 * smp_call_function() if an IPI is sent by the same process we are
1141 * waiting to become inactive.
1143 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1145 unsigned long flags
;
1152 * We do the initial early heuristics without holding
1153 * any task-queue locks at all. We'll only try to get
1154 * the runqueue lock when things look like they will
1160 * If the task is actively running on another CPU
1161 * still, just relax and busy-wait without holding
1164 * NOTE! Since we don't hold any locks, it's not
1165 * even sure that "rq" stays as the right runqueue!
1166 * But we don't care, since "task_running()" will
1167 * return false if the runqueue has changed and p
1168 * is actually now running somewhere else!
1170 while (task_running(rq
, p
)) {
1171 if (match_state
&& unlikely(p
->state
!= match_state
))
1177 * Ok, time to look more closely! We need the rq
1178 * lock now, to be *sure*. If we're wrong, we'll
1179 * just go back and repeat.
1181 rq
= task_rq_lock(p
, &flags
);
1182 trace_sched_wait_task(p
);
1183 running
= task_running(rq
, p
);
1186 if (!match_state
|| p
->state
== match_state
)
1187 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1188 task_rq_unlock(rq
, p
, &flags
);
1191 * If it changed from the expected state, bail out now.
1193 if (unlikely(!ncsw
))
1197 * Was it really running after all now that we
1198 * checked with the proper locks actually held?
1200 * Oops. Go back and try again..
1202 if (unlikely(running
)) {
1208 * It's not enough that it's not actively running,
1209 * it must be off the runqueue _entirely_, and not
1212 * So if it was still runnable (but just not actively
1213 * running right now), it's preempted, and we should
1214 * yield - it could be a while.
1216 if (unlikely(on_rq
)) {
1217 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1219 set_current_state(TASK_UNINTERRUPTIBLE
);
1220 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1225 * Ahh, all good. It wasn't running, and it wasn't
1226 * runnable, which means that it will never become
1227 * running in the future either. We're all done!
1236 * kick_process - kick a running thread to enter/exit the kernel
1237 * @p: the to-be-kicked thread
1239 * Cause a process which is running on another CPU to enter
1240 * kernel-mode, without any delay. (to get signals handled.)
1242 * NOTE: this function doesn't have to take the runqueue lock,
1243 * because all it wants to ensure is that the remote task enters
1244 * the kernel. If the IPI races and the task has been migrated
1245 * to another CPU then no harm is done and the purpose has been
1248 void kick_process(struct task_struct
*p
)
1254 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1255 smp_send_reschedule(cpu
);
1258 EXPORT_SYMBOL_GPL(kick_process
);
1259 #endif /* CONFIG_SMP */
1263 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1265 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1268 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
1272 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1275 /* Any allowed, online CPU? */
1276 dest_cpu
= cpumask_any_and(tsk_cpus_allowed(p
), cpu_active_mask
);
1277 if (dest_cpu
< nr_cpu_ids
)
1280 /* No more Mr. Nice Guy. */
1281 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
1283 * Don't tell them about moving exiting tasks or
1284 * kernel threads (both mm NULL), since they never
1287 if (p
->mm
&& printk_ratelimit()) {
1288 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1289 task_pid_nr(p
), p
->comm
, cpu
);
1296 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1299 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1301 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1304 * In order not to call set_task_cpu() on a blocking task we need
1305 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1308 * Since this is common to all placement strategies, this lives here.
1310 * [ this allows ->select_task() to simply return task_cpu(p) and
1311 * not worry about this generic constraint ]
1313 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1315 cpu
= select_fallback_rq(task_cpu(p
), p
);
1320 static void update_avg(u64
*avg
, u64 sample
)
1322 s64 diff
= sample
- *avg
;
1328 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1330 #ifdef CONFIG_SCHEDSTATS
1331 struct rq
*rq
= this_rq();
1334 int this_cpu
= smp_processor_id();
1336 if (cpu
== this_cpu
) {
1337 schedstat_inc(rq
, ttwu_local
);
1338 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1340 struct sched_domain
*sd
;
1342 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1344 for_each_domain(this_cpu
, sd
) {
1345 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1346 schedstat_inc(sd
, ttwu_wake_remote
);
1353 if (wake_flags
& WF_MIGRATED
)
1354 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1356 #endif /* CONFIG_SMP */
1358 schedstat_inc(rq
, ttwu_count
);
1359 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1361 if (wake_flags
& WF_SYNC
)
1362 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1364 #endif /* CONFIG_SCHEDSTATS */
1367 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1369 activate_task(rq
, p
, en_flags
);
1372 /* if a worker is waking up, notify workqueue */
1373 if (p
->flags
& PF_WQ_WORKER
)
1374 wq_worker_waking_up(p
, cpu_of(rq
));
1378 * Mark the task runnable and perform wakeup-preemption.
1381 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1383 trace_sched_wakeup(p
, true);
1384 check_preempt_curr(rq
, p
, wake_flags
);
1386 p
->state
= TASK_RUNNING
;
1388 if (p
->sched_class
->task_woken
)
1389 p
->sched_class
->task_woken(rq
, p
);
1391 if (rq
->idle_stamp
) {
1392 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1393 u64 max
= 2*sysctl_sched_migration_cost
;
1398 update_avg(&rq
->avg_idle
, delta
);
1405 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1408 if (p
->sched_contributes_to_load
)
1409 rq
->nr_uninterruptible
--;
1412 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1413 ttwu_do_wakeup(rq
, p
, wake_flags
);
1417 * Called in case the task @p isn't fully descheduled from its runqueue,
1418 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1419 * since all we need to do is flip p->state to TASK_RUNNING, since
1420 * the task is still ->on_rq.
1422 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1427 rq
= __task_rq_lock(p
);
1429 ttwu_do_wakeup(rq
, p
, wake_flags
);
1432 __task_rq_unlock(rq
);
1438 static void sched_ttwu_pending(void)
1440 struct rq
*rq
= this_rq();
1441 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1442 struct task_struct
*p
;
1444 raw_spin_lock(&rq
->lock
);
1447 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1448 llist
= llist_next(llist
);
1449 ttwu_do_activate(rq
, p
, 0);
1452 raw_spin_unlock(&rq
->lock
);
1455 void scheduler_ipi(void)
1457 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1461 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1462 * traditionally all their work was done from the interrupt return
1463 * path. Now that we actually do some work, we need to make sure
1466 * Some archs already do call them, luckily irq_enter/exit nest
1469 * Arguably we should visit all archs and update all handlers,
1470 * however a fair share of IPIs are still resched only so this would
1471 * somewhat pessimize the simple resched case.
1474 sched_ttwu_pending();
1477 * Check if someone kicked us for doing the nohz idle load balance.
1479 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1480 this_rq()->idle_balance
= 1;
1481 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1486 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1488 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1489 smp_send_reschedule(cpu
);
1492 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1493 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1498 rq
= __task_rq_lock(p
);
1500 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1501 ttwu_do_wakeup(rq
, p
, wake_flags
);
1504 __task_rq_unlock(rq
);
1509 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1511 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1513 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1515 #endif /* CONFIG_SMP */
1517 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1519 struct rq
*rq
= cpu_rq(cpu
);
1521 #if defined(CONFIG_SMP)
1522 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1523 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1524 ttwu_queue_remote(p
, cpu
);
1529 raw_spin_lock(&rq
->lock
);
1530 ttwu_do_activate(rq
, p
, 0);
1531 raw_spin_unlock(&rq
->lock
);
1535 * try_to_wake_up - wake up a thread
1536 * @p: the thread to be awakened
1537 * @state: the mask of task states that can be woken
1538 * @wake_flags: wake modifier flags (WF_*)
1540 * Put it on the run-queue if it's not already there. The "current"
1541 * thread is always on the run-queue (except when the actual
1542 * re-schedule is in progress), and as such you're allowed to do
1543 * the simpler "current->state = TASK_RUNNING" to mark yourself
1544 * runnable without the overhead of this.
1546 * Returns %true if @p was woken up, %false if it was already running
1547 * or @state didn't match @p's state.
1550 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1552 unsigned long flags
;
1553 int cpu
, success
= 0;
1556 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1557 if (!(p
->state
& state
))
1560 success
= 1; /* we're going to change ->state */
1563 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1568 * If the owning (remote) cpu is still in the middle of schedule() with
1569 * this task as prev, wait until its done referencing the task.
1572 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1574 * In case the architecture enables interrupts in
1575 * context_switch(), we cannot busy wait, since that
1576 * would lead to deadlocks when an interrupt hits and
1577 * tries to wake up @prev. So bail and do a complete
1580 if (ttwu_activate_remote(p
, wake_flags
))
1587 * Pairs with the smp_wmb() in finish_lock_switch().
1591 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1592 p
->state
= TASK_WAKING
;
1594 if (p
->sched_class
->task_waking
)
1595 p
->sched_class
->task_waking(p
);
1597 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1598 if (task_cpu(p
) != cpu
) {
1599 wake_flags
|= WF_MIGRATED
;
1600 set_task_cpu(p
, cpu
);
1602 #endif /* CONFIG_SMP */
1606 ttwu_stat(p
, cpu
, wake_flags
);
1608 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1614 * try_to_wake_up_local - try to wake up a local task with rq lock held
1615 * @p: the thread to be awakened
1617 * Put @p on the run-queue if it's not already there. The caller must
1618 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1621 static void try_to_wake_up_local(struct task_struct
*p
)
1623 struct rq
*rq
= task_rq(p
);
1625 BUG_ON(rq
!= this_rq());
1626 BUG_ON(p
== current
);
1627 lockdep_assert_held(&rq
->lock
);
1629 if (!raw_spin_trylock(&p
->pi_lock
)) {
1630 raw_spin_unlock(&rq
->lock
);
1631 raw_spin_lock(&p
->pi_lock
);
1632 raw_spin_lock(&rq
->lock
);
1635 if (!(p
->state
& TASK_NORMAL
))
1639 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1641 ttwu_do_wakeup(rq
, p
, 0);
1642 ttwu_stat(p
, smp_processor_id(), 0);
1644 raw_spin_unlock(&p
->pi_lock
);
1648 * wake_up_process - Wake up a specific process
1649 * @p: The process to be woken up.
1651 * Attempt to wake up the nominated process and move it to the set of runnable
1652 * processes. Returns 1 if the process was woken up, 0 if it was already
1655 * It may be assumed that this function implies a write memory barrier before
1656 * changing the task state if and only if any tasks are woken up.
1658 int wake_up_process(struct task_struct
*p
)
1660 return try_to_wake_up(p
, TASK_ALL
, 0);
1662 EXPORT_SYMBOL(wake_up_process
);
1664 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1666 return try_to_wake_up(p
, state
, 0);
1670 * Perform scheduler related setup for a newly forked process p.
1671 * p is forked by current.
1673 * __sched_fork() is basic setup used by init_idle() too:
1675 static void __sched_fork(struct task_struct
*p
)
1680 p
->se
.exec_start
= 0;
1681 p
->se
.sum_exec_runtime
= 0;
1682 p
->se
.prev_sum_exec_runtime
= 0;
1683 p
->se
.nr_migrations
= 0;
1685 INIT_LIST_HEAD(&p
->se
.group_node
);
1687 #ifdef CONFIG_SCHEDSTATS
1688 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1691 INIT_LIST_HEAD(&p
->rt
.run_list
);
1693 #ifdef CONFIG_PREEMPT_NOTIFIERS
1694 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1699 * fork()/clone()-time setup:
1701 void sched_fork(struct task_struct
*p
)
1703 unsigned long flags
;
1704 int cpu
= get_cpu();
1708 * We mark the process as running here. This guarantees that
1709 * nobody will actually run it, and a signal or other external
1710 * event cannot wake it up and insert it on the runqueue either.
1712 p
->state
= TASK_RUNNING
;
1715 * Make sure we do not leak PI boosting priority to the child.
1717 p
->prio
= current
->normal_prio
;
1720 * Revert to default priority/policy on fork if requested.
1722 if (unlikely(p
->sched_reset_on_fork
)) {
1723 if (task_has_rt_policy(p
)) {
1724 p
->policy
= SCHED_NORMAL
;
1725 p
->static_prio
= NICE_TO_PRIO(0);
1727 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1728 p
->static_prio
= NICE_TO_PRIO(0);
1730 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1734 * We don't need the reset flag anymore after the fork. It has
1735 * fulfilled its duty:
1737 p
->sched_reset_on_fork
= 0;
1740 if (!rt_prio(p
->prio
))
1741 p
->sched_class
= &fair_sched_class
;
1743 if (p
->sched_class
->task_fork
)
1744 p
->sched_class
->task_fork(p
);
1747 * The child is not yet in the pid-hash so no cgroup attach races,
1748 * and the cgroup is pinned to this child due to cgroup_fork()
1749 * is ran before sched_fork().
1751 * Silence PROVE_RCU.
1753 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1754 set_task_cpu(p
, cpu
);
1755 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1757 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1758 if (likely(sched_info_on()))
1759 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1761 #if defined(CONFIG_SMP)
1764 #ifdef CONFIG_PREEMPT_COUNT
1765 /* Want to start with kernel preemption disabled. */
1766 task_thread_info(p
)->preempt_count
= 1;
1769 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1776 * wake_up_new_task - wake up a newly created task for the first time.
1778 * This function will do some initial scheduler statistics housekeeping
1779 * that must be done for every newly created context, then puts the task
1780 * on the runqueue and wakes it.
1782 void wake_up_new_task(struct task_struct
*p
)
1784 unsigned long flags
;
1787 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1790 * Fork balancing, do it here and not earlier because:
1791 * - cpus_allowed can change in the fork path
1792 * - any previously selected cpu might disappear through hotplug
1794 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1797 rq
= __task_rq_lock(p
);
1798 activate_task(rq
, p
, 0);
1800 trace_sched_wakeup_new(p
, true);
1801 check_preempt_curr(rq
, p
, WF_FORK
);
1803 if (p
->sched_class
->task_woken
)
1804 p
->sched_class
->task_woken(rq
, p
);
1806 task_rq_unlock(rq
, p
, &flags
);
1809 #ifdef CONFIG_PREEMPT_NOTIFIERS
1812 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1813 * @notifier: notifier struct to register
1815 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1817 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1819 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1822 * preempt_notifier_unregister - no longer interested in preemption notifications
1823 * @notifier: notifier struct to unregister
1825 * This is safe to call from within a preemption notifier.
1827 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1829 hlist_del(¬ifier
->link
);
1831 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1833 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1835 struct preempt_notifier
*notifier
;
1836 struct hlist_node
*node
;
1838 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1839 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1843 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1844 struct task_struct
*next
)
1846 struct preempt_notifier
*notifier
;
1847 struct hlist_node
*node
;
1849 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1850 notifier
->ops
->sched_out(notifier
, next
);
1853 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1855 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1860 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1861 struct task_struct
*next
)
1865 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1868 * prepare_task_switch - prepare to switch tasks
1869 * @rq: the runqueue preparing to switch
1870 * @prev: the current task that is being switched out
1871 * @next: the task we are going to switch to.
1873 * This is called with the rq lock held and interrupts off. It must
1874 * be paired with a subsequent finish_task_switch after the context
1877 * prepare_task_switch sets up locking and calls architecture specific
1881 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1882 struct task_struct
*next
)
1884 sched_info_switch(prev
, next
);
1885 perf_event_task_sched_out(prev
, next
);
1886 fire_sched_out_preempt_notifiers(prev
, next
);
1887 prepare_lock_switch(rq
, next
);
1888 prepare_arch_switch(next
);
1889 trace_sched_switch(prev
, next
);
1893 * finish_task_switch - clean up after a task-switch
1894 * @rq: runqueue associated with task-switch
1895 * @prev: the thread we just switched away from.
1897 * finish_task_switch must be called after the context switch, paired
1898 * with a prepare_task_switch call before the context switch.
1899 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1900 * and do any other architecture-specific cleanup actions.
1902 * Note that we may have delayed dropping an mm in context_switch(). If
1903 * so, we finish that here outside of the runqueue lock. (Doing it
1904 * with the lock held can cause deadlocks; see schedule() for
1907 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1908 __releases(rq
->lock
)
1910 struct mm_struct
*mm
= rq
->prev_mm
;
1916 * A task struct has one reference for the use as "current".
1917 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1918 * schedule one last time. The schedule call will never return, and
1919 * the scheduled task must drop that reference.
1920 * The test for TASK_DEAD must occur while the runqueue locks are
1921 * still held, otherwise prev could be scheduled on another cpu, die
1922 * there before we look at prev->state, and then the reference would
1924 * Manfred Spraul <manfred@colorfullife.com>
1926 prev_state
= prev
->state
;
1927 finish_arch_switch(prev
);
1928 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1929 local_irq_disable();
1930 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1931 perf_event_task_sched_in(prev
, current
);
1932 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1934 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1935 finish_lock_switch(rq
, prev
);
1937 fire_sched_in_preempt_notifiers(current
);
1940 if (unlikely(prev_state
== TASK_DEAD
)) {
1942 * Remove function-return probe instances associated with this
1943 * task and put them back on the free list.
1945 kprobe_flush_task(prev
);
1946 put_task_struct(prev
);
1952 /* assumes rq->lock is held */
1953 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1955 if (prev
->sched_class
->pre_schedule
)
1956 prev
->sched_class
->pre_schedule(rq
, prev
);
1959 /* rq->lock is NOT held, but preemption is disabled */
1960 static inline void post_schedule(struct rq
*rq
)
1962 if (rq
->post_schedule
) {
1963 unsigned long flags
;
1965 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1966 if (rq
->curr
->sched_class
->post_schedule
)
1967 rq
->curr
->sched_class
->post_schedule(rq
);
1968 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1970 rq
->post_schedule
= 0;
1976 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1980 static inline void post_schedule(struct rq
*rq
)
1987 * schedule_tail - first thing a freshly forked thread must call.
1988 * @prev: the thread we just switched away from.
1990 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1991 __releases(rq
->lock
)
1993 struct rq
*rq
= this_rq();
1995 finish_task_switch(rq
, prev
);
1998 * FIXME: do we need to worry about rq being invalidated by the
2003 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2004 /* In this case, finish_task_switch does not reenable preemption */
2007 if (current
->set_child_tid
)
2008 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2012 * context_switch - switch to the new MM and the new
2013 * thread's register state.
2016 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2017 struct task_struct
*next
)
2019 struct mm_struct
*mm
, *oldmm
;
2021 prepare_task_switch(rq
, prev
, next
);
2024 oldmm
= prev
->active_mm
;
2026 * For paravirt, this is coupled with an exit in switch_to to
2027 * combine the page table reload and the switch backend into
2030 arch_start_context_switch(prev
);
2033 next
->active_mm
= oldmm
;
2034 atomic_inc(&oldmm
->mm_count
);
2035 enter_lazy_tlb(oldmm
, next
);
2037 switch_mm(oldmm
, mm
, next
);
2040 prev
->active_mm
= NULL
;
2041 rq
->prev_mm
= oldmm
;
2044 * Since the runqueue lock will be released by the next
2045 * task (which is an invalid locking op but in the case
2046 * of the scheduler it's an obvious special-case), so we
2047 * do an early lockdep release here:
2049 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2050 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2053 /* Here we just switch the register state and the stack. */
2054 switch_to(prev
, next
, prev
);
2058 * this_rq must be evaluated again because prev may have moved
2059 * CPUs since it called schedule(), thus the 'rq' on its stack
2060 * frame will be invalid.
2062 finish_task_switch(this_rq(), prev
);
2066 * nr_running, nr_uninterruptible and nr_context_switches:
2068 * externally visible scheduler statistics: current number of runnable
2069 * threads, current number of uninterruptible-sleeping threads, total
2070 * number of context switches performed since bootup.
2072 unsigned long nr_running(void)
2074 unsigned long i
, sum
= 0;
2076 for_each_online_cpu(i
)
2077 sum
+= cpu_rq(i
)->nr_running
;
2082 unsigned long nr_uninterruptible(void)
2084 unsigned long i
, sum
= 0;
2086 for_each_possible_cpu(i
)
2087 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2090 * Since we read the counters lockless, it might be slightly
2091 * inaccurate. Do not allow it to go below zero though:
2093 if (unlikely((long)sum
< 0))
2099 unsigned long long nr_context_switches(void)
2102 unsigned long long sum
= 0;
2104 for_each_possible_cpu(i
)
2105 sum
+= cpu_rq(i
)->nr_switches
;
2110 unsigned long nr_iowait(void)
2112 unsigned long i
, sum
= 0;
2114 for_each_possible_cpu(i
)
2115 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2120 unsigned long nr_iowait_cpu(int cpu
)
2122 struct rq
*this = cpu_rq(cpu
);
2123 return atomic_read(&this->nr_iowait
);
2126 unsigned long this_cpu_load(void)
2128 struct rq
*this = this_rq();
2129 return this->cpu_load
[0];
2133 /* Variables and functions for calc_load */
2134 static atomic_long_t calc_load_tasks
;
2135 static unsigned long calc_load_update
;
2136 unsigned long avenrun
[3];
2137 EXPORT_SYMBOL(avenrun
);
2139 static long calc_load_fold_active(struct rq
*this_rq
)
2141 long nr_active
, delta
= 0;
2143 nr_active
= this_rq
->nr_running
;
2144 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2146 if (nr_active
!= this_rq
->calc_load_active
) {
2147 delta
= nr_active
- this_rq
->calc_load_active
;
2148 this_rq
->calc_load_active
= nr_active
;
2154 static unsigned long
2155 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2158 load
+= active
* (FIXED_1
- exp
);
2159 load
+= 1UL << (FSHIFT
- 1);
2160 return load
>> FSHIFT
;
2165 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2167 * When making the ILB scale, we should try to pull this in as well.
2169 static atomic_long_t calc_load_tasks_idle
;
2171 void calc_load_account_idle(struct rq
*this_rq
)
2175 delta
= calc_load_fold_active(this_rq
);
2177 atomic_long_add(delta
, &calc_load_tasks_idle
);
2180 static long calc_load_fold_idle(void)
2185 * Its got a race, we don't care...
2187 if (atomic_long_read(&calc_load_tasks_idle
))
2188 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2194 * fixed_power_int - compute: x^n, in O(log n) time
2196 * @x: base of the power
2197 * @frac_bits: fractional bits of @x
2198 * @n: power to raise @x to.
2200 * By exploiting the relation between the definition of the natural power
2201 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2202 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2203 * (where: n_i \elem {0, 1}, the binary vector representing n),
2204 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2205 * of course trivially computable in O(log_2 n), the length of our binary
2208 static unsigned long
2209 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2211 unsigned long result
= 1UL << frac_bits
;
2216 result
+= 1UL << (frac_bits
- 1);
2217 result
>>= frac_bits
;
2223 x
+= 1UL << (frac_bits
- 1);
2231 * a1 = a0 * e + a * (1 - e)
2233 * a2 = a1 * e + a * (1 - e)
2234 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2235 * = a0 * e^2 + a * (1 - e) * (1 + e)
2237 * a3 = a2 * e + a * (1 - e)
2238 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2239 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2243 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2244 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2245 * = a0 * e^n + a * (1 - e^n)
2247 * [1] application of the geometric series:
2250 * S_n := \Sum x^i = -------------
2253 static unsigned long
2254 calc_load_n(unsigned long load
, unsigned long exp
,
2255 unsigned long active
, unsigned int n
)
2258 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2262 * NO_HZ can leave us missing all per-cpu ticks calling
2263 * calc_load_account_active(), but since an idle CPU folds its delta into
2264 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2265 * in the pending idle delta if our idle period crossed a load cycle boundary.
2267 * Once we've updated the global active value, we need to apply the exponential
2268 * weights adjusted to the number of cycles missed.
2270 static void calc_global_nohz(void)
2272 long delta
, active
, n
;
2275 * If we crossed a calc_load_update boundary, make sure to fold
2276 * any pending idle changes, the respective CPUs might have
2277 * missed the tick driven calc_load_account_active() update
2280 delta
= calc_load_fold_idle();
2282 atomic_long_add(delta
, &calc_load_tasks
);
2285 * It could be the one fold was all it took, we done!
2287 if (time_before(jiffies
, calc_load_update
+ 10))
2291 * Catch-up, fold however many we are behind still
2293 delta
= jiffies
- calc_load_update
- 10;
2294 n
= 1 + (delta
/ LOAD_FREQ
);
2296 active
= atomic_long_read(&calc_load_tasks
);
2297 active
= active
> 0 ? active
* FIXED_1
: 0;
2299 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2300 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2301 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2303 calc_load_update
+= n
* LOAD_FREQ
;
2306 void calc_load_account_idle(struct rq
*this_rq
)
2310 static inline long calc_load_fold_idle(void)
2315 static void calc_global_nohz(void)
2321 * get_avenrun - get the load average array
2322 * @loads: pointer to dest load array
2323 * @offset: offset to add
2324 * @shift: shift count to shift the result left
2326 * These values are estimates at best, so no need for locking.
2328 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2330 loads
[0] = (avenrun
[0] + offset
) << shift
;
2331 loads
[1] = (avenrun
[1] + offset
) << shift
;
2332 loads
[2] = (avenrun
[2] + offset
) << shift
;
2336 * calc_load - update the avenrun load estimates 10 ticks after the
2337 * CPUs have updated calc_load_tasks.
2339 void calc_global_load(unsigned long ticks
)
2343 if (time_before(jiffies
, calc_load_update
+ 10))
2346 active
= atomic_long_read(&calc_load_tasks
);
2347 active
= active
> 0 ? active
* FIXED_1
: 0;
2349 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2350 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2351 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2353 calc_load_update
+= LOAD_FREQ
;
2356 * Account one period with whatever state we found before
2357 * folding in the nohz state and ageing the entire idle period.
2359 * This avoids loosing a sample when we go idle between
2360 * calc_load_account_active() (10 ticks ago) and now and thus
2367 * Called from update_cpu_load() to periodically update this CPU's
2370 static void calc_load_account_active(struct rq
*this_rq
)
2374 if (time_before(jiffies
, this_rq
->calc_load_update
))
2377 delta
= calc_load_fold_active(this_rq
);
2378 delta
+= calc_load_fold_idle();
2380 atomic_long_add(delta
, &calc_load_tasks
);
2382 this_rq
->calc_load_update
+= LOAD_FREQ
;
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 void update_cpu_load(struct rq
*this_rq
)
2459 unsigned long this_load
= this_rq
->load
.weight
;
2460 unsigned long curr_jiffies
= jiffies
;
2461 unsigned long pending_updates
;
2464 this_rq
->nr_load_updates
++;
2466 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2467 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2470 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2471 this_rq
->last_load_update_tick
= curr_jiffies
;
2473 /* Update our load: */
2474 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2475 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2476 unsigned long old_load
, new_load
;
2478 /* scale is effectively 1 << i now, and >> i divides by scale */
2480 old_load
= this_rq
->cpu_load
[i
];
2481 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2482 new_load
= this_load
;
2484 * Round up the averaging division if load is increasing. This
2485 * prevents us from getting stuck on 9 if the load is 10, for
2488 if (new_load
> old_load
)
2489 new_load
+= scale
- 1;
2491 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2494 sched_avg_update(this_rq
);
2497 static void update_cpu_load_active(struct rq
*this_rq
)
2499 update_cpu_load(this_rq
);
2501 calc_load_account_active(this_rq
);
2507 * sched_exec - execve() is a valuable balancing opportunity, because at
2508 * this point the task has the smallest effective memory and cache footprint.
2510 void sched_exec(void)
2512 struct task_struct
*p
= current
;
2513 unsigned long flags
;
2516 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2517 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2518 if (dest_cpu
== smp_processor_id())
2521 if (likely(cpu_active(dest_cpu
))) {
2522 struct migration_arg arg
= { p
, dest_cpu
};
2524 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2525 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2529 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2534 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2535 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2537 EXPORT_PER_CPU_SYMBOL(kstat
);
2538 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2541 * Return any ns on the sched_clock that have not yet been accounted in
2542 * @p in case that task is currently running.
2544 * Called with task_rq_lock() held on @rq.
2546 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2550 if (task_current(rq
, p
)) {
2551 update_rq_clock(rq
);
2552 ns
= rq
->clock_task
- p
->se
.exec_start
;
2560 unsigned long long task_delta_exec(struct task_struct
*p
)
2562 unsigned long flags
;
2566 rq
= task_rq_lock(p
, &flags
);
2567 ns
= do_task_delta_exec(p
, rq
);
2568 task_rq_unlock(rq
, p
, &flags
);
2574 * Return accounted runtime for the task.
2575 * In case the task is currently running, return the runtime plus current's
2576 * pending runtime that have not been accounted yet.
2578 unsigned long long task_sched_runtime(struct task_struct
*p
)
2580 unsigned long flags
;
2584 rq
= task_rq_lock(p
, &flags
);
2585 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2586 task_rq_unlock(rq
, p
, &flags
);
2591 #ifdef CONFIG_CGROUP_CPUACCT
2592 struct cgroup_subsys cpuacct_subsys
;
2593 struct cpuacct root_cpuacct
;
2596 static inline void task_group_account_field(struct task_struct
*p
, int index
,
2599 #ifdef CONFIG_CGROUP_CPUACCT
2600 struct kernel_cpustat
*kcpustat
;
2604 * Since all updates are sure to touch the root cgroup, we
2605 * get ourselves ahead and touch it first. If the root cgroup
2606 * is the only cgroup, then nothing else should be necessary.
2609 __get_cpu_var(kernel_cpustat
).cpustat
[index
] += tmp
;
2611 #ifdef CONFIG_CGROUP_CPUACCT
2612 if (unlikely(!cpuacct_subsys
.active
))
2617 while (ca
&& (ca
!= &root_cpuacct
)) {
2618 kcpustat
= this_cpu_ptr(ca
->cpustat
);
2619 kcpustat
->cpustat
[index
] += tmp
;
2628 * Account user cpu time to a process.
2629 * @p: the process that the cpu time gets accounted to
2630 * @cputime: the cpu time spent in user space since the last update
2631 * @cputime_scaled: cputime scaled by cpu frequency
2633 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
2634 cputime_t cputime_scaled
)
2638 /* Add user time to process. */
2639 p
->utime
+= cputime
;
2640 p
->utimescaled
+= cputime_scaled
;
2641 account_group_user_time(p
, cputime
);
2643 index
= (TASK_NICE(p
) > 0) ? CPUTIME_NICE
: CPUTIME_USER
;
2645 /* Add user time to cpustat. */
2646 task_group_account_field(p
, index
, (__force u64
) cputime
);
2648 /* Account for user time used */
2649 acct_update_integrals(p
);
2653 * Account guest cpu time to a process.
2654 * @p: the process that the cpu time gets accounted to
2655 * @cputime: the cpu time spent in virtual machine since the last update
2656 * @cputime_scaled: cputime scaled by cpu frequency
2658 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
2659 cputime_t cputime_scaled
)
2661 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2663 /* Add guest time to process. */
2664 p
->utime
+= cputime
;
2665 p
->utimescaled
+= cputime_scaled
;
2666 account_group_user_time(p
, cputime
);
2667 p
->gtime
+= cputime
;
2669 /* Add guest time to cpustat. */
2670 if (TASK_NICE(p
) > 0) {
2671 cpustat
[CPUTIME_NICE
] += (__force u64
) cputime
;
2672 cpustat
[CPUTIME_GUEST_NICE
] += (__force u64
) cputime
;
2674 cpustat
[CPUTIME_USER
] += (__force u64
) cputime
;
2675 cpustat
[CPUTIME_GUEST
] += (__force u64
) cputime
;
2680 * Account system cpu time to a process and desired cpustat field
2681 * @p: the process that the cpu time gets accounted to
2682 * @cputime: the cpu time spent in kernel space since the last update
2683 * @cputime_scaled: cputime scaled by cpu frequency
2684 * @target_cputime64: pointer to cpustat field that has to be updated
2687 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
2688 cputime_t cputime_scaled
, int index
)
2690 /* Add system time to process. */
2691 p
->stime
+= cputime
;
2692 p
->stimescaled
+= cputime_scaled
;
2693 account_group_system_time(p
, cputime
);
2695 /* Add system time to cpustat. */
2696 task_group_account_field(p
, index
, (__force u64
) cputime
);
2698 /* Account for system time used */
2699 acct_update_integrals(p
);
2703 * Account system cpu time to a process.
2704 * @p: the process that the cpu time gets accounted to
2705 * @hardirq_offset: the offset to subtract from hardirq_count()
2706 * @cputime: the cpu time spent in kernel space since the last update
2707 * @cputime_scaled: cputime scaled by cpu frequency
2709 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2710 cputime_t cputime
, cputime_t cputime_scaled
)
2714 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
2715 account_guest_time(p
, cputime
, cputime_scaled
);
2719 if (hardirq_count() - hardirq_offset
)
2720 index
= CPUTIME_IRQ
;
2721 else if (in_serving_softirq())
2722 index
= CPUTIME_SOFTIRQ
;
2724 index
= CPUTIME_SYSTEM
;
2726 __account_system_time(p
, cputime
, cputime_scaled
, index
);
2730 * Account for involuntary wait time.
2731 * @cputime: the cpu time spent in involuntary wait
2733 void account_steal_time(cputime_t cputime
)
2735 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2737 cpustat
[CPUTIME_STEAL
] += (__force u64
) cputime
;
2741 * Account for idle time.
2742 * @cputime: the cpu time spent in idle wait
2744 void account_idle_time(cputime_t cputime
)
2746 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2747 struct rq
*rq
= this_rq();
2749 if (atomic_read(&rq
->nr_iowait
) > 0)
2750 cpustat
[CPUTIME_IOWAIT
] += (__force u64
) cputime
;
2752 cpustat
[CPUTIME_IDLE
] += (__force u64
) cputime
;
2755 static __always_inline
bool steal_account_process_tick(void)
2757 #ifdef CONFIG_PARAVIRT
2758 if (static_key_false(¶virt_steal_enabled
)) {
2761 steal
= paravirt_steal_clock(smp_processor_id());
2762 steal
-= this_rq()->prev_steal_time
;
2764 st
= steal_ticks(steal
);
2765 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
2767 account_steal_time(st
);
2774 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2776 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2778 * Account a tick to a process and cpustat
2779 * @p: the process that the cpu time gets accounted to
2780 * @user_tick: is the tick from userspace
2781 * @rq: the pointer to rq
2783 * Tick demultiplexing follows the order
2784 * - pending hardirq update
2785 * - pending softirq update
2789 * - check for guest_time
2790 * - else account as system_time
2792 * Check for hardirq is done both for system and user time as there is
2793 * no timer going off while we are on hardirq and hence we may never get an
2794 * opportunity to update it solely in system time.
2795 * p->stime and friends are only updated on system time and not on irq
2796 * softirq as those do not count in task exec_runtime any more.
2798 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2801 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2802 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2804 if (steal_account_process_tick())
2807 if (irqtime_account_hi_update()) {
2808 cpustat
[CPUTIME_IRQ
] += (__force u64
) cputime_one_jiffy
;
2809 } else if (irqtime_account_si_update()) {
2810 cpustat
[CPUTIME_SOFTIRQ
] += (__force u64
) cputime_one_jiffy
;
2811 } else if (this_cpu_ksoftirqd() == p
) {
2813 * ksoftirqd time do not get accounted in cpu_softirq_time.
2814 * So, we have to handle it separately here.
2815 * Also, p->stime needs to be updated for ksoftirqd.
2817 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2819 } else if (user_tick
) {
2820 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2821 } else if (p
== rq
->idle
) {
2822 account_idle_time(cputime_one_jiffy
);
2823 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
2824 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2826 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2831 static void irqtime_account_idle_ticks(int ticks
)
2834 struct rq
*rq
= this_rq();
2836 for (i
= 0; i
< ticks
; i
++)
2837 irqtime_account_process_tick(current
, 0, rq
);
2839 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2840 static void irqtime_account_idle_ticks(int ticks
) {}
2841 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2843 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2846 * Account a single tick of cpu time.
2847 * @p: the process that the cpu time gets accounted to
2848 * @user_tick: indicates if the tick is a user or a system tick
2850 void account_process_tick(struct task_struct
*p
, int user_tick
)
2852 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2853 struct rq
*rq
= this_rq();
2855 if (sched_clock_irqtime
) {
2856 irqtime_account_process_tick(p
, user_tick
, rq
);
2860 if (steal_account_process_tick())
2864 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2865 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
2866 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
2869 account_idle_time(cputime_one_jiffy
);
2873 * Account multiple ticks of steal time.
2874 * @p: the process from which the cpu time has been stolen
2875 * @ticks: number of stolen ticks
2877 void account_steal_ticks(unsigned long ticks
)
2879 account_steal_time(jiffies_to_cputime(ticks
));
2883 * Account multiple ticks of idle time.
2884 * @ticks: number of stolen ticks
2886 void account_idle_ticks(unsigned long ticks
)
2889 if (sched_clock_irqtime
) {
2890 irqtime_account_idle_ticks(ticks
);
2894 account_idle_time(jiffies_to_cputime(ticks
));
2900 * Use precise platform statistics if available:
2902 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2903 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2909 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2911 struct task_cputime cputime
;
2913 thread_group_cputime(p
, &cputime
);
2915 *ut
= cputime
.utime
;
2916 *st
= cputime
.stime
;
2920 #ifndef nsecs_to_cputime
2921 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2924 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2926 cputime_t rtime
, utime
= p
->utime
, total
= utime
+ p
->stime
;
2929 * Use CFS's precise accounting:
2931 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
2934 u64 temp
= (__force u64
) rtime
;
2936 temp
*= (__force u64
) utime
;
2937 do_div(temp
, (__force u32
) total
);
2938 utime
= (__force cputime_t
) temp
;
2943 * Compare with previous values, to keep monotonicity:
2945 p
->prev_utime
= max(p
->prev_utime
, utime
);
2946 p
->prev_stime
= max(p
->prev_stime
, rtime
- p
->prev_utime
);
2948 *ut
= p
->prev_utime
;
2949 *st
= p
->prev_stime
;
2953 * Must be called with siglock held.
2955 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2957 struct signal_struct
*sig
= p
->signal
;
2958 struct task_cputime cputime
;
2959 cputime_t rtime
, utime
, total
;
2961 thread_group_cputime(p
, &cputime
);
2963 total
= cputime
.utime
+ cputime
.stime
;
2964 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
2967 u64 temp
= (__force u64
) rtime
;
2969 temp
*= (__force u64
) cputime
.utime
;
2970 do_div(temp
, (__force u32
) total
);
2971 utime
= (__force cputime_t
) temp
;
2975 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
2976 sig
->prev_stime
= max(sig
->prev_stime
, rtime
- sig
->prev_utime
);
2978 *ut
= sig
->prev_utime
;
2979 *st
= sig
->prev_stime
;
2984 * This function gets called by the timer code, with HZ frequency.
2985 * We call it with interrupts disabled.
2987 void scheduler_tick(void)
2989 int cpu
= smp_processor_id();
2990 struct rq
*rq
= cpu_rq(cpu
);
2991 struct task_struct
*curr
= rq
->curr
;
2995 raw_spin_lock(&rq
->lock
);
2996 update_rq_clock(rq
);
2997 update_cpu_load_active(rq
);
2998 curr
->sched_class
->task_tick(rq
, curr
, 0);
2999 raw_spin_unlock(&rq
->lock
);
3001 perf_event_task_tick();
3004 rq
->idle_balance
= idle_cpu(cpu
);
3005 trigger_load_balance(rq
, cpu
);
3009 notrace
unsigned long get_parent_ip(unsigned long addr
)
3011 if (in_lock_functions(addr
)) {
3012 addr
= CALLER_ADDR2
;
3013 if (in_lock_functions(addr
))
3014 addr
= CALLER_ADDR3
;
3019 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3020 defined(CONFIG_PREEMPT_TRACER))
3022 void __kprobes
add_preempt_count(int val
)
3024 #ifdef CONFIG_DEBUG_PREEMPT
3028 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3031 preempt_count() += val
;
3032 #ifdef CONFIG_DEBUG_PREEMPT
3034 * Spinlock count overflowing soon?
3036 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3039 if (preempt_count() == val
)
3040 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3042 EXPORT_SYMBOL(add_preempt_count
);
3044 void __kprobes
sub_preempt_count(int val
)
3046 #ifdef CONFIG_DEBUG_PREEMPT
3050 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3053 * Is the spinlock portion underflowing?
3055 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3056 !(preempt_count() & PREEMPT_MASK
)))
3060 if (preempt_count() == val
)
3061 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3062 preempt_count() -= val
;
3064 EXPORT_SYMBOL(sub_preempt_count
);
3069 * Print scheduling while atomic bug:
3071 static noinline
void __schedule_bug(struct task_struct
*prev
)
3073 struct pt_regs
*regs
= get_irq_regs();
3075 if (oops_in_progress
)
3078 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3079 prev
->comm
, prev
->pid
, preempt_count());
3081 debug_show_held_locks(prev
);
3083 if (irqs_disabled())
3084 print_irqtrace_events(prev
);
3093 * Various schedule()-time debugging checks and statistics:
3095 static inline void schedule_debug(struct task_struct
*prev
)
3098 * Test if we are atomic. Since do_exit() needs to call into
3099 * schedule() atomically, we ignore that path for now.
3100 * Otherwise, whine if we are scheduling when we should not be.
3102 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3103 __schedule_bug(prev
);
3106 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3108 schedstat_inc(this_rq(), sched_count
);
3111 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3113 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
3114 update_rq_clock(rq
);
3115 prev
->sched_class
->put_prev_task(rq
, prev
);
3119 * Pick up the highest-prio task:
3121 static inline struct task_struct
*
3122 pick_next_task(struct rq
*rq
)
3124 const struct sched_class
*class;
3125 struct task_struct
*p
;
3128 * Optimization: we know that if all tasks are in
3129 * the fair class we can call that function directly:
3131 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3132 p
= fair_sched_class
.pick_next_task(rq
);
3137 for_each_class(class) {
3138 p
= class->pick_next_task(rq
);
3143 BUG(); /* the idle class will always have a runnable task */
3147 * __schedule() is the main scheduler function.
3149 static void __sched
__schedule(void)
3151 struct task_struct
*prev
, *next
;
3152 unsigned long *switch_count
;
3158 cpu
= smp_processor_id();
3160 rcu_note_context_switch(cpu
);
3163 schedule_debug(prev
);
3165 if (sched_feat(HRTICK
))
3168 raw_spin_lock_irq(&rq
->lock
);
3170 switch_count
= &prev
->nivcsw
;
3171 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3172 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3173 prev
->state
= TASK_RUNNING
;
3175 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3179 * If a worker went to sleep, notify and ask workqueue
3180 * whether it wants to wake up a task to maintain
3183 if (prev
->flags
& PF_WQ_WORKER
) {
3184 struct task_struct
*to_wakeup
;
3186 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3188 try_to_wake_up_local(to_wakeup
);
3191 switch_count
= &prev
->nvcsw
;
3194 pre_schedule(rq
, prev
);
3196 if (unlikely(!rq
->nr_running
))
3197 idle_balance(cpu
, rq
);
3199 put_prev_task(rq
, prev
);
3200 next
= pick_next_task(rq
);
3201 clear_tsk_need_resched(prev
);
3202 rq
->skip_clock_update
= 0;
3204 if (likely(prev
!= next
)) {
3209 context_switch(rq
, prev
, next
); /* unlocks the rq */
3211 * The context switch have flipped the stack from under us
3212 * and restored the local variables which were saved when
3213 * this task called schedule() in the past. prev == current
3214 * is still correct, but it can be moved to another cpu/rq.
3216 cpu
= smp_processor_id();
3219 raw_spin_unlock_irq(&rq
->lock
);
3223 sched_preempt_enable_no_resched();
3228 static inline void sched_submit_work(struct task_struct
*tsk
)
3230 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3233 * If we are going to sleep and we have plugged IO queued,
3234 * make sure to submit it to avoid deadlocks.
3236 if (blk_needs_flush_plug(tsk
))
3237 blk_schedule_flush_plug(tsk
);
3240 asmlinkage
void __sched
schedule(void)
3242 struct task_struct
*tsk
= current
;
3244 sched_submit_work(tsk
);
3247 EXPORT_SYMBOL(schedule
);
3250 * schedule_preempt_disabled - called with preemption disabled
3252 * Returns with preemption disabled. Note: preempt_count must be 1
3254 void __sched
schedule_preempt_disabled(void)
3256 sched_preempt_enable_no_resched();
3261 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3263 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3265 if (lock
->owner
!= owner
)
3269 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3270 * lock->owner still matches owner, if that fails, owner might
3271 * point to free()d memory, if it still matches, the rcu_read_lock()
3272 * ensures the memory stays valid.
3276 return owner
->on_cpu
;
3280 * Look out! "owner" is an entirely speculative pointer
3281 * access and not reliable.
3283 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3285 if (!sched_feat(OWNER_SPIN
))
3289 while (owner_running(lock
, owner
)) {
3293 arch_mutex_cpu_relax();
3298 * We break out the loop above on need_resched() and when the
3299 * owner changed, which is a sign for heavy contention. Return
3300 * success only when lock->owner is NULL.
3302 return lock
->owner
== NULL
;
3306 #ifdef CONFIG_PREEMPT
3308 * this is the entry point to schedule() from in-kernel preemption
3309 * off of preempt_enable. Kernel preemptions off return from interrupt
3310 * occur there and call schedule directly.
3312 asmlinkage
void __sched notrace
preempt_schedule(void)
3314 struct thread_info
*ti
= current_thread_info();
3317 * If there is a non-zero preempt_count or interrupts are disabled,
3318 * we do not want to preempt the current task. Just return..
3320 if (likely(ti
->preempt_count
|| irqs_disabled()))
3324 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3326 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3329 * Check again in case we missed a preemption opportunity
3330 * between schedule and now.
3333 } while (need_resched());
3335 EXPORT_SYMBOL(preempt_schedule
);
3338 * this is the entry point to schedule() from kernel preemption
3339 * off of irq context.
3340 * Note, that this is called and return with irqs disabled. This will
3341 * protect us against recursive calling from irq.
3343 asmlinkage
void __sched
preempt_schedule_irq(void)
3345 struct thread_info
*ti
= current_thread_info();
3347 /* Catch callers which need to be fixed */
3348 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3351 add_preempt_count(PREEMPT_ACTIVE
);
3354 local_irq_disable();
3355 sub_preempt_count(PREEMPT_ACTIVE
);
3358 * Check again in case we missed a preemption opportunity
3359 * between schedule and now.
3362 } while (need_resched());
3365 #endif /* CONFIG_PREEMPT */
3367 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3370 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3372 EXPORT_SYMBOL(default_wake_function
);
3375 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3376 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3377 * number) then we wake all the non-exclusive tasks and one exclusive task.
3379 * There are circumstances in which we can try to wake a task which has already
3380 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3381 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3383 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3384 int nr_exclusive
, int wake_flags
, void *key
)
3386 wait_queue_t
*curr
, *next
;
3388 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3389 unsigned flags
= curr
->flags
;
3391 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3392 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3398 * __wake_up - wake up threads blocked on a waitqueue.
3400 * @mode: which threads
3401 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3402 * @key: is directly passed to the wakeup function
3404 * It may be assumed that this function implies a write memory barrier before
3405 * changing the task state if and only if any tasks are woken up.
3407 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3408 int nr_exclusive
, void *key
)
3410 unsigned long flags
;
3412 spin_lock_irqsave(&q
->lock
, flags
);
3413 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3414 spin_unlock_irqrestore(&q
->lock
, flags
);
3416 EXPORT_SYMBOL(__wake_up
);
3419 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3421 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3423 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3425 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3427 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3429 __wake_up_common(q
, mode
, 1, 0, key
);
3431 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3434 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3436 * @mode: which threads
3437 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3438 * @key: opaque value to be passed to wakeup targets
3440 * The sync wakeup differs that the waker knows that it will schedule
3441 * away soon, so while the target thread will be woken up, it will not
3442 * be migrated to another CPU - ie. the two threads are 'synchronized'
3443 * with each other. This can prevent needless bouncing between CPUs.
3445 * On UP it can prevent extra preemption.
3447 * It may be assumed that this function implies a write memory barrier before
3448 * changing the task state if and only if any tasks are woken up.
3450 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3451 int nr_exclusive
, void *key
)
3453 unsigned long flags
;
3454 int wake_flags
= WF_SYNC
;
3459 if (unlikely(!nr_exclusive
))
3462 spin_lock_irqsave(&q
->lock
, flags
);
3463 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3464 spin_unlock_irqrestore(&q
->lock
, flags
);
3466 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3469 * __wake_up_sync - see __wake_up_sync_key()
3471 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3473 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3475 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3478 * complete: - signals a single thread waiting on this completion
3479 * @x: holds the state of this particular completion
3481 * This will wake up a single thread waiting on this completion. Threads will be
3482 * awakened in the same order in which they were queued.
3484 * See also complete_all(), wait_for_completion() and related routines.
3486 * It may be assumed that this function implies a write memory barrier before
3487 * changing the task state if and only if any tasks are woken up.
3489 void complete(struct completion
*x
)
3491 unsigned long flags
;
3493 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3495 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3496 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3498 EXPORT_SYMBOL(complete
);
3501 * complete_all: - signals all threads waiting on this completion
3502 * @x: holds the state of this particular completion
3504 * This will wake up all threads waiting on this particular completion event.
3506 * It may be assumed that this function implies a write memory barrier before
3507 * changing the task state if and only if any tasks are woken up.
3509 void complete_all(struct completion
*x
)
3511 unsigned long flags
;
3513 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3514 x
->done
+= UINT_MAX
/2;
3515 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3516 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3518 EXPORT_SYMBOL(complete_all
);
3520 static inline long __sched
3521 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3524 DECLARE_WAITQUEUE(wait
, current
);
3526 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3528 if (signal_pending_state(state
, current
)) {
3529 timeout
= -ERESTARTSYS
;
3532 __set_current_state(state
);
3533 spin_unlock_irq(&x
->wait
.lock
);
3534 timeout
= schedule_timeout(timeout
);
3535 spin_lock_irq(&x
->wait
.lock
);
3536 } while (!x
->done
&& timeout
);
3537 __remove_wait_queue(&x
->wait
, &wait
);
3542 return timeout
?: 1;
3546 wait_for_common(struct completion
*x
, long timeout
, int state
)
3550 spin_lock_irq(&x
->wait
.lock
);
3551 timeout
= do_wait_for_common(x
, timeout
, state
);
3552 spin_unlock_irq(&x
->wait
.lock
);
3557 * wait_for_completion: - waits for completion of a task
3558 * @x: holds the state of this particular completion
3560 * This waits to be signaled for completion of a specific task. It is NOT
3561 * interruptible and there is no timeout.
3563 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3564 * and interrupt capability. Also see complete().
3566 void __sched
wait_for_completion(struct completion
*x
)
3568 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3570 EXPORT_SYMBOL(wait_for_completion
);
3573 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3574 * @x: holds the state of this particular completion
3575 * @timeout: timeout value in jiffies
3577 * This waits for either a completion of a specific task to be signaled or for a
3578 * specified timeout to expire. The timeout is in jiffies. It is not
3581 * The return value is 0 if timed out, and positive (at least 1, or number of
3582 * jiffies left till timeout) if completed.
3584 unsigned long __sched
3585 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3587 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3589 EXPORT_SYMBOL(wait_for_completion_timeout
);
3592 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3593 * @x: holds the state of this particular completion
3595 * This waits for completion of a specific task to be signaled. It is
3598 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3600 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3602 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3603 if (t
== -ERESTARTSYS
)
3607 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3610 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3611 * @x: holds the state of this particular completion
3612 * @timeout: timeout value in jiffies
3614 * This waits for either a completion of a specific task to be signaled or for a
3615 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3617 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3618 * positive (at least 1, or number of jiffies left till timeout) if completed.
3621 wait_for_completion_interruptible_timeout(struct completion
*x
,
3622 unsigned long timeout
)
3624 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3626 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3629 * wait_for_completion_killable: - waits for completion of a task (killable)
3630 * @x: holds the state of this particular completion
3632 * This waits to be signaled for completion of a specific task. It can be
3633 * interrupted by a kill signal.
3635 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3637 int __sched
wait_for_completion_killable(struct completion
*x
)
3639 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3640 if (t
== -ERESTARTSYS
)
3644 EXPORT_SYMBOL(wait_for_completion_killable
);
3647 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3648 * @x: holds the state of this particular completion
3649 * @timeout: timeout value in jiffies
3651 * This waits for either a completion of a specific task to be
3652 * signaled or for a specified timeout to expire. It can be
3653 * interrupted by a kill signal. The timeout is in jiffies.
3655 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3656 * positive (at least 1, or number of jiffies left till timeout) if completed.
3659 wait_for_completion_killable_timeout(struct completion
*x
,
3660 unsigned long timeout
)
3662 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3664 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3667 * try_wait_for_completion - try to decrement a completion without blocking
3668 * @x: completion structure
3670 * Returns: 0 if a decrement cannot be done without blocking
3671 * 1 if a decrement succeeded.
3673 * If a completion is being used as a counting completion,
3674 * attempt to decrement the counter without blocking. This
3675 * enables us to avoid waiting if the resource the completion
3676 * is protecting is not available.
3678 bool try_wait_for_completion(struct completion
*x
)
3680 unsigned long flags
;
3683 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3688 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3691 EXPORT_SYMBOL(try_wait_for_completion
);
3694 * completion_done - Test to see if a completion has any waiters
3695 * @x: completion structure
3697 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3698 * 1 if there are no waiters.
3701 bool completion_done(struct completion
*x
)
3703 unsigned long flags
;
3706 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3709 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3712 EXPORT_SYMBOL(completion_done
);
3715 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3717 unsigned long flags
;
3720 init_waitqueue_entry(&wait
, current
);
3722 __set_current_state(state
);
3724 spin_lock_irqsave(&q
->lock
, flags
);
3725 __add_wait_queue(q
, &wait
);
3726 spin_unlock(&q
->lock
);
3727 timeout
= schedule_timeout(timeout
);
3728 spin_lock_irq(&q
->lock
);
3729 __remove_wait_queue(q
, &wait
);
3730 spin_unlock_irqrestore(&q
->lock
, flags
);
3735 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3737 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3739 EXPORT_SYMBOL(interruptible_sleep_on
);
3742 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3744 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3746 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3748 void __sched
sleep_on(wait_queue_head_t
*q
)
3750 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3752 EXPORT_SYMBOL(sleep_on
);
3754 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3756 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3758 EXPORT_SYMBOL(sleep_on_timeout
);
3760 #ifdef CONFIG_RT_MUTEXES
3763 * rt_mutex_setprio - set the current priority of a task
3765 * @prio: prio value (kernel-internal form)
3767 * This function changes the 'effective' priority of a task. It does
3768 * not touch ->normal_prio like __setscheduler().
3770 * Used by the rt_mutex code to implement priority inheritance logic.
3772 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3774 int oldprio
, on_rq
, running
;
3776 const struct sched_class
*prev_class
;
3778 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3780 rq
= __task_rq_lock(p
);
3783 * Idle task boosting is a nono in general. There is one
3784 * exception, when PREEMPT_RT and NOHZ is active:
3786 * The idle task calls get_next_timer_interrupt() and holds
3787 * the timer wheel base->lock on the CPU and another CPU wants
3788 * to access the timer (probably to cancel it). We can safely
3789 * ignore the boosting request, as the idle CPU runs this code
3790 * with interrupts disabled and will complete the lock
3791 * protected section without being interrupted. So there is no
3792 * real need to boost.
3794 if (unlikely(p
== rq
->idle
)) {
3795 WARN_ON(p
!= rq
->curr
);
3796 WARN_ON(p
->pi_blocked_on
);
3800 trace_sched_pi_setprio(p
, prio
);
3802 prev_class
= p
->sched_class
;
3804 running
= task_current(rq
, p
);
3806 dequeue_task(rq
, p
, 0);
3808 p
->sched_class
->put_prev_task(rq
, p
);
3811 p
->sched_class
= &rt_sched_class
;
3813 p
->sched_class
= &fair_sched_class
;
3818 p
->sched_class
->set_curr_task(rq
);
3820 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3822 check_class_changed(rq
, p
, prev_class
, oldprio
);
3824 __task_rq_unlock(rq
);
3827 void set_user_nice(struct task_struct
*p
, long nice
)
3829 int old_prio
, delta
, on_rq
;
3830 unsigned long flags
;
3833 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3836 * We have to be careful, if called from sys_setpriority(),
3837 * the task might be in the middle of scheduling on another CPU.
3839 rq
= task_rq_lock(p
, &flags
);
3841 * The RT priorities are set via sched_setscheduler(), but we still
3842 * allow the 'normal' nice value to be set - but as expected
3843 * it wont have any effect on scheduling until the task is
3844 * SCHED_FIFO/SCHED_RR:
3846 if (task_has_rt_policy(p
)) {
3847 p
->static_prio
= NICE_TO_PRIO(nice
);
3852 dequeue_task(rq
, p
, 0);
3854 p
->static_prio
= NICE_TO_PRIO(nice
);
3857 p
->prio
= effective_prio(p
);
3858 delta
= p
->prio
- old_prio
;
3861 enqueue_task(rq
, p
, 0);
3863 * If the task increased its priority or is running and
3864 * lowered its priority, then reschedule its CPU:
3866 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3867 resched_task(rq
->curr
);
3870 task_rq_unlock(rq
, p
, &flags
);
3872 EXPORT_SYMBOL(set_user_nice
);
3875 * can_nice - check if a task can reduce its nice value
3879 int can_nice(const struct task_struct
*p
, const int nice
)
3881 /* convert nice value [19,-20] to rlimit style value [1,40] */
3882 int nice_rlim
= 20 - nice
;
3884 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3885 capable(CAP_SYS_NICE
));
3888 #ifdef __ARCH_WANT_SYS_NICE
3891 * sys_nice - change the priority of the current process.
3892 * @increment: priority increment
3894 * sys_setpriority is a more generic, but much slower function that
3895 * does similar things.
3897 SYSCALL_DEFINE1(nice
, int, increment
)
3902 * Setpriority might change our priority at the same moment.
3903 * We don't have to worry. Conceptually one call occurs first
3904 * and we have a single winner.
3906 if (increment
< -40)
3911 nice
= TASK_NICE(current
) + increment
;
3917 if (increment
< 0 && !can_nice(current
, nice
))
3920 retval
= security_task_setnice(current
, nice
);
3924 set_user_nice(current
, nice
);
3931 * task_prio - return the priority value of a given task.
3932 * @p: the task in question.
3934 * This is the priority value as seen by users in /proc.
3935 * RT tasks are offset by -200. Normal tasks are centered
3936 * around 0, value goes from -16 to +15.
3938 int task_prio(const struct task_struct
*p
)
3940 return p
->prio
- MAX_RT_PRIO
;
3944 * task_nice - return the nice value of a given task.
3945 * @p: the task in question.
3947 int task_nice(const struct task_struct
*p
)
3949 return TASK_NICE(p
);
3951 EXPORT_SYMBOL(task_nice
);
3954 * idle_cpu - is a given cpu idle currently?
3955 * @cpu: the processor in question.
3957 int idle_cpu(int cpu
)
3959 struct rq
*rq
= cpu_rq(cpu
);
3961 if (rq
->curr
!= rq
->idle
)
3968 if (!llist_empty(&rq
->wake_list
))
3976 * idle_task - return the idle task for a given cpu.
3977 * @cpu: the processor in question.
3979 struct task_struct
*idle_task(int cpu
)
3981 return cpu_rq(cpu
)->idle
;
3985 * find_process_by_pid - find a process with a matching PID value.
3986 * @pid: the pid in question.
3988 static struct task_struct
*find_process_by_pid(pid_t pid
)
3990 return pid
? find_task_by_vpid(pid
) : current
;
3993 /* Actually do priority change: must hold rq lock. */
3995 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3998 p
->rt_priority
= prio
;
3999 p
->normal_prio
= normal_prio(p
);
4000 /* we are holding p->pi_lock already */
4001 p
->prio
= rt_mutex_getprio(p
);
4002 if (rt_prio(p
->prio
))
4003 p
->sched_class
= &rt_sched_class
;
4005 p
->sched_class
= &fair_sched_class
;
4010 * check the target process has a UID that matches the current process's
4012 static bool check_same_owner(struct task_struct
*p
)
4014 const struct cred
*cred
= current_cred(), *pcred
;
4018 pcred
= __task_cred(p
);
4019 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
4020 match
= (cred
->euid
== pcred
->euid
||
4021 cred
->euid
== pcred
->uid
);
4028 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4029 const struct sched_param
*param
, bool user
)
4031 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4032 unsigned long flags
;
4033 const struct sched_class
*prev_class
;
4037 /* may grab non-irq protected spin_locks */
4038 BUG_ON(in_interrupt());
4040 /* double check policy once rq lock held */
4042 reset_on_fork
= p
->sched_reset_on_fork
;
4043 policy
= oldpolicy
= p
->policy
;
4045 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4046 policy
&= ~SCHED_RESET_ON_FORK
;
4048 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4049 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4050 policy
!= SCHED_IDLE
)
4055 * Valid priorities for SCHED_FIFO and SCHED_RR are
4056 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4057 * SCHED_BATCH and SCHED_IDLE is 0.
4059 if (param
->sched_priority
< 0 ||
4060 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4061 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4063 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4067 * Allow unprivileged RT tasks to decrease priority:
4069 if (user
&& !capable(CAP_SYS_NICE
)) {
4070 if (rt_policy(policy
)) {
4071 unsigned long rlim_rtprio
=
4072 task_rlimit(p
, RLIMIT_RTPRIO
);
4074 /* can't set/change the rt policy */
4075 if (policy
!= p
->policy
&& !rlim_rtprio
)
4078 /* can't increase priority */
4079 if (param
->sched_priority
> p
->rt_priority
&&
4080 param
->sched_priority
> rlim_rtprio
)
4085 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4086 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4088 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
4089 if (!can_nice(p
, TASK_NICE(p
)))
4093 /* can't change other user's priorities */
4094 if (!check_same_owner(p
))
4097 /* Normal users shall not reset the sched_reset_on_fork flag */
4098 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4103 retval
= security_task_setscheduler(p
);
4109 * make sure no PI-waiters arrive (or leave) while we are
4110 * changing the priority of the task:
4112 * To be able to change p->policy safely, the appropriate
4113 * runqueue lock must be held.
4115 rq
= task_rq_lock(p
, &flags
);
4118 * Changing the policy of the stop threads its a very bad idea
4120 if (p
== rq
->stop
) {
4121 task_rq_unlock(rq
, p
, &flags
);
4126 * If not changing anything there's no need to proceed further:
4128 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
4129 param
->sched_priority
== p
->rt_priority
))) {
4131 __task_rq_unlock(rq
);
4132 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4136 #ifdef CONFIG_RT_GROUP_SCHED
4139 * Do not allow realtime tasks into groups that have no runtime
4142 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4143 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4144 !task_group_is_autogroup(task_group(p
))) {
4145 task_rq_unlock(rq
, p
, &flags
);
4151 /* recheck policy now with rq lock held */
4152 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4153 policy
= oldpolicy
= -1;
4154 task_rq_unlock(rq
, p
, &flags
);
4158 running
= task_current(rq
, p
);
4160 dequeue_task(rq
, p
, 0);
4162 p
->sched_class
->put_prev_task(rq
, p
);
4164 p
->sched_reset_on_fork
= reset_on_fork
;
4167 prev_class
= p
->sched_class
;
4168 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4171 p
->sched_class
->set_curr_task(rq
);
4173 enqueue_task(rq
, p
, 0);
4175 check_class_changed(rq
, p
, prev_class
, oldprio
);
4176 task_rq_unlock(rq
, p
, &flags
);
4178 rt_mutex_adjust_pi(p
);
4184 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4185 * @p: the task in question.
4186 * @policy: new policy.
4187 * @param: structure containing the new RT priority.
4189 * NOTE that the task may be already dead.
4191 int sched_setscheduler(struct task_struct
*p
, int policy
,
4192 const struct sched_param
*param
)
4194 return __sched_setscheduler(p
, policy
, param
, true);
4196 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4199 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4200 * @p: the task in question.
4201 * @policy: new policy.
4202 * @param: structure containing the new RT priority.
4204 * Just like sched_setscheduler, only don't bother checking if the
4205 * current context has permission. For example, this is needed in
4206 * stop_machine(): we create temporary high priority worker threads,
4207 * but our caller might not have that capability.
4209 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4210 const struct sched_param
*param
)
4212 return __sched_setscheduler(p
, policy
, param
, false);
4216 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4218 struct sched_param lparam
;
4219 struct task_struct
*p
;
4222 if (!param
|| pid
< 0)
4224 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4229 p
= find_process_by_pid(pid
);
4231 retval
= sched_setscheduler(p
, policy
, &lparam
);
4238 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4239 * @pid: the pid in question.
4240 * @policy: new policy.
4241 * @param: structure containing the new RT priority.
4243 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4244 struct sched_param __user
*, param
)
4246 /* negative values for policy are not valid */
4250 return do_sched_setscheduler(pid
, policy
, param
);
4254 * sys_sched_setparam - set/change the RT priority of a thread
4255 * @pid: the pid in question.
4256 * @param: structure containing the new RT priority.
4258 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4260 return do_sched_setscheduler(pid
, -1, param
);
4264 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4265 * @pid: the pid in question.
4267 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4269 struct task_struct
*p
;
4277 p
= find_process_by_pid(pid
);
4279 retval
= security_task_getscheduler(p
);
4282 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4289 * sys_sched_getparam - get the RT priority of a thread
4290 * @pid: the pid in question.
4291 * @param: structure containing the RT priority.
4293 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4295 struct sched_param lp
;
4296 struct task_struct
*p
;
4299 if (!param
|| pid
< 0)
4303 p
= find_process_by_pid(pid
);
4308 retval
= security_task_getscheduler(p
);
4312 lp
.sched_priority
= p
->rt_priority
;
4316 * This one might sleep, we cannot do it with a spinlock held ...
4318 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4327 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4329 cpumask_var_t cpus_allowed
, new_mask
;
4330 struct task_struct
*p
;
4336 p
= find_process_by_pid(pid
);
4343 /* Prevent p going away */
4347 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4351 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4353 goto out_free_cpus_allowed
;
4356 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4359 retval
= security_task_setscheduler(p
);
4363 cpuset_cpus_allowed(p
, cpus_allowed
);
4364 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4366 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4369 cpuset_cpus_allowed(p
, cpus_allowed
);
4370 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4372 * We must have raced with a concurrent cpuset
4373 * update. Just reset the cpus_allowed to the
4374 * cpuset's cpus_allowed
4376 cpumask_copy(new_mask
, cpus_allowed
);
4381 free_cpumask_var(new_mask
);
4382 out_free_cpus_allowed
:
4383 free_cpumask_var(cpus_allowed
);
4390 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4391 struct cpumask
*new_mask
)
4393 if (len
< cpumask_size())
4394 cpumask_clear(new_mask
);
4395 else if (len
> cpumask_size())
4396 len
= cpumask_size();
4398 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4402 * sys_sched_setaffinity - set the cpu affinity of a process
4403 * @pid: pid of the process
4404 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4405 * @user_mask_ptr: user-space pointer to the new cpu mask
4407 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4408 unsigned long __user
*, user_mask_ptr
)
4410 cpumask_var_t new_mask
;
4413 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4416 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4418 retval
= sched_setaffinity(pid
, new_mask
);
4419 free_cpumask_var(new_mask
);
4423 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4425 struct task_struct
*p
;
4426 unsigned long flags
;
4433 p
= find_process_by_pid(pid
);
4437 retval
= security_task_getscheduler(p
);
4441 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4442 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4443 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4453 * sys_sched_getaffinity - get the cpu affinity of a process
4454 * @pid: pid of the process
4455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4456 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4458 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4459 unsigned long __user
*, user_mask_ptr
)
4464 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4466 if (len
& (sizeof(unsigned long)-1))
4469 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4472 ret
= sched_getaffinity(pid
, mask
);
4474 size_t retlen
= min_t(size_t, len
, cpumask_size());
4476 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4481 free_cpumask_var(mask
);
4487 * sys_sched_yield - yield the current processor to other threads.
4489 * This function yields the current CPU to other tasks. If there are no
4490 * other threads running on this CPU then this function will return.
4492 SYSCALL_DEFINE0(sched_yield
)
4494 struct rq
*rq
= this_rq_lock();
4496 schedstat_inc(rq
, yld_count
);
4497 current
->sched_class
->yield_task(rq
);
4500 * Since we are going to call schedule() anyway, there's
4501 * no need to preempt or enable interrupts:
4503 __release(rq
->lock
);
4504 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4505 do_raw_spin_unlock(&rq
->lock
);
4506 sched_preempt_enable_no_resched();
4513 static inline int should_resched(void)
4515 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4518 static void __cond_resched(void)
4520 add_preempt_count(PREEMPT_ACTIVE
);
4522 sub_preempt_count(PREEMPT_ACTIVE
);
4525 int __sched
_cond_resched(void)
4527 if (should_resched()) {
4533 EXPORT_SYMBOL(_cond_resched
);
4536 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4537 * call schedule, and on return reacquire the lock.
4539 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4540 * operations here to prevent schedule() from being called twice (once via
4541 * spin_unlock(), once by hand).
4543 int __cond_resched_lock(spinlock_t
*lock
)
4545 int resched
= should_resched();
4548 lockdep_assert_held(lock
);
4550 if (spin_needbreak(lock
) || resched
) {
4561 EXPORT_SYMBOL(__cond_resched_lock
);
4563 int __sched
__cond_resched_softirq(void)
4565 BUG_ON(!in_softirq());
4567 if (should_resched()) {
4575 EXPORT_SYMBOL(__cond_resched_softirq
);
4578 * yield - yield the current processor to other threads.
4580 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4582 * The scheduler is at all times free to pick the calling task as the most
4583 * eligible task to run, if removing the yield() call from your code breaks
4584 * it, its already broken.
4586 * Typical broken usage is:
4591 * where one assumes that yield() will let 'the other' process run that will
4592 * make event true. If the current task is a SCHED_FIFO task that will never
4593 * happen. Never use yield() as a progress guarantee!!
4595 * If you want to use yield() to wait for something, use wait_event().
4596 * If you want to use yield() to be 'nice' for others, use cond_resched().
4597 * If you still want to use yield(), do not!
4599 void __sched
yield(void)
4601 set_current_state(TASK_RUNNING
);
4604 EXPORT_SYMBOL(yield
);
4607 * yield_to - yield the current processor to another thread in
4608 * your thread group, or accelerate that thread toward the
4609 * processor it's on.
4611 * @preempt: whether task preemption is allowed or not
4613 * It's the caller's job to ensure that the target task struct
4614 * can't go away on us before we can do any checks.
4616 * Returns true if we indeed boosted the target task.
4618 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4620 struct task_struct
*curr
= current
;
4621 struct rq
*rq
, *p_rq
;
4622 unsigned long flags
;
4625 local_irq_save(flags
);
4630 double_rq_lock(rq
, p_rq
);
4631 while (task_rq(p
) != p_rq
) {
4632 double_rq_unlock(rq
, p_rq
);
4636 if (!curr
->sched_class
->yield_to_task
)
4639 if (curr
->sched_class
!= p
->sched_class
)
4642 if (task_running(p_rq
, p
) || p
->state
)
4645 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4647 schedstat_inc(rq
, yld_count
);
4649 * Make p's CPU reschedule; pick_next_entity takes care of
4652 if (preempt
&& rq
!= p_rq
)
4653 resched_task(p_rq
->curr
);
4656 * We might have set it in task_yield_fair(), but are
4657 * not going to schedule(), so don't want to skip
4660 rq
->skip_clock_update
= 0;
4664 double_rq_unlock(rq
, p_rq
);
4665 local_irq_restore(flags
);
4672 EXPORT_SYMBOL_GPL(yield_to
);
4675 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4676 * that process accounting knows that this is a task in IO wait state.
4678 void __sched
io_schedule(void)
4680 struct rq
*rq
= raw_rq();
4682 delayacct_blkio_start();
4683 atomic_inc(&rq
->nr_iowait
);
4684 blk_flush_plug(current
);
4685 current
->in_iowait
= 1;
4687 current
->in_iowait
= 0;
4688 atomic_dec(&rq
->nr_iowait
);
4689 delayacct_blkio_end();
4691 EXPORT_SYMBOL(io_schedule
);
4693 long __sched
io_schedule_timeout(long timeout
)
4695 struct rq
*rq
= raw_rq();
4698 delayacct_blkio_start();
4699 atomic_inc(&rq
->nr_iowait
);
4700 blk_flush_plug(current
);
4701 current
->in_iowait
= 1;
4702 ret
= schedule_timeout(timeout
);
4703 current
->in_iowait
= 0;
4704 atomic_dec(&rq
->nr_iowait
);
4705 delayacct_blkio_end();
4710 * sys_sched_get_priority_max - return maximum RT priority.
4711 * @policy: scheduling class.
4713 * this syscall returns the maximum rt_priority that can be used
4714 * by a given scheduling class.
4716 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4723 ret
= MAX_USER_RT_PRIO
-1;
4735 * sys_sched_get_priority_min - return minimum RT priority.
4736 * @policy: scheduling class.
4738 * this syscall returns the minimum rt_priority that can be used
4739 * by a given scheduling class.
4741 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4759 * sys_sched_rr_get_interval - return the default timeslice of a process.
4760 * @pid: pid of the process.
4761 * @interval: userspace pointer to the timeslice value.
4763 * this syscall writes the default timeslice value of a given process
4764 * into the user-space timespec buffer. A value of '0' means infinity.
4766 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4767 struct timespec __user
*, interval
)
4769 struct task_struct
*p
;
4770 unsigned int time_slice
;
4771 unsigned long flags
;
4781 p
= find_process_by_pid(pid
);
4785 retval
= security_task_getscheduler(p
);
4789 rq
= task_rq_lock(p
, &flags
);
4790 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4791 task_rq_unlock(rq
, p
, &flags
);
4794 jiffies_to_timespec(time_slice
, &t
);
4795 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4803 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4805 void sched_show_task(struct task_struct
*p
)
4807 unsigned long free
= 0;
4810 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4811 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4812 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4813 #if BITS_PER_LONG == 32
4814 if (state
== TASK_RUNNING
)
4815 printk(KERN_CONT
" running ");
4817 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4819 if (state
== TASK_RUNNING
)
4820 printk(KERN_CONT
" running task ");
4822 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4824 #ifdef CONFIG_DEBUG_STACK_USAGE
4825 free
= stack_not_used(p
);
4827 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4828 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4829 (unsigned long)task_thread_info(p
)->flags
);
4831 show_stack(p
, NULL
);
4834 void show_state_filter(unsigned long state_filter
)
4836 struct task_struct
*g
, *p
;
4838 #if BITS_PER_LONG == 32
4840 " task PC stack pid father\n");
4843 " task PC stack pid father\n");
4846 do_each_thread(g
, p
) {
4848 * reset the NMI-timeout, listing all files on a slow
4849 * console might take a lot of time:
4851 touch_nmi_watchdog();
4852 if (!state_filter
|| (p
->state
& state_filter
))
4854 } while_each_thread(g
, p
);
4856 touch_all_softlockup_watchdogs();
4858 #ifdef CONFIG_SCHED_DEBUG
4859 sysrq_sched_debug_show();
4863 * Only show locks if all tasks are dumped:
4866 debug_show_all_locks();
4869 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4871 idle
->sched_class
= &idle_sched_class
;
4875 * init_idle - set up an idle thread for a given CPU
4876 * @idle: task in question
4877 * @cpu: cpu the idle task belongs to
4879 * NOTE: this function does not set the idle thread's NEED_RESCHED
4880 * flag, to make booting more robust.
4882 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4884 struct rq
*rq
= cpu_rq(cpu
);
4885 unsigned long flags
;
4887 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4890 idle
->state
= TASK_RUNNING
;
4891 idle
->se
.exec_start
= sched_clock();
4893 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4895 * We're having a chicken and egg problem, even though we are
4896 * holding rq->lock, the cpu isn't yet set to this cpu so the
4897 * lockdep check in task_group() will fail.
4899 * Similar case to sched_fork(). / Alternatively we could
4900 * use task_rq_lock() here and obtain the other rq->lock.
4905 __set_task_cpu(idle
, cpu
);
4908 rq
->curr
= rq
->idle
= idle
;
4909 #if defined(CONFIG_SMP)
4912 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4914 /* Set the preempt count _outside_ the spinlocks! */
4915 task_thread_info(idle
)->preempt_count
= 0;
4918 * The idle tasks have their own, simple scheduling class:
4920 idle
->sched_class
= &idle_sched_class
;
4921 ftrace_graph_init_idle_task(idle
, cpu
);
4922 #if defined(CONFIG_SMP)
4923 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4928 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4930 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4931 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4933 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4934 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
4938 * This is how migration works:
4940 * 1) we invoke migration_cpu_stop() on the target CPU using
4942 * 2) stopper starts to run (implicitly forcing the migrated thread
4944 * 3) it checks whether the migrated task is still in the wrong runqueue.
4945 * 4) if it's in the wrong runqueue then the migration thread removes
4946 * it and puts it into the right queue.
4947 * 5) stopper completes and stop_one_cpu() returns and the migration
4952 * Change a given task's CPU affinity. Migrate the thread to a
4953 * proper CPU and schedule it away if the CPU it's executing on
4954 * is removed from the allowed bitmask.
4956 * NOTE: the caller must have a valid reference to the task, the
4957 * task must not exit() & deallocate itself prematurely. The
4958 * call is not atomic; no spinlocks may be held.
4960 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4962 unsigned long flags
;
4964 unsigned int dest_cpu
;
4967 rq
= task_rq_lock(p
, &flags
);
4969 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4972 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4977 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4982 do_set_cpus_allowed(p
, new_mask
);
4984 /* Can the task run on the task's current CPU? If so, we're done */
4985 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4988 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4990 struct migration_arg arg
= { p
, dest_cpu
};
4991 /* Need help from migration thread: drop lock and wait. */
4992 task_rq_unlock(rq
, p
, &flags
);
4993 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4994 tlb_migrate_finish(p
->mm
);
4998 task_rq_unlock(rq
, p
, &flags
);
5002 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5005 * Move (not current) task off this cpu, onto dest cpu. We're doing
5006 * this because either it can't run here any more (set_cpus_allowed()
5007 * away from this CPU, or CPU going down), or because we're
5008 * attempting to rebalance this task on exec (sched_exec).
5010 * So we race with normal scheduler movements, but that's OK, as long
5011 * as the task is no longer on this CPU.
5013 * Returns non-zero if task was successfully migrated.
5015 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5017 struct rq
*rq_dest
, *rq_src
;
5020 if (unlikely(!cpu_active(dest_cpu
)))
5023 rq_src
= cpu_rq(src_cpu
);
5024 rq_dest
= cpu_rq(dest_cpu
);
5026 raw_spin_lock(&p
->pi_lock
);
5027 double_rq_lock(rq_src
, rq_dest
);
5028 /* Already moved. */
5029 if (task_cpu(p
) != src_cpu
)
5031 /* Affinity changed (again). */
5032 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
5036 * If we're not on a rq, the next wake-up will ensure we're
5040 dequeue_task(rq_src
, p
, 0);
5041 set_task_cpu(p
, dest_cpu
);
5042 enqueue_task(rq_dest
, p
, 0);
5043 check_preempt_curr(rq_dest
, p
, 0);
5048 double_rq_unlock(rq_src
, rq_dest
);
5049 raw_spin_unlock(&p
->pi_lock
);
5054 * migration_cpu_stop - this will be executed by a highprio stopper thread
5055 * and performs thread migration by bumping thread off CPU then
5056 * 'pushing' onto another runqueue.
5058 static int migration_cpu_stop(void *data
)
5060 struct migration_arg
*arg
= data
;
5063 * The original target cpu might have gone down and we might
5064 * be on another cpu but it doesn't matter.
5066 local_irq_disable();
5067 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5072 #ifdef CONFIG_HOTPLUG_CPU
5075 * Ensures that the idle task is using init_mm right before its cpu goes
5078 void idle_task_exit(void)
5080 struct mm_struct
*mm
= current
->active_mm
;
5082 BUG_ON(cpu_online(smp_processor_id()));
5085 switch_mm(mm
, &init_mm
, current
);
5090 * While a dead CPU has no uninterruptible tasks queued at this point,
5091 * it might still have a nonzero ->nr_uninterruptible counter, because
5092 * for performance reasons the counter is not stricly tracking tasks to
5093 * their home CPUs. So we just add the counter to another CPU's counter,
5094 * to keep the global sum constant after CPU-down:
5096 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5098 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5100 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5101 rq_src
->nr_uninterruptible
= 0;
5105 * remove the tasks which were accounted by rq from calc_load_tasks.
5107 static void calc_global_load_remove(struct rq
*rq
)
5109 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5110 rq
->calc_load_active
= 0;
5114 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5115 * try_to_wake_up()->select_task_rq().
5117 * Called with rq->lock held even though we'er in stop_machine() and
5118 * there's no concurrency possible, we hold the required locks anyway
5119 * because of lock validation efforts.
5121 static void migrate_tasks(unsigned int dead_cpu
)
5123 struct rq
*rq
= cpu_rq(dead_cpu
);
5124 struct task_struct
*next
, *stop
= rq
->stop
;
5128 * Fudge the rq selection such that the below task selection loop
5129 * doesn't get stuck on the currently eligible stop task.
5131 * We're currently inside stop_machine() and the rq is either stuck
5132 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5133 * either way we should never end up calling schedule() until we're
5138 /* Ensure any throttled groups are reachable by pick_next_task */
5139 unthrottle_offline_cfs_rqs(rq
);
5143 * There's this thread running, bail when that's the only
5146 if (rq
->nr_running
== 1)
5149 next
= pick_next_task(rq
);
5151 next
->sched_class
->put_prev_task(rq
, next
);
5153 /* Find suitable destination for @next, with force if needed. */
5154 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5155 raw_spin_unlock(&rq
->lock
);
5157 __migrate_task(next
, dead_cpu
, dest_cpu
);
5159 raw_spin_lock(&rq
->lock
);
5165 #endif /* CONFIG_HOTPLUG_CPU */
5167 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5169 static struct ctl_table sd_ctl_dir
[] = {
5171 .procname
= "sched_domain",
5177 static struct ctl_table sd_ctl_root
[] = {
5179 .procname
= "kernel",
5181 .child
= sd_ctl_dir
,
5186 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5188 struct ctl_table
*entry
=
5189 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5194 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5196 struct ctl_table
*entry
;
5199 * In the intermediate directories, both the child directory and
5200 * procname are dynamically allocated and could fail but the mode
5201 * will always be set. In the lowest directory the names are
5202 * static strings and all have proc handlers.
5204 for (entry
= *tablep
; entry
->mode
; entry
++) {
5206 sd_free_ctl_entry(&entry
->child
);
5207 if (entry
->proc_handler
== NULL
)
5208 kfree(entry
->procname
);
5216 set_table_entry(struct ctl_table
*entry
,
5217 const char *procname
, void *data
, int maxlen
,
5218 umode_t mode
, proc_handler
*proc_handler
)
5220 entry
->procname
= procname
;
5222 entry
->maxlen
= maxlen
;
5224 entry
->proc_handler
= proc_handler
;
5227 static struct ctl_table
*
5228 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5230 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5235 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5236 sizeof(long), 0644, proc_doulongvec_minmax
);
5237 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5238 sizeof(long), 0644, proc_doulongvec_minmax
);
5239 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5240 sizeof(int), 0644, proc_dointvec_minmax
);
5241 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5242 sizeof(int), 0644, proc_dointvec_minmax
);
5243 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5244 sizeof(int), 0644, proc_dointvec_minmax
);
5245 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5246 sizeof(int), 0644, proc_dointvec_minmax
);
5247 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5248 sizeof(int), 0644, proc_dointvec_minmax
);
5249 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5250 sizeof(int), 0644, proc_dointvec_minmax
);
5251 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5252 sizeof(int), 0644, proc_dointvec_minmax
);
5253 set_table_entry(&table
[9], "cache_nice_tries",
5254 &sd
->cache_nice_tries
,
5255 sizeof(int), 0644, proc_dointvec_minmax
);
5256 set_table_entry(&table
[10], "flags", &sd
->flags
,
5257 sizeof(int), 0644, proc_dointvec_minmax
);
5258 set_table_entry(&table
[11], "name", sd
->name
,
5259 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5260 /* &table[12] is terminator */
5265 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5267 struct ctl_table
*entry
, *table
;
5268 struct sched_domain
*sd
;
5269 int domain_num
= 0, i
;
5272 for_each_domain(cpu
, sd
)
5274 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5279 for_each_domain(cpu
, sd
) {
5280 snprintf(buf
, 32, "domain%d", i
);
5281 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5283 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5290 static struct ctl_table_header
*sd_sysctl_header
;
5291 static void register_sched_domain_sysctl(void)
5293 int i
, cpu_num
= num_possible_cpus();
5294 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5297 WARN_ON(sd_ctl_dir
[0].child
);
5298 sd_ctl_dir
[0].child
= entry
;
5303 for_each_possible_cpu(i
) {
5304 snprintf(buf
, 32, "cpu%d", i
);
5305 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5307 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5311 WARN_ON(sd_sysctl_header
);
5312 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5315 /* may be called multiple times per register */
5316 static void unregister_sched_domain_sysctl(void)
5318 if (sd_sysctl_header
)
5319 unregister_sysctl_table(sd_sysctl_header
);
5320 sd_sysctl_header
= NULL
;
5321 if (sd_ctl_dir
[0].child
)
5322 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5325 static void register_sched_domain_sysctl(void)
5328 static void unregister_sched_domain_sysctl(void)
5333 static void set_rq_online(struct rq
*rq
)
5336 const struct sched_class
*class;
5338 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5341 for_each_class(class) {
5342 if (class->rq_online
)
5343 class->rq_online(rq
);
5348 static void set_rq_offline(struct rq
*rq
)
5351 const struct sched_class
*class;
5353 for_each_class(class) {
5354 if (class->rq_offline
)
5355 class->rq_offline(rq
);
5358 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5364 * migration_call - callback that gets triggered when a CPU is added.
5365 * Here we can start up the necessary migration thread for the new CPU.
5367 static int __cpuinit
5368 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5370 int cpu
= (long)hcpu
;
5371 unsigned long flags
;
5372 struct rq
*rq
= cpu_rq(cpu
);
5374 switch (action
& ~CPU_TASKS_FROZEN
) {
5376 case CPU_UP_PREPARE
:
5377 rq
->calc_load_update
= calc_load_update
;
5381 /* Update our root-domain */
5382 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5384 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5388 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5391 #ifdef CONFIG_HOTPLUG_CPU
5393 sched_ttwu_pending();
5394 /* Update our root-domain */
5395 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5397 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5401 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5402 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5404 migrate_nr_uninterruptible(rq
);
5405 calc_global_load_remove(rq
);
5410 update_max_interval();
5416 * Register at high priority so that task migration (migrate_all_tasks)
5417 * happens before everything else. This has to be lower priority than
5418 * the notifier in the perf_event subsystem, though.
5420 static struct notifier_block __cpuinitdata migration_notifier
= {
5421 .notifier_call
= migration_call
,
5422 .priority
= CPU_PRI_MIGRATION
,
5425 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5426 unsigned long action
, void *hcpu
)
5428 switch (action
& ~CPU_TASKS_FROZEN
) {
5430 case CPU_DOWN_FAILED
:
5431 set_cpu_active((long)hcpu
, true);
5438 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5439 unsigned long action
, void *hcpu
)
5441 switch (action
& ~CPU_TASKS_FROZEN
) {
5442 case CPU_DOWN_PREPARE
:
5443 set_cpu_active((long)hcpu
, false);
5450 static int __init
migration_init(void)
5452 void *cpu
= (void *)(long)smp_processor_id();
5455 /* Initialize migration for the boot CPU */
5456 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5457 BUG_ON(err
== NOTIFY_BAD
);
5458 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5459 register_cpu_notifier(&migration_notifier
);
5461 /* Register cpu active notifiers */
5462 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5463 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5467 early_initcall(migration_init
);
5472 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5474 #ifdef CONFIG_SCHED_DEBUG
5476 static __read_mostly
int sched_domain_debug_enabled
;
5478 static int __init
sched_domain_debug_setup(char *str
)
5480 sched_domain_debug_enabled
= 1;
5484 early_param("sched_debug", sched_domain_debug_setup
);
5486 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5487 struct cpumask
*groupmask
)
5489 struct sched_group
*group
= sd
->groups
;
5492 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5493 cpumask_clear(groupmask
);
5495 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5497 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5498 printk("does not load-balance\n");
5500 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5505 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5507 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5508 printk(KERN_ERR
"ERROR: domain->span does not contain "
5511 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5512 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5516 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5520 printk(KERN_ERR
"ERROR: group is NULL\n");
5524 if (!group
->sgp
->power
) {
5525 printk(KERN_CONT
"\n");
5526 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5531 if (!cpumask_weight(sched_group_cpus(group
))) {
5532 printk(KERN_CONT
"\n");
5533 printk(KERN_ERR
"ERROR: empty group\n");
5537 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5538 printk(KERN_CONT
"\n");
5539 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5543 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5545 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5547 printk(KERN_CONT
" %s", str
);
5548 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5549 printk(KERN_CONT
" (cpu_power = %d)",
5553 group
= group
->next
;
5554 } while (group
!= sd
->groups
);
5555 printk(KERN_CONT
"\n");
5557 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5558 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5561 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5562 printk(KERN_ERR
"ERROR: parent span is not a superset "
5563 "of domain->span\n");
5567 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5571 if (!sched_domain_debug_enabled
)
5575 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5579 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5582 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5590 #else /* !CONFIG_SCHED_DEBUG */
5591 # define sched_domain_debug(sd, cpu) do { } while (0)
5592 #endif /* CONFIG_SCHED_DEBUG */
5594 static int sd_degenerate(struct sched_domain
*sd
)
5596 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5599 /* Following flags need at least 2 groups */
5600 if (sd
->flags
& (SD_LOAD_BALANCE
|
5601 SD_BALANCE_NEWIDLE
|
5605 SD_SHARE_PKG_RESOURCES
)) {
5606 if (sd
->groups
!= sd
->groups
->next
)
5610 /* Following flags don't use groups */
5611 if (sd
->flags
& (SD_WAKE_AFFINE
))
5618 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5620 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5622 if (sd_degenerate(parent
))
5625 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5628 /* Flags needing groups don't count if only 1 group in parent */
5629 if (parent
->groups
== parent
->groups
->next
) {
5630 pflags
&= ~(SD_LOAD_BALANCE
|
5631 SD_BALANCE_NEWIDLE
|
5635 SD_SHARE_PKG_RESOURCES
);
5636 if (nr_node_ids
== 1)
5637 pflags
&= ~SD_SERIALIZE
;
5639 if (~cflags
& pflags
)
5645 static void free_rootdomain(struct rcu_head
*rcu
)
5647 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5649 cpupri_cleanup(&rd
->cpupri
);
5650 free_cpumask_var(rd
->rto_mask
);
5651 free_cpumask_var(rd
->online
);
5652 free_cpumask_var(rd
->span
);
5656 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5658 struct root_domain
*old_rd
= NULL
;
5659 unsigned long flags
;
5661 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5666 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5669 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5672 * If we dont want to free the old_rt yet then
5673 * set old_rd to NULL to skip the freeing later
5676 if (!atomic_dec_and_test(&old_rd
->refcount
))
5680 atomic_inc(&rd
->refcount
);
5683 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5684 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5687 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5690 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5693 static int init_rootdomain(struct root_domain
*rd
)
5695 memset(rd
, 0, sizeof(*rd
));
5697 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5699 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5701 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5704 if (cpupri_init(&rd
->cpupri
) != 0)
5709 free_cpumask_var(rd
->rto_mask
);
5711 free_cpumask_var(rd
->online
);
5713 free_cpumask_var(rd
->span
);
5719 * By default the system creates a single root-domain with all cpus as
5720 * members (mimicking the global state we have today).
5722 struct root_domain def_root_domain
;
5724 static void init_defrootdomain(void)
5726 init_rootdomain(&def_root_domain
);
5728 atomic_set(&def_root_domain
.refcount
, 1);
5731 static struct root_domain
*alloc_rootdomain(void)
5733 struct root_domain
*rd
;
5735 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5739 if (init_rootdomain(rd
) != 0) {
5747 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5749 struct sched_group
*tmp
, *first
;
5758 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5763 } while (sg
!= first
);
5766 static void free_sched_domain(struct rcu_head
*rcu
)
5768 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5771 * If its an overlapping domain it has private groups, iterate and
5774 if (sd
->flags
& SD_OVERLAP
) {
5775 free_sched_groups(sd
->groups
, 1);
5776 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5777 kfree(sd
->groups
->sgp
);
5783 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5785 call_rcu(&sd
->rcu
, free_sched_domain
);
5788 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5790 for (; sd
; sd
= sd
->parent
)
5791 destroy_sched_domain(sd
, cpu
);
5795 * Keep a special pointer to the highest sched_domain that has
5796 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5797 * allows us to avoid some pointer chasing select_idle_sibling().
5799 * Also keep a unique ID per domain (we use the first cpu number in
5800 * the cpumask of the domain), this allows us to quickly tell if
5801 * two cpus are in the same cache domain, see cpus_share_cache().
5803 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5804 DEFINE_PER_CPU(int, sd_llc_id
);
5806 static void update_top_cache_domain(int cpu
)
5808 struct sched_domain
*sd
;
5811 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5813 id
= cpumask_first(sched_domain_span(sd
));
5815 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5816 per_cpu(sd_llc_id
, cpu
) = id
;
5820 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5821 * hold the hotplug lock.
5824 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5826 struct rq
*rq
= cpu_rq(cpu
);
5827 struct sched_domain
*tmp
;
5829 /* Remove the sched domains which do not contribute to scheduling. */
5830 for (tmp
= sd
; tmp
; ) {
5831 struct sched_domain
*parent
= tmp
->parent
;
5835 if (sd_parent_degenerate(tmp
, parent
)) {
5836 tmp
->parent
= parent
->parent
;
5838 parent
->parent
->child
= tmp
;
5839 destroy_sched_domain(parent
, cpu
);
5844 if (sd
&& sd_degenerate(sd
)) {
5847 destroy_sched_domain(tmp
, cpu
);
5852 sched_domain_debug(sd
, cpu
);
5854 rq_attach_root(rq
, rd
);
5856 rcu_assign_pointer(rq
->sd
, sd
);
5857 destroy_sched_domains(tmp
, cpu
);
5859 update_top_cache_domain(cpu
);
5862 /* cpus with isolated domains */
5863 static cpumask_var_t cpu_isolated_map
;
5865 /* Setup the mask of cpus configured for isolated domains */
5866 static int __init
isolated_cpu_setup(char *str
)
5868 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5869 cpulist_parse(str
, cpu_isolated_map
);
5873 __setup("isolcpus=", isolated_cpu_setup
);
5878 * find_next_best_node - find the next node to include in a sched_domain
5879 * @node: node whose sched_domain we're building
5880 * @used_nodes: nodes already in the sched_domain
5882 * Find the next node to include in a given scheduling domain. Simply
5883 * finds the closest node not already in the @used_nodes map.
5885 * Should use nodemask_t.
5887 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
5889 int i
, n
, val
, min_val
, best_node
= -1;
5893 for (i
= 0; i
< nr_node_ids
; i
++) {
5894 /* Start at @node */
5895 n
= (node
+ i
) % nr_node_ids
;
5897 if (!nr_cpus_node(n
))
5900 /* Skip already used nodes */
5901 if (node_isset(n
, *used_nodes
))
5904 /* Simple min distance search */
5905 val
= node_distance(node
, n
);
5907 if (val
< min_val
) {
5913 if (best_node
!= -1)
5914 node_set(best_node
, *used_nodes
);
5919 * sched_domain_node_span - get a cpumask for a node's sched_domain
5920 * @node: node whose cpumask we're constructing
5921 * @span: resulting cpumask
5923 * Given a node, construct a good cpumask for its sched_domain to span. It
5924 * should be one that prevents unnecessary balancing, but also spreads tasks
5927 static void sched_domain_node_span(int node
, struct cpumask
*span
)
5929 nodemask_t used_nodes
;
5932 cpumask_clear(span
);
5933 nodes_clear(used_nodes
);
5935 cpumask_or(span
, span
, cpumask_of_node(node
));
5936 node_set(node
, used_nodes
);
5938 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5939 int next_node
= find_next_best_node(node
, &used_nodes
);
5942 cpumask_or(span
, span
, cpumask_of_node(next_node
));
5946 static const struct cpumask
*cpu_node_mask(int cpu
)
5948 lockdep_assert_held(&sched_domains_mutex
);
5950 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
5952 return sched_domains_tmpmask
;
5955 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
5957 return cpu_possible_mask
;
5959 #endif /* CONFIG_NUMA */
5961 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5963 return cpumask_of_node(cpu_to_node(cpu
));
5966 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5969 struct sched_domain
**__percpu sd
;
5970 struct sched_group
**__percpu sg
;
5971 struct sched_group_power
**__percpu sgp
;
5975 struct sched_domain
** __percpu sd
;
5976 struct root_domain
*rd
;
5986 struct sched_domain_topology_level
;
5988 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5989 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5991 #define SDTL_OVERLAP 0x01
5993 struct sched_domain_topology_level
{
5994 sched_domain_init_f init
;
5995 sched_domain_mask_f mask
;
5997 struct sd_data data
;
6001 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6003 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6004 const struct cpumask
*span
= sched_domain_span(sd
);
6005 struct cpumask
*covered
= sched_domains_tmpmask
;
6006 struct sd_data
*sdd
= sd
->private;
6007 struct sched_domain
*child
;
6010 cpumask_clear(covered
);
6012 for_each_cpu(i
, span
) {
6013 struct cpumask
*sg_span
;
6015 if (cpumask_test_cpu(i
, covered
))
6018 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6019 GFP_KERNEL
, cpu_to_node(cpu
));
6024 sg_span
= sched_group_cpus(sg
);
6026 child
= *per_cpu_ptr(sdd
->sd
, i
);
6028 child
= child
->child
;
6029 cpumask_copy(sg_span
, sched_domain_span(child
));
6031 cpumask_set_cpu(i
, sg_span
);
6033 cpumask_or(covered
, covered
, sg_span
);
6035 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
6036 atomic_inc(&sg
->sgp
->ref
);
6038 if (cpumask_test_cpu(cpu
, sg_span
))
6048 sd
->groups
= groups
;
6053 free_sched_groups(first
, 0);
6058 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6060 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6061 struct sched_domain
*child
= sd
->child
;
6064 cpu
= cpumask_first(sched_domain_span(child
));
6067 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6068 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
6069 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
6076 * build_sched_groups will build a circular linked list of the groups
6077 * covered by the given span, and will set each group's ->cpumask correctly,
6078 * and ->cpu_power to 0.
6080 * Assumes the sched_domain tree is fully constructed
6083 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6085 struct sched_group
*first
= NULL
, *last
= NULL
;
6086 struct sd_data
*sdd
= sd
->private;
6087 const struct cpumask
*span
= sched_domain_span(sd
);
6088 struct cpumask
*covered
;
6091 get_group(cpu
, sdd
, &sd
->groups
);
6092 atomic_inc(&sd
->groups
->ref
);
6094 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
6097 lockdep_assert_held(&sched_domains_mutex
);
6098 covered
= sched_domains_tmpmask
;
6100 cpumask_clear(covered
);
6102 for_each_cpu(i
, span
) {
6103 struct sched_group
*sg
;
6104 int group
= get_group(i
, sdd
, &sg
);
6107 if (cpumask_test_cpu(i
, covered
))
6110 cpumask_clear(sched_group_cpus(sg
));
6113 for_each_cpu(j
, span
) {
6114 if (get_group(j
, sdd
, NULL
) != group
)
6117 cpumask_set_cpu(j
, covered
);
6118 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6133 * Initialize sched groups cpu_power.
6135 * cpu_power indicates the capacity of sched group, which is used while
6136 * distributing the load between different sched groups in a sched domain.
6137 * Typically cpu_power for all the groups in a sched domain will be same unless
6138 * there are asymmetries in the topology. If there are asymmetries, group
6139 * having more cpu_power will pickup more load compared to the group having
6142 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6144 struct sched_group
*sg
= sd
->groups
;
6146 WARN_ON(!sd
|| !sg
);
6149 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6151 } while (sg
!= sd
->groups
);
6153 if (cpu
!= group_first_cpu(sg
))
6156 update_group_power(sd
, cpu
);
6157 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6160 int __weak
arch_sd_sibling_asym_packing(void)
6162 return 0*SD_ASYM_PACKING
;
6166 * Initializers for schedule domains
6167 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6170 #ifdef CONFIG_SCHED_DEBUG
6171 # define SD_INIT_NAME(sd, type) sd->name = #type
6173 # define SD_INIT_NAME(sd, type) do { } while (0)
6176 #define SD_INIT_FUNC(type) \
6177 static noinline struct sched_domain * \
6178 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6180 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6181 *sd = SD_##type##_INIT; \
6182 SD_INIT_NAME(sd, type); \
6183 sd->private = &tl->data; \
6189 SD_INIT_FUNC(ALLNODES
)
6192 #ifdef CONFIG_SCHED_SMT
6193 SD_INIT_FUNC(SIBLING
)
6195 #ifdef CONFIG_SCHED_MC
6198 #ifdef CONFIG_SCHED_BOOK
6202 static int default_relax_domain_level
= -1;
6203 int sched_domain_level_max
;
6205 static int __init
setup_relax_domain_level(char *str
)
6209 val
= simple_strtoul(str
, NULL
, 0);
6210 if (val
< sched_domain_level_max
)
6211 default_relax_domain_level
= val
;
6215 __setup("relax_domain_level=", setup_relax_domain_level
);
6217 static void set_domain_attribute(struct sched_domain
*sd
,
6218 struct sched_domain_attr
*attr
)
6222 if (!attr
|| attr
->relax_domain_level
< 0) {
6223 if (default_relax_domain_level
< 0)
6226 request
= default_relax_domain_level
;
6228 request
= attr
->relax_domain_level
;
6229 if (request
< sd
->level
) {
6230 /* turn off idle balance on this domain */
6231 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6233 /* turn on idle balance on this domain */
6234 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6238 static void __sdt_free(const struct cpumask
*cpu_map
);
6239 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6241 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6242 const struct cpumask
*cpu_map
)
6246 if (!atomic_read(&d
->rd
->refcount
))
6247 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6249 free_percpu(d
->sd
); /* fall through */
6251 __sdt_free(cpu_map
); /* fall through */
6257 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6258 const struct cpumask
*cpu_map
)
6260 memset(d
, 0, sizeof(*d
));
6262 if (__sdt_alloc(cpu_map
))
6263 return sa_sd_storage
;
6264 d
->sd
= alloc_percpu(struct sched_domain
*);
6266 return sa_sd_storage
;
6267 d
->rd
= alloc_rootdomain();
6270 return sa_rootdomain
;
6274 * NULL the sd_data elements we've used to build the sched_domain and
6275 * sched_group structure so that the subsequent __free_domain_allocs()
6276 * will not free the data we're using.
6278 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6280 struct sd_data
*sdd
= sd
->private;
6282 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6283 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6285 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6286 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6288 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6289 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6292 #ifdef CONFIG_SCHED_SMT
6293 static const struct cpumask
*cpu_smt_mask(int cpu
)
6295 return topology_thread_cpumask(cpu
);
6300 * Topology list, bottom-up.
6302 static struct sched_domain_topology_level default_topology
[] = {
6303 #ifdef CONFIG_SCHED_SMT
6304 { sd_init_SIBLING
, cpu_smt_mask
, },
6306 #ifdef CONFIG_SCHED_MC
6307 { sd_init_MC
, cpu_coregroup_mask
, },
6309 #ifdef CONFIG_SCHED_BOOK
6310 { sd_init_BOOK
, cpu_book_mask
, },
6312 { sd_init_CPU
, cpu_cpu_mask
, },
6314 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
6315 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
6320 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6322 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6324 struct sched_domain_topology_level
*tl
;
6327 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6328 struct sd_data
*sdd
= &tl
->data
;
6330 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6334 sdd
->sg
= alloc_percpu(struct sched_group
*);
6338 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6342 for_each_cpu(j
, cpu_map
) {
6343 struct sched_domain
*sd
;
6344 struct sched_group
*sg
;
6345 struct sched_group_power
*sgp
;
6347 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6348 GFP_KERNEL
, cpu_to_node(j
));
6352 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6354 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6355 GFP_KERNEL
, cpu_to_node(j
));
6359 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6361 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
6362 GFP_KERNEL
, cpu_to_node(j
));
6366 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6373 static void __sdt_free(const struct cpumask
*cpu_map
)
6375 struct sched_domain_topology_level
*tl
;
6378 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6379 struct sd_data
*sdd
= &tl
->data
;
6381 for_each_cpu(j
, cpu_map
) {
6382 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, j
);
6383 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6384 free_sched_groups(sd
->groups
, 0);
6385 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6386 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6387 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6389 free_percpu(sdd
->sd
);
6390 free_percpu(sdd
->sg
);
6391 free_percpu(sdd
->sgp
);
6395 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6396 struct s_data
*d
, const struct cpumask
*cpu_map
,
6397 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6400 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6404 set_domain_attribute(sd
, attr
);
6405 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6407 sd
->level
= child
->level
+ 1;
6408 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6417 * Build sched domains for a given set of cpus and attach the sched domains
6418 * to the individual cpus
6420 static int build_sched_domains(const struct cpumask
*cpu_map
,
6421 struct sched_domain_attr
*attr
)
6423 enum s_alloc alloc_state
= sa_none
;
6424 struct sched_domain
*sd
;
6426 int i
, ret
= -ENOMEM
;
6428 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6429 if (alloc_state
!= sa_rootdomain
)
6432 /* Set up domains for cpus specified by the cpu_map. */
6433 for_each_cpu(i
, cpu_map
) {
6434 struct sched_domain_topology_level
*tl
;
6437 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6438 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6439 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6440 sd
->flags
|= SD_OVERLAP
;
6441 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6448 *per_cpu_ptr(d
.sd
, i
) = sd
;
6451 /* Build the groups for the domains */
6452 for_each_cpu(i
, cpu_map
) {
6453 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6454 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6455 if (sd
->flags
& SD_OVERLAP
) {
6456 if (build_overlap_sched_groups(sd
, i
))
6459 if (build_sched_groups(sd
, i
))
6465 /* Calculate CPU power for physical packages and nodes */
6466 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6467 if (!cpumask_test_cpu(i
, cpu_map
))
6470 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6471 claim_allocations(i
, sd
);
6472 init_sched_groups_power(i
, sd
);
6476 /* Attach the domains */
6478 for_each_cpu(i
, cpu_map
) {
6479 sd
= *per_cpu_ptr(d
.sd
, i
);
6480 cpu_attach_domain(sd
, d
.rd
, i
);
6486 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6490 static cpumask_var_t
*doms_cur
; /* current sched domains */
6491 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6492 static struct sched_domain_attr
*dattr_cur
;
6493 /* attribues of custom domains in 'doms_cur' */
6496 * Special case: If a kmalloc of a doms_cur partition (array of
6497 * cpumask) fails, then fallback to a single sched domain,
6498 * as determined by the single cpumask fallback_doms.
6500 static cpumask_var_t fallback_doms
;
6503 * arch_update_cpu_topology lets virtualized architectures update the
6504 * cpu core maps. It is supposed to return 1 if the topology changed
6505 * or 0 if it stayed the same.
6507 int __attribute__((weak
)) arch_update_cpu_topology(void)
6512 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6515 cpumask_var_t
*doms
;
6517 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6520 for (i
= 0; i
< ndoms
; i
++) {
6521 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6522 free_sched_domains(doms
, i
);
6529 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6532 for (i
= 0; i
< ndoms
; i
++)
6533 free_cpumask_var(doms
[i
]);
6538 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6539 * For now this just excludes isolated cpus, but could be used to
6540 * exclude other special cases in the future.
6542 static int init_sched_domains(const struct cpumask
*cpu_map
)
6546 arch_update_cpu_topology();
6548 doms_cur
= alloc_sched_domains(ndoms_cur
);
6550 doms_cur
= &fallback_doms
;
6551 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6553 err
= build_sched_domains(doms_cur
[0], NULL
);
6554 register_sched_domain_sysctl();
6560 * Detach sched domains from a group of cpus specified in cpu_map
6561 * These cpus will now be attached to the NULL domain
6563 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6568 for_each_cpu(i
, cpu_map
)
6569 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6573 /* handle null as "default" */
6574 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6575 struct sched_domain_attr
*new, int idx_new
)
6577 struct sched_domain_attr tmp
;
6584 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6585 new ? (new + idx_new
) : &tmp
,
6586 sizeof(struct sched_domain_attr
));
6590 * Partition sched domains as specified by the 'ndoms_new'
6591 * cpumasks in the array doms_new[] of cpumasks. This compares
6592 * doms_new[] to the current sched domain partitioning, doms_cur[].
6593 * It destroys each deleted domain and builds each new domain.
6595 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6596 * The masks don't intersect (don't overlap.) We should setup one
6597 * sched domain for each mask. CPUs not in any of the cpumasks will
6598 * not be load balanced. If the same cpumask appears both in the
6599 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6602 * The passed in 'doms_new' should be allocated using
6603 * alloc_sched_domains. This routine takes ownership of it and will
6604 * free_sched_domains it when done with it. If the caller failed the
6605 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6606 * and partition_sched_domains() will fallback to the single partition
6607 * 'fallback_doms', it also forces the domains to be rebuilt.
6609 * If doms_new == NULL it will be replaced with cpu_online_mask.
6610 * ndoms_new == 0 is a special case for destroying existing domains,
6611 * and it will not create the default domain.
6613 * Call with hotplug lock held
6615 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6616 struct sched_domain_attr
*dattr_new
)
6621 mutex_lock(&sched_domains_mutex
);
6623 /* always unregister in case we don't destroy any domains */
6624 unregister_sched_domain_sysctl();
6626 /* Let architecture update cpu core mappings. */
6627 new_topology
= arch_update_cpu_topology();
6629 n
= doms_new
? ndoms_new
: 0;
6631 /* Destroy deleted domains */
6632 for (i
= 0; i
< ndoms_cur
; i
++) {
6633 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6634 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6635 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6638 /* no match - a current sched domain not in new doms_new[] */
6639 detach_destroy_domains(doms_cur
[i
]);
6644 if (doms_new
== NULL
) {
6646 doms_new
= &fallback_doms
;
6647 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6648 WARN_ON_ONCE(dattr_new
);
6651 /* Build new domains */
6652 for (i
= 0; i
< ndoms_new
; i
++) {
6653 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6654 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6655 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6658 /* no match - add a new doms_new */
6659 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6664 /* Remember the new sched domains */
6665 if (doms_cur
!= &fallback_doms
)
6666 free_sched_domains(doms_cur
, ndoms_cur
);
6667 kfree(dattr_cur
); /* kfree(NULL) is safe */
6668 doms_cur
= doms_new
;
6669 dattr_cur
= dattr_new
;
6670 ndoms_cur
= ndoms_new
;
6672 register_sched_domain_sysctl();
6674 mutex_unlock(&sched_domains_mutex
);
6677 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6678 static void reinit_sched_domains(void)
6682 /* Destroy domains first to force the rebuild */
6683 partition_sched_domains(0, NULL
, NULL
);
6685 rebuild_sched_domains();
6689 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6691 unsigned int level
= 0;
6693 if (sscanf(buf
, "%u", &level
) != 1)
6697 * level is always be positive so don't check for
6698 * level < POWERSAVINGS_BALANCE_NONE which is 0
6699 * What happens on 0 or 1 byte write,
6700 * need to check for count as well?
6703 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
6707 sched_smt_power_savings
= level
;
6709 sched_mc_power_savings
= level
;
6711 reinit_sched_domains();
6716 #ifdef CONFIG_SCHED_MC
6717 static ssize_t
sched_mc_power_savings_show(struct device
*dev
,
6718 struct device_attribute
*attr
,
6721 return sprintf(buf
, "%u\n", sched_mc_power_savings
);
6723 static ssize_t
sched_mc_power_savings_store(struct device
*dev
,
6724 struct device_attribute
*attr
,
6725 const char *buf
, size_t count
)
6727 return sched_power_savings_store(buf
, count
, 0);
6729 static DEVICE_ATTR(sched_mc_power_savings
, 0644,
6730 sched_mc_power_savings_show
,
6731 sched_mc_power_savings_store
);
6734 #ifdef CONFIG_SCHED_SMT
6735 static ssize_t
sched_smt_power_savings_show(struct device
*dev
,
6736 struct device_attribute
*attr
,
6739 return sprintf(buf
, "%u\n", sched_smt_power_savings
);
6741 static ssize_t
sched_smt_power_savings_store(struct device
*dev
,
6742 struct device_attribute
*attr
,
6743 const char *buf
, size_t count
)
6745 return sched_power_savings_store(buf
, count
, 1);
6747 static DEVICE_ATTR(sched_smt_power_savings
, 0644,
6748 sched_smt_power_savings_show
,
6749 sched_smt_power_savings_store
);
6752 int __init
sched_create_sysfs_power_savings_entries(struct device
*dev
)
6756 #ifdef CONFIG_SCHED_SMT
6758 err
= device_create_file(dev
, &dev_attr_sched_smt_power_savings
);
6760 #ifdef CONFIG_SCHED_MC
6761 if (!err
&& mc_capable())
6762 err
= device_create_file(dev
, &dev_attr_sched_mc_power_savings
);
6766 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6769 * Update cpusets according to cpu_active mask. If cpusets are
6770 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6771 * around partition_sched_domains().
6773 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6776 switch (action
& ~CPU_TASKS_FROZEN
) {
6778 case CPU_DOWN_FAILED
:
6779 cpuset_update_active_cpus();
6786 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6789 switch (action
& ~CPU_TASKS_FROZEN
) {
6790 case CPU_DOWN_PREPARE
:
6791 cpuset_update_active_cpus();
6798 void __init
sched_init_smp(void)
6800 cpumask_var_t non_isolated_cpus
;
6802 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6803 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6806 mutex_lock(&sched_domains_mutex
);
6807 init_sched_domains(cpu_active_mask
);
6808 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6809 if (cpumask_empty(non_isolated_cpus
))
6810 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6811 mutex_unlock(&sched_domains_mutex
);
6814 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6815 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6817 /* RT runtime code needs to handle some hotplug events */
6818 hotcpu_notifier(update_runtime
, 0);
6822 /* Move init over to a non-isolated CPU */
6823 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6825 sched_init_granularity();
6826 free_cpumask_var(non_isolated_cpus
);
6828 init_sched_rt_class();
6831 void __init
sched_init_smp(void)
6833 sched_init_granularity();
6835 #endif /* CONFIG_SMP */
6837 const_debug
unsigned int sysctl_timer_migration
= 1;
6839 int in_sched_functions(unsigned long addr
)
6841 return in_lock_functions(addr
) ||
6842 (addr
>= (unsigned long)__sched_text_start
6843 && addr
< (unsigned long)__sched_text_end
);
6846 #ifdef CONFIG_CGROUP_SCHED
6847 struct task_group root_task_group
;
6850 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6852 void __init
sched_init(void)
6855 unsigned long alloc_size
= 0, ptr
;
6857 #ifdef CONFIG_FAIR_GROUP_SCHED
6858 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6860 #ifdef CONFIG_RT_GROUP_SCHED
6861 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6863 #ifdef CONFIG_CPUMASK_OFFSTACK
6864 alloc_size
+= num_possible_cpus() * cpumask_size();
6867 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6869 #ifdef CONFIG_FAIR_GROUP_SCHED
6870 root_task_group
.se
= (struct sched_entity
**)ptr
;
6871 ptr
+= nr_cpu_ids
* sizeof(void **);
6873 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6874 ptr
+= nr_cpu_ids
* sizeof(void **);
6876 #endif /* CONFIG_FAIR_GROUP_SCHED */
6877 #ifdef CONFIG_RT_GROUP_SCHED
6878 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6879 ptr
+= nr_cpu_ids
* sizeof(void **);
6881 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6882 ptr
+= nr_cpu_ids
* sizeof(void **);
6884 #endif /* CONFIG_RT_GROUP_SCHED */
6885 #ifdef CONFIG_CPUMASK_OFFSTACK
6886 for_each_possible_cpu(i
) {
6887 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6888 ptr
+= cpumask_size();
6890 #endif /* CONFIG_CPUMASK_OFFSTACK */
6894 init_defrootdomain();
6897 init_rt_bandwidth(&def_rt_bandwidth
,
6898 global_rt_period(), global_rt_runtime());
6900 #ifdef CONFIG_RT_GROUP_SCHED
6901 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6902 global_rt_period(), global_rt_runtime());
6903 #endif /* CONFIG_RT_GROUP_SCHED */
6905 #ifdef CONFIG_CGROUP_SCHED
6906 list_add(&root_task_group
.list
, &task_groups
);
6907 INIT_LIST_HEAD(&root_task_group
.children
);
6908 INIT_LIST_HEAD(&root_task_group
.siblings
);
6909 autogroup_init(&init_task
);
6911 #endif /* CONFIG_CGROUP_SCHED */
6913 #ifdef CONFIG_CGROUP_CPUACCT
6914 root_cpuacct
.cpustat
= &kernel_cpustat
;
6915 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6916 /* Too early, not expected to fail */
6917 BUG_ON(!root_cpuacct
.cpuusage
);
6919 for_each_possible_cpu(i
) {
6923 raw_spin_lock_init(&rq
->lock
);
6925 rq
->calc_load_active
= 0;
6926 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6927 init_cfs_rq(&rq
->cfs
);
6928 init_rt_rq(&rq
->rt
, rq
);
6929 #ifdef CONFIG_FAIR_GROUP_SCHED
6930 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6931 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6933 * How much cpu bandwidth does root_task_group get?
6935 * In case of task-groups formed thr' the cgroup filesystem, it
6936 * gets 100% of the cpu resources in the system. This overall
6937 * system cpu resource is divided among the tasks of
6938 * root_task_group and its child task-groups in a fair manner,
6939 * based on each entity's (task or task-group's) weight
6940 * (se->load.weight).
6942 * In other words, if root_task_group has 10 tasks of weight
6943 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6944 * then A0's share of the cpu resource is:
6946 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6948 * We achieve this by letting root_task_group's tasks sit
6949 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6951 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6952 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6953 #endif /* CONFIG_FAIR_GROUP_SCHED */
6955 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6956 #ifdef CONFIG_RT_GROUP_SCHED
6957 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6958 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6961 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6962 rq
->cpu_load
[j
] = 0;
6964 rq
->last_load_update_tick
= jiffies
;
6969 rq
->cpu_power
= SCHED_POWER_SCALE
;
6970 rq
->post_schedule
= 0;
6971 rq
->active_balance
= 0;
6972 rq
->next_balance
= jiffies
;
6977 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6979 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6981 rq_attach_root(rq
, &def_root_domain
);
6987 atomic_set(&rq
->nr_iowait
, 0);
6990 set_load_weight(&init_task
);
6992 #ifdef CONFIG_PREEMPT_NOTIFIERS
6993 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6996 #ifdef CONFIG_RT_MUTEXES
6997 plist_head_init(&init_task
.pi_waiters
);
7001 * The boot idle thread does lazy MMU switching as well:
7003 atomic_inc(&init_mm
.mm_count
);
7004 enter_lazy_tlb(&init_mm
, current
);
7007 * Make us the idle thread. Technically, schedule() should not be
7008 * called from this thread, however somewhere below it might be,
7009 * but because we are the idle thread, we just pick up running again
7010 * when this runqueue becomes "idle".
7012 init_idle(current
, smp_processor_id());
7014 calc_load_update
= jiffies
+ LOAD_FREQ
;
7017 * During early bootup we pretend to be a normal task:
7019 current
->sched_class
= &fair_sched_class
;
7022 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7023 /* May be allocated at isolcpus cmdline parse time */
7024 if (cpu_isolated_map
== NULL
)
7025 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7027 init_sched_fair_class();
7029 scheduler_running
= 1;
7032 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7033 static inline int preempt_count_equals(int preempt_offset
)
7035 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7037 return (nested
== preempt_offset
);
7040 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7042 static unsigned long prev_jiffy
; /* ratelimiting */
7044 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7045 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7046 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7048 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7050 prev_jiffy
= jiffies
;
7053 "BUG: sleeping function called from invalid context at %s:%d\n",
7056 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7057 in_atomic(), irqs_disabled(),
7058 current
->pid
, current
->comm
);
7060 debug_show_held_locks(current
);
7061 if (irqs_disabled())
7062 print_irqtrace_events(current
);
7065 EXPORT_SYMBOL(__might_sleep
);
7068 #ifdef CONFIG_MAGIC_SYSRQ
7069 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7071 const struct sched_class
*prev_class
= p
->sched_class
;
7072 int old_prio
= p
->prio
;
7077 dequeue_task(rq
, p
, 0);
7078 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7080 enqueue_task(rq
, p
, 0);
7081 resched_task(rq
->curr
);
7084 check_class_changed(rq
, p
, prev_class
, old_prio
);
7087 void normalize_rt_tasks(void)
7089 struct task_struct
*g
, *p
;
7090 unsigned long flags
;
7093 read_lock_irqsave(&tasklist_lock
, flags
);
7094 do_each_thread(g
, p
) {
7096 * Only normalize user tasks:
7101 p
->se
.exec_start
= 0;
7102 #ifdef CONFIG_SCHEDSTATS
7103 p
->se
.statistics
.wait_start
= 0;
7104 p
->se
.statistics
.sleep_start
= 0;
7105 p
->se
.statistics
.block_start
= 0;
7110 * Renice negative nice level userspace
7113 if (TASK_NICE(p
) < 0 && p
->mm
)
7114 set_user_nice(p
, 0);
7118 raw_spin_lock(&p
->pi_lock
);
7119 rq
= __task_rq_lock(p
);
7121 normalize_task(rq
, p
);
7123 __task_rq_unlock(rq
);
7124 raw_spin_unlock(&p
->pi_lock
);
7125 } while_each_thread(g
, p
);
7127 read_unlock_irqrestore(&tasklist_lock
, flags
);
7130 #endif /* CONFIG_MAGIC_SYSRQ */
7132 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7134 * These functions are only useful for the IA64 MCA handling, or kdb.
7136 * They can only be called when the whole system has been
7137 * stopped - every CPU needs to be quiescent, and no scheduling
7138 * activity can take place. Using them for anything else would
7139 * be a serious bug, and as a result, they aren't even visible
7140 * under any other configuration.
7144 * curr_task - return the current task for a given cpu.
7145 * @cpu: the processor in question.
7147 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7149 struct task_struct
*curr_task(int cpu
)
7151 return cpu_curr(cpu
);
7154 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7158 * set_curr_task - set the current task for a given cpu.
7159 * @cpu: the processor in question.
7160 * @p: the task pointer to set.
7162 * Description: This function must only be used when non-maskable interrupts
7163 * are serviced on a separate stack. It allows the architecture to switch the
7164 * notion of the current task on a cpu in a non-blocking manner. This function
7165 * must be called with all CPU's synchronized, and interrupts disabled, the
7166 * and caller must save the original value of the current task (see
7167 * curr_task() above) and restore that value before reenabling interrupts and
7168 * re-starting the system.
7170 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7172 void set_curr_task(int cpu
, struct task_struct
*p
)
7179 #ifdef CONFIG_CGROUP_SCHED
7180 /* task_group_lock serializes the addition/removal of task groups */
7181 static DEFINE_SPINLOCK(task_group_lock
);
7183 static void free_sched_group(struct task_group
*tg
)
7185 free_fair_sched_group(tg
);
7186 free_rt_sched_group(tg
);
7191 /* allocate runqueue etc for a new task group */
7192 struct task_group
*sched_create_group(struct task_group
*parent
)
7194 struct task_group
*tg
;
7195 unsigned long flags
;
7197 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7199 return ERR_PTR(-ENOMEM
);
7201 if (!alloc_fair_sched_group(tg
, parent
))
7204 if (!alloc_rt_sched_group(tg
, parent
))
7207 spin_lock_irqsave(&task_group_lock
, flags
);
7208 list_add_rcu(&tg
->list
, &task_groups
);
7210 WARN_ON(!parent
); /* root should already exist */
7212 tg
->parent
= parent
;
7213 INIT_LIST_HEAD(&tg
->children
);
7214 list_add_rcu(&tg
->siblings
, &parent
->children
);
7215 spin_unlock_irqrestore(&task_group_lock
, flags
);
7220 free_sched_group(tg
);
7221 return ERR_PTR(-ENOMEM
);
7224 /* rcu callback to free various structures associated with a task group */
7225 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7227 /* now it should be safe to free those cfs_rqs */
7228 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7231 /* Destroy runqueue etc associated with a task group */
7232 void sched_destroy_group(struct task_group
*tg
)
7234 unsigned long flags
;
7237 /* end participation in shares distribution */
7238 for_each_possible_cpu(i
)
7239 unregister_fair_sched_group(tg
, i
);
7241 spin_lock_irqsave(&task_group_lock
, flags
);
7242 list_del_rcu(&tg
->list
);
7243 list_del_rcu(&tg
->siblings
);
7244 spin_unlock_irqrestore(&task_group_lock
, flags
);
7246 /* wait for possible concurrent references to cfs_rqs complete */
7247 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7250 /* change task's runqueue when it moves between groups.
7251 * The caller of this function should have put the task in its new group
7252 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7253 * reflect its new group.
7255 void sched_move_task(struct task_struct
*tsk
)
7258 unsigned long flags
;
7261 rq
= task_rq_lock(tsk
, &flags
);
7263 running
= task_current(rq
, tsk
);
7267 dequeue_task(rq
, tsk
, 0);
7268 if (unlikely(running
))
7269 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7271 #ifdef CONFIG_FAIR_GROUP_SCHED
7272 if (tsk
->sched_class
->task_move_group
)
7273 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7276 set_task_rq(tsk
, task_cpu(tsk
));
7278 if (unlikely(running
))
7279 tsk
->sched_class
->set_curr_task(rq
);
7281 enqueue_task(rq
, tsk
, 0);
7283 task_rq_unlock(rq
, tsk
, &flags
);
7285 #endif /* CONFIG_CGROUP_SCHED */
7287 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7288 static unsigned long to_ratio(u64 period
, u64 runtime
)
7290 if (runtime
== RUNTIME_INF
)
7293 return div64_u64(runtime
<< 20, period
);
7297 #ifdef CONFIG_RT_GROUP_SCHED
7299 * Ensure that the real time constraints are schedulable.
7301 static DEFINE_MUTEX(rt_constraints_mutex
);
7303 /* Must be called with tasklist_lock held */
7304 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7306 struct task_struct
*g
, *p
;
7308 do_each_thread(g
, p
) {
7309 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7311 } while_each_thread(g
, p
);
7316 struct rt_schedulable_data
{
7317 struct task_group
*tg
;
7322 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7324 struct rt_schedulable_data
*d
= data
;
7325 struct task_group
*child
;
7326 unsigned long total
, sum
= 0;
7327 u64 period
, runtime
;
7329 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7330 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7333 period
= d
->rt_period
;
7334 runtime
= d
->rt_runtime
;
7338 * Cannot have more runtime than the period.
7340 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7344 * Ensure we don't starve existing RT tasks.
7346 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7349 total
= to_ratio(period
, runtime
);
7352 * Nobody can have more than the global setting allows.
7354 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7358 * The sum of our children's runtime should not exceed our own.
7360 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7361 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7362 runtime
= child
->rt_bandwidth
.rt_runtime
;
7364 if (child
== d
->tg
) {
7365 period
= d
->rt_period
;
7366 runtime
= d
->rt_runtime
;
7369 sum
+= to_ratio(period
, runtime
);
7378 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7382 struct rt_schedulable_data data
= {
7384 .rt_period
= period
,
7385 .rt_runtime
= runtime
,
7389 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7395 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7396 u64 rt_period
, u64 rt_runtime
)
7400 mutex_lock(&rt_constraints_mutex
);
7401 read_lock(&tasklist_lock
);
7402 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7406 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7407 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7408 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7410 for_each_possible_cpu(i
) {
7411 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7413 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7414 rt_rq
->rt_runtime
= rt_runtime
;
7415 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7417 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7419 read_unlock(&tasklist_lock
);
7420 mutex_unlock(&rt_constraints_mutex
);
7425 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7427 u64 rt_runtime
, rt_period
;
7429 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7430 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7431 if (rt_runtime_us
< 0)
7432 rt_runtime
= RUNTIME_INF
;
7434 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7437 long sched_group_rt_runtime(struct task_group
*tg
)
7441 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7444 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7445 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7446 return rt_runtime_us
;
7449 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7451 u64 rt_runtime
, rt_period
;
7453 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7454 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7459 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7462 long sched_group_rt_period(struct task_group
*tg
)
7466 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7467 do_div(rt_period_us
, NSEC_PER_USEC
);
7468 return rt_period_us
;
7471 static int sched_rt_global_constraints(void)
7473 u64 runtime
, period
;
7476 if (sysctl_sched_rt_period
<= 0)
7479 runtime
= global_rt_runtime();
7480 period
= global_rt_period();
7483 * Sanity check on the sysctl variables.
7485 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7488 mutex_lock(&rt_constraints_mutex
);
7489 read_lock(&tasklist_lock
);
7490 ret
= __rt_schedulable(NULL
, 0, 0);
7491 read_unlock(&tasklist_lock
);
7492 mutex_unlock(&rt_constraints_mutex
);
7497 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7499 /* Don't accept realtime tasks when there is no way for them to run */
7500 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7506 #else /* !CONFIG_RT_GROUP_SCHED */
7507 static int sched_rt_global_constraints(void)
7509 unsigned long flags
;
7512 if (sysctl_sched_rt_period
<= 0)
7516 * There's always some RT tasks in the root group
7517 * -- migration, kstopmachine etc..
7519 if (sysctl_sched_rt_runtime
== 0)
7522 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7523 for_each_possible_cpu(i
) {
7524 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7526 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7527 rt_rq
->rt_runtime
= global_rt_runtime();
7528 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7530 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7534 #endif /* CONFIG_RT_GROUP_SCHED */
7536 int sched_rt_handler(struct ctl_table
*table
, int write
,
7537 void __user
*buffer
, size_t *lenp
,
7541 int old_period
, old_runtime
;
7542 static DEFINE_MUTEX(mutex
);
7545 old_period
= sysctl_sched_rt_period
;
7546 old_runtime
= sysctl_sched_rt_runtime
;
7548 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7550 if (!ret
&& write
) {
7551 ret
= sched_rt_global_constraints();
7553 sysctl_sched_rt_period
= old_period
;
7554 sysctl_sched_rt_runtime
= old_runtime
;
7556 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7557 def_rt_bandwidth
.rt_period
=
7558 ns_to_ktime(global_rt_period());
7561 mutex_unlock(&mutex
);
7566 #ifdef CONFIG_CGROUP_SCHED
7568 /* return corresponding task_group object of a cgroup */
7569 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7571 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7572 struct task_group
, css
);
7575 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7577 struct task_group
*tg
, *parent
;
7579 if (!cgrp
->parent
) {
7580 /* This is early initialization for the top cgroup */
7581 return &root_task_group
.css
;
7584 parent
= cgroup_tg(cgrp
->parent
);
7585 tg
= sched_create_group(parent
);
7587 return ERR_PTR(-ENOMEM
);
7592 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7594 struct task_group
*tg
= cgroup_tg(cgrp
);
7596 sched_destroy_group(tg
);
7599 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7600 struct cgroup_taskset
*tset
)
7602 struct task_struct
*task
;
7604 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7605 #ifdef CONFIG_RT_GROUP_SCHED
7606 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7609 /* We don't support RT-tasks being in separate groups */
7610 if (task
->sched_class
!= &fair_sched_class
)
7617 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7618 struct cgroup_taskset
*tset
)
7620 struct task_struct
*task
;
7622 cgroup_taskset_for_each(task
, cgrp
, tset
)
7623 sched_move_task(task
);
7627 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7628 struct task_struct
*task
)
7631 * cgroup_exit() is called in the copy_process() failure path.
7632 * Ignore this case since the task hasn't ran yet, this avoids
7633 * trying to poke a half freed task state from generic code.
7635 if (!(task
->flags
& PF_EXITING
))
7638 sched_move_task(task
);
7641 #ifdef CONFIG_FAIR_GROUP_SCHED
7642 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7645 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7648 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7650 struct task_group
*tg
= cgroup_tg(cgrp
);
7652 return (u64
) scale_load_down(tg
->shares
);
7655 #ifdef CONFIG_CFS_BANDWIDTH
7656 static DEFINE_MUTEX(cfs_constraints_mutex
);
7658 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7659 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7661 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7663 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7665 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7666 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7668 if (tg
== &root_task_group
)
7672 * Ensure we have at some amount of bandwidth every period. This is
7673 * to prevent reaching a state of large arrears when throttled via
7674 * entity_tick() resulting in prolonged exit starvation.
7676 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7680 * Likewise, bound things on the otherside by preventing insane quota
7681 * periods. This also allows us to normalize in computing quota
7684 if (period
> max_cfs_quota_period
)
7687 mutex_lock(&cfs_constraints_mutex
);
7688 ret
= __cfs_schedulable(tg
, period
, quota
);
7692 runtime_enabled
= quota
!= RUNTIME_INF
;
7693 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7694 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7695 raw_spin_lock_irq(&cfs_b
->lock
);
7696 cfs_b
->period
= ns_to_ktime(period
);
7697 cfs_b
->quota
= quota
;
7699 __refill_cfs_bandwidth_runtime(cfs_b
);
7700 /* restart the period timer (if active) to handle new period expiry */
7701 if (runtime_enabled
&& cfs_b
->timer_active
) {
7702 /* force a reprogram */
7703 cfs_b
->timer_active
= 0;
7704 __start_cfs_bandwidth(cfs_b
);
7706 raw_spin_unlock_irq(&cfs_b
->lock
);
7708 for_each_possible_cpu(i
) {
7709 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7710 struct rq
*rq
= cfs_rq
->rq
;
7712 raw_spin_lock_irq(&rq
->lock
);
7713 cfs_rq
->runtime_enabled
= runtime_enabled
;
7714 cfs_rq
->runtime_remaining
= 0;
7716 if (cfs_rq
->throttled
)
7717 unthrottle_cfs_rq(cfs_rq
);
7718 raw_spin_unlock_irq(&rq
->lock
);
7721 mutex_unlock(&cfs_constraints_mutex
);
7726 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7730 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7731 if (cfs_quota_us
< 0)
7732 quota
= RUNTIME_INF
;
7734 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7736 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7739 long tg_get_cfs_quota(struct task_group
*tg
)
7743 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7746 quota_us
= tg
->cfs_bandwidth
.quota
;
7747 do_div(quota_us
, NSEC_PER_USEC
);
7752 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7756 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7757 quota
= tg
->cfs_bandwidth
.quota
;
7759 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7762 long tg_get_cfs_period(struct task_group
*tg
)
7766 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7767 do_div(cfs_period_us
, NSEC_PER_USEC
);
7769 return cfs_period_us
;
7772 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7774 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7777 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7780 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7783 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7785 return tg_get_cfs_period(cgroup_tg(cgrp
));
7788 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7791 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7794 struct cfs_schedulable_data
{
7795 struct task_group
*tg
;
7800 * normalize group quota/period to be quota/max_period
7801 * note: units are usecs
7803 static u64
normalize_cfs_quota(struct task_group
*tg
,
7804 struct cfs_schedulable_data
*d
)
7812 period
= tg_get_cfs_period(tg
);
7813 quota
= tg_get_cfs_quota(tg
);
7816 /* note: these should typically be equivalent */
7817 if (quota
== RUNTIME_INF
|| quota
== -1)
7820 return to_ratio(period
, quota
);
7823 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7825 struct cfs_schedulable_data
*d
= data
;
7826 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7827 s64 quota
= 0, parent_quota
= -1;
7830 quota
= RUNTIME_INF
;
7832 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7834 quota
= normalize_cfs_quota(tg
, d
);
7835 parent_quota
= parent_b
->hierarchal_quota
;
7838 * ensure max(child_quota) <= parent_quota, inherit when no
7841 if (quota
== RUNTIME_INF
)
7842 quota
= parent_quota
;
7843 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7846 cfs_b
->hierarchal_quota
= quota
;
7851 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7854 struct cfs_schedulable_data data
= {
7860 if (quota
!= RUNTIME_INF
) {
7861 do_div(data
.period
, NSEC_PER_USEC
);
7862 do_div(data
.quota
, NSEC_PER_USEC
);
7866 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7872 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7873 struct cgroup_map_cb
*cb
)
7875 struct task_group
*tg
= cgroup_tg(cgrp
);
7876 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7878 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7879 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7880 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7884 #endif /* CONFIG_CFS_BANDWIDTH */
7885 #endif /* CONFIG_FAIR_GROUP_SCHED */
7887 #ifdef CONFIG_RT_GROUP_SCHED
7888 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7891 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7894 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7896 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7899 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7902 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7905 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7907 return sched_group_rt_period(cgroup_tg(cgrp
));
7909 #endif /* CONFIG_RT_GROUP_SCHED */
7911 static struct cftype cpu_files
[] = {
7912 #ifdef CONFIG_FAIR_GROUP_SCHED
7915 .read_u64
= cpu_shares_read_u64
,
7916 .write_u64
= cpu_shares_write_u64
,
7919 #ifdef CONFIG_CFS_BANDWIDTH
7921 .name
= "cfs_quota_us",
7922 .read_s64
= cpu_cfs_quota_read_s64
,
7923 .write_s64
= cpu_cfs_quota_write_s64
,
7926 .name
= "cfs_period_us",
7927 .read_u64
= cpu_cfs_period_read_u64
,
7928 .write_u64
= cpu_cfs_period_write_u64
,
7932 .read_map
= cpu_stats_show
,
7935 #ifdef CONFIG_RT_GROUP_SCHED
7937 .name
= "rt_runtime_us",
7938 .read_s64
= cpu_rt_runtime_read
,
7939 .write_s64
= cpu_rt_runtime_write
,
7942 .name
= "rt_period_us",
7943 .read_u64
= cpu_rt_period_read_uint
,
7944 .write_u64
= cpu_rt_period_write_uint
,
7949 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7951 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7954 struct cgroup_subsys cpu_cgroup_subsys
= {
7956 .create
= cpu_cgroup_create
,
7957 .destroy
= cpu_cgroup_destroy
,
7958 .can_attach
= cpu_cgroup_can_attach
,
7959 .attach
= cpu_cgroup_attach
,
7960 .exit
= cpu_cgroup_exit
,
7961 .populate
= cpu_cgroup_populate
,
7962 .subsys_id
= cpu_cgroup_subsys_id
,
7966 #endif /* CONFIG_CGROUP_SCHED */
7968 #ifdef CONFIG_CGROUP_CPUACCT
7971 * CPU accounting code for task groups.
7973 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7974 * (balbir@in.ibm.com).
7977 /* create a new cpu accounting group */
7978 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
7983 return &root_cpuacct
.css
;
7985 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7989 ca
->cpuusage
= alloc_percpu(u64
);
7993 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7995 goto out_free_cpuusage
;
8000 free_percpu(ca
->cpuusage
);
8004 return ERR_PTR(-ENOMEM
);
8007 /* destroy an existing cpu accounting group */
8008 static void cpuacct_destroy(struct cgroup
*cgrp
)
8010 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8012 free_percpu(ca
->cpustat
);
8013 free_percpu(ca
->cpuusage
);
8017 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8019 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8022 #ifndef CONFIG_64BIT
8024 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8026 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8028 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8036 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8038 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8040 #ifndef CONFIG_64BIT
8042 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8044 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8046 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8052 /* return total cpu usage (in nanoseconds) of a group */
8053 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8055 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8056 u64 totalcpuusage
= 0;
8059 for_each_present_cpu(i
)
8060 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8062 return totalcpuusage
;
8065 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8068 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8077 for_each_present_cpu(i
)
8078 cpuacct_cpuusage_write(ca
, i
, 0);
8084 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8087 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8091 for_each_present_cpu(i
) {
8092 percpu
= cpuacct_cpuusage_read(ca
, i
);
8093 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8095 seq_printf(m
, "\n");
8099 static const char *cpuacct_stat_desc
[] = {
8100 [CPUACCT_STAT_USER
] = "user",
8101 [CPUACCT_STAT_SYSTEM
] = "system",
8104 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8105 struct cgroup_map_cb
*cb
)
8107 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8111 for_each_online_cpu(cpu
) {
8112 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8113 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8114 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8116 val
= cputime64_to_clock_t(val
);
8117 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8120 for_each_online_cpu(cpu
) {
8121 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8122 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8123 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8124 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8127 val
= cputime64_to_clock_t(val
);
8128 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8133 static struct cftype files
[] = {
8136 .read_u64
= cpuusage_read
,
8137 .write_u64
= cpuusage_write
,
8140 .name
= "usage_percpu",
8141 .read_seq_string
= cpuacct_percpu_seq_read
,
8145 .read_map
= cpuacct_stats_show
,
8149 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8151 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8155 * charge this task's execution time to its accounting group.
8157 * called with rq->lock held.
8159 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8164 if (unlikely(!cpuacct_subsys
.active
))
8167 cpu
= task_cpu(tsk
);
8173 for (; ca
; ca
= parent_ca(ca
)) {
8174 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8175 *cpuusage
+= cputime
;
8181 struct cgroup_subsys cpuacct_subsys
= {
8183 .create
= cpuacct_create
,
8184 .destroy
= cpuacct_destroy
,
8185 .populate
= cpuacct_populate
,
8186 .subsys_id
= cpuacct_subsys_id
,
8188 #endif /* CONFIG_CGROUP_CPUACCT */