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 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1365 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
1370 rq
= __task_rq_lock(p
);
1372 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1373 ttwu_do_wakeup(rq
, p
, wake_flags
);
1376 __task_rq_unlock(rq
);
1381 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1383 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1385 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1387 #endif /* CONFIG_SMP */
1389 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1391 struct rq
*rq
= cpu_rq(cpu
);
1393 #if defined(CONFIG_SMP)
1394 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1395 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1396 ttwu_queue_remote(p
, cpu
);
1401 raw_spin_lock(&rq
->lock
);
1402 ttwu_do_activate(rq
, p
, 0);
1403 raw_spin_unlock(&rq
->lock
);
1407 * try_to_wake_up - wake up a thread
1408 * @p: the thread to be awakened
1409 * @state: the mask of task states that can be woken
1410 * @wake_flags: wake modifier flags (WF_*)
1412 * Put it on the run-queue if it's not already there. The "current"
1413 * thread is always on the run-queue (except when the actual
1414 * re-schedule is in progress), and as such you're allowed to do
1415 * the simpler "current->state = TASK_RUNNING" to mark yourself
1416 * runnable without the overhead of this.
1418 * Returns %true if @p was woken up, %false if it was already running
1419 * or @state didn't match @p's state.
1422 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1424 unsigned long flags
;
1425 int cpu
, success
= 0;
1428 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1429 if (!(p
->state
& state
))
1432 success
= 1; /* we're going to change ->state */
1435 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1440 * If the owning (remote) cpu is still in the middle of schedule() with
1441 * this task as prev, wait until its done referencing the task.
1444 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1446 * In case the architecture enables interrupts in
1447 * context_switch(), we cannot busy wait, since that
1448 * would lead to deadlocks when an interrupt hits and
1449 * tries to wake up @prev. So bail and do a complete
1452 if (ttwu_activate_remote(p
, wake_flags
))
1459 * Pairs with the smp_wmb() in finish_lock_switch().
1463 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1464 p
->state
= TASK_WAKING
;
1466 if (p
->sched_class
->task_waking
)
1467 p
->sched_class
->task_waking(p
);
1469 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1470 if (task_cpu(p
) != cpu
) {
1471 wake_flags
|= WF_MIGRATED
;
1472 set_task_cpu(p
, cpu
);
1474 #endif /* CONFIG_SMP */
1478 ttwu_stat(p
, cpu
, wake_flags
);
1480 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1486 * try_to_wake_up_local - try to wake up a local task with rq lock held
1487 * @p: the thread to be awakened
1489 * Put @p on the run-queue if it's not already there. The caller must
1490 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1493 static void try_to_wake_up_local(struct task_struct
*p
)
1495 struct rq
*rq
= task_rq(p
);
1497 BUG_ON(rq
!= this_rq());
1498 BUG_ON(p
== current
);
1499 lockdep_assert_held(&rq
->lock
);
1501 if (!raw_spin_trylock(&p
->pi_lock
)) {
1502 raw_spin_unlock(&rq
->lock
);
1503 raw_spin_lock(&p
->pi_lock
);
1504 raw_spin_lock(&rq
->lock
);
1507 if (!(p
->state
& TASK_NORMAL
))
1511 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1513 ttwu_do_wakeup(rq
, p
, 0);
1514 ttwu_stat(p
, smp_processor_id(), 0);
1516 raw_spin_unlock(&p
->pi_lock
);
1520 * wake_up_process - Wake up a specific process
1521 * @p: The process to be woken up.
1523 * Attempt to wake up the nominated process and move it to the set of runnable
1524 * processes. Returns 1 if the process was woken up, 0 if it was already
1527 * It may be assumed that this function implies a write memory barrier before
1528 * changing the task state if and only if any tasks are woken up.
1530 int wake_up_process(struct task_struct
*p
)
1532 return try_to_wake_up(p
, TASK_ALL
, 0);
1534 EXPORT_SYMBOL(wake_up_process
);
1536 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1538 return try_to_wake_up(p
, state
, 0);
1542 * Perform scheduler related setup for a newly forked process p.
1543 * p is forked by current.
1545 * __sched_fork() is basic setup used by init_idle() too:
1547 static void __sched_fork(struct task_struct
*p
)
1552 p
->se
.exec_start
= 0;
1553 p
->se
.sum_exec_runtime
= 0;
1554 p
->se
.prev_sum_exec_runtime
= 0;
1555 p
->se
.nr_migrations
= 0;
1557 INIT_LIST_HEAD(&p
->se
.group_node
);
1559 #ifdef CONFIG_SCHEDSTATS
1560 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1563 INIT_LIST_HEAD(&p
->rt
.run_list
);
1565 #ifdef CONFIG_PREEMPT_NOTIFIERS
1566 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1571 * fork()/clone()-time setup:
1573 void sched_fork(struct task_struct
*p
)
1575 unsigned long flags
;
1576 int cpu
= get_cpu();
1580 * We mark the process as running here. This guarantees that
1581 * nobody will actually run it, and a signal or other external
1582 * event cannot wake it up and insert it on the runqueue either.
1584 p
->state
= TASK_RUNNING
;
1587 * Make sure we do not leak PI boosting priority to the child.
1589 p
->prio
= current
->normal_prio
;
1592 * Revert to default priority/policy on fork if requested.
1594 if (unlikely(p
->sched_reset_on_fork
)) {
1595 if (task_has_rt_policy(p
)) {
1596 p
->policy
= SCHED_NORMAL
;
1597 p
->static_prio
= NICE_TO_PRIO(0);
1599 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1600 p
->static_prio
= NICE_TO_PRIO(0);
1602 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1606 * We don't need the reset flag anymore after the fork. It has
1607 * fulfilled its duty:
1609 p
->sched_reset_on_fork
= 0;
1612 if (!rt_prio(p
->prio
))
1613 p
->sched_class
= &fair_sched_class
;
1615 if (p
->sched_class
->task_fork
)
1616 p
->sched_class
->task_fork(p
);
1619 * The child is not yet in the pid-hash so no cgroup attach races,
1620 * and the cgroup is pinned to this child due to cgroup_fork()
1621 * is ran before sched_fork().
1623 * Silence PROVE_RCU.
1625 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1626 set_task_cpu(p
, cpu
);
1627 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1629 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1630 if (likely(sched_info_on()))
1631 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1633 #if defined(CONFIG_SMP)
1636 #ifdef CONFIG_PREEMPT_COUNT
1637 /* Want to start with kernel preemption disabled. */
1638 task_thread_info(p
)->preempt_count
= 1;
1641 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1648 * wake_up_new_task - wake up a newly created task for the first time.
1650 * This function will do some initial scheduler statistics housekeeping
1651 * that must be done for every newly created context, then puts the task
1652 * on the runqueue and wakes it.
1654 void wake_up_new_task(struct task_struct
*p
)
1656 unsigned long flags
;
1659 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1662 * Fork balancing, do it here and not earlier because:
1663 * - cpus_allowed can change in the fork path
1664 * - any previously selected cpu might disappear through hotplug
1666 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1669 rq
= __task_rq_lock(p
);
1670 activate_task(rq
, p
, 0);
1672 trace_sched_wakeup_new(p
, true);
1673 check_preempt_curr(rq
, p
, WF_FORK
);
1675 if (p
->sched_class
->task_woken
)
1676 p
->sched_class
->task_woken(rq
, p
);
1678 task_rq_unlock(rq
, p
, &flags
);
1681 #ifdef CONFIG_PREEMPT_NOTIFIERS
1684 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1685 * @notifier: notifier struct to register
1687 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1689 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1691 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1694 * preempt_notifier_unregister - no longer interested in preemption notifications
1695 * @notifier: notifier struct to unregister
1697 * This is safe to call from within a preemption notifier.
1699 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1701 hlist_del(¬ifier
->link
);
1703 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1705 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1707 struct preempt_notifier
*notifier
;
1708 struct hlist_node
*node
;
1710 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1711 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1715 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1716 struct task_struct
*next
)
1718 struct preempt_notifier
*notifier
;
1719 struct hlist_node
*node
;
1721 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
1722 notifier
->ops
->sched_out(notifier
, next
);
1725 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1727 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1732 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1733 struct task_struct
*next
)
1737 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1740 * prepare_task_switch - prepare to switch tasks
1741 * @rq: the runqueue preparing to switch
1742 * @prev: the current task that is being switched out
1743 * @next: the task we are going to switch to.
1745 * This is called with the rq lock held and interrupts off. It must
1746 * be paired with a subsequent finish_task_switch after the context
1749 * prepare_task_switch sets up locking and calls architecture specific
1753 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1754 struct task_struct
*next
)
1756 trace_sched_switch(prev
, next
);
1757 sched_info_switch(prev
, next
);
1758 perf_event_task_sched_out(prev
, next
);
1759 fire_sched_out_preempt_notifiers(prev
, next
);
1760 prepare_lock_switch(rq
, next
);
1761 prepare_arch_switch(next
);
1765 * finish_task_switch - clean up after a task-switch
1766 * @rq: runqueue associated with task-switch
1767 * @prev: the thread we just switched away from.
1769 * finish_task_switch must be called after the context switch, paired
1770 * with a prepare_task_switch call before the context switch.
1771 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1772 * and do any other architecture-specific cleanup actions.
1774 * Note that we may have delayed dropping an mm in context_switch(). If
1775 * so, we finish that here outside of the runqueue lock. (Doing it
1776 * with the lock held can cause deadlocks; see schedule() for
1779 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1780 __releases(rq
->lock
)
1782 struct mm_struct
*mm
= rq
->prev_mm
;
1788 * A task struct has one reference for the use as "current".
1789 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1790 * schedule one last time. The schedule call will never return, and
1791 * the scheduled task must drop that reference.
1792 * The test for TASK_DEAD must occur while the runqueue locks are
1793 * still held, otherwise prev could be scheduled on another cpu, die
1794 * there before we look at prev->state, and then the reference would
1796 * Manfred Spraul <manfred@colorfullife.com>
1798 prev_state
= prev
->state
;
1799 account_switch_vtime(prev
);
1800 finish_arch_switch(prev
);
1801 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1802 local_irq_disable();
1803 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1804 perf_event_task_sched_in(prev
, current
);
1805 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1807 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1808 finish_lock_switch(rq
, prev
);
1809 finish_arch_post_lock_switch();
1811 fire_sched_in_preempt_notifiers(current
);
1814 if (unlikely(prev_state
== TASK_DEAD
)) {
1816 * Remove function-return probe instances associated with this
1817 * task and put them back on the free list.
1819 kprobe_flush_task(prev
);
1820 put_task_struct(prev
);
1826 /* assumes rq->lock is held */
1827 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1829 if (prev
->sched_class
->pre_schedule
)
1830 prev
->sched_class
->pre_schedule(rq
, prev
);
1833 /* rq->lock is NOT held, but preemption is disabled */
1834 static inline void post_schedule(struct rq
*rq
)
1836 if (rq
->post_schedule
) {
1837 unsigned long flags
;
1839 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1840 if (rq
->curr
->sched_class
->post_schedule
)
1841 rq
->curr
->sched_class
->post_schedule(rq
);
1842 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1844 rq
->post_schedule
= 0;
1850 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1854 static inline void post_schedule(struct rq
*rq
)
1861 * schedule_tail - first thing a freshly forked thread must call.
1862 * @prev: the thread we just switched away from.
1864 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1865 __releases(rq
->lock
)
1867 struct rq
*rq
= this_rq();
1869 finish_task_switch(rq
, prev
);
1872 * FIXME: do we need to worry about rq being invalidated by the
1877 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1878 /* In this case, finish_task_switch does not reenable preemption */
1881 if (current
->set_child_tid
)
1882 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1886 * context_switch - switch to the new MM and the new
1887 * thread's register state.
1890 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1891 struct task_struct
*next
)
1893 struct mm_struct
*mm
, *oldmm
;
1895 prepare_task_switch(rq
, prev
, next
);
1898 oldmm
= prev
->active_mm
;
1900 * For paravirt, this is coupled with an exit in switch_to to
1901 * combine the page table reload and the switch backend into
1904 arch_start_context_switch(prev
);
1907 next
->active_mm
= oldmm
;
1908 atomic_inc(&oldmm
->mm_count
);
1909 enter_lazy_tlb(oldmm
, next
);
1911 switch_mm(oldmm
, mm
, next
);
1914 prev
->active_mm
= NULL
;
1915 rq
->prev_mm
= oldmm
;
1918 * Since the runqueue lock will be released by the next
1919 * task (which is an invalid locking op but in the case
1920 * of the scheduler it's an obvious special-case), so we
1921 * do an early lockdep release here:
1923 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1924 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1927 /* Here we just switch the register state and the stack. */
1928 switch_to(prev
, next
, prev
);
1932 * this_rq must be evaluated again because prev may have moved
1933 * CPUs since it called schedule(), thus the 'rq' on its stack
1934 * frame will be invalid.
1936 finish_task_switch(this_rq(), prev
);
1940 * nr_running, nr_uninterruptible and nr_context_switches:
1942 * externally visible scheduler statistics: current number of runnable
1943 * threads, current number of uninterruptible-sleeping threads, total
1944 * number of context switches performed since bootup.
1946 unsigned long nr_running(void)
1948 unsigned long i
, sum
= 0;
1950 for_each_online_cpu(i
)
1951 sum
+= cpu_rq(i
)->nr_running
;
1956 unsigned long nr_uninterruptible(void)
1958 unsigned long i
, sum
= 0;
1960 for_each_possible_cpu(i
)
1961 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1964 * Since we read the counters lockless, it might be slightly
1965 * inaccurate. Do not allow it to go below zero though:
1967 if (unlikely((long)sum
< 0))
1973 unsigned long long nr_context_switches(void)
1976 unsigned long long sum
= 0;
1978 for_each_possible_cpu(i
)
1979 sum
+= cpu_rq(i
)->nr_switches
;
1984 unsigned long nr_iowait(void)
1986 unsigned long i
, sum
= 0;
1988 for_each_possible_cpu(i
)
1989 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1994 unsigned long nr_iowait_cpu(int cpu
)
1996 struct rq
*this = cpu_rq(cpu
);
1997 return atomic_read(&this->nr_iowait
);
2000 unsigned long this_cpu_load(void)
2002 struct rq
*this = this_rq();
2003 return this->cpu_load
[0];
2008 * Global load-average calculations
2010 * We take a distributed and async approach to calculating the global load-avg
2011 * in order to minimize overhead.
2013 * The global load average is an exponentially decaying average of nr_running +
2014 * nr_uninterruptible.
2016 * Once every LOAD_FREQ:
2019 * for_each_possible_cpu(cpu)
2020 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2022 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2024 * Due to a number of reasons the above turns in the mess below:
2026 * - for_each_possible_cpu() is prohibitively expensive on machines with
2027 * serious number of cpus, therefore we need to take a distributed approach
2028 * to calculating nr_active.
2030 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2031 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2033 * So assuming nr_active := 0 when we start out -- true per definition, we
2034 * can simply take per-cpu deltas and fold those into a global accumulate
2035 * to obtain the same result. See calc_load_fold_active().
2037 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2038 * across the machine, we assume 10 ticks is sufficient time for every
2039 * cpu to have completed this task.
2041 * This places an upper-bound on the IRQ-off latency of the machine. Then
2042 * again, being late doesn't loose the delta, just wrecks the sample.
2044 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2045 * this would add another cross-cpu cacheline miss and atomic operation
2046 * to the wakeup path. Instead we increment on whatever cpu the task ran
2047 * when it went into uninterruptible state and decrement on whatever cpu
2048 * did the wakeup. This means that only the sum of nr_uninterruptible over
2049 * all cpus yields the correct result.
2051 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2054 /* Variables and functions for calc_load */
2055 static atomic_long_t calc_load_tasks
;
2056 static unsigned long calc_load_update
;
2057 unsigned long avenrun
[3];
2058 EXPORT_SYMBOL(avenrun
); /* should be removed */
2061 * get_avenrun - get the load average array
2062 * @loads: pointer to dest load array
2063 * @offset: offset to add
2064 * @shift: shift count to shift the result left
2066 * These values are estimates at best, so no need for locking.
2068 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2070 loads
[0] = (avenrun
[0] + offset
) << shift
;
2071 loads
[1] = (avenrun
[1] + offset
) << shift
;
2072 loads
[2] = (avenrun
[2] + offset
) << shift
;
2075 static long calc_load_fold_active(struct rq
*this_rq
)
2077 long nr_active
, delta
= 0;
2079 nr_active
= this_rq
->nr_running
;
2080 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2082 if (nr_active
!= this_rq
->calc_load_active
) {
2083 delta
= nr_active
- this_rq
->calc_load_active
;
2084 this_rq
->calc_load_active
= nr_active
;
2091 * a1 = a0 * e + a * (1 - e)
2093 static unsigned long
2094 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2097 load
+= active
* (FIXED_1
- exp
);
2098 load
+= 1UL << (FSHIFT
- 1);
2099 return load
>> FSHIFT
;
2104 * Handle NO_HZ for the global load-average.
2106 * Since the above described distributed algorithm to compute the global
2107 * load-average relies on per-cpu sampling from the tick, it is affected by
2110 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2111 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2112 * when we read the global state.
2114 * Obviously reality has to ruin such a delightfully simple scheme:
2116 * - When we go NO_HZ idle during the window, we can negate our sample
2117 * contribution, causing under-accounting.
2119 * We avoid this by keeping two idle-delta counters and flipping them
2120 * when the window starts, thus separating old and new NO_HZ load.
2122 * The only trick is the slight shift in index flip for read vs write.
2126 * |-|-----------|-|-----------|-|-----------|-|
2127 * r:0 0 1 1 0 0 1 1 0
2128 * w:0 1 1 0 0 1 1 0 0
2130 * This ensures we'll fold the old idle contribution in this window while
2131 * accumlating the new one.
2133 * - When we wake up from NO_HZ idle during the window, we push up our
2134 * contribution, since we effectively move our sample point to a known
2137 * This is solved by pushing the window forward, and thus skipping the
2138 * sample, for this cpu (effectively using the idle-delta for this cpu which
2139 * was in effect at the time the window opened). This also solves the issue
2140 * of having to deal with a cpu having been in NOHZ idle for multiple
2141 * LOAD_FREQ intervals.
2143 * When making the ILB scale, we should try to pull this in as well.
2145 static atomic_long_t calc_load_idle
[2];
2146 static int calc_load_idx
;
2148 static inline int calc_load_write_idx(void)
2150 int idx
= calc_load_idx
;
2153 * See calc_global_nohz(), if we observe the new index, we also
2154 * need to observe the new update time.
2159 * If the folding window started, make sure we start writing in the
2162 if (!time_before(jiffies
, calc_load_update
))
2168 static inline int calc_load_read_idx(void)
2170 return calc_load_idx
& 1;
2173 void calc_load_enter_idle(void)
2175 struct rq
*this_rq
= this_rq();
2179 * We're going into NOHZ mode, if there's any pending delta, fold it
2180 * into the pending idle delta.
2182 delta
= calc_load_fold_active(this_rq
);
2184 int idx
= calc_load_write_idx();
2185 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2189 void calc_load_exit_idle(void)
2191 struct rq
*this_rq
= this_rq();
2194 * If we're still before the sample window, we're done.
2196 if (time_before(jiffies
, this_rq
->calc_load_update
))
2200 * We woke inside or after the sample window, this means we're already
2201 * accounted through the nohz accounting, so skip the entire deal and
2202 * sync up for the next window.
2204 this_rq
->calc_load_update
= calc_load_update
;
2205 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2206 this_rq
->calc_load_update
+= LOAD_FREQ
;
2209 static long calc_load_fold_idle(void)
2211 int idx
= calc_load_read_idx();
2214 if (atomic_long_read(&calc_load_idle
[idx
]))
2215 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2221 * fixed_power_int - compute: x^n, in O(log n) time
2223 * @x: base of the power
2224 * @frac_bits: fractional bits of @x
2225 * @n: power to raise @x to.
2227 * By exploiting the relation between the definition of the natural power
2228 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2229 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2230 * (where: n_i \elem {0, 1}, the binary vector representing n),
2231 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2232 * of course trivially computable in O(log_2 n), the length of our binary
2235 static unsigned long
2236 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2238 unsigned long result
= 1UL << frac_bits
;
2243 result
+= 1UL << (frac_bits
- 1);
2244 result
>>= frac_bits
;
2250 x
+= 1UL << (frac_bits
- 1);
2258 * a1 = a0 * e + a * (1 - e)
2260 * a2 = a1 * e + a * (1 - e)
2261 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2262 * = a0 * e^2 + a * (1 - e) * (1 + e)
2264 * a3 = a2 * e + a * (1 - e)
2265 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2266 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2270 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2271 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2272 * = a0 * e^n + a * (1 - e^n)
2274 * [1] application of the geometric series:
2277 * S_n := \Sum x^i = -------------
2280 static unsigned long
2281 calc_load_n(unsigned long load
, unsigned long exp
,
2282 unsigned long active
, unsigned int n
)
2285 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2289 * NO_HZ can leave us missing all per-cpu ticks calling
2290 * calc_load_account_active(), but since an idle CPU folds its delta into
2291 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2292 * in the pending idle delta if our idle period crossed a load cycle boundary.
2294 * Once we've updated the global active value, we need to apply the exponential
2295 * weights adjusted to the number of cycles missed.
2297 static void calc_global_nohz(void)
2299 long delta
, active
, n
;
2301 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2303 * Catch-up, fold however many we are behind still
2305 delta
= jiffies
- calc_load_update
- 10;
2306 n
= 1 + (delta
/ LOAD_FREQ
);
2308 active
= atomic_long_read(&calc_load_tasks
);
2309 active
= active
> 0 ? active
* FIXED_1
: 0;
2311 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2312 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2313 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2315 calc_load_update
+= n
* LOAD_FREQ
;
2319 * Flip the idle index...
2321 * Make sure we first write the new time then flip the index, so that
2322 * calc_load_write_idx() will see the new time when it reads the new
2323 * index, this avoids a double flip messing things up.
2328 #else /* !CONFIG_NO_HZ */
2330 static inline long calc_load_fold_idle(void) { return 0; }
2331 static inline void calc_global_nohz(void) { }
2333 #endif /* CONFIG_NO_HZ */
2336 * calc_load - update the avenrun load estimates 10 ticks after the
2337 * CPUs have updated calc_load_tasks.
2339 void calc_global_load(unsigned long ticks
)
2343 if (time_before(jiffies
, calc_load_update
+ 10))
2347 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2349 delta
= calc_load_fold_idle();
2351 atomic_long_add(delta
, &calc_load_tasks
);
2353 active
= atomic_long_read(&calc_load_tasks
);
2354 active
= active
> 0 ? active
* FIXED_1
: 0;
2356 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2357 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2358 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2360 calc_load_update
+= LOAD_FREQ
;
2363 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2369 * Called from update_cpu_load() to periodically update this CPU's
2372 static void calc_load_account_active(struct rq
*this_rq
)
2376 if (time_before(jiffies
, this_rq
->calc_load_update
))
2379 delta
= calc_load_fold_active(this_rq
);
2381 atomic_long_add(delta
, &calc_load_tasks
);
2383 this_rq
->calc_load_update
+= LOAD_FREQ
;
2387 * End of global load-average stuff
2391 * The exact cpuload at various idx values, calculated at every tick would be
2392 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2394 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2395 * on nth tick when cpu may be busy, then we have:
2396 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2397 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2399 * decay_load_missed() below does efficient calculation of
2400 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2401 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2403 * The calculation is approximated on a 128 point scale.
2404 * degrade_zero_ticks is the number of ticks after which load at any
2405 * particular idx is approximated to be zero.
2406 * degrade_factor is a precomputed table, a row for each load idx.
2407 * Each column corresponds to degradation factor for a power of two ticks,
2408 * based on 128 point scale.
2410 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2411 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2413 * With this power of 2 load factors, we can degrade the load n times
2414 * by looking at 1 bits in n and doing as many mult/shift instead of
2415 * n mult/shifts needed by the exact degradation.
2417 #define DEGRADE_SHIFT 7
2418 static const unsigned char
2419 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2420 static const unsigned char
2421 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2422 {0, 0, 0, 0, 0, 0, 0, 0},
2423 {64, 32, 8, 0, 0, 0, 0, 0},
2424 {96, 72, 40, 12, 1, 0, 0},
2425 {112, 98, 75, 43, 15, 1, 0},
2426 {120, 112, 98, 76, 45, 16, 2} };
2429 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2430 * would be when CPU is idle and so we just decay the old load without
2431 * adding any new load.
2433 static unsigned long
2434 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2438 if (!missed_updates
)
2441 if (missed_updates
>= degrade_zero_ticks
[idx
])
2445 return load
>> missed_updates
;
2447 while (missed_updates
) {
2448 if (missed_updates
% 2)
2449 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2451 missed_updates
>>= 1;
2458 * Update rq->cpu_load[] statistics. This function is usually called every
2459 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2460 * every tick. We fix it up based on jiffies.
2462 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2463 unsigned long pending_updates
)
2467 this_rq
->nr_load_updates
++;
2469 /* Update our load: */
2470 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2471 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2472 unsigned long old_load
, new_load
;
2474 /* scale is effectively 1 << i now, and >> i divides by scale */
2476 old_load
= this_rq
->cpu_load
[i
];
2477 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2478 new_load
= this_load
;
2480 * Round up the averaging division if load is increasing. This
2481 * prevents us from getting stuck on 9 if the load is 10, for
2484 if (new_load
> old_load
)
2485 new_load
+= scale
- 1;
2487 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2490 sched_avg_update(this_rq
);
2495 * There is no sane way to deal with nohz on smp when using jiffies because the
2496 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2497 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2499 * Therefore we cannot use the delta approach from the regular tick since that
2500 * would seriously skew the load calculation. However we'll make do for those
2501 * updates happening while idle (nohz_idle_balance) or coming out of idle
2502 * (tick_nohz_idle_exit).
2504 * This means we might still be one tick off for nohz periods.
2508 * Called from nohz_idle_balance() to update the load ratings before doing the
2511 void update_idle_cpu_load(struct rq
*this_rq
)
2513 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2514 unsigned long load
= this_rq
->load
.weight
;
2515 unsigned long pending_updates
;
2518 * bail if there's load or we're actually up-to-date.
2520 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2523 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2524 this_rq
->last_load_update_tick
= curr_jiffies
;
2526 __update_cpu_load(this_rq
, load
, pending_updates
);
2530 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2532 void update_cpu_load_nohz(void)
2534 struct rq
*this_rq
= this_rq();
2535 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2536 unsigned long pending_updates
;
2538 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2541 raw_spin_lock(&this_rq
->lock
);
2542 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2543 if (pending_updates
) {
2544 this_rq
->last_load_update_tick
= curr_jiffies
;
2546 * We were idle, this means load 0, the current load might be
2547 * !0 due to remote wakeups and the sort.
2549 __update_cpu_load(this_rq
, 0, pending_updates
);
2551 raw_spin_unlock(&this_rq
->lock
);
2553 #endif /* CONFIG_NO_HZ */
2556 * Called from scheduler_tick()
2558 static void update_cpu_load_active(struct rq
*this_rq
)
2561 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2563 this_rq
->last_load_update_tick
= jiffies
;
2564 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2566 calc_load_account_active(this_rq
);
2572 * sched_exec - execve() is a valuable balancing opportunity, because at
2573 * this point the task has the smallest effective memory and cache footprint.
2575 void sched_exec(void)
2577 struct task_struct
*p
= current
;
2578 unsigned long flags
;
2581 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2582 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2583 if (dest_cpu
== smp_processor_id())
2586 if (likely(cpu_active(dest_cpu
))) {
2587 struct migration_arg arg
= { p
, dest_cpu
};
2589 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2590 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2594 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2599 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2600 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2602 EXPORT_PER_CPU_SYMBOL(kstat
);
2603 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2606 * Return any ns on the sched_clock that have not yet been accounted in
2607 * @p in case that task is currently running.
2609 * Called with task_rq_lock() held on @rq.
2611 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2615 if (task_current(rq
, p
)) {
2616 update_rq_clock(rq
);
2617 ns
= rq
->clock_task
- p
->se
.exec_start
;
2625 unsigned long long task_delta_exec(struct task_struct
*p
)
2627 unsigned long flags
;
2631 rq
= task_rq_lock(p
, &flags
);
2632 ns
= do_task_delta_exec(p
, rq
);
2633 task_rq_unlock(rq
, p
, &flags
);
2639 * Return accounted runtime for the task.
2640 * In case the task is currently running, return the runtime plus current's
2641 * pending runtime that have not been accounted yet.
2643 unsigned long long task_sched_runtime(struct task_struct
*p
)
2645 unsigned long flags
;
2649 rq
= task_rq_lock(p
, &flags
);
2650 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2651 task_rq_unlock(rq
, p
, &flags
);
2657 * This function gets called by the timer code, with HZ frequency.
2658 * We call it with interrupts disabled.
2660 void scheduler_tick(void)
2662 int cpu
= smp_processor_id();
2663 struct rq
*rq
= cpu_rq(cpu
);
2664 struct task_struct
*curr
= rq
->curr
;
2668 raw_spin_lock(&rq
->lock
);
2669 update_rq_clock(rq
);
2670 update_cpu_load_active(rq
);
2671 curr
->sched_class
->task_tick(rq
, curr
, 0);
2672 raw_spin_unlock(&rq
->lock
);
2674 perf_event_task_tick();
2677 rq
->idle_balance
= idle_cpu(cpu
);
2678 trigger_load_balance(rq
, cpu
);
2682 notrace
unsigned long get_parent_ip(unsigned long addr
)
2684 if (in_lock_functions(addr
)) {
2685 addr
= CALLER_ADDR2
;
2686 if (in_lock_functions(addr
))
2687 addr
= CALLER_ADDR3
;
2692 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2693 defined(CONFIG_PREEMPT_TRACER))
2695 void __kprobes
add_preempt_count(int val
)
2697 #ifdef CONFIG_DEBUG_PREEMPT
2701 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2704 preempt_count() += val
;
2705 #ifdef CONFIG_DEBUG_PREEMPT
2707 * Spinlock count overflowing soon?
2709 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2712 if (preempt_count() == val
)
2713 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2715 EXPORT_SYMBOL(add_preempt_count
);
2717 void __kprobes
sub_preempt_count(int val
)
2719 #ifdef CONFIG_DEBUG_PREEMPT
2723 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2726 * Is the spinlock portion underflowing?
2728 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2729 !(preempt_count() & PREEMPT_MASK
)))
2733 if (preempt_count() == val
)
2734 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2735 preempt_count() -= val
;
2737 EXPORT_SYMBOL(sub_preempt_count
);
2742 * Print scheduling while atomic bug:
2744 static noinline
void __schedule_bug(struct task_struct
*prev
)
2746 if (oops_in_progress
)
2749 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2750 prev
->comm
, prev
->pid
, preempt_count());
2752 debug_show_held_locks(prev
);
2754 if (irqs_disabled())
2755 print_irqtrace_events(prev
);
2757 add_taint(TAINT_WARN
);
2761 * Various schedule()-time debugging checks and statistics:
2763 static inline void schedule_debug(struct task_struct
*prev
)
2766 * Test if we are atomic. Since do_exit() needs to call into
2767 * schedule() atomically, we ignore that path for now.
2768 * Otherwise, whine if we are scheduling when we should not be.
2770 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2771 __schedule_bug(prev
);
2774 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2776 schedstat_inc(this_rq(), sched_count
);
2779 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2781 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2782 update_rq_clock(rq
);
2783 prev
->sched_class
->put_prev_task(rq
, prev
);
2787 * Pick up the highest-prio task:
2789 static inline struct task_struct
*
2790 pick_next_task(struct rq
*rq
)
2792 const struct sched_class
*class;
2793 struct task_struct
*p
;
2796 * Optimization: we know that if all tasks are in
2797 * the fair class we can call that function directly:
2799 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2800 p
= fair_sched_class
.pick_next_task(rq
);
2805 for_each_class(class) {
2806 p
= class->pick_next_task(rq
);
2811 BUG(); /* the idle class will always have a runnable task */
2815 * __schedule() is the main scheduler function.
2817 * The main means of driving the scheduler and thus entering this function are:
2819 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2821 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2822 * paths. For example, see arch/x86/entry_64.S.
2824 * To drive preemption between tasks, the scheduler sets the flag in timer
2825 * interrupt handler scheduler_tick().
2827 * 3. Wakeups don't really cause entry into schedule(). They add a
2828 * task to the run-queue and that's it.
2830 * Now, if the new task added to the run-queue preempts the current
2831 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2832 * called on the nearest possible occasion:
2834 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2836 * - in syscall or exception context, at the next outmost
2837 * preempt_enable(). (this might be as soon as the wake_up()'s
2840 * - in IRQ context, return from interrupt-handler to
2841 * preemptible context
2843 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2846 * - cond_resched() call
2847 * - explicit schedule() call
2848 * - return from syscall or exception to user-space
2849 * - return from interrupt-handler to user-space
2851 static void __sched
__schedule(void)
2853 struct task_struct
*prev
, *next
;
2854 unsigned long *switch_count
;
2860 cpu
= smp_processor_id();
2862 rcu_note_context_switch(cpu
);
2865 schedule_debug(prev
);
2867 if (sched_feat(HRTICK
))
2870 raw_spin_lock_irq(&rq
->lock
);
2872 switch_count
= &prev
->nivcsw
;
2873 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2874 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2875 prev
->state
= TASK_RUNNING
;
2877 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2881 * If a worker went to sleep, notify and ask workqueue
2882 * whether it wants to wake up a task to maintain
2885 if (prev
->flags
& PF_WQ_WORKER
) {
2886 struct task_struct
*to_wakeup
;
2888 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2890 try_to_wake_up_local(to_wakeup
);
2893 switch_count
= &prev
->nvcsw
;
2896 pre_schedule(rq
, prev
);
2898 if (unlikely(!rq
->nr_running
))
2899 idle_balance(cpu
, rq
);
2901 put_prev_task(rq
, prev
);
2902 next
= pick_next_task(rq
);
2903 clear_tsk_need_resched(prev
);
2904 rq
->skip_clock_update
= 0;
2906 if (likely(prev
!= next
)) {
2911 context_switch(rq
, prev
, next
); /* unlocks the rq */
2913 * The context switch have flipped the stack from under us
2914 * and restored the local variables which were saved when
2915 * this task called schedule() in the past. prev == current
2916 * is still correct, but it can be moved to another cpu/rq.
2918 cpu
= smp_processor_id();
2921 raw_spin_unlock_irq(&rq
->lock
);
2925 sched_preempt_enable_no_resched();
2930 static inline void sched_submit_work(struct task_struct
*tsk
)
2932 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2935 * If we are going to sleep and we have plugged IO queued,
2936 * make sure to submit it to avoid deadlocks.
2938 if (blk_needs_flush_plug(tsk
))
2939 blk_schedule_flush_plug(tsk
);
2942 asmlinkage
void __sched
schedule(void)
2944 struct task_struct
*tsk
= current
;
2946 sched_submit_work(tsk
);
2949 EXPORT_SYMBOL(schedule
);
2952 * schedule_preempt_disabled - called with preemption disabled
2954 * Returns with preemption disabled. Note: preempt_count must be 1
2956 void __sched
schedule_preempt_disabled(void)
2958 sched_preempt_enable_no_resched();
2963 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2965 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
2967 if (lock
->owner
!= owner
)
2971 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2972 * lock->owner still matches owner, if that fails, owner might
2973 * point to free()d memory, if it still matches, the rcu_read_lock()
2974 * ensures the memory stays valid.
2978 return owner
->on_cpu
;
2982 * Look out! "owner" is an entirely speculative pointer
2983 * access and not reliable.
2985 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
2987 if (!sched_feat(OWNER_SPIN
))
2991 while (owner_running(lock
, owner
)) {
2995 arch_mutex_cpu_relax();
3000 * We break out the loop above on need_resched() and when the
3001 * owner changed, which is a sign for heavy contention. Return
3002 * success only when lock->owner is NULL.
3004 return lock
->owner
== NULL
;
3008 #ifdef CONFIG_PREEMPT
3010 * this is the entry point to schedule() from in-kernel preemption
3011 * off of preempt_enable. Kernel preemptions off return from interrupt
3012 * occur there and call schedule directly.
3014 asmlinkage
void __sched notrace
preempt_schedule(void)
3016 struct thread_info
*ti
= current_thread_info();
3019 * If there is a non-zero preempt_count or interrupts are disabled,
3020 * we do not want to preempt the current task. Just return..
3022 if (likely(ti
->preempt_count
|| irqs_disabled()))
3026 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3028 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3031 * Check again in case we missed a preemption opportunity
3032 * between schedule and now.
3035 } while (need_resched());
3037 EXPORT_SYMBOL(preempt_schedule
);
3040 * this is the entry point to schedule() from kernel preemption
3041 * off of irq context.
3042 * Note, that this is called and return with irqs disabled. This will
3043 * protect us against recursive calling from irq.
3045 asmlinkage
void __sched
preempt_schedule_irq(void)
3047 struct thread_info
*ti
= current_thread_info();
3049 /* Catch callers which need to be fixed */
3050 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3053 add_preempt_count(PREEMPT_ACTIVE
);
3056 local_irq_disable();
3057 sub_preempt_count(PREEMPT_ACTIVE
);
3060 * Check again in case we missed a preemption opportunity
3061 * between schedule and now.
3064 } while (need_resched());
3067 #endif /* CONFIG_PREEMPT */
3069 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3072 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3074 EXPORT_SYMBOL(default_wake_function
);
3077 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3078 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3079 * number) then we wake all the non-exclusive tasks and one exclusive task.
3081 * There are circumstances in which we can try to wake a task which has already
3082 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3083 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3085 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3086 int nr_exclusive
, int wake_flags
, void *key
)
3088 wait_queue_t
*curr
, *next
;
3090 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3091 unsigned flags
= curr
->flags
;
3093 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3094 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3100 * __wake_up - wake up threads blocked on a waitqueue.
3102 * @mode: which threads
3103 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3104 * @key: is directly passed to the wakeup function
3106 * It may be assumed that this function implies a write memory barrier before
3107 * changing the task state if and only if any tasks are woken up.
3109 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3110 int nr_exclusive
, void *key
)
3112 unsigned long flags
;
3114 spin_lock_irqsave(&q
->lock
, flags
);
3115 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3116 spin_unlock_irqrestore(&q
->lock
, flags
);
3118 EXPORT_SYMBOL(__wake_up
);
3121 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3123 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3125 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3127 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3129 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3131 __wake_up_common(q
, mode
, 1, 0, key
);
3133 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3136 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3138 * @mode: which threads
3139 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3140 * @key: opaque value to be passed to wakeup targets
3142 * The sync wakeup differs that the waker knows that it will schedule
3143 * away soon, so while the target thread will be woken up, it will not
3144 * be migrated to another CPU - ie. the two threads are 'synchronized'
3145 * with each other. This can prevent needless bouncing between CPUs.
3147 * On UP it can prevent extra preemption.
3149 * It may be assumed that this function implies a write memory barrier before
3150 * changing the task state if and only if any tasks are woken up.
3152 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3153 int nr_exclusive
, void *key
)
3155 unsigned long flags
;
3156 int wake_flags
= WF_SYNC
;
3161 if (unlikely(!nr_exclusive
))
3164 spin_lock_irqsave(&q
->lock
, flags
);
3165 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3166 spin_unlock_irqrestore(&q
->lock
, flags
);
3168 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3171 * __wake_up_sync - see __wake_up_sync_key()
3173 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3175 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3177 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3180 * complete: - signals a single thread waiting on this completion
3181 * @x: holds the state of this particular completion
3183 * This will wake up a single thread waiting on this completion. Threads will be
3184 * awakened in the same order in which they were queued.
3186 * See also complete_all(), wait_for_completion() and related routines.
3188 * It may be assumed that this function implies a write memory barrier before
3189 * changing the task state if and only if any tasks are woken up.
3191 void complete(struct completion
*x
)
3193 unsigned long flags
;
3195 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3197 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3198 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3200 EXPORT_SYMBOL(complete
);
3203 * complete_all: - signals all threads waiting on this completion
3204 * @x: holds the state of this particular completion
3206 * This will wake up all threads waiting on this particular completion event.
3208 * It may be assumed that this function implies a write memory barrier before
3209 * changing the task state if and only if any tasks are woken up.
3211 void complete_all(struct completion
*x
)
3213 unsigned long flags
;
3215 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3216 x
->done
+= UINT_MAX
/2;
3217 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3218 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3220 EXPORT_SYMBOL(complete_all
);
3222 static inline long __sched
3223 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3226 DECLARE_WAITQUEUE(wait
, current
);
3228 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3230 if (signal_pending_state(state
, current
)) {
3231 timeout
= -ERESTARTSYS
;
3234 __set_current_state(state
);
3235 spin_unlock_irq(&x
->wait
.lock
);
3236 timeout
= schedule_timeout(timeout
);
3237 spin_lock_irq(&x
->wait
.lock
);
3238 } while (!x
->done
&& timeout
);
3239 __remove_wait_queue(&x
->wait
, &wait
);
3244 return timeout
?: 1;
3248 wait_for_common(struct completion
*x
, long timeout
, int state
)
3252 spin_lock_irq(&x
->wait
.lock
);
3253 timeout
= do_wait_for_common(x
, timeout
, state
);
3254 spin_unlock_irq(&x
->wait
.lock
);
3259 * wait_for_completion: - waits for completion of a task
3260 * @x: holds the state of this particular completion
3262 * This waits to be signaled for completion of a specific task. It is NOT
3263 * interruptible and there is no timeout.
3265 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3266 * and interrupt capability. Also see complete().
3268 void __sched
wait_for_completion(struct completion
*x
)
3270 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3272 EXPORT_SYMBOL(wait_for_completion
);
3275 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3276 * @x: holds the state of this particular completion
3277 * @timeout: timeout value in jiffies
3279 * This waits for either a completion of a specific task to be signaled or for a
3280 * specified timeout to expire. The timeout is in jiffies. It is not
3283 * The return value is 0 if timed out, and positive (at least 1, or number of
3284 * jiffies left till timeout) if completed.
3286 unsigned long __sched
3287 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3289 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3291 EXPORT_SYMBOL(wait_for_completion_timeout
);
3294 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3295 * @x: holds the state of this particular completion
3297 * This waits for completion of a specific task to be signaled. It is
3300 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3302 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3304 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3305 if (t
== -ERESTARTSYS
)
3309 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3312 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3313 * @x: holds the state of this particular completion
3314 * @timeout: timeout value in jiffies
3316 * This waits for either a completion of a specific task to be signaled or for a
3317 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3319 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3320 * positive (at least 1, or number of jiffies left till timeout) if completed.
3323 wait_for_completion_interruptible_timeout(struct completion
*x
,
3324 unsigned long timeout
)
3326 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3328 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3331 * wait_for_completion_killable: - waits for completion of a task (killable)
3332 * @x: holds the state of this particular completion
3334 * This waits to be signaled for completion of a specific task. It can be
3335 * interrupted by a kill signal.
3337 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3339 int __sched
wait_for_completion_killable(struct completion
*x
)
3341 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3342 if (t
== -ERESTARTSYS
)
3346 EXPORT_SYMBOL(wait_for_completion_killable
);
3349 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3350 * @x: holds the state of this particular completion
3351 * @timeout: timeout value in jiffies
3353 * This waits for either a completion of a specific task to be
3354 * signaled or for a specified timeout to expire. It can be
3355 * interrupted by a kill signal. The timeout is in jiffies.
3357 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3358 * positive (at least 1, or number of jiffies left till timeout) if completed.
3361 wait_for_completion_killable_timeout(struct completion
*x
,
3362 unsigned long timeout
)
3364 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3366 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3369 * try_wait_for_completion - try to decrement a completion without blocking
3370 * @x: completion structure
3372 * Returns: 0 if a decrement cannot be done without blocking
3373 * 1 if a decrement succeeded.
3375 * If a completion is being used as a counting completion,
3376 * attempt to decrement the counter without blocking. This
3377 * enables us to avoid waiting if the resource the completion
3378 * is protecting is not available.
3380 bool try_wait_for_completion(struct completion
*x
)
3382 unsigned long flags
;
3385 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3390 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3393 EXPORT_SYMBOL(try_wait_for_completion
);
3396 * completion_done - Test to see if a completion has any waiters
3397 * @x: completion structure
3399 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3400 * 1 if there are no waiters.
3403 bool completion_done(struct completion
*x
)
3405 unsigned long flags
;
3408 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3411 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3414 EXPORT_SYMBOL(completion_done
);
3417 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3419 unsigned long flags
;
3422 init_waitqueue_entry(&wait
, current
);
3424 __set_current_state(state
);
3426 spin_lock_irqsave(&q
->lock
, flags
);
3427 __add_wait_queue(q
, &wait
);
3428 spin_unlock(&q
->lock
);
3429 timeout
= schedule_timeout(timeout
);
3430 spin_lock_irq(&q
->lock
);
3431 __remove_wait_queue(q
, &wait
);
3432 spin_unlock_irqrestore(&q
->lock
, flags
);
3437 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3439 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3441 EXPORT_SYMBOL(interruptible_sleep_on
);
3444 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3446 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3448 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3450 void __sched
sleep_on(wait_queue_head_t
*q
)
3452 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3454 EXPORT_SYMBOL(sleep_on
);
3456 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3458 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3460 EXPORT_SYMBOL(sleep_on_timeout
);
3462 #ifdef CONFIG_RT_MUTEXES
3465 * rt_mutex_setprio - set the current priority of a task
3467 * @prio: prio value (kernel-internal form)
3469 * This function changes the 'effective' priority of a task. It does
3470 * not touch ->normal_prio like __setscheduler().
3472 * Used by the rt_mutex code to implement priority inheritance logic.
3474 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3476 int oldprio
, on_rq
, running
;
3478 const struct sched_class
*prev_class
;
3480 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3482 rq
= __task_rq_lock(p
);
3485 * Idle task boosting is a nono in general. There is one
3486 * exception, when PREEMPT_RT and NOHZ is active:
3488 * The idle task calls get_next_timer_interrupt() and holds
3489 * the timer wheel base->lock on the CPU and another CPU wants
3490 * to access the timer (probably to cancel it). We can safely
3491 * ignore the boosting request, as the idle CPU runs this code
3492 * with interrupts disabled and will complete the lock
3493 * protected section without being interrupted. So there is no
3494 * real need to boost.
3496 if (unlikely(p
== rq
->idle
)) {
3497 WARN_ON(p
!= rq
->curr
);
3498 WARN_ON(p
->pi_blocked_on
);
3502 trace_sched_pi_setprio(p
, prio
);
3504 prev_class
= p
->sched_class
;
3506 running
= task_current(rq
, p
);
3508 dequeue_task(rq
, p
, 0);
3510 p
->sched_class
->put_prev_task(rq
, p
);
3513 p
->sched_class
= &rt_sched_class
;
3515 p
->sched_class
= &fair_sched_class
;
3520 p
->sched_class
->set_curr_task(rq
);
3522 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3524 check_class_changed(rq
, p
, prev_class
, oldprio
);
3526 __task_rq_unlock(rq
);
3529 void set_user_nice(struct task_struct
*p
, long nice
)
3531 int old_prio
, delta
, on_rq
;
3532 unsigned long flags
;
3535 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3538 * We have to be careful, if called from sys_setpriority(),
3539 * the task might be in the middle of scheduling on another CPU.
3541 rq
= task_rq_lock(p
, &flags
);
3543 * The RT priorities are set via sched_setscheduler(), but we still
3544 * allow the 'normal' nice value to be set - but as expected
3545 * it wont have any effect on scheduling until the task is
3546 * SCHED_FIFO/SCHED_RR:
3548 if (task_has_rt_policy(p
)) {
3549 p
->static_prio
= NICE_TO_PRIO(nice
);
3554 dequeue_task(rq
, p
, 0);
3556 p
->static_prio
= NICE_TO_PRIO(nice
);
3559 p
->prio
= effective_prio(p
);
3560 delta
= p
->prio
- old_prio
;
3563 enqueue_task(rq
, p
, 0);
3565 * If the task increased its priority or is running and
3566 * lowered its priority, then reschedule its CPU:
3568 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3569 resched_task(rq
->curr
);
3572 task_rq_unlock(rq
, p
, &flags
);
3574 EXPORT_SYMBOL(set_user_nice
);
3577 * can_nice - check if a task can reduce its nice value
3581 int can_nice(const struct task_struct
*p
, const int nice
)
3583 /* convert nice value [19,-20] to rlimit style value [1,40] */
3584 int nice_rlim
= 20 - nice
;
3586 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3587 capable(CAP_SYS_NICE
));
3590 #ifdef __ARCH_WANT_SYS_NICE
3593 * sys_nice - change the priority of the current process.
3594 * @increment: priority increment
3596 * sys_setpriority is a more generic, but much slower function that
3597 * does similar things.
3599 SYSCALL_DEFINE1(nice
, int, increment
)
3604 * Setpriority might change our priority at the same moment.
3605 * We don't have to worry. Conceptually one call occurs first
3606 * and we have a single winner.
3608 if (increment
< -40)
3613 nice
= TASK_NICE(current
) + increment
;
3619 if (increment
< 0 && !can_nice(current
, nice
))
3622 retval
= security_task_setnice(current
, nice
);
3626 set_user_nice(current
, nice
);
3633 * task_prio - return the priority value of a given task.
3634 * @p: the task in question.
3636 * This is the priority value as seen by users in /proc.
3637 * RT tasks are offset by -200. Normal tasks are centered
3638 * around 0, value goes from -16 to +15.
3640 int task_prio(const struct task_struct
*p
)
3642 return p
->prio
- MAX_RT_PRIO
;
3646 * task_nice - return the nice value of a given task.
3647 * @p: the task in question.
3649 int task_nice(const struct task_struct
*p
)
3651 return TASK_NICE(p
);
3653 EXPORT_SYMBOL(task_nice
);
3656 * idle_cpu - is a given cpu idle currently?
3657 * @cpu: the processor in question.
3659 int idle_cpu(int cpu
)
3661 struct rq
*rq
= cpu_rq(cpu
);
3663 if (rq
->curr
!= rq
->idle
)
3670 if (!llist_empty(&rq
->wake_list
))
3678 * idle_task - return the idle task for a given cpu.
3679 * @cpu: the processor in question.
3681 struct task_struct
*idle_task(int cpu
)
3683 return cpu_rq(cpu
)->idle
;
3687 * find_process_by_pid - find a process with a matching PID value.
3688 * @pid: the pid in question.
3690 static struct task_struct
*find_process_by_pid(pid_t pid
)
3692 return pid
? find_task_by_vpid(pid
) : current
;
3695 /* Actually do priority change: must hold rq lock. */
3697 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3700 p
->rt_priority
= prio
;
3701 p
->normal_prio
= normal_prio(p
);
3702 /* we are holding p->pi_lock already */
3703 p
->prio
= rt_mutex_getprio(p
);
3704 if (rt_prio(p
->prio
))
3705 p
->sched_class
= &rt_sched_class
;
3707 p
->sched_class
= &fair_sched_class
;
3712 * check the target process has a UID that matches the current process's
3714 static bool check_same_owner(struct task_struct
*p
)
3716 const struct cred
*cred
= current_cred(), *pcred
;
3720 pcred
= __task_cred(p
);
3721 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3722 uid_eq(cred
->euid
, pcred
->uid
));
3727 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3728 const struct sched_param
*param
, bool user
)
3730 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3731 unsigned long flags
;
3732 const struct sched_class
*prev_class
;
3736 /* may grab non-irq protected spin_locks */
3737 BUG_ON(in_interrupt());
3739 /* double check policy once rq lock held */
3741 reset_on_fork
= p
->sched_reset_on_fork
;
3742 policy
= oldpolicy
= p
->policy
;
3744 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3745 policy
&= ~SCHED_RESET_ON_FORK
;
3747 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3748 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3749 policy
!= SCHED_IDLE
)
3754 * Valid priorities for SCHED_FIFO and SCHED_RR are
3755 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3756 * SCHED_BATCH and SCHED_IDLE is 0.
3758 if (param
->sched_priority
< 0 ||
3759 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3760 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3762 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3766 * Allow unprivileged RT tasks to decrease priority:
3768 if (user
&& !capable(CAP_SYS_NICE
)) {
3769 if (rt_policy(policy
)) {
3770 unsigned long rlim_rtprio
=
3771 task_rlimit(p
, RLIMIT_RTPRIO
);
3773 /* can't set/change the rt policy */
3774 if (policy
!= p
->policy
&& !rlim_rtprio
)
3777 /* can't increase priority */
3778 if (param
->sched_priority
> p
->rt_priority
&&
3779 param
->sched_priority
> rlim_rtprio
)
3784 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3785 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3787 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3788 if (!can_nice(p
, TASK_NICE(p
)))
3792 /* can't change other user's priorities */
3793 if (!check_same_owner(p
))
3796 /* Normal users shall not reset the sched_reset_on_fork flag */
3797 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3802 retval
= security_task_setscheduler(p
);
3808 * make sure no PI-waiters arrive (or leave) while we are
3809 * changing the priority of the task:
3811 * To be able to change p->policy safely, the appropriate
3812 * runqueue lock must be held.
3814 rq
= task_rq_lock(p
, &flags
);
3817 * Changing the policy of the stop threads its a very bad idea
3819 if (p
== rq
->stop
) {
3820 task_rq_unlock(rq
, p
, &flags
);
3825 * If not changing anything there's no need to proceed further:
3827 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3828 param
->sched_priority
== p
->rt_priority
))) {
3829 task_rq_unlock(rq
, p
, &flags
);
3833 #ifdef CONFIG_RT_GROUP_SCHED
3836 * Do not allow realtime tasks into groups that have no runtime
3839 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3840 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3841 !task_group_is_autogroup(task_group(p
))) {
3842 task_rq_unlock(rq
, p
, &flags
);
3848 /* recheck policy now with rq lock held */
3849 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3850 policy
= oldpolicy
= -1;
3851 task_rq_unlock(rq
, p
, &flags
);
3855 running
= task_current(rq
, p
);
3857 dequeue_task(rq
, p
, 0);
3859 p
->sched_class
->put_prev_task(rq
, p
);
3861 p
->sched_reset_on_fork
= reset_on_fork
;
3864 prev_class
= p
->sched_class
;
3865 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3868 p
->sched_class
->set_curr_task(rq
);
3870 enqueue_task(rq
, p
, 0);
3872 check_class_changed(rq
, p
, prev_class
, oldprio
);
3873 task_rq_unlock(rq
, p
, &flags
);
3875 rt_mutex_adjust_pi(p
);
3881 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3882 * @p: the task in question.
3883 * @policy: new policy.
3884 * @param: structure containing the new RT priority.
3886 * NOTE that the task may be already dead.
3888 int sched_setscheduler(struct task_struct
*p
, int policy
,
3889 const struct sched_param
*param
)
3891 return __sched_setscheduler(p
, policy
, param
, true);
3893 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3896 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3897 * @p: the task in question.
3898 * @policy: new policy.
3899 * @param: structure containing the new RT priority.
3901 * Just like sched_setscheduler, only don't bother checking if the
3902 * current context has permission. For example, this is needed in
3903 * stop_machine(): we create temporary high priority worker threads,
3904 * but our caller might not have that capability.
3906 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3907 const struct sched_param
*param
)
3909 return __sched_setscheduler(p
, policy
, param
, false);
3913 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3915 struct sched_param lparam
;
3916 struct task_struct
*p
;
3919 if (!param
|| pid
< 0)
3921 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3926 p
= find_process_by_pid(pid
);
3928 retval
= sched_setscheduler(p
, policy
, &lparam
);
3935 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3936 * @pid: the pid in question.
3937 * @policy: new policy.
3938 * @param: structure containing the new RT priority.
3940 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3941 struct sched_param __user
*, param
)
3943 /* negative values for policy are not valid */
3947 return do_sched_setscheduler(pid
, policy
, param
);
3951 * sys_sched_setparam - set/change the RT priority of a thread
3952 * @pid: the pid in question.
3953 * @param: structure containing the new RT priority.
3955 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3957 return do_sched_setscheduler(pid
, -1, param
);
3961 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3962 * @pid: the pid in question.
3964 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3966 struct task_struct
*p
;
3974 p
= find_process_by_pid(pid
);
3976 retval
= security_task_getscheduler(p
);
3979 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3986 * sys_sched_getparam - get the RT priority of a thread
3987 * @pid: the pid in question.
3988 * @param: structure containing the RT priority.
3990 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3992 struct sched_param lp
;
3993 struct task_struct
*p
;
3996 if (!param
|| pid
< 0)
4000 p
= find_process_by_pid(pid
);
4005 retval
= security_task_getscheduler(p
);
4009 lp
.sched_priority
= p
->rt_priority
;
4013 * This one might sleep, we cannot do it with a spinlock held ...
4015 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4024 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4026 cpumask_var_t cpus_allowed
, new_mask
;
4027 struct task_struct
*p
;
4033 p
= find_process_by_pid(pid
);
4040 /* Prevent p going away */
4044 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4048 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4050 goto out_free_cpus_allowed
;
4053 if (!check_same_owner(p
) && !ns_capable(task_user_ns(p
), CAP_SYS_NICE
))
4056 retval
= security_task_setscheduler(p
);
4060 cpuset_cpus_allowed(p
, cpus_allowed
);
4061 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4063 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4066 cpuset_cpus_allowed(p
, cpus_allowed
);
4067 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4069 * We must have raced with a concurrent cpuset
4070 * update. Just reset the cpus_allowed to the
4071 * cpuset's cpus_allowed
4073 cpumask_copy(new_mask
, cpus_allowed
);
4078 free_cpumask_var(new_mask
);
4079 out_free_cpus_allowed
:
4080 free_cpumask_var(cpus_allowed
);
4087 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4088 struct cpumask
*new_mask
)
4090 if (len
< cpumask_size())
4091 cpumask_clear(new_mask
);
4092 else if (len
> cpumask_size())
4093 len
= cpumask_size();
4095 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4099 * sys_sched_setaffinity - set the cpu affinity of a process
4100 * @pid: pid of the process
4101 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4102 * @user_mask_ptr: user-space pointer to the new cpu mask
4104 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4105 unsigned long __user
*, user_mask_ptr
)
4107 cpumask_var_t new_mask
;
4110 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4113 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4115 retval
= sched_setaffinity(pid
, new_mask
);
4116 free_cpumask_var(new_mask
);
4120 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4122 struct task_struct
*p
;
4123 unsigned long flags
;
4130 p
= find_process_by_pid(pid
);
4134 retval
= security_task_getscheduler(p
);
4138 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4139 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4140 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4150 * sys_sched_getaffinity - get the cpu affinity of a process
4151 * @pid: pid of the process
4152 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4153 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4155 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4156 unsigned long __user
*, user_mask_ptr
)
4161 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4163 if (len
& (sizeof(unsigned long)-1))
4166 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4169 ret
= sched_getaffinity(pid
, mask
);
4171 size_t retlen
= min_t(size_t, len
, cpumask_size());
4173 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4178 free_cpumask_var(mask
);
4184 * sys_sched_yield - yield the current processor to other threads.
4186 * This function yields the current CPU to other tasks. If there are no
4187 * other threads running on this CPU then this function will return.
4189 SYSCALL_DEFINE0(sched_yield
)
4191 struct rq
*rq
= this_rq_lock();
4193 schedstat_inc(rq
, yld_count
);
4194 current
->sched_class
->yield_task(rq
);
4197 * Since we are going to call schedule() anyway, there's
4198 * no need to preempt or enable interrupts:
4200 __release(rq
->lock
);
4201 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4202 do_raw_spin_unlock(&rq
->lock
);
4203 sched_preempt_enable_no_resched();
4210 static inline int should_resched(void)
4212 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4215 static void __cond_resched(void)
4217 add_preempt_count(PREEMPT_ACTIVE
);
4219 sub_preempt_count(PREEMPT_ACTIVE
);
4222 int __sched
_cond_resched(void)
4224 if (should_resched()) {
4230 EXPORT_SYMBOL(_cond_resched
);
4233 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4234 * call schedule, and on return reacquire the lock.
4236 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4237 * operations here to prevent schedule() from being called twice (once via
4238 * spin_unlock(), once by hand).
4240 int __cond_resched_lock(spinlock_t
*lock
)
4242 int resched
= should_resched();
4245 lockdep_assert_held(lock
);
4247 if (spin_needbreak(lock
) || resched
) {
4258 EXPORT_SYMBOL(__cond_resched_lock
);
4260 int __sched
__cond_resched_softirq(void)
4262 BUG_ON(!in_softirq());
4264 if (should_resched()) {
4272 EXPORT_SYMBOL(__cond_resched_softirq
);
4275 * yield - yield the current processor to other threads.
4277 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4279 * The scheduler is at all times free to pick the calling task as the most
4280 * eligible task to run, if removing the yield() call from your code breaks
4281 * it, its already broken.
4283 * Typical broken usage is:
4288 * where one assumes that yield() will let 'the other' process run that will
4289 * make event true. If the current task is a SCHED_FIFO task that will never
4290 * happen. Never use yield() as a progress guarantee!!
4292 * If you want to use yield() to wait for something, use wait_event().
4293 * If you want to use yield() to be 'nice' for others, use cond_resched().
4294 * If you still want to use yield(), do not!
4296 void __sched
yield(void)
4298 set_current_state(TASK_RUNNING
);
4301 EXPORT_SYMBOL(yield
);
4304 * yield_to - yield the current processor to another thread in
4305 * your thread group, or accelerate that thread toward the
4306 * processor it's on.
4308 * @preempt: whether task preemption is allowed or not
4310 * It's the caller's job to ensure that the target task struct
4311 * can't go away on us before we can do any checks.
4313 * Returns true if we indeed boosted the target task.
4315 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4317 struct task_struct
*curr
= current
;
4318 struct rq
*rq
, *p_rq
;
4319 unsigned long flags
;
4322 local_irq_save(flags
);
4327 double_rq_lock(rq
, p_rq
);
4328 while (task_rq(p
) != p_rq
) {
4329 double_rq_unlock(rq
, p_rq
);
4333 if (!curr
->sched_class
->yield_to_task
)
4336 if (curr
->sched_class
!= p
->sched_class
)
4339 if (task_running(p_rq
, p
) || p
->state
)
4342 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4344 schedstat_inc(rq
, yld_count
);
4346 * Make p's CPU reschedule; pick_next_entity takes care of
4349 if (preempt
&& rq
!= p_rq
)
4350 resched_task(p_rq
->curr
);
4353 * We might have set it in task_yield_fair(), but are
4354 * not going to schedule(), so don't want to skip
4357 rq
->skip_clock_update
= 0;
4361 double_rq_unlock(rq
, p_rq
);
4362 local_irq_restore(flags
);
4369 EXPORT_SYMBOL_GPL(yield_to
);
4372 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4373 * that process accounting knows that this is a task in IO wait state.
4375 void __sched
io_schedule(void)
4377 struct rq
*rq
= raw_rq();
4379 delayacct_blkio_start();
4380 atomic_inc(&rq
->nr_iowait
);
4381 blk_flush_plug(current
);
4382 current
->in_iowait
= 1;
4384 current
->in_iowait
= 0;
4385 atomic_dec(&rq
->nr_iowait
);
4386 delayacct_blkio_end();
4388 EXPORT_SYMBOL(io_schedule
);
4390 long __sched
io_schedule_timeout(long timeout
)
4392 struct rq
*rq
= raw_rq();
4395 delayacct_blkio_start();
4396 atomic_inc(&rq
->nr_iowait
);
4397 blk_flush_plug(current
);
4398 current
->in_iowait
= 1;
4399 ret
= schedule_timeout(timeout
);
4400 current
->in_iowait
= 0;
4401 atomic_dec(&rq
->nr_iowait
);
4402 delayacct_blkio_end();
4407 * sys_sched_get_priority_max - return maximum RT priority.
4408 * @policy: scheduling class.
4410 * this syscall returns the maximum rt_priority that can be used
4411 * by a given scheduling class.
4413 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4420 ret
= MAX_USER_RT_PRIO
-1;
4432 * sys_sched_get_priority_min - return minimum RT priority.
4433 * @policy: scheduling class.
4435 * this syscall returns the minimum rt_priority that can be used
4436 * by a given scheduling class.
4438 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4456 * sys_sched_rr_get_interval - return the default timeslice of a process.
4457 * @pid: pid of the process.
4458 * @interval: userspace pointer to the timeslice value.
4460 * this syscall writes the default timeslice value of a given process
4461 * into the user-space timespec buffer. A value of '0' means infinity.
4463 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4464 struct timespec __user
*, interval
)
4466 struct task_struct
*p
;
4467 unsigned int time_slice
;
4468 unsigned long flags
;
4478 p
= find_process_by_pid(pid
);
4482 retval
= security_task_getscheduler(p
);
4486 rq
= task_rq_lock(p
, &flags
);
4487 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4488 task_rq_unlock(rq
, p
, &flags
);
4491 jiffies_to_timespec(time_slice
, &t
);
4492 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4500 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4502 void sched_show_task(struct task_struct
*p
)
4504 unsigned long free
= 0;
4507 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4508 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4509 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4510 #if BITS_PER_LONG == 32
4511 if (state
== TASK_RUNNING
)
4512 printk(KERN_CONT
" running ");
4514 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4516 if (state
== TASK_RUNNING
)
4517 printk(KERN_CONT
" running task ");
4519 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4521 #ifdef CONFIG_DEBUG_STACK_USAGE
4522 free
= stack_not_used(p
);
4524 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4525 task_pid_nr(p
), task_pid_nr(rcu_dereference(p
->real_parent
)),
4526 (unsigned long)task_thread_info(p
)->flags
);
4528 show_stack(p
, NULL
);
4531 void show_state_filter(unsigned long state_filter
)
4533 struct task_struct
*g
, *p
;
4535 #if BITS_PER_LONG == 32
4537 " task PC stack pid father\n");
4540 " task PC stack pid father\n");
4543 do_each_thread(g
, p
) {
4545 * reset the NMI-timeout, listing all files on a slow
4546 * console might take a lot of time:
4548 touch_nmi_watchdog();
4549 if (!state_filter
|| (p
->state
& state_filter
))
4551 } while_each_thread(g
, p
);
4553 touch_all_softlockup_watchdogs();
4555 #ifdef CONFIG_SCHED_DEBUG
4556 sysrq_sched_debug_show();
4560 * Only show locks if all tasks are dumped:
4563 debug_show_all_locks();
4566 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4568 idle
->sched_class
= &idle_sched_class
;
4572 * init_idle - set up an idle thread for a given CPU
4573 * @idle: task in question
4574 * @cpu: cpu the idle task belongs to
4576 * NOTE: this function does not set the idle thread's NEED_RESCHED
4577 * flag, to make booting more robust.
4579 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4581 struct rq
*rq
= cpu_rq(cpu
);
4582 unsigned long flags
;
4584 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4587 idle
->state
= TASK_RUNNING
;
4588 idle
->se
.exec_start
= sched_clock();
4590 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4592 * We're having a chicken and egg problem, even though we are
4593 * holding rq->lock, the cpu isn't yet set to this cpu so the
4594 * lockdep check in task_group() will fail.
4596 * Similar case to sched_fork(). / Alternatively we could
4597 * use task_rq_lock() here and obtain the other rq->lock.
4602 __set_task_cpu(idle
, cpu
);
4605 rq
->curr
= rq
->idle
= idle
;
4606 #if defined(CONFIG_SMP)
4609 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4611 /* Set the preempt count _outside_ the spinlocks! */
4612 task_thread_info(idle
)->preempt_count
= 0;
4615 * The idle tasks have their own, simple scheduling class:
4617 idle
->sched_class
= &idle_sched_class
;
4618 ftrace_graph_init_idle_task(idle
, cpu
);
4619 #if defined(CONFIG_SMP)
4620 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4625 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4627 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4628 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4630 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4631 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4635 * This is how migration works:
4637 * 1) we invoke migration_cpu_stop() on the target CPU using
4639 * 2) stopper starts to run (implicitly forcing the migrated thread
4641 * 3) it checks whether the migrated task is still in the wrong runqueue.
4642 * 4) if it's in the wrong runqueue then the migration thread removes
4643 * it and puts it into the right queue.
4644 * 5) stopper completes and stop_one_cpu() returns and the migration
4649 * Change a given task's CPU affinity. Migrate the thread to a
4650 * proper CPU and schedule it away if the CPU it's executing on
4651 * is removed from the allowed bitmask.
4653 * NOTE: the caller must have a valid reference to the task, the
4654 * task must not exit() & deallocate itself prematurely. The
4655 * call is not atomic; no spinlocks may be held.
4657 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4659 unsigned long flags
;
4661 unsigned int dest_cpu
;
4664 rq
= task_rq_lock(p
, &flags
);
4666 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4669 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4674 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
4679 do_set_cpus_allowed(p
, new_mask
);
4681 /* Can the task run on the task's current CPU? If so, we're done */
4682 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4685 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4687 struct migration_arg arg
= { p
, dest_cpu
};
4688 /* Need help from migration thread: drop lock and wait. */
4689 task_rq_unlock(rq
, p
, &flags
);
4690 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4691 tlb_migrate_finish(p
->mm
);
4695 task_rq_unlock(rq
, p
, &flags
);
4699 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4702 * Move (not current) task off this cpu, onto dest cpu. We're doing
4703 * this because either it can't run here any more (set_cpus_allowed()
4704 * away from this CPU, or CPU going down), or because we're
4705 * attempting to rebalance this task on exec (sched_exec).
4707 * So we race with normal scheduler movements, but that's OK, as long
4708 * as the task is no longer on this CPU.
4710 * Returns non-zero if task was successfully migrated.
4712 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4714 struct rq
*rq_dest
, *rq_src
;
4717 if (unlikely(!cpu_active(dest_cpu
)))
4720 rq_src
= cpu_rq(src_cpu
);
4721 rq_dest
= cpu_rq(dest_cpu
);
4723 raw_spin_lock(&p
->pi_lock
);
4724 double_rq_lock(rq_src
, rq_dest
);
4725 /* Already moved. */
4726 if (task_cpu(p
) != src_cpu
)
4728 /* Affinity changed (again). */
4729 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4733 * If we're not on a rq, the next wake-up will ensure we're
4737 dequeue_task(rq_src
, p
, 0);
4738 set_task_cpu(p
, dest_cpu
);
4739 enqueue_task(rq_dest
, p
, 0);
4740 check_preempt_curr(rq_dest
, p
, 0);
4745 double_rq_unlock(rq_src
, rq_dest
);
4746 raw_spin_unlock(&p
->pi_lock
);
4751 * migration_cpu_stop - this will be executed by a highprio stopper thread
4752 * and performs thread migration by bumping thread off CPU then
4753 * 'pushing' onto another runqueue.
4755 static int migration_cpu_stop(void *data
)
4757 struct migration_arg
*arg
= data
;
4760 * The original target cpu might have gone down and we might
4761 * be on another cpu but it doesn't matter.
4763 local_irq_disable();
4764 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4769 #ifdef CONFIG_HOTPLUG_CPU
4772 * Ensures that the idle task is using init_mm right before its cpu goes
4775 void idle_task_exit(void)
4777 struct mm_struct
*mm
= current
->active_mm
;
4779 BUG_ON(cpu_online(smp_processor_id()));
4782 switch_mm(mm
, &init_mm
, current
);
4787 * Since this CPU is going 'away' for a while, fold any nr_active delta
4788 * we might have. Assumes we're called after migrate_tasks() so that the
4789 * nr_active count is stable.
4791 * Also see the comment "Global load-average calculations".
4793 static void calc_load_migrate(struct rq
*rq
)
4795 long delta
= calc_load_fold_active(rq
);
4797 atomic_long_add(delta
, &calc_load_tasks
);
4801 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4802 * try_to_wake_up()->select_task_rq().
4804 * Called with rq->lock held even though we'er in stop_machine() and
4805 * there's no concurrency possible, we hold the required locks anyway
4806 * because of lock validation efforts.
4808 static void migrate_tasks(unsigned int dead_cpu
)
4810 struct rq
*rq
= cpu_rq(dead_cpu
);
4811 struct task_struct
*next
, *stop
= rq
->stop
;
4815 * Fudge the rq selection such that the below task selection loop
4816 * doesn't get stuck on the currently eligible stop task.
4818 * We're currently inside stop_machine() and the rq is either stuck
4819 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4820 * either way we should never end up calling schedule() until we're
4827 * There's this thread running, bail when that's the only
4830 if (rq
->nr_running
== 1)
4833 next
= pick_next_task(rq
);
4835 next
->sched_class
->put_prev_task(rq
, next
);
4837 /* Find suitable destination for @next, with force if needed. */
4838 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4839 raw_spin_unlock(&rq
->lock
);
4841 __migrate_task(next
, dead_cpu
, dest_cpu
);
4843 raw_spin_lock(&rq
->lock
);
4849 #endif /* CONFIG_HOTPLUG_CPU */
4851 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4853 static struct ctl_table sd_ctl_dir
[] = {
4855 .procname
= "sched_domain",
4861 static struct ctl_table sd_ctl_root
[] = {
4863 .procname
= "kernel",
4865 .child
= sd_ctl_dir
,
4870 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4872 struct ctl_table
*entry
=
4873 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4878 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4880 struct ctl_table
*entry
;
4883 * In the intermediate directories, both the child directory and
4884 * procname are dynamically allocated and could fail but the mode
4885 * will always be set. In the lowest directory the names are
4886 * static strings and all have proc handlers.
4888 for (entry
= *tablep
; entry
->mode
; entry
++) {
4890 sd_free_ctl_entry(&entry
->child
);
4891 if (entry
->proc_handler
== NULL
)
4892 kfree(entry
->procname
);
4900 set_table_entry(struct ctl_table
*entry
,
4901 const char *procname
, void *data
, int maxlen
,
4902 umode_t mode
, proc_handler
*proc_handler
)
4904 entry
->procname
= procname
;
4906 entry
->maxlen
= maxlen
;
4908 entry
->proc_handler
= proc_handler
;
4911 static struct ctl_table
*
4912 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4914 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4919 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4920 sizeof(long), 0644, proc_doulongvec_minmax
);
4921 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4922 sizeof(long), 0644, proc_doulongvec_minmax
);
4923 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4924 sizeof(int), 0644, proc_dointvec_minmax
);
4925 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4926 sizeof(int), 0644, proc_dointvec_minmax
);
4927 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4928 sizeof(int), 0644, proc_dointvec_minmax
);
4929 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4930 sizeof(int), 0644, proc_dointvec_minmax
);
4931 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4932 sizeof(int), 0644, proc_dointvec_minmax
);
4933 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4934 sizeof(int), 0644, proc_dointvec_minmax
);
4935 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4936 sizeof(int), 0644, proc_dointvec_minmax
);
4937 set_table_entry(&table
[9], "cache_nice_tries",
4938 &sd
->cache_nice_tries
,
4939 sizeof(int), 0644, proc_dointvec_minmax
);
4940 set_table_entry(&table
[10], "flags", &sd
->flags
,
4941 sizeof(int), 0644, proc_dointvec_minmax
);
4942 set_table_entry(&table
[11], "name", sd
->name
,
4943 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
4944 /* &table[12] is terminator */
4949 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4951 struct ctl_table
*entry
, *table
;
4952 struct sched_domain
*sd
;
4953 int domain_num
= 0, i
;
4956 for_each_domain(cpu
, sd
)
4958 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4963 for_each_domain(cpu
, sd
) {
4964 snprintf(buf
, 32, "domain%d", i
);
4965 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4967 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4974 static struct ctl_table_header
*sd_sysctl_header
;
4975 static void register_sched_domain_sysctl(void)
4977 int i
, cpu_num
= num_possible_cpus();
4978 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4981 WARN_ON(sd_ctl_dir
[0].child
);
4982 sd_ctl_dir
[0].child
= entry
;
4987 for_each_possible_cpu(i
) {
4988 snprintf(buf
, 32, "cpu%d", i
);
4989 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4991 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4995 WARN_ON(sd_sysctl_header
);
4996 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4999 /* may be called multiple times per register */
5000 static void unregister_sched_domain_sysctl(void)
5002 if (sd_sysctl_header
)
5003 unregister_sysctl_table(sd_sysctl_header
);
5004 sd_sysctl_header
= NULL
;
5005 if (sd_ctl_dir
[0].child
)
5006 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5009 static void register_sched_domain_sysctl(void)
5012 static void unregister_sched_domain_sysctl(void)
5017 static void set_rq_online(struct rq
*rq
)
5020 const struct sched_class
*class;
5022 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5025 for_each_class(class) {
5026 if (class->rq_online
)
5027 class->rq_online(rq
);
5032 static void set_rq_offline(struct rq
*rq
)
5035 const struct sched_class
*class;
5037 for_each_class(class) {
5038 if (class->rq_offline
)
5039 class->rq_offline(rq
);
5042 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5048 * migration_call - callback that gets triggered when a CPU is added.
5049 * Here we can start up the necessary migration thread for the new CPU.
5051 static int __cpuinit
5052 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5054 int cpu
= (long)hcpu
;
5055 unsigned long flags
;
5056 struct rq
*rq
= cpu_rq(cpu
);
5058 switch (action
& ~CPU_TASKS_FROZEN
) {
5060 case CPU_UP_PREPARE
:
5061 rq
->calc_load_update
= calc_load_update
;
5065 /* Update our root-domain */
5066 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5068 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5072 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5075 #ifdef CONFIG_HOTPLUG_CPU
5077 sched_ttwu_pending();
5078 /* Update our root-domain */
5079 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5081 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5085 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5086 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5088 calc_load_migrate(rq
);
5093 update_max_interval();
5099 * Register at high priority so that task migration (migrate_all_tasks)
5100 * happens before everything else. This has to be lower priority than
5101 * the notifier in the perf_event subsystem, though.
5103 static struct notifier_block __cpuinitdata migration_notifier
= {
5104 .notifier_call
= migration_call
,
5105 .priority
= CPU_PRI_MIGRATION
,
5108 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5109 unsigned long action
, void *hcpu
)
5111 switch (action
& ~CPU_TASKS_FROZEN
) {
5113 case CPU_DOWN_FAILED
:
5114 set_cpu_active((long)hcpu
, true);
5121 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5122 unsigned long action
, void *hcpu
)
5124 switch (action
& ~CPU_TASKS_FROZEN
) {
5125 case CPU_DOWN_PREPARE
:
5126 set_cpu_active((long)hcpu
, false);
5133 static int __init
migration_init(void)
5135 void *cpu
= (void *)(long)smp_processor_id();
5138 /* Initialize migration for the boot CPU */
5139 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5140 BUG_ON(err
== NOTIFY_BAD
);
5141 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5142 register_cpu_notifier(&migration_notifier
);
5144 /* Register cpu active notifiers */
5145 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5146 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5150 early_initcall(migration_init
);
5155 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5157 #ifdef CONFIG_SCHED_DEBUG
5159 static __read_mostly
int sched_debug_enabled
;
5161 static int __init
sched_debug_setup(char *str
)
5163 sched_debug_enabled
= 1;
5167 early_param("sched_debug", sched_debug_setup
);
5169 static inline bool sched_debug(void)
5171 return sched_debug_enabled
;
5174 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5175 struct cpumask
*groupmask
)
5177 struct sched_group
*group
= sd
->groups
;
5180 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5181 cpumask_clear(groupmask
);
5183 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5185 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5186 printk("does not load-balance\n");
5188 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5193 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5195 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5196 printk(KERN_ERR
"ERROR: domain->span does not contain "
5199 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5200 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5204 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5208 printk(KERN_ERR
"ERROR: group is NULL\n");
5213 * Even though we initialize ->power to something semi-sane,
5214 * we leave power_orig unset. This allows us to detect if
5215 * domain iteration is still funny without causing /0 traps.
5217 if (!group
->sgp
->power_orig
) {
5218 printk(KERN_CONT
"\n");
5219 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5224 if (!cpumask_weight(sched_group_cpus(group
))) {
5225 printk(KERN_CONT
"\n");
5226 printk(KERN_ERR
"ERROR: empty group\n");
5230 if (!(sd
->flags
& SD_OVERLAP
) &&
5231 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5232 printk(KERN_CONT
"\n");
5233 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5237 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5239 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5241 printk(KERN_CONT
" %s", str
);
5242 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5243 printk(KERN_CONT
" (cpu_power = %d)",
5247 group
= group
->next
;
5248 } while (group
!= sd
->groups
);
5249 printk(KERN_CONT
"\n");
5251 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5252 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5255 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5256 printk(KERN_ERR
"ERROR: parent span is not a superset "
5257 "of domain->span\n");
5261 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5265 if (!sched_debug_enabled
)
5269 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5273 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5276 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5284 #else /* !CONFIG_SCHED_DEBUG */
5285 # define sched_domain_debug(sd, cpu) do { } while (0)
5286 static inline bool sched_debug(void)
5290 #endif /* CONFIG_SCHED_DEBUG */
5292 static int sd_degenerate(struct sched_domain
*sd
)
5294 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5297 /* Following flags need at least 2 groups */
5298 if (sd
->flags
& (SD_LOAD_BALANCE
|
5299 SD_BALANCE_NEWIDLE
|
5303 SD_SHARE_PKG_RESOURCES
)) {
5304 if (sd
->groups
!= sd
->groups
->next
)
5308 /* Following flags don't use groups */
5309 if (sd
->flags
& (SD_WAKE_AFFINE
))
5316 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5318 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5320 if (sd_degenerate(parent
))
5323 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5326 /* Flags needing groups don't count if only 1 group in parent */
5327 if (parent
->groups
== parent
->groups
->next
) {
5328 pflags
&= ~(SD_LOAD_BALANCE
|
5329 SD_BALANCE_NEWIDLE
|
5333 SD_SHARE_PKG_RESOURCES
);
5334 if (nr_node_ids
== 1)
5335 pflags
&= ~SD_SERIALIZE
;
5337 if (~cflags
& pflags
)
5343 static void free_rootdomain(struct rcu_head
*rcu
)
5345 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5347 cpupri_cleanup(&rd
->cpupri
);
5348 free_cpumask_var(rd
->rto_mask
);
5349 free_cpumask_var(rd
->online
);
5350 free_cpumask_var(rd
->span
);
5354 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5356 struct root_domain
*old_rd
= NULL
;
5357 unsigned long flags
;
5359 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5364 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5367 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5370 * If we dont want to free the old_rt yet then
5371 * set old_rd to NULL to skip the freeing later
5374 if (!atomic_dec_and_test(&old_rd
->refcount
))
5378 atomic_inc(&rd
->refcount
);
5381 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5382 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5385 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5388 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5391 static int init_rootdomain(struct root_domain
*rd
)
5393 memset(rd
, 0, sizeof(*rd
));
5395 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5397 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5399 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5402 if (cpupri_init(&rd
->cpupri
) != 0)
5407 free_cpumask_var(rd
->rto_mask
);
5409 free_cpumask_var(rd
->online
);
5411 free_cpumask_var(rd
->span
);
5417 * By default the system creates a single root-domain with all cpus as
5418 * members (mimicking the global state we have today).
5420 struct root_domain def_root_domain
;
5422 static void init_defrootdomain(void)
5424 init_rootdomain(&def_root_domain
);
5426 atomic_set(&def_root_domain
.refcount
, 1);
5429 static struct root_domain
*alloc_rootdomain(void)
5431 struct root_domain
*rd
;
5433 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5437 if (init_rootdomain(rd
) != 0) {
5445 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5447 struct sched_group
*tmp
, *first
;
5456 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5461 } while (sg
!= first
);
5464 static void free_sched_domain(struct rcu_head
*rcu
)
5466 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5469 * If its an overlapping domain it has private groups, iterate and
5472 if (sd
->flags
& SD_OVERLAP
) {
5473 free_sched_groups(sd
->groups
, 1);
5474 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5475 kfree(sd
->groups
->sgp
);
5481 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5483 call_rcu(&sd
->rcu
, free_sched_domain
);
5486 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5488 for (; sd
; sd
= sd
->parent
)
5489 destroy_sched_domain(sd
, cpu
);
5493 * Keep a special pointer to the highest sched_domain that has
5494 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5495 * allows us to avoid some pointer chasing select_idle_sibling().
5497 * Iterate domains and sched_groups downward, assigning CPUs to be
5498 * select_idle_sibling() hw buddy. Cross-wiring hw makes bouncing
5499 * due to random perturbation self canceling, ie sw buddies pull
5500 * their counterpart to their CPU's hw counterpart.
5502 * Also keep a unique ID per domain (we use the first cpu number in
5503 * the cpumask of the domain), this allows us to quickly tell if
5504 * two cpus are in the same cache domain, see cpus_share_cache().
5506 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5507 DEFINE_PER_CPU(int, sd_llc_id
);
5509 static void update_top_cache_domain(int cpu
)
5511 struct sched_domain
*sd
;
5514 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5516 struct sched_domain
*tmp
= sd
;
5517 struct sched_group
*sg
, *prev
;
5521 * Traverse to first CPU in group, and count hops
5522 * to cpu from there, switching direction on each
5523 * hop, never ever pointing the last CPU rightward.
5526 id
= cpumask_first(sched_domain_span(tmp
));
5527 prev
= sg
= tmp
->groups
;
5530 while (cpumask_first(sched_group_cpus(sg
)) != id
)
5533 while (!cpumask_test_cpu(cpu
, sched_group_cpus(sg
))) {
5539 /* A CPU went down, never point back to domain start. */
5540 if (right
&& cpumask_first(sched_group_cpus(sg
->next
)) == id
)
5543 sg
= right
? sg
->next
: prev
;
5544 tmp
->idle_buddy
= cpumask_first(sched_group_cpus(sg
));
5545 } while ((tmp
= tmp
->child
));
5547 id
= cpumask_first(sched_domain_span(sd
));
5550 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5551 per_cpu(sd_llc_id
, cpu
) = id
;
5555 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5556 * hold the hotplug lock.
5559 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5561 struct rq
*rq
= cpu_rq(cpu
);
5562 struct sched_domain
*tmp
;
5564 /* Remove the sched domains which do not contribute to scheduling. */
5565 for (tmp
= sd
; tmp
; ) {
5566 struct sched_domain
*parent
= tmp
->parent
;
5570 if (sd_parent_degenerate(tmp
, parent
)) {
5571 tmp
->parent
= parent
->parent
;
5573 parent
->parent
->child
= tmp
;
5574 destroy_sched_domain(parent
, cpu
);
5579 if (sd
&& sd_degenerate(sd
)) {
5582 destroy_sched_domain(tmp
, cpu
);
5587 sched_domain_debug(sd
, cpu
);
5589 rq_attach_root(rq
, rd
);
5591 rcu_assign_pointer(rq
->sd
, sd
);
5592 destroy_sched_domains(tmp
, cpu
);
5594 update_top_cache_domain(cpu
);
5597 /* cpus with isolated domains */
5598 static cpumask_var_t cpu_isolated_map
;
5600 /* Setup the mask of cpus configured for isolated domains */
5601 static int __init
isolated_cpu_setup(char *str
)
5603 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5604 cpulist_parse(str
, cpu_isolated_map
);
5608 __setup("isolcpus=", isolated_cpu_setup
);
5610 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5612 return cpumask_of_node(cpu_to_node(cpu
));
5616 struct sched_domain
**__percpu sd
;
5617 struct sched_group
**__percpu sg
;
5618 struct sched_group_power
**__percpu sgp
;
5622 struct sched_domain
** __percpu sd
;
5623 struct root_domain
*rd
;
5633 struct sched_domain_topology_level
;
5635 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5636 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5638 #define SDTL_OVERLAP 0x01
5640 struct sched_domain_topology_level
{
5641 sched_domain_init_f init
;
5642 sched_domain_mask_f mask
;
5645 struct sd_data data
;
5649 * Build an iteration mask that can exclude certain CPUs from the upwards
5652 * Asymmetric node setups can result in situations where the domain tree is of
5653 * unequal depth, make sure to skip domains that already cover the entire
5656 * In that case build_sched_domains() will have terminated the iteration early
5657 * and our sibling sd spans will be empty. Domains should always include the
5658 * cpu they're built on, so check that.
5661 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5663 const struct cpumask
*span
= sched_domain_span(sd
);
5664 struct sd_data
*sdd
= sd
->private;
5665 struct sched_domain
*sibling
;
5668 for_each_cpu(i
, span
) {
5669 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5670 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5673 cpumask_set_cpu(i
, sched_group_mask(sg
));
5678 * Return the canonical balance cpu for this group, this is the first cpu
5679 * of this group that's also in the iteration mask.
5681 int group_balance_cpu(struct sched_group
*sg
)
5683 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5687 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5689 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5690 const struct cpumask
*span
= sched_domain_span(sd
);
5691 struct cpumask
*covered
= sched_domains_tmpmask
;
5692 struct sd_data
*sdd
= sd
->private;
5693 struct sched_domain
*child
;
5696 cpumask_clear(covered
);
5698 for_each_cpu(i
, span
) {
5699 struct cpumask
*sg_span
;
5701 if (cpumask_test_cpu(i
, covered
))
5704 child
= *per_cpu_ptr(sdd
->sd
, i
);
5706 /* See the comment near build_group_mask(). */
5707 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5710 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5711 GFP_KERNEL
, cpu_to_node(cpu
));
5716 sg_span
= sched_group_cpus(sg
);
5718 child
= child
->child
;
5719 cpumask_copy(sg_span
, sched_domain_span(child
));
5721 cpumask_set_cpu(i
, sg_span
);
5723 cpumask_or(covered
, covered
, sg_span
);
5725 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5726 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5727 build_group_mask(sd
, sg
);
5730 * Initialize sgp->power such that even if we mess up the
5731 * domains and no possible iteration will get us here, we won't
5734 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5737 * Make sure the first group of this domain contains the
5738 * canonical balance cpu. Otherwise the sched_domain iteration
5739 * breaks. See update_sg_lb_stats().
5741 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5742 group_balance_cpu(sg
) == cpu
)
5752 sd
->groups
= groups
;
5757 free_sched_groups(first
, 0);
5762 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5764 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5765 struct sched_domain
*child
= sd
->child
;
5768 cpu
= cpumask_first(sched_domain_span(child
));
5771 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5772 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5773 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5780 * build_sched_groups will build a circular linked list of the groups
5781 * covered by the given span, and will set each group's ->cpumask correctly,
5782 * and ->cpu_power to 0.
5784 * Assumes the sched_domain tree is fully constructed
5787 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5789 struct sched_group
*first
= NULL
, *last
= NULL
;
5790 struct sd_data
*sdd
= sd
->private;
5791 const struct cpumask
*span
= sched_domain_span(sd
);
5792 struct cpumask
*covered
;
5795 get_group(cpu
, sdd
, &sd
->groups
);
5796 atomic_inc(&sd
->groups
->ref
);
5798 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5801 lockdep_assert_held(&sched_domains_mutex
);
5802 covered
= sched_domains_tmpmask
;
5804 cpumask_clear(covered
);
5806 for_each_cpu(i
, span
) {
5807 struct sched_group
*sg
;
5808 int group
= get_group(i
, sdd
, &sg
);
5811 if (cpumask_test_cpu(i
, covered
))
5814 cpumask_clear(sched_group_cpus(sg
));
5816 cpumask_setall(sched_group_mask(sg
));
5818 for_each_cpu(j
, span
) {
5819 if (get_group(j
, sdd
, NULL
) != group
)
5822 cpumask_set_cpu(j
, covered
);
5823 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5838 * Initialize sched groups cpu_power.
5840 * cpu_power indicates the capacity of sched group, which is used while
5841 * distributing the load between different sched groups in a sched domain.
5842 * Typically cpu_power for all the groups in a sched domain will be same unless
5843 * there are asymmetries in the topology. If there are asymmetries, group
5844 * having more cpu_power will pickup more load compared to the group having
5847 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5849 struct sched_group
*sg
= sd
->groups
;
5851 WARN_ON(!sd
|| !sg
);
5854 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5856 } while (sg
!= sd
->groups
);
5858 if (cpu
!= group_balance_cpu(sg
))
5861 update_group_power(sd
, cpu
);
5862 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5865 int __weak
arch_sd_sibling_asym_packing(void)
5867 return 0*SD_ASYM_PACKING
;
5871 * Initializers for schedule domains
5872 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5875 #ifdef CONFIG_SCHED_DEBUG
5876 # define SD_INIT_NAME(sd, type) sd->name = #type
5878 # define SD_INIT_NAME(sd, type) do { } while (0)
5881 #define SD_INIT_FUNC(type) \
5882 static noinline struct sched_domain * \
5883 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5885 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5886 *sd = SD_##type##_INIT; \
5887 SD_INIT_NAME(sd, type); \
5888 sd->private = &tl->data; \
5893 #ifdef CONFIG_SCHED_SMT
5894 SD_INIT_FUNC(SIBLING
)
5896 #ifdef CONFIG_SCHED_MC
5899 #ifdef CONFIG_SCHED_BOOK
5903 static int default_relax_domain_level
= -1;
5904 int sched_domain_level_max
;
5906 static int __init
setup_relax_domain_level(char *str
)
5908 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5909 pr_warn("Unable to set relax_domain_level\n");
5913 __setup("relax_domain_level=", setup_relax_domain_level
);
5915 static void set_domain_attribute(struct sched_domain
*sd
,
5916 struct sched_domain_attr
*attr
)
5920 if (!attr
|| attr
->relax_domain_level
< 0) {
5921 if (default_relax_domain_level
< 0)
5924 request
= default_relax_domain_level
;
5926 request
= attr
->relax_domain_level
;
5927 if (request
< sd
->level
) {
5928 /* turn off idle balance on this domain */
5929 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5931 /* turn on idle balance on this domain */
5932 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5936 static void __sdt_free(const struct cpumask
*cpu_map
);
5937 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5939 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5940 const struct cpumask
*cpu_map
)
5944 if (!atomic_read(&d
->rd
->refcount
))
5945 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5947 free_percpu(d
->sd
); /* fall through */
5949 __sdt_free(cpu_map
); /* fall through */
5955 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5956 const struct cpumask
*cpu_map
)
5958 memset(d
, 0, sizeof(*d
));
5960 if (__sdt_alloc(cpu_map
))
5961 return sa_sd_storage
;
5962 d
->sd
= alloc_percpu(struct sched_domain
*);
5964 return sa_sd_storage
;
5965 d
->rd
= alloc_rootdomain();
5968 return sa_rootdomain
;
5972 * NULL the sd_data elements we've used to build the sched_domain and
5973 * sched_group structure so that the subsequent __free_domain_allocs()
5974 * will not free the data we're using.
5976 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5978 struct sd_data
*sdd
= sd
->private;
5980 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5981 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5983 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5984 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5986 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5987 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5990 #ifdef CONFIG_SCHED_SMT
5991 static const struct cpumask
*cpu_smt_mask(int cpu
)
5993 return topology_thread_cpumask(cpu
);
5998 * Topology list, bottom-up.
6000 static struct sched_domain_topology_level default_topology
[] = {
6001 #ifdef CONFIG_SCHED_SMT
6002 { sd_init_SIBLING
, cpu_smt_mask
, },
6004 #ifdef CONFIG_SCHED_MC
6005 { sd_init_MC
, cpu_coregroup_mask
, },
6007 #ifdef CONFIG_SCHED_BOOK
6008 { sd_init_BOOK
, cpu_book_mask
, },
6010 { sd_init_CPU
, cpu_cpu_mask
, },
6014 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6018 static int sched_domains_numa_levels
;
6019 static int *sched_domains_numa_distance
;
6020 static struct cpumask
***sched_domains_numa_masks
;
6021 static int sched_domains_curr_level
;
6023 static inline int sd_local_flags(int level
)
6025 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6028 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6031 static struct sched_domain
*
6032 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6034 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6035 int level
= tl
->numa_level
;
6036 int sd_weight
= cpumask_weight(
6037 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6039 *sd
= (struct sched_domain
){
6040 .min_interval
= sd_weight
,
6041 .max_interval
= 2*sd_weight
,
6043 .imbalance_pct
= 125,
6044 .cache_nice_tries
= 2,
6051 .flags
= 1*SD_LOAD_BALANCE
6052 | 1*SD_BALANCE_NEWIDLE
6057 | 0*SD_SHARE_CPUPOWER
6058 | 0*SD_SHARE_PKG_RESOURCES
6060 | 0*SD_PREFER_SIBLING
6061 | sd_local_flags(level
)
6063 .last_balance
= jiffies
,
6064 .balance_interval
= sd_weight
,
6066 SD_INIT_NAME(sd
, NUMA
);
6067 sd
->private = &tl
->data
;
6070 * Ugly hack to pass state to sd_numa_mask()...
6072 sched_domains_curr_level
= tl
->numa_level
;
6077 static const struct cpumask
*sd_numa_mask(int cpu
)
6079 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6082 static void sched_numa_warn(const char *str
)
6084 static int done
= false;
6092 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6094 for (i
= 0; i
< nr_node_ids
; i
++) {
6095 printk(KERN_WARNING
" ");
6096 for (j
= 0; j
< nr_node_ids
; j
++)
6097 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6098 printk(KERN_CONT
"\n");
6100 printk(KERN_WARNING
"\n");
6103 static bool find_numa_distance(int distance
)
6107 if (distance
== node_distance(0, 0))
6110 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6111 if (sched_domains_numa_distance
[i
] == distance
)
6118 static void sched_init_numa(void)
6120 int next_distance
, curr_distance
= node_distance(0, 0);
6121 struct sched_domain_topology_level
*tl
;
6125 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6126 if (!sched_domains_numa_distance
)
6130 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6131 * unique distances in the node_distance() table.
6133 * Assumes node_distance(0,j) includes all distances in
6134 * node_distance(i,j) in order to avoid cubic time.
6136 next_distance
= curr_distance
;
6137 for (i
= 0; i
< nr_node_ids
; i
++) {
6138 for (j
= 0; j
< nr_node_ids
; j
++) {
6139 for (k
= 0; k
< nr_node_ids
; k
++) {
6140 int distance
= node_distance(i
, k
);
6142 if (distance
> curr_distance
&&
6143 (distance
< next_distance
||
6144 next_distance
== curr_distance
))
6145 next_distance
= distance
;
6148 * While not a strong assumption it would be nice to know
6149 * about cases where if node A is connected to B, B is not
6150 * equally connected to A.
6152 if (sched_debug() && node_distance(k
, i
) != distance
)
6153 sched_numa_warn("Node-distance not symmetric");
6155 if (sched_debug() && i
&& !find_numa_distance(distance
))
6156 sched_numa_warn("Node-0 not representative");
6158 if (next_distance
!= curr_distance
) {
6159 sched_domains_numa_distance
[level
++] = next_distance
;
6160 sched_domains_numa_levels
= level
;
6161 curr_distance
= next_distance
;
6166 * In case of sched_debug() we verify the above assumption.
6172 * 'level' contains the number of unique distances, excluding the
6173 * identity distance node_distance(i,i).
6175 * The sched_domains_nume_distance[] array includes the actual distance
6179 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6180 if (!sched_domains_numa_masks
)
6184 * Now for each level, construct a mask per node which contains all
6185 * cpus of nodes that are that many hops away from us.
6187 for (i
= 0; i
< level
; i
++) {
6188 sched_domains_numa_masks
[i
] =
6189 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6190 if (!sched_domains_numa_masks
[i
])
6193 for (j
= 0; j
< nr_node_ids
; j
++) {
6194 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6198 sched_domains_numa_masks
[i
][j
] = mask
;
6200 for (k
= 0; k
< nr_node_ids
; k
++) {
6201 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6204 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6209 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6210 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6215 * Copy the default topology bits..
6217 for (i
= 0; default_topology
[i
].init
; i
++)
6218 tl
[i
] = default_topology
[i
];
6221 * .. and append 'j' levels of NUMA goodness.
6223 for (j
= 0; j
< level
; i
++, j
++) {
6224 tl
[i
] = (struct sched_domain_topology_level
){
6225 .init
= sd_numa_init
,
6226 .mask
= sd_numa_mask
,
6227 .flags
= SDTL_OVERLAP
,
6232 sched_domain_topology
= tl
;
6235 static inline void sched_init_numa(void)
6238 #endif /* CONFIG_NUMA */
6240 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6242 struct sched_domain_topology_level
*tl
;
6245 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6246 struct sd_data
*sdd
= &tl
->data
;
6248 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6252 sdd
->sg
= alloc_percpu(struct sched_group
*);
6256 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6260 for_each_cpu(j
, cpu_map
) {
6261 struct sched_domain
*sd
;
6262 struct sched_group
*sg
;
6263 struct sched_group_power
*sgp
;
6265 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6266 GFP_KERNEL
, cpu_to_node(j
));
6270 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6272 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6273 GFP_KERNEL
, cpu_to_node(j
));
6279 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6281 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6282 GFP_KERNEL
, cpu_to_node(j
));
6286 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6293 static void __sdt_free(const struct cpumask
*cpu_map
)
6295 struct sched_domain_topology_level
*tl
;
6298 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6299 struct sd_data
*sdd
= &tl
->data
;
6301 for_each_cpu(j
, cpu_map
) {
6302 struct sched_domain
*sd
;
6305 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6306 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6307 free_sched_groups(sd
->groups
, 0);
6308 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6312 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6314 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6316 free_percpu(sdd
->sd
);
6318 free_percpu(sdd
->sg
);
6320 free_percpu(sdd
->sgp
);
6325 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6326 struct s_data
*d
, const struct cpumask
*cpu_map
,
6327 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6330 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6334 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6336 sd
->level
= child
->level
+ 1;
6337 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6341 set_domain_attribute(sd
, attr
);
6347 * Build sched domains for a given set of cpus and attach the sched domains
6348 * to the individual cpus
6350 static int build_sched_domains(const struct cpumask
*cpu_map
,
6351 struct sched_domain_attr
*attr
)
6353 enum s_alloc alloc_state
= sa_none
;
6354 struct sched_domain
*sd
;
6356 int i
, ret
= -ENOMEM
;
6358 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6359 if (alloc_state
!= sa_rootdomain
)
6362 /* Set up domains for cpus specified by the cpu_map. */
6363 for_each_cpu(i
, cpu_map
) {
6364 struct sched_domain_topology_level
*tl
;
6367 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6368 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6369 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6370 sd
->flags
|= SD_OVERLAP
;
6371 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6378 *per_cpu_ptr(d
.sd
, i
) = sd
;
6381 /* Build the groups for the domains */
6382 for_each_cpu(i
, cpu_map
) {
6383 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6384 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6385 if (sd
->flags
& SD_OVERLAP
) {
6386 if (build_overlap_sched_groups(sd
, i
))
6389 if (build_sched_groups(sd
, i
))
6395 /* Calculate CPU power for physical packages and nodes */
6396 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6397 if (!cpumask_test_cpu(i
, cpu_map
))
6400 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6401 claim_allocations(i
, sd
);
6402 init_sched_groups_power(i
, sd
);
6406 /* Attach the domains */
6408 for_each_cpu(i
, cpu_map
) {
6409 sd
= *per_cpu_ptr(d
.sd
, i
);
6410 cpu_attach_domain(sd
, d
.rd
, i
);
6416 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6420 static cpumask_var_t
*doms_cur
; /* current sched domains */
6421 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6422 static struct sched_domain_attr
*dattr_cur
;
6423 /* attribues of custom domains in 'doms_cur' */
6426 * Special case: If a kmalloc of a doms_cur partition (array of
6427 * cpumask) fails, then fallback to a single sched domain,
6428 * as determined by the single cpumask fallback_doms.
6430 static cpumask_var_t fallback_doms
;
6433 * arch_update_cpu_topology lets virtualized architectures update the
6434 * cpu core maps. It is supposed to return 1 if the topology changed
6435 * or 0 if it stayed the same.
6437 int __attribute__((weak
)) arch_update_cpu_topology(void)
6442 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6445 cpumask_var_t
*doms
;
6447 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6450 for (i
= 0; i
< ndoms
; i
++) {
6451 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6452 free_sched_domains(doms
, i
);
6459 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6462 for (i
= 0; i
< ndoms
; i
++)
6463 free_cpumask_var(doms
[i
]);
6468 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6469 * For now this just excludes isolated cpus, but could be used to
6470 * exclude other special cases in the future.
6472 static int init_sched_domains(const struct cpumask
*cpu_map
)
6476 arch_update_cpu_topology();
6478 doms_cur
= alloc_sched_domains(ndoms_cur
);
6480 doms_cur
= &fallback_doms
;
6481 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6482 err
= build_sched_domains(doms_cur
[0], NULL
);
6483 register_sched_domain_sysctl();
6489 * Detach sched domains from a group of cpus specified in cpu_map
6490 * These cpus will now be attached to the NULL domain
6492 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6497 for_each_cpu(i
, cpu_map
)
6498 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6502 /* handle null as "default" */
6503 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6504 struct sched_domain_attr
*new, int idx_new
)
6506 struct sched_domain_attr tmp
;
6513 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6514 new ? (new + idx_new
) : &tmp
,
6515 sizeof(struct sched_domain_attr
));
6519 * Partition sched domains as specified by the 'ndoms_new'
6520 * cpumasks in the array doms_new[] of cpumasks. This compares
6521 * doms_new[] to the current sched domain partitioning, doms_cur[].
6522 * It destroys each deleted domain and builds each new domain.
6524 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6525 * The masks don't intersect (don't overlap.) We should setup one
6526 * sched domain for each mask. CPUs not in any of the cpumasks will
6527 * not be load balanced. If the same cpumask appears both in the
6528 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6531 * The passed in 'doms_new' should be allocated using
6532 * alloc_sched_domains. This routine takes ownership of it and will
6533 * free_sched_domains it when done with it. If the caller failed the
6534 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6535 * and partition_sched_domains() will fallback to the single partition
6536 * 'fallback_doms', it also forces the domains to be rebuilt.
6538 * If doms_new == NULL it will be replaced with cpu_online_mask.
6539 * ndoms_new == 0 is a special case for destroying existing domains,
6540 * and it will not create the default domain.
6542 * Call with hotplug lock held
6544 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6545 struct sched_domain_attr
*dattr_new
)
6550 mutex_lock(&sched_domains_mutex
);
6552 /* always unregister in case we don't destroy any domains */
6553 unregister_sched_domain_sysctl();
6555 /* Let architecture update cpu core mappings. */
6556 new_topology
= arch_update_cpu_topology();
6558 n
= doms_new
? ndoms_new
: 0;
6560 /* Destroy deleted domains */
6561 for (i
= 0; i
< ndoms_cur
; i
++) {
6562 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6563 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6564 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6567 /* no match - a current sched domain not in new doms_new[] */
6568 detach_destroy_domains(doms_cur
[i
]);
6573 if (doms_new
== NULL
) {
6575 doms_new
= &fallback_doms
;
6576 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6577 WARN_ON_ONCE(dattr_new
);
6580 /* Build new domains */
6581 for (i
= 0; i
< ndoms_new
; i
++) {
6582 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6583 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6584 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6587 /* no match - add a new doms_new */
6588 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6593 /* Remember the new sched domains */
6594 if (doms_cur
!= &fallback_doms
)
6595 free_sched_domains(doms_cur
, ndoms_cur
);
6596 kfree(dattr_cur
); /* kfree(NULL) is safe */
6597 doms_cur
= doms_new
;
6598 dattr_cur
= dattr_new
;
6599 ndoms_cur
= ndoms_new
;
6601 register_sched_domain_sysctl();
6603 mutex_unlock(&sched_domains_mutex
);
6606 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6609 * Update cpusets according to cpu_active mask. If cpusets are
6610 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6611 * around partition_sched_domains().
6613 * If we come here as part of a suspend/resume, don't touch cpusets because we
6614 * want to restore it back to its original state upon resume anyway.
6616 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6620 case CPU_ONLINE_FROZEN
:
6621 case CPU_DOWN_FAILED_FROZEN
:
6624 * num_cpus_frozen tracks how many CPUs are involved in suspend
6625 * resume sequence. As long as this is not the last online
6626 * operation in the resume sequence, just build a single sched
6627 * domain, ignoring cpusets.
6630 if (likely(num_cpus_frozen
)) {
6631 partition_sched_domains(1, NULL
, NULL
);
6636 * This is the last CPU online operation. So fall through and
6637 * restore the original sched domains by considering the
6638 * cpuset configurations.
6642 case CPU_DOWN_FAILED
:
6643 cpuset_update_active_cpus(true);
6651 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6655 case CPU_DOWN_PREPARE
:
6656 cpuset_update_active_cpus(false);
6658 case CPU_DOWN_PREPARE_FROZEN
:
6660 partition_sched_domains(1, NULL
, NULL
);
6668 void __init
sched_init_smp(void)
6670 cpumask_var_t non_isolated_cpus
;
6672 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6673 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6678 mutex_lock(&sched_domains_mutex
);
6679 init_sched_domains(cpu_active_mask
);
6680 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6681 if (cpumask_empty(non_isolated_cpus
))
6682 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6683 mutex_unlock(&sched_domains_mutex
);
6686 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6687 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6689 /* RT runtime code needs to handle some hotplug events */
6690 hotcpu_notifier(update_runtime
, 0);
6694 /* Move init over to a non-isolated CPU */
6695 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6697 sched_init_granularity();
6698 free_cpumask_var(non_isolated_cpus
);
6700 init_sched_rt_class();
6703 void __init
sched_init_smp(void)
6705 sched_init_granularity();
6707 #endif /* CONFIG_SMP */
6709 const_debug
unsigned int sysctl_timer_migration
= 1;
6711 int in_sched_functions(unsigned long addr
)
6713 return in_lock_functions(addr
) ||
6714 (addr
>= (unsigned long)__sched_text_start
6715 && addr
< (unsigned long)__sched_text_end
);
6718 #ifdef CONFIG_CGROUP_SCHED
6719 struct task_group root_task_group
;
6720 LIST_HEAD(task_groups
);
6723 DECLARE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
6725 void __init
sched_init(void)
6728 unsigned long alloc_size
= 0, ptr
;
6730 #ifdef CONFIG_FAIR_GROUP_SCHED
6731 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6733 #ifdef CONFIG_RT_GROUP_SCHED
6734 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6736 #ifdef CONFIG_CPUMASK_OFFSTACK
6737 alloc_size
+= num_possible_cpus() * cpumask_size();
6740 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6742 #ifdef CONFIG_FAIR_GROUP_SCHED
6743 root_task_group
.se
= (struct sched_entity
**)ptr
;
6744 ptr
+= nr_cpu_ids
* sizeof(void **);
6746 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6747 ptr
+= nr_cpu_ids
* sizeof(void **);
6749 #endif /* CONFIG_FAIR_GROUP_SCHED */
6750 #ifdef CONFIG_RT_GROUP_SCHED
6751 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6752 ptr
+= nr_cpu_ids
* sizeof(void **);
6754 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6755 ptr
+= nr_cpu_ids
* sizeof(void **);
6757 #endif /* CONFIG_RT_GROUP_SCHED */
6758 #ifdef CONFIG_CPUMASK_OFFSTACK
6759 for_each_possible_cpu(i
) {
6760 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
6761 ptr
+= cpumask_size();
6763 #endif /* CONFIG_CPUMASK_OFFSTACK */
6767 init_defrootdomain();
6770 init_rt_bandwidth(&def_rt_bandwidth
,
6771 global_rt_period(), global_rt_runtime());
6773 #ifdef CONFIG_RT_GROUP_SCHED
6774 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6775 global_rt_period(), global_rt_runtime());
6776 #endif /* CONFIG_RT_GROUP_SCHED */
6778 #ifdef CONFIG_CGROUP_SCHED
6779 list_add(&root_task_group
.list
, &task_groups
);
6780 INIT_LIST_HEAD(&root_task_group
.children
);
6781 INIT_LIST_HEAD(&root_task_group
.siblings
);
6782 autogroup_init(&init_task
);
6784 #endif /* CONFIG_CGROUP_SCHED */
6786 #ifdef CONFIG_CGROUP_CPUACCT
6787 root_cpuacct
.cpustat
= &kernel_cpustat
;
6788 root_cpuacct
.cpuusage
= alloc_percpu(u64
);
6789 /* Too early, not expected to fail */
6790 BUG_ON(!root_cpuacct
.cpuusage
);
6792 for_each_possible_cpu(i
) {
6796 raw_spin_lock_init(&rq
->lock
);
6798 rq
->calc_load_active
= 0;
6799 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6800 init_cfs_rq(&rq
->cfs
);
6801 init_rt_rq(&rq
->rt
, rq
);
6802 #ifdef CONFIG_FAIR_GROUP_SCHED
6803 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6804 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6806 * How much cpu bandwidth does root_task_group get?
6808 * In case of task-groups formed thr' the cgroup filesystem, it
6809 * gets 100% of the cpu resources in the system. This overall
6810 * system cpu resource is divided among the tasks of
6811 * root_task_group and its child task-groups in a fair manner,
6812 * based on each entity's (task or task-group's) weight
6813 * (se->load.weight).
6815 * In other words, if root_task_group has 10 tasks of weight
6816 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6817 * then A0's share of the cpu resource is:
6819 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6821 * We achieve this by letting root_task_group's tasks sit
6822 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6824 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6825 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6826 #endif /* CONFIG_FAIR_GROUP_SCHED */
6828 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6829 #ifdef CONFIG_RT_GROUP_SCHED
6830 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6831 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6834 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6835 rq
->cpu_load
[j
] = 0;
6837 rq
->last_load_update_tick
= jiffies
;
6842 rq
->cpu_power
= SCHED_POWER_SCALE
;
6843 rq
->post_schedule
= 0;
6844 rq
->active_balance
= 0;
6845 rq
->next_balance
= jiffies
;
6850 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6852 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6854 rq_attach_root(rq
, &def_root_domain
);
6860 atomic_set(&rq
->nr_iowait
, 0);
6863 set_load_weight(&init_task
);
6865 #ifdef CONFIG_PREEMPT_NOTIFIERS
6866 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6869 #ifdef CONFIG_RT_MUTEXES
6870 plist_head_init(&init_task
.pi_waiters
);
6874 * The boot idle thread does lazy MMU switching as well:
6876 atomic_inc(&init_mm
.mm_count
);
6877 enter_lazy_tlb(&init_mm
, current
);
6880 * Make us the idle thread. Technically, schedule() should not be
6881 * called from this thread, however somewhere below it might be,
6882 * but because we are the idle thread, we just pick up running again
6883 * when this runqueue becomes "idle".
6885 init_idle(current
, smp_processor_id());
6887 calc_load_update
= jiffies
+ LOAD_FREQ
;
6890 * During early bootup we pretend to be a normal task:
6892 current
->sched_class
= &fair_sched_class
;
6895 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6896 /* May be allocated at isolcpus cmdline parse time */
6897 if (cpu_isolated_map
== NULL
)
6898 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6899 idle_thread_set_boot_cpu();
6901 init_sched_fair_class();
6903 scheduler_running
= 1;
6906 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6907 static inline int preempt_count_equals(int preempt_offset
)
6909 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6911 return (nested
== preempt_offset
);
6914 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6916 static unsigned long prev_jiffy
; /* ratelimiting */
6918 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6919 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6920 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6922 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6924 prev_jiffy
= jiffies
;
6927 "BUG: sleeping function called from invalid context at %s:%d\n",
6930 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6931 in_atomic(), irqs_disabled(),
6932 current
->pid
, current
->comm
);
6934 debug_show_held_locks(current
);
6935 if (irqs_disabled())
6936 print_irqtrace_events(current
);
6939 EXPORT_SYMBOL(__might_sleep
);
6942 #ifdef CONFIG_MAGIC_SYSRQ
6943 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6945 const struct sched_class
*prev_class
= p
->sched_class
;
6946 int old_prio
= p
->prio
;
6951 dequeue_task(rq
, p
, 0);
6952 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6954 enqueue_task(rq
, p
, 0);
6955 resched_task(rq
->curr
);
6958 check_class_changed(rq
, p
, prev_class
, old_prio
);
6961 void normalize_rt_tasks(void)
6963 struct task_struct
*g
, *p
;
6964 unsigned long flags
;
6967 read_lock_irqsave(&tasklist_lock
, flags
);
6968 do_each_thread(g
, p
) {
6970 * Only normalize user tasks:
6975 p
->se
.exec_start
= 0;
6976 #ifdef CONFIG_SCHEDSTATS
6977 p
->se
.statistics
.wait_start
= 0;
6978 p
->se
.statistics
.sleep_start
= 0;
6979 p
->se
.statistics
.block_start
= 0;
6984 * Renice negative nice level userspace
6987 if (TASK_NICE(p
) < 0 && p
->mm
)
6988 set_user_nice(p
, 0);
6992 raw_spin_lock(&p
->pi_lock
);
6993 rq
= __task_rq_lock(p
);
6995 normalize_task(rq
, p
);
6997 __task_rq_unlock(rq
);
6998 raw_spin_unlock(&p
->pi_lock
);
6999 } while_each_thread(g
, p
);
7001 read_unlock_irqrestore(&tasklist_lock
, flags
);
7004 #endif /* CONFIG_MAGIC_SYSRQ */
7006 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7008 * These functions are only useful for the IA64 MCA handling, or kdb.
7010 * They can only be called when the whole system has been
7011 * stopped - every CPU needs to be quiescent, and no scheduling
7012 * activity can take place. Using them for anything else would
7013 * be a serious bug, and as a result, they aren't even visible
7014 * under any other configuration.
7018 * curr_task - return the current task for a given cpu.
7019 * @cpu: the processor in question.
7021 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7023 struct task_struct
*curr_task(int cpu
)
7025 return cpu_curr(cpu
);
7028 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7032 * set_curr_task - set the current task for a given cpu.
7033 * @cpu: the processor in question.
7034 * @p: the task pointer to set.
7036 * Description: This function must only be used when non-maskable interrupts
7037 * are serviced on a separate stack. It allows the architecture to switch the
7038 * notion of the current task on a cpu in a non-blocking manner. This function
7039 * must be called with all CPU's synchronized, and interrupts disabled, the
7040 * and caller must save the original value of the current task (see
7041 * curr_task() above) and restore that value before reenabling interrupts and
7042 * re-starting the system.
7044 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7046 void set_curr_task(int cpu
, struct task_struct
*p
)
7053 #ifdef CONFIG_CGROUP_SCHED
7054 /* task_group_lock serializes the addition/removal of task groups */
7055 static DEFINE_SPINLOCK(task_group_lock
);
7057 static void free_sched_group(struct task_group
*tg
)
7059 free_fair_sched_group(tg
);
7060 free_rt_sched_group(tg
);
7065 /* allocate runqueue etc for a new task group */
7066 struct task_group
*sched_create_group(struct task_group
*parent
)
7068 struct task_group
*tg
;
7069 unsigned long flags
;
7071 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7073 return ERR_PTR(-ENOMEM
);
7075 if (!alloc_fair_sched_group(tg
, parent
))
7078 if (!alloc_rt_sched_group(tg
, parent
))
7081 spin_lock_irqsave(&task_group_lock
, flags
);
7082 list_add_rcu(&tg
->list
, &task_groups
);
7084 WARN_ON(!parent
); /* root should already exist */
7086 tg
->parent
= parent
;
7087 INIT_LIST_HEAD(&tg
->children
);
7088 list_add_rcu(&tg
->siblings
, &parent
->children
);
7089 spin_unlock_irqrestore(&task_group_lock
, flags
);
7094 free_sched_group(tg
);
7095 return ERR_PTR(-ENOMEM
);
7098 /* rcu callback to free various structures associated with a task group */
7099 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7101 /* now it should be safe to free those cfs_rqs */
7102 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7105 /* Destroy runqueue etc associated with a task group */
7106 void sched_destroy_group(struct task_group
*tg
)
7108 unsigned long flags
;
7111 /* end participation in shares distribution */
7112 for_each_possible_cpu(i
)
7113 unregister_fair_sched_group(tg
, i
);
7115 spin_lock_irqsave(&task_group_lock
, flags
);
7116 list_del_rcu(&tg
->list
);
7117 list_del_rcu(&tg
->siblings
);
7118 spin_unlock_irqrestore(&task_group_lock
, flags
);
7120 /* wait for possible concurrent references to cfs_rqs complete */
7121 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7124 /* change task's runqueue when it moves between groups.
7125 * The caller of this function should have put the task in its new group
7126 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7127 * reflect its new group.
7129 void sched_move_task(struct task_struct
*tsk
)
7131 struct task_group
*tg
;
7133 unsigned long flags
;
7136 rq
= task_rq_lock(tsk
, &flags
);
7138 running
= task_current(rq
, tsk
);
7142 dequeue_task(rq
, tsk
, 0);
7143 if (unlikely(running
))
7144 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7146 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7147 lockdep_is_held(&tsk
->sighand
->siglock
)),
7148 struct task_group
, css
);
7149 tg
= autogroup_task_group(tsk
, tg
);
7150 tsk
->sched_task_group
= tg
;
7152 #ifdef CONFIG_FAIR_GROUP_SCHED
7153 if (tsk
->sched_class
->task_move_group
)
7154 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7157 set_task_rq(tsk
, task_cpu(tsk
));
7159 if (unlikely(running
))
7160 tsk
->sched_class
->set_curr_task(rq
);
7162 enqueue_task(rq
, tsk
, 0);
7164 task_rq_unlock(rq
, tsk
, &flags
);
7166 #endif /* CONFIG_CGROUP_SCHED */
7168 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7169 static unsigned long to_ratio(u64 period
, u64 runtime
)
7171 if (runtime
== RUNTIME_INF
)
7174 return div64_u64(runtime
<< 20, period
);
7178 #ifdef CONFIG_RT_GROUP_SCHED
7180 * Ensure that the real time constraints are schedulable.
7182 static DEFINE_MUTEX(rt_constraints_mutex
);
7184 /* Must be called with tasklist_lock held */
7185 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7187 struct task_struct
*g
, *p
;
7189 do_each_thread(g
, p
) {
7190 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7192 } while_each_thread(g
, p
);
7197 struct rt_schedulable_data
{
7198 struct task_group
*tg
;
7203 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7205 struct rt_schedulable_data
*d
= data
;
7206 struct task_group
*child
;
7207 unsigned long total
, sum
= 0;
7208 u64 period
, runtime
;
7210 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7211 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7214 period
= d
->rt_period
;
7215 runtime
= d
->rt_runtime
;
7219 * Cannot have more runtime than the period.
7221 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7225 * Ensure we don't starve existing RT tasks.
7227 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7230 total
= to_ratio(period
, runtime
);
7233 * Nobody can have more than the global setting allows.
7235 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7239 * The sum of our children's runtime should not exceed our own.
7241 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7242 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7243 runtime
= child
->rt_bandwidth
.rt_runtime
;
7245 if (child
== d
->tg
) {
7246 period
= d
->rt_period
;
7247 runtime
= d
->rt_runtime
;
7250 sum
+= to_ratio(period
, runtime
);
7259 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7263 struct rt_schedulable_data data
= {
7265 .rt_period
= period
,
7266 .rt_runtime
= runtime
,
7270 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7276 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7277 u64 rt_period
, u64 rt_runtime
)
7281 mutex_lock(&rt_constraints_mutex
);
7282 read_lock(&tasklist_lock
);
7283 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7287 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7288 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7289 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7291 for_each_possible_cpu(i
) {
7292 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7294 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7295 rt_rq
->rt_runtime
= rt_runtime
;
7296 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7298 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7300 read_unlock(&tasklist_lock
);
7301 mutex_unlock(&rt_constraints_mutex
);
7306 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7308 u64 rt_runtime
, rt_period
;
7310 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7311 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7312 if (rt_runtime_us
< 0)
7313 rt_runtime
= RUNTIME_INF
;
7315 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7318 long sched_group_rt_runtime(struct task_group
*tg
)
7322 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7325 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7326 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7327 return rt_runtime_us
;
7330 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7332 u64 rt_runtime
, rt_period
;
7334 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7335 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7340 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7343 long sched_group_rt_period(struct task_group
*tg
)
7347 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7348 do_div(rt_period_us
, NSEC_PER_USEC
);
7349 return rt_period_us
;
7352 static int sched_rt_global_constraints(void)
7354 u64 runtime
, period
;
7357 if (sysctl_sched_rt_period
<= 0)
7360 runtime
= global_rt_runtime();
7361 period
= global_rt_period();
7364 * Sanity check on the sysctl variables.
7366 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7369 mutex_lock(&rt_constraints_mutex
);
7370 read_lock(&tasklist_lock
);
7371 ret
= __rt_schedulable(NULL
, 0, 0);
7372 read_unlock(&tasklist_lock
);
7373 mutex_unlock(&rt_constraints_mutex
);
7378 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7380 /* Don't accept realtime tasks when there is no way for them to run */
7381 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7387 #else /* !CONFIG_RT_GROUP_SCHED */
7388 static int sched_rt_global_constraints(void)
7390 unsigned long flags
;
7393 if (sysctl_sched_rt_period
<= 0)
7397 * There's always some RT tasks in the root group
7398 * -- migration, kstopmachine etc..
7400 if (sysctl_sched_rt_runtime
== 0)
7403 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7404 for_each_possible_cpu(i
) {
7405 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7407 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7408 rt_rq
->rt_runtime
= global_rt_runtime();
7409 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7411 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7415 #endif /* CONFIG_RT_GROUP_SCHED */
7417 int sched_rt_handler(struct ctl_table
*table
, int write
,
7418 void __user
*buffer
, size_t *lenp
,
7422 int old_period
, old_runtime
;
7423 static DEFINE_MUTEX(mutex
);
7426 old_period
= sysctl_sched_rt_period
;
7427 old_runtime
= sysctl_sched_rt_runtime
;
7429 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7431 if (!ret
&& write
) {
7432 ret
= sched_rt_global_constraints();
7434 sysctl_sched_rt_period
= old_period
;
7435 sysctl_sched_rt_runtime
= old_runtime
;
7437 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7438 def_rt_bandwidth
.rt_period
=
7439 ns_to_ktime(global_rt_period());
7442 mutex_unlock(&mutex
);
7447 #ifdef CONFIG_CGROUP_SCHED
7449 /* return corresponding task_group object of a cgroup */
7450 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7452 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7453 struct task_group
, css
);
7456 static struct cgroup_subsys_state
*cpu_cgroup_create(struct cgroup
*cgrp
)
7458 struct task_group
*tg
, *parent
;
7460 if (!cgrp
->parent
) {
7461 /* This is early initialization for the top cgroup */
7462 return &root_task_group
.css
;
7465 parent
= cgroup_tg(cgrp
->parent
);
7466 tg
= sched_create_group(parent
);
7468 return ERR_PTR(-ENOMEM
);
7473 static void cpu_cgroup_destroy(struct cgroup
*cgrp
)
7475 struct task_group
*tg
= cgroup_tg(cgrp
);
7477 sched_destroy_group(tg
);
7480 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7481 struct cgroup_taskset
*tset
)
7483 struct task_struct
*task
;
7485 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7486 #ifdef CONFIG_RT_GROUP_SCHED
7487 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7490 /* We don't support RT-tasks being in separate groups */
7491 if (task
->sched_class
!= &fair_sched_class
)
7498 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7499 struct cgroup_taskset
*tset
)
7501 struct task_struct
*task
;
7503 cgroup_taskset_for_each(task
, cgrp
, tset
)
7504 sched_move_task(task
);
7508 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7509 struct task_struct
*task
)
7512 * cgroup_exit() is called in the copy_process() failure path.
7513 * Ignore this case since the task hasn't ran yet, this avoids
7514 * trying to poke a half freed task state from generic code.
7516 if (!(task
->flags
& PF_EXITING
))
7519 sched_move_task(task
);
7522 #ifdef CONFIG_FAIR_GROUP_SCHED
7523 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7526 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7529 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7531 struct task_group
*tg
= cgroup_tg(cgrp
);
7533 return (u64
) scale_load_down(tg
->shares
);
7536 #ifdef CONFIG_CFS_BANDWIDTH
7537 static DEFINE_MUTEX(cfs_constraints_mutex
);
7539 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7540 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7542 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7544 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7546 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7547 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7549 if (tg
== &root_task_group
)
7553 * Ensure we have at some amount of bandwidth every period. This is
7554 * to prevent reaching a state of large arrears when throttled via
7555 * entity_tick() resulting in prolonged exit starvation.
7557 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7561 * Likewise, bound things on the otherside by preventing insane quota
7562 * periods. This also allows us to normalize in computing quota
7565 if (period
> max_cfs_quota_period
)
7568 mutex_lock(&cfs_constraints_mutex
);
7569 ret
= __cfs_schedulable(tg
, period
, quota
);
7573 runtime_enabled
= quota
!= RUNTIME_INF
;
7574 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7575 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7576 raw_spin_lock_irq(&cfs_b
->lock
);
7577 cfs_b
->period
= ns_to_ktime(period
);
7578 cfs_b
->quota
= quota
;
7580 __refill_cfs_bandwidth_runtime(cfs_b
);
7581 /* restart the period timer (if active) to handle new period expiry */
7582 if (runtime_enabled
&& cfs_b
->timer_active
) {
7583 /* force a reprogram */
7584 cfs_b
->timer_active
= 0;
7585 __start_cfs_bandwidth(cfs_b
);
7587 raw_spin_unlock_irq(&cfs_b
->lock
);
7589 for_each_possible_cpu(i
) {
7590 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7591 struct rq
*rq
= cfs_rq
->rq
;
7593 raw_spin_lock_irq(&rq
->lock
);
7594 cfs_rq
->runtime_enabled
= runtime_enabled
;
7595 cfs_rq
->runtime_remaining
= 0;
7597 if (cfs_rq
->throttled
)
7598 unthrottle_cfs_rq(cfs_rq
);
7599 raw_spin_unlock_irq(&rq
->lock
);
7602 mutex_unlock(&cfs_constraints_mutex
);
7607 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7611 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7612 if (cfs_quota_us
< 0)
7613 quota
= RUNTIME_INF
;
7615 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7617 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7620 long tg_get_cfs_quota(struct task_group
*tg
)
7624 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7627 quota_us
= tg
->cfs_bandwidth
.quota
;
7628 do_div(quota_us
, NSEC_PER_USEC
);
7633 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7637 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7638 quota
= tg
->cfs_bandwidth
.quota
;
7640 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7643 long tg_get_cfs_period(struct task_group
*tg
)
7647 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7648 do_div(cfs_period_us
, NSEC_PER_USEC
);
7650 return cfs_period_us
;
7653 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7655 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7658 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7661 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7664 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7666 return tg_get_cfs_period(cgroup_tg(cgrp
));
7669 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7672 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7675 struct cfs_schedulable_data
{
7676 struct task_group
*tg
;
7681 * normalize group quota/period to be quota/max_period
7682 * note: units are usecs
7684 static u64
normalize_cfs_quota(struct task_group
*tg
,
7685 struct cfs_schedulable_data
*d
)
7693 period
= tg_get_cfs_period(tg
);
7694 quota
= tg_get_cfs_quota(tg
);
7697 /* note: these should typically be equivalent */
7698 if (quota
== RUNTIME_INF
|| quota
== -1)
7701 return to_ratio(period
, quota
);
7704 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7706 struct cfs_schedulable_data
*d
= data
;
7707 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7708 s64 quota
= 0, parent_quota
= -1;
7711 quota
= RUNTIME_INF
;
7713 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7715 quota
= normalize_cfs_quota(tg
, d
);
7716 parent_quota
= parent_b
->hierarchal_quota
;
7719 * ensure max(child_quota) <= parent_quota, inherit when no
7722 if (quota
== RUNTIME_INF
)
7723 quota
= parent_quota
;
7724 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7727 cfs_b
->hierarchal_quota
= quota
;
7732 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7735 struct cfs_schedulable_data data
= {
7741 if (quota
!= RUNTIME_INF
) {
7742 do_div(data
.period
, NSEC_PER_USEC
);
7743 do_div(data
.quota
, NSEC_PER_USEC
);
7747 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7753 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7754 struct cgroup_map_cb
*cb
)
7756 struct task_group
*tg
= cgroup_tg(cgrp
);
7757 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7759 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7760 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7761 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7765 #endif /* CONFIG_CFS_BANDWIDTH */
7766 #endif /* CONFIG_FAIR_GROUP_SCHED */
7768 #ifdef CONFIG_RT_GROUP_SCHED
7769 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7772 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7775 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7777 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7780 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7783 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7786 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7788 return sched_group_rt_period(cgroup_tg(cgrp
));
7790 #endif /* CONFIG_RT_GROUP_SCHED */
7792 static struct cftype cpu_files
[] = {
7793 #ifdef CONFIG_FAIR_GROUP_SCHED
7796 .read_u64
= cpu_shares_read_u64
,
7797 .write_u64
= cpu_shares_write_u64
,
7800 #ifdef CONFIG_CFS_BANDWIDTH
7802 .name
= "cfs_quota_us",
7803 .read_s64
= cpu_cfs_quota_read_s64
,
7804 .write_s64
= cpu_cfs_quota_write_s64
,
7807 .name
= "cfs_period_us",
7808 .read_u64
= cpu_cfs_period_read_u64
,
7809 .write_u64
= cpu_cfs_period_write_u64
,
7813 .read_map
= cpu_stats_show
,
7816 #ifdef CONFIG_RT_GROUP_SCHED
7818 .name
= "rt_runtime_us",
7819 .read_s64
= cpu_rt_runtime_read
,
7820 .write_s64
= cpu_rt_runtime_write
,
7823 .name
= "rt_period_us",
7824 .read_u64
= cpu_rt_period_read_uint
,
7825 .write_u64
= cpu_rt_period_write_uint
,
7831 struct cgroup_subsys cpu_cgroup_subsys
= {
7833 .create
= cpu_cgroup_create
,
7834 .destroy
= cpu_cgroup_destroy
,
7835 .can_attach
= cpu_cgroup_can_attach
,
7836 .attach
= cpu_cgroup_attach
,
7837 .exit
= cpu_cgroup_exit
,
7838 .subsys_id
= cpu_cgroup_subsys_id
,
7839 .base_cftypes
= cpu_files
,
7843 #endif /* CONFIG_CGROUP_SCHED */
7845 #ifdef CONFIG_CGROUP_CPUACCT
7848 * CPU accounting code for task groups.
7850 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7851 * (balbir@in.ibm.com).
7854 struct cpuacct root_cpuacct
;
7856 /* create a new cpu accounting group */
7857 static struct cgroup_subsys_state
*cpuacct_create(struct cgroup
*cgrp
)
7862 return &root_cpuacct
.css
;
7864 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
7868 ca
->cpuusage
= alloc_percpu(u64
);
7872 ca
->cpustat
= alloc_percpu(struct kernel_cpustat
);
7874 goto out_free_cpuusage
;
7879 free_percpu(ca
->cpuusage
);
7883 return ERR_PTR(-ENOMEM
);
7886 /* destroy an existing cpu accounting group */
7887 static void cpuacct_destroy(struct cgroup
*cgrp
)
7889 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7891 free_percpu(ca
->cpustat
);
7892 free_percpu(ca
->cpuusage
);
7896 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
7898 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7901 #ifndef CONFIG_64BIT
7903 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7905 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7907 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7915 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
7917 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
7919 #ifndef CONFIG_64BIT
7921 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7923 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
7925 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
7931 /* return total cpu usage (in nanoseconds) of a group */
7932 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7934 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7935 u64 totalcpuusage
= 0;
7938 for_each_present_cpu(i
)
7939 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
7941 return totalcpuusage
;
7944 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
7947 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7956 for_each_present_cpu(i
)
7957 cpuacct_cpuusage_write(ca
, i
, 0);
7963 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
7966 struct cpuacct
*ca
= cgroup_ca(cgroup
);
7970 for_each_present_cpu(i
) {
7971 percpu
= cpuacct_cpuusage_read(ca
, i
);
7972 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
7974 seq_printf(m
, "\n");
7978 static const char *cpuacct_stat_desc
[] = {
7979 [CPUACCT_STAT_USER
] = "user",
7980 [CPUACCT_STAT_SYSTEM
] = "system",
7983 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7984 struct cgroup_map_cb
*cb
)
7986 struct cpuacct
*ca
= cgroup_ca(cgrp
);
7990 for_each_online_cpu(cpu
) {
7991 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
7992 val
+= kcpustat
->cpustat
[CPUTIME_USER
];
7993 val
+= kcpustat
->cpustat
[CPUTIME_NICE
];
7995 val
= cputime64_to_clock_t(val
);
7996 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_USER
], val
);
7999 for_each_online_cpu(cpu
) {
8000 struct kernel_cpustat
*kcpustat
= per_cpu_ptr(ca
->cpustat
, cpu
);
8001 val
+= kcpustat
->cpustat
[CPUTIME_SYSTEM
];
8002 val
+= kcpustat
->cpustat
[CPUTIME_IRQ
];
8003 val
+= kcpustat
->cpustat
[CPUTIME_SOFTIRQ
];
8006 val
= cputime64_to_clock_t(val
);
8007 cb
->fill(cb
, cpuacct_stat_desc
[CPUACCT_STAT_SYSTEM
], val
);
8012 static struct cftype files
[] = {
8015 .read_u64
= cpuusage_read
,
8016 .write_u64
= cpuusage_write
,
8019 .name
= "usage_percpu",
8020 .read_seq_string
= cpuacct_percpu_seq_read
,
8024 .read_map
= cpuacct_stats_show
,
8030 * charge this task's execution time to its accounting group.
8032 * called with rq->lock held.
8034 void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8039 if (unlikely(!cpuacct_subsys
.active
))
8042 cpu
= task_cpu(tsk
);
8048 for (; ca
; ca
= parent_ca(ca
)) {
8049 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8050 *cpuusage
+= cputime
;
8056 struct cgroup_subsys cpuacct_subsys
= {
8058 .create
= cpuacct_create
,
8059 .destroy
= cpuacct_destroy
,
8060 .subsys_id
= cpuacct_subsys_id
,
8061 .base_cftypes
= files
,
8063 #endif /* CONFIG_CGROUP_CPUACCT */