4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
90 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
93 ktime_t soft
, hard
, now
;
96 if (hrtimer_active(period_timer
))
99 now
= hrtimer_cb_get_time(period_timer
);
100 hrtimer_forward(period_timer
, now
, period
);
102 soft
= hrtimer_get_softexpires(period_timer
);
103 hard
= hrtimer_get_expires(period_timer
);
104 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
105 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
106 HRTIMER_MODE_ABS_PINNED
, 0);
110 DEFINE_MUTEX(sched_domains_mutex
);
111 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
113 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
115 void update_rq_clock(struct rq
*rq
)
119 if (rq
->skip_clock_update
> 0)
122 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
124 update_rq_clock_task(rq
, delta
);
128 * Debugging: various feature bits
131 #define SCHED_FEAT(name, enabled) \
132 (1UL << __SCHED_FEAT_##name) * enabled |
134 const_debug
unsigned int sysctl_sched_features
=
135 #include "features.h"
140 #ifdef CONFIG_SCHED_DEBUG
141 #define SCHED_FEAT(name, enabled) \
144 static __read_mostly
char *sched_feat_names
[] = {
145 #include "features.h"
151 static int sched_feat_show(struct seq_file
*m
, void *v
)
155 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
156 if (!(sysctl_sched_features
& (1UL << i
)))
158 seq_printf(m
, "%s ", sched_feat_names
[i
]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
174 #include "features.h"
179 static void sched_feat_disable(int i
)
181 if (static_key_enabled(&sched_feat_keys
[i
]))
182 static_key_slow_dec(&sched_feat_keys
[i
]);
185 static void sched_feat_enable(int i
)
187 if (!static_key_enabled(&sched_feat_keys
[i
]))
188 static_key_slow_inc(&sched_feat_keys
[i
]);
191 static void sched_feat_disable(int i
) { };
192 static void sched_feat_enable(int i
) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
197 size_t cnt
, loff_t
*ppos
)
207 if (copy_from_user(&buf
, ubuf
, cnt
))
213 if (strncmp(cmp
, "NO_", 3) == 0) {
218 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
219 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
221 sysctl_sched_features
&= ~(1UL << i
);
222 sched_feat_disable(i
);
224 sysctl_sched_features
|= (1UL << i
);
225 sched_feat_enable(i
);
231 if (i
== __SCHED_FEAT_NR
)
239 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
241 return single_open(filp
, sched_feat_show
, NULL
);
244 static const struct file_operations sched_feat_fops
= {
245 .open
= sched_feat_open
,
246 .write
= sched_feat_write
,
249 .release
= single_release
,
252 static __init
int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
259 late_initcall(sched_init_debug
);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
269 * period over which we average the RT time consumption, measured
274 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period
= 1000000;
282 __read_mostly
int scheduler_running
;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime
= 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
300 lockdep_assert_held(&p
->pi_lock
);
304 raw_spin_lock(&rq
->lock
);
305 if (likely(rq
== task_rq(p
)))
307 raw_spin_unlock(&rq
->lock
);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
315 __acquires(p
->pi_lock
)
321 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
323 raw_spin_lock(&rq
->lock
);
324 if (likely(rq
== task_rq(p
)))
326 raw_spin_unlock(&rq
->lock
);
327 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
331 static void __task_rq_unlock(struct rq
*rq
)
334 raw_spin_unlock(&rq
->lock
);
338 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
340 __releases(p
->pi_lock
)
342 raw_spin_unlock(&rq
->lock
);
343 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq
*this_rq_lock(void)
356 raw_spin_lock(&rq
->lock
);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq
*rq
)
375 if (hrtimer_active(&rq
->hrtick_timer
))
376 hrtimer_cancel(&rq
->hrtick_timer
);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
385 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
387 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
389 raw_spin_lock(&rq
->lock
);
391 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
392 raw_spin_unlock(&rq
->lock
);
394 return HRTIMER_NORESTART
;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg
)
405 raw_spin_lock(&rq
->lock
);
406 hrtimer_restart(&rq
->hrtick_timer
);
407 rq
->hrtick_csd_pending
= 0;
408 raw_spin_unlock(&rq
->lock
);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq
*rq
, u64 delay
)
418 struct hrtimer
*timer
= &rq
->hrtick_timer
;
419 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
421 hrtimer_set_expires(timer
, time
);
423 if (rq
== this_rq()) {
424 hrtimer_restart(timer
);
425 } else if (!rq
->hrtick_csd_pending
) {
426 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
427 rq
->hrtick_csd_pending
= 1;
432 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
434 int cpu
= (int)(long)hcpu
;
437 case CPU_UP_CANCELED
:
438 case CPU_UP_CANCELED_FROZEN
:
439 case CPU_DOWN_PREPARE
:
440 case CPU_DOWN_PREPARE_FROZEN
:
442 case CPU_DEAD_FROZEN
:
443 hrtick_clear(cpu_rq(cpu
));
450 static __init
void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick
, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq
*rq
, u64 delay
)
462 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
463 HRTIMER_MODE_REL_PINNED
, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq
*rq
)
474 rq
->hrtick_csd_pending
= 0;
476 rq
->hrtick_csd
.flags
= 0;
477 rq
->hrtick_csd
.func
= __hrtick_start
;
478 rq
->hrtick_csd
.info
= rq
;
481 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
482 rq
->hrtick_timer
.function
= hrtick
;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq
*rq
)
489 static inline void init_rq_hrtick(struct rq
*rq
)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct
*p
)
515 assert_raw_spin_locked(&task_rq(p
)->lock
);
517 if (test_tsk_need_resched(p
))
520 set_tsk_need_resched(p
);
523 if (cpu
== smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p
))
529 smp_send_reschedule(cpu
);
532 void resched_cpu(int cpu
)
534 struct rq
*rq
= cpu_rq(cpu
);
537 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
539 resched_task(cpu_curr(cpu
));
540 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu
= smp_processor_id();
556 struct sched_domain
*sd
;
559 for_each_domain(cpu
, sd
) {
560 for_each_cpu(i
, sched_domain_span(sd
)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu
)
583 struct rq
*rq
= cpu_rq(cpu
);
585 if (cpu
== smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq
->curr
!= rq
->idle
)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq
->idle
);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq
->idle
))
608 smp_send_reschedule(cpu
);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu
= smp_processor_id();
614 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq
*rq
)
628 s64 period
= sched_avg_period();
630 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq
->age_stamp
));
637 rq
->age_stamp
+= period
;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct
*p
)
645 assert_raw_spin_locked(&task_rq(p
)->lock
);
646 set_tsk_need_resched(p
);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group
*from
,
659 tg_visitor down
, tg_visitor up
, void *data
)
661 struct task_group
*parent
, *child
;
667 ret
= (*down
)(parent
, data
);
670 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
677 ret
= (*up
)(parent
, data
);
678 if (ret
|| parent
== from
)
682 parent
= parent
->parent
;
689 int tg_nop(struct task_group
*tg
, void *data
)
695 void update_cpu_load(struct rq
*this_rq
);
697 static void set_load_weight(struct task_struct
*p
)
699 int prio
= p
->static_prio
- MAX_RT_PRIO
;
700 struct load_weight
*load
= &p
->se
.load
;
703 * SCHED_IDLE tasks get minimal weight:
705 if (p
->policy
== SCHED_IDLE
) {
706 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
707 load
->inv_weight
= WMULT_IDLEPRIO
;
711 load
->weight
= scale_load(prio_to_weight
[prio
]);
712 load
->inv_weight
= prio_to_wmult
[prio
];
715 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
718 sched_info_queued(p
);
719 p
->sched_class
->enqueue_task(rq
, p
, flags
);
722 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
725 sched_info_dequeued(p
);
726 p
->sched_class
->dequeue_task(rq
, p
, flags
);
729 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
731 if (task_contributes_to_load(p
))
732 rq
->nr_uninterruptible
--;
734 enqueue_task(rq
, p
, flags
);
737 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
739 if (task_contributes_to_load(p
))
740 rq
->nr_uninterruptible
++;
742 dequeue_task(rq
, p
, flags
);
745 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
748 * There are no locks covering percpu hardirq/softirq time.
749 * They are only modified in account_system_vtime, on corresponding CPU
750 * with interrupts disabled. So, writes are safe.
751 * They are read and saved off onto struct rq in update_rq_clock().
752 * This may result in other CPU reading this CPU's irq time and can
753 * race with irq/account_system_vtime on this CPU. We would either get old
754 * or new value with a side effect of accounting a slice of irq time to wrong
755 * task when irq is in progress while we read rq->clock. That is a worthy
756 * compromise in place of having locks on each irq in account_system_time.
758 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
759 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
761 static DEFINE_PER_CPU(u64
, irq_start_time
);
762 static int sched_clock_irqtime
;
764 void enable_sched_clock_irqtime(void)
766 sched_clock_irqtime
= 1;
769 void disable_sched_clock_irqtime(void)
771 sched_clock_irqtime
= 0;
775 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
777 static inline void irq_time_write_begin(void)
779 __this_cpu_inc(irq_time_seq
.sequence
);
783 static inline void irq_time_write_end(void)
786 __this_cpu_inc(irq_time_seq
.sequence
);
789 static inline u64
irq_time_read(int cpu
)
795 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
796 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
797 per_cpu(cpu_hardirq_time
, cpu
);
798 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
802 #else /* CONFIG_64BIT */
803 static inline void irq_time_write_begin(void)
807 static inline void irq_time_write_end(void)
811 static inline u64
irq_time_read(int cpu
)
813 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
815 #endif /* CONFIG_64BIT */
818 * Called before incrementing preempt_count on {soft,}irq_enter
819 * and before decrementing preempt_count on {soft,}irq_exit.
821 void account_system_vtime(struct task_struct
*curr
)
827 if (!sched_clock_irqtime
)
830 local_irq_save(flags
);
832 cpu
= smp_processor_id();
833 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
834 __this_cpu_add(irq_start_time
, delta
);
836 irq_time_write_begin();
838 * We do not account for softirq time from ksoftirqd here.
839 * We want to continue accounting softirq time to ksoftirqd thread
840 * in that case, so as not to confuse scheduler with a special task
841 * that do not consume any time, but still wants to run.
844 __this_cpu_add(cpu_hardirq_time
, delta
);
845 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
846 __this_cpu_add(cpu_softirq_time
, delta
);
848 irq_time_write_end();
849 local_irq_restore(flags
);
851 EXPORT_SYMBOL_GPL(account_system_vtime
);
853 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
855 #ifdef CONFIG_PARAVIRT
856 static inline u64
steal_ticks(u64 steal
)
858 if (unlikely(steal
> NSEC_PER_SEC
))
859 return div_u64(steal
, TICK_NSEC
);
861 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
865 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
868 * In theory, the compile should just see 0 here, and optimize out the call
869 * to sched_rt_avg_update. But I don't trust it...
871 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
872 s64 steal
= 0, irq_delta
= 0;
874 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
875 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
878 * Since irq_time is only updated on {soft,}irq_exit, we might run into
879 * this case when a previous update_rq_clock() happened inside a
882 * When this happens, we stop ->clock_task and only update the
883 * prev_irq_time stamp to account for the part that fit, so that a next
884 * update will consume the rest. This ensures ->clock_task is
887 * It does however cause some slight miss-attribution of {soft,}irq
888 * time, a more accurate solution would be to update the irq_time using
889 * the current rq->clock timestamp, except that would require using
892 if (irq_delta
> delta
)
895 rq
->prev_irq_time
+= irq_delta
;
898 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
899 if (static_key_false((¶virt_steal_rq_enabled
))) {
902 steal
= paravirt_steal_clock(cpu_of(rq
));
903 steal
-= rq
->prev_steal_time_rq
;
905 if (unlikely(steal
> delta
))
908 st
= steal_ticks(steal
);
909 steal
= st
* TICK_NSEC
;
911 rq
->prev_steal_time_rq
+= steal
;
917 rq
->clock_task
+= delta
;
919 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
920 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
921 sched_rt_avg_update(rq
, irq_delta
+ steal
);
925 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
926 static int irqtime_account_hi_update(void)
928 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
933 local_irq_save(flags
);
934 latest_ns
= this_cpu_read(cpu_hardirq_time
);
935 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_IRQ
])
937 local_irq_restore(flags
);
941 static int irqtime_account_si_update(void)
943 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
948 local_irq_save(flags
);
949 latest_ns
= this_cpu_read(cpu_softirq_time
);
950 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_SOFTIRQ
])
952 local_irq_restore(flags
);
956 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
958 #define sched_clock_irqtime (0)
962 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
964 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
965 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
969 * Make it appear like a SCHED_FIFO task, its something
970 * userspace knows about and won't get confused about.
972 * Also, it will make PI more or less work without too
973 * much confusion -- but then, stop work should not
974 * rely on PI working anyway.
976 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
978 stop
->sched_class
= &stop_sched_class
;
981 cpu_rq(cpu
)->stop
= stop
;
985 * Reset it back to a normal scheduling class so that
986 * it can die in pieces.
988 old_stop
->sched_class
= &rt_sched_class
;
993 * __normal_prio - return the priority that is based on the static prio
995 static inline int __normal_prio(struct task_struct
*p
)
997 return p
->static_prio
;
1001 * Calculate the expected normal priority: i.e. priority
1002 * without taking RT-inheritance into account. Might be
1003 * boosted by interactivity modifiers. Changes upon fork,
1004 * setprio syscalls, and whenever the interactivity
1005 * estimator recalculates.
1007 static inline int normal_prio(struct task_struct
*p
)
1011 if (task_has_rt_policy(p
))
1012 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1014 prio
= __normal_prio(p
);
1019 * Calculate the current priority, i.e. the priority
1020 * taken into account by the scheduler. This value might
1021 * be boosted by RT tasks, or might be boosted by
1022 * interactivity modifiers. Will be RT if the task got
1023 * RT-boosted. If not then it returns p->normal_prio.
1025 static int effective_prio(struct task_struct
*p
)
1027 p
->normal_prio
= normal_prio(p
);
1029 * If we are RT tasks or we were boosted to RT priority,
1030 * keep the priority unchanged. Otherwise, update priority
1031 * to the normal priority:
1033 if (!rt_prio(p
->prio
))
1034 return p
->normal_prio
;
1039 * task_curr - is this task currently executing on a CPU?
1040 * @p: the task in question.
1042 inline int task_curr(const struct task_struct
*p
)
1044 return cpu_curr(task_cpu(p
)) == p
;
1047 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1048 const struct sched_class
*prev_class
,
1051 if (prev_class
!= p
->sched_class
) {
1052 if (prev_class
->switched_from
)
1053 prev_class
->switched_from(rq
, p
);
1054 p
->sched_class
->switched_to(rq
, p
);
1055 } else if (oldprio
!= p
->prio
)
1056 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1059 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1061 const struct sched_class
*class;
1063 if (p
->sched_class
== rq
->curr
->sched_class
) {
1064 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1066 for_each_class(class) {
1067 if (class == rq
->curr
->sched_class
)
1069 if (class == p
->sched_class
) {
1070 resched_task(rq
->curr
);
1077 * A queue event has occurred, and we're going to schedule. In
1078 * this case, we can save a useless back to back clock update.
1080 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1081 rq
->skip_clock_update
= 1;
1085 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1087 #ifdef CONFIG_SCHED_DEBUG
1089 * We should never call set_task_cpu() on a blocked task,
1090 * ttwu() will sort out the placement.
1092 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1093 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
1095 #ifdef CONFIG_LOCKDEP
1097 * The caller should hold either p->pi_lock or rq->lock, when changing
1098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1100 * sched_move_task() holds both and thus holding either pins the cgroup,
1101 * see set_task_rq().
1103 * Furthermore, all task_rq users should acquire both locks, see
1106 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1107 lockdep_is_held(&task_rq(p
)->lock
)));
1111 trace_sched_migrate_task(p
, new_cpu
);
1113 if (task_cpu(p
) != new_cpu
) {
1114 p
->se
.nr_migrations
++;
1115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1118 __set_task_cpu(p
, new_cpu
);
1121 struct migration_arg
{
1122 struct task_struct
*task
;
1126 static int migration_cpu_stop(void *data
);
1129 * wait_task_inactive - wait for a thread to unschedule.
1131 * If @match_state is nonzero, it's the @p->state value just checked and
1132 * not expected to change. If it changes, i.e. @p might have woken up,
1133 * then return zero. When we succeed in waiting for @p to be off its CPU,
1134 * we return a positive number (its total switch count). If a second call
1135 * a short while later returns the same number, the caller can be sure that
1136 * @p has remained unscheduled the whole time.
1138 * The caller must ensure that the task *will* unschedule sometime soon,
1139 * else this function might spin for a *long* time. This function can't
1140 * be called with interrupts off, or it may introduce deadlock with
1141 * smp_call_function() if an IPI is sent by the same process we are
1142 * waiting to become inactive.
1144 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1146 unsigned long flags
;
1153 * We do the initial early heuristics without holding
1154 * any task-queue locks at all. We'll only try to get
1155 * the runqueue lock when things look like they will
1161 * If the task is actively running on another CPU
1162 * still, just relax and busy-wait without holding
1165 * NOTE! Since we don't hold any locks, it's not
1166 * even sure that "rq" stays as the right runqueue!
1167 * But we don't care, since "task_running()" will
1168 * return false if the runqueue has changed and p
1169 * is actually now running somewhere else!
1171 while (task_running(rq
, p
)) {
1172 if (match_state
&& unlikely(p
->state
!= match_state
))
1178 * Ok, time to look more closely! We need the rq
1179 * lock now, to be *sure*. If we're wrong, we'll
1180 * just go back and repeat.
1182 rq
= task_rq_lock(p
, &flags
);
1183 trace_sched_wait_task(p
);
1184 running
= task_running(rq
, p
);
1187 if (!match_state
|| p
->state
== match_state
)
1188 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1189 task_rq_unlock(rq
, p
, &flags
);
1192 * If it changed from the expected state, bail out now.
1194 if (unlikely(!ncsw
))
1198 * Was it really running after all now that we
1199 * checked with the proper locks actually held?
1201 * Oops. Go back and try again..
1203 if (unlikely(running
)) {
1209 * It's not enough that it's not actively running,
1210 * it must be off the runqueue _entirely_, and not
1213 * So if it was still runnable (but just not actively
1214 * running right now), it's preempted, and we should
1215 * yield - it could be a while.
1217 if (unlikely(on_rq
)) {
1218 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1220 set_current_state(TASK_UNINTERRUPTIBLE
);
1221 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1226 * Ahh, all good. It wasn't running, and it wasn't
1227 * runnable, which means that it will never become
1228 * running in the future either. We're all done!
1237 * kick_process - kick a running thread to enter/exit the kernel
1238 * @p: the to-be-kicked thread
1240 * Cause a process which is running on another CPU to enter
1241 * kernel-mode, without any delay. (to get signals handled.)
1243 * NOTE: this function doesn't have to take the runqueue lock,
1244 * because all it wants to ensure is that the remote task enters
1245 * the kernel. If the IPI races and the task has been migrated
1246 * to another CPU then no harm is done and the purpose has been
1249 void kick_process(struct task_struct
*p
)
1255 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1256 smp_send_reschedule(cpu
);
1259 EXPORT_SYMBOL_GPL(kick_process
);
1260 #endif /* CONFIG_SMP */
1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1266 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1268 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1269 enum { cpuset
, possible
, fail
} state
= cpuset
;
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu_mask(dest_cpu
, *nodemask
) {
1274 if (!cpu_online(dest_cpu
))
1276 if (!cpu_active(dest_cpu
))
1278 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1283 /* Any allowed, online CPU? */
1284 for_each_cpu_mask(dest_cpu
, *tsk_cpus_allowed(p
)) {
1285 if (!cpu_online(dest_cpu
))
1287 if (!cpu_active(dest_cpu
))
1294 /* No more Mr. Nice Guy. */
1295 cpuset_cpus_allowed_fallback(p
);
1300 do_set_cpus_allowed(p
, cpu_possible_mask
);
1311 if (state
!= cpuset
) {
1313 * Don't tell them about moving exiting tasks or
1314 * kernel threads (both mm NULL), since they never
1317 if (p
->mm
&& printk_ratelimit()) {
1318 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1319 task_pid_nr(p
), p
->comm
, cpu
);
1327 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1330 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1332 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1335 * In order not to call set_task_cpu() on a blocking task we need
1336 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1339 * Since this is common to all placement strategies, this lives here.
1341 * [ this allows ->select_task() to simply return task_cpu(p) and
1342 * not worry about this generic constraint ]
1344 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1346 cpu
= select_fallback_rq(task_cpu(p
), p
);
1351 static void update_avg(u64
*avg
, u64 sample
)
1353 s64 diff
= sample
- *avg
;
1359 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1361 #ifdef CONFIG_SCHEDSTATS
1362 struct rq
*rq
= this_rq();
1365 int this_cpu
= smp_processor_id();
1367 if (cpu
== this_cpu
) {
1368 schedstat_inc(rq
, ttwu_local
);
1369 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1371 struct sched_domain
*sd
;
1373 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1375 for_each_domain(this_cpu
, sd
) {
1376 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1377 schedstat_inc(sd
, ttwu_wake_remote
);
1384 if (wake_flags
& WF_MIGRATED
)
1385 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1387 #endif /* CONFIG_SMP */
1389 schedstat_inc(rq
, ttwu_count
);
1390 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1392 if (wake_flags
& WF_SYNC
)
1393 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1395 #endif /* CONFIG_SCHEDSTATS */
1398 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1400 activate_task(rq
, p
, en_flags
);
1403 /* if a worker is waking up, notify workqueue */
1404 if (p
->flags
& PF_WQ_WORKER
)
1405 wq_worker_waking_up(p
, cpu_of(rq
));
1409 * Mark the task runnable and perform wakeup-preemption.
1412 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1414 trace_sched_wakeup(p
, true);
1415 check_preempt_curr(rq
, p
, wake_flags
);
1417 p
->state
= TASK_RUNNING
;
1419 if (p
->sched_class
->task_woken
)
1420 p
->sched_class
->task_woken(rq
, p
);
1422 if (rq
->idle_stamp
) {
1423 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1424 u64 max
= 2*sysctl_sched_migration_cost
;
1429 update_avg(&rq
->avg_idle
, delta
);
1436 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1439 if (p
->sched_contributes_to_load
)
1440 rq
->nr_uninterruptible
--;
1443 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1444 ttwu_do_wakeup(rq
, p
, wake_flags
);
1448 * Called in case the task @p isn't fully descheduled from its runqueue,
1449 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1450 * since all we need to do is flip p->state to TASK_RUNNING, since
1451 * the task is still ->on_rq.
1453 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1458 rq
= __task_rq_lock(p
);
1460 ttwu_do_wakeup(rq
, p
, wake_flags
);
1463 __task_rq_unlock(rq
);
1469 static void sched_ttwu_pending(void)
1471 struct rq
*rq
= this_rq();
1472 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1473 struct task_struct
*p
;
1475 raw_spin_lock(&rq
->lock
);
1478 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1479 llist
= llist_next(llist
);
1480 ttwu_do_activate(rq
, p
, 0);
1483 raw_spin_unlock(&rq
->lock
);
1486 void scheduler_ipi(void)
1488 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1492 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1493 * traditionally all their work was done from the interrupt return
1494 * path. Now that we actually do some work, we need to make sure
1497 * Some archs already do call them, luckily irq_enter/exit nest
1500 * Arguably we should visit all archs and update all handlers,
1501 * however a fair share of IPIs are still resched only so this would
1502 * somewhat pessimize the simple resched case.
1505 sched_ttwu_pending();
1508 * Check if someone kicked us for doing the nohz idle load balance.
1510 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1511 this_rq()->idle_balance
= 1;
1512 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1517 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1519 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1520 smp_send_reschedule(cpu
);
1523 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1524 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1529 rq
= __task_rq_lock(p
);
1531 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1532 ttwu_do_wakeup(rq
, p
, wake_flags
);
1535 __task_rq_unlock(rq
);
1540 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1542 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1544 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1546 #endif /* CONFIG_SMP */
1548 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1550 struct rq
*rq
= cpu_rq(cpu
);
1552 #if defined(CONFIG_SMP)
1553 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1554 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1555 ttwu_queue_remote(p
, cpu
);
1560 raw_spin_lock(&rq
->lock
);
1561 ttwu_do_activate(rq
, p
, 0);
1562 raw_spin_unlock(&rq
->lock
);
1566 * try_to_wake_up - wake up a thread
1567 * @p: the thread to be awakened
1568 * @state: the mask of task states that can be woken
1569 * @wake_flags: wake modifier flags (WF_*)
1571 * Put it on the run-queue if it's not already there. The "current"
1572 * thread is always on the run-queue (except when the actual
1573 * re-schedule is in progress), and as such you're allowed to do
1574 * the simpler "current->state = TASK_RUNNING" to mark yourself
1575 * runnable without the overhead of this.
1577 * Returns %true if @p was woken up, %false if it was already running
1578 * or @state didn't match @p's state.
1581 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1583 unsigned long flags
;
1584 int cpu
, success
= 0;
1587 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1588 if (!(p
->state
& state
))
1591 success
= 1; /* we're going to change ->state */
1594 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1599 * If the owning (remote) cpu is still in the middle of schedule() with
1600 * this task as prev, wait until its done referencing the task.
1603 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1605 * In case the architecture enables interrupts in
1606 * context_switch(), we cannot busy wait, since that
1607 * would lead to deadlocks when an interrupt hits and
1608 * tries to wake up @prev. So bail and do a complete
1611 if (ttwu_activate_remote(p
, wake_flags
))
1618 * Pairs with the smp_wmb() in finish_lock_switch().
1622 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1623 p
->state
= TASK_WAKING
;
1625 if (p
->sched_class
->task_waking
)
1626 p
->sched_class
->task_waking(p
);
1628 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1629 if (task_cpu(p
) != cpu
) {
1630 wake_flags
|= WF_MIGRATED
;
1631 set_task_cpu(p
, cpu
);
1633 #endif /* CONFIG_SMP */
1637 ttwu_stat(p
, cpu
, wake_flags
);
1639 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1645 * try_to_wake_up_local - try to wake up a local task with rq lock held
1646 * @p: the thread to be awakened
1648 * Put @p on the run-queue if it's not already there. The caller must
1649 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1652 static void try_to_wake_up_local(struct task_struct
*p
)
1654 struct rq
*rq
= task_rq(p
);
1656 BUG_ON(rq
!= this_rq());
1657 BUG_ON(p
== current
);
1658 lockdep_assert_held(&rq
->lock
);
1660 if (!raw_spin_trylock(&p
->pi_lock
)) {
1661 raw_spin_unlock(&rq
->lock
);
1662 raw_spin_lock(&p
->pi_lock
);
1663 raw_spin_lock(&rq
->lock
);
1666 if (!(p
->state
& TASK_NORMAL
))
1670 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1672 ttwu_do_wakeup(rq
, p
, 0);
1673 ttwu_stat(p
, smp_processor_id(), 0);
1675 raw_spin_unlock(&p
->pi_lock
);
1679 * wake_up_process - Wake up a specific process
1680 * @p: The process to be woken up.
1682 * Attempt to wake up the nominated process and move it to the set of runnable
1683 * processes. Returns 1 if the process was woken up, 0 if it was already
1686 * It may be assumed that this function implies a write memory barrier before
1687 * changing the task state if and only if any tasks are woken up.
1689 int wake_up_process(struct task_struct
*p
)
1691 return try_to_wake_up(p
, TASK_ALL
, 0);
1693 EXPORT_SYMBOL(wake_up_process
);
1695 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1697 return try_to_wake_up(p
, state
, 0);
1701 * Perform scheduler related setup for a newly forked process p.
1702 * p is forked by current.
1704 * __sched_fork() is basic setup used by init_idle() too:
1706 static void __sched_fork(struct task_struct
*p
)
1711 p
->se
.exec_start
= 0;
1712 p
->se
.sum_exec_runtime
= 0;
1713 p
->se
.prev_sum_exec_runtime
= 0;
1714 p
->se
.nr_migrations
= 0;
1716 INIT_LIST_HEAD(&p
->se
.group_node
);
1718 #ifdef CONFIG_SCHEDSTATS
1719 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1722 INIT_LIST_HEAD(&p
->rt
.run_list
);
1724 #ifdef CONFIG_PREEMPT_NOTIFIERS
1725 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1730 * fork()/clone()-time setup:
1732 void sched_fork(struct task_struct
*p
)
1734 unsigned long flags
;
1735 int cpu
= get_cpu();
1739 * We mark the process as running here. This guarantees that
1740 * nobody will actually run it, and a signal or other external
1741 * event cannot wake it up and insert it on the runqueue either.
1743 p
->state
= TASK_RUNNING
;
1746 * Make sure we do not leak PI boosting priority to the child.
1748 p
->prio
= current
->normal_prio
;
1751 * Revert to default priority/policy on fork if requested.
1753 if (unlikely(p
->sched_reset_on_fork
)) {
1754 if (task_has_rt_policy(p
)) {
1755 p
->policy
= SCHED_NORMAL
;
1756 p
->static_prio
= NICE_TO_PRIO(0);
1758 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1759 p
->static_prio
= NICE_TO_PRIO(0);
1761 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1765 * We don't need the reset flag anymore after the fork. It has
1766 * fulfilled its duty:
1768 p
->sched_reset_on_fork
= 0;
1771 if (!rt_prio(p
->prio
))
1772 p
->sched_class
= &fair_sched_class
;
1774 if (p
->sched_class
->task_fork
)
1775 p
->sched_class
->task_fork(p
);
1778 * The child is not yet in the pid-hash so no cgroup attach races,
1779 * and the cgroup is pinned to this child due to cgroup_fork()
1780 * is ran before sched_fork().
1782 * Silence PROVE_RCU.
1784 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1785 set_task_cpu(p
, cpu
);
1786 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1788 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1789 if (likely(sched_info_on()))
1790 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1792 #if defined(CONFIG_SMP)
1795 #ifdef CONFIG_PREEMPT_COUNT
1796 /* Want to start with kernel preemption disabled. */
1797 task_thread_info(p
)->preempt_count
= 1;
1800 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1807 * wake_up_new_task - wake up a newly created task for the first time.
1809 * This function will do some initial scheduler statistics housekeeping
1810 * that must be done for every newly created context, then puts the task
1811 * on the runqueue and wakes it.
1813 void wake_up_new_task(struct task_struct
*p
)
1815 unsigned long flags
;
1818 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1821 * Fork balancing, do it here and not earlier because:
1822 * - cpus_allowed can change in the fork path
1823 * - any previously selected cpu might disappear through hotplug
1825 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1828 rq
= __task_rq_lock(p
);
1829 activate_task(rq
, p
, 0);
1831 trace_sched_wakeup_new(p
, true);
1832 check_preempt_curr(rq
, p
, WF_FORK
);
1834 if (p
->sched_class
->task_woken
)
1835 p
->sched_class
->task_woken(rq
, p
);
1837 task_rq_unlock(rq
, p
, &flags
);
1840 #ifdef CONFIG_PREEMPT_NOTIFIERS
1843 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1844 * @notifier: notifier struct to register
1846 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1848 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1850 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1853 * preempt_notifier_unregister - no longer interested in preemption notifications
1854 * @notifier: notifier struct to unregister
1856 * This is safe to call from within a preemption notifier.
1858 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1860 hlist_del(¬ifier
->link
);
1862 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1864 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1866 struct preempt_notifier
*notifier
;
1867 struct hlist_node
*node
;
1869 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1870 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1874 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1875 struct task_struct
*next
)
1877 struct preempt_notifier
*notifier
;
1878 struct hlist_node
*node
;
1880 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1881 notifier
->ops
->sched_out(notifier
, next
);
1884 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1886 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1891 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1892 struct task_struct
*next
)
1896 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1899 * prepare_task_switch - prepare to switch tasks
1900 * @rq: the runqueue preparing to switch
1901 * @prev: the current task that is being switched out
1902 * @next: the task we are going to switch to.
1904 * This is called with the rq lock held and interrupts off. It must
1905 * be paired with a subsequent finish_task_switch after the context
1908 * prepare_task_switch sets up locking and calls architecture specific
1912 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1913 struct task_struct
*next
)
1915 sched_info_switch(prev
, next
);
1916 perf_event_task_sched_out(prev
, next
);
1917 fire_sched_out_preempt_notifiers(prev
, next
);
1918 prepare_lock_switch(rq
, next
);
1919 prepare_arch_switch(next
);
1920 trace_sched_switch(prev
, next
);
1924 * finish_task_switch - clean up after a task-switch
1925 * @rq: runqueue associated with task-switch
1926 * @prev: the thread we just switched away from.
1928 * finish_task_switch must be called after the context switch, paired
1929 * with a prepare_task_switch call before the context switch.
1930 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1931 * and do any other architecture-specific cleanup actions.
1933 * Note that we may have delayed dropping an mm in context_switch(). If
1934 * so, we finish that here outside of the runqueue lock. (Doing it
1935 * with the lock held can cause deadlocks; see schedule() for
1938 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1939 __releases(rq
->lock
)
1941 struct mm_struct
*mm
= rq
->prev_mm
;
1947 * A task struct has one reference for the use as "current".
1948 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1949 * schedule one last time. The schedule call will never return, and
1950 * the scheduled task must drop that reference.
1951 * The test for TASK_DEAD must occur while the runqueue locks are
1952 * still held, otherwise prev could be scheduled on another cpu, die
1953 * there before we look at prev->state, and then the reference would
1955 * Manfred Spraul <manfred@colorfullife.com>
1957 prev_state
= prev
->state
;
1958 finish_arch_switch(prev
);
1959 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1960 local_irq_disable();
1961 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1962 perf_event_task_sched_in(prev
, current
);
1963 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1965 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1966 finish_lock_switch(rq
, prev
);
1968 fire_sched_in_preempt_notifiers(current
);
1971 if (unlikely(prev_state
== TASK_DEAD
)) {
1973 * Remove function-return probe instances associated with this
1974 * task and put them back on the free list.
1976 kprobe_flush_task(prev
);
1977 put_task_struct(prev
);
1983 /* assumes rq->lock is held */
1984 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1986 if (prev
->sched_class
->pre_schedule
)
1987 prev
->sched_class
->pre_schedule(rq
, prev
);
1990 /* rq->lock is NOT held, but preemption is disabled */
1991 static inline void post_schedule(struct rq
*rq
)
1993 if (rq
->post_schedule
) {
1994 unsigned long flags
;
1996 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1997 if (rq
->curr
->sched_class
->post_schedule
)
1998 rq
->curr
->sched_class
->post_schedule(rq
);
1999 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2001 rq
->post_schedule
= 0;
2007 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2011 static inline void post_schedule(struct rq
*rq
)
2018 * schedule_tail - first thing a freshly forked thread must call.
2019 * @prev: the thread we just switched away from.
2021 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2022 __releases(rq
->lock
)
2024 struct rq
*rq
= this_rq();
2026 finish_task_switch(rq
, prev
);
2029 * FIXME: do we need to worry about rq being invalidated by the
2034 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2035 /* In this case, finish_task_switch does not reenable preemption */
2038 if (current
->set_child_tid
)
2039 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2043 * context_switch - switch to the new MM and the new
2044 * thread's register state.
2047 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2048 struct task_struct
*next
)
2050 struct mm_struct
*mm
, *oldmm
;
2052 prepare_task_switch(rq
, prev
, next
);
2055 oldmm
= prev
->active_mm
;
2057 * For paravirt, this is coupled with an exit in switch_to to
2058 * combine the page table reload and the switch backend into
2061 arch_start_context_switch(prev
);
2064 next
->active_mm
= oldmm
;
2065 atomic_inc(&oldmm
->mm_count
);
2066 enter_lazy_tlb(oldmm
, next
);
2068 switch_mm(oldmm
, mm
, next
);
2071 prev
->active_mm
= NULL
;
2072 rq
->prev_mm
= oldmm
;
2075 * Since the runqueue lock will be released by the next
2076 * task (which is an invalid locking op but in the case
2077 * of the scheduler it's an obvious special-case), so we
2078 * do an early lockdep release here:
2080 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2081 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2084 /* Here we just switch the register state and the stack. */
2085 switch_to(prev
, next
, prev
);
2089 * this_rq must be evaluated again because prev may have moved
2090 * CPUs since it called schedule(), thus the 'rq' on its stack
2091 * frame will be invalid.
2093 finish_task_switch(this_rq(), prev
);
2097 * nr_running, nr_uninterruptible and nr_context_switches:
2099 * externally visible scheduler statistics: current number of runnable
2100 * threads, current number of uninterruptible-sleeping threads, total
2101 * number of context switches performed since bootup.
2103 unsigned long nr_running(void)
2105 unsigned long i
, sum
= 0;
2107 for_each_online_cpu(i
)
2108 sum
+= cpu_rq(i
)->nr_running
;
2113 unsigned long nr_uninterruptible(void)
2115 unsigned long i
, sum
= 0;
2117 for_each_possible_cpu(i
)
2118 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2121 * Since we read the counters lockless, it might be slightly
2122 * inaccurate. Do not allow it to go below zero though:
2124 if (unlikely((long)sum
< 0))
2130 unsigned long long nr_context_switches(void)
2133 unsigned long long sum
= 0;
2135 for_each_possible_cpu(i
)
2136 sum
+= cpu_rq(i
)->nr_switches
;
2141 unsigned long nr_iowait(void)
2143 unsigned long i
, sum
= 0;
2145 for_each_possible_cpu(i
)
2146 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2151 unsigned long nr_iowait_cpu(int cpu
)
2153 struct rq
*this = cpu_rq(cpu
);
2154 return atomic_read(&this->nr_iowait
);
2157 unsigned long this_cpu_load(void)
2159 struct rq
*this = this_rq();
2160 return this->cpu_load
[0];
2164 /* Variables and functions for calc_load */
2165 static atomic_long_t calc_load_tasks
;
2166 static unsigned long calc_load_update
;
2167 unsigned long avenrun
[3];
2168 EXPORT_SYMBOL(avenrun
);
2170 static long calc_load_fold_active(struct rq
*this_rq
)
2172 long nr_active
, delta
= 0;
2174 nr_active
= this_rq
->nr_running
;
2175 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2177 if (nr_active
!= this_rq
->calc_load_active
) {
2178 delta
= nr_active
- this_rq
->calc_load_active
;
2179 this_rq
->calc_load_active
= nr_active
;
2185 static unsigned long
2186 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2189 load
+= active
* (FIXED_1
- exp
);
2190 load
+= 1UL << (FSHIFT
- 1);
2191 return load
>> FSHIFT
;
2196 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2198 * When making the ILB scale, we should try to pull this in as well.
2200 static atomic_long_t calc_load_tasks_idle
;
2202 void calc_load_account_idle(struct rq
*this_rq
)
2206 delta
= calc_load_fold_active(this_rq
);
2208 atomic_long_add(delta
, &calc_load_tasks_idle
);
2211 static long calc_load_fold_idle(void)
2216 * Its got a race, we don't care...
2218 if (atomic_long_read(&calc_load_tasks_idle
))
2219 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2225 * fixed_power_int - compute: x^n, in O(log n) time
2227 * @x: base of the power
2228 * @frac_bits: fractional bits of @x
2229 * @n: power to raise @x to.
2231 * By exploiting the relation between the definition of the natural power
2232 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2233 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2234 * (where: n_i \elem {0, 1}, the binary vector representing n),
2235 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2236 * of course trivially computable in O(log_2 n), the length of our binary
2239 static unsigned long
2240 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2242 unsigned long result
= 1UL << frac_bits
;
2247 result
+= 1UL << (frac_bits
- 1);
2248 result
>>= frac_bits
;
2254 x
+= 1UL << (frac_bits
- 1);
2262 * a1 = a0 * e + a * (1 - e)
2264 * a2 = a1 * e + a * (1 - e)
2265 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2266 * = a0 * e^2 + a * (1 - e) * (1 + e)
2268 * a3 = a2 * e + a * (1 - e)
2269 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2270 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2274 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2275 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2276 * = a0 * e^n + a * (1 - e^n)
2278 * [1] application of the geometric series:
2281 * S_n := \Sum x^i = -------------
2284 static unsigned long
2285 calc_load_n(unsigned long load
, unsigned long exp
,
2286 unsigned long active
, unsigned int n
)
2289 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2293 * NO_HZ can leave us missing all per-cpu ticks calling
2294 * calc_load_account_active(), but since an idle CPU folds its delta into
2295 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2296 * in the pending idle delta if our idle period crossed a load cycle boundary.
2298 * Once we've updated the global active value, we need to apply the exponential
2299 * weights adjusted to the number of cycles missed.
2301 static void calc_global_nohz(void)
2303 long delta
, active
, n
;
2306 * If we crossed a calc_load_update boundary, make sure to fold
2307 * any pending idle changes, the respective CPUs might have
2308 * missed the tick driven calc_load_account_active() update
2311 delta
= calc_load_fold_idle();
2313 atomic_long_add(delta
, &calc_load_tasks
);
2316 * It could be the one fold was all it took, we done!
2318 if (time_before(jiffies
, calc_load_update
+ 10))
2322 * Catch-up, fold however many we are behind still
2324 delta
= jiffies
- calc_load_update
- 10;
2325 n
= 1 + (delta
/ LOAD_FREQ
);
2327 active
= atomic_long_read(&calc_load_tasks
);
2328 active
= active
> 0 ? active
* FIXED_1
: 0;
2330 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2331 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2332 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2334 calc_load_update
+= n
* LOAD_FREQ
;
2337 void calc_load_account_idle(struct rq
*this_rq
)
2341 static inline long calc_load_fold_idle(void)
2346 static void calc_global_nohz(void)
2352 * get_avenrun - get the load average array
2353 * @loads: pointer to dest load array
2354 * @offset: offset to add
2355 * @shift: shift count to shift the result left
2357 * These values are estimates at best, so no need for locking.
2359 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2361 loads
[0] = (avenrun
[0] + offset
) << shift
;
2362 loads
[1] = (avenrun
[1] + offset
) << shift
;
2363 loads
[2] = (avenrun
[2] + offset
) << shift
;
2367 * calc_load - update the avenrun load estimates 10 ticks after the
2368 * CPUs have updated calc_load_tasks.
2370 void calc_global_load(unsigned long ticks
)
2374 if (time_before(jiffies
, calc_load_update
+ 10))
2377 active
= atomic_long_read(&calc_load_tasks
);
2378 active
= active
> 0 ? active
* FIXED_1
: 0;
2380 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2381 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2382 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2384 calc_load_update
+= LOAD_FREQ
;
2387 * Account one period with whatever state we found before
2388 * folding in the nohz state and ageing the entire idle period.
2390 * This avoids loosing a sample when we go idle between
2391 * calc_load_account_active() (10 ticks ago) and now and thus
2398 * Called from update_cpu_load() to periodically update this CPU's
2401 static void calc_load_account_active(struct rq
*this_rq
)
2405 if (time_before(jiffies
, this_rq
->calc_load_update
))
2408 delta
= calc_load_fold_active(this_rq
);
2409 delta
+= calc_load_fold_idle();
2411 atomic_long_add(delta
, &calc_load_tasks
);
2413 this_rq
->calc_load_update
+= LOAD_FREQ
;
2417 * The exact cpuload at various idx values, calculated at every tick would be
2418 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2420 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2421 * on nth tick when cpu may be busy, then we have:
2422 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2423 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2425 * decay_load_missed() below does efficient calculation of
2426 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2427 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2429 * The calculation is approximated on a 128 point scale.
2430 * degrade_zero_ticks is the number of ticks after which load at any
2431 * particular idx is approximated to be zero.
2432 * degrade_factor is a precomputed table, a row for each load idx.
2433 * Each column corresponds to degradation factor for a power of two ticks,
2434 * based on 128 point scale.
2436 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2437 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2439 * With this power of 2 load factors, we can degrade the load n times
2440 * by looking at 1 bits in n and doing as many mult/shift instead of
2441 * n mult/shifts needed by the exact degradation.
2443 #define DEGRADE_SHIFT 7
2444 static const unsigned char
2445 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2446 static const unsigned char
2447 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2448 {0, 0, 0, 0, 0, 0, 0, 0},
2449 {64, 32, 8, 0, 0, 0, 0, 0},
2450 {96, 72, 40, 12, 1, 0, 0},
2451 {112, 98, 75, 43, 15, 1, 0},
2452 {120, 112, 98, 76, 45, 16, 2} };
2455 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2456 * would be when CPU is idle and so we just decay the old load without
2457 * adding any new load.
2459 static unsigned long
2460 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2464 if (!missed_updates
)
2467 if (missed_updates
>= degrade_zero_ticks
[idx
])
2471 return load
>> missed_updates
;
2473 while (missed_updates
) {
2474 if (missed_updates
% 2)
2475 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2477 missed_updates
>>= 1;
2484 * Update rq->cpu_load[] statistics. This function is usually called every
2485 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2486 * every tick. We fix it up based on jiffies.
2488 void update_cpu_load(struct rq
*this_rq
)
2490 unsigned long this_load
= this_rq
->load
.weight
;
2491 unsigned long curr_jiffies
= jiffies
;
2492 unsigned long pending_updates
;
2495 this_rq
->nr_load_updates
++;
2497 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2498 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2501 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2502 this_rq
->last_load_update_tick
= curr_jiffies
;
2504 /* Update our load: */
2505 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2506 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2507 unsigned long old_load
, new_load
;
2509 /* scale is effectively 1 << i now, and >> i divides by scale */
2511 old_load
= this_rq
->cpu_load
[i
];
2512 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2513 new_load
= this_load
;
2515 * Round up the averaging division if load is increasing. This
2516 * prevents us from getting stuck on 9 if the load is 10, for
2519 if (new_load
> old_load
)
2520 new_load
+= scale
- 1;
2522 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2525 sched_avg_update(this_rq
);
2528 static void update_cpu_load_active(struct rq
*this_rq
)
2530 update_cpu_load(this_rq
);
2532 calc_load_account_active(this_rq
);
2538 * sched_exec - execve() is a valuable balancing opportunity, because at
2539 * this point the task has the smallest effective memory and cache footprint.
2541 void sched_exec(void)
2543 struct task_struct
*p
= current
;
2544 unsigned long flags
;
2547 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2548 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2549 if (dest_cpu
== smp_processor_id())
2552 if (likely(cpu_active(dest_cpu
))) {
2553 struct migration_arg arg
= { p
, dest_cpu
};
2555 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2556 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2560 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2565 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2566 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2568 EXPORT_PER_CPU_SYMBOL(kstat
);
2569 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2572 * Return any ns on the sched_clock that have not yet been accounted in
2573 * @p in case that task is currently running.
2575 * Called with task_rq_lock() held on @rq.
2577 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2581 if (task_current(rq
, p
)) {
2582 update_rq_clock(rq
);
2583 ns
= rq
->clock_task
- p
->se
.exec_start
;
2591 unsigned long long task_delta_exec(struct task_struct
*p
)
2593 unsigned long flags
;
2597 rq
= task_rq_lock(p
, &flags
);
2598 ns
= do_task_delta_exec(p
, rq
);
2599 task_rq_unlock(rq
, p
, &flags
);
2605 * Return accounted runtime for the task.
2606 * In case the task is currently running, return the runtime plus current's
2607 * pending runtime that have not been accounted yet.
2609 unsigned long long task_sched_runtime(struct task_struct
*p
)
2611 unsigned long flags
;
2615 rq
= task_rq_lock(p
, &flags
);
2616 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2617 task_rq_unlock(rq
, p
, &flags
);
2622 #ifdef CONFIG_CGROUP_CPUACCT
2623 struct cgroup_subsys cpuacct_subsys
;
2624 struct cpuacct root_cpuacct
;
2627 static inline void task_group_account_field(struct task_struct
*p
, int index
,
2630 #ifdef CONFIG_CGROUP_CPUACCT
2631 struct kernel_cpustat
*kcpustat
;
2635 * Since all updates are sure to touch the root cgroup, we
2636 * get ourselves ahead and touch it first. If the root cgroup
2637 * is the only cgroup, then nothing else should be necessary.
2640 __get_cpu_var(kernel_cpustat
).cpustat
[index
] += tmp
;
2642 #ifdef CONFIG_CGROUP_CPUACCT
2643 if (unlikely(!cpuacct_subsys
.active
))
2648 while (ca
&& (ca
!= &root_cpuacct
)) {
2649 kcpustat
= this_cpu_ptr(ca
->cpustat
);
2650 kcpustat
->cpustat
[index
] += tmp
;
2659 * Account user cpu time to a process.
2660 * @p: the process that the cpu time gets accounted to
2661 * @cputime: the cpu time spent in user space since the last update
2662 * @cputime_scaled: cputime scaled by cpu frequency
2664 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
2665 cputime_t cputime_scaled
)
2669 /* Add user time to process. */
2670 p
->utime
+= cputime
;
2671 p
->utimescaled
+= cputime_scaled
;
2672 account_group_user_time(p
, cputime
);
2674 index
= (TASK_NICE(p
) > 0) ? CPUTIME_NICE
: CPUTIME_USER
;
2676 /* Add user time to cpustat. */
2677 task_group_account_field(p
, index
, (__force u64
) cputime
);
2679 /* Account for user time used */
2680 acct_update_integrals(p
);
2684 * Account guest cpu time to a process.
2685 * @p: the process that the cpu time gets accounted to
2686 * @cputime: the cpu time spent in virtual machine since the last update
2687 * @cputime_scaled: cputime scaled by cpu frequency
2689 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
2690 cputime_t cputime_scaled
)
2692 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2694 /* Add guest time to process. */
2695 p
->utime
+= cputime
;
2696 p
->utimescaled
+= cputime_scaled
;
2697 account_group_user_time(p
, cputime
);
2698 p
->gtime
+= cputime
;
2700 /* Add guest time to cpustat. */
2701 if (TASK_NICE(p
) > 0) {
2702 cpustat
[CPUTIME_NICE
] += (__force u64
) cputime
;
2703 cpustat
[CPUTIME_GUEST_NICE
] += (__force u64
) cputime
;
2705 cpustat
[CPUTIME_USER
] += (__force u64
) cputime
;
2706 cpustat
[CPUTIME_GUEST
] += (__force u64
) cputime
;
2711 * Account system cpu time to a process and desired cpustat field
2712 * @p: the process that the cpu time gets accounted to
2713 * @cputime: the cpu time spent in kernel space since the last update
2714 * @cputime_scaled: cputime scaled by cpu frequency
2715 * @target_cputime64: pointer to cpustat field that has to be updated
2718 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
2719 cputime_t cputime_scaled
, int index
)
2721 /* Add system time to process. */
2722 p
->stime
+= cputime
;
2723 p
->stimescaled
+= cputime_scaled
;
2724 account_group_system_time(p
, cputime
);
2726 /* Add system time to cpustat. */
2727 task_group_account_field(p
, index
, (__force u64
) cputime
);
2729 /* Account for system time used */
2730 acct_update_integrals(p
);
2734 * Account system cpu time to a process.
2735 * @p: the process that the cpu time gets accounted to
2736 * @hardirq_offset: the offset to subtract from hardirq_count()
2737 * @cputime: the cpu time spent in kernel space since the last update
2738 * @cputime_scaled: cputime scaled by cpu frequency
2740 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2741 cputime_t cputime
, cputime_t cputime_scaled
)
2745 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
2746 account_guest_time(p
, cputime
, cputime_scaled
);
2750 if (hardirq_count() - hardirq_offset
)
2751 index
= CPUTIME_IRQ
;
2752 else if (in_serving_softirq())
2753 index
= CPUTIME_SOFTIRQ
;
2755 index
= CPUTIME_SYSTEM
;
2757 __account_system_time(p
, cputime
, cputime_scaled
, index
);
2761 * Account for involuntary wait time.
2762 * @cputime: the cpu time spent in involuntary wait
2764 void account_steal_time(cputime_t cputime
)
2766 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2768 cpustat
[CPUTIME_STEAL
] += (__force u64
) cputime
;
2772 * Account for idle time.
2773 * @cputime: the cpu time spent in idle wait
2775 void account_idle_time(cputime_t cputime
)
2777 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2778 struct rq
*rq
= this_rq();
2780 if (atomic_read(&rq
->nr_iowait
) > 0)
2781 cpustat
[CPUTIME_IOWAIT
] += (__force u64
) cputime
;
2783 cpustat
[CPUTIME_IDLE
] += (__force u64
) cputime
;
2786 static __always_inline
bool steal_account_process_tick(void)
2788 #ifdef CONFIG_PARAVIRT
2789 if (static_key_false(¶virt_steal_enabled
)) {
2792 steal
= paravirt_steal_clock(smp_processor_id());
2793 steal
-= this_rq()->prev_steal_time
;
2795 st
= steal_ticks(steal
);
2796 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
2798 account_steal_time(st
);
2805 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2807 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2809 * Account a tick to a process and cpustat
2810 * @p: the process that the cpu time gets accounted to
2811 * @user_tick: is the tick from userspace
2812 * @rq: the pointer to rq
2814 * Tick demultiplexing follows the order
2815 * - pending hardirq update
2816 * - pending softirq update
2820 * - check for guest_time
2821 * - else account as system_time
2823 * Check for hardirq is done both for system and user time as there is
2824 * no timer going off while we are on hardirq and hence we may never get an
2825 * opportunity to update it solely in system time.
2826 * p->stime and friends are only updated on system time and not on irq
2827 * softirq as those do not count in task exec_runtime any more.
2829 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2832 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2833 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2835 if (steal_account_process_tick())
2838 if (irqtime_account_hi_update()) {
2839 cpustat
[CPUTIME_IRQ
] += (__force u64
) cputime_one_jiffy
;
2840 } else if (irqtime_account_si_update()) {
2841 cpustat
[CPUTIME_SOFTIRQ
] += (__force u64
) cputime_one_jiffy
;
2842 } else if (this_cpu_ksoftirqd() == p
) {
2844 * ksoftirqd time do not get accounted in cpu_softirq_time.
2845 * So, we have to handle it separately here.
2846 * Also, p->stime needs to be updated for ksoftirqd.
2848 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2850 } else if (user_tick
) {
2851 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2852 } else if (p
== rq
->idle
) {
2853 account_idle_time(cputime_one_jiffy
);
2854 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
2855 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2857 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2862 static void irqtime_account_idle_ticks(int ticks
)
2865 struct rq
*rq
= this_rq();
2867 for (i
= 0; i
< ticks
; i
++)
2868 irqtime_account_process_tick(current
, 0, rq
);
2870 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2871 static void irqtime_account_idle_ticks(int ticks
) {}
2872 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2874 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2877 * Account a single tick of cpu time.
2878 * @p: the process that the cpu time gets accounted to
2879 * @user_tick: indicates if the tick is a user or a system tick
2881 void account_process_tick(struct task_struct
*p
, int user_tick
)
2883 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2884 struct rq
*rq
= this_rq();
2886 if (sched_clock_irqtime
) {
2887 irqtime_account_process_tick(p
, user_tick
, rq
);
2891 if (steal_account_process_tick())
2895 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2896 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
2897 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
2900 account_idle_time(cputime_one_jiffy
);
2904 * Account multiple ticks of steal time.
2905 * @p: the process from which the cpu time has been stolen
2906 * @ticks: number of stolen ticks
2908 void account_steal_ticks(unsigned long ticks
)
2910 account_steal_time(jiffies_to_cputime(ticks
));
2914 * Account multiple ticks of idle time.
2915 * @ticks: number of stolen ticks
2917 void account_idle_ticks(unsigned long ticks
)
2920 if (sched_clock_irqtime
) {
2921 irqtime_account_idle_ticks(ticks
);
2925 account_idle_time(jiffies_to_cputime(ticks
));
2931 * Use precise platform statistics if available:
2933 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2934 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2940 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2942 struct task_cputime cputime
;
2944 thread_group_cputime(p
, &cputime
);
2946 *ut
= cputime
.utime
;
2947 *st
= cputime
.stime
;
2951 #ifndef nsecs_to_cputime
2952 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2955 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2957 cputime_t rtime
, utime
= p
->utime
, total
= utime
+ p
->stime
;
2960 * Use CFS's precise accounting:
2962 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
2965 u64 temp
= (__force u64
) rtime
;
2967 temp
*= (__force u64
) utime
;
2968 do_div(temp
, (__force u32
) total
);
2969 utime
= (__force cputime_t
) temp
;
2974 * Compare with previous values, to keep monotonicity:
2976 p
->prev_utime
= max(p
->prev_utime
, utime
);
2977 p
->prev_stime
= max(p
->prev_stime
, rtime
- p
->prev_utime
);
2979 *ut
= p
->prev_utime
;
2980 *st
= p
->prev_stime
;
2984 * Must be called with siglock held.
2986 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2988 struct signal_struct
*sig
= p
->signal
;
2989 struct task_cputime cputime
;
2990 cputime_t rtime
, utime
, total
;
2992 thread_group_cputime(p
, &cputime
);
2994 total
= cputime
.utime
+ cputime
.stime
;
2995 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
2998 u64 temp
= (__force u64
) rtime
;
3000 temp
*= (__force u64
) cputime
.utime
;
3001 do_div(temp
, (__force u32
) total
);
3002 utime
= (__force cputime_t
) temp
;
3006 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3007 sig
->prev_stime
= max(sig
->prev_stime
, rtime
- sig
->prev_utime
);
3009 *ut
= sig
->prev_utime
;
3010 *st
= sig
->prev_stime
;
3015 * This function gets called by the timer code, with HZ frequency.
3016 * We call it with interrupts disabled.
3018 void scheduler_tick(void)
3020 int cpu
= smp_processor_id();
3021 struct rq
*rq
= cpu_rq(cpu
);
3022 struct task_struct
*curr
= rq
->curr
;
3026 raw_spin_lock(&rq
->lock
);
3027 update_rq_clock(rq
);
3028 update_cpu_load_active(rq
);
3029 curr
->sched_class
->task_tick(rq
, curr
, 0);
3030 raw_spin_unlock(&rq
->lock
);
3032 perf_event_task_tick();
3035 rq
->idle_balance
= idle_cpu(cpu
);
3036 trigger_load_balance(rq
, cpu
);
3040 notrace
unsigned long get_parent_ip(unsigned long addr
)
3042 if (in_lock_functions(addr
)) {
3043 addr
= CALLER_ADDR2
;
3044 if (in_lock_functions(addr
))
3045 addr
= CALLER_ADDR3
;
3050 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3051 defined(CONFIG_PREEMPT_TRACER))
3053 void __kprobes
add_preempt_count(int val
)
3055 #ifdef CONFIG_DEBUG_PREEMPT
3059 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3062 preempt_count() += val
;
3063 #ifdef CONFIG_DEBUG_PREEMPT
3065 * Spinlock count overflowing soon?
3067 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3070 if (preempt_count() == val
)
3071 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3073 EXPORT_SYMBOL(add_preempt_count
);
3075 void __kprobes
sub_preempt_count(int val
)
3077 #ifdef CONFIG_DEBUG_PREEMPT
3081 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3084 * Is the spinlock portion underflowing?
3086 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3087 !(preempt_count() & PREEMPT_MASK
)))
3091 if (preempt_count() == val
)
3092 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3093 preempt_count() -= val
;
3095 EXPORT_SYMBOL(sub_preempt_count
);
3100 * Print scheduling while atomic bug:
3102 static noinline
void __schedule_bug(struct task_struct
*prev
)
3104 struct pt_regs
*regs
= get_irq_regs();
3106 if (oops_in_progress
)
3109 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3110 prev
->comm
, prev
->pid
, preempt_count());
3112 debug_show_held_locks(prev
);
3114 if (irqs_disabled())
3115 print_irqtrace_events(prev
);
3124 * Various schedule()-time debugging checks and statistics:
3126 static inline void schedule_debug(struct task_struct
*prev
)
3129 * Test if we are atomic. Since do_exit() needs to call into
3130 * schedule() atomically, we ignore that path for now.
3131 * Otherwise, whine if we are scheduling when we should not be.
3133 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3134 __schedule_bug(prev
);
3137 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3139 schedstat_inc(this_rq(), sched_count
);
3142 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3144 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
3145 update_rq_clock(rq
);
3146 prev
->sched_class
->put_prev_task(rq
, prev
);
3150 * Pick up the highest-prio task:
3152 static inline struct task_struct
*
3153 pick_next_task(struct rq
*rq
)
3155 const struct sched_class
*class;
3156 struct task_struct
*p
;
3159 * Optimization: we know that if all tasks are in
3160 * the fair class we can call that function directly:
3162 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3163 p
= fair_sched_class
.pick_next_task(rq
);
3168 for_each_class(class) {
3169 p
= class->pick_next_task(rq
);
3174 BUG(); /* the idle class will always have a runnable task */
3178 * __schedule() is the main scheduler function.
3180 static void __sched
__schedule(void)
3182 struct task_struct
*prev
, *next
;
3183 unsigned long *switch_count
;
3189 cpu
= smp_processor_id();
3191 rcu_note_context_switch(cpu
);
3194 schedule_debug(prev
);
3196 if (sched_feat(HRTICK
))
3199 raw_spin_lock_irq(&rq
->lock
);
3201 switch_count
= &prev
->nivcsw
;
3202 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3203 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3204 prev
->state
= TASK_RUNNING
;
3206 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3210 * If a worker went to sleep, notify and ask workqueue
3211 * whether it wants to wake up a task to maintain
3214 if (prev
->flags
& PF_WQ_WORKER
) {
3215 struct task_struct
*to_wakeup
;
3217 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3219 try_to_wake_up_local(to_wakeup
);
3222 switch_count
= &prev
->nvcsw
;
3225 pre_schedule(rq
, prev
);
3227 if (unlikely(!rq
->nr_running
))
3228 idle_balance(cpu
, rq
);
3230 put_prev_task(rq
, prev
);
3231 next
= pick_next_task(rq
);
3232 clear_tsk_need_resched(prev
);
3233 rq
->skip_clock_update
= 0;
3235 if (likely(prev
!= next
)) {
3240 context_switch(rq
, prev
, next
); /* unlocks the rq */
3242 * The context switch have flipped the stack from under us
3243 * and restored the local variables which were saved when
3244 * this task called schedule() in the past. prev == current
3245 * is still correct, but it can be moved to another cpu/rq.
3247 cpu
= smp_processor_id();
3250 raw_spin_unlock_irq(&rq
->lock
);
3254 sched_preempt_enable_no_resched();
3259 static inline void sched_submit_work(struct task_struct
*tsk
)
3261 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3264 * If we are going to sleep and we have plugged IO queued,
3265 * make sure to submit it to avoid deadlocks.
3267 if (blk_needs_flush_plug(tsk
))
3268 blk_schedule_flush_plug(tsk
);
3271 asmlinkage
void __sched
schedule(void)
3273 struct task_struct
*tsk
= current
;
3275 sched_submit_work(tsk
);
3278 EXPORT_SYMBOL(schedule
);
3281 * schedule_preempt_disabled - called with preemption disabled
3283 * Returns with preemption disabled. Note: preempt_count must be 1
3285 void __sched
schedule_preempt_disabled(void)
3287 sched_preempt_enable_no_resched();
3292 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3294 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3296 if (lock
->owner
!= owner
)
3300 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3301 * lock->owner still matches owner, if that fails, owner might
3302 * point to free()d memory, if it still matches, the rcu_read_lock()
3303 * ensures the memory stays valid.
3307 return owner
->on_cpu
;
3311 * Look out! "owner" is an entirely speculative pointer
3312 * access and not reliable.
3314 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3316 if (!sched_feat(OWNER_SPIN
))
3320 while (owner_running(lock
, owner
)) {
3324 arch_mutex_cpu_relax();
3329 * We break out the loop above on need_resched() and when the
3330 * owner changed, which is a sign for heavy contention. Return
3331 * success only when lock->owner is NULL.
3333 return lock
->owner
== NULL
;
3337 #ifdef CONFIG_PREEMPT
3339 * this is the entry point to schedule() from in-kernel preemption
3340 * off of preempt_enable. Kernel preemptions off return from interrupt
3341 * occur there and call schedule directly.
3343 asmlinkage
void __sched notrace
preempt_schedule(void)
3345 struct thread_info
*ti
= current_thread_info();
3348 * If there is a non-zero preempt_count or interrupts are disabled,
3349 * we do not want to preempt the current task. Just return..
3351 if (likely(ti
->preempt_count
|| irqs_disabled()))
3355 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3357 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3360 * Check again in case we missed a preemption opportunity
3361 * between schedule and now.
3364 } while (need_resched());
3366 EXPORT_SYMBOL(preempt_schedule
);
3369 * this is the entry point to schedule() from kernel preemption
3370 * off of irq context.
3371 * Note, that this is called and return with irqs disabled. This will
3372 * protect us against recursive calling from irq.
3374 asmlinkage
void __sched
preempt_schedule_irq(void)
3376 struct thread_info
*ti
= current_thread_info();
3378 /* Catch callers which need to be fixed */
3379 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3382 add_preempt_count(PREEMPT_ACTIVE
);
3385 local_irq_disable();
3386 sub_preempt_count(PREEMPT_ACTIVE
);
3389 * Check again in case we missed a preemption opportunity
3390 * between schedule and now.
3393 } while (need_resched());
3396 #endif /* CONFIG_PREEMPT */
3398 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3401 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3403 EXPORT_SYMBOL(default_wake_function
);
3406 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3407 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3408 * number) then we wake all the non-exclusive tasks and one exclusive task.
3410 * There are circumstances in which we can try to wake a task which has already
3411 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3412 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3414 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3415 int nr_exclusive
, int wake_flags
, void *key
)
3417 wait_queue_t
*curr
, *next
;
3419 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3420 unsigned flags
= curr
->flags
;
3422 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3423 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3429 * __wake_up - wake up threads blocked on a waitqueue.
3431 * @mode: which threads
3432 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3433 * @key: is directly passed to the wakeup function
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(wait_queue_head_t
*q
, unsigned int mode
,
3439 int nr_exclusive
, void *key
)
3441 unsigned long flags
;
3443 spin_lock_irqsave(&q
->lock
, flags
);
3444 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3445 spin_unlock_irqrestore(&q
->lock
, flags
);
3447 EXPORT_SYMBOL(__wake_up
);
3450 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3452 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3454 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3456 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3458 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3460 __wake_up_common(q
, mode
, 1, 0, key
);
3462 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3465 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3467 * @mode: which threads
3468 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3469 * @key: opaque value to be passed to wakeup targets
3471 * The sync wakeup differs that the waker knows that it will schedule
3472 * away soon, so while the target thread will be woken up, it will not
3473 * be migrated to another CPU - ie. the two threads are 'synchronized'
3474 * with each other. This can prevent needless bouncing between CPUs.
3476 * On UP it can prevent extra preemption.
3478 * It may be assumed that this function implies a write memory barrier before
3479 * changing the task state if and only if any tasks are woken up.
3481 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3482 int nr_exclusive
, void *key
)
3484 unsigned long flags
;
3485 int wake_flags
= WF_SYNC
;
3490 if (unlikely(!nr_exclusive
))
3493 spin_lock_irqsave(&q
->lock
, flags
);
3494 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3495 spin_unlock_irqrestore(&q
->lock
, flags
);
3497 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3500 * __wake_up_sync - see __wake_up_sync_key()
3502 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3504 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3506 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3509 * complete: - signals a single thread waiting on this completion
3510 * @x: holds the state of this particular completion
3512 * This will wake up a single thread waiting on this completion. Threads will be
3513 * awakened in the same order in which they were queued.
3515 * See also complete_all(), wait_for_completion() and related routines.
3517 * It may be assumed that this function implies a write memory barrier before
3518 * changing the task state if and only if any tasks are woken up.
3520 void complete(struct completion
*x
)
3522 unsigned long flags
;
3524 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3526 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3527 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3529 EXPORT_SYMBOL(complete
);
3532 * complete_all: - signals all threads waiting on this completion
3533 * @x: holds the state of this particular completion
3535 * This will wake up all threads waiting on this particular completion event.
3537 * It may be assumed that this function implies a write memory barrier before
3538 * changing the task state if and only if any tasks are woken up.
3540 void complete_all(struct completion
*x
)
3542 unsigned long flags
;
3544 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3545 x
->done
+= UINT_MAX
/2;
3546 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3547 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3549 EXPORT_SYMBOL(complete_all
);
3551 static inline long __sched
3552 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3555 DECLARE_WAITQUEUE(wait
, current
);
3557 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3559 if (signal_pending_state(state
, current
)) {
3560 timeout
= -ERESTARTSYS
;
3563 __set_current_state(state
);
3564 spin_unlock_irq(&x
->wait
.lock
);
3565 timeout
= schedule_timeout(timeout
);
3566 spin_lock_irq(&x
->wait
.lock
);
3567 } while (!x
->done
&& timeout
);
3568 __remove_wait_queue(&x
->wait
, &wait
);
3573 return timeout
?: 1;
3577 wait_for_common(struct completion
*x
, long timeout
, int state
)
3581 spin_lock_irq(&x
->wait
.lock
);
3582 timeout
= do_wait_for_common(x
, timeout
, state
);
3583 spin_unlock_irq(&x
->wait
.lock
);
3588 * wait_for_completion: - waits for completion of a task
3589 * @x: holds the state of this particular completion
3591 * This waits to be signaled for completion of a specific task. It is NOT
3592 * interruptible and there is no timeout.
3594 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3595 * and interrupt capability. Also see complete().
3597 void __sched
wait_for_completion(struct completion
*x
)
3599 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3601 EXPORT_SYMBOL(wait_for_completion
);
3604 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3605 * @x: holds the state of this particular completion
3606 * @timeout: timeout value in jiffies
3608 * This waits for either a completion of a specific task to be signaled or for a
3609 * specified timeout to expire. The timeout is in jiffies. It is not
3612 * The return value is 0 if timed out, and positive (at least 1, or number of
3613 * jiffies left till timeout) if completed.
3615 unsigned long __sched
3616 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3618 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3620 EXPORT_SYMBOL(wait_for_completion_timeout
);
3623 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3624 * @x: holds the state of this particular completion
3626 * This waits for completion of a specific task to be signaled. It is
3629 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3631 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3633 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3634 if (t
== -ERESTARTSYS
)
3638 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3641 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3642 * @x: holds the state of this particular completion
3643 * @timeout: timeout value in jiffies
3645 * This waits for either a completion of a specific task to be signaled or for a
3646 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3648 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3649 * positive (at least 1, or number of jiffies left till timeout) if completed.
3652 wait_for_completion_interruptible_timeout(struct completion
*x
,
3653 unsigned long timeout
)
3655 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3657 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3660 * wait_for_completion_killable: - waits for completion of a task (killable)
3661 * @x: holds the state of this particular completion
3663 * This waits to be signaled for completion of a specific task. It can be
3664 * interrupted by a kill signal.
3666 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3668 int __sched
wait_for_completion_killable(struct completion
*x
)
3670 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3671 if (t
== -ERESTARTSYS
)
3675 EXPORT_SYMBOL(wait_for_completion_killable
);
3678 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3679 * @x: holds the state of this particular completion
3680 * @timeout: timeout value in jiffies
3682 * This waits for either a completion of a specific task to be
3683 * signaled or for a specified timeout to expire. It can be
3684 * interrupted by a kill signal. The timeout is in jiffies.
3686 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3687 * positive (at least 1, or number of jiffies left till timeout) if completed.
3690 wait_for_completion_killable_timeout(struct completion
*x
,
3691 unsigned long timeout
)
3693 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3695 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3698 * try_wait_for_completion - try to decrement a completion without blocking
3699 * @x: completion structure
3701 * Returns: 0 if a decrement cannot be done without blocking
3702 * 1 if a decrement succeeded.
3704 * If a completion is being used as a counting completion,
3705 * attempt to decrement the counter without blocking. This
3706 * enables us to avoid waiting if the resource the completion
3707 * is protecting is not available.
3709 bool try_wait_for_completion(struct completion
*x
)
3711 unsigned long flags
;
3714 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3719 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3722 EXPORT_SYMBOL(try_wait_for_completion
);
3725 * completion_done - Test to see if a completion has any waiters
3726 * @x: completion structure
3728 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3729 * 1 if there are no waiters.
3732 bool completion_done(struct completion
*x
)
3734 unsigned long flags
;
3737 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3740 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3743 EXPORT_SYMBOL(completion_done
);
3746 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3748 unsigned long flags
;
3751 init_waitqueue_entry(&wait
, current
);
3753 __set_current_state(state
);
3755 spin_lock_irqsave(&q
->lock
, flags
);
3756 __add_wait_queue(q
, &wait
);
3757 spin_unlock(&q
->lock
);
3758 timeout
= schedule_timeout(timeout
);
3759 spin_lock_irq(&q
->lock
);
3760 __remove_wait_queue(q
, &wait
);
3761 spin_unlock_irqrestore(&q
->lock
, flags
);
3766 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3768 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3770 EXPORT_SYMBOL(interruptible_sleep_on
);
3773 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3775 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3777 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3779 void __sched
sleep_on(wait_queue_head_t
*q
)
3781 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3783 EXPORT_SYMBOL(sleep_on
);
3785 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3787 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3789 EXPORT_SYMBOL(sleep_on_timeout
);
3791 #ifdef CONFIG_RT_MUTEXES
3794 * rt_mutex_setprio - set the current priority of a task
3796 * @prio: prio value (kernel-internal form)
3798 * This function changes the 'effective' priority of a task. It does
3799 * not touch ->normal_prio like __setscheduler().
3801 * Used by the rt_mutex code to implement priority inheritance logic.
3803 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3805 int oldprio
, on_rq
, running
;
3807 const struct sched_class
*prev_class
;
3809 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3811 rq
= __task_rq_lock(p
);
3814 * Idle task boosting is a nono in general. There is one
3815 * exception, when PREEMPT_RT and NOHZ is active:
3817 * The idle task calls get_next_timer_interrupt() and holds
3818 * the timer wheel base->lock on the CPU and another CPU wants
3819 * to access the timer (probably to cancel it). We can safely
3820 * ignore the boosting request, as the idle CPU runs this code
3821 * with interrupts disabled and will complete the lock
3822 * protected section without being interrupted. So there is no
3823 * real need to boost.
3825 if (unlikely(p
== rq
->idle
)) {
3826 WARN_ON(p
!= rq
->curr
);
3827 WARN_ON(p
->pi_blocked_on
);
3831 trace_sched_pi_setprio(p
, prio
);
3833 prev_class
= p
->sched_class
;
3835 running
= task_current(rq
, p
);
3837 dequeue_task(rq
, p
, 0);
3839 p
->sched_class
->put_prev_task(rq
, p
);
3842 p
->sched_class
= &rt_sched_class
;
3844 p
->sched_class
= &fair_sched_class
;
3849 p
->sched_class
->set_curr_task(rq
);
3851 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3853 check_class_changed(rq
, p
, prev_class
, oldprio
);
3855 __task_rq_unlock(rq
);
3858 void set_user_nice(struct task_struct
*p
, long nice
)
3860 int old_prio
, delta
, on_rq
;
3861 unsigned long flags
;
3864 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3867 * We have to be careful, if called from sys_setpriority(),
3868 * the task might be in the middle of scheduling on another CPU.
3870 rq
= task_rq_lock(p
, &flags
);
3872 * The RT priorities are set via sched_setscheduler(), but we still
3873 * allow the 'normal' nice value to be set - but as expected
3874 * it wont have any effect on scheduling until the task is
3875 * SCHED_FIFO/SCHED_RR:
3877 if (task_has_rt_policy(p
)) {
3878 p
->static_prio
= NICE_TO_PRIO(nice
);
3883 dequeue_task(rq
, p
, 0);
3885 p
->static_prio
= NICE_TO_PRIO(nice
);
3888 p
->prio
= effective_prio(p
);
3889 delta
= p
->prio
- old_prio
;
3892 enqueue_task(rq
, p
, 0);
3894 * If the task increased its priority or is running and
3895 * lowered its priority, then reschedule its CPU:
3897 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3898 resched_task(rq
->curr
);
3901 task_rq_unlock(rq
, p
, &flags
);
3903 EXPORT_SYMBOL(set_user_nice
);
3906 * can_nice - check if a task can reduce its nice value
3910 int can_nice(const struct task_struct
*p
, const int nice
)
3912 /* convert nice value [19,-20] to rlimit style value [1,40] */
3913 int nice_rlim
= 20 - nice
;
3915 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3916 capable(CAP_SYS_NICE
));
3919 #ifdef __ARCH_WANT_SYS_NICE
3922 * sys_nice - change the priority of the current process.
3923 * @increment: priority increment
3925 * sys_setpriority is a more generic, but much slower function that
3926 * does similar things.
3928 SYSCALL_DEFINE1(nice
, int, increment
)
3933 * Setpriority might change our priority at the same moment.
3934 * We don't have to worry. Conceptually one call occurs first
3935 * and we have a single winner.
3937 if (increment
< -40)
3942 nice
= TASK_NICE(current
) + increment
;
3948 if (increment
< 0 && !can_nice(current
, nice
))
3951 retval
= security_task_setnice(current
, nice
);
3955 set_user_nice(current
, nice
);
3962 * task_prio - return the priority value of a given task.
3963 * @p: the task in question.
3965 * This is the priority value as seen by users in /proc.
3966 * RT tasks are offset by -200. Normal tasks are centered
3967 * around 0, value goes from -16 to +15.
3969 int task_prio(const struct task_struct
*p
)
3971 return p
->prio
- MAX_RT_PRIO
;
3975 * task_nice - return the nice value of a given task.
3976 * @p: the task in question.
3978 int task_nice(const struct task_struct
*p
)
3980 return TASK_NICE(p
);
3982 EXPORT_SYMBOL(task_nice
);
3985 * idle_cpu - is a given cpu idle currently?
3986 * @cpu: the processor in question.
3988 int idle_cpu(int cpu
)
3990 struct rq
*rq
= cpu_rq(cpu
);
3992 if (rq
->curr
!= rq
->idle
)
3999 if (!llist_empty(&rq
->wake_list
))
4007 * idle_task - return the idle task for a given cpu.
4008 * @cpu: the processor in question.
4010 struct task_struct
*idle_task(int cpu
)
4012 return cpu_rq(cpu
)->idle
;
4016 * find_process_by_pid - find a process with a matching PID value.
4017 * @pid: the pid in question.
4019 static struct task_struct
*find_process_by_pid(pid_t pid
)
4021 return pid
? find_task_by_vpid(pid
) : current
;
4024 /* Actually do priority change: must hold rq lock. */
4026 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4029 p
->rt_priority
= prio
;
4030 p
->normal_prio
= normal_prio(p
);
4031 /* we are holding p->pi_lock already */
4032 p
->prio
= rt_mutex_getprio(p
);
4033 if (rt_prio(p
->prio
))
4034 p
->sched_class
= &rt_sched_class
;
4036 p
->sched_class
= &fair_sched_class
;
4041 * check the target process has a UID that matches the current process's
4043 static bool check_same_owner(struct task_struct
*p
)
4045 const struct cred
*cred
= current_cred(), *pcred
;
4049 pcred
= __task_cred(p
);
4050 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
4051 match
= (cred
->euid
== pcred
->euid
||
4052 cred
->euid
== pcred
->uid
);
4059 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4060 const struct sched_param
*param
, bool user
)
4062 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4063 unsigned long flags
;
4064 const struct sched_class
*prev_class
;
4068 /* may grab non-irq protected spin_locks */
4069 BUG_ON(in_interrupt());
4071 /* double check policy once rq lock held */
4073 reset_on_fork
= p
->sched_reset_on_fork
;
4074 policy
= oldpolicy
= p
->policy
;
4076 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4077 policy
&= ~SCHED_RESET_ON_FORK
;
4079 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4080 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4081 policy
!= SCHED_IDLE
)
4086 * Valid priorities for SCHED_FIFO and SCHED_RR are
4087 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4088 * SCHED_BATCH and SCHED_IDLE is 0.
4090 if (param
->sched_priority
< 0 ||
4091 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4092 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4094 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4098 * Allow unprivileged RT tasks to decrease priority:
4100 if (user
&& !capable(CAP_SYS_NICE
)) {
4101 if (rt_policy(policy
)) {
4102 unsigned long rlim_rtprio
=
4103 task_rlimit(p
, RLIMIT_RTPRIO
);
4105 /* can't set/change the rt policy */
4106 if (policy
!= p
->policy
&& !rlim_rtprio
)
4109 /* can't increase priority */
4110 if (param
->sched_priority
> p
->rt_priority
&&
4111 param
->sched_priority
> rlim_rtprio
)
4116 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4117 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4119 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
4120 if (!can_nice(p
, TASK_NICE(p
)))
4124 /* can't change other user's priorities */
4125 if (!check_same_owner(p
))
4128 /* Normal users shall not reset the sched_reset_on_fork flag */
4129 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4134 retval
= security_task_setscheduler(p
);
4140 * make sure no PI-waiters arrive (or leave) while we are
4141 * changing the priority of the task:
4143 * To be able to change p->policy safely, the appropriate
4144 * runqueue lock must be held.
4146 rq
= task_rq_lock(p
, &flags
);
4149 * Changing the policy of the stop threads its a very bad idea
4151 if (p
== rq
->stop
) {
4152 task_rq_unlock(rq
, p
, &flags
);
4157 * If not changing anything there's no need to proceed further:
4159 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
4160 param
->sched_priority
== p
->rt_priority
))) {
4162 __task_rq_unlock(rq
);
4163 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4167 #ifdef CONFIG_RT_GROUP_SCHED
4170 * Do not allow realtime tasks into groups that have no runtime
4173 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4174 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4175 !task_group_is_autogroup(task_group(p
))) {
4176 task_rq_unlock(rq
, p
, &flags
);
4182 /* recheck policy now with rq lock held */
4183 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4184 policy
= oldpolicy
= -1;
4185 task_rq_unlock(rq
, p
, &flags
);
4189 running
= task_current(rq
, p
);
4191 dequeue_task(rq
, p
, 0);
4193 p
->sched_class
->put_prev_task(rq
, p
);
4195 p
->sched_reset_on_fork
= reset_on_fork
;
4198 prev_class
= p
->sched_class
;
4199 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4202 p
->sched_class
->set_curr_task(rq
);
4204 enqueue_task(rq
, p
, 0);
4206 check_class_changed(rq
, p
, prev_class
, oldprio
);
4207 task_rq_unlock(rq
, p
, &flags
);
4209 rt_mutex_adjust_pi(p
);
4215 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4216 * @p: the task in question.
4217 * @policy: new policy.
4218 * @param: structure containing the new RT priority.
4220 * NOTE that the task may be already dead.
4222 int sched_setscheduler(struct task_struct
*p
, int policy
,
4223 const struct sched_param
*param
)
4225 return __sched_setscheduler(p
, policy
, param
, true);
4227 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4230 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4231 * @p: the task in question.
4232 * @policy: new policy.
4233 * @param: structure containing the new RT priority.
4235 * Just like sched_setscheduler, only don't bother checking if the
4236 * current context has permission. For example, this is needed in
4237 * stop_machine(): we create temporary high priority worker threads,
4238 * but our caller might not have that capability.
4240 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4241 const struct sched_param
*param
)
4243 return __sched_setscheduler(p
, policy
, param
, false);
4247 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4249 struct sched_param lparam
;
4250 struct task_struct
*p
;
4253 if (!param
|| pid
< 0)
4255 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4260 p
= find_process_by_pid(pid
);
4262 retval
= sched_setscheduler(p
, policy
, &lparam
);
4269 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4270 * @pid: the pid in question.
4271 * @policy: new policy.
4272 * @param: structure containing the new RT priority.
4274 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4275 struct sched_param __user
*, param
)
4277 /* negative values for policy are not valid */
4281 return do_sched_setscheduler(pid
, policy
, param
);
4285 * sys_sched_setparam - set/change the RT priority of a thread
4286 * @pid: the pid in question.
4287 * @param: structure containing the new RT priority.
4289 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4291 return do_sched_setscheduler(pid
, -1, param
);
4295 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4296 * @pid: the pid in question.
4298 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4300 struct task_struct
*p
;
4308 p
= find_process_by_pid(pid
);
4310 retval
= security_task_getscheduler(p
);
4313 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4320 * sys_sched_getparam - get the RT priority of a thread
4321 * @pid: the pid in question.
4322 * @param: structure containing the RT priority.
4324 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4326 struct sched_param lp
;
4327 struct task_struct
*p
;
4330 if (!param
|| pid
< 0)
4334 p
= find_process_by_pid(pid
);
4339 retval
= security_task_getscheduler(p
);
4343 lp
.sched_priority
= p
->rt_priority
;
4347 * This one might sleep, we cannot do it with a spinlock held ...
4349 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4358 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4360 cpumask_var_t cpus_allowed
, new_mask
;
4361 struct task_struct
*p
;
4367 p
= find_process_by_pid(pid
);
4374 /* Prevent p going away */
4378 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4382 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4384 goto out_free_cpus_allowed
;
4387 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4390 retval
= security_task_setscheduler(p
);
4394 cpuset_cpus_allowed(p
, cpus_allowed
);
4395 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4397 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4400 cpuset_cpus_allowed(p
, cpus_allowed
);
4401 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4403 * We must have raced with a concurrent cpuset
4404 * update. Just reset the cpus_allowed to the
4405 * cpuset's cpus_allowed
4407 cpumask_copy(new_mask
, cpus_allowed
);
4412 free_cpumask_var(new_mask
);
4413 out_free_cpus_allowed
:
4414 free_cpumask_var(cpus_allowed
);
4421 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4422 struct cpumask
*new_mask
)
4424 if (len
< cpumask_size())
4425 cpumask_clear(new_mask
);
4426 else if (len
> cpumask_size())
4427 len
= cpumask_size();
4429 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4433 * sys_sched_setaffinity - set the cpu affinity of a process
4434 * @pid: pid of the process
4435 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4436 * @user_mask_ptr: user-space pointer to the new cpu mask
4438 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4439 unsigned long __user
*, user_mask_ptr
)
4441 cpumask_var_t new_mask
;
4444 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4447 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4449 retval
= sched_setaffinity(pid
, new_mask
);
4450 free_cpumask_var(new_mask
);
4454 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4456 struct task_struct
*p
;
4457 unsigned long flags
;
4464 p
= find_process_by_pid(pid
);
4468 retval
= security_task_getscheduler(p
);
4472 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4473 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4474 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4484 * sys_sched_getaffinity - get the cpu affinity of a process
4485 * @pid: pid of the process
4486 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4487 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4489 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4490 unsigned long __user
*, user_mask_ptr
)
4495 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4497 if (len
& (sizeof(unsigned long)-1))
4500 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4503 ret
= sched_getaffinity(pid
, mask
);
4505 size_t retlen
= min_t(size_t, len
, cpumask_size());
4507 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4512 free_cpumask_var(mask
);
4518 * sys_sched_yield - yield the current processor to other threads.
4520 * This function yields the current CPU to other tasks. If there are no
4521 * other threads running on this CPU then this function will return.
4523 SYSCALL_DEFINE0(sched_yield
)
4525 struct rq
*rq
= this_rq_lock();
4527 schedstat_inc(rq
, yld_count
);
4528 current
->sched_class
->yield_task(rq
);
4531 * Since we are going to call schedule() anyway, there's
4532 * no need to preempt or enable interrupts:
4534 __release(rq
->lock
);
4535 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4536 do_raw_spin_unlock(&rq
->lock
);
4537 sched_preempt_enable_no_resched();
4544 static inline int should_resched(void)
4546 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4549 static void __cond_resched(void)
4551 add_preempt_count(PREEMPT_ACTIVE
);
4553 sub_preempt_count(PREEMPT_ACTIVE
);
4556 int __sched
_cond_resched(void)
4558 if (should_resched()) {
4564 EXPORT_SYMBOL(_cond_resched
);
4567 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4568 * call schedule, and on return reacquire the lock.
4570 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4571 * operations here to prevent schedule() from being called twice (once via
4572 * spin_unlock(), once by hand).
4574 int __cond_resched_lock(spinlock_t
*lock
)
4576 int resched
= should_resched();
4579 lockdep_assert_held(lock
);
4581 if (spin_needbreak(lock
) || resched
) {
4592 EXPORT_SYMBOL(__cond_resched_lock
);
4594 int __sched
__cond_resched_softirq(void)
4596 BUG_ON(!in_softirq());
4598 if (should_resched()) {
4606 EXPORT_SYMBOL(__cond_resched_softirq
);
4609 * yield - yield the current processor to other threads.
4611 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4613 * The scheduler is at all times free to pick the calling task as the most
4614 * eligible task to run, if removing the yield() call from your code breaks
4615 * it, its already broken.
4617 * Typical broken usage is:
4622 * where one assumes that yield() will let 'the other' process run that will
4623 * make event true. If the current task is a SCHED_FIFO task that will never
4624 * happen. Never use yield() as a progress guarantee!!
4626 * If you want to use yield() to wait for something, use wait_event().
4627 * If you want to use yield() to be 'nice' for others, use cond_resched().
4628 * If you still want to use yield(), do not!
4630 void __sched
yield(void)
4632 set_current_state(TASK_RUNNING
);
4635 EXPORT_SYMBOL(yield
);
4638 * yield_to - yield the current processor to another thread in
4639 * your thread group, or accelerate that thread toward the
4640 * processor it's on.
4642 * @preempt: whether task preemption is allowed or not
4644 * It's the caller's job to ensure that the target task struct
4645 * can't go away on us before we can do any checks.
4647 * Returns true if we indeed boosted the target task.
4649 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4651 struct task_struct
*curr
= current
;
4652 struct rq
*rq
, *p_rq
;
4653 unsigned long flags
;
4656 local_irq_save(flags
);
4661 double_rq_lock(rq
, p_rq
);
4662 while (task_rq(p
) != p_rq
) {
4663 double_rq_unlock(rq
, p_rq
);
4667 if (!curr
->sched_class
->yield_to_task
)
4670 if (curr
->sched_class
!= p
->sched_class
)
4673 if (task_running(p_rq
, p
) || p
->state
)
4676 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4678 schedstat_inc(rq
, yld_count
);
4680 * Make p's CPU reschedule; pick_next_entity takes care of
4683 if (preempt
&& rq
!= p_rq
)
4684 resched_task(p_rq
->curr
);
4687 * We might have set it in task_yield_fair(), but are
4688 * not going to schedule(), so don't want to skip
4691 rq
->skip_clock_update
= 0;
4695 double_rq_unlock(rq
, p_rq
);
4696 local_irq_restore(flags
);
4703 EXPORT_SYMBOL_GPL(yield_to
);
4706 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4707 * that process accounting knows that this is a task in IO wait state.
4709 void __sched
io_schedule(void)
4711 struct rq
*rq
= raw_rq();
4713 delayacct_blkio_start();
4714 atomic_inc(&rq
->nr_iowait
);
4715 blk_flush_plug(current
);
4716 current
->in_iowait
= 1;
4718 current
->in_iowait
= 0;
4719 atomic_dec(&rq
->nr_iowait
);
4720 delayacct_blkio_end();
4722 EXPORT_SYMBOL(io_schedule
);
4724 long __sched
io_schedule_timeout(long timeout
)
4726 struct rq
*rq
= raw_rq();
4729 delayacct_blkio_start();
4730 atomic_inc(&rq
->nr_iowait
);
4731 blk_flush_plug(current
);
4732 current
->in_iowait
= 1;
4733 ret
= schedule_timeout(timeout
);
4734 current
->in_iowait
= 0;
4735 atomic_dec(&rq
->nr_iowait
);
4736 delayacct_blkio_end();
4741 * sys_sched_get_priority_max - return maximum RT priority.
4742 * @policy: scheduling class.
4744 * this syscall returns the maximum rt_priority that can be used
4745 * by a given scheduling class.
4747 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4754 ret
= MAX_USER_RT_PRIO
-1;
4766 * sys_sched_get_priority_min - return minimum RT priority.
4767 * @policy: scheduling class.
4769 * this syscall returns the minimum rt_priority that can be used
4770 * by a given scheduling class.
4772 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4790 * sys_sched_rr_get_interval - return the default timeslice of a process.
4791 * @pid: pid of the process.
4792 * @interval: userspace pointer to the timeslice value.
4794 * this syscall writes the default timeslice value of a given process
4795 * into the user-space timespec buffer. A value of '0' means infinity.
4797 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4798 struct timespec __user
*, interval
)
4800 struct task_struct
*p
;
4801 unsigned int time_slice
;
4802 unsigned long flags
;
4812 p
= find_process_by_pid(pid
);
4816 retval
= security_task_getscheduler(p
);
4820 rq
= task_rq_lock(p
, &flags
);
4821 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4822 task_rq_unlock(rq
, p
, &flags
);
4825 jiffies_to_timespec(time_slice
, &t
);
4826 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4834 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4836 void sched_show_task(struct task_struct
*p
)
4838 unsigned long free
= 0;
4841 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4842 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4843 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4844 #if BITS_PER_LONG == 32
4845 if (state
== TASK_RUNNING
)
4846 printk(KERN_CONT
" running ");
4848 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4850 if (state
== TASK_RUNNING
)
4851 printk(KERN_CONT
" running task ");
4853 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4855 #ifdef CONFIG_DEBUG_STACK_USAGE
4856 free
= stack_not_used(p
);
4858 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4859 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4860 (unsigned long)task_thread_info(p
)->flags
);
4862 show_stack(p
, NULL
);
4865 void show_state_filter(unsigned long state_filter
)
4867 struct task_struct
*g
, *p
;
4869 #if BITS_PER_LONG == 32
4871 " task PC stack pid father\n");
4874 " task PC stack pid father\n");
4877 do_each_thread(g
, p
) {
4879 * reset the NMI-timeout, listing all files on a slow
4880 * console might take a lot of time:
4882 touch_nmi_watchdog();
4883 if (!state_filter
|| (p
->state
& state_filter
))
4885 } while_each_thread(g
, p
);
4887 touch_all_softlockup_watchdogs();
4889 #ifdef CONFIG_SCHED_DEBUG
4890 sysrq_sched_debug_show();
4894 * Only show locks if all tasks are dumped:
4897 debug_show_all_locks();
4900 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4902 idle
->sched_class
= &idle_sched_class
;
4906 * init_idle - set up an idle thread for a given CPU
4907 * @idle: task in question
4908 * @cpu: cpu the idle task belongs to
4910 * NOTE: this function does not set the idle thread's NEED_RESCHED
4911 * flag, to make booting more robust.
4913 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4915 struct rq
*rq
= cpu_rq(cpu
);
4916 unsigned long flags
;
4918 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4921 idle
->state
= TASK_RUNNING
;
4922 idle
->se
.exec_start
= sched_clock();
4924 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4926 * We're having a chicken and egg problem, even though we are
4927 * holding rq->lock, the cpu isn't yet set to this cpu so the
4928 * lockdep check in task_group() will fail.
4930 * Similar case to sched_fork(). / Alternatively we could
4931 * use task_rq_lock() here and obtain the other rq->lock.
4936 __set_task_cpu(idle
, cpu
);
4939 rq
->curr
= rq
->idle
= idle
;
4940 #if defined(CONFIG_SMP)
4943 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4945 /* Set the preempt count _outside_ the spinlocks! */
4946 task_thread_info(idle
)->preempt_count
= 0;
4949 * The idle tasks have their own, simple scheduling class:
4951 idle
->sched_class
= &idle_sched_class
;
4952 ftrace_graph_init_idle_task(idle
, cpu
);
4953 #if defined(CONFIG_SMP)
4954 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4959 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4961 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4962 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4964 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4965 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
4969 * This is how migration works:
4971 * 1) we invoke migration_cpu_stop() on the target CPU using
4973 * 2) stopper starts to run (implicitly forcing the migrated thread
4975 * 3) it checks whether the migrated task is still in the wrong runqueue.
4976 * 4) if it's in the wrong runqueue then the migration thread removes
4977 * it and puts it into the right queue.
4978 * 5) stopper completes and stop_one_cpu() returns and the migration
4983 * Change a given task's CPU affinity. Migrate the thread to a
4984 * proper CPU and schedule it away if the CPU it's executing on
4985 * is removed from the allowed bitmask.
4987 * NOTE: the caller must have a valid reference to the task, the
4988 * task must not exit() & deallocate itself prematurely. The
4989 * call is not atomic; no spinlocks may be held.
4991 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4993 unsigned long flags
;
4995 unsigned int dest_cpu
;
4998 rq
= task_rq_lock(p
, &flags
);
5000 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
5003 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5008 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
5013 do_set_cpus_allowed(p
, new_mask
);
5015 /* Can the task run on the task's current CPU? If so, we're done */
5016 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5019 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5021 struct migration_arg arg
= { p
, dest_cpu
};
5022 /* Need help from migration thread: drop lock and wait. */
5023 task_rq_unlock(rq
, p
, &flags
);
5024 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5025 tlb_migrate_finish(p
->mm
);
5029 task_rq_unlock(rq
, p
, &flags
);
5033 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5036 * Move (not current) task off this cpu, onto dest cpu. We're doing
5037 * this because either it can't run here any more (set_cpus_allowed()
5038 * away from this CPU, or CPU going down), or because we're
5039 * attempting to rebalance this task on exec (sched_exec).
5041 * So we race with normal scheduler movements, but that's OK, as long
5042 * as the task is no longer on this CPU.
5044 * Returns non-zero if task was successfully migrated.
5046 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5048 struct rq
*rq_dest
, *rq_src
;
5051 if (unlikely(!cpu_active(dest_cpu
)))
5054 rq_src
= cpu_rq(src_cpu
);
5055 rq_dest
= cpu_rq(dest_cpu
);
5057 raw_spin_lock(&p
->pi_lock
);
5058 double_rq_lock(rq_src
, rq_dest
);
5059 /* Already moved. */
5060 if (task_cpu(p
) != src_cpu
)
5062 /* Affinity changed (again). */
5063 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
5067 * If we're not on a rq, the next wake-up will ensure we're
5071 dequeue_task(rq_src
, p
, 0);
5072 set_task_cpu(p
, dest_cpu
);
5073 enqueue_task(rq_dest
, p
, 0);
5074 check_preempt_curr(rq_dest
, p
, 0);
5079 double_rq_unlock(rq_src
, rq_dest
);
5080 raw_spin_unlock(&p
->pi_lock
);
5085 * migration_cpu_stop - this will be executed by a highprio stopper thread
5086 * and performs thread migration by bumping thread off CPU then
5087 * 'pushing' onto another runqueue.
5089 static int migration_cpu_stop(void *data
)
5091 struct migration_arg
*arg
= data
;
5094 * The original target cpu might have gone down and we might
5095 * be on another cpu but it doesn't matter.
5097 local_irq_disable();
5098 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5103 #ifdef CONFIG_HOTPLUG_CPU
5106 * Ensures that the idle task is using init_mm right before its cpu goes
5109 void idle_task_exit(void)
5111 struct mm_struct
*mm
= current
->active_mm
;
5113 BUG_ON(cpu_online(smp_processor_id()));
5116 switch_mm(mm
, &init_mm
, current
);
5121 * While a dead CPU has no uninterruptible tasks queued at this point,
5122 * it might still have a nonzero ->nr_uninterruptible counter, because
5123 * for performance reasons the counter is not stricly tracking tasks to
5124 * their home CPUs. So we just add the counter to another CPU's counter,
5125 * to keep the global sum constant after CPU-down:
5127 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5129 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5131 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5132 rq_src
->nr_uninterruptible
= 0;
5136 * remove the tasks which were accounted by rq from calc_load_tasks.
5138 static void calc_global_load_remove(struct rq
*rq
)
5140 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5141 rq
->calc_load_active
= 0;
5145 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5146 * try_to_wake_up()->select_task_rq().
5148 * Called with rq->lock held even though we'er in stop_machine() and
5149 * there's no concurrency possible, we hold the required locks anyway
5150 * because of lock validation efforts.
5152 static void migrate_tasks(unsigned int dead_cpu
)
5154 struct rq
*rq
= cpu_rq(dead_cpu
);
5155 struct task_struct
*next
, *stop
= rq
->stop
;
5159 * Fudge the rq selection such that the below task selection loop
5160 * doesn't get stuck on the currently eligible stop task.
5162 * We're currently inside stop_machine() and the rq is either stuck
5163 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5164 * either way we should never end up calling schedule() until we're
5169 /* Ensure any throttled groups are reachable by pick_next_task */
5170 unthrottle_offline_cfs_rqs(rq
);
5174 * There's this thread running, bail when that's the only
5177 if (rq
->nr_running
== 1)
5180 next
= pick_next_task(rq
);
5182 next
->sched_class
->put_prev_task(rq
, next
);
5184 /* Find suitable destination for @next, with force if needed. */
5185 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5186 raw_spin_unlock(&rq
->lock
);
5188 __migrate_task(next
, dead_cpu
, dest_cpu
);
5190 raw_spin_lock(&rq
->lock
);
5196 #endif /* CONFIG_HOTPLUG_CPU */
5198 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5200 static struct ctl_table sd_ctl_dir
[] = {
5202 .procname
= "sched_domain",
5208 static struct ctl_table sd_ctl_root
[] = {
5210 .procname
= "kernel",
5212 .child
= sd_ctl_dir
,
5217 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5219 struct ctl_table
*entry
=
5220 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5225 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5227 struct ctl_table
*entry
;
5230 * In the intermediate directories, both the child directory and
5231 * procname are dynamically allocated and could fail but the mode
5232 * will always be set. In the lowest directory the names are
5233 * static strings and all have proc handlers.
5235 for (entry
= *tablep
; entry
->mode
; entry
++) {
5237 sd_free_ctl_entry(&entry
->child
);
5238 if (entry
->proc_handler
== NULL
)
5239 kfree(entry
->procname
);
5247 set_table_entry(struct ctl_table
*entry
,
5248 const char *procname
, void *data
, int maxlen
,
5249 umode_t mode
, proc_handler
*proc_handler
)
5251 entry
->procname
= procname
;
5253 entry
->maxlen
= maxlen
;
5255 entry
->proc_handler
= proc_handler
;
5258 static struct ctl_table
*
5259 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5261 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5266 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5267 sizeof(long), 0644, proc_doulongvec_minmax
);
5268 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5269 sizeof(long), 0644, proc_doulongvec_minmax
);
5270 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5271 sizeof(int), 0644, proc_dointvec_minmax
);
5272 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5273 sizeof(int), 0644, proc_dointvec_minmax
);
5274 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5275 sizeof(int), 0644, proc_dointvec_minmax
);
5276 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5277 sizeof(int), 0644, proc_dointvec_minmax
);
5278 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5279 sizeof(int), 0644, proc_dointvec_minmax
);
5280 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5281 sizeof(int), 0644, proc_dointvec_minmax
);
5282 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5283 sizeof(int), 0644, proc_dointvec_minmax
);
5284 set_table_entry(&table
[9], "cache_nice_tries",
5285 &sd
->cache_nice_tries
,
5286 sizeof(int), 0644, proc_dointvec_minmax
);
5287 set_table_entry(&table
[10], "flags", &sd
->flags
,
5288 sizeof(int), 0644, proc_dointvec_minmax
);
5289 set_table_entry(&table
[11], "name", sd
->name
,
5290 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5291 /* &table[12] is terminator */
5296 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5298 struct ctl_table
*entry
, *table
;
5299 struct sched_domain
*sd
;
5300 int domain_num
= 0, i
;
5303 for_each_domain(cpu
, sd
)
5305 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5310 for_each_domain(cpu
, sd
) {
5311 snprintf(buf
, 32, "domain%d", i
);
5312 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5314 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5321 static struct ctl_table_header
*sd_sysctl_header
;
5322 static void register_sched_domain_sysctl(void)
5324 int i
, cpu_num
= num_possible_cpus();
5325 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5328 WARN_ON(sd_ctl_dir
[0].child
);
5329 sd_ctl_dir
[0].child
= entry
;
5334 for_each_possible_cpu(i
) {
5335 snprintf(buf
, 32, "cpu%d", i
);
5336 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5338 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5342 WARN_ON(sd_sysctl_header
);
5343 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5346 /* may be called multiple times per register */
5347 static void unregister_sched_domain_sysctl(void)
5349 if (sd_sysctl_header
)
5350 unregister_sysctl_table(sd_sysctl_header
);
5351 sd_sysctl_header
= NULL
;
5352 if (sd_ctl_dir
[0].child
)
5353 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5356 static void register_sched_domain_sysctl(void)
5359 static void unregister_sched_domain_sysctl(void)
5364 static void set_rq_online(struct rq
*rq
)
5367 const struct sched_class
*class;
5369 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5372 for_each_class(class) {
5373 if (class->rq_online
)
5374 class->rq_online(rq
);
5379 static void set_rq_offline(struct rq
*rq
)
5382 const struct sched_class
*class;
5384 for_each_class(class) {
5385 if (class->rq_offline
)
5386 class->rq_offline(rq
);
5389 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5395 * migration_call - callback that gets triggered when a CPU is added.
5396 * Here we can start up the necessary migration thread for the new CPU.
5398 static int __cpuinit
5399 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5401 int cpu
= (long)hcpu
;
5402 unsigned long flags
;
5403 struct rq
*rq
= cpu_rq(cpu
);
5405 switch (action
& ~CPU_TASKS_FROZEN
) {
5407 case CPU_UP_PREPARE
:
5408 rq
->calc_load_update
= calc_load_update
;
5412 /* Update our root-domain */
5413 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5415 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5419 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5422 #ifdef CONFIG_HOTPLUG_CPU
5424 sched_ttwu_pending();
5425 /* Update our root-domain */
5426 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5428 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5432 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5433 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5435 migrate_nr_uninterruptible(rq
);
5436 calc_global_load_remove(rq
);
5441 update_max_interval();
5447 * Register at high priority so that task migration (migrate_all_tasks)
5448 * happens before everything else. This has to be lower priority than
5449 * the notifier in the perf_event subsystem, though.
5451 static struct notifier_block __cpuinitdata migration_notifier
= {
5452 .notifier_call
= migration_call
,
5453 .priority
= CPU_PRI_MIGRATION
,
5456 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5457 unsigned long action
, void *hcpu
)
5459 switch (action
& ~CPU_TASKS_FROZEN
) {
5461 case CPU_DOWN_FAILED
:
5462 set_cpu_active((long)hcpu
, true);
5469 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5470 unsigned long action
, void *hcpu
)
5472 switch (action
& ~CPU_TASKS_FROZEN
) {
5473 case CPU_DOWN_PREPARE
:
5474 set_cpu_active((long)hcpu
, false);
5481 static int __init
migration_init(void)
5483 void *cpu
= (void *)(long)smp_processor_id();
5486 /* Initialize migration for the boot CPU */
5487 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5488 BUG_ON(err
== NOTIFY_BAD
);
5489 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5490 register_cpu_notifier(&migration_notifier
);
5492 /* Register cpu active notifiers */
5493 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5494 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5498 early_initcall(migration_init
);
5503 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5505 #ifdef CONFIG_SCHED_DEBUG
5507 static __read_mostly
int sched_domain_debug_enabled
;
5509 static int __init
sched_domain_debug_setup(char *str
)
5511 sched_domain_debug_enabled
= 1;
5515 early_param("sched_debug", sched_domain_debug_setup
);
5517 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5518 struct cpumask
*groupmask
)
5520 struct sched_group
*group
= sd
->groups
;
5523 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5524 cpumask_clear(groupmask
);
5526 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5528 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5529 printk("does not load-balance\n");
5531 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5536 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5538 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5539 printk(KERN_ERR
"ERROR: domain->span does not contain "
5542 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5543 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5547 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5551 printk(KERN_ERR
"ERROR: group is NULL\n");
5555 if (!group
->sgp
->power
) {
5556 printk(KERN_CONT
"\n");
5557 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5562 if (!cpumask_weight(sched_group_cpus(group
))) {
5563 printk(KERN_CONT
"\n");
5564 printk(KERN_ERR
"ERROR: empty group\n");
5568 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5569 printk(KERN_CONT
"\n");
5570 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5574 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5576 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5578 printk(KERN_CONT
" %s", str
);
5579 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5580 printk(KERN_CONT
" (cpu_power = %d)",
5584 group
= group
->next
;
5585 } while (group
!= sd
->groups
);
5586 printk(KERN_CONT
"\n");
5588 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5589 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5592 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5593 printk(KERN_ERR
"ERROR: parent span is not a superset "
5594 "of domain->span\n");
5598 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5602 if (!sched_domain_debug_enabled
)
5606 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5610 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5613 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5621 #else /* !CONFIG_SCHED_DEBUG */
5622 # define sched_domain_debug(sd, cpu) do { } while (0)
5623 #endif /* CONFIG_SCHED_DEBUG */
5625 static int sd_degenerate(struct sched_domain
*sd
)
5627 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5630 /* Following flags need at least 2 groups */
5631 if (sd
->flags
& (SD_LOAD_BALANCE
|
5632 SD_BALANCE_NEWIDLE
|
5636 SD_SHARE_PKG_RESOURCES
)) {
5637 if (sd
->groups
!= sd
->groups
->next
)
5641 /* Following flags don't use groups */
5642 if (sd
->flags
& (SD_WAKE_AFFINE
))
5649 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5651 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5653 if (sd_degenerate(parent
))
5656 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5659 /* Flags needing groups don't count if only 1 group in parent */
5660 if (parent
->groups
== parent
->groups
->next
) {
5661 pflags
&= ~(SD_LOAD_BALANCE
|
5662 SD_BALANCE_NEWIDLE
|
5666 SD_SHARE_PKG_RESOURCES
);
5667 if (nr_node_ids
== 1)
5668 pflags
&= ~SD_SERIALIZE
;
5670 if (~cflags
& pflags
)
5676 static void free_rootdomain(struct rcu_head
*rcu
)
5678 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5680 cpupri_cleanup(&rd
->cpupri
);
5681 free_cpumask_var(rd
->rto_mask
);
5682 free_cpumask_var(rd
->online
);
5683 free_cpumask_var(rd
->span
);
5687 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5689 struct root_domain
*old_rd
= NULL
;
5690 unsigned long flags
;
5692 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5697 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5700 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5703 * If we dont want to free the old_rt yet then
5704 * set old_rd to NULL to skip the freeing later
5707 if (!atomic_dec_and_test(&old_rd
->refcount
))
5711 atomic_inc(&rd
->refcount
);
5714 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5715 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5718 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5721 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5724 static int init_rootdomain(struct root_domain
*rd
)
5726 memset(rd
, 0, sizeof(*rd
));
5728 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5730 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5732 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5735 if (cpupri_init(&rd
->cpupri
) != 0)
5740 free_cpumask_var(rd
->rto_mask
);
5742 free_cpumask_var(rd
->online
);
5744 free_cpumask_var(rd
->span
);
5750 * By default the system creates a single root-domain with all cpus as
5751 * members (mimicking the global state we have today).
5753 struct root_domain def_root_domain
;
5755 static void init_defrootdomain(void)
5757 init_rootdomain(&def_root_domain
);
5759 atomic_set(&def_root_domain
.refcount
, 1);
5762 static struct root_domain
*alloc_rootdomain(void)
5764 struct root_domain
*rd
;
5766 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5770 if (init_rootdomain(rd
) != 0) {
5778 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5780 struct sched_group
*tmp
, *first
;
5789 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5794 } while (sg
!= first
);
5797 static void free_sched_domain(struct rcu_head
*rcu
)
5799 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5802 * If its an overlapping domain it has private groups, iterate and
5805 if (sd
->flags
& SD_OVERLAP
) {
5806 free_sched_groups(sd
->groups
, 1);
5807 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5808 kfree(sd
->groups
->sgp
);
5814 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5816 call_rcu(&sd
->rcu
, free_sched_domain
);
5819 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5821 for (; sd
; sd
= sd
->parent
)
5822 destroy_sched_domain(sd
, cpu
);
5826 * Keep a special pointer to the highest sched_domain that has
5827 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5828 * allows us to avoid some pointer chasing select_idle_sibling().
5830 * Also keep a unique ID per domain (we use the first cpu number in
5831 * the cpumask of the domain), this allows us to quickly tell if
5832 * two cpus are in the same cache domain, see cpus_share_cache().
5834 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5835 DEFINE_PER_CPU(int, sd_llc_id
);
5837 static void update_top_cache_domain(int cpu
)
5839 struct sched_domain
*sd
;
5842 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5844 id
= cpumask_first(sched_domain_span(sd
));
5846 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5847 per_cpu(sd_llc_id
, cpu
) = id
;
5851 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5852 * hold the hotplug lock.
5855 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5857 struct rq
*rq
= cpu_rq(cpu
);
5858 struct sched_domain
*tmp
;
5860 /* Remove the sched domains which do not contribute to scheduling. */
5861 for (tmp
= sd
; tmp
; ) {
5862 struct sched_domain
*parent
= tmp
->parent
;
5866 if (sd_parent_degenerate(tmp
, parent
)) {
5867 tmp
->parent
= parent
->parent
;
5869 parent
->parent
->child
= tmp
;
5870 destroy_sched_domain(parent
, cpu
);
5875 if (sd
&& sd_degenerate(sd
)) {
5878 destroy_sched_domain(tmp
, cpu
);
5883 sched_domain_debug(sd
, cpu
);
5885 rq_attach_root(rq
, rd
);
5887 rcu_assign_pointer(rq
->sd
, sd
);
5888 destroy_sched_domains(tmp
, cpu
);
5890 update_top_cache_domain(cpu
);
5893 /* cpus with isolated domains */
5894 static cpumask_var_t cpu_isolated_map
;
5896 /* Setup the mask of cpus configured for isolated domains */
5897 static int __init
isolated_cpu_setup(char *str
)
5899 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5900 cpulist_parse(str
, cpu_isolated_map
);
5904 __setup("isolcpus=", isolated_cpu_setup
);
5909 * find_next_best_node - find the next node to include in a sched_domain
5910 * @node: node whose sched_domain we're building
5911 * @used_nodes: nodes already in the sched_domain
5913 * Find the next node to include in a given scheduling domain. Simply
5914 * finds the closest node not already in the @used_nodes map.
5916 * Should use nodemask_t.
5918 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
5920 int i
, n
, val
, min_val
, best_node
= -1;
5924 for (i
= 0; i
< nr_node_ids
; i
++) {
5925 /* Start at @node */
5926 n
= (node
+ i
) % nr_node_ids
;
5928 if (!nr_cpus_node(n
))
5931 /* Skip already used nodes */
5932 if (node_isset(n
, *used_nodes
))
5935 /* Simple min distance search */
5936 val
= node_distance(node
, n
);
5938 if (val
< min_val
) {
5944 if (best_node
!= -1)
5945 node_set(best_node
, *used_nodes
);
5950 * sched_domain_node_span - get a cpumask for a node's sched_domain
5951 * @node: node whose cpumask we're constructing
5952 * @span: resulting cpumask
5954 * Given a node, construct a good cpumask for its sched_domain to span. It
5955 * should be one that prevents unnecessary balancing, but also spreads tasks
5958 static void sched_domain_node_span(int node
, struct cpumask
*span
)
5960 nodemask_t used_nodes
;
5963 cpumask_clear(span
);
5964 nodes_clear(used_nodes
);
5966 cpumask_or(span
, span
, cpumask_of_node(node
));
5967 node_set(node
, used_nodes
);
5969 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5970 int next_node
= find_next_best_node(node
, &used_nodes
);
5973 cpumask_or(span
, span
, cpumask_of_node(next_node
));
5977 static const struct cpumask
*cpu_node_mask(int cpu
)
5979 lockdep_assert_held(&sched_domains_mutex
);
5981 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
5983 return sched_domains_tmpmask
;
5986 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
5988 return cpu_possible_mask
;
5990 #endif /* CONFIG_NUMA */
5992 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5994 return cpumask_of_node(cpu_to_node(cpu
));
5997 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6000 struct sched_domain
**__percpu sd
;
6001 struct sched_group
**__percpu sg
;
6002 struct sched_group_power
**__percpu sgp
;
6006 struct sched_domain
** __percpu sd
;
6007 struct root_domain
*rd
;
6017 struct sched_domain_topology_level
;
6019 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
6020 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
6022 #define SDTL_OVERLAP 0x01
6024 struct sched_domain_topology_level
{
6025 sched_domain_init_f init
;
6026 sched_domain_mask_f mask
;
6028 struct sd_data data
;
6032 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6034 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6035 const struct cpumask
*span
= sched_domain_span(sd
);
6036 struct cpumask
*covered
= sched_domains_tmpmask
;
6037 struct sd_data
*sdd
= sd
->private;
6038 struct sched_domain
*child
;
6041 cpumask_clear(covered
);
6043 for_each_cpu(i
, span
) {
6044 struct cpumask
*sg_span
;
6046 if (cpumask_test_cpu(i
, covered
))
6049 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6050 GFP_KERNEL
, cpu_to_node(cpu
));
6055 sg_span
= sched_group_cpus(sg
);
6057 child
= *per_cpu_ptr(sdd
->sd
, i
);
6059 child
= child
->child
;
6060 cpumask_copy(sg_span
, sched_domain_span(child
));
6062 cpumask_set_cpu(i
, sg_span
);
6064 cpumask_or(covered
, covered
, sg_span
);
6066 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
6067 atomic_inc(&sg
->sgp
->ref
);
6069 if (cpumask_test_cpu(cpu
, sg_span
))
6079 sd
->groups
= groups
;
6084 free_sched_groups(first
, 0);
6089 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6091 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6092 struct sched_domain
*child
= sd
->child
;
6095 cpu
= cpumask_first(sched_domain_span(child
));
6098 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6099 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
6100 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
6107 * build_sched_groups will build a circular linked list of the groups
6108 * covered by the given span, and will set each group's ->cpumask correctly,
6109 * and ->cpu_power to 0.
6111 * Assumes the sched_domain tree is fully constructed
6114 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6116 struct sched_group
*first
= NULL
, *last
= NULL
;
6117 struct sd_data
*sdd
= sd
->private;
6118 const struct cpumask
*span
= sched_domain_span(sd
);
6119 struct cpumask
*covered
;
6122 get_group(cpu
, sdd
, &sd
->groups
);
6123 atomic_inc(&sd
->groups
->ref
);
6125 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
6128 lockdep_assert_held(&sched_domains_mutex
);
6129 covered
= sched_domains_tmpmask
;
6131 cpumask_clear(covered
);
6133 for_each_cpu(i
, span
) {
6134 struct sched_group
*sg
;
6135 int group
= get_group(i
, sdd
, &sg
);
6138 if (cpumask_test_cpu(i
, covered
))
6141 cpumask_clear(sched_group_cpus(sg
));
6144 for_each_cpu(j
, span
) {
6145 if (get_group(j
, sdd
, NULL
) != group
)
6148 cpumask_set_cpu(j
, covered
);
6149 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6164 * Initialize sched groups cpu_power.
6166 * cpu_power indicates the capacity of sched group, which is used while
6167 * distributing the load between different sched groups in a sched domain.
6168 * Typically cpu_power for all the groups in a sched domain will be same unless
6169 * there are asymmetries in the topology. If there are asymmetries, group
6170 * having more cpu_power will pickup more load compared to the group having
6173 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6175 struct sched_group
*sg
= sd
->groups
;
6177 WARN_ON(!sd
|| !sg
);
6180 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6182 } while (sg
!= sd
->groups
);
6184 if (cpu
!= group_first_cpu(sg
))
6187 update_group_power(sd
, cpu
);
6188 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6191 int __weak
arch_sd_sibling_asym_packing(void)
6193 return 0*SD_ASYM_PACKING
;
6197 * Initializers for schedule domains
6198 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6201 #ifdef CONFIG_SCHED_DEBUG
6202 # define SD_INIT_NAME(sd, type) sd->name = #type
6204 # define SD_INIT_NAME(sd, type) do { } while (0)
6207 #define SD_INIT_FUNC(type) \
6208 static noinline struct sched_domain * \
6209 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6211 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6212 *sd = SD_##type##_INIT; \
6213 SD_INIT_NAME(sd, type); \
6214 sd->private = &tl->data; \
6220 SD_INIT_FUNC(ALLNODES
)
6223 #ifdef CONFIG_SCHED_SMT
6224 SD_INIT_FUNC(SIBLING
)
6226 #ifdef CONFIG_SCHED_MC
6229 #ifdef CONFIG_SCHED_BOOK
6233 static int default_relax_domain_level
= -1;
6234 int sched_domain_level_max
;
6236 static int __init
setup_relax_domain_level(char *str
)
6240 val
= simple_strtoul(str
, NULL
, 0);
6241 if (val
< sched_domain_level_max
)
6242 default_relax_domain_level
= val
;
6246 __setup("relax_domain_level=", setup_relax_domain_level
);
6248 static void set_domain_attribute(struct sched_domain
*sd
,
6249 struct sched_domain_attr
*attr
)
6253 if (!attr
|| attr
->relax_domain_level
< 0) {
6254 if (default_relax_domain_level
< 0)
6257 request
= default_relax_domain_level
;
6259 request
= attr
->relax_domain_level
;
6260 if (request
< sd
->level
) {
6261 /* turn off idle balance on this domain */
6262 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6264 /* turn on idle balance on this domain */
6265 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6269 static void __sdt_free(const struct cpumask
*cpu_map
);
6270 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6272 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6273 const struct cpumask
*cpu_map
)
6277 if (!atomic_read(&d
->rd
->refcount
))
6278 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6280 free_percpu(d
->sd
); /* fall through */
6282 __sdt_free(cpu_map
); /* fall through */
6288 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6289 const struct cpumask
*cpu_map
)
6291 memset(d
, 0, sizeof(*d
));
6293 if (__sdt_alloc(cpu_map
))
6294 return sa_sd_storage
;
6295 d
->sd
= alloc_percpu(struct sched_domain
*);
6297 return sa_sd_storage
;
6298 d
->rd
= alloc_rootdomain();
6301 return sa_rootdomain
;
6305 * NULL the sd_data elements we've used to build the sched_domain and
6306 * sched_group structure so that the subsequent __free_domain_allocs()
6307 * will not free the data we're using.
6309 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6311 struct sd_data
*sdd
= sd
->private;
6313 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6314 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6316 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6317 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6319 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6320 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6323 #ifdef CONFIG_SCHED_SMT
6324 static const struct cpumask
*cpu_smt_mask(int cpu
)
6326 return topology_thread_cpumask(cpu
);
6331 * Topology list, bottom-up.
6333 static struct sched_domain_topology_level default_topology
[] = {
6334 #ifdef CONFIG_SCHED_SMT
6335 { sd_init_SIBLING
, cpu_smt_mask
, },
6337 #ifdef CONFIG_SCHED_MC
6338 { sd_init_MC
, cpu_coregroup_mask
, },
6340 #ifdef CONFIG_SCHED_BOOK
6341 { sd_init_BOOK
, cpu_book_mask
, },
6343 { sd_init_CPU
, cpu_cpu_mask
, },
6345 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
6346 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
6351 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6353 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6355 struct sched_domain_topology_level
*tl
;
6358 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6359 struct sd_data
*sdd
= &tl
->data
;
6361 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6365 sdd
->sg
= alloc_percpu(struct sched_group
*);
6369 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6373 for_each_cpu(j
, cpu_map
) {
6374 struct sched_domain
*sd
;
6375 struct sched_group
*sg
;
6376 struct sched_group_power
*sgp
;
6378 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6379 GFP_KERNEL
, cpu_to_node(j
));
6383 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6385 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6386 GFP_KERNEL
, cpu_to_node(j
));
6390 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6392 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
6393 GFP_KERNEL
, cpu_to_node(j
));
6397 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6404 static void __sdt_free(const struct cpumask
*cpu_map
)
6406 struct sched_domain_topology_level
*tl
;
6409 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6410 struct sd_data
*sdd
= &tl
->data
;
6412 for_each_cpu(j
, cpu_map
) {
6413 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, j
);
6414 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6415 free_sched_groups(sd
->groups
, 0);
6416 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6417 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6418 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6420 free_percpu(sdd
->sd
);
6421 free_percpu(sdd
->sg
);
6422 free_percpu(sdd
->sgp
);
6426 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6427 struct s_data
*d
, const struct cpumask
*cpu_map
,
6428 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6431 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6435 set_domain_attribute(sd
, attr
);
6436 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6438 sd
->level
= child
->level
+ 1;
6439 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6448 * Build sched domains for a given set of cpus and attach the sched domains
6449 * to the individual cpus
6451 static int build_sched_domains(const struct cpumask
*cpu_map
,
6452 struct sched_domain_attr
*attr
)
6454 enum s_alloc alloc_state
= sa_none
;
6455 struct sched_domain
*sd
;
6457 int i
, ret
= -ENOMEM
;
6459 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6460 if (alloc_state
!= sa_rootdomain
)
6463 /* Set up domains for cpus specified by the cpu_map. */
6464 for_each_cpu(i
, cpu_map
) {
6465 struct sched_domain_topology_level
*tl
;
6468 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6469 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6470 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6471 sd
->flags
|= SD_OVERLAP
;
6472 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6479 *per_cpu_ptr(d
.sd
, i
) = sd
;
6482 /* Build the groups for the domains */
6483 for_each_cpu(i
, cpu_map
) {
6484 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6485 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6486 if (sd
->flags
& SD_OVERLAP
) {
6487 if (build_overlap_sched_groups(sd
, i
))
6490 if (build_sched_groups(sd
, i
))
6496 /* Calculate CPU power for physical packages and nodes */
6497 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6498 if (!cpumask_test_cpu(i
, cpu_map
))
6501 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6502 claim_allocations(i
, sd
);
6503 init_sched_groups_power(i
, sd
);
6507 /* Attach the domains */
6509 for_each_cpu(i
, cpu_map
) {
6510 sd
= *per_cpu_ptr(d
.sd
, i
);
6511 cpu_attach_domain(sd
, d
.rd
, i
);
6517 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6521 static cpumask_var_t
*doms_cur
; /* current sched domains */
6522 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6523 static struct sched_domain_attr
*dattr_cur
;
6524 /* attribues of custom domains in 'doms_cur' */
6527 * Special case: If a kmalloc of a doms_cur partition (array of
6528 * cpumask) fails, then fallback to a single sched domain,
6529 * as determined by the single cpumask fallback_doms.
6531 static cpumask_var_t fallback_doms
;
6534 * arch_update_cpu_topology lets virtualized architectures update the
6535 * cpu core maps. It is supposed to return 1 if the topology changed
6536 * or 0 if it stayed the same.
6538 int __attribute__((weak
)) arch_update_cpu_topology(void)
6543 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6546 cpumask_var_t
*doms
;
6548 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6551 for (i
= 0; i
< ndoms
; i
++) {
6552 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6553 free_sched_domains(doms
, i
);
6560 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6563 for (i
= 0; i
< ndoms
; i
++)
6564 free_cpumask_var(doms
[i
]);
6569 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6570 * For now this just excludes isolated cpus, but could be used to
6571 * exclude other special cases in the future.
6573 static int init_sched_domains(const struct cpumask
*cpu_map
)
6577 arch_update_cpu_topology();
6579 doms_cur
= alloc_sched_domains(ndoms_cur
);
6581 doms_cur
= &fallback_doms
;
6582 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6584 err
= build_sched_domains(doms_cur
[0], NULL
);
6585 register_sched_domain_sysctl();
6591 * Detach sched domains from a group of cpus specified in cpu_map
6592 * These cpus will now be attached to the NULL domain
6594 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6599 for_each_cpu(i
, cpu_map
)
6600 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6604 /* handle null as "default" */
6605 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6606 struct sched_domain_attr
*new, int idx_new
)
6608 struct sched_domain_attr tmp
;
6615 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6616 new ? (new + idx_new
) : &tmp
,
6617 sizeof(struct sched_domain_attr
));
6621 * Partition sched domains as specified by the 'ndoms_new'
6622 * cpumasks in the array doms_new[] of cpumasks. This compares
6623 * doms_new[] to the current sched domain partitioning, doms_cur[].
6624 * It destroys each deleted domain and builds each new domain.
6626 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6627 * The masks don't intersect (don't overlap.) We should setup one
6628 * sched domain for each mask. CPUs not in any of the cpumasks will
6629 * not be load balanced. If the same cpumask appears both in the
6630 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6633 * The passed in 'doms_new' should be allocated using
6634 * alloc_sched_domains. This routine takes ownership of it and will
6635 * free_sched_domains it when done with it. If the caller failed the
6636 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6637 * and partition_sched_domains() will fallback to the single partition
6638 * 'fallback_doms', it also forces the domains to be rebuilt.
6640 * If doms_new == NULL it will be replaced with cpu_online_mask.
6641 * ndoms_new == 0 is a special case for destroying existing domains,
6642 * and it will not create the default domain.
6644 * Call with hotplug lock held
6646 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6647 struct sched_domain_attr
*dattr_new
)
6652 mutex_lock(&sched_domains_mutex
);
6654 /* always unregister in case we don't destroy any domains */
6655 unregister_sched_domain_sysctl();
6657 /* Let architecture update cpu core mappings. */
6658 new_topology
= arch_update_cpu_topology();
6660 n
= doms_new
? ndoms_new
: 0;
6662 /* Destroy deleted domains */
6663 for (i
= 0; i
< ndoms_cur
; i
++) {
6664 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6665 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6666 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6669 /* no match - a current sched domain not in new doms_new[] */
6670 detach_destroy_domains(doms_cur
[i
]);
6675 if (doms_new
== NULL
) {
6677 doms_new
= &fallback_doms
;
6678 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6679 WARN_ON_ONCE(dattr_new
);
6682 /* Build new domains */
6683 for (i
= 0; i
< ndoms_new
; i
++) {
6684 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6685 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6686 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6689 /* no match - add a new doms_new */
6690 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6695 /* Remember the new sched domains */
6696 if (doms_cur
!= &fallback_doms
)
6697 free_sched_domains(doms_cur
, ndoms_cur
);
6698 kfree(dattr_cur
); /* kfree(NULL) is safe */
6699 doms_cur
= doms_new
;
6700 dattr_cur
= dattr_new
;
6701 ndoms_cur
= ndoms_new
;
6703 register_sched_domain_sysctl();
6705 mutex_unlock(&sched_domains_mutex
);
6708 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6709 static void reinit_sched_domains(void)
6713 /* Destroy domains first to force the rebuild */
6714 partition_sched_domains(0, NULL
, NULL
);
6716 rebuild_sched_domains();
6720 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6722 unsigned int level
= 0;
6724 if (sscanf(buf
, "%u", &level
) != 1)
6728 * level is always be positive so don't check for
6729 * level < POWERSAVINGS_BALANCE_NONE which is 0
6730 * What happens on 0 or 1 byte write,
6731 * need to check for count as well?
6734 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
6738 sched_smt_power_savings
= level
;
6740 sched_mc_power_savings
= level
;
6742 reinit_sched_domains();
6747 #ifdef CONFIG_SCHED_MC
6748 static ssize_t
sched_mc_power_savings_show(struct device
*dev
,
6749 struct device_attribute
*attr
,
6752 return sprintf(buf
, "%u\n", sched_mc_power_savings
);
6754 static ssize_t
sched_mc_power_savings_store(struct device
*dev
,
6755 struct device_attribute
*attr
,
6756 const char *buf
, size_t count
)
6758 return sched_power_savings_store(buf
, count
, 0);
6760 static DEVICE_ATTR(sched_mc_power_savings
, 0644,
6761 sched_mc_power_savings_show
,
6762 sched_mc_power_savings_store
);
6765 #ifdef CONFIG_SCHED_SMT
6766 static ssize_t
sched_smt_power_savings_show(struct device
*dev
,
6767 struct device_attribute
*attr
,
6770 return sprintf(buf
, "%u\n", sched_smt_power_savings
);
6772 static ssize_t
sched_smt_power_savings_store(struct device
*dev
,
6773 struct device_attribute
*attr
,
6774 const char *buf
, size_t count
)
6776 return sched_power_savings_store(buf
, count
, 1);
6778 static DEVICE_ATTR(sched_smt_power_savings
, 0644,
6779 sched_smt_power_savings_show
,
6780 sched_smt_power_savings_store
);
6783 int __init
sched_create_sysfs_power_savings_entries(struct device
*dev
)
6787 #ifdef CONFIG_SCHED_SMT
6789 err
= device_create_file(dev
, &dev_attr_sched_smt_power_savings
);
6791 #ifdef CONFIG_SCHED_MC
6792 if (!err
&& mc_capable())
6793 err
= device_create_file(dev
, &dev_attr_sched_mc_power_savings
);
6797 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6800 * Update cpusets according to cpu_active mask. If cpusets are
6801 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6802 * around partition_sched_domains().
6804 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6807 switch (action
& ~CPU_TASKS_FROZEN
) {
6809 case CPU_DOWN_FAILED
:
6810 cpuset_update_active_cpus();
6817 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6820 switch (action
& ~CPU_TASKS_FROZEN
) {
6821 case CPU_DOWN_PREPARE
:
6822 cpuset_update_active_cpus();
6829 void __init
sched_init_smp(void)
6831 cpumask_var_t non_isolated_cpus
;
6833 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6834 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6837 mutex_lock(&sched_domains_mutex
);
6838 init_sched_domains(cpu_active_mask
);
6839 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6840 if (cpumask_empty(non_isolated_cpus
))
6841 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6842 mutex_unlock(&sched_domains_mutex
);
6845 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6846 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6848 /* RT runtime code needs to handle some hotplug events */
6849 hotcpu_notifier(update_runtime
, 0);
6853 /* Move init over to a non-isolated CPU */
6854 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6856 sched_init_granularity();
6857 free_cpumask_var(non_isolated_cpus
);
6859 init_sched_rt_class();
6862 void __init
sched_init_smp(void)
6864 sched_init_granularity();
6866 #endif /* CONFIG_SMP */
6868 const_debug
unsigned int sysctl_timer_migration
= 1;
6870 int in_sched_functions(unsigned long addr
)
6872 return in_lock_functions(addr
) ||
6873 (addr
>= (unsigned long)__sched_text_start
6874 && addr
< (unsigned long)__sched_text_end
);
6877 #ifdef CONFIG_CGROUP_SCHED
6878 struct task_group root_task_group
;
6881 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6883 void __init
sched_init(void)
6886 unsigned long alloc_size
= 0, ptr
;
6888 #ifdef CONFIG_FAIR_GROUP_SCHED
6889 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6891 #ifdef CONFIG_RT_GROUP_SCHED
6892 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6894 #ifdef CONFIG_CPUMASK_OFFSTACK
6895 alloc_size
+= num_possible_cpus() * cpumask_size();
6898 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6900 #ifdef CONFIG_FAIR_GROUP_SCHED
6901 root_task_group
.se
= (struct sched_entity
**)ptr
;
6902 ptr
+= nr_cpu_ids
* sizeof(void **);
6904 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6905 ptr
+= nr_cpu_ids
* sizeof(void **);
6907 #endif /* CONFIG_FAIR_GROUP_SCHED */
6908 #ifdef CONFIG_RT_GROUP_SCHED
6909 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6910 ptr
+= nr_cpu_ids
* sizeof(void **);
6912 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6913 ptr
+= nr_cpu_ids
* sizeof(void **);
6915 #endif /* CONFIG_RT_GROUP_SCHED */
6916 #ifdef CONFIG_CPUMASK_OFFSTACK
6917 for_each_possible_cpu(i
) {
6918 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6919 ptr
+= cpumask_size();
6921 #endif /* CONFIG_CPUMASK_OFFSTACK */
6925 init_defrootdomain();
6928 init_rt_bandwidth(&def_rt_bandwidth
,
6929 global_rt_period(), global_rt_runtime());
6931 #ifdef CONFIG_RT_GROUP_SCHED
6932 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6933 global_rt_period(), global_rt_runtime());
6934 #endif /* CONFIG_RT_GROUP_SCHED */
6936 #ifdef CONFIG_CGROUP_SCHED
6937 list_add(&root_task_group
.list
, &task_groups
);
6938 INIT_LIST_HEAD(&root_task_group
.children
);
6939 INIT_LIST_HEAD(&root_task_group
.siblings
);
6940 autogroup_init(&init_task
);
6942 #endif /* CONFIG_CGROUP_SCHED */
6944 #ifdef CONFIG_CGROUP_CPUACCT
6945 root_cpuacct
.cpustat
= &kernel_cpustat
;
6946 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6947 /* Too early, not expected to fail */
6948 BUG_ON(!root_cpuacct
.cpuusage
);
6950 for_each_possible_cpu(i
) {
6954 raw_spin_lock_init(&rq
->lock
);
6956 rq
->calc_load_active
= 0;
6957 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6958 init_cfs_rq(&rq
->cfs
);
6959 init_rt_rq(&rq
->rt
, rq
);
6960 #ifdef CONFIG_FAIR_GROUP_SCHED
6961 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6962 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6964 * How much cpu bandwidth does root_task_group get?
6966 * In case of task-groups formed thr' the cgroup filesystem, it
6967 * gets 100% of the cpu resources in the system. This overall
6968 * system cpu resource is divided among the tasks of
6969 * root_task_group and its child task-groups in a fair manner,
6970 * based on each entity's (task or task-group's) weight
6971 * (se->load.weight).
6973 * In other words, if root_task_group has 10 tasks of weight
6974 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6975 * then A0's share of the cpu resource is:
6977 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6979 * We achieve this by letting root_task_group's tasks sit
6980 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6982 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6983 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6984 #endif /* CONFIG_FAIR_GROUP_SCHED */
6986 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6987 #ifdef CONFIG_RT_GROUP_SCHED
6988 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6989 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6992 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6993 rq
->cpu_load
[j
] = 0;
6995 rq
->last_load_update_tick
= jiffies
;
7000 rq
->cpu_power
= SCHED_POWER_SCALE
;
7001 rq
->post_schedule
= 0;
7002 rq
->active_balance
= 0;
7003 rq
->next_balance
= jiffies
;
7008 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7010 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7012 rq_attach_root(rq
, &def_root_domain
);
7018 atomic_set(&rq
->nr_iowait
, 0);
7021 set_load_weight(&init_task
);
7023 #ifdef CONFIG_PREEMPT_NOTIFIERS
7024 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7027 #ifdef CONFIG_RT_MUTEXES
7028 plist_head_init(&init_task
.pi_waiters
);
7032 * The boot idle thread does lazy MMU switching as well:
7034 atomic_inc(&init_mm
.mm_count
);
7035 enter_lazy_tlb(&init_mm
, current
);
7038 * Make us the idle thread. Technically, schedule() should not be
7039 * called from this thread, however somewhere below it might be,
7040 * but because we are the idle thread, we just pick up running again
7041 * when this runqueue becomes "idle".
7043 init_idle(current
, smp_processor_id());
7045 calc_load_update
= jiffies
+ LOAD_FREQ
;
7048 * During early bootup we pretend to be a normal task:
7050 current
->sched_class
= &fair_sched_class
;
7053 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7054 /* May be allocated at isolcpus cmdline parse time */
7055 if (cpu_isolated_map
== NULL
)
7056 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7058 init_sched_fair_class();
7060 scheduler_running
= 1;
7063 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7064 static inline int preempt_count_equals(int preempt_offset
)
7066 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7068 return (nested
== preempt_offset
);
7071 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7073 static unsigned long prev_jiffy
; /* ratelimiting */
7075 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7076 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7077 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7079 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7081 prev_jiffy
= jiffies
;
7084 "BUG: sleeping function called from invalid context at %s:%d\n",
7087 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7088 in_atomic(), irqs_disabled(),
7089 current
->pid
, current
->comm
);
7091 debug_show_held_locks(current
);
7092 if (irqs_disabled())
7093 print_irqtrace_events(current
);
7096 EXPORT_SYMBOL(__might_sleep
);
7099 #ifdef CONFIG_MAGIC_SYSRQ
7100 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7102 const struct sched_class
*prev_class
= p
->sched_class
;
7103 int old_prio
= p
->prio
;
7108 dequeue_task(rq
, p
, 0);
7109 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7111 enqueue_task(rq
, p
, 0);
7112 resched_task(rq
->curr
);
7115 check_class_changed(rq
, p
, prev_class
, old_prio
);
7118 void normalize_rt_tasks(void)
7120 struct task_struct
*g
, *p
;
7121 unsigned long flags
;
7124 read_lock_irqsave(&tasklist_lock
, flags
);
7125 do_each_thread(g
, p
) {
7127 * Only normalize user tasks:
7132 p
->se
.exec_start
= 0;
7133 #ifdef CONFIG_SCHEDSTATS
7134 p
->se
.statistics
.wait_start
= 0;
7135 p
->se
.statistics
.sleep_start
= 0;
7136 p
->se
.statistics
.block_start
= 0;
7141 * Renice negative nice level userspace
7144 if (TASK_NICE(p
) < 0 && p
->mm
)
7145 set_user_nice(p
, 0);
7149 raw_spin_lock(&p
->pi_lock
);
7150 rq
= __task_rq_lock(p
);
7152 normalize_task(rq
, p
);
7154 __task_rq_unlock(rq
);
7155 raw_spin_unlock(&p
->pi_lock
);
7156 } while_each_thread(g
, p
);
7158 read_unlock_irqrestore(&tasklist_lock
, flags
);
7161 #endif /* CONFIG_MAGIC_SYSRQ */
7163 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7165 * These functions are only useful for the IA64 MCA handling, or kdb.
7167 * They can only be called when the whole system has been
7168 * stopped - every CPU needs to be quiescent, and no scheduling
7169 * activity can take place. Using them for anything else would
7170 * be a serious bug, and as a result, they aren't even visible
7171 * under any other configuration.
7175 * curr_task - return the current task for a given cpu.
7176 * @cpu: the processor in question.
7178 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7180 struct task_struct
*curr_task(int cpu
)
7182 return cpu_curr(cpu
);
7185 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7189 * set_curr_task - set the current task for a given cpu.
7190 * @cpu: the processor in question.
7191 * @p: the task pointer to set.
7193 * Description: This function must only be used when non-maskable interrupts
7194 * are serviced on a separate stack. It allows the architecture to switch the
7195 * notion of the current task on a cpu in a non-blocking manner. This function
7196 * must be called with all CPU's synchronized, and interrupts disabled, the
7197 * and caller must save the original value of the current task (see
7198 * curr_task() above) and restore that value before reenabling interrupts and
7199 * re-starting the system.
7201 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7203 void set_curr_task(int cpu
, struct task_struct
*p
)
7210 #ifdef CONFIG_CGROUP_SCHED
7211 /* task_group_lock serializes the addition/removal of task groups */
7212 static DEFINE_SPINLOCK(task_group_lock
);
7214 static void free_sched_group(struct task_group
*tg
)
7216 free_fair_sched_group(tg
);
7217 free_rt_sched_group(tg
);
7222 /* allocate runqueue etc for a new task group */
7223 struct task_group
*sched_create_group(struct task_group
*parent
)
7225 struct task_group
*tg
;
7226 unsigned long flags
;
7228 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7230 return ERR_PTR(-ENOMEM
);
7232 if (!alloc_fair_sched_group(tg
, parent
))
7235 if (!alloc_rt_sched_group(tg
, parent
))
7238 spin_lock_irqsave(&task_group_lock
, flags
);
7239 list_add_rcu(&tg
->list
, &task_groups
);
7241 WARN_ON(!parent
); /* root should already exist */
7243 tg
->parent
= parent
;
7244 INIT_LIST_HEAD(&tg
->children
);
7245 list_add_rcu(&tg
->siblings
, &parent
->children
);
7246 spin_unlock_irqrestore(&task_group_lock
, flags
);
7251 free_sched_group(tg
);
7252 return ERR_PTR(-ENOMEM
);
7255 /* rcu callback to free various structures associated with a task group */
7256 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7258 /* now it should be safe to free those cfs_rqs */
7259 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7262 /* Destroy runqueue etc associated with a task group */
7263 void sched_destroy_group(struct task_group
*tg
)
7265 unsigned long flags
;
7268 /* end participation in shares distribution */
7269 for_each_possible_cpu(i
)
7270 unregister_fair_sched_group(tg
, i
);
7272 spin_lock_irqsave(&task_group_lock
, flags
);
7273 list_del_rcu(&tg
->list
);
7274 list_del_rcu(&tg
->siblings
);
7275 spin_unlock_irqrestore(&task_group_lock
, flags
);
7277 /* wait for possible concurrent references to cfs_rqs complete */
7278 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7281 /* change task's runqueue when it moves between groups.
7282 * The caller of this function should have put the task in its new group
7283 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7284 * reflect its new group.
7286 void sched_move_task(struct task_struct
*tsk
)
7289 unsigned long flags
;
7292 rq
= task_rq_lock(tsk
, &flags
);
7294 running
= task_current(rq
, tsk
);
7298 dequeue_task(rq
, tsk
, 0);
7299 if (unlikely(running
))
7300 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7302 #ifdef CONFIG_FAIR_GROUP_SCHED
7303 if (tsk
->sched_class
->task_move_group
)
7304 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7307 set_task_rq(tsk
, task_cpu(tsk
));
7309 if (unlikely(running
))
7310 tsk
->sched_class
->set_curr_task(rq
);
7312 enqueue_task(rq
, tsk
, 0);
7314 task_rq_unlock(rq
, tsk
, &flags
);
7316 #endif /* CONFIG_CGROUP_SCHED */
7318 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7319 static unsigned long to_ratio(u64 period
, u64 runtime
)
7321 if (runtime
== RUNTIME_INF
)
7324 return div64_u64(runtime
<< 20, period
);
7328 #ifdef CONFIG_RT_GROUP_SCHED
7330 * Ensure that the real time constraints are schedulable.
7332 static DEFINE_MUTEX(rt_constraints_mutex
);
7334 /* Must be called with tasklist_lock held */
7335 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7337 struct task_struct
*g
, *p
;
7339 do_each_thread(g
, p
) {
7340 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7342 } while_each_thread(g
, p
);
7347 struct rt_schedulable_data
{
7348 struct task_group
*tg
;
7353 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7355 struct rt_schedulable_data
*d
= data
;
7356 struct task_group
*child
;
7357 unsigned long total
, sum
= 0;
7358 u64 period
, runtime
;
7360 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7361 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7364 period
= d
->rt_period
;
7365 runtime
= d
->rt_runtime
;
7369 * Cannot have more runtime than the period.
7371 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7375 * Ensure we don't starve existing RT tasks.
7377 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7380 total
= to_ratio(period
, runtime
);
7383 * Nobody can have more than the global setting allows.
7385 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7389 * The sum of our children's runtime should not exceed our own.
7391 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7392 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7393 runtime
= child
->rt_bandwidth
.rt_runtime
;
7395 if (child
== d
->tg
) {
7396 period
= d
->rt_period
;
7397 runtime
= d
->rt_runtime
;
7400 sum
+= to_ratio(period
, runtime
);
7409 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7413 struct rt_schedulable_data data
= {
7415 .rt_period
= period
,
7416 .rt_runtime
= runtime
,
7420 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7426 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7427 u64 rt_period
, u64 rt_runtime
)
7431 mutex_lock(&rt_constraints_mutex
);
7432 read_lock(&tasklist_lock
);
7433 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7437 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7438 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7439 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7441 for_each_possible_cpu(i
) {
7442 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7444 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7445 rt_rq
->rt_runtime
= rt_runtime
;
7446 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7448 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7450 read_unlock(&tasklist_lock
);
7451 mutex_unlock(&rt_constraints_mutex
);
7456 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7458 u64 rt_runtime
, rt_period
;
7460 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7461 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7462 if (rt_runtime_us
< 0)
7463 rt_runtime
= RUNTIME_INF
;
7465 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7468 long sched_group_rt_runtime(struct task_group
*tg
)
7472 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7475 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7476 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7477 return rt_runtime_us
;
7480 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7482 u64 rt_runtime
, rt_period
;
7484 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7485 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7490 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7493 long sched_group_rt_period(struct task_group
*tg
)
7497 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7498 do_div(rt_period_us
, NSEC_PER_USEC
);
7499 return rt_period_us
;
7502 static int sched_rt_global_constraints(void)
7504 u64 runtime
, period
;
7507 if (sysctl_sched_rt_period
<= 0)
7510 runtime
= global_rt_runtime();
7511 period
= global_rt_period();
7514 * Sanity check on the sysctl variables.
7516 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7519 mutex_lock(&rt_constraints_mutex
);
7520 read_lock(&tasklist_lock
);
7521 ret
= __rt_schedulable(NULL
, 0, 0);
7522 read_unlock(&tasklist_lock
);
7523 mutex_unlock(&rt_constraints_mutex
);
7528 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7530 /* Don't accept realtime tasks when there is no way for them to run */
7531 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7537 #else /* !CONFIG_RT_GROUP_SCHED */
7538 static int sched_rt_global_constraints(void)
7540 unsigned long flags
;
7543 if (sysctl_sched_rt_period
<= 0)
7547 * There's always some RT tasks in the root group
7548 * -- migration, kstopmachine etc..
7550 if (sysctl_sched_rt_runtime
== 0)
7553 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7554 for_each_possible_cpu(i
) {
7555 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7557 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7558 rt_rq
->rt_runtime
= global_rt_runtime();
7559 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7561 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7565 #endif /* CONFIG_RT_GROUP_SCHED */
7567 int sched_rt_handler(struct ctl_table
*table
, int write
,
7568 void __user
*buffer
, size_t *lenp
,
7572 int old_period
, old_runtime
;
7573 static DEFINE_MUTEX(mutex
);
7576 old_period
= sysctl_sched_rt_period
;
7577 old_runtime
= sysctl_sched_rt_runtime
;
7579 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7581 if (!ret
&& write
) {
7582 ret
= sched_rt_global_constraints();
7584 sysctl_sched_rt_period
= old_period
;
7585 sysctl_sched_rt_runtime
= old_runtime
;
7587 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7588 def_rt_bandwidth
.rt_period
=
7589 ns_to_ktime(global_rt_period());
7592 mutex_unlock(&mutex
);
7597 #ifdef CONFIG_CGROUP_SCHED
7599 /* return corresponding task_group object of a cgroup */
7600 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7602 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7603 struct task_group
, css
);
7606 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7608 struct task_group
*tg
, *parent
;
7610 if (!cgrp
->parent
) {
7611 /* This is early initialization for the top cgroup */
7612 return &root_task_group
.css
;
7615 parent
= cgroup_tg(cgrp
->parent
);
7616 tg
= sched_create_group(parent
);
7618 return ERR_PTR(-ENOMEM
);
7623 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7625 struct task_group
*tg
= cgroup_tg(cgrp
);
7627 sched_destroy_group(tg
);
7630 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7631 struct cgroup_taskset
*tset
)
7633 struct task_struct
*task
;
7635 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7636 #ifdef CONFIG_RT_GROUP_SCHED
7637 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7640 /* We don't support RT-tasks being in separate groups */
7641 if (task
->sched_class
!= &fair_sched_class
)
7648 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7649 struct cgroup_taskset
*tset
)
7651 struct task_struct
*task
;
7653 cgroup_taskset_for_each(task
, cgrp
, tset
)
7654 sched_move_task(task
);
7658 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7659 struct task_struct
*task
)
7662 * cgroup_exit() is called in the copy_process() failure path.
7663 * Ignore this case since the task hasn't ran yet, this avoids
7664 * trying to poke a half freed task state from generic code.
7666 if (!(task
->flags
& PF_EXITING
))
7669 sched_move_task(task
);
7672 #ifdef CONFIG_FAIR_GROUP_SCHED
7673 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7676 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7679 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7681 struct task_group
*tg
= cgroup_tg(cgrp
);
7683 return (u64
) scale_load_down(tg
->shares
);
7686 #ifdef CONFIG_CFS_BANDWIDTH
7687 static DEFINE_MUTEX(cfs_constraints_mutex
);
7689 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7690 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7692 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7694 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7696 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7697 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7699 if (tg
== &root_task_group
)
7703 * Ensure we have at some amount of bandwidth every period. This is
7704 * to prevent reaching a state of large arrears when throttled via
7705 * entity_tick() resulting in prolonged exit starvation.
7707 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7711 * Likewise, bound things on the otherside by preventing insane quota
7712 * periods. This also allows us to normalize in computing quota
7715 if (period
> max_cfs_quota_period
)
7718 mutex_lock(&cfs_constraints_mutex
);
7719 ret
= __cfs_schedulable(tg
, period
, quota
);
7723 runtime_enabled
= quota
!= RUNTIME_INF
;
7724 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7725 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7726 raw_spin_lock_irq(&cfs_b
->lock
);
7727 cfs_b
->period
= ns_to_ktime(period
);
7728 cfs_b
->quota
= quota
;
7730 __refill_cfs_bandwidth_runtime(cfs_b
);
7731 /* restart the period timer (if active) to handle new period expiry */
7732 if (runtime_enabled
&& cfs_b
->timer_active
) {
7733 /* force a reprogram */
7734 cfs_b
->timer_active
= 0;
7735 __start_cfs_bandwidth(cfs_b
);
7737 raw_spin_unlock_irq(&cfs_b
->lock
);
7739 for_each_possible_cpu(i
) {
7740 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7741 struct rq
*rq
= cfs_rq
->rq
;
7743 raw_spin_lock_irq(&rq
->lock
);
7744 cfs_rq
->runtime_enabled
= runtime_enabled
;
7745 cfs_rq
->runtime_remaining
= 0;
7747 if (cfs_rq
->throttled
)
7748 unthrottle_cfs_rq(cfs_rq
);
7749 raw_spin_unlock_irq(&rq
->lock
);
7752 mutex_unlock(&cfs_constraints_mutex
);
7757 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7761 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7762 if (cfs_quota_us
< 0)
7763 quota
= RUNTIME_INF
;
7765 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7767 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7770 long tg_get_cfs_quota(struct task_group
*tg
)
7774 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7777 quota_us
= tg
->cfs_bandwidth
.quota
;
7778 do_div(quota_us
, NSEC_PER_USEC
);
7783 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7787 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7788 quota
= tg
->cfs_bandwidth
.quota
;
7790 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7793 long tg_get_cfs_period(struct task_group
*tg
)
7797 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7798 do_div(cfs_period_us
, NSEC_PER_USEC
);
7800 return cfs_period_us
;
7803 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7805 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7808 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7811 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7814 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7816 return tg_get_cfs_period(cgroup_tg(cgrp
));
7819 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7822 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7825 struct cfs_schedulable_data
{
7826 struct task_group
*tg
;
7831 * normalize group quota/period to be quota/max_period
7832 * note: units are usecs
7834 static u64
normalize_cfs_quota(struct task_group
*tg
,
7835 struct cfs_schedulable_data
*d
)
7843 period
= tg_get_cfs_period(tg
);
7844 quota
= tg_get_cfs_quota(tg
);
7847 /* note: these should typically be equivalent */
7848 if (quota
== RUNTIME_INF
|| quota
== -1)
7851 return to_ratio(period
, quota
);
7854 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7856 struct cfs_schedulable_data
*d
= data
;
7857 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7858 s64 quota
= 0, parent_quota
= -1;
7861 quota
= RUNTIME_INF
;
7863 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7865 quota
= normalize_cfs_quota(tg
, d
);
7866 parent_quota
= parent_b
->hierarchal_quota
;
7869 * ensure max(child_quota) <= parent_quota, inherit when no
7872 if (quota
== RUNTIME_INF
)
7873 quota
= parent_quota
;
7874 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7877 cfs_b
->hierarchal_quota
= quota
;
7882 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7885 struct cfs_schedulable_data data
= {
7891 if (quota
!= RUNTIME_INF
) {
7892 do_div(data
.period
, NSEC_PER_USEC
);
7893 do_div(data
.quota
, NSEC_PER_USEC
);
7897 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7903 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7904 struct cgroup_map_cb
*cb
)
7906 struct task_group
*tg
= cgroup_tg(cgrp
);
7907 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7909 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7910 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7911 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7915 #endif /* CONFIG_CFS_BANDWIDTH */
7916 #endif /* CONFIG_FAIR_GROUP_SCHED */
7918 #ifdef CONFIG_RT_GROUP_SCHED
7919 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7922 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7925 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7927 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7930 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7933 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7936 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7938 return sched_group_rt_period(cgroup_tg(cgrp
));
7940 #endif /* CONFIG_RT_GROUP_SCHED */
7942 static struct cftype cpu_files
[] = {
7943 #ifdef CONFIG_FAIR_GROUP_SCHED
7946 .read_u64
= cpu_shares_read_u64
,
7947 .write_u64
= cpu_shares_write_u64
,
7950 #ifdef CONFIG_CFS_BANDWIDTH
7952 .name
= "cfs_quota_us",
7953 .read_s64
= cpu_cfs_quota_read_s64
,
7954 .write_s64
= cpu_cfs_quota_write_s64
,
7957 .name
= "cfs_period_us",
7958 .read_u64
= cpu_cfs_period_read_u64
,
7959 .write_u64
= cpu_cfs_period_write_u64
,
7963 .read_map
= cpu_stats_show
,
7966 #ifdef CONFIG_RT_GROUP_SCHED
7968 .name
= "rt_runtime_us",
7969 .read_s64
= cpu_rt_runtime_read
,
7970 .write_s64
= cpu_rt_runtime_write
,
7973 .name
= "rt_period_us",
7974 .read_u64
= cpu_rt_period_read_uint
,
7975 .write_u64
= cpu_rt_period_write_uint
,
7980 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7982 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7985 struct cgroup_subsys cpu_cgroup_subsys
= {
7987 .create
= cpu_cgroup_create
,
7988 .destroy
= cpu_cgroup_destroy
,
7989 .can_attach
= cpu_cgroup_can_attach
,
7990 .attach
= cpu_cgroup_attach
,
7991 .exit
= cpu_cgroup_exit
,
7992 .populate
= cpu_cgroup_populate
,
7993 .subsys_id
= cpu_cgroup_subsys_id
,
7997 #endif /* CONFIG_CGROUP_SCHED */
7999 #ifdef CONFIG_CGROUP_CPUACCT
8002 * CPU accounting code for task groups.
8004 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8005 * (balbir@in.ibm.com).
8008 /* create a new cpu accounting group */
8009 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
8014 return &root_cpuacct
.css
;
8016 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8020 ca
->cpuusage
= alloc_percpu(u64
);
8024 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
8026 goto out_free_cpuusage
;
8031 free_percpu(ca
->cpuusage
);
8035 return ERR_PTR(-ENOMEM
);
8038 /* destroy an existing cpu accounting group */
8039 static void cpuacct_destroy(struct cgroup
*cgrp
)
8041 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8043 free_percpu(ca
->cpustat
);
8044 free_percpu(ca
->cpuusage
);
8048 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8050 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8053 #ifndef CONFIG_64BIT
8055 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8057 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8059 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8067 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8069 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8071 #ifndef CONFIG_64BIT
8073 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8075 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8077 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8083 /* return total cpu usage (in nanoseconds) of a group */
8084 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8086 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8087 u64 totalcpuusage
= 0;
8090 for_each_present_cpu(i
)
8091 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8093 return totalcpuusage
;
8096 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8099 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8108 for_each_present_cpu(i
)
8109 cpuacct_cpuusage_write(ca
, i
, 0);
8115 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8118 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8122 for_each_present_cpu(i
) {
8123 percpu
= cpuacct_cpuusage_read(ca
, i
);
8124 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8126 seq_printf(m
, "\n");
8130 static const char *cpuacct_stat_desc
[] = {
8131 [CPUACCT_STAT_USER
] = "user",
8132 [CPUACCT_STAT_SYSTEM
] = "system",
8135 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8136 struct cgroup_map_cb
*cb
)
8138 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8142 for_each_online_cpu(cpu
) {
8143 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8144 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8145 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8147 val
= cputime64_to_clock_t(val
);
8148 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8151 for_each_online_cpu(cpu
) {
8152 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8153 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8154 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8155 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8158 val
= cputime64_to_clock_t(val
);
8159 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8164 static struct cftype files
[] = {
8167 .read_u64
= cpuusage_read
,
8168 .write_u64
= cpuusage_write
,
8171 .name
= "usage_percpu",
8172 .read_seq_string
= cpuacct_percpu_seq_read
,
8176 .read_map
= cpuacct_stats_show
,
8180 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8182 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8186 * charge this task's execution time to its accounting group.
8188 * called with rq->lock held.
8190 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8195 if (unlikely(!cpuacct_subsys
.active
))
8198 cpu
= task_cpu(tsk
);
8204 for (; ca
; ca
= parent_ca(ca
)) {
8205 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8206 *cpuusage
+= cputime
;
8212 struct cgroup_subsys cpuacct_subsys
= {
8214 .create
= cpuacct_create
,
8215 .destroy
= cpuacct_destroy
,
8216 .populate
= cpuacct_populate
,
8217 .subsys_id
= cpuacct_subsys_id
,
8219 #endif /* CONFIG_CGROUP_CPUACCT */