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>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
83 #include "../workqueue_sched.h"
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/sched.h>
88 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
91 ktime_t soft
, hard
, now
;
94 if (hrtimer_active(period_timer
))
97 now
= hrtimer_cb_get_time(period_timer
);
98 hrtimer_forward(period_timer
, now
, period
);
100 soft
= hrtimer_get_softexpires(period_timer
);
101 hard
= hrtimer_get_expires(period_timer
);
102 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
103 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
104 HRTIMER_MODE_ABS_PINNED
, 0);
108 DEFINE_MUTEX(sched_domains_mutex
);
109 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
111 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
113 void update_rq_clock(struct rq
*rq
)
117 if (rq
->skip_clock_update
> 0)
120 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
122 update_rq_clock_task(rq
, delta
);
126 * Debugging: various feature bits
129 #define SCHED_FEAT(name, enabled) \
130 (1UL << __SCHED_FEAT_##name) * enabled |
132 const_debug
unsigned int sysctl_sched_features
=
133 #include "features.h"
138 #ifdef CONFIG_SCHED_DEBUG
139 #define SCHED_FEAT(name, enabled) \
142 static __read_mostly
char *sched_feat_names
[] = {
143 #include "features.h"
149 static int sched_feat_show(struct seq_file
*m
, void *v
)
153 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
154 if (!(sysctl_sched_features
& (1UL << i
)))
156 seq_printf(m
, "%s ", sched_feat_names
[i
]);
163 #ifdef HAVE_JUMP_LABEL
165 #define jump_label_key__true jump_label_key_enabled
166 #define jump_label_key__false jump_label_key_disabled
168 #define SCHED_FEAT(name, enabled) \
169 jump_label_key__##enabled ,
171 struct jump_label_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
172 #include "features.h"
177 static void sched_feat_disable(int i
)
179 if (jump_label_enabled(&sched_feat_keys
[i
]))
180 jump_label_dec(&sched_feat_keys
[i
]);
183 static void sched_feat_enable(int i
)
185 if (!jump_label_enabled(&sched_feat_keys
[i
]))
186 jump_label_inc(&sched_feat_keys
[i
]);
189 static void sched_feat_disable(int i
) { };
190 static void sched_feat_enable(int i
) { };
191 #endif /* HAVE_JUMP_LABEL */
194 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
195 size_t cnt
, loff_t
*ppos
)
205 if (copy_from_user(&buf
, ubuf
, cnt
))
211 if (strncmp(cmp
, "NO_", 3) == 0) {
216 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
217 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
219 sysctl_sched_features
&= ~(1UL << i
);
220 sched_feat_disable(i
);
222 sysctl_sched_features
|= (1UL << i
);
223 sched_feat_enable(i
);
229 if (i
== __SCHED_FEAT_NR
)
237 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
239 return single_open(filp
, sched_feat_show
, NULL
);
242 static const struct file_operations sched_feat_fops
= {
243 .open
= sched_feat_open
,
244 .write
= sched_feat_write
,
247 .release
= single_release
,
250 static __init
int sched_init_debug(void)
252 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
257 late_initcall(sched_init_debug
);
258 #endif /* CONFIG_SCHED_DEBUG */
261 * Number of tasks to iterate in a single balance run.
262 * Limited because this is done with IRQs disabled.
264 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
267 * period over which we average the RT time consumption, measured
272 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
275 * period over which we measure -rt task cpu usage in us.
278 unsigned int sysctl_sched_rt_period
= 1000000;
280 __read_mostly
int scheduler_running
;
283 * part of the period that we allow rt tasks to run in us.
286 int sysctl_sched_rt_runtime
= 950000;
291 * __task_rq_lock - lock the rq @p resides on.
293 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
298 lockdep_assert_held(&p
->pi_lock
);
302 raw_spin_lock(&rq
->lock
);
303 if (likely(rq
== task_rq(p
)))
305 raw_spin_unlock(&rq
->lock
);
310 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
312 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
313 __acquires(p
->pi_lock
)
319 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
321 raw_spin_lock(&rq
->lock
);
322 if (likely(rq
== task_rq(p
)))
324 raw_spin_unlock(&rq
->lock
);
325 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
329 static void __task_rq_unlock(struct rq
*rq
)
332 raw_spin_unlock(&rq
->lock
);
336 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
338 __releases(p
->pi_lock
)
340 raw_spin_unlock(&rq
->lock
);
341 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
345 * this_rq_lock - lock this runqueue and disable interrupts.
347 static struct rq
*this_rq_lock(void)
354 raw_spin_lock(&rq
->lock
);
359 #ifdef CONFIG_SCHED_HRTICK
361 * Use HR-timers to deliver accurate preemption points.
363 * Its all a bit involved since we cannot program an hrt while holding the
364 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * When we get rescheduled we reprogram the hrtick_timer outside of the
371 static void hrtick_clear(struct rq
*rq
)
373 if (hrtimer_active(&rq
->hrtick_timer
))
374 hrtimer_cancel(&rq
->hrtick_timer
);
378 * High-resolution timer tick.
379 * Runs from hardirq context with interrupts disabled.
381 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
383 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
385 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
387 raw_spin_lock(&rq
->lock
);
389 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
390 raw_spin_unlock(&rq
->lock
);
392 return HRTIMER_NORESTART
;
397 * called from hardirq (IPI) context
399 static void __hrtick_start(void *arg
)
403 raw_spin_lock(&rq
->lock
);
404 hrtimer_restart(&rq
->hrtick_timer
);
405 rq
->hrtick_csd_pending
= 0;
406 raw_spin_unlock(&rq
->lock
);
410 * Called to set the hrtick timer state.
412 * called with rq->lock held and irqs disabled
414 void hrtick_start(struct rq
*rq
, u64 delay
)
416 struct hrtimer
*timer
= &rq
->hrtick_timer
;
417 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
419 hrtimer_set_expires(timer
, time
);
421 if (rq
== this_rq()) {
422 hrtimer_restart(timer
);
423 } else if (!rq
->hrtick_csd_pending
) {
424 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
425 rq
->hrtick_csd_pending
= 1;
430 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
432 int cpu
= (int)(long)hcpu
;
435 case CPU_UP_CANCELED
:
436 case CPU_UP_CANCELED_FROZEN
:
437 case CPU_DOWN_PREPARE
:
438 case CPU_DOWN_PREPARE_FROZEN
:
440 case CPU_DEAD_FROZEN
:
441 hrtick_clear(cpu_rq(cpu
));
448 static __init
void init_hrtick(void)
450 hotcpu_notifier(hotplug_hrtick
, 0);
454 * Called to set the hrtick timer state.
456 * called with rq->lock held and irqs disabled
458 void hrtick_start(struct rq
*rq
, u64 delay
)
460 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
461 HRTIMER_MODE_REL_PINNED
, 0);
464 static inline void init_hrtick(void)
467 #endif /* CONFIG_SMP */
469 static void init_rq_hrtick(struct rq
*rq
)
472 rq
->hrtick_csd_pending
= 0;
474 rq
->hrtick_csd
.flags
= 0;
475 rq
->hrtick_csd
.func
= __hrtick_start
;
476 rq
->hrtick_csd
.info
= rq
;
479 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
480 rq
->hrtick_timer
.function
= hrtick
;
482 #else /* CONFIG_SCHED_HRTICK */
483 static inline void hrtick_clear(struct rq
*rq
)
487 static inline void init_rq_hrtick(struct rq
*rq
)
491 static inline void init_hrtick(void)
494 #endif /* CONFIG_SCHED_HRTICK */
497 * resched_task - mark a task 'to be rescheduled now'.
499 * On UP this means the setting of the need_resched flag, on SMP it
500 * might also involve a cross-CPU call to trigger the scheduler on
505 #ifndef tsk_is_polling
506 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509 void resched_task(struct task_struct
*p
)
513 assert_raw_spin_locked(&task_rq(p
)->lock
);
515 if (test_tsk_need_resched(p
))
518 set_tsk_need_resched(p
);
521 if (cpu
== smp_processor_id())
524 /* NEED_RESCHED must be visible before we test polling */
526 if (!tsk_is_polling(p
))
527 smp_send_reschedule(cpu
);
530 void resched_cpu(int cpu
)
532 struct rq
*rq
= cpu_rq(cpu
);
535 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
537 resched_task(cpu_curr(cpu
));
538 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
543 * In the semi idle case, use the nearest busy cpu for migrating timers
544 * from an idle cpu. This is good for power-savings.
546 * We don't do similar optimization for completely idle system, as
547 * selecting an idle cpu will add more delays to the timers than intended
548 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
550 int get_nohz_timer_target(void)
552 int cpu
= smp_processor_id();
554 struct sched_domain
*sd
;
557 for_each_domain(cpu
, sd
) {
558 for_each_cpu(i
, sched_domain_span(sd
)) {
570 * When add_timer_on() enqueues a timer into the timer wheel of an
571 * idle CPU then this timer might expire before the next timer event
572 * which is scheduled to wake up that CPU. In case of a completely
573 * idle system the next event might even be infinite time into the
574 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
575 * leaves the inner idle loop so the newly added timer is taken into
576 * account when the CPU goes back to idle and evaluates the timer
577 * wheel for the next timer event.
579 void wake_up_idle_cpu(int cpu
)
581 struct rq
*rq
= cpu_rq(cpu
);
583 if (cpu
== smp_processor_id())
587 * This is safe, as this function is called with the timer
588 * wheel base lock of (cpu) held. When the CPU is on the way
589 * to idle and has not yet set rq->curr to idle then it will
590 * be serialized on the timer wheel base lock and take the new
591 * timer into account automatically.
593 if (rq
->curr
!= rq
->idle
)
597 * We can set TIF_RESCHED on the idle task of the other CPU
598 * lockless. The worst case is that the other CPU runs the
599 * idle task through an additional NOOP schedule()
601 set_tsk_need_resched(rq
->idle
);
603 /* NEED_RESCHED must be visible before we test polling */
605 if (!tsk_is_polling(rq
->idle
))
606 smp_send_reschedule(cpu
);
609 static inline bool got_nohz_idle_kick(void)
611 int cpu
= smp_processor_id();
612 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
615 #else /* CONFIG_NO_HZ */
617 static inline bool got_nohz_idle_kick(void)
622 #endif /* CONFIG_NO_HZ */
624 void sched_avg_update(struct rq
*rq
)
626 s64 period
= sched_avg_period();
628 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
630 * Inline assembly required to prevent the compiler
631 * optimising this loop into a divmod call.
632 * See __iter_div_u64_rem() for another example of this.
634 asm("" : "+rm" (rq
->age_stamp
));
635 rq
->age_stamp
+= period
;
640 #else /* !CONFIG_SMP */
641 void resched_task(struct task_struct
*p
)
643 assert_raw_spin_locked(&task_rq(p
)->lock
);
644 set_tsk_need_resched(p
);
646 #endif /* CONFIG_SMP */
648 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
649 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
651 * Iterate task_group tree rooted at *from, calling @down when first entering a
652 * node and @up when leaving it for the final time.
654 * Caller must hold rcu_lock or sufficient equivalent.
656 int walk_tg_tree_from(struct task_group
*from
,
657 tg_visitor down
, tg_visitor up
, void *data
)
659 struct task_group
*parent
, *child
;
665 ret
= (*down
)(parent
, data
);
668 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
675 ret
= (*up
)(parent
, data
);
676 if (ret
|| parent
== from
)
680 parent
= parent
->parent
;
687 int tg_nop(struct task_group
*tg
, void *data
)
693 void update_cpu_load(struct rq
*this_rq
);
695 static void set_load_weight(struct task_struct
*p
)
697 int prio
= p
->static_prio
- MAX_RT_PRIO
;
698 struct load_weight
*load
= &p
->se
.load
;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p
->policy
== SCHED_IDLE
) {
704 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
705 load
->inv_weight
= WMULT_IDLEPRIO
;
709 load
->weight
= scale_load(prio_to_weight
[prio
]);
710 load
->inv_weight
= prio_to_wmult
[prio
];
713 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
716 sched_info_queued(p
);
717 p
->sched_class
->enqueue_task(rq
, p
, flags
);
720 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
723 sched_info_dequeued(p
);
724 p
->sched_class
->dequeue_task(rq
, p
, flags
);
727 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
729 if (task_contributes_to_load(p
))
730 rq
->nr_uninterruptible
--;
732 enqueue_task(rq
, p
, flags
);
735 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
737 if (task_contributes_to_load(p
))
738 rq
->nr_uninterruptible
++;
740 dequeue_task(rq
, p
, flags
);
743 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
756 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
757 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
759 static DEFINE_PER_CPU(u64
, irq_start_time
);
760 static int sched_clock_irqtime
;
762 void enable_sched_clock_irqtime(void)
764 sched_clock_irqtime
= 1;
767 void disable_sched_clock_irqtime(void)
769 sched_clock_irqtime
= 0;
773 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
775 static inline void irq_time_write_begin(void)
777 __this_cpu_inc(irq_time_seq
.sequence
);
781 static inline void irq_time_write_end(void)
784 __this_cpu_inc(irq_time_seq
.sequence
);
787 static inline u64
irq_time_read(int cpu
)
793 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
794 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
795 per_cpu(cpu_hardirq_time
, cpu
);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
800 #else /* CONFIG_64BIT */
801 static inline void irq_time_write_begin(void)
805 static inline void irq_time_write_end(void)
809 static inline u64
irq_time_read(int cpu
)
811 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
813 #endif /* CONFIG_64BIT */
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
819 void account_system_vtime(struct task_struct
*curr
)
825 if (!sched_clock_irqtime
)
828 local_irq_save(flags
);
830 cpu
= smp_processor_id();
831 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
832 __this_cpu_add(irq_start_time
, delta
);
834 irq_time_write_begin();
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
842 __this_cpu_add(cpu_hardirq_time
, delta
);
843 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time
, delta
);
846 irq_time_write_end();
847 local_irq_restore(flags
);
849 EXPORT_SYMBOL_GPL(account_system_vtime
);
851 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
853 #ifdef CONFIG_PARAVIRT
854 static inline u64
steal_ticks(u64 steal
)
856 if (unlikely(steal
> NSEC_PER_SEC
))
857 return div_u64(steal
, TICK_NSEC
);
859 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
863 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
869 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal
= 0, irq_delta
= 0;
872 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
890 if (irq_delta
> delta
)
893 rq
->prev_irq_time
+= irq_delta
;
896 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_branch((¶virt_steal_rq_enabled
))) {
900 steal
= paravirt_steal_clock(cpu_of(rq
));
901 steal
-= rq
->prev_steal_time_rq
;
903 if (unlikely(steal
> delta
))
906 st
= steal_ticks(steal
);
907 steal
= st
* TICK_NSEC
;
909 rq
->prev_steal_time_rq
+= steal
;
915 rq
->clock_task
+= delta
;
917 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
919 sched_rt_avg_update(rq
, irq_delta
+ steal
);
923 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
924 static int irqtime_account_hi_update(void)
926 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
931 local_irq_save(flags
);
932 latest_ns
= this_cpu_read(cpu_hardirq_time
);
933 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_IRQ
])
935 local_irq_restore(flags
);
939 static int irqtime_account_si_update(void)
941 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
946 local_irq_save(flags
);
947 latest_ns
= this_cpu_read(cpu_softirq_time
);
948 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_SOFTIRQ
])
950 local_irq_restore(flags
);
954 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
956 #define sched_clock_irqtime (0)
960 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
962 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
963 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
974 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
976 stop
->sched_class
= &stop_sched_class
;
979 cpu_rq(cpu
)->stop
= stop
;
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
986 old_stop
->sched_class
= &rt_sched_class
;
991 * __normal_prio - return the priority that is based on the static prio
993 static inline int __normal_prio(struct task_struct
*p
)
995 return p
->static_prio
;
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1005 static inline int normal_prio(struct task_struct
*p
)
1009 if (task_has_rt_policy(p
))
1010 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1012 prio
= __normal_prio(p
);
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1023 static int effective_prio(struct task_struct
*p
)
1025 p
->normal_prio
= normal_prio(p
);
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1031 if (!rt_prio(p
->prio
))
1032 return p
->normal_prio
;
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1040 inline int task_curr(const struct task_struct
*p
)
1042 return cpu_curr(task_cpu(p
)) == p
;
1045 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1046 const struct sched_class
*prev_class
,
1049 if (prev_class
!= p
->sched_class
) {
1050 if (prev_class
->switched_from
)
1051 prev_class
->switched_from(rq
, p
);
1052 p
->sched_class
->switched_to(rq
, p
);
1053 } else if (oldprio
!= p
->prio
)
1054 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1057 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1059 const struct sched_class
*class;
1061 if (p
->sched_class
== rq
->curr
->sched_class
) {
1062 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1064 for_each_class(class) {
1065 if (class == rq
->curr
->sched_class
)
1067 if (class == p
->sched_class
) {
1068 resched_task(rq
->curr
);
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1078 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1079 rq
->skip_clock_update
= 1;
1083 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1085 #ifdef CONFIG_SCHED_DEBUG
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1090 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1091 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
1093 #ifdef CONFIG_LOCKDEP
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see set_task_rq().
1101 * Furthermore, all task_rq users should acquire both locks, see
1104 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1105 lockdep_is_held(&task_rq(p
)->lock
)));
1109 trace_sched_migrate_task(p
, new_cpu
);
1111 if (task_cpu(p
) != new_cpu
) {
1112 p
->se
.nr_migrations
++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1116 __set_task_cpu(p
, new_cpu
);
1119 struct migration_arg
{
1120 struct task_struct
*task
;
1124 static int migration_cpu_stop(void *data
);
1127 * wait_task_inactive - wait for a thread to unschedule.
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1142 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1144 unsigned long flags
;
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1169 while (task_running(rq
, p
)) {
1170 if (match_state
&& unlikely(p
->state
!= match_state
))
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1180 rq
= task_rq_lock(p
, &flags
);
1181 trace_sched_wait_task(p
);
1182 running
= task_running(rq
, p
);
1185 if (!match_state
|| p
->state
== match_state
)
1186 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1187 task_rq_unlock(rq
, p
, &flags
);
1190 * If it changed from the expected state, bail out now.
1192 if (unlikely(!ncsw
))
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1199 * Oops. Go back and try again..
1201 if (unlikely(running
)) {
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1215 if (unlikely(on_rq
)) {
1216 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1218 set_current_state(TASK_UNINTERRUPTIBLE
);
1219 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1247 void kick_process(struct task_struct
*p
)
1253 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1254 smp_send_reschedule(cpu
);
1257 EXPORT_SYMBOL_GPL(kick_process
);
1258 #endif /* CONFIG_SMP */
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1264 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1267 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1269 /* Look for allowed, online CPU in same node. */
1270 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
1271 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1274 /* Any allowed, online CPU? */
1275 dest_cpu
= cpumask_any_and(tsk_cpus_allowed(p
), cpu_active_mask
);
1276 if (dest_cpu
< nr_cpu_ids
)
1279 /* No more Mr. Nice Guy. */
1280 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
1282 * Don't tell them about moving exiting tasks or
1283 * kernel threads (both mm NULL), since they never
1286 if (p
->mm
&& printk_ratelimit()) {
1287 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
1288 task_pid_nr(p
), p
->comm
, cpu
);
1295 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1298 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1300 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1303 * In order not to call set_task_cpu() on a blocking task we need
1304 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1307 * Since this is common to all placement strategies, this lives here.
1309 * [ this allows ->select_task() to simply return task_cpu(p) and
1310 * not worry about this generic constraint ]
1312 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1314 cpu
= select_fallback_rq(task_cpu(p
), p
);
1319 static void update_avg(u64
*avg
, u64 sample
)
1321 s64 diff
= sample
- *avg
;
1327 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1329 #ifdef CONFIG_SCHEDSTATS
1330 struct rq
*rq
= this_rq();
1333 int this_cpu
= smp_processor_id();
1335 if (cpu
== this_cpu
) {
1336 schedstat_inc(rq
, ttwu_local
);
1337 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1339 struct sched_domain
*sd
;
1341 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1343 for_each_domain(this_cpu
, sd
) {
1344 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1345 schedstat_inc(sd
, ttwu_wake_remote
);
1352 if (wake_flags
& WF_MIGRATED
)
1353 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1355 #endif /* CONFIG_SMP */
1357 schedstat_inc(rq
, ttwu_count
);
1358 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1360 if (wake_flags
& WF_SYNC
)
1361 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1363 #endif /* CONFIG_SCHEDSTATS */
1366 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1368 activate_task(rq
, p
, en_flags
);
1371 /* if a worker is waking up, notify workqueue */
1372 if (p
->flags
& PF_WQ_WORKER
)
1373 wq_worker_waking_up(p
, cpu_of(rq
));
1377 * Mark the task runnable and perform wakeup-preemption.
1380 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1382 trace_sched_wakeup(p
, true);
1383 check_preempt_curr(rq
, p
, wake_flags
);
1385 p
->state
= TASK_RUNNING
;
1387 if (p
->sched_class
->task_woken
)
1388 p
->sched_class
->task_woken(rq
, p
);
1390 if (rq
->idle_stamp
) {
1391 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1392 u64 max
= 2*sysctl_sched_migration_cost
;
1397 update_avg(&rq
->avg_idle
, delta
);
1404 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1407 if (p
->sched_contributes_to_load
)
1408 rq
->nr_uninterruptible
--;
1411 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1412 ttwu_do_wakeup(rq
, p
, wake_flags
);
1416 * Called in case the task @p isn't fully descheduled from its runqueue,
1417 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1418 * since all we need to do is flip p->state to TASK_RUNNING, since
1419 * the task is still ->on_rq.
1421 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1426 rq
= __task_rq_lock(p
);
1428 ttwu_do_wakeup(rq
, p
, wake_flags
);
1431 __task_rq_unlock(rq
);
1437 static void sched_ttwu_pending(void)
1439 struct rq
*rq
= this_rq();
1440 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1441 struct task_struct
*p
;
1443 raw_spin_lock(&rq
->lock
);
1446 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1447 llist
= llist_next(llist
);
1448 ttwu_do_activate(rq
, p
, 0);
1451 raw_spin_unlock(&rq
->lock
);
1454 void scheduler_ipi(void)
1456 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1460 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1461 * traditionally all their work was done from the interrupt return
1462 * path. Now that we actually do some work, we need to make sure
1465 * Some archs already do call them, luckily irq_enter/exit nest
1468 * Arguably we should visit all archs and update all handlers,
1469 * however a fair share of IPIs are still resched only so this would
1470 * somewhat pessimize the simple resched case.
1473 sched_ttwu_pending();
1476 * Check if someone kicked us for doing the nohz idle load balance.
1478 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1479 this_rq()->idle_balance
= 1;
1480 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1485 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1487 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1488 smp_send_reschedule(cpu
);
1491 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1492 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1497 rq
= __task_rq_lock(p
);
1499 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1500 ttwu_do_wakeup(rq
, p
, wake_flags
);
1503 __task_rq_unlock(rq
);
1508 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1510 static inline int ttwu_share_cache(int this_cpu
, int that_cpu
)
1512 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1514 #endif /* CONFIG_SMP */
1516 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1518 struct rq
*rq
= cpu_rq(cpu
);
1520 #if defined(CONFIG_SMP)
1521 if (sched_feat(TTWU_QUEUE
) && !ttwu_share_cache(smp_processor_id(), cpu
)) {
1522 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1523 ttwu_queue_remote(p
, cpu
);
1528 raw_spin_lock(&rq
->lock
);
1529 ttwu_do_activate(rq
, p
, 0);
1530 raw_spin_unlock(&rq
->lock
);
1534 * try_to_wake_up - wake up a thread
1535 * @p: the thread to be awakened
1536 * @state: the mask of task states that can be woken
1537 * @wake_flags: wake modifier flags (WF_*)
1539 * Put it on the run-queue if it's not already there. The "current"
1540 * thread is always on the run-queue (except when the actual
1541 * re-schedule is in progress), and as such you're allowed to do
1542 * the simpler "current->state = TASK_RUNNING" to mark yourself
1543 * runnable without the overhead of this.
1545 * Returns %true if @p was woken up, %false if it was already running
1546 * or @state didn't match @p's state.
1549 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1551 unsigned long flags
;
1552 int cpu
, success
= 0;
1555 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1556 if (!(p
->state
& state
))
1559 success
= 1; /* we're going to change ->state */
1562 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1567 * If the owning (remote) cpu is still in the middle of schedule() with
1568 * this task as prev, wait until its done referencing the task.
1571 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1573 * In case the architecture enables interrupts in
1574 * context_switch(), we cannot busy wait, since that
1575 * would lead to deadlocks when an interrupt hits and
1576 * tries to wake up @prev. So bail and do a complete
1579 if (ttwu_activate_remote(p
, wake_flags
))
1586 * Pairs with the smp_wmb() in finish_lock_switch().
1590 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1591 p
->state
= TASK_WAKING
;
1593 if (p
->sched_class
->task_waking
)
1594 p
->sched_class
->task_waking(p
);
1596 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1597 if (task_cpu(p
) != cpu
) {
1598 wake_flags
|= WF_MIGRATED
;
1599 set_task_cpu(p
, cpu
);
1601 #endif /* CONFIG_SMP */
1605 ttwu_stat(p
, cpu
, wake_flags
);
1607 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1613 * try_to_wake_up_local - try to wake up a local task with rq lock held
1614 * @p: the thread to be awakened
1616 * Put @p on the run-queue if it's not already there. The caller must
1617 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1620 static void try_to_wake_up_local(struct task_struct
*p
)
1622 struct rq
*rq
= task_rq(p
);
1624 BUG_ON(rq
!= this_rq());
1625 BUG_ON(p
== current
);
1626 lockdep_assert_held(&rq
->lock
);
1628 if (!raw_spin_trylock(&p
->pi_lock
)) {
1629 raw_spin_unlock(&rq
->lock
);
1630 raw_spin_lock(&p
->pi_lock
);
1631 raw_spin_lock(&rq
->lock
);
1634 if (!(p
->state
& TASK_NORMAL
))
1638 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1640 ttwu_do_wakeup(rq
, p
, 0);
1641 ttwu_stat(p
, smp_processor_id(), 0);
1643 raw_spin_unlock(&p
->pi_lock
);
1647 * wake_up_process - Wake up a specific process
1648 * @p: The process to be woken up.
1650 * Attempt to wake up the nominated process and move it to the set of runnable
1651 * processes. Returns 1 if the process was woken up, 0 if it was already
1654 * It may be assumed that this function implies a write memory barrier before
1655 * changing the task state if and only if any tasks are woken up.
1657 int wake_up_process(struct task_struct
*p
)
1659 return try_to_wake_up(p
, TASK_ALL
, 0);
1661 EXPORT_SYMBOL(wake_up_process
);
1663 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1665 return try_to_wake_up(p
, state
, 0);
1669 * Perform scheduler related setup for a newly forked process p.
1670 * p is forked by current.
1672 * __sched_fork() is basic setup used by init_idle() too:
1674 static void __sched_fork(struct task_struct
*p
)
1679 p
->se
.exec_start
= 0;
1680 p
->se
.sum_exec_runtime
= 0;
1681 p
->se
.prev_sum_exec_runtime
= 0;
1682 p
->se
.nr_migrations
= 0;
1684 INIT_LIST_HEAD(&p
->se
.group_node
);
1686 #ifdef CONFIG_SCHEDSTATS
1687 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1690 INIT_LIST_HEAD(&p
->rt
.run_list
);
1692 #ifdef CONFIG_PREEMPT_NOTIFIERS
1693 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1698 * fork()/clone()-time setup:
1700 void sched_fork(struct task_struct
*p
)
1702 unsigned long flags
;
1703 int cpu
= get_cpu();
1707 * We mark the process as running here. This guarantees that
1708 * nobody will actually run it, and a signal or other external
1709 * event cannot wake it up and insert it on the runqueue either.
1711 p
->state
= TASK_RUNNING
;
1714 * Make sure we do not leak PI boosting priority to the child.
1716 p
->prio
= current
->normal_prio
;
1719 * Revert to default priority/policy on fork if requested.
1721 if (unlikely(p
->sched_reset_on_fork
)) {
1722 if (task_has_rt_policy(p
)) {
1723 p
->policy
= SCHED_NORMAL
;
1724 p
->static_prio
= NICE_TO_PRIO(0);
1726 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1727 p
->static_prio
= NICE_TO_PRIO(0);
1729 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1733 * We don't need the reset flag anymore after the fork. It has
1734 * fulfilled its duty:
1736 p
->sched_reset_on_fork
= 0;
1739 if (!rt_prio(p
->prio
))
1740 p
->sched_class
= &fair_sched_class
;
1742 if (p
->sched_class
->task_fork
)
1743 p
->sched_class
->task_fork(p
);
1746 * The child is not yet in the pid-hash so no cgroup attach races,
1747 * and the cgroup is pinned to this child due to cgroup_fork()
1748 * is ran before sched_fork().
1750 * Silence PROVE_RCU.
1752 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1753 set_task_cpu(p
, cpu
);
1754 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1756 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1757 if (likely(sched_info_on()))
1758 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1760 #if defined(CONFIG_SMP)
1763 #ifdef CONFIG_PREEMPT_COUNT
1764 /* Want to start with kernel preemption disabled. */
1765 task_thread_info(p
)->preempt_count
= 1;
1768 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1775 * wake_up_new_task - wake up a newly created task for the first time.
1777 * This function will do some initial scheduler statistics housekeeping
1778 * that must be done for every newly created context, then puts the task
1779 * on the runqueue and wakes it.
1781 void wake_up_new_task(struct task_struct
*p
)
1783 unsigned long flags
;
1786 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1789 * Fork balancing, do it here and not earlier because:
1790 * - cpus_allowed can change in the fork path
1791 * - any previously selected cpu might disappear through hotplug
1793 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1796 rq
= __task_rq_lock(p
);
1797 activate_task(rq
, p
, 0);
1799 trace_sched_wakeup_new(p
, true);
1800 check_preempt_curr(rq
, p
, WF_FORK
);
1802 if (p
->sched_class
->task_woken
)
1803 p
->sched_class
->task_woken(rq
, p
);
1805 task_rq_unlock(rq
, p
, &flags
);
1808 #ifdef CONFIG_PREEMPT_NOTIFIERS
1811 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1812 * @notifier: notifier struct to register
1814 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1816 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1818 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1821 * preempt_notifier_unregister - no longer interested in preemption notifications
1822 * @notifier: notifier struct to unregister
1824 * This is safe to call from within a preemption notifier.
1826 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1828 hlist_del(¬ifier
->link
);
1830 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1832 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1834 struct preempt_notifier
*notifier
;
1835 struct hlist_node
*node
;
1837 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1838 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1842 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1843 struct task_struct
*next
)
1845 struct preempt_notifier
*notifier
;
1846 struct hlist_node
*node
;
1848 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1849 notifier
->ops
->sched_out(notifier
, next
);
1852 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1854 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1859 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1860 struct task_struct
*next
)
1864 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1867 * prepare_task_switch - prepare to switch tasks
1868 * @rq: the runqueue preparing to switch
1869 * @prev: the current task that is being switched out
1870 * @next: the task we are going to switch to.
1872 * This is called with the rq lock held and interrupts off. It must
1873 * be paired with a subsequent finish_task_switch after the context
1876 * prepare_task_switch sets up locking and calls architecture specific
1880 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1881 struct task_struct
*next
)
1883 sched_info_switch(prev
, next
);
1884 perf_event_task_sched_out(prev
, next
);
1885 fire_sched_out_preempt_notifiers(prev
, next
);
1886 prepare_lock_switch(rq
, next
);
1887 prepare_arch_switch(next
);
1888 trace_sched_switch(prev
, next
);
1892 * finish_task_switch - clean up after a task-switch
1893 * @rq: runqueue associated with task-switch
1894 * @prev: the thread we just switched away from.
1896 * finish_task_switch must be called after the context switch, paired
1897 * with a prepare_task_switch call before the context switch.
1898 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1899 * and do any other architecture-specific cleanup actions.
1901 * Note that we may have delayed dropping an mm in context_switch(). If
1902 * so, we finish that here outside of the runqueue lock. (Doing it
1903 * with the lock held can cause deadlocks; see schedule() for
1906 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1907 __releases(rq
->lock
)
1909 struct mm_struct
*mm
= rq
->prev_mm
;
1915 * A task struct has one reference for the use as "current".
1916 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1917 * schedule one last time. The schedule call will never return, and
1918 * the scheduled task must drop that reference.
1919 * The test for TASK_DEAD must occur while the runqueue locks are
1920 * still held, otherwise prev could be scheduled on another cpu, die
1921 * there before we look at prev->state, and then the reference would
1923 * Manfred Spraul <manfred@colorfullife.com>
1925 prev_state
= prev
->state
;
1926 finish_arch_switch(prev
);
1927 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1928 local_irq_disable();
1929 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1930 perf_event_task_sched_in(prev
, current
);
1931 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1933 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1934 finish_lock_switch(rq
, prev
);
1936 fire_sched_in_preempt_notifiers(current
);
1939 if (unlikely(prev_state
== TASK_DEAD
)) {
1941 * Remove function-return probe instances associated with this
1942 * task and put them back on the free list.
1944 kprobe_flush_task(prev
);
1945 put_task_struct(prev
);
1951 /* assumes rq->lock is held */
1952 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1954 if (prev
->sched_class
->pre_schedule
)
1955 prev
->sched_class
->pre_schedule(rq
, prev
);
1958 /* rq->lock is NOT held, but preemption is disabled */
1959 static inline void post_schedule(struct rq
*rq
)
1961 if (rq
->post_schedule
) {
1962 unsigned long flags
;
1964 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1965 if (rq
->curr
->sched_class
->post_schedule
)
1966 rq
->curr
->sched_class
->post_schedule(rq
);
1967 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1969 rq
->post_schedule
= 0;
1975 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1979 static inline void post_schedule(struct rq
*rq
)
1986 * schedule_tail - first thing a freshly forked thread must call.
1987 * @prev: the thread we just switched away from.
1989 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1990 __releases(rq
->lock
)
1992 struct rq
*rq
= this_rq();
1994 finish_task_switch(rq
, prev
);
1997 * FIXME: do we need to worry about rq being invalidated by the
2002 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2003 /* In this case, finish_task_switch does not reenable preemption */
2006 if (current
->set_child_tid
)
2007 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2011 * context_switch - switch to the new MM and the new
2012 * thread's register state.
2015 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2016 struct task_struct
*next
)
2018 struct mm_struct
*mm
, *oldmm
;
2020 prepare_task_switch(rq
, prev
, next
);
2023 oldmm
= prev
->active_mm
;
2025 * For paravirt, this is coupled with an exit in switch_to to
2026 * combine the page table reload and the switch backend into
2029 arch_start_context_switch(prev
);
2032 next
->active_mm
= oldmm
;
2033 atomic_inc(&oldmm
->mm_count
);
2034 enter_lazy_tlb(oldmm
, next
);
2036 switch_mm(oldmm
, mm
, next
);
2039 prev
->active_mm
= NULL
;
2040 rq
->prev_mm
= oldmm
;
2043 * Since the runqueue lock will be released by the next
2044 * task (which is an invalid locking op but in the case
2045 * of the scheduler it's an obvious special-case), so we
2046 * do an early lockdep release here:
2048 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2049 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2052 /* Here we just switch the register state and the stack. */
2053 switch_to(prev
, next
, prev
);
2057 * this_rq must be evaluated again because prev may have moved
2058 * CPUs since it called schedule(), thus the 'rq' on its stack
2059 * frame will be invalid.
2061 finish_task_switch(this_rq(), prev
);
2065 * nr_running, nr_uninterruptible and nr_context_switches:
2067 * externally visible scheduler statistics: current number of runnable
2068 * threads, current number of uninterruptible-sleeping threads, total
2069 * number of context switches performed since bootup.
2071 unsigned long nr_running(void)
2073 unsigned long i
, sum
= 0;
2075 for_each_online_cpu(i
)
2076 sum
+= cpu_rq(i
)->nr_running
;
2081 unsigned long nr_uninterruptible(void)
2083 unsigned long i
, sum
= 0;
2085 for_each_possible_cpu(i
)
2086 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2089 * Since we read the counters lockless, it might be slightly
2090 * inaccurate. Do not allow it to go below zero though:
2092 if (unlikely((long)sum
< 0))
2098 unsigned long long nr_context_switches(void)
2101 unsigned long long sum
= 0;
2103 for_each_possible_cpu(i
)
2104 sum
+= cpu_rq(i
)->nr_switches
;
2109 unsigned long nr_iowait(void)
2111 unsigned long i
, sum
= 0;
2113 for_each_possible_cpu(i
)
2114 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2119 unsigned long nr_iowait_cpu(int cpu
)
2121 struct rq
*this = cpu_rq(cpu
);
2122 return atomic_read(&this->nr_iowait
);
2125 unsigned long this_cpu_load(void)
2127 struct rq
*this = this_rq();
2128 return this->cpu_load
[0];
2132 /* Variables and functions for calc_load */
2133 static atomic_long_t calc_load_tasks
;
2134 static unsigned long calc_load_update
;
2135 unsigned long avenrun
[3];
2136 EXPORT_SYMBOL(avenrun
);
2138 static long calc_load_fold_active(struct rq
*this_rq
)
2140 long nr_active
, delta
= 0;
2142 nr_active
= this_rq
->nr_running
;
2143 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2145 if (nr_active
!= this_rq
->calc_load_active
) {
2146 delta
= nr_active
- this_rq
->calc_load_active
;
2147 this_rq
->calc_load_active
= nr_active
;
2153 static unsigned long
2154 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2157 load
+= active
* (FIXED_1
- exp
);
2158 load
+= 1UL << (FSHIFT
- 1);
2159 return load
>> FSHIFT
;
2164 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2166 * When making the ILB scale, we should try to pull this in as well.
2168 static atomic_long_t calc_load_tasks_idle
;
2170 void calc_load_account_idle(struct rq
*this_rq
)
2174 delta
= calc_load_fold_active(this_rq
);
2176 atomic_long_add(delta
, &calc_load_tasks_idle
);
2179 static long calc_load_fold_idle(void)
2184 * Its got a race, we don't care...
2186 if (atomic_long_read(&calc_load_tasks_idle
))
2187 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2193 * fixed_power_int - compute: x^n, in O(log n) time
2195 * @x: base of the power
2196 * @frac_bits: fractional bits of @x
2197 * @n: power to raise @x to.
2199 * By exploiting the relation between the definition of the natural power
2200 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2201 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2202 * (where: n_i \elem {0, 1}, the binary vector representing n),
2203 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2204 * of course trivially computable in O(log_2 n), the length of our binary
2207 static unsigned long
2208 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2210 unsigned long result
= 1UL << frac_bits
;
2215 result
+= 1UL << (frac_bits
- 1);
2216 result
>>= frac_bits
;
2222 x
+= 1UL << (frac_bits
- 1);
2230 * a1 = a0 * e + a * (1 - e)
2232 * a2 = a1 * e + a * (1 - e)
2233 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2234 * = a0 * e^2 + a * (1 - e) * (1 + e)
2236 * a3 = a2 * e + a * (1 - e)
2237 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2238 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2242 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2243 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2244 * = a0 * e^n + a * (1 - e^n)
2246 * [1] application of the geometric series:
2249 * S_n := \Sum x^i = -------------
2252 static unsigned long
2253 calc_load_n(unsigned long load
, unsigned long exp
,
2254 unsigned long active
, unsigned int n
)
2257 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2261 * NO_HZ can leave us missing all per-cpu ticks calling
2262 * calc_load_account_active(), but since an idle CPU folds its delta into
2263 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2264 * in the pending idle delta if our idle period crossed a load cycle boundary.
2266 * Once we've updated the global active value, we need to apply the exponential
2267 * weights adjusted to the number of cycles missed.
2269 static void calc_global_nohz(unsigned long ticks
)
2271 long delta
, active
, n
;
2273 if (time_before(jiffies
, calc_load_update
))
2277 * If we crossed a calc_load_update boundary, make sure to fold
2278 * any pending idle changes, the respective CPUs might have
2279 * missed the tick driven calc_load_account_active() update
2282 delta
= calc_load_fold_idle();
2284 atomic_long_add(delta
, &calc_load_tasks
);
2287 * If we were idle for multiple load cycles, apply them.
2289 if (ticks
>= LOAD_FREQ
) {
2290 n
= ticks
/ LOAD_FREQ
;
2292 active
= atomic_long_read(&calc_load_tasks
);
2293 active
= active
> 0 ? active
* FIXED_1
: 0;
2295 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2296 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2297 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2299 calc_load_update
+= n
* LOAD_FREQ
;
2303 * Its possible the remainder of the above division also crosses
2304 * a LOAD_FREQ period, the regular check in calc_global_load()
2305 * which comes after this will take care of that.
2307 * Consider us being 11 ticks before a cycle completion, and us
2308 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2309 * age us 4 cycles, and the test in calc_global_load() will
2310 * pick up the final one.
2314 void calc_load_account_idle(struct rq
*this_rq
)
2318 static inline long calc_load_fold_idle(void)
2323 static void calc_global_nohz(unsigned long ticks
)
2329 * get_avenrun - get the load average array
2330 * @loads: pointer to dest load array
2331 * @offset: offset to add
2332 * @shift: shift count to shift the result left
2334 * These values are estimates at best, so no need for locking.
2336 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2338 loads
[0] = (avenrun
[0] + offset
) << shift
;
2339 loads
[1] = (avenrun
[1] + offset
) << shift
;
2340 loads
[2] = (avenrun
[2] + offset
) << shift
;
2344 * calc_load - update the avenrun load estimates 10 ticks after the
2345 * CPUs have updated calc_load_tasks.
2347 void calc_global_load(unsigned long ticks
)
2351 calc_global_nohz(ticks
);
2353 if (time_before(jiffies
, calc_load_update
+ 10))
2356 active
= atomic_long_read(&calc_load_tasks
);
2357 active
= active
> 0 ? active
* FIXED_1
: 0;
2359 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2360 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2361 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2363 calc_load_update
+= LOAD_FREQ
;
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_branch(¶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 preempt_enable_no_resched();
3228 static inline void sched_submit_work(struct task_struct
*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
);
3249 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3251 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3253 if (lock
->owner
!= owner
)
3257 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3258 * lock->owner still matches owner, if that fails, owner might
3259 * point to free()d memory, if it still matches, the rcu_read_lock()
3260 * ensures the memory stays valid.
3264 return owner
->on_cpu
;
3268 * Look out! "owner" is an entirely speculative pointer
3269 * access and not reliable.
3271 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3273 if (!sched_feat(OWNER_SPIN
))
3277 while (owner_running(lock
, owner
)) {
3281 arch_mutex_cpu_relax();
3286 * We break out the loop above on need_resched() and when the
3287 * owner changed, which is a sign for heavy contention. Return
3288 * success only when lock->owner is NULL.
3290 return lock
->owner
== NULL
;
3294 #ifdef CONFIG_PREEMPT
3296 * this is the entry point to schedule() from in-kernel preemption
3297 * off of preempt_enable. Kernel preemptions off return from interrupt
3298 * occur there and call schedule directly.
3300 asmlinkage
void __sched notrace
preempt_schedule(void)
3302 struct thread_info
*ti
= current_thread_info();
3305 * If there is a non-zero preempt_count or interrupts are disabled,
3306 * we do not want to preempt the current task. Just return..
3308 if (likely(ti
->preempt_count
|| irqs_disabled()))
3312 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3314 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3317 * Check again in case we missed a preemption opportunity
3318 * between schedule and now.
3321 } while (need_resched());
3323 EXPORT_SYMBOL(preempt_schedule
);
3326 * this is the entry point to schedule() from kernel preemption
3327 * off of irq context.
3328 * Note, that this is called and return with irqs disabled. This will
3329 * protect us against recursive calling from irq.
3331 asmlinkage
void __sched
preempt_schedule_irq(void)
3333 struct thread_info
*ti
= current_thread_info();
3335 /* Catch callers which need to be fixed */
3336 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3339 add_preempt_count(PREEMPT_ACTIVE
);
3342 local_irq_disable();
3343 sub_preempt_count(PREEMPT_ACTIVE
);
3346 * Check again in case we missed a preemption opportunity
3347 * between schedule and now.
3350 } while (need_resched());
3353 #endif /* CONFIG_PREEMPT */
3355 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3358 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3360 EXPORT_SYMBOL(default_wake_function
);
3363 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3364 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3365 * number) then we wake all the non-exclusive tasks and one exclusive task.
3367 * There are circumstances in which we can try to wake a task which has already
3368 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3369 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3371 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3372 int nr_exclusive
, int wake_flags
, void *key
)
3374 wait_queue_t
*curr
, *next
;
3376 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3377 unsigned flags
= curr
->flags
;
3379 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3380 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3386 * __wake_up - wake up threads blocked on a waitqueue.
3388 * @mode: which threads
3389 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3390 * @key: is directly passed to the wakeup function
3392 * It may be assumed that this function implies a write memory barrier before
3393 * changing the task state if and only if any tasks are woken up.
3395 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3396 int nr_exclusive
, void *key
)
3398 unsigned long flags
;
3400 spin_lock_irqsave(&q
->lock
, flags
);
3401 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3402 spin_unlock_irqrestore(&q
->lock
, flags
);
3404 EXPORT_SYMBOL(__wake_up
);
3407 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3409 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3411 __wake_up_common(q
, mode
, 1, 0, NULL
);
3413 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3415 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3417 __wake_up_common(q
, mode
, 1, 0, key
);
3419 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3422 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3424 * @mode: which threads
3425 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3426 * @key: opaque value to be passed to wakeup targets
3428 * The sync wakeup differs that the waker knows that it will schedule
3429 * away soon, so while the target thread will be woken up, it will not
3430 * be migrated to another CPU - ie. the two threads are 'synchronized'
3431 * with each other. This can prevent needless bouncing between CPUs.
3433 * On UP it can prevent extra preemption.
3435 * It may be assumed that this function implies a write memory barrier before
3436 * changing the task state if and only if any tasks are woken up.
3438 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3439 int nr_exclusive
, void *key
)
3441 unsigned long flags
;
3442 int wake_flags
= WF_SYNC
;
3447 if (unlikely(!nr_exclusive
))
3450 spin_lock_irqsave(&q
->lock
, flags
);
3451 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3452 spin_unlock_irqrestore(&q
->lock
, flags
);
3454 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3457 * __wake_up_sync - see __wake_up_sync_key()
3459 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3461 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3463 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3466 * complete: - signals a single thread waiting on this completion
3467 * @x: holds the state of this particular completion
3469 * This will wake up a single thread waiting on this completion. Threads will be
3470 * awakened in the same order in which they were queued.
3472 * See also complete_all(), wait_for_completion() and related routines.
3474 * It may be assumed that this function implies a write memory barrier before
3475 * changing the task state if and only if any tasks are woken up.
3477 void complete(struct completion
*x
)
3479 unsigned long flags
;
3481 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3483 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3484 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3486 EXPORT_SYMBOL(complete
);
3489 * complete_all: - signals all threads waiting on this completion
3490 * @x: holds the state of this particular completion
3492 * This will wake up all threads waiting on this particular completion event.
3494 * It may be assumed that this function implies a write memory barrier before
3495 * changing the task state if and only if any tasks are woken up.
3497 void complete_all(struct completion
*x
)
3499 unsigned long flags
;
3501 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3502 x
->done
+= UINT_MAX
/2;
3503 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3504 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3506 EXPORT_SYMBOL(complete_all
);
3508 static inline long __sched
3509 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3512 DECLARE_WAITQUEUE(wait
, current
);
3514 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3516 if (signal_pending_state(state
, current
)) {
3517 timeout
= -ERESTARTSYS
;
3520 __set_current_state(state
);
3521 spin_unlock_irq(&x
->wait
.lock
);
3522 timeout
= schedule_timeout(timeout
);
3523 spin_lock_irq(&x
->wait
.lock
);
3524 } while (!x
->done
&& timeout
);
3525 __remove_wait_queue(&x
->wait
, &wait
);
3530 return timeout
?: 1;
3534 wait_for_common(struct completion
*x
, long timeout
, int state
)
3538 spin_lock_irq(&x
->wait
.lock
);
3539 timeout
= do_wait_for_common(x
, timeout
, state
);
3540 spin_unlock_irq(&x
->wait
.lock
);
3545 * wait_for_completion: - waits for completion of a task
3546 * @x: holds the state of this particular completion
3548 * This waits to be signaled for completion of a specific task. It is NOT
3549 * interruptible and there is no timeout.
3551 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3552 * and interrupt capability. Also see complete().
3554 void __sched
wait_for_completion(struct completion
*x
)
3556 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3558 EXPORT_SYMBOL(wait_for_completion
);
3561 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3562 * @x: holds the state of this particular completion
3563 * @timeout: timeout value in jiffies
3565 * This waits for either a completion of a specific task to be signaled or for a
3566 * specified timeout to expire. The timeout is in jiffies. It is not
3569 * The return value is 0 if timed out, and positive (at least 1, or number of
3570 * jiffies left till timeout) if completed.
3572 unsigned long __sched
3573 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3575 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3577 EXPORT_SYMBOL(wait_for_completion_timeout
);
3580 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3581 * @x: holds the state of this particular completion
3583 * This waits for completion of a specific task to be signaled. It is
3586 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3588 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3590 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3591 if (t
== -ERESTARTSYS
)
3595 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3598 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3599 * @x: holds the state of this particular completion
3600 * @timeout: timeout value in jiffies
3602 * This waits for either a completion of a specific task to be signaled or for a
3603 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3605 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3606 * positive (at least 1, or number of jiffies left till timeout) if completed.
3609 wait_for_completion_interruptible_timeout(struct completion
*x
,
3610 unsigned long timeout
)
3612 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3614 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3617 * wait_for_completion_killable: - waits for completion of a task (killable)
3618 * @x: holds the state of this particular completion
3620 * This waits to be signaled for completion of a specific task. It can be
3621 * interrupted by a kill signal.
3623 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3625 int __sched
wait_for_completion_killable(struct completion
*x
)
3627 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3628 if (t
== -ERESTARTSYS
)
3632 EXPORT_SYMBOL(wait_for_completion_killable
);
3635 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3636 * @x: holds the state of this particular completion
3637 * @timeout: timeout value in jiffies
3639 * This waits for either a completion of a specific task to be
3640 * signaled or for a specified timeout to expire. It can be
3641 * interrupted by a kill signal. The timeout is in jiffies.
3643 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3644 * positive (at least 1, or number of jiffies left till timeout) if completed.
3647 wait_for_completion_killable_timeout(struct completion
*x
,
3648 unsigned long timeout
)
3650 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3652 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3655 * try_wait_for_completion - try to decrement a completion without blocking
3656 * @x: completion structure
3658 * Returns: 0 if a decrement cannot be done without blocking
3659 * 1 if a decrement succeeded.
3661 * If a completion is being used as a counting completion,
3662 * attempt to decrement the counter without blocking. This
3663 * enables us to avoid waiting if the resource the completion
3664 * is protecting is not available.
3666 bool try_wait_for_completion(struct completion
*x
)
3668 unsigned long flags
;
3671 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3676 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3679 EXPORT_SYMBOL(try_wait_for_completion
);
3682 * completion_done - Test to see if a completion has any waiters
3683 * @x: completion structure
3685 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3686 * 1 if there are no waiters.
3689 bool completion_done(struct completion
*x
)
3691 unsigned long flags
;
3694 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3697 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3700 EXPORT_SYMBOL(completion_done
);
3703 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3705 unsigned long flags
;
3708 init_waitqueue_entry(&wait
, current
);
3710 __set_current_state(state
);
3712 spin_lock_irqsave(&q
->lock
, flags
);
3713 __add_wait_queue(q
, &wait
);
3714 spin_unlock(&q
->lock
);
3715 timeout
= schedule_timeout(timeout
);
3716 spin_lock_irq(&q
->lock
);
3717 __remove_wait_queue(q
, &wait
);
3718 spin_unlock_irqrestore(&q
->lock
, flags
);
3723 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3725 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3727 EXPORT_SYMBOL(interruptible_sleep_on
);
3730 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3732 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3734 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3736 void __sched
sleep_on(wait_queue_head_t
*q
)
3738 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3740 EXPORT_SYMBOL(sleep_on
);
3742 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3744 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3746 EXPORT_SYMBOL(sleep_on_timeout
);
3748 #ifdef CONFIG_RT_MUTEXES
3751 * rt_mutex_setprio - set the current priority of a task
3753 * @prio: prio value (kernel-internal form)
3755 * This function changes the 'effective' priority of a task. It does
3756 * not touch ->normal_prio like __setscheduler().
3758 * Used by the rt_mutex code to implement priority inheritance logic.
3760 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3762 int oldprio
, on_rq
, running
;
3764 const struct sched_class
*prev_class
;
3766 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3768 rq
= __task_rq_lock(p
);
3770 trace_sched_pi_setprio(p
, prio
);
3772 prev_class
= p
->sched_class
;
3774 running
= task_current(rq
, p
);
3776 dequeue_task(rq
, p
, 0);
3778 p
->sched_class
->put_prev_task(rq
, p
);
3781 p
->sched_class
= &rt_sched_class
;
3783 p
->sched_class
= &fair_sched_class
;
3788 p
->sched_class
->set_curr_task(rq
);
3790 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3792 check_class_changed(rq
, p
, prev_class
, oldprio
);
3793 __task_rq_unlock(rq
);
3798 void set_user_nice(struct task_struct
*p
, long nice
)
3800 int old_prio
, delta
, on_rq
;
3801 unsigned long flags
;
3804 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3807 * We have to be careful, if called from sys_setpriority(),
3808 * the task might be in the middle of scheduling on another CPU.
3810 rq
= task_rq_lock(p
, &flags
);
3812 * The RT priorities are set via sched_setscheduler(), but we still
3813 * allow the 'normal' nice value to be set - but as expected
3814 * it wont have any effect on scheduling until the task is
3815 * SCHED_FIFO/SCHED_RR:
3817 if (task_has_rt_policy(p
)) {
3818 p
->static_prio
= NICE_TO_PRIO(nice
);
3823 dequeue_task(rq
, p
, 0);
3825 p
->static_prio
= NICE_TO_PRIO(nice
);
3828 p
->prio
= effective_prio(p
);
3829 delta
= p
->prio
- old_prio
;
3832 enqueue_task(rq
, p
, 0);
3834 * If the task increased its priority or is running and
3835 * lowered its priority, then reschedule its CPU:
3837 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3838 resched_task(rq
->curr
);
3841 task_rq_unlock(rq
, p
, &flags
);
3843 EXPORT_SYMBOL(set_user_nice
);
3846 * can_nice - check if a task can reduce its nice value
3850 int can_nice(const struct task_struct
*p
, const int nice
)
3852 /* convert nice value [19,-20] to rlimit style value [1,40] */
3853 int nice_rlim
= 20 - nice
;
3855 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3856 capable(CAP_SYS_NICE
));
3859 #ifdef __ARCH_WANT_SYS_NICE
3862 * sys_nice - change the priority of the current process.
3863 * @increment: priority increment
3865 * sys_setpriority is a more generic, but much slower function that
3866 * does similar things.
3868 SYSCALL_DEFINE1(nice
, int, increment
)
3873 * Setpriority might change our priority at the same moment.
3874 * We don't have to worry. Conceptually one call occurs first
3875 * and we have a single winner.
3877 if (increment
< -40)
3882 nice
= TASK_NICE(current
) + increment
;
3888 if (increment
< 0 && !can_nice(current
, nice
))
3891 retval
= security_task_setnice(current
, nice
);
3895 set_user_nice(current
, nice
);
3902 * task_prio - return the priority value of a given task.
3903 * @p: the task in question.
3905 * This is the priority value as seen by users in /proc.
3906 * RT tasks are offset by -200. Normal tasks are centered
3907 * around 0, value goes from -16 to +15.
3909 int task_prio(const struct task_struct
*p
)
3911 return p
->prio
- MAX_RT_PRIO
;
3915 * task_nice - return the nice value of a given task.
3916 * @p: the task in question.
3918 int task_nice(const struct task_struct
*p
)
3920 return TASK_NICE(p
);
3922 EXPORT_SYMBOL(task_nice
);
3925 * idle_cpu - is a given cpu idle currently?
3926 * @cpu: the processor in question.
3928 int idle_cpu(int cpu
)
3930 struct rq
*rq
= cpu_rq(cpu
);
3932 if (rq
->curr
!= rq
->idle
)
3939 if (!llist_empty(&rq
->wake_list
))
3947 * idle_task - return the idle task for a given cpu.
3948 * @cpu: the processor in question.
3950 struct task_struct
*idle_task(int cpu
)
3952 return cpu_rq(cpu
)->idle
;
3956 * find_process_by_pid - find a process with a matching PID value.
3957 * @pid: the pid in question.
3959 static struct task_struct
*find_process_by_pid(pid_t pid
)
3961 return pid
? find_task_by_vpid(pid
) : current
;
3964 /* Actually do priority change: must hold rq lock. */
3966 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3969 p
->rt_priority
= prio
;
3970 p
->normal_prio
= normal_prio(p
);
3971 /* we are holding p->pi_lock already */
3972 p
->prio
= rt_mutex_getprio(p
);
3973 if (rt_prio(p
->prio
))
3974 p
->sched_class
= &rt_sched_class
;
3976 p
->sched_class
= &fair_sched_class
;
3981 * check the target process has a UID that matches the current process's
3983 static bool check_same_owner(struct task_struct
*p
)
3985 const struct cred
*cred
= current_cred(), *pcred
;
3989 pcred
= __task_cred(p
);
3990 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
3991 match
= (cred
->euid
== pcred
->euid
||
3992 cred
->euid
== pcred
->uid
);
3999 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4000 const struct sched_param
*param
, bool user
)
4002 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4003 unsigned long flags
;
4004 const struct sched_class
*prev_class
;
4008 /* may grab non-irq protected spin_locks */
4009 BUG_ON(in_interrupt());
4011 /* double check policy once rq lock held */
4013 reset_on_fork
= p
->sched_reset_on_fork
;
4014 policy
= oldpolicy
= p
->policy
;
4016 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4017 policy
&= ~SCHED_RESET_ON_FORK
;
4019 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4020 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4021 policy
!= SCHED_IDLE
)
4026 * Valid priorities for SCHED_FIFO and SCHED_RR are
4027 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4028 * SCHED_BATCH and SCHED_IDLE is 0.
4030 if (param
->sched_priority
< 0 ||
4031 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4032 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4034 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4038 * Allow unprivileged RT tasks to decrease priority:
4040 if (user
&& !capable(CAP_SYS_NICE
)) {
4041 if (rt_policy(policy
)) {
4042 unsigned long rlim_rtprio
=
4043 task_rlimit(p
, RLIMIT_RTPRIO
);
4045 /* can't set/change the rt policy */
4046 if (policy
!= p
->policy
&& !rlim_rtprio
)
4049 /* can't increase priority */
4050 if (param
->sched_priority
> p
->rt_priority
&&
4051 param
->sched_priority
> rlim_rtprio
)
4056 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4057 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4059 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
4060 if (!can_nice(p
, TASK_NICE(p
)))
4064 /* can't change other user's priorities */
4065 if (!check_same_owner(p
))
4068 /* Normal users shall not reset the sched_reset_on_fork flag */
4069 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4074 retval
= security_task_setscheduler(p
);
4080 * make sure no PI-waiters arrive (or leave) while we are
4081 * changing the priority of the task:
4083 * To be able to change p->policy safely, the appropriate
4084 * runqueue lock must be held.
4086 rq
= task_rq_lock(p
, &flags
);
4089 * Changing the policy of the stop threads its a very bad idea
4091 if (p
== rq
->stop
) {
4092 task_rq_unlock(rq
, p
, &flags
);
4097 * If not changing anything there's no need to proceed further:
4099 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
4100 param
->sched_priority
== p
->rt_priority
))) {
4102 __task_rq_unlock(rq
);
4103 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4107 #ifdef CONFIG_RT_GROUP_SCHED
4110 * Do not allow realtime tasks into groups that have no runtime
4113 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4114 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4115 !task_group_is_autogroup(task_group(p
))) {
4116 task_rq_unlock(rq
, p
, &flags
);
4122 /* recheck policy now with rq lock held */
4123 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4124 policy
= oldpolicy
= -1;
4125 task_rq_unlock(rq
, p
, &flags
);
4129 running
= task_current(rq
, p
);
4131 dequeue_task(rq
, p
, 0);
4133 p
->sched_class
->put_prev_task(rq
, p
);
4135 p
->sched_reset_on_fork
= reset_on_fork
;
4138 prev_class
= p
->sched_class
;
4139 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4142 p
->sched_class
->set_curr_task(rq
);
4144 enqueue_task(rq
, p
, 0);
4146 check_class_changed(rq
, p
, prev_class
, oldprio
);
4147 task_rq_unlock(rq
, p
, &flags
);
4149 rt_mutex_adjust_pi(p
);
4155 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4156 * @p: the task in question.
4157 * @policy: new policy.
4158 * @param: structure containing the new RT priority.
4160 * NOTE that the task may be already dead.
4162 int sched_setscheduler(struct task_struct
*p
, int policy
,
4163 const struct sched_param
*param
)
4165 return __sched_setscheduler(p
, policy
, param
, true);
4167 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4170 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4171 * @p: the task in question.
4172 * @policy: new policy.
4173 * @param: structure containing the new RT priority.
4175 * Just like sched_setscheduler, only don't bother checking if the
4176 * current context has permission. For example, this is needed in
4177 * stop_machine(): we create temporary high priority worker threads,
4178 * but our caller might not have that capability.
4180 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4181 const struct sched_param
*param
)
4183 return __sched_setscheduler(p
, policy
, param
, false);
4187 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4189 struct sched_param lparam
;
4190 struct task_struct
*p
;
4193 if (!param
|| pid
< 0)
4195 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4200 p
= find_process_by_pid(pid
);
4202 retval
= sched_setscheduler(p
, policy
, &lparam
);
4209 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4210 * @pid: the pid in question.
4211 * @policy: new policy.
4212 * @param: structure containing the new RT priority.
4214 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4215 struct sched_param __user
*, param
)
4217 /* negative values for policy are not valid */
4221 return do_sched_setscheduler(pid
, policy
, param
);
4225 * sys_sched_setparam - set/change the RT priority of a thread
4226 * @pid: the pid in question.
4227 * @param: structure containing the new RT priority.
4229 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4231 return do_sched_setscheduler(pid
, -1, param
);
4235 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4236 * @pid: the pid in question.
4238 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4240 struct task_struct
*p
;
4248 p
= find_process_by_pid(pid
);
4250 retval
= security_task_getscheduler(p
);
4253 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4260 * sys_sched_getparam - get the RT priority of a thread
4261 * @pid: the pid in question.
4262 * @param: structure containing the RT priority.
4264 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4266 struct sched_param lp
;
4267 struct task_struct
*p
;
4270 if (!param
|| pid
< 0)
4274 p
= find_process_by_pid(pid
);
4279 retval
= security_task_getscheduler(p
);
4283 lp
.sched_priority
= p
->rt_priority
;
4287 * This one might sleep, we cannot do it with a spinlock held ...
4289 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4298 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4300 cpumask_var_t cpus_allowed
, new_mask
;
4301 struct task_struct
*p
;
4307 p
= find_process_by_pid(pid
);
4314 /* Prevent p going away */
4318 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4322 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4324 goto out_free_cpus_allowed
;
4327 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4330 retval
= security_task_setscheduler(p
);
4334 cpuset_cpus_allowed(p
, cpus_allowed
);
4335 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4337 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4340 cpuset_cpus_allowed(p
, cpus_allowed
);
4341 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4343 * We must have raced with a concurrent cpuset
4344 * update. Just reset the cpus_allowed to the
4345 * cpuset's cpus_allowed
4347 cpumask_copy(new_mask
, cpus_allowed
);
4352 free_cpumask_var(new_mask
);
4353 out_free_cpus_allowed
:
4354 free_cpumask_var(cpus_allowed
);
4361 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4362 struct cpumask
*new_mask
)
4364 if (len
< cpumask_size())
4365 cpumask_clear(new_mask
);
4366 else if (len
> cpumask_size())
4367 len
= cpumask_size();
4369 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4373 * sys_sched_setaffinity - set the cpu affinity of a process
4374 * @pid: pid of the process
4375 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4376 * @user_mask_ptr: user-space pointer to the new cpu mask
4378 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4379 unsigned long __user
*, user_mask_ptr
)
4381 cpumask_var_t new_mask
;
4384 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4387 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4389 retval
= sched_setaffinity(pid
, new_mask
);
4390 free_cpumask_var(new_mask
);
4394 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4396 struct task_struct
*p
;
4397 unsigned long flags
;
4404 p
= find_process_by_pid(pid
);
4408 retval
= security_task_getscheduler(p
);
4412 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4413 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4414 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4424 * sys_sched_getaffinity - get the cpu affinity of a process
4425 * @pid: pid of the process
4426 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4427 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4429 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4430 unsigned long __user
*, user_mask_ptr
)
4435 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4437 if (len
& (sizeof(unsigned long)-1))
4440 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4443 ret
= sched_getaffinity(pid
, mask
);
4445 size_t retlen
= min_t(size_t, len
, cpumask_size());
4447 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4452 free_cpumask_var(mask
);
4458 * sys_sched_yield - yield the current processor to other threads.
4460 * This function yields the current CPU to other tasks. If there are no
4461 * other threads running on this CPU then this function will return.
4463 SYSCALL_DEFINE0(sched_yield
)
4465 struct rq
*rq
= this_rq_lock();
4467 schedstat_inc(rq
, yld_count
);
4468 current
->sched_class
->yield_task(rq
);
4471 * Since we are going to call schedule() anyway, there's
4472 * no need to preempt or enable interrupts:
4474 __release(rq
->lock
);
4475 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4476 do_raw_spin_unlock(&rq
->lock
);
4477 preempt_enable_no_resched();
4484 static inline int should_resched(void)
4486 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4489 static void __cond_resched(void)
4491 add_preempt_count(PREEMPT_ACTIVE
);
4493 sub_preempt_count(PREEMPT_ACTIVE
);
4496 int __sched
_cond_resched(void)
4498 if (should_resched()) {
4504 EXPORT_SYMBOL(_cond_resched
);
4507 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4508 * call schedule, and on return reacquire the lock.
4510 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4511 * operations here to prevent schedule() from being called twice (once via
4512 * spin_unlock(), once by hand).
4514 int __cond_resched_lock(spinlock_t
*lock
)
4516 int resched
= should_resched();
4519 lockdep_assert_held(lock
);
4521 if (spin_needbreak(lock
) || resched
) {
4532 EXPORT_SYMBOL(__cond_resched_lock
);
4534 int __sched
__cond_resched_softirq(void)
4536 BUG_ON(!in_softirq());
4538 if (should_resched()) {
4546 EXPORT_SYMBOL(__cond_resched_softirq
);
4549 * yield - yield the current processor to other threads.
4551 * This is a shortcut for kernel-space yielding - it marks the
4552 * thread runnable and calls sys_sched_yield().
4554 void __sched
yield(void)
4556 set_current_state(TASK_RUNNING
);
4559 EXPORT_SYMBOL(yield
);
4562 * yield_to - yield the current processor to another thread in
4563 * your thread group, or accelerate that thread toward the
4564 * processor it's on.
4566 * @preempt: whether task preemption is allowed or not
4568 * It's the caller's job to ensure that the target task struct
4569 * can't go away on us before we can do any checks.
4571 * Returns true if we indeed boosted the target task.
4573 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4575 struct task_struct
*curr
= current
;
4576 struct rq
*rq
, *p_rq
;
4577 unsigned long flags
;
4580 local_irq_save(flags
);
4585 double_rq_lock(rq
, p_rq
);
4586 while (task_rq(p
) != p_rq
) {
4587 double_rq_unlock(rq
, p_rq
);
4591 if (!curr
->sched_class
->yield_to_task
)
4594 if (curr
->sched_class
!= p
->sched_class
)
4597 if (task_running(p_rq
, p
) || p
->state
)
4600 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4602 schedstat_inc(rq
, yld_count
);
4604 * Make p's CPU reschedule; pick_next_entity takes care of
4607 if (preempt
&& rq
!= p_rq
)
4608 resched_task(p_rq
->curr
);
4611 * We might have set it in task_yield_fair(), but are
4612 * not going to schedule(), so don't want to skip
4615 rq
->skip_clock_update
= 0;
4619 double_rq_unlock(rq
, p_rq
);
4620 local_irq_restore(flags
);
4627 EXPORT_SYMBOL_GPL(yield_to
);
4630 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4631 * that process accounting knows that this is a task in IO wait state.
4633 void __sched
io_schedule(void)
4635 struct rq
*rq
= raw_rq();
4637 delayacct_blkio_start();
4638 atomic_inc(&rq
->nr_iowait
);
4639 blk_flush_plug(current
);
4640 current
->in_iowait
= 1;
4642 current
->in_iowait
= 0;
4643 atomic_dec(&rq
->nr_iowait
);
4644 delayacct_blkio_end();
4646 EXPORT_SYMBOL(io_schedule
);
4648 long __sched
io_schedule_timeout(long timeout
)
4650 struct rq
*rq
= raw_rq();
4653 delayacct_blkio_start();
4654 atomic_inc(&rq
->nr_iowait
);
4655 blk_flush_plug(current
);
4656 current
->in_iowait
= 1;
4657 ret
= schedule_timeout(timeout
);
4658 current
->in_iowait
= 0;
4659 atomic_dec(&rq
->nr_iowait
);
4660 delayacct_blkio_end();
4665 * sys_sched_get_priority_max - return maximum RT priority.
4666 * @policy: scheduling class.
4668 * this syscall returns the maximum rt_priority that can be used
4669 * by a given scheduling class.
4671 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4678 ret
= MAX_USER_RT_PRIO
-1;
4690 * sys_sched_get_priority_min - return minimum RT priority.
4691 * @policy: scheduling class.
4693 * this syscall returns the minimum rt_priority that can be used
4694 * by a given scheduling class.
4696 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4714 * sys_sched_rr_get_interval - return the default timeslice of a process.
4715 * @pid: pid of the process.
4716 * @interval: userspace pointer to the timeslice value.
4718 * this syscall writes the default timeslice value of a given process
4719 * into the user-space timespec buffer. A value of '0' means infinity.
4721 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4722 struct timespec __user
*, interval
)
4724 struct task_struct
*p
;
4725 unsigned int time_slice
;
4726 unsigned long flags
;
4736 p
= find_process_by_pid(pid
);
4740 retval
= security_task_getscheduler(p
);
4744 rq
= task_rq_lock(p
, &flags
);
4745 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4746 task_rq_unlock(rq
, p
, &flags
);
4749 jiffies_to_timespec(time_slice
, &t
);
4750 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4758 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4760 void sched_show_task(struct task_struct
*p
)
4762 unsigned long free
= 0;
4765 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4766 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4767 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4768 #if BITS_PER_LONG == 32
4769 if (state
== TASK_RUNNING
)
4770 printk(KERN_CONT
" running ");
4772 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4774 if (state
== TASK_RUNNING
)
4775 printk(KERN_CONT
" running task ");
4777 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4779 #ifdef CONFIG_DEBUG_STACK_USAGE
4780 free
= stack_not_used(p
);
4782 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4783 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4784 (unsigned long)task_thread_info(p
)->flags
);
4786 show_stack(p
, NULL
);
4789 void show_state_filter(unsigned long state_filter
)
4791 struct task_struct
*g
, *p
;
4793 #if BITS_PER_LONG == 32
4795 " task PC stack pid father\n");
4798 " task PC stack pid father\n");
4801 do_each_thread(g
, p
) {
4803 * reset the NMI-timeout, listing all files on a slow
4804 * console might take a lot of time:
4806 touch_nmi_watchdog();
4807 if (!state_filter
|| (p
->state
& state_filter
))
4809 } while_each_thread(g
, p
);
4811 touch_all_softlockup_watchdogs();
4813 #ifdef CONFIG_SCHED_DEBUG
4814 sysrq_sched_debug_show();
4818 * Only show locks if all tasks are dumped:
4821 debug_show_all_locks();
4824 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4826 idle
->sched_class
= &idle_sched_class
;
4830 * init_idle - set up an idle thread for a given CPU
4831 * @idle: task in question
4832 * @cpu: cpu the idle task belongs to
4834 * NOTE: this function does not set the idle thread's NEED_RESCHED
4835 * flag, to make booting more robust.
4837 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4839 struct rq
*rq
= cpu_rq(cpu
);
4840 unsigned long flags
;
4842 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4845 idle
->state
= TASK_RUNNING
;
4846 idle
->se
.exec_start
= sched_clock();
4848 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4850 * We're having a chicken and egg problem, even though we are
4851 * holding rq->lock, the cpu isn't yet set to this cpu so the
4852 * lockdep check in task_group() will fail.
4854 * Similar case to sched_fork(). / Alternatively we could
4855 * use task_rq_lock() here and obtain the other rq->lock.
4860 __set_task_cpu(idle
, cpu
);
4863 rq
->curr
= rq
->idle
= idle
;
4864 #if defined(CONFIG_SMP)
4867 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4869 /* Set the preempt count _outside_ the spinlocks! */
4870 task_thread_info(idle
)->preempt_count
= 0;
4873 * The idle tasks have their own, simple scheduling class:
4875 idle
->sched_class
= &idle_sched_class
;
4876 ftrace_graph_init_idle_task(idle
, cpu
);
4877 #if defined(CONFIG_SMP)
4878 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4883 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4885 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4886 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4888 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4889 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
4893 * This is how migration works:
4895 * 1) we invoke migration_cpu_stop() on the target CPU using
4897 * 2) stopper starts to run (implicitly forcing the migrated thread
4899 * 3) it checks whether the migrated task is still in the wrong runqueue.
4900 * 4) if it's in the wrong runqueue then the migration thread removes
4901 * it and puts it into the right queue.
4902 * 5) stopper completes and stop_one_cpu() returns and the migration
4907 * Change a given task's CPU affinity. Migrate the thread to a
4908 * proper CPU and schedule it away if the CPU it's executing on
4909 * is removed from the allowed bitmask.
4911 * NOTE: the caller must have a valid reference to the task, the
4912 * task must not exit() & deallocate itself prematurely. The
4913 * call is not atomic; no spinlocks may be held.
4915 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4917 unsigned long flags
;
4919 unsigned int dest_cpu
;
4922 rq
= task_rq_lock(p
, &flags
);
4924 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4927 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4932 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4937 do_set_cpus_allowed(p
, new_mask
);
4939 /* Can the task run on the task's current CPU? If so, we're done */
4940 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4943 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4945 struct migration_arg arg
= { p
, dest_cpu
};
4946 /* Need help from migration thread: drop lock and wait. */
4947 task_rq_unlock(rq
, p
, &flags
);
4948 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4949 tlb_migrate_finish(p
->mm
);
4953 task_rq_unlock(rq
, p
, &flags
);
4957 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4960 * Move (not current) task off this cpu, onto dest cpu. We're doing
4961 * this because either it can't run here any more (set_cpus_allowed()
4962 * away from this CPU, or CPU going down), or because we're
4963 * attempting to rebalance this task on exec (sched_exec).
4965 * So we race with normal scheduler movements, but that's OK, as long
4966 * as the task is no longer on this CPU.
4968 * Returns non-zero if task was successfully migrated.
4970 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4972 struct rq
*rq_dest
, *rq_src
;
4975 if (unlikely(!cpu_active(dest_cpu
)))
4978 rq_src
= cpu_rq(src_cpu
);
4979 rq_dest
= cpu_rq(dest_cpu
);
4981 raw_spin_lock(&p
->pi_lock
);
4982 double_rq_lock(rq_src
, rq_dest
);
4983 /* Already moved. */
4984 if (task_cpu(p
) != src_cpu
)
4986 /* Affinity changed (again). */
4987 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4991 * If we're not on a rq, the next wake-up will ensure we're
4995 dequeue_task(rq_src
, p
, 0);
4996 set_task_cpu(p
, dest_cpu
);
4997 enqueue_task(rq_dest
, p
, 0);
4998 check_preempt_curr(rq_dest
, p
, 0);
5003 double_rq_unlock(rq_src
, rq_dest
);
5004 raw_spin_unlock(&p
->pi_lock
);
5009 * migration_cpu_stop - this will be executed by a highprio stopper thread
5010 * and performs thread migration by bumping thread off CPU then
5011 * 'pushing' onto another runqueue.
5013 static int migration_cpu_stop(void *data
)
5015 struct migration_arg
*arg
= data
;
5018 * The original target cpu might have gone down and we might
5019 * be on another cpu but it doesn't matter.
5021 local_irq_disable();
5022 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5027 #ifdef CONFIG_HOTPLUG_CPU
5030 * Ensures that the idle task is using init_mm right before its cpu goes
5033 void idle_task_exit(void)
5035 struct mm_struct
*mm
= current
->active_mm
;
5037 BUG_ON(cpu_online(smp_processor_id()));
5040 switch_mm(mm
, &init_mm
, current
);
5045 * While a dead CPU has no uninterruptible tasks queued at this point,
5046 * it might still have a nonzero ->nr_uninterruptible counter, because
5047 * for performance reasons the counter is not stricly tracking tasks to
5048 * their home CPUs. So we just add the counter to another CPU's counter,
5049 * to keep the global sum constant after CPU-down:
5051 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5053 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5055 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5056 rq_src
->nr_uninterruptible
= 0;
5060 * remove the tasks which were accounted by rq from calc_load_tasks.
5062 static void calc_global_load_remove(struct rq
*rq
)
5064 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5065 rq
->calc_load_active
= 0;
5069 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5070 * try_to_wake_up()->select_task_rq().
5072 * Called with rq->lock held even though we'er in stop_machine() and
5073 * there's no concurrency possible, we hold the required locks anyway
5074 * because of lock validation efforts.
5076 static void migrate_tasks(unsigned int dead_cpu
)
5078 struct rq
*rq
= cpu_rq(dead_cpu
);
5079 struct task_struct
*next
, *stop
= rq
->stop
;
5083 * Fudge the rq selection such that the below task selection loop
5084 * doesn't get stuck on the currently eligible stop task.
5086 * We're currently inside stop_machine() and the rq is either stuck
5087 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5088 * either way we should never end up calling schedule() until we're
5093 /* Ensure any throttled groups are reachable by pick_next_task */
5094 unthrottle_offline_cfs_rqs(rq
);
5098 * There's this thread running, bail when that's the only
5101 if (rq
->nr_running
== 1)
5104 next
= pick_next_task(rq
);
5106 next
->sched_class
->put_prev_task(rq
, next
);
5108 /* Find suitable destination for @next, with force if needed. */
5109 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5110 raw_spin_unlock(&rq
->lock
);
5112 __migrate_task(next
, dead_cpu
, dest_cpu
);
5114 raw_spin_lock(&rq
->lock
);
5120 #endif /* CONFIG_HOTPLUG_CPU */
5122 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5124 static struct ctl_table sd_ctl_dir
[] = {
5126 .procname
= "sched_domain",
5132 static struct ctl_table sd_ctl_root
[] = {
5134 .procname
= "kernel",
5136 .child
= sd_ctl_dir
,
5141 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5143 struct ctl_table
*entry
=
5144 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5149 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5151 struct ctl_table
*entry
;
5154 * In the intermediate directories, both the child directory and
5155 * procname are dynamically allocated and could fail but the mode
5156 * will always be set. In the lowest directory the names are
5157 * static strings and all have proc handlers.
5159 for (entry
= *tablep
; entry
->mode
; entry
++) {
5161 sd_free_ctl_entry(&entry
->child
);
5162 if (entry
->proc_handler
== NULL
)
5163 kfree(entry
->procname
);
5171 set_table_entry(struct ctl_table
*entry
,
5172 const char *procname
, void *data
, int maxlen
,
5173 umode_t mode
, proc_handler
*proc_handler
)
5175 entry
->procname
= procname
;
5177 entry
->maxlen
= maxlen
;
5179 entry
->proc_handler
= proc_handler
;
5182 static struct ctl_table
*
5183 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5185 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5190 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5191 sizeof(long), 0644, proc_doulongvec_minmax
);
5192 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5193 sizeof(long), 0644, proc_doulongvec_minmax
);
5194 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5195 sizeof(int), 0644, proc_dointvec_minmax
);
5196 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5197 sizeof(int), 0644, proc_dointvec_minmax
);
5198 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5199 sizeof(int), 0644, proc_dointvec_minmax
);
5200 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5201 sizeof(int), 0644, proc_dointvec_minmax
);
5202 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5203 sizeof(int), 0644, proc_dointvec_minmax
);
5204 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5205 sizeof(int), 0644, proc_dointvec_minmax
);
5206 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5207 sizeof(int), 0644, proc_dointvec_minmax
);
5208 set_table_entry(&table
[9], "cache_nice_tries",
5209 &sd
->cache_nice_tries
,
5210 sizeof(int), 0644, proc_dointvec_minmax
);
5211 set_table_entry(&table
[10], "flags", &sd
->flags
,
5212 sizeof(int), 0644, proc_dointvec_minmax
);
5213 set_table_entry(&table
[11], "name", sd
->name
,
5214 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5215 /* &table[12] is terminator */
5220 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5222 struct ctl_table
*entry
, *table
;
5223 struct sched_domain
*sd
;
5224 int domain_num
= 0, i
;
5227 for_each_domain(cpu
, sd
)
5229 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5234 for_each_domain(cpu
, sd
) {
5235 snprintf(buf
, 32, "domain%d", i
);
5236 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5238 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5245 static struct ctl_table_header
*sd_sysctl_header
;
5246 static void register_sched_domain_sysctl(void)
5248 int i
, cpu_num
= num_possible_cpus();
5249 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5252 WARN_ON(sd_ctl_dir
[0].child
);
5253 sd_ctl_dir
[0].child
= entry
;
5258 for_each_possible_cpu(i
) {
5259 snprintf(buf
, 32, "cpu%d", i
);
5260 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5262 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5266 WARN_ON(sd_sysctl_header
);
5267 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5270 /* may be called multiple times per register */
5271 static void unregister_sched_domain_sysctl(void)
5273 if (sd_sysctl_header
)
5274 unregister_sysctl_table(sd_sysctl_header
);
5275 sd_sysctl_header
= NULL
;
5276 if (sd_ctl_dir
[0].child
)
5277 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5280 static void register_sched_domain_sysctl(void)
5283 static void unregister_sched_domain_sysctl(void)
5288 static void set_rq_online(struct rq
*rq
)
5291 const struct sched_class
*class;
5293 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5296 for_each_class(class) {
5297 if (class->rq_online
)
5298 class->rq_online(rq
);
5303 static void set_rq_offline(struct rq
*rq
)
5306 const struct sched_class
*class;
5308 for_each_class(class) {
5309 if (class->rq_offline
)
5310 class->rq_offline(rq
);
5313 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5319 * migration_call - callback that gets triggered when a CPU is added.
5320 * Here we can start up the necessary migration thread for the new CPU.
5322 static int __cpuinit
5323 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5325 int cpu
= (long)hcpu
;
5326 unsigned long flags
;
5327 struct rq
*rq
= cpu_rq(cpu
);
5329 switch (action
& ~CPU_TASKS_FROZEN
) {
5331 case CPU_UP_PREPARE
:
5332 rq
->calc_load_update
= calc_load_update
;
5336 /* Update our root-domain */
5337 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5339 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5343 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5346 #ifdef CONFIG_HOTPLUG_CPU
5348 sched_ttwu_pending();
5349 /* Update our root-domain */
5350 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5352 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5356 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5357 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5359 migrate_nr_uninterruptible(rq
);
5360 calc_global_load_remove(rq
);
5365 update_max_interval();
5371 * Register at high priority so that task migration (migrate_all_tasks)
5372 * happens before everything else. This has to be lower priority than
5373 * the notifier in the perf_event subsystem, though.
5375 static struct notifier_block __cpuinitdata migration_notifier
= {
5376 .notifier_call
= migration_call
,
5377 .priority
= CPU_PRI_MIGRATION
,
5380 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5381 unsigned long action
, void *hcpu
)
5383 switch (action
& ~CPU_TASKS_FROZEN
) {
5385 case CPU_DOWN_FAILED
:
5386 set_cpu_active((long)hcpu
, true);
5393 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5394 unsigned long action
, void *hcpu
)
5396 switch (action
& ~CPU_TASKS_FROZEN
) {
5397 case CPU_DOWN_PREPARE
:
5398 set_cpu_active((long)hcpu
, false);
5405 static int __init
migration_init(void)
5407 void *cpu
= (void *)(long)smp_processor_id();
5410 /* Initialize migration for the boot CPU */
5411 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5412 BUG_ON(err
== NOTIFY_BAD
);
5413 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5414 register_cpu_notifier(&migration_notifier
);
5416 /* Register cpu active notifiers */
5417 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5418 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5422 early_initcall(migration_init
);
5427 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5429 #ifdef CONFIG_SCHED_DEBUG
5431 static __read_mostly
int sched_domain_debug_enabled
;
5433 static int __init
sched_domain_debug_setup(char *str
)
5435 sched_domain_debug_enabled
= 1;
5439 early_param("sched_debug", sched_domain_debug_setup
);
5441 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5442 struct cpumask
*groupmask
)
5444 struct sched_group
*group
= sd
->groups
;
5447 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5448 cpumask_clear(groupmask
);
5450 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5452 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5453 printk("does not load-balance\n");
5455 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5460 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5462 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5463 printk(KERN_ERR
"ERROR: domain->span does not contain "
5466 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5467 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5471 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5475 printk(KERN_ERR
"ERROR: group is NULL\n");
5479 if (!group
->sgp
->power
) {
5480 printk(KERN_CONT
"\n");
5481 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5486 if (!cpumask_weight(sched_group_cpus(group
))) {
5487 printk(KERN_CONT
"\n");
5488 printk(KERN_ERR
"ERROR: empty group\n");
5492 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5493 printk(KERN_CONT
"\n");
5494 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5498 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5500 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5502 printk(KERN_CONT
" %s", str
);
5503 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5504 printk(KERN_CONT
" (cpu_power = %d)",
5508 group
= group
->next
;
5509 } while (group
!= sd
->groups
);
5510 printk(KERN_CONT
"\n");
5512 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5513 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5516 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5517 printk(KERN_ERR
"ERROR: parent span is not a superset "
5518 "of domain->span\n");
5522 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5526 if (!sched_domain_debug_enabled
)
5530 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5534 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5537 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5545 #else /* !CONFIG_SCHED_DEBUG */
5546 # define sched_domain_debug(sd, cpu) do { } while (0)
5547 #endif /* CONFIG_SCHED_DEBUG */
5549 static int sd_degenerate(struct sched_domain
*sd
)
5551 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5554 /* Following flags need at least 2 groups */
5555 if (sd
->flags
& (SD_LOAD_BALANCE
|
5556 SD_BALANCE_NEWIDLE
|
5560 SD_SHARE_PKG_RESOURCES
)) {
5561 if (sd
->groups
!= sd
->groups
->next
)
5565 /* Following flags don't use groups */
5566 if (sd
->flags
& (SD_WAKE_AFFINE
))
5573 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5575 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5577 if (sd_degenerate(parent
))
5580 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5583 /* Flags needing groups don't count if only 1 group in parent */
5584 if (parent
->groups
== parent
->groups
->next
) {
5585 pflags
&= ~(SD_LOAD_BALANCE
|
5586 SD_BALANCE_NEWIDLE
|
5590 SD_SHARE_PKG_RESOURCES
);
5591 if (nr_node_ids
== 1)
5592 pflags
&= ~SD_SERIALIZE
;
5594 if (~cflags
& pflags
)
5600 static void free_rootdomain(struct rcu_head
*rcu
)
5602 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5604 cpupri_cleanup(&rd
->cpupri
);
5605 free_cpumask_var(rd
->rto_mask
);
5606 free_cpumask_var(rd
->online
);
5607 free_cpumask_var(rd
->span
);
5611 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5613 struct root_domain
*old_rd
= NULL
;
5614 unsigned long flags
;
5616 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5621 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5624 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5627 * If we dont want to free the old_rt yet then
5628 * set old_rd to NULL to skip the freeing later
5631 if (!atomic_dec_and_test(&old_rd
->refcount
))
5635 atomic_inc(&rd
->refcount
);
5638 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5639 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5642 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5645 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5648 static int init_rootdomain(struct root_domain
*rd
)
5650 memset(rd
, 0, sizeof(*rd
));
5652 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5654 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5656 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5659 if (cpupri_init(&rd
->cpupri
) != 0)
5664 free_cpumask_var(rd
->rto_mask
);
5666 free_cpumask_var(rd
->online
);
5668 free_cpumask_var(rd
->span
);
5674 * By default the system creates a single root-domain with all cpus as
5675 * members (mimicking the global state we have today).
5677 struct root_domain def_root_domain
;
5679 static void init_defrootdomain(void)
5681 init_rootdomain(&def_root_domain
);
5683 atomic_set(&def_root_domain
.refcount
, 1);
5686 static struct root_domain
*alloc_rootdomain(void)
5688 struct root_domain
*rd
;
5690 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5694 if (init_rootdomain(rd
) != 0) {
5702 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5704 struct sched_group
*tmp
, *first
;
5713 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5718 } while (sg
!= first
);
5721 static void free_sched_domain(struct rcu_head
*rcu
)
5723 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5726 * If its an overlapping domain it has private groups, iterate and
5729 if (sd
->flags
& SD_OVERLAP
) {
5730 free_sched_groups(sd
->groups
, 1);
5731 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5732 kfree(sd
->groups
->sgp
);
5738 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5740 call_rcu(&sd
->rcu
, free_sched_domain
);
5743 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5745 for (; sd
; sd
= sd
->parent
)
5746 destroy_sched_domain(sd
, cpu
);
5750 * Keep a special pointer to the highest sched_domain that has
5751 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5752 * allows us to avoid some pointer chasing select_idle_sibling().
5754 * Also keep a unique ID per domain (we use the first cpu number in
5755 * the cpumask of the domain), this allows us to quickly tell if
5756 * two cpus are in the same cache domain, see ttwu_share_cache().
5758 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5759 DEFINE_PER_CPU(int, sd_llc_id
);
5761 static void update_top_cache_domain(int cpu
)
5763 struct sched_domain
*sd
;
5766 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5768 id
= cpumask_first(sched_domain_span(sd
));
5770 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5771 per_cpu(sd_llc_id
, cpu
) = id
;
5775 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5776 * hold the hotplug lock.
5779 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5781 struct rq
*rq
= cpu_rq(cpu
);
5782 struct sched_domain
*tmp
;
5784 /* Remove the sched domains which do not contribute to scheduling. */
5785 for (tmp
= sd
; tmp
; ) {
5786 struct sched_domain
*parent
= tmp
->parent
;
5790 if (sd_parent_degenerate(tmp
, parent
)) {
5791 tmp
->parent
= parent
->parent
;
5793 parent
->parent
->child
= tmp
;
5794 destroy_sched_domain(parent
, cpu
);
5799 if (sd
&& sd_degenerate(sd
)) {
5802 destroy_sched_domain(tmp
, cpu
);
5807 sched_domain_debug(sd
, cpu
);
5809 rq_attach_root(rq
, rd
);
5811 rcu_assign_pointer(rq
->sd
, sd
);
5812 destroy_sched_domains(tmp
, cpu
);
5814 update_top_cache_domain(cpu
);
5817 /* cpus with isolated domains */
5818 static cpumask_var_t cpu_isolated_map
;
5820 /* Setup the mask of cpus configured for isolated domains */
5821 static int __init
isolated_cpu_setup(char *str
)
5823 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5824 cpulist_parse(str
, cpu_isolated_map
);
5828 __setup("isolcpus=", isolated_cpu_setup
);
5833 * find_next_best_node - find the next node to include in a sched_domain
5834 * @node: node whose sched_domain we're building
5835 * @used_nodes: nodes already in the sched_domain
5837 * Find the next node to include in a given scheduling domain. Simply
5838 * finds the closest node not already in the @used_nodes map.
5840 * Should use nodemask_t.
5842 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
5844 int i
, n
, val
, min_val
, best_node
= -1;
5848 for (i
= 0; i
< nr_node_ids
; i
++) {
5849 /* Start at @node */
5850 n
= (node
+ i
) % nr_node_ids
;
5852 if (!nr_cpus_node(n
))
5855 /* Skip already used nodes */
5856 if (node_isset(n
, *used_nodes
))
5859 /* Simple min distance search */
5860 val
= node_distance(node
, n
);
5862 if (val
< min_val
) {
5868 if (best_node
!= -1)
5869 node_set(best_node
, *used_nodes
);
5874 * sched_domain_node_span - get a cpumask for a node's sched_domain
5875 * @node: node whose cpumask we're constructing
5876 * @span: resulting cpumask
5878 * Given a node, construct a good cpumask for its sched_domain to span. It
5879 * should be one that prevents unnecessary balancing, but also spreads tasks
5882 static void sched_domain_node_span(int node
, struct cpumask
*span
)
5884 nodemask_t used_nodes
;
5887 cpumask_clear(span
);
5888 nodes_clear(used_nodes
);
5890 cpumask_or(span
, span
, cpumask_of_node(node
));
5891 node_set(node
, used_nodes
);
5893 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5894 int next_node
= find_next_best_node(node
, &used_nodes
);
5897 cpumask_or(span
, span
, cpumask_of_node(next_node
));
5901 static const struct cpumask
*cpu_node_mask(int cpu
)
5903 lockdep_assert_held(&sched_domains_mutex
);
5905 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
5907 return sched_domains_tmpmask
;
5910 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
5912 return cpu_possible_mask
;
5914 #endif /* CONFIG_NUMA */
5916 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5918 return cpumask_of_node(cpu_to_node(cpu
));
5921 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5924 struct sched_domain
**__percpu sd
;
5925 struct sched_group
**__percpu sg
;
5926 struct sched_group_power
**__percpu sgp
;
5930 struct sched_domain
** __percpu sd
;
5931 struct root_domain
*rd
;
5941 struct sched_domain_topology_level
;
5943 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5944 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5946 #define SDTL_OVERLAP 0x01
5948 struct sched_domain_topology_level
{
5949 sched_domain_init_f init
;
5950 sched_domain_mask_f mask
;
5952 struct sd_data data
;
5956 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5958 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5959 const struct cpumask
*span
= sched_domain_span(sd
);
5960 struct cpumask
*covered
= sched_domains_tmpmask
;
5961 struct sd_data
*sdd
= sd
->private;
5962 struct sched_domain
*child
;
5965 cpumask_clear(covered
);
5967 for_each_cpu(i
, span
) {
5968 struct cpumask
*sg_span
;
5970 if (cpumask_test_cpu(i
, covered
))
5973 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5974 GFP_KERNEL
, cpu_to_node(cpu
));
5979 sg_span
= sched_group_cpus(sg
);
5981 child
= *per_cpu_ptr(sdd
->sd
, i
);
5983 child
= child
->child
;
5984 cpumask_copy(sg_span
, sched_domain_span(child
));
5986 cpumask_set_cpu(i
, sg_span
);
5988 cpumask_or(covered
, covered
, sg_span
);
5990 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
5991 atomic_inc(&sg
->sgp
->ref
);
5993 if (cpumask_test_cpu(cpu
, sg_span
))
6003 sd
->groups
= groups
;
6008 free_sched_groups(first
, 0);
6013 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6015 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6016 struct sched_domain
*child
= sd
->child
;
6019 cpu
= cpumask_first(sched_domain_span(child
));
6022 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6023 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
6024 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
6031 * build_sched_groups will build a circular linked list of the groups
6032 * covered by the given span, and will set each group's ->cpumask correctly,
6033 * and ->cpu_power to 0.
6035 * Assumes the sched_domain tree is fully constructed
6038 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6040 struct sched_group
*first
= NULL
, *last
= NULL
;
6041 struct sd_data
*sdd
= sd
->private;
6042 const struct cpumask
*span
= sched_domain_span(sd
);
6043 struct cpumask
*covered
;
6046 get_group(cpu
, sdd
, &sd
->groups
);
6047 atomic_inc(&sd
->groups
->ref
);
6049 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
6052 lockdep_assert_held(&sched_domains_mutex
);
6053 covered
= sched_domains_tmpmask
;
6055 cpumask_clear(covered
);
6057 for_each_cpu(i
, span
) {
6058 struct sched_group
*sg
;
6059 int group
= get_group(i
, sdd
, &sg
);
6062 if (cpumask_test_cpu(i
, covered
))
6065 cpumask_clear(sched_group_cpus(sg
));
6068 for_each_cpu(j
, span
) {
6069 if (get_group(j
, sdd
, NULL
) != group
)
6072 cpumask_set_cpu(j
, covered
);
6073 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6088 * Initialize sched groups cpu_power.
6090 * cpu_power indicates the capacity of sched group, which is used while
6091 * distributing the load between different sched groups in a sched domain.
6092 * Typically cpu_power for all the groups in a sched domain will be same unless
6093 * there are asymmetries in the topology. If there are asymmetries, group
6094 * having more cpu_power will pickup more load compared to the group having
6097 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6099 struct sched_group
*sg
= sd
->groups
;
6101 WARN_ON(!sd
|| !sg
);
6104 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6106 } while (sg
!= sd
->groups
);
6108 if (cpu
!= group_first_cpu(sg
))
6111 update_group_power(sd
, cpu
);
6112 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6115 int __weak
arch_sd_sibling_asym_packing(void)
6117 return 0*SD_ASYM_PACKING
;
6121 * Initializers for schedule domains
6122 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6125 #ifdef CONFIG_SCHED_DEBUG
6126 # define SD_INIT_NAME(sd, type) sd->name = #type
6128 # define SD_INIT_NAME(sd, type) do { } while (0)
6131 #define SD_INIT_FUNC(type) \
6132 static noinline struct sched_domain * \
6133 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6135 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6136 *sd = SD_##type##_INIT; \
6137 SD_INIT_NAME(sd, type); \
6138 sd->private = &tl->data; \
6144 SD_INIT_FUNC(ALLNODES
)
6147 #ifdef CONFIG_SCHED_SMT
6148 SD_INIT_FUNC(SIBLING
)
6150 #ifdef CONFIG_SCHED_MC
6153 #ifdef CONFIG_SCHED_BOOK
6157 static int default_relax_domain_level
= -1;
6158 int sched_domain_level_max
;
6160 static int __init
setup_relax_domain_level(char *str
)
6164 val
= simple_strtoul(str
, NULL
, 0);
6165 if (val
< sched_domain_level_max
)
6166 default_relax_domain_level
= val
;
6170 __setup("relax_domain_level=", setup_relax_domain_level
);
6172 static void set_domain_attribute(struct sched_domain
*sd
,
6173 struct sched_domain_attr
*attr
)
6177 if (!attr
|| attr
->relax_domain_level
< 0) {
6178 if (default_relax_domain_level
< 0)
6181 request
= default_relax_domain_level
;
6183 request
= attr
->relax_domain_level
;
6184 if (request
< sd
->level
) {
6185 /* turn off idle balance on this domain */
6186 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6188 /* turn on idle balance on this domain */
6189 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6193 static void __sdt_free(const struct cpumask
*cpu_map
);
6194 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6196 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6197 const struct cpumask
*cpu_map
)
6201 if (!atomic_read(&d
->rd
->refcount
))
6202 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6204 free_percpu(d
->sd
); /* fall through */
6206 __sdt_free(cpu_map
); /* fall through */
6212 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6213 const struct cpumask
*cpu_map
)
6215 memset(d
, 0, sizeof(*d
));
6217 if (__sdt_alloc(cpu_map
))
6218 return sa_sd_storage
;
6219 d
->sd
= alloc_percpu(struct sched_domain
*);
6221 return sa_sd_storage
;
6222 d
->rd
= alloc_rootdomain();
6225 return sa_rootdomain
;
6229 * NULL the sd_data elements we've used to build the sched_domain and
6230 * sched_group structure so that the subsequent __free_domain_allocs()
6231 * will not free the data we're using.
6233 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6235 struct sd_data
*sdd
= sd
->private;
6237 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6238 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6240 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6241 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6243 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6244 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6247 #ifdef CONFIG_SCHED_SMT
6248 static const struct cpumask
*cpu_smt_mask(int cpu
)
6250 return topology_thread_cpumask(cpu
);
6255 * Topology list, bottom-up.
6257 static struct sched_domain_topology_level default_topology
[] = {
6258 #ifdef CONFIG_SCHED_SMT
6259 { sd_init_SIBLING
, cpu_smt_mask
, },
6261 #ifdef CONFIG_SCHED_MC
6262 { sd_init_MC
, cpu_coregroup_mask
, },
6264 #ifdef CONFIG_SCHED_BOOK
6265 { sd_init_BOOK
, cpu_book_mask
, },
6267 { sd_init_CPU
, cpu_cpu_mask
, },
6269 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
6270 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
6275 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6277 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6279 struct sched_domain_topology_level
*tl
;
6282 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6283 struct sd_data
*sdd
= &tl
->data
;
6285 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6289 sdd
->sg
= alloc_percpu(struct sched_group
*);
6293 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6297 for_each_cpu(j
, cpu_map
) {
6298 struct sched_domain
*sd
;
6299 struct sched_group
*sg
;
6300 struct sched_group_power
*sgp
;
6302 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6303 GFP_KERNEL
, cpu_to_node(j
));
6307 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6309 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6310 GFP_KERNEL
, cpu_to_node(j
));
6314 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6316 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
6317 GFP_KERNEL
, cpu_to_node(j
));
6321 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6328 static void __sdt_free(const struct cpumask
*cpu_map
)
6330 struct sched_domain_topology_level
*tl
;
6333 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6334 struct sd_data
*sdd
= &tl
->data
;
6336 for_each_cpu(j
, cpu_map
) {
6337 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, j
);
6338 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6339 free_sched_groups(sd
->groups
, 0);
6340 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6341 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6342 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6344 free_percpu(sdd
->sd
);
6345 free_percpu(sdd
->sg
);
6346 free_percpu(sdd
->sgp
);
6350 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6351 struct s_data
*d
, const struct cpumask
*cpu_map
,
6352 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6355 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6359 set_domain_attribute(sd
, attr
);
6360 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6362 sd
->level
= child
->level
+ 1;
6363 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6372 * Build sched domains for a given set of cpus and attach the sched domains
6373 * to the individual cpus
6375 static int build_sched_domains(const struct cpumask
*cpu_map
,
6376 struct sched_domain_attr
*attr
)
6378 enum s_alloc alloc_state
= sa_none
;
6379 struct sched_domain
*sd
;
6381 int i
, ret
= -ENOMEM
;
6383 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6384 if (alloc_state
!= sa_rootdomain
)
6387 /* Set up domains for cpus specified by the cpu_map. */
6388 for_each_cpu(i
, cpu_map
) {
6389 struct sched_domain_topology_level
*tl
;
6392 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6393 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6394 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6395 sd
->flags
|= SD_OVERLAP
;
6396 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6403 *per_cpu_ptr(d
.sd
, i
) = sd
;
6406 /* Build the groups for the domains */
6407 for_each_cpu(i
, cpu_map
) {
6408 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6409 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6410 if (sd
->flags
& SD_OVERLAP
) {
6411 if (build_overlap_sched_groups(sd
, i
))
6414 if (build_sched_groups(sd
, i
))
6420 /* Calculate CPU power for physical packages and nodes */
6421 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6422 if (!cpumask_test_cpu(i
, cpu_map
))
6425 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6426 claim_allocations(i
, sd
);
6427 init_sched_groups_power(i
, sd
);
6431 /* Attach the domains */
6433 for_each_cpu(i
, cpu_map
) {
6434 sd
= *per_cpu_ptr(d
.sd
, i
);
6435 cpu_attach_domain(sd
, d
.rd
, i
);
6441 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6445 static cpumask_var_t
*doms_cur
; /* current sched domains */
6446 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6447 static struct sched_domain_attr
*dattr_cur
;
6448 /* attribues of custom domains in 'doms_cur' */
6451 * Special case: If a kmalloc of a doms_cur partition (array of
6452 * cpumask) fails, then fallback to a single sched domain,
6453 * as determined by the single cpumask fallback_doms.
6455 static cpumask_var_t fallback_doms
;
6458 * arch_update_cpu_topology lets virtualized architectures update the
6459 * cpu core maps. It is supposed to return 1 if the topology changed
6460 * or 0 if it stayed the same.
6462 int __attribute__((weak
)) arch_update_cpu_topology(void)
6467 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6470 cpumask_var_t
*doms
;
6472 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6475 for (i
= 0; i
< ndoms
; i
++) {
6476 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6477 free_sched_domains(doms
, i
);
6484 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6487 for (i
= 0; i
< ndoms
; i
++)
6488 free_cpumask_var(doms
[i
]);
6493 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6494 * For now this just excludes isolated cpus, but could be used to
6495 * exclude other special cases in the future.
6497 static int init_sched_domains(const struct cpumask
*cpu_map
)
6501 arch_update_cpu_topology();
6503 doms_cur
= alloc_sched_domains(ndoms_cur
);
6505 doms_cur
= &fallback_doms
;
6506 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6508 err
= build_sched_domains(doms_cur
[0], NULL
);
6509 register_sched_domain_sysctl();
6515 * Detach sched domains from a group of cpus specified in cpu_map
6516 * These cpus will now be attached to the NULL domain
6518 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6523 for_each_cpu(i
, cpu_map
)
6524 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6528 /* handle null as "default" */
6529 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6530 struct sched_domain_attr
*new, int idx_new
)
6532 struct sched_domain_attr tmp
;
6539 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6540 new ? (new + idx_new
) : &tmp
,
6541 sizeof(struct sched_domain_attr
));
6545 * Partition sched domains as specified by the 'ndoms_new'
6546 * cpumasks in the array doms_new[] of cpumasks. This compares
6547 * doms_new[] to the current sched domain partitioning, doms_cur[].
6548 * It destroys each deleted domain and builds each new domain.
6550 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6551 * The masks don't intersect (don't overlap.) We should setup one
6552 * sched domain for each mask. CPUs not in any of the cpumasks will
6553 * not be load balanced. If the same cpumask appears both in the
6554 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6557 * The passed in 'doms_new' should be allocated using
6558 * alloc_sched_domains. This routine takes ownership of it and will
6559 * free_sched_domains it when done with it. If the caller failed the
6560 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6561 * and partition_sched_domains() will fallback to the single partition
6562 * 'fallback_doms', it also forces the domains to be rebuilt.
6564 * If doms_new == NULL it will be replaced with cpu_online_mask.
6565 * ndoms_new == 0 is a special case for destroying existing domains,
6566 * and it will not create the default domain.
6568 * Call with hotplug lock held
6570 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6571 struct sched_domain_attr
*dattr_new
)
6576 mutex_lock(&sched_domains_mutex
);
6578 /* always unregister in case we don't destroy any domains */
6579 unregister_sched_domain_sysctl();
6581 /* Let architecture update cpu core mappings. */
6582 new_topology
= arch_update_cpu_topology();
6584 n
= doms_new
? ndoms_new
: 0;
6586 /* Destroy deleted domains */
6587 for (i
= 0; i
< ndoms_cur
; i
++) {
6588 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6589 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6590 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6593 /* no match - a current sched domain not in new doms_new[] */
6594 detach_destroy_domains(doms_cur
[i
]);
6599 if (doms_new
== NULL
) {
6601 doms_new
= &fallback_doms
;
6602 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6603 WARN_ON_ONCE(dattr_new
);
6606 /* Build new domains */
6607 for (i
= 0; i
< ndoms_new
; i
++) {
6608 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6609 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6610 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6613 /* no match - add a new doms_new */
6614 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6619 /* Remember the new sched domains */
6620 if (doms_cur
!= &fallback_doms
)
6621 free_sched_domains(doms_cur
, ndoms_cur
);
6622 kfree(dattr_cur
); /* kfree(NULL) is safe */
6623 doms_cur
= doms_new
;
6624 dattr_cur
= dattr_new
;
6625 ndoms_cur
= ndoms_new
;
6627 register_sched_domain_sysctl();
6629 mutex_unlock(&sched_domains_mutex
);
6632 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6633 static void reinit_sched_domains(void)
6637 /* Destroy domains first to force the rebuild */
6638 partition_sched_domains(0, NULL
, NULL
);
6640 rebuild_sched_domains();
6644 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6646 unsigned int level
= 0;
6648 if (sscanf(buf
, "%u", &level
) != 1)
6652 * level is always be positive so don't check for
6653 * level < POWERSAVINGS_BALANCE_NONE which is 0
6654 * What happens on 0 or 1 byte write,
6655 * need to check for count as well?
6658 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
6662 sched_smt_power_savings
= level
;
6664 sched_mc_power_savings
= level
;
6666 reinit_sched_domains();
6671 #ifdef CONFIG_SCHED_MC
6672 static ssize_t
sched_mc_power_savings_show(struct device
*dev
,
6673 struct device_attribute
*attr
,
6676 return sprintf(buf
, "%u\n", sched_mc_power_savings
);
6678 static ssize_t
sched_mc_power_savings_store(struct device
*dev
,
6679 struct device_attribute
*attr
,
6680 const char *buf
, size_t count
)
6682 return sched_power_savings_store(buf
, count
, 0);
6684 static DEVICE_ATTR(sched_mc_power_savings
, 0644,
6685 sched_mc_power_savings_show
,
6686 sched_mc_power_savings_store
);
6689 #ifdef CONFIG_SCHED_SMT
6690 static ssize_t
sched_smt_power_savings_show(struct device
*dev
,
6691 struct device_attribute
*attr
,
6694 return sprintf(buf
, "%u\n", sched_smt_power_savings
);
6696 static ssize_t
sched_smt_power_savings_store(struct device
*dev
,
6697 struct device_attribute
*attr
,
6698 const char *buf
, size_t count
)
6700 return sched_power_savings_store(buf
, count
, 1);
6702 static DEVICE_ATTR(sched_smt_power_savings
, 0644,
6703 sched_smt_power_savings_show
,
6704 sched_smt_power_savings_store
);
6707 int __init
sched_create_sysfs_power_savings_entries(struct device
*dev
)
6711 #ifdef CONFIG_SCHED_SMT
6713 err
= device_create_file(dev
, &dev_attr_sched_smt_power_savings
);
6715 #ifdef CONFIG_SCHED_MC
6716 if (!err
&& mc_capable())
6717 err
= device_create_file(dev
, &dev_attr_sched_mc_power_savings
);
6721 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6724 * Update cpusets according to cpu_active mask. If cpusets are
6725 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6726 * around partition_sched_domains().
6728 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6733 case CPU_DOWN_FAILED
:
6734 cpuset_update_active_cpus();
6741 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6745 case CPU_DOWN_PREPARE
:
6746 cpuset_update_active_cpus();
6753 void __init
sched_init_smp(void)
6755 cpumask_var_t non_isolated_cpus
;
6757 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6758 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6761 mutex_lock(&sched_domains_mutex
);
6762 init_sched_domains(cpu_active_mask
);
6763 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6764 if (cpumask_empty(non_isolated_cpus
))
6765 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6766 mutex_unlock(&sched_domains_mutex
);
6769 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6770 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6772 /* RT runtime code needs to handle some hotplug events */
6773 hotcpu_notifier(update_runtime
, 0);
6777 /* Move init over to a non-isolated CPU */
6778 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6780 sched_init_granularity();
6781 free_cpumask_var(non_isolated_cpus
);
6783 init_sched_rt_class();
6786 void __init
sched_init_smp(void)
6788 sched_init_granularity();
6790 #endif /* CONFIG_SMP */
6792 const_debug
unsigned int sysctl_timer_migration
= 1;
6794 int in_sched_functions(unsigned long addr
)
6796 return in_lock_functions(addr
) ||
6797 (addr
>= (unsigned long)__sched_text_start
6798 && addr
< (unsigned long)__sched_text_end
);
6801 #ifdef CONFIG_CGROUP_SCHED
6802 struct task_group root_task_group
;
6805 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6807 void __init
sched_init(void)
6810 unsigned long alloc_size
= 0, ptr
;
6812 #ifdef CONFIG_FAIR_GROUP_SCHED
6813 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6815 #ifdef CONFIG_RT_GROUP_SCHED
6816 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6818 #ifdef CONFIG_CPUMASK_OFFSTACK
6819 alloc_size
+= num_possible_cpus() * cpumask_size();
6822 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6824 #ifdef CONFIG_FAIR_GROUP_SCHED
6825 root_task_group
.se
= (struct sched_entity
**)ptr
;
6826 ptr
+= nr_cpu_ids
* sizeof(void **);
6828 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6829 ptr
+= nr_cpu_ids
* sizeof(void **);
6831 #endif /* CONFIG_FAIR_GROUP_SCHED */
6832 #ifdef CONFIG_RT_GROUP_SCHED
6833 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6834 ptr
+= nr_cpu_ids
* sizeof(void **);
6836 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6837 ptr
+= nr_cpu_ids
* sizeof(void **);
6839 #endif /* CONFIG_RT_GROUP_SCHED */
6840 #ifdef CONFIG_CPUMASK_OFFSTACK
6841 for_each_possible_cpu(i
) {
6842 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6843 ptr
+= cpumask_size();
6845 #endif /* CONFIG_CPUMASK_OFFSTACK */
6849 init_defrootdomain();
6852 init_rt_bandwidth(&def_rt_bandwidth
,
6853 global_rt_period(), global_rt_runtime());
6855 #ifdef CONFIG_RT_GROUP_SCHED
6856 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6857 global_rt_period(), global_rt_runtime());
6858 #endif /* CONFIG_RT_GROUP_SCHED */
6860 #ifdef CONFIG_CGROUP_SCHED
6861 list_add(&root_task_group
.list
, &task_groups
);
6862 INIT_LIST_HEAD(&root_task_group
.children
);
6863 INIT_LIST_HEAD(&root_task_group
.siblings
);
6864 autogroup_init(&init_task
);
6866 #endif /* CONFIG_CGROUP_SCHED */
6868 #ifdef CONFIG_CGROUP_CPUACCT
6869 root_cpuacct
.cpustat
= &kernel_cpustat
;
6870 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6871 /* Too early, not expected to fail */
6872 BUG_ON(!root_cpuacct
.cpuusage
);
6874 for_each_possible_cpu(i
) {
6878 raw_spin_lock_init(&rq
->lock
);
6880 rq
->calc_load_active
= 0;
6881 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6882 init_cfs_rq(&rq
->cfs
);
6883 init_rt_rq(&rq
->rt
, rq
);
6884 #ifdef CONFIG_FAIR_GROUP_SCHED
6885 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6886 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6888 * How much cpu bandwidth does root_task_group get?
6890 * In case of task-groups formed thr' the cgroup filesystem, it
6891 * gets 100% of the cpu resources in the system. This overall
6892 * system cpu resource is divided among the tasks of
6893 * root_task_group and its child task-groups in a fair manner,
6894 * based on each entity's (task or task-group's) weight
6895 * (se->load.weight).
6897 * In other words, if root_task_group has 10 tasks of weight
6898 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6899 * then A0's share of the cpu resource is:
6901 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6903 * We achieve this by letting root_task_group's tasks sit
6904 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6906 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6907 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6908 #endif /* CONFIG_FAIR_GROUP_SCHED */
6910 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6911 #ifdef CONFIG_RT_GROUP_SCHED
6912 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6913 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6916 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6917 rq
->cpu_load
[j
] = 0;
6919 rq
->last_load_update_tick
= jiffies
;
6924 rq
->cpu_power
= SCHED_POWER_SCALE
;
6925 rq
->post_schedule
= 0;
6926 rq
->active_balance
= 0;
6927 rq
->next_balance
= jiffies
;
6932 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6933 rq_attach_root(rq
, &def_root_domain
);
6939 atomic_set(&rq
->nr_iowait
, 0);
6942 set_load_weight(&init_task
);
6944 #ifdef CONFIG_PREEMPT_NOTIFIERS
6945 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6948 #ifdef CONFIG_RT_MUTEXES
6949 plist_head_init(&init_task
.pi_waiters
);
6953 * The boot idle thread does lazy MMU switching as well:
6955 atomic_inc(&init_mm
.mm_count
);
6956 enter_lazy_tlb(&init_mm
, current
);
6959 * Make us the idle thread. Technically, schedule() should not be
6960 * called from this thread, however somewhere below it might be,
6961 * but because we are the idle thread, we just pick up running again
6962 * when this runqueue becomes "idle".
6964 init_idle(current
, smp_processor_id());
6966 calc_load_update
= jiffies
+ LOAD_FREQ
;
6969 * During early bootup we pretend to be a normal task:
6971 current
->sched_class
= &fair_sched_class
;
6974 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6975 /* May be allocated at isolcpus cmdline parse time */
6976 if (cpu_isolated_map
== NULL
)
6977 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6979 init_sched_fair_class();
6981 scheduler_running
= 1;
6984 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6985 static inline int preempt_count_equals(int preempt_offset
)
6987 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6989 return (nested
== preempt_offset
);
6992 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6994 static unsigned long prev_jiffy
; /* ratelimiting */
6996 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6997 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6998 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7000 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7002 prev_jiffy
= jiffies
;
7005 "BUG: sleeping function called from invalid context at %s:%d\n",
7008 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7009 in_atomic(), irqs_disabled(),
7010 current
->pid
, current
->comm
);
7012 debug_show_held_locks(current
);
7013 if (irqs_disabled())
7014 print_irqtrace_events(current
);
7017 EXPORT_SYMBOL(__might_sleep
);
7020 #ifdef CONFIG_MAGIC_SYSRQ
7021 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7023 const struct sched_class
*prev_class
= p
->sched_class
;
7024 int old_prio
= p
->prio
;
7029 dequeue_task(rq
, p
, 0);
7030 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7032 enqueue_task(rq
, p
, 0);
7033 resched_task(rq
->curr
);
7036 check_class_changed(rq
, p
, prev_class
, old_prio
);
7039 void normalize_rt_tasks(void)
7041 struct task_struct
*g
, *p
;
7042 unsigned long flags
;
7045 read_lock_irqsave(&tasklist_lock
, flags
);
7046 do_each_thread(g
, p
) {
7048 * Only normalize user tasks:
7053 p
->se
.exec_start
= 0;
7054 #ifdef CONFIG_SCHEDSTATS
7055 p
->se
.statistics
.wait_start
= 0;
7056 p
->se
.statistics
.sleep_start
= 0;
7057 p
->se
.statistics
.block_start
= 0;
7062 * Renice negative nice level userspace
7065 if (TASK_NICE(p
) < 0 && p
->mm
)
7066 set_user_nice(p
, 0);
7070 raw_spin_lock(&p
->pi_lock
);
7071 rq
= __task_rq_lock(p
);
7073 normalize_task(rq
, p
);
7075 __task_rq_unlock(rq
);
7076 raw_spin_unlock(&p
->pi_lock
);
7077 } while_each_thread(g
, p
);
7079 read_unlock_irqrestore(&tasklist_lock
, flags
);
7082 #endif /* CONFIG_MAGIC_SYSRQ */
7084 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7086 * These functions are only useful for the IA64 MCA handling, or kdb.
7088 * They can only be called when the whole system has been
7089 * stopped - every CPU needs to be quiescent, and no scheduling
7090 * activity can take place. Using them for anything else would
7091 * be a serious bug, and as a result, they aren't even visible
7092 * under any other configuration.
7096 * curr_task - return the current task for a given cpu.
7097 * @cpu: the processor in question.
7099 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7101 struct task_struct
*curr_task(int cpu
)
7103 return cpu_curr(cpu
);
7106 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7110 * set_curr_task - set the current task for a given cpu.
7111 * @cpu: the processor in question.
7112 * @p: the task pointer to set.
7114 * Description: This function must only be used when non-maskable interrupts
7115 * are serviced on a separate stack. It allows the architecture to switch the
7116 * notion of the current task on a cpu in a non-blocking manner. This function
7117 * must be called with all CPU's synchronized, and interrupts disabled, the
7118 * and caller must save the original value of the current task (see
7119 * curr_task() above) and restore that value before reenabling interrupts and
7120 * re-starting the system.
7122 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7124 void set_curr_task(int cpu
, struct task_struct
*p
)
7131 #ifdef CONFIG_CGROUP_SCHED
7132 /* task_group_lock serializes the addition/removal of task groups */
7133 static DEFINE_SPINLOCK(task_group_lock
);
7135 static void free_sched_group(struct task_group
*tg
)
7137 free_fair_sched_group(tg
);
7138 free_rt_sched_group(tg
);
7143 /* allocate runqueue etc for a new task group */
7144 struct task_group
*sched_create_group(struct task_group
*parent
)
7146 struct task_group
*tg
;
7147 unsigned long flags
;
7149 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7151 return ERR_PTR(-ENOMEM
);
7153 if (!alloc_fair_sched_group(tg
, parent
))
7156 if (!alloc_rt_sched_group(tg
, parent
))
7159 spin_lock_irqsave(&task_group_lock
, flags
);
7160 list_add_rcu(&tg
->list
, &task_groups
);
7162 WARN_ON(!parent
); /* root should already exist */
7164 tg
->parent
= parent
;
7165 INIT_LIST_HEAD(&tg
->children
);
7166 list_add_rcu(&tg
->siblings
, &parent
->children
);
7167 spin_unlock_irqrestore(&task_group_lock
, flags
);
7172 free_sched_group(tg
);
7173 return ERR_PTR(-ENOMEM
);
7176 /* rcu callback to free various structures associated with a task group */
7177 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7179 /* now it should be safe to free those cfs_rqs */
7180 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7183 /* Destroy runqueue etc associated with a task group */
7184 void sched_destroy_group(struct task_group
*tg
)
7186 unsigned long flags
;
7189 /* end participation in shares distribution */
7190 for_each_possible_cpu(i
)
7191 unregister_fair_sched_group(tg
, i
);
7193 spin_lock_irqsave(&task_group_lock
, flags
);
7194 list_del_rcu(&tg
->list
);
7195 list_del_rcu(&tg
->siblings
);
7196 spin_unlock_irqrestore(&task_group_lock
, flags
);
7198 /* wait for possible concurrent references to cfs_rqs complete */
7199 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7202 /* change task's runqueue when it moves between groups.
7203 * The caller of this function should have put the task in its new group
7204 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7205 * reflect its new group.
7207 void sched_move_task(struct task_struct
*tsk
)
7210 unsigned long flags
;
7213 rq
= task_rq_lock(tsk
, &flags
);
7215 running
= task_current(rq
, tsk
);
7219 dequeue_task(rq
, tsk
, 0);
7220 if (unlikely(running
))
7221 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7223 #ifdef CONFIG_FAIR_GROUP_SCHED
7224 if (tsk
->sched_class
->task_move_group
)
7225 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7228 set_task_rq(tsk
, task_cpu(tsk
));
7230 if (unlikely(running
))
7231 tsk
->sched_class
->set_curr_task(rq
);
7233 enqueue_task(rq
, tsk
, 0);
7235 task_rq_unlock(rq
, tsk
, &flags
);
7237 #endif /* CONFIG_CGROUP_SCHED */
7239 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7240 static unsigned long to_ratio(u64 period
, u64 runtime
)
7242 if (runtime
== RUNTIME_INF
)
7245 return div64_u64(runtime
<< 20, period
);
7249 #ifdef CONFIG_RT_GROUP_SCHED
7251 * Ensure that the real time constraints are schedulable.
7253 static DEFINE_MUTEX(rt_constraints_mutex
);
7255 /* Must be called with tasklist_lock held */
7256 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7258 struct task_struct
*g
, *p
;
7260 do_each_thread(g
, p
) {
7261 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7263 } while_each_thread(g
, p
);
7268 struct rt_schedulable_data
{
7269 struct task_group
*tg
;
7274 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7276 struct rt_schedulable_data
*d
= data
;
7277 struct task_group
*child
;
7278 unsigned long total
, sum
= 0;
7279 u64 period
, runtime
;
7281 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7282 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7285 period
= d
->rt_period
;
7286 runtime
= d
->rt_runtime
;
7290 * Cannot have more runtime than the period.
7292 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7296 * Ensure we don't starve existing RT tasks.
7298 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7301 total
= to_ratio(period
, runtime
);
7304 * Nobody can have more than the global setting allows.
7306 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7310 * The sum of our children's runtime should not exceed our own.
7312 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7313 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7314 runtime
= child
->rt_bandwidth
.rt_runtime
;
7316 if (child
== d
->tg
) {
7317 period
= d
->rt_period
;
7318 runtime
= d
->rt_runtime
;
7321 sum
+= to_ratio(period
, runtime
);
7330 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7334 struct rt_schedulable_data data
= {
7336 .rt_period
= period
,
7337 .rt_runtime
= runtime
,
7341 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7347 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7348 u64 rt_period
, u64 rt_runtime
)
7352 mutex_lock(&rt_constraints_mutex
);
7353 read_lock(&tasklist_lock
);
7354 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7358 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7359 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7360 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7362 for_each_possible_cpu(i
) {
7363 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7365 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7366 rt_rq
->rt_runtime
= rt_runtime
;
7367 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7369 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7371 read_unlock(&tasklist_lock
);
7372 mutex_unlock(&rt_constraints_mutex
);
7377 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7379 u64 rt_runtime
, rt_period
;
7381 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7382 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7383 if (rt_runtime_us
< 0)
7384 rt_runtime
= RUNTIME_INF
;
7386 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7389 long sched_group_rt_runtime(struct task_group
*tg
)
7393 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7396 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7397 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7398 return rt_runtime_us
;
7401 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7403 u64 rt_runtime
, rt_period
;
7405 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7406 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7411 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7414 long sched_group_rt_period(struct task_group
*tg
)
7418 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7419 do_div(rt_period_us
, NSEC_PER_USEC
);
7420 return rt_period_us
;
7423 static int sched_rt_global_constraints(void)
7425 u64 runtime
, period
;
7428 if (sysctl_sched_rt_period
<= 0)
7431 runtime
= global_rt_runtime();
7432 period
= global_rt_period();
7435 * Sanity check on the sysctl variables.
7437 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7440 mutex_lock(&rt_constraints_mutex
);
7441 read_lock(&tasklist_lock
);
7442 ret
= __rt_schedulable(NULL
, 0, 0);
7443 read_unlock(&tasklist_lock
);
7444 mutex_unlock(&rt_constraints_mutex
);
7449 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7451 /* Don't accept realtime tasks when there is no way for them to run */
7452 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7458 #else /* !CONFIG_RT_GROUP_SCHED */
7459 static int sched_rt_global_constraints(void)
7461 unsigned long flags
;
7464 if (sysctl_sched_rt_period
<= 0)
7468 * There's always some RT tasks in the root group
7469 * -- migration, kstopmachine etc..
7471 if (sysctl_sched_rt_runtime
== 0)
7474 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7475 for_each_possible_cpu(i
) {
7476 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7478 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7479 rt_rq
->rt_runtime
= global_rt_runtime();
7480 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7482 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7486 #endif /* CONFIG_RT_GROUP_SCHED */
7488 int sched_rt_handler(struct ctl_table
*table
, int write
,
7489 void __user
*buffer
, size_t *lenp
,
7493 int old_period
, old_runtime
;
7494 static DEFINE_MUTEX(mutex
);
7497 old_period
= sysctl_sched_rt_period
;
7498 old_runtime
= sysctl_sched_rt_runtime
;
7500 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7502 if (!ret
&& write
) {
7503 ret
= sched_rt_global_constraints();
7505 sysctl_sched_rt_period
= old_period
;
7506 sysctl_sched_rt_runtime
= old_runtime
;
7508 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7509 def_rt_bandwidth
.rt_period
=
7510 ns_to_ktime(global_rt_period());
7513 mutex_unlock(&mutex
);
7518 #ifdef CONFIG_CGROUP_SCHED
7520 /* return corresponding task_group object of a cgroup */
7521 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7523 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7524 struct task_group
, css
);
7527 static struct cgroup_subsys_state
*
7528 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7530 struct task_group
*tg
, *parent
;
7532 if (!cgrp
->parent
) {
7533 /* This is early initialization for the top cgroup */
7534 return &root_task_group
.css
;
7537 parent
= cgroup_tg(cgrp
->parent
);
7538 tg
= sched_create_group(parent
);
7540 return ERR_PTR(-ENOMEM
);
7546 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7548 struct task_group
*tg
= cgroup_tg(cgrp
);
7550 sched_destroy_group(tg
);
7553 static int cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7554 struct cgroup_taskset
*tset
)
7556 struct task_struct
*task
;
7558 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7559 #ifdef CONFIG_RT_GROUP_SCHED
7560 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7563 /* We don't support RT-tasks being in separate groups */
7564 if (task
->sched_class
!= &fair_sched_class
)
7571 static void cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7572 struct cgroup_taskset
*tset
)
7574 struct task_struct
*task
;
7576 cgroup_taskset_for_each(task
, cgrp
, tset
)
7577 sched_move_task(task
);
7581 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
7582 struct cgroup
*old_cgrp
, struct task_struct
*task
)
7585 * cgroup_exit() is called in the copy_process() failure path.
7586 * Ignore this case since the task hasn't ran yet, this avoids
7587 * trying to poke a half freed task state from generic code.
7589 if (!(task
->flags
& PF_EXITING
))
7592 sched_move_task(task
);
7595 #ifdef CONFIG_FAIR_GROUP_SCHED
7596 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7599 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7602 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7604 struct task_group
*tg
= cgroup_tg(cgrp
);
7606 return (u64
) scale_load_down(tg
->shares
);
7609 #ifdef CONFIG_CFS_BANDWIDTH
7610 static DEFINE_MUTEX(cfs_constraints_mutex
);
7612 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7613 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7615 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7617 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7619 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7620 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7622 if (tg
== &root_task_group
)
7626 * Ensure we have at some amount of bandwidth every period. This is
7627 * to prevent reaching a state of large arrears when throttled via
7628 * entity_tick() resulting in prolonged exit starvation.
7630 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7634 * Likewise, bound things on the otherside by preventing insane quota
7635 * periods. This also allows us to normalize in computing quota
7638 if (period
> max_cfs_quota_period
)
7641 mutex_lock(&cfs_constraints_mutex
);
7642 ret
= __cfs_schedulable(tg
, period
, quota
);
7646 runtime_enabled
= quota
!= RUNTIME_INF
;
7647 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7648 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7649 raw_spin_lock_irq(&cfs_b
->lock
);
7650 cfs_b
->period
= ns_to_ktime(period
);
7651 cfs_b
->quota
= quota
;
7653 __refill_cfs_bandwidth_runtime(cfs_b
);
7654 /* restart the period timer (if active) to handle new period expiry */
7655 if (runtime_enabled
&& cfs_b
->timer_active
) {
7656 /* force a reprogram */
7657 cfs_b
->timer_active
= 0;
7658 __start_cfs_bandwidth(cfs_b
);
7660 raw_spin_unlock_irq(&cfs_b
->lock
);
7662 for_each_possible_cpu(i
) {
7663 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7664 struct rq
*rq
= cfs_rq
->rq
;
7666 raw_spin_lock_irq(&rq
->lock
);
7667 cfs_rq
->runtime_enabled
= runtime_enabled
;
7668 cfs_rq
->runtime_remaining
= 0;
7670 if (cfs_rq
->throttled
)
7671 unthrottle_cfs_rq(cfs_rq
);
7672 raw_spin_unlock_irq(&rq
->lock
);
7675 mutex_unlock(&cfs_constraints_mutex
);
7680 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7684 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7685 if (cfs_quota_us
< 0)
7686 quota
= RUNTIME_INF
;
7688 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7690 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7693 long tg_get_cfs_quota(struct task_group
*tg
)
7697 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7700 quota_us
= tg
->cfs_bandwidth
.quota
;
7701 do_div(quota_us
, NSEC_PER_USEC
);
7706 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7710 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7711 quota
= tg
->cfs_bandwidth
.quota
;
7713 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7716 long tg_get_cfs_period(struct task_group
*tg
)
7720 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7721 do_div(cfs_period_us
, NSEC_PER_USEC
);
7723 return cfs_period_us
;
7726 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7728 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7731 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7734 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7737 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7739 return tg_get_cfs_period(cgroup_tg(cgrp
));
7742 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7745 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7748 struct cfs_schedulable_data
{
7749 struct task_group
*tg
;
7754 * normalize group quota/period to be quota/max_period
7755 * note: units are usecs
7757 static u64
normalize_cfs_quota(struct task_group
*tg
,
7758 struct cfs_schedulable_data
*d
)
7766 period
= tg_get_cfs_period(tg
);
7767 quota
= tg_get_cfs_quota(tg
);
7770 /* note: these should typically be equivalent */
7771 if (quota
== RUNTIME_INF
|| quota
== -1)
7774 return to_ratio(period
, quota
);
7777 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7779 struct cfs_schedulable_data
*d
= data
;
7780 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7781 s64 quota
= 0, parent_quota
= -1;
7784 quota
= RUNTIME_INF
;
7786 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7788 quota
= normalize_cfs_quota(tg
, d
);
7789 parent_quota
= parent_b
->hierarchal_quota
;
7792 * ensure max(child_quota) <= parent_quota, inherit when no
7795 if (quota
== RUNTIME_INF
)
7796 quota
= parent_quota
;
7797 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7800 cfs_b
->hierarchal_quota
= quota
;
7805 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7808 struct cfs_schedulable_data data
= {
7814 if (quota
!= RUNTIME_INF
) {
7815 do_div(data
.period
, NSEC_PER_USEC
);
7816 do_div(data
.quota
, NSEC_PER_USEC
);
7820 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7826 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7827 struct cgroup_map_cb
*cb
)
7829 struct task_group
*tg
= cgroup_tg(cgrp
);
7830 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7832 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7833 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7834 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7838 #endif /* CONFIG_CFS_BANDWIDTH */
7839 #endif /* CONFIG_FAIR_GROUP_SCHED */
7841 #ifdef CONFIG_RT_GROUP_SCHED
7842 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7845 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7848 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7850 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7853 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7856 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7859 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7861 return sched_group_rt_period(cgroup_tg(cgrp
));
7863 #endif /* CONFIG_RT_GROUP_SCHED */
7865 static struct cftype cpu_files
[] = {
7866 #ifdef CONFIG_FAIR_GROUP_SCHED
7869 .read_u64
= cpu_shares_read_u64
,
7870 .write_u64
= cpu_shares_write_u64
,
7873 #ifdef CONFIG_CFS_BANDWIDTH
7875 .name
= "cfs_quota_us",
7876 .read_s64
= cpu_cfs_quota_read_s64
,
7877 .write_s64
= cpu_cfs_quota_write_s64
,
7880 .name
= "cfs_period_us",
7881 .read_u64
= cpu_cfs_period_read_u64
,
7882 .write_u64
= cpu_cfs_period_write_u64
,
7886 .read_map
= cpu_stats_show
,
7889 #ifdef CONFIG_RT_GROUP_SCHED
7891 .name
= "rt_runtime_us",
7892 .read_s64
= cpu_rt_runtime_read
,
7893 .write_s64
= cpu_rt_runtime_write
,
7896 .name
= "rt_period_us",
7897 .read_u64
= cpu_rt_period_read_uint
,
7898 .write_u64
= cpu_rt_period_write_uint
,
7903 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7905 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7908 struct cgroup_subsys cpu_cgroup_subsys
= {
7910 .create
= cpu_cgroup_create
,
7911 .destroy
= cpu_cgroup_destroy
,
7912 .can_attach
= cpu_cgroup_can_attach
,
7913 .attach
= cpu_cgroup_attach
,
7914 .exit
= cpu_cgroup_exit
,
7915 .populate
= cpu_cgroup_populate
,
7916 .subsys_id
= cpu_cgroup_subsys_id
,
7920 #endif /* CONFIG_CGROUP_SCHED */
7922 #ifdef CONFIG_CGROUP_CPUACCT
7925 * CPU accounting code for task groups.
7927 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7928 * (balbir@in.ibm.com).
7931 /* create a new cpu accounting group */
7932 static struct cgroup_subsys_state
*cpuacct_create(
7933 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7938 return &root_cpuacct
.css
;
7940 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7944 ca
->cpuusage
= alloc_percpu(u64
);
7948 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7950 goto out_free_cpuusage
;
7955 free_percpu(ca
->cpuusage
);
7959 return ERR_PTR(-ENOMEM
);
7962 /* destroy an existing cpu accounting group */
7964 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
7966 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7968 free_percpu(ca
->cpustat
);
7969 free_percpu(ca
->cpuusage
);
7973 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
7975 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7978 #ifndef CONFIG_64BIT
7980 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7982 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7984 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7992 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
7994 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7996 #ifndef CONFIG_64BIT
7998 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8000 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8002 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8008 /* return total cpu usage (in nanoseconds) of a group */
8009 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8011 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8012 u64 totalcpuusage
= 0;
8015 for_each_present_cpu(i
)
8016 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8018 return totalcpuusage
;
8021 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8024 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8033 for_each_present_cpu(i
)
8034 cpuacct_cpuusage_write(ca
, i
, 0);
8040 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8043 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8047 for_each_present_cpu(i
) {
8048 percpu
= cpuacct_cpuusage_read(ca
, i
);
8049 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8051 seq_printf(m
, "\n");
8055 static const char *cpuacct_stat_desc
[] = {
8056 [CPUACCT_STAT_USER
] = "user",
8057 [CPUACCT_STAT_SYSTEM
] = "system",
8060 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8061 struct cgroup_map_cb
*cb
)
8063 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8067 for_each_online_cpu(cpu
) {
8068 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8069 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8070 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8072 val
= cputime64_to_clock_t(val
);
8073 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8076 for_each_online_cpu(cpu
) {
8077 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8078 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8079 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8080 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8083 val
= cputime64_to_clock_t(val
);
8084 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8089 static struct cftype files
[] = {
8092 .read_u64
= cpuusage_read
,
8093 .write_u64
= cpuusage_write
,
8096 .name
= "usage_percpu",
8097 .read_seq_string
= cpuacct_percpu_seq_read
,
8101 .read_map
= cpuacct_stats_show
,
8105 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8107 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8111 * charge this task's execution time to its accounting group.
8113 * called with rq->lock held.
8115 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8120 if (unlikely(!cpuacct_subsys
.active
))
8123 cpu
= task_cpu(tsk
);
8129 for (; ca
; ca
= parent_ca(ca
)) {
8130 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8131 *cpuusage
+= cputime
;
8137 struct cgroup_subsys cpuacct_subsys
= {
8139 .create
= cpuacct_create
,
8140 .destroy
= cpuacct_destroy
,
8141 .populate
= cpuacct_populate
,
8142 .subsys_id
= cpuacct_subsys_id
,
8144 #endif /* CONFIG_CGROUP_CPUACCT */