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 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp
)
201 if (strncmp(cmp
, "NO_", 3) == 0) {
206 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
207 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
209 sysctl_sched_features
&= ~(1UL << i
);
210 sched_feat_disable(i
);
212 sysctl_sched_features
|= (1UL << i
);
213 sched_feat_enable(i
);
223 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
224 size_t cnt
, loff_t
*ppos
)
233 if (copy_from_user(&buf
, ubuf
, cnt
))
239 i
= sched_feat_set(cmp
);
240 if (i
== __SCHED_FEAT_NR
)
248 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
250 return single_open(filp
, sched_feat_show
, NULL
);
253 static const struct file_operations sched_feat_fops
= {
254 .open
= sched_feat_open
,
255 .write
= sched_feat_write
,
258 .release
= single_release
,
261 static __init
int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
268 late_initcall(sched_init_debug
);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
278 * period over which we average the RT time consumption, measured
283 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period
= 1000000;
291 __read_mostly
int scheduler_running
;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime
= 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
309 lockdep_assert_held(&p
->pi_lock
);
313 raw_spin_lock(&rq
->lock
);
314 if (likely(rq
== task_rq(p
)))
316 raw_spin_unlock(&rq
->lock
);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
324 __acquires(p
->pi_lock
)
330 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
332 raw_spin_lock(&rq
->lock
);
333 if (likely(rq
== task_rq(p
)))
335 raw_spin_unlock(&rq
->lock
);
336 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
340 static void __task_rq_unlock(struct rq
*rq
)
343 raw_spin_unlock(&rq
->lock
);
347 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
349 __releases(p
->pi_lock
)
351 raw_spin_unlock(&rq
->lock
);
352 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq
*this_rq_lock(void)
365 raw_spin_lock(&rq
->lock
);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
375 static void hrtick_clear(struct rq
*rq
)
377 if (hrtimer_active(&rq
->hrtick_timer
))
378 hrtimer_cancel(&rq
->hrtick_timer
);
382 * High-resolution timer tick.
383 * Runs from hardirq context with interrupts disabled.
385 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
387 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
389 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
391 raw_spin_lock(&rq
->lock
);
393 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
394 raw_spin_unlock(&rq
->lock
);
396 return HRTIMER_NORESTART
;
401 static int __hrtick_restart(struct rq
*rq
)
403 struct hrtimer
*timer
= &rq
->hrtick_timer
;
404 ktime_t time
= hrtimer_get_softexpires(timer
);
406 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
410 * called from hardirq (IPI) context
412 static void __hrtick_start(void *arg
)
416 raw_spin_lock(&rq
->lock
);
417 __hrtick_restart(rq
);
418 rq
->hrtick_csd_pending
= 0;
419 raw_spin_unlock(&rq
->lock
);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq
*rq
, u64 delay
)
429 struct hrtimer
*timer
= &rq
->hrtick_timer
;
430 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
432 hrtimer_set_expires(timer
, time
);
434 if (rq
== this_rq()) {
435 __hrtick_restart(rq
);
436 } else if (!rq
->hrtick_csd_pending
) {
437 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
438 rq
->hrtick_csd_pending
= 1;
443 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
445 int cpu
= (int)(long)hcpu
;
448 case CPU_UP_CANCELED
:
449 case CPU_UP_CANCELED_FROZEN
:
450 case CPU_DOWN_PREPARE
:
451 case CPU_DOWN_PREPARE_FROZEN
:
453 case CPU_DEAD_FROZEN
:
454 hrtick_clear(cpu_rq(cpu
));
461 static __init
void init_hrtick(void)
463 hotcpu_notifier(hotplug_hrtick
, 0);
467 * Called to set the hrtick timer state.
469 * called with rq->lock held and irqs disabled
471 void hrtick_start(struct rq
*rq
, u64 delay
)
473 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
474 HRTIMER_MODE_REL_PINNED
, 0);
477 static inline void init_hrtick(void)
480 #endif /* CONFIG_SMP */
482 static void init_rq_hrtick(struct rq
*rq
)
485 rq
->hrtick_csd_pending
= 0;
487 rq
->hrtick_csd
.flags
= 0;
488 rq
->hrtick_csd
.func
= __hrtick_start
;
489 rq
->hrtick_csd
.info
= rq
;
492 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
493 rq
->hrtick_timer
.function
= hrtick
;
495 #else /* CONFIG_SCHED_HRTICK */
496 static inline void hrtick_clear(struct rq
*rq
)
500 static inline void init_rq_hrtick(struct rq
*rq
)
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SCHED_HRTICK */
510 * resched_task - mark a task 'to be rescheduled now'.
512 * On UP this means the setting of the need_resched flag, on SMP it
513 * might also involve a cross-CPU call to trigger the scheduler on
516 void resched_task(struct task_struct
*p
)
520 lockdep_assert_held(&task_rq(p
)->lock
);
522 if (test_tsk_need_resched(p
))
525 set_tsk_need_resched(p
);
528 if (cpu
== smp_processor_id()) {
529 set_preempt_need_resched();
533 /* NEED_RESCHED must be visible before we test polling */
535 if (!tsk_is_polling(p
))
536 smp_send_reschedule(cpu
);
539 void resched_cpu(int cpu
)
541 struct rq
*rq
= cpu_rq(cpu
);
544 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
546 resched_task(cpu_curr(cpu
));
547 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
551 #ifdef CONFIG_NO_HZ_COMMON
553 * In the semi idle case, use the nearest busy cpu for migrating timers
554 * from an idle cpu. This is good for power-savings.
556 * We don't do similar optimization for completely idle system, as
557 * selecting an idle cpu will add more delays to the timers than intended
558 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 int get_nohz_timer_target(void)
562 int cpu
= smp_processor_id();
564 struct sched_domain
*sd
;
567 for_each_domain(cpu
, sd
) {
568 for_each_cpu(i
, sched_domain_span(sd
)) {
580 * When add_timer_on() enqueues a timer into the timer wheel of an
581 * idle CPU then this timer might expire before the next timer event
582 * which is scheduled to wake up that CPU. In case of a completely
583 * idle system the next event might even be infinite time into the
584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
585 * leaves the inner idle loop so the newly added timer is taken into
586 * account when the CPU goes back to idle and evaluates the timer
587 * wheel for the next timer event.
589 static void wake_up_idle_cpu(int cpu
)
591 struct rq
*rq
= cpu_rq(cpu
);
593 if (cpu
== smp_processor_id())
597 * This is safe, as this function is called with the timer
598 * wheel base lock of (cpu) held. When the CPU is on the way
599 * to idle and has not yet set rq->curr to idle then it will
600 * be serialized on the timer wheel base lock and take the new
601 * timer into account automatically.
603 if (rq
->curr
!= rq
->idle
)
607 * We can set TIF_RESCHED on the idle task of the other CPU
608 * lockless. The worst case is that the other CPU runs the
609 * idle task through an additional NOOP schedule()
611 set_tsk_need_resched(rq
->idle
);
613 /* NEED_RESCHED must be visible before we test polling */
615 if (!tsk_is_polling(rq
->idle
))
616 smp_send_reschedule(cpu
);
619 static bool wake_up_full_nohz_cpu(int cpu
)
621 if (tick_nohz_full_cpu(cpu
)) {
622 if (cpu
!= smp_processor_id() ||
623 tick_nohz_tick_stopped())
624 smp_send_reschedule(cpu
);
631 void wake_up_nohz_cpu(int cpu
)
633 if (!wake_up_full_nohz_cpu(cpu
))
634 wake_up_idle_cpu(cpu
);
637 static inline bool got_nohz_idle_kick(void)
639 int cpu
= smp_processor_id();
641 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
644 if (idle_cpu(cpu
) && !need_resched())
648 * We can't run Idle Load Balance on this CPU for this time so we
649 * cancel it and clear NOHZ_BALANCE_KICK
651 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
655 #else /* CONFIG_NO_HZ_COMMON */
657 static inline bool got_nohz_idle_kick(void)
662 #endif /* CONFIG_NO_HZ_COMMON */
664 #ifdef CONFIG_NO_HZ_FULL
665 bool sched_can_stop_tick(void)
671 /* Make sure rq->nr_running update is visible after the IPI */
674 /* More than one running task need preemption */
675 if (rq
->nr_running
> 1)
680 #endif /* CONFIG_NO_HZ_FULL */
682 void sched_avg_update(struct rq
*rq
)
684 s64 period
= sched_avg_period();
686 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
688 * Inline assembly required to prevent the compiler
689 * optimising this loop into a divmod call.
690 * See __iter_div_u64_rem() for another example of this.
692 asm("" : "+rm" (rq
->age_stamp
));
693 rq
->age_stamp
+= period
;
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(rq
, 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(rq
, 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 * Return: 1 if the task is currently executing. 0 otherwise.
935 inline int task_curr(const struct task_struct
*p
)
937 return cpu_curr(task_cpu(p
)) == p
;
940 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
941 const struct sched_class
*prev_class
,
944 if (prev_class
!= p
->sched_class
) {
945 if (prev_class
->switched_from
)
946 prev_class
->switched_from(rq
, p
);
947 p
->sched_class
->switched_to(rq
, p
);
948 } else if (oldprio
!= p
->prio
)
949 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
952 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
954 const struct sched_class
*class;
956 if (p
->sched_class
== rq
->curr
->sched_class
) {
957 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
959 for_each_class(class) {
960 if (class == rq
->curr
->sched_class
)
962 if (class == p
->sched_class
) {
963 resched_task(rq
->curr
);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
974 rq
->skip_clock_update
= 1;
978 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
986 !(task_preempt_count(p
) & PREEMPT_ACTIVE
));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1000 lockdep_is_held(&task_rq(p
)->lock
)));
1004 trace_sched_migrate_task(p
, new_cpu
);
1006 if (task_cpu(p
) != new_cpu
) {
1007 if (p
->sched_class
->migrate_task_rq
)
1008 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1009 p
->se
.nr_migrations
++;
1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1013 __set_task_cpu(p
, new_cpu
);
1016 struct migration_arg
{
1017 struct task_struct
*task
;
1021 static int migration_cpu_stop(void *data
);
1024 * wait_task_inactive - wait for a thread to unschedule.
1026 * If @match_state is nonzero, it's the @p->state value just checked and
1027 * not expected to change. If it changes, i.e. @p might have woken up,
1028 * then return zero. When we succeed in waiting for @p to be off its CPU,
1029 * we return a positive number (its total switch count). If a second call
1030 * a short while later returns the same number, the caller can be sure that
1031 * @p has remained unscheduled the whole time.
1033 * The caller must ensure that the task *will* unschedule sometime soon,
1034 * else this function might spin for a *long* time. This function can't
1035 * be called with interrupts off, or it may introduce deadlock with
1036 * smp_call_function() if an IPI is sent by the same process we are
1037 * waiting to become inactive.
1039 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1041 unsigned long flags
;
1048 * We do the initial early heuristics without holding
1049 * any task-queue locks at all. We'll only try to get
1050 * the runqueue lock when things look like they will
1056 * If the task is actively running on another CPU
1057 * still, just relax and busy-wait without holding
1060 * NOTE! Since we don't hold any locks, it's not
1061 * even sure that "rq" stays as the right runqueue!
1062 * But we don't care, since "task_running()" will
1063 * return false if the runqueue has changed and p
1064 * is actually now running somewhere else!
1066 while (task_running(rq
, p
)) {
1067 if (match_state
&& unlikely(p
->state
!= match_state
))
1073 * Ok, time to look more closely! We need the rq
1074 * lock now, to be *sure*. If we're wrong, we'll
1075 * just go back and repeat.
1077 rq
= task_rq_lock(p
, &flags
);
1078 trace_sched_wait_task(p
);
1079 running
= task_running(rq
, p
);
1082 if (!match_state
|| p
->state
== match_state
)
1083 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1084 task_rq_unlock(rq
, p
, &flags
);
1087 * If it changed from the expected state, bail out now.
1089 if (unlikely(!ncsw
))
1093 * Was it really running after all now that we
1094 * checked with the proper locks actually held?
1096 * Oops. Go back and try again..
1098 if (unlikely(running
)) {
1104 * It's not enough that it's not actively running,
1105 * it must be off the runqueue _entirely_, and not
1108 * So if it was still runnable (but just not actively
1109 * running right now), it's preempted, and we should
1110 * yield - it could be a while.
1112 if (unlikely(on_rq
)) {
1113 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1115 set_current_state(TASK_UNINTERRUPTIBLE
);
1116 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1121 * Ahh, all good. It wasn't running, and it wasn't
1122 * runnable, which means that it will never become
1123 * running in the future either. We're all done!
1132 * kick_process - kick a running thread to enter/exit the kernel
1133 * @p: the to-be-kicked thread
1135 * Cause a process which is running on another CPU to enter
1136 * kernel-mode, without any delay. (to get signals handled.)
1138 * NOTE: this function doesn't have to take the runqueue lock,
1139 * because all it wants to ensure is that the remote task enters
1140 * the kernel. If the IPI races and the task has been migrated
1141 * to another CPU then no harm is done and the purpose has been
1144 void kick_process(struct task_struct
*p
)
1150 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1151 smp_send_reschedule(cpu
);
1154 EXPORT_SYMBOL_GPL(kick_process
);
1155 #endif /* CONFIG_SMP */
1159 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1161 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1163 int nid
= cpu_to_node(cpu
);
1164 const struct cpumask
*nodemask
= NULL
;
1165 enum { cpuset
, possible
, fail
} state
= cpuset
;
1169 * If the node that the cpu is on has been offlined, cpu_to_node()
1170 * will return -1. There is no cpu on the node, and we should
1171 * select the cpu on the other node.
1174 nodemask
= cpumask_of_node(nid
);
1176 /* Look for allowed, online CPU in same node. */
1177 for_each_cpu(dest_cpu
, nodemask
) {
1178 if (!cpu_online(dest_cpu
))
1180 if (!cpu_active(dest_cpu
))
1182 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1188 /* Any allowed, online CPU? */
1189 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1190 if (!cpu_online(dest_cpu
))
1192 if (!cpu_active(dest_cpu
))
1199 /* No more Mr. Nice Guy. */
1200 cpuset_cpus_allowed_fallback(p
);
1205 do_set_cpus_allowed(p
, cpu_possible_mask
);
1216 if (state
!= cpuset
) {
1218 * Don't tell them about moving exiting tasks or
1219 * kernel threads (both mm NULL), since they never
1222 if (p
->mm
&& printk_ratelimit()) {
1223 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1224 task_pid_nr(p
), p
->comm
, cpu
);
1232 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1235 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1237 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1240 * In order not to call set_task_cpu() on a blocking task we need
1241 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1244 * Since this is common to all placement strategies, this lives here.
1246 * [ this allows ->select_task() to simply return task_cpu(p) and
1247 * not worry about this generic constraint ]
1249 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1251 cpu
= select_fallback_rq(task_cpu(p
), p
);
1256 static void update_avg(u64
*avg
, u64 sample
)
1258 s64 diff
= sample
- *avg
;
1264 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1266 #ifdef CONFIG_SCHEDSTATS
1267 struct rq
*rq
= this_rq();
1270 int this_cpu
= smp_processor_id();
1272 if (cpu
== this_cpu
) {
1273 schedstat_inc(rq
, ttwu_local
);
1274 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1276 struct sched_domain
*sd
;
1278 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1280 for_each_domain(this_cpu
, sd
) {
1281 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1282 schedstat_inc(sd
, ttwu_wake_remote
);
1289 if (wake_flags
& WF_MIGRATED
)
1290 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1292 #endif /* CONFIG_SMP */
1294 schedstat_inc(rq
, ttwu_count
);
1295 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1297 if (wake_flags
& WF_SYNC
)
1298 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1300 #endif /* CONFIG_SCHEDSTATS */
1303 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1305 activate_task(rq
, p
, en_flags
);
1308 /* if a worker is waking up, notify workqueue */
1309 if (p
->flags
& PF_WQ_WORKER
)
1310 wq_worker_waking_up(p
, cpu_of(rq
));
1314 * Mark the task runnable and perform wakeup-preemption.
1317 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1319 check_preempt_curr(rq
, p
, wake_flags
);
1320 trace_sched_wakeup(p
, true);
1322 p
->state
= TASK_RUNNING
;
1324 if (p
->sched_class
->task_woken
)
1325 p
->sched_class
->task_woken(rq
, p
);
1327 if (rq
->idle_stamp
) {
1328 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1329 u64 max
= 2*rq
->max_idle_balance_cost
;
1331 update_avg(&rq
->avg_idle
, delta
);
1333 if (rq
->avg_idle
> max
)
1342 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1345 if (p
->sched_contributes_to_load
)
1346 rq
->nr_uninterruptible
--;
1349 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1350 ttwu_do_wakeup(rq
, p
, wake_flags
);
1354 * Called in case the task @p isn't fully descheduled from its runqueue,
1355 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1356 * since all we need to do is flip p->state to TASK_RUNNING, since
1357 * the task is still ->on_rq.
1359 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1364 rq
= __task_rq_lock(p
);
1366 /* check_preempt_curr() may use rq clock */
1367 update_rq_clock(rq
);
1368 ttwu_do_wakeup(rq
, p
, wake_flags
);
1371 __task_rq_unlock(rq
);
1377 static void sched_ttwu_pending(void)
1379 struct rq
*rq
= this_rq();
1380 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1381 struct task_struct
*p
;
1383 raw_spin_lock(&rq
->lock
);
1386 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1387 llist
= llist_next(llist
);
1388 ttwu_do_activate(rq
, p
, 0);
1391 raw_spin_unlock(&rq
->lock
);
1394 void scheduler_ipi(void)
1397 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1398 * TIF_NEED_RESCHED remotely (for the first time) will also send
1401 if (tif_need_resched())
1402 set_preempt_need_resched();
1404 if (llist_empty(&this_rq()->wake_list
)
1405 && !tick_nohz_full_cpu(smp_processor_id())
1406 && !got_nohz_idle_kick())
1410 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1411 * traditionally all their work was done from the interrupt return
1412 * path. Now that we actually do some work, we need to make sure
1415 * Some archs already do call them, luckily irq_enter/exit nest
1418 * Arguably we should visit all archs and update all handlers,
1419 * however a fair share of IPIs are still resched only so this would
1420 * somewhat pessimize the simple resched case.
1423 tick_nohz_full_check();
1424 sched_ttwu_pending();
1427 * Check if someone kicked us for doing the nohz idle load balance.
1429 if (unlikely(got_nohz_idle_kick())) {
1430 this_rq()->idle_balance
= 1;
1431 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1436 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1438 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1439 smp_send_reschedule(cpu
);
1442 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1444 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1446 #endif /* CONFIG_SMP */
1448 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1450 struct rq
*rq
= cpu_rq(cpu
);
1452 #if defined(CONFIG_SMP)
1453 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1454 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1455 ttwu_queue_remote(p
, cpu
);
1460 raw_spin_lock(&rq
->lock
);
1461 ttwu_do_activate(rq
, p
, 0);
1462 raw_spin_unlock(&rq
->lock
);
1466 * try_to_wake_up - wake up a thread
1467 * @p: the thread to be awakened
1468 * @state: the mask of task states that can be woken
1469 * @wake_flags: wake modifier flags (WF_*)
1471 * Put it on the run-queue if it's not already there. The "current"
1472 * thread is always on the run-queue (except when the actual
1473 * re-schedule is in progress), and as such you're allowed to do
1474 * the simpler "current->state = TASK_RUNNING" to mark yourself
1475 * runnable without the overhead of this.
1477 * Return: %true if @p was woken up, %false if it was already running.
1478 * or @state didn't match @p's state.
1481 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1483 unsigned long flags
;
1484 int cpu
, success
= 0;
1487 * If we are going to wake up a thread waiting for CONDITION we
1488 * need to ensure that CONDITION=1 done by the caller can not be
1489 * reordered with p->state check below. This pairs with mb() in
1490 * set_current_state() the waiting thread does.
1492 smp_mb__before_spinlock();
1493 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1494 if (!(p
->state
& state
))
1497 success
= 1; /* we're going to change ->state */
1500 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1505 * If the owning (remote) cpu is still in the middle of schedule() with
1506 * this task as prev, wait until its done referencing the task.
1511 * Pairs with the smp_wmb() in finish_lock_switch().
1515 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1516 p
->state
= TASK_WAKING
;
1518 if (p
->sched_class
->task_waking
)
1519 p
->sched_class
->task_waking(p
);
1521 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1522 if (task_cpu(p
) != cpu
) {
1523 wake_flags
|= WF_MIGRATED
;
1524 set_task_cpu(p
, cpu
);
1526 #endif /* CONFIG_SMP */
1530 ttwu_stat(p
, cpu
, wake_flags
);
1532 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1538 * try_to_wake_up_local - try to wake up a local task with rq lock held
1539 * @p: the thread to be awakened
1541 * Put @p on the run-queue if it's not already there. The caller must
1542 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1545 static void try_to_wake_up_local(struct task_struct
*p
)
1547 struct rq
*rq
= task_rq(p
);
1549 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1550 WARN_ON_ONCE(p
== current
))
1553 lockdep_assert_held(&rq
->lock
);
1555 if (!raw_spin_trylock(&p
->pi_lock
)) {
1556 raw_spin_unlock(&rq
->lock
);
1557 raw_spin_lock(&p
->pi_lock
);
1558 raw_spin_lock(&rq
->lock
);
1561 if (!(p
->state
& TASK_NORMAL
))
1565 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1567 ttwu_do_wakeup(rq
, p
, 0);
1568 ttwu_stat(p
, smp_processor_id(), 0);
1570 raw_spin_unlock(&p
->pi_lock
);
1574 * wake_up_process - Wake up a specific process
1575 * @p: The process to be woken up.
1577 * Attempt to wake up the nominated process and move it to the set of runnable
1580 * Return: 1 if the process was woken up, 0 if it was already running.
1582 * It may be assumed that this function implies a write memory barrier before
1583 * changing the task state if and only if any tasks are woken up.
1585 int wake_up_process(struct task_struct
*p
)
1587 WARN_ON(task_is_stopped_or_traced(p
));
1588 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1590 EXPORT_SYMBOL(wake_up_process
);
1592 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1594 return try_to_wake_up(p
, state
, 0);
1598 * Perform scheduler related setup for a newly forked process p.
1599 * p is forked by current.
1601 * __sched_fork() is basic setup used by init_idle() too:
1603 static void __sched_fork(struct task_struct
*p
)
1608 p
->se
.exec_start
= 0;
1609 p
->se
.sum_exec_runtime
= 0;
1610 p
->se
.prev_sum_exec_runtime
= 0;
1611 p
->se
.nr_migrations
= 0;
1613 INIT_LIST_HEAD(&p
->se
.group_node
);
1615 #ifdef CONFIG_SCHEDSTATS
1616 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1619 INIT_LIST_HEAD(&p
->rt
.run_list
);
1621 #ifdef CONFIG_PREEMPT_NOTIFIERS
1622 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1625 #ifdef CONFIG_NUMA_BALANCING
1626 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1627 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1628 p
->mm
->numa_next_reset
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
1629 p
->mm
->numa_scan_seq
= 0;
1632 p
->node_stamp
= 0ULL;
1633 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1634 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1635 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1636 p
->numa_work
.next
= &p
->numa_work
;
1637 p
->numa_faults
= NULL
;
1638 #endif /* CONFIG_NUMA_BALANCING */
1641 #ifdef CONFIG_NUMA_BALANCING
1642 #ifdef CONFIG_SCHED_DEBUG
1643 void set_numabalancing_state(bool enabled
)
1646 sched_feat_set("NUMA");
1648 sched_feat_set("NO_NUMA");
1651 __read_mostly
bool numabalancing_enabled
;
1653 void set_numabalancing_state(bool enabled
)
1655 numabalancing_enabled
= enabled
;
1657 #endif /* CONFIG_SCHED_DEBUG */
1658 #endif /* CONFIG_NUMA_BALANCING */
1661 * fork()/clone()-time setup:
1663 void sched_fork(struct task_struct
*p
)
1665 unsigned long flags
;
1666 int cpu
= get_cpu();
1670 * We mark the process as running here. This guarantees that
1671 * nobody will actually run it, and a signal or other external
1672 * event cannot wake it up and insert it on the runqueue either.
1674 p
->state
= TASK_RUNNING
;
1677 * Make sure we do not leak PI boosting priority to the child.
1679 p
->prio
= current
->normal_prio
;
1682 * Revert to default priority/policy on fork if requested.
1684 if (unlikely(p
->sched_reset_on_fork
)) {
1685 if (task_has_rt_policy(p
)) {
1686 p
->policy
= SCHED_NORMAL
;
1687 p
->static_prio
= NICE_TO_PRIO(0);
1689 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1690 p
->static_prio
= NICE_TO_PRIO(0);
1692 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1696 * We don't need the reset flag anymore after the fork. It has
1697 * fulfilled its duty:
1699 p
->sched_reset_on_fork
= 0;
1702 if (!rt_prio(p
->prio
))
1703 p
->sched_class
= &fair_sched_class
;
1705 if (p
->sched_class
->task_fork
)
1706 p
->sched_class
->task_fork(p
);
1709 * The child is not yet in the pid-hash so no cgroup attach races,
1710 * and the cgroup is pinned to this child due to cgroup_fork()
1711 * is ran before sched_fork().
1713 * Silence PROVE_RCU.
1715 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1716 set_task_cpu(p
, cpu
);
1717 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1719 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1720 if (likely(sched_info_on()))
1721 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1723 #if defined(CONFIG_SMP)
1726 init_task_preempt_count(p
);
1728 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1735 * wake_up_new_task - wake up a newly created task for the first time.
1737 * This function will do some initial scheduler statistics housekeeping
1738 * that must be done for every newly created context, then puts the task
1739 * on the runqueue and wakes it.
1741 void wake_up_new_task(struct task_struct
*p
)
1743 unsigned long flags
;
1746 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1749 * Fork balancing, do it here and not earlier because:
1750 * - cpus_allowed can change in the fork path
1751 * - any previously selected cpu might disappear through hotplug
1753 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1756 /* Initialize new task's runnable average */
1757 init_task_runnable_average(p
);
1758 rq
= __task_rq_lock(p
);
1759 activate_task(rq
, p
, 0);
1761 trace_sched_wakeup_new(p
, true);
1762 check_preempt_curr(rq
, p
, WF_FORK
);
1764 if (p
->sched_class
->task_woken
)
1765 p
->sched_class
->task_woken(rq
, p
);
1767 task_rq_unlock(rq
, p
, &flags
);
1770 #ifdef CONFIG_PREEMPT_NOTIFIERS
1773 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1774 * @notifier: notifier struct to register
1776 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1778 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1780 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1783 * preempt_notifier_unregister - no longer interested in preemption notifications
1784 * @notifier: notifier struct to unregister
1786 * This is safe to call from within a preemption notifier.
1788 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1790 hlist_del(¬ifier
->link
);
1792 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1794 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1796 struct preempt_notifier
*notifier
;
1798 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1799 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1803 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1804 struct task_struct
*next
)
1806 struct preempt_notifier
*notifier
;
1808 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1809 notifier
->ops
->sched_out(notifier
, next
);
1812 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1814 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1819 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1820 struct task_struct
*next
)
1824 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1827 * prepare_task_switch - prepare to switch tasks
1828 * @rq: the runqueue preparing to switch
1829 * @prev: the current task that is being switched out
1830 * @next: the task we are going to switch to.
1832 * This is called with the rq lock held and interrupts off. It must
1833 * be paired with a subsequent finish_task_switch after the context
1836 * prepare_task_switch sets up locking and calls architecture specific
1840 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1841 struct task_struct
*next
)
1843 trace_sched_switch(prev
, next
);
1844 sched_info_switch(rq
, prev
, next
);
1845 perf_event_task_sched_out(prev
, next
);
1846 fire_sched_out_preempt_notifiers(prev
, next
);
1847 prepare_lock_switch(rq
, next
);
1848 prepare_arch_switch(next
);
1852 * finish_task_switch - clean up after a task-switch
1853 * @rq: runqueue associated with task-switch
1854 * @prev: the thread we just switched away from.
1856 * finish_task_switch must be called after the context switch, paired
1857 * with a prepare_task_switch call before the context switch.
1858 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1859 * and do any other architecture-specific cleanup actions.
1861 * Note that we may have delayed dropping an mm in context_switch(). If
1862 * so, we finish that here outside of the runqueue lock. (Doing it
1863 * with the lock held can cause deadlocks; see schedule() for
1866 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1867 __releases(rq
->lock
)
1869 struct mm_struct
*mm
= rq
->prev_mm
;
1875 * A task struct has one reference for the use as "current".
1876 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1877 * schedule one last time. The schedule call will never return, and
1878 * the scheduled task must drop that reference.
1879 * The test for TASK_DEAD must occur while the runqueue locks are
1880 * still held, otherwise prev could be scheduled on another cpu, die
1881 * there before we look at prev->state, and then the reference would
1883 * Manfred Spraul <manfred@colorfullife.com>
1885 prev_state
= prev
->state
;
1886 vtime_task_switch(prev
);
1887 finish_arch_switch(prev
);
1888 perf_event_task_sched_in(prev
, current
);
1889 finish_lock_switch(rq
, prev
);
1890 finish_arch_post_lock_switch();
1892 fire_sched_in_preempt_notifiers(current
);
1895 if (unlikely(prev_state
== TASK_DEAD
)) {
1896 task_numa_free(prev
);
1899 * Remove function-return probe instances associated with this
1900 * task and put them back on the free list.
1902 kprobe_flush_task(prev
);
1903 put_task_struct(prev
);
1906 tick_nohz_task_switch(current
);
1911 /* assumes rq->lock is held */
1912 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1914 if (prev
->sched_class
->pre_schedule
)
1915 prev
->sched_class
->pre_schedule(rq
, prev
);
1918 /* rq->lock is NOT held, but preemption is disabled */
1919 static inline void post_schedule(struct rq
*rq
)
1921 if (rq
->post_schedule
) {
1922 unsigned long flags
;
1924 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1925 if (rq
->curr
->sched_class
->post_schedule
)
1926 rq
->curr
->sched_class
->post_schedule(rq
);
1927 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1929 rq
->post_schedule
= 0;
1935 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1939 static inline void post_schedule(struct rq
*rq
)
1946 * schedule_tail - first thing a freshly forked thread must call.
1947 * @prev: the thread we just switched away from.
1949 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1950 __releases(rq
->lock
)
1952 struct rq
*rq
= this_rq();
1954 finish_task_switch(rq
, prev
);
1957 * FIXME: do we need to worry about rq being invalidated by the
1962 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1963 /* In this case, finish_task_switch does not reenable preemption */
1966 if (current
->set_child_tid
)
1967 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1971 * context_switch - switch to the new MM and the new
1972 * thread's register state.
1975 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1976 struct task_struct
*next
)
1978 struct mm_struct
*mm
, *oldmm
;
1980 prepare_task_switch(rq
, prev
, next
);
1983 oldmm
= prev
->active_mm
;
1985 * For paravirt, this is coupled with an exit in switch_to to
1986 * combine the page table reload and the switch backend into
1989 arch_start_context_switch(prev
);
1992 next
->active_mm
= oldmm
;
1993 atomic_inc(&oldmm
->mm_count
);
1994 enter_lazy_tlb(oldmm
, next
);
1996 switch_mm(oldmm
, mm
, next
);
1999 prev
->active_mm
= NULL
;
2000 rq
->prev_mm
= oldmm
;
2003 * Since the runqueue lock will be released by the next
2004 * task (which is an invalid locking op but in the case
2005 * of the scheduler it's an obvious special-case), so we
2006 * do an early lockdep release here:
2008 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2009 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2012 context_tracking_task_switch(prev
, next
);
2013 /* Here we just switch the register state and the stack. */
2014 switch_to(prev
, next
, prev
);
2018 * this_rq must be evaluated again because prev may have moved
2019 * CPUs since it called schedule(), thus the 'rq' on its stack
2020 * frame will be invalid.
2022 finish_task_switch(this_rq(), prev
);
2026 * nr_running and nr_context_switches:
2028 * externally visible scheduler statistics: current number of runnable
2029 * threads, total number of context switches performed since bootup.
2031 unsigned long nr_running(void)
2033 unsigned long i
, sum
= 0;
2035 for_each_online_cpu(i
)
2036 sum
+= cpu_rq(i
)->nr_running
;
2041 unsigned long long nr_context_switches(void)
2044 unsigned long long sum
= 0;
2046 for_each_possible_cpu(i
)
2047 sum
+= cpu_rq(i
)->nr_switches
;
2052 unsigned long nr_iowait(void)
2054 unsigned long i
, sum
= 0;
2056 for_each_possible_cpu(i
)
2057 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2062 unsigned long nr_iowait_cpu(int cpu
)
2064 struct rq
*this = cpu_rq(cpu
);
2065 return atomic_read(&this->nr_iowait
);
2071 * sched_exec - execve() is a valuable balancing opportunity, because at
2072 * this point the task has the smallest effective memory and cache footprint.
2074 void sched_exec(void)
2076 struct task_struct
*p
= current
;
2077 unsigned long flags
;
2080 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2081 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2082 if (dest_cpu
== smp_processor_id())
2085 if (likely(cpu_active(dest_cpu
))) {
2086 struct migration_arg arg
= { p
, dest_cpu
};
2088 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2089 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2093 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2098 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2099 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2101 EXPORT_PER_CPU_SYMBOL(kstat
);
2102 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2105 * Return any ns on the sched_clock that have not yet been accounted in
2106 * @p in case that task is currently running.
2108 * Called with task_rq_lock() held on @rq.
2110 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2114 if (task_current(rq
, p
)) {
2115 update_rq_clock(rq
);
2116 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2124 unsigned long long task_delta_exec(struct task_struct
*p
)
2126 unsigned long flags
;
2130 rq
= task_rq_lock(p
, &flags
);
2131 ns
= do_task_delta_exec(p
, rq
);
2132 task_rq_unlock(rq
, p
, &flags
);
2138 * Return accounted runtime for the task.
2139 * In case the task is currently running, return the runtime plus current's
2140 * pending runtime that have not been accounted yet.
2142 unsigned long long task_sched_runtime(struct task_struct
*p
)
2144 unsigned long flags
;
2148 rq
= task_rq_lock(p
, &flags
);
2149 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2150 task_rq_unlock(rq
, p
, &flags
);
2156 * This function gets called by the timer code, with HZ frequency.
2157 * We call it with interrupts disabled.
2159 void scheduler_tick(void)
2161 int cpu
= smp_processor_id();
2162 struct rq
*rq
= cpu_rq(cpu
);
2163 struct task_struct
*curr
= rq
->curr
;
2167 raw_spin_lock(&rq
->lock
);
2168 update_rq_clock(rq
);
2169 curr
->sched_class
->task_tick(rq
, curr
, 0);
2170 update_cpu_load_active(rq
);
2171 raw_spin_unlock(&rq
->lock
);
2173 perf_event_task_tick();
2176 rq
->idle_balance
= idle_cpu(cpu
);
2177 trigger_load_balance(rq
, cpu
);
2179 rq_last_tick_reset(rq
);
2182 #ifdef CONFIG_NO_HZ_FULL
2184 * scheduler_tick_max_deferment
2186 * Keep at least one tick per second when a single
2187 * active task is running because the scheduler doesn't
2188 * yet completely support full dynticks environment.
2190 * This makes sure that uptime, CFS vruntime, load
2191 * balancing, etc... continue to move forward, even
2192 * with a very low granularity.
2194 * Return: Maximum deferment in nanoseconds.
2196 u64
scheduler_tick_max_deferment(void)
2198 struct rq
*rq
= this_rq();
2199 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2201 next
= rq
->last_sched_tick
+ HZ
;
2203 if (time_before_eq(next
, now
))
2206 return jiffies_to_usecs(next
- now
) * NSEC_PER_USEC
;
2210 notrace
unsigned long get_parent_ip(unsigned long addr
)
2212 if (in_lock_functions(addr
)) {
2213 addr
= CALLER_ADDR2
;
2214 if (in_lock_functions(addr
))
2215 addr
= CALLER_ADDR3
;
2220 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2221 defined(CONFIG_PREEMPT_TRACER))
2223 void __kprobes
preempt_count_add(int val
)
2225 #ifdef CONFIG_DEBUG_PREEMPT
2229 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2232 __preempt_count_add(val
);
2233 #ifdef CONFIG_DEBUG_PREEMPT
2235 * Spinlock count overflowing soon?
2237 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2240 if (preempt_count() == val
)
2241 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2243 EXPORT_SYMBOL(preempt_count_add
);
2245 void __kprobes
preempt_count_sub(int val
)
2247 #ifdef CONFIG_DEBUG_PREEMPT
2251 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2254 * Is the spinlock portion underflowing?
2256 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2257 !(preempt_count() & PREEMPT_MASK
)))
2261 if (preempt_count() == val
)
2262 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2263 __preempt_count_sub(val
);
2265 EXPORT_SYMBOL(preempt_count_sub
);
2270 * Print scheduling while atomic bug:
2272 static noinline
void __schedule_bug(struct task_struct
*prev
)
2274 if (oops_in_progress
)
2277 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2278 prev
->comm
, prev
->pid
, preempt_count());
2280 debug_show_held_locks(prev
);
2282 if (irqs_disabled())
2283 print_irqtrace_events(prev
);
2285 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2289 * Various schedule()-time debugging checks and statistics:
2291 static inline void schedule_debug(struct task_struct
*prev
)
2294 * Test if we are atomic. Since do_exit() needs to call into
2295 * schedule() atomically, we ignore that path for now.
2296 * Otherwise, whine if we are scheduling when we should not be.
2298 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2299 __schedule_bug(prev
);
2302 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2304 schedstat_inc(this_rq(), sched_count
);
2307 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2309 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2310 update_rq_clock(rq
);
2311 prev
->sched_class
->put_prev_task(rq
, prev
);
2315 * Pick up the highest-prio task:
2317 static inline struct task_struct
*
2318 pick_next_task(struct rq
*rq
)
2320 const struct sched_class
*class;
2321 struct task_struct
*p
;
2324 * Optimization: we know that if all tasks are in
2325 * the fair class we can call that function directly:
2327 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2328 p
= fair_sched_class
.pick_next_task(rq
);
2333 for_each_class(class) {
2334 p
= class->pick_next_task(rq
);
2339 BUG(); /* the idle class will always have a runnable task */
2343 * __schedule() is the main scheduler function.
2345 * The main means of driving the scheduler and thus entering this function are:
2347 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2349 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2350 * paths. For example, see arch/x86/entry_64.S.
2352 * To drive preemption between tasks, the scheduler sets the flag in timer
2353 * interrupt handler scheduler_tick().
2355 * 3. Wakeups don't really cause entry into schedule(). They add a
2356 * task to the run-queue and that's it.
2358 * Now, if the new task added to the run-queue preempts the current
2359 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2360 * called on the nearest possible occasion:
2362 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2364 * - in syscall or exception context, at the next outmost
2365 * preempt_enable(). (this might be as soon as the wake_up()'s
2368 * - in IRQ context, return from interrupt-handler to
2369 * preemptible context
2371 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2374 * - cond_resched() call
2375 * - explicit schedule() call
2376 * - return from syscall or exception to user-space
2377 * - return from interrupt-handler to user-space
2379 static void __sched
__schedule(void)
2381 struct task_struct
*prev
, *next
;
2382 unsigned long *switch_count
;
2388 cpu
= smp_processor_id();
2390 rcu_note_context_switch(cpu
);
2393 schedule_debug(prev
);
2395 if (sched_feat(HRTICK
))
2399 * Make sure that signal_pending_state()->signal_pending() below
2400 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2401 * done by the caller to avoid the race with signal_wake_up().
2403 smp_mb__before_spinlock();
2404 raw_spin_lock_irq(&rq
->lock
);
2406 switch_count
= &prev
->nivcsw
;
2407 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2408 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2409 prev
->state
= TASK_RUNNING
;
2411 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2415 * If a worker went to sleep, notify and ask workqueue
2416 * whether it wants to wake up a task to maintain
2419 if (prev
->flags
& PF_WQ_WORKER
) {
2420 struct task_struct
*to_wakeup
;
2422 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2424 try_to_wake_up_local(to_wakeup
);
2427 switch_count
= &prev
->nvcsw
;
2430 pre_schedule(rq
, prev
);
2432 if (unlikely(!rq
->nr_running
))
2433 idle_balance(cpu
, rq
);
2435 put_prev_task(rq
, prev
);
2436 next
= pick_next_task(rq
);
2437 clear_tsk_need_resched(prev
);
2438 clear_preempt_need_resched();
2439 rq
->skip_clock_update
= 0;
2441 if (likely(prev
!= next
)) {
2446 context_switch(rq
, prev
, next
); /* unlocks the rq */
2448 * The context switch have flipped the stack from under us
2449 * and restored the local variables which were saved when
2450 * this task called schedule() in the past. prev == current
2451 * is still correct, but it can be moved to another cpu/rq.
2453 cpu
= smp_processor_id();
2456 raw_spin_unlock_irq(&rq
->lock
);
2460 sched_preempt_enable_no_resched();
2465 static inline void sched_submit_work(struct task_struct
*tsk
)
2467 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2470 * If we are going to sleep and we have plugged IO queued,
2471 * make sure to submit it to avoid deadlocks.
2473 if (blk_needs_flush_plug(tsk
))
2474 blk_schedule_flush_plug(tsk
);
2477 asmlinkage
void __sched
schedule(void)
2479 struct task_struct
*tsk
= current
;
2481 sched_submit_work(tsk
);
2484 EXPORT_SYMBOL(schedule
);
2486 #ifdef CONFIG_CONTEXT_TRACKING
2487 asmlinkage
void __sched
schedule_user(void)
2490 * If we come here after a random call to set_need_resched(),
2491 * or we have been woken up remotely but the IPI has not yet arrived,
2492 * we haven't yet exited the RCU idle mode. Do it here manually until
2493 * we find a better solution.
2502 * schedule_preempt_disabled - called with preemption disabled
2504 * Returns with preemption disabled. Note: preempt_count must be 1
2506 void __sched
schedule_preempt_disabled(void)
2508 sched_preempt_enable_no_resched();
2513 #ifdef CONFIG_PREEMPT
2515 * this is the entry point to schedule() from in-kernel preemption
2516 * off of preempt_enable. Kernel preemptions off return from interrupt
2517 * occur there and call schedule directly.
2519 asmlinkage
void __sched notrace
preempt_schedule(void)
2522 * If there is a non-zero preempt_count or interrupts are disabled,
2523 * we do not want to preempt the current task. Just return..
2525 if (likely(!preemptible()))
2529 __preempt_count_add(PREEMPT_ACTIVE
);
2531 __preempt_count_sub(PREEMPT_ACTIVE
);
2534 * Check again in case we missed a preemption opportunity
2535 * between schedule and now.
2538 } while (need_resched());
2540 EXPORT_SYMBOL(preempt_schedule
);
2543 * this is the entry point to schedule() from kernel preemption
2544 * off of irq context.
2545 * Note, that this is called and return with irqs disabled. This will
2546 * protect us against recursive calling from irq.
2548 asmlinkage
void __sched
preempt_schedule_irq(void)
2550 enum ctx_state prev_state
;
2552 /* Catch callers which need to be fixed */
2553 BUG_ON(preempt_count() || !irqs_disabled());
2555 prev_state
= exception_enter();
2558 __preempt_count_add(PREEMPT_ACTIVE
);
2561 local_irq_disable();
2562 __preempt_count_sub(PREEMPT_ACTIVE
);
2565 * Check again in case we missed a preemption opportunity
2566 * between schedule and now.
2569 } while (need_resched());
2571 exception_exit(prev_state
);
2574 #endif /* CONFIG_PREEMPT */
2576 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2579 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2581 EXPORT_SYMBOL(default_wake_function
);
2584 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2585 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2586 * number) then we wake all the non-exclusive tasks and one exclusive task.
2588 * There are circumstances in which we can try to wake a task which has already
2589 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2590 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2592 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
2593 int nr_exclusive
, int wake_flags
, void *key
)
2595 wait_queue_t
*curr
, *next
;
2597 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
2598 unsigned flags
= curr
->flags
;
2600 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
2601 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
2607 * __wake_up - wake up threads blocked on a waitqueue.
2609 * @mode: which threads
2610 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2611 * @key: is directly passed to the wakeup function
2613 * It may be assumed that this function implies a write memory barrier before
2614 * changing the task state if and only if any tasks are woken up.
2616 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
2617 int nr_exclusive
, void *key
)
2619 unsigned long flags
;
2621 spin_lock_irqsave(&q
->lock
, flags
);
2622 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
2623 spin_unlock_irqrestore(&q
->lock
, flags
);
2625 EXPORT_SYMBOL(__wake_up
);
2628 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2630 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
2632 __wake_up_common(q
, mode
, nr
, 0, NULL
);
2634 EXPORT_SYMBOL_GPL(__wake_up_locked
);
2636 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
2638 __wake_up_common(q
, mode
, 1, 0, key
);
2640 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
2643 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2645 * @mode: which threads
2646 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2647 * @key: opaque value to be passed to wakeup targets
2649 * The sync wakeup differs that the waker knows that it will schedule
2650 * away soon, so while the target thread will be woken up, it will not
2651 * be migrated to another CPU - ie. the two threads are 'synchronized'
2652 * with each other. This can prevent needless bouncing between CPUs.
2654 * On UP it can prevent extra preemption.
2656 * It may be assumed that this function implies a write memory barrier before
2657 * changing the task state if and only if any tasks are woken up.
2659 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
2660 int nr_exclusive
, void *key
)
2662 unsigned long flags
;
2663 int wake_flags
= WF_SYNC
;
2668 if (unlikely(nr_exclusive
!= 1))
2671 spin_lock_irqsave(&q
->lock
, flags
);
2672 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
2673 spin_unlock_irqrestore(&q
->lock
, flags
);
2675 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
2678 * __wake_up_sync - see __wake_up_sync_key()
2680 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
2682 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
2684 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
2687 * complete: - signals a single thread waiting on this completion
2688 * @x: holds the state of this particular completion
2690 * This will wake up a single thread waiting on this completion. Threads will be
2691 * awakened in the same order in which they were queued.
2693 * See also complete_all(), wait_for_completion() and related routines.
2695 * It may be assumed that this function implies a write memory barrier before
2696 * changing the task state if and only if any tasks are woken up.
2698 void complete(struct completion
*x
)
2700 unsigned long flags
;
2702 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2704 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
2705 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2707 EXPORT_SYMBOL(complete
);
2710 * complete_all: - signals all threads waiting on this completion
2711 * @x: holds the state of this particular completion
2713 * This will wake up all threads waiting on this particular completion event.
2715 * It may be assumed that this function implies a write memory barrier before
2716 * changing the task state if and only if any tasks are woken up.
2718 void complete_all(struct completion
*x
)
2720 unsigned long flags
;
2722 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2723 x
->done
+= UINT_MAX
/2;
2724 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
2725 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2727 EXPORT_SYMBOL(complete_all
);
2729 static inline long __sched
2730 do_wait_for_common(struct completion
*x
,
2731 long (*action
)(long), long timeout
, int state
)
2734 DECLARE_WAITQUEUE(wait
, current
);
2736 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
2738 if (signal_pending_state(state
, current
)) {
2739 timeout
= -ERESTARTSYS
;
2742 __set_current_state(state
);
2743 spin_unlock_irq(&x
->wait
.lock
);
2744 timeout
= action(timeout
);
2745 spin_lock_irq(&x
->wait
.lock
);
2746 } while (!x
->done
&& timeout
);
2747 __remove_wait_queue(&x
->wait
, &wait
);
2752 return timeout
?: 1;
2755 static inline long __sched
2756 __wait_for_common(struct completion
*x
,
2757 long (*action
)(long), long timeout
, int state
)
2761 spin_lock_irq(&x
->wait
.lock
);
2762 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
2763 spin_unlock_irq(&x
->wait
.lock
);
2768 wait_for_common(struct completion
*x
, long timeout
, int state
)
2770 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
2774 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
2776 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
2780 * wait_for_completion: - waits for completion of a task
2781 * @x: holds the state of this particular completion
2783 * This waits to be signaled for completion of a specific task. It is NOT
2784 * interruptible and there is no timeout.
2786 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2787 * and interrupt capability. Also see complete().
2789 void __sched
wait_for_completion(struct completion
*x
)
2791 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
2793 EXPORT_SYMBOL(wait_for_completion
);
2796 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2797 * @x: holds the state of this particular completion
2798 * @timeout: timeout value in jiffies
2800 * This waits for either a completion of a specific task to be signaled or for a
2801 * specified timeout to expire. The timeout is in jiffies. It is not
2804 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2805 * till timeout) if completed.
2807 unsigned long __sched
2808 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
2810 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
2812 EXPORT_SYMBOL(wait_for_completion_timeout
);
2815 * wait_for_completion_io: - waits for completion of a task
2816 * @x: holds the state of this particular completion
2818 * This waits to be signaled for completion of a specific task. It is NOT
2819 * interruptible and there is no timeout. The caller is accounted as waiting
2822 void __sched
wait_for_completion_io(struct completion
*x
)
2824 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
2826 EXPORT_SYMBOL(wait_for_completion_io
);
2829 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2830 * @x: holds the state of this particular completion
2831 * @timeout: timeout value in jiffies
2833 * This waits for either a completion of a specific task to be signaled or for a
2834 * specified timeout to expire. The timeout is in jiffies. It is not
2835 * interruptible. The caller is accounted as waiting for IO.
2837 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2838 * till timeout) if completed.
2840 unsigned long __sched
2841 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
2843 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
2845 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
2848 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2849 * @x: holds the state of this particular completion
2851 * This waits for completion of a specific task to be signaled. It is
2854 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2856 int __sched
wait_for_completion_interruptible(struct completion
*x
)
2858 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
2859 if (t
== -ERESTARTSYS
)
2863 EXPORT_SYMBOL(wait_for_completion_interruptible
);
2866 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2867 * @x: holds the state of this particular completion
2868 * @timeout: timeout value in jiffies
2870 * This waits for either a completion of a specific task to be signaled or for a
2871 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2873 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2874 * or number of jiffies left till timeout) if completed.
2877 wait_for_completion_interruptible_timeout(struct completion
*x
,
2878 unsigned long timeout
)
2880 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
2882 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
2885 * wait_for_completion_killable: - waits for completion of a task (killable)
2886 * @x: holds the state of this particular completion
2888 * This waits to be signaled for completion of a specific task. It can be
2889 * interrupted by a kill signal.
2891 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2893 int __sched
wait_for_completion_killable(struct completion
*x
)
2895 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
2896 if (t
== -ERESTARTSYS
)
2900 EXPORT_SYMBOL(wait_for_completion_killable
);
2903 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2904 * @x: holds the state of this particular completion
2905 * @timeout: timeout value in jiffies
2907 * This waits for either a completion of a specific task to be
2908 * signaled or for a specified timeout to expire. It can be
2909 * interrupted by a kill signal. The timeout is in jiffies.
2911 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2912 * or number of jiffies left till timeout) if completed.
2915 wait_for_completion_killable_timeout(struct completion
*x
,
2916 unsigned long timeout
)
2918 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
2920 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
2923 * try_wait_for_completion - try to decrement a completion without blocking
2924 * @x: completion structure
2926 * Return: 0 if a decrement cannot be done without blocking
2927 * 1 if a decrement succeeded.
2929 * If a completion is being used as a counting completion,
2930 * attempt to decrement the counter without blocking. This
2931 * enables us to avoid waiting if the resource the completion
2932 * is protecting is not available.
2934 bool try_wait_for_completion(struct completion
*x
)
2936 unsigned long flags
;
2939 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2944 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2947 EXPORT_SYMBOL(try_wait_for_completion
);
2950 * completion_done - Test to see if a completion has any waiters
2951 * @x: completion structure
2953 * Return: 0 if there are waiters (wait_for_completion() in progress)
2954 * 1 if there are no waiters.
2957 bool completion_done(struct completion
*x
)
2959 unsigned long flags
;
2962 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2965 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2968 EXPORT_SYMBOL(completion_done
);
2971 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
2973 unsigned long flags
;
2976 init_waitqueue_entry(&wait
, current
);
2978 __set_current_state(state
);
2980 spin_lock_irqsave(&q
->lock
, flags
);
2981 __add_wait_queue(q
, &wait
);
2982 spin_unlock(&q
->lock
);
2983 timeout
= schedule_timeout(timeout
);
2984 spin_lock_irq(&q
->lock
);
2985 __remove_wait_queue(q
, &wait
);
2986 spin_unlock_irqrestore(&q
->lock
, flags
);
2991 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
2993 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2995 EXPORT_SYMBOL(interruptible_sleep_on
);
2998 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3000 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3002 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3004 void __sched
sleep_on(wait_queue_head_t
*q
)
3006 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3008 EXPORT_SYMBOL(sleep_on
);
3010 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3012 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3014 EXPORT_SYMBOL(sleep_on_timeout
);
3016 #ifdef CONFIG_RT_MUTEXES
3019 * rt_mutex_setprio - set the current priority of a task
3021 * @prio: prio value (kernel-internal form)
3023 * This function changes the 'effective' priority of a task. It does
3024 * not touch ->normal_prio like __setscheduler().
3026 * Used by the rt_mutex code to implement priority inheritance logic.
3028 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3030 int oldprio
, on_rq
, running
;
3032 const struct sched_class
*prev_class
;
3034 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3036 rq
= __task_rq_lock(p
);
3039 * Idle task boosting is a nono in general. There is one
3040 * exception, when PREEMPT_RT and NOHZ is active:
3042 * The idle task calls get_next_timer_interrupt() and holds
3043 * the timer wheel base->lock on the CPU and another CPU wants
3044 * to access the timer (probably to cancel it). We can safely
3045 * ignore the boosting request, as the idle CPU runs this code
3046 * with interrupts disabled and will complete the lock
3047 * protected section without being interrupted. So there is no
3048 * real need to boost.
3050 if (unlikely(p
== rq
->idle
)) {
3051 WARN_ON(p
!= rq
->curr
);
3052 WARN_ON(p
->pi_blocked_on
);
3056 trace_sched_pi_setprio(p
, prio
);
3058 prev_class
= p
->sched_class
;
3060 running
= task_current(rq
, p
);
3062 dequeue_task(rq
, p
, 0);
3064 p
->sched_class
->put_prev_task(rq
, p
);
3067 p
->sched_class
= &rt_sched_class
;
3069 p
->sched_class
= &fair_sched_class
;
3074 p
->sched_class
->set_curr_task(rq
);
3076 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3078 check_class_changed(rq
, p
, prev_class
, oldprio
);
3080 __task_rq_unlock(rq
);
3083 void set_user_nice(struct task_struct
*p
, long nice
)
3085 int old_prio
, delta
, on_rq
;
3086 unsigned long flags
;
3089 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3092 * We have to be careful, if called from sys_setpriority(),
3093 * the task might be in the middle of scheduling on another CPU.
3095 rq
= task_rq_lock(p
, &flags
);
3097 * The RT priorities are set via sched_setscheduler(), but we still
3098 * allow the 'normal' nice value to be set - but as expected
3099 * it wont have any effect on scheduling until the task is
3100 * SCHED_FIFO/SCHED_RR:
3102 if (task_has_rt_policy(p
)) {
3103 p
->static_prio
= NICE_TO_PRIO(nice
);
3108 dequeue_task(rq
, p
, 0);
3110 p
->static_prio
= NICE_TO_PRIO(nice
);
3113 p
->prio
= effective_prio(p
);
3114 delta
= p
->prio
- old_prio
;
3117 enqueue_task(rq
, p
, 0);
3119 * If the task increased its priority or is running and
3120 * lowered its priority, then reschedule its CPU:
3122 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3123 resched_task(rq
->curr
);
3126 task_rq_unlock(rq
, p
, &flags
);
3128 EXPORT_SYMBOL(set_user_nice
);
3131 * can_nice - check if a task can reduce its nice value
3135 int can_nice(const struct task_struct
*p
, const int nice
)
3137 /* convert nice value [19,-20] to rlimit style value [1,40] */
3138 int nice_rlim
= 20 - nice
;
3140 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3141 capable(CAP_SYS_NICE
));
3144 #ifdef __ARCH_WANT_SYS_NICE
3147 * sys_nice - change the priority of the current process.
3148 * @increment: priority increment
3150 * sys_setpriority is a more generic, but much slower function that
3151 * does similar things.
3153 SYSCALL_DEFINE1(nice
, int, increment
)
3158 * Setpriority might change our priority at the same moment.
3159 * We don't have to worry. Conceptually one call occurs first
3160 * and we have a single winner.
3162 if (increment
< -40)
3167 nice
= TASK_NICE(current
) + increment
;
3173 if (increment
< 0 && !can_nice(current
, nice
))
3176 retval
= security_task_setnice(current
, nice
);
3180 set_user_nice(current
, nice
);
3187 * task_prio - return the priority value of a given task.
3188 * @p: the task in question.
3190 * Return: The priority value as seen by users in /proc.
3191 * RT tasks are offset by -200. Normal tasks are centered
3192 * around 0, value goes from -16 to +15.
3194 int task_prio(const struct task_struct
*p
)
3196 return p
->prio
- MAX_RT_PRIO
;
3200 * task_nice - return the nice value of a given task.
3201 * @p: the task in question.
3203 * Return: The nice value [ -20 ... 0 ... 19 ].
3205 int task_nice(const struct task_struct
*p
)
3207 return TASK_NICE(p
);
3209 EXPORT_SYMBOL(task_nice
);
3212 * idle_cpu - is a given cpu idle currently?
3213 * @cpu: the processor in question.
3215 * Return: 1 if the CPU is currently idle. 0 otherwise.
3217 int idle_cpu(int cpu
)
3219 struct rq
*rq
= cpu_rq(cpu
);
3221 if (rq
->curr
!= rq
->idle
)
3228 if (!llist_empty(&rq
->wake_list
))
3236 * idle_task - return the idle task for a given cpu.
3237 * @cpu: the processor in question.
3239 * Return: The idle task for the cpu @cpu.
3241 struct task_struct
*idle_task(int cpu
)
3243 return cpu_rq(cpu
)->idle
;
3247 * find_process_by_pid - find a process with a matching PID value.
3248 * @pid: the pid in question.
3250 * The task of @pid, if found. %NULL otherwise.
3252 static struct task_struct
*find_process_by_pid(pid_t pid
)
3254 return pid
? find_task_by_vpid(pid
) : current
;
3257 /* Actually do priority change: must hold rq lock. */
3259 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3262 p
->rt_priority
= prio
;
3263 p
->normal_prio
= normal_prio(p
);
3264 /* we are holding p->pi_lock already */
3265 p
->prio
= rt_mutex_getprio(p
);
3266 if (rt_prio(p
->prio
))
3267 p
->sched_class
= &rt_sched_class
;
3269 p
->sched_class
= &fair_sched_class
;
3274 * check the target process has a UID that matches the current process's
3276 static bool check_same_owner(struct task_struct
*p
)
3278 const struct cred
*cred
= current_cred(), *pcred
;
3282 pcred
= __task_cred(p
);
3283 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3284 uid_eq(cred
->euid
, pcred
->uid
));
3289 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3290 const struct sched_param
*param
, bool user
)
3292 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3293 unsigned long flags
;
3294 const struct sched_class
*prev_class
;
3298 /* may grab non-irq protected spin_locks */
3299 BUG_ON(in_interrupt());
3301 /* double check policy once rq lock held */
3303 reset_on_fork
= p
->sched_reset_on_fork
;
3304 policy
= oldpolicy
= p
->policy
;
3306 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3307 policy
&= ~SCHED_RESET_ON_FORK
;
3309 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3310 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3311 policy
!= SCHED_IDLE
)
3316 * Valid priorities for SCHED_FIFO and SCHED_RR are
3317 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3318 * SCHED_BATCH and SCHED_IDLE is 0.
3320 if (param
->sched_priority
< 0 ||
3321 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3322 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3324 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3328 * Allow unprivileged RT tasks to decrease priority:
3330 if (user
&& !capable(CAP_SYS_NICE
)) {
3331 if (rt_policy(policy
)) {
3332 unsigned long rlim_rtprio
=
3333 task_rlimit(p
, RLIMIT_RTPRIO
);
3335 /* can't set/change the rt policy */
3336 if (policy
!= p
->policy
&& !rlim_rtprio
)
3339 /* can't increase priority */
3340 if (param
->sched_priority
> p
->rt_priority
&&
3341 param
->sched_priority
> rlim_rtprio
)
3346 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3347 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3349 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3350 if (!can_nice(p
, TASK_NICE(p
)))
3354 /* can't change other user's priorities */
3355 if (!check_same_owner(p
))
3358 /* Normal users shall not reset the sched_reset_on_fork flag */
3359 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3364 retval
= security_task_setscheduler(p
);
3370 * make sure no PI-waiters arrive (or leave) while we are
3371 * changing the priority of the task:
3373 * To be able to change p->policy safely, the appropriate
3374 * runqueue lock must be held.
3376 rq
= task_rq_lock(p
, &flags
);
3379 * Changing the policy of the stop threads its a very bad idea
3381 if (p
== rq
->stop
) {
3382 task_rq_unlock(rq
, p
, &flags
);
3387 * If not changing anything there's no need to proceed further:
3389 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3390 param
->sched_priority
== p
->rt_priority
))) {
3391 task_rq_unlock(rq
, p
, &flags
);
3395 #ifdef CONFIG_RT_GROUP_SCHED
3398 * Do not allow realtime tasks into groups that have no runtime
3401 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3402 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3403 !task_group_is_autogroup(task_group(p
))) {
3404 task_rq_unlock(rq
, p
, &flags
);
3410 /* recheck policy now with rq lock held */
3411 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3412 policy
= oldpolicy
= -1;
3413 task_rq_unlock(rq
, p
, &flags
);
3417 running
= task_current(rq
, p
);
3419 dequeue_task(rq
, p
, 0);
3421 p
->sched_class
->put_prev_task(rq
, p
);
3423 p
->sched_reset_on_fork
= reset_on_fork
;
3426 prev_class
= p
->sched_class
;
3427 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3430 p
->sched_class
->set_curr_task(rq
);
3432 enqueue_task(rq
, p
, 0);
3434 check_class_changed(rq
, p
, prev_class
, oldprio
);
3435 task_rq_unlock(rq
, p
, &flags
);
3437 rt_mutex_adjust_pi(p
);
3443 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3444 * @p: the task in question.
3445 * @policy: new policy.
3446 * @param: structure containing the new RT priority.
3448 * Return: 0 on success. An error code otherwise.
3450 * NOTE that the task may be already dead.
3452 int sched_setscheduler(struct task_struct
*p
, int policy
,
3453 const struct sched_param
*param
)
3455 return __sched_setscheduler(p
, policy
, param
, true);
3457 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3460 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3461 * @p: the task in question.
3462 * @policy: new policy.
3463 * @param: structure containing the new RT priority.
3465 * Just like sched_setscheduler, only don't bother checking if the
3466 * current context has permission. For example, this is needed in
3467 * stop_machine(): we create temporary high priority worker threads,
3468 * but our caller might not have that capability.
3470 * Return: 0 on success. An error code otherwise.
3472 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3473 const struct sched_param
*param
)
3475 return __sched_setscheduler(p
, policy
, param
, false);
3479 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3481 struct sched_param lparam
;
3482 struct task_struct
*p
;
3485 if (!param
|| pid
< 0)
3487 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3492 p
= find_process_by_pid(pid
);
3494 retval
= sched_setscheduler(p
, policy
, &lparam
);
3501 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3502 * @pid: the pid in question.
3503 * @policy: new policy.
3504 * @param: structure containing the new RT priority.
3506 * Return: 0 on success. An error code otherwise.
3508 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3509 struct sched_param __user
*, param
)
3511 /* negative values for policy are not valid */
3515 return do_sched_setscheduler(pid
, policy
, param
);
3519 * sys_sched_setparam - set/change the RT priority of a thread
3520 * @pid: the pid in question.
3521 * @param: structure containing the new RT priority.
3523 * Return: 0 on success. An error code otherwise.
3525 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3527 return do_sched_setscheduler(pid
, -1, param
);
3531 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3532 * @pid: the pid in question.
3534 * Return: On success, the policy of the thread. Otherwise, a negative error
3537 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3539 struct task_struct
*p
;
3547 p
= find_process_by_pid(pid
);
3549 retval
= security_task_getscheduler(p
);
3552 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3559 * sys_sched_getparam - get the RT priority of a thread
3560 * @pid: the pid in question.
3561 * @param: structure containing the RT priority.
3563 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3566 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3568 struct sched_param lp
;
3569 struct task_struct
*p
;
3572 if (!param
|| pid
< 0)
3576 p
= find_process_by_pid(pid
);
3581 retval
= security_task_getscheduler(p
);
3585 lp
.sched_priority
= p
->rt_priority
;
3589 * This one might sleep, we cannot do it with a spinlock held ...
3591 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3600 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3602 cpumask_var_t cpus_allowed
, new_mask
;
3603 struct task_struct
*p
;
3609 p
= find_process_by_pid(pid
);
3616 /* Prevent p going away */
3620 if (p
->flags
& PF_NO_SETAFFINITY
) {
3624 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3628 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3630 goto out_free_cpus_allowed
;
3633 if (!check_same_owner(p
)) {
3635 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3642 retval
= security_task_setscheduler(p
);
3646 cpuset_cpus_allowed(p
, cpus_allowed
);
3647 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3649 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3652 cpuset_cpus_allowed(p
, cpus_allowed
);
3653 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3655 * We must have raced with a concurrent cpuset
3656 * update. Just reset the cpus_allowed to the
3657 * cpuset's cpus_allowed
3659 cpumask_copy(new_mask
, cpus_allowed
);
3664 free_cpumask_var(new_mask
);
3665 out_free_cpus_allowed
:
3666 free_cpumask_var(cpus_allowed
);
3673 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3674 struct cpumask
*new_mask
)
3676 if (len
< cpumask_size())
3677 cpumask_clear(new_mask
);
3678 else if (len
> cpumask_size())
3679 len
= cpumask_size();
3681 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3685 * sys_sched_setaffinity - set the cpu affinity of a process
3686 * @pid: pid of the process
3687 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3688 * @user_mask_ptr: user-space pointer to the new cpu mask
3690 * Return: 0 on success. An error code otherwise.
3692 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3693 unsigned long __user
*, user_mask_ptr
)
3695 cpumask_var_t new_mask
;
3698 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3701 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3703 retval
= sched_setaffinity(pid
, new_mask
);
3704 free_cpumask_var(new_mask
);
3708 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
3710 struct task_struct
*p
;
3711 unsigned long flags
;
3718 p
= find_process_by_pid(pid
);
3722 retval
= security_task_getscheduler(p
);
3726 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3727 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
3728 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3738 * sys_sched_getaffinity - get the cpu affinity of a process
3739 * @pid: pid of the process
3740 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3741 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3743 * Return: 0 on success. An error code otherwise.
3745 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
3746 unsigned long __user
*, user_mask_ptr
)
3751 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
3753 if (len
& (sizeof(unsigned long)-1))
3756 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
3759 ret
= sched_getaffinity(pid
, mask
);
3761 size_t retlen
= min_t(size_t, len
, cpumask_size());
3763 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
3768 free_cpumask_var(mask
);
3774 * sys_sched_yield - yield the current processor to other threads.
3776 * This function yields the current CPU to other tasks. If there are no
3777 * other threads running on this CPU then this function will return.
3781 SYSCALL_DEFINE0(sched_yield
)
3783 struct rq
*rq
= this_rq_lock();
3785 schedstat_inc(rq
, yld_count
);
3786 current
->sched_class
->yield_task(rq
);
3789 * Since we are going to call schedule() anyway, there's
3790 * no need to preempt or enable interrupts:
3792 __release(rq
->lock
);
3793 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3794 do_raw_spin_unlock(&rq
->lock
);
3795 sched_preempt_enable_no_resched();
3802 static void __cond_resched(void)
3804 __preempt_count_add(PREEMPT_ACTIVE
);
3806 __preempt_count_sub(PREEMPT_ACTIVE
);
3809 int __sched
_cond_resched(void)
3811 if (should_resched()) {
3817 EXPORT_SYMBOL(_cond_resched
);
3820 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3821 * call schedule, and on return reacquire the lock.
3823 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3824 * operations here to prevent schedule() from being called twice (once via
3825 * spin_unlock(), once by hand).
3827 int __cond_resched_lock(spinlock_t
*lock
)
3829 int resched
= should_resched();
3832 lockdep_assert_held(lock
);
3834 if (spin_needbreak(lock
) || resched
) {
3845 EXPORT_SYMBOL(__cond_resched_lock
);
3847 int __sched
__cond_resched_softirq(void)
3849 BUG_ON(!in_softirq());
3851 if (should_resched()) {
3859 EXPORT_SYMBOL(__cond_resched_softirq
);
3862 * yield - yield the current processor to other threads.
3864 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3866 * The scheduler is at all times free to pick the calling task as the most
3867 * eligible task to run, if removing the yield() call from your code breaks
3868 * it, its already broken.
3870 * Typical broken usage is:
3875 * where one assumes that yield() will let 'the other' process run that will
3876 * make event true. If the current task is a SCHED_FIFO task that will never
3877 * happen. Never use yield() as a progress guarantee!!
3879 * If you want to use yield() to wait for something, use wait_event().
3880 * If you want to use yield() to be 'nice' for others, use cond_resched().
3881 * If you still want to use yield(), do not!
3883 void __sched
yield(void)
3885 set_current_state(TASK_RUNNING
);
3888 EXPORT_SYMBOL(yield
);
3891 * yield_to - yield the current processor to another thread in
3892 * your thread group, or accelerate that thread toward the
3893 * processor it's on.
3895 * @preempt: whether task preemption is allowed or not
3897 * It's the caller's job to ensure that the target task struct
3898 * can't go away on us before we can do any checks.
3901 * true (>0) if we indeed boosted the target task.
3902 * false (0) if we failed to boost the target.
3903 * -ESRCH if there's no task to yield to.
3905 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
3907 struct task_struct
*curr
= current
;
3908 struct rq
*rq
, *p_rq
;
3909 unsigned long flags
;
3912 local_irq_save(flags
);
3918 * If we're the only runnable task on the rq and target rq also
3919 * has only one task, there's absolutely no point in yielding.
3921 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
3926 double_rq_lock(rq
, p_rq
);
3927 while (task_rq(p
) != p_rq
) {
3928 double_rq_unlock(rq
, p_rq
);
3932 if (!curr
->sched_class
->yield_to_task
)
3935 if (curr
->sched_class
!= p
->sched_class
)
3938 if (task_running(p_rq
, p
) || p
->state
)
3941 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
3943 schedstat_inc(rq
, yld_count
);
3945 * Make p's CPU reschedule; pick_next_entity takes care of
3948 if (preempt
&& rq
!= p_rq
)
3949 resched_task(p_rq
->curr
);
3953 double_rq_unlock(rq
, p_rq
);
3955 local_irq_restore(flags
);
3962 EXPORT_SYMBOL_GPL(yield_to
);
3965 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3966 * that process accounting knows that this is a task in IO wait state.
3968 void __sched
io_schedule(void)
3970 struct rq
*rq
= raw_rq();
3972 delayacct_blkio_start();
3973 atomic_inc(&rq
->nr_iowait
);
3974 blk_flush_plug(current
);
3975 current
->in_iowait
= 1;
3977 current
->in_iowait
= 0;
3978 atomic_dec(&rq
->nr_iowait
);
3979 delayacct_blkio_end();
3981 EXPORT_SYMBOL(io_schedule
);
3983 long __sched
io_schedule_timeout(long timeout
)
3985 struct rq
*rq
= raw_rq();
3988 delayacct_blkio_start();
3989 atomic_inc(&rq
->nr_iowait
);
3990 blk_flush_plug(current
);
3991 current
->in_iowait
= 1;
3992 ret
= schedule_timeout(timeout
);
3993 current
->in_iowait
= 0;
3994 atomic_dec(&rq
->nr_iowait
);
3995 delayacct_blkio_end();
4000 * sys_sched_get_priority_max - return maximum RT priority.
4001 * @policy: scheduling class.
4003 * Return: On success, this syscall returns the maximum
4004 * rt_priority that can be used by a given scheduling class.
4005 * On failure, a negative error code is returned.
4007 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4014 ret
= MAX_USER_RT_PRIO
-1;
4026 * sys_sched_get_priority_min - return minimum RT priority.
4027 * @policy: scheduling class.
4029 * Return: On success, this syscall returns the minimum
4030 * rt_priority that can be used by a given scheduling class.
4031 * On failure, a negative error code is returned.
4033 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4051 * sys_sched_rr_get_interval - return the default timeslice of a process.
4052 * @pid: pid of the process.
4053 * @interval: userspace pointer to the timeslice value.
4055 * this syscall writes the default timeslice value of a given process
4056 * into the user-space timespec buffer. A value of '0' means infinity.
4058 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4061 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4062 struct timespec __user
*, interval
)
4064 struct task_struct
*p
;
4065 unsigned int time_slice
;
4066 unsigned long flags
;
4076 p
= find_process_by_pid(pid
);
4080 retval
= security_task_getscheduler(p
);
4084 rq
= task_rq_lock(p
, &flags
);
4085 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4086 task_rq_unlock(rq
, p
, &flags
);
4089 jiffies_to_timespec(time_slice
, &t
);
4090 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4098 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4100 void sched_show_task(struct task_struct
*p
)
4102 unsigned long free
= 0;
4106 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4107 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4108 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4109 #if BITS_PER_LONG == 32
4110 if (state
== TASK_RUNNING
)
4111 printk(KERN_CONT
" running ");
4113 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4115 if (state
== TASK_RUNNING
)
4116 printk(KERN_CONT
" running task ");
4118 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4120 #ifdef CONFIG_DEBUG_STACK_USAGE
4121 free
= stack_not_used(p
);
4124 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4126 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4127 task_pid_nr(p
), ppid
,
4128 (unsigned long)task_thread_info(p
)->flags
);
4130 print_worker_info(KERN_INFO
, p
);
4131 show_stack(p
, NULL
);
4134 void show_state_filter(unsigned long state_filter
)
4136 struct task_struct
*g
, *p
;
4138 #if BITS_PER_LONG == 32
4140 " task PC stack pid father\n");
4143 " task PC stack pid father\n");
4146 do_each_thread(g
, p
) {
4148 * reset the NMI-timeout, listing all files on a slow
4149 * console might take a lot of time:
4151 touch_nmi_watchdog();
4152 if (!state_filter
|| (p
->state
& state_filter
))
4154 } while_each_thread(g
, p
);
4156 touch_all_softlockup_watchdogs();
4158 #ifdef CONFIG_SCHED_DEBUG
4159 sysrq_sched_debug_show();
4163 * Only show locks if all tasks are dumped:
4166 debug_show_all_locks();
4169 void init_idle_bootup_task(struct task_struct
*idle
)
4171 idle
->sched_class
= &idle_sched_class
;
4175 * init_idle - set up an idle thread for a given CPU
4176 * @idle: task in question
4177 * @cpu: cpu the idle task belongs to
4179 * NOTE: this function does not set the idle thread's NEED_RESCHED
4180 * flag, to make booting more robust.
4182 void init_idle(struct task_struct
*idle
, int cpu
)
4184 struct rq
*rq
= cpu_rq(cpu
);
4185 unsigned long flags
;
4187 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4190 idle
->state
= TASK_RUNNING
;
4191 idle
->se
.exec_start
= sched_clock();
4193 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4195 * We're having a chicken and egg problem, even though we are
4196 * holding rq->lock, the cpu isn't yet set to this cpu so the
4197 * lockdep check in task_group() will fail.
4199 * Similar case to sched_fork(). / Alternatively we could
4200 * use task_rq_lock() here and obtain the other rq->lock.
4205 __set_task_cpu(idle
, cpu
);
4208 rq
->curr
= rq
->idle
= idle
;
4209 #if defined(CONFIG_SMP)
4212 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4214 /* Set the preempt count _outside_ the spinlocks! */
4215 init_idle_preempt_count(idle
, cpu
);
4218 * The idle tasks have their own, simple scheduling class:
4220 idle
->sched_class
= &idle_sched_class
;
4221 ftrace_graph_init_idle_task(idle
, cpu
);
4222 vtime_init_idle(idle
, cpu
);
4223 #if defined(CONFIG_SMP)
4224 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4229 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4231 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4232 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4234 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4235 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4239 * This is how migration works:
4241 * 1) we invoke migration_cpu_stop() on the target CPU using
4243 * 2) stopper starts to run (implicitly forcing the migrated thread
4245 * 3) it checks whether the migrated task is still in the wrong runqueue.
4246 * 4) if it's in the wrong runqueue then the migration thread removes
4247 * it and puts it into the right queue.
4248 * 5) stopper completes and stop_one_cpu() returns and the migration
4253 * Change a given task's CPU affinity. Migrate the thread to a
4254 * proper CPU and schedule it away if the CPU it's executing on
4255 * is removed from the allowed bitmask.
4257 * NOTE: the caller must have a valid reference to the task, the
4258 * task must not exit() & deallocate itself prematurely. The
4259 * call is not atomic; no spinlocks may be held.
4261 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4263 unsigned long flags
;
4265 unsigned int dest_cpu
;
4268 rq
= task_rq_lock(p
, &flags
);
4270 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4273 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4278 do_set_cpus_allowed(p
, new_mask
);
4280 /* Can the task run on the task's current CPU? If so, we're done */
4281 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4284 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4286 struct migration_arg arg
= { p
, dest_cpu
};
4287 /* Need help from migration thread: drop lock and wait. */
4288 task_rq_unlock(rq
, p
, &flags
);
4289 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4290 tlb_migrate_finish(p
->mm
);
4294 task_rq_unlock(rq
, p
, &flags
);
4298 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4301 * Move (not current) task off this cpu, onto dest cpu. We're doing
4302 * this because either it can't run here any more (set_cpus_allowed()
4303 * away from this CPU, or CPU going down), or because we're
4304 * attempting to rebalance this task on exec (sched_exec).
4306 * So we race with normal scheduler movements, but that's OK, as long
4307 * as the task is no longer on this CPU.
4309 * Returns non-zero if task was successfully migrated.
4311 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4313 struct rq
*rq_dest
, *rq_src
;
4316 if (unlikely(!cpu_active(dest_cpu
)))
4319 rq_src
= cpu_rq(src_cpu
);
4320 rq_dest
= cpu_rq(dest_cpu
);
4322 raw_spin_lock(&p
->pi_lock
);
4323 double_rq_lock(rq_src
, rq_dest
);
4324 /* Already moved. */
4325 if (task_cpu(p
) != src_cpu
)
4327 /* Affinity changed (again). */
4328 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4332 * If we're not on a rq, the next wake-up will ensure we're
4336 dequeue_task(rq_src
, p
, 0);
4337 set_task_cpu(p
, dest_cpu
);
4338 enqueue_task(rq_dest
, p
, 0);
4339 check_preempt_curr(rq_dest
, p
, 0);
4344 double_rq_unlock(rq_src
, rq_dest
);
4345 raw_spin_unlock(&p
->pi_lock
);
4350 * migration_cpu_stop - this will be executed by a highprio stopper thread
4351 * and performs thread migration by bumping thread off CPU then
4352 * 'pushing' onto another runqueue.
4354 static int migration_cpu_stop(void *data
)
4356 struct migration_arg
*arg
= data
;
4359 * The original target cpu might have gone down and we might
4360 * be on another cpu but it doesn't matter.
4362 local_irq_disable();
4363 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4368 #ifdef CONFIG_HOTPLUG_CPU
4371 * Ensures that the idle task is using init_mm right before its cpu goes
4374 void idle_task_exit(void)
4376 struct mm_struct
*mm
= current
->active_mm
;
4378 BUG_ON(cpu_online(smp_processor_id()));
4381 switch_mm(mm
, &init_mm
, current
);
4386 * Since this CPU is going 'away' for a while, fold any nr_active delta
4387 * we might have. Assumes we're called after migrate_tasks() so that the
4388 * nr_active count is stable.
4390 * Also see the comment "Global load-average calculations".
4392 static void calc_load_migrate(struct rq
*rq
)
4394 long delta
= calc_load_fold_active(rq
);
4396 atomic_long_add(delta
, &calc_load_tasks
);
4400 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4401 * try_to_wake_up()->select_task_rq().
4403 * Called with rq->lock held even though we'er in stop_machine() and
4404 * there's no concurrency possible, we hold the required locks anyway
4405 * because of lock validation efforts.
4407 static void migrate_tasks(unsigned int dead_cpu
)
4409 struct rq
*rq
= cpu_rq(dead_cpu
);
4410 struct task_struct
*next
, *stop
= rq
->stop
;
4414 * Fudge the rq selection such that the below task selection loop
4415 * doesn't get stuck on the currently eligible stop task.
4417 * We're currently inside stop_machine() and the rq is either stuck
4418 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4419 * either way we should never end up calling schedule() until we're
4425 * put_prev_task() and pick_next_task() sched
4426 * class method both need to have an up-to-date
4427 * value of rq->clock[_task]
4429 update_rq_clock(rq
);
4433 * There's this thread running, bail when that's the only
4436 if (rq
->nr_running
== 1)
4439 next
= pick_next_task(rq
);
4441 next
->sched_class
->put_prev_task(rq
, next
);
4443 /* Find suitable destination for @next, with force if needed. */
4444 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4445 raw_spin_unlock(&rq
->lock
);
4447 __migrate_task(next
, dead_cpu
, dest_cpu
);
4449 raw_spin_lock(&rq
->lock
);
4455 #endif /* CONFIG_HOTPLUG_CPU */
4457 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4459 static struct ctl_table sd_ctl_dir
[] = {
4461 .procname
= "sched_domain",
4467 static struct ctl_table sd_ctl_root
[] = {
4469 .procname
= "kernel",
4471 .child
= sd_ctl_dir
,
4476 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4478 struct ctl_table
*entry
=
4479 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4484 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4486 struct ctl_table
*entry
;
4489 * In the intermediate directories, both the child directory and
4490 * procname are dynamically allocated and could fail but the mode
4491 * will always be set. In the lowest directory the names are
4492 * static strings and all have proc handlers.
4494 for (entry
= *tablep
; entry
->mode
; entry
++) {
4496 sd_free_ctl_entry(&entry
->child
);
4497 if (entry
->proc_handler
== NULL
)
4498 kfree(entry
->procname
);
4505 static int min_load_idx
= 0;
4506 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4509 set_table_entry(struct ctl_table
*entry
,
4510 const char *procname
, void *data
, int maxlen
,
4511 umode_t mode
, proc_handler
*proc_handler
,
4514 entry
->procname
= procname
;
4516 entry
->maxlen
= maxlen
;
4518 entry
->proc_handler
= proc_handler
;
4521 entry
->extra1
= &min_load_idx
;
4522 entry
->extra2
= &max_load_idx
;
4526 static struct ctl_table
*
4527 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4529 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4534 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4535 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4536 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4537 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4538 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4539 sizeof(int), 0644, proc_dointvec_minmax
, true);
4540 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4541 sizeof(int), 0644, proc_dointvec_minmax
, true);
4542 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4543 sizeof(int), 0644, proc_dointvec_minmax
, true);
4544 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4545 sizeof(int), 0644, proc_dointvec_minmax
, true);
4546 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4547 sizeof(int), 0644, proc_dointvec_minmax
, true);
4548 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4549 sizeof(int), 0644, proc_dointvec_minmax
, false);
4550 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4551 sizeof(int), 0644, proc_dointvec_minmax
, false);
4552 set_table_entry(&table
[9], "cache_nice_tries",
4553 &sd
->cache_nice_tries
,
4554 sizeof(int), 0644, proc_dointvec_minmax
, false);
4555 set_table_entry(&table
[10], "flags", &sd
->flags
,
4556 sizeof(int), 0644, proc_dointvec_minmax
, false);
4557 set_table_entry(&table
[11], "name", sd
->name
,
4558 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4559 /* &table[12] is terminator */
4564 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4566 struct ctl_table
*entry
, *table
;
4567 struct sched_domain
*sd
;
4568 int domain_num
= 0, i
;
4571 for_each_domain(cpu
, sd
)
4573 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4578 for_each_domain(cpu
, sd
) {
4579 snprintf(buf
, 32, "domain%d", i
);
4580 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4582 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4589 static struct ctl_table_header
*sd_sysctl_header
;
4590 static void register_sched_domain_sysctl(void)
4592 int i
, cpu_num
= num_possible_cpus();
4593 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4596 WARN_ON(sd_ctl_dir
[0].child
);
4597 sd_ctl_dir
[0].child
= entry
;
4602 for_each_possible_cpu(i
) {
4603 snprintf(buf
, 32, "cpu%d", i
);
4604 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4606 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4610 WARN_ON(sd_sysctl_header
);
4611 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4614 /* may be called multiple times per register */
4615 static void unregister_sched_domain_sysctl(void)
4617 if (sd_sysctl_header
)
4618 unregister_sysctl_table(sd_sysctl_header
);
4619 sd_sysctl_header
= NULL
;
4620 if (sd_ctl_dir
[0].child
)
4621 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4624 static void register_sched_domain_sysctl(void)
4627 static void unregister_sched_domain_sysctl(void)
4632 static void set_rq_online(struct rq
*rq
)
4635 const struct sched_class
*class;
4637 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
4640 for_each_class(class) {
4641 if (class->rq_online
)
4642 class->rq_online(rq
);
4647 static void set_rq_offline(struct rq
*rq
)
4650 const struct sched_class
*class;
4652 for_each_class(class) {
4653 if (class->rq_offline
)
4654 class->rq_offline(rq
);
4657 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
4663 * migration_call - callback that gets triggered when a CPU is added.
4664 * Here we can start up the necessary migration thread for the new CPU.
4667 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
4669 int cpu
= (long)hcpu
;
4670 unsigned long flags
;
4671 struct rq
*rq
= cpu_rq(cpu
);
4673 switch (action
& ~CPU_TASKS_FROZEN
) {
4675 case CPU_UP_PREPARE
:
4676 rq
->calc_load_update
= calc_load_update
;
4680 /* Update our root-domain */
4681 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4683 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
4687 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4690 #ifdef CONFIG_HOTPLUG_CPU
4692 sched_ttwu_pending();
4693 /* Update our root-domain */
4694 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4696 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
4700 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
4701 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4705 calc_load_migrate(rq
);
4710 update_max_interval();
4716 * Register at high priority so that task migration (migrate_all_tasks)
4717 * happens before everything else. This has to be lower priority than
4718 * the notifier in the perf_event subsystem, though.
4720 static struct notifier_block migration_notifier
= {
4721 .notifier_call
= migration_call
,
4722 .priority
= CPU_PRI_MIGRATION
,
4725 static int sched_cpu_active(struct notifier_block
*nfb
,
4726 unsigned long action
, void *hcpu
)
4728 switch (action
& ~CPU_TASKS_FROZEN
) {
4730 case CPU_DOWN_FAILED
:
4731 set_cpu_active((long)hcpu
, true);
4738 static int sched_cpu_inactive(struct notifier_block
*nfb
,
4739 unsigned long action
, void *hcpu
)
4741 switch (action
& ~CPU_TASKS_FROZEN
) {
4742 case CPU_DOWN_PREPARE
:
4743 set_cpu_active((long)hcpu
, false);
4750 static int __init
migration_init(void)
4752 void *cpu
= (void *)(long)smp_processor_id();
4755 /* Initialize migration for the boot CPU */
4756 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4757 BUG_ON(err
== NOTIFY_BAD
);
4758 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4759 register_cpu_notifier(&migration_notifier
);
4761 /* Register cpu active notifiers */
4762 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
4763 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
4767 early_initcall(migration_init
);
4772 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
4774 #ifdef CONFIG_SCHED_DEBUG
4776 static __read_mostly
int sched_debug_enabled
;
4778 static int __init
sched_debug_setup(char *str
)
4780 sched_debug_enabled
= 1;
4784 early_param("sched_debug", sched_debug_setup
);
4786 static inline bool sched_debug(void)
4788 return sched_debug_enabled
;
4791 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
4792 struct cpumask
*groupmask
)
4794 struct sched_group
*group
= sd
->groups
;
4797 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
4798 cpumask_clear(groupmask
);
4800 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
4802 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4803 printk("does not load-balance\n");
4805 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
4810 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
4812 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
4813 printk(KERN_ERR
"ERROR: domain->span does not contain "
4816 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
4817 printk(KERN_ERR
"ERROR: domain->groups does not contain"
4821 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
4825 printk(KERN_ERR
"ERROR: group is NULL\n");
4830 * Even though we initialize ->power to something semi-sane,
4831 * we leave power_orig unset. This allows us to detect if
4832 * domain iteration is still funny without causing /0 traps.
4834 if (!group
->sgp
->power_orig
) {
4835 printk(KERN_CONT
"\n");
4836 printk(KERN_ERR
"ERROR: domain->cpu_power not "
4841 if (!cpumask_weight(sched_group_cpus(group
))) {
4842 printk(KERN_CONT
"\n");
4843 printk(KERN_ERR
"ERROR: empty group\n");
4847 if (!(sd
->flags
& SD_OVERLAP
) &&
4848 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
4849 printk(KERN_CONT
"\n");
4850 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4854 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
4856 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
4858 printk(KERN_CONT
" %s", str
);
4859 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
4860 printk(KERN_CONT
" (cpu_power = %d)",
4864 group
= group
->next
;
4865 } while (group
!= sd
->groups
);
4866 printk(KERN_CONT
"\n");
4868 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
4869 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4872 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
4873 printk(KERN_ERR
"ERROR: parent span is not a superset "
4874 "of domain->span\n");
4878 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4882 if (!sched_debug_enabled
)
4886 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4890 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4893 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
4901 #else /* !CONFIG_SCHED_DEBUG */
4902 # define sched_domain_debug(sd, cpu) do { } while (0)
4903 static inline bool sched_debug(void)
4907 #endif /* CONFIG_SCHED_DEBUG */
4909 static int sd_degenerate(struct sched_domain
*sd
)
4911 if (cpumask_weight(sched_domain_span(sd
)) == 1)
4914 /* Following flags need at least 2 groups */
4915 if (sd
->flags
& (SD_LOAD_BALANCE
|
4916 SD_BALANCE_NEWIDLE
|
4920 SD_SHARE_PKG_RESOURCES
)) {
4921 if (sd
->groups
!= sd
->groups
->next
)
4925 /* Following flags don't use groups */
4926 if (sd
->flags
& (SD_WAKE_AFFINE
))
4933 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
4935 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4937 if (sd_degenerate(parent
))
4940 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
4943 /* Flags needing groups don't count if only 1 group in parent */
4944 if (parent
->groups
== parent
->groups
->next
) {
4945 pflags
&= ~(SD_LOAD_BALANCE
|
4946 SD_BALANCE_NEWIDLE
|
4950 SD_SHARE_PKG_RESOURCES
|
4952 if (nr_node_ids
== 1)
4953 pflags
&= ~SD_SERIALIZE
;
4955 if (~cflags
& pflags
)
4961 static void free_rootdomain(struct rcu_head
*rcu
)
4963 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
4965 cpupri_cleanup(&rd
->cpupri
);
4966 free_cpumask_var(rd
->rto_mask
);
4967 free_cpumask_var(rd
->online
);
4968 free_cpumask_var(rd
->span
);
4972 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
4974 struct root_domain
*old_rd
= NULL
;
4975 unsigned long flags
;
4977 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4982 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
4985 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
4988 * If we dont want to free the old_rt yet then
4989 * set old_rd to NULL to skip the freeing later
4992 if (!atomic_dec_and_test(&old_rd
->refcount
))
4996 atomic_inc(&rd
->refcount
);
4999 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5000 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5003 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5006 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5009 static int init_rootdomain(struct root_domain
*rd
)
5011 memset(rd
, 0, sizeof(*rd
));
5013 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5015 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5017 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5020 if (cpupri_init(&rd
->cpupri
) != 0)
5025 free_cpumask_var(rd
->rto_mask
);
5027 free_cpumask_var(rd
->online
);
5029 free_cpumask_var(rd
->span
);
5035 * By default the system creates a single root-domain with all cpus as
5036 * members (mimicking the global state we have today).
5038 struct root_domain def_root_domain
;
5040 static void init_defrootdomain(void)
5042 init_rootdomain(&def_root_domain
);
5044 atomic_set(&def_root_domain
.refcount
, 1);
5047 static struct root_domain
*alloc_rootdomain(void)
5049 struct root_domain
*rd
;
5051 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5055 if (init_rootdomain(rd
) != 0) {
5063 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5065 struct sched_group
*tmp
, *first
;
5074 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5079 } while (sg
!= first
);
5082 static void free_sched_domain(struct rcu_head
*rcu
)
5084 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5087 * If its an overlapping domain it has private groups, iterate and
5090 if (sd
->flags
& SD_OVERLAP
) {
5091 free_sched_groups(sd
->groups
, 1);
5092 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5093 kfree(sd
->groups
->sgp
);
5099 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5101 call_rcu(&sd
->rcu
, free_sched_domain
);
5104 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5106 for (; sd
; sd
= sd
->parent
)
5107 destroy_sched_domain(sd
, cpu
);
5111 * Keep a special pointer to the highest sched_domain that has
5112 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5113 * allows us to avoid some pointer chasing select_idle_sibling().
5115 * Also keep a unique ID per domain (we use the first cpu number in
5116 * the cpumask of the domain), this allows us to quickly tell if
5117 * two cpus are in the same cache domain, see cpus_share_cache().
5119 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5120 DEFINE_PER_CPU(int, sd_llc_size
);
5121 DEFINE_PER_CPU(int, sd_llc_id
);
5123 static void update_top_cache_domain(int cpu
)
5125 struct sched_domain
*sd
;
5129 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5131 id
= cpumask_first(sched_domain_span(sd
));
5132 size
= cpumask_weight(sched_domain_span(sd
));
5135 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5136 per_cpu(sd_llc_size
, cpu
) = size
;
5137 per_cpu(sd_llc_id
, cpu
) = id
;
5141 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5142 * hold the hotplug lock.
5145 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5147 struct rq
*rq
= cpu_rq(cpu
);
5148 struct sched_domain
*tmp
;
5150 /* Remove the sched domains which do not contribute to scheduling. */
5151 for (tmp
= sd
; tmp
; ) {
5152 struct sched_domain
*parent
= tmp
->parent
;
5156 if (sd_parent_degenerate(tmp
, parent
)) {
5157 tmp
->parent
= parent
->parent
;
5159 parent
->parent
->child
= tmp
;
5161 * Transfer SD_PREFER_SIBLING down in case of a
5162 * degenerate parent; the spans match for this
5163 * so the property transfers.
5165 if (parent
->flags
& SD_PREFER_SIBLING
)
5166 tmp
->flags
|= SD_PREFER_SIBLING
;
5167 destroy_sched_domain(parent
, cpu
);
5172 if (sd
&& sd_degenerate(sd
)) {
5175 destroy_sched_domain(tmp
, cpu
);
5180 sched_domain_debug(sd
, cpu
);
5182 rq_attach_root(rq
, rd
);
5184 rcu_assign_pointer(rq
->sd
, sd
);
5185 destroy_sched_domains(tmp
, cpu
);
5187 update_top_cache_domain(cpu
);
5190 /* cpus with isolated domains */
5191 static cpumask_var_t cpu_isolated_map
;
5193 /* Setup the mask of cpus configured for isolated domains */
5194 static int __init
isolated_cpu_setup(char *str
)
5196 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5197 cpulist_parse(str
, cpu_isolated_map
);
5201 __setup("isolcpus=", isolated_cpu_setup
);
5203 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5205 return cpumask_of_node(cpu_to_node(cpu
));
5209 struct sched_domain
**__percpu sd
;
5210 struct sched_group
**__percpu sg
;
5211 struct sched_group_power
**__percpu sgp
;
5215 struct sched_domain
** __percpu sd
;
5216 struct root_domain
*rd
;
5226 struct sched_domain_topology_level
;
5228 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5229 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5231 #define SDTL_OVERLAP 0x01
5233 struct sched_domain_topology_level
{
5234 sched_domain_init_f init
;
5235 sched_domain_mask_f mask
;
5238 struct sd_data data
;
5242 * Build an iteration mask that can exclude certain CPUs from the upwards
5245 * Asymmetric node setups can result in situations where the domain tree is of
5246 * unequal depth, make sure to skip domains that already cover the entire
5249 * In that case build_sched_domains() will have terminated the iteration early
5250 * and our sibling sd spans will be empty. Domains should always include the
5251 * cpu they're built on, so check that.
5254 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5256 const struct cpumask
*span
= sched_domain_span(sd
);
5257 struct sd_data
*sdd
= sd
->private;
5258 struct sched_domain
*sibling
;
5261 for_each_cpu(i
, span
) {
5262 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5263 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5266 cpumask_set_cpu(i
, sched_group_mask(sg
));
5271 * Return the canonical balance cpu for this group, this is the first cpu
5272 * of this group that's also in the iteration mask.
5274 int group_balance_cpu(struct sched_group
*sg
)
5276 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5280 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5282 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5283 const struct cpumask
*span
= sched_domain_span(sd
);
5284 struct cpumask
*covered
= sched_domains_tmpmask
;
5285 struct sd_data
*sdd
= sd
->private;
5286 struct sched_domain
*child
;
5289 cpumask_clear(covered
);
5291 for_each_cpu(i
, span
) {
5292 struct cpumask
*sg_span
;
5294 if (cpumask_test_cpu(i
, covered
))
5297 child
= *per_cpu_ptr(sdd
->sd
, i
);
5299 /* See the comment near build_group_mask(). */
5300 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5303 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5304 GFP_KERNEL
, cpu_to_node(cpu
));
5309 sg_span
= sched_group_cpus(sg
);
5311 child
= child
->child
;
5312 cpumask_copy(sg_span
, sched_domain_span(child
));
5314 cpumask_set_cpu(i
, sg_span
);
5316 cpumask_or(covered
, covered
, sg_span
);
5318 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5319 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5320 build_group_mask(sd
, sg
);
5323 * Initialize sgp->power such that even if we mess up the
5324 * domains and no possible iteration will get us here, we won't
5327 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5330 * Make sure the first group of this domain contains the
5331 * canonical balance cpu. Otherwise the sched_domain iteration
5332 * breaks. See update_sg_lb_stats().
5334 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5335 group_balance_cpu(sg
) == cpu
)
5345 sd
->groups
= groups
;
5350 free_sched_groups(first
, 0);
5355 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5357 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5358 struct sched_domain
*child
= sd
->child
;
5361 cpu
= cpumask_first(sched_domain_span(child
));
5364 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5365 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5366 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5373 * build_sched_groups will build a circular linked list of the groups
5374 * covered by the given span, and will set each group's ->cpumask correctly,
5375 * and ->cpu_power to 0.
5377 * Assumes the sched_domain tree is fully constructed
5380 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5382 struct sched_group
*first
= NULL
, *last
= NULL
;
5383 struct sd_data
*sdd
= sd
->private;
5384 const struct cpumask
*span
= sched_domain_span(sd
);
5385 struct cpumask
*covered
;
5388 get_group(cpu
, sdd
, &sd
->groups
);
5389 atomic_inc(&sd
->groups
->ref
);
5391 if (cpu
!= cpumask_first(span
))
5394 lockdep_assert_held(&sched_domains_mutex
);
5395 covered
= sched_domains_tmpmask
;
5397 cpumask_clear(covered
);
5399 for_each_cpu(i
, span
) {
5400 struct sched_group
*sg
;
5403 if (cpumask_test_cpu(i
, covered
))
5406 group
= get_group(i
, sdd
, &sg
);
5407 cpumask_clear(sched_group_cpus(sg
));
5409 cpumask_setall(sched_group_mask(sg
));
5411 for_each_cpu(j
, span
) {
5412 if (get_group(j
, sdd
, NULL
) != group
)
5415 cpumask_set_cpu(j
, covered
);
5416 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5431 * Initialize sched groups cpu_power.
5433 * cpu_power indicates the capacity of sched group, which is used while
5434 * distributing the load between different sched groups in a sched domain.
5435 * Typically cpu_power for all the groups in a sched domain will be same unless
5436 * there are asymmetries in the topology. If there are asymmetries, group
5437 * having more cpu_power will pickup more load compared to the group having
5440 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5442 struct sched_group
*sg
= sd
->groups
;
5447 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5449 } while (sg
!= sd
->groups
);
5451 if (cpu
!= group_balance_cpu(sg
))
5454 update_group_power(sd
, cpu
);
5455 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5458 int __weak
arch_sd_sibling_asym_packing(void)
5460 return 0*SD_ASYM_PACKING
;
5464 * Initializers for schedule domains
5465 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5468 #ifdef CONFIG_SCHED_DEBUG
5469 # define SD_INIT_NAME(sd, type) sd->name = #type
5471 # define SD_INIT_NAME(sd, type) do { } while (0)
5474 #define SD_INIT_FUNC(type) \
5475 static noinline struct sched_domain * \
5476 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5478 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5479 *sd = SD_##type##_INIT; \
5480 SD_INIT_NAME(sd, type); \
5481 sd->private = &tl->data; \
5486 #ifdef CONFIG_SCHED_SMT
5487 SD_INIT_FUNC(SIBLING
)
5489 #ifdef CONFIG_SCHED_MC
5492 #ifdef CONFIG_SCHED_BOOK
5496 static int default_relax_domain_level
= -1;
5497 int sched_domain_level_max
;
5499 static int __init
setup_relax_domain_level(char *str
)
5501 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5502 pr_warn("Unable to set relax_domain_level\n");
5506 __setup("relax_domain_level=", setup_relax_domain_level
);
5508 static void set_domain_attribute(struct sched_domain
*sd
,
5509 struct sched_domain_attr
*attr
)
5513 if (!attr
|| attr
->relax_domain_level
< 0) {
5514 if (default_relax_domain_level
< 0)
5517 request
= default_relax_domain_level
;
5519 request
= attr
->relax_domain_level
;
5520 if (request
< sd
->level
) {
5521 /* turn off idle balance on this domain */
5522 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5524 /* turn on idle balance on this domain */
5525 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5529 static void __sdt_free(const struct cpumask
*cpu_map
);
5530 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5532 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5533 const struct cpumask
*cpu_map
)
5537 if (!atomic_read(&d
->rd
->refcount
))
5538 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5540 free_percpu(d
->sd
); /* fall through */
5542 __sdt_free(cpu_map
); /* fall through */
5548 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5549 const struct cpumask
*cpu_map
)
5551 memset(d
, 0, sizeof(*d
));
5553 if (__sdt_alloc(cpu_map
))
5554 return sa_sd_storage
;
5555 d
->sd
= alloc_percpu(struct sched_domain
*);
5557 return sa_sd_storage
;
5558 d
->rd
= alloc_rootdomain();
5561 return sa_rootdomain
;
5565 * NULL the sd_data elements we've used to build the sched_domain and
5566 * sched_group structure so that the subsequent __free_domain_allocs()
5567 * will not free the data we're using.
5569 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5571 struct sd_data
*sdd
= sd
->private;
5573 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5574 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5576 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5577 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5579 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5580 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5583 #ifdef CONFIG_SCHED_SMT
5584 static const struct cpumask
*cpu_smt_mask(int cpu
)
5586 return topology_thread_cpumask(cpu
);
5591 * Topology list, bottom-up.
5593 static struct sched_domain_topology_level default_topology
[] = {
5594 #ifdef CONFIG_SCHED_SMT
5595 { sd_init_SIBLING
, cpu_smt_mask
, },
5597 #ifdef CONFIG_SCHED_MC
5598 { sd_init_MC
, cpu_coregroup_mask
, },
5600 #ifdef CONFIG_SCHED_BOOK
5601 { sd_init_BOOK
, cpu_book_mask
, },
5603 { sd_init_CPU
, cpu_cpu_mask
, },
5607 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5609 #define for_each_sd_topology(tl) \
5610 for (tl = sched_domain_topology; tl->init; tl++)
5614 static int sched_domains_numa_levels
;
5615 static int *sched_domains_numa_distance
;
5616 static struct cpumask
***sched_domains_numa_masks
;
5617 static int sched_domains_curr_level
;
5619 static inline int sd_local_flags(int level
)
5621 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5624 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
5627 static struct sched_domain
*
5628 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
5630 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5631 int level
= tl
->numa_level
;
5632 int sd_weight
= cpumask_weight(
5633 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
5635 *sd
= (struct sched_domain
){
5636 .min_interval
= sd_weight
,
5637 .max_interval
= 2*sd_weight
,
5639 .imbalance_pct
= 125,
5640 .cache_nice_tries
= 2,
5647 .flags
= 1*SD_LOAD_BALANCE
5648 | 1*SD_BALANCE_NEWIDLE
5653 | 0*SD_SHARE_CPUPOWER
5654 | 0*SD_SHARE_PKG_RESOURCES
5656 | 0*SD_PREFER_SIBLING
5657 | sd_local_flags(level
)
5659 .last_balance
= jiffies
,
5660 .balance_interval
= sd_weight
,
5662 SD_INIT_NAME(sd
, NUMA
);
5663 sd
->private = &tl
->data
;
5666 * Ugly hack to pass state to sd_numa_mask()...
5668 sched_domains_curr_level
= tl
->numa_level
;
5673 static const struct cpumask
*sd_numa_mask(int cpu
)
5675 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
5678 static void sched_numa_warn(const char *str
)
5680 static int done
= false;
5688 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
5690 for (i
= 0; i
< nr_node_ids
; i
++) {
5691 printk(KERN_WARNING
" ");
5692 for (j
= 0; j
< nr_node_ids
; j
++)
5693 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
5694 printk(KERN_CONT
"\n");
5696 printk(KERN_WARNING
"\n");
5699 static bool find_numa_distance(int distance
)
5703 if (distance
== node_distance(0, 0))
5706 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5707 if (sched_domains_numa_distance
[i
] == distance
)
5714 static void sched_init_numa(void)
5716 int next_distance
, curr_distance
= node_distance(0, 0);
5717 struct sched_domain_topology_level
*tl
;
5721 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
5722 if (!sched_domains_numa_distance
)
5726 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5727 * unique distances in the node_distance() table.
5729 * Assumes node_distance(0,j) includes all distances in
5730 * node_distance(i,j) in order to avoid cubic time.
5732 next_distance
= curr_distance
;
5733 for (i
= 0; i
< nr_node_ids
; i
++) {
5734 for (j
= 0; j
< nr_node_ids
; j
++) {
5735 for (k
= 0; k
< nr_node_ids
; k
++) {
5736 int distance
= node_distance(i
, k
);
5738 if (distance
> curr_distance
&&
5739 (distance
< next_distance
||
5740 next_distance
== curr_distance
))
5741 next_distance
= distance
;
5744 * While not a strong assumption it would be nice to know
5745 * about cases where if node A is connected to B, B is not
5746 * equally connected to A.
5748 if (sched_debug() && node_distance(k
, i
) != distance
)
5749 sched_numa_warn("Node-distance not symmetric");
5751 if (sched_debug() && i
&& !find_numa_distance(distance
))
5752 sched_numa_warn("Node-0 not representative");
5754 if (next_distance
!= curr_distance
) {
5755 sched_domains_numa_distance
[level
++] = next_distance
;
5756 sched_domains_numa_levels
= level
;
5757 curr_distance
= next_distance
;
5762 * In case of sched_debug() we verify the above assumption.
5768 * 'level' contains the number of unique distances, excluding the
5769 * identity distance node_distance(i,i).
5771 * The sched_domains_numa_distance[] array includes the actual distance
5776 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5777 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5778 * the array will contain less then 'level' members. This could be
5779 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5780 * in other functions.
5782 * We reset it to 'level' at the end of this function.
5784 sched_domains_numa_levels
= 0;
5786 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
5787 if (!sched_domains_numa_masks
)
5791 * Now for each level, construct a mask per node which contains all
5792 * cpus of nodes that are that many hops away from us.
5794 for (i
= 0; i
< level
; i
++) {
5795 sched_domains_numa_masks
[i
] =
5796 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
5797 if (!sched_domains_numa_masks
[i
])
5800 for (j
= 0; j
< nr_node_ids
; j
++) {
5801 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
5805 sched_domains_numa_masks
[i
][j
] = mask
;
5807 for (k
= 0; k
< nr_node_ids
; k
++) {
5808 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
5811 cpumask_or(mask
, mask
, cpumask_of_node(k
));
5816 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
5817 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
5822 * Copy the default topology bits..
5824 for (i
= 0; default_topology
[i
].init
; i
++)
5825 tl
[i
] = default_topology
[i
];
5828 * .. and append 'j' levels of NUMA goodness.
5830 for (j
= 0; j
< level
; i
++, j
++) {
5831 tl
[i
] = (struct sched_domain_topology_level
){
5832 .init
= sd_numa_init
,
5833 .mask
= sd_numa_mask
,
5834 .flags
= SDTL_OVERLAP
,
5839 sched_domain_topology
= tl
;
5841 sched_domains_numa_levels
= level
;
5844 static void sched_domains_numa_masks_set(int cpu
)
5847 int node
= cpu_to_node(cpu
);
5849 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5850 for (j
= 0; j
< nr_node_ids
; j
++) {
5851 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
5852 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
5857 static void sched_domains_numa_masks_clear(int cpu
)
5860 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5861 for (j
= 0; j
< nr_node_ids
; j
++)
5862 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
5867 * Update sched_domains_numa_masks[level][node] array when new cpus
5870 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
5871 unsigned long action
,
5874 int cpu
= (long)hcpu
;
5876 switch (action
& ~CPU_TASKS_FROZEN
) {
5878 sched_domains_numa_masks_set(cpu
);
5882 sched_domains_numa_masks_clear(cpu
);
5892 static inline void sched_init_numa(void)
5896 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
5897 unsigned long action
,
5902 #endif /* CONFIG_NUMA */
5904 static int __sdt_alloc(const struct cpumask
*cpu_map
)
5906 struct sched_domain_topology_level
*tl
;
5909 for_each_sd_topology(tl
) {
5910 struct sd_data
*sdd
= &tl
->data
;
5912 sdd
->sd
= alloc_percpu(struct sched_domain
*);
5916 sdd
->sg
= alloc_percpu(struct sched_group
*);
5920 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
5924 for_each_cpu(j
, cpu_map
) {
5925 struct sched_domain
*sd
;
5926 struct sched_group
*sg
;
5927 struct sched_group_power
*sgp
;
5929 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
5930 GFP_KERNEL
, cpu_to_node(j
));
5934 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
5936 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5937 GFP_KERNEL
, cpu_to_node(j
));
5943 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
5945 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
5946 GFP_KERNEL
, cpu_to_node(j
));
5950 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
5957 static void __sdt_free(const struct cpumask
*cpu_map
)
5959 struct sched_domain_topology_level
*tl
;
5962 for_each_sd_topology(tl
) {
5963 struct sd_data
*sdd
= &tl
->data
;
5965 for_each_cpu(j
, cpu_map
) {
5966 struct sched_domain
*sd
;
5969 sd
= *per_cpu_ptr(sdd
->sd
, j
);
5970 if (sd
&& (sd
->flags
& SD_OVERLAP
))
5971 free_sched_groups(sd
->groups
, 0);
5972 kfree(*per_cpu_ptr(sdd
->sd
, j
));
5976 kfree(*per_cpu_ptr(sdd
->sg
, j
));
5978 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
5980 free_percpu(sdd
->sd
);
5982 free_percpu(sdd
->sg
);
5984 free_percpu(sdd
->sgp
);
5989 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
5990 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
5991 struct sched_domain
*child
, int cpu
)
5993 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
5997 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
5999 sd
->level
= child
->level
+ 1;
6000 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6004 set_domain_attribute(sd
, attr
);
6010 * Build sched domains for a given set of cpus and attach the sched domains
6011 * to the individual cpus
6013 static int build_sched_domains(const struct cpumask
*cpu_map
,
6014 struct sched_domain_attr
*attr
)
6016 enum s_alloc alloc_state
;
6017 struct sched_domain
*sd
;
6019 int i
, ret
= -ENOMEM
;
6021 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6022 if (alloc_state
!= sa_rootdomain
)
6025 /* Set up domains for cpus specified by the cpu_map. */
6026 for_each_cpu(i
, cpu_map
) {
6027 struct sched_domain_topology_level
*tl
;
6030 for_each_sd_topology(tl
) {
6031 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6032 if (tl
== sched_domain_topology
)
6033 *per_cpu_ptr(d
.sd
, i
) = sd
;
6034 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6035 sd
->flags
|= SD_OVERLAP
;
6036 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6041 /* Build the groups for the domains */
6042 for_each_cpu(i
, cpu_map
) {
6043 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6044 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6045 if (sd
->flags
& SD_OVERLAP
) {
6046 if (build_overlap_sched_groups(sd
, i
))
6049 if (build_sched_groups(sd
, i
))
6055 /* Calculate CPU power for physical packages and nodes */
6056 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6057 if (!cpumask_test_cpu(i
, cpu_map
))
6060 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6061 claim_allocations(i
, sd
);
6062 init_sched_groups_power(i
, sd
);
6066 /* Attach the domains */
6068 for_each_cpu(i
, cpu_map
) {
6069 sd
= *per_cpu_ptr(d
.sd
, i
);
6070 cpu_attach_domain(sd
, d
.rd
, i
);
6076 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6080 static cpumask_var_t
*doms_cur
; /* current sched domains */
6081 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6082 static struct sched_domain_attr
*dattr_cur
;
6083 /* attribues of custom domains in 'doms_cur' */
6086 * Special case: If a kmalloc of a doms_cur partition (array of
6087 * cpumask) fails, then fallback to a single sched domain,
6088 * as determined by the single cpumask fallback_doms.
6090 static cpumask_var_t fallback_doms
;
6093 * arch_update_cpu_topology lets virtualized architectures update the
6094 * cpu core maps. It is supposed to return 1 if the topology changed
6095 * or 0 if it stayed the same.
6097 int __attribute__((weak
)) arch_update_cpu_topology(void)
6102 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6105 cpumask_var_t
*doms
;
6107 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6110 for (i
= 0; i
< ndoms
; i
++) {
6111 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6112 free_sched_domains(doms
, i
);
6119 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6122 for (i
= 0; i
< ndoms
; i
++)
6123 free_cpumask_var(doms
[i
]);
6128 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6129 * For now this just excludes isolated cpus, but could be used to
6130 * exclude other special cases in the future.
6132 static int init_sched_domains(const struct cpumask
*cpu_map
)
6136 arch_update_cpu_topology();
6138 doms_cur
= alloc_sched_domains(ndoms_cur
);
6140 doms_cur
= &fallback_doms
;
6141 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6142 err
= build_sched_domains(doms_cur
[0], NULL
);
6143 register_sched_domain_sysctl();
6149 * Detach sched domains from a group of cpus specified in cpu_map
6150 * These cpus will now be attached to the NULL domain
6152 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6157 for_each_cpu(i
, cpu_map
)
6158 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6162 /* handle null as "default" */
6163 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6164 struct sched_domain_attr
*new, int idx_new
)
6166 struct sched_domain_attr tmp
;
6173 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6174 new ? (new + idx_new
) : &tmp
,
6175 sizeof(struct sched_domain_attr
));
6179 * Partition sched domains as specified by the 'ndoms_new'
6180 * cpumasks in the array doms_new[] of cpumasks. This compares
6181 * doms_new[] to the current sched domain partitioning, doms_cur[].
6182 * It destroys each deleted domain and builds each new domain.
6184 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6185 * The masks don't intersect (don't overlap.) We should setup one
6186 * sched domain for each mask. CPUs not in any of the cpumasks will
6187 * not be load balanced. If the same cpumask appears both in the
6188 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6191 * The passed in 'doms_new' should be allocated using
6192 * alloc_sched_domains. This routine takes ownership of it and will
6193 * free_sched_domains it when done with it. If the caller failed the
6194 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6195 * and partition_sched_domains() will fallback to the single partition
6196 * 'fallback_doms', it also forces the domains to be rebuilt.
6198 * If doms_new == NULL it will be replaced with cpu_online_mask.
6199 * ndoms_new == 0 is a special case for destroying existing domains,
6200 * and it will not create the default domain.
6202 * Call with hotplug lock held
6204 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6205 struct sched_domain_attr
*dattr_new
)
6210 mutex_lock(&sched_domains_mutex
);
6212 /* always unregister in case we don't destroy any domains */
6213 unregister_sched_domain_sysctl();
6215 /* Let architecture update cpu core mappings. */
6216 new_topology
= arch_update_cpu_topology();
6218 n
= doms_new
? ndoms_new
: 0;
6220 /* Destroy deleted domains */
6221 for (i
= 0; i
< ndoms_cur
; i
++) {
6222 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6223 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6224 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6227 /* no match - a current sched domain not in new doms_new[] */
6228 detach_destroy_domains(doms_cur
[i
]);
6234 if (doms_new
== NULL
) {
6236 doms_new
= &fallback_doms
;
6237 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6238 WARN_ON_ONCE(dattr_new
);
6241 /* Build new domains */
6242 for (i
= 0; i
< ndoms_new
; i
++) {
6243 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6244 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6245 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6248 /* no match - add a new doms_new */
6249 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6254 /* Remember the new sched domains */
6255 if (doms_cur
!= &fallback_doms
)
6256 free_sched_domains(doms_cur
, ndoms_cur
);
6257 kfree(dattr_cur
); /* kfree(NULL) is safe */
6258 doms_cur
= doms_new
;
6259 dattr_cur
= dattr_new
;
6260 ndoms_cur
= ndoms_new
;
6262 register_sched_domain_sysctl();
6264 mutex_unlock(&sched_domains_mutex
);
6267 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6270 * Update cpusets according to cpu_active mask. If cpusets are
6271 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6272 * around partition_sched_domains().
6274 * If we come here as part of a suspend/resume, don't touch cpusets because we
6275 * want to restore it back to its original state upon resume anyway.
6277 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6281 case CPU_ONLINE_FROZEN
:
6282 case CPU_DOWN_FAILED_FROZEN
:
6285 * num_cpus_frozen tracks how many CPUs are involved in suspend
6286 * resume sequence. As long as this is not the last online
6287 * operation in the resume sequence, just build a single sched
6288 * domain, ignoring cpusets.
6291 if (likely(num_cpus_frozen
)) {
6292 partition_sched_domains(1, NULL
, NULL
);
6297 * This is the last CPU online operation. So fall through and
6298 * restore the original sched domains by considering the
6299 * cpuset configurations.
6303 case CPU_DOWN_FAILED
:
6304 cpuset_update_active_cpus(true);
6312 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6316 case CPU_DOWN_PREPARE
:
6317 cpuset_update_active_cpus(false);
6319 case CPU_DOWN_PREPARE_FROZEN
:
6321 partition_sched_domains(1, NULL
, NULL
);
6329 void __init
sched_init_smp(void)
6331 cpumask_var_t non_isolated_cpus
;
6333 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6334 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6339 mutex_lock(&sched_domains_mutex
);
6340 init_sched_domains(cpu_active_mask
);
6341 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6342 if (cpumask_empty(non_isolated_cpus
))
6343 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6344 mutex_unlock(&sched_domains_mutex
);
6347 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6348 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6349 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6353 /* Move init over to a non-isolated CPU */
6354 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6356 sched_init_granularity();
6357 free_cpumask_var(non_isolated_cpus
);
6359 init_sched_rt_class();
6362 void __init
sched_init_smp(void)
6364 sched_init_granularity();
6366 #endif /* CONFIG_SMP */
6368 const_debug
unsigned int sysctl_timer_migration
= 1;
6370 int in_sched_functions(unsigned long addr
)
6372 return in_lock_functions(addr
) ||
6373 (addr
>= (unsigned long)__sched_text_start
6374 && addr
< (unsigned long)__sched_text_end
);
6377 #ifdef CONFIG_CGROUP_SCHED
6379 * Default task group.
6380 * Every task in system belongs to this group at bootup.
6382 struct task_group root_task_group
;
6383 LIST_HEAD(task_groups
);
6386 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6388 void __init
sched_init(void)
6391 unsigned long alloc_size
= 0, ptr
;
6393 #ifdef CONFIG_FAIR_GROUP_SCHED
6394 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6396 #ifdef CONFIG_RT_GROUP_SCHED
6397 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6399 #ifdef CONFIG_CPUMASK_OFFSTACK
6400 alloc_size
+= num_possible_cpus() * cpumask_size();
6403 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6405 #ifdef CONFIG_FAIR_GROUP_SCHED
6406 root_task_group
.se
= (struct sched_entity
**)ptr
;
6407 ptr
+= nr_cpu_ids
* sizeof(void **);
6409 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6410 ptr
+= nr_cpu_ids
* sizeof(void **);
6412 #endif /* CONFIG_FAIR_GROUP_SCHED */
6413 #ifdef CONFIG_RT_GROUP_SCHED
6414 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6415 ptr
+= nr_cpu_ids
* sizeof(void **);
6417 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6418 ptr
+= nr_cpu_ids
* sizeof(void **);
6420 #endif /* CONFIG_RT_GROUP_SCHED */
6421 #ifdef CONFIG_CPUMASK_OFFSTACK
6422 for_each_possible_cpu(i
) {
6423 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6424 ptr
+= cpumask_size();
6426 #endif /* CONFIG_CPUMASK_OFFSTACK */
6430 init_defrootdomain();
6433 init_rt_bandwidth(&def_rt_bandwidth
,
6434 global_rt_period(), global_rt_runtime());
6436 #ifdef CONFIG_RT_GROUP_SCHED
6437 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6438 global_rt_period(), global_rt_runtime());
6439 #endif /* CONFIG_RT_GROUP_SCHED */
6441 #ifdef CONFIG_CGROUP_SCHED
6442 list_add(&root_task_group
.list
, &task_groups
);
6443 INIT_LIST_HEAD(&root_task_group
.children
);
6444 INIT_LIST_HEAD(&root_task_group
.siblings
);
6445 autogroup_init(&init_task
);
6447 #endif /* CONFIG_CGROUP_SCHED */
6449 for_each_possible_cpu(i
) {
6453 raw_spin_lock_init(&rq
->lock
);
6455 rq
->calc_load_active
= 0;
6456 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6457 init_cfs_rq(&rq
->cfs
);
6458 init_rt_rq(&rq
->rt
, rq
);
6459 #ifdef CONFIG_FAIR_GROUP_SCHED
6460 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6461 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6463 * How much cpu bandwidth does root_task_group get?
6465 * In case of task-groups formed thr' the cgroup filesystem, it
6466 * gets 100% of the cpu resources in the system. This overall
6467 * system cpu resource is divided among the tasks of
6468 * root_task_group and its child task-groups in a fair manner,
6469 * based on each entity's (task or task-group's) weight
6470 * (se->load.weight).
6472 * In other words, if root_task_group has 10 tasks of weight
6473 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6474 * then A0's share of the cpu resource is:
6476 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6478 * We achieve this by letting root_task_group's tasks sit
6479 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6481 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6482 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6483 #endif /* CONFIG_FAIR_GROUP_SCHED */
6485 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6486 #ifdef CONFIG_RT_GROUP_SCHED
6487 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6488 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6491 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6492 rq
->cpu_load
[j
] = 0;
6494 rq
->last_load_update_tick
= jiffies
;
6499 rq
->cpu_power
= SCHED_POWER_SCALE
;
6500 rq
->post_schedule
= 0;
6501 rq
->active_balance
= 0;
6502 rq
->next_balance
= jiffies
;
6507 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6508 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6510 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6512 rq_attach_root(rq
, &def_root_domain
);
6513 #ifdef CONFIG_NO_HZ_COMMON
6516 #ifdef CONFIG_NO_HZ_FULL
6517 rq
->last_sched_tick
= 0;
6521 atomic_set(&rq
->nr_iowait
, 0);
6524 set_load_weight(&init_task
);
6526 #ifdef CONFIG_PREEMPT_NOTIFIERS
6527 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6530 #ifdef CONFIG_RT_MUTEXES
6531 plist_head_init(&init_task
.pi_waiters
);
6535 * The boot idle thread does lazy MMU switching as well:
6537 atomic_inc(&init_mm
.mm_count
);
6538 enter_lazy_tlb(&init_mm
, current
);
6541 * Make us the idle thread. Technically, schedule() should not be
6542 * called from this thread, however somewhere below it might be,
6543 * but because we are the idle thread, we just pick up running again
6544 * when this runqueue becomes "idle".
6546 init_idle(current
, smp_processor_id());
6548 calc_load_update
= jiffies
+ LOAD_FREQ
;
6551 * During early bootup we pretend to be a normal task:
6553 current
->sched_class
= &fair_sched_class
;
6556 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6557 /* May be allocated at isolcpus cmdline parse time */
6558 if (cpu_isolated_map
== NULL
)
6559 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6560 idle_thread_set_boot_cpu();
6562 init_sched_fair_class();
6564 scheduler_running
= 1;
6567 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6568 static inline int preempt_count_equals(int preempt_offset
)
6570 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6572 return (nested
== preempt_offset
);
6575 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6577 static unsigned long prev_jiffy
; /* ratelimiting */
6579 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6580 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6581 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6583 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6585 prev_jiffy
= jiffies
;
6588 "BUG: sleeping function called from invalid context at %s:%d\n",
6591 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6592 in_atomic(), irqs_disabled(),
6593 current
->pid
, current
->comm
);
6595 debug_show_held_locks(current
);
6596 if (irqs_disabled())
6597 print_irqtrace_events(current
);
6600 EXPORT_SYMBOL(__might_sleep
);
6603 #ifdef CONFIG_MAGIC_SYSRQ
6604 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6606 const struct sched_class
*prev_class
= p
->sched_class
;
6607 int old_prio
= p
->prio
;
6612 dequeue_task(rq
, p
, 0);
6613 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6615 enqueue_task(rq
, p
, 0);
6616 resched_task(rq
->curr
);
6619 check_class_changed(rq
, p
, prev_class
, old_prio
);
6622 void normalize_rt_tasks(void)
6624 struct task_struct
*g
, *p
;
6625 unsigned long flags
;
6628 read_lock_irqsave(&tasklist_lock
, flags
);
6629 do_each_thread(g
, p
) {
6631 * Only normalize user tasks:
6636 p
->se
.exec_start
= 0;
6637 #ifdef CONFIG_SCHEDSTATS
6638 p
->se
.statistics
.wait_start
= 0;
6639 p
->se
.statistics
.sleep_start
= 0;
6640 p
->se
.statistics
.block_start
= 0;
6645 * Renice negative nice level userspace
6648 if (TASK_NICE(p
) < 0 && p
->mm
)
6649 set_user_nice(p
, 0);
6653 raw_spin_lock(&p
->pi_lock
);
6654 rq
= __task_rq_lock(p
);
6656 normalize_task(rq
, p
);
6658 __task_rq_unlock(rq
);
6659 raw_spin_unlock(&p
->pi_lock
);
6660 } while_each_thread(g
, p
);
6662 read_unlock_irqrestore(&tasklist_lock
, flags
);
6665 #endif /* CONFIG_MAGIC_SYSRQ */
6667 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6669 * These functions are only useful for the IA64 MCA handling, or kdb.
6671 * They can only be called when the whole system has been
6672 * stopped - every CPU needs to be quiescent, and no scheduling
6673 * activity can take place. Using them for anything else would
6674 * be a serious bug, and as a result, they aren't even visible
6675 * under any other configuration.
6679 * curr_task - return the current task for a given cpu.
6680 * @cpu: the processor in question.
6682 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6684 * Return: The current task for @cpu.
6686 struct task_struct
*curr_task(int cpu
)
6688 return cpu_curr(cpu
);
6691 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6695 * set_curr_task - set the current task for a given cpu.
6696 * @cpu: the processor in question.
6697 * @p: the task pointer to set.
6699 * Description: This function must only be used when non-maskable interrupts
6700 * are serviced on a separate stack. It allows the architecture to switch the
6701 * notion of the current task on a cpu in a non-blocking manner. This function
6702 * must be called with all CPU's synchronized, and interrupts disabled, the
6703 * and caller must save the original value of the current task (see
6704 * curr_task() above) and restore that value before reenabling interrupts and
6705 * re-starting the system.
6707 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6709 void set_curr_task(int cpu
, struct task_struct
*p
)
6716 #ifdef CONFIG_CGROUP_SCHED
6717 /* task_group_lock serializes the addition/removal of task groups */
6718 static DEFINE_SPINLOCK(task_group_lock
);
6720 static void free_sched_group(struct task_group
*tg
)
6722 free_fair_sched_group(tg
);
6723 free_rt_sched_group(tg
);
6728 /* allocate runqueue etc for a new task group */
6729 struct task_group
*sched_create_group(struct task_group
*parent
)
6731 struct task_group
*tg
;
6733 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6735 return ERR_PTR(-ENOMEM
);
6737 if (!alloc_fair_sched_group(tg
, parent
))
6740 if (!alloc_rt_sched_group(tg
, parent
))
6746 free_sched_group(tg
);
6747 return ERR_PTR(-ENOMEM
);
6750 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6752 unsigned long flags
;
6754 spin_lock_irqsave(&task_group_lock
, flags
);
6755 list_add_rcu(&tg
->list
, &task_groups
);
6757 WARN_ON(!parent
); /* root should already exist */
6759 tg
->parent
= parent
;
6760 INIT_LIST_HEAD(&tg
->children
);
6761 list_add_rcu(&tg
->siblings
, &parent
->children
);
6762 spin_unlock_irqrestore(&task_group_lock
, flags
);
6765 /* rcu callback to free various structures associated with a task group */
6766 static void free_sched_group_rcu(struct rcu_head
*rhp
)
6768 /* now it should be safe to free those cfs_rqs */
6769 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
6772 /* Destroy runqueue etc associated with a task group */
6773 void sched_destroy_group(struct task_group
*tg
)
6775 /* wait for possible concurrent references to cfs_rqs complete */
6776 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
6779 void sched_offline_group(struct task_group
*tg
)
6781 unsigned long flags
;
6784 /* end participation in shares distribution */
6785 for_each_possible_cpu(i
)
6786 unregister_fair_sched_group(tg
, i
);
6788 spin_lock_irqsave(&task_group_lock
, flags
);
6789 list_del_rcu(&tg
->list
);
6790 list_del_rcu(&tg
->siblings
);
6791 spin_unlock_irqrestore(&task_group_lock
, flags
);
6794 /* change task's runqueue when it moves between groups.
6795 * The caller of this function should have put the task in its new group
6796 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6797 * reflect its new group.
6799 void sched_move_task(struct task_struct
*tsk
)
6801 struct task_group
*tg
;
6803 unsigned long flags
;
6806 rq
= task_rq_lock(tsk
, &flags
);
6808 running
= task_current(rq
, tsk
);
6812 dequeue_task(rq
, tsk
, 0);
6813 if (unlikely(running
))
6814 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6816 tg
= container_of(task_css_check(tsk
, cpu_cgroup_subsys_id
,
6817 lockdep_is_held(&tsk
->sighand
->siglock
)),
6818 struct task_group
, css
);
6819 tg
= autogroup_task_group(tsk
, tg
);
6820 tsk
->sched_task_group
= tg
;
6822 #ifdef CONFIG_FAIR_GROUP_SCHED
6823 if (tsk
->sched_class
->task_move_group
)
6824 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
6827 set_task_rq(tsk
, task_cpu(tsk
));
6829 if (unlikely(running
))
6830 tsk
->sched_class
->set_curr_task(rq
);
6832 enqueue_task(rq
, tsk
, 0);
6834 task_rq_unlock(rq
, tsk
, &flags
);
6836 #endif /* CONFIG_CGROUP_SCHED */
6838 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6839 static unsigned long to_ratio(u64 period
, u64 runtime
)
6841 if (runtime
== RUNTIME_INF
)
6844 return div64_u64(runtime
<< 20, period
);
6848 #ifdef CONFIG_RT_GROUP_SCHED
6850 * Ensure that the real time constraints are schedulable.
6852 static DEFINE_MUTEX(rt_constraints_mutex
);
6854 /* Must be called with tasklist_lock held */
6855 static inline int tg_has_rt_tasks(struct task_group
*tg
)
6857 struct task_struct
*g
, *p
;
6859 do_each_thread(g
, p
) {
6860 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
6862 } while_each_thread(g
, p
);
6867 struct rt_schedulable_data
{
6868 struct task_group
*tg
;
6873 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
6875 struct rt_schedulable_data
*d
= data
;
6876 struct task_group
*child
;
6877 unsigned long total
, sum
= 0;
6878 u64 period
, runtime
;
6880 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6881 runtime
= tg
->rt_bandwidth
.rt_runtime
;
6884 period
= d
->rt_period
;
6885 runtime
= d
->rt_runtime
;
6889 * Cannot have more runtime than the period.
6891 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
6895 * Ensure we don't starve existing RT tasks.
6897 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
6900 total
= to_ratio(period
, runtime
);
6903 * Nobody can have more than the global setting allows.
6905 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
6909 * The sum of our children's runtime should not exceed our own.
6911 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
6912 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
6913 runtime
= child
->rt_bandwidth
.rt_runtime
;
6915 if (child
== d
->tg
) {
6916 period
= d
->rt_period
;
6917 runtime
= d
->rt_runtime
;
6920 sum
+= to_ratio(period
, runtime
);
6929 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
6933 struct rt_schedulable_data data
= {
6935 .rt_period
= period
,
6936 .rt_runtime
= runtime
,
6940 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
6946 static int tg_set_rt_bandwidth(struct task_group
*tg
,
6947 u64 rt_period
, u64 rt_runtime
)
6951 mutex_lock(&rt_constraints_mutex
);
6952 read_lock(&tasklist_lock
);
6953 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
6957 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6958 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
6959 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
6961 for_each_possible_cpu(i
) {
6962 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
6964 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6965 rt_rq
->rt_runtime
= rt_runtime
;
6966 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6968 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6970 read_unlock(&tasklist_lock
);
6971 mutex_unlock(&rt_constraints_mutex
);
6976 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
6978 u64 rt_runtime
, rt_period
;
6980 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6981 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
6982 if (rt_runtime_us
< 0)
6983 rt_runtime
= RUNTIME_INF
;
6985 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6988 static long sched_group_rt_runtime(struct task_group
*tg
)
6992 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
6995 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
6996 do_div(rt_runtime_us
, NSEC_PER_USEC
);
6997 return rt_runtime_us
;
7000 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7002 u64 rt_runtime
, rt_period
;
7004 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7005 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7010 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7013 static long sched_group_rt_period(struct task_group
*tg
)
7017 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7018 do_div(rt_period_us
, NSEC_PER_USEC
);
7019 return rt_period_us
;
7022 static int sched_rt_global_constraints(void)
7024 u64 runtime
, period
;
7027 if (sysctl_sched_rt_period
<= 0)
7030 runtime
= global_rt_runtime();
7031 period
= global_rt_period();
7034 * Sanity check on the sysctl variables.
7036 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7039 mutex_lock(&rt_constraints_mutex
);
7040 read_lock(&tasklist_lock
);
7041 ret
= __rt_schedulable(NULL
, 0, 0);
7042 read_unlock(&tasklist_lock
);
7043 mutex_unlock(&rt_constraints_mutex
);
7048 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7050 /* Don't accept realtime tasks when there is no way for them to run */
7051 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7057 #else /* !CONFIG_RT_GROUP_SCHED */
7058 static int sched_rt_global_constraints(void)
7060 unsigned long flags
;
7063 if (sysctl_sched_rt_period
<= 0)
7067 * There's always some RT tasks in the root group
7068 * -- migration, kstopmachine etc..
7070 if (sysctl_sched_rt_runtime
== 0)
7073 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7074 for_each_possible_cpu(i
) {
7075 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7077 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7078 rt_rq
->rt_runtime
= global_rt_runtime();
7079 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7081 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7085 #endif /* CONFIG_RT_GROUP_SCHED */
7087 int sched_rr_handler(struct ctl_table
*table
, int write
,
7088 void __user
*buffer
, size_t *lenp
,
7092 static DEFINE_MUTEX(mutex
);
7095 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7096 /* make sure that internally we keep jiffies */
7097 /* also, writing zero resets timeslice to default */
7098 if (!ret
&& write
) {
7099 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7100 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7102 mutex_unlock(&mutex
);
7106 int sched_rt_handler(struct ctl_table
*table
, int write
,
7107 void __user
*buffer
, size_t *lenp
,
7111 int old_period
, old_runtime
;
7112 static DEFINE_MUTEX(mutex
);
7115 old_period
= sysctl_sched_rt_period
;
7116 old_runtime
= sysctl_sched_rt_runtime
;
7118 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7120 if (!ret
&& write
) {
7121 ret
= sched_rt_global_constraints();
7123 sysctl_sched_rt_period
= old_period
;
7124 sysctl_sched_rt_runtime
= old_runtime
;
7126 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7127 def_rt_bandwidth
.rt_period
=
7128 ns_to_ktime(global_rt_period());
7131 mutex_unlock(&mutex
);
7136 #ifdef CONFIG_CGROUP_SCHED
7138 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7140 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7143 static struct cgroup_subsys_state
*
7144 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7146 struct task_group
*parent
= css_tg(parent_css
);
7147 struct task_group
*tg
;
7150 /* This is early initialization for the top cgroup */
7151 return &root_task_group
.css
;
7154 tg
= sched_create_group(parent
);
7156 return ERR_PTR(-ENOMEM
);
7161 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7163 struct task_group
*tg
= css_tg(css
);
7164 struct task_group
*parent
= css_tg(css_parent(css
));
7167 sched_online_group(tg
, parent
);
7171 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7173 struct task_group
*tg
= css_tg(css
);
7175 sched_destroy_group(tg
);
7178 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7180 struct task_group
*tg
= css_tg(css
);
7182 sched_offline_group(tg
);
7185 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7186 struct cgroup_taskset
*tset
)
7188 struct task_struct
*task
;
7190 cgroup_taskset_for_each(task
, css
, tset
) {
7191 #ifdef CONFIG_RT_GROUP_SCHED
7192 if (!sched_rt_can_attach(css_tg(css
), task
))
7195 /* We don't support RT-tasks being in separate groups */
7196 if (task
->sched_class
!= &fair_sched_class
)
7203 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7204 struct cgroup_taskset
*tset
)
7206 struct task_struct
*task
;
7208 cgroup_taskset_for_each(task
, css
, tset
)
7209 sched_move_task(task
);
7212 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7213 struct cgroup_subsys_state
*old_css
,
7214 struct task_struct
*task
)
7217 * cgroup_exit() is called in the copy_process() failure path.
7218 * Ignore this case since the task hasn't ran yet, this avoids
7219 * trying to poke a half freed task state from generic code.
7221 if (!(task
->flags
& PF_EXITING
))
7224 sched_move_task(task
);
7227 #ifdef CONFIG_FAIR_GROUP_SCHED
7228 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7229 struct cftype
*cftype
, u64 shareval
)
7231 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7234 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7237 struct task_group
*tg
= css_tg(css
);
7239 return (u64
) scale_load_down(tg
->shares
);
7242 #ifdef CONFIG_CFS_BANDWIDTH
7243 static DEFINE_MUTEX(cfs_constraints_mutex
);
7245 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7246 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7248 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7250 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7252 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7253 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7255 if (tg
== &root_task_group
)
7259 * Ensure we have at some amount of bandwidth every period. This is
7260 * to prevent reaching a state of large arrears when throttled via
7261 * entity_tick() resulting in prolonged exit starvation.
7263 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7267 * Likewise, bound things on the otherside by preventing insane quota
7268 * periods. This also allows us to normalize in computing quota
7271 if (period
> max_cfs_quota_period
)
7274 mutex_lock(&cfs_constraints_mutex
);
7275 ret
= __cfs_schedulable(tg
, period
, quota
);
7279 runtime_enabled
= quota
!= RUNTIME_INF
;
7280 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7281 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7282 raw_spin_lock_irq(&cfs_b
->lock
);
7283 cfs_b
->period
= ns_to_ktime(period
);
7284 cfs_b
->quota
= quota
;
7286 __refill_cfs_bandwidth_runtime(cfs_b
);
7287 /* restart the period timer (if active) to handle new period expiry */
7288 if (runtime_enabled
&& cfs_b
->timer_active
) {
7289 /* force a reprogram */
7290 cfs_b
->timer_active
= 0;
7291 __start_cfs_bandwidth(cfs_b
);
7293 raw_spin_unlock_irq(&cfs_b
->lock
);
7295 for_each_possible_cpu(i
) {
7296 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7297 struct rq
*rq
= cfs_rq
->rq
;
7299 raw_spin_lock_irq(&rq
->lock
);
7300 cfs_rq
->runtime_enabled
= runtime_enabled
;
7301 cfs_rq
->runtime_remaining
= 0;
7303 if (cfs_rq
->throttled
)
7304 unthrottle_cfs_rq(cfs_rq
);
7305 raw_spin_unlock_irq(&rq
->lock
);
7308 mutex_unlock(&cfs_constraints_mutex
);
7313 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7317 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7318 if (cfs_quota_us
< 0)
7319 quota
= RUNTIME_INF
;
7321 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7323 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7326 long tg_get_cfs_quota(struct task_group
*tg
)
7330 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7333 quota_us
= tg
->cfs_bandwidth
.quota
;
7334 do_div(quota_us
, NSEC_PER_USEC
);
7339 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7343 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7344 quota
= tg
->cfs_bandwidth
.quota
;
7346 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7349 long tg_get_cfs_period(struct task_group
*tg
)
7353 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7354 do_div(cfs_period_us
, NSEC_PER_USEC
);
7356 return cfs_period_us
;
7359 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7362 return tg_get_cfs_quota(css_tg(css
));
7365 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7366 struct cftype
*cftype
, s64 cfs_quota_us
)
7368 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7371 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7374 return tg_get_cfs_period(css_tg(css
));
7377 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7378 struct cftype
*cftype
, u64 cfs_period_us
)
7380 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7383 struct cfs_schedulable_data
{
7384 struct task_group
*tg
;
7389 * normalize group quota/period to be quota/max_period
7390 * note: units are usecs
7392 static u64
normalize_cfs_quota(struct task_group
*tg
,
7393 struct cfs_schedulable_data
*d
)
7401 period
= tg_get_cfs_period(tg
);
7402 quota
= tg_get_cfs_quota(tg
);
7405 /* note: these should typically be equivalent */
7406 if (quota
== RUNTIME_INF
|| quota
== -1)
7409 return to_ratio(period
, quota
);
7412 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7414 struct cfs_schedulable_data
*d
= data
;
7415 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7416 s64 quota
= 0, parent_quota
= -1;
7419 quota
= RUNTIME_INF
;
7421 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7423 quota
= normalize_cfs_quota(tg
, d
);
7424 parent_quota
= parent_b
->hierarchal_quota
;
7427 * ensure max(child_quota) <= parent_quota, inherit when no
7430 if (quota
== RUNTIME_INF
)
7431 quota
= parent_quota
;
7432 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7435 cfs_b
->hierarchal_quota
= quota
;
7440 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7443 struct cfs_schedulable_data data
= {
7449 if (quota
!= RUNTIME_INF
) {
7450 do_div(data
.period
, NSEC_PER_USEC
);
7451 do_div(data
.quota
, NSEC_PER_USEC
);
7455 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7461 static int cpu_stats_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
7462 struct cgroup_map_cb
*cb
)
7464 struct task_group
*tg
= css_tg(css
);
7465 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7467 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7468 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7469 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7473 #endif /* CONFIG_CFS_BANDWIDTH */
7474 #endif /* CONFIG_FAIR_GROUP_SCHED */
7476 #ifdef CONFIG_RT_GROUP_SCHED
7477 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7478 struct cftype
*cft
, s64 val
)
7480 return sched_group_set_rt_runtime(css_tg(css
), val
);
7483 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7486 return sched_group_rt_runtime(css_tg(css
));
7489 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7490 struct cftype
*cftype
, u64 rt_period_us
)
7492 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7495 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7498 return sched_group_rt_period(css_tg(css
));
7500 #endif /* CONFIG_RT_GROUP_SCHED */
7502 static struct cftype cpu_files
[] = {
7503 #ifdef CONFIG_FAIR_GROUP_SCHED
7506 .read_u64
= cpu_shares_read_u64
,
7507 .write_u64
= cpu_shares_write_u64
,
7510 #ifdef CONFIG_CFS_BANDWIDTH
7512 .name
= "cfs_quota_us",
7513 .read_s64
= cpu_cfs_quota_read_s64
,
7514 .write_s64
= cpu_cfs_quota_write_s64
,
7517 .name
= "cfs_period_us",
7518 .read_u64
= cpu_cfs_period_read_u64
,
7519 .write_u64
= cpu_cfs_period_write_u64
,
7523 .read_map
= cpu_stats_show
,
7526 #ifdef CONFIG_RT_GROUP_SCHED
7528 .name
= "rt_runtime_us",
7529 .read_s64
= cpu_rt_runtime_read
,
7530 .write_s64
= cpu_rt_runtime_write
,
7533 .name
= "rt_period_us",
7534 .read_u64
= cpu_rt_period_read_uint
,
7535 .write_u64
= cpu_rt_period_write_uint
,
7541 struct cgroup_subsys cpu_cgroup_subsys
= {
7543 .css_alloc
= cpu_cgroup_css_alloc
,
7544 .css_free
= cpu_cgroup_css_free
,
7545 .css_online
= cpu_cgroup_css_online
,
7546 .css_offline
= cpu_cgroup_css_offline
,
7547 .can_attach
= cpu_cgroup_can_attach
,
7548 .attach
= cpu_cgroup_attach
,
7549 .exit
= cpu_cgroup_exit
,
7550 .subsys_id
= cpu_cgroup_subsys_id
,
7551 .base_cftypes
= cpu_files
,
7555 #endif /* CONFIG_CGROUP_SCHED */
7557 void dump_cpu_task(int cpu
)
7559 pr_info("Task dump for CPU %d:\n", cpu
);
7560 sched_show_task(cpu_curr(cpu
));