4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
90 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
93 ktime_t soft
, hard
, now
;
96 if (hrtimer_active(period_timer
))
99 now
= hrtimer_cb_get_time(period_timer
);
100 hrtimer_forward(period_timer
, now
, period
);
102 soft
= hrtimer_get_softexpires(period_timer
);
103 hard
= hrtimer_get_expires(period_timer
);
104 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
105 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
106 HRTIMER_MODE_ABS_PINNED
, 0);
110 DEFINE_MUTEX(sched_domains_mutex
);
111 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
113 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
115 void update_rq_clock(struct rq
*rq
)
119 if (rq
->skip_clock_update
> 0)
122 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
124 update_rq_clock_task(rq
, delta
);
128 * Debugging: various feature bits
131 #define SCHED_FEAT(name, enabled) \
132 (1UL << __SCHED_FEAT_##name) * enabled |
134 const_debug
unsigned int sysctl_sched_features
=
135 #include "features.h"
140 #ifdef CONFIG_SCHED_DEBUG
141 #define SCHED_FEAT(name, enabled) \
144 static __read_mostly
char *sched_feat_names
[] = {
145 #include "features.h"
151 static int sched_feat_show(struct seq_file
*m
, void *v
)
155 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
156 if (!(sysctl_sched_features
& (1UL << i
)))
158 seq_printf(m
, "%s ", sched_feat_names
[i
]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
174 #include "features.h"
179 static void sched_feat_disable(int i
)
181 if (static_key_enabled(&sched_feat_keys
[i
]))
182 static_key_slow_dec(&sched_feat_keys
[i
]);
185 static void sched_feat_enable(int i
)
187 if (!static_key_enabled(&sched_feat_keys
[i
]))
188 static_key_slow_inc(&sched_feat_keys
[i
]);
191 static void sched_feat_disable(int i
) { };
192 static void sched_feat_enable(int i
) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
197 size_t cnt
, loff_t
*ppos
)
207 if (copy_from_user(&buf
, ubuf
, cnt
))
213 if (strncmp(cmp
, "NO_", 3) == 0) {
218 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
219 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
221 sysctl_sched_features
&= ~(1UL << i
);
222 sched_feat_disable(i
);
224 sysctl_sched_features
|= (1UL << i
);
225 sched_feat_enable(i
);
231 if (i
== __SCHED_FEAT_NR
)
239 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
241 return single_open(filp
, sched_feat_show
, NULL
);
244 static const struct file_operations sched_feat_fops
= {
245 .open
= sched_feat_open
,
246 .write
= sched_feat_write
,
249 .release
= single_release
,
252 static __init
int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
259 late_initcall(sched_init_debug
);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
269 * period over which we average the RT time consumption, measured
274 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period
= 1000000;
282 __read_mostly
int scheduler_running
;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime
= 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
300 lockdep_assert_held(&p
->pi_lock
);
304 raw_spin_lock(&rq
->lock
);
305 if (likely(rq
== task_rq(p
)))
307 raw_spin_unlock(&rq
->lock
);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
315 __acquires(p
->pi_lock
)
321 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
323 raw_spin_lock(&rq
->lock
);
324 if (likely(rq
== task_rq(p
)))
326 raw_spin_unlock(&rq
->lock
);
327 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
331 static void __task_rq_unlock(struct rq
*rq
)
334 raw_spin_unlock(&rq
->lock
);
338 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
340 __releases(p
->pi_lock
)
342 raw_spin_unlock(&rq
->lock
);
343 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq
*this_rq_lock(void)
356 raw_spin_lock(&rq
->lock
);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq
*rq
)
375 if (hrtimer_active(&rq
->hrtick_timer
))
376 hrtimer_cancel(&rq
->hrtick_timer
);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
385 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
387 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
389 raw_spin_lock(&rq
->lock
);
391 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
392 raw_spin_unlock(&rq
->lock
);
394 return HRTIMER_NORESTART
;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg
)
405 raw_spin_lock(&rq
->lock
);
406 hrtimer_restart(&rq
->hrtick_timer
);
407 rq
->hrtick_csd_pending
= 0;
408 raw_spin_unlock(&rq
->lock
);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq
*rq
, u64 delay
)
418 struct hrtimer
*timer
= &rq
->hrtick_timer
;
419 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
421 hrtimer_set_expires(timer
, time
);
423 if (rq
== this_rq()) {
424 hrtimer_restart(timer
);
425 } else if (!rq
->hrtick_csd_pending
) {
426 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
427 rq
->hrtick_csd_pending
= 1;
432 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
434 int cpu
= (int)(long)hcpu
;
437 case CPU_UP_CANCELED
:
438 case CPU_UP_CANCELED_FROZEN
:
439 case CPU_DOWN_PREPARE
:
440 case CPU_DOWN_PREPARE_FROZEN
:
442 case CPU_DEAD_FROZEN
:
443 hrtick_clear(cpu_rq(cpu
));
450 static __init
void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick
, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq
*rq
, u64 delay
)
462 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
463 HRTIMER_MODE_REL_PINNED
, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq
*rq
)
474 rq
->hrtick_csd_pending
= 0;
476 rq
->hrtick_csd
.flags
= 0;
477 rq
->hrtick_csd
.func
= __hrtick_start
;
478 rq
->hrtick_csd
.info
= rq
;
481 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
482 rq
->hrtick_timer
.function
= hrtick
;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq
*rq
)
489 static inline void init_rq_hrtick(struct rq
*rq
)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct
*p
)
515 assert_raw_spin_locked(&task_rq(p
)->lock
);
517 if (test_tsk_need_resched(p
))
520 set_tsk_need_resched(p
);
523 if (cpu
== smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p
))
529 smp_send_reschedule(cpu
);
532 void resched_cpu(int cpu
)
534 struct rq
*rq
= cpu_rq(cpu
);
537 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
539 resched_task(cpu_curr(cpu
));
540 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu
= smp_processor_id();
556 struct sched_domain
*sd
;
559 for_each_domain(cpu
, sd
) {
560 for_each_cpu(i
, sched_domain_span(sd
)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu
)
583 struct rq
*rq
= cpu_rq(cpu
);
585 if (cpu
== smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq
->curr
!= rq
->idle
)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq
->idle
);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq
->idle
))
608 smp_send_reschedule(cpu
);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu
= smp_processor_id();
614 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq
*rq
)
628 s64 period
= sched_avg_period();
630 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq
->age_stamp
));
637 rq
->age_stamp
+= period
;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct
*p
)
645 assert_raw_spin_locked(&task_rq(p
)->lock
);
646 set_tsk_need_resched(p
);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group
*from
,
659 tg_visitor down
, tg_visitor up
, void *data
)
661 struct task_group
*parent
, *child
;
667 ret
= (*down
)(parent
, data
);
670 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
677 ret
= (*up
)(parent
, data
);
678 if (ret
|| parent
== from
)
682 parent
= parent
->parent
;
689 int tg_nop(struct task_group
*tg
, void *data
)
695 void update_cpu_load(struct rq
*this_rq
);
697 static void set_load_weight(struct task_struct
*p
)
699 int prio
= p
->static_prio
- MAX_RT_PRIO
;
700 struct load_weight
*load
= &p
->se
.load
;
703 * SCHED_IDLE tasks get minimal weight:
705 if (p
->policy
== SCHED_IDLE
) {
706 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
707 load
->inv_weight
= WMULT_IDLEPRIO
;
711 load
->weight
= scale_load(prio_to_weight
[prio
]);
712 load
->inv_weight
= prio_to_wmult
[prio
];
715 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
718 sched_info_queued(p
);
719 p
->sched_class
->enqueue_task(rq
, p
, flags
);
722 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
725 sched_info_dequeued(p
);
726 p
->sched_class
->dequeue_task(rq
, p
, flags
);
729 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
731 if (task_contributes_to_load(p
))
732 rq
->nr_uninterruptible
--;
734 enqueue_task(rq
, p
, flags
);
737 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
739 if (task_contributes_to_load(p
))
740 rq
->nr_uninterruptible
++;
742 dequeue_task(rq
, p
, flags
);
745 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
748 * There are no locks covering percpu hardirq/softirq time.
749 * They are only modified in account_system_vtime, on corresponding CPU
750 * with interrupts disabled. So, writes are safe.
751 * They are read and saved off onto struct rq in update_rq_clock().
752 * This may result in other CPU reading this CPU's irq time and can
753 * race with irq/account_system_vtime on this CPU. We would either get old
754 * or new value with a side effect of accounting a slice of irq time to wrong
755 * task when irq is in progress while we read rq->clock. That is a worthy
756 * compromise in place of having locks on each irq in account_system_time.
758 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
759 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
761 static DEFINE_PER_CPU(u64
, irq_start_time
);
762 static int sched_clock_irqtime
;
764 void enable_sched_clock_irqtime(void)
766 sched_clock_irqtime
= 1;
769 void disable_sched_clock_irqtime(void)
771 sched_clock_irqtime
= 0;
775 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
777 static inline void irq_time_write_begin(void)
779 __this_cpu_inc(irq_time_seq
.sequence
);
783 static inline void irq_time_write_end(void)
786 __this_cpu_inc(irq_time_seq
.sequence
);
789 static inline u64
irq_time_read(int cpu
)
795 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
796 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
797 per_cpu(cpu_hardirq_time
, cpu
);
798 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
802 #else /* CONFIG_64BIT */
803 static inline void irq_time_write_begin(void)
807 static inline void irq_time_write_end(void)
811 static inline u64
irq_time_read(int cpu
)
813 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
815 #endif /* CONFIG_64BIT */
818 * Called before incrementing preempt_count on {soft,}irq_enter
819 * and before decrementing preempt_count on {soft,}irq_exit.
821 void account_system_vtime(struct task_struct
*curr
)
827 if (!sched_clock_irqtime
)
830 local_irq_save(flags
);
832 cpu
= smp_processor_id();
833 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
834 __this_cpu_add(irq_start_time
, delta
);
836 irq_time_write_begin();
838 * We do not account for softirq time from ksoftirqd here.
839 * We want to continue accounting softirq time to ksoftirqd thread
840 * in that case, so as not to confuse scheduler with a special task
841 * that do not consume any time, but still wants to run.
844 __this_cpu_add(cpu_hardirq_time
, delta
);
845 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
846 __this_cpu_add(cpu_softirq_time
, delta
);
848 irq_time_write_end();
849 local_irq_restore(flags
);
851 EXPORT_SYMBOL_GPL(account_system_vtime
);
853 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
855 #ifdef CONFIG_PARAVIRT
856 static inline u64
steal_ticks(u64 steal
)
858 if (unlikely(steal
> NSEC_PER_SEC
))
859 return div_u64(steal
, TICK_NSEC
);
861 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
865 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
868 * In theory, the compile should just see 0 here, and optimize out the call
869 * to sched_rt_avg_update. But I don't trust it...
871 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
872 s64 steal
= 0, irq_delta
= 0;
874 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
875 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
878 * Since irq_time is only updated on {soft,}irq_exit, we might run into
879 * this case when a previous update_rq_clock() happened inside a
882 * When this happens, we stop ->clock_task and only update the
883 * prev_irq_time stamp to account for the part that fit, so that a next
884 * update will consume the rest. This ensures ->clock_task is
887 * It does however cause some slight miss-attribution of {soft,}irq
888 * time, a more accurate solution would be to update the irq_time using
889 * the current rq->clock timestamp, except that would require using
892 if (irq_delta
> delta
)
895 rq
->prev_irq_time
+= irq_delta
;
898 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
899 if (static_key_false((¶virt_steal_rq_enabled
))) {
902 steal
= paravirt_steal_clock(cpu_of(rq
));
903 steal
-= rq
->prev_steal_time_rq
;
905 if (unlikely(steal
> delta
))
908 st
= steal_ticks(steal
);
909 steal
= st
* TICK_NSEC
;
911 rq
->prev_steal_time_rq
+= steal
;
917 rq
->clock_task
+= delta
;
919 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
920 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
921 sched_rt_avg_update(rq
, irq_delta
+ steal
);
925 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
926 static int irqtime_account_hi_update(void)
928 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
933 local_irq_save(flags
);
934 latest_ns
= this_cpu_read(cpu_hardirq_time
);
935 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_IRQ
])
937 local_irq_restore(flags
);
941 static int irqtime_account_si_update(void)
943 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
948 local_irq_save(flags
);
949 latest_ns
= this_cpu_read(cpu_softirq_time
);
950 if (nsecs_to_cputime64(latest_ns
) > cpustat
[CPUTIME_SOFTIRQ
])
952 local_irq_restore(flags
);
956 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
958 #define sched_clock_irqtime (0)
962 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
964 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
965 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
969 * Make it appear like a SCHED_FIFO task, its something
970 * userspace knows about and won't get confused about.
972 * Also, it will make PI more or less work without too
973 * much confusion -- but then, stop work should not
974 * rely on PI working anyway.
976 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
978 stop
->sched_class
= &stop_sched_class
;
981 cpu_rq(cpu
)->stop
= stop
;
985 * Reset it back to a normal scheduling class so that
986 * it can die in pieces.
988 old_stop
->sched_class
= &rt_sched_class
;
993 * __normal_prio - return the priority that is based on the static prio
995 static inline int __normal_prio(struct task_struct
*p
)
997 return p
->static_prio
;
1001 * Calculate the expected normal priority: i.e. priority
1002 * without taking RT-inheritance into account. Might be
1003 * boosted by interactivity modifiers. Changes upon fork,
1004 * setprio syscalls, and whenever the interactivity
1005 * estimator recalculates.
1007 static inline int normal_prio(struct task_struct
*p
)
1011 if (task_has_rt_policy(p
))
1012 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1014 prio
= __normal_prio(p
);
1019 * Calculate the current priority, i.e. the priority
1020 * taken into account by the scheduler. This value might
1021 * be boosted by RT tasks, or might be boosted by
1022 * interactivity modifiers. Will be RT if the task got
1023 * RT-boosted. If not then it returns p->normal_prio.
1025 static int effective_prio(struct task_struct
*p
)
1027 p
->normal_prio
= normal_prio(p
);
1029 * If we are RT tasks or we were boosted to RT priority,
1030 * keep the priority unchanged. Otherwise, update priority
1031 * to the normal priority:
1033 if (!rt_prio(p
->prio
))
1034 return p
->normal_prio
;
1039 * task_curr - is this task currently executing on a CPU?
1040 * @p: the task in question.
1042 inline int task_curr(const struct task_struct
*p
)
1044 return cpu_curr(task_cpu(p
)) == p
;
1047 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1048 const struct sched_class
*prev_class
,
1051 if (prev_class
!= p
->sched_class
) {
1052 if (prev_class
->switched_from
)
1053 prev_class
->switched_from(rq
, p
);
1054 p
->sched_class
->switched_to(rq
, p
);
1055 } else if (oldprio
!= p
->prio
)
1056 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1059 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1061 const struct sched_class
*class;
1063 if (p
->sched_class
== rq
->curr
->sched_class
) {
1064 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1066 for_each_class(class) {
1067 if (class == rq
->curr
->sched_class
)
1069 if (class == p
->sched_class
) {
1070 resched_task(rq
->curr
);
1077 * A queue event has occurred, and we're going to schedule. In
1078 * this case, we can save a useless back to back clock update.
1080 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1081 rq
->skip_clock_update
= 1;
1085 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1087 #ifdef CONFIG_SCHED_DEBUG
1089 * We should never call set_task_cpu() on a blocked task,
1090 * ttwu() will sort out the placement.
1092 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1093 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
1095 #ifdef CONFIG_LOCKDEP
1097 * The caller should hold either p->pi_lock or rq->lock, when changing
1098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1100 * sched_move_task() holds both and thus holding either pins the cgroup,
1101 * see set_task_rq().
1103 * Furthermore, all task_rq users should acquire both locks, see
1106 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1107 lockdep_is_held(&task_rq(p
)->lock
)));
1111 trace_sched_migrate_task(p
, new_cpu
);
1113 if (task_cpu(p
) != new_cpu
) {
1114 p
->se
.nr_migrations
++;
1115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1118 __set_task_cpu(p
, new_cpu
);
1121 struct migration_arg
{
1122 struct task_struct
*task
;
1126 static int migration_cpu_stop(void *data
);
1129 * wait_task_inactive - wait for a thread to unschedule.
1131 * If @match_state is nonzero, it's the @p->state value just checked and
1132 * not expected to change. If it changes, i.e. @p might have woken up,
1133 * then return zero. When we succeed in waiting for @p to be off its CPU,
1134 * we return a positive number (its total switch count). If a second call
1135 * a short while later returns the same number, the caller can be sure that
1136 * @p has remained unscheduled the whole time.
1138 * The caller must ensure that the task *will* unschedule sometime soon,
1139 * else this function might spin for a *long* time. This function can't
1140 * be called with interrupts off, or it may introduce deadlock with
1141 * smp_call_function() if an IPI is sent by the same process we are
1142 * waiting to become inactive.
1144 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1146 unsigned long flags
;
1153 * We do the initial early heuristics without holding
1154 * any task-queue locks at all. We'll only try to get
1155 * the runqueue lock when things look like they will
1161 * If the task is actively running on another CPU
1162 * still, just relax and busy-wait without holding
1165 * NOTE! Since we don't hold any locks, it's not
1166 * even sure that "rq" stays as the right runqueue!
1167 * But we don't care, since "task_running()" will
1168 * return false if the runqueue has changed and p
1169 * is actually now running somewhere else!
1171 while (task_running(rq
, p
)) {
1172 if (match_state
&& unlikely(p
->state
!= match_state
))
1178 * Ok, time to look more closely! We need the rq
1179 * lock now, to be *sure*. If we're wrong, we'll
1180 * just go back and repeat.
1182 rq
= task_rq_lock(p
, &flags
);
1183 trace_sched_wait_task(p
);
1184 running
= task_running(rq
, p
);
1187 if (!match_state
|| p
->state
== match_state
)
1188 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1189 task_rq_unlock(rq
, p
, &flags
);
1192 * If it changed from the expected state, bail out now.
1194 if (unlikely(!ncsw
))
1198 * Was it really running after all now that we
1199 * checked with the proper locks actually held?
1201 * Oops. Go back and try again..
1203 if (unlikely(running
)) {
1209 * It's not enough that it's not actively running,
1210 * it must be off the runqueue _entirely_, and not
1213 * So if it was still runnable (but just not actively
1214 * running right now), it's preempted, and we should
1215 * yield - it could be a while.
1217 if (unlikely(on_rq
)) {
1218 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1220 set_current_state(TASK_UNINTERRUPTIBLE
);
1221 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1226 * Ahh, all good. It wasn't running, and it wasn't
1227 * runnable, which means that it will never become
1228 * running in the future either. We're all done!
1237 * kick_process - kick a running thread to enter/exit the kernel
1238 * @p: the to-be-kicked thread
1240 * Cause a process which is running on another CPU to enter
1241 * kernel-mode, without any delay. (to get signals handled.)
1243 * NOTE: this function doesn't have to take the runqueue lock,
1244 * because all it wants to ensure is that the remote task enters
1245 * the kernel. If the IPI races and the task has been migrated
1246 * to another CPU then no harm is done and the purpose has been
1249 void kick_process(struct task_struct
*p
)
1255 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1256 smp_send_reschedule(cpu
);
1259 EXPORT_SYMBOL_GPL(kick_process
);
1260 #endif /* CONFIG_SMP */
1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1266 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1268 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1269 enum { cpuset
, possible
, fail
} state
= cpuset
;
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu(dest_cpu
, nodemask
) {
1274 if (!cpu_online(dest_cpu
))
1276 if (!cpu_active(dest_cpu
))
1278 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1283 /* Any allowed, online CPU? */
1284 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1285 if (!cpu_online(dest_cpu
))
1287 if (!cpu_active(dest_cpu
))
1294 /* No more Mr. Nice Guy. */
1295 cpuset_cpus_allowed_fallback(p
);
1300 do_set_cpus_allowed(p
, cpu_possible_mask
);
1311 if (state
!= cpuset
) {
1313 * Don't tell them about moving exiting tasks or
1314 * kernel threads (both mm NULL), since they never
1317 if (p
->mm
&& printk_ratelimit()) {
1318 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1319 task_pid_nr(p
), p
->comm
, cpu
);
1327 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1330 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1332 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1335 * In order not to call set_task_cpu() on a blocking task we need
1336 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1339 * Since this is common to all placement strategies, this lives here.
1341 * [ this allows ->select_task() to simply return task_cpu(p) and
1342 * not worry about this generic constraint ]
1344 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1346 cpu
= select_fallback_rq(task_cpu(p
), p
);
1351 static void update_avg(u64
*avg
, u64 sample
)
1353 s64 diff
= sample
- *avg
;
1359 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1361 #ifdef CONFIG_SCHEDSTATS
1362 struct rq
*rq
= this_rq();
1365 int this_cpu
= smp_processor_id();
1367 if (cpu
== this_cpu
) {
1368 schedstat_inc(rq
, ttwu_local
);
1369 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1371 struct sched_domain
*sd
;
1373 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1375 for_each_domain(this_cpu
, sd
) {
1376 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1377 schedstat_inc(sd
, ttwu_wake_remote
);
1384 if (wake_flags
& WF_MIGRATED
)
1385 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1387 #endif /* CONFIG_SMP */
1389 schedstat_inc(rq
, ttwu_count
);
1390 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1392 if (wake_flags
& WF_SYNC
)
1393 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1395 #endif /* CONFIG_SCHEDSTATS */
1398 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1400 activate_task(rq
, p
, en_flags
);
1403 /* if a worker is waking up, notify workqueue */
1404 if (p
->flags
& PF_WQ_WORKER
)
1405 wq_worker_waking_up(p
, cpu_of(rq
));
1409 * Mark the task runnable and perform wakeup-preemption.
1412 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1414 trace_sched_wakeup(p
, true);
1415 check_preempt_curr(rq
, p
, wake_flags
);
1417 p
->state
= TASK_RUNNING
;
1419 if (p
->sched_class
->task_woken
)
1420 p
->sched_class
->task_woken(rq
, p
);
1422 if (rq
->idle_stamp
) {
1423 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1424 u64 max
= 2*sysctl_sched_migration_cost
;
1429 update_avg(&rq
->avg_idle
, delta
);
1436 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1439 if (p
->sched_contributes_to_load
)
1440 rq
->nr_uninterruptible
--;
1443 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1444 ttwu_do_wakeup(rq
, p
, wake_flags
);
1448 * Called in case the task @p isn't fully descheduled from its runqueue,
1449 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1450 * since all we need to do is flip p->state to TASK_RUNNING, since
1451 * the task is still ->on_rq.
1453 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1458 rq
= __task_rq_lock(p
);
1460 ttwu_do_wakeup(rq
, p
, wake_flags
);
1463 __task_rq_unlock(rq
);
1469 static void sched_ttwu_pending(void)
1471 struct rq
*rq
= this_rq();
1472 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1473 struct task_struct
*p
;
1475 raw_spin_lock(&rq
->lock
);
1478 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1479 llist
= llist_next(llist
);
1480 ttwu_do_activate(rq
, p
, 0);
1483 raw_spin_unlock(&rq
->lock
);
1486 void scheduler_ipi(void)
1488 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1492 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1493 * traditionally all their work was done from the interrupt return
1494 * path. Now that we actually do some work, we need to make sure
1497 * Some archs already do call them, luckily irq_enter/exit nest
1500 * Arguably we should visit all archs and update all handlers,
1501 * however a fair share of IPIs are still resched only so this would
1502 * somewhat pessimize the simple resched case.
1505 sched_ttwu_pending();
1508 * Check if someone kicked us for doing the nohz idle load balance.
1510 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1511 this_rq()->idle_balance
= 1;
1512 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1517 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1519 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1520 smp_send_reschedule(cpu
);
1523 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1524 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1529 rq
= __task_rq_lock(p
);
1531 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1532 ttwu_do_wakeup(rq
, p
, wake_flags
);
1535 __task_rq_unlock(rq
);
1540 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1542 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1544 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1546 #endif /* CONFIG_SMP */
1548 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1550 struct rq
*rq
= cpu_rq(cpu
);
1552 #if defined(CONFIG_SMP)
1553 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1554 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1555 ttwu_queue_remote(p
, cpu
);
1560 raw_spin_lock(&rq
->lock
);
1561 ttwu_do_activate(rq
, p
, 0);
1562 raw_spin_unlock(&rq
->lock
);
1566 * try_to_wake_up - wake up a thread
1567 * @p: the thread to be awakened
1568 * @state: the mask of task states that can be woken
1569 * @wake_flags: wake modifier flags (WF_*)
1571 * Put it on the run-queue if it's not already there. The "current"
1572 * thread is always on the run-queue (except when the actual
1573 * re-schedule is in progress), and as such you're allowed to do
1574 * the simpler "current->state = TASK_RUNNING" to mark yourself
1575 * runnable without the overhead of this.
1577 * Returns %true if @p was woken up, %false if it was already running
1578 * or @state didn't match @p's state.
1581 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1583 unsigned long flags
;
1584 int cpu
, success
= 0;
1587 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1588 if (!(p
->state
& state
))
1591 success
= 1; /* we're going to change ->state */
1594 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1599 * If the owning (remote) cpu is still in the middle of schedule() with
1600 * this task as prev, wait until its done referencing the task.
1603 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1605 * In case the architecture enables interrupts in
1606 * context_switch(), we cannot busy wait, since that
1607 * would lead to deadlocks when an interrupt hits and
1608 * tries to wake up @prev. So bail and do a complete
1611 if (ttwu_activate_remote(p
, wake_flags
))
1618 * Pairs with the smp_wmb() in finish_lock_switch().
1622 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1623 p
->state
= TASK_WAKING
;
1625 if (p
->sched_class
->task_waking
)
1626 p
->sched_class
->task_waking(p
);
1628 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1629 if (task_cpu(p
) != cpu
) {
1630 wake_flags
|= WF_MIGRATED
;
1631 set_task_cpu(p
, cpu
);
1633 #endif /* CONFIG_SMP */
1637 ttwu_stat(p
, cpu
, wake_flags
);
1639 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1645 * try_to_wake_up_local - try to wake up a local task with rq lock held
1646 * @p: the thread to be awakened
1648 * Put @p on the run-queue if it's not already there. The caller must
1649 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1652 static void try_to_wake_up_local(struct task_struct
*p
)
1654 struct rq
*rq
= task_rq(p
);
1656 BUG_ON(rq
!= this_rq());
1657 BUG_ON(p
== current
);
1658 lockdep_assert_held(&rq
->lock
);
1660 if (!raw_spin_trylock(&p
->pi_lock
)) {
1661 raw_spin_unlock(&rq
->lock
);
1662 raw_spin_lock(&p
->pi_lock
);
1663 raw_spin_lock(&rq
->lock
);
1666 if (!(p
->state
& TASK_NORMAL
))
1670 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1672 ttwu_do_wakeup(rq
, p
, 0);
1673 ttwu_stat(p
, smp_processor_id(), 0);
1675 raw_spin_unlock(&p
->pi_lock
);
1679 * wake_up_process - Wake up a specific process
1680 * @p: The process to be woken up.
1682 * Attempt to wake up the nominated process and move it to the set of runnable
1683 * processes. Returns 1 if the process was woken up, 0 if it was already
1686 * It may be assumed that this function implies a write memory barrier before
1687 * changing the task state if and only if any tasks are woken up.
1689 int wake_up_process(struct task_struct
*p
)
1691 return try_to_wake_up(p
, TASK_ALL
, 0);
1693 EXPORT_SYMBOL(wake_up_process
);
1695 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1697 return try_to_wake_up(p
, state
, 0);
1701 * Perform scheduler related setup for a newly forked process p.
1702 * p is forked by current.
1704 * __sched_fork() is basic setup used by init_idle() too:
1706 static void __sched_fork(struct task_struct
*p
)
1711 p
->se
.exec_start
= 0;
1712 p
->se
.sum_exec_runtime
= 0;
1713 p
->se
.prev_sum_exec_runtime
= 0;
1714 p
->se
.nr_migrations
= 0;
1716 INIT_LIST_HEAD(&p
->se
.group_node
);
1718 #ifdef CONFIG_SCHEDSTATS
1719 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1722 INIT_LIST_HEAD(&p
->rt
.run_list
);
1724 #ifdef CONFIG_PREEMPT_NOTIFIERS
1725 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1730 * fork()/clone()-time setup:
1732 void sched_fork(struct task_struct
*p
)
1734 unsigned long flags
;
1735 int cpu
= get_cpu();
1739 * We mark the process as running here. This guarantees that
1740 * nobody will actually run it, and a signal or other external
1741 * event cannot wake it up and insert it on the runqueue either.
1743 p
->state
= TASK_RUNNING
;
1746 * Make sure we do not leak PI boosting priority to the child.
1748 p
->prio
= current
->normal_prio
;
1751 * Revert to default priority/policy on fork if requested.
1753 if (unlikely(p
->sched_reset_on_fork
)) {
1754 if (task_has_rt_policy(p
)) {
1755 p
->policy
= SCHED_NORMAL
;
1756 p
->static_prio
= NICE_TO_PRIO(0);
1758 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1759 p
->static_prio
= NICE_TO_PRIO(0);
1761 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1765 * We don't need the reset flag anymore after the fork. It has
1766 * fulfilled its duty:
1768 p
->sched_reset_on_fork
= 0;
1771 if (!rt_prio(p
->prio
))
1772 p
->sched_class
= &fair_sched_class
;
1774 if (p
->sched_class
->task_fork
)
1775 p
->sched_class
->task_fork(p
);
1778 * The child is not yet in the pid-hash so no cgroup attach races,
1779 * and the cgroup is pinned to this child due to cgroup_fork()
1780 * is ran before sched_fork().
1782 * Silence PROVE_RCU.
1784 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1785 set_task_cpu(p
, cpu
);
1786 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1788 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1789 if (likely(sched_info_on()))
1790 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1792 #if defined(CONFIG_SMP)
1795 #ifdef CONFIG_PREEMPT_COUNT
1796 /* Want to start with kernel preemption disabled. */
1797 task_thread_info(p
)->preempt_count
= 1;
1800 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1807 * wake_up_new_task - wake up a newly created task for the first time.
1809 * This function will do some initial scheduler statistics housekeeping
1810 * that must be done for every newly created context, then puts the task
1811 * on the runqueue and wakes it.
1813 void wake_up_new_task(struct task_struct
*p
)
1815 unsigned long flags
;
1818 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1821 * Fork balancing, do it here and not earlier because:
1822 * - cpus_allowed can change in the fork path
1823 * - any previously selected cpu might disappear through hotplug
1825 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1828 rq
= __task_rq_lock(p
);
1829 activate_task(rq
, p
, 0);
1831 trace_sched_wakeup_new(p
, true);
1832 check_preempt_curr(rq
, p
, WF_FORK
);
1834 if (p
->sched_class
->task_woken
)
1835 p
->sched_class
->task_woken(rq
, p
);
1837 task_rq_unlock(rq
, p
, &flags
);
1840 #ifdef CONFIG_PREEMPT_NOTIFIERS
1843 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1844 * @notifier: notifier struct to register
1846 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1848 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1850 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1853 * preempt_notifier_unregister - no longer interested in preemption notifications
1854 * @notifier: notifier struct to unregister
1856 * This is safe to call from within a preemption notifier.
1858 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1860 hlist_del(¬ifier
->link
);
1862 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1864 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1866 struct preempt_notifier
*notifier
;
1867 struct hlist_node
*node
;
1869 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1870 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1874 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1875 struct task_struct
*next
)
1877 struct preempt_notifier
*notifier
;
1878 struct hlist_node
*node
;
1880 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1881 notifier
->ops
->sched_out(notifier
, next
);
1884 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1886 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1891 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1892 struct task_struct
*next
)
1896 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1899 * prepare_task_switch - prepare to switch tasks
1900 * @rq: the runqueue preparing to switch
1901 * @prev: the current task that is being switched out
1902 * @next: the task we are going to switch to.
1904 * This is called with the rq lock held and interrupts off. It must
1905 * be paired with a subsequent finish_task_switch after the context
1908 * prepare_task_switch sets up locking and calls architecture specific
1912 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1913 struct task_struct
*next
)
1915 sched_info_switch(prev
, next
);
1916 perf_event_task_sched_out(prev
, next
);
1917 fire_sched_out_preempt_notifiers(prev
, next
);
1918 prepare_lock_switch(rq
, next
);
1919 prepare_arch_switch(next
);
1920 trace_sched_switch(prev
, next
);
1924 * finish_task_switch - clean up after a task-switch
1925 * @rq: runqueue associated with task-switch
1926 * @prev: the thread we just switched away from.
1928 * finish_task_switch must be called after the context switch, paired
1929 * with a prepare_task_switch call before the context switch.
1930 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1931 * and do any other architecture-specific cleanup actions.
1933 * Note that we may have delayed dropping an mm in context_switch(). If
1934 * so, we finish that here outside of the runqueue lock. (Doing it
1935 * with the lock held can cause deadlocks; see schedule() for
1938 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1939 __releases(rq
->lock
)
1941 struct mm_struct
*mm
= rq
->prev_mm
;
1947 * A task struct has one reference for the use as "current".
1948 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1949 * schedule one last time. The schedule call will never return, and
1950 * the scheduled task must drop that reference.
1951 * The test for TASK_DEAD must occur while the runqueue locks are
1952 * still held, otherwise prev could be scheduled on another cpu, die
1953 * there before we look at prev->state, and then the reference would
1955 * Manfred Spraul <manfred@colorfullife.com>
1957 prev_state
= prev
->state
;
1958 finish_arch_switch(prev
);
1959 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1960 local_irq_disable();
1961 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1962 perf_event_task_sched_in(prev
, current
);
1963 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1965 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1966 finish_lock_switch(rq
, prev
);
1967 finish_arch_post_lock_switch();
1969 fire_sched_in_preempt_notifiers(current
);
1972 if (unlikely(prev_state
== TASK_DEAD
)) {
1974 * Remove function-return probe instances associated with this
1975 * task and put them back on the free list.
1977 kprobe_flush_task(prev
);
1978 put_task_struct(prev
);
1984 /* assumes rq->lock is held */
1985 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1987 if (prev
->sched_class
->pre_schedule
)
1988 prev
->sched_class
->pre_schedule(rq
, prev
);
1991 /* rq->lock is NOT held, but preemption is disabled */
1992 static inline void post_schedule(struct rq
*rq
)
1994 if (rq
->post_schedule
) {
1995 unsigned long flags
;
1997 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1998 if (rq
->curr
->sched_class
->post_schedule
)
1999 rq
->curr
->sched_class
->post_schedule(rq
);
2000 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2002 rq
->post_schedule
= 0;
2008 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2012 static inline void post_schedule(struct rq
*rq
)
2019 * schedule_tail - first thing a freshly forked thread must call.
2020 * @prev: the thread we just switched away from.
2022 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2023 __releases(rq
->lock
)
2025 struct rq
*rq
= this_rq();
2027 finish_task_switch(rq
, prev
);
2030 * FIXME: do we need to worry about rq being invalidated by the
2035 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2036 /* In this case, finish_task_switch does not reenable preemption */
2039 if (current
->set_child_tid
)
2040 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2044 * context_switch - switch to the new MM and the new
2045 * thread's register state.
2048 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2049 struct task_struct
*next
)
2051 struct mm_struct
*mm
, *oldmm
;
2053 prepare_task_switch(rq
, prev
, next
);
2056 oldmm
= prev
->active_mm
;
2058 * For paravirt, this is coupled with an exit in switch_to to
2059 * combine the page table reload and the switch backend into
2062 arch_start_context_switch(prev
);
2065 next
->active_mm
= oldmm
;
2066 atomic_inc(&oldmm
->mm_count
);
2067 enter_lazy_tlb(oldmm
, next
);
2069 switch_mm(oldmm
, mm
, next
);
2072 prev
->active_mm
= NULL
;
2073 rq
->prev_mm
= oldmm
;
2076 * Since the runqueue lock will be released by the next
2077 * task (which is an invalid locking op but in the case
2078 * of the scheduler it's an obvious special-case), so we
2079 * do an early lockdep release here:
2081 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2082 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2085 /* Here we just switch the register state and the stack. */
2086 switch_to(prev
, next
, prev
);
2090 * this_rq must be evaluated again because prev may have moved
2091 * CPUs since it called schedule(), thus the 'rq' on its stack
2092 * frame will be invalid.
2094 finish_task_switch(this_rq(), prev
);
2098 * nr_running, nr_uninterruptible and nr_context_switches:
2100 * externally visible scheduler statistics: current number of runnable
2101 * threads, current number of uninterruptible-sleeping threads, total
2102 * number of context switches performed since bootup.
2104 unsigned long nr_running(void)
2106 unsigned long i
, sum
= 0;
2108 for_each_online_cpu(i
)
2109 sum
+= cpu_rq(i
)->nr_running
;
2114 unsigned long nr_uninterruptible(void)
2116 unsigned long i
, sum
= 0;
2118 for_each_possible_cpu(i
)
2119 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2122 * Since we read the counters lockless, it might be slightly
2123 * inaccurate. Do not allow it to go below zero though:
2125 if (unlikely((long)sum
< 0))
2131 unsigned long long nr_context_switches(void)
2134 unsigned long long sum
= 0;
2136 for_each_possible_cpu(i
)
2137 sum
+= cpu_rq(i
)->nr_switches
;
2142 unsigned long nr_iowait(void)
2144 unsigned long i
, sum
= 0;
2146 for_each_possible_cpu(i
)
2147 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2152 unsigned long nr_iowait_cpu(int cpu
)
2154 struct rq
*this = cpu_rq(cpu
);
2155 return atomic_read(&this->nr_iowait
);
2158 unsigned long this_cpu_load(void)
2160 struct rq
*this = this_rq();
2161 return this->cpu_load
[0];
2165 /* Variables and functions for calc_load */
2166 static atomic_long_t calc_load_tasks
;
2167 static unsigned long calc_load_update
;
2168 unsigned long avenrun
[3];
2169 EXPORT_SYMBOL(avenrun
);
2171 static long calc_load_fold_active(struct rq
*this_rq
)
2173 long nr_active
, delta
= 0;
2175 nr_active
= this_rq
->nr_running
;
2176 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2178 if (nr_active
!= this_rq
->calc_load_active
) {
2179 delta
= nr_active
- this_rq
->calc_load_active
;
2180 this_rq
->calc_load_active
= nr_active
;
2186 static unsigned long
2187 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2190 load
+= active
* (FIXED_1
- exp
);
2191 load
+= 1UL << (FSHIFT
- 1);
2192 return load
>> FSHIFT
;
2197 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2199 * When making the ILB scale, we should try to pull this in as well.
2201 static atomic_long_t calc_load_tasks_idle
;
2203 void calc_load_account_idle(struct rq
*this_rq
)
2207 delta
= calc_load_fold_active(this_rq
);
2209 atomic_long_add(delta
, &calc_load_tasks_idle
);
2212 static long calc_load_fold_idle(void)
2217 * Its got a race, we don't care...
2219 if (atomic_long_read(&calc_load_tasks_idle
))
2220 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2226 * fixed_power_int - compute: x^n, in O(log n) time
2228 * @x: base of the power
2229 * @frac_bits: fractional bits of @x
2230 * @n: power to raise @x to.
2232 * By exploiting the relation between the definition of the natural power
2233 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2234 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2235 * (where: n_i \elem {0, 1}, the binary vector representing n),
2236 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2237 * of course trivially computable in O(log_2 n), the length of our binary
2240 static unsigned long
2241 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2243 unsigned long result
= 1UL << frac_bits
;
2248 result
+= 1UL << (frac_bits
- 1);
2249 result
>>= frac_bits
;
2255 x
+= 1UL << (frac_bits
- 1);
2263 * a1 = a0 * e + a * (1 - e)
2265 * a2 = a1 * e + a * (1 - e)
2266 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2267 * = a0 * e^2 + a * (1 - e) * (1 + e)
2269 * a3 = a2 * e + a * (1 - e)
2270 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2271 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2275 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2276 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2277 * = a0 * e^n + a * (1 - e^n)
2279 * [1] application of the geometric series:
2282 * S_n := \Sum x^i = -------------
2285 static unsigned long
2286 calc_load_n(unsigned long load
, unsigned long exp
,
2287 unsigned long active
, unsigned int n
)
2290 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2294 * NO_HZ can leave us missing all per-cpu ticks calling
2295 * calc_load_account_active(), but since an idle CPU folds its delta into
2296 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2297 * in the pending idle delta if our idle period crossed a load cycle boundary.
2299 * Once we've updated the global active value, we need to apply the exponential
2300 * weights adjusted to the number of cycles missed.
2302 static void calc_global_nohz(void)
2304 long delta
, active
, n
;
2307 * If we crossed a calc_load_update boundary, make sure to fold
2308 * any pending idle changes, the respective CPUs might have
2309 * missed the tick driven calc_load_account_active() update
2312 delta
= calc_load_fold_idle();
2314 atomic_long_add(delta
, &calc_load_tasks
);
2317 * It could be the one fold was all it took, we done!
2319 if (time_before(jiffies
, calc_load_update
+ 10))
2323 * Catch-up, fold however many we are behind still
2325 delta
= jiffies
- calc_load_update
- 10;
2326 n
= 1 + (delta
/ LOAD_FREQ
);
2328 active
= atomic_long_read(&calc_load_tasks
);
2329 active
= active
> 0 ? active
* FIXED_1
: 0;
2331 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2332 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2333 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2335 calc_load_update
+= n
* LOAD_FREQ
;
2338 void calc_load_account_idle(struct rq
*this_rq
)
2342 static inline long calc_load_fold_idle(void)
2347 static void calc_global_nohz(void)
2353 * get_avenrun - get the load average array
2354 * @loads: pointer to dest load array
2355 * @offset: offset to add
2356 * @shift: shift count to shift the result left
2358 * These values are estimates at best, so no need for locking.
2360 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2362 loads
[0] = (avenrun
[0] + offset
) << shift
;
2363 loads
[1] = (avenrun
[1] + offset
) << shift
;
2364 loads
[2] = (avenrun
[2] + offset
) << shift
;
2368 * calc_load - update the avenrun load estimates 10 ticks after the
2369 * CPUs have updated calc_load_tasks.
2371 void calc_global_load(unsigned long ticks
)
2375 if (time_before(jiffies
, calc_load_update
+ 10))
2378 active
= atomic_long_read(&calc_load_tasks
);
2379 active
= active
> 0 ? active
* FIXED_1
: 0;
2381 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2382 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2383 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2385 calc_load_update
+= LOAD_FREQ
;
2388 * Account one period with whatever state we found before
2389 * folding in the nohz state and ageing the entire idle period.
2391 * This avoids loosing a sample when we go idle between
2392 * calc_load_account_active() (10 ticks ago) and now and thus
2399 * Called from update_cpu_load() to periodically update this CPU's
2402 static void calc_load_account_active(struct rq
*this_rq
)
2406 if (time_before(jiffies
, this_rq
->calc_load_update
))
2409 delta
= calc_load_fold_active(this_rq
);
2410 delta
+= calc_load_fold_idle();
2412 atomic_long_add(delta
, &calc_load_tasks
);
2414 this_rq
->calc_load_update
+= LOAD_FREQ
;
2418 * The exact cpuload at various idx values, calculated at every tick would be
2419 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2421 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2422 * on nth tick when cpu may be busy, then we have:
2423 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2424 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2426 * decay_load_missed() below does efficient calculation of
2427 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2428 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2430 * The calculation is approximated on a 128 point scale.
2431 * degrade_zero_ticks is the number of ticks after which load at any
2432 * particular idx is approximated to be zero.
2433 * degrade_factor is a precomputed table, a row for each load idx.
2434 * Each column corresponds to degradation factor for a power of two ticks,
2435 * based on 128 point scale.
2437 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2438 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2440 * With this power of 2 load factors, we can degrade the load n times
2441 * by looking at 1 bits in n and doing as many mult/shift instead of
2442 * n mult/shifts needed by the exact degradation.
2444 #define DEGRADE_SHIFT 7
2445 static const unsigned char
2446 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2447 static const unsigned char
2448 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2449 {0, 0, 0, 0, 0, 0, 0, 0},
2450 {64, 32, 8, 0, 0, 0, 0, 0},
2451 {96, 72, 40, 12, 1, 0, 0},
2452 {112, 98, 75, 43, 15, 1, 0},
2453 {120, 112, 98, 76, 45, 16, 2} };
2456 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2457 * would be when CPU is idle and so we just decay the old load without
2458 * adding any new load.
2460 static unsigned long
2461 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2465 if (!missed_updates
)
2468 if (missed_updates
>= degrade_zero_ticks
[idx
])
2472 return load
>> missed_updates
;
2474 while (missed_updates
) {
2475 if (missed_updates
% 2)
2476 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2478 missed_updates
>>= 1;
2485 * Update rq->cpu_load[] statistics. This function is usually called every
2486 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2487 * every tick. We fix it up based on jiffies.
2489 void update_cpu_load(struct rq
*this_rq
)
2491 unsigned long this_load
= this_rq
->load
.weight
;
2492 unsigned long curr_jiffies
= jiffies
;
2493 unsigned long pending_updates
;
2496 this_rq
->nr_load_updates
++;
2498 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2499 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2502 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2503 this_rq
->last_load_update_tick
= curr_jiffies
;
2505 /* Update our load: */
2506 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2507 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2508 unsigned long old_load
, new_load
;
2510 /* scale is effectively 1 << i now, and >> i divides by scale */
2512 old_load
= this_rq
->cpu_load
[i
];
2513 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2514 new_load
= this_load
;
2516 * Round up the averaging division if load is increasing. This
2517 * prevents us from getting stuck on 9 if the load is 10, for
2520 if (new_load
> old_load
)
2521 new_load
+= scale
- 1;
2523 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2526 sched_avg_update(this_rq
);
2529 static void update_cpu_load_active(struct rq
*this_rq
)
2531 update_cpu_load(this_rq
);
2533 calc_load_account_active(this_rq
);
2539 * sched_exec - execve() is a valuable balancing opportunity, because at
2540 * this point the task has the smallest effective memory and cache footprint.
2542 void sched_exec(void)
2544 struct task_struct
*p
= current
;
2545 unsigned long flags
;
2548 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2549 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2550 if (dest_cpu
== smp_processor_id())
2553 if (likely(cpu_active(dest_cpu
))) {
2554 struct migration_arg arg
= { p
, dest_cpu
};
2556 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2557 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2561 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2566 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2567 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2569 EXPORT_PER_CPU_SYMBOL(kstat
);
2570 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2573 * Return any ns on the sched_clock that have not yet been accounted in
2574 * @p in case that task is currently running.
2576 * Called with task_rq_lock() held on @rq.
2578 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2582 if (task_current(rq
, p
)) {
2583 update_rq_clock(rq
);
2584 ns
= rq
->clock_task
- p
->se
.exec_start
;
2592 unsigned long long task_delta_exec(struct task_struct
*p
)
2594 unsigned long flags
;
2598 rq
= task_rq_lock(p
, &flags
);
2599 ns
= do_task_delta_exec(p
, rq
);
2600 task_rq_unlock(rq
, p
, &flags
);
2606 * Return accounted runtime for the task.
2607 * In case the task is currently running, return the runtime plus current's
2608 * pending runtime that have not been accounted yet.
2610 unsigned long long task_sched_runtime(struct task_struct
*p
)
2612 unsigned long flags
;
2616 rq
= task_rq_lock(p
, &flags
);
2617 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2618 task_rq_unlock(rq
, p
, &flags
);
2623 #ifdef CONFIG_CGROUP_CPUACCT
2624 struct cgroup_subsys cpuacct_subsys
;
2625 struct cpuacct root_cpuacct
;
2628 static inline void task_group_account_field(struct task_struct
*p
, int index
,
2631 #ifdef CONFIG_CGROUP_CPUACCT
2632 struct kernel_cpustat
*kcpustat
;
2636 * Since all updates are sure to touch the root cgroup, we
2637 * get ourselves ahead and touch it first. If the root cgroup
2638 * is the only cgroup, then nothing else should be necessary.
2641 __get_cpu_var(kernel_cpustat
).cpustat
[index
] += tmp
;
2643 #ifdef CONFIG_CGROUP_CPUACCT
2644 if (unlikely(!cpuacct_subsys
.active
))
2649 while (ca
&& (ca
!= &root_cpuacct
)) {
2650 kcpustat
= this_cpu_ptr(ca
->cpustat
);
2651 kcpustat
->cpustat
[index
] += tmp
;
2660 * Account user cpu time to a process.
2661 * @p: the process that the cpu time gets accounted to
2662 * @cputime: the cpu time spent in user space since the last update
2663 * @cputime_scaled: cputime scaled by cpu frequency
2665 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
2666 cputime_t cputime_scaled
)
2670 /* Add user time to process. */
2671 p
->utime
+= cputime
;
2672 p
->utimescaled
+= cputime_scaled
;
2673 account_group_user_time(p
, cputime
);
2675 index
= (TASK_NICE(p
) > 0) ? CPUTIME_NICE
: CPUTIME_USER
;
2677 /* Add user time to cpustat. */
2678 task_group_account_field(p
, index
, (__force u64
) cputime
);
2680 /* Account for user time used */
2681 acct_update_integrals(p
);
2685 * Account guest cpu time to a process.
2686 * @p: the process that the cpu time gets accounted to
2687 * @cputime: the cpu time spent in virtual machine since the last update
2688 * @cputime_scaled: cputime scaled by cpu frequency
2690 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
2691 cputime_t cputime_scaled
)
2693 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2695 /* Add guest time to process. */
2696 p
->utime
+= cputime
;
2697 p
->utimescaled
+= cputime_scaled
;
2698 account_group_user_time(p
, cputime
);
2699 p
->gtime
+= cputime
;
2701 /* Add guest time to cpustat. */
2702 if (TASK_NICE(p
) > 0) {
2703 cpustat
[CPUTIME_NICE
] += (__force u64
) cputime
;
2704 cpustat
[CPUTIME_GUEST_NICE
] += (__force u64
) cputime
;
2706 cpustat
[CPUTIME_USER
] += (__force u64
) cputime
;
2707 cpustat
[CPUTIME_GUEST
] += (__force u64
) cputime
;
2712 * Account system cpu time to a process and desired cpustat field
2713 * @p: the process that the cpu time gets accounted to
2714 * @cputime: the cpu time spent in kernel space since the last update
2715 * @cputime_scaled: cputime scaled by cpu frequency
2716 * @target_cputime64: pointer to cpustat field that has to be updated
2719 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
2720 cputime_t cputime_scaled
, int index
)
2722 /* Add system time to process. */
2723 p
->stime
+= cputime
;
2724 p
->stimescaled
+= cputime_scaled
;
2725 account_group_system_time(p
, cputime
);
2727 /* Add system time to cpustat. */
2728 task_group_account_field(p
, index
, (__force u64
) cputime
);
2730 /* Account for system time used */
2731 acct_update_integrals(p
);
2735 * Account system cpu time to a process.
2736 * @p: the process that the cpu time gets accounted to
2737 * @hardirq_offset: the offset to subtract from hardirq_count()
2738 * @cputime: the cpu time spent in kernel space since the last update
2739 * @cputime_scaled: cputime scaled by cpu frequency
2741 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
2742 cputime_t cputime
, cputime_t cputime_scaled
)
2746 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
2747 account_guest_time(p
, cputime
, cputime_scaled
);
2751 if (hardirq_count() - hardirq_offset
)
2752 index
= CPUTIME_IRQ
;
2753 else if (in_serving_softirq())
2754 index
= CPUTIME_SOFTIRQ
;
2756 index
= CPUTIME_SYSTEM
;
2758 __account_system_time(p
, cputime
, cputime_scaled
, index
);
2762 * Account for involuntary wait time.
2763 * @cputime: the cpu time spent in involuntary wait
2765 void account_steal_time(cputime_t cputime
)
2767 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2769 cpustat
[CPUTIME_STEAL
] += (__force u64
) cputime
;
2773 * Account for idle time.
2774 * @cputime: the cpu time spent in idle wait
2776 void account_idle_time(cputime_t cputime
)
2778 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2779 struct rq
*rq
= this_rq();
2781 if (atomic_read(&rq
->nr_iowait
) > 0)
2782 cpustat
[CPUTIME_IOWAIT
] += (__force u64
) cputime
;
2784 cpustat
[CPUTIME_IDLE
] += (__force u64
) cputime
;
2787 static __always_inline
bool steal_account_process_tick(void)
2789 #ifdef CONFIG_PARAVIRT
2790 if (static_key_false(¶virt_steal_enabled
)) {
2793 steal
= paravirt_steal_clock(smp_processor_id());
2794 steal
-= this_rq()->prev_steal_time
;
2796 st
= steal_ticks(steal
);
2797 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
2799 account_steal_time(st
);
2806 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2808 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2810 * Account a tick to a process and cpustat
2811 * @p: the process that the cpu time gets accounted to
2812 * @user_tick: is the tick from userspace
2813 * @rq: the pointer to rq
2815 * Tick demultiplexing follows the order
2816 * - pending hardirq update
2817 * - pending softirq update
2821 * - check for guest_time
2822 * - else account as system_time
2824 * Check for hardirq is done both for system and user time as there is
2825 * no timer going off while we are on hardirq and hence we may never get an
2826 * opportunity to update it solely in system time.
2827 * p->stime and friends are only updated on system time and not on irq
2828 * softirq as those do not count in task exec_runtime any more.
2830 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2833 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2834 u64
*cpustat
= kcpustat_this_cpu
->cpustat
;
2836 if (steal_account_process_tick())
2839 if (irqtime_account_hi_update()) {
2840 cpustat
[CPUTIME_IRQ
] += (__force u64
) cputime_one_jiffy
;
2841 } else if (irqtime_account_si_update()) {
2842 cpustat
[CPUTIME_SOFTIRQ
] += (__force u64
) cputime_one_jiffy
;
2843 } else if (this_cpu_ksoftirqd() == p
) {
2845 * ksoftirqd time do not get accounted in cpu_softirq_time.
2846 * So, we have to handle it separately here.
2847 * Also, p->stime needs to be updated for ksoftirqd.
2849 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2851 } else if (user_tick
) {
2852 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2853 } else if (p
== rq
->idle
) {
2854 account_idle_time(cputime_one_jiffy
);
2855 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
2856 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2858 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
2863 static void irqtime_account_idle_ticks(int ticks
)
2866 struct rq
*rq
= this_rq();
2868 for (i
= 0; i
< ticks
; i
++)
2869 irqtime_account_process_tick(current
, 0, rq
);
2871 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2872 static void irqtime_account_idle_ticks(int ticks
) {}
2873 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
2875 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2878 * Account a single tick of cpu time.
2879 * @p: the process that the cpu time gets accounted to
2880 * @user_tick: indicates if the tick is a user or a system tick
2882 void account_process_tick(struct task_struct
*p
, int user_tick
)
2884 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
2885 struct rq
*rq
= this_rq();
2887 if (sched_clock_irqtime
) {
2888 irqtime_account_process_tick(p
, user_tick
, rq
);
2892 if (steal_account_process_tick())
2896 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
2897 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
2898 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
2901 account_idle_time(cputime_one_jiffy
);
2905 * Account multiple ticks of steal time.
2906 * @p: the process from which the cpu time has been stolen
2907 * @ticks: number of stolen ticks
2909 void account_steal_ticks(unsigned long ticks
)
2911 account_steal_time(jiffies_to_cputime(ticks
));
2915 * Account multiple ticks of idle time.
2916 * @ticks: number of stolen ticks
2918 void account_idle_ticks(unsigned long ticks
)
2921 if (sched_clock_irqtime
) {
2922 irqtime_account_idle_ticks(ticks
);
2926 account_idle_time(jiffies_to_cputime(ticks
));
2932 * Use precise platform statistics if available:
2934 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2935 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2941 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2943 struct task_cputime cputime
;
2945 thread_group_cputime(p
, &cputime
);
2947 *ut
= cputime
.utime
;
2948 *st
= cputime
.stime
;
2952 #ifndef nsecs_to_cputime
2953 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2956 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2958 cputime_t rtime
, utime
= p
->utime
, total
= utime
+ p
->stime
;
2961 * Use CFS's precise accounting:
2963 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
2966 u64 temp
= (__force u64
) rtime
;
2968 temp
*= (__force u64
) utime
;
2969 do_div(temp
, (__force u32
) total
);
2970 utime
= (__force cputime_t
) temp
;
2975 * Compare with previous values, to keep monotonicity:
2977 p
->prev_utime
= max(p
->prev_utime
, utime
);
2978 p
->prev_stime
= max(p
->prev_stime
, rtime
- p
->prev_utime
);
2980 *ut
= p
->prev_utime
;
2981 *st
= p
->prev_stime
;
2985 * Must be called with siglock held.
2987 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
2989 struct signal_struct
*sig
= p
->signal
;
2990 struct task_cputime cputime
;
2991 cputime_t rtime
, utime
, total
;
2993 thread_group_cputime(p
, &cputime
);
2995 total
= cputime
.utime
+ cputime
.stime
;
2996 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
2999 u64 temp
= (__force u64
) rtime
;
3001 temp
*= (__force u64
) cputime
.utime
;
3002 do_div(temp
, (__force u32
) total
);
3003 utime
= (__force cputime_t
) temp
;
3007 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3008 sig
->prev_stime
= max(sig
->prev_stime
, rtime
- sig
->prev_utime
);
3010 *ut
= sig
->prev_utime
;
3011 *st
= sig
->prev_stime
;
3016 * This function gets called by the timer code, with HZ frequency.
3017 * We call it with interrupts disabled.
3019 void scheduler_tick(void)
3021 int cpu
= smp_processor_id();
3022 struct rq
*rq
= cpu_rq(cpu
);
3023 struct task_struct
*curr
= rq
->curr
;
3027 raw_spin_lock(&rq
->lock
);
3028 update_rq_clock(rq
);
3029 update_cpu_load_active(rq
);
3030 curr
->sched_class
->task_tick(rq
, curr
, 0);
3031 raw_spin_unlock(&rq
->lock
);
3033 perf_event_task_tick();
3036 rq
->idle_balance
= idle_cpu(cpu
);
3037 trigger_load_balance(rq
, cpu
);
3041 notrace
unsigned long get_parent_ip(unsigned long addr
)
3043 if (in_lock_functions(addr
)) {
3044 addr
= CALLER_ADDR2
;
3045 if (in_lock_functions(addr
))
3046 addr
= CALLER_ADDR3
;
3051 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3052 defined(CONFIG_PREEMPT_TRACER))
3054 void __kprobes
add_preempt_count(int val
)
3056 #ifdef CONFIG_DEBUG_PREEMPT
3060 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3063 preempt_count() += val
;
3064 #ifdef CONFIG_DEBUG_PREEMPT
3066 * Spinlock count overflowing soon?
3068 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3071 if (preempt_count() == val
)
3072 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3074 EXPORT_SYMBOL(add_preempt_count
);
3076 void __kprobes
sub_preempt_count(int val
)
3078 #ifdef CONFIG_DEBUG_PREEMPT
3082 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3085 * Is the spinlock portion underflowing?
3087 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3088 !(preempt_count() & PREEMPT_MASK
)))
3092 if (preempt_count() == val
)
3093 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3094 preempt_count() -= val
;
3096 EXPORT_SYMBOL(sub_preempt_count
);
3101 * Print scheduling while atomic bug:
3103 static noinline
void __schedule_bug(struct task_struct
*prev
)
3105 if (oops_in_progress
)
3108 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3109 prev
->comm
, prev
->pid
, preempt_count());
3111 debug_show_held_locks(prev
);
3113 if (irqs_disabled())
3114 print_irqtrace_events(prev
);
3119 * Various schedule()-time debugging checks and statistics:
3121 static inline void schedule_debug(struct task_struct
*prev
)
3124 * Test if we are atomic. Since do_exit() needs to call into
3125 * schedule() atomically, we ignore that path for now.
3126 * Otherwise, whine if we are scheduling when we should not be.
3128 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3129 __schedule_bug(prev
);
3132 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3134 schedstat_inc(this_rq(), sched_count
);
3137 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3139 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
3140 update_rq_clock(rq
);
3141 prev
->sched_class
->put_prev_task(rq
, prev
);
3145 * Pick up the highest-prio task:
3147 static inline struct task_struct
*
3148 pick_next_task(struct rq
*rq
)
3150 const struct sched_class
*class;
3151 struct task_struct
*p
;
3154 * Optimization: we know that if all tasks are in
3155 * the fair class we can call that function directly:
3157 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3158 p
= fair_sched_class
.pick_next_task(rq
);
3163 for_each_class(class) {
3164 p
= class->pick_next_task(rq
);
3169 BUG(); /* the idle class will always have a runnable task */
3173 * __schedule() is the main scheduler function.
3175 static void __sched
__schedule(void)
3177 struct task_struct
*prev
, *next
;
3178 unsigned long *switch_count
;
3184 cpu
= smp_processor_id();
3186 rcu_note_context_switch(cpu
);
3189 schedule_debug(prev
);
3191 if (sched_feat(HRTICK
))
3194 raw_spin_lock_irq(&rq
->lock
);
3196 switch_count
= &prev
->nivcsw
;
3197 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3198 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3199 prev
->state
= TASK_RUNNING
;
3201 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3205 * If a worker went to sleep, notify and ask workqueue
3206 * whether it wants to wake up a task to maintain
3209 if (prev
->flags
& PF_WQ_WORKER
) {
3210 struct task_struct
*to_wakeup
;
3212 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3214 try_to_wake_up_local(to_wakeup
);
3217 switch_count
= &prev
->nvcsw
;
3220 pre_schedule(rq
, prev
);
3222 if (unlikely(!rq
->nr_running
))
3223 idle_balance(cpu
, rq
);
3225 put_prev_task(rq
, prev
);
3226 next
= pick_next_task(rq
);
3227 clear_tsk_need_resched(prev
);
3228 rq
->skip_clock_update
= 0;
3230 if (likely(prev
!= next
)) {
3235 context_switch(rq
, prev
, next
); /* unlocks the rq */
3237 * The context switch have flipped the stack from under us
3238 * and restored the local variables which were saved when
3239 * this task called schedule() in the past. prev == current
3240 * is still correct, but it can be moved to another cpu/rq.
3242 cpu
= smp_processor_id();
3245 raw_spin_unlock_irq(&rq
->lock
);
3249 sched_preempt_enable_no_resched();
3254 static inline void sched_submit_work(struct task_struct
*tsk
)
3256 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3259 * If we are going to sleep and we have plugged IO queued,
3260 * make sure to submit it to avoid deadlocks.
3262 if (blk_needs_flush_plug(tsk
))
3263 blk_schedule_flush_plug(tsk
);
3266 asmlinkage
void __sched
schedule(void)
3268 struct task_struct
*tsk
= current
;
3270 sched_submit_work(tsk
);
3273 EXPORT_SYMBOL(schedule
);
3276 * schedule_preempt_disabled - called with preemption disabled
3278 * Returns with preemption disabled. Note: preempt_count must be 1
3280 void __sched
schedule_preempt_disabled(void)
3282 sched_preempt_enable_no_resched();
3287 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3289 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
3291 if (lock
->owner
!= owner
)
3295 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3296 * lock->owner still matches owner, if that fails, owner might
3297 * point to free()d memory, if it still matches, the rcu_read_lock()
3298 * ensures the memory stays valid.
3302 return owner
->on_cpu
;
3306 * Look out! "owner" is an entirely speculative pointer
3307 * access and not reliable.
3309 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
3311 if (!sched_feat(OWNER_SPIN
))
3315 while (owner_running(lock
, owner
)) {
3319 arch_mutex_cpu_relax();
3324 * We break out the loop above on need_resched() and when the
3325 * owner changed, which is a sign for heavy contention. Return
3326 * success only when lock->owner is NULL.
3328 return lock
->owner
== NULL
;
3332 #ifdef CONFIG_PREEMPT
3334 * this is the entry point to schedule() from in-kernel preemption
3335 * off of preempt_enable. Kernel preemptions off return from interrupt
3336 * occur there and call schedule directly.
3338 asmlinkage
void __sched notrace
preempt_schedule(void)
3340 struct thread_info
*ti
= current_thread_info();
3343 * If there is a non-zero preempt_count or interrupts are disabled,
3344 * we do not want to preempt the current task. Just return..
3346 if (likely(ti
->preempt_count
|| irqs_disabled()))
3350 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3352 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3355 * Check again in case we missed a preemption opportunity
3356 * between schedule and now.
3359 } while (need_resched());
3361 EXPORT_SYMBOL(preempt_schedule
);
3364 * this is the entry point to schedule() from kernel preemption
3365 * off of irq context.
3366 * Note, that this is called and return with irqs disabled. This will
3367 * protect us against recursive calling from irq.
3369 asmlinkage
void __sched
preempt_schedule_irq(void)
3371 struct thread_info
*ti
= current_thread_info();
3373 /* Catch callers which need to be fixed */
3374 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3377 add_preempt_count(PREEMPT_ACTIVE
);
3380 local_irq_disable();
3381 sub_preempt_count(PREEMPT_ACTIVE
);
3384 * Check again in case we missed a preemption opportunity
3385 * between schedule and now.
3388 } while (need_resched());
3391 #endif /* CONFIG_PREEMPT */
3393 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3396 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3398 EXPORT_SYMBOL(default_wake_function
);
3401 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3402 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3403 * number) then we wake all the non-exclusive tasks and one exclusive task.
3405 * There are circumstances in which we can try to wake a task which has already
3406 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3407 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3409 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3410 int nr_exclusive
, int wake_flags
, void *key
)
3412 wait_queue_t
*curr
, *next
;
3414 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3415 unsigned flags
= curr
->flags
;
3417 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3418 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3424 * __wake_up - wake up threads blocked on a waitqueue.
3426 * @mode: which threads
3427 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3428 * @key: is directly passed to the wakeup function
3430 * It may be assumed that this function implies a write memory barrier before
3431 * changing the task state if and only if any tasks are woken up.
3433 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3434 int nr_exclusive
, void *key
)
3436 unsigned long flags
;
3438 spin_lock_irqsave(&q
->lock
, flags
);
3439 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3440 spin_unlock_irqrestore(&q
->lock
, flags
);
3442 EXPORT_SYMBOL(__wake_up
);
3445 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3447 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3449 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3451 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3453 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3455 __wake_up_common(q
, mode
, 1, 0, key
);
3457 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3460 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3462 * @mode: which threads
3463 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3464 * @key: opaque value to be passed to wakeup targets
3466 * The sync wakeup differs that the waker knows that it will schedule
3467 * away soon, so while the target thread will be woken up, it will not
3468 * be migrated to another CPU - ie. the two threads are 'synchronized'
3469 * with each other. This can prevent needless bouncing between CPUs.
3471 * On UP it can prevent extra preemption.
3473 * It may be assumed that this function implies a write memory barrier before
3474 * changing the task state if and only if any tasks are woken up.
3476 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3477 int nr_exclusive
, void *key
)
3479 unsigned long flags
;
3480 int wake_flags
= WF_SYNC
;
3485 if (unlikely(!nr_exclusive
))
3488 spin_lock_irqsave(&q
->lock
, flags
);
3489 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3490 spin_unlock_irqrestore(&q
->lock
, flags
);
3492 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3495 * __wake_up_sync - see __wake_up_sync_key()
3497 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3499 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3501 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3504 * complete: - signals a single thread waiting on this completion
3505 * @x: holds the state of this particular completion
3507 * This will wake up a single thread waiting on this completion. Threads will be
3508 * awakened in the same order in which they were queued.
3510 * See also complete_all(), wait_for_completion() and related routines.
3512 * It may be assumed that this function implies a write memory barrier before
3513 * changing the task state if and only if any tasks are woken up.
3515 void complete(struct completion
*x
)
3517 unsigned long flags
;
3519 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3521 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3522 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3524 EXPORT_SYMBOL(complete
);
3527 * complete_all: - signals all threads waiting on this completion
3528 * @x: holds the state of this particular completion
3530 * This will wake up all threads waiting on this particular completion event.
3532 * It may be assumed that this function implies a write memory barrier before
3533 * changing the task state if and only if any tasks are woken up.
3535 void complete_all(struct completion
*x
)
3537 unsigned long flags
;
3539 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3540 x
->done
+= UINT_MAX
/2;
3541 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3542 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3544 EXPORT_SYMBOL(complete_all
);
3546 static inline long __sched
3547 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3550 DECLARE_WAITQUEUE(wait
, current
);
3552 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3554 if (signal_pending_state(state
, current
)) {
3555 timeout
= -ERESTARTSYS
;
3558 __set_current_state(state
);
3559 spin_unlock_irq(&x
->wait
.lock
);
3560 timeout
= schedule_timeout(timeout
);
3561 spin_lock_irq(&x
->wait
.lock
);
3562 } while (!x
->done
&& timeout
);
3563 __remove_wait_queue(&x
->wait
, &wait
);
3568 return timeout
?: 1;
3572 wait_for_common(struct completion
*x
, long timeout
, int state
)
3576 spin_lock_irq(&x
->wait
.lock
);
3577 timeout
= do_wait_for_common(x
, timeout
, state
);
3578 spin_unlock_irq(&x
->wait
.lock
);
3583 * wait_for_completion: - waits for completion of a task
3584 * @x: holds the state of this particular completion
3586 * This waits to be signaled for completion of a specific task. It is NOT
3587 * interruptible and there is no timeout.
3589 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3590 * and interrupt capability. Also see complete().
3592 void __sched
wait_for_completion(struct completion
*x
)
3594 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3596 EXPORT_SYMBOL(wait_for_completion
);
3599 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3600 * @x: holds the state of this particular completion
3601 * @timeout: timeout value in jiffies
3603 * This waits for either a completion of a specific task to be signaled or for a
3604 * specified timeout to expire. The timeout is in jiffies. It is not
3607 * The return value is 0 if timed out, and positive (at least 1, or number of
3608 * jiffies left till timeout) if completed.
3610 unsigned long __sched
3611 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3613 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3615 EXPORT_SYMBOL(wait_for_completion_timeout
);
3618 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3619 * @x: holds the state of this particular completion
3621 * This waits for completion of a specific task to be signaled. It is
3624 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3626 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3628 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3629 if (t
== -ERESTARTSYS
)
3633 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3636 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3637 * @x: holds the state of this particular completion
3638 * @timeout: timeout value in jiffies
3640 * This waits for either a completion of a specific task to be signaled or for a
3641 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3643 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3644 * positive (at least 1, or number of jiffies left till timeout) if completed.
3647 wait_for_completion_interruptible_timeout(struct completion
*x
,
3648 unsigned long timeout
)
3650 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3652 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3655 * wait_for_completion_killable: - waits for completion of a task (killable)
3656 * @x: holds the state of this particular completion
3658 * This waits to be signaled for completion of a specific task. It can be
3659 * interrupted by a kill signal.
3661 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3663 int __sched
wait_for_completion_killable(struct completion
*x
)
3665 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3666 if (t
== -ERESTARTSYS
)
3670 EXPORT_SYMBOL(wait_for_completion_killable
);
3673 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3674 * @x: holds the state of this particular completion
3675 * @timeout: timeout value in jiffies
3677 * This waits for either a completion of a specific task to be
3678 * signaled or for a specified timeout to expire. It can be
3679 * interrupted by a kill signal. The timeout is in jiffies.
3681 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3682 * positive (at least 1, or number of jiffies left till timeout) if completed.
3685 wait_for_completion_killable_timeout(struct completion
*x
,
3686 unsigned long timeout
)
3688 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3690 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3693 * try_wait_for_completion - try to decrement a completion without blocking
3694 * @x: completion structure
3696 * Returns: 0 if a decrement cannot be done without blocking
3697 * 1 if a decrement succeeded.
3699 * If a completion is being used as a counting completion,
3700 * attempt to decrement the counter without blocking. This
3701 * enables us to avoid waiting if the resource the completion
3702 * is protecting is not available.
3704 bool try_wait_for_completion(struct completion
*x
)
3706 unsigned long flags
;
3709 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3714 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3717 EXPORT_SYMBOL(try_wait_for_completion
);
3720 * completion_done - Test to see if a completion has any waiters
3721 * @x: completion structure
3723 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3724 * 1 if there are no waiters.
3727 bool completion_done(struct completion
*x
)
3729 unsigned long flags
;
3732 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3735 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3738 EXPORT_SYMBOL(completion_done
);
3741 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3743 unsigned long flags
;
3746 init_waitqueue_entry(&wait
, current
);
3748 __set_current_state(state
);
3750 spin_lock_irqsave(&q
->lock
, flags
);
3751 __add_wait_queue(q
, &wait
);
3752 spin_unlock(&q
->lock
);
3753 timeout
= schedule_timeout(timeout
);
3754 spin_lock_irq(&q
->lock
);
3755 __remove_wait_queue(q
, &wait
);
3756 spin_unlock_irqrestore(&q
->lock
, flags
);
3761 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3763 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3765 EXPORT_SYMBOL(interruptible_sleep_on
);
3768 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3770 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3772 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3774 void __sched
sleep_on(wait_queue_head_t
*q
)
3776 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3778 EXPORT_SYMBOL(sleep_on
);
3780 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3782 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3784 EXPORT_SYMBOL(sleep_on_timeout
);
3786 #ifdef CONFIG_RT_MUTEXES
3789 * rt_mutex_setprio - set the current priority of a task
3791 * @prio: prio value (kernel-internal form)
3793 * This function changes the 'effective' priority of a task. It does
3794 * not touch ->normal_prio like __setscheduler().
3796 * Used by the rt_mutex code to implement priority inheritance logic.
3798 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3800 int oldprio
, on_rq
, running
;
3802 const struct sched_class
*prev_class
;
3804 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3806 rq
= __task_rq_lock(p
);
3809 * Idle task boosting is a nono in general. There is one
3810 * exception, when PREEMPT_RT and NOHZ is active:
3812 * The idle task calls get_next_timer_interrupt() and holds
3813 * the timer wheel base->lock on the CPU and another CPU wants
3814 * to access the timer (probably to cancel it). We can safely
3815 * ignore the boosting request, as the idle CPU runs this code
3816 * with interrupts disabled and will complete the lock
3817 * protected section without being interrupted. So there is no
3818 * real need to boost.
3820 if (unlikely(p
== rq
->idle
)) {
3821 WARN_ON(p
!= rq
->curr
);
3822 WARN_ON(p
->pi_blocked_on
);
3826 trace_sched_pi_setprio(p
, prio
);
3828 prev_class
= p
->sched_class
;
3830 running
= task_current(rq
, p
);
3832 dequeue_task(rq
, p
, 0);
3834 p
->sched_class
->put_prev_task(rq
, p
);
3837 p
->sched_class
= &rt_sched_class
;
3839 p
->sched_class
= &fair_sched_class
;
3844 p
->sched_class
->set_curr_task(rq
);
3846 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3848 check_class_changed(rq
, p
, prev_class
, oldprio
);
3850 __task_rq_unlock(rq
);
3853 void set_user_nice(struct task_struct
*p
, long nice
)
3855 int old_prio
, delta
, on_rq
;
3856 unsigned long flags
;
3859 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3862 * We have to be careful, if called from sys_setpriority(),
3863 * the task might be in the middle of scheduling on another CPU.
3865 rq
= task_rq_lock(p
, &flags
);
3867 * The RT priorities are set via sched_setscheduler(), but we still
3868 * allow the 'normal' nice value to be set - but as expected
3869 * it wont have any effect on scheduling until the task is
3870 * SCHED_FIFO/SCHED_RR:
3872 if (task_has_rt_policy(p
)) {
3873 p
->static_prio
= NICE_TO_PRIO(nice
);
3878 dequeue_task(rq
, p
, 0);
3880 p
->static_prio
= NICE_TO_PRIO(nice
);
3883 p
->prio
= effective_prio(p
);
3884 delta
= p
->prio
- old_prio
;
3887 enqueue_task(rq
, p
, 0);
3889 * If the task increased its priority or is running and
3890 * lowered its priority, then reschedule its CPU:
3892 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3893 resched_task(rq
->curr
);
3896 task_rq_unlock(rq
, p
, &flags
);
3898 EXPORT_SYMBOL(set_user_nice
);
3901 * can_nice - check if a task can reduce its nice value
3905 int can_nice(const struct task_struct
*p
, const int nice
)
3907 /* convert nice value [19,-20] to rlimit style value [1,40] */
3908 int nice_rlim
= 20 - nice
;
3910 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3911 capable(CAP_SYS_NICE
));
3914 #ifdef __ARCH_WANT_SYS_NICE
3917 * sys_nice - change the priority of the current process.
3918 * @increment: priority increment
3920 * sys_setpriority is a more generic, but much slower function that
3921 * does similar things.
3923 SYSCALL_DEFINE1(nice
, int, increment
)
3928 * Setpriority might change our priority at the same moment.
3929 * We don't have to worry. Conceptually one call occurs first
3930 * and we have a single winner.
3932 if (increment
< -40)
3937 nice
= TASK_NICE(current
) + increment
;
3943 if (increment
< 0 && !can_nice(current
, nice
))
3946 retval
= security_task_setnice(current
, nice
);
3950 set_user_nice(current
, nice
);
3957 * task_prio - return the priority value of a given task.
3958 * @p: the task in question.
3960 * This is the priority value as seen by users in /proc.
3961 * RT tasks are offset by -200. Normal tasks are centered
3962 * around 0, value goes from -16 to +15.
3964 int task_prio(const struct task_struct
*p
)
3966 return p
->prio
- MAX_RT_PRIO
;
3970 * task_nice - return the nice value of a given task.
3971 * @p: the task in question.
3973 int task_nice(const struct task_struct
*p
)
3975 return TASK_NICE(p
);
3977 EXPORT_SYMBOL(task_nice
);
3980 * idle_cpu - is a given cpu idle currently?
3981 * @cpu: the processor in question.
3983 int idle_cpu(int cpu
)
3985 struct rq
*rq
= cpu_rq(cpu
);
3987 if (rq
->curr
!= rq
->idle
)
3994 if (!llist_empty(&rq
->wake_list
))
4002 * idle_task - return the idle task for a given cpu.
4003 * @cpu: the processor in question.
4005 struct task_struct
*idle_task(int cpu
)
4007 return cpu_rq(cpu
)->idle
;
4011 * find_process_by_pid - find a process with a matching PID value.
4012 * @pid: the pid in question.
4014 static struct task_struct
*find_process_by_pid(pid_t pid
)
4016 return pid
? find_task_by_vpid(pid
) : current
;
4019 /* Actually do priority change: must hold rq lock. */
4021 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4024 p
->rt_priority
= prio
;
4025 p
->normal_prio
= normal_prio(p
);
4026 /* we are holding p->pi_lock already */
4027 p
->prio
= rt_mutex_getprio(p
);
4028 if (rt_prio(p
->prio
))
4029 p
->sched_class
= &rt_sched_class
;
4031 p
->sched_class
= &fair_sched_class
;
4036 * check the target process has a UID that matches the current process's
4038 static bool check_same_owner(struct task_struct
*p
)
4040 const struct cred
*cred
= current_cred(), *pcred
;
4044 pcred
= __task_cred(p
);
4045 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
4046 match
= (cred
->euid
== pcred
->euid
||
4047 cred
->euid
== pcred
->uid
);
4054 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4055 const struct sched_param
*param
, bool user
)
4057 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4058 unsigned long flags
;
4059 const struct sched_class
*prev_class
;
4063 /* may grab non-irq protected spin_locks */
4064 BUG_ON(in_interrupt());
4066 /* double check policy once rq lock held */
4068 reset_on_fork
= p
->sched_reset_on_fork
;
4069 policy
= oldpolicy
= p
->policy
;
4071 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4072 policy
&= ~SCHED_RESET_ON_FORK
;
4074 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4075 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4076 policy
!= SCHED_IDLE
)
4081 * Valid priorities for SCHED_FIFO and SCHED_RR are
4082 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4083 * SCHED_BATCH and SCHED_IDLE is 0.
4085 if (param
->sched_priority
< 0 ||
4086 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4087 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4089 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4093 * Allow unprivileged RT tasks to decrease priority:
4095 if (user
&& !capable(CAP_SYS_NICE
)) {
4096 if (rt_policy(policy
)) {
4097 unsigned long rlim_rtprio
=
4098 task_rlimit(p
, RLIMIT_RTPRIO
);
4100 /* can't set/change the rt policy */
4101 if (policy
!= p
->policy
&& !rlim_rtprio
)
4104 /* can't increase priority */
4105 if (param
->sched_priority
> p
->rt_priority
&&
4106 param
->sched_priority
> rlim_rtprio
)
4111 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4112 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4114 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
4115 if (!can_nice(p
, TASK_NICE(p
)))
4119 /* can't change other user's priorities */
4120 if (!check_same_owner(p
))
4123 /* Normal users shall not reset the sched_reset_on_fork flag */
4124 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4129 retval
= security_task_setscheduler(p
);
4135 * make sure no PI-waiters arrive (or leave) while we are
4136 * changing the priority of the task:
4138 * To be able to change p->policy safely, the appropriate
4139 * runqueue lock must be held.
4141 rq
= task_rq_lock(p
, &flags
);
4144 * Changing the policy of the stop threads its a very bad idea
4146 if (p
== rq
->stop
) {
4147 task_rq_unlock(rq
, p
, &flags
);
4152 * If not changing anything there's no need to proceed further:
4154 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
4155 param
->sched_priority
== p
->rt_priority
))) {
4157 __task_rq_unlock(rq
);
4158 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4162 #ifdef CONFIG_RT_GROUP_SCHED
4165 * Do not allow realtime tasks into groups that have no runtime
4168 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4169 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4170 !task_group_is_autogroup(task_group(p
))) {
4171 task_rq_unlock(rq
, p
, &flags
);
4177 /* recheck policy now with rq lock held */
4178 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4179 policy
= oldpolicy
= -1;
4180 task_rq_unlock(rq
, p
, &flags
);
4184 running
= task_current(rq
, p
);
4186 dequeue_task(rq
, p
, 0);
4188 p
->sched_class
->put_prev_task(rq
, p
);
4190 p
->sched_reset_on_fork
= reset_on_fork
;
4193 prev_class
= p
->sched_class
;
4194 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4197 p
->sched_class
->set_curr_task(rq
);
4199 enqueue_task(rq
, p
, 0);
4201 check_class_changed(rq
, p
, prev_class
, oldprio
);
4202 task_rq_unlock(rq
, p
, &flags
);
4204 rt_mutex_adjust_pi(p
);
4210 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4211 * @p: the task in question.
4212 * @policy: new policy.
4213 * @param: structure containing the new RT priority.
4215 * NOTE that the task may be already dead.
4217 int sched_setscheduler(struct task_struct
*p
, int policy
,
4218 const struct sched_param
*param
)
4220 return __sched_setscheduler(p
, policy
, param
, true);
4222 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4225 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4226 * @p: the task in question.
4227 * @policy: new policy.
4228 * @param: structure containing the new RT priority.
4230 * Just like sched_setscheduler, only don't bother checking if the
4231 * current context has permission. For example, this is needed in
4232 * stop_machine(): we create temporary high priority worker threads,
4233 * but our caller might not have that capability.
4235 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4236 const struct sched_param
*param
)
4238 return __sched_setscheduler(p
, policy
, param
, false);
4242 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4244 struct sched_param lparam
;
4245 struct task_struct
*p
;
4248 if (!param
|| pid
< 0)
4250 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4255 p
= find_process_by_pid(pid
);
4257 retval
= sched_setscheduler(p
, policy
, &lparam
);
4264 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4265 * @pid: the pid in question.
4266 * @policy: new policy.
4267 * @param: structure containing the new RT priority.
4269 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4270 struct sched_param __user
*, param
)
4272 /* negative values for policy are not valid */
4276 return do_sched_setscheduler(pid
, policy
, param
);
4280 * sys_sched_setparam - set/change the RT priority of a thread
4281 * @pid: the pid in question.
4282 * @param: structure containing the new RT priority.
4284 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4286 return do_sched_setscheduler(pid
, -1, param
);
4290 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4291 * @pid: the pid in question.
4293 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4295 struct task_struct
*p
;
4303 p
= find_process_by_pid(pid
);
4305 retval
= security_task_getscheduler(p
);
4308 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4315 * sys_sched_getparam - get the RT priority of a thread
4316 * @pid: the pid in question.
4317 * @param: structure containing the RT priority.
4319 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4321 struct sched_param lp
;
4322 struct task_struct
*p
;
4325 if (!param
|| pid
< 0)
4329 p
= find_process_by_pid(pid
);
4334 retval
= security_task_getscheduler(p
);
4338 lp
.sched_priority
= p
->rt_priority
;
4342 * This one might sleep, we cannot do it with a spinlock held ...
4344 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4353 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4355 cpumask_var_t cpus_allowed
, new_mask
;
4356 struct task_struct
*p
;
4362 p
= find_process_by_pid(pid
);
4369 /* Prevent p going away */
4373 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4377 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4379 goto out_free_cpus_allowed
;
4382 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4385 retval
= security_task_setscheduler(p
);
4389 cpuset_cpus_allowed(p
, cpus_allowed
);
4390 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4392 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4395 cpuset_cpus_allowed(p
, cpus_allowed
);
4396 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4398 * We must have raced with a concurrent cpuset
4399 * update. Just reset the cpus_allowed to the
4400 * cpuset's cpus_allowed
4402 cpumask_copy(new_mask
, cpus_allowed
);
4407 free_cpumask_var(new_mask
);
4408 out_free_cpus_allowed
:
4409 free_cpumask_var(cpus_allowed
);
4416 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4417 struct cpumask
*new_mask
)
4419 if (len
< cpumask_size())
4420 cpumask_clear(new_mask
);
4421 else if (len
> cpumask_size())
4422 len
= cpumask_size();
4424 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4428 * sys_sched_setaffinity - set the cpu affinity of a process
4429 * @pid: pid of the process
4430 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4431 * @user_mask_ptr: user-space pointer to the new cpu mask
4433 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4434 unsigned long __user
*, user_mask_ptr
)
4436 cpumask_var_t new_mask
;
4439 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4442 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4444 retval
= sched_setaffinity(pid
, new_mask
);
4445 free_cpumask_var(new_mask
);
4449 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4451 struct task_struct
*p
;
4452 unsigned long flags
;
4459 p
= find_process_by_pid(pid
);
4463 retval
= security_task_getscheduler(p
);
4467 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4468 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4469 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4479 * sys_sched_getaffinity - get the cpu affinity of a process
4480 * @pid: pid of the process
4481 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4482 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4484 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4485 unsigned long __user
*, user_mask_ptr
)
4490 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4492 if (len
& (sizeof(unsigned long)-1))
4495 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4498 ret
= sched_getaffinity(pid
, mask
);
4500 size_t retlen
= min_t(size_t, len
, cpumask_size());
4502 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4507 free_cpumask_var(mask
);
4513 * sys_sched_yield - yield the current processor to other threads.
4515 * This function yields the current CPU to other tasks. If there are no
4516 * other threads running on this CPU then this function will return.
4518 SYSCALL_DEFINE0(sched_yield
)
4520 struct rq
*rq
= this_rq_lock();
4522 schedstat_inc(rq
, yld_count
);
4523 current
->sched_class
->yield_task(rq
);
4526 * Since we are going to call schedule() anyway, there's
4527 * no need to preempt or enable interrupts:
4529 __release(rq
->lock
);
4530 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4531 do_raw_spin_unlock(&rq
->lock
);
4532 sched_preempt_enable_no_resched();
4539 static inline int should_resched(void)
4541 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4544 static void __cond_resched(void)
4546 add_preempt_count(PREEMPT_ACTIVE
);
4548 sub_preempt_count(PREEMPT_ACTIVE
);
4551 int __sched
_cond_resched(void)
4553 if (should_resched()) {
4559 EXPORT_SYMBOL(_cond_resched
);
4562 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4563 * call schedule, and on return reacquire the lock.
4565 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4566 * operations here to prevent schedule() from being called twice (once via
4567 * spin_unlock(), once by hand).
4569 int __cond_resched_lock(spinlock_t
*lock
)
4571 int resched
= should_resched();
4574 lockdep_assert_held(lock
);
4576 if (spin_needbreak(lock
) || resched
) {
4587 EXPORT_SYMBOL(__cond_resched_lock
);
4589 int __sched
__cond_resched_softirq(void)
4591 BUG_ON(!in_softirq());
4593 if (should_resched()) {
4601 EXPORT_SYMBOL(__cond_resched_softirq
);
4604 * yield - yield the current processor to other threads.
4606 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4608 * The scheduler is at all times free to pick the calling task as the most
4609 * eligible task to run, if removing the yield() call from your code breaks
4610 * it, its already broken.
4612 * Typical broken usage is:
4617 * where one assumes that yield() will let 'the other' process run that will
4618 * make event true. If the current task is a SCHED_FIFO task that will never
4619 * happen. Never use yield() as a progress guarantee!!
4621 * If you want to use yield() to wait for something, use wait_event().
4622 * If you want to use yield() to be 'nice' for others, use cond_resched().
4623 * If you still want to use yield(), do not!
4625 void __sched
yield(void)
4627 set_current_state(TASK_RUNNING
);
4630 EXPORT_SYMBOL(yield
);
4633 * yield_to - yield the current processor to another thread in
4634 * your thread group, or accelerate that thread toward the
4635 * processor it's on.
4637 * @preempt: whether task preemption is allowed or not
4639 * It's the caller's job to ensure that the target task struct
4640 * can't go away on us before we can do any checks.
4642 * Returns true if we indeed boosted the target task.
4644 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4646 struct task_struct
*curr
= current
;
4647 struct rq
*rq
, *p_rq
;
4648 unsigned long flags
;
4651 local_irq_save(flags
);
4656 double_rq_lock(rq
, p_rq
);
4657 while (task_rq(p
) != p_rq
) {
4658 double_rq_unlock(rq
, p_rq
);
4662 if (!curr
->sched_class
->yield_to_task
)
4665 if (curr
->sched_class
!= p
->sched_class
)
4668 if (task_running(p_rq
, p
) || p
->state
)
4671 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4673 schedstat_inc(rq
, yld_count
);
4675 * Make p's CPU reschedule; pick_next_entity takes care of
4678 if (preempt
&& rq
!= p_rq
)
4679 resched_task(p_rq
->curr
);
4682 * We might have set it in task_yield_fair(), but are
4683 * not going to schedule(), so don't want to skip
4686 rq
->skip_clock_update
= 0;
4690 double_rq_unlock(rq
, p_rq
);
4691 local_irq_restore(flags
);
4698 EXPORT_SYMBOL_GPL(yield_to
);
4701 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4702 * that process accounting knows that this is a task in IO wait state.
4704 void __sched
io_schedule(void)
4706 struct rq
*rq
= raw_rq();
4708 delayacct_blkio_start();
4709 atomic_inc(&rq
->nr_iowait
);
4710 blk_flush_plug(current
);
4711 current
->in_iowait
= 1;
4713 current
->in_iowait
= 0;
4714 atomic_dec(&rq
->nr_iowait
);
4715 delayacct_blkio_end();
4717 EXPORT_SYMBOL(io_schedule
);
4719 long __sched
io_schedule_timeout(long timeout
)
4721 struct rq
*rq
= raw_rq();
4724 delayacct_blkio_start();
4725 atomic_inc(&rq
->nr_iowait
);
4726 blk_flush_plug(current
);
4727 current
->in_iowait
= 1;
4728 ret
= schedule_timeout(timeout
);
4729 current
->in_iowait
= 0;
4730 atomic_dec(&rq
->nr_iowait
);
4731 delayacct_blkio_end();
4736 * sys_sched_get_priority_max - return maximum RT priority.
4737 * @policy: scheduling class.
4739 * this syscall returns the maximum rt_priority that can be used
4740 * by a given scheduling class.
4742 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4749 ret
= MAX_USER_RT_PRIO
-1;
4761 * sys_sched_get_priority_min - return minimum RT priority.
4762 * @policy: scheduling class.
4764 * this syscall returns the minimum rt_priority that can be used
4765 * by a given scheduling class.
4767 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4785 * sys_sched_rr_get_interval - return the default timeslice of a process.
4786 * @pid: pid of the process.
4787 * @interval: userspace pointer to the timeslice value.
4789 * this syscall writes the default timeslice value of a given process
4790 * into the user-space timespec buffer. A value of '0' means infinity.
4792 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4793 struct timespec __user
*, interval
)
4795 struct task_struct
*p
;
4796 unsigned int time_slice
;
4797 unsigned long flags
;
4807 p
= find_process_by_pid(pid
);
4811 retval
= security_task_getscheduler(p
);
4815 rq
= task_rq_lock(p
, &flags
);
4816 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4817 task_rq_unlock(rq
, p
, &flags
);
4820 jiffies_to_timespec(time_slice
, &t
);
4821 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4829 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4831 void sched_show_task(struct task_struct
*p
)
4833 unsigned long free
= 0;
4836 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4837 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4838 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4839 #if BITS_PER_LONG == 32
4840 if (state
== TASK_RUNNING
)
4841 printk(KERN_CONT
" running ");
4843 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4845 if (state
== TASK_RUNNING
)
4846 printk(KERN_CONT
" running task ");
4848 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4850 #ifdef CONFIG_DEBUG_STACK_USAGE
4851 free
= stack_not_used(p
);
4853 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4854 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4855 (unsigned long)task_thread_info(p
)->flags
);
4857 show_stack(p
, NULL
);
4860 void show_state_filter(unsigned long state_filter
)
4862 struct task_struct
*g
, *p
;
4864 #if BITS_PER_LONG == 32
4866 " task PC stack pid father\n");
4869 " task PC stack pid father\n");
4872 do_each_thread(g
, p
) {
4874 * reset the NMI-timeout, listing all files on a slow
4875 * console might take a lot of time:
4877 touch_nmi_watchdog();
4878 if (!state_filter
|| (p
->state
& state_filter
))
4880 } while_each_thread(g
, p
);
4882 touch_all_softlockup_watchdogs();
4884 #ifdef CONFIG_SCHED_DEBUG
4885 sysrq_sched_debug_show();
4889 * Only show locks if all tasks are dumped:
4892 debug_show_all_locks();
4895 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4897 idle
->sched_class
= &idle_sched_class
;
4901 * init_idle - set up an idle thread for a given CPU
4902 * @idle: task in question
4903 * @cpu: cpu the idle task belongs to
4905 * NOTE: this function does not set the idle thread's NEED_RESCHED
4906 * flag, to make booting more robust.
4908 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4910 struct rq
*rq
= cpu_rq(cpu
);
4911 unsigned long flags
;
4913 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4916 idle
->state
= TASK_RUNNING
;
4917 idle
->se
.exec_start
= sched_clock();
4919 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4921 * We're having a chicken and egg problem, even though we are
4922 * holding rq->lock, the cpu isn't yet set to this cpu so the
4923 * lockdep check in task_group() will fail.
4925 * Similar case to sched_fork(). / Alternatively we could
4926 * use task_rq_lock() here and obtain the other rq->lock.
4931 __set_task_cpu(idle
, cpu
);
4934 rq
->curr
= rq
->idle
= idle
;
4935 #if defined(CONFIG_SMP)
4938 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4940 /* Set the preempt count _outside_ the spinlocks! */
4941 task_thread_info(idle
)->preempt_count
= 0;
4944 * The idle tasks have their own, simple scheduling class:
4946 idle
->sched_class
= &idle_sched_class
;
4947 ftrace_graph_init_idle_task(idle
, cpu
);
4948 #if defined(CONFIG_SMP)
4949 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4954 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4956 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4957 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4959 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4960 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
4964 * This is how migration works:
4966 * 1) we invoke migration_cpu_stop() on the target CPU using
4968 * 2) stopper starts to run (implicitly forcing the migrated thread
4970 * 3) it checks whether the migrated task is still in the wrong runqueue.
4971 * 4) if it's in the wrong runqueue then the migration thread removes
4972 * it and puts it into the right queue.
4973 * 5) stopper completes and stop_one_cpu() returns and the migration
4978 * Change a given task's CPU affinity. Migrate the thread to a
4979 * proper CPU and schedule it away if the CPU it's executing on
4980 * is removed from the allowed bitmask.
4982 * NOTE: the caller must have a valid reference to the task, the
4983 * task must not exit() & deallocate itself prematurely. The
4984 * call is not atomic; no spinlocks may be held.
4986 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4988 unsigned long flags
;
4990 unsigned int dest_cpu
;
4993 rq
= task_rq_lock(p
, &flags
);
4995 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4998 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5003 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
5008 do_set_cpus_allowed(p
, new_mask
);
5010 /* Can the task run on the task's current CPU? If so, we're done */
5011 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5014 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5016 struct migration_arg arg
= { p
, dest_cpu
};
5017 /* Need help from migration thread: drop lock and wait. */
5018 task_rq_unlock(rq
, p
, &flags
);
5019 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5020 tlb_migrate_finish(p
->mm
);
5024 task_rq_unlock(rq
, p
, &flags
);
5028 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5031 * Move (not current) task off this cpu, onto dest cpu. We're doing
5032 * this because either it can't run here any more (set_cpus_allowed()
5033 * away from this CPU, or CPU going down), or because we're
5034 * attempting to rebalance this task on exec (sched_exec).
5036 * So we race with normal scheduler movements, but that's OK, as long
5037 * as the task is no longer on this CPU.
5039 * Returns non-zero if task was successfully migrated.
5041 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5043 struct rq
*rq_dest
, *rq_src
;
5046 if (unlikely(!cpu_active(dest_cpu
)))
5049 rq_src
= cpu_rq(src_cpu
);
5050 rq_dest
= cpu_rq(dest_cpu
);
5052 raw_spin_lock(&p
->pi_lock
);
5053 double_rq_lock(rq_src
, rq_dest
);
5054 /* Already moved. */
5055 if (task_cpu(p
) != src_cpu
)
5057 /* Affinity changed (again). */
5058 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
5062 * If we're not on a rq, the next wake-up will ensure we're
5066 dequeue_task(rq_src
, p
, 0);
5067 set_task_cpu(p
, dest_cpu
);
5068 enqueue_task(rq_dest
, p
, 0);
5069 check_preempt_curr(rq_dest
, p
, 0);
5074 double_rq_unlock(rq_src
, rq_dest
);
5075 raw_spin_unlock(&p
->pi_lock
);
5080 * migration_cpu_stop - this will be executed by a highprio stopper thread
5081 * and performs thread migration by bumping thread off CPU then
5082 * 'pushing' onto another runqueue.
5084 static int migration_cpu_stop(void *data
)
5086 struct migration_arg
*arg
= data
;
5089 * The original target cpu might have gone down and we might
5090 * be on another cpu but it doesn't matter.
5092 local_irq_disable();
5093 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5098 #ifdef CONFIG_HOTPLUG_CPU
5101 * Ensures that the idle task is using init_mm right before its cpu goes
5104 void idle_task_exit(void)
5106 struct mm_struct
*mm
= current
->active_mm
;
5108 BUG_ON(cpu_online(smp_processor_id()));
5111 switch_mm(mm
, &init_mm
, current
);
5116 * While a dead CPU has no uninterruptible tasks queued at this point,
5117 * it might still have a nonzero ->nr_uninterruptible counter, because
5118 * for performance reasons the counter is not stricly tracking tasks to
5119 * their home CPUs. So we just add the counter to another CPU's counter,
5120 * to keep the global sum constant after CPU-down:
5122 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5124 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5126 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5127 rq_src
->nr_uninterruptible
= 0;
5131 * remove the tasks which were accounted by rq from calc_load_tasks.
5133 static void calc_global_load_remove(struct rq
*rq
)
5135 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5136 rq
->calc_load_active
= 0;
5140 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5141 * try_to_wake_up()->select_task_rq().
5143 * Called with rq->lock held even though we'er in stop_machine() and
5144 * there's no concurrency possible, we hold the required locks anyway
5145 * because of lock validation efforts.
5147 static void migrate_tasks(unsigned int dead_cpu
)
5149 struct rq
*rq
= cpu_rq(dead_cpu
);
5150 struct task_struct
*next
, *stop
= rq
->stop
;
5154 * Fudge the rq selection such that the below task selection loop
5155 * doesn't get stuck on the currently eligible stop task.
5157 * We're currently inside stop_machine() and the rq is either stuck
5158 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5159 * either way we should never end up calling schedule() until we're
5164 /* Ensure any throttled groups are reachable by pick_next_task */
5165 unthrottle_offline_cfs_rqs(rq
);
5169 * There's this thread running, bail when that's the only
5172 if (rq
->nr_running
== 1)
5175 next
= pick_next_task(rq
);
5177 next
->sched_class
->put_prev_task(rq
, next
);
5179 /* Find suitable destination for @next, with force if needed. */
5180 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5181 raw_spin_unlock(&rq
->lock
);
5183 __migrate_task(next
, dead_cpu
, dest_cpu
);
5185 raw_spin_lock(&rq
->lock
);
5191 #endif /* CONFIG_HOTPLUG_CPU */
5193 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5195 static struct ctl_table sd_ctl_dir
[] = {
5197 .procname
= "sched_domain",
5203 static struct ctl_table sd_ctl_root
[] = {
5205 .procname
= "kernel",
5207 .child
= sd_ctl_dir
,
5212 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5214 struct ctl_table
*entry
=
5215 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5220 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5222 struct ctl_table
*entry
;
5225 * In the intermediate directories, both the child directory and
5226 * procname are dynamically allocated and could fail but the mode
5227 * will always be set. In the lowest directory the names are
5228 * static strings and all have proc handlers.
5230 for (entry
= *tablep
; entry
->mode
; entry
++) {
5232 sd_free_ctl_entry(&entry
->child
);
5233 if (entry
->proc_handler
== NULL
)
5234 kfree(entry
->procname
);
5242 set_table_entry(struct ctl_table
*entry
,
5243 const char *procname
, void *data
, int maxlen
,
5244 umode_t mode
, proc_handler
*proc_handler
)
5246 entry
->procname
= procname
;
5248 entry
->maxlen
= maxlen
;
5250 entry
->proc_handler
= proc_handler
;
5253 static struct ctl_table
*
5254 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5256 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5261 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5262 sizeof(long), 0644, proc_doulongvec_minmax
);
5263 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5264 sizeof(long), 0644, proc_doulongvec_minmax
);
5265 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5266 sizeof(int), 0644, proc_dointvec_minmax
);
5267 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5268 sizeof(int), 0644, proc_dointvec_minmax
);
5269 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5270 sizeof(int), 0644, proc_dointvec_minmax
);
5271 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5272 sizeof(int), 0644, proc_dointvec_minmax
);
5273 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5274 sizeof(int), 0644, proc_dointvec_minmax
);
5275 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5276 sizeof(int), 0644, proc_dointvec_minmax
);
5277 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5278 sizeof(int), 0644, proc_dointvec_minmax
);
5279 set_table_entry(&table
[9], "cache_nice_tries",
5280 &sd
->cache_nice_tries
,
5281 sizeof(int), 0644, proc_dointvec_minmax
);
5282 set_table_entry(&table
[10], "flags", &sd
->flags
,
5283 sizeof(int), 0644, proc_dointvec_minmax
);
5284 set_table_entry(&table
[11], "name", sd
->name
,
5285 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5286 /* &table[12] is terminator */
5291 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5293 struct ctl_table
*entry
, *table
;
5294 struct sched_domain
*sd
;
5295 int domain_num
= 0, i
;
5298 for_each_domain(cpu
, sd
)
5300 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5305 for_each_domain(cpu
, sd
) {
5306 snprintf(buf
, 32, "domain%d", i
);
5307 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5309 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5316 static struct ctl_table_header
*sd_sysctl_header
;
5317 static void register_sched_domain_sysctl(void)
5319 int i
, cpu_num
= num_possible_cpus();
5320 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5323 WARN_ON(sd_ctl_dir
[0].child
);
5324 sd_ctl_dir
[0].child
= entry
;
5329 for_each_possible_cpu(i
) {
5330 snprintf(buf
, 32, "cpu%d", i
);
5331 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5333 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5337 WARN_ON(sd_sysctl_header
);
5338 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5341 /* may be called multiple times per register */
5342 static void unregister_sched_domain_sysctl(void)
5344 if (sd_sysctl_header
)
5345 unregister_sysctl_table(sd_sysctl_header
);
5346 sd_sysctl_header
= NULL
;
5347 if (sd_ctl_dir
[0].child
)
5348 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5351 static void register_sched_domain_sysctl(void)
5354 static void unregister_sched_domain_sysctl(void)
5359 static void set_rq_online(struct rq
*rq
)
5362 const struct sched_class
*class;
5364 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5367 for_each_class(class) {
5368 if (class->rq_online
)
5369 class->rq_online(rq
);
5374 static void set_rq_offline(struct rq
*rq
)
5377 const struct sched_class
*class;
5379 for_each_class(class) {
5380 if (class->rq_offline
)
5381 class->rq_offline(rq
);
5384 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5390 * migration_call - callback that gets triggered when a CPU is added.
5391 * Here we can start up the necessary migration thread for the new CPU.
5393 static int __cpuinit
5394 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5396 int cpu
= (long)hcpu
;
5397 unsigned long flags
;
5398 struct rq
*rq
= cpu_rq(cpu
);
5400 switch (action
& ~CPU_TASKS_FROZEN
) {
5402 case CPU_UP_PREPARE
:
5403 rq
->calc_load_update
= calc_load_update
;
5407 /* Update our root-domain */
5408 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5410 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5414 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5417 #ifdef CONFIG_HOTPLUG_CPU
5419 sched_ttwu_pending();
5420 /* Update our root-domain */
5421 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5423 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5427 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5428 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5430 migrate_nr_uninterruptible(rq
);
5431 calc_global_load_remove(rq
);
5436 update_max_interval();
5442 * Register at high priority so that task migration (migrate_all_tasks)
5443 * happens before everything else. This has to be lower priority than
5444 * the notifier in the perf_event subsystem, though.
5446 static struct notifier_block __cpuinitdata migration_notifier
= {
5447 .notifier_call
= migration_call
,
5448 .priority
= CPU_PRI_MIGRATION
,
5451 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5452 unsigned long action
, void *hcpu
)
5454 switch (action
& ~CPU_TASKS_FROZEN
) {
5456 case CPU_DOWN_FAILED
:
5457 set_cpu_active((long)hcpu
, true);
5464 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5465 unsigned long action
, void *hcpu
)
5467 switch (action
& ~CPU_TASKS_FROZEN
) {
5468 case CPU_DOWN_PREPARE
:
5469 set_cpu_active((long)hcpu
, false);
5476 static int __init
migration_init(void)
5478 void *cpu
= (void *)(long)smp_processor_id();
5481 /* Initialize migration for the boot CPU */
5482 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5483 BUG_ON(err
== NOTIFY_BAD
);
5484 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5485 register_cpu_notifier(&migration_notifier
);
5487 /* Register cpu active notifiers */
5488 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5489 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5493 early_initcall(migration_init
);
5498 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5500 #ifdef CONFIG_SCHED_DEBUG
5502 static __read_mostly
int sched_domain_debug_enabled
;
5504 static int __init
sched_domain_debug_setup(char *str
)
5506 sched_domain_debug_enabled
= 1;
5510 early_param("sched_debug", sched_domain_debug_setup
);
5512 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5513 struct cpumask
*groupmask
)
5515 struct sched_group
*group
= sd
->groups
;
5518 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5519 cpumask_clear(groupmask
);
5521 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5523 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5524 printk("does not load-balance\n");
5526 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5531 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5533 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5534 printk(KERN_ERR
"ERROR: domain->span does not contain "
5537 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5538 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5542 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5546 printk(KERN_ERR
"ERROR: group is NULL\n");
5550 if (!group
->sgp
->power
) {
5551 printk(KERN_CONT
"\n");
5552 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5557 if (!cpumask_weight(sched_group_cpus(group
))) {
5558 printk(KERN_CONT
"\n");
5559 printk(KERN_ERR
"ERROR: empty group\n");
5563 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5564 printk(KERN_CONT
"\n");
5565 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5569 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5571 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5573 printk(KERN_CONT
" %s", str
);
5574 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5575 printk(KERN_CONT
" (cpu_power = %d)",
5579 group
= group
->next
;
5580 } while (group
!= sd
->groups
);
5581 printk(KERN_CONT
"\n");
5583 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5584 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5587 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5588 printk(KERN_ERR
"ERROR: parent span is not a superset "
5589 "of domain->span\n");
5593 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5597 if (!sched_domain_debug_enabled
)
5601 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5605 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5608 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5616 #else /* !CONFIG_SCHED_DEBUG */
5617 # define sched_domain_debug(sd, cpu) do { } while (0)
5618 #endif /* CONFIG_SCHED_DEBUG */
5620 static int sd_degenerate(struct sched_domain
*sd
)
5622 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5625 /* Following flags need at least 2 groups */
5626 if (sd
->flags
& (SD_LOAD_BALANCE
|
5627 SD_BALANCE_NEWIDLE
|
5631 SD_SHARE_PKG_RESOURCES
)) {
5632 if (sd
->groups
!= sd
->groups
->next
)
5636 /* Following flags don't use groups */
5637 if (sd
->flags
& (SD_WAKE_AFFINE
))
5644 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5646 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5648 if (sd_degenerate(parent
))
5651 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5654 /* Flags needing groups don't count if only 1 group in parent */
5655 if (parent
->groups
== parent
->groups
->next
) {
5656 pflags
&= ~(SD_LOAD_BALANCE
|
5657 SD_BALANCE_NEWIDLE
|
5661 SD_SHARE_PKG_RESOURCES
);
5662 if (nr_node_ids
== 1)
5663 pflags
&= ~SD_SERIALIZE
;
5665 if (~cflags
& pflags
)
5671 static void free_rootdomain(struct rcu_head
*rcu
)
5673 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5675 cpupri_cleanup(&rd
->cpupri
);
5676 free_cpumask_var(rd
->rto_mask
);
5677 free_cpumask_var(rd
->online
);
5678 free_cpumask_var(rd
->span
);
5682 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5684 struct root_domain
*old_rd
= NULL
;
5685 unsigned long flags
;
5687 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5692 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5695 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5698 * If we dont want to free the old_rt yet then
5699 * set old_rd to NULL to skip the freeing later
5702 if (!atomic_dec_and_test(&old_rd
->refcount
))
5706 atomic_inc(&rd
->refcount
);
5709 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5710 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5713 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5716 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5719 static int init_rootdomain(struct root_domain
*rd
)
5721 memset(rd
, 0, sizeof(*rd
));
5723 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5725 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5727 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5730 if (cpupri_init(&rd
->cpupri
) != 0)
5735 free_cpumask_var(rd
->rto_mask
);
5737 free_cpumask_var(rd
->online
);
5739 free_cpumask_var(rd
->span
);
5745 * By default the system creates a single root-domain with all cpus as
5746 * members (mimicking the global state we have today).
5748 struct root_domain def_root_domain
;
5750 static void init_defrootdomain(void)
5752 init_rootdomain(&def_root_domain
);
5754 atomic_set(&def_root_domain
.refcount
, 1);
5757 static struct root_domain
*alloc_rootdomain(void)
5759 struct root_domain
*rd
;
5761 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5765 if (init_rootdomain(rd
) != 0) {
5773 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5775 struct sched_group
*tmp
, *first
;
5784 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5789 } while (sg
!= first
);
5792 static void free_sched_domain(struct rcu_head
*rcu
)
5794 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5797 * If its an overlapping domain it has private groups, iterate and
5800 if (sd
->flags
& SD_OVERLAP
) {
5801 free_sched_groups(sd
->groups
, 1);
5802 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5803 kfree(sd
->groups
->sgp
);
5809 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5811 call_rcu(&sd
->rcu
, free_sched_domain
);
5814 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5816 for (; sd
; sd
= sd
->parent
)
5817 destroy_sched_domain(sd
, cpu
);
5821 * Keep a special pointer to the highest sched_domain that has
5822 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5823 * allows us to avoid some pointer chasing select_idle_sibling().
5825 * Also keep a unique ID per domain (we use the first cpu number in
5826 * the cpumask of the domain), this allows us to quickly tell if
5827 * two cpus are in the same cache domain, see cpus_share_cache().
5829 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5830 DEFINE_PER_CPU(int, sd_llc_id
);
5832 static void update_top_cache_domain(int cpu
)
5834 struct sched_domain
*sd
;
5837 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5839 id
= cpumask_first(sched_domain_span(sd
));
5841 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5842 per_cpu(sd_llc_id
, cpu
) = id
;
5846 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5847 * hold the hotplug lock.
5850 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5852 struct rq
*rq
= cpu_rq(cpu
);
5853 struct sched_domain
*tmp
;
5855 /* Remove the sched domains which do not contribute to scheduling. */
5856 for (tmp
= sd
; tmp
; ) {
5857 struct sched_domain
*parent
= tmp
->parent
;
5861 if (sd_parent_degenerate(tmp
, parent
)) {
5862 tmp
->parent
= parent
->parent
;
5864 parent
->parent
->child
= tmp
;
5865 destroy_sched_domain(parent
, cpu
);
5870 if (sd
&& sd_degenerate(sd
)) {
5873 destroy_sched_domain(tmp
, cpu
);
5878 sched_domain_debug(sd
, cpu
);
5880 rq_attach_root(rq
, rd
);
5882 rcu_assign_pointer(rq
->sd
, sd
);
5883 destroy_sched_domains(tmp
, cpu
);
5885 update_top_cache_domain(cpu
);
5888 /* cpus with isolated domains */
5889 static cpumask_var_t cpu_isolated_map
;
5891 /* Setup the mask of cpus configured for isolated domains */
5892 static int __init
isolated_cpu_setup(char *str
)
5894 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5895 cpulist_parse(str
, cpu_isolated_map
);
5899 __setup("isolcpus=", isolated_cpu_setup
);
5904 * find_next_best_node - find the next node to include in a sched_domain
5905 * @node: node whose sched_domain we're building
5906 * @used_nodes: nodes already in the sched_domain
5908 * Find the next node to include in a given scheduling domain. Simply
5909 * finds the closest node not already in the @used_nodes map.
5911 * Should use nodemask_t.
5913 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
5915 int i
, n
, val
, min_val
, best_node
= -1;
5919 for (i
= 0; i
< nr_node_ids
; i
++) {
5920 /* Start at @node */
5921 n
= (node
+ i
) % nr_node_ids
;
5923 if (!nr_cpus_node(n
))
5926 /* Skip already used nodes */
5927 if (node_isset(n
, *used_nodes
))
5930 /* Simple min distance search */
5931 val
= node_distance(node
, n
);
5933 if (val
< min_val
) {
5939 if (best_node
!= -1)
5940 node_set(best_node
, *used_nodes
);
5945 * sched_domain_node_span - get a cpumask for a node's sched_domain
5946 * @node: node whose cpumask we're constructing
5947 * @span: resulting cpumask
5949 * Given a node, construct a good cpumask for its sched_domain to span. It
5950 * should be one that prevents unnecessary balancing, but also spreads tasks
5953 static void sched_domain_node_span(int node
, struct cpumask
*span
)
5955 nodemask_t used_nodes
;
5958 cpumask_clear(span
);
5959 nodes_clear(used_nodes
);
5961 cpumask_or(span
, span
, cpumask_of_node(node
));
5962 node_set(node
, used_nodes
);
5964 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
5965 int next_node
= find_next_best_node(node
, &used_nodes
);
5968 cpumask_or(span
, span
, cpumask_of_node(next_node
));
5972 static const struct cpumask
*cpu_node_mask(int cpu
)
5974 lockdep_assert_held(&sched_domains_mutex
);
5976 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
5978 return sched_domains_tmpmask
;
5981 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
5983 return cpu_possible_mask
;
5985 #endif /* CONFIG_NUMA */
5987 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5989 return cpumask_of_node(cpu_to_node(cpu
));
5992 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
5995 struct sched_domain
**__percpu sd
;
5996 struct sched_group
**__percpu sg
;
5997 struct sched_group_power
**__percpu sgp
;
6001 struct sched_domain
** __percpu sd
;
6002 struct root_domain
*rd
;
6012 struct sched_domain_topology_level
;
6014 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
6015 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
6017 #define SDTL_OVERLAP 0x01
6019 struct sched_domain_topology_level
{
6020 sched_domain_init_f init
;
6021 sched_domain_mask_f mask
;
6023 struct sd_data data
;
6027 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6029 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6030 const struct cpumask
*span
= sched_domain_span(sd
);
6031 struct cpumask
*covered
= sched_domains_tmpmask
;
6032 struct sd_data
*sdd
= sd
->private;
6033 struct sched_domain
*child
;
6036 cpumask_clear(covered
);
6038 for_each_cpu(i
, span
) {
6039 struct cpumask
*sg_span
;
6041 if (cpumask_test_cpu(i
, covered
))
6044 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6045 GFP_KERNEL
, cpu_to_node(cpu
));
6050 sg_span
= sched_group_cpus(sg
);
6052 child
= *per_cpu_ptr(sdd
->sd
, i
);
6054 child
= child
->child
;
6055 cpumask_copy(sg_span
, sched_domain_span(child
));
6057 cpumask_set_cpu(i
, sg_span
);
6059 cpumask_or(covered
, covered
, sg_span
);
6061 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
6062 atomic_inc(&sg
->sgp
->ref
);
6064 if (cpumask_test_cpu(cpu
, sg_span
))
6074 sd
->groups
= groups
;
6079 free_sched_groups(first
, 0);
6084 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6086 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6087 struct sched_domain
*child
= sd
->child
;
6090 cpu
= cpumask_first(sched_domain_span(child
));
6093 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6094 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
6095 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
6102 * build_sched_groups will build a circular linked list of the groups
6103 * covered by the given span, and will set each group's ->cpumask correctly,
6104 * and ->cpu_power to 0.
6106 * Assumes the sched_domain tree is fully constructed
6109 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6111 struct sched_group
*first
= NULL
, *last
= NULL
;
6112 struct sd_data
*sdd
= sd
->private;
6113 const struct cpumask
*span
= sched_domain_span(sd
);
6114 struct cpumask
*covered
;
6117 get_group(cpu
, sdd
, &sd
->groups
);
6118 atomic_inc(&sd
->groups
->ref
);
6120 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
6123 lockdep_assert_held(&sched_domains_mutex
);
6124 covered
= sched_domains_tmpmask
;
6126 cpumask_clear(covered
);
6128 for_each_cpu(i
, span
) {
6129 struct sched_group
*sg
;
6130 int group
= get_group(i
, sdd
, &sg
);
6133 if (cpumask_test_cpu(i
, covered
))
6136 cpumask_clear(sched_group_cpus(sg
));
6139 for_each_cpu(j
, span
) {
6140 if (get_group(j
, sdd
, NULL
) != group
)
6143 cpumask_set_cpu(j
, covered
);
6144 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6159 * Initialize sched groups cpu_power.
6161 * cpu_power indicates the capacity of sched group, which is used while
6162 * distributing the load between different sched groups in a sched domain.
6163 * Typically cpu_power for all the groups in a sched domain will be same unless
6164 * there are asymmetries in the topology. If there are asymmetries, group
6165 * having more cpu_power will pickup more load compared to the group having
6168 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6170 struct sched_group
*sg
= sd
->groups
;
6172 WARN_ON(!sd
|| !sg
);
6175 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6177 } while (sg
!= sd
->groups
);
6179 if (cpu
!= group_first_cpu(sg
))
6182 update_group_power(sd
, cpu
);
6183 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6186 int __weak
arch_sd_sibling_asym_packing(void)
6188 return 0*SD_ASYM_PACKING
;
6192 * Initializers for schedule domains
6193 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6196 #ifdef CONFIG_SCHED_DEBUG
6197 # define SD_INIT_NAME(sd, type) sd->name = #type
6199 # define SD_INIT_NAME(sd, type) do { } while (0)
6202 #define SD_INIT_FUNC(type) \
6203 static noinline struct sched_domain * \
6204 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6206 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6207 *sd = SD_##type##_INIT; \
6208 SD_INIT_NAME(sd, type); \
6209 sd->private = &tl->data; \
6215 SD_INIT_FUNC(ALLNODES
)
6218 #ifdef CONFIG_SCHED_SMT
6219 SD_INIT_FUNC(SIBLING
)
6221 #ifdef CONFIG_SCHED_MC
6224 #ifdef CONFIG_SCHED_BOOK
6228 static int default_relax_domain_level
= -1;
6229 int sched_domain_level_max
;
6231 static int __init
setup_relax_domain_level(char *str
)
6235 val
= simple_strtoul(str
, NULL
, 0);
6236 if (val
< sched_domain_level_max
)
6237 default_relax_domain_level
= val
;
6241 __setup("relax_domain_level=", setup_relax_domain_level
);
6243 static void set_domain_attribute(struct sched_domain
*sd
,
6244 struct sched_domain_attr
*attr
)
6248 if (!attr
|| attr
->relax_domain_level
< 0) {
6249 if (default_relax_domain_level
< 0)
6252 request
= default_relax_domain_level
;
6254 request
= attr
->relax_domain_level
;
6255 if (request
< sd
->level
) {
6256 /* turn off idle balance on this domain */
6257 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6259 /* turn on idle balance on this domain */
6260 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6264 static void __sdt_free(const struct cpumask
*cpu_map
);
6265 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6267 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6268 const struct cpumask
*cpu_map
)
6272 if (!atomic_read(&d
->rd
->refcount
))
6273 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6275 free_percpu(d
->sd
); /* fall through */
6277 __sdt_free(cpu_map
); /* fall through */
6283 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6284 const struct cpumask
*cpu_map
)
6286 memset(d
, 0, sizeof(*d
));
6288 if (__sdt_alloc(cpu_map
))
6289 return sa_sd_storage
;
6290 d
->sd
= alloc_percpu(struct sched_domain
*);
6292 return sa_sd_storage
;
6293 d
->rd
= alloc_rootdomain();
6296 return sa_rootdomain
;
6300 * NULL the sd_data elements we've used to build the sched_domain and
6301 * sched_group structure so that the subsequent __free_domain_allocs()
6302 * will not free the data we're using.
6304 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6306 struct sd_data
*sdd
= sd
->private;
6308 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6309 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6311 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6312 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6314 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6315 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6318 #ifdef CONFIG_SCHED_SMT
6319 static const struct cpumask
*cpu_smt_mask(int cpu
)
6321 return topology_thread_cpumask(cpu
);
6326 * Topology list, bottom-up.
6328 static struct sched_domain_topology_level default_topology
[] = {
6329 #ifdef CONFIG_SCHED_SMT
6330 { sd_init_SIBLING
, cpu_smt_mask
, },
6332 #ifdef CONFIG_SCHED_MC
6333 { sd_init_MC
, cpu_coregroup_mask
, },
6335 #ifdef CONFIG_SCHED_BOOK
6336 { sd_init_BOOK
, cpu_book_mask
, },
6338 { sd_init_CPU
, cpu_cpu_mask
, },
6340 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
6341 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
6346 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6348 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6350 struct sched_domain_topology_level
*tl
;
6353 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6354 struct sd_data
*sdd
= &tl
->data
;
6356 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6360 sdd
->sg
= alloc_percpu(struct sched_group
*);
6364 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6368 for_each_cpu(j
, cpu_map
) {
6369 struct sched_domain
*sd
;
6370 struct sched_group
*sg
;
6371 struct sched_group_power
*sgp
;
6373 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6374 GFP_KERNEL
, cpu_to_node(j
));
6378 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6380 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6381 GFP_KERNEL
, cpu_to_node(j
));
6385 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6387 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
6388 GFP_KERNEL
, cpu_to_node(j
));
6392 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6399 static void __sdt_free(const struct cpumask
*cpu_map
)
6401 struct sched_domain_topology_level
*tl
;
6404 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6405 struct sd_data
*sdd
= &tl
->data
;
6407 for_each_cpu(j
, cpu_map
) {
6408 struct sched_domain
*sd
;
6411 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6412 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6413 free_sched_groups(sd
->groups
, 0);
6414 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6418 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6420 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6422 free_percpu(sdd
->sd
);
6424 free_percpu(sdd
->sg
);
6426 free_percpu(sdd
->sgp
);
6431 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6432 struct s_data
*d
, const struct cpumask
*cpu_map
,
6433 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6436 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6440 set_domain_attribute(sd
, attr
);
6441 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6443 sd
->level
= child
->level
+ 1;
6444 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6453 * Build sched domains for a given set of cpus and attach the sched domains
6454 * to the individual cpus
6456 static int build_sched_domains(const struct cpumask
*cpu_map
,
6457 struct sched_domain_attr
*attr
)
6459 enum s_alloc alloc_state
= sa_none
;
6460 struct sched_domain
*sd
;
6462 int i
, ret
= -ENOMEM
;
6464 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6465 if (alloc_state
!= sa_rootdomain
)
6468 /* Set up domains for cpus specified by the cpu_map. */
6469 for_each_cpu(i
, cpu_map
) {
6470 struct sched_domain_topology_level
*tl
;
6473 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6474 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6475 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6476 sd
->flags
|= SD_OVERLAP
;
6477 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6484 *per_cpu_ptr(d
.sd
, i
) = sd
;
6487 /* Build the groups for the domains */
6488 for_each_cpu(i
, cpu_map
) {
6489 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6490 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6491 if (sd
->flags
& SD_OVERLAP
) {
6492 if (build_overlap_sched_groups(sd
, i
))
6495 if (build_sched_groups(sd
, i
))
6501 /* Calculate CPU power for physical packages and nodes */
6502 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6503 if (!cpumask_test_cpu(i
, cpu_map
))
6506 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6507 claim_allocations(i
, sd
);
6508 init_sched_groups_power(i
, sd
);
6512 /* Attach the domains */
6514 for_each_cpu(i
, cpu_map
) {
6515 sd
= *per_cpu_ptr(d
.sd
, i
);
6516 cpu_attach_domain(sd
, d
.rd
, i
);
6522 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6526 static cpumask_var_t
*doms_cur
; /* current sched domains */
6527 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6528 static struct sched_domain_attr
*dattr_cur
;
6529 /* attribues of custom domains in 'doms_cur' */
6532 * Special case: If a kmalloc of a doms_cur partition (array of
6533 * cpumask) fails, then fallback to a single sched domain,
6534 * as determined by the single cpumask fallback_doms.
6536 static cpumask_var_t fallback_doms
;
6539 * arch_update_cpu_topology lets virtualized architectures update the
6540 * cpu core maps. It is supposed to return 1 if the topology changed
6541 * or 0 if it stayed the same.
6543 int __attribute__((weak
)) arch_update_cpu_topology(void)
6548 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6551 cpumask_var_t
*doms
;
6553 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6556 for (i
= 0; i
< ndoms
; i
++) {
6557 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6558 free_sched_domains(doms
, i
);
6565 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6568 for (i
= 0; i
< ndoms
; i
++)
6569 free_cpumask_var(doms
[i
]);
6574 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6575 * For now this just excludes isolated cpus, but could be used to
6576 * exclude other special cases in the future.
6578 static int init_sched_domains(const struct cpumask
*cpu_map
)
6582 arch_update_cpu_topology();
6584 doms_cur
= alloc_sched_domains(ndoms_cur
);
6586 doms_cur
= &fallback_doms
;
6587 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6589 err
= build_sched_domains(doms_cur
[0], NULL
);
6590 register_sched_domain_sysctl();
6596 * Detach sched domains from a group of cpus specified in cpu_map
6597 * These cpus will now be attached to the NULL domain
6599 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6604 for_each_cpu(i
, cpu_map
)
6605 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6609 /* handle null as "default" */
6610 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6611 struct sched_domain_attr
*new, int idx_new
)
6613 struct sched_domain_attr tmp
;
6620 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6621 new ? (new + idx_new
) : &tmp
,
6622 sizeof(struct sched_domain_attr
));
6626 * Partition sched domains as specified by the 'ndoms_new'
6627 * cpumasks in the array doms_new[] of cpumasks. This compares
6628 * doms_new[] to the current sched domain partitioning, doms_cur[].
6629 * It destroys each deleted domain and builds each new domain.
6631 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6632 * The masks don't intersect (don't overlap.) We should setup one
6633 * sched domain for each mask. CPUs not in any of the cpumasks will
6634 * not be load balanced. If the same cpumask appears both in the
6635 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6638 * The passed in 'doms_new' should be allocated using
6639 * alloc_sched_domains. This routine takes ownership of it and will
6640 * free_sched_domains it when done with it. If the caller failed the
6641 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6642 * and partition_sched_domains() will fallback to the single partition
6643 * 'fallback_doms', it also forces the domains to be rebuilt.
6645 * If doms_new == NULL it will be replaced with cpu_online_mask.
6646 * ndoms_new == 0 is a special case for destroying existing domains,
6647 * and it will not create the default domain.
6649 * Call with hotplug lock held
6651 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6652 struct sched_domain_attr
*dattr_new
)
6657 mutex_lock(&sched_domains_mutex
);
6659 /* always unregister in case we don't destroy any domains */
6660 unregister_sched_domain_sysctl();
6662 /* Let architecture update cpu core mappings. */
6663 new_topology
= arch_update_cpu_topology();
6665 n
= doms_new
? ndoms_new
: 0;
6667 /* Destroy deleted domains */
6668 for (i
= 0; i
< ndoms_cur
; i
++) {
6669 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6670 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6671 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6674 /* no match - a current sched domain not in new doms_new[] */
6675 detach_destroy_domains(doms_cur
[i
]);
6680 if (doms_new
== NULL
) {
6682 doms_new
= &fallback_doms
;
6683 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6684 WARN_ON_ONCE(dattr_new
);
6687 /* Build new domains */
6688 for (i
= 0; i
< ndoms_new
; i
++) {
6689 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6690 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6691 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6694 /* no match - add a new doms_new */
6695 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6700 /* Remember the new sched domains */
6701 if (doms_cur
!= &fallback_doms
)
6702 free_sched_domains(doms_cur
, ndoms_cur
);
6703 kfree(dattr_cur
); /* kfree(NULL) is safe */
6704 doms_cur
= doms_new
;
6705 dattr_cur
= dattr_new
;
6706 ndoms_cur
= ndoms_new
;
6708 register_sched_domain_sysctl();
6710 mutex_unlock(&sched_domains_mutex
);
6713 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6714 static void reinit_sched_domains(void)
6718 /* Destroy domains first to force the rebuild */
6719 partition_sched_domains(0, NULL
, NULL
);
6721 rebuild_sched_domains();
6725 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6727 unsigned int level
= 0;
6729 if (sscanf(buf
, "%u", &level
) != 1)
6733 * level is always be positive so don't check for
6734 * level < POWERSAVINGS_BALANCE_NONE which is 0
6735 * What happens on 0 or 1 byte write,
6736 * need to check for count as well?
6739 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
6743 sched_smt_power_savings
= level
;
6745 sched_mc_power_savings
= level
;
6747 reinit_sched_domains();
6752 #ifdef CONFIG_SCHED_MC
6753 static ssize_t
sched_mc_power_savings_show(struct device
*dev
,
6754 struct device_attribute
*attr
,
6757 return sprintf(buf
, "%u\n", sched_mc_power_savings
);
6759 static ssize_t
sched_mc_power_savings_store(struct device
*dev
,
6760 struct device_attribute
*attr
,
6761 const char *buf
, size_t count
)
6763 return sched_power_savings_store(buf
, count
, 0);
6765 static DEVICE_ATTR(sched_mc_power_savings
, 0644,
6766 sched_mc_power_savings_show
,
6767 sched_mc_power_savings_store
);
6770 #ifdef CONFIG_SCHED_SMT
6771 static ssize_t
sched_smt_power_savings_show(struct device
*dev
,
6772 struct device_attribute
*attr
,
6775 return sprintf(buf
, "%u\n", sched_smt_power_savings
);
6777 static ssize_t
sched_smt_power_savings_store(struct device
*dev
,
6778 struct device_attribute
*attr
,
6779 const char *buf
, size_t count
)
6781 return sched_power_savings_store(buf
, count
, 1);
6783 static DEVICE_ATTR(sched_smt_power_savings
, 0644,
6784 sched_smt_power_savings_show
,
6785 sched_smt_power_savings_store
);
6788 int __init
sched_create_sysfs_power_savings_entries(struct device
*dev
)
6792 #ifdef CONFIG_SCHED_SMT
6794 err
= device_create_file(dev
, &dev_attr_sched_smt_power_savings
);
6796 #ifdef CONFIG_SCHED_MC
6797 if (!err
&& mc_capable())
6798 err
= device_create_file(dev
, &dev_attr_sched_mc_power_savings
);
6802 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6805 * Update cpusets according to cpu_active mask. If cpusets are
6806 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6807 * around partition_sched_domains().
6809 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6812 switch (action
& ~CPU_TASKS_FROZEN
) {
6814 case CPU_DOWN_FAILED
:
6815 cpuset_update_active_cpus();
6822 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6825 switch (action
& ~CPU_TASKS_FROZEN
) {
6826 case CPU_DOWN_PREPARE
:
6827 cpuset_update_active_cpus();
6834 void __init
sched_init_smp(void)
6836 cpumask_var_t non_isolated_cpus
;
6838 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6839 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6842 mutex_lock(&sched_domains_mutex
);
6843 init_sched_domains(cpu_active_mask
);
6844 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6845 if (cpumask_empty(non_isolated_cpus
))
6846 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6847 mutex_unlock(&sched_domains_mutex
);
6850 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6851 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6853 /* RT runtime code needs to handle some hotplug events */
6854 hotcpu_notifier(update_runtime
, 0);
6858 /* Move init over to a non-isolated CPU */
6859 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6861 sched_init_granularity();
6862 free_cpumask_var(non_isolated_cpus
);
6864 init_sched_rt_class();
6867 void __init
sched_init_smp(void)
6869 sched_init_granularity();
6871 #endif /* CONFIG_SMP */
6873 const_debug
unsigned int sysctl_timer_migration
= 1;
6875 int in_sched_functions(unsigned long addr
)
6877 return in_lock_functions(addr
) ||
6878 (addr
>= (unsigned long)__sched_text_start
6879 && addr
< (unsigned long)__sched_text_end
);
6882 #ifdef CONFIG_CGROUP_SCHED
6883 struct task_group root_task_group
;
6886 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6888 void __init
sched_init(void)
6891 unsigned long alloc_size
= 0, ptr
;
6893 #ifdef CONFIG_FAIR_GROUP_SCHED
6894 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6896 #ifdef CONFIG_RT_GROUP_SCHED
6897 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6899 #ifdef CONFIG_CPUMASK_OFFSTACK
6900 alloc_size
+= num_possible_cpus() * cpumask_size();
6903 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6905 #ifdef CONFIG_FAIR_GROUP_SCHED
6906 root_task_group
.se
= (struct sched_entity
**)ptr
;
6907 ptr
+= nr_cpu_ids
* sizeof(void **);
6909 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6910 ptr
+= nr_cpu_ids
* sizeof(void **);
6912 #endif /* CONFIG_FAIR_GROUP_SCHED */
6913 #ifdef CONFIG_RT_GROUP_SCHED
6914 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6915 ptr
+= nr_cpu_ids
* sizeof(void **);
6917 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6918 ptr
+= nr_cpu_ids
* sizeof(void **);
6920 #endif /* CONFIG_RT_GROUP_SCHED */
6921 #ifdef CONFIG_CPUMASK_OFFSTACK
6922 for_each_possible_cpu(i
) {
6923 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6924 ptr
+= cpumask_size();
6926 #endif /* CONFIG_CPUMASK_OFFSTACK */
6930 init_defrootdomain();
6933 init_rt_bandwidth(&def_rt_bandwidth
,
6934 global_rt_period(), global_rt_runtime());
6936 #ifdef CONFIG_RT_GROUP_SCHED
6937 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6938 global_rt_period(), global_rt_runtime());
6939 #endif /* CONFIG_RT_GROUP_SCHED */
6941 #ifdef CONFIG_CGROUP_SCHED
6942 list_add(&root_task_group
.list
, &task_groups
);
6943 INIT_LIST_HEAD(&root_task_group
.children
);
6944 INIT_LIST_HEAD(&root_task_group
.siblings
);
6945 autogroup_init(&init_task
);
6947 #endif /* CONFIG_CGROUP_SCHED */
6949 #ifdef CONFIG_CGROUP_CPUACCT
6950 root_cpuacct
.cpustat
= &kernel_cpustat
;
6951 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6952 /* Too early, not expected to fail */
6953 BUG_ON(!root_cpuacct
.cpuusage
);
6955 for_each_possible_cpu(i
) {
6959 raw_spin_lock_init(&rq
->lock
);
6961 rq
->calc_load_active
= 0;
6962 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6963 init_cfs_rq(&rq
->cfs
);
6964 init_rt_rq(&rq
->rt
, rq
);
6965 #ifdef CONFIG_FAIR_GROUP_SCHED
6966 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6967 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6969 * How much cpu bandwidth does root_task_group get?
6971 * In case of task-groups formed thr' the cgroup filesystem, it
6972 * gets 100% of the cpu resources in the system. This overall
6973 * system cpu resource is divided among the tasks of
6974 * root_task_group and its child task-groups in a fair manner,
6975 * based on each entity's (task or task-group's) weight
6976 * (se->load.weight).
6978 * In other words, if root_task_group has 10 tasks of weight
6979 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6980 * then A0's share of the cpu resource is:
6982 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6984 * We achieve this by letting root_task_group's tasks sit
6985 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6987 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6988 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6989 #endif /* CONFIG_FAIR_GROUP_SCHED */
6991 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6992 #ifdef CONFIG_RT_GROUP_SCHED
6993 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6994 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6997 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6998 rq
->cpu_load
[j
] = 0;
7000 rq
->last_load_update_tick
= jiffies
;
7005 rq
->cpu_power
= SCHED_POWER_SCALE
;
7006 rq
->post_schedule
= 0;
7007 rq
->active_balance
= 0;
7008 rq
->next_balance
= jiffies
;
7013 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7015 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7017 rq_attach_root(rq
, &def_root_domain
);
7023 atomic_set(&rq
->nr_iowait
, 0);
7026 set_load_weight(&init_task
);
7028 #ifdef CONFIG_PREEMPT_NOTIFIERS
7029 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7032 #ifdef CONFIG_RT_MUTEXES
7033 plist_head_init(&init_task
.pi_waiters
);
7037 * The boot idle thread does lazy MMU switching as well:
7039 atomic_inc(&init_mm
.mm_count
);
7040 enter_lazy_tlb(&init_mm
, current
);
7043 * Make us the idle thread. Technically, schedule() should not be
7044 * called from this thread, however somewhere below it might be,
7045 * but because we are the idle thread, we just pick up running again
7046 * when this runqueue becomes "idle".
7048 init_idle(current
, smp_processor_id());
7050 calc_load_update
= jiffies
+ LOAD_FREQ
;
7053 * During early bootup we pretend to be a normal task:
7055 current
->sched_class
= &fair_sched_class
;
7058 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7059 /* May be allocated at isolcpus cmdline parse time */
7060 if (cpu_isolated_map
== NULL
)
7061 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7063 init_sched_fair_class();
7065 scheduler_running
= 1;
7068 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7069 static inline int preempt_count_equals(int preempt_offset
)
7071 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7073 return (nested
== preempt_offset
);
7076 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7078 static unsigned long prev_jiffy
; /* ratelimiting */
7080 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7081 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7082 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7084 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7086 prev_jiffy
= jiffies
;
7089 "BUG: sleeping function called from invalid context at %s:%d\n",
7092 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7093 in_atomic(), irqs_disabled(),
7094 current
->pid
, current
->comm
);
7096 debug_show_held_locks(current
);
7097 if (irqs_disabled())
7098 print_irqtrace_events(current
);
7101 EXPORT_SYMBOL(__might_sleep
);
7104 #ifdef CONFIG_MAGIC_SYSRQ
7105 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7107 const struct sched_class
*prev_class
= p
->sched_class
;
7108 int old_prio
= p
->prio
;
7113 dequeue_task(rq
, p
, 0);
7114 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7116 enqueue_task(rq
, p
, 0);
7117 resched_task(rq
->curr
);
7120 check_class_changed(rq
, p
, prev_class
, old_prio
);
7123 void normalize_rt_tasks(void)
7125 struct task_struct
*g
, *p
;
7126 unsigned long flags
;
7129 read_lock_irqsave(&tasklist_lock
, flags
);
7130 do_each_thread(g
, p
) {
7132 * Only normalize user tasks:
7137 p
->se
.exec_start
= 0;
7138 #ifdef CONFIG_SCHEDSTATS
7139 p
->se
.statistics
.wait_start
= 0;
7140 p
->se
.statistics
.sleep_start
= 0;
7141 p
->se
.statistics
.block_start
= 0;
7146 * Renice negative nice level userspace
7149 if (TASK_NICE(p
) < 0 && p
->mm
)
7150 set_user_nice(p
, 0);
7154 raw_spin_lock(&p
->pi_lock
);
7155 rq
= __task_rq_lock(p
);
7157 normalize_task(rq
, p
);
7159 __task_rq_unlock(rq
);
7160 raw_spin_unlock(&p
->pi_lock
);
7161 } while_each_thread(g
, p
);
7163 read_unlock_irqrestore(&tasklist_lock
, flags
);
7166 #endif /* CONFIG_MAGIC_SYSRQ */
7168 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7170 * These functions are only useful for the IA64 MCA handling, or kdb.
7172 * They can only be called when the whole system has been
7173 * stopped - every CPU needs to be quiescent, and no scheduling
7174 * activity can take place. Using them for anything else would
7175 * be a serious bug, and as a result, they aren't even visible
7176 * under any other configuration.
7180 * curr_task - return the current task for a given cpu.
7181 * @cpu: the processor in question.
7183 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7185 struct task_struct
*curr_task(int cpu
)
7187 return cpu_curr(cpu
);
7190 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7194 * set_curr_task - set the current task for a given cpu.
7195 * @cpu: the processor in question.
7196 * @p: the task pointer to set.
7198 * Description: This function must only be used when non-maskable interrupts
7199 * are serviced on a separate stack. It allows the architecture to switch the
7200 * notion of the current task on a cpu in a non-blocking manner. This function
7201 * must be called with all CPU's synchronized, and interrupts disabled, the
7202 * and caller must save the original value of the current task (see
7203 * curr_task() above) and restore that value before reenabling interrupts and
7204 * re-starting the system.
7206 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7208 void set_curr_task(int cpu
, struct task_struct
*p
)
7215 #ifdef CONFIG_CGROUP_SCHED
7216 /* task_group_lock serializes the addition/removal of task groups */
7217 static DEFINE_SPINLOCK(task_group_lock
);
7219 static void free_sched_group(struct task_group
*tg
)
7221 free_fair_sched_group(tg
);
7222 free_rt_sched_group(tg
);
7227 /* allocate runqueue etc for a new task group */
7228 struct task_group
*sched_create_group(struct task_group
*parent
)
7230 struct task_group
*tg
;
7231 unsigned long flags
;
7233 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7235 return ERR_PTR(-ENOMEM
);
7237 if (!alloc_fair_sched_group(tg
, parent
))
7240 if (!alloc_rt_sched_group(tg
, parent
))
7243 spin_lock_irqsave(&task_group_lock
, flags
);
7244 list_add_rcu(&tg
->list
, &task_groups
);
7246 WARN_ON(!parent
); /* root should already exist */
7248 tg
->parent
= parent
;
7249 INIT_LIST_HEAD(&tg
->children
);
7250 list_add_rcu(&tg
->siblings
, &parent
->children
);
7251 spin_unlock_irqrestore(&task_group_lock
, flags
);
7256 free_sched_group(tg
);
7257 return ERR_PTR(-ENOMEM
);
7260 /* rcu callback to free various structures associated with a task group */
7261 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7263 /* now it should be safe to free those cfs_rqs */
7264 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7267 /* Destroy runqueue etc associated with a task group */
7268 void sched_destroy_group(struct task_group
*tg
)
7270 unsigned long flags
;
7273 /* end participation in shares distribution */
7274 for_each_possible_cpu(i
)
7275 unregister_fair_sched_group(tg
, i
);
7277 spin_lock_irqsave(&task_group_lock
, flags
);
7278 list_del_rcu(&tg
->list
);
7279 list_del_rcu(&tg
->siblings
);
7280 spin_unlock_irqrestore(&task_group_lock
, flags
);
7282 /* wait for possible concurrent references to cfs_rqs complete */
7283 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7286 /* change task's runqueue when it moves between groups.
7287 * The caller of this function should have put the task in its new group
7288 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7289 * reflect its new group.
7291 void sched_move_task(struct task_struct
*tsk
)
7294 unsigned long flags
;
7297 rq
= task_rq_lock(tsk
, &flags
);
7299 running
= task_current(rq
, tsk
);
7303 dequeue_task(rq
, tsk
, 0);
7304 if (unlikely(running
))
7305 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7307 #ifdef CONFIG_FAIR_GROUP_SCHED
7308 if (tsk
->sched_class
->task_move_group
)
7309 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7312 set_task_rq(tsk
, task_cpu(tsk
));
7314 if (unlikely(running
))
7315 tsk
->sched_class
->set_curr_task(rq
);
7317 enqueue_task(rq
, tsk
, 0);
7319 task_rq_unlock(rq
, tsk
, &flags
);
7321 #endif /* CONFIG_CGROUP_SCHED */
7323 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7324 static unsigned long to_ratio(u64 period
, u64 runtime
)
7326 if (runtime
== RUNTIME_INF
)
7329 return div64_u64(runtime
<< 20, period
);
7333 #ifdef CONFIG_RT_GROUP_SCHED
7335 * Ensure that the real time constraints are schedulable.
7337 static DEFINE_MUTEX(rt_constraints_mutex
);
7339 /* Must be called with tasklist_lock held */
7340 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7342 struct task_struct
*g
, *p
;
7344 do_each_thread(g
, p
) {
7345 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7347 } while_each_thread(g
, p
);
7352 struct rt_schedulable_data
{
7353 struct task_group
*tg
;
7358 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7360 struct rt_schedulable_data
*d
= data
;
7361 struct task_group
*child
;
7362 unsigned long total
, sum
= 0;
7363 u64 period
, runtime
;
7365 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7366 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7369 period
= d
->rt_period
;
7370 runtime
= d
->rt_runtime
;
7374 * Cannot have more runtime than the period.
7376 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7380 * Ensure we don't starve existing RT tasks.
7382 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7385 total
= to_ratio(period
, runtime
);
7388 * Nobody can have more than the global setting allows.
7390 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7394 * The sum of our children's runtime should not exceed our own.
7396 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7397 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7398 runtime
= child
->rt_bandwidth
.rt_runtime
;
7400 if (child
== d
->tg
) {
7401 period
= d
->rt_period
;
7402 runtime
= d
->rt_runtime
;
7405 sum
+= to_ratio(period
, runtime
);
7414 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7418 struct rt_schedulable_data data
= {
7420 .rt_period
= period
,
7421 .rt_runtime
= runtime
,
7425 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7431 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7432 u64 rt_period
, u64 rt_runtime
)
7436 mutex_lock(&rt_constraints_mutex
);
7437 read_lock(&tasklist_lock
);
7438 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7442 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7443 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7444 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7446 for_each_possible_cpu(i
) {
7447 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7449 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7450 rt_rq
->rt_runtime
= rt_runtime
;
7451 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7453 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7455 read_unlock(&tasklist_lock
);
7456 mutex_unlock(&rt_constraints_mutex
);
7461 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7463 u64 rt_runtime
, rt_period
;
7465 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7466 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7467 if (rt_runtime_us
< 0)
7468 rt_runtime
= RUNTIME_INF
;
7470 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7473 long sched_group_rt_runtime(struct task_group
*tg
)
7477 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7480 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7481 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7482 return rt_runtime_us
;
7485 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7487 u64 rt_runtime
, rt_period
;
7489 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7490 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7495 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7498 long sched_group_rt_period(struct task_group
*tg
)
7502 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7503 do_div(rt_period_us
, NSEC_PER_USEC
);
7504 return rt_period_us
;
7507 static int sched_rt_global_constraints(void)
7509 u64 runtime
, period
;
7512 if (sysctl_sched_rt_period
<= 0)
7515 runtime
= global_rt_runtime();
7516 period
= global_rt_period();
7519 * Sanity check on the sysctl variables.
7521 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7524 mutex_lock(&rt_constraints_mutex
);
7525 read_lock(&tasklist_lock
);
7526 ret
= __rt_schedulable(NULL
, 0, 0);
7527 read_unlock(&tasklist_lock
);
7528 mutex_unlock(&rt_constraints_mutex
);
7533 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7535 /* Don't accept realtime tasks when there is no way for them to run */
7536 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7542 #else /* !CONFIG_RT_GROUP_SCHED */
7543 static int sched_rt_global_constraints(void)
7545 unsigned long flags
;
7548 if (sysctl_sched_rt_period
<= 0)
7552 * There's always some RT tasks in the root group
7553 * -- migration, kstopmachine etc..
7555 if (sysctl_sched_rt_runtime
== 0)
7558 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7559 for_each_possible_cpu(i
) {
7560 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7562 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7563 rt_rq
->rt_runtime
= global_rt_runtime();
7564 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7566 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7570 #endif /* CONFIG_RT_GROUP_SCHED */
7572 int sched_rt_handler(struct ctl_table
*table
, int write
,
7573 void __user
*buffer
, size_t *lenp
,
7577 int old_period
, old_runtime
;
7578 static DEFINE_MUTEX(mutex
);
7581 old_period
= sysctl_sched_rt_period
;
7582 old_runtime
= sysctl_sched_rt_runtime
;
7584 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7586 if (!ret
&& write
) {
7587 ret
= sched_rt_global_constraints();
7589 sysctl_sched_rt_period
= old_period
;
7590 sysctl_sched_rt_runtime
= old_runtime
;
7592 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7593 def_rt_bandwidth
.rt_period
=
7594 ns_to_ktime(global_rt_period());
7597 mutex_unlock(&mutex
);
7602 #ifdef CONFIG_CGROUP_SCHED
7604 /* return corresponding task_group object of a cgroup */
7605 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7607 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7608 struct task_group
, css
);
7611 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7613 struct task_group
*tg
, *parent
;
7615 if (!cgrp
->parent
) {
7616 /* This is early initialization for the top cgroup */
7617 return &root_task_group
.css
;
7620 parent
= cgroup_tg(cgrp
->parent
);
7621 tg
= sched_create_group(parent
);
7623 return ERR_PTR(-ENOMEM
);
7628 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7630 struct task_group
*tg
= cgroup_tg(cgrp
);
7632 sched_destroy_group(tg
);
7635 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7636 struct cgroup_taskset
*tset
)
7638 struct task_struct
*task
;
7640 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7641 #ifdef CONFIG_RT_GROUP_SCHED
7642 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7645 /* We don't support RT-tasks being in separate groups */
7646 if (task
->sched_class
!= &fair_sched_class
)
7653 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7654 struct cgroup_taskset
*tset
)
7656 struct task_struct
*task
;
7658 cgroup_taskset_for_each(task
, cgrp
, tset
)
7659 sched_move_task(task
);
7663 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7664 struct task_struct
*task
)
7667 * cgroup_exit() is called in the copy_process() failure path.
7668 * Ignore this case since the task hasn't ran yet, this avoids
7669 * trying to poke a half freed task state from generic code.
7671 if (!(task
->flags
& PF_EXITING
))
7674 sched_move_task(task
);
7677 #ifdef CONFIG_FAIR_GROUP_SCHED
7678 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7681 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7684 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7686 struct task_group
*tg
= cgroup_tg(cgrp
);
7688 return (u64
) scale_load_down(tg
->shares
);
7691 #ifdef CONFIG_CFS_BANDWIDTH
7692 static DEFINE_MUTEX(cfs_constraints_mutex
);
7694 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7695 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7697 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7699 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7701 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7702 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7704 if (tg
== &root_task_group
)
7708 * Ensure we have at some amount of bandwidth every period. This is
7709 * to prevent reaching a state of large arrears when throttled via
7710 * entity_tick() resulting in prolonged exit starvation.
7712 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7716 * Likewise, bound things on the otherside by preventing insane quota
7717 * periods. This also allows us to normalize in computing quota
7720 if (period
> max_cfs_quota_period
)
7723 mutex_lock(&cfs_constraints_mutex
);
7724 ret
= __cfs_schedulable(tg
, period
, quota
);
7728 runtime_enabled
= quota
!= RUNTIME_INF
;
7729 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7730 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7731 raw_spin_lock_irq(&cfs_b
->lock
);
7732 cfs_b
->period
= ns_to_ktime(period
);
7733 cfs_b
->quota
= quota
;
7735 __refill_cfs_bandwidth_runtime(cfs_b
);
7736 /* restart the period timer (if active) to handle new period expiry */
7737 if (runtime_enabled
&& cfs_b
->timer_active
) {
7738 /* force a reprogram */
7739 cfs_b
->timer_active
= 0;
7740 __start_cfs_bandwidth(cfs_b
);
7742 raw_spin_unlock_irq(&cfs_b
->lock
);
7744 for_each_possible_cpu(i
) {
7745 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7746 struct rq
*rq
= cfs_rq
->rq
;
7748 raw_spin_lock_irq(&rq
->lock
);
7749 cfs_rq
->runtime_enabled
= runtime_enabled
;
7750 cfs_rq
->runtime_remaining
= 0;
7752 if (cfs_rq
->throttled
)
7753 unthrottle_cfs_rq(cfs_rq
);
7754 raw_spin_unlock_irq(&rq
->lock
);
7757 mutex_unlock(&cfs_constraints_mutex
);
7762 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7766 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7767 if (cfs_quota_us
< 0)
7768 quota
= RUNTIME_INF
;
7770 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7772 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7775 long tg_get_cfs_quota(struct task_group
*tg
)
7779 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7782 quota_us
= tg
->cfs_bandwidth
.quota
;
7783 do_div(quota_us
, NSEC_PER_USEC
);
7788 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7792 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7793 quota
= tg
->cfs_bandwidth
.quota
;
7795 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7798 long tg_get_cfs_period(struct task_group
*tg
)
7802 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7803 do_div(cfs_period_us
, NSEC_PER_USEC
);
7805 return cfs_period_us
;
7808 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7810 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7813 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7816 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7819 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7821 return tg_get_cfs_period(cgroup_tg(cgrp
));
7824 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7827 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7830 struct cfs_schedulable_data
{
7831 struct task_group
*tg
;
7836 * normalize group quota/period to be quota/max_period
7837 * note: units are usecs
7839 static u64
normalize_cfs_quota(struct task_group
*tg
,
7840 struct cfs_schedulable_data
*d
)
7848 period
= tg_get_cfs_period(tg
);
7849 quota
= tg_get_cfs_quota(tg
);
7852 /* note: these should typically be equivalent */
7853 if (quota
== RUNTIME_INF
|| quota
== -1)
7856 return to_ratio(period
, quota
);
7859 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7861 struct cfs_schedulable_data
*d
= data
;
7862 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7863 s64 quota
= 0, parent_quota
= -1;
7866 quota
= RUNTIME_INF
;
7868 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7870 quota
= normalize_cfs_quota(tg
, d
);
7871 parent_quota
= parent_b
->hierarchal_quota
;
7874 * ensure max(child_quota) <= parent_quota, inherit when no
7877 if (quota
== RUNTIME_INF
)
7878 quota
= parent_quota
;
7879 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7882 cfs_b
->hierarchal_quota
= quota
;
7887 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7890 struct cfs_schedulable_data data
= {
7896 if (quota
!= RUNTIME_INF
) {
7897 do_div(data
.period
, NSEC_PER_USEC
);
7898 do_div(data
.quota
, NSEC_PER_USEC
);
7902 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7908 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7909 struct cgroup_map_cb
*cb
)
7911 struct task_group
*tg
= cgroup_tg(cgrp
);
7912 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7914 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7915 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7916 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7920 #endif /* CONFIG_CFS_BANDWIDTH */
7921 #endif /* CONFIG_FAIR_GROUP_SCHED */
7923 #ifdef CONFIG_RT_GROUP_SCHED
7924 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7927 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7930 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7932 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7935 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7938 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7941 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7943 return sched_group_rt_period(cgroup_tg(cgrp
));
7945 #endif /* CONFIG_RT_GROUP_SCHED */
7947 static struct cftype cpu_files
[] = {
7948 #ifdef CONFIG_FAIR_GROUP_SCHED
7951 .read_u64
= cpu_shares_read_u64
,
7952 .write_u64
= cpu_shares_write_u64
,
7955 #ifdef CONFIG_CFS_BANDWIDTH
7957 .name
= "cfs_quota_us",
7958 .read_s64
= cpu_cfs_quota_read_s64
,
7959 .write_s64
= cpu_cfs_quota_write_s64
,
7962 .name
= "cfs_period_us",
7963 .read_u64
= cpu_cfs_period_read_u64
,
7964 .write_u64
= cpu_cfs_period_write_u64
,
7968 .read_map
= cpu_stats_show
,
7971 #ifdef CONFIG_RT_GROUP_SCHED
7973 .name
= "rt_runtime_us",
7974 .read_s64
= cpu_rt_runtime_read
,
7975 .write_s64
= cpu_rt_runtime_write
,
7978 .name
= "rt_period_us",
7979 .read_u64
= cpu_rt_period_read_uint
,
7980 .write_u64
= cpu_rt_period_write_uint
,
7985 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
7987 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
7990 struct cgroup_subsys cpu_cgroup_subsys
= {
7992 .create
= cpu_cgroup_create
,
7993 .destroy
= cpu_cgroup_destroy
,
7994 .can_attach
= cpu_cgroup_can_attach
,
7995 .attach
= cpu_cgroup_attach
,
7996 .exit
= cpu_cgroup_exit
,
7997 .populate
= cpu_cgroup_populate
,
7998 .subsys_id
= cpu_cgroup_subsys_id
,
8002 #endif /* CONFIG_CGROUP_SCHED */
8004 #ifdef CONFIG_CGROUP_CPUACCT
8007 * CPU accounting code for task groups.
8009 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8010 * (balbir@in.ibm.com).
8013 /* create a new cpu accounting group */
8014 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
8019 return &root_cpuacct
.css
;
8021 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8025 ca
->cpuusage
= alloc_percpu(u64
);
8029 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
8031 goto out_free_cpuusage
;
8036 free_percpu(ca
->cpuusage
);
8040 return ERR_PTR(-ENOMEM
);
8043 /* destroy an existing cpu accounting group */
8044 static void cpuacct_destroy(struct cgroup
*cgrp
)
8046 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8048 free_percpu(ca
->cpustat
);
8049 free_percpu(ca
->cpuusage
);
8053 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8055 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8058 #ifndef CONFIG_64BIT
8060 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8062 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8064 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8072 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8074 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8076 #ifndef CONFIG_64BIT
8078 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8080 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8082 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8088 /* return total cpu usage (in nanoseconds) of a group */
8089 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8091 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8092 u64 totalcpuusage
= 0;
8095 for_each_present_cpu(i
)
8096 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8098 return totalcpuusage
;
8101 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8104 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8113 for_each_present_cpu(i
)
8114 cpuacct_cpuusage_write(ca
, i
, 0);
8120 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8123 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8127 for_each_present_cpu(i
) {
8128 percpu
= cpuacct_cpuusage_read(ca
, i
);
8129 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8131 seq_printf(m
, "\n");
8135 static const char *cpuacct_stat_desc
[] = {
8136 [CPUACCT_STAT_USER
] = "user",
8137 [CPUACCT_STAT_SYSTEM
] = "system",
8140 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8141 struct cgroup_map_cb
*cb
)
8143 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8147 for_each_online_cpu(cpu
) {
8148 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8149 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
8150 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
8152 val
= cputime64_to_clock_t(val
);
8153 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
8156 for_each_online_cpu(cpu
) {
8157 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8158 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8159 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8160 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8163 val
= cputime64_to_clock_t(val
);
8164 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8169 static struct cftype files
[] = {
8172 .read_u64
= cpuusage_read
,
8173 .write_u64
= cpuusage_write
,
8176 .name
= "usage_percpu",
8177 .read_seq_string
= cpuacct_percpu_seq_read
,
8181 .read_map
= cpuacct_stats_show
,
8185 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8187 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8191 * charge this task's execution time to its accounting group.
8193 * called with rq->lock held.
8195 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8200 if (unlikely(!cpuacct_subsys
.active
))
8203 cpu
= task_cpu(tsk
);
8209 for (; ca
; ca
= parent_ca(ca
)) {
8210 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8211 *cpuusage
+= cputime
;
8217 struct cgroup_subsys cpuacct_subsys
= {
8219 .create
= cpuacct_create
,
8220 .destroy
= cpuacct_destroy
,
8221 .populate
= cpuacct_populate
,
8222 .subsys_id
= cpuacct_subsys_id
,
8224 #endif /* CONFIG_CGROUP_CPUACCT */