4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
94 ktime_t soft
, hard
, now
;
97 if (hrtimer_active(period_timer
))
100 now
= hrtimer_cb_get_time(period_timer
);
101 hrtimer_forward(period_timer
, now
, period
);
103 soft
= hrtimer_get_softexpires(period_timer
);
104 hard
= hrtimer_get_expires(period_timer
);
105 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
106 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
107 HRTIMER_MODE_ABS_PINNED
, 0);
111 DEFINE_MUTEX(sched_domains_mutex
);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
114 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
116 void update_rq_clock(struct rq
*rq
)
120 if (rq
->skip_clock_update
> 0)
123 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
125 update_rq_clock_task(rq
, delta
);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug
unsigned int sysctl_sched_features
=
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static __read_mostly
char *sched_feat_names
[] = {
146 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
197 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
198 size_t cnt
, loff_t
*ppos
)
208 if (copy_from_user(&buf
, ubuf
, cnt
))
214 if (strncmp(cmp
, "NO_", 3) == 0) {
219 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
220 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
222 sysctl_sched_features
&= ~(1UL << i
);
223 sched_feat_disable(i
);
225 sysctl_sched_features
|= (1UL << i
);
226 sched_feat_enable(i
);
232 if (i
== __SCHED_FEAT_NR
)
240 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
242 return single_open(filp
, sched_feat_show
, NULL
);
245 static const struct file_operations sched_feat_fops
= {
246 .open
= sched_feat_open
,
247 .write
= sched_feat_write
,
250 .release
= single_release
,
253 static __init
int sched_init_debug(void)
255 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
260 late_initcall(sched_init_debug
);
261 #endif /* CONFIG_SCHED_DEBUG */
264 * Number of tasks to iterate in a single balance run.
265 * Limited because this is done with IRQs disabled.
267 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
270 * period over which we average the RT time consumption, measured
275 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
278 * period over which we measure -rt task cpu usage in us.
281 unsigned int sysctl_sched_rt_period
= 1000000;
283 __read_mostly
int scheduler_running
;
286 * part of the period that we allow rt tasks to run in us.
289 int sysctl_sched_rt_runtime
= 950000;
294 * __task_rq_lock - lock the rq @p resides on.
296 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
301 lockdep_assert_held(&p
->pi_lock
);
305 raw_spin_lock(&rq
->lock
);
306 if (likely(rq
== task_rq(p
)))
308 raw_spin_unlock(&rq
->lock
);
313 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
315 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
316 __acquires(p
->pi_lock
)
322 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
324 raw_spin_lock(&rq
->lock
);
325 if (likely(rq
== task_rq(p
)))
327 raw_spin_unlock(&rq
->lock
);
328 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
332 static void __task_rq_unlock(struct rq
*rq
)
335 raw_spin_unlock(&rq
->lock
);
339 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
341 __releases(p
->pi_lock
)
343 raw_spin_unlock(&rq
->lock
);
344 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
348 * this_rq_lock - lock this runqueue and disable interrupts.
350 static struct rq
*this_rq_lock(void)
357 raw_spin_lock(&rq
->lock
);
362 #ifdef CONFIG_SCHED_HRTICK
364 * Use HR-timers to deliver accurate preemption points.
366 * Its all a bit involved since we cannot program an hrt while holding the
367 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
370 * When we get rescheduled we reprogram the hrtick_timer outside of the
374 static void hrtick_clear(struct rq
*rq
)
376 if (hrtimer_active(&rq
->hrtick_timer
))
377 hrtimer_cancel(&rq
->hrtick_timer
);
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
384 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
386 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
388 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
390 raw_spin_lock(&rq
->lock
);
392 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
393 raw_spin_unlock(&rq
->lock
);
395 return HRTIMER_NORESTART
;
400 * called from hardirq (IPI) context
402 static void __hrtick_start(void *arg
)
406 raw_spin_lock(&rq
->lock
);
407 hrtimer_restart(&rq
->hrtick_timer
);
408 rq
->hrtick_csd_pending
= 0;
409 raw_spin_unlock(&rq
->lock
);
413 * Called to set the hrtick timer state.
415 * called with rq->lock held and irqs disabled
417 void hrtick_start(struct rq
*rq
, u64 delay
)
419 struct hrtimer
*timer
= &rq
->hrtick_timer
;
420 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
422 hrtimer_set_expires(timer
, time
);
424 if (rq
== this_rq()) {
425 hrtimer_restart(timer
);
426 } else if (!rq
->hrtick_csd_pending
) {
427 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
428 rq
->hrtick_csd_pending
= 1;
433 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
435 int cpu
= (int)(long)hcpu
;
438 case CPU_UP_CANCELED
:
439 case CPU_UP_CANCELED_FROZEN
:
440 case CPU_DOWN_PREPARE
:
441 case CPU_DOWN_PREPARE_FROZEN
:
443 case CPU_DEAD_FROZEN
:
444 hrtick_clear(cpu_rq(cpu
));
451 static __init
void init_hrtick(void)
453 hotcpu_notifier(hotplug_hrtick
, 0);
457 * Called to set the hrtick timer state.
459 * called with rq->lock held and irqs disabled
461 void hrtick_start(struct rq
*rq
, u64 delay
)
463 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
464 HRTIMER_MODE_REL_PINNED
, 0);
467 static inline void init_hrtick(void)
470 #endif /* CONFIG_SMP */
472 static void init_rq_hrtick(struct rq
*rq
)
475 rq
->hrtick_csd_pending
= 0;
477 rq
->hrtick_csd
.flags
= 0;
478 rq
->hrtick_csd
.func
= __hrtick_start
;
479 rq
->hrtick_csd
.info
= rq
;
482 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
483 rq
->hrtick_timer
.function
= hrtick
;
485 #else /* CONFIG_SCHED_HRTICK */
486 static inline void hrtick_clear(struct rq
*rq
)
490 static inline void init_rq_hrtick(struct rq
*rq
)
494 static inline void init_hrtick(void)
497 #endif /* CONFIG_SCHED_HRTICK */
500 * resched_task - mark a task 'to be rescheduled now'.
502 * On UP this means the setting of the need_resched flag, on SMP it
503 * might also involve a cross-CPU call to trigger the scheduler on
508 #ifndef tsk_is_polling
509 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
512 void resched_task(struct task_struct
*p
)
516 assert_raw_spin_locked(&task_rq(p
)->lock
);
518 if (test_tsk_need_resched(p
))
521 set_tsk_need_resched(p
);
524 if (cpu
== smp_processor_id())
527 /* NEED_RESCHED must be visible before we test polling */
529 if (!tsk_is_polling(p
))
530 smp_send_reschedule(cpu
);
533 void resched_cpu(int cpu
)
535 struct rq
*rq
= cpu_rq(cpu
);
538 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
540 resched_task(cpu_curr(cpu
));
541 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
546 * In the semi idle case, use the nearest busy cpu for migrating timers
547 * from an idle cpu. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle cpu will add more delays to the timers than intended
551 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int cpu
= smp_processor_id();
557 struct sched_domain
*sd
;
560 for_each_domain(cpu
, sd
) {
561 for_each_cpu(i
, sched_domain_span(sd
)) {
573 * When add_timer_on() enqueues a timer into the timer wheel of an
574 * idle CPU then this timer might expire before the next timer event
575 * which is scheduled to wake up that CPU. In case of a completely
576 * idle system the next event might even be infinite time into the
577 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
578 * leaves the inner idle loop so the newly added timer is taken into
579 * account when the CPU goes back to idle and evaluates the timer
580 * wheel for the next timer event.
582 void wake_up_idle_cpu(int cpu
)
584 struct rq
*rq
= cpu_rq(cpu
);
586 if (cpu
== smp_processor_id())
590 * This is safe, as this function is called with the timer
591 * wheel base lock of (cpu) held. When the CPU is on the way
592 * to idle and has not yet set rq->curr to idle then it will
593 * be serialized on the timer wheel base lock and take the new
594 * timer into account automatically.
596 if (rq
->curr
!= rq
->idle
)
600 * We can set TIF_RESCHED on the idle task of the other CPU
601 * lockless. The worst case is that the other CPU runs the
602 * idle task through an additional NOOP schedule()
604 set_tsk_need_resched(rq
->idle
);
606 /* NEED_RESCHED must be visible before we test polling */
608 if (!tsk_is_polling(rq
->idle
))
609 smp_send_reschedule(cpu
);
612 static inline bool got_nohz_idle_kick(void)
614 int cpu
= smp_processor_id();
615 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
618 #else /* CONFIG_NO_HZ */
620 static inline bool got_nohz_idle_kick(void)
625 #endif /* CONFIG_NO_HZ */
627 void sched_avg_update(struct rq
*rq
)
629 s64 period
= sched_avg_period();
631 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
633 * Inline assembly required to prevent the compiler
634 * optimising this loop into a divmod call.
635 * See __iter_div_u64_rem() for another example of this.
637 asm("" : "+rm" (rq
->age_stamp
));
638 rq
->age_stamp
+= period
;
643 #else /* !CONFIG_SMP */
644 void resched_task(struct task_struct
*p
)
646 assert_raw_spin_locked(&task_rq(p
)->lock
);
647 set_tsk_need_resched(p
);
649 #endif /* CONFIG_SMP */
651 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
652 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
654 * Iterate task_group tree rooted at *from, calling @down when first entering a
655 * node and @up when leaving it for the final time.
657 * Caller must hold rcu_lock or sufficient equivalent.
659 int walk_tg_tree_from(struct task_group
*from
,
660 tg_visitor down
, tg_visitor up
, void *data
)
662 struct task_group
*parent
, *child
;
668 ret
= (*down
)(parent
, data
);
671 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
678 ret
= (*up
)(parent
, data
);
679 if (ret
|| parent
== from
)
683 parent
= parent
->parent
;
690 int tg_nop(struct task_group
*tg
, void *data
)
696 static void set_load_weight(struct task_struct
*p
)
698 int prio
= p
->static_prio
- MAX_RT_PRIO
;
699 struct load_weight
*load
= &p
->se
.load
;
702 * SCHED_IDLE tasks get minimal weight:
704 if (p
->policy
== SCHED_IDLE
) {
705 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
706 load
->inv_weight
= WMULT_IDLEPRIO
;
710 load
->weight
= scale_load(prio_to_weight
[prio
]);
711 load
->inv_weight
= prio_to_wmult
[prio
];
714 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
717 sched_info_queued(p
);
718 p
->sched_class
->enqueue_task(rq
, p
, flags
);
721 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
724 sched_info_dequeued(p
);
725 p
->sched_class
->dequeue_task(rq
, p
, flags
);
728 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
730 if (task_contributes_to_load(p
))
731 rq
->nr_uninterruptible
--;
733 enqueue_task(rq
, p
, flags
);
736 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
738 if (task_contributes_to_load(p
))
739 rq
->nr_uninterruptible
++;
741 dequeue_task(rq
, p
, flags
);
744 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
747 * There are no locks covering percpu hardirq/softirq time.
748 * They are only modified in account_system_vtime, on corresponding CPU
749 * with interrupts disabled. So, writes are safe.
750 * They are read and saved off onto struct rq in update_rq_clock().
751 * This may result in other CPU reading this CPU's irq time and can
752 * race with irq/account_system_vtime on this CPU. We would either get old
753 * or new value with a side effect of accounting a slice of irq time to wrong
754 * task when irq is in progress while we read rq->clock. That is a worthy
755 * compromise in place of having locks on each irq in account_system_time.
757 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
758 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
760 static DEFINE_PER_CPU(u64
, irq_start_time
);
761 static int sched_clock_irqtime
;
763 void enable_sched_clock_irqtime(void)
765 sched_clock_irqtime
= 1;
768 void disable_sched_clock_irqtime(void)
770 sched_clock_irqtime
= 0;
774 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
776 static inline void irq_time_write_begin(void)
778 __this_cpu_inc(irq_time_seq
.sequence
);
782 static inline void irq_time_write_end(void)
785 __this_cpu_inc(irq_time_seq
.sequence
);
788 static inline u64
irq_time_read(int cpu
)
794 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
795 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
796 per_cpu(cpu_hardirq_time
, cpu
);
797 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
801 #else /* CONFIG_64BIT */
802 static inline void irq_time_write_begin(void)
806 static inline void irq_time_write_end(void)
810 static inline u64
irq_time_read(int cpu
)
812 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
814 #endif /* CONFIG_64BIT */
817 * Called before incrementing preempt_count on {soft,}irq_enter
818 * and before decrementing preempt_count on {soft,}irq_exit.
820 void account_system_vtime(struct task_struct
*curr
)
826 if (!sched_clock_irqtime
)
829 local_irq_save(flags
);
831 cpu
= smp_processor_id();
832 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
833 __this_cpu_add(irq_start_time
, delta
);
835 irq_time_write_begin();
837 * We do not account for softirq time from ksoftirqd here.
838 * We want to continue accounting softirq time to ksoftirqd thread
839 * in that case, so as not to confuse scheduler with a special task
840 * that do not consume any time, but still wants to run.
843 __this_cpu_add(cpu_hardirq_time
, delta
);
844 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
845 __this_cpu_add(cpu_softirq_time
, delta
);
847 irq_time_write_end();
848 local_irq_restore(flags
);
850 EXPORT_SYMBOL_GPL(account_system_vtime
);
852 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
854 #ifdef CONFIG_PARAVIRT
855 static inline u64
steal_ticks(u64 steal
)
857 if (unlikely(steal
> NSEC_PER_SEC
))
858 return div_u64(steal
, TICK_NSEC
);
860 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
864 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
867 * In theory, the compile should just see 0 here, and optimize out the call
868 * to sched_rt_avg_update. But I don't trust it...
870 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
871 s64 steal
= 0, irq_delta
= 0;
873 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
874 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
877 * Since irq_time is only updated on {soft,}irq_exit, we might run into
878 * this case when a previous update_rq_clock() happened inside a
881 * When this happens, we stop ->clock_task and only update the
882 * prev_irq_time stamp to account for the part that fit, so that a next
883 * update will consume the rest. This ensures ->clock_task is
886 * It does however cause some slight miss-attribution of {soft,}irq
887 * time, a more accurate solution would be to update the irq_time using
888 * the current rq->clock timestamp, except that would require using
891 if (irq_delta
> delta
)
894 rq
->prev_irq_time
+= irq_delta
;
897 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
898 if (static_key_false((¶virt_steal_rq_enabled
))) {
901 steal
= paravirt_steal_clock(cpu_of(rq
));
902 steal
-= rq
->prev_steal_time_rq
;
904 if (unlikely(steal
> delta
))
907 st
= steal_ticks(steal
);
908 steal
= st
* TICK_NSEC
;
910 rq
->prev_steal_time_rq
+= steal
;
916 rq
->clock_task
+= delta
;
918 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
919 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
920 sched_rt_avg_update(rq
, irq_delta
+ steal
);
924 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
925 static int irqtime_account_hi_update(void)
927 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
932 local_irq_save(flags
);
933 latest_ns
= this_cpu_read(cpu_hardirq_time
);
934 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_IRQ
])
936 local_irq_restore(flags
);
940 static int irqtime_account_si_update(void)
942 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
947 local_irq_save(flags
);
948 latest_ns
= this_cpu_read(cpu_softirq_time
);
949 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_SOFTIRQ
])
951 local_irq_restore(flags
);
955 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
957 #define sched_clock_irqtime (0)
961 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
963 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
964 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
968 * Make it appear like a SCHED_FIFO task, its something
969 * userspace knows about and won't get confused about.
971 * Also, it will make PI more or less work without too
972 * much confusion -- but then, stop work should not
973 * rely on PI working anyway.
975 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
977 stop
->sched_class
= &stop_sched_class
;
980 cpu_rq(cpu
)->stop
= stop
;
984 * Reset it back to a normal scheduling class so that
985 * it can die in pieces.
987 old_stop
->sched_class
= &rt_sched_class
;
992 * __normal_prio - return the priority that is based on the static prio
994 static inline int __normal_prio(struct task_struct
*p
)
996 return p
->static_prio
;
1000 * Calculate the expected normal priority: i.e. priority
1001 * without taking RT-inheritance into account. Might be
1002 * boosted by interactivity modifiers. Changes upon fork,
1003 * setprio syscalls, and whenever the interactivity
1004 * estimator recalculates.
1006 static inline int normal_prio(struct task_struct
*p
)
1010 if (task_has_rt_policy(p
))
1011 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1013 prio
= __normal_prio(p
);
1018 * Calculate the current priority, i.e. the priority
1019 * taken into account by the scheduler. This value might
1020 * be boosted by RT tasks, or might be boosted by
1021 * interactivity modifiers. Will be RT if the task got
1022 * RT-boosted. If not then it returns p->normal_prio.
1024 static int effective_prio(struct task_struct
*p
)
1026 p
->normal_prio
= normal_prio(p
);
1028 * If we are RT tasks or we were boosted to RT priority,
1029 * keep the priority unchanged. Otherwise, update priority
1030 * to the normal priority:
1032 if (!rt_prio(p
->prio
))
1033 return p
->normal_prio
;
1038 * task_curr - is this task currently executing on a CPU?
1039 * @p: the task in question.
1041 inline int task_curr(const struct task_struct
*p
)
1043 return cpu_curr(task_cpu(p
)) == p
;
1046 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1047 const struct sched_class
*prev_class
,
1050 if (prev_class
!= p
->sched_class
) {
1051 if (prev_class
->switched_from
)
1052 prev_class
->switched_from(rq
, p
);
1053 p
->sched_class
->switched_to(rq
, p
);
1054 } else if (oldprio
!= p
->prio
)
1055 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1058 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1060 const struct sched_class
*class;
1062 if (p
->sched_class
== rq
->curr
->sched_class
) {
1063 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1065 for_each_class(class) {
1066 if (class == rq
->curr
->sched_class
)
1068 if (class == p
->sched_class
) {
1069 resched_task(rq
->curr
);
1076 * A queue event has occurred, and we're going to schedule. In
1077 * this case, we can save a useless back to back clock update.
1079 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1080 rq
->skip_clock_update
= 1;
1084 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1086 #ifdef CONFIG_SCHED_DEBUG
1088 * We should never call set_task_cpu() on a blocked task,
1089 * ttwu() will sort out the placement.
1091 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1092 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
1094 #ifdef CONFIG_LOCKDEP
1096 * The caller should hold either p->pi_lock or rq->lock, when changing
1097 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1099 * sched_move_task() holds both and thus holding either pins the cgroup,
1100 * see set_task_rq().
1102 * Furthermore, all task_rq users should acquire both locks, see
1105 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1106 lockdep_is_held(&task_rq(p
)->lock
)));
1110 trace_sched_migrate_task(p
, new_cpu
);
1112 if (task_cpu(p
) != new_cpu
) {
1113 p
->se
.nr_migrations
++;
1114 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1117 __set_task_cpu(p
, new_cpu
);
1120 struct migration_arg
{
1121 struct task_struct
*task
;
1125 static int migration_cpu_stop(void *data
);
1128 * wait_task_inactive - wait for a thread to unschedule.
1130 * If @match_state is nonzero, it's the @p->state value just checked and
1131 * not expected to change. If it changes, i.e. @p might have woken up,
1132 * then return zero. When we succeed in waiting for @p to be off its CPU,
1133 * we return a positive number (its total switch count). If a second call
1134 * a short while later returns the same number, the caller can be sure that
1135 * @p has remained unscheduled the whole time.
1137 * The caller must ensure that the task *will* unschedule sometime soon,
1138 * else this function might spin for a *long* time. This function can't
1139 * be called with interrupts off, or it may introduce deadlock with
1140 * smp_call_function() if an IPI is sent by the same process we are
1141 * waiting to become inactive.
1143 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1145 unsigned long flags
;
1152 * We do the initial early heuristics without holding
1153 * any task-queue locks at all. We'll only try to get
1154 * the runqueue lock when things look like they will
1160 * If the task is actively running on another CPU
1161 * still, just relax and busy-wait without holding
1164 * NOTE! Since we don't hold any locks, it's not
1165 * even sure that "rq" stays as the right runqueue!
1166 * But we don't care, since "task_running()" will
1167 * return false if the runqueue has changed and p
1168 * is actually now running somewhere else!
1170 while (task_running(rq
, p
)) {
1171 if (match_state
&& unlikely(p
->state
!= match_state
))
1177 * Ok, time to look more closely! We need the rq
1178 * lock now, to be *sure*. If we're wrong, we'll
1179 * just go back and repeat.
1181 rq
= task_rq_lock(p
, &flags
);
1182 trace_sched_wait_task(p
);
1183 running
= task_running(rq
, p
);
1186 if (!match_state
|| p
->state
== match_state
)
1187 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1188 task_rq_unlock(rq
, p
, &flags
);
1191 * If it changed from the expected state, bail out now.
1193 if (unlikely(!ncsw
))
1197 * Was it really running after all now that we
1198 * checked with the proper locks actually held?
1200 * Oops. Go back and try again..
1202 if (unlikely(running
)) {
1208 * It's not enough that it's not actively running,
1209 * it must be off the runqueue _entirely_, and not
1212 * So if it was still runnable (but just not actively
1213 * running right now), it's preempted, and we should
1214 * yield - it could be a while.
1216 if (unlikely(on_rq
)) {
1217 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1219 set_current_state(TASK_UNINTERRUPTIBLE
);
1220 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1225 * Ahh, all good. It wasn't running, and it wasn't
1226 * runnable, which means that it will never become
1227 * running in the future either. We're all done!
1236 * kick_process - kick a running thread to enter/exit the kernel
1237 * @p: the to-be-kicked thread
1239 * Cause a process which is running on another CPU to enter
1240 * kernel-mode, without any delay. (to get signals handled.)
1242 * NOTE: this function doesn't have to take the runqueue lock,
1243 * because all it wants to ensure is that the remote task enters
1244 * the kernel. If the IPI races and the task has been migrated
1245 * to another CPU then no harm is done and the purpose has been
1248 void kick_process(struct task_struct
*p
)
1254 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1255 smp_send_reschedule(cpu
);
1258 EXPORT_SYMBOL_GPL(kick_process
);
1259 #endif /* CONFIG_SMP */
1263 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1265 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1267 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1268 enum { cpuset
, possible
, fail
} state
= cpuset
;
1271 /* Look for allowed, online CPU in same node. */
1272 for_each_cpu(dest_cpu
, nodemask
) {
1273 if (!cpu_online(dest_cpu
))
1275 if (!cpu_active(dest_cpu
))
1277 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1282 /* Any allowed, online CPU? */
1283 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1284 if (!cpu_online(dest_cpu
))
1286 if (!cpu_active(dest_cpu
))
1293 /* No more Mr. Nice Guy. */
1294 cpuset_cpus_allowed_fallback(p
);
1299 do_set_cpus_allowed(p
, cpu_possible_mask
);
1310 if (state
!= cpuset
) {
1312 * Don't tell them about moving exiting tasks or
1313 * kernel threads (both mm NULL), since they never
1316 if (p
->mm
&& printk_ratelimit()) {
1317 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1318 task_pid_nr(p
), p
->comm
, cpu
);
1326 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1329 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1331 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1334 * In order not to call set_task_cpu() on a blocking task we need
1335 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1338 * Since this is common to all placement strategies, this lives here.
1340 * [ this allows ->select_task() to simply return task_cpu(p) and
1341 * not worry about this generic constraint ]
1343 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1345 cpu
= select_fallback_rq(task_cpu(p
), p
);
1350 static void update_avg(u64
*avg
, u64 sample
)
1352 s64 diff
= sample
- *avg
;
1358 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1360 #ifdef CONFIG_SCHEDSTATS
1361 struct rq
*rq
= this_rq();
1364 int this_cpu
= smp_processor_id();
1366 if (cpu
== this_cpu
) {
1367 schedstat_inc(rq
, ttwu_local
);
1368 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1370 struct sched_domain
*sd
;
1372 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1374 for_each_domain(this_cpu
, sd
) {
1375 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1376 schedstat_inc(sd
, ttwu_wake_remote
);
1383 if (wake_flags
& WF_MIGRATED
)
1384 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1386 #endif /* CONFIG_SMP */
1388 schedstat_inc(rq
, ttwu_count
);
1389 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1391 if (wake_flags
& WF_SYNC
)
1392 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1394 #endif /* CONFIG_SCHEDSTATS */
1397 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1399 activate_task(rq
, p
, en_flags
);
1402 /* if a worker is waking up, notify workqueue */
1403 if (p
->flags
& PF_WQ_WORKER
)
1404 wq_worker_waking_up(p
, cpu_of(rq
));
1408 * Mark the task runnable and perform wakeup-preemption.
1411 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1413 trace_sched_wakeup(p
, true);
1414 check_preempt_curr(rq
, p
, wake_flags
);
1416 p
->state
= TASK_RUNNING
;
1418 if (p
->sched_class
->task_woken
)
1419 p
->sched_class
->task_woken(rq
, p
);
1421 if (rq
->idle_stamp
) {
1422 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1423 u64 max
= 2*sysctl_sched_migration_cost
;
1428 update_avg(&rq
->avg_idle
, delta
);
1435 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1438 if (p
->sched_contributes_to_load
)
1439 rq
->nr_uninterruptible
--;
1442 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1443 ttwu_do_wakeup(rq
, p
, wake_flags
);
1447 * Called in case the task @p isn't fully descheduled from its runqueue,
1448 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1449 * since all we need to do is flip p->state to TASK_RUNNING, since
1450 * the task is still ->on_rq.
1452 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1457 rq
= __task_rq_lock(p
);
1459 ttwu_do_wakeup(rq
, p
, wake_flags
);
1462 __task_rq_unlock(rq
);
1468 static void sched_ttwu_pending(void)
1470 struct rq
*rq
= this_rq();
1471 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1472 struct task_struct
*p
;
1474 raw_spin_lock(&rq
->lock
);
1477 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1478 llist
= llist_next(llist
);
1479 ttwu_do_activate(rq
, p
, 0);
1482 raw_spin_unlock(&rq
->lock
);
1485 void scheduler_ipi(void)
1487 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1491 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1492 * traditionally all their work was done from the interrupt return
1493 * path. Now that we actually do some work, we need to make sure
1496 * Some archs already do call them, luckily irq_enter/exit nest
1499 * Arguably we should visit all archs and update all handlers,
1500 * however a fair share of IPIs are still resched only so this would
1501 * somewhat pessimize the simple resched case.
1504 sched_ttwu_pending();
1507 * Check if someone kicked us for doing the nohz idle load balance.
1509 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1510 this_rq()->idle_balance
= 1;
1511 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1516 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1518 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1519 smp_send_reschedule(cpu
);
1522 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1523 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1528 rq
= __task_rq_lock(p
);
1530 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1531 ttwu_do_wakeup(rq
, p
, wake_flags
);
1534 __task_rq_unlock(rq
);
1539 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1541 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1543 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1545 #endif /* CONFIG_SMP */
1547 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1549 struct rq
*rq
= cpu_rq(cpu
);
1551 #if defined(CONFIG_SMP)
1552 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1553 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1554 ttwu_queue_remote(p
, cpu
);
1559 raw_spin_lock(&rq
->lock
);
1560 ttwu_do_activate(rq
, p
, 0);
1561 raw_spin_unlock(&rq
->lock
);
1565 * try_to_wake_up - wake up a thread
1566 * @p: the thread to be awakened
1567 * @state: the mask of task states that can be woken
1568 * @wake_flags: wake modifier flags (WF_*)
1570 * Put it on the run-queue if it's not already there. The "current"
1571 * thread is always on the run-queue (except when the actual
1572 * re-schedule is in progress), and as such you're allowed to do
1573 * the simpler "current->state = TASK_RUNNING" to mark yourself
1574 * runnable without the overhead of this.
1576 * Returns %true if @p was woken up, %false if it was already running
1577 * or @state didn't match @p's state.
1580 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1582 unsigned long flags
;
1583 int cpu
, success
= 0;
1586 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1587 if (!(p
->state
& state
))
1590 success
= 1; /* we're going to change ->state */
1593 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1598 * If the owning (remote) cpu is still in the middle of schedule() with
1599 * this task as prev, wait until its done referencing the task.
1602 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1604 * In case the architecture enables interrupts in
1605 * context_switch(), we cannot busy wait, since that
1606 * would lead to deadlocks when an interrupt hits and
1607 * tries to wake up @prev. So bail and do a complete
1610 if (ttwu_activate_remote(p
, wake_flags
))
1617 * Pairs with the smp_wmb() in finish_lock_switch().
1621 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1622 p
->state
= TASK_WAKING
;
1624 if (p
->sched_class
->task_waking
)
1625 p
->sched_class
->task_waking(p
);
1627 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1628 if (task_cpu(p
) != cpu
) {
1629 wake_flags
|= WF_MIGRATED
;
1630 set_task_cpu(p
, cpu
);
1632 #endif /* CONFIG_SMP */
1636 ttwu_stat(p
, cpu
, wake_flags
);
1638 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1644 * try_to_wake_up_local - try to wake up a local task with rq lock held
1645 * @p: the thread to be awakened
1647 * Put @p on the run-queue if it's not already there. The caller must
1648 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1651 static void try_to_wake_up_local(struct task_struct
*p
)
1653 struct rq
*rq
= task_rq(p
);
1655 BUG_ON(rq
!= this_rq());
1656 BUG_ON(p
== current
);
1657 lockdep_assert_held(&rq
->lock
);
1659 if (!raw_spin_trylock(&p
->pi_lock
)) {
1660 raw_spin_unlock(&rq
->lock
);
1661 raw_spin_lock(&p
->pi_lock
);
1662 raw_spin_lock(&rq
->lock
);
1665 if (!(p
->state
& TASK_NORMAL
))
1669 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1671 ttwu_do_wakeup(rq
, p
, 0);
1672 ttwu_stat(p
, smp_processor_id(), 0);
1674 raw_spin_unlock(&p
->pi_lock
);
1678 * wake_up_process - Wake up a specific process
1679 * @p: The process to be woken up.
1681 * Attempt to wake up the nominated process and move it to the set of runnable
1682 * processes. Returns 1 if the process was woken up, 0 if it was already
1685 * It may be assumed that this function implies a write memory barrier before
1686 * changing the task state if and only if any tasks are woken up.
1688 int wake_up_process(struct task_struct
*p
)
1690 return try_to_wake_up(p
, TASK_ALL
, 0);
1692 EXPORT_SYMBOL(wake_up_process
);
1694 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1696 return try_to_wake_up(p
, state
, 0);
1700 * Perform scheduler related setup for a newly forked process p.
1701 * p is forked by current.
1703 * __sched_fork() is basic setup used by init_idle() too:
1705 static void __sched_fork(struct task_struct
*p
)
1710 p
->se
.exec_start
= 0;
1711 p
->se
.sum_exec_runtime
= 0;
1712 p
->se
.prev_sum_exec_runtime
= 0;
1713 p
->se
.nr_migrations
= 0;
1715 INIT_LIST_HEAD(&p
->se
.group_node
);
1717 #ifdef CONFIG_SCHEDSTATS
1718 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1721 INIT_LIST_HEAD(&p
->rt
.run_list
);
1723 #ifdef CONFIG_PREEMPT_NOTIFIERS
1724 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1729 * fork()/clone()-time setup:
1731 void sched_fork(struct task_struct
*p
)
1733 unsigned long flags
;
1734 int cpu
= get_cpu();
1738 * We mark the process as running here. This guarantees that
1739 * nobody will actually run it, and a signal or other external
1740 * event cannot wake it up and insert it on the runqueue either.
1742 p
->state
= TASK_RUNNING
;
1745 * Make sure we do not leak PI boosting priority to the child.
1747 p
->prio
= current
->normal_prio
;
1750 * Revert to default priority/policy on fork if requested.
1752 if (unlikely(p
->sched_reset_on_fork
)) {
1753 if (task_has_rt_policy(p
)) {
1754 p
->policy
= SCHED_NORMAL
;
1755 p
->static_prio
= NICE_TO_PRIO(0);
1757 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1758 p
->static_prio
= NICE_TO_PRIO(0);
1760 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1764 * We don't need the reset flag anymore after the fork. It has
1765 * fulfilled its duty:
1767 p
->sched_reset_on_fork
= 0;
1770 if (!rt_prio(p
->prio
))
1771 p
->sched_class
= &fair_sched_class
;
1773 if (p
->sched_class
->task_fork
)
1774 p
->sched_class
->task_fork(p
);
1777 * The child is not yet in the pid-hash so no cgroup attach races,
1778 * and the cgroup is pinned to this child due to cgroup_fork()
1779 * is ran before sched_fork().
1781 * Silence PROVE_RCU.
1783 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1784 set_task_cpu(p
, cpu
);
1785 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1787 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1788 if (likely(sched_info_on()))
1789 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1791 #if defined(CONFIG_SMP)
1794 #ifdef CONFIG_PREEMPT_COUNT
1795 /* Want to start with kernel preemption disabled. */
1796 task_thread_info(p
)->preempt_count
= 1;
1799 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1806 * wake_up_new_task - wake up a newly created task for the first time.
1808 * This function will do some initial scheduler statistics housekeeping
1809 * that must be done for every newly created context, then puts the task
1810 * on the runqueue and wakes it.
1812 void wake_up_new_task(struct task_struct
*p
)
1814 unsigned long flags
;
1817 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1820 * Fork balancing, do it here and not earlier because:
1821 * - cpus_allowed can change in the fork path
1822 * - any previously selected cpu might disappear through hotplug
1824 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1827 rq
= __task_rq_lock(p
);
1828 activate_task(rq
, p
, 0);
1830 trace_sched_wakeup_new(p
, true);
1831 check_preempt_curr(rq
, p
, WF_FORK
);
1833 if (p
->sched_class
->task_woken
)
1834 p
->sched_class
->task_woken(rq
, p
);
1836 task_rq_unlock(rq
, p
, &flags
);
1839 #ifdef CONFIG_PREEMPT_NOTIFIERS
1842 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1843 * @notifier: notifier struct to register
1845 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1847 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1849 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1852 * preempt_notifier_unregister - no longer interested in preemption notifications
1853 * @notifier: notifier struct to unregister
1855 * This is safe to call from within a preemption notifier.
1857 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1859 hlist_del(¬ifier
->link
);
1861 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1863 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1865 struct preempt_notifier
*notifier
;
1866 struct hlist_node
*node
;
1868 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1869 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1873 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1874 struct task_struct
*next
)
1876 struct preempt_notifier
*notifier
;
1877 struct hlist_node
*node
;
1879 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1880 notifier
->ops
->sched_out(notifier
, next
);
1883 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1885 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1890 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1891 struct task_struct
*next
)
1895 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1898 * prepare_task_switch - prepare to switch tasks
1899 * @rq: the runqueue preparing to switch
1900 * @prev: the current task that is being switched out
1901 * @next: the task we are going to switch to.
1903 * This is called with the rq lock held and interrupts off. It must
1904 * be paired with a subsequent finish_task_switch after the context
1907 * prepare_task_switch sets up locking and calls architecture specific
1911 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1912 struct task_struct
*next
)
1914 sched_info_switch(prev
, next
);
1915 perf_event_task_sched_out(prev
, next
);
1916 fire_sched_out_preempt_notifiers(prev
, next
);
1917 prepare_lock_switch(rq
, next
);
1918 prepare_arch_switch(next
);
1919 trace_sched_switch(prev
, next
);
1923 * finish_task_switch - clean up after a task-switch
1924 * @rq: runqueue associated with task-switch
1925 * @prev: the thread we just switched away from.
1927 * finish_task_switch must be called after the context switch, paired
1928 * with a prepare_task_switch call before the context switch.
1929 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1930 * and do any other architecture-specific cleanup actions.
1932 * Note that we may have delayed dropping an mm in context_switch(). If
1933 * so, we finish that here outside of the runqueue lock. (Doing it
1934 * with the lock held can cause deadlocks; see schedule() for
1937 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1938 __releases(rq
->lock
)
1940 struct mm_struct
*mm
= rq
->prev_mm
;
1946 * A task struct has one reference for the use as "current".
1947 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1948 * schedule one last time. The schedule call will never return, and
1949 * the scheduled task must drop that reference.
1950 * The test for TASK_DEAD must occur while the runqueue locks are
1951 * still held, otherwise prev could be scheduled on another cpu, die
1952 * there before we look at prev->state, and then the reference would
1954 * Manfred Spraul <manfred@colorfullife.com>
1956 prev_state
= prev
->state
;
1957 finish_arch_switch(prev
);
1958 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1959 local_irq_disable();
1960 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1961 perf_event_task_sched_in(prev
, current
);
1962 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1964 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1965 finish_lock_switch(rq
, prev
);
1966 finish_arch_post_lock_switch();
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 rcu_switch_from(prev
);
2086 switch_to(prev
, next
, prev
);
2090 * this_rq must be evaluated again because prev may have moved
2091 * CPUs since it called schedule(), thus the 'rq' on its stack
2092 * frame will be invalid.
2094 finish_task_switch(this_rq(), prev
);
2098 * nr_running, nr_uninterruptible and nr_context_switches:
2100 * externally visible scheduler statistics: current number of runnable
2101 * threads, current number of uninterruptible-sleeping threads, total
2102 * number of context switches performed since bootup.
2104 unsigned long nr_running(void)
2106 unsigned long i
, sum
= 0;
2108 for_each_online_cpu(i
)
2109 sum
+= cpu_rq(i
)->nr_running
;
2114 unsigned long nr_uninterruptible(void)
2116 unsigned long i
, sum
= 0;
2118 for_each_possible_cpu(i
)
2119 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2122 * Since we read the counters lockless, it might be slightly
2123 * inaccurate. Do not allow it to go below zero though:
2125 if (unlikely((long)sum
< 0))
2131 unsigned long long nr_context_switches(void)
2134 unsigned long long sum
= 0;
2136 for_each_possible_cpu(i
)
2137 sum
+= cpu_rq(i
)->nr_switches
;
2142 unsigned long nr_iowait(void)
2144 unsigned long i
, sum
= 0;
2146 for_each_possible_cpu(i
)
2147 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2152 unsigned long nr_iowait_cpu(int cpu
)
2154 struct rq
*this = cpu_rq(cpu
);
2155 return atomic_read(&this->nr_iowait
);
2158 unsigned long this_cpu_load(void)
2160 struct rq
*this = this_rq();
2161 return this->cpu_load
[0];
2165 /* Variables and functions for calc_load */
2166 static atomic_long_t calc_load_tasks
;
2167 static unsigned long calc_load_update
;
2168 unsigned long avenrun
[3];
2169 EXPORT_SYMBOL(avenrun
);
2171 static long calc_load_fold_active(struct rq
*this_rq
)
2173 long nr_active
, delta
= 0;
2175 nr_active
= this_rq
->nr_running
;
2176 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2178 if (nr_active
!= this_rq
->calc_load_active
) {
2179 delta
= nr_active
- this_rq
->calc_load_active
;
2180 this_rq
->calc_load_active
= nr_active
;
2186 static unsigned long
2187 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2190 load
+= active
* (FIXED_1
- exp
);
2191 load
+= 1UL << (FSHIFT
- 1);
2192 return load
>> FSHIFT
;
2197 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2199 * When making the ILB scale, we should try to pull this in as well.
2201 static atomic_long_t calc_load_tasks_idle
;
2203 void calc_load_account_idle(struct rq
*this_rq
)
2207 delta
= calc_load_fold_active(this_rq
);
2209 atomic_long_add(delta
, &calc_load_tasks_idle
);
2212 static long calc_load_fold_idle(void)
2217 * Its got a race, we don't care...
2219 if (atomic_long_read(&calc_load_tasks_idle
))
2220 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2226 * fixed_power_int - compute: x^n, in O(log n) time
2228 * @x: base of the power
2229 * @frac_bits: fractional bits of @x
2230 * @n: power to raise @x to.
2232 * By exploiting the relation between the definition of the natural power
2233 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2234 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2235 * (where: n_i \elem {0, 1}, the binary vector representing n),
2236 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2237 * of course trivially computable in O(log_2 n), the length of our binary
2240 static unsigned long
2241 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2243 unsigned long result
= 1UL << frac_bits
;
2248 result
+= 1UL << (frac_bits
- 1);
2249 result
>>= frac_bits
;
2255 x
+= 1UL << (frac_bits
- 1);
2263 * a1 = a0 * e + a * (1 - e)
2265 * a2 = a1 * e + a * (1 - e)
2266 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2267 * = a0 * e^2 + a * (1 - e) * (1 + e)
2269 * a3 = a2 * e + a * (1 - e)
2270 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2271 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2275 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2276 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2277 * = a0 * e^n + a * (1 - e^n)
2279 * [1] application of the geometric series:
2282 * S_n := \Sum x^i = -------------
2285 static unsigned long
2286 calc_load_n(unsigned long load
, unsigned long exp
,
2287 unsigned long active
, unsigned int n
)
2290 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2294 * NO_HZ can leave us missing all per-cpu ticks calling
2295 * calc_load_account_active(), but since an idle CPU folds its delta into
2296 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2297 * in the pending idle delta if our idle period crossed a load cycle boundary.
2299 * Once we've updated the global active value, we need to apply the exponential
2300 * weights adjusted to the number of cycles missed.
2302 static void calc_global_nohz(void)
2304 long delta
, active
, n
;
2307 * If we crossed a calc_load_update boundary, make sure to fold
2308 * any pending idle changes, the respective CPUs might have
2309 * missed the tick driven calc_load_account_active() update
2312 delta
= calc_load_fold_idle();
2314 atomic_long_add(delta
, &calc_load_tasks
);
2317 * It could be the one fold was all it took, we done!
2319 if (time_before(jiffies
, calc_load_update
+ 10))
2323 * Catch-up, fold however many we are behind still
2325 delta
= jiffies
- calc_load_update
- 10;
2326 n
= 1 + (delta
/ LOAD_FREQ
);
2328 active
= atomic_long_read(&calc_load_tasks
);
2329 active
= active
> 0 ? active
* FIXED_1
: 0;
2331 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2332 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2333 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2335 calc_load_update
+= n
* LOAD_FREQ
;
2338 void calc_load_account_idle(struct rq
*this_rq
)
2342 static inline long calc_load_fold_idle(void)
2347 static void calc_global_nohz(void)
2353 * get_avenrun - get the load average array
2354 * @loads: pointer to dest load array
2355 * @offset: offset to add
2356 * @shift: shift count to shift the result left
2358 * These values are estimates at best, so no need for locking.
2360 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2362 loads
[0] = (avenrun
[0] + offset
) << shift
;
2363 loads
[1] = (avenrun
[1] + offset
) << shift
;
2364 loads
[2] = (avenrun
[2] + offset
) << shift
;
2368 * calc_load - update the avenrun load estimates 10 ticks after the
2369 * CPUs have updated calc_load_tasks.
2371 void calc_global_load(unsigned long ticks
)
2375 if (time_before(jiffies
, calc_load_update
+ 10))
2378 active
= atomic_long_read(&calc_load_tasks
);
2379 active
= active
> 0 ? active
* FIXED_1
: 0;
2381 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2382 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2383 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2385 calc_load_update
+= LOAD_FREQ
;
2388 * Account one period with whatever state we found before
2389 * folding in the nohz state and ageing the entire idle period.
2391 * This avoids loosing a sample when we go idle between
2392 * calc_load_account_active() (10 ticks ago) and now and thus
2399 * Called from update_cpu_load() to periodically update this CPU's
2402 static void calc_load_account_active(struct rq
*this_rq
)
2406 if (time_before(jiffies
, this_rq
->calc_load_update
))
2409 delta
= calc_load_fold_active(this_rq
);
2410 delta
+= calc_load_fold_idle();
2412 atomic_long_add(delta
, &calc_load_tasks
);
2414 this_rq
->calc_load_update
+= LOAD_FREQ
;
2418 * The exact cpuload at various idx values, calculated at every tick would be
2419 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2421 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2422 * on nth tick when cpu may be busy, then we have:
2423 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2424 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2426 * decay_load_missed() below does efficient calculation of
2427 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2428 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2430 * The calculation is approximated on a 128 point scale.
2431 * degrade_zero_ticks is the number of ticks after which load at any
2432 * particular idx is approximated to be zero.
2433 * degrade_factor is a precomputed table, a row for each load idx.
2434 * Each column corresponds to degradation factor for a power of two ticks,
2435 * based on 128 point scale.
2437 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2438 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2440 * With this power of 2 load factors, we can degrade the load n times
2441 * by looking at 1 bits in n and doing as many mult/shift instead of
2442 * n mult/shifts needed by the exact degradation.
2444 #define DEGRADE_SHIFT 7
2445 static const unsigned char
2446 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2447 static const unsigned char
2448 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2449 {0, 0, 0, 0, 0, 0, 0, 0},
2450 {64, 32, 8, 0, 0, 0, 0, 0},
2451 {96, 72, 40, 12, 1, 0, 0},
2452 {112, 98, 75, 43, 15, 1, 0},
2453 {120, 112, 98, 76, 45, 16, 2} };
2456 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2457 * would be when CPU is idle and so we just decay the old load without
2458 * adding any new load.
2460 static unsigned long
2461 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2465 if (!missed_updates
)
2468 if (missed_updates
>= degrade_zero_ticks
[idx
])
2472 return load
>> missed_updates
;
2474 while (missed_updates
) {
2475 if (missed_updates
% 2)
2476 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2478 missed_updates
>>= 1;
2485 * Update rq->cpu_load[] statistics. This function is usually called every
2486 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2487 * every tick. We fix it up based on jiffies.
2489 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2490 unsigned long pending_updates
)
2494 this_rq
->nr_load_updates
++;
2496 /* Update our load: */
2497 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2498 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2499 unsigned long old_load
, new_load
;
2501 /* scale is effectively 1 << i now, and >> i divides by scale */
2503 old_load
= this_rq
->cpu_load
[i
];
2504 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2505 new_load
= this_load
;
2507 * Round up the averaging division if load is increasing. This
2508 * prevents us from getting stuck on 9 if the load is 10, for
2511 if (new_load
> old_load
)
2512 new_load
+= scale
- 1;
2514 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2517 sched_avg_update(this_rq
);
2521 * Called from nohz_idle_balance() to update the load ratings before doing the
2524 void update_idle_cpu_load(struct rq
*this_rq
)
2526 unsigned long curr_jiffies
= jiffies
;
2527 unsigned long load
= this_rq
->load
.weight
;
2528 unsigned long pending_updates
;
2531 * Bloody broken means of dealing with nohz, but better than nothing..
2532 * jiffies is updated by one cpu, another cpu can drift wrt the jiffy
2533 * update and see 0 difference the one time and 2 the next, even though
2534 * we ticked at roughtly the same rate.
2536 * Hence we only use this from nohz_idle_balance() and skip this
2537 * nonsense when called from the scheduler_tick() since that's
2538 * guaranteed a stable rate.
2540 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2543 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2544 this_rq
->last_load_update_tick
= curr_jiffies
;
2546 __update_cpu_load(this_rq
, load
, pending_updates
);
2550 * Called from scheduler_tick()
2552 static void update_cpu_load_active(struct rq
*this_rq
)
2555 * See the mess in update_idle_cpu_load().
2557 this_rq
->last_load_update_tick
= jiffies
;
2558 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2560 calc_load_account_active(this_rq
);
2566 * sched_exec - execve() is a valuable balancing opportunity, because at
2567 * this point the task has the smallest effective memory and cache footprint.
2569 void sched_exec(void)
2571 struct task_struct
*p
= current
;
2572 unsigned long flags
;
2575 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2576 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2577 if (dest_cpu
== smp_processor_id())
2580 if (likely(cpu_active(dest_cpu
))) {
2581 struct migration_arg arg
= { p
, dest_cpu
};
2583 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2584 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2588 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2593 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2594 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2596 EXPORT_PER_CPU_SYMBOL(kstat
);
2597 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2600 * Return any ns on the sched_clock that have not yet been accounted in
2601 * @p in case that task is currently running.
2603 * Called with task_rq_lock() held on @rq.
2605 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2609 if (task_current(rq
, p
)) {
2610 update_rq_clock(rq
);
2611 ns
= rq
->clock_task
- p
->se
.exec_start
;
2619 unsigned long long task_delta_exec(struct task_struct
*p
)
2621 unsigned long flags
;
2625 rq
= task_rq_lock(p
, &flags
);
2626 ns
= do_task_delta_exec(p
, rq
);
2627 task_rq_unlock(rq
, p
, &flags
);
2633 * Return accounted runtime for the task.
2634 * In case the task is currently running, return the runtime plus current's
2635 * pending runtime that have not been accounted yet.
2637 unsigned long long task_sched_runtime(struct task_struct
*p
)
2639 unsigned long flags
;
2643 rq
= task_rq_lock(p
, &flags
);
2644 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2645 task_rq_unlock(rq
, p
, &flags
);
2650 #ifdef CONFIG_CGROUP_CPUACCT
2651 struct cgroup_subsys cpuacct_subsys
;
2652 struct cpuacct root_cpuacct
;
2655 static inline void task_group_account_field(struct task_struct
*p
, int index
,
2658 #ifdef CONFIG_CGROUP_CPUACCT
2659 struct kernel_cpustat
*kcpustat
;
2663 * Since all updates are sure to touch the root cgroup, we
2664 * get ourselves ahead and touch it first. If the root cgroup
2665 * is the only cgroup, then nothing else should be necessary.
2668 __get_cpu_var(kernel_cpustat
).cpustat
[index
] += tmp
;
2670 #ifdef CONFIG_CGROUP_CPUACCT
2671 if (unlikely(!cpuacct_subsys
.active
))
2676 while (ca
&& (ca
!= &root_cpuacct
)) {
2677 kcpustat
= this_cpu_ptr(ca
->cpustat
);
2678 kcpustat
->cpustat
[index
] += tmp
;
2687 * Account user cpu time to a process.
2688 * @p: the process that the cpu time gets accounted to
2689 * @cputime: the cpu time spent in user space since the last update
2690 * @cputime_scaled: cputime scaled by cpu frequency
2692 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
2693 cputime_t cputime_scaled
)
2697 /* Add user time to process. */
2698 p
->utime
+= cputime
;
2699 p
->utimescaled
+= cputime_scaled
;
2700 account_group_user_time(p
, cputime
);
2702 index
= (TASK_NICE(p
) > 0) ? CPUTIME_NICE
: CPUTIME_USER
;
2704 /* Add user time to cpustat. */
2705 task_group_account_field(p
, index
, (__force u64
) cputime
);
2707 /* Account for user time used */
2708 acct_update_integrals(p
);
2712 * Account guest cpu time to a process.
2713 * @p: the process that the cpu time gets accounted to
2714 * @cputime: the cpu time spent in virtual machine since the last update
2715 * @cputime_scaled: cputime scaled by cpu frequency
2717 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
2718 cputime_t cputime_scaled
)
2720 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2722 /* Add guest time to process. */
2723 p
->utime
+= cputime
;
2724 p
->utimescaled
+= cputime_scaled
;
2725 account_group_user_time(p
, cputime
);
2726 p
->gtime
+= cputime
;
2728 /* Add guest time to cpustat. */
2729 if (TASK_NICE(p
) > 0) {
2730 cpustat
[CPUTIME_NICE
] += (__force u64
) cputime
;
2731 cpustat
[CPUTIME_GUEST_NICE
] += (__force u64
) cputime
;
2733 cpustat
[CPUTIME_USER
] += (__force u64
) cputime
;
2734 cpustat
[CPUTIME_GUEST
] += (__force u64
) cputime
;
2739 * Account system cpu time to a process and desired cpustat field
2740 * @p: the process that the cpu time gets accounted to
2741 * @cputime: the cpu time spent in kernel space since the last update
2742 * @cputime_scaled: cputime scaled by cpu frequency
2743 * @target_cputime64: pointer to cpustat field that has to be updated
2746 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
2747 cputime_t cputime_scaled
, int index
)
2749 /* Add system time to process. */
2750 p
->stime
+= cputime
;
2751 p
->stimescaled
+= cputime_scaled
;
2752 account_group_system_time(p
, cputime
);
2754 /* Add system time to cpustat. */
2755 task_group_account_field(p
, index
, (__force u64
) cputime
);
2757 /* Account for system time used */
2758 acct_update_integrals(p
);
2762 * Account system cpu time to a process.
2763 * @p: the process that the cpu time gets accounted to
2764 * @hardirq_offset: the offset to subtract from hardirq_count()
2765 * @cputime: the cpu time spent in kernel space since the last update
2766 * @cputime_scaled: cputime scaled by cpu frequency
2768 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2769 cputime_t cputime
, cputime_t cputime_scaled
)
2773 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
2774 account_guest_time(p
, cputime
, cputime_scaled
);
2778 if (hardirq_count() - hardirq_offset
)
2779 index
= CPUTIME_IRQ
;
2780 else if (in_serving_softirq())
2781 index
= CPUTIME_SOFTIRQ
;
2783 index
= CPUTIME_SYSTEM
;
2785 __account_system_time(p
, cputime
, cputime_scaled
, index
);
2789 * Account for involuntary wait time.
2790 * @cputime: the cpu time spent in involuntary wait
2792 void account_steal_time(cputime_t cputime
)
2794 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2796 cpustat
[CPUTIME_STEAL
] += (__force u64
) cputime
;
2800 * Account for idle time.
2801 * @cputime: the cpu time spent in idle wait
2803 void account_idle_time(cputime_t cputime
)
2805 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2806 struct rq
*rq
= this_rq();
2808 if (atomic_read(&rq
->nr_iowait
) > 0)
2809 cpustat
[CPUTIME_IOWAIT
] += (__force u64
) cputime
;
2811 cpustat
[CPUTIME_IDLE
] += (__force u64
) cputime
;
2814 static __always_inline
bool steal_account_process_tick(void)
2816 #ifdef CONFIG_PARAVIRT
2817 if (static_key_false(¶virt_steal_enabled
)) {
2820 steal
= paravirt_steal_clock(smp_processor_id());
2821 steal
-= this_rq()->prev_steal_time
;
2823 st
= steal_ticks(steal
);
2824 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
2826 account_steal_time(st
);
2833 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2835 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2837 * Account a tick to a process and cpustat
2838 * @p: the process that the cpu time gets accounted to
2839 * @user_tick: is the tick from userspace
2840 * @rq: the pointer to rq
2842 * Tick demultiplexing follows the order
2843 * - pending hardirq update
2844 * - pending softirq update
2848 * - check for guest_time
2849 * - else account as system_time
2851 * Check for hardirq is done both for system and user time as there is
2852 * no timer going off while we are on hardirq and hence we may never get an
2853 * opportunity to update it solely in system time.
2854 * p->stime and friends are only updated on system time and not on irq
2855 * softirq as those do not count in task exec_runtime any more.
2857 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2860 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2861 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2863 if (steal_account_process_tick())
2866 if (irqtime_account_hi_update()) {
2867 cpustat
[CPUTIME_IRQ
] += (__force u64
) cputime_one_jiffy
;
2868 } else if (irqtime_account_si_update()) {
2869 cpustat
[CPUTIME_SOFTIRQ
] += (__force u64
) cputime_one_jiffy
;
2870 } else if (this_cpu_ksoftirqd() == p
) {
2872 * ksoftirqd time do not get accounted in cpu_softirq_time.
2873 * So, we have to handle it separately here.
2874 * Also, p->stime needs to be updated for ksoftirqd.
2876 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2878 } else if (user_tick
) {
2879 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2880 } else if (p
== rq
->idle
) {
2881 account_idle_time(cputime_one_jiffy
);
2882 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
2883 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2885 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2890 static void irqtime_account_idle_ticks(int ticks
)
2893 struct rq
*rq
= this_rq();
2895 for (i
= 0; i
< ticks
; i
++)
2896 irqtime_account_process_tick(current
, 0, rq
);
2898 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2899 static void irqtime_account_idle_ticks(int ticks
) {}
2900 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2902 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2905 * Account a single tick of cpu time.
2906 * @p: the process that the cpu time gets accounted to
2907 * @user_tick: indicates if the tick is a user or a system tick
2909 void account_process_tick(struct task_struct
*p
, int user_tick
)
2911 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2912 struct rq
*rq
= this_rq();
2914 if (sched_clock_irqtime
) {
2915 irqtime_account_process_tick(p
, user_tick
, rq
);
2919 if (steal_account_process_tick())
2923 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2924 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
2925 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
2928 account_idle_time(cputime_one_jiffy
);
2932 * Account multiple ticks of steal time.
2933 * @p: the process from which the cpu time has been stolen
2934 * @ticks: number of stolen ticks
2936 void account_steal_ticks(unsigned long ticks
)
2938 account_steal_time(jiffies_to_cputime(ticks
));
2942 * Account multiple ticks of idle time.
2943 * @ticks: number of stolen ticks
2945 void account_idle_ticks(unsigned long ticks
)
2948 if (sched_clock_irqtime
) {
2949 irqtime_account_idle_ticks(ticks
);
2953 account_idle_time(jiffies_to_cputime(ticks
));
2959 * Use precise platform statistics if available:
2961 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2962 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2968 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2970 struct task_cputime cputime
;
2972 thread_group_cputime(p
, &cputime
);
2974 *ut
= cputime
.utime
;
2975 *st
= cputime
.stime
;
2979 #ifndef nsecs_to_cputime
2980 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2983 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2985 cputime_t rtime
, utime
= p
->utime
, total
= utime
+ p
->stime
;
2988 * Use CFS's precise accounting:
2990 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
2993 u64 temp
= (__force u64
) rtime
;
2995 temp
*= (__force u64
) utime
;
2996 do_div(temp
, (__force u32
) total
);
2997 utime
= (__force cputime_t
) temp
;
3002 * Compare with previous values, to keep monotonicity:
3004 p
->prev_utime
= max(p
->prev_utime
, utime
);
3005 p
->prev_stime
= max(p
->prev_stime
, rtime
- p
->prev_utime
);
3007 *ut
= p
->prev_utime
;
3008 *st
= p
->prev_stime
;
3012 * Must be called with siglock held.
3014 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3016 struct signal_struct
*sig
= p
->signal
;
3017 struct task_cputime cputime
;
3018 cputime_t rtime
, utime
, total
;
3020 thread_group_cputime(p
, &cputime
);
3022 total
= cputime
.utime
+ cputime
.stime
;
3023 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3026 u64 temp
= (__force u64
) rtime
;
3028 temp
*= (__force u64
) cputime
.utime
;
3029 do_div(temp
, (__force u32
) total
);
3030 utime
= (__force cputime_t
) temp
;
3034 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3035 sig
->prev_stime
= max(sig
->prev_stime
, rtime
- sig
->prev_utime
);
3037 *ut
= sig
->prev_utime
;
3038 *st
= sig
->prev_stime
;
3043 * This function gets called by the timer code, with HZ frequency.
3044 * We call it with interrupts disabled.
3046 void scheduler_tick(void)
3048 int cpu
= smp_processor_id();
3049 struct rq
*rq
= cpu_rq(cpu
);
3050 struct task_struct
*curr
= rq
->curr
;
3054 raw_spin_lock(&rq
->lock
);
3055 update_rq_clock(rq
);
3056 update_cpu_load_active(rq
);
3057 curr
->sched_class
->task_tick(rq
, curr
, 0);
3058 raw_spin_unlock(&rq
->lock
);
3060 perf_event_task_tick();
3063 rq
->idle_balance
= idle_cpu(cpu
);
3064 trigger_load_balance(rq
, cpu
);
3068 notrace
unsigned long get_parent_ip(unsigned long addr
)
3070 if (in_lock_functions(addr
)) {
3071 addr
= CALLER_ADDR2
;
3072 if (in_lock_functions(addr
))
3073 addr
= CALLER_ADDR3
;
3078 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3079 defined(CONFIG_PREEMPT_TRACER))
3081 void __kprobes
add_preempt_count(int val
)
3083 #ifdef CONFIG_DEBUG_PREEMPT
3087 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3090 preempt_count() += val
;
3091 #ifdef CONFIG_DEBUG_PREEMPT
3093 * Spinlock count overflowing soon?
3095 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3098 if (preempt_count() == val
)
3099 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3101 EXPORT_SYMBOL(add_preempt_count
);
3103 void __kprobes
sub_preempt_count(int val
)
3105 #ifdef CONFIG_DEBUG_PREEMPT
3109 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3112 * Is the spinlock portion underflowing?
3114 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3115 !(preempt_count() & PREEMPT_MASK
)))
3119 if (preempt_count() == val
)
3120 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3121 preempt_count() -= val
;
3123 EXPORT_SYMBOL(sub_preempt_count
);
3128 * Print scheduling while atomic bug:
3130 static noinline
void __schedule_bug(struct task_struct
*prev
)
3132 if (oops_in_progress
)
3135 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3136 prev
->comm
, prev
->pid
, preempt_count());
3138 debug_show_held_locks(prev
);
3140 if (irqs_disabled())
3141 print_irqtrace_events(prev
);
3143 add_taint(TAINT_WARN
);
3147 * Various schedule()-time debugging checks and statistics:
3149 static inline void schedule_debug(struct task_struct
*prev
)
3152 * Test if we are atomic. Since do_exit() needs to call into
3153 * schedule() atomically, we ignore that path for now.
3154 * Otherwise, whine if we are scheduling when we should not be.
3156 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3157 __schedule_bug(prev
);
3160 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3162 schedstat_inc(this_rq(), sched_count
);
3165 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3167 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
3168 update_rq_clock(rq
);
3169 prev
->sched_class
->put_prev_task(rq
, prev
);
3173 * Pick up the highest-prio task:
3175 static inline struct task_struct
*
3176 pick_next_task(struct rq
*rq
)
3178 const struct sched_class
*class;
3179 struct task_struct
*p
;
3182 * Optimization: we know that if all tasks are in
3183 * the fair class we can call that function directly:
3185 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3186 p
= fair_sched_class
.pick_next_task(rq
);
3191 for_each_class(class) {
3192 p
= class->pick_next_task(rq
);
3197 BUG(); /* the idle class will always have a runnable task */
3201 * __schedule() is the main scheduler function.
3203 static void __sched
__schedule(void)
3205 struct task_struct
*prev
, *next
;
3206 unsigned long *switch_count
;
3212 cpu
= smp_processor_id();
3214 rcu_note_context_switch(cpu
);
3217 schedule_debug(prev
);
3219 if (sched_feat(HRTICK
))
3222 raw_spin_lock_irq(&rq
->lock
);
3224 switch_count
= &prev
->nivcsw
;
3225 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3226 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3227 prev
->state
= TASK_RUNNING
;
3229 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3233 * If a worker went to sleep, notify and ask workqueue
3234 * whether it wants to wake up a task to maintain
3237 if (prev
->flags
& PF_WQ_WORKER
) {
3238 struct task_struct
*to_wakeup
;
3240 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3242 try_to_wake_up_local(to_wakeup
);
3245 switch_count
= &prev
->nvcsw
;
3248 pre_schedule(rq
, prev
);
3250 if (unlikely(!rq
->nr_running
))
3251 idle_balance(cpu
, rq
);
3253 put_prev_task(rq
, prev
);
3254 next
= pick_next_task(rq
);
3255 clear_tsk_need_resched(prev
);
3256 rq
->skip_clock_update
= 0;
3258 if (likely(prev
!= next
)) {
3263 context_switch(rq
, prev
, next
); /* unlocks the rq */
3265 * The context switch have flipped the stack from under us
3266 * and restored the local variables which were saved when
3267 * this task called schedule() in the past. prev == current
3268 * is still correct, but it can be moved to another cpu/rq.
3270 cpu
= smp_processor_id();
3273 raw_spin_unlock_irq(&rq
->lock
);
3277 sched_preempt_enable_no_resched();
3282 static inline void sched_submit_work(struct task_struct
*tsk
)
3284 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3287 * If we are going to sleep and we have plugged IO queued,
3288 * make sure to submit it to avoid deadlocks.
3290 if (blk_needs_flush_plug(tsk
))
3291 blk_schedule_flush_plug(tsk
);
3294 asmlinkage
void __sched
schedule(void)
3296 struct task_struct
*tsk
= current
;
3298 sched_submit_work(tsk
);
3301 EXPORT_SYMBOL(schedule
);
3304 * schedule_preempt_disabled - called with preemption disabled
3306 * Returns with preemption disabled. Note: preempt_count must be 1
3308 void __sched
schedule_preempt_disabled(void)
3310 sched_preempt_enable_no_resched();
3315 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3317 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3319 if (lock
->owner
!= owner
)
3323 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3324 * lock->owner still matches owner, if that fails, owner might
3325 * point to free()d memory, if it still matches, the rcu_read_lock()
3326 * ensures the memory stays valid.
3330 return owner
->on_cpu
;
3334 * Look out! "owner" is an entirely speculative pointer
3335 * access and not reliable.
3337 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3339 if (!sched_feat(OWNER_SPIN
))
3343 while (owner_running(lock
, owner
)) {
3347 arch_mutex_cpu_relax();
3352 * We break out the loop above on need_resched() and when the
3353 * owner changed, which is a sign for heavy contention. Return
3354 * success only when lock->owner is NULL.
3356 return lock
->owner
== NULL
;
3360 #ifdef CONFIG_PREEMPT
3362 * this is the entry point to schedule() from in-kernel preemption
3363 * off of preempt_enable. Kernel preemptions off return from interrupt
3364 * occur there and call schedule directly.
3366 asmlinkage
void __sched notrace
preempt_schedule(void)
3368 struct thread_info
*ti
= current_thread_info();
3371 * If there is a non-zero preempt_count or interrupts are disabled,
3372 * we do not want to preempt the current task. Just return..
3374 if (likely(ti
->preempt_count
|| irqs_disabled()))
3378 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3380 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3383 * Check again in case we missed a preemption opportunity
3384 * between schedule and now.
3387 } while (need_resched());
3389 EXPORT_SYMBOL(preempt_schedule
);
3392 * this is the entry point to schedule() from kernel preemption
3393 * off of irq context.
3394 * Note, that this is called and return with irqs disabled. This will
3395 * protect us against recursive calling from irq.
3397 asmlinkage
void __sched
preempt_schedule_irq(void)
3399 struct thread_info
*ti
= current_thread_info();
3401 /* Catch callers which need to be fixed */
3402 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3405 add_preempt_count(PREEMPT_ACTIVE
);
3408 local_irq_disable();
3409 sub_preempt_count(PREEMPT_ACTIVE
);
3412 * Check again in case we missed a preemption opportunity
3413 * between schedule and now.
3416 } while (need_resched());
3419 #endif /* CONFIG_PREEMPT */
3421 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3424 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3426 EXPORT_SYMBOL(default_wake_function
);
3429 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3430 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3431 * number) then we wake all the non-exclusive tasks and one exclusive task.
3433 * There are circumstances in which we can try to wake a task which has already
3434 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3435 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3437 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3438 int nr_exclusive
, int wake_flags
, void *key
)
3440 wait_queue_t
*curr
, *next
;
3442 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3443 unsigned flags
= curr
->flags
;
3445 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3446 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3452 * __wake_up - wake up threads blocked on a waitqueue.
3454 * @mode: which threads
3455 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3456 * @key: is directly passed to the wakeup function
3458 * It may be assumed that this function implies a write memory barrier before
3459 * changing the task state if and only if any tasks are woken up.
3461 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3462 int nr_exclusive
, void *key
)
3464 unsigned long flags
;
3466 spin_lock_irqsave(&q
->lock
, flags
);
3467 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3468 spin_unlock_irqrestore(&q
->lock
, flags
);
3470 EXPORT_SYMBOL(__wake_up
);
3473 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3475 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3477 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3479 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3481 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3483 __wake_up_common(q
, mode
, 1, 0, key
);
3485 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3488 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3490 * @mode: which threads
3491 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3492 * @key: opaque value to be passed to wakeup targets
3494 * The sync wakeup differs that the waker knows that it will schedule
3495 * away soon, so while the target thread will be woken up, it will not
3496 * be migrated to another CPU - ie. the two threads are 'synchronized'
3497 * with each other. This can prevent needless bouncing between CPUs.
3499 * On UP it can prevent extra preemption.
3501 * It may be assumed that this function implies a write memory barrier before
3502 * changing the task state if and only if any tasks are woken up.
3504 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3505 int nr_exclusive
, void *key
)
3507 unsigned long flags
;
3508 int wake_flags
= WF_SYNC
;
3513 if (unlikely(!nr_exclusive
))
3516 spin_lock_irqsave(&q
->lock
, flags
);
3517 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3518 spin_unlock_irqrestore(&q
->lock
, flags
);
3520 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3523 * __wake_up_sync - see __wake_up_sync_key()
3525 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3527 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3529 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3532 * complete: - signals a single thread waiting on this completion
3533 * @x: holds the state of this particular completion
3535 * This will wake up a single thread waiting on this completion. Threads will be
3536 * awakened in the same order in which they were queued.
3538 * See also complete_all(), wait_for_completion() and related routines.
3540 * It may be assumed that this function implies a write memory barrier before
3541 * changing the task state if and only if any tasks are woken up.
3543 void complete(struct completion
*x
)
3545 unsigned long flags
;
3547 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3549 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3550 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3552 EXPORT_SYMBOL(complete
);
3555 * complete_all: - signals all threads waiting on this completion
3556 * @x: holds the state of this particular completion
3558 * This will wake up all threads waiting on this particular completion event.
3560 * It may be assumed that this function implies a write memory barrier before
3561 * changing the task state if and only if any tasks are woken up.
3563 void complete_all(struct completion
*x
)
3565 unsigned long flags
;
3567 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3568 x
->done
+= UINT_MAX
/2;
3569 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3570 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3572 EXPORT_SYMBOL(complete_all
);
3574 static inline long __sched
3575 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3578 DECLARE_WAITQUEUE(wait
, current
);
3580 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3582 if (signal_pending_state(state
, current
)) {
3583 timeout
= -ERESTARTSYS
;
3586 __set_current_state(state
);
3587 spin_unlock_irq(&x
->wait
.lock
);
3588 timeout
= schedule_timeout(timeout
);
3589 spin_lock_irq(&x
->wait
.lock
);
3590 } while (!x
->done
&& timeout
);
3591 __remove_wait_queue(&x
->wait
, &wait
);
3596 return timeout
?: 1;
3600 wait_for_common(struct completion
*x
, long timeout
, int state
)
3604 spin_lock_irq(&x
->wait
.lock
);
3605 timeout
= do_wait_for_common(x
, timeout
, state
);
3606 spin_unlock_irq(&x
->wait
.lock
);
3611 * wait_for_completion: - waits for completion of a task
3612 * @x: holds the state of this particular completion
3614 * This waits to be signaled for completion of a specific task. It is NOT
3615 * interruptible and there is no timeout.
3617 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3618 * and interrupt capability. Also see complete().
3620 void __sched
wait_for_completion(struct completion
*x
)
3622 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3624 EXPORT_SYMBOL(wait_for_completion
);
3627 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3628 * @x: holds the state of this particular completion
3629 * @timeout: timeout value in jiffies
3631 * This waits for either a completion of a specific task to be signaled or for a
3632 * specified timeout to expire. The timeout is in jiffies. It is not
3635 * The return value is 0 if timed out, and positive (at least 1, or number of
3636 * jiffies left till timeout) if completed.
3638 unsigned long __sched
3639 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3641 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3643 EXPORT_SYMBOL(wait_for_completion_timeout
);
3646 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3647 * @x: holds the state of this particular completion
3649 * This waits for completion of a specific task to be signaled. It is
3652 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3654 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3656 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3657 if (t
== -ERESTARTSYS
)
3661 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3664 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3665 * @x: holds the state of this particular completion
3666 * @timeout: timeout value in jiffies
3668 * This waits for either a completion of a specific task to be signaled or for a
3669 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3671 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3672 * positive (at least 1, or number of jiffies left till timeout) if completed.
3675 wait_for_completion_interruptible_timeout(struct completion
*x
,
3676 unsigned long timeout
)
3678 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3680 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3683 * wait_for_completion_killable: - waits for completion of a task (killable)
3684 * @x: holds the state of this particular completion
3686 * This waits to be signaled for completion of a specific task. It can be
3687 * interrupted by a kill signal.
3689 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3691 int __sched
wait_for_completion_killable(struct completion
*x
)
3693 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3694 if (t
== -ERESTARTSYS
)
3698 EXPORT_SYMBOL(wait_for_completion_killable
);
3701 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3702 * @x: holds the state of this particular completion
3703 * @timeout: timeout value in jiffies
3705 * This waits for either a completion of a specific task to be
3706 * signaled or for a specified timeout to expire. It can be
3707 * interrupted by a kill signal. The timeout is in jiffies.
3709 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3710 * positive (at least 1, or number of jiffies left till timeout) if completed.
3713 wait_for_completion_killable_timeout(struct completion
*x
,
3714 unsigned long timeout
)
3716 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3718 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3721 * try_wait_for_completion - try to decrement a completion without blocking
3722 * @x: completion structure
3724 * Returns: 0 if a decrement cannot be done without blocking
3725 * 1 if a decrement succeeded.
3727 * If a completion is being used as a counting completion,
3728 * attempt to decrement the counter without blocking. This
3729 * enables us to avoid waiting if the resource the completion
3730 * is protecting is not available.
3732 bool try_wait_for_completion(struct completion
*x
)
3734 unsigned long flags
;
3737 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3742 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3745 EXPORT_SYMBOL(try_wait_for_completion
);
3748 * completion_done - Test to see if a completion has any waiters
3749 * @x: completion structure
3751 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3752 * 1 if there are no waiters.
3755 bool completion_done(struct completion
*x
)
3757 unsigned long flags
;
3760 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3763 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3766 EXPORT_SYMBOL(completion_done
);
3769 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3771 unsigned long flags
;
3774 init_waitqueue_entry(&wait
, current
);
3776 __set_current_state(state
);
3778 spin_lock_irqsave(&q
->lock
, flags
);
3779 __add_wait_queue(q
, &wait
);
3780 spin_unlock(&q
->lock
);
3781 timeout
= schedule_timeout(timeout
);
3782 spin_lock_irq(&q
->lock
);
3783 __remove_wait_queue(q
, &wait
);
3784 spin_unlock_irqrestore(&q
->lock
, flags
);
3789 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3791 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3793 EXPORT_SYMBOL(interruptible_sleep_on
);
3796 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3798 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3800 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3802 void __sched
sleep_on(wait_queue_head_t
*q
)
3804 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3806 EXPORT_SYMBOL(sleep_on
);
3808 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3810 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3812 EXPORT_SYMBOL(sleep_on_timeout
);
3814 #ifdef CONFIG_RT_MUTEXES
3817 * rt_mutex_setprio - set the current priority of a task
3819 * @prio: prio value (kernel-internal form)
3821 * This function changes the 'effective' priority of a task. It does
3822 * not touch ->normal_prio like __setscheduler().
3824 * Used by the rt_mutex code to implement priority inheritance logic.
3826 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3828 int oldprio
, on_rq
, running
;
3830 const struct sched_class
*prev_class
;
3832 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3834 rq
= __task_rq_lock(p
);
3837 * Idle task boosting is a nono in general. There is one
3838 * exception, when PREEMPT_RT and NOHZ is active:
3840 * The idle task calls get_next_timer_interrupt() and holds
3841 * the timer wheel base->lock on the CPU and another CPU wants
3842 * to access the timer (probably to cancel it). We can safely
3843 * ignore the boosting request, as the idle CPU runs this code
3844 * with interrupts disabled and will complete the lock
3845 * protected section without being interrupted. So there is no
3846 * real need to boost.
3848 if (unlikely(p
== rq
->idle
)) {
3849 WARN_ON(p
!= rq
->curr
);
3850 WARN_ON(p
->pi_blocked_on
);
3854 trace_sched_pi_setprio(p
, prio
);
3856 prev_class
= p
->sched_class
;
3858 running
= task_current(rq
, p
);
3860 dequeue_task(rq
, p
, 0);
3862 p
->sched_class
->put_prev_task(rq
, p
);
3865 p
->sched_class
= &rt_sched_class
;
3867 p
->sched_class
= &fair_sched_class
;
3872 p
->sched_class
->set_curr_task(rq
);
3874 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3876 check_class_changed(rq
, p
, prev_class
, oldprio
);
3878 __task_rq_unlock(rq
);
3881 void set_user_nice(struct task_struct
*p
, long nice
)
3883 int old_prio
, delta
, on_rq
;
3884 unsigned long flags
;
3887 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3890 * We have to be careful, if called from sys_setpriority(),
3891 * the task might be in the middle of scheduling on another CPU.
3893 rq
= task_rq_lock(p
, &flags
);
3895 * The RT priorities are set via sched_setscheduler(), but we still
3896 * allow the 'normal' nice value to be set - but as expected
3897 * it wont have any effect on scheduling until the task is
3898 * SCHED_FIFO/SCHED_RR:
3900 if (task_has_rt_policy(p
)) {
3901 p
->static_prio
= NICE_TO_PRIO(nice
);
3906 dequeue_task(rq
, p
, 0);
3908 p
->static_prio
= NICE_TO_PRIO(nice
);
3911 p
->prio
= effective_prio(p
);
3912 delta
= p
->prio
- old_prio
;
3915 enqueue_task(rq
, p
, 0);
3917 * If the task increased its priority or is running and
3918 * lowered its priority, then reschedule its CPU:
3920 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3921 resched_task(rq
->curr
);
3924 task_rq_unlock(rq
, p
, &flags
);
3926 EXPORT_SYMBOL(set_user_nice
);
3929 * can_nice - check if a task can reduce its nice value
3933 int can_nice(const struct task_struct
*p
, const int nice
)
3935 /* convert nice value [19,-20] to rlimit style value [1,40] */
3936 int nice_rlim
= 20 - nice
;
3938 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3939 capable(CAP_SYS_NICE
));
3942 #ifdef __ARCH_WANT_SYS_NICE
3945 * sys_nice - change the priority of the current process.
3946 * @increment: priority increment
3948 * sys_setpriority is a more generic, but much slower function that
3949 * does similar things.
3951 SYSCALL_DEFINE1(nice
, int, increment
)
3956 * Setpriority might change our priority at the same moment.
3957 * We don't have to worry. Conceptually one call occurs first
3958 * and we have a single winner.
3960 if (increment
< -40)
3965 nice
= TASK_NICE(current
) + increment
;
3971 if (increment
< 0 && !can_nice(current
, nice
))
3974 retval
= security_task_setnice(current
, nice
);
3978 set_user_nice(current
, nice
);
3985 * task_prio - return the priority value of a given task.
3986 * @p: the task in question.
3988 * This is the priority value as seen by users in /proc.
3989 * RT tasks are offset by -200. Normal tasks are centered
3990 * around 0, value goes from -16 to +15.
3992 int task_prio(const struct task_struct
*p
)
3994 return p
->prio
- MAX_RT_PRIO
;
3998 * task_nice - return the nice value of a given task.
3999 * @p: the task in question.
4001 int task_nice(const struct task_struct
*p
)
4003 return TASK_NICE(p
);
4005 EXPORT_SYMBOL(task_nice
);
4008 * idle_cpu - is a given cpu idle currently?
4009 * @cpu: the processor in question.
4011 int idle_cpu(int cpu
)
4013 struct rq
*rq
= cpu_rq(cpu
);
4015 if (rq
->curr
!= rq
->idle
)
4022 if (!llist_empty(&rq
->wake_list
))
4030 * idle_task - return the idle task for a given cpu.
4031 * @cpu: the processor in question.
4033 struct task_struct
*idle_task(int cpu
)
4035 return cpu_rq(cpu
)->idle
;
4039 * find_process_by_pid - find a process with a matching PID value.
4040 * @pid: the pid in question.
4042 static struct task_struct
*find_process_by_pid(pid_t pid
)
4044 return pid
? find_task_by_vpid(pid
) : current
;
4047 /* Actually do priority change: must hold rq lock. */
4049 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4052 p
->rt_priority
= prio
;
4053 p
->normal_prio
= normal_prio(p
);
4054 /* we are holding p->pi_lock already */
4055 p
->prio
= rt_mutex_getprio(p
);
4056 if (rt_prio(p
->prio
))
4057 p
->sched_class
= &rt_sched_class
;
4059 p
->sched_class
= &fair_sched_class
;
4064 * check the target process has a UID that matches the current process's
4066 static bool check_same_owner(struct task_struct
*p
)
4068 const struct cred
*cred
= current_cred(), *pcred
;
4072 pcred
= __task_cred(p
);
4073 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4074 uid_eq(cred
->euid
, pcred
->uid
));
4079 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4080 const struct sched_param
*param
, bool user
)
4082 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4083 unsigned long flags
;
4084 const struct sched_class
*prev_class
;
4088 /* may grab non-irq protected spin_locks */
4089 BUG_ON(in_interrupt());
4091 /* double check policy once rq lock held */
4093 reset_on_fork
= p
->sched_reset_on_fork
;
4094 policy
= oldpolicy
= p
->policy
;
4096 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4097 policy
&= ~SCHED_RESET_ON_FORK
;
4099 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4100 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4101 policy
!= SCHED_IDLE
)
4106 * Valid priorities for SCHED_FIFO and SCHED_RR are
4107 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4108 * SCHED_BATCH and SCHED_IDLE is 0.
4110 if (param
->sched_priority
< 0 ||
4111 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4112 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4114 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4118 * Allow unprivileged RT tasks to decrease priority:
4120 if (user
&& !capable(CAP_SYS_NICE
)) {
4121 if (rt_policy(policy
)) {
4122 unsigned long rlim_rtprio
=
4123 task_rlimit(p
, RLIMIT_RTPRIO
);
4125 /* can't set/change the rt policy */
4126 if (policy
!= p
->policy
&& !rlim_rtprio
)
4129 /* can't increase priority */
4130 if (param
->sched_priority
> p
->rt_priority
&&
4131 param
->sched_priority
> rlim_rtprio
)
4136 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4137 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4139 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
4140 if (!can_nice(p
, TASK_NICE(p
)))
4144 /* can't change other user's priorities */
4145 if (!check_same_owner(p
))
4148 /* Normal users shall not reset the sched_reset_on_fork flag */
4149 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4154 retval
= security_task_setscheduler(p
);
4160 * make sure no PI-waiters arrive (or leave) while we are
4161 * changing the priority of the task:
4163 * To be able to change p->policy safely, the appropriate
4164 * runqueue lock must be held.
4166 rq
= task_rq_lock(p
, &flags
);
4169 * Changing the policy of the stop threads its a very bad idea
4171 if (p
== rq
->stop
) {
4172 task_rq_unlock(rq
, p
, &flags
);
4177 * If not changing anything there's no need to proceed further:
4179 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
4180 param
->sched_priority
== p
->rt_priority
))) {
4182 __task_rq_unlock(rq
);
4183 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4187 #ifdef CONFIG_RT_GROUP_SCHED
4190 * Do not allow realtime tasks into groups that have no runtime
4193 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4194 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4195 !task_group_is_autogroup(task_group(p
))) {
4196 task_rq_unlock(rq
, p
, &flags
);
4202 /* recheck policy now with rq lock held */
4203 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4204 policy
= oldpolicy
= -1;
4205 task_rq_unlock(rq
, p
, &flags
);
4209 running
= task_current(rq
, p
);
4211 dequeue_task(rq
, p
, 0);
4213 p
->sched_class
->put_prev_task(rq
, p
);
4215 p
->sched_reset_on_fork
= reset_on_fork
;
4218 prev_class
= p
->sched_class
;
4219 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4222 p
->sched_class
->set_curr_task(rq
);
4224 enqueue_task(rq
, p
, 0);
4226 check_class_changed(rq
, p
, prev_class
, oldprio
);
4227 task_rq_unlock(rq
, p
, &flags
);
4229 rt_mutex_adjust_pi(p
);
4235 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4236 * @p: the task in question.
4237 * @policy: new policy.
4238 * @param: structure containing the new RT priority.
4240 * NOTE that the task may be already dead.
4242 int sched_setscheduler(struct task_struct
*p
, int policy
,
4243 const struct sched_param
*param
)
4245 return __sched_setscheduler(p
, policy
, param
, true);
4247 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4250 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4251 * @p: the task in question.
4252 * @policy: new policy.
4253 * @param: structure containing the new RT priority.
4255 * Just like sched_setscheduler, only don't bother checking if the
4256 * current context has permission. For example, this is needed in
4257 * stop_machine(): we create temporary high priority worker threads,
4258 * but our caller might not have that capability.
4260 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4261 const struct sched_param
*param
)
4263 return __sched_setscheduler(p
, policy
, param
, false);
4267 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4269 struct sched_param lparam
;
4270 struct task_struct
*p
;
4273 if (!param
|| pid
< 0)
4275 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4280 p
= find_process_by_pid(pid
);
4282 retval
= sched_setscheduler(p
, policy
, &lparam
);
4289 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4290 * @pid: the pid in question.
4291 * @policy: new policy.
4292 * @param: structure containing the new RT priority.
4294 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4295 struct sched_param __user
*, param
)
4297 /* negative values for policy are not valid */
4301 return do_sched_setscheduler(pid
, policy
, param
);
4305 * sys_sched_setparam - set/change the RT priority of a thread
4306 * @pid: the pid in question.
4307 * @param: structure containing the new RT priority.
4309 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4311 return do_sched_setscheduler(pid
, -1, param
);
4315 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4316 * @pid: the pid in question.
4318 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4320 struct task_struct
*p
;
4328 p
= find_process_by_pid(pid
);
4330 retval
= security_task_getscheduler(p
);
4333 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4340 * sys_sched_getparam - get the RT priority of a thread
4341 * @pid: the pid in question.
4342 * @param: structure containing the RT priority.
4344 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4346 struct sched_param lp
;
4347 struct task_struct
*p
;
4350 if (!param
|| pid
< 0)
4354 p
= find_process_by_pid(pid
);
4359 retval
= security_task_getscheduler(p
);
4363 lp
.sched_priority
= p
->rt_priority
;
4367 * This one might sleep, we cannot do it with a spinlock held ...
4369 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4378 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4380 cpumask_var_t cpus_allowed
, new_mask
;
4381 struct task_struct
*p
;
4387 p
= find_process_by_pid(pid
);
4394 /* Prevent p going away */
4398 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4402 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4404 goto out_free_cpus_allowed
;
4407 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4410 retval
= security_task_setscheduler(p
);
4414 cpuset_cpus_allowed(p
, cpus_allowed
);
4415 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4417 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4420 cpuset_cpus_allowed(p
, cpus_allowed
);
4421 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4423 * We must have raced with a concurrent cpuset
4424 * update. Just reset the cpus_allowed to the
4425 * cpuset's cpus_allowed
4427 cpumask_copy(new_mask
, cpus_allowed
);
4432 free_cpumask_var(new_mask
);
4433 out_free_cpus_allowed
:
4434 free_cpumask_var(cpus_allowed
);
4441 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4442 struct cpumask
*new_mask
)
4444 if (len
< cpumask_size())
4445 cpumask_clear(new_mask
);
4446 else if (len
> cpumask_size())
4447 len
= cpumask_size();
4449 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4453 * sys_sched_setaffinity - set the cpu affinity of a process
4454 * @pid: pid of the process
4455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4456 * @user_mask_ptr: user-space pointer to the new cpu mask
4458 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4459 unsigned long __user
*, user_mask_ptr
)
4461 cpumask_var_t new_mask
;
4464 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4467 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4469 retval
= sched_setaffinity(pid
, new_mask
);
4470 free_cpumask_var(new_mask
);
4474 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4476 struct task_struct
*p
;
4477 unsigned long flags
;
4484 p
= find_process_by_pid(pid
);
4488 retval
= security_task_getscheduler(p
);
4492 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4493 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4494 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4504 * sys_sched_getaffinity - get the cpu affinity of a process
4505 * @pid: pid of the process
4506 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4507 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4509 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4510 unsigned long __user
*, user_mask_ptr
)
4515 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4517 if (len
& (sizeof(unsigned long)-1))
4520 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4523 ret
= sched_getaffinity(pid
, mask
);
4525 size_t retlen
= min_t(size_t, len
, cpumask_size());
4527 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4532 free_cpumask_var(mask
);
4538 * sys_sched_yield - yield the current processor to other threads.
4540 * This function yields the current CPU to other tasks. If there are no
4541 * other threads running on this CPU then this function will return.
4543 SYSCALL_DEFINE0(sched_yield
)
4545 struct rq
*rq
= this_rq_lock();
4547 schedstat_inc(rq
, yld_count
);
4548 current
->sched_class
->yield_task(rq
);
4551 * Since we are going to call schedule() anyway, there's
4552 * no need to preempt or enable interrupts:
4554 __release(rq
->lock
);
4555 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4556 do_raw_spin_unlock(&rq
->lock
);
4557 sched_preempt_enable_no_resched();
4564 static inline int should_resched(void)
4566 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4569 static void __cond_resched(void)
4571 add_preempt_count(PREEMPT_ACTIVE
);
4573 sub_preempt_count(PREEMPT_ACTIVE
);
4576 int __sched
_cond_resched(void)
4578 if (should_resched()) {
4584 EXPORT_SYMBOL(_cond_resched
);
4587 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4588 * call schedule, and on return reacquire the lock.
4590 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4591 * operations here to prevent schedule() from being called twice (once via
4592 * spin_unlock(), once by hand).
4594 int __cond_resched_lock(spinlock_t
*lock
)
4596 int resched
= should_resched();
4599 lockdep_assert_held(lock
);
4601 if (spin_needbreak(lock
) || resched
) {
4612 EXPORT_SYMBOL(__cond_resched_lock
);
4614 int __sched
__cond_resched_softirq(void)
4616 BUG_ON(!in_softirq());
4618 if (should_resched()) {
4626 EXPORT_SYMBOL(__cond_resched_softirq
);
4629 * yield - yield the current processor to other threads.
4631 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4633 * The scheduler is at all times free to pick the calling task as the most
4634 * eligible task to run, if removing the yield() call from your code breaks
4635 * it, its already broken.
4637 * Typical broken usage is:
4642 * where one assumes that yield() will let 'the other' process run that will
4643 * make event true. If the current task is a SCHED_FIFO task that will never
4644 * happen. Never use yield() as a progress guarantee!!
4646 * If you want to use yield() to wait for something, use wait_event().
4647 * If you want to use yield() to be 'nice' for others, use cond_resched().
4648 * If you still want to use yield(), do not!
4650 void __sched
yield(void)
4652 set_current_state(TASK_RUNNING
);
4655 EXPORT_SYMBOL(yield
);
4658 * yield_to - yield the current processor to another thread in
4659 * your thread group, or accelerate that thread toward the
4660 * processor it's on.
4662 * @preempt: whether task preemption is allowed or not
4664 * It's the caller's job to ensure that the target task struct
4665 * can't go away on us before we can do any checks.
4667 * Returns true if we indeed boosted the target task.
4669 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4671 struct task_struct
*curr
= current
;
4672 struct rq
*rq
, *p_rq
;
4673 unsigned long flags
;
4676 local_irq_save(flags
);
4681 double_rq_lock(rq
, p_rq
);
4682 while (task_rq(p
) != p_rq
) {
4683 double_rq_unlock(rq
, p_rq
);
4687 if (!curr
->sched_class
->yield_to_task
)
4690 if (curr
->sched_class
!= p
->sched_class
)
4693 if (task_running(p_rq
, p
) || p
->state
)
4696 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4698 schedstat_inc(rq
, yld_count
);
4700 * Make p's CPU reschedule; pick_next_entity takes care of
4703 if (preempt
&& rq
!= p_rq
)
4704 resched_task(p_rq
->curr
);
4707 * We might have set it in task_yield_fair(), but are
4708 * not going to schedule(), so don't want to skip
4711 rq
->skip_clock_update
= 0;
4715 double_rq_unlock(rq
, p_rq
);
4716 local_irq_restore(flags
);
4723 EXPORT_SYMBOL_GPL(yield_to
);
4726 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4727 * that process accounting knows that this is a task in IO wait state.
4729 void __sched
io_schedule(void)
4731 struct rq
*rq
= raw_rq();
4733 delayacct_blkio_start();
4734 atomic_inc(&rq
->nr_iowait
);
4735 blk_flush_plug(current
);
4736 current
->in_iowait
= 1;
4738 current
->in_iowait
= 0;
4739 atomic_dec(&rq
->nr_iowait
);
4740 delayacct_blkio_end();
4742 EXPORT_SYMBOL(io_schedule
);
4744 long __sched
io_schedule_timeout(long timeout
)
4746 struct rq
*rq
= raw_rq();
4749 delayacct_blkio_start();
4750 atomic_inc(&rq
->nr_iowait
);
4751 blk_flush_plug(current
);
4752 current
->in_iowait
= 1;
4753 ret
= schedule_timeout(timeout
);
4754 current
->in_iowait
= 0;
4755 atomic_dec(&rq
->nr_iowait
);
4756 delayacct_blkio_end();
4761 * sys_sched_get_priority_max - return maximum RT priority.
4762 * @policy: scheduling class.
4764 * this syscall returns the maximum rt_priority that can be used
4765 * by a given scheduling class.
4767 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4774 ret
= MAX_USER_RT_PRIO
-1;
4786 * sys_sched_get_priority_min - return minimum RT priority.
4787 * @policy: scheduling class.
4789 * this syscall returns the minimum rt_priority that can be used
4790 * by a given scheduling class.
4792 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4810 * sys_sched_rr_get_interval - return the default timeslice of a process.
4811 * @pid: pid of the process.
4812 * @interval: userspace pointer to the timeslice value.
4814 * this syscall writes the default timeslice value of a given process
4815 * into the user-space timespec buffer. A value of '0' means infinity.
4817 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4818 struct timespec __user
*, interval
)
4820 struct task_struct
*p
;
4821 unsigned int time_slice
;
4822 unsigned long flags
;
4832 p
= find_process_by_pid(pid
);
4836 retval
= security_task_getscheduler(p
);
4840 rq
= task_rq_lock(p
, &flags
);
4841 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4842 task_rq_unlock(rq
, p
, &flags
);
4845 jiffies_to_timespec(time_slice
, &t
);
4846 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4854 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4856 void sched_show_task(struct task_struct
*p
)
4858 unsigned long free
= 0;
4861 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4862 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4863 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4864 #if BITS_PER_LONG == 32
4865 if (state
== TASK_RUNNING
)
4866 printk(KERN_CONT
" running ");
4868 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4870 if (state
== TASK_RUNNING
)
4871 printk(KERN_CONT
" running task ");
4873 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4875 #ifdef CONFIG_DEBUG_STACK_USAGE
4876 free
= stack_not_used(p
);
4878 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4879 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4880 (unsigned long)task_thread_info(p
)->flags
);
4882 show_stack(p
, NULL
);
4885 void show_state_filter(unsigned long state_filter
)
4887 struct task_struct
*g
, *p
;
4889 #if BITS_PER_LONG == 32
4891 " task PC stack pid father\n");
4894 " task PC stack pid father\n");
4897 do_each_thread(g
, p
) {
4899 * reset the NMI-timeout, listing all files on a slow
4900 * console might take a lot of time:
4902 touch_nmi_watchdog();
4903 if (!state_filter
|| (p
->state
& state_filter
))
4905 } while_each_thread(g
, p
);
4907 touch_all_softlockup_watchdogs();
4909 #ifdef CONFIG_SCHED_DEBUG
4910 sysrq_sched_debug_show();
4914 * Only show locks if all tasks are dumped:
4917 debug_show_all_locks();
4920 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4922 idle
->sched_class
= &idle_sched_class
;
4926 * init_idle - set up an idle thread for a given CPU
4927 * @idle: task in question
4928 * @cpu: cpu the idle task belongs to
4930 * NOTE: this function does not set the idle thread's NEED_RESCHED
4931 * flag, to make booting more robust.
4933 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4935 struct rq
*rq
= cpu_rq(cpu
);
4936 unsigned long flags
;
4938 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4941 idle
->state
= TASK_RUNNING
;
4942 idle
->se
.exec_start
= sched_clock();
4944 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4946 * We're having a chicken and egg problem, even though we are
4947 * holding rq->lock, the cpu isn't yet set to this cpu so the
4948 * lockdep check in task_group() will fail.
4950 * Similar case to sched_fork(). / Alternatively we could
4951 * use task_rq_lock() here and obtain the other rq->lock.
4956 __set_task_cpu(idle
, cpu
);
4959 rq
->curr
= rq
->idle
= idle
;
4960 #if defined(CONFIG_SMP)
4963 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4965 /* Set the preempt count _outside_ the spinlocks! */
4966 task_thread_info(idle
)->preempt_count
= 0;
4969 * The idle tasks have their own, simple scheduling class:
4971 idle
->sched_class
= &idle_sched_class
;
4972 ftrace_graph_init_idle_task(idle
, cpu
);
4973 #if defined(CONFIG_SMP)
4974 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4979 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4981 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4982 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4984 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4985 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
4989 * This is how migration works:
4991 * 1) we invoke migration_cpu_stop() on the target CPU using
4993 * 2) stopper starts to run (implicitly forcing the migrated thread
4995 * 3) it checks whether the migrated task is still in the wrong runqueue.
4996 * 4) if it's in the wrong runqueue then the migration thread removes
4997 * it and puts it into the right queue.
4998 * 5) stopper completes and stop_one_cpu() returns and the migration
5003 * Change a given task's CPU affinity. Migrate the thread to a
5004 * proper CPU and schedule it away if the CPU it's executing on
5005 * is removed from the allowed bitmask.
5007 * NOTE: the caller must have a valid reference to the task, the
5008 * task must not exit() & deallocate itself prematurely. The
5009 * call is not atomic; no spinlocks may be held.
5011 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5013 unsigned long flags
;
5015 unsigned int dest_cpu
;
5018 rq
= task_rq_lock(p
, &flags
);
5020 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
5023 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5028 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
5033 do_set_cpus_allowed(p
, new_mask
);
5035 /* Can the task run on the task's current CPU? If so, we're done */
5036 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5039 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5041 struct migration_arg arg
= { p
, dest_cpu
};
5042 /* Need help from migration thread: drop lock and wait. */
5043 task_rq_unlock(rq
, p
, &flags
);
5044 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5045 tlb_migrate_finish(p
->mm
);
5049 task_rq_unlock(rq
, p
, &flags
);
5053 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5056 * Move (not current) task off this cpu, onto dest cpu. We're doing
5057 * this because either it can't run here any more (set_cpus_allowed()
5058 * away from this CPU, or CPU going down), or because we're
5059 * attempting to rebalance this task on exec (sched_exec).
5061 * So we race with normal scheduler movements, but that's OK, as long
5062 * as the task is no longer on this CPU.
5064 * Returns non-zero if task was successfully migrated.
5066 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5068 struct rq
*rq_dest
, *rq_src
;
5071 if (unlikely(!cpu_active(dest_cpu
)))
5074 rq_src
= cpu_rq(src_cpu
);
5075 rq_dest
= cpu_rq(dest_cpu
);
5077 raw_spin_lock(&p
->pi_lock
);
5078 double_rq_lock(rq_src
, rq_dest
);
5079 /* Already moved. */
5080 if (task_cpu(p
) != src_cpu
)
5082 /* Affinity changed (again). */
5083 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
5087 * If we're not on a rq, the next wake-up will ensure we're
5091 dequeue_task(rq_src
, p
, 0);
5092 set_task_cpu(p
, dest_cpu
);
5093 enqueue_task(rq_dest
, p
, 0);
5094 check_preempt_curr(rq_dest
, p
, 0);
5099 double_rq_unlock(rq_src
, rq_dest
);
5100 raw_spin_unlock(&p
->pi_lock
);
5105 * migration_cpu_stop - this will be executed by a highprio stopper thread
5106 * and performs thread migration by bumping thread off CPU then
5107 * 'pushing' onto another runqueue.
5109 static int migration_cpu_stop(void *data
)
5111 struct migration_arg
*arg
= data
;
5114 * The original target cpu might have gone down and we might
5115 * be on another cpu but it doesn't matter.
5117 local_irq_disable();
5118 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5123 #ifdef CONFIG_HOTPLUG_CPU
5126 * Ensures that the idle task is using init_mm right before its cpu goes
5129 void idle_task_exit(void)
5131 struct mm_struct
*mm
= current
->active_mm
;
5133 BUG_ON(cpu_online(smp_processor_id()));
5136 switch_mm(mm
, &init_mm
, current
);
5141 * While a dead CPU has no uninterruptible tasks queued at this point,
5142 * it might still have a nonzero ->nr_uninterruptible counter, because
5143 * for performance reasons the counter is not stricly tracking tasks to
5144 * their home CPUs. So we just add the counter to another CPU's counter,
5145 * to keep the global sum constant after CPU-down:
5147 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5149 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5151 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5152 rq_src
->nr_uninterruptible
= 0;
5156 * remove the tasks which were accounted by rq from calc_load_tasks.
5158 static void calc_global_load_remove(struct rq
*rq
)
5160 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5161 rq
->calc_load_active
= 0;
5165 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5166 * try_to_wake_up()->select_task_rq().
5168 * Called with rq->lock held even though we'er in stop_machine() and
5169 * there's no concurrency possible, we hold the required locks anyway
5170 * because of lock validation efforts.
5172 static void migrate_tasks(unsigned int dead_cpu
)
5174 struct rq
*rq
= cpu_rq(dead_cpu
);
5175 struct task_struct
*next
, *stop
= rq
->stop
;
5179 * Fudge the rq selection such that the below task selection loop
5180 * doesn't get stuck on the currently eligible stop task.
5182 * We're currently inside stop_machine() and the rq is either stuck
5183 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5184 * either way we should never end up calling schedule() until we're
5189 /* Ensure any throttled groups are reachable by pick_next_task */
5190 unthrottle_offline_cfs_rqs(rq
);
5194 * There's this thread running, bail when that's the only
5197 if (rq
->nr_running
== 1)
5200 next
= pick_next_task(rq
);
5202 next
->sched_class
->put_prev_task(rq
, next
);
5204 /* Find suitable destination for @next, with force if needed. */
5205 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5206 raw_spin_unlock(&rq
->lock
);
5208 __migrate_task(next
, dead_cpu
, dest_cpu
);
5210 raw_spin_lock(&rq
->lock
);
5216 #endif /* CONFIG_HOTPLUG_CPU */
5218 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5220 static struct ctl_table sd_ctl_dir
[] = {
5222 .procname
= "sched_domain",
5228 static struct ctl_table sd_ctl_root
[] = {
5230 .procname
= "kernel",
5232 .child
= sd_ctl_dir
,
5237 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5239 struct ctl_table
*entry
=
5240 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5245 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5247 struct ctl_table
*entry
;
5250 * In the intermediate directories, both the child directory and
5251 * procname are dynamically allocated and could fail but the mode
5252 * will always be set. In the lowest directory the names are
5253 * static strings and all have proc handlers.
5255 for (entry
= *tablep
; entry
->mode
; entry
++) {
5257 sd_free_ctl_entry(&entry
->child
);
5258 if (entry
->proc_handler
== NULL
)
5259 kfree(entry
->procname
);
5267 set_table_entry(struct ctl_table
*entry
,
5268 const char *procname
, void *data
, int maxlen
,
5269 umode_t mode
, proc_handler
*proc_handler
)
5271 entry
->procname
= procname
;
5273 entry
->maxlen
= maxlen
;
5275 entry
->proc_handler
= proc_handler
;
5278 static struct ctl_table
*
5279 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5281 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5286 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5287 sizeof(long), 0644, proc_doulongvec_minmax
);
5288 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5289 sizeof(long), 0644, proc_doulongvec_minmax
);
5290 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5291 sizeof(int), 0644, proc_dointvec_minmax
);
5292 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5293 sizeof(int), 0644, proc_dointvec_minmax
);
5294 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5295 sizeof(int), 0644, proc_dointvec_minmax
);
5296 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5297 sizeof(int), 0644, proc_dointvec_minmax
);
5298 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5299 sizeof(int), 0644, proc_dointvec_minmax
);
5300 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5301 sizeof(int), 0644, proc_dointvec_minmax
);
5302 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5303 sizeof(int), 0644, proc_dointvec_minmax
);
5304 set_table_entry(&table
[9], "cache_nice_tries",
5305 &sd
->cache_nice_tries
,
5306 sizeof(int), 0644, proc_dointvec_minmax
);
5307 set_table_entry(&table
[10], "flags", &sd
->flags
,
5308 sizeof(int), 0644, proc_dointvec_minmax
);
5309 set_table_entry(&table
[11], "name", sd
->name
,
5310 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5311 /* &table[12] is terminator */
5316 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5318 struct ctl_table
*entry
, *table
;
5319 struct sched_domain
*sd
;
5320 int domain_num
= 0, i
;
5323 for_each_domain(cpu
, sd
)
5325 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5330 for_each_domain(cpu
, sd
) {
5331 snprintf(buf
, 32, "domain%d", i
);
5332 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5334 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5341 static struct ctl_table_header
*sd_sysctl_header
;
5342 static void register_sched_domain_sysctl(void)
5344 int i
, cpu_num
= num_possible_cpus();
5345 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5348 WARN_ON(sd_ctl_dir
[0].child
);
5349 sd_ctl_dir
[0].child
= entry
;
5354 for_each_possible_cpu(i
) {
5355 snprintf(buf
, 32, "cpu%d", i
);
5356 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5358 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5362 WARN_ON(sd_sysctl_header
);
5363 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5366 /* may be called multiple times per register */
5367 static void unregister_sched_domain_sysctl(void)
5369 if (sd_sysctl_header
)
5370 unregister_sysctl_table(sd_sysctl_header
);
5371 sd_sysctl_header
= NULL
;
5372 if (sd_ctl_dir
[0].child
)
5373 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5376 static void register_sched_domain_sysctl(void)
5379 static void unregister_sched_domain_sysctl(void)
5384 static void set_rq_online(struct rq
*rq
)
5387 const struct sched_class
*class;
5389 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5392 for_each_class(class) {
5393 if (class->rq_online
)
5394 class->rq_online(rq
);
5399 static void set_rq_offline(struct rq
*rq
)
5402 const struct sched_class
*class;
5404 for_each_class(class) {
5405 if (class->rq_offline
)
5406 class->rq_offline(rq
);
5409 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5415 * migration_call - callback that gets triggered when a CPU is added.
5416 * Here we can start up the necessary migration thread for the new CPU.
5418 static int __cpuinit
5419 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5421 int cpu
= (long)hcpu
;
5422 unsigned long flags
;
5423 struct rq
*rq
= cpu_rq(cpu
);
5425 switch (action
& ~CPU_TASKS_FROZEN
) {
5427 case CPU_UP_PREPARE
:
5428 rq
->calc_load_update
= calc_load_update
;
5432 /* Update our root-domain */
5433 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5435 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5439 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5442 #ifdef CONFIG_HOTPLUG_CPU
5444 sched_ttwu_pending();
5445 /* Update our root-domain */
5446 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5448 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5452 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5453 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5455 migrate_nr_uninterruptible(rq
);
5456 calc_global_load_remove(rq
);
5461 update_max_interval();
5467 * Register at high priority so that task migration (migrate_all_tasks)
5468 * happens before everything else. This has to be lower priority than
5469 * the notifier in the perf_event subsystem, though.
5471 static struct notifier_block __cpuinitdata migration_notifier
= {
5472 .notifier_call
= migration_call
,
5473 .priority
= CPU_PRI_MIGRATION
,
5476 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5477 unsigned long action
, void *hcpu
)
5479 switch (action
& ~CPU_TASKS_FROZEN
) {
5481 case CPU_DOWN_FAILED
:
5482 set_cpu_active((long)hcpu
, true);
5489 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5490 unsigned long action
, void *hcpu
)
5492 switch (action
& ~CPU_TASKS_FROZEN
) {
5493 case CPU_DOWN_PREPARE
:
5494 set_cpu_active((long)hcpu
, false);
5501 static int __init
migration_init(void)
5503 void *cpu
= (void *)(long)smp_processor_id();
5506 /* Initialize migration for the boot CPU */
5507 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5508 BUG_ON(err
== NOTIFY_BAD
);
5509 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5510 register_cpu_notifier(&migration_notifier
);
5512 /* Register cpu active notifiers */
5513 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5514 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5518 early_initcall(migration_init
);
5523 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5525 #ifdef CONFIG_SCHED_DEBUG
5527 static __read_mostly
int sched_domain_debug_enabled
;
5529 static int __init
sched_domain_debug_setup(char *str
)
5531 sched_domain_debug_enabled
= 1;
5535 early_param("sched_debug", sched_domain_debug_setup
);
5537 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5538 struct cpumask
*groupmask
)
5540 struct sched_group
*group
= sd
->groups
;
5543 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5544 cpumask_clear(groupmask
);
5546 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5548 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5549 printk("does not load-balance\n");
5551 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5556 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5558 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5559 printk(KERN_ERR
"ERROR: domain->span does not contain "
5562 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5563 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5567 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5571 printk(KERN_ERR
"ERROR: group is NULL\n");
5575 if (!group
->sgp
->power
) {
5576 printk(KERN_CONT
"\n");
5577 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5582 if (!cpumask_weight(sched_group_cpus(group
))) {
5583 printk(KERN_CONT
"\n");
5584 printk(KERN_ERR
"ERROR: empty group\n");
5588 if (!(sd
->flags
& SD_OVERLAP
) &&
5589 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5590 printk(KERN_CONT
"\n");
5591 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5595 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5597 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5599 printk(KERN_CONT
" %s", str
);
5600 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5601 printk(KERN_CONT
" (cpu_power = %d)",
5605 group
= group
->next
;
5606 } while (group
!= sd
->groups
);
5607 printk(KERN_CONT
"\n");
5609 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5610 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5613 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5614 printk(KERN_ERR
"ERROR: parent span is not a superset "
5615 "of domain->span\n");
5619 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5623 if (!sched_domain_debug_enabled
)
5627 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5631 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5634 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5642 #else /* !CONFIG_SCHED_DEBUG */
5643 # define sched_domain_debug(sd, cpu) do { } while (0)
5644 #endif /* CONFIG_SCHED_DEBUG */
5646 static int sd_degenerate(struct sched_domain
*sd
)
5648 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5651 /* Following flags need at least 2 groups */
5652 if (sd
->flags
& (SD_LOAD_BALANCE
|
5653 SD_BALANCE_NEWIDLE
|
5657 SD_SHARE_PKG_RESOURCES
)) {
5658 if (sd
->groups
!= sd
->groups
->next
)
5662 /* Following flags don't use groups */
5663 if (sd
->flags
& (SD_WAKE_AFFINE
))
5670 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5672 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5674 if (sd_degenerate(parent
))
5677 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5680 /* Flags needing groups don't count if only 1 group in parent */
5681 if (parent
->groups
== parent
->groups
->next
) {
5682 pflags
&= ~(SD_LOAD_BALANCE
|
5683 SD_BALANCE_NEWIDLE
|
5687 SD_SHARE_PKG_RESOURCES
);
5688 if (nr_node_ids
== 1)
5689 pflags
&= ~SD_SERIALIZE
;
5691 if (~cflags
& pflags
)
5697 static void free_rootdomain(struct rcu_head
*rcu
)
5699 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5701 cpupri_cleanup(&rd
->cpupri
);
5702 free_cpumask_var(rd
->rto_mask
);
5703 free_cpumask_var(rd
->online
);
5704 free_cpumask_var(rd
->span
);
5708 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5710 struct root_domain
*old_rd
= NULL
;
5711 unsigned long flags
;
5713 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5718 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5721 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5724 * If we dont want to free the old_rt yet then
5725 * set old_rd to NULL to skip the freeing later
5728 if (!atomic_dec_and_test(&old_rd
->refcount
))
5732 atomic_inc(&rd
->refcount
);
5735 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5736 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5739 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5742 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5745 static int init_rootdomain(struct root_domain
*rd
)
5747 memset(rd
, 0, sizeof(*rd
));
5749 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5751 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5753 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5756 if (cpupri_init(&rd
->cpupri
) != 0)
5761 free_cpumask_var(rd
->rto_mask
);
5763 free_cpumask_var(rd
->online
);
5765 free_cpumask_var(rd
->span
);
5771 * By default the system creates a single root-domain with all cpus as
5772 * members (mimicking the global state we have today).
5774 struct root_domain def_root_domain
;
5776 static void init_defrootdomain(void)
5778 init_rootdomain(&def_root_domain
);
5780 atomic_set(&def_root_domain
.refcount
, 1);
5783 static struct root_domain
*alloc_rootdomain(void)
5785 struct root_domain
*rd
;
5787 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5791 if (init_rootdomain(rd
) != 0) {
5799 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5801 struct sched_group
*tmp
, *first
;
5810 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5815 } while (sg
!= first
);
5818 static void free_sched_domain(struct rcu_head
*rcu
)
5820 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5823 * If its an overlapping domain it has private groups, iterate and
5826 if (sd
->flags
& SD_OVERLAP
) {
5827 free_sched_groups(sd
->groups
, 1);
5828 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5829 kfree(sd
->groups
->sgp
);
5835 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5837 call_rcu(&sd
->rcu
, free_sched_domain
);
5840 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5842 for (; sd
; sd
= sd
->parent
)
5843 destroy_sched_domain(sd
, cpu
);
5847 * Keep a special pointer to the highest sched_domain that has
5848 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5849 * allows us to avoid some pointer chasing select_idle_sibling().
5851 * Also keep a unique ID per domain (we use the first cpu number in
5852 * the cpumask of the domain), this allows us to quickly tell if
5853 * two cpus are in the same cache domain, see cpus_share_cache().
5855 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5856 DEFINE_PER_CPU(int, sd_llc_id
);
5858 static void update_top_cache_domain(int cpu
)
5860 struct sched_domain
*sd
;
5863 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5865 id
= cpumask_first(sched_domain_span(sd
));
5867 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5868 per_cpu(sd_llc_id
, cpu
) = id
;
5872 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5873 * hold the hotplug lock.
5876 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5878 struct rq
*rq
= cpu_rq(cpu
);
5879 struct sched_domain
*tmp
;
5881 /* Remove the sched domains which do not contribute to scheduling. */
5882 for (tmp
= sd
; tmp
; ) {
5883 struct sched_domain
*parent
= tmp
->parent
;
5887 if (sd_parent_degenerate(tmp
, parent
)) {
5888 tmp
->parent
= parent
->parent
;
5890 parent
->parent
->child
= tmp
;
5891 destroy_sched_domain(parent
, cpu
);
5896 if (sd
&& sd_degenerate(sd
)) {
5899 destroy_sched_domain(tmp
, cpu
);
5904 sched_domain_debug(sd
, cpu
);
5906 rq_attach_root(rq
, rd
);
5908 rcu_assign_pointer(rq
->sd
, sd
);
5909 destroy_sched_domains(tmp
, cpu
);
5911 update_top_cache_domain(cpu
);
5914 /* cpus with isolated domains */
5915 static cpumask_var_t cpu_isolated_map
;
5917 /* Setup the mask of cpus configured for isolated domains */
5918 static int __init
isolated_cpu_setup(char *str
)
5920 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5921 cpulist_parse(str
, cpu_isolated_map
);
5925 __setup("isolcpus=", isolated_cpu_setup
);
5927 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5929 return cpumask_of_node(cpu_to_node(cpu
));
5933 struct sched_domain
**__percpu sd
;
5934 struct sched_group
**__percpu sg
;
5935 struct sched_group_power
**__percpu sgp
;
5939 struct sched_domain
** __percpu sd
;
5940 struct root_domain
*rd
;
5950 struct sched_domain_topology_level
;
5952 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5953 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5955 #define SDTL_OVERLAP 0x01
5957 struct sched_domain_topology_level
{
5958 sched_domain_init_f init
;
5959 sched_domain_mask_f mask
;
5962 struct sd_data data
;
5966 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5968 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5969 const struct cpumask
*span
= sched_domain_span(sd
);
5970 struct cpumask
*covered
= sched_domains_tmpmask
;
5971 struct sd_data
*sdd
= sd
->private;
5972 struct sched_domain
*child
;
5975 cpumask_clear(covered
);
5977 for_each_cpu(i
, span
) {
5978 struct cpumask
*sg_span
;
5980 if (cpumask_test_cpu(i
, covered
))
5983 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5984 GFP_KERNEL
, cpu_to_node(cpu
));
5989 sg_span
= sched_group_cpus(sg
);
5991 child
= *per_cpu_ptr(sdd
->sd
, i
);
5993 child
= child
->child
;
5994 cpumask_copy(sg_span
, sched_domain_span(child
));
5996 cpumask_set_cpu(i
, sg_span
);
5998 cpumask_or(covered
, covered
, sg_span
);
6000 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
6001 atomic_inc(&sg
->sgp
->ref
);
6003 if (cpumask_test_cpu(cpu
, sg_span
))
6013 sd
->groups
= groups
;
6018 free_sched_groups(first
, 0);
6023 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6025 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6026 struct sched_domain
*child
= sd
->child
;
6029 cpu
= cpumask_first(sched_domain_span(child
));
6032 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6033 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
6034 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
6041 * build_sched_groups will build a circular linked list of the groups
6042 * covered by the given span, and will set each group's ->cpumask correctly,
6043 * and ->cpu_power to 0.
6045 * Assumes the sched_domain tree is fully constructed
6048 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6050 struct sched_group
*first
= NULL
, *last
= NULL
;
6051 struct sd_data
*sdd
= sd
->private;
6052 const struct cpumask
*span
= sched_domain_span(sd
);
6053 struct cpumask
*covered
;
6056 get_group(cpu
, sdd
, &sd
->groups
);
6057 atomic_inc(&sd
->groups
->ref
);
6059 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
6062 lockdep_assert_held(&sched_domains_mutex
);
6063 covered
= sched_domains_tmpmask
;
6065 cpumask_clear(covered
);
6067 for_each_cpu(i
, span
) {
6068 struct sched_group
*sg
;
6069 int group
= get_group(i
, sdd
, &sg
);
6072 if (cpumask_test_cpu(i
, covered
))
6075 cpumask_clear(sched_group_cpus(sg
));
6078 for_each_cpu(j
, span
) {
6079 if (get_group(j
, sdd
, NULL
) != group
)
6082 cpumask_set_cpu(j
, covered
);
6083 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6098 * Initialize sched groups cpu_power.
6100 * cpu_power indicates the capacity of sched group, which is used while
6101 * distributing the load between different sched groups in a sched domain.
6102 * Typically cpu_power for all the groups in a sched domain will be same unless
6103 * there are asymmetries in the topology. If there are asymmetries, group
6104 * having more cpu_power will pickup more load compared to the group having
6107 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6109 struct sched_group
*sg
= sd
->groups
;
6111 WARN_ON(!sd
|| !sg
);
6114 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6116 } while (sg
!= sd
->groups
);
6118 if (cpu
!= group_first_cpu(sg
))
6121 update_group_power(sd
, cpu
);
6122 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6125 int __weak
arch_sd_sibling_asym_packing(void)
6127 return 0*SD_ASYM_PACKING
;
6131 * Initializers for schedule domains
6132 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6135 #ifdef CONFIG_SCHED_DEBUG
6136 # define SD_INIT_NAME(sd, type) sd->name = #type
6138 # define SD_INIT_NAME(sd, type) do { } while (0)
6141 #define SD_INIT_FUNC(type) \
6142 static noinline struct sched_domain * \
6143 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6145 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6146 *sd = SD_##type##_INIT; \
6147 SD_INIT_NAME(sd, type); \
6148 sd->private = &tl->data; \
6153 #ifdef CONFIG_SCHED_SMT
6154 SD_INIT_FUNC(SIBLING
)
6156 #ifdef CONFIG_SCHED_MC
6159 #ifdef CONFIG_SCHED_BOOK
6163 static int default_relax_domain_level
= -1;
6164 int sched_domain_level_max
;
6166 static int __init
setup_relax_domain_level(char *str
)
6170 val
= simple_strtoul(str
, NULL
, 0);
6171 if (val
< sched_domain_level_max
)
6172 default_relax_domain_level
= val
;
6176 __setup("relax_domain_level=", setup_relax_domain_level
);
6178 static void set_domain_attribute(struct sched_domain
*sd
,
6179 struct sched_domain_attr
*attr
)
6183 if (!attr
|| attr
->relax_domain_level
< 0) {
6184 if (default_relax_domain_level
< 0)
6187 request
= default_relax_domain_level
;
6189 request
= attr
->relax_domain_level
;
6190 if (request
< sd
->level
) {
6191 /* turn off idle balance on this domain */
6192 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6194 /* turn on idle balance on this domain */
6195 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6199 static void __sdt_free(const struct cpumask
*cpu_map
);
6200 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6202 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6203 const struct cpumask
*cpu_map
)
6207 if (!atomic_read(&d
->rd
->refcount
))
6208 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6210 free_percpu(d
->sd
); /* fall through */
6212 __sdt_free(cpu_map
); /* fall through */
6218 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6219 const struct cpumask
*cpu_map
)
6221 memset(d
, 0, sizeof(*d
));
6223 if (__sdt_alloc(cpu_map
))
6224 return sa_sd_storage
;
6225 d
->sd
= alloc_percpu(struct sched_domain
*);
6227 return sa_sd_storage
;
6228 d
->rd
= alloc_rootdomain();
6231 return sa_rootdomain
;
6235 * NULL the sd_data elements we've used to build the sched_domain and
6236 * sched_group structure so that the subsequent __free_domain_allocs()
6237 * will not free the data we're using.
6239 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6241 struct sd_data
*sdd
= sd
->private;
6243 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6244 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6246 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6247 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6249 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6250 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6253 #ifdef CONFIG_SCHED_SMT
6254 static const struct cpumask
*cpu_smt_mask(int cpu
)
6256 return topology_thread_cpumask(cpu
);
6261 * Topology list, bottom-up.
6263 static struct sched_domain_topology_level default_topology
[] = {
6264 #ifdef CONFIG_SCHED_SMT
6265 { sd_init_SIBLING
, cpu_smt_mask
, },
6267 #ifdef CONFIG_SCHED_MC
6268 { sd_init_MC
, cpu_coregroup_mask
, },
6270 #ifdef CONFIG_SCHED_BOOK
6271 { sd_init_BOOK
, cpu_book_mask
, },
6273 { sd_init_CPU
, cpu_cpu_mask
, },
6277 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6281 static int sched_domains_numa_levels
;
6282 static int sched_domains_numa_scale
;
6283 static int *sched_domains_numa_distance
;
6284 static struct cpumask
***sched_domains_numa_masks
;
6285 static int sched_domains_curr_level
;
6287 static inline int sd_local_flags(int level
)
6289 if (sched_domains_numa_distance
[level
] > REMOTE_DISTANCE
)
6292 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6295 static struct sched_domain
*
6296 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6298 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6299 int level
= tl
->numa_level
;
6300 int sd_weight
= cpumask_weight(
6301 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6303 *sd
= (struct sched_domain
){
6304 .min_interval
= sd_weight
,
6305 .max_interval
= 2*sd_weight
,
6307 .imbalance_pct
= 125,
6308 .cache_nice_tries
= 2,
6315 .flags
= 1*SD_LOAD_BALANCE
6316 | 1*SD_BALANCE_NEWIDLE
6322 | 0*SD_SHARE_CPUPOWER
6323 | 0*SD_SHARE_PKG_RESOURCES
6325 | 0*SD_PREFER_SIBLING
6326 | sd_local_flags(level
)
6328 .last_balance
= jiffies
,
6329 .balance_interval
= sd_weight
,
6331 SD_INIT_NAME(sd
, NUMA
);
6332 sd
->private = &tl
->data
;
6335 * Ugly hack to pass state to sd_numa_mask()...
6337 sched_domains_curr_level
= tl
->numa_level
;
6342 static const struct cpumask
*sd_numa_mask(int cpu
)
6344 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6347 static void sched_init_numa(void)
6349 int next_distance
, curr_distance
= node_distance(0, 0);
6350 struct sched_domain_topology_level
*tl
;
6354 sched_domains_numa_scale
= curr_distance
;
6355 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6356 if (!sched_domains_numa_distance
)
6360 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6361 * unique distances in the node_distance() table.
6363 * Assumes node_distance(0,j) includes all distances in
6364 * node_distance(i,j) in order to avoid cubic time.
6366 * XXX: could be optimized to O(n log n) by using sort()
6368 next_distance
= curr_distance
;
6369 for (i
= 0; i
< nr_node_ids
; i
++) {
6370 for (j
= 0; j
< nr_node_ids
; j
++) {
6371 int distance
= node_distance(0, j
);
6372 if (distance
> curr_distance
&&
6373 (distance
< next_distance
||
6374 next_distance
== curr_distance
))
6375 next_distance
= distance
;
6377 if (next_distance
!= curr_distance
) {
6378 sched_domains_numa_distance
[level
++] = next_distance
;
6379 sched_domains_numa_levels
= level
;
6380 curr_distance
= next_distance
;
6384 * 'level' contains the number of unique distances, excluding the
6385 * identity distance node_distance(i,i).
6387 * The sched_domains_nume_distance[] array includes the actual distance
6391 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6392 if (!sched_domains_numa_masks
)
6396 * Now for each level, construct a mask per node which contains all
6397 * cpus of nodes that are that many hops away from us.
6399 for (i
= 0; i
< level
; i
++) {
6400 sched_domains_numa_masks
[i
] =
6401 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6402 if (!sched_domains_numa_masks
[i
])
6405 for (j
= 0; j
< nr_node_ids
; j
++) {
6406 struct cpumask
*mask
= kzalloc_node(cpumask_size(), GFP_KERNEL
, j
);
6410 sched_domains_numa_masks
[i
][j
] = mask
;
6412 for (k
= 0; k
< nr_node_ids
; k
++) {
6413 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6416 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6421 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6422 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6427 * Copy the default topology bits..
6429 for (i
= 0; default_topology
[i
].init
; i
++)
6430 tl
[i
] = default_topology
[i
];
6433 * .. and append 'j' levels of NUMA goodness.
6435 for (j
= 0; j
< level
; i
++, j
++) {
6436 tl
[i
] = (struct sched_domain_topology_level
){
6437 .init
= sd_numa_init
,
6438 .mask
= sd_numa_mask
,
6439 .flags
= SDTL_OVERLAP
,
6444 sched_domain_topology
= tl
;
6447 static inline void sched_init_numa(void)
6450 #endif /* CONFIG_NUMA */
6452 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6454 struct sched_domain_topology_level
*tl
;
6457 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6458 struct sd_data
*sdd
= &tl
->data
;
6460 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6464 sdd
->sg
= alloc_percpu(struct sched_group
*);
6468 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6472 for_each_cpu(j
, cpu_map
) {
6473 struct sched_domain
*sd
;
6474 struct sched_group
*sg
;
6475 struct sched_group_power
*sgp
;
6477 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6478 GFP_KERNEL
, cpu_to_node(j
));
6482 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6484 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6485 GFP_KERNEL
, cpu_to_node(j
));
6491 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6493 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
6494 GFP_KERNEL
, cpu_to_node(j
));
6498 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6505 static void __sdt_free(const struct cpumask
*cpu_map
)
6507 struct sched_domain_topology_level
*tl
;
6510 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6511 struct sd_data
*sdd
= &tl
->data
;
6513 for_each_cpu(j
, cpu_map
) {
6514 struct sched_domain
*sd
;
6517 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6518 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6519 free_sched_groups(sd
->groups
, 0);
6520 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6524 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6526 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6528 free_percpu(sdd
->sd
);
6530 free_percpu(sdd
->sg
);
6532 free_percpu(sdd
->sgp
);
6537 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6538 struct s_data
*d
, const struct cpumask
*cpu_map
,
6539 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6542 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6546 set_domain_attribute(sd
, attr
);
6547 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6549 sd
->level
= child
->level
+ 1;
6550 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6559 * Build sched domains for a given set of cpus and attach the sched domains
6560 * to the individual cpus
6562 static int build_sched_domains(const struct cpumask
*cpu_map
,
6563 struct sched_domain_attr
*attr
)
6565 enum s_alloc alloc_state
= sa_none
;
6566 struct sched_domain
*sd
;
6568 int i
, ret
= -ENOMEM
;
6570 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6571 if (alloc_state
!= sa_rootdomain
)
6574 /* Set up domains for cpus specified by the cpu_map. */
6575 for_each_cpu(i
, cpu_map
) {
6576 struct sched_domain_topology_level
*tl
;
6579 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6580 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6581 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6582 sd
->flags
|= SD_OVERLAP
;
6583 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6590 *per_cpu_ptr(d
.sd
, i
) = sd
;
6593 /* Build the groups for the domains */
6594 for_each_cpu(i
, cpu_map
) {
6595 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6596 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6597 if (sd
->flags
& SD_OVERLAP
) {
6598 if (build_overlap_sched_groups(sd
, i
))
6601 if (build_sched_groups(sd
, i
))
6607 /* Calculate CPU power for physical packages and nodes */
6608 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6609 if (!cpumask_test_cpu(i
, cpu_map
))
6612 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6613 claim_allocations(i
, sd
);
6614 init_sched_groups_power(i
, sd
);
6618 /* Attach the domains */
6620 for_each_cpu(i
, cpu_map
) {
6621 sd
= *per_cpu_ptr(d
.sd
, i
);
6622 cpu_attach_domain(sd
, d
.rd
, i
);
6628 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6632 static cpumask_var_t
*doms_cur
; /* current sched domains */
6633 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6634 static struct sched_domain_attr
*dattr_cur
;
6635 /* attribues of custom domains in 'doms_cur' */
6638 * Special case: If a kmalloc of a doms_cur partition (array of
6639 * cpumask) fails, then fallback to a single sched domain,
6640 * as determined by the single cpumask fallback_doms.
6642 static cpumask_var_t fallback_doms
;
6645 * arch_update_cpu_topology lets virtualized architectures update the
6646 * cpu core maps. It is supposed to return 1 if the topology changed
6647 * or 0 if it stayed the same.
6649 int __attribute__((weak
)) arch_update_cpu_topology(void)
6654 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6657 cpumask_var_t
*doms
;
6659 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6662 for (i
= 0; i
< ndoms
; i
++) {
6663 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6664 free_sched_domains(doms
, i
);
6671 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6674 for (i
= 0; i
< ndoms
; i
++)
6675 free_cpumask_var(doms
[i
]);
6680 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6681 * For now this just excludes isolated cpus, but could be used to
6682 * exclude other special cases in the future.
6684 static int init_sched_domains(const struct cpumask
*cpu_map
)
6688 arch_update_cpu_topology();
6690 doms_cur
= alloc_sched_domains(ndoms_cur
);
6692 doms_cur
= &fallback_doms
;
6693 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6695 err
= build_sched_domains(doms_cur
[0], NULL
);
6696 register_sched_domain_sysctl();
6702 * Detach sched domains from a group of cpus specified in cpu_map
6703 * These cpus will now be attached to the NULL domain
6705 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6710 for_each_cpu(i
, cpu_map
)
6711 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6715 /* handle null as "default" */
6716 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6717 struct sched_domain_attr
*new, int idx_new
)
6719 struct sched_domain_attr tmp
;
6726 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6727 new ? (new + idx_new
) : &tmp
,
6728 sizeof(struct sched_domain_attr
));
6732 * Partition sched domains as specified by the 'ndoms_new'
6733 * cpumasks in the array doms_new[] of cpumasks. This compares
6734 * doms_new[] to the current sched domain partitioning, doms_cur[].
6735 * It destroys each deleted domain and builds each new domain.
6737 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6738 * The masks don't intersect (don't overlap.) We should setup one
6739 * sched domain for each mask. CPUs not in any of the cpumasks will
6740 * not be load balanced. If the same cpumask appears both in the
6741 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6744 * The passed in 'doms_new' should be allocated using
6745 * alloc_sched_domains. This routine takes ownership of it and will
6746 * free_sched_domains it when done with it. If the caller failed the
6747 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6748 * and partition_sched_domains() will fallback to the single partition
6749 * 'fallback_doms', it also forces the domains to be rebuilt.
6751 * If doms_new == NULL it will be replaced with cpu_online_mask.
6752 * ndoms_new == 0 is a special case for destroying existing domains,
6753 * and it will not create the default domain.
6755 * Call with hotplug lock held
6757 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6758 struct sched_domain_attr
*dattr_new
)
6763 mutex_lock(&sched_domains_mutex
);
6765 /* always unregister in case we don't destroy any domains */
6766 unregister_sched_domain_sysctl();
6768 /* Let architecture update cpu core mappings. */
6769 new_topology
= arch_update_cpu_topology();
6771 n
= doms_new
? ndoms_new
: 0;
6773 /* Destroy deleted domains */
6774 for (i
= 0; i
< ndoms_cur
; i
++) {
6775 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6776 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6777 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6780 /* no match - a current sched domain not in new doms_new[] */
6781 detach_destroy_domains(doms_cur
[i
]);
6786 if (doms_new
== NULL
) {
6788 doms_new
= &fallback_doms
;
6789 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6790 WARN_ON_ONCE(dattr_new
);
6793 /* Build new domains */
6794 for (i
= 0; i
< ndoms_new
; i
++) {
6795 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6796 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6797 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6800 /* no match - add a new doms_new */
6801 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6806 /* Remember the new sched domains */
6807 if (doms_cur
!= &fallback_doms
)
6808 free_sched_domains(doms_cur
, ndoms_cur
);
6809 kfree(dattr_cur
); /* kfree(NULL) is safe */
6810 doms_cur
= doms_new
;
6811 dattr_cur
= dattr_new
;
6812 ndoms_cur
= ndoms_new
;
6814 register_sched_domain_sysctl();
6816 mutex_unlock(&sched_domains_mutex
);
6820 * Update cpusets according to cpu_active mask. If cpusets are
6821 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6822 * around partition_sched_domains().
6824 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6827 switch (action
& ~CPU_TASKS_FROZEN
) {
6829 case CPU_DOWN_FAILED
:
6830 cpuset_update_active_cpus();
6837 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6840 switch (action
& ~CPU_TASKS_FROZEN
) {
6841 case CPU_DOWN_PREPARE
:
6842 cpuset_update_active_cpus();
6849 void __init
sched_init_smp(void)
6851 cpumask_var_t non_isolated_cpus
;
6853 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6854 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6859 mutex_lock(&sched_domains_mutex
);
6860 init_sched_domains(cpu_active_mask
);
6861 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6862 if (cpumask_empty(non_isolated_cpus
))
6863 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6864 mutex_unlock(&sched_domains_mutex
);
6867 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6868 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6870 /* RT runtime code needs to handle some hotplug events */
6871 hotcpu_notifier(update_runtime
, 0);
6875 /* Move init over to a non-isolated CPU */
6876 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6878 sched_init_granularity();
6879 free_cpumask_var(non_isolated_cpus
);
6881 init_sched_rt_class();
6884 void __init
sched_init_smp(void)
6886 sched_init_granularity();
6888 #endif /* CONFIG_SMP */
6890 const_debug
unsigned int sysctl_timer_migration
= 1;
6892 int in_sched_functions(unsigned long addr
)
6894 return in_lock_functions(addr
) ||
6895 (addr
>= (unsigned long)__sched_text_start
6896 && addr
< (unsigned long)__sched_text_end
);
6899 #ifdef CONFIG_CGROUP_SCHED
6900 struct task_group root_task_group
;
6903 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6905 void __init
sched_init(void)
6908 unsigned long alloc_size
= 0, ptr
;
6910 #ifdef CONFIG_FAIR_GROUP_SCHED
6911 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6913 #ifdef CONFIG_RT_GROUP_SCHED
6914 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6916 #ifdef CONFIG_CPUMASK_OFFSTACK
6917 alloc_size
+= num_possible_cpus() * cpumask_size();
6920 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6922 #ifdef CONFIG_FAIR_GROUP_SCHED
6923 root_task_group
.se
= (struct sched_entity
**)ptr
;
6924 ptr
+= nr_cpu_ids
* sizeof(void **);
6926 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6927 ptr
+= nr_cpu_ids
* sizeof(void **);
6929 #endif /* CONFIG_FAIR_GROUP_SCHED */
6930 #ifdef CONFIG_RT_GROUP_SCHED
6931 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6932 ptr
+= nr_cpu_ids
* sizeof(void **);
6934 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6935 ptr
+= nr_cpu_ids
* sizeof(void **);
6937 #endif /* CONFIG_RT_GROUP_SCHED */
6938 #ifdef CONFIG_CPUMASK_OFFSTACK
6939 for_each_possible_cpu(i
) {
6940 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6941 ptr
+= cpumask_size();
6943 #endif /* CONFIG_CPUMASK_OFFSTACK */
6947 init_defrootdomain();
6950 init_rt_bandwidth(&def_rt_bandwidth
,
6951 global_rt_period(), global_rt_runtime());
6953 #ifdef CONFIG_RT_GROUP_SCHED
6954 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6955 global_rt_period(), global_rt_runtime());
6956 #endif /* CONFIG_RT_GROUP_SCHED */
6958 #ifdef CONFIG_CGROUP_SCHED
6959 list_add(&root_task_group
.list
, &task_groups
);
6960 INIT_LIST_HEAD(&root_task_group
.children
);
6961 INIT_LIST_HEAD(&root_task_group
.siblings
);
6962 autogroup_init(&init_task
);
6964 #endif /* CONFIG_CGROUP_SCHED */
6966 #ifdef CONFIG_CGROUP_CPUACCT
6967 root_cpuacct
.cpustat
= &kernel_cpustat
;
6968 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6969 /* Too early, not expected to fail */
6970 BUG_ON(!root_cpuacct
.cpuusage
);
6972 for_each_possible_cpu(i
) {
6976 raw_spin_lock_init(&rq
->lock
);
6978 rq
->calc_load_active
= 0;
6979 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6980 init_cfs_rq(&rq
->cfs
);
6981 init_rt_rq(&rq
->rt
, rq
);
6982 #ifdef CONFIG_FAIR_GROUP_SCHED
6983 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6984 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6986 * How much cpu bandwidth does root_task_group get?
6988 * In case of task-groups formed thr' the cgroup filesystem, it
6989 * gets 100% of the cpu resources in the system. This overall
6990 * system cpu resource is divided among the tasks of
6991 * root_task_group and its child task-groups in a fair manner,
6992 * based on each entity's (task or task-group's) weight
6993 * (se->load.weight).
6995 * In other words, if root_task_group has 10 tasks of weight
6996 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6997 * then A0's share of the cpu resource is:
6999 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7001 * We achieve this by letting root_task_group's tasks sit
7002 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7004 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7005 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7006 #endif /* CONFIG_FAIR_GROUP_SCHED */
7008 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7009 #ifdef CONFIG_RT_GROUP_SCHED
7010 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7011 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7014 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7015 rq
->cpu_load
[j
] = 0;
7017 rq
->last_load_update_tick
= jiffies
;
7022 rq
->cpu_power
= SCHED_POWER_SCALE
;
7023 rq
->post_schedule
= 0;
7024 rq
->active_balance
= 0;
7025 rq
->next_balance
= jiffies
;
7030 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7032 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7034 rq_attach_root(rq
, &def_root_domain
);
7040 atomic_set(&rq
->nr_iowait
, 0);
7043 set_load_weight(&init_task
);
7045 #ifdef CONFIG_PREEMPT_NOTIFIERS
7046 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7049 #ifdef CONFIG_RT_MUTEXES
7050 plist_head_init(&init_task
.pi_waiters
);
7054 * The boot idle thread does lazy MMU switching as well:
7056 atomic_inc(&init_mm
.mm_count
);
7057 enter_lazy_tlb(&init_mm
, current
);
7060 * Make us the idle thread. Technically, schedule() should not be
7061 * called from this thread, however somewhere below it might be,
7062 * but because we are the idle thread, we just pick up running again
7063 * when this runqueue becomes "idle".
7065 init_idle(current
, smp_processor_id());
7067 calc_load_update
= jiffies
+ LOAD_FREQ
;
7070 * During early bootup we pretend to be a normal task:
7072 current
->sched_class
= &fair_sched_class
;
7075 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7076 /* May be allocated at isolcpus cmdline parse time */
7077 if (cpu_isolated_map
== NULL
)
7078 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7079 idle_thread_set_boot_cpu();
7081 init_sched_fair_class();
7083 scheduler_running
= 1;
7086 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7087 static inline int preempt_count_equals(int preempt_offset
)
7089 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7091 return (nested
== preempt_offset
);
7094 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7096 static unsigned long prev_jiffy
; /* ratelimiting */
7098 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7099 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7100 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7102 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7104 prev_jiffy
= jiffies
;
7107 "BUG: sleeping function called from invalid context at %s:%d\n",
7110 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7111 in_atomic(), irqs_disabled(),
7112 current
->pid
, current
->comm
);
7114 debug_show_held_locks(current
);
7115 if (irqs_disabled())
7116 print_irqtrace_events(current
);
7119 EXPORT_SYMBOL(__might_sleep
);
7122 #ifdef CONFIG_MAGIC_SYSRQ
7123 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7125 const struct sched_class
*prev_class
= p
->sched_class
;
7126 int old_prio
= p
->prio
;
7131 dequeue_task(rq
, p
, 0);
7132 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7134 enqueue_task(rq
, p
, 0);
7135 resched_task(rq
->curr
);
7138 check_class_changed(rq
, p
, prev_class
, old_prio
);
7141 void normalize_rt_tasks(void)
7143 struct task_struct
*g
, *p
;
7144 unsigned long flags
;
7147 read_lock_irqsave(&tasklist_lock
, flags
);
7148 do_each_thread(g
, p
) {
7150 * Only normalize user tasks:
7155 p
->se
.exec_start
= 0;
7156 #ifdef CONFIG_SCHEDSTATS
7157 p
->se
.statistics
.wait_start
= 0;
7158 p
->se
.statistics
.sleep_start
= 0;
7159 p
->se
.statistics
.block_start
= 0;
7164 * Renice negative nice level userspace
7167 if (TASK_NICE(p
) < 0 && p
->mm
)
7168 set_user_nice(p
, 0);
7172 raw_spin_lock(&p
->pi_lock
);
7173 rq
= __task_rq_lock(p
);
7175 normalize_task(rq
, p
);
7177 __task_rq_unlock(rq
);
7178 raw_spin_unlock(&p
->pi_lock
);
7179 } while_each_thread(g
, p
);
7181 read_unlock_irqrestore(&tasklist_lock
, flags
);
7184 #endif /* CONFIG_MAGIC_SYSRQ */
7186 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7188 * These functions are only useful for the IA64 MCA handling, or kdb.
7190 * They can only be called when the whole system has been
7191 * stopped - every CPU needs to be quiescent, and no scheduling
7192 * activity can take place. Using them for anything else would
7193 * be a serious bug, and as a result, they aren't even visible
7194 * under any other configuration.
7198 * curr_task - return the current task for a given cpu.
7199 * @cpu: the processor in question.
7201 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7203 struct task_struct
*curr_task(int cpu
)
7205 return cpu_curr(cpu
);
7208 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7212 * set_curr_task - set the current task for a given cpu.
7213 * @cpu: the processor in question.
7214 * @p: the task pointer to set.
7216 * Description: This function must only be used when non-maskable interrupts
7217 * are serviced on a separate stack. It allows the architecture to switch the
7218 * notion of the current task on a cpu in a non-blocking manner. This function
7219 * must be called with all CPU's synchronized, and interrupts disabled, the
7220 * and caller must save the original value of the current task (see
7221 * curr_task() above) and restore that value before reenabling interrupts and
7222 * re-starting the system.
7224 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7226 void set_curr_task(int cpu
, struct task_struct
*p
)
7233 #ifdef CONFIG_CGROUP_SCHED
7234 /* task_group_lock serializes the addition/removal of task groups */
7235 static DEFINE_SPINLOCK(task_group_lock
);
7237 static void free_sched_group(struct task_group
*tg
)
7239 free_fair_sched_group(tg
);
7240 free_rt_sched_group(tg
);
7245 /* allocate runqueue etc for a new task group */
7246 struct task_group
*sched_create_group(struct task_group
*parent
)
7248 struct task_group
*tg
;
7249 unsigned long flags
;
7251 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7253 return ERR_PTR(-ENOMEM
);
7255 if (!alloc_fair_sched_group(tg
, parent
))
7258 if (!alloc_rt_sched_group(tg
, parent
))
7261 spin_lock_irqsave(&task_group_lock
, flags
);
7262 list_add_rcu(&tg
->list
, &task_groups
);
7264 WARN_ON(!parent
); /* root should already exist */
7266 tg
->parent
= parent
;
7267 INIT_LIST_HEAD(&tg
->children
);
7268 list_add_rcu(&tg
->siblings
, &parent
->children
);
7269 spin_unlock_irqrestore(&task_group_lock
, flags
);
7274 free_sched_group(tg
);
7275 return ERR_PTR(-ENOMEM
);
7278 /* rcu callback to free various structures associated with a task group */
7279 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7281 /* now it should be safe to free those cfs_rqs */
7282 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7285 /* Destroy runqueue etc associated with a task group */
7286 void sched_destroy_group(struct task_group
*tg
)
7288 unsigned long flags
;
7291 /* end participation in shares distribution */
7292 for_each_possible_cpu(i
)
7293 unregister_fair_sched_group(tg
, i
);
7295 spin_lock_irqsave(&task_group_lock
, flags
);
7296 list_del_rcu(&tg
->list
);
7297 list_del_rcu(&tg
->siblings
);
7298 spin_unlock_irqrestore(&task_group_lock
, flags
);
7300 /* wait for possible concurrent references to cfs_rqs complete */
7301 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7304 /* change task's runqueue when it moves between groups.
7305 * The caller of this function should have put the task in its new group
7306 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7307 * reflect its new group.
7309 void sched_move_task(struct task_struct
*tsk
)
7312 unsigned long flags
;
7315 rq
= task_rq_lock(tsk
, &flags
);
7317 running
= task_current(rq
, tsk
);
7321 dequeue_task(rq
, tsk
, 0);
7322 if (unlikely(running
))
7323 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7325 #ifdef CONFIG_FAIR_GROUP_SCHED
7326 if (tsk
->sched_class
->task_move_group
)
7327 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7330 set_task_rq(tsk
, task_cpu(tsk
));
7332 if (unlikely(running
))
7333 tsk
->sched_class
->set_curr_task(rq
);
7335 enqueue_task(rq
, tsk
, 0);
7337 task_rq_unlock(rq
, tsk
, &flags
);
7339 #endif /* CONFIG_CGROUP_SCHED */
7341 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7342 static unsigned long to_ratio(u64 period
, u64 runtime
)
7344 if (runtime
== RUNTIME_INF
)
7347 return div64_u64(runtime
<< 20, period
);
7351 #ifdef CONFIG_RT_GROUP_SCHED
7353 * Ensure that the real time constraints are schedulable.
7355 static DEFINE_MUTEX(rt_constraints_mutex
);
7357 /* Must be called with tasklist_lock held */
7358 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7360 struct task_struct
*g
, *p
;
7362 do_each_thread(g
, p
) {
7363 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7365 } while_each_thread(g
, p
);
7370 struct rt_schedulable_data
{
7371 struct task_group
*tg
;
7376 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7378 struct rt_schedulable_data
*d
= data
;
7379 struct task_group
*child
;
7380 unsigned long total
, sum
= 0;
7381 u64 period
, runtime
;
7383 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7384 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7387 period
= d
->rt_period
;
7388 runtime
= d
->rt_runtime
;
7392 * Cannot have more runtime than the period.
7394 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7398 * Ensure we don't starve existing RT tasks.
7400 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7403 total
= to_ratio(period
, runtime
);
7406 * Nobody can have more than the global setting allows.
7408 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7412 * The sum of our children's runtime should not exceed our own.
7414 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7415 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7416 runtime
= child
->rt_bandwidth
.rt_runtime
;
7418 if (child
== d
->tg
) {
7419 period
= d
->rt_period
;
7420 runtime
= d
->rt_runtime
;
7423 sum
+= to_ratio(period
, runtime
);
7432 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7436 struct rt_schedulable_data data
= {
7438 .rt_period
= period
,
7439 .rt_runtime
= runtime
,
7443 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7449 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7450 u64 rt_period
, u64 rt_runtime
)
7454 mutex_lock(&rt_constraints_mutex
);
7455 read_lock(&tasklist_lock
);
7456 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7460 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7461 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7462 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7464 for_each_possible_cpu(i
) {
7465 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7467 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7468 rt_rq
->rt_runtime
= rt_runtime
;
7469 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7471 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7473 read_unlock(&tasklist_lock
);
7474 mutex_unlock(&rt_constraints_mutex
);
7479 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7481 u64 rt_runtime
, rt_period
;
7483 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7484 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7485 if (rt_runtime_us
< 0)
7486 rt_runtime
= RUNTIME_INF
;
7488 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7491 long sched_group_rt_runtime(struct task_group
*tg
)
7495 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7498 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7499 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7500 return rt_runtime_us
;
7503 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7505 u64 rt_runtime
, rt_period
;
7507 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7508 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7513 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7516 long sched_group_rt_period(struct task_group
*tg
)
7520 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7521 do_div(rt_period_us
, NSEC_PER_USEC
);
7522 return rt_period_us
;
7525 static int sched_rt_global_constraints(void)
7527 u64 runtime
, period
;
7530 if (sysctl_sched_rt_period
<= 0)
7533 runtime
= global_rt_runtime();
7534 period
= global_rt_period();
7537 * Sanity check on the sysctl variables.
7539 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7542 mutex_lock(&rt_constraints_mutex
);
7543 read_lock(&tasklist_lock
);
7544 ret
= __rt_schedulable(NULL
, 0, 0);
7545 read_unlock(&tasklist_lock
);
7546 mutex_unlock(&rt_constraints_mutex
);
7551 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7553 /* Don't accept realtime tasks when there is no way for them to run */
7554 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7560 #else /* !CONFIG_RT_GROUP_SCHED */
7561 static int sched_rt_global_constraints(void)
7563 unsigned long flags
;
7566 if (sysctl_sched_rt_period
<= 0)
7570 * There's always some RT tasks in the root group
7571 * -- migration, kstopmachine etc..
7573 if (sysctl_sched_rt_runtime
== 0)
7576 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7577 for_each_possible_cpu(i
) {
7578 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7580 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7581 rt_rq
->rt_runtime
= global_rt_runtime();
7582 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7584 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7588 #endif /* CONFIG_RT_GROUP_SCHED */
7590 int sched_rt_handler(struct ctl_table
*table
, int write
,
7591 void __user
*buffer
, size_t *lenp
,
7595 int old_period
, old_runtime
;
7596 static DEFINE_MUTEX(mutex
);
7599 old_period
= sysctl_sched_rt_period
;
7600 old_runtime
= sysctl_sched_rt_runtime
;
7602 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7604 if (!ret
&& write
) {
7605 ret
= sched_rt_global_constraints();
7607 sysctl_sched_rt_period
= old_period
;
7608 sysctl_sched_rt_runtime
= old_runtime
;
7610 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7611 def_rt_bandwidth
.rt_period
=
7612 ns_to_ktime(global_rt_period());
7615 mutex_unlock(&mutex
);
7620 #ifdef CONFIG_CGROUP_SCHED
7622 /* return corresponding task_group object of a cgroup */
7623 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7625 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7626 struct task_group
, css
);
7629 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7631 struct task_group
*tg
, *parent
;
7633 if (!cgrp
->parent
) {
7634 /* This is early initialization for the top cgroup */
7635 return &root_task_group
.css
;
7638 parent
= cgroup_tg(cgrp
->parent
);
7639 tg
= sched_create_group(parent
);
7641 return ERR_PTR(-ENOMEM
);
7646 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7648 struct task_group
*tg
= cgroup_tg(cgrp
);
7650 sched_destroy_group(tg
);
7653 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7654 struct cgroup_taskset
*tset
)
7656 struct task_struct
*task
;
7658 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7659 #ifdef CONFIG_RT_GROUP_SCHED
7660 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7663 /* We don't support RT-tasks being in separate groups */
7664 if (task
->sched_class
!= &fair_sched_class
)
7671 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7672 struct cgroup_taskset
*tset
)
7674 struct task_struct
*task
;
7676 cgroup_taskset_for_each(task
, cgrp
, tset
)
7677 sched_move_task(task
);
7681 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7682 struct task_struct
*task
)
7685 * cgroup_exit() is called in the copy_process() failure path.
7686 * Ignore this case since the task hasn't ran yet, this avoids
7687 * trying to poke a half freed task state from generic code.
7689 if (!(task
->flags
& PF_EXITING
))
7692 sched_move_task(task
);
7695 #ifdef CONFIG_FAIR_GROUP_SCHED
7696 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7699 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7702 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7704 struct task_group
*tg
= cgroup_tg(cgrp
);
7706 return (u64
) scale_load_down(tg
->shares
);
7709 #ifdef CONFIG_CFS_BANDWIDTH
7710 static DEFINE_MUTEX(cfs_constraints_mutex
);
7712 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7713 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7715 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7717 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7719 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7720 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7722 if (tg
== &root_task_group
)
7726 * Ensure we have at some amount of bandwidth every period. This is
7727 * to prevent reaching a state of large arrears when throttled via
7728 * entity_tick() resulting in prolonged exit starvation.
7730 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7734 * Likewise, bound things on the otherside by preventing insane quota
7735 * periods. This also allows us to normalize in computing quota
7738 if (period
> max_cfs_quota_period
)
7741 mutex_lock(&cfs_constraints_mutex
);
7742 ret
= __cfs_schedulable(tg
, period
, quota
);
7746 runtime_enabled
= quota
!= RUNTIME_INF
;
7747 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7748 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7749 raw_spin_lock_irq(&cfs_b
->lock
);
7750 cfs_b
->period
= ns_to_ktime(period
);
7751 cfs_b
->quota
= quota
;
7753 __refill_cfs_bandwidth_runtime(cfs_b
);
7754 /* restart the period timer (if active) to handle new period expiry */
7755 if (runtime_enabled
&& cfs_b
->timer_active
) {
7756 /* force a reprogram */
7757 cfs_b
->timer_active
= 0;
7758 __start_cfs_bandwidth(cfs_b
);
7760 raw_spin_unlock_irq(&cfs_b
->lock
);
7762 for_each_possible_cpu(i
) {
7763 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7764 struct rq
*rq
= cfs_rq
->rq
;
7766 raw_spin_lock_irq(&rq
->lock
);
7767 cfs_rq
->runtime_enabled
= runtime_enabled
;
7768 cfs_rq
->runtime_remaining
= 0;
7770 if (cfs_rq
->throttled
)
7771 unthrottle_cfs_rq(cfs_rq
);
7772 raw_spin_unlock_irq(&rq
->lock
);
7775 mutex_unlock(&cfs_constraints_mutex
);
7780 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7784 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7785 if (cfs_quota_us
< 0)
7786 quota
= RUNTIME_INF
;
7788 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7790 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7793 long tg_get_cfs_quota(struct task_group
*tg
)
7797 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7800 quota_us
= tg
->cfs_bandwidth
.quota
;
7801 do_div(quota_us
, NSEC_PER_USEC
);
7806 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7810 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7811 quota
= tg
->cfs_bandwidth
.quota
;
7813 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7816 long tg_get_cfs_period(struct task_group
*tg
)
7820 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7821 do_div(cfs_period_us
, NSEC_PER_USEC
);
7823 return cfs_period_us
;
7826 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7828 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7831 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7834 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7837 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7839 return tg_get_cfs_period(cgroup_tg(cgrp
));
7842 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7845 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7848 struct cfs_schedulable_data
{
7849 struct task_group
*tg
;
7854 * normalize group quota/period to be quota/max_period
7855 * note: units are usecs
7857 static u64
normalize_cfs_quota(struct task_group
*tg
,
7858 struct cfs_schedulable_data
*d
)
7866 period
= tg_get_cfs_period(tg
);
7867 quota
= tg_get_cfs_quota(tg
);
7870 /* note: these should typically be equivalent */
7871 if (quota
== RUNTIME_INF
|| quota
== -1)
7874 return to_ratio(period
, quota
);
7877 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7879 struct cfs_schedulable_data
*d
= data
;
7880 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7881 s64 quota
= 0, parent_quota
= -1;
7884 quota
= RUNTIME_INF
;
7886 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7888 quota
= normalize_cfs_quota(tg
, d
);
7889 parent_quota
= parent_b
->hierarchal_quota
;
7892 * ensure max(child_quota) <= parent_quota, inherit when no
7895 if (quota
== RUNTIME_INF
)
7896 quota
= parent_quota
;
7897 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7900 cfs_b
->hierarchal_quota
= quota
;
7905 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7908 struct cfs_schedulable_data data
= {
7914 if (quota
!= RUNTIME_INF
) {
7915 do_div(data
.period
, NSEC_PER_USEC
);
7916 do_div(data
.quota
, NSEC_PER_USEC
);
7920 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7926 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7927 struct cgroup_map_cb
*cb
)
7929 struct task_group
*tg
= cgroup_tg(cgrp
);
7930 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7932 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7933 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7934 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7938 #endif /* CONFIG_CFS_BANDWIDTH */
7939 #endif /* CONFIG_FAIR_GROUP_SCHED */
7941 #ifdef CONFIG_RT_GROUP_SCHED
7942 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7945 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7948 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7950 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7953 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7956 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7959 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7961 return sched_group_rt_period(cgroup_tg(cgrp
));
7963 #endif /* CONFIG_RT_GROUP_SCHED */
7965 static struct cftype cpu_files
[] = {
7966 #ifdef CONFIG_FAIR_GROUP_SCHED
7969 .read_u64
= cpu_shares_read_u64
,
7970 .write_u64
= cpu_shares_write_u64
,
7973 #ifdef CONFIG_CFS_BANDWIDTH
7975 .name
= "cfs_quota_us",
7976 .read_s64
= cpu_cfs_quota_read_s64
,
7977 .write_s64
= cpu_cfs_quota_write_s64
,
7980 .name
= "cfs_period_us",
7981 .read_u64
= cpu_cfs_period_read_u64
,
7982 .write_u64
= cpu_cfs_period_write_u64
,
7986 .read_map
= cpu_stats_show
,
7989 #ifdef CONFIG_RT_GROUP_SCHED
7991 .name
= "rt_runtime_us",
7992 .read_s64
= cpu_rt_runtime_read
,
7993 .write_s64
= cpu_rt_runtime_write
,
7996 .name
= "rt_period_us",
7997 .read_u64
= cpu_rt_period_read_uint
,
7998 .write_u64
= cpu_rt_period_write_uint
,
8004 struct cgroup_subsys cpu_cgroup_subsys
= {
8006 .create
= cpu_cgroup_create
,
8007 .destroy
= cpu_cgroup_destroy
,
8008 .can_attach
= cpu_cgroup_can_attach
,
8009 .attach
= cpu_cgroup_attach
,
8010 .exit
= cpu_cgroup_exit
,
8011 .subsys_id
= cpu_cgroup_subsys_id
,
8012 .base_cftypes
= cpu_files
,
8016 #endif /* CONFIG_CGROUP_SCHED */
8018 #ifdef CONFIG_CGROUP_CPUACCT
8021 * CPU accounting code for task groups.
8023 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8024 * (balbir@in.ibm.com).
8027 /* create a new cpu accounting group */
8028 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
8033 return &root_cpuacct
.css
;
8035 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8039 ca
->cpuusage
= alloc_percpu(u64
);
8043 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
8045 goto out_free_cpuusage
;
8050 free_percpu(ca
->cpuusage
);
8054 return ERR_PTR(-ENOMEM
);
8057 /* destroy an existing cpu accounting group */
8058 static void cpuacct_destroy(struct cgroup
*cgrp
)
8060 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8062 free_percpu(ca
->cpustat
);
8063 free_percpu(ca
->cpuusage
);
8067 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8069 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8072 #ifndef CONFIG_64BIT
8074 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8076 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8078 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8086 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8088 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8090 #ifndef CONFIG_64BIT
8092 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8094 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8096 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8102 /* return total cpu usage (in nanoseconds) of a group */
8103 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8105 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8106 u64 totalcpuusage
= 0;
8109 for_each_present_cpu(i
)
8110 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8112 return totalcpuusage
;
8115 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8118 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8127 for_each_present_cpu(i
)
8128 cpuacct_cpuusage_write(ca
, i
, 0);
8134 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8137 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8141 for_each_present_cpu(i
) {
8142 percpu
= cpuacct_cpuusage_read(ca
, i
);
8143 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8145 seq_printf(m
, "\n");
8149 static const char *cpuacct_stat_desc
[] = {
8150 [CPUACCT_STAT_USER
] = "user",
8151 [CPUACCT_STAT_SYSTEM
] = "system",
8154 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8155 struct cgroup_map_cb
*cb
)
8157 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8161 for_each_online_cpu(cpu
) {
8162 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8163 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8164 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8166 val
= cputime64_to_clock_t(val
);
8167 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8170 for_each_online_cpu(cpu
) {
8171 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8172 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8173 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8174 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8177 val
= cputime64_to_clock_t(val
);
8178 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8183 static struct cftype files
[] = {
8186 .read_u64
= cpuusage_read
,
8187 .write_u64
= cpuusage_write
,
8190 .name
= "usage_percpu",
8191 .read_seq_string
= cpuacct_percpu_seq_read
,
8195 .read_map
= cpuacct_stats_show
,
8201 * charge this task's execution time to its accounting group.
8203 * called with rq->lock held.
8205 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8210 if (unlikely(!cpuacct_subsys
.active
))
8213 cpu
= task_cpu(tsk
);
8219 for (; ca
; ca
= parent_ca(ca
)) {
8220 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8221 *cpuusage
+= cputime
;
8227 struct cgroup_subsys cpuacct_subsys
= {
8229 .create
= cpuacct_create
,
8230 .destroy
= cpuacct_destroy
,
8231 .subsys_id
= cpuacct_subsys_id
,
8232 .base_cftypes
= files
,
8234 #endif /* CONFIG_CGROUP_CPUACCT */