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 void update_cpu_load(struct rq
*this_rq
);
698 static void set_load_weight(struct task_struct
*p
)
700 int prio
= p
->static_prio
- MAX_RT_PRIO
;
701 struct load_weight
*load
= &p
->se
.load
;
704 * SCHED_IDLE tasks get minimal weight:
706 if (p
->policy
== SCHED_IDLE
) {
707 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
708 load
->inv_weight
= WMULT_IDLEPRIO
;
712 load
->weight
= scale_load(prio_to_weight
[prio
]);
713 load
->inv_weight
= prio_to_wmult
[prio
];
716 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
719 sched_info_queued(p
);
720 p
->sched_class
->enqueue_task(rq
, p
, flags
);
723 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
726 sched_info_dequeued(p
);
727 p
->sched_class
->dequeue_task(rq
, p
, flags
);
730 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
732 if (task_contributes_to_load(p
))
733 rq
->nr_uninterruptible
--;
735 enqueue_task(rq
, p
, flags
);
738 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
740 if (task_contributes_to_load(p
))
741 rq
->nr_uninterruptible
++;
743 dequeue_task(rq
, p
, flags
);
746 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
749 * There are no locks covering percpu hardirq/softirq time.
750 * They are only modified in account_system_vtime, on corresponding CPU
751 * with interrupts disabled. So, writes are safe.
752 * They are read and saved off onto struct rq in update_rq_clock().
753 * This may result in other CPU reading this CPU's irq time and can
754 * race with irq/account_system_vtime on this CPU. We would either get old
755 * or new value with a side effect of accounting a slice of irq time to wrong
756 * task when irq is in progress while we read rq->clock. That is a worthy
757 * compromise in place of having locks on each irq in account_system_time.
759 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
760 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
762 static DEFINE_PER_CPU(u64
, irq_start_time
);
763 static int sched_clock_irqtime
;
765 void enable_sched_clock_irqtime(void)
767 sched_clock_irqtime
= 1;
770 void disable_sched_clock_irqtime(void)
772 sched_clock_irqtime
= 0;
776 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
778 static inline void irq_time_write_begin(void)
780 __this_cpu_inc(irq_time_seq
.sequence
);
784 static inline void irq_time_write_end(void)
787 __this_cpu_inc(irq_time_seq
.sequence
);
790 static inline u64
irq_time_read(int cpu
)
796 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
797 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
798 per_cpu(cpu_hardirq_time
, cpu
);
799 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
803 #else /* CONFIG_64BIT */
804 static inline void irq_time_write_begin(void)
808 static inline void irq_time_write_end(void)
812 static inline u64
irq_time_read(int cpu
)
814 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
816 #endif /* CONFIG_64BIT */
819 * Called before incrementing preempt_count on {soft,}irq_enter
820 * and before decrementing preempt_count on {soft,}irq_exit.
822 void account_system_vtime(struct task_struct
*curr
)
828 if (!sched_clock_irqtime
)
831 local_irq_save(flags
);
833 cpu
= smp_processor_id();
834 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
835 __this_cpu_add(irq_start_time
, delta
);
837 irq_time_write_begin();
839 * We do not account for softirq time from ksoftirqd here.
840 * We want to continue accounting softirq time to ksoftirqd thread
841 * in that case, so as not to confuse scheduler with a special task
842 * that do not consume any time, but still wants to run.
845 __this_cpu_add(cpu_hardirq_time
, delta
);
846 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
847 __this_cpu_add(cpu_softirq_time
, delta
);
849 irq_time_write_end();
850 local_irq_restore(flags
);
852 EXPORT_SYMBOL_GPL(account_system_vtime
);
854 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
856 #ifdef CONFIG_PARAVIRT
857 static inline u64
steal_ticks(u64 steal
)
859 if (unlikely(steal
> NSEC_PER_SEC
))
860 return div_u64(steal
, TICK_NSEC
);
862 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
866 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
869 * In theory, the compile should just see 0 here, and optimize out the call
870 * to sched_rt_avg_update. But I don't trust it...
872 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
873 s64 steal
= 0, irq_delta
= 0;
875 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
876 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
879 * Since irq_time is only updated on {soft,}irq_exit, we might run into
880 * this case when a previous update_rq_clock() happened inside a
883 * When this happens, we stop ->clock_task and only update the
884 * prev_irq_time stamp to account for the part that fit, so that a next
885 * update will consume the rest. This ensures ->clock_task is
888 * It does however cause some slight miss-attribution of {soft,}irq
889 * time, a more accurate solution would be to update the irq_time using
890 * the current rq->clock timestamp, except that would require using
893 if (irq_delta
> delta
)
896 rq
->prev_irq_time
+= irq_delta
;
899 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
900 if (static_key_false((¶virt_steal_rq_enabled
))) {
903 steal
= paravirt_steal_clock(cpu_of(rq
));
904 steal
-= rq
->prev_steal_time_rq
;
906 if (unlikely(steal
> delta
))
909 st
= steal_ticks(steal
);
910 steal
= st
* TICK_NSEC
;
912 rq
->prev_steal_time_rq
+= steal
;
918 rq
->clock_task
+= delta
;
920 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
921 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
922 sched_rt_avg_update(rq
, irq_delta
+ steal
);
926 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
927 static int irqtime_account_hi_update(void)
929 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
934 local_irq_save(flags
);
935 latest_ns
= this_cpu_read(cpu_hardirq_time
);
936 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_IRQ
])
938 local_irq_restore(flags
);
942 static int irqtime_account_si_update(void)
944 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
949 local_irq_save(flags
);
950 latest_ns
= this_cpu_read(cpu_softirq_time
);
951 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_SOFTIRQ
])
953 local_irq_restore(flags
);
957 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
959 #define sched_clock_irqtime (0)
963 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
965 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
966 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
970 * Make it appear like a SCHED_FIFO task, its something
971 * userspace knows about and won't get confused about.
973 * Also, it will make PI more or less work without too
974 * much confusion -- but then, stop work should not
975 * rely on PI working anyway.
977 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
979 stop
->sched_class
= &stop_sched_class
;
982 cpu_rq(cpu
)->stop
= stop
;
986 * Reset it back to a normal scheduling class so that
987 * it can die in pieces.
989 old_stop
->sched_class
= &rt_sched_class
;
994 * __normal_prio - return the priority that is based on the static prio
996 static inline int __normal_prio(struct task_struct
*p
)
998 return p
->static_prio
;
1002 * Calculate the expected normal priority: i.e. priority
1003 * without taking RT-inheritance into account. Might be
1004 * boosted by interactivity modifiers. Changes upon fork,
1005 * setprio syscalls, and whenever the interactivity
1006 * estimator recalculates.
1008 static inline int normal_prio(struct task_struct
*p
)
1012 if (task_has_rt_policy(p
))
1013 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1015 prio
= __normal_prio(p
);
1020 * Calculate the current priority, i.e. the priority
1021 * taken into account by the scheduler. This value might
1022 * be boosted by RT tasks, or might be boosted by
1023 * interactivity modifiers. Will be RT if the task got
1024 * RT-boosted. If not then it returns p->normal_prio.
1026 static int effective_prio(struct task_struct
*p
)
1028 p
->normal_prio
= normal_prio(p
);
1030 * If we are RT tasks or we were boosted to RT priority,
1031 * keep the priority unchanged. Otherwise, update priority
1032 * to the normal priority:
1034 if (!rt_prio(p
->prio
))
1035 return p
->normal_prio
;
1040 * task_curr - is this task currently executing on a CPU?
1041 * @p: the task in question.
1043 inline int task_curr(const struct task_struct
*p
)
1045 return cpu_curr(task_cpu(p
)) == p
;
1048 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1049 const struct sched_class
*prev_class
,
1052 if (prev_class
!= p
->sched_class
) {
1053 if (prev_class
->switched_from
)
1054 prev_class
->switched_from(rq
, p
);
1055 p
->sched_class
->switched_to(rq
, p
);
1056 } else if (oldprio
!= p
->prio
)
1057 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1060 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1062 const struct sched_class
*class;
1064 if (p
->sched_class
== rq
->curr
->sched_class
) {
1065 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1067 for_each_class(class) {
1068 if (class == rq
->curr
->sched_class
)
1070 if (class == p
->sched_class
) {
1071 resched_task(rq
->curr
);
1078 * A queue event has occurred, and we're going to schedule. In
1079 * this case, we can save a useless back to back clock update.
1081 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1082 rq
->skip_clock_update
= 1;
1086 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1088 #ifdef CONFIG_SCHED_DEBUG
1090 * We should never call set_task_cpu() on a blocked task,
1091 * ttwu() will sort out the placement.
1093 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1094 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
1096 #ifdef CONFIG_LOCKDEP
1098 * The caller should hold either p->pi_lock or rq->lock, when changing
1099 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1101 * sched_move_task() holds both and thus holding either pins the cgroup,
1102 * see set_task_rq().
1104 * Furthermore, all task_rq users should acquire both locks, see
1107 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1108 lockdep_is_held(&task_rq(p
)->lock
)));
1112 trace_sched_migrate_task(p
, new_cpu
);
1114 if (task_cpu(p
) != new_cpu
) {
1115 p
->se
.nr_migrations
++;
1116 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1119 __set_task_cpu(p
, new_cpu
);
1122 struct migration_arg
{
1123 struct task_struct
*task
;
1127 static int migration_cpu_stop(void *data
);
1130 * wait_task_inactive - wait for a thread to unschedule.
1132 * If @match_state is nonzero, it's the @p->state value just checked and
1133 * not expected to change. If it changes, i.e. @p might have woken up,
1134 * then return zero. When we succeed in waiting for @p to be off its CPU,
1135 * we return a positive number (its total switch count). If a second call
1136 * a short while later returns the same number, the caller can be sure that
1137 * @p has remained unscheduled the whole time.
1139 * The caller must ensure that the task *will* unschedule sometime soon,
1140 * else this function might spin for a *long* time. This function can't
1141 * be called with interrupts off, or it may introduce deadlock with
1142 * smp_call_function() if an IPI is sent by the same process we are
1143 * waiting to become inactive.
1145 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1147 unsigned long flags
;
1154 * We do the initial early heuristics without holding
1155 * any task-queue locks at all. We'll only try to get
1156 * the runqueue lock when things look like they will
1162 * If the task is actively running on another CPU
1163 * still, just relax and busy-wait without holding
1166 * NOTE! Since we don't hold any locks, it's not
1167 * even sure that "rq" stays as the right runqueue!
1168 * But we don't care, since "task_running()" will
1169 * return false if the runqueue has changed and p
1170 * is actually now running somewhere else!
1172 while (task_running(rq
, p
)) {
1173 if (match_state
&& unlikely(p
->state
!= match_state
))
1179 * Ok, time to look more closely! We need the rq
1180 * lock now, to be *sure*. If we're wrong, we'll
1181 * just go back and repeat.
1183 rq
= task_rq_lock(p
, &flags
);
1184 trace_sched_wait_task(p
);
1185 running
= task_running(rq
, p
);
1188 if (!match_state
|| p
->state
== match_state
)
1189 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1190 task_rq_unlock(rq
, p
, &flags
);
1193 * If it changed from the expected state, bail out now.
1195 if (unlikely(!ncsw
))
1199 * Was it really running after all now that we
1200 * checked with the proper locks actually held?
1202 * Oops. Go back and try again..
1204 if (unlikely(running
)) {
1210 * It's not enough that it's not actively running,
1211 * it must be off the runqueue _entirely_, and not
1214 * So if it was still runnable (but just not actively
1215 * running right now), it's preempted, and we should
1216 * yield - it could be a while.
1218 if (unlikely(on_rq
)) {
1219 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1221 set_current_state(TASK_UNINTERRUPTIBLE
);
1222 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1227 * Ahh, all good. It wasn't running, and it wasn't
1228 * runnable, which means that it will never become
1229 * running in the future either. We're all done!
1238 * kick_process - kick a running thread to enter/exit the kernel
1239 * @p: the to-be-kicked thread
1241 * Cause a process which is running on another CPU to enter
1242 * kernel-mode, without any delay. (to get signals handled.)
1244 * NOTE: this function doesn't have to take the runqueue lock,
1245 * because all it wants to ensure is that the remote task enters
1246 * the kernel. If the IPI races and the task has been migrated
1247 * to another CPU then no harm is done and the purpose has been
1250 void kick_process(struct task_struct
*p
)
1256 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1257 smp_send_reschedule(cpu
);
1260 EXPORT_SYMBOL_GPL(kick_process
);
1261 #endif /* CONFIG_SMP */
1265 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1267 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1269 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1270 enum { cpuset
, possible
, fail
} state
= cpuset
;
1273 /* Look for allowed, online CPU in same node. */
1274 for_each_cpu(dest_cpu
, nodemask
) {
1275 if (!cpu_online(dest_cpu
))
1277 if (!cpu_active(dest_cpu
))
1279 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1284 /* Any allowed, online CPU? */
1285 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1286 if (!cpu_online(dest_cpu
))
1288 if (!cpu_active(dest_cpu
))
1295 /* No more Mr. Nice Guy. */
1296 cpuset_cpus_allowed_fallback(p
);
1301 do_set_cpus_allowed(p
, cpu_possible_mask
);
1312 if (state
!= cpuset
) {
1314 * Don't tell them about moving exiting tasks or
1315 * kernel threads (both mm NULL), since they never
1318 if (p
->mm
&& printk_ratelimit()) {
1319 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1320 task_pid_nr(p
), p
->comm
, cpu
);
1328 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1331 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1333 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1336 * In order not to call set_task_cpu() on a blocking task we need
1337 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1340 * Since this is common to all placement strategies, this lives here.
1342 * [ this allows ->select_task() to simply return task_cpu(p) and
1343 * not worry about this generic constraint ]
1345 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1347 cpu
= select_fallback_rq(task_cpu(p
), p
);
1352 static void update_avg(u64
*avg
, u64 sample
)
1354 s64 diff
= sample
- *avg
;
1360 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1362 #ifdef CONFIG_SCHEDSTATS
1363 struct rq
*rq
= this_rq();
1366 int this_cpu
= smp_processor_id();
1368 if (cpu
== this_cpu
) {
1369 schedstat_inc(rq
, ttwu_local
);
1370 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1372 struct sched_domain
*sd
;
1374 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1376 for_each_domain(this_cpu
, sd
) {
1377 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1378 schedstat_inc(sd
, ttwu_wake_remote
);
1385 if (wake_flags
& WF_MIGRATED
)
1386 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1388 #endif /* CONFIG_SMP */
1390 schedstat_inc(rq
, ttwu_count
);
1391 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1393 if (wake_flags
& WF_SYNC
)
1394 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1396 #endif /* CONFIG_SCHEDSTATS */
1399 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1401 activate_task(rq
, p
, en_flags
);
1404 /* if a worker is waking up, notify workqueue */
1405 if (p
->flags
& PF_WQ_WORKER
)
1406 wq_worker_waking_up(p
, cpu_of(rq
));
1410 * Mark the task runnable and perform wakeup-preemption.
1413 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1415 trace_sched_wakeup(p
, true);
1416 check_preempt_curr(rq
, p
, wake_flags
);
1418 p
->state
= TASK_RUNNING
;
1420 if (p
->sched_class
->task_woken
)
1421 p
->sched_class
->task_woken(rq
, p
);
1423 if (rq
->idle_stamp
) {
1424 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1425 u64 max
= 2*sysctl_sched_migration_cost
;
1430 update_avg(&rq
->avg_idle
, delta
);
1437 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1440 if (p
->sched_contributes_to_load
)
1441 rq
->nr_uninterruptible
--;
1444 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1445 ttwu_do_wakeup(rq
, p
, wake_flags
);
1449 * Called in case the task @p isn't fully descheduled from its runqueue,
1450 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1451 * since all we need to do is flip p->state to TASK_RUNNING, since
1452 * the task is still ->on_rq.
1454 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1459 rq
= __task_rq_lock(p
);
1461 ttwu_do_wakeup(rq
, p
, wake_flags
);
1464 __task_rq_unlock(rq
);
1470 static void sched_ttwu_pending(void)
1472 struct rq
*rq
= this_rq();
1473 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1474 struct task_struct
*p
;
1476 raw_spin_lock(&rq
->lock
);
1479 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1480 llist
= llist_next(llist
);
1481 ttwu_do_activate(rq
, p
, 0);
1484 raw_spin_unlock(&rq
->lock
);
1487 void scheduler_ipi(void)
1489 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1493 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1494 * traditionally all their work was done from the interrupt return
1495 * path. Now that we actually do some work, we need to make sure
1498 * Some archs already do call them, luckily irq_enter/exit nest
1501 * Arguably we should visit all archs and update all handlers,
1502 * however a fair share of IPIs are still resched only so this would
1503 * somewhat pessimize the simple resched case.
1506 sched_ttwu_pending();
1509 * Check if someone kicked us for doing the nohz idle load balance.
1511 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1512 this_rq()->idle_balance
= 1;
1513 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1518 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1520 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1521 smp_send_reschedule(cpu
);
1524 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1525 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1530 rq
= __task_rq_lock(p
);
1532 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1533 ttwu_do_wakeup(rq
, p
, wake_flags
);
1536 __task_rq_unlock(rq
);
1541 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1543 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1545 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1547 #endif /* CONFIG_SMP */
1549 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1551 struct rq
*rq
= cpu_rq(cpu
);
1553 #if defined(CONFIG_SMP)
1554 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1555 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1556 ttwu_queue_remote(p
, cpu
);
1561 raw_spin_lock(&rq
->lock
);
1562 ttwu_do_activate(rq
, p
, 0);
1563 raw_spin_unlock(&rq
->lock
);
1567 * try_to_wake_up - wake up a thread
1568 * @p: the thread to be awakened
1569 * @state: the mask of task states that can be woken
1570 * @wake_flags: wake modifier flags (WF_*)
1572 * Put it on the run-queue if it's not already there. The "current"
1573 * thread is always on the run-queue (except when the actual
1574 * re-schedule is in progress), and as such you're allowed to do
1575 * the simpler "current->state = TASK_RUNNING" to mark yourself
1576 * runnable without the overhead of this.
1578 * Returns %true if @p was woken up, %false if it was already running
1579 * or @state didn't match @p's state.
1582 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1584 unsigned long flags
;
1585 int cpu
, success
= 0;
1588 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1589 if (!(p
->state
& state
))
1592 success
= 1; /* we're going to change ->state */
1595 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1600 * If the owning (remote) cpu is still in the middle of schedule() with
1601 * this task as prev, wait until its done referencing the task.
1604 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1606 * In case the architecture enables interrupts in
1607 * context_switch(), we cannot busy wait, since that
1608 * would lead to deadlocks when an interrupt hits and
1609 * tries to wake up @prev. So bail and do a complete
1612 if (ttwu_activate_remote(p
, wake_flags
))
1619 * Pairs with the smp_wmb() in finish_lock_switch().
1623 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1624 p
->state
= TASK_WAKING
;
1626 if (p
->sched_class
->task_waking
)
1627 p
->sched_class
->task_waking(p
);
1629 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1630 if (task_cpu(p
) != cpu
) {
1631 wake_flags
|= WF_MIGRATED
;
1632 set_task_cpu(p
, cpu
);
1634 #endif /* CONFIG_SMP */
1638 ttwu_stat(p
, cpu
, wake_flags
);
1640 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1646 * try_to_wake_up_local - try to wake up a local task with rq lock held
1647 * @p: the thread to be awakened
1649 * Put @p on the run-queue if it's not already there. The caller must
1650 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1653 static void try_to_wake_up_local(struct task_struct
*p
)
1655 struct rq
*rq
= task_rq(p
);
1657 BUG_ON(rq
!= this_rq());
1658 BUG_ON(p
== current
);
1659 lockdep_assert_held(&rq
->lock
);
1661 if (!raw_spin_trylock(&p
->pi_lock
)) {
1662 raw_spin_unlock(&rq
->lock
);
1663 raw_spin_lock(&p
->pi_lock
);
1664 raw_spin_lock(&rq
->lock
);
1667 if (!(p
->state
& TASK_NORMAL
))
1671 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1673 ttwu_do_wakeup(rq
, p
, 0);
1674 ttwu_stat(p
, smp_processor_id(), 0);
1676 raw_spin_unlock(&p
->pi_lock
);
1680 * wake_up_process - Wake up a specific process
1681 * @p: The process to be woken up.
1683 * Attempt to wake up the nominated process and move it to the set of runnable
1684 * processes. Returns 1 if the process was woken up, 0 if it was already
1687 * It may be assumed that this function implies a write memory barrier before
1688 * changing the task state if and only if any tasks are woken up.
1690 int wake_up_process(struct task_struct
*p
)
1692 return try_to_wake_up(p
, TASK_ALL
, 0);
1694 EXPORT_SYMBOL(wake_up_process
);
1696 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1698 return try_to_wake_up(p
, state
, 0);
1702 * Perform scheduler related setup for a newly forked process p.
1703 * p is forked by current.
1705 * __sched_fork() is basic setup used by init_idle() too:
1707 static void __sched_fork(struct task_struct
*p
)
1712 p
->se
.exec_start
= 0;
1713 p
->se
.sum_exec_runtime
= 0;
1714 p
->se
.prev_sum_exec_runtime
= 0;
1715 p
->se
.nr_migrations
= 0;
1717 INIT_LIST_HEAD(&p
->se
.group_node
);
1719 #ifdef CONFIG_SCHEDSTATS
1720 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1723 INIT_LIST_HEAD(&p
->rt
.run_list
);
1725 #ifdef CONFIG_PREEMPT_NOTIFIERS
1726 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1731 * fork()/clone()-time setup:
1733 void sched_fork(struct task_struct
*p
)
1735 unsigned long flags
;
1736 int cpu
= get_cpu();
1740 * We mark the process as running here. This guarantees that
1741 * nobody will actually run it, and a signal or other external
1742 * event cannot wake it up and insert it on the runqueue either.
1744 p
->state
= TASK_RUNNING
;
1747 * Make sure we do not leak PI boosting priority to the child.
1749 p
->prio
= current
->normal_prio
;
1752 * Revert to default priority/policy on fork if requested.
1754 if (unlikely(p
->sched_reset_on_fork
)) {
1755 if (task_has_rt_policy(p
)) {
1756 p
->policy
= SCHED_NORMAL
;
1757 p
->static_prio
= NICE_TO_PRIO(0);
1759 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1760 p
->static_prio
= NICE_TO_PRIO(0);
1762 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1766 * We don't need the reset flag anymore after the fork. It has
1767 * fulfilled its duty:
1769 p
->sched_reset_on_fork
= 0;
1772 if (!rt_prio(p
->prio
))
1773 p
->sched_class
= &fair_sched_class
;
1775 if (p
->sched_class
->task_fork
)
1776 p
->sched_class
->task_fork(p
);
1779 * The child is not yet in the pid-hash so no cgroup attach races,
1780 * and the cgroup is pinned to this child due to cgroup_fork()
1781 * is ran before sched_fork().
1783 * Silence PROVE_RCU.
1785 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1786 set_task_cpu(p
, cpu
);
1787 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1789 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1790 if (likely(sched_info_on()))
1791 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1793 #if defined(CONFIG_SMP)
1796 #ifdef CONFIG_PREEMPT_COUNT
1797 /* Want to start with kernel preemption disabled. */
1798 task_thread_info(p
)->preempt_count
= 1;
1801 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1808 * wake_up_new_task - wake up a newly created task for the first time.
1810 * This function will do some initial scheduler statistics housekeeping
1811 * that must be done for every newly created context, then puts the task
1812 * on the runqueue and wakes it.
1814 void wake_up_new_task(struct task_struct
*p
)
1816 unsigned long flags
;
1819 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1822 * Fork balancing, do it here and not earlier because:
1823 * - cpus_allowed can change in the fork path
1824 * - any previously selected cpu might disappear through hotplug
1826 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1829 rq
= __task_rq_lock(p
);
1830 activate_task(rq
, p
, 0);
1832 trace_sched_wakeup_new(p
, true);
1833 check_preempt_curr(rq
, p
, WF_FORK
);
1835 if (p
->sched_class
->task_woken
)
1836 p
->sched_class
->task_woken(rq
, p
);
1838 task_rq_unlock(rq
, p
, &flags
);
1841 #ifdef CONFIG_PREEMPT_NOTIFIERS
1844 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1845 * @notifier: notifier struct to register
1847 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1849 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1851 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1854 * preempt_notifier_unregister - no longer interested in preemption notifications
1855 * @notifier: notifier struct to unregister
1857 * This is safe to call from within a preemption notifier.
1859 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1861 hlist_del(¬ifier
->link
);
1863 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1865 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1867 struct preempt_notifier
*notifier
;
1868 struct hlist_node
*node
;
1870 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1871 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1875 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1876 struct task_struct
*next
)
1878 struct preempt_notifier
*notifier
;
1879 struct hlist_node
*node
;
1881 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1882 notifier
->ops
->sched_out(notifier
, next
);
1885 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1887 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1892 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1893 struct task_struct
*next
)
1897 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1900 * prepare_task_switch - prepare to switch tasks
1901 * @rq: the runqueue preparing to switch
1902 * @prev: the current task that is being switched out
1903 * @next: the task we are going to switch to.
1905 * This is called with the rq lock held and interrupts off. It must
1906 * be paired with a subsequent finish_task_switch after the context
1909 * prepare_task_switch sets up locking and calls architecture specific
1913 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1914 struct task_struct
*next
)
1916 sched_info_switch(prev
, next
);
1917 perf_event_task_sched_out(prev
, next
);
1918 fire_sched_out_preempt_notifiers(prev
, next
);
1919 prepare_lock_switch(rq
, next
);
1920 prepare_arch_switch(next
);
1921 trace_sched_switch(prev
, next
);
1925 * finish_task_switch - clean up after a task-switch
1926 * @rq: runqueue associated with task-switch
1927 * @prev: the thread we just switched away from.
1929 * finish_task_switch must be called after the context switch, paired
1930 * with a prepare_task_switch call before the context switch.
1931 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1932 * and do any other architecture-specific cleanup actions.
1934 * Note that we may have delayed dropping an mm in context_switch(). If
1935 * so, we finish that here outside of the runqueue lock. (Doing it
1936 * with the lock held can cause deadlocks; see schedule() for
1939 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1940 __releases(rq
->lock
)
1942 struct mm_struct
*mm
= rq
->prev_mm
;
1948 * A task struct has one reference for the use as "current".
1949 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1950 * schedule one last time. The schedule call will never return, and
1951 * the scheduled task must drop that reference.
1952 * The test for TASK_DEAD must occur while the runqueue locks are
1953 * still held, otherwise prev could be scheduled on another cpu, die
1954 * there before we look at prev->state, and then the reference would
1956 * Manfred Spraul <manfred@colorfullife.com>
1958 prev_state
= prev
->state
;
1959 finish_arch_switch(prev
);
1960 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1961 local_irq_disable();
1962 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1963 perf_event_task_sched_in(prev
, current
);
1964 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1966 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1967 finish_lock_switch(rq
, prev
);
1968 finish_arch_post_lock_switch();
1970 fire_sched_in_preempt_notifiers(current
);
1973 if (unlikely(prev_state
== TASK_DEAD
)) {
1975 * Remove function-return probe instances associated with this
1976 * task and put them back on the free list.
1978 kprobe_flush_task(prev
);
1979 put_task_struct(prev
);
1985 /* assumes rq->lock is held */
1986 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1988 if (prev
->sched_class
->pre_schedule
)
1989 prev
->sched_class
->pre_schedule(rq
, prev
);
1992 /* rq->lock is NOT held, but preemption is disabled */
1993 static inline void post_schedule(struct rq
*rq
)
1995 if (rq
->post_schedule
) {
1996 unsigned long flags
;
1998 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1999 if (rq
->curr
->sched_class
->post_schedule
)
2000 rq
->curr
->sched_class
->post_schedule(rq
);
2001 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2003 rq
->post_schedule
= 0;
2009 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2013 static inline void post_schedule(struct rq
*rq
)
2020 * schedule_tail - first thing a freshly forked thread must call.
2021 * @prev: the thread we just switched away from.
2023 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2024 __releases(rq
->lock
)
2026 struct rq
*rq
= this_rq();
2028 finish_task_switch(rq
, prev
);
2031 * FIXME: do we need to worry about rq being invalidated by the
2036 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2037 /* In this case, finish_task_switch does not reenable preemption */
2040 if (current
->set_child_tid
)
2041 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2045 * context_switch - switch to the new MM and the new
2046 * thread's register state.
2049 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2050 struct task_struct
*next
)
2052 struct mm_struct
*mm
, *oldmm
;
2054 prepare_task_switch(rq
, prev
, next
);
2057 oldmm
= prev
->active_mm
;
2059 * For paravirt, this is coupled with an exit in switch_to to
2060 * combine the page table reload and the switch backend into
2063 arch_start_context_switch(prev
);
2066 next
->active_mm
= oldmm
;
2067 atomic_inc(&oldmm
->mm_count
);
2068 enter_lazy_tlb(oldmm
, next
);
2070 switch_mm(oldmm
, mm
, next
);
2073 prev
->active_mm
= NULL
;
2074 rq
->prev_mm
= oldmm
;
2077 * Since the runqueue lock will be released by the next
2078 * task (which is an invalid locking op but in the case
2079 * of the scheduler it's an obvious special-case), so we
2080 * do an early lockdep release here:
2082 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2083 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2086 /* Here we just switch the register state and the stack. */
2087 rcu_switch_from(prev
);
2088 switch_to(prev
, next
, prev
);
2092 * this_rq must be evaluated again because prev may have moved
2093 * CPUs since it called schedule(), thus the 'rq' on its stack
2094 * frame will be invalid.
2096 finish_task_switch(this_rq(), prev
);
2100 * nr_running, nr_uninterruptible and nr_context_switches:
2102 * externally visible scheduler statistics: current number of runnable
2103 * threads, current number of uninterruptible-sleeping threads, total
2104 * number of context switches performed since bootup.
2106 unsigned long nr_running(void)
2108 unsigned long i
, sum
= 0;
2110 for_each_online_cpu(i
)
2111 sum
+= cpu_rq(i
)->nr_running
;
2116 unsigned long nr_uninterruptible(void)
2118 unsigned long i
, sum
= 0;
2120 for_each_possible_cpu(i
)
2121 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2124 * Since we read the counters lockless, it might be slightly
2125 * inaccurate. Do not allow it to go below zero though:
2127 if (unlikely((long)sum
< 0))
2133 unsigned long long nr_context_switches(void)
2136 unsigned long long sum
= 0;
2138 for_each_possible_cpu(i
)
2139 sum
+= cpu_rq(i
)->nr_switches
;
2144 unsigned long nr_iowait(void)
2146 unsigned long i
, sum
= 0;
2148 for_each_possible_cpu(i
)
2149 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2154 unsigned long nr_iowait_cpu(int cpu
)
2156 struct rq
*this = cpu_rq(cpu
);
2157 return atomic_read(&this->nr_iowait
);
2160 unsigned long this_cpu_load(void)
2162 struct rq
*this = this_rq();
2163 return this->cpu_load
[0];
2167 /* Variables and functions for calc_load */
2168 static atomic_long_t calc_load_tasks
;
2169 static unsigned long calc_load_update
;
2170 unsigned long avenrun
[3];
2171 EXPORT_SYMBOL(avenrun
);
2173 static long calc_load_fold_active(struct rq
*this_rq
)
2175 long nr_active
, delta
= 0;
2177 nr_active
= this_rq
->nr_running
;
2178 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2180 if (nr_active
!= this_rq
->calc_load_active
) {
2181 delta
= nr_active
- this_rq
->calc_load_active
;
2182 this_rq
->calc_load_active
= nr_active
;
2188 static unsigned long
2189 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2192 load
+= active
* (FIXED_1
- exp
);
2193 load
+= 1UL << (FSHIFT
- 1);
2194 return load
>> FSHIFT
;
2199 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2201 * When making the ILB scale, we should try to pull this in as well.
2203 static atomic_long_t calc_load_tasks_idle
;
2205 void calc_load_account_idle(struct rq
*this_rq
)
2209 delta
= calc_load_fold_active(this_rq
);
2211 atomic_long_add(delta
, &calc_load_tasks_idle
);
2214 static long calc_load_fold_idle(void)
2219 * Its got a race, we don't care...
2221 if (atomic_long_read(&calc_load_tasks_idle
))
2222 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2228 * fixed_power_int - compute: x^n, in O(log n) time
2230 * @x: base of the power
2231 * @frac_bits: fractional bits of @x
2232 * @n: power to raise @x to.
2234 * By exploiting the relation between the definition of the natural power
2235 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2236 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2237 * (where: n_i \elem {0, 1}, the binary vector representing n),
2238 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2239 * of course trivially computable in O(log_2 n), the length of our binary
2242 static unsigned long
2243 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2245 unsigned long result
= 1UL << frac_bits
;
2250 result
+= 1UL << (frac_bits
- 1);
2251 result
>>= frac_bits
;
2257 x
+= 1UL << (frac_bits
- 1);
2265 * a1 = a0 * e + a * (1 - e)
2267 * a2 = a1 * e + a * (1 - e)
2268 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2269 * = a0 * e^2 + a * (1 - e) * (1 + e)
2271 * a3 = a2 * e + a * (1 - e)
2272 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2273 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2277 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2278 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2279 * = a0 * e^n + a * (1 - e^n)
2281 * [1] application of the geometric series:
2284 * S_n := \Sum x^i = -------------
2287 static unsigned long
2288 calc_load_n(unsigned long load
, unsigned long exp
,
2289 unsigned long active
, unsigned int n
)
2292 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2296 * NO_HZ can leave us missing all per-cpu ticks calling
2297 * calc_load_account_active(), but since an idle CPU folds its delta into
2298 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2299 * in the pending idle delta if our idle period crossed a load cycle boundary.
2301 * Once we've updated the global active value, we need to apply the exponential
2302 * weights adjusted to the number of cycles missed.
2304 static void calc_global_nohz(void)
2306 long delta
, active
, n
;
2309 * If we crossed a calc_load_update boundary, make sure to fold
2310 * any pending idle changes, the respective CPUs might have
2311 * missed the tick driven calc_load_account_active() update
2314 delta
= calc_load_fold_idle();
2316 atomic_long_add(delta
, &calc_load_tasks
);
2319 * It could be the one fold was all it took, we done!
2321 if (time_before(jiffies
, calc_load_update
+ 10))
2325 * Catch-up, fold however many we are behind still
2327 delta
= jiffies
- calc_load_update
- 10;
2328 n
= 1 + (delta
/ LOAD_FREQ
);
2330 active
= atomic_long_read(&calc_load_tasks
);
2331 active
= active
> 0 ? active
* FIXED_1
: 0;
2333 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2334 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2335 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2337 calc_load_update
+= n
* LOAD_FREQ
;
2340 void calc_load_account_idle(struct rq
*this_rq
)
2344 static inline long calc_load_fold_idle(void)
2349 static void calc_global_nohz(void)
2355 * get_avenrun - get the load average array
2356 * @loads: pointer to dest load array
2357 * @offset: offset to add
2358 * @shift: shift count to shift the result left
2360 * These values are estimates at best, so no need for locking.
2362 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2364 loads
[0] = (avenrun
[0] + offset
) << shift
;
2365 loads
[1] = (avenrun
[1] + offset
) << shift
;
2366 loads
[2] = (avenrun
[2] + offset
) << shift
;
2370 * calc_load - update the avenrun load estimates 10 ticks after the
2371 * CPUs have updated calc_load_tasks.
2373 void calc_global_load(unsigned long ticks
)
2377 if (time_before(jiffies
, calc_load_update
+ 10))
2380 active
= atomic_long_read(&calc_load_tasks
);
2381 active
= active
> 0 ? active
* FIXED_1
: 0;
2383 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2384 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2385 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2387 calc_load_update
+= LOAD_FREQ
;
2390 * Account one period with whatever state we found before
2391 * folding in the nohz state and ageing the entire idle period.
2393 * This avoids loosing a sample when we go idle between
2394 * calc_load_account_active() (10 ticks ago) and now and thus
2401 * Called from update_cpu_load() to periodically update this CPU's
2404 static void calc_load_account_active(struct rq
*this_rq
)
2408 if (time_before(jiffies
, this_rq
->calc_load_update
))
2411 delta
= calc_load_fold_active(this_rq
);
2412 delta
+= calc_load_fold_idle();
2414 atomic_long_add(delta
, &calc_load_tasks
);
2416 this_rq
->calc_load_update
+= LOAD_FREQ
;
2420 * The exact cpuload at various idx values, calculated at every tick would be
2421 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2423 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2424 * on nth tick when cpu may be busy, then we have:
2425 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2426 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2428 * decay_load_missed() below does efficient calculation of
2429 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2430 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2432 * The calculation is approximated on a 128 point scale.
2433 * degrade_zero_ticks is the number of ticks after which load at any
2434 * particular idx is approximated to be zero.
2435 * degrade_factor is a precomputed table, a row for each load idx.
2436 * Each column corresponds to degradation factor for a power of two ticks,
2437 * based on 128 point scale.
2439 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2440 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2442 * With this power of 2 load factors, we can degrade the load n times
2443 * by looking at 1 bits in n and doing as many mult/shift instead of
2444 * n mult/shifts needed by the exact degradation.
2446 #define DEGRADE_SHIFT 7
2447 static const unsigned char
2448 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2449 static const unsigned char
2450 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2451 {0, 0, 0, 0, 0, 0, 0, 0},
2452 {64, 32, 8, 0, 0, 0, 0, 0},
2453 {96, 72, 40, 12, 1, 0, 0},
2454 {112, 98, 75, 43, 15, 1, 0},
2455 {120, 112, 98, 76, 45, 16, 2} };
2458 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2459 * would be when CPU is idle and so we just decay the old load without
2460 * adding any new load.
2462 static unsigned long
2463 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2467 if (!missed_updates
)
2470 if (missed_updates
>= degrade_zero_ticks
[idx
])
2474 return load
>> missed_updates
;
2476 while (missed_updates
) {
2477 if (missed_updates
% 2)
2478 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2480 missed_updates
>>= 1;
2487 * Update rq->cpu_load[] statistics. This function is usually called every
2488 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2489 * every tick. We fix it up based on jiffies.
2491 void update_cpu_load(struct rq
*this_rq
)
2493 unsigned long this_load
= this_rq
->load
.weight
;
2494 unsigned long curr_jiffies
= jiffies
;
2495 unsigned long pending_updates
;
2498 this_rq
->nr_load_updates
++;
2500 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2501 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2504 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2505 this_rq
->last_load_update_tick
= curr_jiffies
;
2507 /* Update our load: */
2508 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2509 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2510 unsigned long old_load
, new_load
;
2512 /* scale is effectively 1 << i now, and >> i divides by scale */
2514 old_load
= this_rq
->cpu_load
[i
];
2515 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2516 new_load
= this_load
;
2518 * Round up the averaging division if load is increasing. This
2519 * prevents us from getting stuck on 9 if the load is 10, for
2522 if (new_load
> old_load
)
2523 new_load
+= scale
- 1;
2525 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2528 sched_avg_update(this_rq
);
2531 static void update_cpu_load_active(struct rq
*this_rq
)
2533 update_cpu_load(this_rq
);
2535 calc_load_account_active(this_rq
);
2541 * sched_exec - execve() is a valuable balancing opportunity, because at
2542 * this point the task has the smallest effective memory and cache footprint.
2544 void sched_exec(void)
2546 struct task_struct
*p
= current
;
2547 unsigned long flags
;
2550 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2551 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2552 if (dest_cpu
== smp_processor_id())
2555 if (likely(cpu_active(dest_cpu
))) {
2556 struct migration_arg arg
= { p
, dest_cpu
};
2558 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2559 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2563 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2568 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2569 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2571 EXPORT_PER_CPU_SYMBOL(kstat
);
2572 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2575 * Return any ns on the sched_clock that have not yet been accounted in
2576 * @p in case that task is currently running.
2578 * Called with task_rq_lock() held on @rq.
2580 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2584 if (task_current(rq
, p
)) {
2585 update_rq_clock(rq
);
2586 ns
= rq
->clock_task
- p
->se
.exec_start
;
2594 unsigned long long task_delta_exec(struct task_struct
*p
)
2596 unsigned long flags
;
2600 rq
= task_rq_lock(p
, &flags
);
2601 ns
= do_task_delta_exec(p
, rq
);
2602 task_rq_unlock(rq
, p
, &flags
);
2608 * Return accounted runtime for the task.
2609 * In case the task is currently running, return the runtime plus current's
2610 * pending runtime that have not been accounted yet.
2612 unsigned long long task_sched_runtime(struct task_struct
*p
)
2614 unsigned long flags
;
2618 rq
= task_rq_lock(p
, &flags
);
2619 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2620 task_rq_unlock(rq
, p
, &flags
);
2625 #ifdef CONFIG_CGROUP_CPUACCT
2626 struct cgroup_subsys cpuacct_subsys
;
2627 struct cpuacct root_cpuacct
;
2630 static inline void task_group_account_field(struct task_struct
*p
, int index
,
2633 #ifdef CONFIG_CGROUP_CPUACCT
2634 struct kernel_cpustat
*kcpustat
;
2638 * Since all updates are sure to touch the root cgroup, we
2639 * get ourselves ahead and touch it first. If the root cgroup
2640 * is the only cgroup, then nothing else should be necessary.
2643 __get_cpu_var(kernel_cpustat
).cpustat
[index
] += tmp
;
2645 #ifdef CONFIG_CGROUP_CPUACCT
2646 if (unlikely(!cpuacct_subsys
.active
))
2651 while (ca
&& (ca
!= &root_cpuacct
)) {
2652 kcpustat
= this_cpu_ptr(ca
->cpustat
);
2653 kcpustat
->cpustat
[index
] += tmp
;
2662 * Account user cpu time to a process.
2663 * @p: the process that the cpu time gets accounted to
2664 * @cputime: the cpu time spent in user space since the last update
2665 * @cputime_scaled: cputime scaled by cpu frequency
2667 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
2668 cputime_t cputime_scaled
)
2672 /* Add user time to process. */
2673 p
->utime
+= cputime
;
2674 p
->utimescaled
+= cputime_scaled
;
2675 account_group_user_time(p
, cputime
);
2677 index
= (TASK_NICE(p
) > 0) ? CPUTIME_NICE
: CPUTIME_USER
;
2679 /* Add user time to cpustat. */
2680 task_group_account_field(p
, index
, (__force u64
) cputime
);
2682 /* Account for user time used */
2683 acct_update_integrals(p
);
2687 * Account guest cpu time to a process.
2688 * @p: the process that the cpu time gets accounted to
2689 * @cputime: the cpu time spent in virtual machine since the last update
2690 * @cputime_scaled: cputime scaled by cpu frequency
2692 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
2693 cputime_t cputime_scaled
)
2695 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2697 /* Add guest time to process. */
2698 p
->utime
+= cputime
;
2699 p
->utimescaled
+= cputime_scaled
;
2700 account_group_user_time(p
, cputime
);
2701 p
->gtime
+= cputime
;
2703 /* Add guest time to cpustat. */
2704 if (TASK_NICE(p
) > 0) {
2705 cpustat
[CPUTIME_NICE
] += (__force u64
) cputime
;
2706 cpustat
[CPUTIME_GUEST_NICE
] += (__force u64
) cputime
;
2708 cpustat
[CPUTIME_USER
] += (__force u64
) cputime
;
2709 cpustat
[CPUTIME_GUEST
] += (__force u64
) cputime
;
2714 * Account system cpu time to a process and desired cpustat field
2715 * @p: the process that the cpu time gets accounted to
2716 * @cputime: the cpu time spent in kernel space since the last update
2717 * @cputime_scaled: cputime scaled by cpu frequency
2718 * @target_cputime64: pointer to cpustat field that has to be updated
2721 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
2722 cputime_t cputime_scaled
, int index
)
2724 /* Add system time to process. */
2725 p
->stime
+= cputime
;
2726 p
->stimescaled
+= cputime_scaled
;
2727 account_group_system_time(p
, cputime
);
2729 /* Add system time to cpustat. */
2730 task_group_account_field(p
, index
, (__force u64
) cputime
);
2732 /* Account for system time used */
2733 acct_update_integrals(p
);
2737 * Account system cpu time to a process.
2738 * @p: the process that the cpu time gets accounted to
2739 * @hardirq_offset: the offset to subtract from hardirq_count()
2740 * @cputime: the cpu time spent in kernel space since the last update
2741 * @cputime_scaled: cputime scaled by cpu frequency
2743 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2744 cputime_t cputime
, cputime_t cputime_scaled
)
2748 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
2749 account_guest_time(p
, cputime
, cputime_scaled
);
2753 if (hardirq_count() - hardirq_offset
)
2754 index
= CPUTIME_IRQ
;
2755 else if (in_serving_softirq())
2756 index
= CPUTIME_SOFTIRQ
;
2758 index
= CPUTIME_SYSTEM
;
2760 __account_system_time(p
, cputime
, cputime_scaled
, index
);
2764 * Account for involuntary wait time.
2765 * @cputime: the cpu time spent in involuntary wait
2767 void account_steal_time(cputime_t cputime
)
2769 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2771 cpustat
[CPUTIME_STEAL
] += (__force u64
) cputime
;
2775 * Account for idle time.
2776 * @cputime: the cpu time spent in idle wait
2778 void account_idle_time(cputime_t cputime
)
2780 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2781 struct rq
*rq
= this_rq();
2783 if (atomic_read(&rq
->nr_iowait
) > 0)
2784 cpustat
[CPUTIME_IOWAIT
] += (__force u64
) cputime
;
2786 cpustat
[CPUTIME_IDLE
] += (__force u64
) cputime
;
2789 static __always_inline
bool steal_account_process_tick(void)
2791 #ifdef CONFIG_PARAVIRT
2792 if (static_key_false(¶virt_steal_enabled
)) {
2795 steal
= paravirt_steal_clock(smp_processor_id());
2796 steal
-= this_rq()->prev_steal_time
;
2798 st
= steal_ticks(steal
);
2799 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
2801 account_steal_time(st
);
2808 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2810 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2812 * Account a tick to a process and cpustat
2813 * @p: the process that the cpu time gets accounted to
2814 * @user_tick: is the tick from userspace
2815 * @rq: the pointer to rq
2817 * Tick demultiplexing follows the order
2818 * - pending hardirq update
2819 * - pending softirq update
2823 * - check for guest_time
2824 * - else account as system_time
2826 * Check for hardirq is done both for system and user time as there is
2827 * no timer going off while we are on hardirq and hence we may never get an
2828 * opportunity to update it solely in system time.
2829 * p->stime and friends are only updated on system time and not on irq
2830 * softirq as those do not count in task exec_runtime any more.
2832 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2835 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2836 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2838 if (steal_account_process_tick())
2841 if (irqtime_account_hi_update()) {
2842 cpustat
[CPUTIME_IRQ
] += (__force u64
) cputime_one_jiffy
;
2843 } else if (irqtime_account_si_update()) {
2844 cpustat
[CPUTIME_SOFTIRQ
] += (__force u64
) cputime_one_jiffy
;
2845 } else if (this_cpu_ksoftirqd() == p
) {
2847 * ksoftirqd time do not get accounted in cpu_softirq_time.
2848 * So, we have to handle it separately here.
2849 * Also, p->stime needs to be updated for ksoftirqd.
2851 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2853 } else if (user_tick
) {
2854 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2855 } else if (p
== rq
->idle
) {
2856 account_idle_time(cputime_one_jiffy
);
2857 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
2858 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2860 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2865 static void irqtime_account_idle_ticks(int ticks
)
2868 struct rq
*rq
= this_rq();
2870 for (i
= 0; i
< ticks
; i
++)
2871 irqtime_account_process_tick(current
, 0, rq
);
2873 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2874 static void irqtime_account_idle_ticks(int ticks
) {}
2875 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2877 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2880 * Account a single tick of cpu time.
2881 * @p: the process that the cpu time gets accounted to
2882 * @user_tick: indicates if the tick is a user or a system tick
2884 void account_process_tick(struct task_struct
*p
, int user_tick
)
2886 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2887 struct rq
*rq
= this_rq();
2889 if (sched_clock_irqtime
) {
2890 irqtime_account_process_tick(p
, user_tick
, rq
);
2894 if (steal_account_process_tick())
2898 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2899 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
2900 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
2903 account_idle_time(cputime_one_jiffy
);
2907 * Account multiple ticks of steal time.
2908 * @p: the process from which the cpu time has been stolen
2909 * @ticks: number of stolen ticks
2911 void account_steal_ticks(unsigned long ticks
)
2913 account_steal_time(jiffies_to_cputime(ticks
));
2917 * Account multiple ticks of idle time.
2918 * @ticks: number of stolen ticks
2920 void account_idle_ticks(unsigned long ticks
)
2923 if (sched_clock_irqtime
) {
2924 irqtime_account_idle_ticks(ticks
);
2928 account_idle_time(jiffies_to_cputime(ticks
));
2934 * Use precise platform statistics if available:
2936 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2937 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2943 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2945 struct task_cputime cputime
;
2947 thread_group_cputime(p
, &cputime
);
2949 *ut
= cputime
.utime
;
2950 *st
= cputime
.stime
;
2954 #ifndef nsecs_to_cputime
2955 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2958 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2960 cputime_t rtime
, utime
= p
->utime
, total
= utime
+ p
->stime
;
2963 * Use CFS's precise accounting:
2965 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
2968 u64 temp
= (__force u64
) rtime
;
2970 temp
*= (__force u64
) utime
;
2971 do_div(temp
, (__force u32
) total
);
2972 utime
= (__force cputime_t
) temp
;
2977 * Compare with previous values, to keep monotonicity:
2979 p
->prev_utime
= max(p
->prev_utime
, utime
);
2980 p
->prev_stime
= max(p
->prev_stime
, rtime
- p
->prev_utime
);
2982 *ut
= p
->prev_utime
;
2983 *st
= p
->prev_stime
;
2987 * Must be called with siglock held.
2989 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2991 struct signal_struct
*sig
= p
->signal
;
2992 struct task_cputime cputime
;
2993 cputime_t rtime
, utime
, total
;
2995 thread_group_cputime(p
, &cputime
);
2997 total
= cputime
.utime
+ cputime
.stime
;
2998 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3001 u64 temp
= (__force u64
) rtime
;
3003 temp
*= (__force u64
) cputime
.utime
;
3004 do_div(temp
, (__force u32
) total
);
3005 utime
= (__force cputime_t
) temp
;
3009 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3010 sig
->prev_stime
= max(sig
->prev_stime
, rtime
- sig
->prev_utime
);
3012 *ut
= sig
->prev_utime
;
3013 *st
= sig
->prev_stime
;
3018 * This function gets called by the timer code, with HZ frequency.
3019 * We call it with interrupts disabled.
3021 void scheduler_tick(void)
3023 int cpu
= smp_processor_id();
3024 struct rq
*rq
= cpu_rq(cpu
);
3025 struct task_struct
*curr
= rq
->curr
;
3029 raw_spin_lock(&rq
->lock
);
3030 update_rq_clock(rq
);
3031 update_cpu_load_active(rq
);
3032 curr
->sched_class
->task_tick(rq
, curr
, 0);
3033 raw_spin_unlock(&rq
->lock
);
3035 perf_event_task_tick();
3038 rq
->idle_balance
= idle_cpu(cpu
);
3039 trigger_load_balance(rq
, cpu
);
3043 notrace
unsigned long get_parent_ip(unsigned long addr
)
3045 if (in_lock_functions(addr
)) {
3046 addr
= CALLER_ADDR2
;
3047 if (in_lock_functions(addr
))
3048 addr
= CALLER_ADDR3
;
3053 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3054 defined(CONFIG_PREEMPT_TRACER))
3056 void __kprobes
add_preempt_count(int val
)
3058 #ifdef CONFIG_DEBUG_PREEMPT
3062 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3065 preempt_count() += val
;
3066 #ifdef CONFIG_DEBUG_PREEMPT
3068 * Spinlock count overflowing soon?
3070 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3073 if (preempt_count() == val
)
3074 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3076 EXPORT_SYMBOL(add_preempt_count
);
3078 void __kprobes
sub_preempt_count(int val
)
3080 #ifdef CONFIG_DEBUG_PREEMPT
3084 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3087 * Is the spinlock portion underflowing?
3089 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3090 !(preempt_count() & PREEMPT_MASK
)))
3094 if (preempt_count() == val
)
3095 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3096 preempt_count() -= val
;
3098 EXPORT_SYMBOL(sub_preempt_count
);
3103 * Print scheduling while atomic bug:
3105 static noinline
void __schedule_bug(struct task_struct
*prev
)
3107 if (oops_in_progress
)
3110 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3111 prev
->comm
, prev
->pid
, preempt_count());
3113 debug_show_held_locks(prev
);
3115 if (irqs_disabled())
3116 print_irqtrace_events(prev
);
3121 * Various schedule()-time debugging checks and statistics:
3123 static inline void schedule_debug(struct task_struct
*prev
)
3126 * Test if we are atomic. Since do_exit() needs to call into
3127 * schedule() atomically, we ignore that path for now.
3128 * Otherwise, whine if we are scheduling when we should not be.
3130 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3131 __schedule_bug(prev
);
3134 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3136 schedstat_inc(this_rq(), sched_count
);
3139 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3141 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
3142 update_rq_clock(rq
);
3143 prev
->sched_class
->put_prev_task(rq
, prev
);
3147 * Pick up the highest-prio task:
3149 static inline struct task_struct
*
3150 pick_next_task(struct rq
*rq
)
3152 const struct sched_class
*class;
3153 struct task_struct
*p
;
3156 * Optimization: we know that if all tasks are in
3157 * the fair class we can call that function directly:
3159 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3160 p
= fair_sched_class
.pick_next_task(rq
);
3165 for_each_class(class) {
3166 p
= class->pick_next_task(rq
);
3171 BUG(); /* the idle class will always have a runnable task */
3175 * __schedule() is the main scheduler function.
3177 static void __sched
__schedule(void)
3179 struct task_struct
*prev
, *next
;
3180 unsigned long *switch_count
;
3186 cpu
= smp_processor_id();
3188 rcu_note_context_switch(cpu
);
3191 schedule_debug(prev
);
3193 if (sched_feat(HRTICK
))
3196 raw_spin_lock_irq(&rq
->lock
);
3198 switch_count
= &prev
->nivcsw
;
3199 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3200 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3201 prev
->state
= TASK_RUNNING
;
3203 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3207 * If a worker went to sleep, notify and ask workqueue
3208 * whether it wants to wake up a task to maintain
3211 if (prev
->flags
& PF_WQ_WORKER
) {
3212 struct task_struct
*to_wakeup
;
3214 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3216 try_to_wake_up_local(to_wakeup
);
3219 switch_count
= &prev
->nvcsw
;
3222 pre_schedule(rq
, prev
);
3224 if (unlikely(!rq
->nr_running
))
3225 idle_balance(cpu
, rq
);
3227 put_prev_task(rq
, prev
);
3228 next
= pick_next_task(rq
);
3229 clear_tsk_need_resched(prev
);
3230 rq
->skip_clock_update
= 0;
3232 if (likely(prev
!= next
)) {
3237 context_switch(rq
, prev
, next
); /* unlocks the rq */
3239 * The context switch have flipped the stack from under us
3240 * and restored the local variables which were saved when
3241 * this task called schedule() in the past. prev == current
3242 * is still correct, but it can be moved to another cpu/rq.
3244 cpu
= smp_processor_id();
3247 raw_spin_unlock_irq(&rq
->lock
);
3251 sched_preempt_enable_no_resched();
3256 static inline void sched_submit_work(struct task_struct
*tsk
)
3258 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3261 * If we are going to sleep and we have plugged IO queued,
3262 * make sure to submit it to avoid deadlocks.
3264 if (blk_needs_flush_plug(tsk
))
3265 blk_schedule_flush_plug(tsk
);
3268 asmlinkage
void __sched
schedule(void)
3270 struct task_struct
*tsk
= current
;
3272 sched_submit_work(tsk
);
3275 EXPORT_SYMBOL(schedule
);
3278 * schedule_preempt_disabled - called with preemption disabled
3280 * Returns with preemption disabled. Note: preempt_count must be 1
3282 void __sched
schedule_preempt_disabled(void)
3284 sched_preempt_enable_no_resched();
3289 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3291 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3293 if (lock
->owner
!= owner
)
3297 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3298 * lock->owner still matches owner, if that fails, owner might
3299 * point to free()d memory, if it still matches, the rcu_read_lock()
3300 * ensures the memory stays valid.
3304 return owner
->on_cpu
;
3308 * Look out! "owner" is an entirely speculative pointer
3309 * access and not reliable.
3311 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3313 if (!sched_feat(OWNER_SPIN
))
3317 while (owner_running(lock
, owner
)) {
3321 arch_mutex_cpu_relax();
3326 * We break out the loop above on need_resched() and when the
3327 * owner changed, which is a sign for heavy contention. Return
3328 * success only when lock->owner is NULL.
3330 return lock
->owner
== NULL
;
3334 #ifdef CONFIG_PREEMPT
3336 * this is the entry point to schedule() from in-kernel preemption
3337 * off of preempt_enable. Kernel preemptions off return from interrupt
3338 * occur there and call schedule directly.
3340 asmlinkage
void __sched notrace
preempt_schedule(void)
3342 struct thread_info
*ti
= current_thread_info();
3345 * If there is a non-zero preempt_count or interrupts are disabled,
3346 * we do not want to preempt the current task. Just return..
3348 if (likely(ti
->preempt_count
|| irqs_disabled()))
3352 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3354 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3357 * Check again in case we missed a preemption opportunity
3358 * between schedule and now.
3361 } while (need_resched());
3363 EXPORT_SYMBOL(preempt_schedule
);
3366 * this is the entry point to schedule() from kernel preemption
3367 * off of irq context.
3368 * Note, that this is called and return with irqs disabled. This will
3369 * protect us against recursive calling from irq.
3371 asmlinkage
void __sched
preempt_schedule_irq(void)
3373 struct thread_info
*ti
= current_thread_info();
3375 /* Catch callers which need to be fixed */
3376 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3379 add_preempt_count(PREEMPT_ACTIVE
);
3382 local_irq_disable();
3383 sub_preempt_count(PREEMPT_ACTIVE
);
3386 * Check again in case we missed a preemption opportunity
3387 * between schedule and now.
3390 } while (need_resched());
3393 #endif /* CONFIG_PREEMPT */
3395 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3398 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3400 EXPORT_SYMBOL(default_wake_function
);
3403 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3404 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3405 * number) then we wake all the non-exclusive tasks and one exclusive task.
3407 * There are circumstances in which we can try to wake a task which has already
3408 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3409 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3411 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3412 int nr_exclusive
, int wake_flags
, void *key
)
3414 wait_queue_t
*curr
, *next
;
3416 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3417 unsigned flags
= curr
->flags
;
3419 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3420 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3426 * __wake_up - wake up threads blocked on a waitqueue.
3428 * @mode: which threads
3429 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3430 * @key: is directly passed to the wakeup function
3432 * It may be assumed that this function implies a write memory barrier before
3433 * changing the task state if and only if any tasks are woken up.
3435 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3436 int nr_exclusive
, void *key
)
3438 unsigned long flags
;
3440 spin_lock_irqsave(&q
->lock
, flags
);
3441 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3442 spin_unlock_irqrestore(&q
->lock
, flags
);
3444 EXPORT_SYMBOL(__wake_up
);
3447 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3449 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3451 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3453 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3455 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3457 __wake_up_common(q
, mode
, 1, 0, key
);
3459 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3462 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3464 * @mode: which threads
3465 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3466 * @key: opaque value to be passed to wakeup targets
3468 * The sync wakeup differs that the waker knows that it will schedule
3469 * away soon, so while the target thread will be woken up, it will not
3470 * be migrated to another CPU - ie. the two threads are 'synchronized'
3471 * with each other. This can prevent needless bouncing between CPUs.
3473 * On UP it can prevent extra preemption.
3475 * It may be assumed that this function implies a write memory barrier before
3476 * changing the task state if and only if any tasks are woken up.
3478 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3479 int nr_exclusive
, void *key
)
3481 unsigned long flags
;
3482 int wake_flags
= WF_SYNC
;
3487 if (unlikely(!nr_exclusive
))
3490 spin_lock_irqsave(&q
->lock
, flags
);
3491 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3492 spin_unlock_irqrestore(&q
->lock
, flags
);
3494 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3497 * __wake_up_sync - see __wake_up_sync_key()
3499 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3501 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3503 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3506 * complete: - signals a single thread waiting on this completion
3507 * @x: holds the state of this particular completion
3509 * This will wake up a single thread waiting on this completion. Threads will be
3510 * awakened in the same order in which they were queued.
3512 * See also complete_all(), wait_for_completion() and related routines.
3514 * It may be assumed that this function implies a write memory barrier before
3515 * changing the task state if and only if any tasks are woken up.
3517 void complete(struct completion
*x
)
3519 unsigned long flags
;
3521 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3523 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3524 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3526 EXPORT_SYMBOL(complete
);
3529 * complete_all: - signals all threads waiting on this completion
3530 * @x: holds the state of this particular completion
3532 * This will wake up all threads waiting on this particular completion event.
3534 * It may be assumed that this function implies a write memory barrier before
3535 * changing the task state if and only if any tasks are woken up.
3537 void complete_all(struct completion
*x
)
3539 unsigned long flags
;
3541 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3542 x
->done
+= UINT_MAX
/2;
3543 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3544 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3546 EXPORT_SYMBOL(complete_all
);
3548 static inline long __sched
3549 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3552 DECLARE_WAITQUEUE(wait
, current
);
3554 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3556 if (signal_pending_state(state
, current
)) {
3557 timeout
= -ERESTARTSYS
;
3560 __set_current_state(state
);
3561 spin_unlock_irq(&x
->wait
.lock
);
3562 timeout
= schedule_timeout(timeout
);
3563 spin_lock_irq(&x
->wait
.lock
);
3564 } while (!x
->done
&& timeout
);
3565 __remove_wait_queue(&x
->wait
, &wait
);
3570 return timeout
?: 1;
3574 wait_for_common(struct completion
*x
, long timeout
, int state
)
3578 spin_lock_irq(&x
->wait
.lock
);
3579 timeout
= do_wait_for_common(x
, timeout
, state
);
3580 spin_unlock_irq(&x
->wait
.lock
);
3585 * wait_for_completion: - waits for completion of a task
3586 * @x: holds the state of this particular completion
3588 * This waits to be signaled for completion of a specific task. It is NOT
3589 * interruptible and there is no timeout.
3591 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3592 * and interrupt capability. Also see complete().
3594 void __sched
wait_for_completion(struct completion
*x
)
3596 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3598 EXPORT_SYMBOL(wait_for_completion
);
3601 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3602 * @x: holds the state of this particular completion
3603 * @timeout: timeout value in jiffies
3605 * This waits for either a completion of a specific task to be signaled or for a
3606 * specified timeout to expire. The timeout is in jiffies. It is not
3609 * The return value is 0 if timed out, and positive (at least 1, or number of
3610 * jiffies left till timeout) if completed.
3612 unsigned long __sched
3613 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3615 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3617 EXPORT_SYMBOL(wait_for_completion_timeout
);
3620 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3621 * @x: holds the state of this particular completion
3623 * This waits for completion of a specific task to be signaled. It is
3626 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3628 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3630 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3631 if (t
== -ERESTARTSYS
)
3635 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3638 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3639 * @x: holds the state of this particular completion
3640 * @timeout: timeout value in jiffies
3642 * This waits for either a completion of a specific task to be signaled or for a
3643 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3645 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3646 * positive (at least 1, or number of jiffies left till timeout) if completed.
3649 wait_for_completion_interruptible_timeout(struct completion
*x
,
3650 unsigned long timeout
)
3652 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3654 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3657 * wait_for_completion_killable: - waits for completion of a task (killable)
3658 * @x: holds the state of this particular completion
3660 * This waits to be signaled for completion of a specific task. It can be
3661 * interrupted by a kill signal.
3663 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3665 int __sched
wait_for_completion_killable(struct completion
*x
)
3667 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3668 if (t
== -ERESTARTSYS
)
3672 EXPORT_SYMBOL(wait_for_completion_killable
);
3675 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3676 * @x: holds the state of this particular completion
3677 * @timeout: timeout value in jiffies
3679 * This waits for either a completion of a specific task to be
3680 * signaled or for a specified timeout to expire. It can be
3681 * interrupted by a kill signal. The timeout is in jiffies.
3683 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3684 * positive (at least 1, or number of jiffies left till timeout) if completed.
3687 wait_for_completion_killable_timeout(struct completion
*x
,
3688 unsigned long timeout
)
3690 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3692 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3695 * try_wait_for_completion - try to decrement a completion without blocking
3696 * @x: completion structure
3698 * Returns: 0 if a decrement cannot be done without blocking
3699 * 1 if a decrement succeeded.
3701 * If a completion is being used as a counting completion,
3702 * attempt to decrement the counter without blocking. This
3703 * enables us to avoid waiting if the resource the completion
3704 * is protecting is not available.
3706 bool try_wait_for_completion(struct completion
*x
)
3708 unsigned long flags
;
3711 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3716 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3719 EXPORT_SYMBOL(try_wait_for_completion
);
3722 * completion_done - Test to see if a completion has any waiters
3723 * @x: completion structure
3725 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3726 * 1 if there are no waiters.
3729 bool completion_done(struct completion
*x
)
3731 unsigned long flags
;
3734 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3737 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3740 EXPORT_SYMBOL(completion_done
);
3743 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3745 unsigned long flags
;
3748 init_waitqueue_entry(&wait
, current
);
3750 __set_current_state(state
);
3752 spin_lock_irqsave(&q
->lock
, flags
);
3753 __add_wait_queue(q
, &wait
);
3754 spin_unlock(&q
->lock
);
3755 timeout
= schedule_timeout(timeout
);
3756 spin_lock_irq(&q
->lock
);
3757 __remove_wait_queue(q
, &wait
);
3758 spin_unlock_irqrestore(&q
->lock
, flags
);
3763 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3765 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3767 EXPORT_SYMBOL(interruptible_sleep_on
);
3770 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3772 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3774 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3776 void __sched
sleep_on(wait_queue_head_t
*q
)
3778 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3780 EXPORT_SYMBOL(sleep_on
);
3782 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3784 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3786 EXPORT_SYMBOL(sleep_on_timeout
);
3788 #ifdef CONFIG_RT_MUTEXES
3791 * rt_mutex_setprio - set the current priority of a task
3793 * @prio: prio value (kernel-internal form)
3795 * This function changes the 'effective' priority of a task. It does
3796 * not touch ->normal_prio like __setscheduler().
3798 * Used by the rt_mutex code to implement priority inheritance logic.
3800 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3802 int oldprio
, on_rq
, running
;
3804 const struct sched_class
*prev_class
;
3806 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3808 rq
= __task_rq_lock(p
);
3811 * Idle task boosting is a nono in general. There is one
3812 * exception, when PREEMPT_RT and NOHZ is active:
3814 * The idle task calls get_next_timer_interrupt() and holds
3815 * the timer wheel base->lock on the CPU and another CPU wants
3816 * to access the timer (probably to cancel it). We can safely
3817 * ignore the boosting request, as the idle CPU runs this code
3818 * with interrupts disabled and will complete the lock
3819 * protected section without being interrupted. So there is no
3820 * real need to boost.
3822 if (unlikely(p
== rq
->idle
)) {
3823 WARN_ON(p
!= rq
->curr
);
3824 WARN_ON(p
->pi_blocked_on
);
3828 trace_sched_pi_setprio(p
, prio
);
3830 prev_class
= p
->sched_class
;
3832 running
= task_current(rq
, p
);
3834 dequeue_task(rq
, p
, 0);
3836 p
->sched_class
->put_prev_task(rq
, p
);
3839 p
->sched_class
= &rt_sched_class
;
3841 p
->sched_class
= &fair_sched_class
;
3846 p
->sched_class
->set_curr_task(rq
);
3848 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3850 check_class_changed(rq
, p
, prev_class
, oldprio
);
3852 __task_rq_unlock(rq
);
3855 void set_user_nice(struct task_struct
*p
, long nice
)
3857 int old_prio
, delta
, on_rq
;
3858 unsigned long flags
;
3861 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3864 * We have to be careful, if called from sys_setpriority(),
3865 * the task might be in the middle of scheduling on another CPU.
3867 rq
= task_rq_lock(p
, &flags
);
3869 * The RT priorities are set via sched_setscheduler(), but we still
3870 * allow the 'normal' nice value to be set - but as expected
3871 * it wont have any effect on scheduling until the task is
3872 * SCHED_FIFO/SCHED_RR:
3874 if (task_has_rt_policy(p
)) {
3875 p
->static_prio
= NICE_TO_PRIO(nice
);
3880 dequeue_task(rq
, p
, 0);
3882 p
->static_prio
= NICE_TO_PRIO(nice
);
3885 p
->prio
= effective_prio(p
);
3886 delta
= p
->prio
- old_prio
;
3889 enqueue_task(rq
, p
, 0);
3891 * If the task increased its priority or is running and
3892 * lowered its priority, then reschedule its CPU:
3894 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3895 resched_task(rq
->curr
);
3898 task_rq_unlock(rq
, p
, &flags
);
3900 EXPORT_SYMBOL(set_user_nice
);
3903 * can_nice - check if a task can reduce its nice value
3907 int can_nice(const struct task_struct
*p
, const int nice
)
3909 /* convert nice value [19,-20] to rlimit style value [1,40] */
3910 int nice_rlim
= 20 - nice
;
3912 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3913 capable(CAP_SYS_NICE
));
3916 #ifdef __ARCH_WANT_SYS_NICE
3919 * sys_nice - change the priority of the current process.
3920 * @increment: priority increment
3922 * sys_setpriority is a more generic, but much slower function that
3923 * does similar things.
3925 SYSCALL_DEFINE1(nice
, int, increment
)
3930 * Setpriority might change our priority at the same moment.
3931 * We don't have to worry. Conceptually one call occurs first
3932 * and we have a single winner.
3934 if (increment
< -40)
3939 nice
= TASK_NICE(current
) + increment
;
3945 if (increment
< 0 && !can_nice(current
, nice
))
3948 retval
= security_task_setnice(current
, nice
);
3952 set_user_nice(current
, nice
);
3959 * task_prio - return the priority value of a given task.
3960 * @p: the task in question.
3962 * This is the priority value as seen by users in /proc.
3963 * RT tasks are offset by -200. Normal tasks are centered
3964 * around 0, value goes from -16 to +15.
3966 int task_prio(const struct task_struct
*p
)
3968 return p
->prio
- MAX_RT_PRIO
;
3972 * task_nice - return the nice value of a given task.
3973 * @p: the task in question.
3975 int task_nice(const struct task_struct
*p
)
3977 return TASK_NICE(p
);
3979 EXPORT_SYMBOL(task_nice
);
3982 * idle_cpu - is a given cpu idle currently?
3983 * @cpu: the processor in question.
3985 int idle_cpu(int cpu
)
3987 struct rq
*rq
= cpu_rq(cpu
);
3989 if (rq
->curr
!= rq
->idle
)
3996 if (!llist_empty(&rq
->wake_list
))
4004 * idle_task - return the idle task for a given cpu.
4005 * @cpu: the processor in question.
4007 struct task_struct
*idle_task(int cpu
)
4009 return cpu_rq(cpu
)->idle
;
4013 * find_process_by_pid - find a process with a matching PID value.
4014 * @pid: the pid in question.
4016 static struct task_struct
*find_process_by_pid(pid_t pid
)
4018 return pid
? find_task_by_vpid(pid
) : current
;
4021 /* Actually do priority change: must hold rq lock. */
4023 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4026 p
->rt_priority
= prio
;
4027 p
->normal_prio
= normal_prio(p
);
4028 /* we are holding p->pi_lock already */
4029 p
->prio
= rt_mutex_getprio(p
);
4030 if (rt_prio(p
->prio
))
4031 p
->sched_class
= &rt_sched_class
;
4033 p
->sched_class
= &fair_sched_class
;
4038 * check the target process has a UID that matches the current process's
4040 static bool check_same_owner(struct task_struct
*p
)
4042 const struct cred
*cred
= current_cred(), *pcred
;
4046 pcred
= __task_cred(p
);
4047 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
4048 match
= (cred
->euid
== pcred
->euid
||
4049 cred
->euid
== pcred
->uid
);
4056 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4057 const struct sched_param
*param
, bool user
)
4059 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4060 unsigned long flags
;
4061 const struct sched_class
*prev_class
;
4065 /* may grab non-irq protected spin_locks */
4066 BUG_ON(in_interrupt());
4068 /* double check policy once rq lock held */
4070 reset_on_fork
= p
->sched_reset_on_fork
;
4071 policy
= oldpolicy
= p
->policy
;
4073 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4074 policy
&= ~SCHED_RESET_ON_FORK
;
4076 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4077 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4078 policy
!= SCHED_IDLE
)
4083 * Valid priorities for SCHED_FIFO and SCHED_RR are
4084 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4085 * SCHED_BATCH and SCHED_IDLE is 0.
4087 if (param
->sched_priority
< 0 ||
4088 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4089 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4091 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4095 * Allow unprivileged RT tasks to decrease priority:
4097 if (user
&& !capable(CAP_SYS_NICE
)) {
4098 if (rt_policy(policy
)) {
4099 unsigned long rlim_rtprio
=
4100 task_rlimit(p
, RLIMIT_RTPRIO
);
4102 /* can't set/change the rt policy */
4103 if (policy
!= p
->policy
&& !rlim_rtprio
)
4106 /* can't increase priority */
4107 if (param
->sched_priority
> p
->rt_priority
&&
4108 param
->sched_priority
> rlim_rtprio
)
4113 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4114 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4116 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
4117 if (!can_nice(p
, TASK_NICE(p
)))
4121 /* can't change other user's priorities */
4122 if (!check_same_owner(p
))
4125 /* Normal users shall not reset the sched_reset_on_fork flag */
4126 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4131 retval
= security_task_setscheduler(p
);
4137 * make sure no PI-waiters arrive (or leave) while we are
4138 * changing the priority of the task:
4140 * To be able to change p->policy safely, the appropriate
4141 * runqueue lock must be held.
4143 rq
= task_rq_lock(p
, &flags
);
4146 * Changing the policy of the stop threads its a very bad idea
4148 if (p
== rq
->stop
) {
4149 task_rq_unlock(rq
, p
, &flags
);
4154 * If not changing anything there's no need to proceed further:
4156 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
4157 param
->sched_priority
== p
->rt_priority
))) {
4159 __task_rq_unlock(rq
);
4160 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4164 #ifdef CONFIG_RT_GROUP_SCHED
4167 * Do not allow realtime tasks into groups that have no runtime
4170 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4171 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4172 !task_group_is_autogroup(task_group(p
))) {
4173 task_rq_unlock(rq
, p
, &flags
);
4179 /* recheck policy now with rq lock held */
4180 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4181 policy
= oldpolicy
= -1;
4182 task_rq_unlock(rq
, p
, &flags
);
4186 running
= task_current(rq
, p
);
4188 dequeue_task(rq
, p
, 0);
4190 p
->sched_class
->put_prev_task(rq
, p
);
4192 p
->sched_reset_on_fork
= reset_on_fork
;
4195 prev_class
= p
->sched_class
;
4196 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4199 p
->sched_class
->set_curr_task(rq
);
4201 enqueue_task(rq
, p
, 0);
4203 check_class_changed(rq
, p
, prev_class
, oldprio
);
4204 task_rq_unlock(rq
, p
, &flags
);
4206 rt_mutex_adjust_pi(p
);
4212 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4213 * @p: the task in question.
4214 * @policy: new policy.
4215 * @param: structure containing the new RT priority.
4217 * NOTE that the task may be already dead.
4219 int sched_setscheduler(struct task_struct
*p
, int policy
,
4220 const struct sched_param
*param
)
4222 return __sched_setscheduler(p
, policy
, param
, true);
4224 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4227 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4228 * @p: the task in question.
4229 * @policy: new policy.
4230 * @param: structure containing the new RT priority.
4232 * Just like sched_setscheduler, only don't bother checking if the
4233 * current context has permission. For example, this is needed in
4234 * stop_machine(): we create temporary high priority worker threads,
4235 * but our caller might not have that capability.
4237 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4238 const struct sched_param
*param
)
4240 return __sched_setscheduler(p
, policy
, param
, false);
4244 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4246 struct sched_param lparam
;
4247 struct task_struct
*p
;
4250 if (!param
|| pid
< 0)
4252 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4257 p
= find_process_by_pid(pid
);
4259 retval
= sched_setscheduler(p
, policy
, &lparam
);
4266 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4267 * @pid: the pid in question.
4268 * @policy: new policy.
4269 * @param: structure containing the new RT priority.
4271 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4272 struct sched_param __user
*, param
)
4274 /* negative values for policy are not valid */
4278 return do_sched_setscheduler(pid
, policy
, param
);
4282 * sys_sched_setparam - set/change the RT priority of a thread
4283 * @pid: the pid in question.
4284 * @param: structure containing the new RT priority.
4286 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4288 return do_sched_setscheduler(pid
, -1, param
);
4292 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4293 * @pid: the pid in question.
4295 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4297 struct task_struct
*p
;
4305 p
= find_process_by_pid(pid
);
4307 retval
= security_task_getscheduler(p
);
4310 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4317 * sys_sched_getparam - get the RT priority of a thread
4318 * @pid: the pid in question.
4319 * @param: structure containing the RT priority.
4321 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4323 struct sched_param lp
;
4324 struct task_struct
*p
;
4327 if (!param
|| pid
< 0)
4331 p
= find_process_by_pid(pid
);
4336 retval
= security_task_getscheduler(p
);
4340 lp
.sched_priority
= p
->rt_priority
;
4344 * This one might sleep, we cannot do it with a spinlock held ...
4346 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4355 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4357 cpumask_var_t cpus_allowed
, new_mask
;
4358 struct task_struct
*p
;
4364 p
= find_process_by_pid(pid
);
4371 /* Prevent p going away */
4375 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4379 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4381 goto out_free_cpus_allowed
;
4384 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4387 retval
= security_task_setscheduler(p
);
4391 cpuset_cpus_allowed(p
, cpus_allowed
);
4392 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4394 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4397 cpuset_cpus_allowed(p
, cpus_allowed
);
4398 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4400 * We must have raced with a concurrent cpuset
4401 * update. Just reset the cpus_allowed to the
4402 * cpuset's cpus_allowed
4404 cpumask_copy(new_mask
, cpus_allowed
);
4409 free_cpumask_var(new_mask
);
4410 out_free_cpus_allowed
:
4411 free_cpumask_var(cpus_allowed
);
4418 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4419 struct cpumask
*new_mask
)
4421 if (len
< cpumask_size())
4422 cpumask_clear(new_mask
);
4423 else if (len
> cpumask_size())
4424 len
= cpumask_size();
4426 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4430 * sys_sched_setaffinity - set the cpu affinity of a process
4431 * @pid: pid of the process
4432 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4433 * @user_mask_ptr: user-space pointer to the new cpu mask
4435 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4436 unsigned long __user
*, user_mask_ptr
)
4438 cpumask_var_t new_mask
;
4441 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4444 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4446 retval
= sched_setaffinity(pid
, new_mask
);
4447 free_cpumask_var(new_mask
);
4451 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4453 struct task_struct
*p
;
4454 unsigned long flags
;
4461 p
= find_process_by_pid(pid
);
4465 retval
= security_task_getscheduler(p
);
4469 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4470 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4471 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4481 * sys_sched_getaffinity - get the cpu affinity of a process
4482 * @pid: pid of the process
4483 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4484 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4486 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4487 unsigned long __user
*, user_mask_ptr
)
4492 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4494 if (len
& (sizeof(unsigned long)-1))
4497 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4500 ret
= sched_getaffinity(pid
, mask
);
4502 size_t retlen
= min_t(size_t, len
, cpumask_size());
4504 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4509 free_cpumask_var(mask
);
4515 * sys_sched_yield - yield the current processor to other threads.
4517 * This function yields the current CPU to other tasks. If there are no
4518 * other threads running on this CPU then this function will return.
4520 SYSCALL_DEFINE0(sched_yield
)
4522 struct rq
*rq
= this_rq_lock();
4524 schedstat_inc(rq
, yld_count
);
4525 current
->sched_class
->yield_task(rq
);
4528 * Since we are going to call schedule() anyway, there's
4529 * no need to preempt or enable interrupts:
4531 __release(rq
->lock
);
4532 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4533 do_raw_spin_unlock(&rq
->lock
);
4534 sched_preempt_enable_no_resched();
4541 static inline int should_resched(void)
4543 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4546 static void __cond_resched(void)
4548 add_preempt_count(PREEMPT_ACTIVE
);
4550 sub_preempt_count(PREEMPT_ACTIVE
);
4553 int __sched
_cond_resched(void)
4555 if (should_resched()) {
4561 EXPORT_SYMBOL(_cond_resched
);
4564 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4565 * call schedule, and on return reacquire the lock.
4567 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4568 * operations here to prevent schedule() from being called twice (once via
4569 * spin_unlock(), once by hand).
4571 int __cond_resched_lock(spinlock_t
*lock
)
4573 int resched
= should_resched();
4576 lockdep_assert_held(lock
);
4578 if (spin_needbreak(lock
) || resched
) {
4589 EXPORT_SYMBOL(__cond_resched_lock
);
4591 int __sched
__cond_resched_softirq(void)
4593 BUG_ON(!in_softirq());
4595 if (should_resched()) {
4603 EXPORT_SYMBOL(__cond_resched_softirq
);
4606 * yield - yield the current processor to other threads.
4608 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4610 * The scheduler is at all times free to pick the calling task as the most
4611 * eligible task to run, if removing the yield() call from your code breaks
4612 * it, its already broken.
4614 * Typical broken usage is:
4619 * where one assumes that yield() will let 'the other' process run that will
4620 * make event true. If the current task is a SCHED_FIFO task that will never
4621 * happen. Never use yield() as a progress guarantee!!
4623 * If you want to use yield() to wait for something, use wait_event().
4624 * If you want to use yield() to be 'nice' for others, use cond_resched().
4625 * If you still want to use yield(), do not!
4627 void __sched
yield(void)
4629 set_current_state(TASK_RUNNING
);
4632 EXPORT_SYMBOL(yield
);
4635 * yield_to - yield the current processor to another thread in
4636 * your thread group, or accelerate that thread toward the
4637 * processor it's on.
4639 * @preempt: whether task preemption is allowed or not
4641 * It's the caller's job to ensure that the target task struct
4642 * can't go away on us before we can do any checks.
4644 * Returns true if we indeed boosted the target task.
4646 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4648 struct task_struct
*curr
= current
;
4649 struct rq
*rq
, *p_rq
;
4650 unsigned long flags
;
4653 local_irq_save(flags
);
4658 double_rq_lock(rq
, p_rq
);
4659 while (task_rq(p
) != p_rq
) {
4660 double_rq_unlock(rq
, p_rq
);
4664 if (!curr
->sched_class
->yield_to_task
)
4667 if (curr
->sched_class
!= p
->sched_class
)
4670 if (task_running(p_rq
, p
) || p
->state
)
4673 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4675 schedstat_inc(rq
, yld_count
);
4677 * Make p's CPU reschedule; pick_next_entity takes care of
4680 if (preempt
&& rq
!= p_rq
)
4681 resched_task(p_rq
->curr
);
4684 * We might have set it in task_yield_fair(), but are
4685 * not going to schedule(), so don't want to skip
4688 rq
->skip_clock_update
= 0;
4692 double_rq_unlock(rq
, p_rq
);
4693 local_irq_restore(flags
);
4700 EXPORT_SYMBOL_GPL(yield_to
);
4703 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4704 * that process accounting knows that this is a task in IO wait state.
4706 void __sched
io_schedule(void)
4708 struct rq
*rq
= raw_rq();
4710 delayacct_blkio_start();
4711 atomic_inc(&rq
->nr_iowait
);
4712 blk_flush_plug(current
);
4713 current
->in_iowait
= 1;
4715 current
->in_iowait
= 0;
4716 atomic_dec(&rq
->nr_iowait
);
4717 delayacct_blkio_end();
4719 EXPORT_SYMBOL(io_schedule
);
4721 long __sched
io_schedule_timeout(long timeout
)
4723 struct rq
*rq
= raw_rq();
4726 delayacct_blkio_start();
4727 atomic_inc(&rq
->nr_iowait
);
4728 blk_flush_plug(current
);
4729 current
->in_iowait
= 1;
4730 ret
= schedule_timeout(timeout
);
4731 current
->in_iowait
= 0;
4732 atomic_dec(&rq
->nr_iowait
);
4733 delayacct_blkio_end();
4738 * sys_sched_get_priority_max - return maximum RT priority.
4739 * @policy: scheduling class.
4741 * this syscall returns the maximum rt_priority that can be used
4742 * by a given scheduling class.
4744 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4751 ret
= MAX_USER_RT_PRIO
-1;
4763 * sys_sched_get_priority_min - return minimum RT priority.
4764 * @policy: scheduling class.
4766 * this syscall returns the minimum rt_priority that can be used
4767 * by a given scheduling class.
4769 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4787 * sys_sched_rr_get_interval - return the default timeslice of a process.
4788 * @pid: pid of the process.
4789 * @interval: userspace pointer to the timeslice value.
4791 * this syscall writes the default timeslice value of a given process
4792 * into the user-space timespec buffer. A value of '0' means infinity.
4794 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4795 struct timespec __user
*, interval
)
4797 struct task_struct
*p
;
4798 unsigned int time_slice
;
4799 unsigned long flags
;
4809 p
= find_process_by_pid(pid
);
4813 retval
= security_task_getscheduler(p
);
4817 rq
= task_rq_lock(p
, &flags
);
4818 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4819 task_rq_unlock(rq
, p
, &flags
);
4822 jiffies_to_timespec(time_slice
, &t
);
4823 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4831 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4833 void sched_show_task(struct task_struct
*p
)
4835 unsigned long free
= 0;
4838 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4839 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4840 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4841 #if BITS_PER_LONG == 32
4842 if (state
== TASK_RUNNING
)
4843 printk(KERN_CONT
" running ");
4845 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4847 if (state
== TASK_RUNNING
)
4848 printk(KERN_CONT
" running task ");
4850 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4852 #ifdef CONFIG_DEBUG_STACK_USAGE
4853 free
= stack_not_used(p
);
4855 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4856 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4857 (unsigned long)task_thread_info(p
)->flags
);
4859 show_stack(p
, NULL
);
4862 void show_state_filter(unsigned long state_filter
)
4864 struct task_struct
*g
, *p
;
4866 #if BITS_PER_LONG == 32
4868 " task PC stack pid father\n");
4871 " task PC stack pid father\n");
4874 do_each_thread(g
, p
) {
4876 * reset the NMI-timeout, listing all files on a slow
4877 * console might take a lot of time:
4879 touch_nmi_watchdog();
4880 if (!state_filter
|| (p
->state
& state_filter
))
4882 } while_each_thread(g
, p
);
4884 touch_all_softlockup_watchdogs();
4886 #ifdef CONFIG_SCHED_DEBUG
4887 sysrq_sched_debug_show();
4891 * Only show locks if all tasks are dumped:
4894 debug_show_all_locks();
4897 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4899 idle
->sched_class
= &idle_sched_class
;
4903 * init_idle - set up an idle thread for a given CPU
4904 * @idle: task in question
4905 * @cpu: cpu the idle task belongs to
4907 * NOTE: this function does not set the idle thread's NEED_RESCHED
4908 * flag, to make booting more robust.
4910 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4912 struct rq
*rq
= cpu_rq(cpu
);
4913 unsigned long flags
;
4915 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4918 idle
->state
= TASK_RUNNING
;
4919 idle
->se
.exec_start
= sched_clock();
4921 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4923 * We're having a chicken and egg problem, even though we are
4924 * holding rq->lock, the cpu isn't yet set to this cpu so the
4925 * lockdep check in task_group() will fail.
4927 * Similar case to sched_fork(). / Alternatively we could
4928 * use task_rq_lock() here and obtain the other rq->lock.
4933 __set_task_cpu(idle
, cpu
);
4936 rq
->curr
= rq
->idle
= idle
;
4937 #if defined(CONFIG_SMP)
4940 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4942 /* Set the preempt count _outside_ the spinlocks! */
4943 task_thread_info(idle
)->preempt_count
= 0;
4946 * The idle tasks have their own, simple scheduling class:
4948 idle
->sched_class
= &idle_sched_class
;
4949 ftrace_graph_init_idle_task(idle
, cpu
);
4950 #if defined(CONFIG_SMP)
4951 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4956 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4958 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4959 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4961 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4962 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
4966 * This is how migration works:
4968 * 1) we invoke migration_cpu_stop() on the target CPU using
4970 * 2) stopper starts to run (implicitly forcing the migrated thread
4972 * 3) it checks whether the migrated task is still in the wrong runqueue.
4973 * 4) if it's in the wrong runqueue then the migration thread removes
4974 * it and puts it into the right queue.
4975 * 5) stopper completes and stop_one_cpu() returns and the migration
4980 * Change a given task's CPU affinity. Migrate the thread to a
4981 * proper CPU and schedule it away if the CPU it's executing on
4982 * is removed from the allowed bitmask.
4984 * NOTE: the caller must have a valid reference to the task, the
4985 * task must not exit() & deallocate itself prematurely. The
4986 * call is not atomic; no spinlocks may be held.
4988 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4990 unsigned long flags
;
4992 unsigned int dest_cpu
;
4995 rq
= task_rq_lock(p
, &flags
);
4997 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
5000 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5005 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
5010 do_set_cpus_allowed(p
, new_mask
);
5012 /* Can the task run on the task's current CPU? If so, we're done */
5013 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5016 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5018 struct migration_arg arg
= { p
, dest_cpu
};
5019 /* Need help from migration thread: drop lock and wait. */
5020 task_rq_unlock(rq
, p
, &flags
);
5021 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5022 tlb_migrate_finish(p
->mm
);
5026 task_rq_unlock(rq
, p
, &flags
);
5030 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5033 * Move (not current) task off this cpu, onto dest cpu. We're doing
5034 * this because either it can't run here any more (set_cpus_allowed()
5035 * away from this CPU, or CPU going down), or because we're
5036 * attempting to rebalance this task on exec (sched_exec).
5038 * So we race with normal scheduler movements, but that's OK, as long
5039 * as the task is no longer on this CPU.
5041 * Returns non-zero if task was successfully migrated.
5043 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5045 struct rq
*rq_dest
, *rq_src
;
5048 if (unlikely(!cpu_active(dest_cpu
)))
5051 rq_src
= cpu_rq(src_cpu
);
5052 rq_dest
= cpu_rq(dest_cpu
);
5054 raw_spin_lock(&p
->pi_lock
);
5055 double_rq_lock(rq_src
, rq_dest
);
5056 /* Already moved. */
5057 if (task_cpu(p
) != src_cpu
)
5059 /* Affinity changed (again). */
5060 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
5064 * If we're not on a rq, the next wake-up will ensure we're
5068 dequeue_task(rq_src
, p
, 0);
5069 set_task_cpu(p
, dest_cpu
);
5070 enqueue_task(rq_dest
, p
, 0);
5071 check_preempt_curr(rq_dest
, p
, 0);
5076 double_rq_unlock(rq_src
, rq_dest
);
5077 raw_spin_unlock(&p
->pi_lock
);
5082 * migration_cpu_stop - this will be executed by a highprio stopper thread
5083 * and performs thread migration by bumping thread off CPU then
5084 * 'pushing' onto another runqueue.
5086 static int migration_cpu_stop(void *data
)
5088 struct migration_arg
*arg
= data
;
5091 * The original target cpu might have gone down and we might
5092 * be on another cpu but it doesn't matter.
5094 local_irq_disable();
5095 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5100 #ifdef CONFIG_HOTPLUG_CPU
5103 * Ensures that the idle task is using init_mm right before its cpu goes
5106 void idle_task_exit(void)
5108 struct mm_struct
*mm
= current
->active_mm
;
5110 BUG_ON(cpu_online(smp_processor_id()));
5113 switch_mm(mm
, &init_mm
, current
);
5118 * While a dead CPU has no uninterruptible tasks queued at this point,
5119 * it might still have a nonzero ->nr_uninterruptible counter, because
5120 * for performance reasons the counter is not stricly tracking tasks to
5121 * their home CPUs. So we just add the counter to another CPU's counter,
5122 * to keep the global sum constant after CPU-down:
5124 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5126 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5128 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5129 rq_src
->nr_uninterruptible
= 0;
5133 * remove the tasks which were accounted by rq from calc_load_tasks.
5135 static void calc_global_load_remove(struct rq
*rq
)
5137 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5138 rq
->calc_load_active
= 0;
5142 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5143 * try_to_wake_up()->select_task_rq().
5145 * Called with rq->lock held even though we'er in stop_machine() and
5146 * there's no concurrency possible, we hold the required locks anyway
5147 * because of lock validation efforts.
5149 static void migrate_tasks(unsigned int dead_cpu
)
5151 struct rq
*rq
= cpu_rq(dead_cpu
);
5152 struct task_struct
*next
, *stop
= rq
->stop
;
5156 * Fudge the rq selection such that the below task selection loop
5157 * doesn't get stuck on the currently eligible stop task.
5159 * We're currently inside stop_machine() and the rq is either stuck
5160 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5161 * either way we should never end up calling schedule() until we're
5166 /* Ensure any throttled groups are reachable by pick_next_task */
5167 unthrottle_offline_cfs_rqs(rq
);
5171 * There's this thread running, bail when that's the only
5174 if (rq
->nr_running
== 1)
5177 next
= pick_next_task(rq
);
5179 next
->sched_class
->put_prev_task(rq
, next
);
5181 /* Find suitable destination for @next, with force if needed. */
5182 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5183 raw_spin_unlock(&rq
->lock
);
5185 __migrate_task(next
, dead_cpu
, dest_cpu
);
5187 raw_spin_lock(&rq
->lock
);
5193 #endif /* CONFIG_HOTPLUG_CPU */
5195 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5197 static struct ctl_table sd_ctl_dir
[] = {
5199 .procname
= "sched_domain",
5205 static struct ctl_table sd_ctl_root
[] = {
5207 .procname
= "kernel",
5209 .child
= sd_ctl_dir
,
5214 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5216 struct ctl_table
*entry
=
5217 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5222 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5224 struct ctl_table
*entry
;
5227 * In the intermediate directories, both the child directory and
5228 * procname are dynamically allocated and could fail but the mode
5229 * will always be set. In the lowest directory the names are
5230 * static strings and all have proc handlers.
5232 for (entry
= *tablep
; entry
->mode
; entry
++) {
5234 sd_free_ctl_entry(&entry
->child
);
5235 if (entry
->proc_handler
== NULL
)
5236 kfree(entry
->procname
);
5244 set_table_entry(struct ctl_table
*entry
,
5245 const char *procname
, void *data
, int maxlen
,
5246 umode_t mode
, proc_handler
*proc_handler
)
5248 entry
->procname
= procname
;
5250 entry
->maxlen
= maxlen
;
5252 entry
->proc_handler
= proc_handler
;
5255 static struct ctl_table
*
5256 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5258 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5263 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5264 sizeof(long), 0644, proc_doulongvec_minmax
);
5265 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5266 sizeof(long), 0644, proc_doulongvec_minmax
);
5267 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5268 sizeof(int), 0644, proc_dointvec_minmax
);
5269 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5270 sizeof(int), 0644, proc_dointvec_minmax
);
5271 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5272 sizeof(int), 0644, proc_dointvec_minmax
);
5273 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5274 sizeof(int), 0644, proc_dointvec_minmax
);
5275 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5276 sizeof(int), 0644, proc_dointvec_minmax
);
5277 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5278 sizeof(int), 0644, proc_dointvec_minmax
);
5279 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5280 sizeof(int), 0644, proc_dointvec_minmax
);
5281 set_table_entry(&table
[9], "cache_nice_tries",
5282 &sd
->cache_nice_tries
,
5283 sizeof(int), 0644, proc_dointvec_minmax
);
5284 set_table_entry(&table
[10], "flags", &sd
->flags
,
5285 sizeof(int), 0644, proc_dointvec_minmax
);
5286 set_table_entry(&table
[11], "name", sd
->name
,
5287 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5288 /* &table[12] is terminator */
5293 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5295 struct ctl_table
*entry
, *table
;
5296 struct sched_domain
*sd
;
5297 int domain_num
= 0, i
;
5300 for_each_domain(cpu
, sd
)
5302 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5307 for_each_domain(cpu
, sd
) {
5308 snprintf(buf
, 32, "domain%d", i
);
5309 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5311 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5318 static struct ctl_table_header
*sd_sysctl_header
;
5319 static void register_sched_domain_sysctl(void)
5321 int i
, cpu_num
= num_possible_cpus();
5322 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5325 WARN_ON(sd_ctl_dir
[0].child
);
5326 sd_ctl_dir
[0].child
= entry
;
5331 for_each_possible_cpu(i
) {
5332 snprintf(buf
, 32, "cpu%d", i
);
5333 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5335 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5339 WARN_ON(sd_sysctl_header
);
5340 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5343 /* may be called multiple times per register */
5344 static void unregister_sched_domain_sysctl(void)
5346 if (sd_sysctl_header
)
5347 unregister_sysctl_table(sd_sysctl_header
);
5348 sd_sysctl_header
= NULL
;
5349 if (sd_ctl_dir
[0].child
)
5350 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5353 static void register_sched_domain_sysctl(void)
5356 static void unregister_sched_domain_sysctl(void)
5361 static void set_rq_online(struct rq
*rq
)
5364 const struct sched_class
*class;
5366 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5369 for_each_class(class) {
5370 if (class->rq_online
)
5371 class->rq_online(rq
);
5376 static void set_rq_offline(struct rq
*rq
)
5379 const struct sched_class
*class;
5381 for_each_class(class) {
5382 if (class->rq_offline
)
5383 class->rq_offline(rq
);
5386 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5392 * migration_call - callback that gets triggered when a CPU is added.
5393 * Here we can start up the necessary migration thread for the new CPU.
5395 static int __cpuinit
5396 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5398 int cpu
= (long)hcpu
;
5399 unsigned long flags
;
5400 struct rq
*rq
= cpu_rq(cpu
);
5402 switch (action
& ~CPU_TASKS_FROZEN
) {
5404 case CPU_UP_PREPARE
:
5405 rq
->calc_load_update
= calc_load_update
;
5409 /* Update our root-domain */
5410 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5412 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5416 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5419 #ifdef CONFIG_HOTPLUG_CPU
5421 sched_ttwu_pending();
5422 /* Update our root-domain */
5423 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5425 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5429 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5430 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5432 migrate_nr_uninterruptible(rq
);
5433 calc_global_load_remove(rq
);
5438 update_max_interval();
5444 * Register at high priority so that task migration (migrate_all_tasks)
5445 * happens before everything else. This has to be lower priority than
5446 * the notifier in the perf_event subsystem, though.
5448 static struct notifier_block __cpuinitdata migration_notifier
= {
5449 .notifier_call
= migration_call
,
5450 .priority
= CPU_PRI_MIGRATION
,
5453 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5454 unsigned long action
, void *hcpu
)
5456 switch (action
& ~CPU_TASKS_FROZEN
) {
5458 case CPU_DOWN_FAILED
:
5459 set_cpu_active((long)hcpu
, true);
5466 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5467 unsigned long action
, void *hcpu
)
5469 switch (action
& ~CPU_TASKS_FROZEN
) {
5470 case CPU_DOWN_PREPARE
:
5471 set_cpu_active((long)hcpu
, false);
5478 static int __init
migration_init(void)
5480 void *cpu
= (void *)(long)smp_processor_id();
5483 /* Initialize migration for the boot CPU */
5484 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5485 BUG_ON(err
== NOTIFY_BAD
);
5486 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5487 register_cpu_notifier(&migration_notifier
);
5489 /* Register cpu active notifiers */
5490 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5491 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5495 early_initcall(migration_init
);
5500 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5502 #ifdef CONFIG_SCHED_DEBUG
5504 static __read_mostly
int sched_domain_debug_enabled
;
5506 static int __init
sched_domain_debug_setup(char *str
)
5508 sched_domain_debug_enabled
= 1;
5512 early_param("sched_debug", sched_domain_debug_setup
);
5514 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5515 struct cpumask
*groupmask
)
5517 struct sched_group
*group
= sd
->groups
;
5520 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5521 cpumask_clear(groupmask
);
5523 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5525 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5526 printk("does not load-balance\n");
5528 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5533 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5535 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5536 printk(KERN_ERR
"ERROR: domain->span does not contain "
5539 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5540 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5544 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5548 printk(KERN_ERR
"ERROR: group is NULL\n");
5552 if (!group
->sgp
->power
) {
5553 printk(KERN_CONT
"\n");
5554 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5559 if (!cpumask_weight(sched_group_cpus(group
))) {
5560 printk(KERN_CONT
"\n");
5561 printk(KERN_ERR
"ERROR: empty group\n");
5565 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5566 printk(KERN_CONT
"\n");
5567 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5571 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5573 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5575 printk(KERN_CONT
" %s", str
);
5576 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5577 printk(KERN_CONT
" (cpu_power = %d)",
5581 group
= group
->next
;
5582 } while (group
!= sd
->groups
);
5583 printk(KERN_CONT
"\n");
5585 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5586 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5589 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5590 printk(KERN_ERR
"ERROR: parent span is not a superset "
5591 "of domain->span\n");
5595 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5599 if (!sched_domain_debug_enabled
)
5603 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5607 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5610 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5618 #else /* !CONFIG_SCHED_DEBUG */
5619 # define sched_domain_debug(sd, cpu) do { } while (0)
5620 #endif /* CONFIG_SCHED_DEBUG */
5622 static int sd_degenerate(struct sched_domain
*sd
)
5624 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5627 /* Following flags need at least 2 groups */
5628 if (sd
->flags
& (SD_LOAD_BALANCE
|
5629 SD_BALANCE_NEWIDLE
|
5633 SD_SHARE_PKG_RESOURCES
)) {
5634 if (sd
->groups
!= sd
->groups
->next
)
5638 /* Following flags don't use groups */
5639 if (sd
->flags
& (SD_WAKE_AFFINE
))
5646 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5648 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5650 if (sd_degenerate(parent
))
5653 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5656 /* Flags needing groups don't count if only 1 group in parent */
5657 if (parent
->groups
== parent
->groups
->next
) {
5658 pflags
&= ~(SD_LOAD_BALANCE
|
5659 SD_BALANCE_NEWIDLE
|
5663 SD_SHARE_PKG_RESOURCES
);
5664 if (nr_node_ids
== 1)
5665 pflags
&= ~SD_SERIALIZE
;
5667 if (~cflags
& pflags
)
5673 static void free_rootdomain(struct rcu_head
*rcu
)
5675 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5677 cpupri_cleanup(&rd
->cpupri
);
5678 free_cpumask_var(rd
->rto_mask
);
5679 free_cpumask_var(rd
->online
);
5680 free_cpumask_var(rd
->span
);
5684 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5686 struct root_domain
*old_rd
= NULL
;
5687 unsigned long flags
;
5689 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5694 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5697 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5700 * If we dont want to free the old_rt yet then
5701 * set old_rd to NULL to skip the freeing later
5704 if (!atomic_dec_and_test(&old_rd
->refcount
))
5708 atomic_inc(&rd
->refcount
);
5711 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5712 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5715 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5718 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5721 static int init_rootdomain(struct root_domain
*rd
)
5723 memset(rd
, 0, sizeof(*rd
));
5725 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5727 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5729 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5732 if (cpupri_init(&rd
->cpupri
) != 0)
5737 free_cpumask_var(rd
->rto_mask
);
5739 free_cpumask_var(rd
->online
);
5741 free_cpumask_var(rd
->span
);
5747 * By default the system creates a single root-domain with all cpus as
5748 * members (mimicking the global state we have today).
5750 struct root_domain def_root_domain
;
5752 static void init_defrootdomain(void)
5754 init_rootdomain(&def_root_domain
);
5756 atomic_set(&def_root_domain
.refcount
, 1);
5759 static struct root_domain
*alloc_rootdomain(void)
5761 struct root_domain
*rd
;
5763 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5767 if (init_rootdomain(rd
) != 0) {
5775 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5777 struct sched_group
*tmp
, *first
;
5786 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5791 } while (sg
!= first
);
5794 static void free_sched_domain(struct rcu_head
*rcu
)
5796 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5799 * If its an overlapping domain it has private groups, iterate and
5802 if (sd
->flags
& SD_OVERLAP
) {
5803 free_sched_groups(sd
->groups
, 1);
5804 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5805 kfree(sd
->groups
->sgp
);
5811 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5813 call_rcu(&sd
->rcu
, free_sched_domain
);
5816 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5818 for (; sd
; sd
= sd
->parent
)
5819 destroy_sched_domain(sd
, cpu
);
5823 * Keep a special pointer to the highest sched_domain that has
5824 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5825 * allows us to avoid some pointer chasing select_idle_sibling().
5827 * Also keep a unique ID per domain (we use the first cpu number in
5828 * the cpumask of the domain), this allows us to quickly tell if
5829 * two cpus are in the same cache domain, see cpus_share_cache().
5831 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5832 DEFINE_PER_CPU(int, sd_llc_id
);
5834 static void update_top_cache_domain(int cpu
)
5836 struct sched_domain
*sd
;
5839 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5841 id
= cpumask_first(sched_domain_span(sd
));
5843 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5844 per_cpu(sd_llc_id
, cpu
) = id
;
5848 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5849 * hold the hotplug lock.
5852 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5854 struct rq
*rq
= cpu_rq(cpu
);
5855 struct sched_domain
*tmp
;
5857 /* Remove the sched domains which do not contribute to scheduling. */
5858 for (tmp
= sd
; tmp
; ) {
5859 struct sched_domain
*parent
= tmp
->parent
;
5863 if (sd_parent_degenerate(tmp
, parent
)) {
5864 tmp
->parent
= parent
->parent
;
5866 parent
->parent
->child
= tmp
;
5867 destroy_sched_domain(parent
, cpu
);
5872 if (sd
&& sd_degenerate(sd
)) {
5875 destroy_sched_domain(tmp
, cpu
);
5880 sched_domain_debug(sd
, cpu
);
5882 rq_attach_root(rq
, rd
);
5884 rcu_assign_pointer(rq
->sd
, sd
);
5885 destroy_sched_domains(tmp
, cpu
);
5887 update_top_cache_domain(cpu
);
5890 /* cpus with isolated domains */
5891 static cpumask_var_t cpu_isolated_map
;
5893 /* Setup the mask of cpus configured for isolated domains */
5894 static int __init
isolated_cpu_setup(char *str
)
5896 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5897 cpulist_parse(str
, cpu_isolated_map
);
5901 __setup("isolcpus=", isolated_cpu_setup
);
5906 * find_next_best_node - find the next node to include in a sched_domain
5907 * @node: node whose sched_domain we're building
5908 * @used_nodes: nodes already in the sched_domain
5910 * Find the next node to include in a given scheduling domain. Simply
5911 * finds the closest node not already in the @used_nodes map.
5913 * Should use nodemask_t.
5915 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
5917 int i
, n
, val
, min_val
, best_node
= -1;
5921 for (i
= 0; i
< nr_node_ids
; i
++) {
5922 /* Start at @node */
5923 n
= (node
+ i
) % nr_node_ids
;
5925 if (!nr_cpus_node(n
))
5928 /* Skip already used nodes */
5929 if (node_isset(n
, *used_nodes
))
5932 /* Simple min distance search */
5933 val
= node_distance(node
, n
);
5935 if (val
< min_val
) {
5941 if (best_node
!= -1)
5942 node_set(best_node
, *used_nodes
);
5947 * sched_domain_node_span - get a cpumask for a node's sched_domain
5948 * @node: node whose cpumask we're constructing
5949 * @span: resulting cpumask
5951 * Given a node, construct a good cpumask for its sched_domain to span. It
5952 * should be one that prevents unnecessary balancing, but also spreads tasks
5955 static void sched_domain_node_span(int node
, struct cpumask
*span
)
5957 nodemask_t used_nodes
;
5960 cpumask_clear(span
);
5961 nodes_clear(used_nodes
);
5963 cpumask_or(span
, span
, cpumask_of_node(node
));
5964 node_set(node
, used_nodes
);
5966 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5967 int next_node
= find_next_best_node(node
, &used_nodes
);
5970 cpumask_or(span
, span
, cpumask_of_node(next_node
));
5974 static const struct cpumask
*cpu_node_mask(int cpu
)
5976 lockdep_assert_held(&sched_domains_mutex
);
5978 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
5980 return sched_domains_tmpmask
;
5983 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
5985 return cpu_possible_mask
;
5987 #endif /* CONFIG_NUMA */
5989 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5991 return cpumask_of_node(cpu_to_node(cpu
));
5994 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5997 struct sched_domain
**__percpu sd
;
5998 struct sched_group
**__percpu sg
;
5999 struct sched_group_power
**__percpu sgp
;
6003 struct sched_domain
** __percpu sd
;
6004 struct root_domain
*rd
;
6014 struct sched_domain_topology_level
;
6016 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
6017 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
6019 #define SDTL_OVERLAP 0x01
6021 struct sched_domain_topology_level
{
6022 sched_domain_init_f init
;
6023 sched_domain_mask_f mask
;
6025 struct sd_data data
;
6029 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6031 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6032 const struct cpumask
*span
= sched_domain_span(sd
);
6033 struct cpumask
*covered
= sched_domains_tmpmask
;
6034 struct sd_data
*sdd
= sd
->private;
6035 struct sched_domain
*child
;
6038 cpumask_clear(covered
);
6040 for_each_cpu(i
, span
) {
6041 struct cpumask
*sg_span
;
6043 if (cpumask_test_cpu(i
, covered
))
6046 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6047 GFP_KERNEL
, cpu_to_node(cpu
));
6052 sg_span
= sched_group_cpus(sg
);
6054 child
= *per_cpu_ptr(sdd
->sd
, i
);
6056 child
= child
->child
;
6057 cpumask_copy(sg_span
, sched_domain_span(child
));
6059 cpumask_set_cpu(i
, sg_span
);
6061 cpumask_or(covered
, covered
, sg_span
);
6063 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
6064 atomic_inc(&sg
->sgp
->ref
);
6066 if (cpumask_test_cpu(cpu
, sg_span
))
6076 sd
->groups
= groups
;
6081 free_sched_groups(first
, 0);
6086 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6088 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6089 struct sched_domain
*child
= sd
->child
;
6092 cpu
= cpumask_first(sched_domain_span(child
));
6095 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6096 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
6097 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
6104 * build_sched_groups will build a circular linked list of the groups
6105 * covered by the given span, and will set each group's ->cpumask correctly,
6106 * and ->cpu_power to 0.
6108 * Assumes the sched_domain tree is fully constructed
6111 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6113 struct sched_group
*first
= NULL
, *last
= NULL
;
6114 struct sd_data
*sdd
= sd
->private;
6115 const struct cpumask
*span
= sched_domain_span(sd
);
6116 struct cpumask
*covered
;
6119 get_group(cpu
, sdd
, &sd
->groups
);
6120 atomic_inc(&sd
->groups
->ref
);
6122 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
6125 lockdep_assert_held(&sched_domains_mutex
);
6126 covered
= sched_domains_tmpmask
;
6128 cpumask_clear(covered
);
6130 for_each_cpu(i
, span
) {
6131 struct sched_group
*sg
;
6132 int group
= get_group(i
, sdd
, &sg
);
6135 if (cpumask_test_cpu(i
, covered
))
6138 cpumask_clear(sched_group_cpus(sg
));
6141 for_each_cpu(j
, span
) {
6142 if (get_group(j
, sdd
, NULL
) != group
)
6145 cpumask_set_cpu(j
, covered
);
6146 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6161 * Initialize sched groups cpu_power.
6163 * cpu_power indicates the capacity of sched group, which is used while
6164 * distributing the load between different sched groups in a sched domain.
6165 * Typically cpu_power for all the groups in a sched domain will be same unless
6166 * there are asymmetries in the topology. If there are asymmetries, group
6167 * having more cpu_power will pickup more load compared to the group having
6170 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6172 struct sched_group
*sg
= sd
->groups
;
6174 WARN_ON(!sd
|| !sg
);
6177 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6179 } while (sg
!= sd
->groups
);
6181 if (cpu
!= group_first_cpu(sg
))
6184 update_group_power(sd
, cpu
);
6185 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6188 int __weak
arch_sd_sibling_asym_packing(void)
6190 return 0*SD_ASYM_PACKING
;
6194 * Initializers for schedule domains
6195 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6198 #ifdef CONFIG_SCHED_DEBUG
6199 # define SD_INIT_NAME(sd, type) sd->name = #type
6201 # define SD_INIT_NAME(sd, type) do { } while (0)
6204 #define SD_INIT_FUNC(type) \
6205 static noinline struct sched_domain * \
6206 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6208 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6209 *sd = SD_##type##_INIT; \
6210 SD_INIT_NAME(sd, type); \
6211 sd->private = &tl->data; \
6217 SD_INIT_FUNC(ALLNODES
)
6220 #ifdef CONFIG_SCHED_SMT
6221 SD_INIT_FUNC(SIBLING
)
6223 #ifdef CONFIG_SCHED_MC
6226 #ifdef CONFIG_SCHED_BOOK
6230 static int default_relax_domain_level
= -1;
6231 int sched_domain_level_max
;
6233 static int __init
setup_relax_domain_level(char *str
)
6237 val
= simple_strtoul(str
, NULL
, 0);
6238 if (val
< sched_domain_level_max
)
6239 default_relax_domain_level
= val
;
6243 __setup("relax_domain_level=", setup_relax_domain_level
);
6245 static void set_domain_attribute(struct sched_domain
*sd
,
6246 struct sched_domain_attr
*attr
)
6250 if (!attr
|| attr
->relax_domain_level
< 0) {
6251 if (default_relax_domain_level
< 0)
6254 request
= default_relax_domain_level
;
6256 request
= attr
->relax_domain_level
;
6257 if (request
< sd
->level
) {
6258 /* turn off idle balance on this domain */
6259 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6261 /* turn on idle balance on this domain */
6262 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6266 static void __sdt_free(const struct cpumask
*cpu_map
);
6267 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6269 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6270 const struct cpumask
*cpu_map
)
6274 if (!atomic_read(&d
->rd
->refcount
))
6275 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6277 free_percpu(d
->sd
); /* fall through */
6279 __sdt_free(cpu_map
); /* fall through */
6285 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6286 const struct cpumask
*cpu_map
)
6288 memset(d
, 0, sizeof(*d
));
6290 if (__sdt_alloc(cpu_map
))
6291 return sa_sd_storage
;
6292 d
->sd
= alloc_percpu(struct sched_domain
*);
6294 return sa_sd_storage
;
6295 d
->rd
= alloc_rootdomain();
6298 return sa_rootdomain
;
6302 * NULL the sd_data elements we've used to build the sched_domain and
6303 * sched_group structure so that the subsequent __free_domain_allocs()
6304 * will not free the data we're using.
6306 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6308 struct sd_data
*sdd
= sd
->private;
6310 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6311 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6313 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6314 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6316 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6317 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6320 #ifdef CONFIG_SCHED_SMT
6321 static const struct cpumask
*cpu_smt_mask(int cpu
)
6323 return topology_thread_cpumask(cpu
);
6328 * Topology list, bottom-up.
6330 static struct sched_domain_topology_level default_topology
[] = {
6331 #ifdef CONFIG_SCHED_SMT
6332 { sd_init_SIBLING
, cpu_smt_mask
, },
6334 #ifdef CONFIG_SCHED_MC
6335 { sd_init_MC
, cpu_coregroup_mask
, },
6337 #ifdef CONFIG_SCHED_BOOK
6338 { sd_init_BOOK
, cpu_book_mask
, },
6340 { sd_init_CPU
, cpu_cpu_mask
, },
6342 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
6343 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
6348 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6350 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6352 struct sched_domain_topology_level
*tl
;
6355 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6356 struct sd_data
*sdd
= &tl
->data
;
6358 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6362 sdd
->sg
= alloc_percpu(struct sched_group
*);
6366 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6370 for_each_cpu(j
, cpu_map
) {
6371 struct sched_domain
*sd
;
6372 struct sched_group
*sg
;
6373 struct sched_group_power
*sgp
;
6375 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6376 GFP_KERNEL
, cpu_to_node(j
));
6380 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6382 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6383 GFP_KERNEL
, cpu_to_node(j
));
6389 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6391 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
6392 GFP_KERNEL
, cpu_to_node(j
));
6396 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6403 static void __sdt_free(const struct cpumask
*cpu_map
)
6405 struct sched_domain_topology_level
*tl
;
6408 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6409 struct sd_data
*sdd
= &tl
->data
;
6411 for_each_cpu(j
, cpu_map
) {
6412 struct sched_domain
*sd
;
6415 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6416 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6417 free_sched_groups(sd
->groups
, 0);
6418 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6422 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6424 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6426 free_percpu(sdd
->sd
);
6428 free_percpu(sdd
->sg
);
6430 free_percpu(sdd
->sgp
);
6435 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6436 struct s_data
*d
, const struct cpumask
*cpu_map
,
6437 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6440 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6444 set_domain_attribute(sd
, attr
);
6445 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6447 sd
->level
= child
->level
+ 1;
6448 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6457 * Build sched domains for a given set of cpus and attach the sched domains
6458 * to the individual cpus
6460 static int build_sched_domains(const struct cpumask
*cpu_map
,
6461 struct sched_domain_attr
*attr
)
6463 enum s_alloc alloc_state
= sa_none
;
6464 struct sched_domain
*sd
;
6466 int i
, ret
= -ENOMEM
;
6468 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6469 if (alloc_state
!= sa_rootdomain
)
6472 /* Set up domains for cpus specified by the cpu_map. */
6473 for_each_cpu(i
, cpu_map
) {
6474 struct sched_domain_topology_level
*tl
;
6477 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6478 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6479 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6480 sd
->flags
|= SD_OVERLAP
;
6481 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6488 *per_cpu_ptr(d
.sd
, i
) = sd
;
6491 /* Build the groups for the domains */
6492 for_each_cpu(i
, cpu_map
) {
6493 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6494 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6495 if (sd
->flags
& SD_OVERLAP
) {
6496 if (build_overlap_sched_groups(sd
, i
))
6499 if (build_sched_groups(sd
, i
))
6505 /* Calculate CPU power for physical packages and nodes */
6506 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6507 if (!cpumask_test_cpu(i
, cpu_map
))
6510 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6511 claim_allocations(i
, sd
);
6512 init_sched_groups_power(i
, sd
);
6516 /* Attach the domains */
6518 for_each_cpu(i
, cpu_map
) {
6519 sd
= *per_cpu_ptr(d
.sd
, i
);
6520 cpu_attach_domain(sd
, d
.rd
, i
);
6526 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6530 static cpumask_var_t
*doms_cur
; /* current sched domains */
6531 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6532 static struct sched_domain_attr
*dattr_cur
;
6533 /* attribues of custom domains in 'doms_cur' */
6536 * Special case: If a kmalloc of a doms_cur partition (array of
6537 * cpumask) fails, then fallback to a single sched domain,
6538 * as determined by the single cpumask fallback_doms.
6540 static cpumask_var_t fallback_doms
;
6543 * arch_update_cpu_topology lets virtualized architectures update the
6544 * cpu core maps. It is supposed to return 1 if the topology changed
6545 * or 0 if it stayed the same.
6547 int __attribute__((weak
)) arch_update_cpu_topology(void)
6552 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6555 cpumask_var_t
*doms
;
6557 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6560 for (i
= 0; i
< ndoms
; i
++) {
6561 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6562 free_sched_domains(doms
, i
);
6569 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6572 for (i
= 0; i
< ndoms
; i
++)
6573 free_cpumask_var(doms
[i
]);
6578 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6579 * For now this just excludes isolated cpus, but could be used to
6580 * exclude other special cases in the future.
6582 static int init_sched_domains(const struct cpumask
*cpu_map
)
6586 arch_update_cpu_topology();
6588 doms_cur
= alloc_sched_domains(ndoms_cur
);
6590 doms_cur
= &fallback_doms
;
6591 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6593 err
= build_sched_domains(doms_cur
[0], NULL
);
6594 register_sched_domain_sysctl();
6600 * Detach sched domains from a group of cpus specified in cpu_map
6601 * These cpus will now be attached to the NULL domain
6603 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6608 for_each_cpu(i
, cpu_map
)
6609 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6613 /* handle null as "default" */
6614 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6615 struct sched_domain_attr
*new, int idx_new
)
6617 struct sched_domain_attr tmp
;
6624 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6625 new ? (new + idx_new
) : &tmp
,
6626 sizeof(struct sched_domain_attr
));
6630 * Partition sched domains as specified by the 'ndoms_new'
6631 * cpumasks in the array doms_new[] of cpumasks. This compares
6632 * doms_new[] to the current sched domain partitioning, doms_cur[].
6633 * It destroys each deleted domain and builds each new domain.
6635 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6636 * The masks don't intersect (don't overlap.) We should setup one
6637 * sched domain for each mask. CPUs not in any of the cpumasks will
6638 * not be load balanced. If the same cpumask appears both in the
6639 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6642 * The passed in 'doms_new' should be allocated using
6643 * alloc_sched_domains. This routine takes ownership of it and will
6644 * free_sched_domains it when done with it. If the caller failed the
6645 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6646 * and partition_sched_domains() will fallback to the single partition
6647 * 'fallback_doms', it also forces the domains to be rebuilt.
6649 * If doms_new == NULL it will be replaced with cpu_online_mask.
6650 * ndoms_new == 0 is a special case for destroying existing domains,
6651 * and it will not create the default domain.
6653 * Call with hotplug lock held
6655 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6656 struct sched_domain_attr
*dattr_new
)
6661 mutex_lock(&sched_domains_mutex
);
6663 /* always unregister in case we don't destroy any domains */
6664 unregister_sched_domain_sysctl();
6666 /* Let architecture update cpu core mappings. */
6667 new_topology
= arch_update_cpu_topology();
6669 n
= doms_new
? ndoms_new
: 0;
6671 /* Destroy deleted domains */
6672 for (i
= 0; i
< ndoms_cur
; i
++) {
6673 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6674 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6675 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6678 /* no match - a current sched domain not in new doms_new[] */
6679 detach_destroy_domains(doms_cur
[i
]);
6684 if (doms_new
== NULL
) {
6686 doms_new
= &fallback_doms
;
6687 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6688 WARN_ON_ONCE(dattr_new
);
6691 /* Build new domains */
6692 for (i
= 0; i
< ndoms_new
; i
++) {
6693 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6694 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6695 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6698 /* no match - add a new doms_new */
6699 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6704 /* Remember the new sched domains */
6705 if (doms_cur
!= &fallback_doms
)
6706 free_sched_domains(doms_cur
, ndoms_cur
);
6707 kfree(dattr_cur
); /* kfree(NULL) is safe */
6708 doms_cur
= doms_new
;
6709 dattr_cur
= dattr_new
;
6710 ndoms_cur
= ndoms_new
;
6712 register_sched_domain_sysctl();
6714 mutex_unlock(&sched_domains_mutex
);
6717 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6718 static void reinit_sched_domains(void)
6722 /* Destroy domains first to force the rebuild */
6723 partition_sched_domains(0, NULL
, NULL
);
6725 rebuild_sched_domains();
6729 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6731 unsigned int level
= 0;
6733 if (sscanf(buf
, "%u", &level
) != 1)
6737 * level is always be positive so don't check for
6738 * level < POWERSAVINGS_BALANCE_NONE which is 0
6739 * What happens on 0 or 1 byte write,
6740 * need to check for count as well?
6743 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
6747 sched_smt_power_savings
= level
;
6749 sched_mc_power_savings
= level
;
6751 reinit_sched_domains();
6756 #ifdef CONFIG_SCHED_MC
6757 static ssize_t
sched_mc_power_savings_show(struct device
*dev
,
6758 struct device_attribute
*attr
,
6761 return sprintf(buf
, "%u\n", sched_mc_power_savings
);
6763 static ssize_t
sched_mc_power_savings_store(struct device
*dev
,
6764 struct device_attribute
*attr
,
6765 const char *buf
, size_t count
)
6767 return sched_power_savings_store(buf
, count
, 0);
6769 static DEVICE_ATTR(sched_mc_power_savings
, 0644,
6770 sched_mc_power_savings_show
,
6771 sched_mc_power_savings_store
);
6774 #ifdef CONFIG_SCHED_SMT
6775 static ssize_t
sched_smt_power_savings_show(struct device
*dev
,
6776 struct device_attribute
*attr
,
6779 return sprintf(buf
, "%u\n", sched_smt_power_savings
);
6781 static ssize_t
sched_smt_power_savings_store(struct device
*dev
,
6782 struct device_attribute
*attr
,
6783 const char *buf
, size_t count
)
6785 return sched_power_savings_store(buf
, count
, 1);
6787 static DEVICE_ATTR(sched_smt_power_savings
, 0644,
6788 sched_smt_power_savings_show
,
6789 sched_smt_power_savings_store
);
6792 int __init
sched_create_sysfs_power_savings_entries(struct device
*dev
)
6796 #ifdef CONFIG_SCHED_SMT
6798 err
= device_create_file(dev
, &dev_attr_sched_smt_power_savings
);
6800 #ifdef CONFIG_SCHED_MC
6801 if (!err
&& mc_capable())
6802 err
= device_create_file(dev
, &dev_attr_sched_mc_power_savings
);
6806 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6809 * Update cpusets according to cpu_active mask. If cpusets are
6810 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6811 * around partition_sched_domains().
6813 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6816 switch (action
& ~CPU_TASKS_FROZEN
) {
6818 case CPU_DOWN_FAILED
:
6819 cpuset_update_active_cpus();
6826 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6829 switch (action
& ~CPU_TASKS_FROZEN
) {
6830 case CPU_DOWN_PREPARE
:
6831 cpuset_update_active_cpus();
6838 void __init
sched_init_smp(void)
6840 cpumask_var_t non_isolated_cpus
;
6842 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6843 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6846 mutex_lock(&sched_domains_mutex
);
6847 init_sched_domains(cpu_active_mask
);
6848 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6849 if (cpumask_empty(non_isolated_cpus
))
6850 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6851 mutex_unlock(&sched_domains_mutex
);
6854 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6855 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6857 /* RT runtime code needs to handle some hotplug events */
6858 hotcpu_notifier(update_runtime
, 0);
6862 /* Move init over to a non-isolated CPU */
6863 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6865 sched_init_granularity();
6866 free_cpumask_var(non_isolated_cpus
);
6868 init_sched_rt_class();
6871 void __init
sched_init_smp(void)
6873 sched_init_granularity();
6875 #endif /* CONFIG_SMP */
6877 const_debug
unsigned int sysctl_timer_migration
= 1;
6879 int in_sched_functions(unsigned long addr
)
6881 return in_lock_functions(addr
) ||
6882 (addr
>= (unsigned long)__sched_text_start
6883 && addr
< (unsigned long)__sched_text_end
);
6886 #ifdef CONFIG_CGROUP_SCHED
6887 struct task_group root_task_group
;
6890 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6892 void __init
sched_init(void)
6895 unsigned long alloc_size
= 0, ptr
;
6897 #ifdef CONFIG_FAIR_GROUP_SCHED
6898 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6900 #ifdef CONFIG_RT_GROUP_SCHED
6901 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6903 #ifdef CONFIG_CPUMASK_OFFSTACK
6904 alloc_size
+= num_possible_cpus() * cpumask_size();
6907 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6909 #ifdef CONFIG_FAIR_GROUP_SCHED
6910 root_task_group
.se
= (struct sched_entity
**)ptr
;
6911 ptr
+= nr_cpu_ids
* sizeof(void **);
6913 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6914 ptr
+= nr_cpu_ids
* sizeof(void **);
6916 #endif /* CONFIG_FAIR_GROUP_SCHED */
6917 #ifdef CONFIG_RT_GROUP_SCHED
6918 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6919 ptr
+= nr_cpu_ids
* sizeof(void **);
6921 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6922 ptr
+= nr_cpu_ids
* sizeof(void **);
6924 #endif /* CONFIG_RT_GROUP_SCHED */
6925 #ifdef CONFIG_CPUMASK_OFFSTACK
6926 for_each_possible_cpu(i
) {
6927 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6928 ptr
+= cpumask_size();
6930 #endif /* CONFIG_CPUMASK_OFFSTACK */
6934 init_defrootdomain();
6937 init_rt_bandwidth(&def_rt_bandwidth
,
6938 global_rt_period(), global_rt_runtime());
6940 #ifdef CONFIG_RT_GROUP_SCHED
6941 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6942 global_rt_period(), global_rt_runtime());
6943 #endif /* CONFIG_RT_GROUP_SCHED */
6945 #ifdef CONFIG_CGROUP_SCHED
6946 list_add(&root_task_group
.list
, &task_groups
);
6947 INIT_LIST_HEAD(&root_task_group
.children
);
6948 INIT_LIST_HEAD(&root_task_group
.siblings
);
6949 autogroup_init(&init_task
);
6951 #endif /* CONFIG_CGROUP_SCHED */
6953 #ifdef CONFIG_CGROUP_CPUACCT
6954 root_cpuacct
.cpustat
= &kernel_cpustat
;
6955 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6956 /* Too early, not expected to fail */
6957 BUG_ON(!root_cpuacct
.cpuusage
);
6959 for_each_possible_cpu(i
) {
6963 raw_spin_lock_init(&rq
->lock
);
6965 rq
->calc_load_active
= 0;
6966 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6967 init_cfs_rq(&rq
->cfs
);
6968 init_rt_rq(&rq
->rt
, rq
);
6969 #ifdef CONFIG_FAIR_GROUP_SCHED
6970 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6971 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6973 * How much cpu bandwidth does root_task_group get?
6975 * In case of task-groups formed thr' the cgroup filesystem, it
6976 * gets 100% of the cpu resources in the system. This overall
6977 * system cpu resource is divided among the tasks of
6978 * root_task_group and its child task-groups in a fair manner,
6979 * based on each entity's (task or task-group's) weight
6980 * (se->load.weight).
6982 * In other words, if root_task_group has 10 tasks of weight
6983 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6984 * then A0's share of the cpu resource is:
6986 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6988 * We achieve this by letting root_task_group's tasks sit
6989 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6991 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6992 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6993 #endif /* CONFIG_FAIR_GROUP_SCHED */
6995 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6996 #ifdef CONFIG_RT_GROUP_SCHED
6997 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6998 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7001 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7002 rq
->cpu_load
[j
] = 0;
7004 rq
->last_load_update_tick
= jiffies
;
7009 rq
->cpu_power
= SCHED_POWER_SCALE
;
7010 rq
->post_schedule
= 0;
7011 rq
->active_balance
= 0;
7012 rq
->next_balance
= jiffies
;
7017 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7019 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7021 rq_attach_root(rq
, &def_root_domain
);
7027 atomic_set(&rq
->nr_iowait
, 0);
7030 set_load_weight(&init_task
);
7032 #ifdef CONFIG_PREEMPT_NOTIFIERS
7033 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7036 #ifdef CONFIG_RT_MUTEXES
7037 plist_head_init(&init_task
.pi_waiters
);
7041 * The boot idle thread does lazy MMU switching as well:
7043 atomic_inc(&init_mm
.mm_count
);
7044 enter_lazy_tlb(&init_mm
, current
);
7047 * Make us the idle thread. Technically, schedule() should not be
7048 * called from this thread, however somewhere below it might be,
7049 * but because we are the idle thread, we just pick up running again
7050 * when this runqueue becomes "idle".
7052 init_idle(current
, smp_processor_id());
7054 calc_load_update
= jiffies
+ LOAD_FREQ
;
7057 * During early bootup we pretend to be a normal task:
7059 current
->sched_class
= &fair_sched_class
;
7062 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7063 /* May be allocated at isolcpus cmdline parse time */
7064 if (cpu_isolated_map
== NULL
)
7065 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7066 idle_thread_set_boot_cpu();
7068 init_sched_fair_class();
7070 scheduler_running
= 1;
7073 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7074 static inline int preempt_count_equals(int preempt_offset
)
7076 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7078 return (nested
== preempt_offset
);
7081 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7083 static unsigned long prev_jiffy
; /* ratelimiting */
7085 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7086 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7087 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7089 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7091 prev_jiffy
= jiffies
;
7094 "BUG: sleeping function called from invalid context at %s:%d\n",
7097 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7098 in_atomic(), irqs_disabled(),
7099 current
->pid
, current
->comm
);
7101 debug_show_held_locks(current
);
7102 if (irqs_disabled())
7103 print_irqtrace_events(current
);
7106 EXPORT_SYMBOL(__might_sleep
);
7109 #ifdef CONFIG_MAGIC_SYSRQ
7110 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7112 const struct sched_class
*prev_class
= p
->sched_class
;
7113 int old_prio
= p
->prio
;
7118 dequeue_task(rq
, p
, 0);
7119 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7121 enqueue_task(rq
, p
, 0);
7122 resched_task(rq
->curr
);
7125 check_class_changed(rq
, p
, prev_class
, old_prio
);
7128 void normalize_rt_tasks(void)
7130 struct task_struct
*g
, *p
;
7131 unsigned long flags
;
7134 read_lock_irqsave(&tasklist_lock
, flags
);
7135 do_each_thread(g
, p
) {
7137 * Only normalize user tasks:
7142 p
->se
.exec_start
= 0;
7143 #ifdef CONFIG_SCHEDSTATS
7144 p
->se
.statistics
.wait_start
= 0;
7145 p
->se
.statistics
.sleep_start
= 0;
7146 p
->se
.statistics
.block_start
= 0;
7151 * Renice negative nice level userspace
7154 if (TASK_NICE(p
) < 0 && p
->mm
)
7155 set_user_nice(p
, 0);
7159 raw_spin_lock(&p
->pi_lock
);
7160 rq
= __task_rq_lock(p
);
7162 normalize_task(rq
, p
);
7164 __task_rq_unlock(rq
);
7165 raw_spin_unlock(&p
->pi_lock
);
7166 } while_each_thread(g
, p
);
7168 read_unlock_irqrestore(&tasklist_lock
, flags
);
7171 #endif /* CONFIG_MAGIC_SYSRQ */
7173 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7175 * These functions are only useful for the IA64 MCA handling, or kdb.
7177 * They can only be called when the whole system has been
7178 * stopped - every CPU needs to be quiescent, and no scheduling
7179 * activity can take place. Using them for anything else would
7180 * be a serious bug, and as a result, they aren't even visible
7181 * under any other configuration.
7185 * curr_task - return the current task for a given cpu.
7186 * @cpu: the processor in question.
7188 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7190 struct task_struct
*curr_task(int cpu
)
7192 return cpu_curr(cpu
);
7195 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7199 * set_curr_task - set the current task for a given cpu.
7200 * @cpu: the processor in question.
7201 * @p: the task pointer to set.
7203 * Description: This function must only be used when non-maskable interrupts
7204 * are serviced on a separate stack. It allows the architecture to switch the
7205 * notion of the current task on a cpu in a non-blocking manner. This function
7206 * must be called with all CPU's synchronized, and interrupts disabled, the
7207 * and caller must save the original value of the current task (see
7208 * curr_task() above) and restore that value before reenabling interrupts and
7209 * re-starting the system.
7211 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7213 void set_curr_task(int cpu
, struct task_struct
*p
)
7220 #ifdef CONFIG_CGROUP_SCHED
7221 /* task_group_lock serializes the addition/removal of task groups */
7222 static DEFINE_SPINLOCK(task_group_lock
);
7224 static void free_sched_group(struct task_group
*tg
)
7226 free_fair_sched_group(tg
);
7227 free_rt_sched_group(tg
);
7232 /* allocate runqueue etc for a new task group */
7233 struct task_group
*sched_create_group(struct task_group
*parent
)
7235 struct task_group
*tg
;
7236 unsigned long flags
;
7238 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7240 return ERR_PTR(-ENOMEM
);
7242 if (!alloc_fair_sched_group(tg
, parent
))
7245 if (!alloc_rt_sched_group(tg
, parent
))
7248 spin_lock_irqsave(&task_group_lock
, flags
);
7249 list_add_rcu(&tg
->list
, &task_groups
);
7251 WARN_ON(!parent
); /* root should already exist */
7253 tg
->parent
= parent
;
7254 INIT_LIST_HEAD(&tg
->children
);
7255 list_add_rcu(&tg
->siblings
, &parent
->children
);
7256 spin_unlock_irqrestore(&task_group_lock
, flags
);
7261 free_sched_group(tg
);
7262 return ERR_PTR(-ENOMEM
);
7265 /* rcu callback to free various structures associated with a task group */
7266 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7268 /* now it should be safe to free those cfs_rqs */
7269 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7272 /* Destroy runqueue etc associated with a task group */
7273 void sched_destroy_group(struct task_group
*tg
)
7275 unsigned long flags
;
7278 /* end participation in shares distribution */
7279 for_each_possible_cpu(i
)
7280 unregister_fair_sched_group(tg
, i
);
7282 spin_lock_irqsave(&task_group_lock
, flags
);
7283 list_del_rcu(&tg
->list
);
7284 list_del_rcu(&tg
->siblings
);
7285 spin_unlock_irqrestore(&task_group_lock
, flags
);
7287 /* wait for possible concurrent references to cfs_rqs complete */
7288 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7291 /* change task's runqueue when it moves between groups.
7292 * The caller of this function should have put the task in its new group
7293 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7294 * reflect its new group.
7296 void sched_move_task(struct task_struct
*tsk
)
7299 unsigned long flags
;
7302 rq
= task_rq_lock(tsk
, &flags
);
7304 running
= task_current(rq
, tsk
);
7308 dequeue_task(rq
, tsk
, 0);
7309 if (unlikely(running
))
7310 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7312 #ifdef CONFIG_FAIR_GROUP_SCHED
7313 if (tsk
->sched_class
->task_move_group
)
7314 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7317 set_task_rq(tsk
, task_cpu(tsk
));
7319 if (unlikely(running
))
7320 tsk
->sched_class
->set_curr_task(rq
);
7322 enqueue_task(rq
, tsk
, 0);
7324 task_rq_unlock(rq
, tsk
, &flags
);
7326 #endif /* CONFIG_CGROUP_SCHED */
7328 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7329 static unsigned long to_ratio(u64 period
, u64 runtime
)
7331 if (runtime
== RUNTIME_INF
)
7334 return div64_u64(runtime
<< 20, period
);
7338 #ifdef CONFIG_RT_GROUP_SCHED
7340 * Ensure that the real time constraints are schedulable.
7342 static DEFINE_MUTEX(rt_constraints_mutex
);
7344 /* Must be called with tasklist_lock held */
7345 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7347 struct task_struct
*g
, *p
;
7349 do_each_thread(g
, p
) {
7350 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7352 } while_each_thread(g
, p
);
7357 struct rt_schedulable_data
{
7358 struct task_group
*tg
;
7363 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7365 struct rt_schedulable_data
*d
= data
;
7366 struct task_group
*child
;
7367 unsigned long total
, sum
= 0;
7368 u64 period
, runtime
;
7370 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7371 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7374 period
= d
->rt_period
;
7375 runtime
= d
->rt_runtime
;
7379 * Cannot have more runtime than the period.
7381 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7385 * Ensure we don't starve existing RT tasks.
7387 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7390 total
= to_ratio(period
, runtime
);
7393 * Nobody can have more than the global setting allows.
7395 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7399 * The sum of our children's runtime should not exceed our own.
7401 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7402 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7403 runtime
= child
->rt_bandwidth
.rt_runtime
;
7405 if (child
== d
->tg
) {
7406 period
= d
->rt_period
;
7407 runtime
= d
->rt_runtime
;
7410 sum
+= to_ratio(period
, runtime
);
7419 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7423 struct rt_schedulable_data data
= {
7425 .rt_period
= period
,
7426 .rt_runtime
= runtime
,
7430 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7436 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7437 u64 rt_period
, u64 rt_runtime
)
7441 mutex_lock(&rt_constraints_mutex
);
7442 read_lock(&tasklist_lock
);
7443 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7447 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7448 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7449 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7451 for_each_possible_cpu(i
) {
7452 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7454 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7455 rt_rq
->rt_runtime
= rt_runtime
;
7456 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7458 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7460 read_unlock(&tasklist_lock
);
7461 mutex_unlock(&rt_constraints_mutex
);
7466 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7468 u64 rt_runtime
, rt_period
;
7470 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7471 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7472 if (rt_runtime_us
< 0)
7473 rt_runtime
= RUNTIME_INF
;
7475 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7478 long sched_group_rt_runtime(struct task_group
*tg
)
7482 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7485 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7486 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7487 return rt_runtime_us
;
7490 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7492 u64 rt_runtime
, rt_period
;
7494 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7495 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7500 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7503 long sched_group_rt_period(struct task_group
*tg
)
7507 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7508 do_div(rt_period_us
, NSEC_PER_USEC
);
7509 return rt_period_us
;
7512 static int sched_rt_global_constraints(void)
7514 u64 runtime
, period
;
7517 if (sysctl_sched_rt_period
<= 0)
7520 runtime
= global_rt_runtime();
7521 period
= global_rt_period();
7524 * Sanity check on the sysctl variables.
7526 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7529 mutex_lock(&rt_constraints_mutex
);
7530 read_lock(&tasklist_lock
);
7531 ret
= __rt_schedulable(NULL
, 0, 0);
7532 read_unlock(&tasklist_lock
);
7533 mutex_unlock(&rt_constraints_mutex
);
7538 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7540 /* Don't accept realtime tasks when there is no way for them to run */
7541 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7547 #else /* !CONFIG_RT_GROUP_SCHED */
7548 static int sched_rt_global_constraints(void)
7550 unsigned long flags
;
7553 if (sysctl_sched_rt_period
<= 0)
7557 * There's always some RT tasks in the root group
7558 * -- migration, kstopmachine etc..
7560 if (sysctl_sched_rt_runtime
== 0)
7563 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7564 for_each_possible_cpu(i
) {
7565 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7567 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7568 rt_rq
->rt_runtime
= global_rt_runtime();
7569 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7571 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7575 #endif /* CONFIG_RT_GROUP_SCHED */
7577 int sched_rt_handler(struct ctl_table
*table
, int write
,
7578 void __user
*buffer
, size_t *lenp
,
7582 int old_period
, old_runtime
;
7583 static DEFINE_MUTEX(mutex
);
7586 old_period
= sysctl_sched_rt_period
;
7587 old_runtime
= sysctl_sched_rt_runtime
;
7589 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7591 if (!ret
&& write
) {
7592 ret
= sched_rt_global_constraints();
7594 sysctl_sched_rt_period
= old_period
;
7595 sysctl_sched_rt_runtime
= old_runtime
;
7597 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7598 def_rt_bandwidth
.rt_period
=
7599 ns_to_ktime(global_rt_period());
7602 mutex_unlock(&mutex
);
7607 #ifdef CONFIG_CGROUP_SCHED
7609 /* return corresponding task_group object of a cgroup */
7610 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7612 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7613 struct task_group
, css
);
7616 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7618 struct task_group
*tg
, *parent
;
7620 if (!cgrp
->parent
) {
7621 /* This is early initialization for the top cgroup */
7622 return &root_task_group
.css
;
7625 parent
= cgroup_tg(cgrp
->parent
);
7626 tg
= sched_create_group(parent
);
7628 return ERR_PTR(-ENOMEM
);
7633 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7635 struct task_group
*tg
= cgroup_tg(cgrp
);
7637 sched_destroy_group(tg
);
7640 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7641 struct cgroup_taskset
*tset
)
7643 struct task_struct
*task
;
7645 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7646 #ifdef CONFIG_RT_GROUP_SCHED
7647 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7650 /* We don't support RT-tasks being in separate groups */
7651 if (task
->sched_class
!= &fair_sched_class
)
7658 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7659 struct cgroup_taskset
*tset
)
7661 struct task_struct
*task
;
7663 cgroup_taskset_for_each(task
, cgrp
, tset
)
7664 sched_move_task(task
);
7668 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7669 struct task_struct
*task
)
7672 * cgroup_exit() is called in the copy_process() failure path.
7673 * Ignore this case since the task hasn't ran yet, this avoids
7674 * trying to poke a half freed task state from generic code.
7676 if (!(task
->flags
& PF_EXITING
))
7679 sched_move_task(task
);
7682 #ifdef CONFIG_FAIR_GROUP_SCHED
7683 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7686 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7689 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7691 struct task_group
*tg
= cgroup_tg(cgrp
);
7693 return (u64
) scale_load_down(tg
->shares
);
7696 #ifdef CONFIG_CFS_BANDWIDTH
7697 static DEFINE_MUTEX(cfs_constraints_mutex
);
7699 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7700 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7702 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7704 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7706 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7707 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7709 if (tg
== &root_task_group
)
7713 * Ensure we have at some amount of bandwidth every period. This is
7714 * to prevent reaching a state of large arrears when throttled via
7715 * entity_tick() resulting in prolonged exit starvation.
7717 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7721 * Likewise, bound things on the otherside by preventing insane quota
7722 * periods. This also allows us to normalize in computing quota
7725 if (period
> max_cfs_quota_period
)
7728 mutex_lock(&cfs_constraints_mutex
);
7729 ret
= __cfs_schedulable(tg
, period
, quota
);
7733 runtime_enabled
= quota
!= RUNTIME_INF
;
7734 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7735 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7736 raw_spin_lock_irq(&cfs_b
->lock
);
7737 cfs_b
->period
= ns_to_ktime(period
);
7738 cfs_b
->quota
= quota
;
7740 __refill_cfs_bandwidth_runtime(cfs_b
);
7741 /* restart the period timer (if active) to handle new period expiry */
7742 if (runtime_enabled
&& cfs_b
->timer_active
) {
7743 /* force a reprogram */
7744 cfs_b
->timer_active
= 0;
7745 __start_cfs_bandwidth(cfs_b
);
7747 raw_spin_unlock_irq(&cfs_b
->lock
);
7749 for_each_possible_cpu(i
) {
7750 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7751 struct rq
*rq
= cfs_rq
->rq
;
7753 raw_spin_lock_irq(&rq
->lock
);
7754 cfs_rq
->runtime_enabled
= runtime_enabled
;
7755 cfs_rq
->runtime_remaining
= 0;
7757 if (cfs_rq
->throttled
)
7758 unthrottle_cfs_rq(cfs_rq
);
7759 raw_spin_unlock_irq(&rq
->lock
);
7762 mutex_unlock(&cfs_constraints_mutex
);
7767 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7771 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7772 if (cfs_quota_us
< 0)
7773 quota
= RUNTIME_INF
;
7775 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7777 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7780 long tg_get_cfs_quota(struct task_group
*tg
)
7784 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7787 quota_us
= tg
->cfs_bandwidth
.quota
;
7788 do_div(quota_us
, NSEC_PER_USEC
);
7793 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7797 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7798 quota
= tg
->cfs_bandwidth
.quota
;
7800 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7803 long tg_get_cfs_period(struct task_group
*tg
)
7807 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7808 do_div(cfs_period_us
, NSEC_PER_USEC
);
7810 return cfs_period_us
;
7813 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7815 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7818 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7821 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7824 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7826 return tg_get_cfs_period(cgroup_tg(cgrp
));
7829 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7832 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7835 struct cfs_schedulable_data
{
7836 struct task_group
*tg
;
7841 * normalize group quota/period to be quota/max_period
7842 * note: units are usecs
7844 static u64
normalize_cfs_quota(struct task_group
*tg
,
7845 struct cfs_schedulable_data
*d
)
7853 period
= tg_get_cfs_period(tg
);
7854 quota
= tg_get_cfs_quota(tg
);
7857 /* note: these should typically be equivalent */
7858 if (quota
== RUNTIME_INF
|| quota
== -1)
7861 return to_ratio(period
, quota
);
7864 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7866 struct cfs_schedulable_data
*d
= data
;
7867 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7868 s64 quota
= 0, parent_quota
= -1;
7871 quota
= RUNTIME_INF
;
7873 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7875 quota
= normalize_cfs_quota(tg
, d
);
7876 parent_quota
= parent_b
->hierarchal_quota
;
7879 * ensure max(child_quota) <= parent_quota, inherit when no
7882 if (quota
== RUNTIME_INF
)
7883 quota
= parent_quota
;
7884 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7887 cfs_b
->hierarchal_quota
= quota
;
7892 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7895 struct cfs_schedulable_data data
= {
7901 if (quota
!= RUNTIME_INF
) {
7902 do_div(data
.period
, NSEC_PER_USEC
);
7903 do_div(data
.quota
, NSEC_PER_USEC
);
7907 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7913 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7914 struct cgroup_map_cb
*cb
)
7916 struct task_group
*tg
= cgroup_tg(cgrp
);
7917 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7919 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7920 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7921 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7925 #endif /* CONFIG_CFS_BANDWIDTH */
7926 #endif /* CONFIG_FAIR_GROUP_SCHED */
7928 #ifdef CONFIG_RT_GROUP_SCHED
7929 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7932 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7935 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7937 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7940 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7943 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7946 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7948 return sched_group_rt_period(cgroup_tg(cgrp
));
7950 #endif /* CONFIG_RT_GROUP_SCHED */
7952 static struct cftype cpu_files
[] = {
7953 #ifdef CONFIG_FAIR_GROUP_SCHED
7956 .read_u64
= cpu_shares_read_u64
,
7957 .write_u64
= cpu_shares_write_u64
,
7960 #ifdef CONFIG_CFS_BANDWIDTH
7962 .name
= "cfs_quota_us",
7963 .read_s64
= cpu_cfs_quota_read_s64
,
7964 .write_s64
= cpu_cfs_quota_write_s64
,
7967 .name
= "cfs_period_us",
7968 .read_u64
= cpu_cfs_period_read_u64
,
7969 .write_u64
= cpu_cfs_period_write_u64
,
7973 .read_map
= cpu_stats_show
,
7976 #ifdef CONFIG_RT_GROUP_SCHED
7978 .name
= "rt_runtime_us",
7979 .read_s64
= cpu_rt_runtime_read
,
7980 .write_s64
= cpu_rt_runtime_write
,
7983 .name
= "rt_period_us",
7984 .read_u64
= cpu_rt_period_read_uint
,
7985 .write_u64
= cpu_rt_period_write_uint
,
7991 struct cgroup_subsys cpu_cgroup_subsys
= {
7993 .create
= cpu_cgroup_create
,
7994 .destroy
= cpu_cgroup_destroy
,
7995 .can_attach
= cpu_cgroup_can_attach
,
7996 .attach
= cpu_cgroup_attach
,
7997 .exit
= cpu_cgroup_exit
,
7998 .subsys_id
= cpu_cgroup_subsys_id
,
7999 .base_cftypes
= cpu_files
,
8003 #endif /* CONFIG_CGROUP_SCHED */
8005 #ifdef CONFIG_CGROUP_CPUACCT
8008 * CPU accounting code for task groups.
8010 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8011 * (balbir@in.ibm.com).
8014 /* create a new cpu accounting group */
8015 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
8020 return &root_cpuacct
.css
;
8022 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8026 ca
->cpuusage
= alloc_percpu(u64
);
8030 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
8032 goto out_free_cpuusage
;
8037 free_percpu(ca
->cpuusage
);
8041 return ERR_PTR(-ENOMEM
);
8044 /* destroy an existing cpu accounting group */
8045 static void cpuacct_destroy(struct cgroup
*cgrp
)
8047 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8049 free_percpu(ca
->cpustat
);
8050 free_percpu(ca
->cpuusage
);
8054 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8056 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8059 #ifndef CONFIG_64BIT
8061 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8063 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8065 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8073 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8075 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8077 #ifndef CONFIG_64BIT
8079 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8081 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8083 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8089 /* return total cpu usage (in nanoseconds) of a group */
8090 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8092 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8093 u64 totalcpuusage
= 0;
8096 for_each_present_cpu(i
)
8097 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8099 return totalcpuusage
;
8102 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8105 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8114 for_each_present_cpu(i
)
8115 cpuacct_cpuusage_write(ca
, i
, 0);
8121 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8124 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8128 for_each_present_cpu(i
) {
8129 percpu
= cpuacct_cpuusage_read(ca
, i
);
8130 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8132 seq_printf(m
, "\n");
8136 static const char *cpuacct_stat_desc
[] = {
8137 [CPUACCT_STAT_USER
] = "user",
8138 [CPUACCT_STAT_SYSTEM
] = "system",
8141 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8142 struct cgroup_map_cb
*cb
)
8144 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8148 for_each_online_cpu(cpu
) {
8149 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8150 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8151 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8153 val
= cputime64_to_clock_t(val
);
8154 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8157 for_each_online_cpu(cpu
) {
8158 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8159 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8160 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8161 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8164 val
= cputime64_to_clock_t(val
);
8165 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8170 static struct cftype files
[] = {
8173 .read_u64
= cpuusage_read
,
8174 .write_u64
= cpuusage_write
,
8177 .name
= "usage_percpu",
8178 .read_seq_string
= cpuacct_percpu_seq_read
,
8182 .read_map
= cpuacct_stats_show
,
8188 * charge this task's execution time to its accounting group.
8190 * called with rq->lock held.
8192 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8197 if (unlikely(!cpuacct_subsys
.active
))
8200 cpu
= task_cpu(tsk
);
8206 for (; ca
; ca
= parent_ca(ca
)) {
8207 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8208 *cpuusage
+= cputime
;
8214 struct cgroup_subsys cpuacct_subsys
= {
8216 .create
= cpuacct_create
,
8217 .destroy
= cpuacct_destroy
,
8218 .subsys_id
= cpuacct_subsys_id
,
8219 .base_cftypes
= files
,
8221 #endif /* CONFIG_CGROUP_CPUACCT */