4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
94 ktime_t soft
, hard
, now
;
97 if (hrtimer_active(period_timer
))
100 now
= hrtimer_cb_get_time(period_timer
);
101 hrtimer_forward(period_timer
, now
, period
);
103 soft
= hrtimer_get_softexpires(period_timer
);
104 hard
= hrtimer_get_expires(period_timer
);
105 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
106 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
107 HRTIMER_MODE_ABS_PINNED
, 0);
111 DEFINE_MUTEX(sched_domains_mutex
);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
114 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
116 void update_rq_clock(struct rq
*rq
)
120 if (rq
->skip_clock_update
> 0)
123 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
125 update_rq_clock_task(rq
, delta
);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug
unsigned int sysctl_sched_features
=
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static const char * const sched_feat_names
[] = {
146 #include "features.h"
151 static int sched_feat_show(struct seq_file
*m
, void *v
)
155 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
156 if (!(sysctl_sched_features
& (1UL << i
)))
158 seq_printf(m
, "%s ", sched_feat_names
[i
]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
174 #include "features.h"
179 static void sched_feat_disable(int i
)
181 if (static_key_enabled(&sched_feat_keys
[i
]))
182 static_key_slow_dec(&sched_feat_keys
[i
]);
185 static void sched_feat_enable(int i
)
187 if (!static_key_enabled(&sched_feat_keys
[i
]))
188 static_key_slow_inc(&sched_feat_keys
[i
]);
191 static void sched_feat_disable(int i
) { };
192 static void sched_feat_enable(int i
) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
197 size_t cnt
, loff_t
*ppos
)
207 if (copy_from_user(&buf
, ubuf
, cnt
))
213 if (strncmp(cmp
, "NO_", 3) == 0) {
218 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
219 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
221 sysctl_sched_features
&= ~(1UL << i
);
222 sched_feat_disable(i
);
224 sysctl_sched_features
|= (1UL << i
);
225 sched_feat_enable(i
);
231 if (i
== __SCHED_FEAT_NR
)
239 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
241 return single_open(filp
, sched_feat_show
, NULL
);
244 static const struct file_operations sched_feat_fops
= {
245 .open
= sched_feat_open
,
246 .write
= sched_feat_write
,
249 .release
= single_release
,
252 static __init
int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
259 late_initcall(sched_init_debug
);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
269 * period over which we average the RT time consumption, measured
274 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period
= 1000000;
282 __read_mostly
int scheduler_running
;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime
= 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
300 lockdep_assert_held(&p
->pi_lock
);
304 raw_spin_lock(&rq
->lock
);
305 if (likely(rq
== task_rq(p
)))
307 raw_spin_unlock(&rq
->lock
);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
315 __acquires(p
->pi_lock
)
321 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
323 raw_spin_lock(&rq
->lock
);
324 if (likely(rq
== task_rq(p
)))
326 raw_spin_unlock(&rq
->lock
);
327 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
331 static void __task_rq_unlock(struct rq
*rq
)
334 raw_spin_unlock(&rq
->lock
);
338 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
340 __releases(p
->pi_lock
)
342 raw_spin_unlock(&rq
->lock
);
343 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq
*this_rq_lock(void)
356 raw_spin_lock(&rq
->lock
);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq
*rq
)
375 if (hrtimer_active(&rq
->hrtick_timer
))
376 hrtimer_cancel(&rq
->hrtick_timer
);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
385 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
387 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
389 raw_spin_lock(&rq
->lock
);
391 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
392 raw_spin_unlock(&rq
->lock
);
394 return HRTIMER_NORESTART
;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg
)
405 raw_spin_lock(&rq
->lock
);
406 hrtimer_restart(&rq
->hrtick_timer
);
407 rq
->hrtick_csd_pending
= 0;
408 raw_spin_unlock(&rq
->lock
);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq
*rq
, u64 delay
)
418 struct hrtimer
*timer
= &rq
->hrtick_timer
;
419 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
421 hrtimer_set_expires(timer
, time
);
423 if (rq
== this_rq()) {
424 hrtimer_restart(timer
);
425 } else if (!rq
->hrtick_csd_pending
) {
426 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
427 rq
->hrtick_csd_pending
= 1;
432 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
434 int cpu
= (int)(long)hcpu
;
437 case CPU_UP_CANCELED
:
438 case CPU_UP_CANCELED_FROZEN
:
439 case CPU_DOWN_PREPARE
:
440 case CPU_DOWN_PREPARE_FROZEN
:
442 case CPU_DEAD_FROZEN
:
443 hrtick_clear(cpu_rq(cpu
));
450 static __init
void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick
, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq
*rq
, u64 delay
)
462 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
463 HRTIMER_MODE_REL_PINNED
, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq
*rq
)
474 rq
->hrtick_csd_pending
= 0;
476 rq
->hrtick_csd
.flags
= 0;
477 rq
->hrtick_csd
.func
= __hrtick_start
;
478 rq
->hrtick_csd
.info
= rq
;
481 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
482 rq
->hrtick_timer
.function
= hrtick
;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq
*rq
)
489 static inline void init_rq_hrtick(struct rq
*rq
)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct
*p
)
515 assert_raw_spin_locked(&task_rq(p
)->lock
);
517 if (test_tsk_need_resched(p
))
520 set_tsk_need_resched(p
);
523 if (cpu
== smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p
))
529 smp_send_reschedule(cpu
);
532 void resched_cpu(int cpu
)
534 struct rq
*rq
= cpu_rq(cpu
);
537 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
539 resched_task(cpu_curr(cpu
));
540 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu
= smp_processor_id();
556 struct sched_domain
*sd
;
559 for_each_domain(cpu
, sd
) {
560 for_each_cpu(i
, sched_domain_span(sd
)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu
)
583 struct rq
*rq
= cpu_rq(cpu
);
585 if (cpu
== smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq
->curr
!= rq
->idle
)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq
->idle
);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq
->idle
))
608 smp_send_reschedule(cpu
);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu
= smp_processor_id();
614 return idle_cpu(cpu
) && test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq
*rq
)
628 s64 period
= sched_avg_period();
630 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq
->age_stamp
));
637 rq
->age_stamp
+= period
;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct
*p
)
645 assert_raw_spin_locked(&task_rq(p
)->lock
);
646 set_tsk_need_resched(p
);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group
*from
,
659 tg_visitor down
, tg_visitor up
, void *data
)
661 struct task_group
*parent
, *child
;
667 ret
= (*down
)(parent
, data
);
670 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
677 ret
= (*up
)(parent
, data
);
678 if (ret
|| parent
== from
)
682 parent
= parent
->parent
;
689 int tg_nop(struct task_group
*tg
, void *data
)
695 static void set_load_weight(struct task_struct
*p
)
697 int prio
= p
->static_prio
- MAX_RT_PRIO
;
698 struct load_weight
*load
= &p
->se
.load
;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p
->policy
== SCHED_IDLE
) {
704 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
705 load
->inv_weight
= WMULT_IDLEPRIO
;
709 load
->weight
= scale_load(prio_to_weight
[prio
]);
710 load
->inv_weight
= prio_to_wmult
[prio
];
713 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
716 sched_info_queued(p
);
717 p
->sched_class
->enqueue_task(rq
, p
, flags
);
720 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
723 sched_info_dequeued(p
);
724 p
->sched_class
->dequeue_task(rq
, p
, flags
);
727 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
729 if (task_contributes_to_load(p
))
730 rq
->nr_uninterruptible
--;
732 enqueue_task(rq
, p
, flags
);
735 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
737 if (task_contributes_to_load(p
))
738 rq
->nr_uninterruptible
++;
740 dequeue_task(rq
, p
, flags
);
743 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
746 * In theory, the compile should just see 0 here, and optimize out the call
747 * to sched_rt_avg_update. But I don't trust it...
749 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
750 s64 steal
= 0, irq_delta
= 0;
752 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
753 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
756 * Since irq_time is only updated on {soft,}irq_exit, we might run into
757 * this case when a previous update_rq_clock() happened inside a
760 * When this happens, we stop ->clock_task and only update the
761 * prev_irq_time stamp to account for the part that fit, so that a next
762 * update will consume the rest. This ensures ->clock_task is
765 * It does however cause some slight miss-attribution of {soft,}irq
766 * time, a more accurate solution would be to update the irq_time using
767 * the current rq->clock timestamp, except that would require using
770 if (irq_delta
> delta
)
773 rq
->prev_irq_time
+= irq_delta
;
776 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
777 if (static_key_false((¶virt_steal_rq_enabled
))) {
780 steal
= paravirt_steal_clock(cpu_of(rq
));
781 steal
-= rq
->prev_steal_time_rq
;
783 if (unlikely(steal
> delta
))
786 st
= steal_ticks(steal
);
787 steal
= st
* TICK_NSEC
;
789 rq
->prev_steal_time_rq
+= steal
;
795 rq
->clock_task
+= delta
;
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
799 sched_rt_avg_update(rq
, irq_delta
+ steal
);
803 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
805 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
806 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
810 * Make it appear like a SCHED_FIFO task, its something
811 * userspace knows about and won't get confused about.
813 * Also, it will make PI more or less work without too
814 * much confusion -- but then, stop work should not
815 * rely on PI working anyway.
817 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
819 stop
->sched_class
= &stop_sched_class
;
822 cpu_rq(cpu
)->stop
= stop
;
826 * Reset it back to a normal scheduling class so that
827 * it can die in pieces.
829 old_stop
->sched_class
= &rt_sched_class
;
834 * __normal_prio - return the priority that is based on the static prio
836 static inline int __normal_prio(struct task_struct
*p
)
838 return p
->static_prio
;
842 * Calculate the expected normal priority: i.e. priority
843 * without taking RT-inheritance into account. Might be
844 * boosted by interactivity modifiers. Changes upon fork,
845 * setprio syscalls, and whenever the interactivity
846 * estimator recalculates.
848 static inline int normal_prio(struct task_struct
*p
)
852 if (task_has_rt_policy(p
))
853 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
855 prio
= __normal_prio(p
);
860 * Calculate the current priority, i.e. the priority
861 * taken into account by the scheduler. This value might
862 * be boosted by RT tasks, or might be boosted by
863 * interactivity modifiers. Will be RT if the task got
864 * RT-boosted. If not then it returns p->normal_prio.
866 static int effective_prio(struct task_struct
*p
)
868 p
->normal_prio
= normal_prio(p
);
870 * If we are RT tasks or we were boosted to RT priority,
871 * keep the priority unchanged. Otherwise, update priority
872 * to the normal priority:
874 if (!rt_prio(p
->prio
))
875 return p
->normal_prio
;
880 * task_curr - is this task currently executing on a CPU?
881 * @p: the task in question.
883 inline int task_curr(const struct task_struct
*p
)
885 return cpu_curr(task_cpu(p
)) == p
;
888 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
889 const struct sched_class
*prev_class
,
892 if (prev_class
!= p
->sched_class
) {
893 if (prev_class
->switched_from
)
894 prev_class
->switched_from(rq
, p
);
895 p
->sched_class
->switched_to(rq
, p
);
896 } else if (oldprio
!= p
->prio
)
897 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
900 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
902 const struct sched_class
*class;
904 if (p
->sched_class
== rq
->curr
->sched_class
) {
905 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
907 for_each_class(class) {
908 if (class == rq
->curr
->sched_class
)
910 if (class == p
->sched_class
) {
911 resched_task(rq
->curr
);
918 * A queue event has occurred, and we're going to schedule. In
919 * this case, we can save a useless back to back clock update.
921 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
922 rq
->skip_clock_update
= 1;
926 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
928 #ifdef CONFIG_SCHED_DEBUG
930 * We should never call set_task_cpu() on a blocked task,
931 * ttwu() will sort out the placement.
933 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
934 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
936 #ifdef CONFIG_LOCKDEP
938 * The caller should hold either p->pi_lock or rq->lock, when changing
939 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
941 * sched_move_task() holds both and thus holding either pins the cgroup,
944 * Furthermore, all task_rq users should acquire both locks, see
947 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
948 lockdep_is_held(&task_rq(p
)->lock
)));
952 trace_sched_migrate_task(p
, new_cpu
);
954 if (task_cpu(p
) != new_cpu
) {
955 p
->se
.nr_migrations
++;
956 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
959 __set_task_cpu(p
, new_cpu
);
962 struct migration_arg
{
963 struct task_struct
*task
;
967 static int migration_cpu_stop(void *data
);
970 * wait_task_inactive - wait for a thread to unschedule.
972 * If @match_state is nonzero, it's the @p->state value just checked and
973 * not expected to change. If it changes, i.e. @p might have woken up,
974 * then return zero. When we succeed in waiting for @p to be off its CPU,
975 * we return a positive number (its total switch count). If a second call
976 * a short while later returns the same number, the caller can be sure that
977 * @p has remained unscheduled the whole time.
979 * The caller must ensure that the task *will* unschedule sometime soon,
980 * else this function might spin for a *long* time. This function can't
981 * be called with interrupts off, or it may introduce deadlock with
982 * smp_call_function() if an IPI is sent by the same process we are
983 * waiting to become inactive.
985 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
994 * We do the initial early heuristics without holding
995 * any task-queue locks at all. We'll only try to get
996 * the runqueue lock when things look like they will
1002 * If the task is actively running on another CPU
1003 * still, just relax and busy-wait without holding
1006 * NOTE! Since we don't hold any locks, it's not
1007 * even sure that "rq" stays as the right runqueue!
1008 * But we don't care, since "task_running()" will
1009 * return false if the runqueue has changed and p
1010 * is actually now running somewhere else!
1012 while (task_running(rq
, p
)) {
1013 if (match_state
&& unlikely(p
->state
!= match_state
))
1019 * Ok, time to look more closely! We need the rq
1020 * lock now, to be *sure*. If we're wrong, we'll
1021 * just go back and repeat.
1023 rq
= task_rq_lock(p
, &flags
);
1024 trace_sched_wait_task(p
);
1025 running
= task_running(rq
, p
);
1028 if (!match_state
|| p
->state
== match_state
)
1029 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1030 task_rq_unlock(rq
, p
, &flags
);
1033 * If it changed from the expected state, bail out now.
1035 if (unlikely(!ncsw
))
1039 * Was it really running after all now that we
1040 * checked with the proper locks actually held?
1042 * Oops. Go back and try again..
1044 if (unlikely(running
)) {
1050 * It's not enough that it's not actively running,
1051 * it must be off the runqueue _entirely_, and not
1054 * So if it was still runnable (but just not actively
1055 * running right now), it's preempted, and we should
1056 * yield - it could be a while.
1058 if (unlikely(on_rq
)) {
1059 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1061 set_current_state(TASK_UNINTERRUPTIBLE
);
1062 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1067 * Ahh, all good. It wasn't running, and it wasn't
1068 * runnable, which means that it will never become
1069 * running in the future either. We're all done!
1078 * kick_process - kick a running thread to enter/exit the kernel
1079 * @p: the to-be-kicked thread
1081 * Cause a process which is running on another CPU to enter
1082 * kernel-mode, without any delay. (to get signals handled.)
1084 * NOTE: this function doesn't have to take the runqueue lock,
1085 * because all it wants to ensure is that the remote task enters
1086 * the kernel. If the IPI races and the task has been migrated
1087 * to another CPU then no harm is done and the purpose has been
1090 void kick_process(struct task_struct
*p
)
1096 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1097 smp_send_reschedule(cpu
);
1100 EXPORT_SYMBOL_GPL(kick_process
);
1101 #endif /* CONFIG_SMP */
1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1107 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1109 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
1110 enum { cpuset
, possible
, fail
} state
= cpuset
;
1113 /* Look for allowed, online CPU in same node. */
1114 for_each_cpu(dest_cpu
, nodemask
) {
1115 if (!cpu_online(dest_cpu
))
1117 if (!cpu_active(dest_cpu
))
1119 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1124 /* Any allowed, online CPU? */
1125 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1126 if (!cpu_online(dest_cpu
))
1128 if (!cpu_active(dest_cpu
))
1135 /* No more Mr. Nice Guy. */
1136 cpuset_cpus_allowed_fallback(p
);
1141 do_set_cpus_allowed(p
, cpu_possible_mask
);
1152 if (state
!= cpuset
) {
1154 * Don't tell them about moving exiting tasks or
1155 * kernel threads (both mm NULL), since they never
1158 if (p
->mm
&& printk_ratelimit()) {
1159 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1160 task_pid_nr(p
), p
->comm
, cpu
);
1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1171 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1173 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1176 * In order not to call set_task_cpu() on a blocking task we need
1177 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1180 * Since this is common to all placement strategies, this lives here.
1182 * [ this allows ->select_task() to simply return task_cpu(p) and
1183 * not worry about this generic constraint ]
1185 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1187 cpu
= select_fallback_rq(task_cpu(p
), p
);
1192 static void update_avg(u64
*avg
, u64 sample
)
1194 s64 diff
= sample
- *avg
;
1200 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1202 #ifdef CONFIG_SCHEDSTATS
1203 struct rq
*rq
= this_rq();
1206 int this_cpu
= smp_processor_id();
1208 if (cpu
== this_cpu
) {
1209 schedstat_inc(rq
, ttwu_local
);
1210 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1212 struct sched_domain
*sd
;
1214 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1216 for_each_domain(this_cpu
, sd
) {
1217 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1218 schedstat_inc(sd
, ttwu_wake_remote
);
1225 if (wake_flags
& WF_MIGRATED
)
1226 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1228 #endif /* CONFIG_SMP */
1230 schedstat_inc(rq
, ttwu_count
);
1231 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1233 if (wake_flags
& WF_SYNC
)
1234 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1236 #endif /* CONFIG_SCHEDSTATS */
1239 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1241 activate_task(rq
, p
, en_flags
);
1244 /* if a worker is waking up, notify workqueue */
1245 if (p
->flags
& PF_WQ_WORKER
)
1246 wq_worker_waking_up(p
, cpu_of(rq
));
1250 * Mark the task runnable and perform wakeup-preemption.
1253 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1255 trace_sched_wakeup(p
, true);
1256 check_preempt_curr(rq
, p
, wake_flags
);
1258 p
->state
= TASK_RUNNING
;
1260 if (p
->sched_class
->task_woken
)
1261 p
->sched_class
->task_woken(rq
, p
);
1263 if (rq
->idle_stamp
) {
1264 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1265 u64 max
= 2*sysctl_sched_migration_cost
;
1270 update_avg(&rq
->avg_idle
, delta
);
1277 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1280 if (p
->sched_contributes_to_load
)
1281 rq
->nr_uninterruptible
--;
1284 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1285 ttwu_do_wakeup(rq
, p
, wake_flags
);
1289 * Called in case the task @p isn't fully descheduled from its runqueue,
1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1291 * since all we need to do is flip p->state to TASK_RUNNING, since
1292 * the task is still ->on_rq.
1294 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1299 rq
= __task_rq_lock(p
);
1301 ttwu_do_wakeup(rq
, p
, wake_flags
);
1304 __task_rq_unlock(rq
);
1310 static void sched_ttwu_pending(void)
1312 struct rq
*rq
= this_rq();
1313 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1314 struct task_struct
*p
;
1316 raw_spin_lock(&rq
->lock
);
1319 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1320 llist
= llist_next(llist
);
1321 ttwu_do_activate(rq
, p
, 0);
1324 raw_spin_unlock(&rq
->lock
);
1327 void scheduler_ipi(void)
1329 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1333 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1334 * traditionally all their work was done from the interrupt return
1335 * path. Now that we actually do some work, we need to make sure
1338 * Some archs already do call them, luckily irq_enter/exit nest
1341 * Arguably we should visit all archs and update all handlers,
1342 * however a fair share of IPIs are still resched only so this would
1343 * somewhat pessimize the simple resched case.
1346 sched_ttwu_pending();
1349 * Check if someone kicked us for doing the nohz idle load balance.
1351 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1352 this_rq()->idle_balance
= 1;
1353 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1358 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1360 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1361 smp_send_reschedule(cpu
);
1364 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1366 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1368 #endif /* CONFIG_SMP */
1370 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1372 struct rq
*rq
= cpu_rq(cpu
);
1374 #if defined(CONFIG_SMP)
1375 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1376 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1377 ttwu_queue_remote(p
, cpu
);
1382 raw_spin_lock(&rq
->lock
);
1383 ttwu_do_activate(rq
, p
, 0);
1384 raw_spin_unlock(&rq
->lock
);
1388 * try_to_wake_up - wake up a thread
1389 * @p: the thread to be awakened
1390 * @state: the mask of task states that can be woken
1391 * @wake_flags: wake modifier flags (WF_*)
1393 * Put it on the run-queue if it's not already there. The "current"
1394 * thread is always on the run-queue (except when the actual
1395 * re-schedule is in progress), and as such you're allowed to do
1396 * the simpler "current->state = TASK_RUNNING" to mark yourself
1397 * runnable without the overhead of this.
1399 * Returns %true if @p was woken up, %false if it was already running
1400 * or @state didn't match @p's state.
1403 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1405 unsigned long flags
;
1406 int cpu
, success
= 0;
1409 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1410 if (!(p
->state
& state
))
1413 success
= 1; /* we're going to change ->state */
1416 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1421 * If the owning (remote) cpu is still in the middle of schedule() with
1422 * this task as prev, wait until its done referencing the task.
1427 * Pairs with the smp_wmb() in finish_lock_switch().
1431 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1432 p
->state
= TASK_WAKING
;
1434 if (p
->sched_class
->task_waking
)
1435 p
->sched_class
->task_waking(p
);
1437 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1438 if (task_cpu(p
) != cpu
) {
1439 wake_flags
|= WF_MIGRATED
;
1440 set_task_cpu(p
, cpu
);
1442 #endif /* CONFIG_SMP */
1446 ttwu_stat(p
, cpu
, wake_flags
);
1448 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1454 * try_to_wake_up_local - try to wake up a local task with rq lock held
1455 * @p: the thread to be awakened
1457 * Put @p on the run-queue if it's not already there. The caller must
1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1461 static void try_to_wake_up_local(struct task_struct
*p
)
1463 struct rq
*rq
= task_rq(p
);
1465 BUG_ON(rq
!= this_rq());
1466 BUG_ON(p
== current
);
1467 lockdep_assert_held(&rq
->lock
);
1469 if (!raw_spin_trylock(&p
->pi_lock
)) {
1470 raw_spin_unlock(&rq
->lock
);
1471 raw_spin_lock(&p
->pi_lock
);
1472 raw_spin_lock(&rq
->lock
);
1475 if (!(p
->state
& TASK_NORMAL
))
1479 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1481 ttwu_do_wakeup(rq
, p
, 0);
1482 ttwu_stat(p
, smp_processor_id(), 0);
1484 raw_spin_unlock(&p
->pi_lock
);
1488 * wake_up_process - Wake up a specific process
1489 * @p: The process to be woken up.
1491 * Attempt to wake up the nominated process and move it to the set of runnable
1492 * processes. Returns 1 if the process was woken up, 0 if it was already
1495 * It may be assumed that this function implies a write memory barrier before
1496 * changing the task state if and only if any tasks are woken up.
1498 int wake_up_process(struct task_struct
*p
)
1500 return try_to_wake_up(p
, TASK_ALL
, 0);
1502 EXPORT_SYMBOL(wake_up_process
);
1504 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1506 return try_to_wake_up(p
, state
, 0);
1510 * Perform scheduler related setup for a newly forked process p.
1511 * p is forked by current.
1513 * __sched_fork() is basic setup used by init_idle() too:
1515 static void __sched_fork(struct task_struct
*p
)
1520 p
->se
.exec_start
= 0;
1521 p
->se
.sum_exec_runtime
= 0;
1522 p
->se
.prev_sum_exec_runtime
= 0;
1523 p
->se
.nr_migrations
= 0;
1525 INIT_LIST_HEAD(&p
->se
.group_node
);
1527 #ifdef CONFIG_SCHEDSTATS
1528 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1531 INIT_LIST_HEAD(&p
->rt
.run_list
);
1533 #ifdef CONFIG_PREEMPT_NOTIFIERS
1534 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1539 * fork()/clone()-time setup:
1541 void sched_fork(struct task_struct
*p
)
1543 unsigned long flags
;
1544 int cpu
= get_cpu();
1548 * We mark the process as running here. This guarantees that
1549 * nobody will actually run it, and a signal or other external
1550 * event cannot wake it up and insert it on the runqueue either.
1552 p
->state
= TASK_RUNNING
;
1555 * Make sure we do not leak PI boosting priority to the child.
1557 p
->prio
= current
->normal_prio
;
1560 * Revert to default priority/policy on fork if requested.
1562 if (unlikely(p
->sched_reset_on_fork
)) {
1563 if (task_has_rt_policy(p
)) {
1564 p
->policy
= SCHED_NORMAL
;
1565 p
->static_prio
= NICE_TO_PRIO(0);
1567 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1568 p
->static_prio
= NICE_TO_PRIO(0);
1570 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1574 * We don't need the reset flag anymore after the fork. It has
1575 * fulfilled its duty:
1577 p
->sched_reset_on_fork
= 0;
1580 if (!rt_prio(p
->prio
))
1581 p
->sched_class
= &fair_sched_class
;
1583 if (p
->sched_class
->task_fork
)
1584 p
->sched_class
->task_fork(p
);
1587 * The child is not yet in the pid-hash so no cgroup attach races,
1588 * and the cgroup is pinned to this child due to cgroup_fork()
1589 * is ran before sched_fork().
1591 * Silence PROVE_RCU.
1593 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1594 set_task_cpu(p
, cpu
);
1595 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1597 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1598 if (likely(sched_info_on()))
1599 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1601 #if defined(CONFIG_SMP)
1604 #ifdef CONFIG_PREEMPT_COUNT
1605 /* Want to start with kernel preemption disabled. */
1606 task_thread_info(p
)->preempt_count
= 1;
1609 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1616 * wake_up_new_task - wake up a newly created task for the first time.
1618 * This function will do some initial scheduler statistics housekeeping
1619 * that must be done for every newly created context, then puts the task
1620 * on the runqueue and wakes it.
1622 void wake_up_new_task(struct task_struct
*p
)
1624 unsigned long flags
;
1627 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1630 * Fork balancing, do it here and not earlier because:
1631 * - cpus_allowed can change in the fork path
1632 * - any previously selected cpu might disappear through hotplug
1634 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1637 rq
= __task_rq_lock(p
);
1638 activate_task(rq
, p
, 0);
1640 trace_sched_wakeup_new(p
, true);
1641 check_preempt_curr(rq
, p
, WF_FORK
);
1643 if (p
->sched_class
->task_woken
)
1644 p
->sched_class
->task_woken(rq
, p
);
1646 task_rq_unlock(rq
, p
, &flags
);
1649 #ifdef CONFIG_PREEMPT_NOTIFIERS
1652 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1653 * @notifier: notifier struct to register
1655 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1657 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1659 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1662 * preempt_notifier_unregister - no longer interested in preemption notifications
1663 * @notifier: notifier struct to unregister
1665 * This is safe to call from within a preemption notifier.
1667 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1669 hlist_del(¬ifier
->link
);
1671 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1673 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1675 struct preempt_notifier
*notifier
;
1676 struct hlist_node
*node
;
1678 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1679 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1683 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1684 struct task_struct
*next
)
1686 struct preempt_notifier
*notifier
;
1687 struct hlist_node
*node
;
1689 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1690 notifier
->ops
->sched_out(notifier
, next
);
1693 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1695 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1700 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1701 struct task_struct
*next
)
1705 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1708 * prepare_task_switch - prepare to switch tasks
1709 * @rq: the runqueue preparing to switch
1710 * @prev: the current task that is being switched out
1711 * @next: the task we are going to switch to.
1713 * This is called with the rq lock held and interrupts off. It must
1714 * be paired with a subsequent finish_task_switch after the context
1717 * prepare_task_switch sets up locking and calls architecture specific
1721 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1722 struct task_struct
*next
)
1724 trace_sched_switch(prev
, next
);
1725 sched_info_switch(prev
, next
);
1726 perf_event_task_sched_out(prev
, next
);
1727 fire_sched_out_preempt_notifiers(prev
, next
);
1728 prepare_lock_switch(rq
, next
);
1729 prepare_arch_switch(next
);
1733 * finish_task_switch - clean up after a task-switch
1734 * @rq: runqueue associated with task-switch
1735 * @prev: the thread we just switched away from.
1737 * finish_task_switch must be called after the context switch, paired
1738 * with a prepare_task_switch call before the context switch.
1739 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1740 * and do any other architecture-specific cleanup actions.
1742 * Note that we may have delayed dropping an mm in context_switch(). If
1743 * so, we finish that here outside of the runqueue lock. (Doing it
1744 * with the lock held can cause deadlocks; see schedule() for
1747 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1748 __releases(rq
->lock
)
1750 struct mm_struct
*mm
= rq
->prev_mm
;
1756 * A task struct has one reference for the use as "current".
1757 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1758 * schedule one last time. The schedule call will never return, and
1759 * the scheduled task must drop that reference.
1760 * The test for TASK_DEAD must occur while the runqueue locks are
1761 * still held, otherwise prev could be scheduled on another cpu, die
1762 * there before we look at prev->state, and then the reference would
1764 * Manfred Spraul <manfred@colorfullife.com>
1766 prev_state
= prev
->state
;
1767 vtime_task_switch(prev
);
1768 finish_arch_switch(prev
);
1769 perf_event_task_sched_in(prev
, current
);
1770 finish_lock_switch(rq
, prev
);
1771 finish_arch_post_lock_switch();
1773 fire_sched_in_preempt_notifiers(current
);
1776 if (unlikely(prev_state
== TASK_DEAD
)) {
1778 * Remove function-return probe instances associated with this
1779 * task and put them back on the free list.
1781 kprobe_flush_task(prev
);
1782 put_task_struct(prev
);
1788 /* assumes rq->lock is held */
1789 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1791 if (prev
->sched_class
->pre_schedule
)
1792 prev
->sched_class
->pre_schedule(rq
, prev
);
1795 /* rq->lock is NOT held, but preemption is disabled */
1796 static inline void post_schedule(struct rq
*rq
)
1798 if (rq
->post_schedule
) {
1799 unsigned long flags
;
1801 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1802 if (rq
->curr
->sched_class
->post_schedule
)
1803 rq
->curr
->sched_class
->post_schedule(rq
);
1804 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1806 rq
->post_schedule
= 0;
1812 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1816 static inline void post_schedule(struct rq
*rq
)
1823 * schedule_tail - first thing a freshly forked thread must call.
1824 * @prev: the thread we just switched away from.
1826 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1827 __releases(rq
->lock
)
1829 struct rq
*rq
= this_rq();
1831 finish_task_switch(rq
, prev
);
1834 * FIXME: do we need to worry about rq being invalidated by the
1839 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1840 /* In this case, finish_task_switch does not reenable preemption */
1843 if (current
->set_child_tid
)
1844 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1848 * context_switch - switch to the new MM and the new
1849 * thread's register state.
1852 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1853 struct task_struct
*next
)
1855 struct mm_struct
*mm
, *oldmm
;
1857 prepare_task_switch(rq
, prev
, next
);
1860 oldmm
= prev
->active_mm
;
1862 * For paravirt, this is coupled with an exit in switch_to to
1863 * combine the page table reload and the switch backend into
1866 arch_start_context_switch(prev
);
1869 next
->active_mm
= oldmm
;
1870 atomic_inc(&oldmm
->mm_count
);
1871 enter_lazy_tlb(oldmm
, next
);
1873 switch_mm(oldmm
, mm
, next
);
1876 prev
->active_mm
= NULL
;
1877 rq
->prev_mm
= oldmm
;
1880 * Since the runqueue lock will be released by the next
1881 * task (which is an invalid locking op but in the case
1882 * of the scheduler it's an obvious special-case), so we
1883 * do an early lockdep release here:
1885 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1886 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1889 /* Here we just switch the register state and the stack. */
1890 rcu_switch(prev
, next
);
1891 switch_to(prev
, next
, prev
);
1895 * this_rq must be evaluated again because prev may have moved
1896 * CPUs since it called schedule(), thus the 'rq' on its stack
1897 * frame will be invalid.
1899 finish_task_switch(this_rq(), prev
);
1903 * nr_running, nr_uninterruptible and nr_context_switches:
1905 * externally visible scheduler statistics: current number of runnable
1906 * threads, current number of uninterruptible-sleeping threads, total
1907 * number of context switches performed since bootup.
1909 unsigned long nr_running(void)
1911 unsigned long i
, sum
= 0;
1913 for_each_online_cpu(i
)
1914 sum
+= cpu_rq(i
)->nr_running
;
1919 unsigned long nr_uninterruptible(void)
1921 unsigned long i
, sum
= 0;
1923 for_each_possible_cpu(i
)
1924 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1927 * Since we read the counters lockless, it might be slightly
1928 * inaccurate. Do not allow it to go below zero though:
1930 if (unlikely((long)sum
< 0))
1936 unsigned long long nr_context_switches(void)
1939 unsigned long long sum
= 0;
1941 for_each_possible_cpu(i
)
1942 sum
+= cpu_rq(i
)->nr_switches
;
1947 unsigned long nr_iowait(void)
1949 unsigned long i
, sum
= 0;
1951 for_each_possible_cpu(i
)
1952 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1957 unsigned long nr_iowait_cpu(int cpu
)
1959 struct rq
*this = cpu_rq(cpu
);
1960 return atomic_read(&this->nr_iowait
);
1963 unsigned long this_cpu_load(void)
1965 struct rq
*this = this_rq();
1966 return this->cpu_load
[0];
1971 * Global load-average calculations
1973 * We take a distributed and async approach to calculating the global load-avg
1974 * in order to minimize overhead.
1976 * The global load average is an exponentially decaying average of nr_running +
1977 * nr_uninterruptible.
1979 * Once every LOAD_FREQ:
1982 * for_each_possible_cpu(cpu)
1983 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1985 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
1987 * Due to a number of reasons the above turns in the mess below:
1989 * - for_each_possible_cpu() is prohibitively expensive on machines with
1990 * serious number of cpus, therefore we need to take a distributed approach
1991 * to calculating nr_active.
1993 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
1994 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
1996 * So assuming nr_active := 0 when we start out -- true per definition, we
1997 * can simply take per-cpu deltas and fold those into a global accumulate
1998 * to obtain the same result. See calc_load_fold_active().
2000 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2001 * across the machine, we assume 10 ticks is sufficient time for every
2002 * cpu to have completed this task.
2004 * This places an upper-bound on the IRQ-off latency of the machine. Then
2005 * again, being late doesn't loose the delta, just wrecks the sample.
2007 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2008 * this would add another cross-cpu cacheline miss and atomic operation
2009 * to the wakeup path. Instead we increment on whatever cpu the task ran
2010 * when it went into uninterruptible state and decrement on whatever cpu
2011 * did the wakeup. This means that only the sum of nr_uninterruptible over
2012 * all cpus yields the correct result.
2014 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2017 /* Variables and functions for calc_load */
2018 static atomic_long_t calc_load_tasks
;
2019 static unsigned long calc_load_update
;
2020 unsigned long avenrun
[3];
2021 EXPORT_SYMBOL(avenrun
); /* should be removed */
2024 * get_avenrun - get the load average array
2025 * @loads: pointer to dest load array
2026 * @offset: offset to add
2027 * @shift: shift count to shift the result left
2029 * These values are estimates at best, so no need for locking.
2031 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2033 loads
[0] = (avenrun
[0] + offset
) << shift
;
2034 loads
[1] = (avenrun
[1] + offset
) << shift
;
2035 loads
[2] = (avenrun
[2] + offset
) << shift
;
2038 static long calc_load_fold_active(struct rq
*this_rq
)
2040 long nr_active
, delta
= 0;
2042 nr_active
= this_rq
->nr_running
;
2043 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2045 if (nr_active
!= this_rq
->calc_load_active
) {
2046 delta
= nr_active
- this_rq
->calc_load_active
;
2047 this_rq
->calc_load_active
= nr_active
;
2054 * a1 = a0 * e + a * (1 - e)
2056 static unsigned long
2057 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2060 load
+= active
* (FIXED_1
- exp
);
2061 load
+= 1UL << (FSHIFT
- 1);
2062 return load
>> FSHIFT
;
2067 * Handle NO_HZ for the global load-average.
2069 * Since the above described distributed algorithm to compute the global
2070 * load-average relies on per-cpu sampling from the tick, it is affected by
2073 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2074 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2075 * when we read the global state.
2077 * Obviously reality has to ruin such a delightfully simple scheme:
2079 * - When we go NO_HZ idle during the window, we can negate our sample
2080 * contribution, causing under-accounting.
2082 * We avoid this by keeping two idle-delta counters and flipping them
2083 * when the window starts, thus separating old and new NO_HZ load.
2085 * The only trick is the slight shift in index flip for read vs write.
2089 * |-|-----------|-|-----------|-|-----------|-|
2090 * r:0 0 1 1 0 0 1 1 0
2091 * w:0 1 1 0 0 1 1 0 0
2093 * This ensures we'll fold the old idle contribution in this window while
2094 * accumlating the new one.
2096 * - When we wake up from NO_HZ idle during the window, we push up our
2097 * contribution, since we effectively move our sample point to a known
2100 * This is solved by pushing the window forward, and thus skipping the
2101 * sample, for this cpu (effectively using the idle-delta for this cpu which
2102 * was in effect at the time the window opened). This also solves the issue
2103 * of having to deal with a cpu having been in NOHZ idle for multiple
2104 * LOAD_FREQ intervals.
2106 * When making the ILB scale, we should try to pull this in as well.
2108 static atomic_long_t calc_load_idle
[2];
2109 static int calc_load_idx
;
2111 static inline int calc_load_write_idx(void)
2113 int idx
= calc_load_idx
;
2116 * See calc_global_nohz(), if we observe the new index, we also
2117 * need to observe the new update time.
2122 * If the folding window started, make sure we start writing in the
2125 if (!time_before(jiffies
, calc_load_update
))
2131 static inline int calc_load_read_idx(void)
2133 return calc_load_idx
& 1;
2136 void calc_load_enter_idle(void)
2138 struct rq
*this_rq
= this_rq();
2142 * We're going into NOHZ mode, if there's any pending delta, fold it
2143 * into the pending idle delta.
2145 delta
= calc_load_fold_active(this_rq
);
2147 int idx
= calc_load_write_idx();
2148 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2152 void calc_load_exit_idle(void)
2154 struct rq
*this_rq
= this_rq();
2157 * If we're still before the sample window, we're done.
2159 if (time_before(jiffies
, this_rq
->calc_load_update
))
2163 * We woke inside or after the sample window, this means we're already
2164 * accounted through the nohz accounting, so skip the entire deal and
2165 * sync up for the next window.
2167 this_rq
->calc_load_update
= calc_load_update
;
2168 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2169 this_rq
->calc_load_update
+= LOAD_FREQ
;
2172 static long calc_load_fold_idle(void)
2174 int idx
= calc_load_read_idx();
2177 if (atomic_long_read(&calc_load_idle
[idx
]))
2178 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2184 * fixed_power_int - compute: x^n, in O(log n) time
2186 * @x: base of the power
2187 * @frac_bits: fractional bits of @x
2188 * @n: power to raise @x to.
2190 * By exploiting the relation between the definition of the natural power
2191 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2192 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2193 * (where: n_i \elem {0, 1}, the binary vector representing n),
2194 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2195 * of course trivially computable in O(log_2 n), the length of our binary
2198 static unsigned long
2199 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2201 unsigned long result
= 1UL << frac_bits
;
2206 result
+= 1UL << (frac_bits
- 1);
2207 result
>>= frac_bits
;
2213 x
+= 1UL << (frac_bits
- 1);
2221 * a1 = a0 * e + a * (1 - e)
2223 * a2 = a1 * e + a * (1 - e)
2224 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2225 * = a0 * e^2 + a * (1 - e) * (1 + e)
2227 * a3 = a2 * e + a * (1 - e)
2228 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2229 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2233 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2234 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2235 * = a0 * e^n + a * (1 - e^n)
2237 * [1] application of the geometric series:
2240 * S_n := \Sum x^i = -------------
2243 static unsigned long
2244 calc_load_n(unsigned long load
, unsigned long exp
,
2245 unsigned long active
, unsigned int n
)
2248 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2252 * NO_HZ can leave us missing all per-cpu ticks calling
2253 * calc_load_account_active(), but since an idle CPU folds its delta into
2254 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2255 * in the pending idle delta if our idle period crossed a load cycle boundary.
2257 * Once we've updated the global active value, we need to apply the exponential
2258 * weights adjusted to the number of cycles missed.
2260 static void calc_global_nohz(void)
2262 long delta
, active
, n
;
2264 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2266 * Catch-up, fold however many we are behind still
2268 delta
= jiffies
- calc_load_update
- 10;
2269 n
= 1 + (delta
/ LOAD_FREQ
);
2271 active
= atomic_long_read(&calc_load_tasks
);
2272 active
= active
> 0 ? active
* FIXED_1
: 0;
2274 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2275 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2276 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2278 calc_load_update
+= n
* LOAD_FREQ
;
2282 * Flip the idle index...
2284 * Make sure we first write the new time then flip the index, so that
2285 * calc_load_write_idx() will see the new time when it reads the new
2286 * index, this avoids a double flip messing things up.
2291 #else /* !CONFIG_NO_HZ */
2293 static inline long calc_load_fold_idle(void) { return 0; }
2294 static inline void calc_global_nohz(void) { }
2296 #endif /* CONFIG_NO_HZ */
2299 * calc_load - update the avenrun load estimates 10 ticks after the
2300 * CPUs have updated calc_load_tasks.
2302 void calc_global_load(unsigned long ticks
)
2306 if (time_before(jiffies
, calc_load_update
+ 10))
2310 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2312 delta
= calc_load_fold_idle();
2314 atomic_long_add(delta
, &calc_load_tasks
);
2316 active
= atomic_long_read(&calc_load_tasks
);
2317 active
= active
> 0 ? active
* FIXED_1
: 0;
2319 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2320 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2321 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2323 calc_load_update
+= LOAD_FREQ
;
2326 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2332 * Called from update_cpu_load() to periodically update this CPU's
2335 static void calc_load_account_active(struct rq
*this_rq
)
2339 if (time_before(jiffies
, this_rq
->calc_load_update
))
2342 delta
= calc_load_fold_active(this_rq
);
2344 atomic_long_add(delta
, &calc_load_tasks
);
2346 this_rq
->calc_load_update
+= LOAD_FREQ
;
2350 * End of global load-average stuff
2354 * The exact cpuload at various idx values, calculated at every tick would be
2355 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2357 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2358 * on nth tick when cpu may be busy, then we have:
2359 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2360 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2362 * decay_load_missed() below does efficient calculation of
2363 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2364 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2366 * The calculation is approximated on a 128 point scale.
2367 * degrade_zero_ticks is the number of ticks after which load at any
2368 * particular idx is approximated to be zero.
2369 * degrade_factor is a precomputed table, a row for each load idx.
2370 * Each column corresponds to degradation factor for a power of two ticks,
2371 * based on 128 point scale.
2373 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2374 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2376 * With this power of 2 load factors, we can degrade the load n times
2377 * by looking at 1 bits in n and doing as many mult/shift instead of
2378 * n mult/shifts needed by the exact degradation.
2380 #define DEGRADE_SHIFT 7
2381 static const unsigned char
2382 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2383 static const unsigned char
2384 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2385 {0, 0, 0, 0, 0, 0, 0, 0},
2386 {64, 32, 8, 0, 0, 0, 0, 0},
2387 {96, 72, 40, 12, 1, 0, 0},
2388 {112, 98, 75, 43, 15, 1, 0},
2389 {120, 112, 98, 76, 45, 16, 2} };
2392 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2393 * would be when CPU is idle and so we just decay the old load without
2394 * adding any new load.
2396 static unsigned long
2397 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2401 if (!missed_updates
)
2404 if (missed_updates
>= degrade_zero_ticks
[idx
])
2408 return load
>> missed_updates
;
2410 while (missed_updates
) {
2411 if (missed_updates
% 2)
2412 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2414 missed_updates
>>= 1;
2421 * Update rq->cpu_load[] statistics. This function is usually called every
2422 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2423 * every tick. We fix it up based on jiffies.
2425 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2426 unsigned long pending_updates
)
2430 this_rq
->nr_load_updates
++;
2432 /* Update our load: */
2433 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2434 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2435 unsigned long old_load
, new_load
;
2437 /* scale is effectively 1 << i now, and >> i divides by scale */
2439 old_load
= this_rq
->cpu_load
[i
];
2440 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2441 new_load
= this_load
;
2443 * Round up the averaging division if load is increasing. This
2444 * prevents us from getting stuck on 9 if the load is 10, for
2447 if (new_load
> old_load
)
2448 new_load
+= scale
- 1;
2450 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2453 sched_avg_update(this_rq
);
2458 * There is no sane way to deal with nohz on smp when using jiffies because the
2459 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2460 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2462 * Therefore we cannot use the delta approach from the regular tick since that
2463 * would seriously skew the load calculation. However we'll make do for those
2464 * updates happening while idle (nohz_idle_balance) or coming out of idle
2465 * (tick_nohz_idle_exit).
2467 * This means we might still be one tick off for nohz periods.
2471 * Called from nohz_idle_balance() to update the load ratings before doing the
2474 void update_idle_cpu_load(struct rq
*this_rq
)
2476 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2477 unsigned long load
= this_rq
->load
.weight
;
2478 unsigned long pending_updates
;
2481 * bail if there's load or we're actually up-to-date.
2483 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2486 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2487 this_rq
->last_load_update_tick
= curr_jiffies
;
2489 __update_cpu_load(this_rq
, load
, pending_updates
);
2493 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2495 void update_cpu_load_nohz(void)
2497 struct rq
*this_rq
= this_rq();
2498 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2499 unsigned long pending_updates
;
2501 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2504 raw_spin_lock(&this_rq
->lock
);
2505 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2506 if (pending_updates
) {
2507 this_rq
->last_load_update_tick
= curr_jiffies
;
2509 * We were idle, this means load 0, the current load might be
2510 * !0 due to remote wakeups and the sort.
2512 __update_cpu_load(this_rq
, 0, pending_updates
);
2514 raw_spin_unlock(&this_rq
->lock
);
2516 #endif /* CONFIG_NO_HZ */
2519 * Called from scheduler_tick()
2521 static void update_cpu_load_active(struct rq
*this_rq
)
2524 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2526 this_rq
->last_load_update_tick
= jiffies
;
2527 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2529 calc_load_account_active(this_rq
);
2535 * sched_exec - execve() is a valuable balancing opportunity, because at
2536 * this point the task has the smallest effective memory and cache footprint.
2538 void sched_exec(void)
2540 struct task_struct
*p
= current
;
2541 unsigned long flags
;
2544 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2545 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2546 if (dest_cpu
== smp_processor_id())
2549 if (likely(cpu_active(dest_cpu
))) {
2550 struct migration_arg arg
= { p
, dest_cpu
};
2552 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2553 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2557 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2562 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2563 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2565 EXPORT_PER_CPU_SYMBOL(kstat
);
2566 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2569 * Return any ns on the sched_clock that have not yet been accounted in
2570 * @p in case that task is currently running.
2572 * Called with task_rq_lock() held on @rq.
2574 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2578 if (task_current(rq
, p
)) {
2579 update_rq_clock(rq
);
2580 ns
= rq
->clock_task
- p
->se
.exec_start
;
2588 unsigned long long task_delta_exec(struct task_struct
*p
)
2590 unsigned long flags
;
2594 rq
= task_rq_lock(p
, &flags
);
2595 ns
= do_task_delta_exec(p
, rq
);
2596 task_rq_unlock(rq
, p
, &flags
);
2602 * Return accounted runtime for the task.
2603 * In case the task is currently running, return the runtime plus current's
2604 * pending runtime that have not been accounted yet.
2606 unsigned long long task_sched_runtime(struct task_struct
*p
)
2608 unsigned long flags
;
2612 rq
= task_rq_lock(p
, &flags
);
2613 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2614 task_rq_unlock(rq
, p
, &flags
);
2620 * This function gets called by the timer code, with HZ frequency.
2621 * We call it with interrupts disabled.
2623 void scheduler_tick(void)
2625 int cpu
= smp_processor_id();
2626 struct rq
*rq
= cpu_rq(cpu
);
2627 struct task_struct
*curr
= rq
->curr
;
2631 raw_spin_lock(&rq
->lock
);
2632 update_rq_clock(rq
);
2633 update_cpu_load_active(rq
);
2634 curr
->sched_class
->task_tick(rq
, curr
, 0);
2635 raw_spin_unlock(&rq
->lock
);
2637 perf_event_task_tick();
2640 rq
->idle_balance
= idle_cpu(cpu
);
2641 trigger_load_balance(rq
, cpu
);
2645 notrace
unsigned long get_parent_ip(unsigned long addr
)
2647 if (in_lock_functions(addr
)) {
2648 addr
= CALLER_ADDR2
;
2649 if (in_lock_functions(addr
))
2650 addr
= CALLER_ADDR3
;
2655 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2656 defined(CONFIG_PREEMPT_TRACER))
2658 void __kprobes
add_preempt_count(int val
)
2660 #ifdef CONFIG_DEBUG_PREEMPT
2664 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2667 preempt_count() += val
;
2668 #ifdef CONFIG_DEBUG_PREEMPT
2670 * Spinlock count overflowing soon?
2672 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2675 if (preempt_count() == val
)
2676 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2678 EXPORT_SYMBOL(add_preempt_count
);
2680 void __kprobes
sub_preempt_count(int val
)
2682 #ifdef CONFIG_DEBUG_PREEMPT
2686 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2689 * Is the spinlock portion underflowing?
2691 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2692 !(preempt_count() & PREEMPT_MASK
)))
2696 if (preempt_count() == val
)
2697 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2698 preempt_count() -= val
;
2700 EXPORT_SYMBOL(sub_preempt_count
);
2705 * Print scheduling while atomic bug:
2707 static noinline
void __schedule_bug(struct task_struct
*prev
)
2709 if (oops_in_progress
)
2712 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2713 prev
->comm
, prev
->pid
, preempt_count());
2715 debug_show_held_locks(prev
);
2717 if (irqs_disabled())
2718 print_irqtrace_events(prev
);
2720 add_taint(TAINT_WARN
);
2724 * Various schedule()-time debugging checks and statistics:
2726 static inline void schedule_debug(struct task_struct
*prev
)
2729 * Test if we are atomic. Since do_exit() needs to call into
2730 * schedule() atomically, we ignore that path for now.
2731 * Otherwise, whine if we are scheduling when we should not be.
2733 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2734 __schedule_bug(prev
);
2737 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2739 schedstat_inc(this_rq(), sched_count
);
2742 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2744 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2745 update_rq_clock(rq
);
2746 prev
->sched_class
->put_prev_task(rq
, prev
);
2750 * Pick up the highest-prio task:
2752 static inline struct task_struct
*
2753 pick_next_task(struct rq
*rq
)
2755 const struct sched_class
*class;
2756 struct task_struct
*p
;
2759 * Optimization: we know that if all tasks are in
2760 * the fair class we can call that function directly:
2762 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2763 p
= fair_sched_class
.pick_next_task(rq
);
2768 for_each_class(class) {
2769 p
= class->pick_next_task(rq
);
2774 BUG(); /* the idle class will always have a runnable task */
2778 * __schedule() is the main scheduler function.
2780 * The main means of driving the scheduler and thus entering this function are:
2782 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2784 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2785 * paths. For example, see arch/x86/entry_64.S.
2787 * To drive preemption between tasks, the scheduler sets the flag in timer
2788 * interrupt handler scheduler_tick().
2790 * 3. Wakeups don't really cause entry into schedule(). They add a
2791 * task to the run-queue and that's it.
2793 * Now, if the new task added to the run-queue preempts the current
2794 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2795 * called on the nearest possible occasion:
2797 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2799 * - in syscall or exception context, at the next outmost
2800 * preempt_enable(). (this might be as soon as the wake_up()'s
2803 * - in IRQ context, return from interrupt-handler to
2804 * preemptible context
2806 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2809 * - cond_resched() call
2810 * - explicit schedule() call
2811 * - return from syscall or exception to user-space
2812 * - return from interrupt-handler to user-space
2814 static void __sched
__schedule(void)
2816 struct task_struct
*prev
, *next
;
2817 unsigned long *switch_count
;
2823 cpu
= smp_processor_id();
2825 rcu_note_context_switch(cpu
);
2828 schedule_debug(prev
);
2830 if (sched_feat(HRTICK
))
2833 raw_spin_lock_irq(&rq
->lock
);
2835 switch_count
= &prev
->nivcsw
;
2836 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2837 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2838 prev
->state
= TASK_RUNNING
;
2840 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2844 * If a worker went to sleep, notify and ask workqueue
2845 * whether it wants to wake up a task to maintain
2848 if (prev
->flags
& PF_WQ_WORKER
) {
2849 struct task_struct
*to_wakeup
;
2851 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2853 try_to_wake_up_local(to_wakeup
);
2856 switch_count
= &prev
->nvcsw
;
2859 pre_schedule(rq
, prev
);
2861 if (unlikely(!rq
->nr_running
))
2862 idle_balance(cpu
, rq
);
2864 put_prev_task(rq
, prev
);
2865 next
= pick_next_task(rq
);
2866 clear_tsk_need_resched(prev
);
2867 rq
->skip_clock_update
= 0;
2869 if (likely(prev
!= next
)) {
2874 context_switch(rq
, prev
, next
); /* unlocks the rq */
2876 * The context switch have flipped the stack from under us
2877 * and restored the local variables which were saved when
2878 * this task called schedule() in the past. prev == current
2879 * is still correct, but it can be moved to another cpu/rq.
2881 cpu
= smp_processor_id();
2884 raw_spin_unlock_irq(&rq
->lock
);
2888 sched_preempt_enable_no_resched();
2893 static inline void sched_submit_work(struct task_struct
*tsk
)
2895 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2898 * If we are going to sleep and we have plugged IO queued,
2899 * make sure to submit it to avoid deadlocks.
2901 if (blk_needs_flush_plug(tsk
))
2902 blk_schedule_flush_plug(tsk
);
2905 asmlinkage
void __sched
schedule(void)
2907 struct task_struct
*tsk
= current
;
2909 sched_submit_work(tsk
);
2912 EXPORT_SYMBOL(schedule
);
2914 #ifdef CONFIG_RCU_USER_QS
2915 asmlinkage
void __sched
schedule_user(void)
2918 * If we come here after a random call to set_need_resched(),
2919 * or we have been woken up remotely but the IPI has not yet arrived,
2920 * we haven't yet exited the RCU idle mode. Do it here manually until
2921 * we find a better solution.
2930 * schedule_preempt_disabled - called with preemption disabled
2932 * Returns with preemption disabled. Note: preempt_count must be 1
2934 void __sched
schedule_preempt_disabled(void)
2936 sched_preempt_enable_no_resched();
2941 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2943 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
2945 if (lock
->owner
!= owner
)
2949 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2950 * lock->owner still matches owner, if that fails, owner might
2951 * point to free()d memory, if it still matches, the rcu_read_lock()
2952 * ensures the memory stays valid.
2956 return owner
->on_cpu
;
2960 * Look out! "owner" is an entirely speculative pointer
2961 * access and not reliable.
2963 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
2965 if (!sched_feat(OWNER_SPIN
))
2969 while (owner_running(lock
, owner
)) {
2973 arch_mutex_cpu_relax();
2978 * We break out the loop above on need_resched() and when the
2979 * owner changed, which is a sign for heavy contention. Return
2980 * success only when lock->owner is NULL.
2982 return lock
->owner
== NULL
;
2986 #ifdef CONFIG_PREEMPT
2988 * this is the entry point to schedule() from in-kernel preemption
2989 * off of preempt_enable. Kernel preemptions off return from interrupt
2990 * occur there and call schedule directly.
2992 asmlinkage
void __sched notrace
preempt_schedule(void)
2994 struct thread_info
*ti
= current_thread_info();
2997 * If there is a non-zero preempt_count or interrupts are disabled,
2998 * we do not want to preempt the current task. Just return..
3000 if (likely(ti
->preempt_count
|| irqs_disabled()))
3004 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3006 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3009 * Check again in case we missed a preemption opportunity
3010 * between schedule and now.
3013 } while (need_resched());
3015 EXPORT_SYMBOL(preempt_schedule
);
3018 * this is the entry point to schedule() from kernel preemption
3019 * off of irq context.
3020 * Note, that this is called and return with irqs disabled. This will
3021 * protect us against recursive calling from irq.
3023 asmlinkage
void __sched
preempt_schedule_irq(void)
3025 struct thread_info
*ti
= current_thread_info();
3027 /* Catch callers which need to be fixed */
3028 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3032 add_preempt_count(PREEMPT_ACTIVE
);
3035 local_irq_disable();
3036 sub_preempt_count(PREEMPT_ACTIVE
);
3039 * Check again in case we missed a preemption opportunity
3040 * between schedule and now.
3043 } while (need_resched());
3046 #endif /* CONFIG_PREEMPT */
3048 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3051 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3053 EXPORT_SYMBOL(default_wake_function
);
3056 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3057 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3058 * number) then we wake all the non-exclusive tasks and one exclusive task.
3060 * There are circumstances in which we can try to wake a task which has already
3061 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3062 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3064 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3065 int nr_exclusive
, int wake_flags
, void *key
)
3067 wait_queue_t
*curr
, *next
;
3069 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3070 unsigned flags
= curr
->flags
;
3072 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3073 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3079 * __wake_up - wake up threads blocked on a waitqueue.
3081 * @mode: which threads
3082 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3083 * @key: is directly passed to the wakeup function
3085 * It may be assumed that this function implies a write memory barrier before
3086 * changing the task state if and only if any tasks are woken up.
3088 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3089 int nr_exclusive
, void *key
)
3091 unsigned long flags
;
3093 spin_lock_irqsave(&q
->lock
, flags
);
3094 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3095 spin_unlock_irqrestore(&q
->lock
, flags
);
3097 EXPORT_SYMBOL(__wake_up
);
3100 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3102 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3104 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3106 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3108 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3110 __wake_up_common(q
, mode
, 1, 0, key
);
3112 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3115 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3117 * @mode: which threads
3118 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3119 * @key: opaque value to be passed to wakeup targets
3121 * The sync wakeup differs that the waker knows that it will schedule
3122 * away soon, so while the target thread will be woken up, it will not
3123 * be migrated to another CPU - ie. the two threads are 'synchronized'
3124 * with each other. This can prevent needless bouncing between CPUs.
3126 * On UP it can prevent extra preemption.
3128 * It may be assumed that this function implies a write memory barrier before
3129 * changing the task state if and only if any tasks are woken up.
3131 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3132 int nr_exclusive
, void *key
)
3134 unsigned long flags
;
3135 int wake_flags
= WF_SYNC
;
3140 if (unlikely(!nr_exclusive
))
3143 spin_lock_irqsave(&q
->lock
, flags
);
3144 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3145 spin_unlock_irqrestore(&q
->lock
, flags
);
3147 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3150 * __wake_up_sync - see __wake_up_sync_key()
3152 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3154 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3156 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3159 * complete: - signals a single thread waiting on this completion
3160 * @x: holds the state of this particular completion
3162 * This will wake up a single thread waiting on this completion. Threads will be
3163 * awakened in the same order in which they were queued.
3165 * See also complete_all(), wait_for_completion() and related routines.
3167 * It may be assumed that this function implies a write memory barrier before
3168 * changing the task state if and only if any tasks are woken up.
3170 void complete(struct completion
*x
)
3172 unsigned long flags
;
3174 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3176 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3177 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3179 EXPORT_SYMBOL(complete
);
3182 * complete_all: - signals all threads waiting on this completion
3183 * @x: holds the state of this particular completion
3185 * This will wake up all threads waiting on this particular completion event.
3187 * It may be assumed that this function implies a write memory barrier before
3188 * changing the task state if and only if any tasks are woken up.
3190 void complete_all(struct completion
*x
)
3192 unsigned long flags
;
3194 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3195 x
->done
+= UINT_MAX
/2;
3196 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3197 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3199 EXPORT_SYMBOL(complete_all
);
3201 static inline long __sched
3202 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3205 DECLARE_WAITQUEUE(wait
, current
);
3207 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3209 if (signal_pending_state(state
, current
)) {
3210 timeout
= -ERESTARTSYS
;
3213 __set_current_state(state
);
3214 spin_unlock_irq(&x
->wait
.lock
);
3215 timeout
= schedule_timeout(timeout
);
3216 spin_lock_irq(&x
->wait
.lock
);
3217 } while (!x
->done
&& timeout
);
3218 __remove_wait_queue(&x
->wait
, &wait
);
3223 return timeout
?: 1;
3227 wait_for_common(struct completion
*x
, long timeout
, int state
)
3231 spin_lock_irq(&x
->wait
.lock
);
3232 timeout
= do_wait_for_common(x
, timeout
, state
);
3233 spin_unlock_irq(&x
->wait
.lock
);
3238 * wait_for_completion: - waits for completion of a task
3239 * @x: holds the state of this particular completion
3241 * This waits to be signaled for completion of a specific task. It is NOT
3242 * interruptible and there is no timeout.
3244 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3245 * and interrupt capability. Also see complete().
3247 void __sched
wait_for_completion(struct completion
*x
)
3249 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3251 EXPORT_SYMBOL(wait_for_completion
);
3254 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3255 * @x: holds the state of this particular completion
3256 * @timeout: timeout value in jiffies
3258 * This waits for either a completion of a specific task to be signaled or for a
3259 * specified timeout to expire. The timeout is in jiffies. It is not
3262 * The return value is 0 if timed out, and positive (at least 1, or number of
3263 * jiffies left till timeout) if completed.
3265 unsigned long __sched
3266 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3268 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3270 EXPORT_SYMBOL(wait_for_completion_timeout
);
3273 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3274 * @x: holds the state of this particular completion
3276 * This waits for completion of a specific task to be signaled. It is
3279 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3281 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3283 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3284 if (t
== -ERESTARTSYS
)
3288 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3291 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3292 * @x: holds the state of this particular completion
3293 * @timeout: timeout value in jiffies
3295 * This waits for either a completion of a specific task to be signaled or for a
3296 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3298 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3299 * positive (at least 1, or number of jiffies left till timeout) if completed.
3302 wait_for_completion_interruptible_timeout(struct completion
*x
,
3303 unsigned long timeout
)
3305 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3307 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3310 * wait_for_completion_killable: - waits for completion of a task (killable)
3311 * @x: holds the state of this particular completion
3313 * This waits to be signaled for completion of a specific task. It can be
3314 * interrupted by a kill signal.
3316 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3318 int __sched
wait_for_completion_killable(struct completion
*x
)
3320 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3321 if (t
== -ERESTARTSYS
)
3325 EXPORT_SYMBOL(wait_for_completion_killable
);
3328 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3329 * @x: holds the state of this particular completion
3330 * @timeout: timeout value in jiffies
3332 * This waits for either a completion of a specific task to be
3333 * signaled or for a specified timeout to expire. It can be
3334 * interrupted by a kill signal. The timeout is in jiffies.
3336 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3337 * positive (at least 1, or number of jiffies left till timeout) if completed.
3340 wait_for_completion_killable_timeout(struct completion
*x
,
3341 unsigned long timeout
)
3343 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3345 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3348 * try_wait_for_completion - try to decrement a completion without blocking
3349 * @x: completion structure
3351 * Returns: 0 if a decrement cannot be done without blocking
3352 * 1 if a decrement succeeded.
3354 * If a completion is being used as a counting completion,
3355 * attempt to decrement the counter without blocking. This
3356 * enables us to avoid waiting if the resource the completion
3357 * is protecting is not available.
3359 bool try_wait_for_completion(struct completion
*x
)
3361 unsigned long flags
;
3364 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3369 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3372 EXPORT_SYMBOL(try_wait_for_completion
);
3375 * completion_done - Test to see if a completion has any waiters
3376 * @x: completion structure
3378 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3379 * 1 if there are no waiters.
3382 bool completion_done(struct completion
*x
)
3384 unsigned long flags
;
3387 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3390 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3393 EXPORT_SYMBOL(completion_done
);
3396 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3398 unsigned long flags
;
3401 init_waitqueue_entry(&wait
, current
);
3403 __set_current_state(state
);
3405 spin_lock_irqsave(&q
->lock
, flags
);
3406 __add_wait_queue(q
, &wait
);
3407 spin_unlock(&q
->lock
);
3408 timeout
= schedule_timeout(timeout
);
3409 spin_lock_irq(&q
->lock
);
3410 __remove_wait_queue(q
, &wait
);
3411 spin_unlock_irqrestore(&q
->lock
, flags
);
3416 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3418 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3420 EXPORT_SYMBOL(interruptible_sleep_on
);
3423 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3425 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3427 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3429 void __sched
sleep_on(wait_queue_head_t
*q
)
3431 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3433 EXPORT_SYMBOL(sleep_on
);
3435 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3437 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3439 EXPORT_SYMBOL(sleep_on_timeout
);
3441 #ifdef CONFIG_RT_MUTEXES
3444 * rt_mutex_setprio - set the current priority of a task
3446 * @prio: prio value (kernel-internal form)
3448 * This function changes the 'effective' priority of a task. It does
3449 * not touch ->normal_prio like __setscheduler().
3451 * Used by the rt_mutex code to implement priority inheritance logic.
3453 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3455 int oldprio
, on_rq
, running
;
3457 const struct sched_class
*prev_class
;
3459 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3461 rq
= __task_rq_lock(p
);
3464 * Idle task boosting is a nono in general. There is one
3465 * exception, when PREEMPT_RT and NOHZ is active:
3467 * The idle task calls get_next_timer_interrupt() and holds
3468 * the timer wheel base->lock on the CPU and another CPU wants
3469 * to access the timer (probably to cancel it). We can safely
3470 * ignore the boosting request, as the idle CPU runs this code
3471 * with interrupts disabled and will complete the lock
3472 * protected section without being interrupted. So there is no
3473 * real need to boost.
3475 if (unlikely(p
== rq
->idle
)) {
3476 WARN_ON(p
!= rq
->curr
);
3477 WARN_ON(p
->pi_blocked_on
);
3481 trace_sched_pi_setprio(p
, prio
);
3483 prev_class
= p
->sched_class
;
3485 running
= task_current(rq
, p
);
3487 dequeue_task(rq
, p
, 0);
3489 p
->sched_class
->put_prev_task(rq
, p
);
3492 p
->sched_class
= &rt_sched_class
;
3494 p
->sched_class
= &fair_sched_class
;
3499 p
->sched_class
->set_curr_task(rq
);
3501 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3503 check_class_changed(rq
, p
, prev_class
, oldprio
);
3505 __task_rq_unlock(rq
);
3508 void set_user_nice(struct task_struct
*p
, long nice
)
3510 int old_prio
, delta
, on_rq
;
3511 unsigned long flags
;
3514 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3517 * We have to be careful, if called from sys_setpriority(),
3518 * the task might be in the middle of scheduling on another CPU.
3520 rq
= task_rq_lock(p
, &flags
);
3522 * The RT priorities are set via sched_setscheduler(), but we still
3523 * allow the 'normal' nice value to be set - but as expected
3524 * it wont have any effect on scheduling until the task is
3525 * SCHED_FIFO/SCHED_RR:
3527 if (task_has_rt_policy(p
)) {
3528 p
->static_prio
= NICE_TO_PRIO(nice
);
3533 dequeue_task(rq
, p
, 0);
3535 p
->static_prio
= NICE_TO_PRIO(nice
);
3538 p
->prio
= effective_prio(p
);
3539 delta
= p
->prio
- old_prio
;
3542 enqueue_task(rq
, p
, 0);
3544 * If the task increased its priority or is running and
3545 * lowered its priority, then reschedule its CPU:
3547 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3548 resched_task(rq
->curr
);
3551 task_rq_unlock(rq
, p
, &flags
);
3553 EXPORT_SYMBOL(set_user_nice
);
3556 * can_nice - check if a task can reduce its nice value
3560 int can_nice(const struct task_struct
*p
, const int nice
)
3562 /* convert nice value [19,-20] to rlimit style value [1,40] */
3563 int nice_rlim
= 20 - nice
;
3565 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3566 capable(CAP_SYS_NICE
));
3569 #ifdef __ARCH_WANT_SYS_NICE
3572 * sys_nice - change the priority of the current process.
3573 * @increment: priority increment
3575 * sys_setpriority is a more generic, but much slower function that
3576 * does similar things.
3578 SYSCALL_DEFINE1(nice
, int, increment
)
3583 * Setpriority might change our priority at the same moment.
3584 * We don't have to worry. Conceptually one call occurs first
3585 * and we have a single winner.
3587 if (increment
< -40)
3592 nice
= TASK_NICE(current
) + increment
;
3598 if (increment
< 0 && !can_nice(current
, nice
))
3601 retval
= security_task_setnice(current
, nice
);
3605 set_user_nice(current
, nice
);
3612 * task_prio - return the priority value of a given task.
3613 * @p: the task in question.
3615 * This is the priority value as seen by users in /proc.
3616 * RT tasks are offset by -200. Normal tasks are centered
3617 * around 0, value goes from -16 to +15.
3619 int task_prio(const struct task_struct
*p
)
3621 return p
->prio
- MAX_RT_PRIO
;
3625 * task_nice - return the nice value of a given task.
3626 * @p: the task in question.
3628 int task_nice(const struct task_struct
*p
)
3630 return TASK_NICE(p
);
3632 EXPORT_SYMBOL(task_nice
);
3635 * idle_cpu - is a given cpu idle currently?
3636 * @cpu: the processor in question.
3638 int idle_cpu(int cpu
)
3640 struct rq
*rq
= cpu_rq(cpu
);
3642 if (rq
->curr
!= rq
->idle
)
3649 if (!llist_empty(&rq
->wake_list
))
3657 * idle_task - return the idle task for a given cpu.
3658 * @cpu: the processor in question.
3660 struct task_struct
*idle_task(int cpu
)
3662 return cpu_rq(cpu
)->idle
;
3666 * find_process_by_pid - find a process with a matching PID value.
3667 * @pid: the pid in question.
3669 static struct task_struct
*find_process_by_pid(pid_t pid
)
3671 return pid
? find_task_by_vpid(pid
) : current
;
3674 /* Actually do priority change: must hold rq lock. */
3676 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3679 p
->rt_priority
= prio
;
3680 p
->normal_prio
= normal_prio(p
);
3681 /* we are holding p->pi_lock already */
3682 p
->prio
= rt_mutex_getprio(p
);
3683 if (rt_prio(p
->prio
))
3684 p
->sched_class
= &rt_sched_class
;
3686 p
->sched_class
= &fair_sched_class
;
3691 * check the target process has a UID that matches the current process's
3693 static bool check_same_owner(struct task_struct
*p
)
3695 const struct cred
*cred
= current_cred(), *pcred
;
3699 pcred
= __task_cred(p
);
3700 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3701 uid_eq(cred
->euid
, pcred
->uid
));
3706 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3707 const struct sched_param
*param
, bool user
)
3709 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3710 unsigned long flags
;
3711 const struct sched_class
*prev_class
;
3715 /* may grab non-irq protected spin_locks */
3716 BUG_ON(in_interrupt());
3718 /* double check policy once rq lock held */
3720 reset_on_fork
= p
->sched_reset_on_fork
;
3721 policy
= oldpolicy
= p
->policy
;
3723 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3724 policy
&= ~SCHED_RESET_ON_FORK
;
3726 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3727 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3728 policy
!= SCHED_IDLE
)
3733 * Valid priorities for SCHED_FIFO and SCHED_RR are
3734 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3735 * SCHED_BATCH and SCHED_IDLE is 0.
3737 if (param
->sched_priority
< 0 ||
3738 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3739 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3741 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3745 * Allow unprivileged RT tasks to decrease priority:
3747 if (user
&& !capable(CAP_SYS_NICE
)) {
3748 if (rt_policy(policy
)) {
3749 unsigned long rlim_rtprio
=
3750 task_rlimit(p
, RLIMIT_RTPRIO
);
3752 /* can't set/change the rt policy */
3753 if (policy
!= p
->policy
&& !rlim_rtprio
)
3756 /* can't increase priority */
3757 if (param
->sched_priority
> p
->rt_priority
&&
3758 param
->sched_priority
> rlim_rtprio
)
3763 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3764 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3766 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3767 if (!can_nice(p
, TASK_NICE(p
)))
3771 /* can't change other user's priorities */
3772 if (!check_same_owner(p
))
3775 /* Normal users shall not reset the sched_reset_on_fork flag */
3776 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3781 retval
= security_task_setscheduler(p
);
3787 * make sure no PI-waiters arrive (or leave) while we are
3788 * changing the priority of the task:
3790 * To be able to change p->policy safely, the appropriate
3791 * runqueue lock must be held.
3793 rq
= task_rq_lock(p
, &flags
);
3796 * Changing the policy of the stop threads its a very bad idea
3798 if (p
== rq
->stop
) {
3799 task_rq_unlock(rq
, p
, &flags
);
3804 * If not changing anything there's no need to proceed further:
3806 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3807 param
->sched_priority
== p
->rt_priority
))) {
3808 task_rq_unlock(rq
, p
, &flags
);
3812 #ifdef CONFIG_RT_GROUP_SCHED
3815 * Do not allow realtime tasks into groups that have no runtime
3818 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3819 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3820 !task_group_is_autogroup(task_group(p
))) {
3821 task_rq_unlock(rq
, p
, &flags
);
3827 /* recheck policy now with rq lock held */
3828 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3829 policy
= oldpolicy
= -1;
3830 task_rq_unlock(rq
, p
, &flags
);
3834 running
= task_current(rq
, p
);
3836 dequeue_task(rq
, p
, 0);
3838 p
->sched_class
->put_prev_task(rq
, p
);
3840 p
->sched_reset_on_fork
= reset_on_fork
;
3843 prev_class
= p
->sched_class
;
3844 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3847 p
->sched_class
->set_curr_task(rq
);
3849 enqueue_task(rq
, p
, 0);
3851 check_class_changed(rq
, p
, prev_class
, oldprio
);
3852 task_rq_unlock(rq
, p
, &flags
);
3854 rt_mutex_adjust_pi(p
);
3860 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3861 * @p: the task in question.
3862 * @policy: new policy.
3863 * @param: structure containing the new RT priority.
3865 * NOTE that the task may be already dead.
3867 int sched_setscheduler(struct task_struct
*p
, int policy
,
3868 const struct sched_param
*param
)
3870 return __sched_setscheduler(p
, policy
, param
, true);
3872 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3875 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3876 * @p: the task in question.
3877 * @policy: new policy.
3878 * @param: structure containing the new RT priority.
3880 * Just like sched_setscheduler, only don't bother checking if the
3881 * current context has permission. For example, this is needed in
3882 * stop_machine(): we create temporary high priority worker threads,
3883 * but our caller might not have that capability.
3885 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3886 const struct sched_param
*param
)
3888 return __sched_setscheduler(p
, policy
, param
, false);
3892 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3894 struct sched_param lparam
;
3895 struct task_struct
*p
;
3898 if (!param
|| pid
< 0)
3900 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3905 p
= find_process_by_pid(pid
);
3907 retval
= sched_setscheduler(p
, policy
, &lparam
);
3914 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3915 * @pid: the pid in question.
3916 * @policy: new policy.
3917 * @param: structure containing the new RT priority.
3919 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3920 struct sched_param __user
*, param
)
3922 /* negative values for policy are not valid */
3926 return do_sched_setscheduler(pid
, policy
, param
);
3930 * sys_sched_setparam - set/change the RT priority of a thread
3931 * @pid: the pid in question.
3932 * @param: structure containing the new RT priority.
3934 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3936 return do_sched_setscheduler(pid
, -1, param
);
3940 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3941 * @pid: the pid in question.
3943 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3945 struct task_struct
*p
;
3953 p
= find_process_by_pid(pid
);
3955 retval
= security_task_getscheduler(p
);
3958 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3965 * sys_sched_getparam - get the RT priority of a thread
3966 * @pid: the pid in question.
3967 * @param: structure containing the RT priority.
3969 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3971 struct sched_param lp
;
3972 struct task_struct
*p
;
3975 if (!param
|| pid
< 0)
3979 p
= find_process_by_pid(pid
);
3984 retval
= security_task_getscheduler(p
);
3988 lp
.sched_priority
= p
->rt_priority
;
3992 * This one might sleep, we cannot do it with a spinlock held ...
3994 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4003 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4005 cpumask_var_t cpus_allowed
, new_mask
;
4006 struct task_struct
*p
;
4012 p
= find_process_by_pid(pid
);
4019 /* Prevent p going away */
4023 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4027 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4029 goto out_free_cpus_allowed
;
4032 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4035 retval
= security_task_setscheduler(p
);
4039 cpuset_cpus_allowed(p
, cpus_allowed
);
4040 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4042 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4045 cpuset_cpus_allowed(p
, cpus_allowed
);
4046 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4048 * We must have raced with a concurrent cpuset
4049 * update. Just reset the cpus_allowed to the
4050 * cpuset's cpus_allowed
4052 cpumask_copy(new_mask
, cpus_allowed
);
4057 free_cpumask_var(new_mask
);
4058 out_free_cpus_allowed
:
4059 free_cpumask_var(cpus_allowed
);
4066 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4067 struct cpumask
*new_mask
)
4069 if (len
< cpumask_size())
4070 cpumask_clear(new_mask
);
4071 else if (len
> cpumask_size())
4072 len
= cpumask_size();
4074 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4078 * sys_sched_setaffinity - set the cpu affinity of a process
4079 * @pid: pid of the process
4080 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4081 * @user_mask_ptr: user-space pointer to the new cpu mask
4083 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4084 unsigned long __user
*, user_mask_ptr
)
4086 cpumask_var_t new_mask
;
4089 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4092 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4094 retval
= sched_setaffinity(pid
, new_mask
);
4095 free_cpumask_var(new_mask
);
4099 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4101 struct task_struct
*p
;
4102 unsigned long flags
;
4109 p
= find_process_by_pid(pid
);
4113 retval
= security_task_getscheduler(p
);
4117 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4118 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4119 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4129 * sys_sched_getaffinity - get the cpu affinity of a process
4130 * @pid: pid of the process
4131 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4132 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4134 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4135 unsigned long __user
*, user_mask_ptr
)
4140 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4142 if (len
& (sizeof(unsigned long)-1))
4145 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4148 ret
= sched_getaffinity(pid
, mask
);
4150 size_t retlen
= min_t(size_t, len
, cpumask_size());
4152 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4157 free_cpumask_var(mask
);
4163 * sys_sched_yield - yield the current processor to other threads.
4165 * This function yields the current CPU to other tasks. If there are no
4166 * other threads running on this CPU then this function will return.
4168 SYSCALL_DEFINE0(sched_yield
)
4170 struct rq
*rq
= this_rq_lock();
4172 schedstat_inc(rq
, yld_count
);
4173 current
->sched_class
->yield_task(rq
);
4176 * Since we are going to call schedule() anyway, there's
4177 * no need to preempt or enable interrupts:
4179 __release(rq
->lock
);
4180 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4181 do_raw_spin_unlock(&rq
->lock
);
4182 sched_preempt_enable_no_resched();
4189 static inline int should_resched(void)
4191 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4194 static void __cond_resched(void)
4196 add_preempt_count(PREEMPT_ACTIVE
);
4198 sub_preempt_count(PREEMPT_ACTIVE
);
4201 int __sched
_cond_resched(void)
4203 if (should_resched()) {
4209 EXPORT_SYMBOL(_cond_resched
);
4212 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4213 * call schedule, and on return reacquire the lock.
4215 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4216 * operations here to prevent schedule() from being called twice (once via
4217 * spin_unlock(), once by hand).
4219 int __cond_resched_lock(spinlock_t
*lock
)
4221 int resched
= should_resched();
4224 lockdep_assert_held(lock
);
4226 if (spin_needbreak(lock
) || resched
) {
4237 EXPORT_SYMBOL(__cond_resched_lock
);
4239 int __sched
__cond_resched_softirq(void)
4241 BUG_ON(!in_softirq());
4243 if (should_resched()) {
4251 EXPORT_SYMBOL(__cond_resched_softirq
);
4254 * yield - yield the current processor to other threads.
4256 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4258 * The scheduler is at all times free to pick the calling task as the most
4259 * eligible task to run, if removing the yield() call from your code breaks
4260 * it, its already broken.
4262 * Typical broken usage is:
4267 * where one assumes that yield() will let 'the other' process run that will
4268 * make event true. If the current task is a SCHED_FIFO task that will never
4269 * happen. Never use yield() as a progress guarantee!!
4271 * If you want to use yield() to wait for something, use wait_event().
4272 * If you want to use yield() to be 'nice' for others, use cond_resched().
4273 * If you still want to use yield(), do not!
4275 void __sched
yield(void)
4277 set_current_state(TASK_RUNNING
);
4280 EXPORT_SYMBOL(yield
);
4283 * yield_to - yield the current processor to another thread in
4284 * your thread group, or accelerate that thread toward the
4285 * processor it's on.
4287 * @preempt: whether task preemption is allowed or not
4289 * It's the caller's job to ensure that the target task struct
4290 * can't go away on us before we can do any checks.
4292 * Returns true if we indeed boosted the target task.
4294 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4296 struct task_struct
*curr
= current
;
4297 struct rq
*rq
, *p_rq
;
4298 unsigned long flags
;
4301 local_irq_save(flags
);
4306 double_rq_lock(rq
, p_rq
);
4307 while (task_rq(p
) != p_rq
) {
4308 double_rq_unlock(rq
, p_rq
);
4312 if (!curr
->sched_class
->yield_to_task
)
4315 if (curr
->sched_class
!= p
->sched_class
)
4318 if (task_running(p_rq
, p
) || p
->state
)
4321 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4323 schedstat_inc(rq
, yld_count
);
4325 * Make p's CPU reschedule; pick_next_entity takes care of
4328 if (preempt
&& rq
!= p_rq
)
4329 resched_task(p_rq
->curr
);
4333 double_rq_unlock(rq
, p_rq
);
4334 local_irq_restore(flags
);
4341 EXPORT_SYMBOL_GPL(yield_to
);
4344 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4345 * that process accounting knows that this is a task in IO wait state.
4347 void __sched
io_schedule(void)
4349 struct rq
*rq
= raw_rq();
4351 delayacct_blkio_start();
4352 atomic_inc(&rq
->nr_iowait
);
4353 blk_flush_plug(current
);
4354 current
->in_iowait
= 1;
4356 current
->in_iowait
= 0;
4357 atomic_dec(&rq
->nr_iowait
);
4358 delayacct_blkio_end();
4360 EXPORT_SYMBOL(io_schedule
);
4362 long __sched
io_schedule_timeout(long timeout
)
4364 struct rq
*rq
= raw_rq();
4367 delayacct_blkio_start();
4368 atomic_inc(&rq
->nr_iowait
);
4369 blk_flush_plug(current
);
4370 current
->in_iowait
= 1;
4371 ret
= schedule_timeout(timeout
);
4372 current
->in_iowait
= 0;
4373 atomic_dec(&rq
->nr_iowait
);
4374 delayacct_blkio_end();
4379 * sys_sched_get_priority_max - return maximum RT priority.
4380 * @policy: scheduling class.
4382 * this syscall returns the maximum rt_priority that can be used
4383 * by a given scheduling class.
4385 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4392 ret
= MAX_USER_RT_PRIO
-1;
4404 * sys_sched_get_priority_min - return minimum RT priority.
4405 * @policy: scheduling class.
4407 * this syscall returns the minimum rt_priority that can be used
4408 * by a given scheduling class.
4410 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4428 * sys_sched_rr_get_interval - return the default timeslice of a process.
4429 * @pid: pid of the process.
4430 * @interval: userspace pointer to the timeslice value.
4432 * this syscall writes the default timeslice value of a given process
4433 * into the user-space timespec buffer. A value of '0' means infinity.
4435 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4436 struct timespec __user
*, interval
)
4438 struct task_struct
*p
;
4439 unsigned int time_slice
;
4440 unsigned long flags
;
4450 p
= find_process_by_pid(pid
);
4454 retval
= security_task_getscheduler(p
);
4458 rq
= task_rq_lock(p
, &flags
);
4459 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4460 task_rq_unlock(rq
, p
, &flags
);
4463 jiffies_to_timespec(time_slice
, &t
);
4464 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4472 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4474 void sched_show_task(struct task_struct
*p
)
4476 unsigned long free
= 0;
4479 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4480 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4481 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4482 #if BITS_PER_LONG == 32
4483 if (state
== TASK_RUNNING
)
4484 printk(KERN_CONT
" running ");
4486 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4488 if (state
== TASK_RUNNING
)
4489 printk(KERN_CONT
" running task ");
4491 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4493 #ifdef CONFIG_DEBUG_STACK_USAGE
4494 free
= stack_not_used(p
);
4496 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4497 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4498 (unsigned long)task_thread_info(p
)->flags
);
4500 show_stack(p
, NULL
);
4503 void show_state_filter(unsigned long state_filter
)
4505 struct task_struct
*g
, *p
;
4507 #if BITS_PER_LONG == 32
4509 " task PC stack pid father\n");
4512 " task PC stack pid father\n");
4515 do_each_thread(g
, p
) {
4517 * reset the NMI-timeout, listing all files on a slow
4518 * console might take a lot of time:
4520 touch_nmi_watchdog();
4521 if (!state_filter
|| (p
->state
& state_filter
))
4523 } while_each_thread(g
, p
);
4525 touch_all_softlockup_watchdogs();
4527 #ifdef CONFIG_SCHED_DEBUG
4528 sysrq_sched_debug_show();
4532 * Only show locks if all tasks are dumped:
4535 debug_show_all_locks();
4538 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4540 idle
->sched_class
= &idle_sched_class
;
4544 * init_idle - set up an idle thread for a given CPU
4545 * @idle: task in question
4546 * @cpu: cpu the idle task belongs to
4548 * NOTE: this function does not set the idle thread's NEED_RESCHED
4549 * flag, to make booting more robust.
4551 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4553 struct rq
*rq
= cpu_rq(cpu
);
4554 unsigned long flags
;
4556 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4559 idle
->state
= TASK_RUNNING
;
4560 idle
->se
.exec_start
= sched_clock();
4562 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4564 * We're having a chicken and egg problem, even though we are
4565 * holding rq->lock, the cpu isn't yet set to this cpu so the
4566 * lockdep check in task_group() will fail.
4568 * Similar case to sched_fork(). / Alternatively we could
4569 * use task_rq_lock() here and obtain the other rq->lock.
4574 __set_task_cpu(idle
, cpu
);
4577 rq
->curr
= rq
->idle
= idle
;
4578 #if defined(CONFIG_SMP)
4581 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4583 /* Set the preempt count _outside_ the spinlocks! */
4584 task_thread_info(idle
)->preempt_count
= 0;
4587 * The idle tasks have their own, simple scheduling class:
4589 idle
->sched_class
= &idle_sched_class
;
4590 ftrace_graph_init_idle_task(idle
, cpu
);
4591 #if defined(CONFIG_SMP)
4592 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4597 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4599 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4600 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4602 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4603 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4607 * This is how migration works:
4609 * 1) we invoke migration_cpu_stop() on the target CPU using
4611 * 2) stopper starts to run (implicitly forcing the migrated thread
4613 * 3) it checks whether the migrated task is still in the wrong runqueue.
4614 * 4) if it's in the wrong runqueue then the migration thread removes
4615 * it and puts it into the right queue.
4616 * 5) stopper completes and stop_one_cpu() returns and the migration
4621 * Change a given task's CPU affinity. Migrate the thread to a
4622 * proper CPU and schedule it away if the CPU it's executing on
4623 * is removed from the allowed bitmask.
4625 * NOTE: the caller must have a valid reference to the task, the
4626 * task must not exit() & deallocate itself prematurely. The
4627 * call is not atomic; no spinlocks may be held.
4629 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4631 unsigned long flags
;
4633 unsigned int dest_cpu
;
4636 rq
= task_rq_lock(p
, &flags
);
4638 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4641 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4646 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4651 do_set_cpus_allowed(p
, new_mask
);
4653 /* Can the task run on the task's current CPU? If so, we're done */
4654 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4657 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4659 struct migration_arg arg
= { p
, dest_cpu
};
4660 /* Need help from migration thread: drop lock and wait. */
4661 task_rq_unlock(rq
, p
, &flags
);
4662 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4663 tlb_migrate_finish(p
->mm
);
4667 task_rq_unlock(rq
, p
, &flags
);
4671 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4674 * Move (not current) task off this cpu, onto dest cpu. We're doing
4675 * this because either it can't run here any more (set_cpus_allowed()
4676 * away from this CPU, or CPU going down), or because we're
4677 * attempting to rebalance this task on exec (sched_exec).
4679 * So we race with normal scheduler movements, but that's OK, as long
4680 * as the task is no longer on this CPU.
4682 * Returns non-zero if task was successfully migrated.
4684 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4686 struct rq
*rq_dest
, *rq_src
;
4689 if (unlikely(!cpu_active(dest_cpu
)))
4692 rq_src
= cpu_rq(src_cpu
);
4693 rq_dest
= cpu_rq(dest_cpu
);
4695 raw_spin_lock(&p
->pi_lock
);
4696 double_rq_lock(rq_src
, rq_dest
);
4697 /* Already moved. */
4698 if (task_cpu(p
) != src_cpu
)
4700 /* Affinity changed (again). */
4701 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4705 * If we're not on a rq, the next wake-up will ensure we're
4709 dequeue_task(rq_src
, p
, 0);
4710 set_task_cpu(p
, dest_cpu
);
4711 enqueue_task(rq_dest
, p
, 0);
4712 check_preempt_curr(rq_dest
, p
, 0);
4717 double_rq_unlock(rq_src
, rq_dest
);
4718 raw_spin_unlock(&p
->pi_lock
);
4723 * migration_cpu_stop - this will be executed by a highprio stopper thread
4724 * and performs thread migration by bumping thread off CPU then
4725 * 'pushing' onto another runqueue.
4727 static int migration_cpu_stop(void *data
)
4729 struct migration_arg
*arg
= data
;
4732 * The original target cpu might have gone down and we might
4733 * be on another cpu but it doesn't matter.
4735 local_irq_disable();
4736 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4741 #ifdef CONFIG_HOTPLUG_CPU
4744 * Ensures that the idle task is using init_mm right before its cpu goes
4747 void idle_task_exit(void)
4749 struct mm_struct
*mm
= current
->active_mm
;
4751 BUG_ON(cpu_online(smp_processor_id()));
4754 switch_mm(mm
, &init_mm
, current
);
4759 * Since this CPU is going 'away' for a while, fold any nr_active delta
4760 * we might have. Assumes we're called after migrate_tasks() so that the
4761 * nr_active count is stable.
4763 * Also see the comment "Global load-average calculations".
4765 static void calc_load_migrate(struct rq
*rq
)
4767 long delta
= calc_load_fold_active(rq
);
4769 atomic_long_add(delta
, &calc_load_tasks
);
4773 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4774 * try_to_wake_up()->select_task_rq().
4776 * Called with rq->lock held even though we'er in stop_machine() and
4777 * there's no concurrency possible, we hold the required locks anyway
4778 * because of lock validation efforts.
4780 static void migrate_tasks(unsigned int dead_cpu
)
4782 struct rq
*rq
= cpu_rq(dead_cpu
);
4783 struct task_struct
*next
, *stop
= rq
->stop
;
4787 * Fudge the rq selection such that the below task selection loop
4788 * doesn't get stuck on the currently eligible stop task.
4790 * We're currently inside stop_machine() and the rq is either stuck
4791 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4792 * either way we should never end up calling schedule() until we're
4799 * There's this thread running, bail when that's the only
4802 if (rq
->nr_running
== 1)
4805 next
= pick_next_task(rq
);
4807 next
->sched_class
->put_prev_task(rq
, next
);
4809 /* Find suitable destination for @next, with force if needed. */
4810 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4811 raw_spin_unlock(&rq
->lock
);
4813 __migrate_task(next
, dead_cpu
, dest_cpu
);
4815 raw_spin_lock(&rq
->lock
);
4821 #endif /* CONFIG_HOTPLUG_CPU */
4823 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4825 static struct ctl_table sd_ctl_dir
[] = {
4827 .procname
= "sched_domain",
4833 static struct ctl_table sd_ctl_root
[] = {
4835 .procname
= "kernel",
4837 .child
= sd_ctl_dir
,
4842 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4844 struct ctl_table
*entry
=
4845 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4850 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4852 struct ctl_table
*entry
;
4855 * In the intermediate directories, both the child directory and
4856 * procname are dynamically allocated and could fail but the mode
4857 * will always be set. In the lowest directory the names are
4858 * static strings and all have proc handlers.
4860 for (entry
= *tablep
; entry
->mode
; entry
++) {
4862 sd_free_ctl_entry(&entry
->child
);
4863 if (entry
->proc_handler
== NULL
)
4864 kfree(entry
->procname
);
4871 static int min_load_idx
= 0;
4872 static int max_load_idx
= CPU_LOAD_IDX_MAX
;
4875 set_table_entry(struct ctl_table
*entry
,
4876 const char *procname
, void *data
, int maxlen
,
4877 umode_t mode
, proc_handler
*proc_handler
,
4880 entry
->procname
= procname
;
4882 entry
->maxlen
= maxlen
;
4884 entry
->proc_handler
= proc_handler
;
4887 entry
->extra1
= &min_load_idx
;
4888 entry
->extra2
= &max_load_idx
;
4892 static struct ctl_table
*
4893 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4895 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4900 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4901 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4902 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4903 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4904 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4905 sizeof(int), 0644, proc_dointvec_minmax
, true);
4906 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4907 sizeof(int), 0644, proc_dointvec_minmax
, true);
4908 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4909 sizeof(int), 0644, proc_dointvec_minmax
, true);
4910 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4911 sizeof(int), 0644, proc_dointvec_minmax
, true);
4912 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4913 sizeof(int), 0644, proc_dointvec_minmax
, true);
4914 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4915 sizeof(int), 0644, proc_dointvec_minmax
, false);
4916 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4917 sizeof(int), 0644, proc_dointvec_minmax
, false);
4918 set_table_entry(&table
[9], "cache_nice_tries",
4919 &sd
->cache_nice_tries
,
4920 sizeof(int), 0644, proc_dointvec_minmax
, false);
4921 set_table_entry(&table
[10], "flags", &sd
->flags
,
4922 sizeof(int), 0644, proc_dointvec_minmax
, false);
4923 set_table_entry(&table
[11], "name", sd
->name
,
4924 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4925 /* &table[12] is terminator */
4930 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4932 struct ctl_table
*entry
, *table
;
4933 struct sched_domain
*sd
;
4934 int domain_num
= 0, i
;
4937 for_each_domain(cpu
, sd
)
4939 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4944 for_each_domain(cpu
, sd
) {
4945 snprintf(buf
, 32, "domain%d", i
);
4946 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4948 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4955 static struct ctl_table_header
*sd_sysctl_header
;
4956 static void register_sched_domain_sysctl(void)
4958 int i
, cpu_num
= num_possible_cpus();
4959 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4962 WARN_ON(sd_ctl_dir
[0].child
);
4963 sd_ctl_dir
[0].child
= entry
;
4968 for_each_possible_cpu(i
) {
4969 snprintf(buf
, 32, "cpu%d", i
);
4970 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4972 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4976 WARN_ON(sd_sysctl_header
);
4977 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4980 /* may be called multiple times per register */
4981 static void unregister_sched_domain_sysctl(void)
4983 if (sd_sysctl_header
)
4984 unregister_sysctl_table(sd_sysctl_header
);
4985 sd_sysctl_header
= NULL
;
4986 if (sd_ctl_dir
[0].child
)
4987 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4990 static void register_sched_domain_sysctl(void)
4993 static void unregister_sched_domain_sysctl(void)
4998 static void set_rq_online(struct rq
*rq
)
5001 const struct sched_class
*class;
5003 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5006 for_each_class(class) {
5007 if (class->rq_online
)
5008 class->rq_online(rq
);
5013 static void set_rq_offline(struct rq
*rq
)
5016 const struct sched_class
*class;
5018 for_each_class(class) {
5019 if (class->rq_offline
)
5020 class->rq_offline(rq
);
5023 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5029 * migration_call - callback that gets triggered when a CPU is added.
5030 * Here we can start up the necessary migration thread for the new CPU.
5032 static int __cpuinit
5033 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5035 int cpu
= (long)hcpu
;
5036 unsigned long flags
;
5037 struct rq
*rq
= cpu_rq(cpu
);
5039 switch (action
& ~CPU_TASKS_FROZEN
) {
5041 case CPU_UP_PREPARE
:
5042 rq
->calc_load_update
= calc_load_update
;
5046 /* Update our root-domain */
5047 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5049 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5053 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5056 #ifdef CONFIG_HOTPLUG_CPU
5058 sched_ttwu_pending();
5059 /* Update our root-domain */
5060 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5062 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5066 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5067 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5071 calc_load_migrate(rq
);
5076 update_max_interval();
5082 * Register at high priority so that task migration (migrate_all_tasks)
5083 * happens before everything else. This has to be lower priority than
5084 * the notifier in the perf_event subsystem, though.
5086 static struct notifier_block __cpuinitdata migration_notifier
= {
5087 .notifier_call
= migration_call
,
5088 .priority
= CPU_PRI_MIGRATION
,
5091 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5092 unsigned long action
, void *hcpu
)
5094 switch (action
& ~CPU_TASKS_FROZEN
) {
5096 case CPU_DOWN_FAILED
:
5097 set_cpu_active((long)hcpu
, true);
5104 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5105 unsigned long action
, void *hcpu
)
5107 switch (action
& ~CPU_TASKS_FROZEN
) {
5108 case CPU_DOWN_PREPARE
:
5109 set_cpu_active((long)hcpu
, false);
5116 static int __init
migration_init(void)
5118 void *cpu
= (void *)(long)smp_processor_id();
5121 /* Initialize migration for the boot CPU */
5122 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5123 BUG_ON(err
== NOTIFY_BAD
);
5124 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5125 register_cpu_notifier(&migration_notifier
);
5127 /* Register cpu active notifiers */
5128 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5129 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5133 early_initcall(migration_init
);
5138 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5140 #ifdef CONFIG_SCHED_DEBUG
5142 static __read_mostly
int sched_debug_enabled
;
5144 static int __init
sched_debug_setup(char *str
)
5146 sched_debug_enabled
= 1;
5150 early_param("sched_debug", sched_debug_setup
);
5152 static inline bool sched_debug(void)
5154 return sched_debug_enabled
;
5157 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5158 struct cpumask
*groupmask
)
5160 struct sched_group
*group
= sd
->groups
;
5163 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5164 cpumask_clear(groupmask
);
5166 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5168 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5169 printk("does not load-balance\n");
5171 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5176 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5178 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5179 printk(KERN_ERR
"ERROR: domain->span does not contain "
5182 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5183 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5187 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5191 printk(KERN_ERR
"ERROR: group is NULL\n");
5196 * Even though we initialize ->power to something semi-sane,
5197 * we leave power_orig unset. This allows us to detect if
5198 * domain iteration is still funny without causing /0 traps.
5200 if (!group
->sgp
->power_orig
) {
5201 printk(KERN_CONT
"\n");
5202 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5207 if (!cpumask_weight(sched_group_cpus(group
))) {
5208 printk(KERN_CONT
"\n");
5209 printk(KERN_ERR
"ERROR: empty group\n");
5213 if (!(sd
->flags
& SD_OVERLAP
) &&
5214 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5215 printk(KERN_CONT
"\n");
5216 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5220 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5222 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5224 printk(KERN_CONT
" %s", str
);
5225 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5226 printk(KERN_CONT
" (cpu_power = %d)",
5230 group
= group
->next
;
5231 } while (group
!= sd
->groups
);
5232 printk(KERN_CONT
"\n");
5234 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5235 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5238 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5239 printk(KERN_ERR
"ERROR: parent span is not a superset "
5240 "of domain->span\n");
5244 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5248 if (!sched_debug_enabled
)
5252 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5256 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5259 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5267 #else /* !CONFIG_SCHED_DEBUG */
5268 # define sched_domain_debug(sd, cpu) do { } while (0)
5269 static inline bool sched_debug(void)
5273 #endif /* CONFIG_SCHED_DEBUG */
5275 static int sd_degenerate(struct sched_domain
*sd
)
5277 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5280 /* Following flags need at least 2 groups */
5281 if (sd
->flags
& (SD_LOAD_BALANCE
|
5282 SD_BALANCE_NEWIDLE
|
5286 SD_SHARE_PKG_RESOURCES
)) {
5287 if (sd
->groups
!= sd
->groups
->next
)
5291 /* Following flags don't use groups */
5292 if (sd
->flags
& (SD_WAKE_AFFINE
))
5299 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5301 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5303 if (sd_degenerate(parent
))
5306 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5309 /* Flags needing groups don't count if only 1 group in parent */
5310 if (parent
->groups
== parent
->groups
->next
) {
5311 pflags
&= ~(SD_LOAD_BALANCE
|
5312 SD_BALANCE_NEWIDLE
|
5316 SD_SHARE_PKG_RESOURCES
);
5317 if (nr_node_ids
== 1)
5318 pflags
&= ~SD_SERIALIZE
;
5320 if (~cflags
& pflags
)
5326 static void free_rootdomain(struct rcu_head
*rcu
)
5328 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5330 cpupri_cleanup(&rd
->cpupri
);
5331 free_cpumask_var(rd
->rto_mask
);
5332 free_cpumask_var(rd
->online
);
5333 free_cpumask_var(rd
->span
);
5337 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5339 struct root_domain
*old_rd
= NULL
;
5340 unsigned long flags
;
5342 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5347 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5350 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5353 * If we dont want to free the old_rt yet then
5354 * set old_rd to NULL to skip the freeing later
5357 if (!atomic_dec_and_test(&old_rd
->refcount
))
5361 atomic_inc(&rd
->refcount
);
5364 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5365 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5368 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5371 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5374 static int init_rootdomain(struct root_domain
*rd
)
5376 memset(rd
, 0, sizeof(*rd
));
5378 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5380 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5382 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5385 if (cpupri_init(&rd
->cpupri
) != 0)
5390 free_cpumask_var(rd
->rto_mask
);
5392 free_cpumask_var(rd
->online
);
5394 free_cpumask_var(rd
->span
);
5400 * By default the system creates a single root-domain with all cpus as
5401 * members (mimicking the global state we have today).
5403 struct root_domain def_root_domain
;
5405 static void init_defrootdomain(void)
5407 init_rootdomain(&def_root_domain
);
5409 atomic_set(&def_root_domain
.refcount
, 1);
5412 static struct root_domain
*alloc_rootdomain(void)
5414 struct root_domain
*rd
;
5416 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5420 if (init_rootdomain(rd
) != 0) {
5428 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5430 struct sched_group
*tmp
, *first
;
5439 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5444 } while (sg
!= first
);
5447 static void free_sched_domain(struct rcu_head
*rcu
)
5449 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5452 * If its an overlapping domain it has private groups, iterate and
5455 if (sd
->flags
& SD_OVERLAP
) {
5456 free_sched_groups(sd
->groups
, 1);
5457 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5458 kfree(sd
->groups
->sgp
);
5464 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5466 call_rcu(&sd
->rcu
, free_sched_domain
);
5469 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5471 for (; sd
; sd
= sd
->parent
)
5472 destroy_sched_domain(sd
, cpu
);
5476 * Keep a special pointer to the highest sched_domain that has
5477 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5478 * allows us to avoid some pointer chasing select_idle_sibling().
5480 * Also keep a unique ID per domain (we use the first cpu number in
5481 * the cpumask of the domain), this allows us to quickly tell if
5482 * two cpus are in the same cache domain, see cpus_share_cache().
5484 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5485 DEFINE_PER_CPU(int, sd_llc_id
);
5487 static void update_top_cache_domain(int cpu
)
5489 struct sched_domain
*sd
;
5492 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5494 id
= cpumask_first(sched_domain_span(sd
));
5496 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5497 per_cpu(sd_llc_id
, cpu
) = id
;
5501 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5502 * hold the hotplug lock.
5505 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5507 struct rq
*rq
= cpu_rq(cpu
);
5508 struct sched_domain
*tmp
;
5510 /* Remove the sched domains which do not contribute to scheduling. */
5511 for (tmp
= sd
; tmp
; ) {
5512 struct sched_domain
*parent
= tmp
->parent
;
5516 if (sd_parent_degenerate(tmp
, parent
)) {
5517 tmp
->parent
= parent
->parent
;
5519 parent
->parent
->child
= tmp
;
5520 destroy_sched_domain(parent
, cpu
);
5525 if (sd
&& sd_degenerate(sd
)) {
5528 destroy_sched_domain(tmp
, cpu
);
5533 sched_domain_debug(sd
, cpu
);
5535 rq_attach_root(rq
, rd
);
5537 rcu_assign_pointer(rq
->sd
, sd
);
5538 destroy_sched_domains(tmp
, cpu
);
5540 update_top_cache_domain(cpu
);
5543 /* cpus with isolated domains */
5544 static cpumask_var_t cpu_isolated_map
;
5546 /* Setup the mask of cpus configured for isolated domains */
5547 static int __init
isolated_cpu_setup(char *str
)
5549 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5550 cpulist_parse(str
, cpu_isolated_map
);
5554 __setup("isolcpus=", isolated_cpu_setup
);
5556 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5558 return cpumask_of_node(cpu_to_node(cpu
));
5562 struct sched_domain
**__percpu sd
;
5563 struct sched_group
**__percpu sg
;
5564 struct sched_group_power
**__percpu sgp
;
5568 struct sched_domain
** __percpu sd
;
5569 struct root_domain
*rd
;
5579 struct sched_domain_topology_level
;
5581 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5582 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5584 #define SDTL_OVERLAP 0x01
5586 struct sched_domain_topology_level
{
5587 sched_domain_init_f init
;
5588 sched_domain_mask_f mask
;
5591 struct sd_data data
;
5595 * Build an iteration mask that can exclude certain CPUs from the upwards
5598 * Asymmetric node setups can result in situations where the domain tree is of
5599 * unequal depth, make sure to skip domains that already cover the entire
5602 * In that case build_sched_domains() will have terminated the iteration early
5603 * and our sibling sd spans will be empty. Domains should always include the
5604 * cpu they're built on, so check that.
5607 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5609 const struct cpumask
*span
= sched_domain_span(sd
);
5610 struct sd_data
*sdd
= sd
->private;
5611 struct sched_domain
*sibling
;
5614 for_each_cpu(i
, span
) {
5615 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5616 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5619 cpumask_set_cpu(i
, sched_group_mask(sg
));
5624 * Return the canonical balance cpu for this group, this is the first cpu
5625 * of this group that's also in the iteration mask.
5627 int group_balance_cpu(struct sched_group
*sg
)
5629 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5633 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5635 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5636 const struct cpumask
*span
= sched_domain_span(sd
);
5637 struct cpumask
*covered
= sched_domains_tmpmask
;
5638 struct sd_data
*sdd
= sd
->private;
5639 struct sched_domain
*child
;
5642 cpumask_clear(covered
);
5644 for_each_cpu(i
, span
) {
5645 struct cpumask
*sg_span
;
5647 if (cpumask_test_cpu(i
, covered
))
5650 child
= *per_cpu_ptr(sdd
->sd
, i
);
5652 /* See the comment near build_group_mask(). */
5653 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5656 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5657 GFP_KERNEL
, cpu_to_node(cpu
));
5662 sg_span
= sched_group_cpus(sg
);
5664 child
= child
->child
;
5665 cpumask_copy(sg_span
, sched_domain_span(child
));
5667 cpumask_set_cpu(i
, sg_span
);
5669 cpumask_or(covered
, covered
, sg_span
);
5671 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5672 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5673 build_group_mask(sd
, sg
);
5676 * Initialize sgp->power such that even if we mess up the
5677 * domains and no possible iteration will get us here, we won't
5680 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5683 * Make sure the first group of this domain contains the
5684 * canonical balance cpu. Otherwise the sched_domain iteration
5685 * breaks. See update_sg_lb_stats().
5687 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5688 group_balance_cpu(sg
) == cpu
)
5698 sd
->groups
= groups
;
5703 free_sched_groups(first
, 0);
5708 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5710 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5711 struct sched_domain
*child
= sd
->child
;
5714 cpu
= cpumask_first(sched_domain_span(child
));
5717 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5718 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5719 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5726 * build_sched_groups will build a circular linked list of the groups
5727 * covered by the given span, and will set each group's ->cpumask correctly,
5728 * and ->cpu_power to 0.
5730 * Assumes the sched_domain tree is fully constructed
5733 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5735 struct sched_group
*first
= NULL
, *last
= NULL
;
5736 struct sd_data
*sdd
= sd
->private;
5737 const struct cpumask
*span
= sched_domain_span(sd
);
5738 struct cpumask
*covered
;
5741 get_group(cpu
, sdd
, &sd
->groups
);
5742 atomic_inc(&sd
->groups
->ref
);
5744 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5747 lockdep_assert_held(&sched_domains_mutex
);
5748 covered
= sched_domains_tmpmask
;
5750 cpumask_clear(covered
);
5752 for_each_cpu(i
, span
) {
5753 struct sched_group
*sg
;
5754 int group
= get_group(i
, sdd
, &sg
);
5757 if (cpumask_test_cpu(i
, covered
))
5760 cpumask_clear(sched_group_cpus(sg
));
5762 cpumask_setall(sched_group_mask(sg
));
5764 for_each_cpu(j
, span
) {
5765 if (get_group(j
, sdd
, NULL
) != group
)
5768 cpumask_set_cpu(j
, covered
);
5769 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5784 * Initialize sched groups cpu_power.
5786 * cpu_power indicates the capacity of sched group, which is used while
5787 * distributing the load between different sched groups in a sched domain.
5788 * Typically cpu_power for all the groups in a sched domain will be same unless
5789 * there are asymmetries in the topology. If there are asymmetries, group
5790 * having more cpu_power will pickup more load compared to the group having
5793 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5795 struct sched_group
*sg
= sd
->groups
;
5797 WARN_ON(!sd
|| !sg
);
5800 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5802 } while (sg
!= sd
->groups
);
5804 if (cpu
!= group_balance_cpu(sg
))
5807 update_group_power(sd
, cpu
);
5808 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5811 int __weak
arch_sd_sibling_asym_packing(void)
5813 return 0*SD_ASYM_PACKING
;
5817 * Initializers for schedule domains
5818 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5821 #ifdef CONFIG_SCHED_DEBUG
5822 # define SD_INIT_NAME(sd, type) sd->name = #type
5824 # define SD_INIT_NAME(sd, type) do { } while (0)
5827 #define SD_INIT_FUNC(type) \
5828 static noinline struct sched_domain * \
5829 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5831 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5832 *sd = SD_##type##_INIT; \
5833 SD_INIT_NAME(sd, type); \
5834 sd->private = &tl->data; \
5839 #ifdef CONFIG_SCHED_SMT
5840 SD_INIT_FUNC(SIBLING
)
5842 #ifdef CONFIG_SCHED_MC
5845 #ifdef CONFIG_SCHED_BOOK
5849 static int default_relax_domain_level
= -1;
5850 int sched_domain_level_max
;
5852 static int __init
setup_relax_domain_level(char *str
)
5854 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5855 pr_warn("Unable to set relax_domain_level\n");
5859 __setup("relax_domain_level=", setup_relax_domain_level
);
5861 static void set_domain_attribute(struct sched_domain
*sd
,
5862 struct sched_domain_attr
*attr
)
5866 if (!attr
|| attr
->relax_domain_level
< 0) {
5867 if (default_relax_domain_level
< 0)
5870 request
= default_relax_domain_level
;
5872 request
= attr
->relax_domain_level
;
5873 if (request
< sd
->level
) {
5874 /* turn off idle balance on this domain */
5875 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5877 /* turn on idle balance on this domain */
5878 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5882 static void __sdt_free(const struct cpumask
*cpu_map
);
5883 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5885 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5886 const struct cpumask
*cpu_map
)
5890 if (!atomic_read(&d
->rd
->refcount
))
5891 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5893 free_percpu(d
->sd
); /* fall through */
5895 __sdt_free(cpu_map
); /* fall through */
5901 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5902 const struct cpumask
*cpu_map
)
5904 memset(d
, 0, sizeof(*d
));
5906 if (__sdt_alloc(cpu_map
))
5907 return sa_sd_storage
;
5908 d
->sd
= alloc_percpu(struct sched_domain
*);
5910 return sa_sd_storage
;
5911 d
->rd
= alloc_rootdomain();
5914 return sa_rootdomain
;
5918 * NULL the sd_data elements we've used to build the sched_domain and
5919 * sched_group structure so that the subsequent __free_domain_allocs()
5920 * will not free the data we're using.
5922 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5924 struct sd_data
*sdd
= sd
->private;
5926 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5927 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5929 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5930 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5932 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5933 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5936 #ifdef CONFIG_SCHED_SMT
5937 static const struct cpumask
*cpu_smt_mask(int cpu
)
5939 return topology_thread_cpumask(cpu
);
5944 * Topology list, bottom-up.
5946 static struct sched_domain_topology_level default_topology
[] = {
5947 #ifdef CONFIG_SCHED_SMT
5948 { sd_init_SIBLING
, cpu_smt_mask
, },
5950 #ifdef CONFIG_SCHED_MC
5951 { sd_init_MC
, cpu_coregroup_mask
, },
5953 #ifdef CONFIG_SCHED_BOOK
5954 { sd_init_BOOK
, cpu_book_mask
, },
5956 { sd_init_CPU
, cpu_cpu_mask
, },
5960 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5964 static int sched_domains_numa_levels
;
5965 static int *sched_domains_numa_distance
;
5966 static struct cpumask
***sched_domains_numa_masks
;
5967 static int sched_domains_curr_level
;
5969 static inline int sd_local_flags(int level
)
5971 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5974 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
5977 static struct sched_domain
*
5978 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
5980 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5981 int level
= tl
->numa_level
;
5982 int sd_weight
= cpumask_weight(
5983 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
5985 *sd
= (struct sched_domain
){
5986 .min_interval
= sd_weight
,
5987 .max_interval
= 2*sd_weight
,
5989 .imbalance_pct
= 125,
5990 .cache_nice_tries
= 2,
5997 .flags
= 1*SD_LOAD_BALANCE
5998 | 1*SD_BALANCE_NEWIDLE
6003 | 0*SD_SHARE_CPUPOWER
6004 | 0*SD_SHARE_PKG_RESOURCES
6006 | 0*SD_PREFER_SIBLING
6007 | sd_local_flags(level
)
6009 .last_balance
= jiffies
,
6010 .balance_interval
= sd_weight
,
6012 SD_INIT_NAME(sd
, NUMA
);
6013 sd
->private = &tl
->data
;
6016 * Ugly hack to pass state to sd_numa_mask()...
6018 sched_domains_curr_level
= tl
->numa_level
;
6023 static const struct cpumask
*sd_numa_mask(int cpu
)
6025 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6028 static void sched_numa_warn(const char *str
)
6030 static int done
= false;
6038 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6040 for (i
= 0; i
< nr_node_ids
; i
++) {
6041 printk(KERN_WARNING
" ");
6042 for (j
= 0; j
< nr_node_ids
; j
++)
6043 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6044 printk(KERN_CONT
"\n");
6046 printk(KERN_WARNING
"\n");
6049 static bool find_numa_distance(int distance
)
6053 if (distance
== node_distance(0, 0))
6056 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6057 if (sched_domains_numa_distance
[i
] == distance
)
6064 static void sched_init_numa(void)
6066 int next_distance
, curr_distance
= node_distance(0, 0);
6067 struct sched_domain_topology_level
*tl
;
6071 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6072 if (!sched_domains_numa_distance
)
6076 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6077 * unique distances in the node_distance() table.
6079 * Assumes node_distance(0,j) includes all distances in
6080 * node_distance(i,j) in order to avoid cubic time.
6082 next_distance
= curr_distance
;
6083 for (i
= 0; i
< nr_node_ids
; i
++) {
6084 for (j
= 0; j
< nr_node_ids
; j
++) {
6085 for (k
= 0; k
< nr_node_ids
; k
++) {
6086 int distance
= node_distance(i
, k
);
6088 if (distance
> curr_distance
&&
6089 (distance
< next_distance
||
6090 next_distance
== curr_distance
))
6091 next_distance
= distance
;
6094 * While not a strong assumption it would be nice to know
6095 * about cases where if node A is connected to B, B is not
6096 * equally connected to A.
6098 if (sched_debug() && node_distance(k
, i
) != distance
)
6099 sched_numa_warn("Node-distance not symmetric");
6101 if (sched_debug() && i
&& !find_numa_distance(distance
))
6102 sched_numa_warn("Node-0 not representative");
6104 if (next_distance
!= curr_distance
) {
6105 sched_domains_numa_distance
[level
++] = next_distance
;
6106 sched_domains_numa_levels
= level
;
6107 curr_distance
= next_distance
;
6112 * In case of sched_debug() we verify the above assumption.
6118 * 'level' contains the number of unique distances, excluding the
6119 * identity distance node_distance(i,i).
6121 * The sched_domains_nume_distance[] array includes the actual distance
6125 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6126 if (!sched_domains_numa_masks
)
6130 * Now for each level, construct a mask per node which contains all
6131 * cpus of nodes that are that many hops away from us.
6133 for (i
= 0; i
< level
; i
++) {
6134 sched_domains_numa_masks
[i
] =
6135 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6136 if (!sched_domains_numa_masks
[i
])
6139 for (j
= 0; j
< nr_node_ids
; j
++) {
6140 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6144 sched_domains_numa_masks
[i
][j
] = mask
;
6146 for (k
= 0; k
< nr_node_ids
; k
++) {
6147 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6150 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6155 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6156 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6161 * Copy the default topology bits..
6163 for (i
= 0; default_topology
[i
].init
; i
++)
6164 tl
[i
] = default_topology
[i
];
6167 * .. and append 'j' levels of NUMA goodness.
6169 for (j
= 0; j
< level
; i
++, j
++) {
6170 tl
[i
] = (struct sched_domain_topology_level
){
6171 .init
= sd_numa_init
,
6172 .mask
= sd_numa_mask
,
6173 .flags
= SDTL_OVERLAP
,
6178 sched_domain_topology
= tl
;
6181 static inline void sched_init_numa(void)
6184 #endif /* CONFIG_NUMA */
6186 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6188 struct sched_domain_topology_level
*tl
;
6191 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6192 struct sd_data
*sdd
= &tl
->data
;
6194 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6198 sdd
->sg
= alloc_percpu(struct sched_group
*);
6202 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6206 for_each_cpu(j
, cpu_map
) {
6207 struct sched_domain
*sd
;
6208 struct sched_group
*sg
;
6209 struct sched_group_power
*sgp
;
6211 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6212 GFP_KERNEL
, cpu_to_node(j
));
6216 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6218 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6219 GFP_KERNEL
, cpu_to_node(j
));
6225 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6227 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6228 GFP_KERNEL
, cpu_to_node(j
));
6232 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6239 static void __sdt_free(const struct cpumask
*cpu_map
)
6241 struct sched_domain_topology_level
*tl
;
6244 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6245 struct sd_data
*sdd
= &tl
->data
;
6247 for_each_cpu(j
, cpu_map
) {
6248 struct sched_domain
*sd
;
6251 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6252 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6253 free_sched_groups(sd
->groups
, 0);
6254 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6258 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6260 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6262 free_percpu(sdd
->sd
);
6264 free_percpu(sdd
->sg
);
6266 free_percpu(sdd
->sgp
);
6271 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6272 struct s_data
*d
, const struct cpumask
*cpu_map
,
6273 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6276 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6280 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6282 sd
->level
= child
->level
+ 1;
6283 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6287 set_domain_attribute(sd
, attr
);
6293 * Build sched domains for a given set of cpus and attach the sched domains
6294 * to the individual cpus
6296 static int build_sched_domains(const struct cpumask
*cpu_map
,
6297 struct sched_domain_attr
*attr
)
6299 enum s_alloc alloc_state
= sa_none
;
6300 struct sched_domain
*sd
;
6302 int i
, ret
= -ENOMEM
;
6304 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6305 if (alloc_state
!= sa_rootdomain
)
6308 /* Set up domains for cpus specified by the cpu_map. */
6309 for_each_cpu(i
, cpu_map
) {
6310 struct sched_domain_topology_level
*tl
;
6313 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6314 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6315 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6316 sd
->flags
|= SD_OVERLAP
;
6317 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6324 *per_cpu_ptr(d
.sd
, i
) = sd
;
6327 /* Build the groups for the domains */
6328 for_each_cpu(i
, cpu_map
) {
6329 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6330 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6331 if (sd
->flags
& SD_OVERLAP
) {
6332 if (build_overlap_sched_groups(sd
, i
))
6335 if (build_sched_groups(sd
, i
))
6341 /* Calculate CPU power for physical packages and nodes */
6342 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6343 if (!cpumask_test_cpu(i
, cpu_map
))
6346 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6347 claim_allocations(i
, sd
);
6348 init_sched_groups_power(i
, sd
);
6352 /* Attach the domains */
6354 for_each_cpu(i
, cpu_map
) {
6355 sd
= *per_cpu_ptr(d
.sd
, i
);
6356 cpu_attach_domain(sd
, d
.rd
, i
);
6362 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6366 static cpumask_var_t
*doms_cur
; /* current sched domains */
6367 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6368 static struct sched_domain_attr
*dattr_cur
;
6369 /* attribues of custom domains in 'doms_cur' */
6372 * Special case: If a kmalloc of a doms_cur partition (array of
6373 * cpumask) fails, then fallback to a single sched domain,
6374 * as determined by the single cpumask fallback_doms.
6376 static cpumask_var_t fallback_doms
;
6379 * arch_update_cpu_topology lets virtualized architectures update the
6380 * cpu core maps. It is supposed to return 1 if the topology changed
6381 * or 0 if it stayed the same.
6383 int __attribute__((weak
)) arch_update_cpu_topology(void)
6388 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6391 cpumask_var_t
*doms
;
6393 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6396 for (i
= 0; i
< ndoms
; i
++) {
6397 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6398 free_sched_domains(doms
, i
);
6405 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6408 for (i
= 0; i
< ndoms
; i
++)
6409 free_cpumask_var(doms
[i
]);
6414 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6415 * For now this just excludes isolated cpus, but could be used to
6416 * exclude other special cases in the future.
6418 static int init_sched_domains(const struct cpumask
*cpu_map
)
6422 arch_update_cpu_topology();
6424 doms_cur
= alloc_sched_domains(ndoms_cur
);
6426 doms_cur
= &fallback_doms
;
6427 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6428 err
= build_sched_domains(doms_cur
[0], NULL
);
6429 register_sched_domain_sysctl();
6435 * Detach sched domains from a group of cpus specified in cpu_map
6436 * These cpus will now be attached to the NULL domain
6438 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6443 for_each_cpu(i
, cpu_map
)
6444 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6448 /* handle null as "default" */
6449 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6450 struct sched_domain_attr
*new, int idx_new
)
6452 struct sched_domain_attr tmp
;
6459 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6460 new ? (new + idx_new
) : &tmp
,
6461 sizeof(struct sched_domain_attr
));
6465 * Partition sched domains as specified by the 'ndoms_new'
6466 * cpumasks in the array doms_new[] of cpumasks. This compares
6467 * doms_new[] to the current sched domain partitioning, doms_cur[].
6468 * It destroys each deleted domain and builds each new domain.
6470 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6471 * The masks don't intersect (don't overlap.) We should setup one
6472 * sched domain for each mask. CPUs not in any of the cpumasks will
6473 * not be load balanced. If the same cpumask appears both in the
6474 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6477 * The passed in 'doms_new' should be allocated using
6478 * alloc_sched_domains. This routine takes ownership of it and will
6479 * free_sched_domains it when done with it. If the caller failed the
6480 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6481 * and partition_sched_domains() will fallback to the single partition
6482 * 'fallback_doms', it also forces the domains to be rebuilt.
6484 * If doms_new == NULL it will be replaced with cpu_online_mask.
6485 * ndoms_new == 0 is a special case for destroying existing domains,
6486 * and it will not create the default domain.
6488 * Call with hotplug lock held
6490 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6491 struct sched_domain_attr
*dattr_new
)
6496 mutex_lock(&sched_domains_mutex
);
6498 /* always unregister in case we don't destroy any domains */
6499 unregister_sched_domain_sysctl();
6501 /* Let architecture update cpu core mappings. */
6502 new_topology
= arch_update_cpu_topology();
6504 n
= doms_new
? ndoms_new
: 0;
6506 /* Destroy deleted domains */
6507 for (i
= 0; i
< ndoms_cur
; i
++) {
6508 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6509 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6510 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6513 /* no match - a current sched domain not in new doms_new[] */
6514 detach_destroy_domains(doms_cur
[i
]);
6519 if (doms_new
== NULL
) {
6521 doms_new
= &fallback_doms
;
6522 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6523 WARN_ON_ONCE(dattr_new
);
6526 /* Build new domains */
6527 for (i
= 0; i
< ndoms_new
; i
++) {
6528 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6529 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6530 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6533 /* no match - add a new doms_new */
6534 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6539 /* Remember the new sched domains */
6540 if (doms_cur
!= &fallback_doms
)
6541 free_sched_domains(doms_cur
, ndoms_cur
);
6542 kfree(dattr_cur
); /* kfree(NULL) is safe */
6543 doms_cur
= doms_new
;
6544 dattr_cur
= dattr_new
;
6545 ndoms_cur
= ndoms_new
;
6547 register_sched_domain_sysctl();
6549 mutex_unlock(&sched_domains_mutex
);
6552 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6555 * Update cpusets according to cpu_active mask. If cpusets are
6556 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6557 * around partition_sched_domains().
6559 * If we come here as part of a suspend/resume, don't touch cpusets because we
6560 * want to restore it back to its original state upon resume anyway.
6562 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6566 case CPU_ONLINE_FROZEN
:
6567 case CPU_DOWN_FAILED_FROZEN
:
6570 * num_cpus_frozen tracks how many CPUs are involved in suspend
6571 * resume sequence. As long as this is not the last online
6572 * operation in the resume sequence, just build a single sched
6573 * domain, ignoring cpusets.
6576 if (likely(num_cpus_frozen
)) {
6577 partition_sched_domains(1, NULL
, NULL
);
6582 * This is the last CPU online operation. So fall through and
6583 * restore the original sched domains by considering the
6584 * cpuset configurations.
6588 case CPU_DOWN_FAILED
:
6589 cpuset_update_active_cpus(true);
6597 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6601 case CPU_DOWN_PREPARE
:
6602 cpuset_update_active_cpus(false);
6604 case CPU_DOWN_PREPARE_FROZEN
:
6606 partition_sched_domains(1, NULL
, NULL
);
6614 void __init
sched_init_smp(void)
6616 cpumask_var_t non_isolated_cpus
;
6618 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6619 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6624 mutex_lock(&sched_domains_mutex
);
6625 init_sched_domains(cpu_active_mask
);
6626 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6627 if (cpumask_empty(non_isolated_cpus
))
6628 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6629 mutex_unlock(&sched_domains_mutex
);
6632 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6633 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6635 /* RT runtime code needs to handle some hotplug events */
6636 hotcpu_notifier(update_runtime
, 0);
6640 /* Move init over to a non-isolated CPU */
6641 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6643 sched_init_granularity();
6644 free_cpumask_var(non_isolated_cpus
);
6646 init_sched_rt_class();
6649 void __init
sched_init_smp(void)
6651 sched_init_granularity();
6653 #endif /* CONFIG_SMP */
6655 const_debug
unsigned int sysctl_timer_migration
= 1;
6657 int in_sched_functions(unsigned long addr
)
6659 return in_lock_functions(addr
) ||
6660 (addr
>= (unsigned long)__sched_text_start
6661 && addr
< (unsigned long)__sched_text_end
);
6664 #ifdef CONFIG_CGROUP_SCHED
6665 struct task_group root_task_group
;
6666 LIST_HEAD(task_groups
);
6669 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6671 void __init
sched_init(void)
6674 unsigned long alloc_size
= 0, ptr
;
6676 #ifdef CONFIG_FAIR_GROUP_SCHED
6677 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6679 #ifdef CONFIG_RT_GROUP_SCHED
6680 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6682 #ifdef CONFIG_CPUMASK_OFFSTACK
6683 alloc_size
+= num_possible_cpus() * cpumask_size();
6686 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6688 #ifdef CONFIG_FAIR_GROUP_SCHED
6689 root_task_group
.se
= (struct sched_entity
**)ptr
;
6690 ptr
+= nr_cpu_ids
* sizeof(void **);
6692 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6693 ptr
+= nr_cpu_ids
* sizeof(void **);
6695 #endif /* CONFIG_FAIR_GROUP_SCHED */
6696 #ifdef CONFIG_RT_GROUP_SCHED
6697 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6698 ptr
+= nr_cpu_ids
* sizeof(void **);
6700 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6701 ptr
+= nr_cpu_ids
* sizeof(void **);
6703 #endif /* CONFIG_RT_GROUP_SCHED */
6704 #ifdef CONFIG_CPUMASK_OFFSTACK
6705 for_each_possible_cpu(i
) {
6706 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6707 ptr
+= cpumask_size();
6709 #endif /* CONFIG_CPUMASK_OFFSTACK */
6713 init_defrootdomain();
6716 init_rt_bandwidth(&def_rt_bandwidth
,
6717 global_rt_period(), global_rt_runtime());
6719 #ifdef CONFIG_RT_GROUP_SCHED
6720 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6721 global_rt_period(), global_rt_runtime());
6722 #endif /* CONFIG_RT_GROUP_SCHED */
6724 #ifdef CONFIG_CGROUP_SCHED
6725 list_add(&root_task_group
.list
, &task_groups
);
6726 INIT_LIST_HEAD(&root_task_group
.children
);
6727 INIT_LIST_HEAD(&root_task_group
.siblings
);
6728 autogroup_init(&init_task
);
6730 #endif /* CONFIG_CGROUP_SCHED */
6732 #ifdef CONFIG_CGROUP_CPUACCT
6733 root_cpuacct
.cpustat
= &kernel_cpustat
;
6734 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6735 /* Too early, not expected to fail */
6736 BUG_ON(!root_cpuacct
.cpuusage
);
6738 for_each_possible_cpu(i
) {
6742 raw_spin_lock_init(&rq
->lock
);
6744 rq
->calc_load_active
= 0;
6745 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6746 init_cfs_rq(&rq
->cfs
);
6747 init_rt_rq(&rq
->rt
, rq
);
6748 #ifdef CONFIG_FAIR_GROUP_SCHED
6749 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6750 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6752 * How much cpu bandwidth does root_task_group get?
6754 * In case of task-groups formed thr' the cgroup filesystem, it
6755 * gets 100% of the cpu resources in the system. This overall
6756 * system cpu resource is divided among the tasks of
6757 * root_task_group and its child task-groups in a fair manner,
6758 * based on each entity's (task or task-group's) weight
6759 * (se->load.weight).
6761 * In other words, if root_task_group has 10 tasks of weight
6762 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6763 * then A0's share of the cpu resource is:
6765 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6767 * We achieve this by letting root_task_group's tasks sit
6768 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6770 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6771 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6772 #endif /* CONFIG_FAIR_GROUP_SCHED */
6774 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6775 #ifdef CONFIG_RT_GROUP_SCHED
6776 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6777 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6780 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6781 rq
->cpu_load
[j
] = 0;
6783 rq
->last_load_update_tick
= jiffies
;
6788 rq
->cpu_power
= SCHED_POWER_SCALE
;
6789 rq
->post_schedule
= 0;
6790 rq
->active_balance
= 0;
6791 rq
->next_balance
= jiffies
;
6796 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6798 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6800 rq_attach_root(rq
, &def_root_domain
);
6806 atomic_set(&rq
->nr_iowait
, 0);
6809 set_load_weight(&init_task
);
6811 #ifdef CONFIG_PREEMPT_NOTIFIERS
6812 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6815 #ifdef CONFIG_RT_MUTEXES
6816 plist_head_init(&init_task
.pi_waiters
);
6820 * The boot idle thread does lazy MMU switching as well:
6822 atomic_inc(&init_mm
.mm_count
);
6823 enter_lazy_tlb(&init_mm
, current
);
6826 * Make us the idle thread. Technically, schedule() should not be
6827 * called from this thread, however somewhere below it might be,
6828 * but because we are the idle thread, we just pick up running again
6829 * when this runqueue becomes "idle".
6831 init_idle(current
, smp_processor_id());
6833 calc_load_update
= jiffies
+ LOAD_FREQ
;
6836 * During early bootup we pretend to be a normal task:
6838 current
->sched_class
= &fair_sched_class
;
6841 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6842 /* May be allocated at isolcpus cmdline parse time */
6843 if (cpu_isolated_map
== NULL
)
6844 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6845 idle_thread_set_boot_cpu();
6847 init_sched_fair_class();
6849 scheduler_running
= 1;
6852 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6853 static inline int preempt_count_equals(int preempt_offset
)
6855 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6857 return (nested
== preempt_offset
);
6860 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6862 static unsigned long prev_jiffy
; /* ratelimiting */
6864 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6865 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6866 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6868 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6870 prev_jiffy
= jiffies
;
6873 "BUG: sleeping function called from invalid context at %s:%d\n",
6876 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6877 in_atomic(), irqs_disabled(),
6878 current
->pid
, current
->comm
);
6880 debug_show_held_locks(current
);
6881 if (irqs_disabled())
6882 print_irqtrace_events(current
);
6885 EXPORT_SYMBOL(__might_sleep
);
6888 #ifdef CONFIG_MAGIC_SYSRQ
6889 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6891 const struct sched_class
*prev_class
= p
->sched_class
;
6892 int old_prio
= p
->prio
;
6897 dequeue_task(rq
, p
, 0);
6898 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6900 enqueue_task(rq
, p
, 0);
6901 resched_task(rq
->curr
);
6904 check_class_changed(rq
, p
, prev_class
, old_prio
);
6907 void normalize_rt_tasks(void)
6909 struct task_struct
*g
, *p
;
6910 unsigned long flags
;
6913 read_lock_irqsave(&tasklist_lock
, flags
);
6914 do_each_thread(g
, p
) {
6916 * Only normalize user tasks:
6921 p
->se
.exec_start
= 0;
6922 #ifdef CONFIG_SCHEDSTATS
6923 p
->se
.statistics
.wait_start
= 0;
6924 p
->se
.statistics
.sleep_start
= 0;
6925 p
->se
.statistics
.block_start
= 0;
6930 * Renice negative nice level userspace
6933 if (TASK_NICE(p
) < 0 && p
->mm
)
6934 set_user_nice(p
, 0);
6938 raw_spin_lock(&p
->pi_lock
);
6939 rq
= __task_rq_lock(p
);
6941 normalize_task(rq
, p
);
6943 __task_rq_unlock(rq
);
6944 raw_spin_unlock(&p
->pi_lock
);
6945 } while_each_thread(g
, p
);
6947 read_unlock_irqrestore(&tasklist_lock
, flags
);
6950 #endif /* CONFIG_MAGIC_SYSRQ */
6952 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6954 * These functions are only useful for the IA64 MCA handling, or kdb.
6956 * They can only be called when the whole system has been
6957 * stopped - every CPU needs to be quiescent, and no scheduling
6958 * activity can take place. Using them for anything else would
6959 * be a serious bug, and as a result, they aren't even visible
6960 * under any other configuration.
6964 * curr_task - return the current task for a given cpu.
6965 * @cpu: the processor in question.
6967 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6969 struct task_struct
*curr_task(int cpu
)
6971 return cpu_curr(cpu
);
6974 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6978 * set_curr_task - set the current task for a given cpu.
6979 * @cpu: the processor in question.
6980 * @p: the task pointer to set.
6982 * Description: This function must only be used when non-maskable interrupts
6983 * are serviced on a separate stack. It allows the architecture to switch the
6984 * notion of the current task on a cpu in a non-blocking manner. This function
6985 * must be called with all CPU's synchronized, and interrupts disabled, the
6986 * and caller must save the original value of the current task (see
6987 * curr_task() above) and restore that value before reenabling interrupts and
6988 * re-starting the system.
6990 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6992 void set_curr_task(int cpu
, struct task_struct
*p
)
6999 #ifdef CONFIG_CGROUP_SCHED
7000 /* task_group_lock serializes the addition/removal of task groups */
7001 static DEFINE_SPINLOCK(task_group_lock
);
7003 static void free_sched_group(struct task_group
*tg
)
7005 free_fair_sched_group(tg
);
7006 free_rt_sched_group(tg
);
7011 /* allocate runqueue etc for a new task group */
7012 struct task_group
*sched_create_group(struct task_group
*parent
)
7014 struct task_group
*tg
;
7015 unsigned long flags
;
7017 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7019 return ERR_PTR(-ENOMEM
);
7021 if (!alloc_fair_sched_group(tg
, parent
))
7024 if (!alloc_rt_sched_group(tg
, parent
))
7027 spin_lock_irqsave(&task_group_lock
, flags
);
7028 list_add_rcu(&tg
->list
, &task_groups
);
7030 WARN_ON(!parent
); /* root should already exist */
7032 tg
->parent
= parent
;
7033 INIT_LIST_HEAD(&tg
->children
);
7034 list_add_rcu(&tg
->siblings
, &parent
->children
);
7035 spin_unlock_irqrestore(&task_group_lock
, flags
);
7040 free_sched_group(tg
);
7041 return ERR_PTR(-ENOMEM
);
7044 /* rcu callback to free various structures associated with a task group */
7045 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7047 /* now it should be safe to free those cfs_rqs */
7048 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7051 /* Destroy runqueue etc associated with a task group */
7052 void sched_destroy_group(struct task_group
*tg
)
7054 unsigned long flags
;
7057 /* end participation in shares distribution */
7058 for_each_possible_cpu(i
)
7059 unregister_fair_sched_group(tg
, i
);
7061 spin_lock_irqsave(&task_group_lock
, flags
);
7062 list_del_rcu(&tg
->list
);
7063 list_del_rcu(&tg
->siblings
);
7064 spin_unlock_irqrestore(&task_group_lock
, flags
);
7066 /* wait for possible concurrent references to cfs_rqs complete */
7067 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7070 /* change task's runqueue when it moves between groups.
7071 * The caller of this function should have put the task in its new group
7072 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7073 * reflect its new group.
7075 void sched_move_task(struct task_struct
*tsk
)
7077 struct task_group
*tg
;
7079 unsigned long flags
;
7082 rq
= task_rq_lock(tsk
, &flags
);
7084 running
= task_current(rq
, tsk
);
7088 dequeue_task(rq
, tsk
, 0);
7089 if (unlikely(running
))
7090 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7092 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7093 lockdep_is_held(&tsk
->sighand
->siglock
)),
7094 struct task_group
, css
);
7095 tg
= autogroup_task_group(tsk
, tg
);
7096 tsk
->sched_task_group
= tg
;
7098 #ifdef CONFIG_FAIR_GROUP_SCHED
7099 if (tsk
->sched_class
->task_move_group
)
7100 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7103 set_task_rq(tsk
, task_cpu(tsk
));
7105 if (unlikely(running
))
7106 tsk
->sched_class
->set_curr_task(rq
);
7108 enqueue_task(rq
, tsk
, 0);
7110 task_rq_unlock(rq
, tsk
, &flags
);
7112 #endif /* CONFIG_CGROUP_SCHED */
7114 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7115 static unsigned long to_ratio(u64 period
, u64 runtime
)
7117 if (runtime
== RUNTIME_INF
)
7120 return div64_u64(runtime
<< 20, period
);
7124 #ifdef CONFIG_RT_GROUP_SCHED
7126 * Ensure that the real time constraints are schedulable.
7128 static DEFINE_MUTEX(rt_constraints_mutex
);
7130 /* Must be called with tasklist_lock held */
7131 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7133 struct task_struct
*g
, *p
;
7135 do_each_thread(g
, p
) {
7136 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7138 } while_each_thread(g
, p
);
7143 struct rt_schedulable_data
{
7144 struct task_group
*tg
;
7149 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7151 struct rt_schedulable_data
*d
= data
;
7152 struct task_group
*child
;
7153 unsigned long total
, sum
= 0;
7154 u64 period
, runtime
;
7156 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7157 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7160 period
= d
->rt_period
;
7161 runtime
= d
->rt_runtime
;
7165 * Cannot have more runtime than the period.
7167 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7171 * Ensure we don't starve existing RT tasks.
7173 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7176 total
= to_ratio(period
, runtime
);
7179 * Nobody can have more than the global setting allows.
7181 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7185 * The sum of our children's runtime should not exceed our own.
7187 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7188 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7189 runtime
= child
->rt_bandwidth
.rt_runtime
;
7191 if (child
== d
->tg
) {
7192 period
= d
->rt_period
;
7193 runtime
= d
->rt_runtime
;
7196 sum
+= to_ratio(period
, runtime
);
7205 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7209 struct rt_schedulable_data data
= {
7211 .rt_period
= period
,
7212 .rt_runtime
= runtime
,
7216 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7222 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7223 u64 rt_period
, u64 rt_runtime
)
7227 mutex_lock(&rt_constraints_mutex
);
7228 read_lock(&tasklist_lock
);
7229 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7233 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7234 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7235 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7237 for_each_possible_cpu(i
) {
7238 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7240 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7241 rt_rq
->rt_runtime
= rt_runtime
;
7242 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7244 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7246 read_unlock(&tasklist_lock
);
7247 mutex_unlock(&rt_constraints_mutex
);
7252 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7254 u64 rt_runtime
, rt_period
;
7256 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7257 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7258 if (rt_runtime_us
< 0)
7259 rt_runtime
= RUNTIME_INF
;
7261 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7264 long sched_group_rt_runtime(struct task_group
*tg
)
7268 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7271 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7272 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7273 return rt_runtime_us
;
7276 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7278 u64 rt_runtime
, rt_period
;
7280 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7281 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7286 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7289 long sched_group_rt_period(struct task_group
*tg
)
7293 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7294 do_div(rt_period_us
, NSEC_PER_USEC
);
7295 return rt_period_us
;
7298 static int sched_rt_global_constraints(void)
7300 u64 runtime
, period
;
7303 if (sysctl_sched_rt_period
<= 0)
7306 runtime
= global_rt_runtime();
7307 period
= global_rt_period();
7310 * Sanity check on the sysctl variables.
7312 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7315 mutex_lock(&rt_constraints_mutex
);
7316 read_lock(&tasklist_lock
);
7317 ret
= __rt_schedulable(NULL
, 0, 0);
7318 read_unlock(&tasklist_lock
);
7319 mutex_unlock(&rt_constraints_mutex
);
7324 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7326 /* Don't accept realtime tasks when there is no way for them to run */
7327 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7333 #else /* !CONFIG_RT_GROUP_SCHED */
7334 static int sched_rt_global_constraints(void)
7336 unsigned long flags
;
7339 if (sysctl_sched_rt_period
<= 0)
7343 * There's always some RT tasks in the root group
7344 * -- migration, kstopmachine etc..
7346 if (sysctl_sched_rt_runtime
== 0)
7349 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7350 for_each_possible_cpu(i
) {
7351 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7353 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7354 rt_rq
->rt_runtime
= global_rt_runtime();
7355 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7357 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7361 #endif /* CONFIG_RT_GROUP_SCHED */
7363 int sched_rt_handler(struct ctl_table
*table
, int write
,
7364 void __user
*buffer
, size_t *lenp
,
7368 int old_period
, old_runtime
;
7369 static DEFINE_MUTEX(mutex
);
7372 old_period
= sysctl_sched_rt_period
;
7373 old_runtime
= sysctl_sched_rt_runtime
;
7375 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7377 if (!ret
&& write
) {
7378 ret
= sched_rt_global_constraints();
7380 sysctl_sched_rt_period
= old_period
;
7381 sysctl_sched_rt_runtime
= old_runtime
;
7383 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7384 def_rt_bandwidth
.rt_period
=
7385 ns_to_ktime(global_rt_period());
7388 mutex_unlock(&mutex
);
7393 #ifdef CONFIG_CGROUP_SCHED
7395 /* return corresponding task_group object of a cgroup */
7396 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7398 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7399 struct task_group
, css
);
7402 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7404 struct task_group
*tg
, *parent
;
7406 if (!cgrp
->parent
) {
7407 /* This is early initialization for the top cgroup */
7408 return &root_task_group
.css
;
7411 parent
= cgroup_tg(cgrp
->parent
);
7412 tg
= sched_create_group(parent
);
7414 return ERR_PTR(-ENOMEM
);
7419 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7421 struct task_group
*tg
= cgroup_tg(cgrp
);
7423 sched_destroy_group(tg
);
7426 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7427 struct cgroup_taskset
*tset
)
7429 struct task_struct
*task
;
7431 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7432 #ifdef CONFIG_RT_GROUP_SCHED
7433 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7436 /* We don't support RT-tasks being in separate groups */
7437 if (task
->sched_class
!= &fair_sched_class
)
7444 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7445 struct cgroup_taskset
*tset
)
7447 struct task_struct
*task
;
7449 cgroup_taskset_for_each(task
, cgrp
, tset
)
7450 sched_move_task(task
);
7454 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7455 struct task_struct
*task
)
7458 * cgroup_exit() is called in the copy_process() failure path.
7459 * Ignore this case since the task hasn't ran yet, this avoids
7460 * trying to poke a half freed task state from generic code.
7462 if (!(task
->flags
& PF_EXITING
))
7465 sched_move_task(task
);
7468 #ifdef CONFIG_FAIR_GROUP_SCHED
7469 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7472 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7475 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7477 struct task_group
*tg
= cgroup_tg(cgrp
);
7479 return (u64
) scale_load_down(tg
->shares
);
7482 #ifdef CONFIG_CFS_BANDWIDTH
7483 static DEFINE_MUTEX(cfs_constraints_mutex
);
7485 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7486 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7488 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7490 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7492 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7493 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7495 if (tg
== &root_task_group
)
7499 * Ensure we have at some amount of bandwidth every period. This is
7500 * to prevent reaching a state of large arrears when throttled via
7501 * entity_tick() resulting in prolonged exit starvation.
7503 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7507 * Likewise, bound things on the otherside by preventing insane quota
7508 * periods. This also allows us to normalize in computing quota
7511 if (period
> max_cfs_quota_period
)
7514 mutex_lock(&cfs_constraints_mutex
);
7515 ret
= __cfs_schedulable(tg
, period
, quota
);
7519 runtime_enabled
= quota
!= RUNTIME_INF
;
7520 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7521 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7522 raw_spin_lock_irq(&cfs_b
->lock
);
7523 cfs_b
->period
= ns_to_ktime(period
);
7524 cfs_b
->quota
= quota
;
7526 __refill_cfs_bandwidth_runtime(cfs_b
);
7527 /* restart the period timer (if active) to handle new period expiry */
7528 if (runtime_enabled
&& cfs_b
->timer_active
) {
7529 /* force a reprogram */
7530 cfs_b
->timer_active
= 0;
7531 __start_cfs_bandwidth(cfs_b
);
7533 raw_spin_unlock_irq(&cfs_b
->lock
);
7535 for_each_possible_cpu(i
) {
7536 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7537 struct rq
*rq
= cfs_rq
->rq
;
7539 raw_spin_lock_irq(&rq
->lock
);
7540 cfs_rq
->runtime_enabled
= runtime_enabled
;
7541 cfs_rq
->runtime_remaining
= 0;
7543 if (cfs_rq
->throttled
)
7544 unthrottle_cfs_rq(cfs_rq
);
7545 raw_spin_unlock_irq(&rq
->lock
);
7548 mutex_unlock(&cfs_constraints_mutex
);
7553 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7557 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7558 if (cfs_quota_us
< 0)
7559 quota
= RUNTIME_INF
;
7561 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7563 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7566 long tg_get_cfs_quota(struct task_group
*tg
)
7570 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7573 quota_us
= tg
->cfs_bandwidth
.quota
;
7574 do_div(quota_us
, NSEC_PER_USEC
);
7579 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7583 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7584 quota
= tg
->cfs_bandwidth
.quota
;
7586 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7589 long tg_get_cfs_period(struct task_group
*tg
)
7593 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7594 do_div(cfs_period_us
, NSEC_PER_USEC
);
7596 return cfs_period_us
;
7599 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7601 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7604 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7607 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7610 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7612 return tg_get_cfs_period(cgroup_tg(cgrp
));
7615 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7618 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7621 struct cfs_schedulable_data
{
7622 struct task_group
*tg
;
7627 * normalize group quota/period to be quota/max_period
7628 * note: units are usecs
7630 static u64
normalize_cfs_quota(struct task_group
*tg
,
7631 struct cfs_schedulable_data
*d
)
7639 period
= tg_get_cfs_period(tg
);
7640 quota
= tg_get_cfs_quota(tg
);
7643 /* note: these should typically be equivalent */
7644 if (quota
== RUNTIME_INF
|| quota
== -1)
7647 return to_ratio(period
, quota
);
7650 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7652 struct cfs_schedulable_data
*d
= data
;
7653 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7654 s64 quota
= 0, parent_quota
= -1;
7657 quota
= RUNTIME_INF
;
7659 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7661 quota
= normalize_cfs_quota(tg
, d
);
7662 parent_quota
= parent_b
->hierarchal_quota
;
7665 * ensure max(child_quota) <= parent_quota, inherit when no
7668 if (quota
== RUNTIME_INF
)
7669 quota
= parent_quota
;
7670 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7673 cfs_b
->hierarchal_quota
= quota
;
7678 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7681 struct cfs_schedulable_data data
= {
7687 if (quota
!= RUNTIME_INF
) {
7688 do_div(data
.period
, NSEC_PER_USEC
);
7689 do_div(data
.quota
, NSEC_PER_USEC
);
7693 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7699 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7700 struct cgroup_map_cb
*cb
)
7702 struct task_group
*tg
= cgroup_tg(cgrp
);
7703 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7705 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7706 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7707 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7711 #endif /* CONFIG_CFS_BANDWIDTH */
7712 #endif /* CONFIG_FAIR_GROUP_SCHED */
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7718 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7721 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7723 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7726 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7729 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7732 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7734 return sched_group_rt_period(cgroup_tg(cgrp
));
7736 #endif /* CONFIG_RT_GROUP_SCHED */
7738 static struct cftype cpu_files
[] = {
7739 #ifdef CONFIG_FAIR_GROUP_SCHED
7742 .read_u64
= cpu_shares_read_u64
,
7743 .write_u64
= cpu_shares_write_u64
,
7746 #ifdef CONFIG_CFS_BANDWIDTH
7748 .name
= "cfs_quota_us",
7749 .read_s64
= cpu_cfs_quota_read_s64
,
7750 .write_s64
= cpu_cfs_quota_write_s64
,
7753 .name
= "cfs_period_us",
7754 .read_u64
= cpu_cfs_period_read_u64
,
7755 .write_u64
= cpu_cfs_period_write_u64
,
7759 .read_map
= cpu_stats_show
,
7762 #ifdef CONFIG_RT_GROUP_SCHED
7764 .name
= "rt_runtime_us",
7765 .read_s64
= cpu_rt_runtime_read
,
7766 .write_s64
= cpu_rt_runtime_write
,
7769 .name
= "rt_period_us",
7770 .read_u64
= cpu_rt_period_read_uint
,
7771 .write_u64
= cpu_rt_period_write_uint
,
7777 struct cgroup_subsys cpu_cgroup_subsys
= {
7779 .create
= cpu_cgroup_create
,
7780 .destroy
= cpu_cgroup_destroy
,
7781 .can_attach
= cpu_cgroup_can_attach
,
7782 .attach
= cpu_cgroup_attach
,
7783 .exit
= cpu_cgroup_exit
,
7784 .subsys_id
= cpu_cgroup_subsys_id
,
7785 .base_cftypes
= cpu_files
,
7789 #endif /* CONFIG_CGROUP_SCHED */
7791 #ifdef CONFIG_CGROUP_CPUACCT
7794 * CPU accounting code for task groups.
7796 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7797 * (balbir@in.ibm.com).
7800 struct cpuacct root_cpuacct
;
7802 /* create a new cpu accounting group */
7803 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
7808 return &root_cpuacct
.css
;
7810 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7814 ca
->cpuusage
= alloc_percpu(u64
);
7818 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7820 goto out_free_cpuusage
;
7825 free_percpu(ca
->cpuusage
);
7829 return ERR_PTR(-ENOMEM
);
7832 /* destroy an existing cpu accounting group */
7833 static void cpuacct_destroy(struct cgroup
*cgrp
)
7835 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7837 free_percpu(ca
->cpustat
);
7838 free_percpu(ca
->cpuusage
);
7842 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
7844 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7847 #ifndef CONFIG_64BIT
7849 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7851 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7853 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7861 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
7863 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7865 #ifndef CONFIG_64BIT
7867 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7869 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7871 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7877 /* return total cpu usage (in nanoseconds) of a group */
7878 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7880 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7881 u64 totalcpuusage
= 0;
7884 for_each_present_cpu(i
)
7885 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
7887 return totalcpuusage
;
7890 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
7893 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7902 for_each_present_cpu(i
)
7903 cpuacct_cpuusage_write(ca
, i
, 0);
7909 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
7912 struct cpuacct
*ca
= cgroup_ca(cgroup
);
7916 for_each_present_cpu(i
) {
7917 percpu
= cpuacct_cpuusage_read(ca
, i
);
7918 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
7920 seq_printf(m
, "\n");
7924 static const char *cpuacct_stat_desc
[] = {
7925 [CPUACCT_STAT_USER
] = "user",
7926 [CPUACCT_STAT_SYSTEM
] = "system",
7929 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7930 struct cgroup_map_cb
*cb
)
7932 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7936 for_each_online_cpu(cpu
) {
7937 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
7938 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
7939 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
7941 val
= cputime64_to_clock_t(val
);
7942 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
7945 for_each_online_cpu(cpu
) {
7946 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
7947 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
7948 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
7949 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
7952 val
= cputime64_to_clock_t(val
);
7953 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
7958 static struct cftype files
[] = {
7961 .read_u64
= cpuusage_read
,
7962 .write_u64
= cpuusage_write
,
7965 .name
= "usage_percpu",
7966 .read_seq_string
= cpuacct_percpu_seq_read
,
7970 .read_map
= cpuacct_stats_show
,
7976 * charge this task's execution time to its accounting group.
7978 * called with rq->lock held.
7980 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
7985 if (unlikely(!cpuacct_subsys
.active
))
7988 cpu
= task_cpu(tsk
);
7994 for (; ca
; ca
= parent_ca(ca
)) {
7995 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7996 *cpuusage
+= cputime
;
8002 struct cgroup_subsys cpuacct_subsys
= {
8004 .create
= cpuacct_create
,
8005 .destroy
= cpuacct_destroy
,
8006 .subsys_id
= cpuacct_subsys_id
,
8007 .base_cftypes
= files
,
8009 #endif /* CONFIG_CGROUP_CPUACCT */