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>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
95 ktime_t soft
, hard
, now
;
98 if (hrtimer_active(period_timer
))
101 now
= hrtimer_cb_get_time(period_timer
);
102 hrtimer_forward(period_timer
, now
, period
);
104 soft
= hrtimer_get_softexpires(period_timer
);
105 hard
= hrtimer_get_expires(period_timer
);
106 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
107 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
108 HRTIMER_MODE_ABS_PINNED
, 0);
112 DEFINE_MUTEX(sched_domains_mutex
);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
115 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
117 void update_rq_clock(struct rq
*rq
)
121 if (rq
->skip_clock_update
> 0)
124 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
126 update_rq_clock_task(rq
, delta
);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug
unsigned int sysctl_sched_features
=
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names
[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 static_key_disable(&sched_feat_keys
[i
]);
185 static void sched_feat_enable(int i
)
187 static_key_enable(&sched_feat_keys
[i
]);
190 static void sched_feat_disable(int i
) { };
191 static void sched_feat_enable(int i
) { };
192 #endif /* HAVE_JUMP_LABEL */
194 static int sched_feat_set(char *cmp
)
199 if (strncmp(cmp
, "NO_", 3) == 0) {
204 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
205 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
207 sysctl_sched_features
&= ~(1UL << i
);
208 sched_feat_disable(i
);
210 sysctl_sched_features
|= (1UL << i
);
211 sched_feat_enable(i
);
221 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
222 size_t cnt
, loff_t
*ppos
)
231 if (copy_from_user(&buf
, ubuf
, cnt
))
237 i
= sched_feat_set(cmp
);
238 if (i
== __SCHED_FEAT_NR
)
246 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
248 return single_open(filp
, sched_feat_show
, NULL
);
251 static const struct file_operations sched_feat_fops
= {
252 .open
= sched_feat_open
,
253 .write
= sched_feat_write
,
256 .release
= single_release
,
259 static __init
int sched_init_debug(void)
261 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
266 late_initcall(sched_init_debug
);
267 #endif /* CONFIG_SCHED_DEBUG */
270 * Number of tasks to iterate in a single balance run.
271 * Limited because this is done with IRQs disabled.
273 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
276 * period over which we average the RT time consumption, measured
281 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
284 * period over which we measure -rt task cpu usage in us.
287 unsigned int sysctl_sched_rt_period
= 1000000;
289 __read_mostly
int scheduler_running
;
292 * part of the period that we allow rt tasks to run in us.
295 int sysctl_sched_rt_runtime
= 950000;
300 * __task_rq_lock - lock the rq @p resides on.
302 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
307 lockdep_assert_held(&p
->pi_lock
);
311 raw_spin_lock(&rq
->lock
);
312 if (likely(rq
== task_rq(p
)))
314 raw_spin_unlock(&rq
->lock
);
319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
322 __acquires(p
->pi_lock
)
328 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
330 raw_spin_lock(&rq
->lock
);
331 if (likely(rq
== task_rq(p
)))
333 raw_spin_unlock(&rq
->lock
);
334 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
338 static void __task_rq_unlock(struct rq
*rq
)
341 raw_spin_unlock(&rq
->lock
);
345 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
347 __releases(p
->pi_lock
)
349 raw_spin_unlock(&rq
->lock
);
350 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
354 * this_rq_lock - lock this runqueue and disable interrupts.
356 static struct rq
*this_rq_lock(void)
363 raw_spin_lock(&rq
->lock
);
368 #ifdef CONFIG_SCHED_HRTICK
370 * Use HR-timers to deliver accurate preemption points.
372 * Its all a bit involved since we cannot program an hrt while holding the
373 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
376 * When we get rescheduled we reprogram the hrtick_timer outside of the
380 static void hrtick_clear(struct rq
*rq
)
382 if (hrtimer_active(&rq
->hrtick_timer
))
383 hrtimer_cancel(&rq
->hrtick_timer
);
387 * High-resolution timer tick.
388 * Runs from hardirq context with interrupts disabled.
390 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
392 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
394 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
396 raw_spin_lock(&rq
->lock
);
398 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
399 raw_spin_unlock(&rq
->lock
);
401 return HRTIMER_NORESTART
;
406 * called from hardirq (IPI) context
408 static void __hrtick_start(void *arg
)
412 raw_spin_lock(&rq
->lock
);
413 hrtimer_restart(&rq
->hrtick_timer
);
414 rq
->hrtick_csd_pending
= 0;
415 raw_spin_unlock(&rq
->lock
);
419 * Called to set the hrtick timer state.
421 * called with rq->lock held and irqs disabled
423 void hrtick_start(struct rq
*rq
, u64 delay
)
425 struct hrtimer
*timer
= &rq
->hrtick_timer
;
426 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
428 hrtimer_set_expires(timer
, time
);
430 if (rq
== this_rq()) {
431 hrtimer_restart(timer
);
432 } else if (!rq
->hrtick_csd_pending
) {
433 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
434 rq
->hrtick_csd_pending
= 1;
439 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
441 int cpu
= (int)(long)hcpu
;
444 case CPU_UP_CANCELED
:
445 case CPU_UP_CANCELED_FROZEN
:
446 case CPU_DOWN_PREPARE
:
447 case CPU_DOWN_PREPARE_FROZEN
:
449 case CPU_DEAD_FROZEN
:
450 hrtick_clear(cpu_rq(cpu
));
457 static __init
void init_hrtick(void)
459 hotcpu_notifier(hotplug_hrtick
, 0);
463 * Called to set the hrtick timer state.
465 * called with rq->lock held and irqs disabled
467 void hrtick_start(struct rq
*rq
, u64 delay
)
469 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
470 HRTIMER_MODE_REL_PINNED
, 0);
473 static inline void init_hrtick(void)
476 #endif /* CONFIG_SMP */
478 static void init_rq_hrtick(struct rq
*rq
)
481 rq
->hrtick_csd_pending
= 0;
483 rq
->hrtick_csd
.flags
= 0;
484 rq
->hrtick_csd
.func
= __hrtick_start
;
485 rq
->hrtick_csd
.info
= rq
;
488 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
489 rq
->hrtick_timer
.function
= hrtick
;
491 #else /* CONFIG_SCHED_HRTICK */
492 static inline void hrtick_clear(struct rq
*rq
)
496 static inline void init_rq_hrtick(struct rq
*rq
)
500 static inline void init_hrtick(void)
503 #endif /* CONFIG_SCHED_HRTICK */
506 * resched_task - mark a task 'to be rescheduled now'.
508 * On UP this means the setting of the need_resched flag, on SMP it
509 * might also involve a cross-CPU call to trigger the scheduler on
513 void resched_task(struct task_struct
*p
)
517 assert_raw_spin_locked(&task_rq(p
)->lock
);
519 if (test_tsk_need_resched(p
))
522 set_tsk_need_resched(p
);
525 if (cpu
== smp_processor_id())
528 /* NEED_RESCHED must be visible before we test polling */
530 if (!tsk_is_polling(p
))
531 smp_send_reschedule(cpu
);
534 void resched_cpu(int cpu
)
536 struct rq
*rq
= cpu_rq(cpu
);
539 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
541 resched_task(cpu_curr(cpu
));
542 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
545 #ifdef CONFIG_NO_HZ_COMMON
547 * In the semi idle case, use the nearest busy cpu for migrating timers
548 * from an idle cpu. This is good for power-savings.
550 * We don't do similar optimization for completely idle system, as
551 * selecting an idle cpu will add more delays to the timers than intended
552 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
554 int get_nohz_timer_target(void)
556 int cpu
= smp_processor_id();
558 struct sched_domain
*sd
;
561 for_each_domain(cpu
, sd
) {
562 for_each_cpu(i
, sched_domain_span(sd
)) {
574 * When add_timer_on() enqueues a timer into the timer wheel of an
575 * idle CPU then this timer might expire before the next timer event
576 * which is scheduled to wake up that CPU. In case of a completely
577 * idle system the next event might even be infinite time into the
578 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
579 * leaves the inner idle loop so the newly added timer is taken into
580 * account when the CPU goes back to idle and evaluates the timer
581 * wheel for the next timer event.
583 static void wake_up_idle_cpu(int cpu
)
585 struct rq
*rq
= cpu_rq(cpu
);
587 if (cpu
== smp_processor_id())
591 * This is safe, as this function is called with the timer
592 * wheel base lock of (cpu) held. When the CPU is on the way
593 * to idle and has not yet set rq->curr to idle then it will
594 * be serialized on the timer wheel base lock and take the new
595 * timer into account automatically.
597 if (rq
->curr
!= rq
->idle
)
601 * We can set TIF_RESCHED on the idle task of the other CPU
602 * lockless. The worst case is that the other CPU runs the
603 * idle task through an additional NOOP schedule()
605 set_tsk_need_resched(rq
->idle
);
607 /* NEED_RESCHED must be visible before we test polling */
609 if (!tsk_is_polling(rq
->idle
))
610 smp_send_reschedule(cpu
);
613 static bool wake_up_full_nohz_cpu(int cpu
)
615 if (tick_nohz_full_cpu(cpu
)) {
616 if (cpu
!= smp_processor_id() ||
617 tick_nohz_tick_stopped())
618 smp_send_reschedule(cpu
);
625 void wake_up_nohz_cpu(int cpu
)
627 if (!wake_up_full_nohz_cpu(cpu
))
628 wake_up_idle_cpu(cpu
);
631 static inline bool got_nohz_idle_kick(void)
633 int cpu
= smp_processor_id();
635 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
638 if (idle_cpu(cpu
) && !need_resched())
642 * We can't run Idle Load Balance on this CPU for this time so we
643 * cancel it and clear NOHZ_BALANCE_KICK
645 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
649 #else /* CONFIG_NO_HZ_COMMON */
651 static inline bool got_nohz_idle_kick(void)
656 #endif /* CONFIG_NO_HZ_COMMON */
658 #ifdef CONFIG_NO_HZ_FULL
659 bool sched_can_stop_tick(void)
665 /* Make sure rq->nr_running update is visible after the IPI */
668 /* More than one running task need preemption */
669 if (rq
->nr_running
> 1)
674 #endif /* CONFIG_NO_HZ_FULL */
676 void sched_avg_update(struct rq
*rq
)
678 s64 period
= sched_avg_period();
680 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
682 * Inline assembly required to prevent the compiler
683 * optimising this loop into a divmod call.
684 * See __iter_div_u64_rem() for another example of this.
686 asm("" : "+rm" (rq
->age_stamp
));
687 rq
->age_stamp
+= period
;
692 #else /* !CONFIG_SMP */
693 void resched_task(struct task_struct
*p
)
695 assert_raw_spin_locked(&task_rq(p
)->lock
);
696 set_tsk_need_resched(p
);
698 #endif /* CONFIG_SMP */
700 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
703 * Iterate task_group tree rooted at *from, calling @down when first entering a
704 * node and @up when leaving it for the final time.
706 * Caller must hold rcu_lock or sufficient equivalent.
708 int walk_tg_tree_from(struct task_group
*from
,
709 tg_visitor down
, tg_visitor up
, void *data
)
711 struct task_group
*parent
, *child
;
717 ret
= (*down
)(parent
, data
);
720 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
727 ret
= (*up
)(parent
, data
);
728 if (ret
|| parent
== from
)
732 parent
= parent
->parent
;
739 int tg_nop(struct task_group
*tg
, void *data
)
745 static void set_load_weight(struct task_struct
*p
)
747 int prio
= p
->static_prio
- MAX_RT_PRIO
;
748 struct load_weight
*load
= &p
->se
.load
;
751 * SCHED_IDLE tasks get minimal weight:
753 if (p
->policy
== SCHED_IDLE
) {
754 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
755 load
->inv_weight
= WMULT_IDLEPRIO
;
759 load
->weight
= scale_load(prio_to_weight
[prio
]);
760 load
->inv_weight
= prio_to_wmult
[prio
];
763 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 sched_info_queued(p
);
767 p
->sched_class
->enqueue_task(rq
, p
, flags
);
770 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
773 sched_info_dequeued(p
);
774 p
->sched_class
->dequeue_task(rq
, p
, flags
);
777 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
779 if (task_contributes_to_load(p
))
780 rq
->nr_uninterruptible
--;
782 enqueue_task(rq
, p
, flags
);
785 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
787 if (task_contributes_to_load(p
))
788 rq
->nr_uninterruptible
++;
790 dequeue_task(rq
, p
, flags
);
793 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
796 * In theory, the compile should just see 0 here, and optimize out the call
797 * to sched_rt_avg_update. But I don't trust it...
799 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
800 s64 steal
= 0, irq_delta
= 0;
802 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
803 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
806 * Since irq_time is only updated on {soft,}irq_exit, we might run into
807 * this case when a previous update_rq_clock() happened inside a
810 * When this happens, we stop ->clock_task and only update the
811 * prev_irq_time stamp to account for the part that fit, so that a next
812 * update will consume the rest. This ensures ->clock_task is
815 * It does however cause some slight miss-attribution of {soft,}irq
816 * time, a more accurate solution would be to update the irq_time using
817 * the current rq->clock timestamp, except that would require using
820 if (irq_delta
> delta
)
823 rq
->prev_irq_time
+= irq_delta
;
826 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
827 if (static_key_false((¶virt_steal_rq_enabled
))) {
830 steal
= paravirt_steal_clock(cpu_of(rq
));
831 steal
-= rq
->prev_steal_time_rq
;
833 if (unlikely(steal
> delta
))
836 st
= steal_ticks(steal
);
837 steal
= st
* TICK_NSEC
;
839 rq
->prev_steal_time_rq
+= steal
;
845 rq
->clock_task
+= delta
;
847 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
848 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
849 sched_rt_avg_update(rq
, irq_delta
+ steal
);
853 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
855 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
856 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
860 * Make it appear like a SCHED_FIFO task, its something
861 * userspace knows about and won't get confused about.
863 * Also, it will make PI more or less work without too
864 * much confusion -- but then, stop work should not
865 * rely on PI working anyway.
867 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
869 stop
->sched_class
= &stop_sched_class
;
872 cpu_rq(cpu
)->stop
= stop
;
876 * Reset it back to a normal scheduling class so that
877 * it can die in pieces.
879 old_stop
->sched_class
= &rt_sched_class
;
884 * __normal_prio - return the priority that is based on the static prio
886 static inline int __normal_prio(struct task_struct
*p
)
888 return p
->static_prio
;
892 * Calculate the expected normal priority: i.e. priority
893 * without taking RT-inheritance into account. Might be
894 * boosted by interactivity modifiers. Changes upon fork,
895 * setprio syscalls, and whenever the interactivity
896 * estimator recalculates.
898 static inline int normal_prio(struct task_struct
*p
)
902 if (task_has_rt_policy(p
))
903 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
905 prio
= __normal_prio(p
);
910 * Calculate the current priority, i.e. the priority
911 * taken into account by the scheduler. This value might
912 * be boosted by RT tasks, or might be boosted by
913 * interactivity modifiers. Will be RT if the task got
914 * RT-boosted. If not then it returns p->normal_prio.
916 static int effective_prio(struct task_struct
*p
)
918 p
->normal_prio
= normal_prio(p
);
920 * If we are RT tasks or we were boosted to RT priority,
921 * keep the priority unchanged. Otherwise, update priority
922 * to the normal priority:
924 if (!rt_prio(p
->prio
))
925 return p
->normal_prio
;
930 * task_curr - is this task currently executing on a CPU?
931 * @p: the task in question.
933 inline int task_curr(const struct task_struct
*p
)
935 return cpu_curr(task_cpu(p
)) == p
;
938 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
939 const struct sched_class
*prev_class
,
942 if (prev_class
!= p
->sched_class
) {
943 if (prev_class
->switched_from
)
944 prev_class
->switched_from(rq
, p
);
945 p
->sched_class
->switched_to(rq
, p
);
946 } else if (oldprio
!= p
->prio
)
947 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
950 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
952 const struct sched_class
*class;
954 if (p
->sched_class
== rq
->curr
->sched_class
) {
955 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
957 for_each_class(class) {
958 if (class == rq
->curr
->sched_class
)
960 if (class == p
->sched_class
) {
961 resched_task(rq
->curr
);
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
972 rq
->skip_clock_update
= 1;
975 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
977 void register_task_migration_notifier(struct notifier_block
*n
)
979 atomic_notifier_chain_register(&task_migration_notifier
, n
);
983 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
985 #ifdef CONFIG_SCHED_DEBUG
987 * We should never call set_task_cpu() on a blocked task,
988 * ttwu() will sort out the placement.
990 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
991 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
993 #ifdef CONFIG_LOCKDEP
995 * The caller should hold either p->pi_lock or rq->lock, when changing
996 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
998 * sched_move_task() holds both and thus holding either pins the cgroup,
1001 * Furthermore, all task_rq users should acquire both locks, see
1004 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1005 lockdep_is_held(&task_rq(p
)->lock
)));
1009 trace_sched_migrate_task(p
, new_cpu
);
1011 if (task_cpu(p
) != new_cpu
) {
1012 struct task_migration_notifier tmn
;
1014 if (p
->sched_class
->migrate_task_rq
)
1015 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1016 p
->se
.nr_migrations
++;
1017 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1020 tmn
.from_cpu
= task_cpu(p
);
1021 tmn
.to_cpu
= new_cpu
;
1023 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
1026 __set_task_cpu(p
, new_cpu
);
1029 struct migration_arg
{
1030 struct task_struct
*task
;
1034 static int migration_cpu_stop(void *data
);
1037 * wait_task_inactive - wait for a thread to unschedule.
1039 * If @match_state is nonzero, it's the @p->state value just checked and
1040 * not expected to change. If it changes, i.e. @p might have woken up,
1041 * then return zero. When we succeed in waiting for @p to be off its CPU,
1042 * we return a positive number (its total switch count). If a second call
1043 * a short while later returns the same number, the caller can be sure that
1044 * @p has remained unscheduled the whole time.
1046 * The caller must ensure that the task *will* unschedule sometime soon,
1047 * else this function might spin for a *long* time. This function can't
1048 * be called with interrupts off, or it may introduce deadlock with
1049 * smp_call_function() if an IPI is sent by the same process we are
1050 * waiting to become inactive.
1052 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1054 unsigned long flags
;
1061 * We do the initial early heuristics without holding
1062 * any task-queue locks at all. We'll only try to get
1063 * the runqueue lock when things look like they will
1069 * If the task is actively running on another CPU
1070 * still, just relax and busy-wait without holding
1073 * NOTE! Since we don't hold any locks, it's not
1074 * even sure that "rq" stays as the right runqueue!
1075 * But we don't care, since "task_running()" will
1076 * return false if the runqueue has changed and p
1077 * is actually now running somewhere else!
1079 while (task_running(rq
, p
)) {
1080 if (match_state
&& unlikely(p
->state
!= match_state
))
1086 * Ok, time to look more closely! We need the rq
1087 * lock now, to be *sure*. If we're wrong, we'll
1088 * just go back and repeat.
1090 rq
= task_rq_lock(p
, &flags
);
1091 trace_sched_wait_task(p
);
1092 running
= task_running(rq
, p
);
1095 if (!match_state
|| p
->state
== match_state
)
1096 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1097 task_rq_unlock(rq
, p
, &flags
);
1100 * If it changed from the expected state, bail out now.
1102 if (unlikely(!ncsw
))
1106 * Was it really running after all now that we
1107 * checked with the proper locks actually held?
1109 * Oops. Go back and try again..
1111 if (unlikely(running
)) {
1117 * It's not enough that it's not actively running,
1118 * it must be off the runqueue _entirely_, and not
1121 * So if it was still runnable (but just not actively
1122 * running right now), it's preempted, and we should
1123 * yield - it could be a while.
1125 if (unlikely(on_rq
)) {
1126 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1128 set_current_state(TASK_UNINTERRUPTIBLE
);
1129 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1134 * Ahh, all good. It wasn't running, and it wasn't
1135 * runnable, which means that it will never become
1136 * running in the future either. We're all done!
1145 * kick_process - kick a running thread to enter/exit the kernel
1146 * @p: the to-be-kicked thread
1148 * Cause a process which is running on another CPU to enter
1149 * kernel-mode, without any delay. (to get signals handled.)
1151 * NOTE: this function doesn't have to take the runqueue lock,
1152 * because all it wants to ensure is that the remote task enters
1153 * the kernel. If the IPI races and the task has been migrated
1154 * to another CPU then no harm is done and the purpose has been
1157 void kick_process(struct task_struct
*p
)
1163 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1164 smp_send_reschedule(cpu
);
1167 EXPORT_SYMBOL_GPL(kick_process
);
1168 #endif /* CONFIG_SMP */
1172 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1174 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1176 int nid
= cpu_to_node(cpu
);
1177 const struct cpumask
*nodemask
= NULL
;
1178 enum { cpuset
, possible
, fail
} state
= cpuset
;
1182 * If the node that the cpu is on has been offlined, cpu_to_node()
1183 * will return -1. There is no cpu on the node, and we should
1184 * select the cpu on the other node.
1187 nodemask
= cpumask_of_node(nid
);
1189 /* Look for allowed, online CPU in same node. */
1190 for_each_cpu(dest_cpu
, nodemask
) {
1191 if (!cpu_online(dest_cpu
))
1193 if (!cpu_active(dest_cpu
))
1195 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1201 /* Any allowed, online CPU? */
1202 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1203 if (!cpu_online(dest_cpu
))
1205 if (!cpu_active(dest_cpu
))
1212 /* No more Mr. Nice Guy. */
1213 cpuset_cpus_allowed_fallback(p
);
1218 do_set_cpus_allowed(p
, cpu_possible_mask
);
1229 if (state
!= cpuset
) {
1231 * Don't tell them about moving exiting tasks or
1232 * kernel threads (both mm NULL), since they never
1235 if (p
->mm
&& printk_ratelimit()) {
1236 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1237 task_pid_nr(p
), p
->comm
, cpu
);
1245 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1248 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1250 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1253 * In order not to call set_task_cpu() on a blocking task we need
1254 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1257 * Since this is common to all placement strategies, this lives here.
1259 * [ this allows ->select_task() to simply return task_cpu(p) and
1260 * not worry about this generic constraint ]
1262 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1264 cpu
= select_fallback_rq(task_cpu(p
), p
);
1269 static void update_avg(u64
*avg
, u64 sample
)
1271 s64 diff
= sample
- *avg
;
1277 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1279 #ifdef CONFIG_SCHEDSTATS
1280 struct rq
*rq
= this_rq();
1283 int this_cpu
= smp_processor_id();
1285 if (cpu
== this_cpu
) {
1286 schedstat_inc(rq
, ttwu_local
);
1287 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1289 struct sched_domain
*sd
;
1291 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1293 for_each_domain(this_cpu
, sd
) {
1294 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1295 schedstat_inc(sd
, ttwu_wake_remote
);
1302 if (wake_flags
& WF_MIGRATED
)
1303 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1305 #endif /* CONFIG_SMP */
1307 schedstat_inc(rq
, ttwu_count
);
1308 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1310 if (wake_flags
& WF_SYNC
)
1311 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1313 #endif /* CONFIG_SCHEDSTATS */
1316 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1318 activate_task(rq
, p
, en_flags
);
1321 /* if a worker is waking up, notify workqueue */
1322 if (p
->flags
& PF_WQ_WORKER
)
1323 wq_worker_waking_up(p
, cpu_of(rq
));
1327 * Mark the task runnable and perform wakeup-preemption.
1330 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1332 check_preempt_curr(rq
, p
, wake_flags
);
1333 trace_sched_wakeup(p
, true);
1335 p
->state
= TASK_RUNNING
;
1337 if (p
->sched_class
->task_woken
)
1338 p
->sched_class
->task_woken(rq
, p
);
1340 if (rq
->idle_stamp
) {
1341 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1342 u64 max
= 2*sysctl_sched_migration_cost
;
1347 update_avg(&rq
->avg_idle
, delta
);
1354 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1357 if (p
->sched_contributes_to_load
)
1358 rq
->nr_uninterruptible
--;
1361 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1362 ttwu_do_wakeup(rq
, p
, wake_flags
);
1366 * Called in case the task @p isn't fully descheduled from its runqueue,
1367 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1368 * since all we need to do is flip p->state to TASK_RUNNING, since
1369 * the task is still ->on_rq.
1371 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1376 rq
= __task_rq_lock(p
);
1378 ttwu_do_wakeup(rq
, p
, wake_flags
);
1381 __task_rq_unlock(rq
);
1387 static void sched_ttwu_pending(void)
1389 struct rq
*rq
= this_rq();
1390 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1391 struct task_struct
*p
;
1393 raw_spin_lock(&rq
->lock
);
1396 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1397 llist
= llist_next(llist
);
1398 ttwu_do_activate(rq
, p
, 0);
1401 raw_spin_unlock(&rq
->lock
);
1404 void scheduler_ipi(void)
1406 if (llist_empty(&this_rq()->wake_list
)
1407 && !tick_nohz_full_cpu(smp_processor_id())
1408 && !got_nohz_idle_kick())
1412 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1413 * traditionally all their work was done from the interrupt return
1414 * path. Now that we actually do some work, we need to make sure
1417 * Some archs already do call them, luckily irq_enter/exit nest
1420 * Arguably we should visit all archs and update all handlers,
1421 * however a fair share of IPIs are still resched only so this would
1422 * somewhat pessimize the simple resched case.
1425 tick_nohz_full_check();
1426 sched_ttwu_pending();
1429 * Check if someone kicked us for doing the nohz idle load balance.
1431 if (unlikely(got_nohz_idle_kick())) {
1432 this_rq()->idle_balance
= 1;
1433 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1438 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1440 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1441 smp_send_reschedule(cpu
);
1444 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1446 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1448 #endif /* CONFIG_SMP */
1450 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1452 struct rq
*rq
= cpu_rq(cpu
);
1454 #if defined(CONFIG_SMP)
1455 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1456 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1457 ttwu_queue_remote(p
, cpu
);
1462 raw_spin_lock(&rq
->lock
);
1463 ttwu_do_activate(rq
, p
, 0);
1464 raw_spin_unlock(&rq
->lock
);
1468 * try_to_wake_up - wake up a thread
1469 * @p: the thread to be awakened
1470 * @state: the mask of task states that can be woken
1471 * @wake_flags: wake modifier flags (WF_*)
1473 * Put it on the run-queue if it's not already there. The "current"
1474 * thread is always on the run-queue (except when the actual
1475 * re-schedule is in progress), and as such you're allowed to do
1476 * the simpler "current->state = TASK_RUNNING" to mark yourself
1477 * runnable without the overhead of this.
1479 * Returns %true if @p was woken up, %false if it was already running
1480 * or @state didn't match @p's state.
1483 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1485 unsigned long flags
;
1486 int cpu
, success
= 0;
1489 * If we are going to wake up a thread waiting for CONDITION we
1490 * need to ensure that CONDITION=1 done by the caller can not be
1491 * reordered with p->state check below. This pairs with mb() in
1492 * set_current_state() the waiting thread does.
1494 smp_mb__before_spinlock();
1495 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1496 if (!(p
->state
& state
))
1499 success
= 1; /* we're going to change ->state */
1503 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1504 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1505 * in smp_cond_load_acquire() below.
1507 * sched_ttwu_pending() try_to_wake_up()
1508 * [S] p->on_rq = 1; [L] P->state
1509 * UNLOCK rq->lock -----.
1513 * LOCK rq->lock -----'
1517 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1519 * Pairs with the UNLOCK+LOCK on rq->lock from the
1520 * last wakeup of our task and the schedule that got our task
1524 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1529 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1530 * possible to, falsely, observe p->on_cpu == 0.
1532 * One must be running (->on_cpu == 1) in order to remove oneself
1533 * from the runqueue.
1535 * [S] ->on_cpu = 1; [L] ->on_rq
1539 * [S] ->on_rq = 0; [L] ->on_cpu
1541 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1542 * from the consecutive calls to schedule(); the first switching to our
1543 * task, the second putting it to sleep.
1548 * If the owning (remote) cpu is still in the middle of schedule() with
1549 * this task as prev, wait until its done referencing the task.
1554 * Pairs with the smp_wmb() in finish_lock_switch().
1558 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1559 p
->state
= TASK_WAKING
;
1561 if (p
->sched_class
->task_waking
)
1562 p
->sched_class
->task_waking(p
);
1564 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1565 if (task_cpu(p
) != cpu
) {
1566 wake_flags
|= WF_MIGRATED
;
1567 set_task_cpu(p
, cpu
);
1569 #endif /* CONFIG_SMP */
1573 ttwu_stat(p
, cpu
, wake_flags
);
1575 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1581 * try_to_wake_up_local - try to wake up a local task with rq lock held
1582 * @p: the thread to be awakened
1584 * Put @p on the run-queue if it's not already there. The caller must
1585 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1588 static void try_to_wake_up_local(struct task_struct
*p
)
1590 struct rq
*rq
= task_rq(p
);
1592 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1593 WARN_ON_ONCE(p
== current
))
1596 lockdep_assert_held(&rq
->lock
);
1598 if (!raw_spin_trylock(&p
->pi_lock
)) {
1599 raw_spin_unlock(&rq
->lock
);
1600 raw_spin_lock(&p
->pi_lock
);
1601 raw_spin_lock(&rq
->lock
);
1604 if (!(p
->state
& TASK_NORMAL
))
1608 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1610 ttwu_do_wakeup(rq
, p
, 0);
1611 ttwu_stat(p
, smp_processor_id(), 0);
1613 raw_spin_unlock(&p
->pi_lock
);
1617 * wake_up_process - Wake up a specific process
1618 * @p: The process to be woken up.
1620 * Attempt to wake up the nominated process and move it to the set of runnable
1621 * processes. Returns 1 if the process was woken up, 0 if it was already
1624 * It may be assumed that this function implies a write memory barrier before
1625 * changing the task state if and only if any tasks are woken up.
1627 int wake_up_process(struct task_struct
*p
)
1629 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1631 EXPORT_SYMBOL(wake_up_process
);
1633 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1635 return try_to_wake_up(p
, state
, 0);
1639 * Perform scheduler related setup for a newly forked process p.
1640 * p is forked by current.
1642 * __sched_fork() is basic setup used by init_idle() too:
1644 static void __sched_fork(struct task_struct
*p
)
1649 p
->se
.exec_start
= 0;
1650 p
->se
.sum_exec_runtime
= 0;
1651 p
->se
.prev_sum_exec_runtime
= 0;
1652 p
->se
.nr_migrations
= 0;
1654 INIT_LIST_HEAD(&p
->se
.group_node
);
1657 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1658 * removed when useful for applications beyond shares distribution (e.g.
1661 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1662 p
->se
.avg
.runnable_avg_period
= 0;
1663 p
->se
.avg
.runnable_avg_sum
= 0;
1665 #ifdef CONFIG_SCHEDSTATS
1666 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1669 INIT_LIST_HEAD(&p
->rt
.run_list
);
1671 #ifdef CONFIG_PREEMPT_NOTIFIERS
1672 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1675 #ifdef CONFIG_NUMA_BALANCING
1676 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1677 p
->mm
->numa_next_scan
= jiffies
;
1678 p
->mm
->numa_next_reset
= jiffies
;
1679 p
->mm
->numa_scan_seq
= 0;
1682 p
->node_stamp
= 0ULL;
1683 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1684 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1685 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1686 p
->numa_work
.next
= &p
->numa_work
;
1687 #endif /* CONFIG_NUMA_BALANCING */
1690 #ifdef CONFIG_NUMA_BALANCING
1691 #ifdef CONFIG_SCHED_DEBUG
1692 void set_numabalancing_state(bool enabled
)
1695 sched_feat_set("NUMA");
1697 sched_feat_set("NO_NUMA");
1700 __read_mostly
bool numabalancing_enabled
;
1702 void set_numabalancing_state(bool enabled
)
1704 numabalancing_enabled
= enabled
;
1706 #endif /* CONFIG_SCHED_DEBUG */
1707 #endif /* CONFIG_NUMA_BALANCING */
1710 * fork()/clone()-time setup:
1712 void sched_fork(struct task_struct
*p
)
1714 unsigned long flags
;
1715 int cpu
= get_cpu();
1719 * We mark the process as running here. This guarantees that
1720 * nobody will actually run it, and a signal or other external
1721 * event cannot wake it up and insert it on the runqueue either.
1723 p
->state
= TASK_RUNNING
;
1726 * Make sure we do not leak PI boosting priority to the child.
1728 p
->prio
= current
->normal_prio
;
1731 * Revert to default priority/policy on fork if requested.
1733 if (unlikely(p
->sched_reset_on_fork
)) {
1734 if (task_has_rt_policy(p
)) {
1735 p
->policy
= SCHED_NORMAL
;
1736 p
->static_prio
= NICE_TO_PRIO(0);
1738 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1739 p
->static_prio
= NICE_TO_PRIO(0);
1741 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1745 * We don't need the reset flag anymore after the fork. It has
1746 * fulfilled its duty:
1748 p
->sched_reset_on_fork
= 0;
1751 if (!rt_prio(p
->prio
))
1752 p
->sched_class
= &fair_sched_class
;
1754 if (p
->sched_class
->task_fork
)
1755 p
->sched_class
->task_fork(p
);
1758 * The child is not yet in the pid-hash so no cgroup attach races,
1759 * and the cgroup is pinned to this child due to cgroup_fork()
1760 * is ran before sched_fork().
1762 * Silence PROVE_RCU.
1764 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1765 set_task_cpu(p
, cpu
);
1766 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1768 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1769 if (likely(sched_info_on()))
1770 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1772 #if defined(CONFIG_SMP)
1775 #ifdef CONFIG_PREEMPT_COUNT
1776 /* Want to start with kernel preemption disabled. */
1777 task_thread_info(p
)->preempt_count
= 1;
1780 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1787 * wake_up_new_task - wake up a newly created task for the first time.
1789 * This function will do some initial scheduler statistics housekeeping
1790 * that must be done for every newly created context, then puts the task
1791 * on the runqueue and wakes it.
1793 void wake_up_new_task(struct task_struct
*p
)
1795 unsigned long flags
;
1798 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1801 * Fork balancing, do it here and not earlier because:
1802 * - cpus_allowed can change in the fork path
1803 * - any previously selected cpu might disappear through hotplug
1805 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1808 rq
= __task_rq_lock(p
);
1809 activate_task(rq
, p
, 0);
1811 trace_sched_wakeup_new(p
, true);
1812 check_preempt_curr(rq
, p
, WF_FORK
);
1814 if (p
->sched_class
->task_woken
)
1815 p
->sched_class
->task_woken(rq
, p
);
1817 task_rq_unlock(rq
, p
, &flags
);
1820 #ifdef CONFIG_PREEMPT_NOTIFIERS
1823 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1824 * @notifier: notifier struct to register
1826 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1828 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1830 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1833 * preempt_notifier_unregister - no longer interested in preemption notifications
1834 * @notifier: notifier struct to unregister
1836 * This is safe to call from within a preemption notifier.
1838 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1840 hlist_del(¬ifier
->link
);
1842 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1844 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1846 struct preempt_notifier
*notifier
;
1848 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1849 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1853 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1854 struct task_struct
*next
)
1856 struct preempt_notifier
*notifier
;
1858 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1859 notifier
->ops
->sched_out(notifier
, next
);
1862 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1864 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1869 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1870 struct task_struct
*next
)
1874 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1877 * prepare_task_switch - prepare to switch tasks
1878 * @rq: the runqueue preparing to switch
1879 * @prev: the current task that is being switched out
1880 * @next: the task we are going to switch to.
1882 * This is called with the rq lock held and interrupts off. It must
1883 * be paired with a subsequent finish_task_switch after the context
1886 * prepare_task_switch sets up locking and calls architecture specific
1890 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1891 struct task_struct
*next
)
1893 trace_sched_switch(prev
, next
);
1894 sched_info_switch(prev
, next
);
1895 perf_event_task_sched_out(prev
, next
);
1896 fire_sched_out_preempt_notifiers(prev
, next
);
1897 prepare_lock_switch(rq
, next
);
1898 prepare_arch_switch(next
);
1902 * finish_task_switch - clean up after a task-switch
1903 * @rq: runqueue associated with task-switch
1904 * @prev: the thread we just switched away from.
1906 * finish_task_switch must be called after the context switch, paired
1907 * with a prepare_task_switch call before the context switch.
1908 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1909 * and do any other architecture-specific cleanup actions.
1911 * Note that we may have delayed dropping an mm in context_switch(). If
1912 * so, we finish that here outside of the runqueue lock. (Doing it
1913 * with the lock held can cause deadlocks; see schedule() for
1916 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1917 __releases(rq
->lock
)
1919 struct mm_struct
*mm
= rq
->prev_mm
;
1925 * A task struct has one reference for the use as "current".
1926 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1927 * schedule one last time. The schedule call will never return, and
1928 * the scheduled task must drop that reference.
1929 * The test for TASK_DEAD must occur while the runqueue locks are
1930 * still held, otherwise prev could be scheduled on another cpu, die
1931 * there before we look at prev->state, and then the reference would
1933 * Manfred Spraul <manfred@colorfullife.com>
1935 prev_state
= prev
->state
;
1936 vtime_task_switch(prev
);
1937 finish_arch_switch(prev
);
1938 perf_event_task_sched_in(prev
, current
);
1939 finish_lock_switch(rq
, prev
);
1940 finish_arch_post_lock_switch();
1942 fire_sched_in_preempt_notifiers(current
);
1945 if (unlikely(prev_state
== TASK_DEAD
)) {
1947 * Remove function-return probe instances associated with this
1948 * task and put them back on the free list.
1950 kprobe_flush_task(prev
);
1951 put_task_struct(prev
);
1954 tick_nohz_task_switch(current
);
1959 /* assumes rq->lock is held */
1960 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1962 if (prev
->sched_class
->pre_schedule
)
1963 prev
->sched_class
->pre_schedule(rq
, prev
);
1966 /* rq->lock is NOT held, but preemption is disabled */
1967 static inline void post_schedule(struct rq
*rq
)
1969 if (rq
->post_schedule
) {
1970 unsigned long flags
;
1972 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1973 if (rq
->curr
->sched_class
->post_schedule
)
1974 rq
->curr
->sched_class
->post_schedule(rq
);
1975 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1977 rq
->post_schedule
= 0;
1983 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1987 static inline void post_schedule(struct rq
*rq
)
1994 * schedule_tail - first thing a freshly forked thread must call.
1995 * @prev: the thread we just switched away from.
1997 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1998 __releases(rq
->lock
)
2000 struct rq
*rq
= this_rq();
2002 finish_task_switch(rq
, prev
);
2005 * FIXME: do we need to worry about rq being invalidated by the
2010 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2011 /* In this case, finish_task_switch does not reenable preemption */
2014 if (current
->set_child_tid
)
2015 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2019 * context_switch - switch to the new MM and the new
2020 * thread's register state.
2023 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2024 struct task_struct
*next
)
2026 struct mm_struct
*mm
, *oldmm
;
2028 prepare_task_switch(rq
, prev
, next
);
2031 oldmm
= prev
->active_mm
;
2033 * For paravirt, this is coupled with an exit in switch_to to
2034 * combine the page table reload and the switch backend into
2037 arch_start_context_switch(prev
);
2040 next
->active_mm
= oldmm
;
2041 atomic_inc(&oldmm
->mm_count
);
2042 enter_lazy_tlb(oldmm
, next
);
2044 switch_mm(oldmm
, mm
, next
);
2047 prev
->active_mm
= NULL
;
2048 rq
->prev_mm
= oldmm
;
2051 * Since the runqueue lock will be released by the next
2052 * task (which is an invalid locking op but in the case
2053 * of the scheduler it's an obvious special-case), so we
2054 * do an early lockdep release here:
2056 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2057 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2060 context_tracking_task_switch(prev
, next
);
2061 /* Here we just switch the register state and the stack. */
2062 switch_to(prev
, next
, prev
);
2066 * this_rq must be evaluated again because prev may have moved
2067 * CPUs since it called schedule(), thus the 'rq' on its stack
2068 * frame will be invalid.
2070 finish_task_switch(this_rq(), prev
);
2074 * nr_running and nr_context_switches:
2076 * externally visible scheduler statistics: current number of runnable
2077 * threads, total number of context switches performed since bootup.
2079 unsigned long nr_running(void)
2081 unsigned long i
, sum
= 0;
2083 for_each_online_cpu(i
)
2084 sum
+= cpu_rq(i
)->nr_running
;
2089 unsigned long long nr_context_switches(void)
2092 unsigned long long sum
= 0;
2094 for_each_possible_cpu(i
)
2095 sum
+= cpu_rq(i
)->nr_switches
;
2100 unsigned long nr_iowait(void)
2102 unsigned long i
, sum
= 0;
2104 for_each_possible_cpu(i
)
2105 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2110 unsigned long nr_iowait_cpu(int cpu
)
2112 struct rq
*this = cpu_rq(cpu
);
2113 return atomic_read(&this->nr_iowait
);
2116 unsigned long this_cpu_load(void)
2118 struct rq
*this = this_rq();
2119 return this->cpu_load
[0];
2124 * Global load-average calculations
2126 * We take a distributed and async approach to calculating the global load-avg
2127 * in order to minimize overhead.
2129 * The global load average is an exponentially decaying average of nr_running +
2130 * nr_uninterruptible.
2132 * Once every LOAD_FREQ:
2135 * for_each_possible_cpu(cpu)
2136 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2138 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2140 * Due to a number of reasons the above turns in the mess below:
2142 * - for_each_possible_cpu() is prohibitively expensive on machines with
2143 * serious number of cpus, therefore we need to take a distributed approach
2144 * to calculating nr_active.
2146 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2147 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2149 * So assuming nr_active := 0 when we start out -- true per definition, we
2150 * can simply take per-cpu deltas and fold those into a global accumulate
2151 * to obtain the same result. See calc_load_fold_active().
2153 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2154 * across the machine, we assume 10 ticks is sufficient time for every
2155 * cpu to have completed this task.
2157 * This places an upper-bound on the IRQ-off latency of the machine. Then
2158 * again, being late doesn't loose the delta, just wrecks the sample.
2160 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2161 * this would add another cross-cpu cacheline miss and atomic operation
2162 * to the wakeup path. Instead we increment on whatever cpu the task ran
2163 * when it went into uninterruptible state and decrement on whatever cpu
2164 * did the wakeup. This means that only the sum of nr_uninterruptible over
2165 * all cpus yields the correct result.
2167 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2170 /* Variables and functions for calc_load */
2171 static atomic_long_t calc_load_tasks
;
2172 static unsigned long calc_load_update
;
2173 unsigned long avenrun
[3];
2174 EXPORT_SYMBOL(avenrun
); /* should be removed */
2177 * get_avenrun - get the load average array
2178 * @loads: pointer to dest load array
2179 * @offset: offset to add
2180 * @shift: shift count to shift the result left
2182 * These values are estimates at best, so no need for locking.
2184 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2186 loads
[0] = (avenrun
[0] + offset
) << shift
;
2187 loads
[1] = (avenrun
[1] + offset
) << shift
;
2188 loads
[2] = (avenrun
[2] + offset
) << shift
;
2191 static long calc_load_fold_active(struct rq
*this_rq
)
2193 long nr_active
, delta
= 0;
2195 nr_active
= this_rq
->nr_running
;
2196 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2198 if (nr_active
!= this_rq
->calc_load_active
) {
2199 delta
= nr_active
- this_rq
->calc_load_active
;
2200 this_rq
->calc_load_active
= nr_active
;
2207 * a1 = a0 * e + a * (1 - e)
2209 static unsigned long
2210 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2213 load
+= active
* (FIXED_1
- exp
);
2214 load
+= 1UL << (FSHIFT
- 1);
2215 return load
>> FSHIFT
;
2218 #ifdef CONFIG_NO_HZ_COMMON
2220 * Handle NO_HZ for the global load-average.
2222 * Since the above described distributed algorithm to compute the global
2223 * load-average relies on per-cpu sampling from the tick, it is affected by
2226 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2227 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2228 * when we read the global state.
2230 * Obviously reality has to ruin such a delightfully simple scheme:
2232 * - When we go NO_HZ idle during the window, we can negate our sample
2233 * contribution, causing under-accounting.
2235 * We avoid this by keeping two idle-delta counters and flipping them
2236 * when the window starts, thus separating old and new NO_HZ load.
2238 * The only trick is the slight shift in index flip for read vs write.
2242 * |-|-----------|-|-----------|-|-----------|-|
2243 * r:0 0 1 1 0 0 1 1 0
2244 * w:0 1 1 0 0 1 1 0 0
2246 * This ensures we'll fold the old idle contribution in this window while
2247 * accumlating the new one.
2249 * - When we wake up from NO_HZ idle during the window, we push up our
2250 * contribution, since we effectively move our sample point to a known
2253 * This is solved by pushing the window forward, and thus skipping the
2254 * sample, for this cpu (effectively using the idle-delta for this cpu which
2255 * was in effect at the time the window opened). This also solves the issue
2256 * of having to deal with a cpu having been in NOHZ idle for multiple
2257 * LOAD_FREQ intervals.
2259 * When making the ILB scale, we should try to pull this in as well.
2261 static atomic_long_t calc_load_idle
[2];
2262 static int calc_load_idx
;
2264 static inline int calc_load_write_idx(void)
2266 int idx
= calc_load_idx
;
2269 * See calc_global_nohz(), if we observe the new index, we also
2270 * need to observe the new update time.
2275 * If the folding window started, make sure we start writing in the
2278 if (!time_before(jiffies
, calc_load_update
))
2284 static inline int calc_load_read_idx(void)
2286 return calc_load_idx
& 1;
2289 void calc_load_enter_idle(void)
2291 struct rq
*this_rq
= this_rq();
2295 * We're going into NOHZ mode, if there's any pending delta, fold it
2296 * into the pending idle delta.
2298 delta
= calc_load_fold_active(this_rq
);
2300 int idx
= calc_load_write_idx();
2301 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2305 void calc_load_exit_idle(void)
2307 struct rq
*this_rq
= this_rq();
2310 * If we're still before the sample window, we're done.
2312 if (time_before(jiffies
, this_rq
->calc_load_update
))
2316 * We woke inside or after the sample window, this means we're already
2317 * accounted through the nohz accounting, so skip the entire deal and
2318 * sync up for the next window.
2320 this_rq
->calc_load_update
= calc_load_update
;
2321 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2322 this_rq
->calc_load_update
+= LOAD_FREQ
;
2325 static long calc_load_fold_idle(void)
2327 int idx
= calc_load_read_idx();
2330 if (atomic_long_read(&calc_load_idle
[idx
]))
2331 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2337 * fixed_power_int - compute: x^n, in O(log n) time
2339 * @x: base of the power
2340 * @frac_bits: fractional bits of @x
2341 * @n: power to raise @x to.
2343 * By exploiting the relation between the definition of the natural power
2344 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2345 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2346 * (where: n_i \elem {0, 1}, the binary vector representing n),
2347 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2348 * of course trivially computable in O(log_2 n), the length of our binary
2351 static unsigned long
2352 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2354 unsigned long result
= 1UL << frac_bits
;
2359 result
+= 1UL << (frac_bits
- 1);
2360 result
>>= frac_bits
;
2366 x
+= 1UL << (frac_bits
- 1);
2374 * a1 = a0 * e + a * (1 - e)
2376 * a2 = a1 * e + a * (1 - e)
2377 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2378 * = a0 * e^2 + a * (1 - e) * (1 + e)
2380 * a3 = a2 * e + a * (1 - e)
2381 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2382 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2386 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2387 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2388 * = a0 * e^n + a * (1 - e^n)
2390 * [1] application of the geometric series:
2393 * S_n := \Sum x^i = -------------
2396 static unsigned long
2397 calc_load_n(unsigned long load
, unsigned long exp
,
2398 unsigned long active
, unsigned int n
)
2401 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2405 * NO_HZ can leave us missing all per-cpu ticks calling
2406 * calc_load_account_active(), but since an idle CPU folds its delta into
2407 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2408 * in the pending idle delta if our idle period crossed a load cycle boundary.
2410 * Once we've updated the global active value, we need to apply the exponential
2411 * weights adjusted to the number of cycles missed.
2413 static void calc_global_nohz(void)
2415 long delta
, active
, n
;
2417 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2419 * Catch-up, fold however many we are behind still
2421 delta
= jiffies
- calc_load_update
- 10;
2422 n
= 1 + (delta
/ LOAD_FREQ
);
2424 active
= atomic_long_read(&calc_load_tasks
);
2425 active
= active
> 0 ? active
* FIXED_1
: 0;
2427 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2428 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2429 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2431 calc_load_update
+= n
* LOAD_FREQ
;
2435 * Flip the idle index...
2437 * Make sure we first write the new time then flip the index, so that
2438 * calc_load_write_idx() will see the new time when it reads the new
2439 * index, this avoids a double flip messing things up.
2444 #else /* !CONFIG_NO_HZ_COMMON */
2446 static inline long calc_load_fold_idle(void) { return 0; }
2447 static inline void calc_global_nohz(void) { }
2449 #endif /* CONFIG_NO_HZ_COMMON */
2452 * calc_load - update the avenrun load estimates 10 ticks after the
2453 * CPUs have updated calc_load_tasks.
2455 void calc_global_load(unsigned long ticks
)
2459 if (time_before(jiffies
, calc_load_update
+ 10))
2463 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2465 delta
= calc_load_fold_idle();
2467 atomic_long_add(delta
, &calc_load_tasks
);
2469 active
= atomic_long_read(&calc_load_tasks
);
2470 active
= active
> 0 ? active
* FIXED_1
: 0;
2472 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2473 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2474 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2476 calc_load_update
+= LOAD_FREQ
;
2479 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2485 * Called from update_cpu_load() to periodically update this CPU's
2488 static void calc_load_account_active(struct rq
*this_rq
)
2492 if (time_before(jiffies
, this_rq
->calc_load_update
))
2495 delta
= calc_load_fold_active(this_rq
);
2497 atomic_long_add(delta
, &calc_load_tasks
);
2499 this_rq
->calc_load_update
+= LOAD_FREQ
;
2503 * End of global load-average stuff
2507 * The exact cpuload at various idx values, calculated at every tick would be
2508 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2510 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2511 * on nth tick when cpu may be busy, then we have:
2512 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2513 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2515 * decay_load_missed() below does efficient calculation of
2516 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2517 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2519 * The calculation is approximated on a 128 point scale.
2520 * degrade_zero_ticks is the number of ticks after which load at any
2521 * particular idx is approximated to be zero.
2522 * degrade_factor is a precomputed table, a row for each load idx.
2523 * Each column corresponds to degradation factor for a power of two ticks,
2524 * based on 128 point scale.
2526 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2527 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2529 * With this power of 2 load factors, we can degrade the load n times
2530 * by looking at 1 bits in n and doing as many mult/shift instead of
2531 * n mult/shifts needed by the exact degradation.
2533 #define DEGRADE_SHIFT 7
2534 static const unsigned char
2535 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2536 static const unsigned char
2537 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2538 {0, 0, 0, 0, 0, 0, 0, 0},
2539 {64, 32, 8, 0, 0, 0, 0, 0},
2540 {96, 72, 40, 12, 1, 0, 0},
2541 {112, 98, 75, 43, 15, 1, 0},
2542 {120, 112, 98, 76, 45, 16, 2} };
2545 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2546 * would be when CPU is idle and so we just decay the old load without
2547 * adding any new load.
2549 static unsigned long
2550 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2554 if (!missed_updates
)
2557 if (missed_updates
>= degrade_zero_ticks
[idx
])
2561 return load
>> missed_updates
;
2563 while (missed_updates
) {
2564 if (missed_updates
% 2)
2565 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2567 missed_updates
>>= 1;
2574 * Update rq->cpu_load[] statistics. This function is usually called every
2575 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2576 * every tick. We fix it up based on jiffies.
2578 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2579 unsigned long pending_updates
)
2583 this_rq
->nr_load_updates
++;
2585 /* Update our load: */
2586 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2587 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2588 unsigned long old_load
, new_load
;
2590 /* scale is effectively 1 << i now, and >> i divides by scale */
2592 old_load
= this_rq
->cpu_load
[i
];
2593 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2594 new_load
= this_load
;
2596 * Round up the averaging division if load is increasing. This
2597 * prevents us from getting stuck on 9 if the load is 10, for
2600 if (new_load
> old_load
)
2601 new_load
+= scale
- 1;
2603 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2606 sched_avg_update(this_rq
);
2609 #ifdef CONFIG_NO_HZ_COMMON
2611 * There is no sane way to deal with nohz on smp when using jiffies because the
2612 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2613 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2615 * Therefore we cannot use the delta approach from the regular tick since that
2616 * would seriously skew the load calculation. However we'll make do for those
2617 * updates happening while idle (nohz_idle_balance) or coming out of idle
2618 * (tick_nohz_idle_exit).
2620 * This means we might still be one tick off for nohz periods.
2624 * Called from nohz_idle_balance() to update the load ratings before doing the
2627 void update_idle_cpu_load(struct rq
*this_rq
)
2629 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2630 unsigned long load
= this_rq
->load
.weight
;
2631 unsigned long pending_updates
;
2634 * bail if there's load or we're actually up-to-date.
2636 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2639 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2640 this_rq
->last_load_update_tick
= curr_jiffies
;
2642 __update_cpu_load(this_rq
, load
, pending_updates
);
2646 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2648 void update_cpu_load_nohz(void)
2650 struct rq
*this_rq
= this_rq();
2651 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2652 unsigned long pending_updates
;
2654 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2657 raw_spin_lock(&this_rq
->lock
);
2658 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2659 if (pending_updates
) {
2660 this_rq
->last_load_update_tick
= curr_jiffies
;
2662 * We were idle, this means load 0, the current load might be
2663 * !0 due to remote wakeups and the sort.
2665 __update_cpu_load(this_rq
, 0, pending_updates
);
2667 raw_spin_unlock(&this_rq
->lock
);
2669 #endif /* CONFIG_NO_HZ_COMMON */
2672 * Called from scheduler_tick()
2674 static void update_cpu_load_active(struct rq
*this_rq
)
2677 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2679 this_rq
->last_load_update_tick
= jiffies
;
2680 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2682 calc_load_account_active(this_rq
);
2688 * sched_exec - execve() is a valuable balancing opportunity, because at
2689 * this point the task has the smallest effective memory and cache footprint.
2691 void sched_exec(void)
2693 struct task_struct
*p
= current
;
2694 unsigned long flags
;
2697 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2698 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2699 if (dest_cpu
== smp_processor_id())
2702 if (likely(cpu_active(dest_cpu
))) {
2703 struct migration_arg arg
= { p
, dest_cpu
};
2705 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2706 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2710 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2715 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2716 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2718 EXPORT_PER_CPU_SYMBOL(kstat
);
2719 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2722 * Return any ns on the sched_clock that have not yet been accounted in
2723 * @p in case that task is currently running.
2725 * Called with task_rq_lock() held on @rq.
2727 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2731 if (task_current(rq
, p
)) {
2732 update_rq_clock(rq
);
2733 ns
= rq
->clock_task
- p
->se
.exec_start
;
2741 unsigned long long task_delta_exec(struct task_struct
*p
)
2743 unsigned long flags
;
2747 rq
= task_rq_lock(p
, &flags
);
2748 ns
= do_task_delta_exec(p
, rq
);
2749 task_rq_unlock(rq
, p
, &flags
);
2755 * Return accounted runtime for the task.
2756 * In case the task is currently running, return the runtime plus current's
2757 * pending runtime that have not been accounted yet.
2759 unsigned long long task_sched_runtime(struct task_struct
*p
)
2761 unsigned long flags
;
2765 rq
= task_rq_lock(p
, &flags
);
2766 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2767 task_rq_unlock(rq
, p
, &flags
);
2773 * This function gets called by the timer code, with HZ frequency.
2774 * We call it with interrupts disabled.
2776 void scheduler_tick(void)
2778 int cpu
= smp_processor_id();
2779 struct rq
*rq
= cpu_rq(cpu
);
2780 struct task_struct
*curr
= rq
->curr
;
2784 raw_spin_lock(&rq
->lock
);
2785 update_rq_clock(rq
);
2786 update_cpu_load_active(rq
);
2787 curr
->sched_class
->task_tick(rq
, curr
, 0);
2788 raw_spin_unlock(&rq
->lock
);
2790 perf_event_task_tick();
2793 rq
->idle_balance
= idle_cpu(cpu
);
2794 trigger_load_balance(rq
, cpu
);
2796 rq_last_tick_reset(rq
);
2799 #ifdef CONFIG_NO_HZ_FULL
2801 * scheduler_tick_max_deferment
2803 * Keep at least one tick per second when a single
2804 * active task is running because the scheduler doesn't
2805 * yet completely support full dynticks environment.
2807 * This makes sure that uptime, CFS vruntime, load
2808 * balancing, etc... continue to move forward, even
2809 * with a very low granularity.
2811 u64
scheduler_tick_max_deferment(void)
2813 struct rq
*rq
= this_rq();
2814 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2816 next
= rq
->last_sched_tick
+ HZ
;
2818 if (time_before_eq(next
, now
))
2821 return jiffies_to_usecs(next
- now
) * NSEC_PER_USEC
;
2825 notrace
unsigned long get_parent_ip(unsigned long addr
)
2827 if (in_lock_functions(addr
)) {
2828 addr
= CALLER_ADDR2
;
2829 if (in_lock_functions(addr
))
2830 addr
= CALLER_ADDR3
;
2835 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2836 defined(CONFIG_PREEMPT_TRACER))
2838 void __kprobes
add_preempt_count(int val
)
2840 #ifdef CONFIG_DEBUG_PREEMPT
2844 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2847 preempt_count() += val
;
2848 #ifdef CONFIG_DEBUG_PREEMPT
2850 * Spinlock count overflowing soon?
2852 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2855 if (preempt_count() == val
)
2856 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2858 EXPORT_SYMBOL(add_preempt_count
);
2860 void __kprobes
sub_preempt_count(int val
)
2862 #ifdef CONFIG_DEBUG_PREEMPT
2866 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2869 * Is the spinlock portion underflowing?
2871 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2872 !(preempt_count() & PREEMPT_MASK
)))
2876 if (preempt_count() == val
)
2877 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2878 preempt_count() -= val
;
2880 EXPORT_SYMBOL(sub_preempt_count
);
2885 * Print scheduling while atomic bug:
2887 static noinline
void __schedule_bug(struct task_struct
*prev
)
2889 if (oops_in_progress
)
2892 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2893 prev
->comm
, prev
->pid
, preempt_count());
2895 debug_show_held_locks(prev
);
2897 if (irqs_disabled())
2898 print_irqtrace_events(prev
);
2900 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2904 * Various schedule()-time debugging checks and statistics:
2906 static inline void schedule_debug(struct task_struct
*prev
)
2909 * Test if we are atomic. Since do_exit() needs to call into
2910 * schedule() atomically, we ignore that path for now.
2911 * Otherwise, whine if we are scheduling when we should not be.
2913 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2914 __schedule_bug(prev
);
2917 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2919 schedstat_inc(this_rq(), sched_count
);
2922 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2924 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2925 update_rq_clock(rq
);
2926 prev
->sched_class
->put_prev_task(rq
, prev
);
2930 * Pick up the highest-prio task:
2932 static inline struct task_struct
*
2933 pick_next_task(struct rq
*rq
)
2935 const struct sched_class
*class;
2936 struct task_struct
*p
;
2939 * Optimization: we know that if all tasks are in
2940 * the fair class we can call that function directly:
2942 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2943 p
= fair_sched_class
.pick_next_task(rq
);
2948 for_each_class(class) {
2949 p
= class->pick_next_task(rq
);
2954 BUG(); /* the idle class will always have a runnable task */
2958 * __schedule() is the main scheduler function.
2960 * The main means of driving the scheduler and thus entering this function are:
2962 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2964 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2965 * paths. For example, see arch/x86/entry_64.S.
2967 * To drive preemption between tasks, the scheduler sets the flag in timer
2968 * interrupt handler scheduler_tick().
2970 * 3. Wakeups don't really cause entry into schedule(). They add a
2971 * task to the run-queue and that's it.
2973 * Now, if the new task added to the run-queue preempts the current
2974 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2975 * called on the nearest possible occasion:
2977 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2979 * - in syscall or exception context, at the next outmost
2980 * preempt_enable(). (this might be as soon as the wake_up()'s
2983 * - in IRQ context, return from interrupt-handler to
2984 * preemptible context
2986 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2989 * - cond_resched() call
2990 * - explicit schedule() call
2991 * - return from syscall or exception to user-space
2992 * - return from interrupt-handler to user-space
2994 static void __sched
__schedule(void)
2996 struct task_struct
*prev
, *next
;
2997 unsigned long *switch_count
;
3003 cpu
= smp_processor_id();
3005 rcu_note_context_switch(cpu
);
3008 schedule_debug(prev
);
3010 if (sched_feat(HRTICK
))
3014 * Make sure that signal_pending_state()->signal_pending() below
3015 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3016 * done by the caller to avoid the race with signal_wake_up().
3018 smp_mb__before_spinlock();
3019 raw_spin_lock_irq(&rq
->lock
);
3021 switch_count
= &prev
->nivcsw
;
3022 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3023 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3024 prev
->state
= TASK_RUNNING
;
3026 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3030 * If a worker went to sleep, notify and ask workqueue
3031 * whether it wants to wake up a task to maintain
3034 if (prev
->flags
& PF_WQ_WORKER
) {
3035 struct task_struct
*to_wakeup
;
3037 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3039 try_to_wake_up_local(to_wakeup
);
3042 switch_count
= &prev
->nvcsw
;
3045 pre_schedule(rq
, prev
);
3047 if (unlikely(!rq
->nr_running
))
3048 idle_balance(cpu
, rq
);
3050 put_prev_task(rq
, prev
);
3051 next
= pick_next_task(rq
);
3052 clear_tsk_need_resched(prev
);
3053 rq
->skip_clock_update
= 0;
3055 if (likely(prev
!= next
)) {
3060 context_switch(rq
, prev
, next
); /* unlocks the rq */
3062 * The context switch have flipped the stack from under us
3063 * and restored the local variables which were saved when
3064 * this task called schedule() in the past. prev == current
3065 * is still correct, but it can be moved to another cpu/rq.
3067 cpu
= smp_processor_id();
3070 raw_spin_unlock_irq(&rq
->lock
);
3074 sched_preempt_enable_no_resched();
3079 static inline void sched_submit_work(struct task_struct
*tsk
)
3081 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3084 * If we are going to sleep and we have plugged IO queued,
3085 * make sure to submit it to avoid deadlocks.
3087 if (blk_needs_flush_plug(tsk
))
3088 blk_schedule_flush_plug(tsk
);
3091 asmlinkage
void __sched
schedule(void)
3093 struct task_struct
*tsk
= current
;
3095 sched_submit_work(tsk
);
3098 EXPORT_SYMBOL(schedule
);
3100 #ifdef CONFIG_CONTEXT_TRACKING
3101 asmlinkage
void __sched
schedule_user(void)
3104 * If we come here after a random call to set_need_resched(),
3105 * or we have been woken up remotely but the IPI has not yet arrived,
3106 * we haven't yet exited the RCU idle mode. Do it here manually until
3107 * we find a better solution.
3116 * schedule_preempt_disabled - called with preemption disabled
3118 * Returns with preemption disabled. Note: preempt_count must be 1
3120 void __sched
schedule_preempt_disabled(void)
3122 sched_preempt_enable_no_resched();
3127 #ifdef CONFIG_PREEMPT
3129 * this is the entry point to schedule() from in-kernel preemption
3130 * off of preempt_enable. Kernel preemptions off return from interrupt
3131 * occur there and call schedule directly.
3133 asmlinkage
void __sched notrace
preempt_schedule(void)
3135 struct thread_info
*ti
= current_thread_info();
3138 * If there is a non-zero preempt_count or interrupts are disabled,
3139 * we do not want to preempt the current task. Just return..
3141 if (likely(ti
->preempt_count
|| irqs_disabled()))
3145 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3147 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3150 * Check again in case we missed a preemption opportunity
3151 * between schedule and now.
3154 } while (need_resched());
3156 EXPORT_SYMBOL(preempt_schedule
);
3159 * this is the entry point to schedule() from kernel preemption
3160 * off of irq context.
3161 * Note, that this is called and return with irqs disabled. This will
3162 * protect us against recursive calling from irq.
3164 asmlinkage
void __sched
preempt_schedule_irq(void)
3166 struct thread_info
*ti
= current_thread_info();
3167 enum ctx_state prev_state
;
3169 /* Catch callers which need to be fixed */
3170 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3172 prev_state
= exception_enter();
3175 add_preempt_count(PREEMPT_ACTIVE
);
3178 local_irq_disable();
3179 sub_preempt_count(PREEMPT_ACTIVE
);
3182 * Check again in case we missed a preemption opportunity
3183 * between schedule and now.
3186 } while (need_resched());
3188 exception_exit(prev_state
);
3191 #endif /* CONFIG_PREEMPT */
3193 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3196 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3198 EXPORT_SYMBOL(default_wake_function
);
3201 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3202 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3203 * number) then we wake all the non-exclusive tasks and one exclusive task.
3205 * There are circumstances in which we can try to wake a task which has already
3206 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3207 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3209 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3210 int nr_exclusive
, int wake_flags
, void *key
)
3212 wait_queue_t
*curr
, *next
;
3214 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3215 unsigned flags
= curr
->flags
;
3217 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3218 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3224 * __wake_up - wake up threads blocked on a waitqueue.
3226 * @mode: which threads
3227 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3228 * @key: is directly passed to the wakeup function
3230 * It may be assumed that this function implies a write memory barrier before
3231 * changing the task state if and only if any tasks are woken up.
3233 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3234 int nr_exclusive
, void *key
)
3236 unsigned long flags
;
3238 spin_lock_irqsave(&q
->lock
, flags
);
3239 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3240 spin_unlock_irqrestore(&q
->lock
, flags
);
3242 EXPORT_SYMBOL(__wake_up
);
3245 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3247 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3249 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3251 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3253 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3255 __wake_up_common(q
, mode
, 1, 0, key
);
3257 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3260 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3262 * @mode: which threads
3263 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3264 * @key: opaque value to be passed to wakeup targets
3266 * The sync wakeup differs that the waker knows that it will schedule
3267 * away soon, so while the target thread will be woken up, it will not
3268 * be migrated to another CPU - ie. the two threads are 'synchronized'
3269 * with each other. This can prevent needless bouncing between CPUs.
3271 * On UP it can prevent extra preemption.
3273 * It may be assumed that this function implies a write memory barrier before
3274 * changing the task state if and only if any tasks are woken up.
3276 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3277 int nr_exclusive
, void *key
)
3279 unsigned long flags
;
3280 int wake_flags
= WF_SYNC
;
3285 if (unlikely(!nr_exclusive
))
3288 spin_lock_irqsave(&q
->lock
, flags
);
3289 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3290 spin_unlock_irqrestore(&q
->lock
, flags
);
3292 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3295 * __wake_up_sync - see __wake_up_sync_key()
3297 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3299 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3301 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3304 * complete: - signals a single thread waiting on this completion
3305 * @x: holds the state of this particular completion
3307 * This will wake up a single thread waiting on this completion. Threads will be
3308 * awakened in the same order in which they were queued.
3310 * See also complete_all(), wait_for_completion() and related routines.
3312 * It may be assumed that this function implies a write memory barrier before
3313 * changing the task state if and only if any tasks are woken up.
3315 void complete(struct completion
*x
)
3317 unsigned long flags
;
3319 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3321 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3322 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3324 EXPORT_SYMBOL(complete
);
3327 * complete_all: - signals all threads waiting on this completion
3328 * @x: holds the state of this particular completion
3330 * This will wake up all threads waiting on this particular completion event.
3332 * It may be assumed that this function implies a write memory barrier before
3333 * changing the task state if and only if any tasks are woken up.
3335 void complete_all(struct completion
*x
)
3337 unsigned long flags
;
3339 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3340 x
->done
+= UINT_MAX
/2;
3341 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3342 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3344 EXPORT_SYMBOL(complete_all
);
3346 static inline long __sched
3347 do_wait_for_common(struct completion
*x
,
3348 long (*action
)(long), long timeout
, int state
)
3351 DECLARE_WAITQUEUE(wait
, current
);
3353 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3355 if (signal_pending_state(state
, current
)) {
3356 timeout
= -ERESTARTSYS
;
3359 __set_current_state(state
);
3360 spin_unlock_irq(&x
->wait
.lock
);
3361 timeout
= action(timeout
);
3362 spin_lock_irq(&x
->wait
.lock
);
3363 } while (!x
->done
&& timeout
);
3364 __remove_wait_queue(&x
->wait
, &wait
);
3369 return timeout
?: 1;
3372 static inline long __sched
3373 __wait_for_common(struct completion
*x
,
3374 long (*action
)(long), long timeout
, int state
)
3378 spin_lock_irq(&x
->wait
.lock
);
3379 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
3380 spin_unlock_irq(&x
->wait
.lock
);
3385 wait_for_common(struct completion
*x
, long timeout
, int state
)
3387 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
3391 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
3393 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
3397 * wait_for_completion: - waits for completion of a task
3398 * @x: holds the state of this particular completion
3400 * This waits to be signaled for completion of a specific task. It is NOT
3401 * interruptible and there is no timeout.
3403 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3404 * and interrupt capability. Also see complete().
3406 void __sched
wait_for_completion(struct completion
*x
)
3408 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3410 EXPORT_SYMBOL(wait_for_completion
);
3413 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3414 * @x: holds the state of this particular completion
3415 * @timeout: timeout value in jiffies
3417 * This waits for either a completion of a specific task to be signaled or for a
3418 * specified timeout to expire. The timeout is in jiffies. It is not
3421 * The return value is 0 if timed out, and positive (at least 1, or number of
3422 * jiffies left till timeout) if completed.
3424 unsigned long __sched
3425 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3427 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3429 EXPORT_SYMBOL(wait_for_completion_timeout
);
3432 * wait_for_completion_io: - waits for completion of a task
3433 * @x: holds the state of this particular completion
3435 * This waits to be signaled for completion of a specific task. It is NOT
3436 * interruptible and there is no timeout. The caller is accounted as waiting
3439 void __sched
wait_for_completion_io(struct completion
*x
)
3441 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3443 EXPORT_SYMBOL(wait_for_completion_io
);
3446 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3447 * @x: holds the state of this particular completion
3448 * @timeout: timeout value in jiffies
3450 * This waits for either a completion of a specific task to be signaled or for a
3451 * specified timeout to expire. The timeout is in jiffies. It is not
3452 * interruptible. The caller is accounted as waiting for IO.
3454 * The return value is 0 if timed out, and positive (at least 1, or number of
3455 * jiffies left till timeout) if completed.
3457 unsigned long __sched
3458 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
3460 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3462 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
3465 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3466 * @x: holds the state of this particular completion
3468 * This waits for completion of a specific task to be signaled. It is
3471 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3473 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3475 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3476 if (t
== -ERESTARTSYS
)
3480 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3483 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3484 * @x: holds the state of this particular completion
3485 * @timeout: timeout value in jiffies
3487 * This waits for either a completion of a specific task to be signaled or for a
3488 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3490 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3491 * positive (at least 1, or number of jiffies left till timeout) if completed.
3494 wait_for_completion_interruptible_timeout(struct completion
*x
,
3495 unsigned long timeout
)
3497 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3499 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3502 * wait_for_completion_killable: - waits for completion of a task (killable)
3503 * @x: holds the state of this particular completion
3505 * This waits to be signaled for completion of a specific task. It can be
3506 * interrupted by a kill signal.
3508 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3510 int __sched
wait_for_completion_killable(struct completion
*x
)
3512 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3513 if (t
== -ERESTARTSYS
)
3517 EXPORT_SYMBOL(wait_for_completion_killable
);
3520 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3521 * @x: holds the state of this particular completion
3522 * @timeout: timeout value in jiffies
3524 * This waits for either a completion of a specific task to be
3525 * signaled or for a specified timeout to expire. It can be
3526 * interrupted by a kill signal. The timeout is in jiffies.
3528 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3529 * positive (at least 1, or number of jiffies left till timeout) if completed.
3532 wait_for_completion_killable_timeout(struct completion
*x
,
3533 unsigned long timeout
)
3535 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3537 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3540 * try_wait_for_completion - try to decrement a completion without blocking
3541 * @x: completion structure
3543 * Returns: 0 if a decrement cannot be done without blocking
3544 * 1 if a decrement succeeded.
3546 * If a completion is being used as a counting completion,
3547 * attempt to decrement the counter without blocking. This
3548 * enables us to avoid waiting if the resource the completion
3549 * is protecting is not available.
3551 bool try_wait_for_completion(struct completion
*x
)
3553 unsigned long flags
;
3556 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3561 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3564 EXPORT_SYMBOL(try_wait_for_completion
);
3567 * completion_done - Test to see if a completion has any waiters
3568 * @x: completion structure
3570 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3571 * 1 if there are no waiters.
3574 bool completion_done(struct completion
*x
)
3576 unsigned long flags
;
3579 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3582 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3585 EXPORT_SYMBOL(completion_done
);
3588 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3590 unsigned long flags
;
3593 init_waitqueue_entry(&wait
, current
);
3595 __set_current_state(state
);
3597 spin_lock_irqsave(&q
->lock
, flags
);
3598 __add_wait_queue(q
, &wait
);
3599 spin_unlock(&q
->lock
);
3600 timeout
= schedule_timeout(timeout
);
3601 spin_lock_irq(&q
->lock
);
3602 __remove_wait_queue(q
, &wait
);
3603 spin_unlock_irqrestore(&q
->lock
, flags
);
3608 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3610 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3612 EXPORT_SYMBOL(interruptible_sleep_on
);
3615 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3617 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3619 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3621 void __sched
sleep_on(wait_queue_head_t
*q
)
3623 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3625 EXPORT_SYMBOL(sleep_on
);
3627 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3629 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3631 EXPORT_SYMBOL(sleep_on_timeout
);
3633 #ifdef CONFIG_RT_MUTEXES
3636 * rt_mutex_setprio - set the current priority of a task
3638 * @prio: prio value (kernel-internal form)
3640 * This function changes the 'effective' priority of a task. It does
3641 * not touch ->normal_prio like __setscheduler().
3643 * Used by the rt_mutex code to implement priority inheritance logic.
3645 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3647 int oldprio
, on_rq
, running
;
3649 const struct sched_class
*prev_class
;
3651 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3653 rq
= __task_rq_lock(p
);
3656 * Idle task boosting is a nono in general. There is one
3657 * exception, when PREEMPT_RT and NOHZ is active:
3659 * The idle task calls get_next_timer_interrupt() and holds
3660 * the timer wheel base->lock on the CPU and another CPU wants
3661 * to access the timer (probably to cancel it). We can safely
3662 * ignore the boosting request, as the idle CPU runs this code
3663 * with interrupts disabled and will complete the lock
3664 * protected section without being interrupted. So there is no
3665 * real need to boost.
3667 if (unlikely(p
== rq
->idle
)) {
3668 WARN_ON(p
!= rq
->curr
);
3669 WARN_ON(p
->pi_blocked_on
);
3673 trace_sched_pi_setprio(p
, prio
);
3675 prev_class
= p
->sched_class
;
3677 running
= task_current(rq
, p
);
3679 dequeue_task(rq
, p
, 0);
3681 p
->sched_class
->put_prev_task(rq
, p
);
3684 p
->sched_class
= &rt_sched_class
;
3686 p
->sched_class
= &fair_sched_class
;
3691 p
->sched_class
->set_curr_task(rq
);
3693 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3695 check_class_changed(rq
, p
, prev_class
, oldprio
);
3697 __task_rq_unlock(rq
);
3700 void set_user_nice(struct task_struct
*p
, long nice
)
3702 int old_prio
, delta
, on_rq
;
3703 unsigned long flags
;
3706 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3709 * We have to be careful, if called from sys_setpriority(),
3710 * the task might be in the middle of scheduling on another CPU.
3712 rq
= task_rq_lock(p
, &flags
);
3714 * The RT priorities are set via sched_setscheduler(), but we still
3715 * allow the 'normal' nice value to be set - but as expected
3716 * it wont have any effect on scheduling until the task is
3717 * SCHED_FIFO/SCHED_RR:
3719 if (task_has_rt_policy(p
)) {
3720 p
->static_prio
= NICE_TO_PRIO(nice
);
3725 dequeue_task(rq
, p
, 0);
3727 p
->static_prio
= NICE_TO_PRIO(nice
);
3730 p
->prio
= effective_prio(p
);
3731 delta
= p
->prio
- old_prio
;
3734 enqueue_task(rq
, p
, 0);
3736 * If the task increased its priority or is running and
3737 * lowered its priority, then reschedule its CPU:
3739 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3740 resched_task(rq
->curr
);
3743 task_rq_unlock(rq
, p
, &flags
);
3745 EXPORT_SYMBOL(set_user_nice
);
3748 * can_nice - check if a task can reduce its nice value
3752 int can_nice(const struct task_struct
*p
, const int nice
)
3754 /* convert nice value [19,-20] to rlimit style value [1,40] */
3755 int nice_rlim
= 20 - nice
;
3757 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3758 capable(CAP_SYS_NICE
));
3761 #ifdef __ARCH_WANT_SYS_NICE
3764 * sys_nice - change the priority of the current process.
3765 * @increment: priority increment
3767 * sys_setpriority is a more generic, but much slower function that
3768 * does similar things.
3770 SYSCALL_DEFINE1(nice
, int, increment
)
3775 * Setpriority might change our priority at the same moment.
3776 * We don't have to worry. Conceptually one call occurs first
3777 * and we have a single winner.
3779 if (increment
< -40)
3784 nice
= TASK_NICE(current
) + increment
;
3790 if (increment
< 0 && !can_nice(current
, nice
))
3793 retval
= security_task_setnice(current
, nice
);
3797 set_user_nice(current
, nice
);
3804 * task_prio - return the priority value of a given task.
3805 * @p: the task in question.
3807 * This is the priority value as seen by users in /proc.
3808 * RT tasks are offset by -200. Normal tasks are centered
3809 * around 0, value goes from -16 to +15.
3811 int task_prio(const struct task_struct
*p
)
3813 return p
->prio
- MAX_RT_PRIO
;
3817 * task_nice - return the nice value of a given task.
3818 * @p: the task in question.
3820 int task_nice(const struct task_struct
*p
)
3822 return TASK_NICE(p
);
3824 EXPORT_SYMBOL(task_nice
);
3827 * idle_cpu - is a given cpu idle currently?
3828 * @cpu: the processor in question.
3830 int idle_cpu(int cpu
)
3832 struct rq
*rq
= cpu_rq(cpu
);
3834 if (rq
->curr
!= rq
->idle
)
3841 if (!llist_empty(&rq
->wake_list
))
3849 * idle_task - return the idle task for a given cpu.
3850 * @cpu: the processor in question.
3852 struct task_struct
*idle_task(int cpu
)
3854 return cpu_rq(cpu
)->idle
;
3858 * find_process_by_pid - find a process with a matching PID value.
3859 * @pid: the pid in question.
3861 static struct task_struct
*find_process_by_pid(pid_t pid
)
3863 return pid
? find_task_by_vpid(pid
) : current
;
3866 /* Actually do priority change: must hold rq lock. */
3868 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3871 p
->rt_priority
= prio
;
3872 p
->normal_prio
= normal_prio(p
);
3873 /* we are holding p->pi_lock already */
3874 p
->prio
= rt_mutex_getprio(p
);
3875 if (rt_prio(p
->prio
))
3876 p
->sched_class
= &rt_sched_class
;
3878 p
->sched_class
= &fair_sched_class
;
3883 * check the target process has a UID that matches the current process's
3885 static bool check_same_owner(struct task_struct
*p
)
3887 const struct cred
*cred
= current_cred(), *pcred
;
3891 pcred
= __task_cred(p
);
3892 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3893 uid_eq(cred
->euid
, pcred
->uid
));
3898 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3899 const struct sched_param
*param
, bool user
)
3901 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3902 unsigned long flags
;
3903 const struct sched_class
*prev_class
;
3907 /* may grab non-irq protected spin_locks */
3908 BUG_ON(in_interrupt());
3910 /* double check policy once rq lock held */
3912 reset_on_fork
= p
->sched_reset_on_fork
;
3913 policy
= oldpolicy
= p
->policy
;
3915 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3916 policy
&= ~SCHED_RESET_ON_FORK
;
3918 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3919 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3920 policy
!= SCHED_IDLE
)
3925 * Valid priorities for SCHED_FIFO and SCHED_RR are
3926 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3927 * SCHED_BATCH and SCHED_IDLE is 0.
3929 if (param
->sched_priority
< 0 ||
3930 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3931 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3933 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3937 * Allow unprivileged RT tasks to decrease priority:
3939 if (user
&& !capable(CAP_SYS_NICE
)) {
3940 if (rt_policy(policy
)) {
3941 unsigned long rlim_rtprio
=
3942 task_rlimit(p
, RLIMIT_RTPRIO
);
3944 /* can't set/change the rt policy */
3945 if (policy
!= p
->policy
&& !rlim_rtprio
)
3948 /* can't increase priority */
3949 if (param
->sched_priority
> p
->rt_priority
&&
3950 param
->sched_priority
> rlim_rtprio
)
3955 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3956 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3958 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3959 if (!can_nice(p
, TASK_NICE(p
)))
3963 /* can't change other user's priorities */
3964 if (!check_same_owner(p
))
3967 /* Normal users shall not reset the sched_reset_on_fork flag */
3968 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3973 retval
= security_task_setscheduler(p
);
3979 * make sure no PI-waiters arrive (or leave) while we are
3980 * changing the priority of the task:
3982 * To be able to change p->policy safely, the appropriate
3983 * runqueue lock must be held.
3985 rq
= task_rq_lock(p
, &flags
);
3988 * Changing the policy of the stop threads its a very bad idea
3990 if (p
== rq
->stop
) {
3991 task_rq_unlock(rq
, p
, &flags
);
3996 * If not changing anything there's no need to proceed further:
3998 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3999 param
->sched_priority
== p
->rt_priority
))) {
4000 task_rq_unlock(rq
, p
, &flags
);
4004 #ifdef CONFIG_RT_GROUP_SCHED
4007 * Do not allow realtime tasks into groups that have no runtime
4010 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4011 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4012 !task_group_is_autogroup(task_group(p
))) {
4013 task_rq_unlock(rq
, p
, &flags
);
4019 /* recheck policy now with rq lock held */
4020 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4021 policy
= oldpolicy
= -1;
4022 task_rq_unlock(rq
, p
, &flags
);
4026 running
= task_current(rq
, p
);
4028 dequeue_task(rq
, p
, 0);
4030 p
->sched_class
->put_prev_task(rq
, p
);
4032 p
->sched_reset_on_fork
= reset_on_fork
;
4035 prev_class
= p
->sched_class
;
4036 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4039 p
->sched_class
->set_curr_task(rq
);
4041 enqueue_task(rq
, p
, 0);
4043 check_class_changed(rq
, p
, prev_class
, oldprio
);
4044 task_rq_unlock(rq
, p
, &flags
);
4046 rt_mutex_adjust_pi(p
);
4052 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4053 * @p: the task in question.
4054 * @policy: new policy.
4055 * @param: structure containing the new RT priority.
4057 * NOTE that the task may be already dead.
4059 int sched_setscheduler(struct task_struct
*p
, int policy
,
4060 const struct sched_param
*param
)
4062 return __sched_setscheduler(p
, policy
, param
, true);
4064 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4067 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4068 * @p: the task in question.
4069 * @policy: new policy.
4070 * @param: structure containing the new RT priority.
4072 * Just like sched_setscheduler, only don't bother checking if the
4073 * current context has permission. For example, this is needed in
4074 * stop_machine(): we create temporary high priority worker threads,
4075 * but our caller might not have that capability.
4077 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4078 const struct sched_param
*param
)
4080 return __sched_setscheduler(p
, policy
, param
, false);
4084 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4086 struct sched_param lparam
;
4087 struct task_struct
*p
;
4090 if (!param
|| pid
< 0)
4092 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4097 p
= find_process_by_pid(pid
);
4099 retval
= sched_setscheduler(p
, policy
, &lparam
);
4106 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4107 * @pid: the pid in question.
4108 * @policy: new policy.
4109 * @param: structure containing the new RT priority.
4111 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4112 struct sched_param __user
*, param
)
4114 /* negative values for policy are not valid */
4118 return do_sched_setscheduler(pid
, policy
, param
);
4122 * sys_sched_setparam - set/change the RT priority of a thread
4123 * @pid: the pid in question.
4124 * @param: structure containing the new RT priority.
4126 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4128 return do_sched_setscheduler(pid
, -1, param
);
4132 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4133 * @pid: the pid in question.
4135 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4137 struct task_struct
*p
;
4145 p
= find_process_by_pid(pid
);
4147 retval
= security_task_getscheduler(p
);
4150 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4157 * sys_sched_getparam - get the RT priority of a thread
4158 * @pid: the pid in question.
4159 * @param: structure containing the RT priority.
4161 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4163 struct sched_param lp
;
4164 struct task_struct
*p
;
4167 if (!param
|| pid
< 0)
4171 p
= find_process_by_pid(pid
);
4176 retval
= security_task_getscheduler(p
);
4180 lp
.sched_priority
= p
->rt_priority
;
4184 * This one might sleep, we cannot do it with a spinlock held ...
4186 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4195 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4197 cpumask_var_t cpus_allowed
, new_mask
;
4198 struct task_struct
*p
;
4204 p
= find_process_by_pid(pid
);
4211 /* Prevent p going away */
4215 if (p
->flags
& PF_NO_SETAFFINITY
) {
4219 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4223 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4225 goto out_free_cpus_allowed
;
4228 if (!check_same_owner(p
)) {
4230 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4237 retval
= security_task_setscheduler(p
);
4241 cpuset_cpus_allowed(p
, cpus_allowed
);
4242 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4244 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4247 cpuset_cpus_allowed(p
, cpus_allowed
);
4248 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4250 * We must have raced with a concurrent cpuset
4251 * update. Just reset the cpus_allowed to the
4252 * cpuset's cpus_allowed
4254 cpumask_copy(new_mask
, cpus_allowed
);
4259 free_cpumask_var(new_mask
);
4260 out_free_cpus_allowed
:
4261 free_cpumask_var(cpus_allowed
);
4268 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4269 struct cpumask
*new_mask
)
4271 if (len
< cpumask_size())
4272 cpumask_clear(new_mask
);
4273 else if (len
> cpumask_size())
4274 len
= cpumask_size();
4276 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4280 * sys_sched_setaffinity - set the cpu affinity of a process
4281 * @pid: pid of the process
4282 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4283 * @user_mask_ptr: user-space pointer to the new cpu mask
4285 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4286 unsigned long __user
*, user_mask_ptr
)
4288 cpumask_var_t new_mask
;
4291 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4294 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4296 retval
= sched_setaffinity(pid
, new_mask
);
4297 free_cpumask_var(new_mask
);
4301 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4303 struct task_struct
*p
;
4304 unsigned long flags
;
4311 p
= find_process_by_pid(pid
);
4315 retval
= security_task_getscheduler(p
);
4319 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4320 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4321 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4331 * sys_sched_getaffinity - get the cpu affinity of a process
4332 * @pid: pid of the process
4333 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4334 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4336 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4337 unsigned long __user
*, user_mask_ptr
)
4342 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4344 if (len
& (sizeof(unsigned long)-1))
4347 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4350 ret
= sched_getaffinity(pid
, mask
);
4352 size_t retlen
= min_t(size_t, len
, cpumask_size());
4354 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4359 free_cpumask_var(mask
);
4365 * sys_sched_yield - yield the current processor to other threads.
4367 * This function yields the current CPU to other tasks. If there are no
4368 * other threads running on this CPU then this function will return.
4370 SYSCALL_DEFINE0(sched_yield
)
4372 struct rq
*rq
= this_rq_lock();
4374 schedstat_inc(rq
, yld_count
);
4375 current
->sched_class
->yield_task(rq
);
4378 * Since we are going to call schedule() anyway, there's
4379 * no need to preempt or enable interrupts:
4381 __release(rq
->lock
);
4382 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4383 do_raw_spin_unlock(&rq
->lock
);
4384 sched_preempt_enable_no_resched();
4391 static inline int should_resched(void)
4393 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4396 static void __cond_resched(void)
4398 add_preempt_count(PREEMPT_ACTIVE
);
4400 sub_preempt_count(PREEMPT_ACTIVE
);
4403 int __sched
_cond_resched(void)
4405 if (should_resched()) {
4411 EXPORT_SYMBOL(_cond_resched
);
4414 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4415 * call schedule, and on return reacquire the lock.
4417 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4418 * operations here to prevent schedule() from being called twice (once via
4419 * spin_unlock(), once by hand).
4421 int __cond_resched_lock(spinlock_t
*lock
)
4423 int resched
= should_resched();
4426 lockdep_assert_held(lock
);
4428 if (spin_needbreak(lock
) || resched
) {
4439 EXPORT_SYMBOL(__cond_resched_lock
);
4441 int __sched
__cond_resched_softirq(void)
4443 BUG_ON(!in_softirq());
4445 if (should_resched()) {
4453 EXPORT_SYMBOL(__cond_resched_softirq
);
4456 * yield - yield the current processor to other threads.
4458 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4460 * The scheduler is at all times free to pick the calling task as the most
4461 * eligible task to run, if removing the yield() call from your code breaks
4462 * it, its already broken.
4464 * Typical broken usage is:
4469 * where one assumes that yield() will let 'the other' process run that will
4470 * make event true. If the current task is a SCHED_FIFO task that will never
4471 * happen. Never use yield() as a progress guarantee!!
4473 * If you want to use yield() to wait for something, use wait_event().
4474 * If you want to use yield() to be 'nice' for others, use cond_resched().
4475 * If you still want to use yield(), do not!
4477 void __sched
yield(void)
4479 set_current_state(TASK_RUNNING
);
4482 EXPORT_SYMBOL(yield
);
4485 * yield_to - yield the current processor to another thread in
4486 * your thread group, or accelerate that thread toward the
4487 * processor it's on.
4489 * @preempt: whether task preemption is allowed or not
4491 * It's the caller's job to ensure that the target task struct
4492 * can't go away on us before we can do any checks.
4495 * true (>0) if we indeed boosted the target task.
4496 * false (0) if we failed to boost the target.
4497 * -ESRCH if there's no task to yield to.
4499 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4501 struct task_struct
*curr
= current
;
4502 struct rq
*rq
, *p_rq
;
4503 unsigned long flags
;
4506 local_irq_save(flags
);
4512 * If we're the only runnable task on the rq and target rq also
4513 * has only one task, there's absolutely no point in yielding.
4515 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4520 double_rq_lock(rq
, p_rq
);
4521 while (task_rq(p
) != p_rq
) {
4522 double_rq_unlock(rq
, p_rq
);
4526 if (!curr
->sched_class
->yield_to_task
)
4529 if (curr
->sched_class
!= p
->sched_class
)
4532 if (task_running(p_rq
, p
) || p
->state
)
4535 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4537 schedstat_inc(rq
, yld_count
);
4539 * Make p's CPU reschedule; pick_next_entity takes care of
4542 if (preempt
&& rq
!= p_rq
)
4543 resched_task(p_rq
->curr
);
4547 double_rq_unlock(rq
, p_rq
);
4549 local_irq_restore(flags
);
4556 EXPORT_SYMBOL_GPL(yield_to
);
4559 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4560 * that process accounting knows that this is a task in IO wait state.
4562 void __sched
io_schedule(void)
4564 struct rq
*rq
= raw_rq();
4566 delayacct_blkio_start();
4567 atomic_inc(&rq
->nr_iowait
);
4568 blk_flush_plug(current
);
4569 current
->in_iowait
= 1;
4571 current
->in_iowait
= 0;
4572 atomic_dec(&rq
->nr_iowait
);
4573 delayacct_blkio_end();
4575 EXPORT_SYMBOL(io_schedule
);
4577 long __sched
io_schedule_timeout(long timeout
)
4579 struct rq
*rq
= raw_rq();
4582 delayacct_blkio_start();
4583 atomic_inc(&rq
->nr_iowait
);
4584 blk_flush_plug(current
);
4585 current
->in_iowait
= 1;
4586 ret
= schedule_timeout(timeout
);
4587 current
->in_iowait
= 0;
4588 atomic_dec(&rq
->nr_iowait
);
4589 delayacct_blkio_end();
4594 * sys_sched_get_priority_max - return maximum RT priority.
4595 * @policy: scheduling class.
4597 * this syscall returns the maximum rt_priority that can be used
4598 * by a given scheduling class.
4600 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4607 ret
= MAX_USER_RT_PRIO
-1;
4619 * sys_sched_get_priority_min - return minimum RT priority.
4620 * @policy: scheduling class.
4622 * this syscall returns the minimum rt_priority that can be used
4623 * by a given scheduling class.
4625 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4643 * sys_sched_rr_get_interval - return the default timeslice of a process.
4644 * @pid: pid of the process.
4645 * @interval: userspace pointer to the timeslice value.
4647 * this syscall writes the default timeslice value of a given process
4648 * into the user-space timespec buffer. A value of '0' means infinity.
4650 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4651 struct timespec __user
*, interval
)
4653 struct task_struct
*p
;
4654 unsigned int time_slice
;
4655 unsigned long flags
;
4665 p
= find_process_by_pid(pid
);
4669 retval
= security_task_getscheduler(p
);
4673 rq
= task_rq_lock(p
, &flags
);
4674 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4675 task_rq_unlock(rq
, p
, &flags
);
4678 jiffies_to_timespec(time_slice
, &t
);
4679 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4687 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4689 void sched_show_task(struct task_struct
*p
)
4691 unsigned long free
= 0;
4695 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4696 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4697 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4698 #if BITS_PER_LONG == 32
4699 if (state
== TASK_RUNNING
)
4700 printk(KERN_CONT
" running ");
4702 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4704 if (state
== TASK_RUNNING
)
4705 printk(KERN_CONT
" running task ");
4707 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4709 #ifdef CONFIG_DEBUG_STACK_USAGE
4710 free
= stack_not_used(p
);
4713 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4715 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4716 task_pid_nr(p
), ppid
,
4717 (unsigned long)task_thread_info(p
)->flags
);
4719 print_worker_info(KERN_INFO
, p
);
4720 show_stack(p
, NULL
);
4723 void show_state_filter(unsigned long state_filter
)
4725 struct task_struct
*g
, *p
;
4727 #if BITS_PER_LONG == 32
4729 " task PC stack pid father\n");
4732 " task PC stack pid father\n");
4735 do_each_thread(g
, p
) {
4737 * reset the NMI-timeout, listing all files on a slow
4738 * console might take a lot of time:
4740 touch_nmi_watchdog();
4741 if (!state_filter
|| (p
->state
& state_filter
))
4743 } while_each_thread(g
, p
);
4745 touch_all_softlockup_watchdogs();
4747 #ifdef CONFIG_SCHED_DEBUG
4749 sysrq_sched_debug_show();
4753 * Only show locks if all tasks are dumped:
4756 debug_show_all_locks();
4759 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4761 idle
->sched_class
= &idle_sched_class
;
4765 * init_idle - set up an idle thread for a given CPU
4766 * @idle: task in question
4767 * @cpu: cpu the idle task belongs to
4769 * NOTE: this function does not set the idle thread's NEED_RESCHED
4770 * flag, to make booting more robust.
4772 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4774 struct rq
*rq
= cpu_rq(cpu
);
4775 unsigned long flags
;
4777 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4780 idle
->state
= TASK_RUNNING
;
4781 idle
->se
.exec_start
= sched_clock();
4783 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4785 * We're having a chicken and egg problem, even though we are
4786 * holding rq->lock, the cpu isn't yet set to this cpu so the
4787 * lockdep check in task_group() will fail.
4789 * Similar case to sched_fork(). / Alternatively we could
4790 * use task_rq_lock() here and obtain the other rq->lock.
4795 __set_task_cpu(idle
, cpu
);
4798 rq
->curr
= rq
->idle
= idle
;
4799 #if defined(CONFIG_SMP)
4802 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4804 /* Set the preempt count _outside_ the spinlocks! */
4805 task_thread_info(idle
)->preempt_count
= 0;
4808 * The idle tasks have their own, simple scheduling class:
4810 idle
->sched_class
= &idle_sched_class
;
4811 ftrace_graph_init_idle_task(idle
, cpu
);
4812 vtime_init_idle(idle
, cpu
);
4813 #if defined(CONFIG_SMP)
4814 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4819 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4821 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4822 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4824 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4825 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4829 * This is how migration works:
4831 * 1) we invoke migration_cpu_stop() on the target CPU using
4833 * 2) stopper starts to run (implicitly forcing the migrated thread
4835 * 3) it checks whether the migrated task is still in the wrong runqueue.
4836 * 4) if it's in the wrong runqueue then the migration thread removes
4837 * it and puts it into the right queue.
4838 * 5) stopper completes and stop_one_cpu() returns and the migration
4843 * Change a given task's CPU affinity. Migrate the thread to a
4844 * proper CPU and schedule it away if the CPU it's executing on
4845 * is removed from the allowed bitmask.
4847 * NOTE: the caller must have a valid reference to the task, the
4848 * task must not exit() & deallocate itself prematurely. The
4849 * call is not atomic; no spinlocks may be held.
4851 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4853 unsigned long flags
;
4855 unsigned int dest_cpu
;
4858 rq
= task_rq_lock(p
, &flags
);
4860 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4863 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4868 do_set_cpus_allowed(p
, new_mask
);
4870 /* Can the task run on the task's current CPU? If so, we're done */
4871 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4874 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4876 struct migration_arg arg
= { p
, dest_cpu
};
4877 /* Need help from migration thread: drop lock and wait. */
4878 task_rq_unlock(rq
, p
, &flags
);
4879 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4880 tlb_migrate_finish(p
->mm
);
4884 task_rq_unlock(rq
, p
, &flags
);
4888 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4891 * Move (not current) task off this cpu, onto dest cpu. We're doing
4892 * this because either it can't run here any more (set_cpus_allowed()
4893 * away from this CPU, or CPU going down), or because we're
4894 * attempting to rebalance this task on exec (sched_exec).
4896 * So we race with normal scheduler movements, but that's OK, as long
4897 * as the task is no longer on this CPU.
4899 * Returns non-zero if task was successfully migrated.
4901 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4903 struct rq
*rq_dest
, *rq_src
;
4906 if (unlikely(!cpu_active(dest_cpu
)))
4909 rq_src
= cpu_rq(src_cpu
);
4910 rq_dest
= cpu_rq(dest_cpu
);
4912 raw_spin_lock(&p
->pi_lock
);
4913 double_rq_lock(rq_src
, rq_dest
);
4914 /* Already moved. */
4915 if (task_cpu(p
) != src_cpu
)
4917 /* Affinity changed (again). */
4918 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4922 * If we're not on a rq, the next wake-up will ensure we're
4926 dequeue_task(rq_src
, p
, 0);
4927 set_task_cpu(p
, dest_cpu
);
4928 enqueue_task(rq_dest
, p
, 0);
4929 check_preempt_curr(rq_dest
, p
, 0);
4934 double_rq_unlock(rq_src
, rq_dest
);
4935 raw_spin_unlock(&p
->pi_lock
);
4940 * migration_cpu_stop - this will be executed by a highprio stopper thread
4941 * and performs thread migration by bumping thread off CPU then
4942 * 'pushing' onto another runqueue.
4944 static int migration_cpu_stop(void *data
)
4946 struct migration_arg
*arg
= data
;
4949 * The original target cpu might have gone down and we might
4950 * be on another cpu but it doesn't matter.
4952 local_irq_disable();
4953 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4958 #ifdef CONFIG_HOTPLUG_CPU
4961 * Ensures that the idle task is using init_mm right before its cpu goes
4964 void idle_task_exit(void)
4966 struct mm_struct
*mm
= current
->active_mm
;
4968 BUG_ON(cpu_online(smp_processor_id()));
4971 switch_mm(mm
, &init_mm
, current
);
4976 * Since this CPU is going 'away' for a while, fold any nr_active delta
4977 * we might have. Assumes we're called after migrate_tasks() so that the
4978 * nr_active count is stable.
4980 * Also see the comment "Global load-average calculations".
4982 static void calc_load_migrate(struct rq
*rq
)
4984 long delta
= calc_load_fold_active(rq
);
4986 atomic_long_add(delta
, &calc_load_tasks
);
4990 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4991 * try_to_wake_up()->select_task_rq().
4993 * Called with rq->lock held even though we'er in stop_machine() and
4994 * there's no concurrency possible, we hold the required locks anyway
4995 * because of lock validation efforts.
4997 static void migrate_tasks(unsigned int dead_cpu
)
4999 struct rq
*rq
= cpu_rq(dead_cpu
);
5000 struct task_struct
*next
, *stop
= rq
->stop
;
5004 * Fudge the rq selection such that the below task selection loop
5005 * doesn't get stuck on the currently eligible stop task.
5007 * We're currently inside stop_machine() and the rq is either stuck
5008 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5009 * either way we should never end up calling schedule() until we're
5016 * There's this thread running, bail when that's the only
5019 if (rq
->nr_running
== 1)
5022 next
= pick_next_task(rq
);
5024 next
->sched_class
->put_prev_task(rq
, next
);
5026 /* Find suitable destination for @next, with force if needed. */
5027 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5028 raw_spin_unlock(&rq
->lock
);
5030 __migrate_task(next
, dead_cpu
, dest_cpu
);
5032 raw_spin_lock(&rq
->lock
);
5038 #endif /* CONFIG_HOTPLUG_CPU */
5040 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5042 static struct ctl_table sd_ctl_dir
[] = {
5044 .procname
= "sched_domain",
5050 static struct ctl_table sd_ctl_root
[] = {
5052 .procname
= "kernel",
5054 .child
= sd_ctl_dir
,
5059 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5061 struct ctl_table
*entry
=
5062 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5067 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5069 struct ctl_table
*entry
;
5072 * In the intermediate directories, both the child directory and
5073 * procname are dynamically allocated and could fail but the mode
5074 * will always be set. In the lowest directory the names are
5075 * static strings and all have proc handlers.
5077 for (entry
= *tablep
; entry
->mode
; entry
++) {
5079 sd_free_ctl_entry(&entry
->child
);
5080 if (entry
->proc_handler
== NULL
)
5081 kfree(entry
->procname
);
5088 static int min_load_idx
= 0;
5089 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5092 set_table_entry(struct ctl_table
*entry
,
5093 const char *procname
, void *data
, int maxlen
,
5094 umode_t mode
, proc_handler
*proc_handler
,
5097 entry
->procname
= procname
;
5099 entry
->maxlen
= maxlen
;
5101 entry
->proc_handler
= proc_handler
;
5104 entry
->extra1
= &min_load_idx
;
5105 entry
->extra2
= &max_load_idx
;
5109 static struct ctl_table
*
5110 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5112 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5117 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5118 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5119 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5120 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5121 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5122 sizeof(int), 0644, proc_dointvec_minmax
, true);
5123 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5124 sizeof(int), 0644, proc_dointvec_minmax
, true);
5125 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5126 sizeof(int), 0644, proc_dointvec_minmax
, true);
5127 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5128 sizeof(int), 0644, proc_dointvec_minmax
, true);
5129 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5130 sizeof(int), 0644, proc_dointvec_minmax
, true);
5131 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5132 sizeof(int), 0644, proc_dointvec_minmax
, false);
5133 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5134 sizeof(int), 0644, proc_dointvec_minmax
, false);
5135 set_table_entry(&table
[9], "cache_nice_tries",
5136 &sd
->cache_nice_tries
,
5137 sizeof(int), 0644, proc_dointvec_minmax
, false);
5138 set_table_entry(&table
[10], "flags", &sd
->flags
,
5139 sizeof(int), 0644, proc_dointvec_minmax
, false);
5140 set_table_entry(&table
[11], "name", sd
->name
,
5141 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5142 /* &table[12] is terminator */
5147 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5149 struct ctl_table
*entry
, *table
;
5150 struct sched_domain
*sd
;
5151 int domain_num
= 0, i
;
5154 for_each_domain(cpu
, sd
)
5156 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5161 for_each_domain(cpu
, sd
) {
5162 snprintf(buf
, 32, "domain%d", i
);
5163 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5165 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5172 static struct ctl_table_header
*sd_sysctl_header
;
5173 static void register_sched_domain_sysctl(void)
5175 int i
, cpu_num
= num_possible_cpus();
5176 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5179 WARN_ON(sd_ctl_dir
[0].child
);
5180 sd_ctl_dir
[0].child
= entry
;
5185 for_each_possible_cpu(i
) {
5186 snprintf(buf
, 32, "cpu%d", i
);
5187 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5189 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5193 WARN_ON(sd_sysctl_header
);
5194 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5197 /* may be called multiple times per register */
5198 static void unregister_sched_domain_sysctl(void)
5200 if (sd_sysctl_header
)
5201 unregister_sysctl_table(sd_sysctl_header
);
5202 sd_sysctl_header
= NULL
;
5203 if (sd_ctl_dir
[0].child
)
5204 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5207 static void register_sched_domain_sysctl(void)
5210 static void unregister_sched_domain_sysctl(void)
5215 static void set_rq_online(struct rq
*rq
)
5218 const struct sched_class
*class;
5220 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5223 for_each_class(class) {
5224 if (class->rq_online
)
5225 class->rq_online(rq
);
5230 static void set_rq_offline(struct rq
*rq
)
5233 const struct sched_class
*class;
5235 for_each_class(class) {
5236 if (class->rq_offline
)
5237 class->rq_offline(rq
);
5240 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5246 * migration_call - callback that gets triggered when a CPU is added.
5247 * Here we can start up the necessary migration thread for the new CPU.
5249 static int __cpuinit
5250 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5252 int cpu
= (long)hcpu
;
5253 unsigned long flags
;
5254 struct rq
*rq
= cpu_rq(cpu
);
5256 switch (action
& ~CPU_TASKS_FROZEN
) {
5258 case CPU_UP_PREPARE
:
5259 rq
->calc_load_update
= calc_load_update
;
5260 account_reset_rq(rq
);
5264 /* Update our root-domain */
5265 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5267 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5271 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5274 #ifdef CONFIG_HOTPLUG_CPU
5276 sched_ttwu_pending();
5277 /* Update our root-domain */
5278 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5280 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5284 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5285 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5289 calc_load_migrate(rq
);
5294 update_max_interval();
5300 * Register at high priority so that task migration (migrate_all_tasks)
5301 * happens before everything else. This has to be lower priority than
5302 * the notifier in the perf_event subsystem, though.
5304 static struct notifier_block __cpuinitdata migration_notifier
= {
5305 .notifier_call
= migration_call
,
5306 .priority
= CPU_PRI_MIGRATION
,
5309 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5310 unsigned long action
, void *hcpu
)
5312 switch (action
& ~CPU_TASKS_FROZEN
) {
5313 case CPU_DOWN_FAILED
:
5314 set_cpu_active((long)hcpu
, true);
5321 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5322 unsigned long action
, void *hcpu
)
5324 switch (action
& ~CPU_TASKS_FROZEN
) {
5325 case CPU_DOWN_PREPARE
:
5326 set_cpu_active((long)hcpu
, false);
5333 static int __init
migration_init(void)
5335 void *cpu
= (void *)(long)smp_processor_id();
5338 /* Initialize migration for the boot CPU */
5339 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5340 BUG_ON(err
== NOTIFY_BAD
);
5341 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5342 register_cpu_notifier(&migration_notifier
);
5344 /* Register cpu active notifiers */
5345 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5346 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5350 early_initcall(migration_init
);
5355 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5357 #ifdef CONFIG_SCHED_DEBUG
5359 static __read_mostly
int sched_debug_enabled
;
5361 static int __init
sched_debug_setup(char *str
)
5363 sched_debug_enabled
= 1;
5367 early_param("sched_debug", sched_debug_setup
);
5369 static inline bool sched_debug(void)
5371 return sched_debug_enabled
;
5374 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5375 struct cpumask
*groupmask
)
5377 struct sched_group
*group
= sd
->groups
;
5380 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5381 cpumask_clear(groupmask
);
5383 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5385 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5386 printk("does not load-balance\n");
5388 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5393 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5395 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5396 printk(KERN_ERR
"ERROR: domain->span does not contain "
5399 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5400 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5404 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5408 printk(KERN_ERR
"ERROR: group is NULL\n");
5413 * Even though we initialize ->power to something semi-sane,
5414 * we leave power_orig unset. This allows us to detect if
5415 * domain iteration is still funny without causing /0 traps.
5417 if (!group
->sgp
->power_orig
) {
5418 printk(KERN_CONT
"\n");
5419 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5424 if (!cpumask_weight(sched_group_cpus(group
))) {
5425 printk(KERN_CONT
"\n");
5426 printk(KERN_ERR
"ERROR: empty group\n");
5430 if (!(sd
->flags
& SD_OVERLAP
) &&
5431 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5432 printk(KERN_CONT
"\n");
5433 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5437 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5439 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5441 printk(KERN_CONT
" %s", str
);
5442 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5443 printk(KERN_CONT
" (cpu_power = %d)",
5447 group
= group
->next
;
5448 } while (group
!= sd
->groups
);
5449 printk(KERN_CONT
"\n");
5451 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5452 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5455 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5456 printk(KERN_ERR
"ERROR: parent span is not a superset "
5457 "of domain->span\n");
5461 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5465 if (!sched_debug_enabled
)
5469 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5473 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5476 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5484 #else /* !CONFIG_SCHED_DEBUG */
5485 # define sched_domain_debug(sd, cpu) do { } while (0)
5486 static inline bool sched_debug(void)
5490 #endif /* CONFIG_SCHED_DEBUG */
5492 static int sd_degenerate(struct sched_domain
*sd
)
5494 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5497 /* Following flags need at least 2 groups */
5498 if (sd
->flags
& (SD_LOAD_BALANCE
|
5499 SD_BALANCE_NEWIDLE
|
5503 SD_SHARE_PKG_RESOURCES
)) {
5504 if (sd
->groups
!= sd
->groups
->next
)
5508 /* Following flags don't use groups */
5509 if (sd
->flags
& (SD_WAKE_AFFINE
))
5516 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5518 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5520 if (sd_degenerate(parent
))
5523 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5526 /* Flags needing groups don't count if only 1 group in parent */
5527 if (parent
->groups
== parent
->groups
->next
) {
5528 pflags
&= ~(SD_LOAD_BALANCE
|
5529 SD_BALANCE_NEWIDLE
|
5533 SD_SHARE_PKG_RESOURCES
);
5534 if (nr_node_ids
== 1)
5535 pflags
&= ~SD_SERIALIZE
;
5537 if (~cflags
& pflags
)
5543 static void free_rootdomain(struct rcu_head
*rcu
)
5545 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5547 cpupri_cleanup(&rd
->cpupri
);
5548 free_cpumask_var(rd
->rto_mask
);
5549 free_cpumask_var(rd
->online
);
5550 free_cpumask_var(rd
->span
);
5554 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5556 struct root_domain
*old_rd
= NULL
;
5557 unsigned long flags
;
5559 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5564 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5567 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5570 * If we dont want to free the old_rt yet then
5571 * set old_rd to NULL to skip the freeing later
5574 if (!atomic_dec_and_test(&old_rd
->refcount
))
5578 atomic_inc(&rd
->refcount
);
5581 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5582 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5585 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5588 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5591 static int init_rootdomain(struct root_domain
*rd
)
5593 memset(rd
, 0, sizeof(*rd
));
5595 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5597 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5599 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5602 if (cpupri_init(&rd
->cpupri
) != 0)
5607 free_cpumask_var(rd
->rto_mask
);
5609 free_cpumask_var(rd
->online
);
5611 free_cpumask_var(rd
->span
);
5617 * By default the system creates a single root-domain with all cpus as
5618 * members (mimicking the global state we have today).
5620 struct root_domain def_root_domain
;
5622 static void init_defrootdomain(void)
5624 init_rootdomain(&def_root_domain
);
5626 atomic_set(&def_root_domain
.refcount
, 1);
5629 static struct root_domain
*alloc_rootdomain(void)
5631 struct root_domain
*rd
;
5633 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5637 if (init_rootdomain(rd
) != 0) {
5645 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5647 struct sched_group
*tmp
, *first
;
5656 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5661 } while (sg
!= first
);
5664 static void free_sched_domain(struct rcu_head
*rcu
)
5666 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5669 * If its an overlapping domain it has private groups, iterate and
5672 if (sd
->flags
& SD_OVERLAP
) {
5673 free_sched_groups(sd
->groups
, 1);
5674 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5675 kfree(sd
->groups
->sgp
);
5681 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5683 call_rcu(&sd
->rcu
, free_sched_domain
);
5686 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5688 for (; sd
; sd
= sd
->parent
)
5689 destroy_sched_domain(sd
, cpu
);
5693 * Keep a special pointer to the highest sched_domain that has
5694 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5695 * allows us to avoid some pointer chasing select_idle_sibling().
5697 * Also keep a unique ID per domain (we use the first cpu number in
5698 * the cpumask of the domain), this allows us to quickly tell if
5699 * two cpus are in the same cache domain, see cpus_share_cache().
5701 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5702 DEFINE_PER_CPU(int, sd_llc_id
);
5704 static void update_top_cache_domain(int cpu
)
5706 struct sched_domain
*sd
;
5709 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5711 id
= cpumask_first(sched_domain_span(sd
));
5713 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5714 per_cpu(sd_llc_id
, cpu
) = id
;
5718 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5719 * hold the hotplug lock.
5722 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5724 struct rq
*rq
= cpu_rq(cpu
);
5725 struct sched_domain
*tmp
;
5727 /* Remove the sched domains which do not contribute to scheduling. */
5728 for (tmp
= sd
; tmp
; ) {
5729 struct sched_domain
*parent
= tmp
->parent
;
5733 if (sd_parent_degenerate(tmp
, parent
)) {
5734 tmp
->parent
= parent
->parent
;
5736 parent
->parent
->child
= tmp
;
5737 destroy_sched_domain(parent
, cpu
);
5742 if (sd
&& sd_degenerate(sd
)) {
5745 destroy_sched_domain(tmp
, cpu
);
5750 sched_domain_debug(sd
, cpu
);
5752 rq_attach_root(rq
, rd
);
5754 rcu_assign_pointer(rq
->sd
, sd
);
5755 destroy_sched_domains(tmp
, cpu
);
5757 update_top_cache_domain(cpu
);
5760 /* cpus with isolated domains */
5761 static cpumask_var_t cpu_isolated_map
;
5763 /* Setup the mask of cpus configured for isolated domains */
5764 static int __init
isolated_cpu_setup(char *str
)
5766 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5767 cpulist_parse(str
, cpu_isolated_map
);
5771 __setup("isolcpus=", isolated_cpu_setup
);
5773 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5775 return cpumask_of_node(cpu_to_node(cpu
));
5779 struct sched_domain
**__percpu sd
;
5780 struct sched_group
**__percpu sg
;
5781 struct sched_group_power
**__percpu sgp
;
5785 struct sched_domain
** __percpu sd
;
5786 struct root_domain
*rd
;
5796 struct sched_domain_topology_level
;
5798 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5799 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5801 #define SDTL_OVERLAP 0x01
5803 struct sched_domain_topology_level
{
5804 sched_domain_init_f init
;
5805 sched_domain_mask_f mask
;
5808 struct sd_data data
;
5812 * Build an iteration mask that can exclude certain CPUs from the upwards
5815 * Asymmetric node setups can result in situations where the domain tree is of
5816 * unequal depth, make sure to skip domains that already cover the entire
5819 * In that case build_sched_domains() will have terminated the iteration early
5820 * and our sibling sd spans will be empty. Domains should always include the
5821 * cpu they're built on, so check that.
5824 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5826 const struct cpumask
*span
= sched_domain_span(sd
);
5827 struct sd_data
*sdd
= sd
->private;
5828 struct sched_domain
*sibling
;
5831 for_each_cpu(i
, span
) {
5832 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5833 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5836 cpumask_set_cpu(i
, sched_group_mask(sg
));
5841 * Return the canonical balance cpu for this group, this is the first cpu
5842 * of this group that's also in the iteration mask.
5844 int group_balance_cpu(struct sched_group
*sg
)
5846 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5850 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5852 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5853 const struct cpumask
*span
= sched_domain_span(sd
);
5854 struct cpumask
*covered
= sched_domains_tmpmask
;
5855 struct sd_data
*sdd
= sd
->private;
5856 struct sched_domain
*child
;
5859 cpumask_clear(covered
);
5861 for_each_cpu(i
, span
) {
5862 struct cpumask
*sg_span
;
5864 if (cpumask_test_cpu(i
, covered
))
5867 child
= *per_cpu_ptr(sdd
->sd
, i
);
5869 /* See the comment near build_group_mask(). */
5870 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5873 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5874 GFP_KERNEL
, cpu_to_node(cpu
));
5879 sg_span
= sched_group_cpus(sg
);
5881 child
= child
->child
;
5882 cpumask_copy(sg_span
, sched_domain_span(child
));
5884 cpumask_set_cpu(i
, sg_span
);
5886 cpumask_or(covered
, covered
, sg_span
);
5888 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5889 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5890 build_group_mask(sd
, sg
);
5893 * Initialize sgp->power such that even if we mess up the
5894 * domains and no possible iteration will get us here, we won't
5897 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5900 * Make sure the first group of this domain contains the
5901 * canonical balance cpu. Otherwise the sched_domain iteration
5902 * breaks. See update_sg_lb_stats().
5904 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5905 group_balance_cpu(sg
) == cpu
)
5915 sd
->groups
= groups
;
5920 free_sched_groups(first
, 0);
5925 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5927 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5928 struct sched_domain
*child
= sd
->child
;
5931 cpu
= cpumask_first(sched_domain_span(child
));
5934 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5935 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5936 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5943 * build_sched_groups will build a circular linked list of the groups
5944 * covered by the given span, and will set each group's ->cpumask correctly,
5945 * and ->cpu_power to 0.
5947 * Assumes the sched_domain tree is fully constructed
5950 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5952 struct sched_group
*first
= NULL
, *last
= NULL
;
5953 struct sd_data
*sdd
= sd
->private;
5954 const struct cpumask
*span
= sched_domain_span(sd
);
5955 struct cpumask
*covered
;
5958 get_group(cpu
, sdd
, &sd
->groups
);
5959 atomic_inc(&sd
->groups
->ref
);
5961 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5964 lockdep_assert_held(&sched_domains_mutex
);
5965 covered
= sched_domains_tmpmask
;
5967 cpumask_clear(covered
);
5969 for_each_cpu(i
, span
) {
5970 struct sched_group
*sg
;
5971 int group
= get_group(i
, sdd
, &sg
);
5974 if (cpumask_test_cpu(i
, covered
))
5977 cpumask_clear(sched_group_cpus(sg
));
5979 cpumask_setall(sched_group_mask(sg
));
5981 for_each_cpu(j
, span
) {
5982 if (get_group(j
, sdd
, NULL
) != group
)
5985 cpumask_set_cpu(j
, covered
);
5986 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6001 * Initialize sched groups cpu_power.
6003 * cpu_power indicates the capacity of sched group, which is used while
6004 * distributing the load between different sched groups in a sched domain.
6005 * Typically cpu_power for all the groups in a sched domain will be same unless
6006 * there are asymmetries in the topology. If there are asymmetries, group
6007 * having more cpu_power will pickup more load compared to the group having
6010 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6012 struct sched_group
*sg
= sd
->groups
;
6014 WARN_ON(!sd
|| !sg
);
6017 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6019 } while (sg
!= sd
->groups
);
6021 if (cpu
!= group_balance_cpu(sg
))
6024 update_group_power(sd
, cpu
);
6025 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6028 int __weak
arch_sd_sibling_asym_packing(void)
6030 return 0*SD_ASYM_PACKING
;
6034 * Initializers for schedule domains
6035 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6038 #ifdef CONFIG_SCHED_DEBUG
6039 # define SD_INIT_NAME(sd, type) sd->name = #type
6041 # define SD_INIT_NAME(sd, type) do { } while (0)
6044 #define SD_INIT_FUNC(type) \
6045 static noinline struct sched_domain * \
6046 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6048 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6049 *sd = SD_##type##_INIT; \
6050 SD_INIT_NAME(sd, type); \
6051 sd->private = &tl->data; \
6056 #ifdef CONFIG_SCHED_SMT
6057 SD_INIT_FUNC(SIBLING
)
6059 #ifdef CONFIG_SCHED_MC
6062 #ifdef CONFIG_SCHED_BOOK
6066 static int default_relax_domain_level
= -1;
6067 int sched_domain_level_max
;
6069 static int __init
setup_relax_domain_level(char *str
)
6071 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6072 pr_warn("Unable to set relax_domain_level\n");
6076 __setup("relax_domain_level=", setup_relax_domain_level
);
6078 static void set_domain_attribute(struct sched_domain
*sd
,
6079 struct sched_domain_attr
*attr
)
6083 if (!attr
|| attr
->relax_domain_level
< 0) {
6084 if (default_relax_domain_level
< 0)
6087 request
= default_relax_domain_level
;
6089 request
= attr
->relax_domain_level
;
6090 if (request
< sd
->level
) {
6091 /* turn off idle balance on this domain */
6092 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6094 /* turn on idle balance on this domain */
6095 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6099 static void __sdt_free(const struct cpumask
*cpu_map
);
6100 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6102 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6103 const struct cpumask
*cpu_map
)
6107 if (!atomic_read(&d
->rd
->refcount
))
6108 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6110 free_percpu(d
->sd
); /* fall through */
6112 __sdt_free(cpu_map
); /* fall through */
6118 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6119 const struct cpumask
*cpu_map
)
6121 memset(d
, 0, sizeof(*d
));
6123 if (__sdt_alloc(cpu_map
))
6124 return sa_sd_storage
;
6125 d
->sd
= alloc_percpu(struct sched_domain
*);
6127 return sa_sd_storage
;
6128 d
->rd
= alloc_rootdomain();
6131 return sa_rootdomain
;
6135 * NULL the sd_data elements we've used to build the sched_domain and
6136 * sched_group structure so that the subsequent __free_domain_allocs()
6137 * will not free the data we're using.
6139 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6141 struct sd_data
*sdd
= sd
->private;
6143 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6144 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6146 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6147 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6149 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6150 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6153 #ifdef CONFIG_SCHED_SMT
6154 static const struct cpumask
*cpu_smt_mask(int cpu
)
6156 return topology_thread_cpumask(cpu
);
6161 * Topology list, bottom-up.
6163 static struct sched_domain_topology_level default_topology
[] = {
6164 #ifdef CONFIG_SCHED_SMT
6165 { sd_init_SIBLING
, cpu_smt_mask
, },
6167 #ifdef CONFIG_SCHED_MC
6168 { sd_init_MC
, cpu_coregroup_mask
, },
6170 #ifdef CONFIG_SCHED_BOOK
6171 { sd_init_BOOK
, cpu_book_mask
, },
6173 { sd_init_CPU
, cpu_cpu_mask
, },
6177 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6181 static int sched_domains_numa_levels
;
6182 static int *sched_domains_numa_distance
;
6183 static struct cpumask
***sched_domains_numa_masks
;
6184 static int sched_domains_curr_level
;
6186 static inline int sd_local_flags(int level
)
6188 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6191 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6194 static struct sched_domain
*
6195 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6197 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6198 int level
= tl
->numa_level
;
6199 int sd_weight
= cpumask_weight(
6200 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6202 *sd
= (struct sched_domain
){
6203 .min_interval
= sd_weight
,
6204 .max_interval
= 2*sd_weight
,
6206 .imbalance_pct
= 125,
6207 .cache_nice_tries
= 2,
6214 .flags
= 1*SD_LOAD_BALANCE
6215 | 1*SD_BALANCE_NEWIDLE
6220 | 0*SD_SHARE_CPUPOWER
6221 | 0*SD_SHARE_PKG_RESOURCES
6223 | 0*SD_PREFER_SIBLING
6224 | sd_local_flags(level
)
6226 .last_balance
= jiffies
,
6227 .balance_interval
= sd_weight
,
6229 SD_INIT_NAME(sd
, NUMA
);
6230 sd
->private = &tl
->data
;
6233 * Ugly hack to pass state to sd_numa_mask()...
6235 sched_domains_curr_level
= tl
->numa_level
;
6240 static const struct cpumask
*sd_numa_mask(int cpu
)
6242 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6245 static void sched_numa_warn(const char *str
)
6247 static int done
= false;
6255 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6257 for (i
= 0; i
< nr_node_ids
; i
++) {
6258 printk(KERN_WARNING
" ");
6259 for (j
= 0; j
< nr_node_ids
; j
++)
6260 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6261 printk(KERN_CONT
"\n");
6263 printk(KERN_WARNING
"\n");
6266 static bool find_numa_distance(int distance
)
6270 if (distance
== node_distance(0, 0))
6273 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6274 if (sched_domains_numa_distance
[i
] == distance
)
6281 static void sched_init_numa(void)
6283 int next_distance
, curr_distance
= node_distance(0, 0);
6284 struct sched_domain_topology_level
*tl
;
6288 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6289 if (!sched_domains_numa_distance
)
6293 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6294 * unique distances in the node_distance() table.
6296 * Assumes node_distance(0,j) includes all distances in
6297 * node_distance(i,j) in order to avoid cubic time.
6299 next_distance
= curr_distance
;
6300 for (i
= 0; i
< nr_node_ids
; i
++) {
6301 for (j
= 0; j
< nr_node_ids
; j
++) {
6302 for (k
= 0; k
< nr_node_ids
; k
++) {
6303 int distance
= node_distance(i
, k
);
6305 if (distance
> curr_distance
&&
6306 (distance
< next_distance
||
6307 next_distance
== curr_distance
))
6308 next_distance
= distance
;
6311 * While not a strong assumption it would be nice to know
6312 * about cases where if node A is connected to B, B is not
6313 * equally connected to A.
6315 if (sched_debug() && node_distance(k
, i
) != distance
)
6316 sched_numa_warn("Node-distance not symmetric");
6318 if (sched_debug() && i
&& !find_numa_distance(distance
))
6319 sched_numa_warn("Node-0 not representative");
6321 if (next_distance
!= curr_distance
) {
6322 sched_domains_numa_distance
[level
++] = next_distance
;
6323 sched_domains_numa_levels
= level
;
6324 curr_distance
= next_distance
;
6329 * In case of sched_debug() we verify the above assumption.
6335 * 'level' contains the number of unique distances, excluding the
6336 * identity distance node_distance(i,i).
6338 * The sched_domains_numa_distance[] array includes the actual distance
6343 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6344 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6345 * the array will contain less then 'level' members. This could be
6346 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6347 * in other functions.
6349 * We reset it to 'level' at the end of this function.
6351 sched_domains_numa_levels
= 0;
6353 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6354 if (!sched_domains_numa_masks
)
6358 * Now for each level, construct a mask per node which contains all
6359 * cpus of nodes that are that many hops away from us.
6361 for (i
= 0; i
< level
; i
++) {
6362 sched_domains_numa_masks
[i
] =
6363 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6364 if (!sched_domains_numa_masks
[i
])
6367 for (j
= 0; j
< nr_node_ids
; j
++) {
6368 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6372 sched_domains_numa_masks
[i
][j
] = mask
;
6374 for (k
= 0; k
< nr_node_ids
; k
++) {
6375 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6378 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6383 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6384 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6389 * Copy the default topology bits..
6391 for (i
= 0; default_topology
[i
].init
; i
++)
6392 tl
[i
] = default_topology
[i
];
6395 * .. and append 'j' levels of NUMA goodness.
6397 for (j
= 0; j
< level
; i
++, j
++) {
6398 tl
[i
] = (struct sched_domain_topology_level
){
6399 .init
= sd_numa_init
,
6400 .mask
= sd_numa_mask
,
6401 .flags
= SDTL_OVERLAP
,
6406 sched_domain_topology
= tl
;
6408 sched_domains_numa_levels
= level
;
6411 static void sched_domains_numa_masks_set(int cpu
)
6414 int node
= cpu_to_node(cpu
);
6416 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6417 for (j
= 0; j
< nr_node_ids
; j
++) {
6418 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6419 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6424 static void sched_domains_numa_masks_clear(int cpu
)
6427 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6428 for (j
= 0; j
< nr_node_ids
; j
++)
6429 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6434 * Update sched_domains_numa_masks[level][node] array when new cpus
6437 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6438 unsigned long action
,
6441 int cpu
= (long)hcpu
;
6443 switch (action
& ~CPU_TASKS_FROZEN
) {
6445 sched_domains_numa_masks_set(cpu
);
6449 sched_domains_numa_masks_clear(cpu
);
6459 static inline void sched_init_numa(void)
6463 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6464 unsigned long action
,
6469 #endif /* CONFIG_NUMA */
6471 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6473 struct sched_domain_topology_level
*tl
;
6476 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6477 struct sd_data
*sdd
= &tl
->data
;
6479 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6483 sdd
->sg
= alloc_percpu(struct sched_group
*);
6487 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6491 for_each_cpu(j
, cpu_map
) {
6492 struct sched_domain
*sd
;
6493 struct sched_group
*sg
;
6494 struct sched_group_power
*sgp
;
6496 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6497 GFP_KERNEL
, cpu_to_node(j
));
6501 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6503 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6504 GFP_KERNEL
, cpu_to_node(j
));
6510 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6512 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6513 GFP_KERNEL
, cpu_to_node(j
));
6517 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6524 static void __sdt_free(const struct cpumask
*cpu_map
)
6526 struct sched_domain_topology_level
*tl
;
6529 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6530 struct sd_data
*sdd
= &tl
->data
;
6532 for_each_cpu(j
, cpu_map
) {
6533 struct sched_domain
*sd
;
6536 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6537 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6538 free_sched_groups(sd
->groups
, 0);
6539 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6543 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6545 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6547 free_percpu(sdd
->sd
);
6549 free_percpu(sdd
->sg
);
6551 free_percpu(sdd
->sgp
);
6556 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6557 struct s_data
*d
, const struct cpumask
*cpu_map
,
6558 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6561 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6565 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6567 sd
->level
= child
->level
+ 1;
6568 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6572 set_domain_attribute(sd
, attr
);
6578 * Build sched domains for a given set of cpus and attach the sched domains
6579 * to the individual cpus
6581 static int build_sched_domains(const struct cpumask
*cpu_map
,
6582 struct sched_domain_attr
*attr
)
6584 enum s_alloc alloc_state
= sa_none
;
6585 struct sched_domain
*sd
;
6587 int i
, ret
= -ENOMEM
;
6589 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6590 if (alloc_state
!= sa_rootdomain
)
6593 /* Set up domains for cpus specified by the cpu_map. */
6594 for_each_cpu(i
, cpu_map
) {
6595 struct sched_domain_topology_level
*tl
;
6598 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6599 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6600 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6601 sd
->flags
|= SD_OVERLAP
;
6602 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6609 *per_cpu_ptr(d
.sd
, i
) = sd
;
6612 /* Build the groups for the domains */
6613 for_each_cpu(i
, cpu_map
) {
6614 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6615 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6616 if (sd
->flags
& SD_OVERLAP
) {
6617 if (build_overlap_sched_groups(sd
, i
))
6620 if (build_sched_groups(sd
, i
))
6626 /* Calculate CPU power for physical packages and nodes */
6627 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6628 if (!cpumask_test_cpu(i
, cpu_map
))
6631 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6632 claim_allocations(i
, sd
);
6633 init_sched_groups_power(i
, sd
);
6637 /* Attach the domains */
6639 for_each_cpu(i
, cpu_map
) {
6640 sd
= *per_cpu_ptr(d
.sd
, i
);
6641 cpu_attach_domain(sd
, d
.rd
, i
);
6647 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6651 static cpumask_var_t
*doms_cur
; /* current sched domains */
6652 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6653 static struct sched_domain_attr
*dattr_cur
;
6654 /* attribues of custom domains in 'doms_cur' */
6657 * Special case: If a kmalloc of a doms_cur partition (array of
6658 * cpumask) fails, then fallback to a single sched domain,
6659 * as determined by the single cpumask fallback_doms.
6661 static cpumask_var_t fallback_doms
;
6664 * arch_update_cpu_topology lets virtualized architectures update the
6665 * cpu core maps. It is supposed to return 1 if the topology changed
6666 * or 0 if it stayed the same.
6668 int __attribute__((weak
)) arch_update_cpu_topology(void)
6673 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6676 cpumask_var_t
*doms
;
6678 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6681 for (i
= 0; i
< ndoms
; i
++) {
6682 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6683 free_sched_domains(doms
, i
);
6690 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6693 for (i
= 0; i
< ndoms
; i
++)
6694 free_cpumask_var(doms
[i
]);
6699 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6700 * For now this just excludes isolated cpus, but could be used to
6701 * exclude other special cases in the future.
6703 static int init_sched_domains(const struct cpumask
*cpu_map
)
6707 arch_update_cpu_topology();
6709 doms_cur
= alloc_sched_domains(ndoms_cur
);
6711 doms_cur
= &fallback_doms
;
6712 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6713 err
= build_sched_domains(doms_cur
[0], NULL
);
6714 register_sched_domain_sysctl();
6720 * Detach sched domains from a group of cpus specified in cpu_map
6721 * These cpus will now be attached to the NULL domain
6723 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6728 for_each_cpu(i
, cpu_map
)
6729 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6733 /* handle null as "default" */
6734 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6735 struct sched_domain_attr
*new, int idx_new
)
6737 struct sched_domain_attr tmp
;
6744 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6745 new ? (new + idx_new
) : &tmp
,
6746 sizeof(struct sched_domain_attr
));
6750 * Partition sched domains as specified by the 'ndoms_new'
6751 * cpumasks in the array doms_new[] of cpumasks. This compares
6752 * doms_new[] to the current sched domain partitioning, doms_cur[].
6753 * It destroys each deleted domain and builds each new domain.
6755 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6756 * The masks don't intersect (don't overlap.) We should setup one
6757 * sched domain for each mask. CPUs not in any of the cpumasks will
6758 * not be load balanced. If the same cpumask appears both in the
6759 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6762 * The passed in 'doms_new' should be allocated using
6763 * alloc_sched_domains. This routine takes ownership of it and will
6764 * free_sched_domains it when done with it. If the caller failed the
6765 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6766 * and partition_sched_domains() will fallback to the single partition
6767 * 'fallback_doms', it also forces the domains to be rebuilt.
6769 * If doms_new == NULL it will be replaced with cpu_online_mask.
6770 * ndoms_new == 0 is a special case for destroying existing domains,
6771 * and it will not create the default domain.
6773 * Call with hotplug lock held
6775 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6776 struct sched_domain_attr
*dattr_new
)
6781 mutex_lock(&sched_domains_mutex
);
6783 /* always unregister in case we don't destroy any domains */
6784 unregister_sched_domain_sysctl();
6786 /* Let architecture update cpu core mappings. */
6787 new_topology
= arch_update_cpu_topology();
6789 n
= doms_new
? ndoms_new
: 0;
6791 /* Destroy deleted domains */
6792 for (i
= 0; i
< ndoms_cur
; i
++) {
6793 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6794 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6795 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6798 /* no match - a current sched domain not in new doms_new[] */
6799 detach_destroy_domains(doms_cur
[i
]);
6804 if (doms_new
== NULL
) {
6806 doms_new
= &fallback_doms
;
6807 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6808 WARN_ON_ONCE(dattr_new
);
6811 /* Build new domains */
6812 for (i
= 0; i
< ndoms_new
; i
++) {
6813 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6814 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6815 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6818 /* no match - add a new doms_new */
6819 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6824 /* Remember the new sched domains */
6825 if (doms_cur
!= &fallback_doms
)
6826 free_sched_domains(doms_cur
, ndoms_cur
);
6827 kfree(dattr_cur
); /* kfree(NULL) is safe */
6828 doms_cur
= doms_new
;
6829 dattr_cur
= dattr_new
;
6830 ndoms_cur
= ndoms_new
;
6832 register_sched_domain_sysctl();
6834 mutex_unlock(&sched_domains_mutex
);
6837 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6840 * Update cpusets according to cpu_active mask. If cpusets are
6841 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6842 * around partition_sched_domains().
6844 * If we come here as part of a suspend/resume, don't touch cpusets because we
6845 * want to restore it back to its original state upon resume anyway.
6847 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6851 case CPU_ONLINE_FROZEN
:
6852 case CPU_DOWN_FAILED_FROZEN
:
6855 * num_cpus_frozen tracks how many CPUs are involved in suspend
6856 * resume sequence. As long as this is not the last online
6857 * operation in the resume sequence, just build a single sched
6858 * domain, ignoring cpusets.
6861 if (likely(num_cpus_frozen
)) {
6862 partition_sched_domains(1, NULL
, NULL
);
6867 * This is the last CPU online operation. So fall through and
6868 * restore the original sched domains by considering the
6869 * cpuset configurations.
6873 case CPU_DOWN_FAILED
:
6874 cpuset_update_active_cpus(true);
6882 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6886 case CPU_DOWN_PREPARE
:
6887 cpuset_update_active_cpus(false);
6889 case CPU_DOWN_PREPARE_FROZEN
:
6891 partition_sched_domains(1, NULL
, NULL
);
6899 void __init
sched_init_smp(void)
6901 cpumask_var_t non_isolated_cpus
;
6903 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6904 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6909 mutex_lock(&sched_domains_mutex
);
6910 init_sched_domains(cpu_active_mask
);
6911 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6912 if (cpumask_empty(non_isolated_cpus
))
6913 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6914 mutex_unlock(&sched_domains_mutex
);
6917 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6918 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6919 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6921 /* RT runtime code needs to handle some hotplug events */
6922 hotcpu_notifier(update_runtime
, 0);
6926 /* Move init over to a non-isolated CPU */
6927 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6929 sched_init_granularity();
6930 free_cpumask_var(non_isolated_cpus
);
6932 init_sched_rt_class();
6935 void __init
sched_init_smp(void)
6937 sched_init_granularity();
6939 #endif /* CONFIG_SMP */
6941 const_debug
unsigned int sysctl_timer_migration
= 1;
6943 int in_sched_functions(unsigned long addr
)
6945 return in_lock_functions(addr
) ||
6946 (addr
>= (unsigned long)__sched_text_start
6947 && addr
< (unsigned long)__sched_text_end
);
6950 #ifdef CONFIG_CGROUP_SCHED
6952 * Default task group.
6953 * Every task in system belongs to this group at bootup.
6955 struct task_group root_task_group
;
6956 LIST_HEAD(task_groups
);
6959 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6961 void __init
sched_init(void)
6964 unsigned long alloc_size
= 0, ptr
;
6966 #ifdef CONFIG_FAIR_GROUP_SCHED
6967 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6969 #ifdef CONFIG_RT_GROUP_SCHED
6970 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6972 #ifdef CONFIG_CPUMASK_OFFSTACK
6973 alloc_size
+= num_possible_cpus() * cpumask_size();
6976 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6978 #ifdef CONFIG_FAIR_GROUP_SCHED
6979 root_task_group
.se
= (struct sched_entity
**)ptr
;
6980 ptr
+= nr_cpu_ids
* sizeof(void **);
6982 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6983 ptr
+= nr_cpu_ids
* sizeof(void **);
6985 #endif /* CONFIG_FAIR_GROUP_SCHED */
6986 #ifdef CONFIG_RT_GROUP_SCHED
6987 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6988 ptr
+= nr_cpu_ids
* sizeof(void **);
6990 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6991 ptr
+= nr_cpu_ids
* sizeof(void **);
6993 #endif /* CONFIG_RT_GROUP_SCHED */
6994 #ifdef CONFIG_CPUMASK_OFFSTACK
6995 for_each_possible_cpu(i
) {
6996 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6997 ptr
+= cpumask_size();
6999 #endif /* CONFIG_CPUMASK_OFFSTACK */
7003 init_defrootdomain();
7006 init_rt_bandwidth(&def_rt_bandwidth
,
7007 global_rt_period(), global_rt_runtime());
7009 #ifdef CONFIG_RT_GROUP_SCHED
7010 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7011 global_rt_period(), global_rt_runtime());
7012 #endif /* CONFIG_RT_GROUP_SCHED */
7014 #ifdef CONFIG_CGROUP_SCHED
7015 list_add(&root_task_group
.list
, &task_groups
);
7016 INIT_LIST_HEAD(&root_task_group
.children
);
7017 INIT_LIST_HEAD(&root_task_group
.siblings
);
7018 autogroup_init(&init_task
);
7020 #endif /* CONFIG_CGROUP_SCHED */
7022 for_each_possible_cpu(i
) {
7026 raw_spin_lock_init(&rq
->lock
);
7028 rq
->calc_load_active
= 0;
7029 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7030 init_cfs_rq(&rq
->cfs
);
7031 init_rt_rq(&rq
->rt
, rq
);
7032 #ifdef CONFIG_FAIR_GROUP_SCHED
7033 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7034 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7036 * How much cpu bandwidth does root_task_group get?
7038 * In case of task-groups formed thr' the cgroup filesystem, it
7039 * gets 100% of the cpu resources in the system. This overall
7040 * system cpu resource is divided among the tasks of
7041 * root_task_group and its child task-groups in a fair manner,
7042 * based on each entity's (task or task-group's) weight
7043 * (se->load.weight).
7045 * In other words, if root_task_group has 10 tasks of weight
7046 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7047 * then A0's share of the cpu resource is:
7049 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7051 * We achieve this by letting root_task_group's tasks sit
7052 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7054 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7055 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7056 #endif /* CONFIG_FAIR_GROUP_SCHED */
7058 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7059 #ifdef CONFIG_RT_GROUP_SCHED
7060 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7061 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7064 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7065 rq
->cpu_load
[j
] = 0;
7067 rq
->last_load_update_tick
= jiffies
;
7072 rq
->cpu_power
= SCHED_POWER_SCALE
;
7073 rq
->post_schedule
= 0;
7074 rq
->active_balance
= 0;
7075 rq
->next_balance
= jiffies
;
7080 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7082 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7084 rq_attach_root(rq
, &def_root_domain
);
7085 #ifdef CONFIG_NO_HZ_COMMON
7088 #ifdef CONFIG_NO_HZ_FULL
7089 rq
->last_sched_tick
= 0;
7093 atomic_set(&rq
->nr_iowait
, 0);
7096 set_load_weight(&init_task
);
7098 #ifdef CONFIG_PREEMPT_NOTIFIERS
7099 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7102 #ifdef CONFIG_RT_MUTEXES
7103 plist_head_init(&init_task
.pi_waiters
);
7107 * The boot idle thread does lazy MMU switching as well:
7109 atomic_inc(&init_mm
.mm_count
);
7110 enter_lazy_tlb(&init_mm
, current
);
7113 * Make us the idle thread. Technically, schedule() should not be
7114 * called from this thread, however somewhere below it might be,
7115 * but because we are the idle thread, we just pick up running again
7116 * when this runqueue becomes "idle".
7118 init_idle(current
, smp_processor_id());
7120 calc_load_update
= jiffies
+ LOAD_FREQ
;
7123 * During early bootup we pretend to be a normal task:
7125 current
->sched_class
= &fair_sched_class
;
7128 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7129 /* May be allocated at isolcpus cmdline parse time */
7130 if (cpu_isolated_map
== NULL
)
7131 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7132 idle_thread_set_boot_cpu();
7134 init_sched_fair_class();
7136 scheduler_running
= 1;
7139 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7140 static inline int preempt_count_equals(int preempt_offset
)
7142 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7144 return (nested
== preempt_offset
);
7147 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7149 static unsigned long prev_jiffy
; /* ratelimiting */
7151 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7152 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7153 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7155 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7157 prev_jiffy
= jiffies
;
7160 "BUG: sleeping function called from invalid context at %s:%d\n",
7163 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7164 in_atomic(), irqs_disabled(),
7165 current
->pid
, current
->comm
);
7167 debug_show_held_locks(current
);
7168 if (irqs_disabled())
7169 print_irqtrace_events(current
);
7172 EXPORT_SYMBOL(__might_sleep
);
7175 #ifdef CONFIG_MAGIC_SYSRQ
7176 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7178 const struct sched_class
*prev_class
= p
->sched_class
;
7179 int old_prio
= p
->prio
;
7184 dequeue_task(rq
, p
, 0);
7185 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7187 enqueue_task(rq
, p
, 0);
7188 resched_task(rq
->curr
);
7191 check_class_changed(rq
, p
, prev_class
, old_prio
);
7194 void normalize_rt_tasks(void)
7196 struct task_struct
*g
, *p
;
7197 unsigned long flags
;
7200 read_lock_irqsave(&tasklist_lock
, flags
);
7201 do_each_thread(g
, p
) {
7203 * Only normalize user tasks:
7208 p
->se
.exec_start
= 0;
7209 #ifdef CONFIG_SCHEDSTATS
7210 p
->se
.statistics
.wait_start
= 0;
7211 p
->se
.statistics
.sleep_start
= 0;
7212 p
->se
.statistics
.block_start
= 0;
7217 * Renice negative nice level userspace
7220 if (TASK_NICE(p
) < 0 && p
->mm
)
7221 set_user_nice(p
, 0);
7225 raw_spin_lock(&p
->pi_lock
);
7226 rq
= __task_rq_lock(p
);
7228 normalize_task(rq
, p
);
7230 __task_rq_unlock(rq
);
7231 raw_spin_unlock(&p
->pi_lock
);
7232 } while_each_thread(g
, p
);
7234 read_unlock_irqrestore(&tasklist_lock
, flags
);
7237 #endif /* CONFIG_MAGIC_SYSRQ */
7239 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7241 * These functions are only useful for the IA64 MCA handling, or kdb.
7243 * They can only be called when the whole system has been
7244 * stopped - every CPU needs to be quiescent, and no scheduling
7245 * activity can take place. Using them for anything else would
7246 * be a serious bug, and as a result, they aren't even visible
7247 * under any other configuration.
7251 * curr_task - return the current task for a given cpu.
7252 * @cpu: the processor in question.
7254 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7256 struct task_struct
*curr_task(int cpu
)
7258 return cpu_curr(cpu
);
7261 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7265 * set_curr_task - set the current task for a given cpu.
7266 * @cpu: the processor in question.
7267 * @p: the task pointer to set.
7269 * Description: This function must only be used when non-maskable interrupts
7270 * are serviced on a separate stack. It allows the architecture to switch the
7271 * notion of the current task on a cpu in a non-blocking manner. This function
7272 * must be called with all CPU's synchronized, and interrupts disabled, the
7273 * and caller must save the original value of the current task (see
7274 * curr_task() above) and restore that value before reenabling interrupts and
7275 * re-starting the system.
7277 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7279 void set_curr_task(int cpu
, struct task_struct
*p
)
7286 #ifdef CONFIG_CGROUP_SCHED
7287 /* task_group_lock serializes the addition/removal of task groups */
7288 static DEFINE_SPINLOCK(task_group_lock
);
7290 static void free_sched_group(struct task_group
*tg
)
7292 free_fair_sched_group(tg
);
7293 free_rt_sched_group(tg
);
7298 /* allocate runqueue etc for a new task group */
7299 struct task_group
*sched_create_group(struct task_group
*parent
)
7301 struct task_group
*tg
;
7303 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7305 return ERR_PTR(-ENOMEM
);
7307 if (!alloc_fair_sched_group(tg
, parent
))
7310 if (!alloc_rt_sched_group(tg
, parent
))
7316 free_sched_group(tg
);
7317 return ERR_PTR(-ENOMEM
);
7320 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7322 unsigned long flags
;
7324 spin_lock_irqsave(&task_group_lock
, flags
);
7325 list_add_rcu(&tg
->list
, &task_groups
);
7327 WARN_ON(!parent
); /* root should already exist */
7329 tg
->parent
= parent
;
7330 INIT_LIST_HEAD(&tg
->children
);
7331 list_add_rcu(&tg
->siblings
, &parent
->children
);
7332 spin_unlock_irqrestore(&task_group_lock
, flags
);
7335 /* rcu callback to free various structures associated with a task group */
7336 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7338 /* now it should be safe to free those cfs_rqs */
7339 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7342 /* Destroy runqueue etc associated with a task group */
7343 void sched_destroy_group(struct task_group
*tg
)
7345 /* wait for possible concurrent references to cfs_rqs complete */
7346 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7349 void sched_offline_group(struct task_group
*tg
)
7351 unsigned long flags
;
7354 /* end participation in shares distribution */
7355 for_each_possible_cpu(i
)
7356 unregister_fair_sched_group(tg
, i
);
7358 spin_lock_irqsave(&task_group_lock
, flags
);
7359 list_del_rcu(&tg
->list
);
7360 list_del_rcu(&tg
->siblings
);
7361 spin_unlock_irqrestore(&task_group_lock
, flags
);
7364 /* change task's runqueue when it moves between groups.
7365 * The caller of this function should have put the task in its new group
7366 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7367 * reflect its new group.
7369 void sched_move_task(struct task_struct
*tsk
)
7371 struct task_group
*tg
;
7373 unsigned long flags
;
7376 rq
= task_rq_lock(tsk
, &flags
);
7378 running
= task_current(rq
, tsk
);
7382 dequeue_task(rq
, tsk
, 0);
7383 if (unlikely(running
))
7384 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7386 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7387 lockdep_is_held(&tsk
->sighand
->siglock
)),
7388 struct task_group
, css
);
7389 tg
= autogroup_task_group(tsk
, tg
);
7390 tsk
->sched_task_group
= tg
;
7392 #ifdef CONFIG_FAIR_GROUP_SCHED
7393 if (tsk
->sched_class
->task_move_group
)
7394 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7397 set_task_rq(tsk
, task_cpu(tsk
));
7399 if (unlikely(running
))
7400 tsk
->sched_class
->set_curr_task(rq
);
7402 enqueue_task(rq
, tsk
, 0);
7404 task_rq_unlock(rq
, tsk
, &flags
);
7406 #endif /* CONFIG_CGROUP_SCHED */
7408 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7409 static unsigned long to_ratio(u64 period
, u64 runtime
)
7411 if (runtime
== RUNTIME_INF
)
7414 return div64_u64(runtime
<< 20, period
);
7418 #ifdef CONFIG_RT_GROUP_SCHED
7420 * Ensure that the real time constraints are schedulable.
7422 static DEFINE_MUTEX(rt_constraints_mutex
);
7424 /* Must be called with tasklist_lock held */
7425 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7427 struct task_struct
*g
, *p
;
7429 do_each_thread(g
, p
) {
7430 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7432 } while_each_thread(g
, p
);
7437 struct rt_schedulable_data
{
7438 struct task_group
*tg
;
7443 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7445 struct rt_schedulable_data
*d
= data
;
7446 struct task_group
*child
;
7447 unsigned long total
, sum
= 0;
7448 u64 period
, runtime
;
7450 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7451 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7454 period
= d
->rt_period
;
7455 runtime
= d
->rt_runtime
;
7459 * Cannot have more runtime than the period.
7461 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7465 * Ensure we don't starve existing RT tasks.
7467 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7470 total
= to_ratio(period
, runtime
);
7473 * Nobody can have more than the global setting allows.
7475 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7479 * The sum of our children's runtime should not exceed our own.
7481 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7482 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7483 runtime
= child
->rt_bandwidth
.rt_runtime
;
7485 if (child
== d
->tg
) {
7486 period
= d
->rt_period
;
7487 runtime
= d
->rt_runtime
;
7490 sum
+= to_ratio(period
, runtime
);
7499 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7503 struct rt_schedulable_data data
= {
7505 .rt_period
= period
,
7506 .rt_runtime
= runtime
,
7510 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7516 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7517 u64 rt_period
, u64 rt_runtime
)
7521 mutex_lock(&rt_constraints_mutex
);
7522 read_lock(&tasklist_lock
);
7523 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7527 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7528 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7529 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7531 for_each_possible_cpu(i
) {
7532 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7534 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7535 rt_rq
->rt_runtime
= rt_runtime
;
7536 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7538 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7540 read_unlock(&tasklist_lock
);
7541 mutex_unlock(&rt_constraints_mutex
);
7546 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7548 u64 rt_runtime
, rt_period
;
7550 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7551 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7552 if (rt_runtime_us
< 0)
7553 rt_runtime
= RUNTIME_INF
;
7555 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7558 static long sched_group_rt_runtime(struct task_group
*tg
)
7562 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7565 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7566 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7567 return rt_runtime_us
;
7570 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7572 u64 rt_runtime
, rt_period
;
7574 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7575 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7580 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7583 static long sched_group_rt_period(struct task_group
*tg
)
7587 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7588 do_div(rt_period_us
, NSEC_PER_USEC
);
7589 return rt_period_us
;
7592 static int sched_rt_global_constraints(void)
7594 u64 runtime
, period
;
7597 if (sysctl_sched_rt_period
<= 0)
7600 runtime
= global_rt_runtime();
7601 period
= global_rt_period();
7604 * Sanity check on the sysctl variables.
7606 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7609 mutex_lock(&rt_constraints_mutex
);
7610 read_lock(&tasklist_lock
);
7611 ret
= __rt_schedulable(NULL
, 0, 0);
7612 read_unlock(&tasklist_lock
);
7613 mutex_unlock(&rt_constraints_mutex
);
7618 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7620 /* Don't accept realtime tasks when there is no way for them to run */
7621 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7627 #else /* !CONFIG_RT_GROUP_SCHED */
7628 static int sched_rt_global_constraints(void)
7630 unsigned long flags
;
7633 if (sysctl_sched_rt_period
<= 0)
7637 * There's always some RT tasks in the root group
7638 * -- migration, kstopmachine etc..
7640 if (sysctl_sched_rt_runtime
== 0)
7643 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7644 for_each_possible_cpu(i
) {
7645 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7647 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7648 rt_rq
->rt_runtime
= global_rt_runtime();
7649 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7651 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7655 #endif /* CONFIG_RT_GROUP_SCHED */
7657 int sched_rr_handler(struct ctl_table
*table
, int write
,
7658 void __user
*buffer
, size_t *lenp
,
7662 static DEFINE_MUTEX(mutex
);
7665 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7666 /* make sure that internally we keep jiffies */
7667 /* also, writing zero resets timeslice to default */
7668 if (!ret
&& write
) {
7669 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7670 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7672 mutex_unlock(&mutex
);
7676 int sched_rt_handler(struct ctl_table
*table
, int write
,
7677 void __user
*buffer
, size_t *lenp
,
7681 int old_period
, old_runtime
;
7682 static DEFINE_MUTEX(mutex
);
7685 old_period
= sysctl_sched_rt_period
;
7686 old_runtime
= sysctl_sched_rt_runtime
;
7688 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7690 if (!ret
&& write
) {
7691 ret
= sched_rt_global_constraints();
7693 sysctl_sched_rt_period
= old_period
;
7694 sysctl_sched_rt_runtime
= old_runtime
;
7696 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7697 def_rt_bandwidth
.rt_period
=
7698 ns_to_ktime(global_rt_period());
7701 mutex_unlock(&mutex
);
7706 #ifdef CONFIG_CGROUP_SCHED
7708 /* return corresponding task_group object of a cgroup */
7709 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7711 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7712 struct task_group
, css
);
7715 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7717 struct task_group
*tg
, *parent
;
7719 if (!cgrp
->parent
) {
7720 /* This is early initialization for the top cgroup */
7721 return &root_task_group
.css
;
7724 parent
= cgroup_tg(cgrp
->parent
);
7725 tg
= sched_create_group(parent
);
7727 return ERR_PTR(-ENOMEM
);
7732 static int cpu_cgroup_css_online(struct cgroup
*cgrp
)
7734 struct task_group
*tg
= cgroup_tg(cgrp
);
7735 struct task_group
*parent
;
7740 parent
= cgroup_tg(cgrp
->parent
);
7741 sched_online_group(tg
, parent
);
7745 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7747 struct task_group
*tg
= cgroup_tg(cgrp
);
7749 sched_destroy_group(tg
);
7752 static void cpu_cgroup_css_offline(struct cgroup
*cgrp
)
7754 struct task_group
*tg
= cgroup_tg(cgrp
);
7756 sched_offline_group(tg
);
7759 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7760 struct cgroup_taskset
*tset
)
7762 struct task_struct
*task
;
7764 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7765 #ifdef CONFIG_RT_GROUP_SCHED
7766 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7769 /* We don't support RT-tasks being in separate groups */
7770 if (task
->sched_class
!= &fair_sched_class
)
7777 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7778 struct cgroup_taskset
*tset
)
7780 struct task_struct
*task
;
7782 cgroup_taskset_for_each(task
, cgrp
, tset
)
7783 sched_move_task(task
);
7787 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7788 struct task_struct
*task
)
7791 * cgroup_exit() is called in the copy_process() failure path.
7792 * Ignore this case since the task hasn't ran yet, this avoids
7793 * trying to poke a half freed task state from generic code.
7795 if (!(task
->flags
& PF_EXITING
))
7798 sched_move_task(task
);
7801 #ifdef CONFIG_FAIR_GROUP_SCHED
7802 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7805 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7808 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7810 struct task_group
*tg
= cgroup_tg(cgrp
);
7812 return (u64
) scale_load_down(tg
->shares
);
7815 #ifdef CONFIG_CFS_BANDWIDTH
7816 static DEFINE_MUTEX(cfs_constraints_mutex
);
7818 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7819 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7821 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7823 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7825 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7826 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7828 if (tg
== &root_task_group
)
7832 * Ensure we have at some amount of bandwidth every period. This is
7833 * to prevent reaching a state of large arrears when throttled via
7834 * entity_tick() resulting in prolonged exit starvation.
7836 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7840 * Likewise, bound things on the otherside by preventing insane quota
7841 * periods. This also allows us to normalize in computing quota
7844 if (period
> max_cfs_quota_period
)
7847 mutex_lock(&cfs_constraints_mutex
);
7848 ret
= __cfs_schedulable(tg
, period
, quota
);
7852 runtime_enabled
= quota
!= RUNTIME_INF
;
7853 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7855 * If we need to toggle cfs_bandwidth_used, off->on must occur
7856 * before making related changes, and on->off must occur afterwards
7858 if (runtime_enabled
&& !runtime_was_enabled
)
7859 cfs_bandwidth_usage_inc();
7860 raw_spin_lock_irq(&cfs_b
->lock
);
7861 cfs_b
->period
= ns_to_ktime(period
);
7862 cfs_b
->quota
= quota
;
7864 __refill_cfs_bandwidth_runtime(cfs_b
);
7865 /* restart the period timer (if active) to handle new period expiry */
7866 if (runtime_enabled
&& cfs_b
->timer_active
) {
7867 /* force a reprogram */
7868 cfs_b
->timer_active
= 0;
7869 __start_cfs_bandwidth(cfs_b
);
7871 raw_spin_unlock_irq(&cfs_b
->lock
);
7873 for_each_possible_cpu(i
) {
7874 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7875 struct rq
*rq
= cfs_rq
->rq
;
7877 raw_spin_lock_irq(&rq
->lock
);
7878 cfs_rq
->runtime_enabled
= runtime_enabled
;
7879 cfs_rq
->runtime_remaining
= 0;
7881 if (cfs_rq
->throttled
)
7882 unthrottle_cfs_rq(cfs_rq
);
7883 raw_spin_unlock_irq(&rq
->lock
);
7885 if (runtime_was_enabled
&& !runtime_enabled
)
7886 cfs_bandwidth_usage_dec();
7888 mutex_unlock(&cfs_constraints_mutex
);
7893 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7897 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7898 if (cfs_quota_us
< 0)
7899 quota
= RUNTIME_INF
;
7901 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7903 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7906 long tg_get_cfs_quota(struct task_group
*tg
)
7910 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7913 quota_us
= tg
->cfs_bandwidth
.quota
;
7914 do_div(quota_us
, NSEC_PER_USEC
);
7919 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7923 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7924 quota
= tg
->cfs_bandwidth
.quota
;
7926 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7929 long tg_get_cfs_period(struct task_group
*tg
)
7933 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7934 do_div(cfs_period_us
, NSEC_PER_USEC
);
7936 return cfs_period_us
;
7939 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7941 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7944 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7947 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7950 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7952 return tg_get_cfs_period(cgroup_tg(cgrp
));
7955 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7958 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7961 struct cfs_schedulable_data
{
7962 struct task_group
*tg
;
7967 * normalize group quota/period to be quota/max_period
7968 * note: units are usecs
7970 static u64
normalize_cfs_quota(struct task_group
*tg
,
7971 struct cfs_schedulable_data
*d
)
7979 period
= tg_get_cfs_period(tg
);
7980 quota
= tg_get_cfs_quota(tg
);
7983 /* note: these should typically be equivalent */
7984 if (quota
== RUNTIME_INF
|| quota
== -1)
7987 return to_ratio(period
, quota
);
7990 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7992 struct cfs_schedulable_data
*d
= data
;
7993 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7994 s64 quota
= 0, parent_quota
= -1;
7997 quota
= RUNTIME_INF
;
7999 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8001 quota
= normalize_cfs_quota(tg
, d
);
8002 parent_quota
= parent_b
->hierarchal_quota
;
8005 * ensure max(child_quota) <= parent_quota, inherit when no
8008 if (quota
== RUNTIME_INF
)
8009 quota
= parent_quota
;
8010 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8013 cfs_b
->hierarchal_quota
= quota
;
8018 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8021 struct cfs_schedulable_data data
= {
8027 if (quota
!= RUNTIME_INF
) {
8028 do_div(data
.period
, NSEC_PER_USEC
);
8029 do_div(data
.quota
, NSEC_PER_USEC
);
8033 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8039 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8040 struct cgroup_map_cb
*cb
)
8042 struct task_group
*tg
= cgroup_tg(cgrp
);
8043 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8045 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
8046 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
8047 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
8051 #endif /* CONFIG_CFS_BANDWIDTH */
8052 #endif /* CONFIG_FAIR_GROUP_SCHED */
8054 #ifdef CONFIG_RT_GROUP_SCHED
8055 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8058 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8061 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8063 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8066 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8069 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8072 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8074 return sched_group_rt_period(cgroup_tg(cgrp
));
8076 #endif /* CONFIG_RT_GROUP_SCHED */
8078 static struct cftype cpu_files
[] = {
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8082 .read_u64
= cpu_shares_read_u64
,
8083 .write_u64
= cpu_shares_write_u64
,
8086 #ifdef CONFIG_CFS_BANDWIDTH
8088 .name
= "cfs_quota_us",
8089 .read_s64
= cpu_cfs_quota_read_s64
,
8090 .write_s64
= cpu_cfs_quota_write_s64
,
8093 .name
= "cfs_period_us",
8094 .read_u64
= cpu_cfs_period_read_u64
,
8095 .write_u64
= cpu_cfs_period_write_u64
,
8099 .read_map
= cpu_stats_show
,
8102 #ifdef CONFIG_RT_GROUP_SCHED
8104 .name
= "rt_runtime_us",
8105 .read_s64
= cpu_rt_runtime_read
,
8106 .write_s64
= cpu_rt_runtime_write
,
8109 .name
= "rt_period_us",
8110 .read_u64
= cpu_rt_period_read_uint
,
8111 .write_u64
= cpu_rt_period_write_uint
,
8117 struct cgroup_subsys cpu_cgroup_subsys
= {
8119 .css_alloc
= cpu_cgroup_css_alloc
,
8120 .css_free
= cpu_cgroup_css_free
,
8121 .css_online
= cpu_cgroup_css_online
,
8122 .css_offline
= cpu_cgroup_css_offline
,
8123 .can_attach
= cpu_cgroup_can_attach
,
8124 .attach
= cpu_cgroup_attach
,
8125 .exit
= cpu_cgroup_exit
,
8126 .subsys_id
= cpu_cgroup_subsys_id
,
8127 .base_cftypes
= cpu_files
,
8131 #endif /* CONFIG_CGROUP_SCHED */
8133 void dump_cpu_task(int cpu
)
8135 pr_info("Task dump for CPU %d:\n", cpu
);
8136 sched_show_task(cpu_curr(cpu
));