4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
95 ktime_t soft
, hard
, now
;
98 if (hrtimer_active(period_timer
))
101 now
= hrtimer_cb_get_time(period_timer
);
102 hrtimer_forward(period_timer
, now
, period
);
104 soft
= hrtimer_get_softexpires(period_timer
);
105 hard
= hrtimer_get_expires(period_timer
);
106 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
107 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
108 HRTIMER_MODE_ABS_PINNED
, 0);
112 DEFINE_MUTEX(sched_domains_mutex
);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
115 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
117 void update_rq_clock(struct rq
*rq
)
121 if (rq
->skip_clock_update
> 0)
124 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
126 update_rq_clock_task(rq
, delta
);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug
unsigned int sysctl_sched_features
=
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names
[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp
)
201 if (strncmp(cmp
, "NO_", 3) == 0) {
206 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
207 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
209 sysctl_sched_features
&= ~(1UL << i
);
210 sched_feat_disable(i
);
212 sysctl_sched_features
|= (1UL << i
);
213 sched_feat_enable(i
);
223 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
224 size_t cnt
, loff_t
*ppos
)
233 if (copy_from_user(&buf
, ubuf
, cnt
))
239 i
= sched_feat_set(cmp
);
240 if (i
== __SCHED_FEAT_NR
)
248 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
250 return single_open(filp
, sched_feat_show
, NULL
);
253 static const struct file_operations sched_feat_fops
= {
254 .open
= sched_feat_open
,
255 .write
= sched_feat_write
,
258 .release
= single_release
,
261 static __init
int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
268 late_initcall(sched_init_debug
);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
278 * period over which we average the RT time consumption, measured
283 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period
= 1000000;
291 __read_mostly
int scheduler_running
;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime
= 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
309 lockdep_assert_held(&p
->pi_lock
);
313 raw_spin_lock(&rq
->lock
);
314 if (likely(rq
== task_rq(p
)))
316 raw_spin_unlock(&rq
->lock
);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
324 __acquires(p
->pi_lock
)
330 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
332 raw_spin_lock(&rq
->lock
);
333 if (likely(rq
== task_rq(p
)))
335 raw_spin_unlock(&rq
->lock
);
336 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
340 static void __task_rq_unlock(struct rq
*rq
)
343 raw_spin_unlock(&rq
->lock
);
347 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
349 __releases(p
->pi_lock
)
351 raw_spin_unlock(&rq
->lock
);
352 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq
*this_rq_lock(void)
365 raw_spin_lock(&rq
->lock
);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq
*rq
)
384 if (hrtimer_active(&rq
->hrtick_timer
))
385 hrtimer_cancel(&rq
->hrtick_timer
);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
394 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
396 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
398 raw_spin_lock(&rq
->lock
);
400 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
401 raw_spin_unlock(&rq
->lock
);
403 return HRTIMER_NORESTART
;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg
)
414 raw_spin_lock(&rq
->lock
);
415 hrtimer_restart(&rq
->hrtick_timer
);
416 rq
->hrtick_csd_pending
= 0;
417 raw_spin_unlock(&rq
->lock
);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq
*rq
, u64 delay
)
427 struct hrtimer
*timer
= &rq
->hrtick_timer
;
428 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
430 hrtimer_set_expires(timer
, time
);
432 if (rq
== this_rq()) {
433 hrtimer_restart(timer
);
434 } else if (!rq
->hrtick_csd_pending
) {
435 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
436 rq
->hrtick_csd_pending
= 1;
441 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
443 int cpu
= (int)(long)hcpu
;
446 case CPU_UP_CANCELED
:
447 case CPU_UP_CANCELED_FROZEN
:
448 case CPU_DOWN_PREPARE
:
449 case CPU_DOWN_PREPARE_FROZEN
:
451 case CPU_DEAD_FROZEN
:
452 hrtick_clear(cpu_rq(cpu
));
459 static __init
void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick
, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq
*rq
, u64 delay
)
471 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
472 HRTIMER_MODE_REL_PINNED
, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq
*rq
)
483 rq
->hrtick_csd_pending
= 0;
485 rq
->hrtick_csd
.flags
= 0;
486 rq
->hrtick_csd
.func
= __hrtick_start
;
487 rq
->hrtick_csd
.info
= rq
;
490 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
491 rq
->hrtick_timer
.function
= hrtick
;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq
*rq
)
498 static inline void init_rq_hrtick(struct rq
*rq
)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct
*p
)
519 assert_raw_spin_locked(&task_rq(p
)->lock
);
521 if (test_tsk_need_resched(p
))
524 set_tsk_need_resched(p
);
527 if (cpu
== smp_processor_id())
530 /* NEED_RESCHED must be visible before we test polling */
532 if (!tsk_is_polling(p
))
533 smp_send_reschedule(cpu
);
536 void resched_cpu(int cpu
)
538 struct rq
*rq
= cpu_rq(cpu
);
541 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
543 resched_task(cpu_curr(cpu
));
544 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int cpu
= smp_processor_id();
560 struct sched_domain
*sd
;
563 for_each_domain(cpu
, sd
) {
564 for_each_cpu(i
, sched_domain_span(sd
)) {
576 * When add_timer_on() enqueues a timer into the timer wheel of an
577 * idle CPU then this timer might expire before the next timer event
578 * which is scheduled to wake up that CPU. In case of a completely
579 * idle system the next event might even be infinite time into the
580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
581 * leaves the inner idle loop so the newly added timer is taken into
582 * account when the CPU goes back to idle and evaluates the timer
583 * wheel for the next timer event.
585 static void wake_up_idle_cpu(int cpu
)
587 struct rq
*rq
= cpu_rq(cpu
);
589 if (cpu
== smp_processor_id())
593 * This is safe, as this function is called with the timer
594 * wheel base lock of (cpu) held. When the CPU is on the way
595 * to idle and has not yet set rq->curr to idle then it will
596 * be serialized on the timer wheel base lock and take the new
597 * timer into account automatically.
599 if (rq
->curr
!= rq
->idle
)
603 * We can set TIF_RESCHED on the idle task of the other CPU
604 * lockless. The worst case is that the other CPU runs the
605 * idle task through an additional NOOP schedule()
607 set_tsk_need_resched(rq
->idle
);
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(rq
->idle
))
612 smp_send_reschedule(cpu
);
615 static bool wake_up_full_nohz_cpu(int cpu
)
617 if (tick_nohz_full_cpu(cpu
)) {
618 if (cpu
!= smp_processor_id() ||
619 tick_nohz_tick_stopped())
620 smp_send_reschedule(cpu
);
627 void wake_up_nohz_cpu(int cpu
)
629 if (!wake_up_full_nohz_cpu(cpu
))
630 wake_up_idle_cpu(cpu
);
633 static inline bool got_nohz_idle_kick(void)
635 int cpu
= smp_processor_id();
637 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
640 if (idle_cpu(cpu
) && !need_resched())
644 * We can't run Idle Load Balance on this CPU for this time so we
645 * cancel it and clear NOHZ_BALANCE_KICK
647 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
651 #else /* CONFIG_NO_HZ_COMMON */
653 static inline bool got_nohz_idle_kick(void)
658 #endif /* CONFIG_NO_HZ_COMMON */
660 #ifdef CONFIG_NO_HZ_FULL
661 bool sched_can_stop_tick(void)
667 /* Make sure rq->nr_running update is visible after the IPI */
670 /* More than one running task need preemption */
671 if (rq
->nr_running
> 1)
676 #endif /* CONFIG_NO_HZ_FULL */
678 void sched_avg_update(struct rq
*rq
)
680 s64 period
= sched_avg_period();
682 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
684 * Inline assembly required to prevent the compiler
685 * optimising this loop into a divmod call.
686 * See __iter_div_u64_rem() for another example of this.
688 asm("" : "+rm" (rq
->age_stamp
));
689 rq
->age_stamp
+= period
;
694 #else /* !CONFIG_SMP */
695 void resched_task(struct task_struct
*p
)
697 assert_raw_spin_locked(&task_rq(p
)->lock
);
698 set_tsk_need_resched(p
);
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group
*from
,
711 tg_visitor down
, tg_visitor up
, void *data
)
713 struct task_group
*parent
, *child
;
719 ret
= (*down
)(parent
, data
);
722 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
729 ret
= (*up
)(parent
, data
);
730 if (ret
|| parent
== from
)
734 parent
= parent
->parent
;
741 int tg_nop(struct task_group
*tg
, void *data
)
747 static void set_load_weight(struct task_struct
*p
)
749 int prio
= p
->static_prio
- MAX_RT_PRIO
;
750 struct load_weight
*load
= &p
->se
.load
;
753 * SCHED_IDLE tasks get minimal weight:
755 if (p
->policy
== SCHED_IDLE
) {
756 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
757 load
->inv_weight
= WMULT_IDLEPRIO
;
761 load
->weight
= scale_load(prio_to_weight
[prio
]);
762 load
->inv_weight
= prio_to_wmult
[prio
];
765 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
768 sched_info_queued(p
);
769 p
->sched_class
->enqueue_task(rq
, p
, flags
);
772 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
775 sched_info_dequeued(p
);
776 p
->sched_class
->dequeue_task(rq
, p
, flags
);
779 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
781 if (task_contributes_to_load(p
))
782 rq
->nr_uninterruptible
--;
784 enqueue_task(rq
, p
, flags
);
787 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
789 if (task_contributes_to_load(p
))
790 rq
->nr_uninterruptible
++;
792 dequeue_task(rq
, p
, flags
);
795 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal
= 0, irq_delta
= 0;
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
822 if (irq_delta
> delta
)
825 rq
->prev_irq_time
+= irq_delta
;
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled
))) {
832 steal
= paravirt_steal_clock(cpu_of(rq
));
833 steal
-= rq
->prev_steal_time_rq
;
835 if (unlikely(steal
> delta
))
838 st
= steal_ticks(steal
);
839 steal
= st
* TICK_NSEC
;
841 rq
->prev_steal_time_rq
+= steal
;
847 rq
->clock_task
+= delta
;
849 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
850 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
851 sched_rt_avg_update(rq
, irq_delta
+ steal
);
855 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
857 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
858 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
862 * Make it appear like a SCHED_FIFO task, its something
863 * userspace knows about and won't get confused about.
865 * Also, it will make PI more or less work without too
866 * much confusion -- but then, stop work should not
867 * rely on PI working anyway.
869 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
871 stop
->sched_class
= &stop_sched_class
;
874 cpu_rq(cpu
)->stop
= stop
;
878 * Reset it back to a normal scheduling class so that
879 * it can die in pieces.
881 old_stop
->sched_class
= &rt_sched_class
;
886 * __normal_prio - return the priority that is based on the static prio
888 static inline int __normal_prio(struct task_struct
*p
)
890 return p
->static_prio
;
894 * Calculate the expected normal priority: i.e. priority
895 * without taking RT-inheritance into account. Might be
896 * boosted by interactivity modifiers. Changes upon fork,
897 * setprio syscalls, and whenever the interactivity
898 * estimator recalculates.
900 static inline int normal_prio(struct task_struct
*p
)
904 if (task_has_rt_policy(p
))
905 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
907 prio
= __normal_prio(p
);
912 * Calculate the current priority, i.e. the priority
913 * taken into account by the scheduler. This value might
914 * be boosted by RT tasks, or might be boosted by
915 * interactivity modifiers. Will be RT if the task got
916 * RT-boosted. If not then it returns p->normal_prio.
918 static int effective_prio(struct task_struct
*p
)
920 p
->normal_prio
= normal_prio(p
);
922 * If we are RT tasks or we were boosted to RT priority,
923 * keep the priority unchanged. Otherwise, update priority
924 * to the normal priority:
926 if (!rt_prio(p
->prio
))
927 return p
->normal_prio
;
932 * task_curr - is this task currently executing on a CPU?
933 * @p: the task in question.
935 inline int task_curr(const struct task_struct
*p
)
937 return cpu_curr(task_cpu(p
)) == p
;
940 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
941 const struct sched_class
*prev_class
,
944 if (prev_class
!= p
->sched_class
) {
945 if (prev_class
->switched_from
)
946 prev_class
->switched_from(rq
, p
);
947 p
->sched_class
->switched_to(rq
, p
);
948 } else if (oldprio
!= p
->prio
)
949 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
952 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
954 const struct sched_class
*class;
956 if (p
->sched_class
== rq
->curr
->sched_class
) {
957 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
959 for_each_class(class) {
960 if (class == rq
->curr
->sched_class
)
962 if (class == p
->sched_class
) {
963 resched_task(rq
->curr
);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
974 rq
->skip_clock_update
= 1;
977 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
979 void register_task_migration_notifier(struct notifier_block
*n
)
981 atomic_notifier_chain_register(&task_migration_notifier
, n
);
985 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
987 #ifdef CONFIG_SCHED_DEBUG
989 * We should never call set_task_cpu() on a blocked task,
990 * ttwu() will sort out the placement.
992 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
993 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
995 #ifdef CONFIG_LOCKDEP
997 * The caller should hold either p->pi_lock or rq->lock, when changing
998 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1000 * sched_move_task() holds both and thus holding either pins the cgroup,
1003 * Furthermore, all task_rq users should acquire both locks, see
1006 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1007 lockdep_is_held(&task_rq(p
)->lock
)));
1011 trace_sched_migrate_task(p
, new_cpu
);
1013 if (task_cpu(p
) != new_cpu
) {
1014 struct task_migration_notifier tmn
;
1016 if (p
->sched_class
->migrate_task_rq
)
1017 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1018 p
->se
.nr_migrations
++;
1019 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1022 tmn
.from_cpu
= task_cpu(p
);
1023 tmn
.to_cpu
= new_cpu
;
1025 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
1028 __set_task_cpu(p
, new_cpu
);
1031 struct migration_arg
{
1032 struct task_struct
*task
;
1036 static int migration_cpu_stop(void *data
);
1039 * wait_task_inactive - wait for a thread to unschedule.
1041 * If @match_state is nonzero, it's the @p->state value just checked and
1042 * not expected to change. If it changes, i.e. @p might have woken up,
1043 * then return zero. When we succeed in waiting for @p to be off its CPU,
1044 * we return a positive number (its total switch count). If a second call
1045 * a short while later returns the same number, the caller can be sure that
1046 * @p has remained unscheduled the whole time.
1048 * The caller must ensure that the task *will* unschedule sometime soon,
1049 * else this function might spin for a *long* time. This function can't
1050 * be called with interrupts off, or it may introduce deadlock with
1051 * smp_call_function() if an IPI is sent by the same process we are
1052 * waiting to become inactive.
1054 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1056 unsigned long flags
;
1063 * We do the initial early heuristics without holding
1064 * any task-queue locks at all. We'll only try to get
1065 * the runqueue lock when things look like they will
1071 * If the task is actively running on another CPU
1072 * still, just relax and busy-wait without holding
1075 * NOTE! Since we don't hold any locks, it's not
1076 * even sure that "rq" stays as the right runqueue!
1077 * But we don't care, since "task_running()" will
1078 * return false if the runqueue has changed and p
1079 * is actually now running somewhere else!
1081 while (task_running(rq
, p
)) {
1082 if (match_state
&& unlikely(p
->state
!= match_state
))
1088 * Ok, time to look more closely! We need the rq
1089 * lock now, to be *sure*. If we're wrong, we'll
1090 * just go back and repeat.
1092 rq
= task_rq_lock(p
, &flags
);
1093 trace_sched_wait_task(p
);
1094 running
= task_running(rq
, p
);
1097 if (!match_state
|| p
->state
== match_state
)
1098 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1099 task_rq_unlock(rq
, p
, &flags
);
1102 * If it changed from the expected state, bail out now.
1104 if (unlikely(!ncsw
))
1108 * Was it really running after all now that we
1109 * checked with the proper locks actually held?
1111 * Oops. Go back and try again..
1113 if (unlikely(running
)) {
1119 * It's not enough that it's not actively running,
1120 * it must be off the runqueue _entirely_, and not
1123 * So if it was still runnable (but just not actively
1124 * running right now), it's preempted, and we should
1125 * yield - it could be a while.
1127 if (unlikely(on_rq
)) {
1128 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1130 set_current_state(TASK_UNINTERRUPTIBLE
);
1131 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1136 * Ahh, all good. It wasn't running, and it wasn't
1137 * runnable, which means that it will never become
1138 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesn't have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct
*p
)
1165 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1166 smp_send_reschedule(cpu
);
1169 EXPORT_SYMBOL_GPL(kick_process
);
1170 #endif /* CONFIG_SMP */
1174 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1176 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1178 int nid
= cpu_to_node(cpu
);
1179 const struct cpumask
*nodemask
= NULL
;
1180 enum { cpuset
, possible
, fail
} state
= cpuset
;
1184 * If the node that the cpu is on has been offlined, cpu_to_node()
1185 * will return -1. There is no cpu on the node, and we should
1186 * select the cpu on the other node.
1189 nodemask
= cpumask_of_node(nid
);
1191 /* Look for allowed, online CPU in same node. */
1192 for_each_cpu(dest_cpu
, nodemask
) {
1193 if (!cpu_online(dest_cpu
))
1195 if (!cpu_active(dest_cpu
))
1197 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1203 /* Any allowed, online CPU? */
1204 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1205 if (!cpu_online(dest_cpu
))
1207 if (!cpu_active(dest_cpu
))
1214 /* No more Mr. Nice Guy. */
1215 cpuset_cpus_allowed_fallback(p
);
1220 do_set_cpus_allowed(p
, cpu_possible_mask
);
1231 if (state
!= cpuset
) {
1233 * Don't tell them about moving exiting tasks or
1234 * kernel threads (both mm NULL), since they never
1237 if (p
->mm
&& printk_ratelimit()) {
1238 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1239 task_pid_nr(p
), p
->comm
, cpu
);
1247 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1250 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1252 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1255 * In order not to call set_task_cpu() on a blocking task we need
1256 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1259 * Since this is common to all placement strategies, this lives here.
1261 * [ this allows ->select_task() to simply return task_cpu(p) and
1262 * not worry about this generic constraint ]
1264 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1266 cpu
= select_fallback_rq(task_cpu(p
), p
);
1271 static void update_avg(u64
*avg
, u64 sample
)
1273 s64 diff
= sample
- *avg
;
1279 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1281 #ifdef CONFIG_SCHEDSTATS
1282 struct rq
*rq
= this_rq();
1285 int this_cpu
= smp_processor_id();
1287 if (cpu
== this_cpu
) {
1288 schedstat_inc(rq
, ttwu_local
);
1289 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1291 struct sched_domain
*sd
;
1293 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1295 for_each_domain(this_cpu
, sd
) {
1296 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1297 schedstat_inc(sd
, ttwu_wake_remote
);
1304 if (wake_flags
& WF_MIGRATED
)
1305 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1307 #endif /* CONFIG_SMP */
1309 schedstat_inc(rq
, ttwu_count
);
1310 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1312 if (wake_flags
& WF_SYNC
)
1313 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1315 #endif /* CONFIG_SCHEDSTATS */
1318 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1320 activate_task(rq
, p
, en_flags
);
1323 /* if a worker is waking up, notify workqueue */
1324 if (p
->flags
& PF_WQ_WORKER
)
1325 wq_worker_waking_up(p
, cpu_of(rq
));
1329 * Mark the task runnable and perform wakeup-preemption.
1332 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1334 check_preempt_curr(rq
, p
, wake_flags
);
1335 trace_sched_wakeup(p
, true);
1337 p
->state
= TASK_RUNNING
;
1339 if (p
->sched_class
->task_woken
)
1340 p
->sched_class
->task_woken(rq
, p
);
1342 if (rq
->idle_stamp
) {
1343 u64 delta
= rq
->clock
- rq
->idle_stamp
;
1344 u64 max
= 2*sysctl_sched_migration_cost
;
1349 update_avg(&rq
->avg_idle
, delta
);
1356 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1359 if (p
->sched_contributes_to_load
)
1360 rq
->nr_uninterruptible
--;
1363 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1364 ttwu_do_wakeup(rq
, p
, wake_flags
);
1368 * Called in case the task @p isn't fully descheduled from its runqueue,
1369 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1370 * since all we need to do is flip p->state to TASK_RUNNING, since
1371 * the task is still ->on_rq.
1373 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1378 rq
= __task_rq_lock(p
);
1380 ttwu_do_wakeup(rq
, p
, wake_flags
);
1383 __task_rq_unlock(rq
);
1389 static void sched_ttwu_pending(void)
1391 struct rq
*rq
= this_rq();
1392 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1393 struct task_struct
*p
;
1395 raw_spin_lock(&rq
->lock
);
1398 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1399 llist
= llist_next(llist
);
1400 ttwu_do_activate(rq
, p
, 0);
1403 raw_spin_unlock(&rq
->lock
);
1406 void scheduler_ipi(void)
1408 if (llist_empty(&this_rq()->wake_list
)
1409 && !tick_nohz_full_cpu(smp_processor_id())
1410 && !got_nohz_idle_kick())
1414 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1415 * traditionally all their work was done from the interrupt return
1416 * path. Now that we actually do some work, we need to make sure
1419 * Some archs already do call them, luckily irq_enter/exit nest
1422 * Arguably we should visit all archs and update all handlers,
1423 * however a fair share of IPIs are still resched only so this would
1424 * somewhat pessimize the simple resched case.
1427 tick_nohz_full_check();
1428 sched_ttwu_pending();
1431 * Check if someone kicked us for doing the nohz idle load balance.
1433 if (unlikely(got_nohz_idle_kick())) {
1434 this_rq()->idle_balance
= 1;
1435 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1440 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1442 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1443 smp_send_reschedule(cpu
);
1446 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1448 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1450 #endif /* CONFIG_SMP */
1452 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1454 struct rq
*rq
= cpu_rq(cpu
);
1456 #if defined(CONFIG_SMP)
1457 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1458 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1459 ttwu_queue_remote(p
, cpu
);
1464 raw_spin_lock(&rq
->lock
);
1465 ttwu_do_activate(rq
, p
, 0);
1466 raw_spin_unlock(&rq
->lock
);
1470 * try_to_wake_up - wake up a thread
1471 * @p: the thread to be awakened
1472 * @state: the mask of task states that can be woken
1473 * @wake_flags: wake modifier flags (WF_*)
1475 * Put it on the run-queue if it's not already there. The "current"
1476 * thread is always on the run-queue (except when the actual
1477 * re-schedule is in progress), and as such you're allowed to do
1478 * the simpler "current->state = TASK_RUNNING" to mark yourself
1479 * runnable without the overhead of this.
1481 * Returns %true if @p was woken up, %false if it was already running
1482 * or @state didn't match @p's state.
1485 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1487 unsigned long flags
;
1488 int cpu
, success
= 0;
1491 * If we are going to wake up a thread waiting for CONDITION we
1492 * need to ensure that CONDITION=1 done by the caller can not be
1493 * reordered with p->state check below. This pairs with mb() in
1494 * set_current_state() the waiting thread does.
1496 smp_mb__before_spinlock();
1497 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1498 if (!(p
->state
& state
))
1501 success
= 1; /* we're going to change ->state */
1505 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1506 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1507 * in smp_cond_load_acquire() below.
1509 * sched_ttwu_pending() try_to_wake_up()
1510 * [S] p->on_rq = 1; [L] P->state
1511 * UNLOCK rq->lock -----.
1515 * LOCK rq->lock -----'
1519 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1521 * Pairs with the UNLOCK+LOCK on rq->lock from the
1522 * last wakeup of our task and the schedule that got our task
1526 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1531 * If the owning (remote) cpu is still in the middle of schedule() with
1532 * this task as prev, wait until its done referencing the task.
1537 * Pairs with the smp_wmb() in finish_lock_switch().
1541 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1542 p
->state
= TASK_WAKING
;
1544 if (p
->sched_class
->task_waking
)
1545 p
->sched_class
->task_waking(p
);
1547 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1548 if (task_cpu(p
) != cpu
) {
1549 wake_flags
|= WF_MIGRATED
;
1550 set_task_cpu(p
, cpu
);
1552 #endif /* CONFIG_SMP */
1556 ttwu_stat(p
, cpu
, wake_flags
);
1558 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1564 * try_to_wake_up_local - try to wake up a local task with rq lock held
1565 * @p: the thread to be awakened
1567 * Put @p on the run-queue if it's not already there. The caller must
1568 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1571 static void try_to_wake_up_local(struct task_struct
*p
)
1573 struct rq
*rq
= task_rq(p
);
1575 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1576 WARN_ON_ONCE(p
== current
))
1579 lockdep_assert_held(&rq
->lock
);
1581 if (!raw_spin_trylock(&p
->pi_lock
)) {
1582 raw_spin_unlock(&rq
->lock
);
1583 raw_spin_lock(&p
->pi_lock
);
1584 raw_spin_lock(&rq
->lock
);
1587 if (!(p
->state
& TASK_NORMAL
))
1591 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1593 ttwu_do_wakeup(rq
, p
, 0);
1594 ttwu_stat(p
, smp_processor_id(), 0);
1596 raw_spin_unlock(&p
->pi_lock
);
1600 * wake_up_process - Wake up a specific process
1601 * @p: The process to be woken up.
1603 * Attempt to wake up the nominated process and move it to the set of runnable
1604 * processes. Returns 1 if the process was woken up, 0 if it was already
1607 * It may be assumed that this function implies a write memory barrier before
1608 * changing the task state if and only if any tasks are woken up.
1610 int wake_up_process(struct task_struct
*p
)
1612 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1614 EXPORT_SYMBOL(wake_up_process
);
1616 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1618 return try_to_wake_up(p
, state
, 0);
1622 * Perform scheduler related setup for a newly forked process p.
1623 * p is forked by current.
1625 * __sched_fork() is basic setup used by init_idle() too:
1627 static void __sched_fork(struct task_struct
*p
)
1632 p
->se
.exec_start
= 0;
1633 p
->se
.sum_exec_runtime
= 0;
1634 p
->se
.prev_sum_exec_runtime
= 0;
1635 p
->se
.nr_migrations
= 0;
1637 INIT_LIST_HEAD(&p
->se
.group_node
);
1640 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1641 * removed when useful for applications beyond shares distribution (e.g.
1644 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1645 p
->se
.avg
.runnable_avg_period
= 0;
1646 p
->se
.avg
.runnable_avg_sum
= 0;
1648 #ifdef CONFIG_SCHEDSTATS
1649 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1652 INIT_LIST_HEAD(&p
->rt
.run_list
);
1654 #ifdef CONFIG_PREEMPT_NOTIFIERS
1655 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1658 #ifdef CONFIG_NUMA_BALANCING
1659 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1660 p
->mm
->numa_next_scan
= jiffies
;
1661 p
->mm
->numa_next_reset
= jiffies
;
1662 p
->mm
->numa_scan_seq
= 0;
1665 p
->node_stamp
= 0ULL;
1666 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1667 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1668 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1669 p
->numa_work
.next
= &p
->numa_work
;
1670 #endif /* CONFIG_NUMA_BALANCING */
1673 #ifdef CONFIG_NUMA_BALANCING
1674 #ifdef CONFIG_SCHED_DEBUG
1675 void set_numabalancing_state(bool enabled
)
1678 sched_feat_set("NUMA");
1680 sched_feat_set("NO_NUMA");
1683 __read_mostly
bool numabalancing_enabled
;
1685 void set_numabalancing_state(bool enabled
)
1687 numabalancing_enabled
= enabled
;
1689 #endif /* CONFIG_SCHED_DEBUG */
1690 #endif /* CONFIG_NUMA_BALANCING */
1693 * fork()/clone()-time setup:
1695 void sched_fork(struct task_struct
*p
)
1697 unsigned long flags
;
1698 int cpu
= get_cpu();
1702 * We mark the process as running here. This guarantees that
1703 * nobody will actually run it, and a signal or other external
1704 * event cannot wake it up and insert it on the runqueue either.
1706 p
->state
= TASK_RUNNING
;
1709 * Make sure we do not leak PI boosting priority to the child.
1711 p
->prio
= current
->normal_prio
;
1714 * Revert to default priority/policy on fork if requested.
1716 if (unlikely(p
->sched_reset_on_fork
)) {
1717 if (task_has_rt_policy(p
)) {
1718 p
->policy
= SCHED_NORMAL
;
1719 p
->static_prio
= NICE_TO_PRIO(0);
1721 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1722 p
->static_prio
= NICE_TO_PRIO(0);
1724 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1728 * We don't need the reset flag anymore after the fork. It has
1729 * fulfilled its duty:
1731 p
->sched_reset_on_fork
= 0;
1734 if (!rt_prio(p
->prio
))
1735 p
->sched_class
= &fair_sched_class
;
1737 if (p
->sched_class
->task_fork
)
1738 p
->sched_class
->task_fork(p
);
1741 * The child is not yet in the pid-hash so no cgroup attach races,
1742 * and the cgroup is pinned to this child due to cgroup_fork()
1743 * is ran before sched_fork().
1745 * Silence PROVE_RCU.
1747 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1748 set_task_cpu(p
, cpu
);
1749 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1751 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1752 if (likely(sched_info_on()))
1753 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1755 #if defined(CONFIG_SMP)
1758 #ifdef CONFIG_PREEMPT_COUNT
1759 /* Want to start with kernel preemption disabled. */
1760 task_thread_info(p
)->preempt_count
= 1;
1763 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1770 * wake_up_new_task - wake up a newly created task for the first time.
1772 * This function will do some initial scheduler statistics housekeeping
1773 * that must be done for every newly created context, then puts the task
1774 * on the runqueue and wakes it.
1776 void wake_up_new_task(struct task_struct
*p
)
1778 unsigned long flags
;
1781 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1784 * Fork balancing, do it here and not earlier because:
1785 * - cpus_allowed can change in the fork path
1786 * - any previously selected cpu might disappear through hotplug
1788 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1791 rq
= __task_rq_lock(p
);
1792 activate_task(rq
, p
, 0);
1794 trace_sched_wakeup_new(p
, true);
1795 check_preempt_curr(rq
, p
, WF_FORK
);
1797 if (p
->sched_class
->task_woken
)
1798 p
->sched_class
->task_woken(rq
, p
);
1800 task_rq_unlock(rq
, p
, &flags
);
1803 #ifdef CONFIG_PREEMPT_NOTIFIERS
1806 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1807 * @notifier: notifier struct to register
1809 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1811 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1813 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1816 * preempt_notifier_unregister - no longer interested in preemption notifications
1817 * @notifier: notifier struct to unregister
1819 * This is safe to call from within a preemption notifier.
1821 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1823 hlist_del(¬ifier
->link
);
1825 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1827 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1829 struct preempt_notifier
*notifier
;
1831 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1832 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1836 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1837 struct task_struct
*next
)
1839 struct preempt_notifier
*notifier
;
1841 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1842 notifier
->ops
->sched_out(notifier
, next
);
1845 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1847 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1852 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1853 struct task_struct
*next
)
1857 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1860 * prepare_task_switch - prepare to switch tasks
1861 * @rq: the runqueue preparing to switch
1862 * @prev: the current task that is being switched out
1863 * @next: the task we are going to switch to.
1865 * This is called with the rq lock held and interrupts off. It must
1866 * be paired with a subsequent finish_task_switch after the context
1869 * prepare_task_switch sets up locking and calls architecture specific
1873 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1874 struct task_struct
*next
)
1876 trace_sched_switch(prev
, next
);
1877 sched_info_switch(prev
, next
);
1878 perf_event_task_sched_out(prev
, next
);
1879 fire_sched_out_preempt_notifiers(prev
, next
);
1880 prepare_lock_switch(rq
, next
);
1881 prepare_arch_switch(next
);
1885 * finish_task_switch - clean up after a task-switch
1886 * @rq: runqueue associated with task-switch
1887 * @prev: the thread we just switched away from.
1889 * finish_task_switch must be called after the context switch, paired
1890 * with a prepare_task_switch call before the context switch.
1891 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1892 * and do any other architecture-specific cleanup actions.
1894 * Note that we may have delayed dropping an mm in context_switch(). If
1895 * so, we finish that here outside of the runqueue lock. (Doing it
1896 * with the lock held can cause deadlocks; see schedule() for
1899 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1900 __releases(rq
->lock
)
1902 struct mm_struct
*mm
= rq
->prev_mm
;
1908 * A task struct has one reference for the use as "current".
1909 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1910 * schedule one last time. The schedule call will never return, and
1911 * the scheduled task must drop that reference.
1912 * The test for TASK_DEAD must occur while the runqueue locks are
1913 * still held, otherwise prev could be scheduled on another cpu, die
1914 * there before we look at prev->state, and then the reference would
1916 * Manfred Spraul <manfred@colorfullife.com>
1918 prev_state
= prev
->state
;
1919 vtime_task_switch(prev
);
1920 finish_arch_switch(prev
);
1921 perf_event_task_sched_in(prev
, current
);
1922 finish_lock_switch(rq
, prev
);
1923 finish_arch_post_lock_switch();
1925 fire_sched_in_preempt_notifiers(current
);
1928 if (unlikely(prev_state
== TASK_DEAD
)) {
1930 * Remove function-return probe instances associated with this
1931 * task and put them back on the free list.
1933 kprobe_flush_task(prev
);
1934 put_task_struct(prev
);
1937 tick_nohz_task_switch(current
);
1942 /* assumes rq->lock is held */
1943 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1945 if (prev
->sched_class
->pre_schedule
)
1946 prev
->sched_class
->pre_schedule(rq
, prev
);
1949 /* rq->lock is NOT held, but preemption is disabled */
1950 static inline void post_schedule(struct rq
*rq
)
1952 if (rq
->post_schedule
) {
1953 unsigned long flags
;
1955 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1956 if (rq
->curr
->sched_class
->post_schedule
)
1957 rq
->curr
->sched_class
->post_schedule(rq
);
1958 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1960 rq
->post_schedule
= 0;
1966 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1970 static inline void post_schedule(struct rq
*rq
)
1977 * schedule_tail - first thing a freshly forked thread must call.
1978 * @prev: the thread we just switched away from.
1980 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1981 __releases(rq
->lock
)
1983 struct rq
*rq
= this_rq();
1985 finish_task_switch(rq
, prev
);
1988 * FIXME: do we need to worry about rq being invalidated by the
1993 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1994 /* In this case, finish_task_switch does not reenable preemption */
1997 if (current
->set_child_tid
)
1998 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2002 * context_switch - switch to the new MM and the new
2003 * thread's register state.
2006 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2007 struct task_struct
*next
)
2009 struct mm_struct
*mm
, *oldmm
;
2011 prepare_task_switch(rq
, prev
, next
);
2014 oldmm
= prev
->active_mm
;
2016 * For paravirt, this is coupled with an exit in switch_to to
2017 * combine the page table reload and the switch backend into
2020 arch_start_context_switch(prev
);
2023 next
->active_mm
= oldmm
;
2024 atomic_inc(&oldmm
->mm_count
);
2025 enter_lazy_tlb(oldmm
, next
);
2027 switch_mm(oldmm
, mm
, next
);
2030 prev
->active_mm
= NULL
;
2031 rq
->prev_mm
= oldmm
;
2034 * Since the runqueue lock will be released by the next
2035 * task (which is an invalid locking op but in the case
2036 * of the scheduler it's an obvious special-case), so we
2037 * do an early lockdep release here:
2039 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2040 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2043 context_tracking_task_switch(prev
, next
);
2044 /* Here we just switch the register state and the stack. */
2045 switch_to(prev
, next
, prev
);
2049 * this_rq must be evaluated again because prev may have moved
2050 * CPUs since it called schedule(), thus the 'rq' on its stack
2051 * frame will be invalid.
2053 finish_task_switch(this_rq(), prev
);
2057 * nr_running and nr_context_switches:
2059 * externally visible scheduler statistics: current number of runnable
2060 * threads, total number of context switches performed since bootup.
2062 unsigned long nr_running(void)
2064 unsigned long i
, sum
= 0;
2066 for_each_online_cpu(i
)
2067 sum
+= cpu_rq(i
)->nr_running
;
2072 unsigned long long nr_context_switches(void)
2075 unsigned long long sum
= 0;
2077 for_each_possible_cpu(i
)
2078 sum
+= cpu_rq(i
)->nr_switches
;
2083 unsigned long nr_iowait(void)
2085 unsigned long i
, sum
= 0;
2087 for_each_possible_cpu(i
)
2088 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2093 unsigned long nr_iowait_cpu(int cpu
)
2095 struct rq
*this = cpu_rq(cpu
);
2096 return atomic_read(&this->nr_iowait
);
2099 unsigned long this_cpu_load(void)
2101 struct rq
*this = this_rq();
2102 return this->cpu_load
[0];
2107 * Global load-average calculations
2109 * We take a distributed and async approach to calculating the global load-avg
2110 * in order to minimize overhead.
2112 * The global load average is an exponentially decaying average of nr_running +
2113 * nr_uninterruptible.
2115 * Once every LOAD_FREQ:
2118 * for_each_possible_cpu(cpu)
2119 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2121 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2123 * Due to a number of reasons the above turns in the mess below:
2125 * - for_each_possible_cpu() is prohibitively expensive on machines with
2126 * serious number of cpus, therefore we need to take a distributed approach
2127 * to calculating nr_active.
2129 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2130 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2132 * So assuming nr_active := 0 when we start out -- true per definition, we
2133 * can simply take per-cpu deltas and fold those into a global accumulate
2134 * to obtain the same result. See calc_load_fold_active().
2136 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2137 * across the machine, we assume 10 ticks is sufficient time for every
2138 * cpu to have completed this task.
2140 * This places an upper-bound on the IRQ-off latency of the machine. Then
2141 * again, being late doesn't loose the delta, just wrecks the sample.
2143 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2144 * this would add another cross-cpu cacheline miss and atomic operation
2145 * to the wakeup path. Instead we increment on whatever cpu the task ran
2146 * when it went into uninterruptible state and decrement on whatever cpu
2147 * did the wakeup. This means that only the sum of nr_uninterruptible over
2148 * all cpus yields the correct result.
2150 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2153 /* Variables and functions for calc_load */
2154 static atomic_long_t calc_load_tasks
;
2155 static unsigned long calc_load_update
;
2156 unsigned long avenrun
[3];
2157 EXPORT_SYMBOL(avenrun
); /* should be removed */
2160 * get_avenrun - get the load average array
2161 * @loads: pointer to dest load array
2162 * @offset: offset to add
2163 * @shift: shift count to shift the result left
2165 * These values are estimates at best, so no need for locking.
2167 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2169 loads
[0] = (avenrun
[0] + offset
) << shift
;
2170 loads
[1] = (avenrun
[1] + offset
) << shift
;
2171 loads
[2] = (avenrun
[2] + offset
) << shift
;
2174 static long calc_load_fold_active(struct rq
*this_rq
)
2176 long nr_active
, delta
= 0;
2178 nr_active
= this_rq
->nr_running
;
2179 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2181 if (nr_active
!= this_rq
->calc_load_active
) {
2182 delta
= nr_active
- this_rq
->calc_load_active
;
2183 this_rq
->calc_load_active
= nr_active
;
2190 * a1 = a0 * e + a * (1 - e)
2192 static unsigned long
2193 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2196 load
+= active
* (FIXED_1
- exp
);
2197 load
+= 1UL << (FSHIFT
- 1);
2198 return load
>> FSHIFT
;
2201 #ifdef CONFIG_NO_HZ_COMMON
2203 * Handle NO_HZ for the global load-average.
2205 * Since the above described distributed algorithm to compute the global
2206 * load-average relies on per-cpu sampling from the tick, it is affected by
2209 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2210 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2211 * when we read the global state.
2213 * Obviously reality has to ruin such a delightfully simple scheme:
2215 * - When we go NO_HZ idle during the window, we can negate our sample
2216 * contribution, causing under-accounting.
2218 * We avoid this by keeping two idle-delta counters and flipping them
2219 * when the window starts, thus separating old and new NO_HZ load.
2221 * The only trick is the slight shift in index flip for read vs write.
2225 * |-|-----------|-|-----------|-|-----------|-|
2226 * r:0 0 1 1 0 0 1 1 0
2227 * w:0 1 1 0 0 1 1 0 0
2229 * This ensures we'll fold the old idle contribution in this window while
2230 * accumlating the new one.
2232 * - When we wake up from NO_HZ idle during the window, we push up our
2233 * contribution, since we effectively move our sample point to a known
2236 * This is solved by pushing the window forward, and thus skipping the
2237 * sample, for this cpu (effectively using the idle-delta for this cpu which
2238 * was in effect at the time the window opened). This also solves the issue
2239 * of having to deal with a cpu having been in NOHZ idle for multiple
2240 * LOAD_FREQ intervals.
2242 * When making the ILB scale, we should try to pull this in as well.
2244 static atomic_long_t calc_load_idle
[2];
2245 static int calc_load_idx
;
2247 static inline int calc_load_write_idx(void)
2249 int idx
= calc_load_idx
;
2252 * See calc_global_nohz(), if we observe the new index, we also
2253 * need to observe the new update time.
2258 * If the folding window started, make sure we start writing in the
2261 if (!time_before(jiffies
, calc_load_update
))
2267 static inline int calc_load_read_idx(void)
2269 return calc_load_idx
& 1;
2272 void calc_load_enter_idle(void)
2274 struct rq
*this_rq
= this_rq();
2278 * We're going into NOHZ mode, if there's any pending delta, fold it
2279 * into the pending idle delta.
2281 delta
= calc_load_fold_active(this_rq
);
2283 int idx
= calc_load_write_idx();
2284 atomic_long_add(delta
, &calc_load_idle
[idx
]);
2288 void calc_load_exit_idle(void)
2290 struct rq
*this_rq
= this_rq();
2293 * If we're still before the sample window, we're done.
2295 if (time_before(jiffies
, this_rq
->calc_load_update
))
2299 * We woke inside or after the sample window, this means we're already
2300 * accounted through the nohz accounting, so skip the entire deal and
2301 * sync up for the next window.
2303 this_rq
->calc_load_update
= calc_load_update
;
2304 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
2305 this_rq
->calc_load_update
+= LOAD_FREQ
;
2308 static long calc_load_fold_idle(void)
2310 int idx
= calc_load_read_idx();
2313 if (atomic_long_read(&calc_load_idle
[idx
]))
2314 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
2320 * fixed_power_int - compute: x^n, in O(log n) time
2322 * @x: base of the power
2323 * @frac_bits: fractional bits of @x
2324 * @n: power to raise @x to.
2326 * By exploiting the relation between the definition of the natural power
2327 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2328 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2329 * (where: n_i \elem {0, 1}, the binary vector representing n),
2330 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2331 * of course trivially computable in O(log_2 n), the length of our binary
2334 static unsigned long
2335 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2337 unsigned long result
= 1UL << frac_bits
;
2342 result
+= 1UL << (frac_bits
- 1);
2343 result
>>= frac_bits
;
2349 x
+= 1UL << (frac_bits
- 1);
2357 * a1 = a0 * e + a * (1 - e)
2359 * a2 = a1 * e + a * (1 - e)
2360 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2361 * = a0 * e^2 + a * (1 - e) * (1 + e)
2363 * a3 = a2 * e + a * (1 - e)
2364 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2365 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2369 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2370 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2371 * = a0 * e^n + a * (1 - e^n)
2373 * [1] application of the geometric series:
2376 * S_n := \Sum x^i = -------------
2379 static unsigned long
2380 calc_load_n(unsigned long load
, unsigned long exp
,
2381 unsigned long active
, unsigned int n
)
2384 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
2388 * NO_HZ can leave us missing all per-cpu ticks calling
2389 * calc_load_account_active(), but since an idle CPU folds its delta into
2390 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2391 * in the pending idle delta if our idle period crossed a load cycle boundary.
2393 * Once we've updated the global active value, we need to apply the exponential
2394 * weights adjusted to the number of cycles missed.
2396 static void calc_global_nohz(void)
2398 long delta
, active
, n
;
2400 if (!time_before(jiffies
, calc_load_update
+ 10)) {
2402 * Catch-up, fold however many we are behind still
2404 delta
= jiffies
- calc_load_update
- 10;
2405 n
= 1 + (delta
/ LOAD_FREQ
);
2407 active
= atomic_long_read(&calc_load_tasks
);
2408 active
= active
> 0 ? active
* FIXED_1
: 0;
2410 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
2411 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
2412 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
2414 calc_load_update
+= n
* LOAD_FREQ
;
2418 * Flip the idle index...
2420 * Make sure we first write the new time then flip the index, so that
2421 * calc_load_write_idx() will see the new time when it reads the new
2422 * index, this avoids a double flip messing things up.
2427 #else /* !CONFIG_NO_HZ_COMMON */
2429 static inline long calc_load_fold_idle(void) { return 0; }
2430 static inline void calc_global_nohz(void) { }
2432 #endif /* CONFIG_NO_HZ_COMMON */
2435 * calc_load - update the avenrun load estimates 10 ticks after the
2436 * CPUs have updated calc_load_tasks.
2438 void calc_global_load(unsigned long ticks
)
2442 if (time_before(jiffies
, calc_load_update
+ 10))
2446 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2448 delta
= calc_load_fold_idle();
2450 atomic_long_add(delta
, &calc_load_tasks
);
2452 active
= atomic_long_read(&calc_load_tasks
);
2453 active
= active
> 0 ? active
* FIXED_1
: 0;
2455 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2456 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2457 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2459 calc_load_update
+= LOAD_FREQ
;
2462 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2468 * Called from update_cpu_load() to periodically update this CPU's
2471 static void calc_load_account_active(struct rq
*this_rq
)
2475 if (time_before(jiffies
, this_rq
->calc_load_update
))
2478 delta
= calc_load_fold_active(this_rq
);
2480 atomic_long_add(delta
, &calc_load_tasks
);
2482 this_rq
->calc_load_update
+= LOAD_FREQ
;
2486 * End of global load-average stuff
2490 * The exact cpuload at various idx values, calculated at every tick would be
2491 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2493 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2494 * on nth tick when cpu may be busy, then we have:
2495 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2496 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2498 * decay_load_missed() below does efficient calculation of
2499 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2500 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2502 * The calculation is approximated on a 128 point scale.
2503 * degrade_zero_ticks is the number of ticks after which load at any
2504 * particular idx is approximated to be zero.
2505 * degrade_factor is a precomputed table, a row for each load idx.
2506 * Each column corresponds to degradation factor for a power of two ticks,
2507 * based on 128 point scale.
2509 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2510 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2512 * With this power of 2 load factors, we can degrade the load n times
2513 * by looking at 1 bits in n and doing as many mult/shift instead of
2514 * n mult/shifts needed by the exact degradation.
2516 #define DEGRADE_SHIFT 7
2517 static const unsigned char
2518 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
2519 static const unsigned char
2520 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
2521 {0, 0, 0, 0, 0, 0, 0, 0},
2522 {64, 32, 8, 0, 0, 0, 0, 0},
2523 {96, 72, 40, 12, 1, 0, 0},
2524 {112, 98, 75, 43, 15, 1, 0},
2525 {120, 112, 98, 76, 45, 16, 2} };
2528 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2529 * would be when CPU is idle and so we just decay the old load without
2530 * adding any new load.
2532 static unsigned long
2533 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
2537 if (!missed_updates
)
2540 if (missed_updates
>= degrade_zero_ticks
[idx
])
2544 return load
>> missed_updates
;
2546 while (missed_updates
) {
2547 if (missed_updates
% 2)
2548 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
2550 missed_updates
>>= 1;
2557 * Update rq->cpu_load[] statistics. This function is usually called every
2558 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2559 * every tick. We fix it up based on jiffies.
2561 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
2562 unsigned long pending_updates
)
2566 this_rq
->nr_load_updates
++;
2568 /* Update our load: */
2569 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
2570 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2571 unsigned long old_load
, new_load
;
2573 /* scale is effectively 1 << i now, and >> i divides by scale */
2575 old_load
= this_rq
->cpu_load
[i
];
2576 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
2577 new_load
= this_load
;
2579 * Round up the averaging division if load is increasing. This
2580 * prevents us from getting stuck on 9 if the load is 10, for
2583 if (new_load
> old_load
)
2584 new_load
+= scale
- 1;
2586 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
2589 sched_avg_update(this_rq
);
2592 #ifdef CONFIG_NO_HZ_COMMON
2594 * There is no sane way to deal with nohz on smp when using jiffies because the
2595 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2596 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2598 * Therefore we cannot use the delta approach from the regular tick since that
2599 * would seriously skew the load calculation. However we'll make do for those
2600 * updates happening while idle (nohz_idle_balance) or coming out of idle
2601 * (tick_nohz_idle_exit).
2603 * This means we might still be one tick off for nohz periods.
2607 * Called from nohz_idle_balance() to update the load ratings before doing the
2610 void update_idle_cpu_load(struct rq
*this_rq
)
2612 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2613 unsigned long load
= this_rq
->load
.weight
;
2614 unsigned long pending_updates
;
2617 * bail if there's load or we're actually up-to-date.
2619 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
2622 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2623 this_rq
->last_load_update_tick
= curr_jiffies
;
2625 __update_cpu_load(this_rq
, load
, pending_updates
);
2629 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2631 void update_cpu_load_nohz(void)
2633 struct rq
*this_rq
= this_rq();
2634 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
2635 unsigned long pending_updates
;
2637 if (curr_jiffies
== this_rq
->last_load_update_tick
)
2640 raw_spin_lock(&this_rq
->lock
);
2641 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
2642 if (pending_updates
) {
2643 this_rq
->last_load_update_tick
= curr_jiffies
;
2645 * We were idle, this means load 0, the current load might be
2646 * !0 due to remote wakeups and the sort.
2648 __update_cpu_load(this_rq
, 0, pending_updates
);
2650 raw_spin_unlock(&this_rq
->lock
);
2652 #endif /* CONFIG_NO_HZ_COMMON */
2655 * Called from scheduler_tick()
2657 static void update_cpu_load_active(struct rq
*this_rq
)
2660 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2662 this_rq
->last_load_update_tick
= jiffies
;
2663 __update_cpu_load(this_rq
, this_rq
->load
.weight
, 1);
2665 calc_load_account_active(this_rq
);
2671 * sched_exec - execve() is a valuable balancing opportunity, because at
2672 * this point the task has the smallest effective memory and cache footprint.
2674 void sched_exec(void)
2676 struct task_struct
*p
= current
;
2677 unsigned long flags
;
2680 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2681 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2682 if (dest_cpu
== smp_processor_id())
2685 if (likely(cpu_active(dest_cpu
))) {
2686 struct migration_arg arg
= { p
, dest_cpu
};
2688 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2689 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2693 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2698 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2699 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2701 EXPORT_PER_CPU_SYMBOL(kstat
);
2702 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2705 * Return any ns on the sched_clock that have not yet been accounted in
2706 * @p in case that task is currently running.
2708 * Called with task_rq_lock() held on @rq.
2710 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2714 if (task_current(rq
, p
)) {
2715 update_rq_clock(rq
);
2716 ns
= rq
->clock_task
- p
->se
.exec_start
;
2724 unsigned long long task_delta_exec(struct task_struct
*p
)
2726 unsigned long flags
;
2730 rq
= task_rq_lock(p
, &flags
);
2731 ns
= do_task_delta_exec(p
, rq
);
2732 task_rq_unlock(rq
, p
, &flags
);
2738 * Return accounted runtime for the task.
2739 * In case the task is currently running, return the runtime plus current's
2740 * pending runtime that have not been accounted yet.
2742 unsigned long long task_sched_runtime(struct task_struct
*p
)
2744 unsigned long flags
;
2748 rq
= task_rq_lock(p
, &flags
);
2749 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2750 task_rq_unlock(rq
, p
, &flags
);
2756 * This function gets called by the timer code, with HZ frequency.
2757 * We call it with interrupts disabled.
2759 void scheduler_tick(void)
2761 int cpu
= smp_processor_id();
2762 struct rq
*rq
= cpu_rq(cpu
);
2763 struct task_struct
*curr
= rq
->curr
;
2767 raw_spin_lock(&rq
->lock
);
2768 update_rq_clock(rq
);
2769 update_cpu_load_active(rq
);
2770 curr
->sched_class
->task_tick(rq
, curr
, 0);
2771 raw_spin_unlock(&rq
->lock
);
2773 perf_event_task_tick();
2776 rq
->idle_balance
= idle_cpu(cpu
);
2777 trigger_load_balance(rq
, cpu
);
2779 rq_last_tick_reset(rq
);
2782 #ifdef CONFIG_NO_HZ_FULL
2784 * scheduler_tick_max_deferment
2786 * Keep at least one tick per second when a single
2787 * active task is running because the scheduler doesn't
2788 * yet completely support full dynticks environment.
2790 * This makes sure that uptime, CFS vruntime, load
2791 * balancing, etc... continue to move forward, even
2792 * with a very low granularity.
2794 u64
scheduler_tick_max_deferment(void)
2796 struct rq
*rq
= this_rq();
2797 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2799 next
= rq
->last_sched_tick
+ HZ
;
2801 if (time_before_eq(next
, now
))
2804 return jiffies_to_usecs(next
- now
) * NSEC_PER_USEC
;
2808 notrace
unsigned long get_parent_ip(unsigned long addr
)
2810 if (in_lock_functions(addr
)) {
2811 addr
= CALLER_ADDR2
;
2812 if (in_lock_functions(addr
))
2813 addr
= CALLER_ADDR3
;
2818 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2819 defined(CONFIG_PREEMPT_TRACER))
2821 void __kprobes
add_preempt_count(int val
)
2823 #ifdef CONFIG_DEBUG_PREEMPT
2827 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2830 preempt_count() += val
;
2831 #ifdef CONFIG_DEBUG_PREEMPT
2833 * Spinlock count overflowing soon?
2835 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2838 if (preempt_count() == val
)
2839 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2841 EXPORT_SYMBOL(add_preempt_count
);
2843 void __kprobes
sub_preempt_count(int val
)
2845 #ifdef CONFIG_DEBUG_PREEMPT
2849 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2852 * Is the spinlock portion underflowing?
2854 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2855 !(preempt_count() & PREEMPT_MASK
)))
2859 if (preempt_count() == val
)
2860 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2861 preempt_count() -= val
;
2863 EXPORT_SYMBOL(sub_preempt_count
);
2868 * Print scheduling while atomic bug:
2870 static noinline
void __schedule_bug(struct task_struct
*prev
)
2872 if (oops_in_progress
)
2875 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2876 prev
->comm
, prev
->pid
, preempt_count());
2878 debug_show_held_locks(prev
);
2880 if (irqs_disabled())
2881 print_irqtrace_events(prev
);
2883 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2887 * Various schedule()-time debugging checks and statistics:
2889 static inline void schedule_debug(struct task_struct
*prev
)
2892 * Test if we are atomic. Since do_exit() needs to call into
2893 * schedule() atomically, we ignore that path for now.
2894 * Otherwise, whine if we are scheduling when we should not be.
2896 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2897 __schedule_bug(prev
);
2900 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2902 schedstat_inc(this_rq(), sched_count
);
2905 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2907 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2908 update_rq_clock(rq
);
2909 prev
->sched_class
->put_prev_task(rq
, prev
);
2913 * Pick up the highest-prio task:
2915 static inline struct task_struct
*
2916 pick_next_task(struct rq
*rq
)
2918 const struct sched_class
*class;
2919 struct task_struct
*p
;
2922 * Optimization: we know that if all tasks are in
2923 * the fair class we can call that function directly:
2925 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2926 p
= fair_sched_class
.pick_next_task(rq
);
2931 for_each_class(class) {
2932 p
= class->pick_next_task(rq
);
2937 BUG(); /* the idle class will always have a runnable task */
2941 * __schedule() is the main scheduler function.
2943 * The main means of driving the scheduler and thus entering this function are:
2945 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2947 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2948 * paths. For example, see arch/x86/entry_64.S.
2950 * To drive preemption between tasks, the scheduler sets the flag in timer
2951 * interrupt handler scheduler_tick().
2953 * 3. Wakeups don't really cause entry into schedule(). They add a
2954 * task to the run-queue and that's it.
2956 * Now, if the new task added to the run-queue preempts the current
2957 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2958 * called on the nearest possible occasion:
2960 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2962 * - in syscall or exception context, at the next outmost
2963 * preempt_enable(). (this might be as soon as the wake_up()'s
2966 * - in IRQ context, return from interrupt-handler to
2967 * preemptible context
2969 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2972 * - cond_resched() call
2973 * - explicit schedule() call
2974 * - return from syscall or exception to user-space
2975 * - return from interrupt-handler to user-space
2977 static void __sched
__schedule(void)
2979 struct task_struct
*prev
, *next
;
2980 unsigned long *switch_count
;
2986 cpu
= smp_processor_id();
2988 rcu_note_context_switch(cpu
);
2991 schedule_debug(prev
);
2993 if (sched_feat(HRTICK
))
2997 * Make sure that signal_pending_state()->signal_pending() below
2998 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2999 * done by the caller to avoid the race with signal_wake_up().
3001 smp_mb__before_spinlock();
3002 raw_spin_lock_irq(&rq
->lock
);
3004 switch_count
= &prev
->nivcsw
;
3005 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3006 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3007 prev
->state
= TASK_RUNNING
;
3009 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3013 * If a worker went to sleep, notify and ask workqueue
3014 * whether it wants to wake up a task to maintain
3017 if (prev
->flags
& PF_WQ_WORKER
) {
3018 struct task_struct
*to_wakeup
;
3020 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3022 try_to_wake_up_local(to_wakeup
);
3025 switch_count
= &prev
->nvcsw
;
3028 pre_schedule(rq
, prev
);
3030 if (unlikely(!rq
->nr_running
))
3031 idle_balance(cpu
, rq
);
3033 put_prev_task(rq
, prev
);
3034 next
= pick_next_task(rq
);
3035 clear_tsk_need_resched(prev
);
3036 rq
->skip_clock_update
= 0;
3038 if (likely(prev
!= next
)) {
3043 context_switch(rq
, prev
, next
); /* unlocks the rq */
3045 * The context switch have flipped the stack from under us
3046 * and restored the local variables which were saved when
3047 * this task called schedule() in the past. prev == current
3048 * is still correct, but it can be moved to another cpu/rq.
3050 cpu
= smp_processor_id();
3053 raw_spin_unlock_irq(&rq
->lock
);
3057 sched_preempt_enable_no_resched();
3062 static inline void sched_submit_work(struct task_struct
*tsk
)
3064 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3067 * If we are going to sleep and we have plugged IO queued,
3068 * make sure to submit it to avoid deadlocks.
3070 if (blk_needs_flush_plug(tsk
))
3071 blk_schedule_flush_plug(tsk
);
3074 asmlinkage
void __sched
schedule(void)
3076 struct task_struct
*tsk
= current
;
3078 sched_submit_work(tsk
);
3081 EXPORT_SYMBOL(schedule
);
3083 #ifdef CONFIG_CONTEXT_TRACKING
3084 asmlinkage
void __sched
schedule_user(void)
3087 * If we come here after a random call to set_need_resched(),
3088 * or we have been woken up remotely but the IPI has not yet arrived,
3089 * we haven't yet exited the RCU idle mode. Do it here manually until
3090 * we find a better solution.
3099 * schedule_preempt_disabled - called with preemption disabled
3101 * Returns with preemption disabled. Note: preempt_count must be 1
3103 void __sched
schedule_preempt_disabled(void)
3105 sched_preempt_enable_no_resched();
3110 #ifdef CONFIG_PREEMPT
3112 * this is the entry point to schedule() from in-kernel preemption
3113 * off of preempt_enable. Kernel preemptions off return from interrupt
3114 * occur there and call schedule directly.
3116 asmlinkage
void __sched notrace
preempt_schedule(void)
3118 struct thread_info
*ti
= current_thread_info();
3121 * If there is a non-zero preempt_count or interrupts are disabled,
3122 * we do not want to preempt the current task. Just return..
3124 if (likely(ti
->preempt_count
|| irqs_disabled()))
3128 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3130 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3133 * Check again in case we missed a preemption opportunity
3134 * between schedule and now.
3137 } while (need_resched());
3139 EXPORT_SYMBOL(preempt_schedule
);
3142 * this is the entry point to schedule() from kernel preemption
3143 * off of irq context.
3144 * Note, that this is called and return with irqs disabled. This will
3145 * protect us against recursive calling from irq.
3147 asmlinkage
void __sched
preempt_schedule_irq(void)
3149 struct thread_info
*ti
= current_thread_info();
3150 enum ctx_state prev_state
;
3152 /* Catch callers which need to be fixed */
3153 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3155 prev_state
= exception_enter();
3158 add_preempt_count(PREEMPT_ACTIVE
);
3161 local_irq_disable();
3162 sub_preempt_count(PREEMPT_ACTIVE
);
3165 * Check again in case we missed a preemption opportunity
3166 * between schedule and now.
3169 } while (need_resched());
3171 exception_exit(prev_state
);
3174 #endif /* CONFIG_PREEMPT */
3176 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3179 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3181 EXPORT_SYMBOL(default_wake_function
);
3184 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3185 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3186 * number) then we wake all the non-exclusive tasks and one exclusive task.
3188 * There are circumstances in which we can try to wake a task which has already
3189 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3190 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3192 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3193 int nr_exclusive
, int wake_flags
, void *key
)
3195 wait_queue_t
*curr
, *next
;
3197 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3198 unsigned flags
= curr
->flags
;
3200 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3201 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3207 * __wake_up - wake up threads blocked on a waitqueue.
3209 * @mode: which threads
3210 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3211 * @key: is directly passed to the wakeup function
3213 * It may be assumed that this function implies a write memory barrier before
3214 * changing the task state if and only if any tasks are woken up.
3216 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3217 int nr_exclusive
, void *key
)
3219 unsigned long flags
;
3221 spin_lock_irqsave(&q
->lock
, flags
);
3222 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3223 spin_unlock_irqrestore(&q
->lock
, flags
);
3225 EXPORT_SYMBOL(__wake_up
);
3228 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3230 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
3232 __wake_up_common(q
, mode
, nr
, 0, NULL
);
3234 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3236 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3238 __wake_up_common(q
, mode
, 1, 0, key
);
3240 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
3243 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3245 * @mode: which threads
3246 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3247 * @key: opaque value to be passed to wakeup targets
3249 * The sync wakeup differs that the waker knows that it will schedule
3250 * away soon, so while the target thread will be woken up, it will not
3251 * be migrated to another CPU - ie. the two threads are 'synchronized'
3252 * with each other. This can prevent needless bouncing between CPUs.
3254 * On UP it can prevent extra preemption.
3256 * It may be assumed that this function implies a write memory barrier before
3257 * changing the task state if and only if any tasks are woken up.
3259 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3260 int nr_exclusive
, void *key
)
3262 unsigned long flags
;
3263 int wake_flags
= WF_SYNC
;
3268 if (unlikely(!nr_exclusive
))
3271 spin_lock_irqsave(&q
->lock
, flags
);
3272 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3273 spin_unlock_irqrestore(&q
->lock
, flags
);
3275 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3278 * __wake_up_sync - see __wake_up_sync_key()
3280 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3282 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3284 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3287 * complete: - signals a single thread waiting on this completion
3288 * @x: holds the state of this particular completion
3290 * This will wake up a single thread waiting on this completion. Threads will be
3291 * awakened in the same order in which they were queued.
3293 * See also complete_all(), wait_for_completion() and related routines.
3295 * It may be assumed that this function implies a write memory barrier before
3296 * changing the task state if and only if any tasks are woken up.
3298 void complete(struct completion
*x
)
3300 unsigned long flags
;
3302 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3304 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3305 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3307 EXPORT_SYMBOL(complete
);
3310 * complete_all: - signals all threads waiting on this completion
3311 * @x: holds the state of this particular completion
3313 * This will wake up all threads waiting on this particular completion event.
3315 * It may be assumed that this function implies a write memory barrier before
3316 * changing the task state if and only if any tasks are woken up.
3318 void complete_all(struct completion
*x
)
3320 unsigned long flags
;
3322 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3323 x
->done
+= UINT_MAX
/2;
3324 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3325 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3327 EXPORT_SYMBOL(complete_all
);
3329 static inline long __sched
3330 do_wait_for_common(struct completion
*x
,
3331 long (*action
)(long), long timeout
, int state
)
3334 DECLARE_WAITQUEUE(wait
, current
);
3336 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3338 if (signal_pending_state(state
, current
)) {
3339 timeout
= -ERESTARTSYS
;
3342 __set_current_state(state
);
3343 spin_unlock_irq(&x
->wait
.lock
);
3344 timeout
= action(timeout
);
3345 spin_lock_irq(&x
->wait
.lock
);
3346 } while (!x
->done
&& timeout
);
3347 __remove_wait_queue(&x
->wait
, &wait
);
3352 return timeout
?: 1;
3355 static inline long __sched
3356 __wait_for_common(struct completion
*x
,
3357 long (*action
)(long), long timeout
, int state
)
3361 spin_lock_irq(&x
->wait
.lock
);
3362 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
3363 spin_unlock_irq(&x
->wait
.lock
);
3368 wait_for_common(struct completion
*x
, long timeout
, int state
)
3370 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
3374 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
3376 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
3380 * wait_for_completion: - waits for completion of a task
3381 * @x: holds the state of this particular completion
3383 * This waits to be signaled for completion of a specific task. It is NOT
3384 * interruptible and there is no timeout.
3386 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3387 * and interrupt capability. Also see complete().
3389 void __sched
wait_for_completion(struct completion
*x
)
3391 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3393 EXPORT_SYMBOL(wait_for_completion
);
3396 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3397 * @x: holds the state of this particular completion
3398 * @timeout: timeout value in jiffies
3400 * This waits for either a completion of a specific task to be signaled or for a
3401 * specified timeout to expire. The timeout is in jiffies. It is not
3404 * The return value is 0 if timed out, and positive (at least 1, or number of
3405 * jiffies left till timeout) if completed.
3407 unsigned long __sched
3408 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3410 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3412 EXPORT_SYMBOL(wait_for_completion_timeout
);
3415 * wait_for_completion_io: - waits for completion of a task
3416 * @x: holds the state of this particular completion
3418 * This waits to be signaled for completion of a specific task. It is NOT
3419 * interruptible and there is no timeout. The caller is accounted as waiting
3422 void __sched
wait_for_completion_io(struct completion
*x
)
3424 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3426 EXPORT_SYMBOL(wait_for_completion_io
);
3429 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3430 * @x: holds the state of this particular completion
3431 * @timeout: timeout value in jiffies
3433 * This waits for either a completion of a specific task to be signaled or for a
3434 * specified timeout to expire. The timeout is in jiffies. It is not
3435 * interruptible. The caller is accounted as waiting for IO.
3437 * The return value is 0 if timed out, and positive (at least 1, or number of
3438 * jiffies left till timeout) if completed.
3440 unsigned long __sched
3441 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
3443 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3445 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
3448 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3449 * @x: holds the state of this particular completion
3451 * This waits for completion of a specific task to be signaled. It is
3454 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3456 int __sched
wait_for_completion_interruptible(struct completion
*x
)
3458 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
3459 if (t
== -ERESTARTSYS
)
3463 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3466 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3467 * @x: holds the state of this particular completion
3468 * @timeout: timeout value in jiffies
3470 * This waits for either a completion of a specific task to be signaled or for a
3471 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3473 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3474 * positive (at least 1, or number of jiffies left till timeout) if completed.
3477 wait_for_completion_interruptible_timeout(struct completion
*x
,
3478 unsigned long timeout
)
3480 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
3482 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3485 * wait_for_completion_killable: - waits for completion of a task (killable)
3486 * @x: holds the state of this particular completion
3488 * This waits to be signaled for completion of a specific task. It can be
3489 * interrupted by a kill signal.
3491 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3493 int __sched
wait_for_completion_killable(struct completion
*x
)
3495 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
3496 if (t
== -ERESTARTSYS
)
3500 EXPORT_SYMBOL(wait_for_completion_killable
);
3503 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3504 * @x: holds the state of this particular completion
3505 * @timeout: timeout value in jiffies
3507 * This waits for either a completion of a specific task to be
3508 * signaled or for a specified timeout to expire. It can be
3509 * interrupted by a kill signal. The timeout is in jiffies.
3511 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3512 * positive (at least 1, or number of jiffies left till timeout) if completed.
3515 wait_for_completion_killable_timeout(struct completion
*x
,
3516 unsigned long timeout
)
3518 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
3520 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
3523 * try_wait_for_completion - try to decrement a completion without blocking
3524 * @x: completion structure
3526 * Returns: 0 if a decrement cannot be done without blocking
3527 * 1 if a decrement succeeded.
3529 * If a completion is being used as a counting completion,
3530 * attempt to decrement the counter without blocking. This
3531 * enables us to avoid waiting if the resource the completion
3532 * is protecting is not available.
3534 bool try_wait_for_completion(struct completion
*x
)
3536 unsigned long flags
;
3539 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3544 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3547 EXPORT_SYMBOL(try_wait_for_completion
);
3550 * completion_done - Test to see if a completion has any waiters
3551 * @x: completion structure
3553 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3554 * 1 if there are no waiters.
3557 bool completion_done(struct completion
*x
)
3559 unsigned long flags
;
3562 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3565 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3568 EXPORT_SYMBOL(completion_done
);
3571 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
3573 unsigned long flags
;
3576 init_waitqueue_entry(&wait
, current
);
3578 __set_current_state(state
);
3580 spin_lock_irqsave(&q
->lock
, flags
);
3581 __add_wait_queue(q
, &wait
);
3582 spin_unlock(&q
->lock
);
3583 timeout
= schedule_timeout(timeout
);
3584 spin_lock_irq(&q
->lock
);
3585 __remove_wait_queue(q
, &wait
);
3586 spin_unlock_irqrestore(&q
->lock
, flags
);
3591 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3593 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3595 EXPORT_SYMBOL(interruptible_sleep_on
);
3598 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3600 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3602 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3604 void __sched
sleep_on(wait_queue_head_t
*q
)
3606 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3608 EXPORT_SYMBOL(sleep_on
);
3610 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3612 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3614 EXPORT_SYMBOL(sleep_on_timeout
);
3616 #ifdef CONFIG_RT_MUTEXES
3619 * rt_mutex_setprio - set the current priority of a task
3621 * @prio: prio value (kernel-internal form)
3623 * This function changes the 'effective' priority of a task. It does
3624 * not touch ->normal_prio like __setscheduler().
3626 * Used by the rt_mutex code to implement priority inheritance logic.
3628 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3630 int oldprio
, on_rq
, running
;
3632 const struct sched_class
*prev_class
;
3634 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3636 rq
= __task_rq_lock(p
);
3639 * Idle task boosting is a nono in general. There is one
3640 * exception, when PREEMPT_RT and NOHZ is active:
3642 * The idle task calls get_next_timer_interrupt() and holds
3643 * the timer wheel base->lock on the CPU and another CPU wants
3644 * to access the timer (probably to cancel it). We can safely
3645 * ignore the boosting request, as the idle CPU runs this code
3646 * with interrupts disabled and will complete the lock
3647 * protected section without being interrupted. So there is no
3648 * real need to boost.
3650 if (unlikely(p
== rq
->idle
)) {
3651 WARN_ON(p
!= rq
->curr
);
3652 WARN_ON(p
->pi_blocked_on
);
3656 trace_sched_pi_setprio(p
, prio
);
3658 prev_class
= p
->sched_class
;
3660 running
= task_current(rq
, p
);
3662 dequeue_task(rq
, p
, 0);
3664 p
->sched_class
->put_prev_task(rq
, p
);
3667 p
->sched_class
= &rt_sched_class
;
3669 p
->sched_class
= &fair_sched_class
;
3674 p
->sched_class
->set_curr_task(rq
);
3676 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3678 check_class_changed(rq
, p
, prev_class
, oldprio
);
3680 __task_rq_unlock(rq
);
3683 void set_user_nice(struct task_struct
*p
, long nice
)
3685 int old_prio
, delta
, on_rq
;
3686 unsigned long flags
;
3689 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3692 * We have to be careful, if called from sys_setpriority(),
3693 * the task might be in the middle of scheduling on another CPU.
3695 rq
= task_rq_lock(p
, &flags
);
3697 * The RT priorities are set via sched_setscheduler(), but we still
3698 * allow the 'normal' nice value to be set - but as expected
3699 * it wont have any effect on scheduling until the task is
3700 * SCHED_FIFO/SCHED_RR:
3702 if (task_has_rt_policy(p
)) {
3703 p
->static_prio
= NICE_TO_PRIO(nice
);
3708 dequeue_task(rq
, p
, 0);
3710 p
->static_prio
= NICE_TO_PRIO(nice
);
3713 p
->prio
= effective_prio(p
);
3714 delta
= p
->prio
- old_prio
;
3717 enqueue_task(rq
, p
, 0);
3719 * If the task increased its priority or is running and
3720 * lowered its priority, then reschedule its CPU:
3722 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3723 resched_task(rq
->curr
);
3726 task_rq_unlock(rq
, p
, &flags
);
3728 EXPORT_SYMBOL(set_user_nice
);
3731 * can_nice - check if a task can reduce its nice value
3735 int can_nice(const struct task_struct
*p
, const int nice
)
3737 /* convert nice value [19,-20] to rlimit style value [1,40] */
3738 int nice_rlim
= 20 - nice
;
3740 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3741 capable(CAP_SYS_NICE
));
3744 #ifdef __ARCH_WANT_SYS_NICE
3747 * sys_nice - change the priority of the current process.
3748 * @increment: priority increment
3750 * sys_setpriority is a more generic, but much slower function that
3751 * does similar things.
3753 SYSCALL_DEFINE1(nice
, int, increment
)
3758 * Setpriority might change our priority at the same moment.
3759 * We don't have to worry. Conceptually one call occurs first
3760 * and we have a single winner.
3762 if (increment
< -40)
3767 nice
= TASK_NICE(current
) + increment
;
3773 if (increment
< 0 && !can_nice(current
, nice
))
3776 retval
= security_task_setnice(current
, nice
);
3780 set_user_nice(current
, nice
);
3787 * task_prio - return the priority value of a given task.
3788 * @p: the task in question.
3790 * This is the priority value as seen by users in /proc.
3791 * RT tasks are offset by -200. Normal tasks are centered
3792 * around 0, value goes from -16 to +15.
3794 int task_prio(const struct task_struct
*p
)
3796 return p
->prio
- MAX_RT_PRIO
;
3800 * task_nice - return the nice value of a given task.
3801 * @p: the task in question.
3803 int task_nice(const struct task_struct
*p
)
3805 return TASK_NICE(p
);
3807 EXPORT_SYMBOL(task_nice
);
3810 * idle_cpu - is a given cpu idle currently?
3811 * @cpu: the processor in question.
3813 int idle_cpu(int cpu
)
3815 struct rq
*rq
= cpu_rq(cpu
);
3817 if (rq
->curr
!= rq
->idle
)
3824 if (!llist_empty(&rq
->wake_list
))
3832 * idle_task - return the idle task for a given cpu.
3833 * @cpu: the processor in question.
3835 struct task_struct
*idle_task(int cpu
)
3837 return cpu_rq(cpu
)->idle
;
3841 * find_process_by_pid - find a process with a matching PID value.
3842 * @pid: the pid in question.
3844 static struct task_struct
*find_process_by_pid(pid_t pid
)
3846 return pid
? find_task_by_vpid(pid
) : current
;
3849 /* Actually do priority change: must hold rq lock. */
3851 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3854 p
->rt_priority
= prio
;
3855 p
->normal_prio
= normal_prio(p
);
3856 /* we are holding p->pi_lock already */
3857 p
->prio
= rt_mutex_getprio(p
);
3858 if (rt_prio(p
->prio
))
3859 p
->sched_class
= &rt_sched_class
;
3861 p
->sched_class
= &fair_sched_class
;
3866 * check the target process has a UID that matches the current process's
3868 static bool check_same_owner(struct task_struct
*p
)
3870 const struct cred
*cred
= current_cred(), *pcred
;
3874 pcred
= __task_cred(p
);
3875 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3876 uid_eq(cred
->euid
, pcred
->uid
));
3881 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3882 const struct sched_param
*param
, bool user
)
3884 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3885 unsigned long flags
;
3886 const struct sched_class
*prev_class
;
3890 /* may grab non-irq protected spin_locks */
3891 BUG_ON(in_interrupt());
3893 /* double check policy once rq lock held */
3895 reset_on_fork
= p
->sched_reset_on_fork
;
3896 policy
= oldpolicy
= p
->policy
;
3898 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3899 policy
&= ~SCHED_RESET_ON_FORK
;
3901 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3902 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3903 policy
!= SCHED_IDLE
)
3908 * Valid priorities for SCHED_FIFO and SCHED_RR are
3909 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3910 * SCHED_BATCH and SCHED_IDLE is 0.
3912 if (param
->sched_priority
< 0 ||
3913 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3914 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3916 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3920 * Allow unprivileged RT tasks to decrease priority:
3922 if (user
&& !capable(CAP_SYS_NICE
)) {
3923 if (rt_policy(policy
)) {
3924 unsigned long rlim_rtprio
=
3925 task_rlimit(p
, RLIMIT_RTPRIO
);
3927 /* can't set/change the rt policy */
3928 if (policy
!= p
->policy
&& !rlim_rtprio
)
3931 /* can't increase priority */
3932 if (param
->sched_priority
> p
->rt_priority
&&
3933 param
->sched_priority
> rlim_rtprio
)
3938 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3939 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3941 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3942 if (!can_nice(p
, TASK_NICE(p
)))
3946 /* can't change other user's priorities */
3947 if (!check_same_owner(p
))
3950 /* Normal users shall not reset the sched_reset_on_fork flag */
3951 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3956 retval
= security_task_setscheduler(p
);
3962 * make sure no PI-waiters arrive (or leave) while we are
3963 * changing the priority of the task:
3965 * To be able to change p->policy safely, the appropriate
3966 * runqueue lock must be held.
3968 rq
= task_rq_lock(p
, &flags
);
3971 * Changing the policy of the stop threads its a very bad idea
3973 if (p
== rq
->stop
) {
3974 task_rq_unlock(rq
, p
, &flags
);
3979 * If not changing anything there's no need to proceed further:
3981 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3982 param
->sched_priority
== p
->rt_priority
))) {
3983 task_rq_unlock(rq
, p
, &flags
);
3987 #ifdef CONFIG_RT_GROUP_SCHED
3990 * Do not allow realtime tasks into groups that have no runtime
3993 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3994 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3995 !task_group_is_autogroup(task_group(p
))) {
3996 task_rq_unlock(rq
, p
, &flags
);
4002 /* recheck policy now with rq lock held */
4003 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4004 policy
= oldpolicy
= -1;
4005 task_rq_unlock(rq
, p
, &flags
);
4009 running
= task_current(rq
, p
);
4011 dequeue_task(rq
, p
, 0);
4013 p
->sched_class
->put_prev_task(rq
, p
);
4015 p
->sched_reset_on_fork
= reset_on_fork
;
4018 prev_class
= p
->sched_class
;
4019 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4022 p
->sched_class
->set_curr_task(rq
);
4024 enqueue_task(rq
, p
, 0);
4026 check_class_changed(rq
, p
, prev_class
, oldprio
);
4027 task_rq_unlock(rq
, p
, &flags
);
4029 rt_mutex_adjust_pi(p
);
4035 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4036 * @p: the task in question.
4037 * @policy: new policy.
4038 * @param: structure containing the new RT priority.
4040 * NOTE that the task may be already dead.
4042 int sched_setscheduler(struct task_struct
*p
, int policy
,
4043 const struct sched_param
*param
)
4045 return __sched_setscheduler(p
, policy
, param
, true);
4047 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4050 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4051 * @p: the task in question.
4052 * @policy: new policy.
4053 * @param: structure containing the new RT priority.
4055 * Just like sched_setscheduler, only don't bother checking if the
4056 * current context has permission. For example, this is needed in
4057 * stop_machine(): we create temporary high priority worker threads,
4058 * but our caller might not have that capability.
4060 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4061 const struct sched_param
*param
)
4063 return __sched_setscheduler(p
, policy
, param
, false);
4067 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4069 struct sched_param lparam
;
4070 struct task_struct
*p
;
4073 if (!param
|| pid
< 0)
4075 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4080 p
= find_process_by_pid(pid
);
4082 retval
= sched_setscheduler(p
, policy
, &lparam
);
4089 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4090 * @pid: the pid in question.
4091 * @policy: new policy.
4092 * @param: structure containing the new RT priority.
4094 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4095 struct sched_param __user
*, param
)
4097 /* negative values for policy are not valid */
4101 return do_sched_setscheduler(pid
, policy
, param
);
4105 * sys_sched_setparam - set/change the RT priority of a thread
4106 * @pid: the pid in question.
4107 * @param: structure containing the new RT priority.
4109 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4111 return do_sched_setscheduler(pid
, -1, param
);
4115 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4116 * @pid: the pid in question.
4118 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4120 struct task_struct
*p
;
4128 p
= find_process_by_pid(pid
);
4130 retval
= security_task_getscheduler(p
);
4133 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4140 * sys_sched_getparam - get the RT priority of a thread
4141 * @pid: the pid in question.
4142 * @param: structure containing the RT priority.
4144 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4146 struct sched_param lp
;
4147 struct task_struct
*p
;
4150 if (!param
|| pid
< 0)
4154 p
= find_process_by_pid(pid
);
4159 retval
= security_task_getscheduler(p
);
4163 lp
.sched_priority
= p
->rt_priority
;
4167 * This one might sleep, we cannot do it with a spinlock held ...
4169 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4178 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4180 cpumask_var_t cpus_allowed
, new_mask
;
4181 struct task_struct
*p
;
4187 p
= find_process_by_pid(pid
);
4194 /* Prevent p going away */
4198 if (p
->flags
& PF_NO_SETAFFINITY
) {
4202 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4206 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4208 goto out_free_cpus_allowed
;
4211 if (!check_same_owner(p
)) {
4213 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4220 retval
= security_task_setscheduler(p
);
4224 cpuset_cpus_allowed(p
, cpus_allowed
);
4225 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4227 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4230 cpuset_cpus_allowed(p
, cpus_allowed
);
4231 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4233 * We must have raced with a concurrent cpuset
4234 * update. Just reset the cpus_allowed to the
4235 * cpuset's cpus_allowed
4237 cpumask_copy(new_mask
, cpus_allowed
);
4242 free_cpumask_var(new_mask
);
4243 out_free_cpus_allowed
:
4244 free_cpumask_var(cpus_allowed
);
4251 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4252 struct cpumask
*new_mask
)
4254 if (len
< cpumask_size())
4255 cpumask_clear(new_mask
);
4256 else if (len
> cpumask_size())
4257 len
= cpumask_size();
4259 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4263 * sys_sched_setaffinity - set the cpu affinity of a process
4264 * @pid: pid of the process
4265 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4266 * @user_mask_ptr: user-space pointer to the new cpu mask
4268 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4269 unsigned long __user
*, user_mask_ptr
)
4271 cpumask_var_t new_mask
;
4274 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4277 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4279 retval
= sched_setaffinity(pid
, new_mask
);
4280 free_cpumask_var(new_mask
);
4284 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4286 struct task_struct
*p
;
4287 unsigned long flags
;
4294 p
= find_process_by_pid(pid
);
4298 retval
= security_task_getscheduler(p
);
4302 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4303 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4304 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4314 * sys_sched_getaffinity - get the cpu affinity of a process
4315 * @pid: pid of the process
4316 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4317 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4319 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4320 unsigned long __user
*, user_mask_ptr
)
4325 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4327 if (len
& (sizeof(unsigned long)-1))
4330 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4333 ret
= sched_getaffinity(pid
, mask
);
4335 size_t retlen
= min_t(size_t, len
, cpumask_size());
4337 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4342 free_cpumask_var(mask
);
4348 * sys_sched_yield - yield the current processor to other threads.
4350 * This function yields the current CPU to other tasks. If there are no
4351 * other threads running on this CPU then this function will return.
4353 SYSCALL_DEFINE0(sched_yield
)
4355 struct rq
*rq
= this_rq_lock();
4357 schedstat_inc(rq
, yld_count
);
4358 current
->sched_class
->yield_task(rq
);
4361 * Since we are going to call schedule() anyway, there's
4362 * no need to preempt or enable interrupts:
4364 __release(rq
->lock
);
4365 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4366 do_raw_spin_unlock(&rq
->lock
);
4367 sched_preempt_enable_no_resched();
4374 static inline int should_resched(void)
4376 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4379 static void __cond_resched(void)
4381 add_preempt_count(PREEMPT_ACTIVE
);
4383 sub_preempt_count(PREEMPT_ACTIVE
);
4386 int __sched
_cond_resched(void)
4388 if (should_resched()) {
4394 EXPORT_SYMBOL(_cond_resched
);
4397 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4398 * call schedule, and on return reacquire the lock.
4400 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4401 * operations here to prevent schedule() from being called twice (once via
4402 * spin_unlock(), once by hand).
4404 int __cond_resched_lock(spinlock_t
*lock
)
4406 int resched
= should_resched();
4409 lockdep_assert_held(lock
);
4411 if (spin_needbreak(lock
) || resched
) {
4422 EXPORT_SYMBOL(__cond_resched_lock
);
4424 int __sched
__cond_resched_softirq(void)
4426 BUG_ON(!in_softirq());
4428 if (should_resched()) {
4436 EXPORT_SYMBOL(__cond_resched_softirq
);
4439 * yield - yield the current processor to other threads.
4441 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4443 * The scheduler is at all times free to pick the calling task as the most
4444 * eligible task to run, if removing the yield() call from your code breaks
4445 * it, its already broken.
4447 * Typical broken usage is:
4452 * where one assumes that yield() will let 'the other' process run that will
4453 * make event true. If the current task is a SCHED_FIFO task that will never
4454 * happen. Never use yield() as a progress guarantee!!
4456 * If you want to use yield() to wait for something, use wait_event().
4457 * If you want to use yield() to be 'nice' for others, use cond_resched().
4458 * If you still want to use yield(), do not!
4460 void __sched
yield(void)
4462 set_current_state(TASK_RUNNING
);
4465 EXPORT_SYMBOL(yield
);
4468 * yield_to - yield the current processor to another thread in
4469 * your thread group, or accelerate that thread toward the
4470 * processor it's on.
4472 * @preempt: whether task preemption is allowed or not
4474 * It's the caller's job to ensure that the target task struct
4475 * can't go away on us before we can do any checks.
4478 * true (>0) if we indeed boosted the target task.
4479 * false (0) if we failed to boost the target.
4480 * -ESRCH if there's no task to yield to.
4482 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4484 struct task_struct
*curr
= current
;
4485 struct rq
*rq
, *p_rq
;
4486 unsigned long flags
;
4489 local_irq_save(flags
);
4495 * If we're the only runnable task on the rq and target rq also
4496 * has only one task, there's absolutely no point in yielding.
4498 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4503 double_rq_lock(rq
, p_rq
);
4504 while (task_rq(p
) != p_rq
) {
4505 double_rq_unlock(rq
, p_rq
);
4509 if (!curr
->sched_class
->yield_to_task
)
4512 if (curr
->sched_class
!= p
->sched_class
)
4515 if (task_running(p_rq
, p
) || p
->state
)
4518 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4520 schedstat_inc(rq
, yld_count
);
4522 * Make p's CPU reschedule; pick_next_entity takes care of
4525 if (preempt
&& rq
!= p_rq
)
4526 resched_task(p_rq
->curr
);
4530 double_rq_unlock(rq
, p_rq
);
4532 local_irq_restore(flags
);
4539 EXPORT_SYMBOL_GPL(yield_to
);
4542 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4543 * that process accounting knows that this is a task in IO wait state.
4545 void __sched
io_schedule(void)
4547 struct rq
*rq
= raw_rq();
4549 delayacct_blkio_start();
4550 atomic_inc(&rq
->nr_iowait
);
4551 blk_flush_plug(current
);
4552 current
->in_iowait
= 1;
4554 current
->in_iowait
= 0;
4555 atomic_dec(&rq
->nr_iowait
);
4556 delayacct_blkio_end();
4558 EXPORT_SYMBOL(io_schedule
);
4560 long __sched
io_schedule_timeout(long timeout
)
4562 struct rq
*rq
= raw_rq();
4565 delayacct_blkio_start();
4566 atomic_inc(&rq
->nr_iowait
);
4567 blk_flush_plug(current
);
4568 current
->in_iowait
= 1;
4569 ret
= schedule_timeout(timeout
);
4570 current
->in_iowait
= 0;
4571 atomic_dec(&rq
->nr_iowait
);
4572 delayacct_blkio_end();
4577 * sys_sched_get_priority_max - return maximum RT priority.
4578 * @policy: scheduling class.
4580 * this syscall returns the maximum rt_priority that can be used
4581 * by a given scheduling class.
4583 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4590 ret
= MAX_USER_RT_PRIO
-1;
4602 * sys_sched_get_priority_min - return minimum RT priority.
4603 * @policy: scheduling class.
4605 * this syscall returns the minimum rt_priority that can be used
4606 * by a given scheduling class.
4608 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4626 * sys_sched_rr_get_interval - return the default timeslice of a process.
4627 * @pid: pid of the process.
4628 * @interval: userspace pointer to the timeslice value.
4630 * this syscall writes the default timeslice value of a given process
4631 * into the user-space timespec buffer. A value of '0' means infinity.
4633 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4634 struct timespec __user
*, interval
)
4636 struct task_struct
*p
;
4637 unsigned int time_slice
;
4638 unsigned long flags
;
4648 p
= find_process_by_pid(pid
);
4652 retval
= security_task_getscheduler(p
);
4656 rq
= task_rq_lock(p
, &flags
);
4657 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4658 task_rq_unlock(rq
, p
, &flags
);
4661 jiffies_to_timespec(time_slice
, &t
);
4662 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4670 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4672 void sched_show_task(struct task_struct
*p
)
4674 unsigned long free
= 0;
4678 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4679 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4680 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4681 #if BITS_PER_LONG == 32
4682 if (state
== TASK_RUNNING
)
4683 printk(KERN_CONT
" running ");
4685 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4687 if (state
== TASK_RUNNING
)
4688 printk(KERN_CONT
" running task ");
4690 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4692 #ifdef CONFIG_DEBUG_STACK_USAGE
4693 free
= stack_not_used(p
);
4696 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4698 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4699 task_pid_nr(p
), ppid
,
4700 (unsigned long)task_thread_info(p
)->flags
);
4702 print_worker_info(KERN_INFO
, p
);
4703 show_stack(p
, NULL
);
4706 void show_state_filter(unsigned long state_filter
)
4708 struct task_struct
*g
, *p
;
4710 #if BITS_PER_LONG == 32
4712 " task PC stack pid father\n");
4715 " task PC stack pid father\n");
4718 do_each_thread(g
, p
) {
4720 * reset the NMI-timeout, listing all files on a slow
4721 * console might take a lot of time:
4723 touch_nmi_watchdog();
4724 if (!state_filter
|| (p
->state
& state_filter
))
4726 } while_each_thread(g
, p
);
4728 touch_all_softlockup_watchdogs();
4730 #ifdef CONFIG_SCHED_DEBUG
4731 sysrq_sched_debug_show();
4735 * Only show locks if all tasks are dumped:
4738 debug_show_all_locks();
4741 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4743 idle
->sched_class
= &idle_sched_class
;
4747 * init_idle - set up an idle thread for a given CPU
4748 * @idle: task in question
4749 * @cpu: cpu the idle task belongs to
4751 * NOTE: this function does not set the idle thread's NEED_RESCHED
4752 * flag, to make booting more robust.
4754 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4756 struct rq
*rq
= cpu_rq(cpu
);
4757 unsigned long flags
;
4759 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4762 idle
->state
= TASK_RUNNING
;
4763 idle
->se
.exec_start
= sched_clock();
4765 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4767 * We're having a chicken and egg problem, even though we are
4768 * holding rq->lock, the cpu isn't yet set to this cpu so the
4769 * lockdep check in task_group() will fail.
4771 * Similar case to sched_fork(). / Alternatively we could
4772 * use task_rq_lock() here and obtain the other rq->lock.
4777 __set_task_cpu(idle
, cpu
);
4780 rq
->curr
= rq
->idle
= idle
;
4781 #if defined(CONFIG_SMP)
4784 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4786 /* Set the preempt count _outside_ the spinlocks! */
4787 task_thread_info(idle
)->preempt_count
= 0;
4790 * The idle tasks have their own, simple scheduling class:
4792 idle
->sched_class
= &idle_sched_class
;
4793 ftrace_graph_init_idle_task(idle
, cpu
);
4794 vtime_init_idle(idle
, cpu
);
4795 #if defined(CONFIG_SMP)
4796 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4801 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4803 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4804 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4806 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4807 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4811 * This is how migration works:
4813 * 1) we invoke migration_cpu_stop() on the target CPU using
4815 * 2) stopper starts to run (implicitly forcing the migrated thread
4817 * 3) it checks whether the migrated task is still in the wrong runqueue.
4818 * 4) if it's in the wrong runqueue then the migration thread removes
4819 * it and puts it into the right queue.
4820 * 5) stopper completes and stop_one_cpu() returns and the migration
4825 * Change a given task's CPU affinity. Migrate the thread to a
4826 * proper CPU and schedule it away if the CPU it's executing on
4827 * is removed from the allowed bitmask.
4829 * NOTE: the caller must have a valid reference to the task, the
4830 * task must not exit() & deallocate itself prematurely. The
4831 * call is not atomic; no spinlocks may be held.
4833 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4835 unsigned long flags
;
4837 unsigned int dest_cpu
;
4840 rq
= task_rq_lock(p
, &flags
);
4842 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4845 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4850 do_set_cpus_allowed(p
, new_mask
);
4852 /* Can the task run on the task's current CPU? If so, we're done */
4853 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4856 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4858 struct migration_arg arg
= { p
, dest_cpu
};
4859 /* Need help from migration thread: drop lock and wait. */
4860 task_rq_unlock(rq
, p
, &flags
);
4861 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4862 tlb_migrate_finish(p
->mm
);
4866 task_rq_unlock(rq
, p
, &flags
);
4870 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4873 * Move (not current) task off this cpu, onto dest cpu. We're doing
4874 * this because either it can't run here any more (set_cpus_allowed()
4875 * away from this CPU, or CPU going down), or because we're
4876 * attempting to rebalance this task on exec (sched_exec).
4878 * So we race with normal scheduler movements, but that's OK, as long
4879 * as the task is no longer on this CPU.
4881 * Returns non-zero if task was successfully migrated.
4883 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4885 struct rq
*rq_dest
, *rq_src
;
4888 if (unlikely(!cpu_active(dest_cpu
)))
4891 rq_src
= cpu_rq(src_cpu
);
4892 rq_dest
= cpu_rq(dest_cpu
);
4894 raw_spin_lock(&p
->pi_lock
);
4895 double_rq_lock(rq_src
, rq_dest
);
4896 /* Already moved. */
4897 if (task_cpu(p
) != src_cpu
)
4899 /* Affinity changed (again). */
4900 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4904 * If we're not on a rq, the next wake-up will ensure we're
4908 dequeue_task(rq_src
, p
, 0);
4909 set_task_cpu(p
, dest_cpu
);
4910 enqueue_task(rq_dest
, p
, 0);
4911 check_preempt_curr(rq_dest
, p
, 0);
4916 double_rq_unlock(rq_src
, rq_dest
);
4917 raw_spin_unlock(&p
->pi_lock
);
4922 * migration_cpu_stop - this will be executed by a highprio stopper thread
4923 * and performs thread migration by bumping thread off CPU then
4924 * 'pushing' onto another runqueue.
4926 static int migration_cpu_stop(void *data
)
4928 struct migration_arg
*arg
= data
;
4931 * The original target cpu might have gone down and we might
4932 * be on another cpu but it doesn't matter.
4934 local_irq_disable();
4935 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4940 #ifdef CONFIG_HOTPLUG_CPU
4943 * Ensures that the idle task is using init_mm right before its cpu goes
4946 void idle_task_exit(void)
4948 struct mm_struct
*mm
= current
->active_mm
;
4950 BUG_ON(cpu_online(smp_processor_id()));
4953 switch_mm(mm
, &init_mm
, current
);
4958 * Since this CPU is going 'away' for a while, fold any nr_active delta
4959 * we might have. Assumes we're called after migrate_tasks() so that the
4960 * nr_active count is stable.
4962 * Also see the comment "Global load-average calculations".
4964 static void calc_load_migrate(struct rq
*rq
)
4966 long delta
= calc_load_fold_active(rq
);
4968 atomic_long_add(delta
, &calc_load_tasks
);
4972 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4973 * try_to_wake_up()->select_task_rq().
4975 * Called with rq->lock held even though we'er in stop_machine() and
4976 * there's no concurrency possible, we hold the required locks anyway
4977 * because of lock validation efforts.
4979 static void migrate_tasks(unsigned int dead_cpu
)
4981 struct rq
*rq
= cpu_rq(dead_cpu
);
4982 struct task_struct
*next
, *stop
= rq
->stop
;
4986 * Fudge the rq selection such that the below task selection loop
4987 * doesn't get stuck on the currently eligible stop task.
4989 * We're currently inside stop_machine() and the rq is either stuck
4990 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4991 * either way we should never end up calling schedule() until we're
4998 * There's this thread running, bail when that's the only
5001 if (rq
->nr_running
== 1)
5004 next
= pick_next_task(rq
);
5006 next
->sched_class
->put_prev_task(rq
, next
);
5008 /* Find suitable destination for @next, with force if needed. */
5009 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5010 raw_spin_unlock(&rq
->lock
);
5012 __migrate_task(next
, dead_cpu
, dest_cpu
);
5014 raw_spin_lock(&rq
->lock
);
5020 #endif /* CONFIG_HOTPLUG_CPU */
5022 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5024 static struct ctl_table sd_ctl_dir
[] = {
5026 .procname
= "sched_domain",
5032 static struct ctl_table sd_ctl_root
[] = {
5034 .procname
= "kernel",
5036 .child
= sd_ctl_dir
,
5041 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5043 struct ctl_table
*entry
=
5044 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5049 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5051 struct ctl_table
*entry
;
5054 * In the intermediate directories, both the child directory and
5055 * procname are dynamically allocated and could fail but the mode
5056 * will always be set. In the lowest directory the names are
5057 * static strings and all have proc handlers.
5059 for (entry
= *tablep
; entry
->mode
; entry
++) {
5061 sd_free_ctl_entry(&entry
->child
);
5062 if (entry
->proc_handler
== NULL
)
5063 kfree(entry
->procname
);
5070 static int min_load_idx
= 0;
5071 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5074 set_table_entry(struct ctl_table
*entry
,
5075 const char *procname
, void *data
, int maxlen
,
5076 umode_t mode
, proc_handler
*proc_handler
,
5079 entry
->procname
= procname
;
5081 entry
->maxlen
= maxlen
;
5083 entry
->proc_handler
= proc_handler
;
5086 entry
->extra1
= &min_load_idx
;
5087 entry
->extra2
= &max_load_idx
;
5091 static struct ctl_table
*
5092 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5094 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5099 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5100 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5101 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5102 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5103 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5104 sizeof(int), 0644, proc_dointvec_minmax
, true);
5105 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5106 sizeof(int), 0644, proc_dointvec_minmax
, true);
5107 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5108 sizeof(int), 0644, proc_dointvec_minmax
, true);
5109 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5110 sizeof(int), 0644, proc_dointvec_minmax
, true);
5111 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5112 sizeof(int), 0644, proc_dointvec_minmax
, true);
5113 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5114 sizeof(int), 0644, proc_dointvec_minmax
, false);
5115 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5116 sizeof(int), 0644, proc_dointvec_minmax
, false);
5117 set_table_entry(&table
[9], "cache_nice_tries",
5118 &sd
->cache_nice_tries
,
5119 sizeof(int), 0644, proc_dointvec_minmax
, false);
5120 set_table_entry(&table
[10], "flags", &sd
->flags
,
5121 sizeof(int), 0644, proc_dointvec_minmax
, false);
5122 set_table_entry(&table
[11], "name", sd
->name
,
5123 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5124 /* &table[12] is terminator */
5129 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5131 struct ctl_table
*entry
, *table
;
5132 struct sched_domain
*sd
;
5133 int domain_num
= 0, i
;
5136 for_each_domain(cpu
, sd
)
5138 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5143 for_each_domain(cpu
, sd
) {
5144 snprintf(buf
, 32, "domain%d", i
);
5145 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5147 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5154 static struct ctl_table_header
*sd_sysctl_header
;
5155 static void register_sched_domain_sysctl(void)
5157 int i
, cpu_num
= num_possible_cpus();
5158 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5161 WARN_ON(sd_ctl_dir
[0].child
);
5162 sd_ctl_dir
[0].child
= entry
;
5167 for_each_possible_cpu(i
) {
5168 snprintf(buf
, 32, "cpu%d", i
);
5169 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5171 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5175 WARN_ON(sd_sysctl_header
);
5176 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5179 /* may be called multiple times per register */
5180 static void unregister_sched_domain_sysctl(void)
5182 if (sd_sysctl_header
)
5183 unregister_sysctl_table(sd_sysctl_header
);
5184 sd_sysctl_header
= NULL
;
5185 if (sd_ctl_dir
[0].child
)
5186 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5189 static void register_sched_domain_sysctl(void)
5192 static void unregister_sched_domain_sysctl(void)
5197 static void set_rq_online(struct rq
*rq
)
5200 const struct sched_class
*class;
5202 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5205 for_each_class(class) {
5206 if (class->rq_online
)
5207 class->rq_online(rq
);
5212 static void set_rq_offline(struct rq
*rq
)
5215 const struct sched_class
*class;
5217 for_each_class(class) {
5218 if (class->rq_offline
)
5219 class->rq_offline(rq
);
5222 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5228 * migration_call - callback that gets triggered when a CPU is added.
5229 * Here we can start up the necessary migration thread for the new CPU.
5231 static int __cpuinit
5232 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5234 int cpu
= (long)hcpu
;
5235 unsigned long flags
;
5236 struct rq
*rq
= cpu_rq(cpu
);
5238 switch (action
& ~CPU_TASKS_FROZEN
) {
5240 case CPU_UP_PREPARE
:
5241 rq
->calc_load_update
= calc_load_update
;
5242 account_reset_rq(rq
);
5246 /* Update our root-domain */
5247 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5249 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5253 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5256 #ifdef CONFIG_HOTPLUG_CPU
5258 sched_ttwu_pending();
5259 /* Update our root-domain */
5260 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5262 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5266 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5267 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5271 calc_load_migrate(rq
);
5276 update_max_interval();
5282 * Register at high priority so that task migration (migrate_all_tasks)
5283 * happens before everything else. This has to be lower priority than
5284 * the notifier in the perf_event subsystem, though.
5286 static struct notifier_block __cpuinitdata migration_notifier
= {
5287 .notifier_call
= migration_call
,
5288 .priority
= CPU_PRI_MIGRATION
,
5291 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5292 unsigned long action
, void *hcpu
)
5294 switch (action
& ~CPU_TASKS_FROZEN
) {
5295 case CPU_DOWN_FAILED
:
5296 set_cpu_active((long)hcpu
, true);
5303 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5304 unsigned long action
, void *hcpu
)
5306 switch (action
& ~CPU_TASKS_FROZEN
) {
5307 case CPU_DOWN_PREPARE
:
5308 set_cpu_active((long)hcpu
, false);
5315 static int __init
migration_init(void)
5317 void *cpu
= (void *)(long)smp_processor_id();
5320 /* Initialize migration for the boot CPU */
5321 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5322 BUG_ON(err
== NOTIFY_BAD
);
5323 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5324 register_cpu_notifier(&migration_notifier
);
5326 /* Register cpu active notifiers */
5327 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5328 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5332 early_initcall(migration_init
);
5337 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5339 #ifdef CONFIG_SCHED_DEBUG
5341 static __read_mostly
int sched_debug_enabled
;
5343 static int __init
sched_debug_setup(char *str
)
5345 sched_debug_enabled
= 1;
5349 early_param("sched_debug", sched_debug_setup
);
5351 static inline bool sched_debug(void)
5353 return sched_debug_enabled
;
5356 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5357 struct cpumask
*groupmask
)
5359 struct sched_group
*group
= sd
->groups
;
5362 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5363 cpumask_clear(groupmask
);
5365 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5367 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5368 printk("does not load-balance\n");
5370 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5375 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5377 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5378 printk(KERN_ERR
"ERROR: domain->span does not contain "
5381 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5382 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5386 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5390 printk(KERN_ERR
"ERROR: group is NULL\n");
5395 * Even though we initialize ->power to something semi-sane,
5396 * we leave power_orig unset. This allows us to detect if
5397 * domain iteration is still funny without causing /0 traps.
5399 if (!group
->sgp
->power_orig
) {
5400 printk(KERN_CONT
"\n");
5401 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5406 if (!cpumask_weight(sched_group_cpus(group
))) {
5407 printk(KERN_CONT
"\n");
5408 printk(KERN_ERR
"ERROR: empty group\n");
5412 if (!(sd
->flags
& SD_OVERLAP
) &&
5413 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5414 printk(KERN_CONT
"\n");
5415 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5419 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5421 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5423 printk(KERN_CONT
" %s", str
);
5424 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5425 printk(KERN_CONT
" (cpu_power = %d)",
5429 group
= group
->next
;
5430 } while (group
!= sd
->groups
);
5431 printk(KERN_CONT
"\n");
5433 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5434 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5437 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5438 printk(KERN_ERR
"ERROR: parent span is not a superset "
5439 "of domain->span\n");
5443 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5447 if (!sched_debug_enabled
)
5451 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5455 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5458 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5466 #else /* !CONFIG_SCHED_DEBUG */
5467 # define sched_domain_debug(sd, cpu) do { } while (0)
5468 static inline bool sched_debug(void)
5472 #endif /* CONFIG_SCHED_DEBUG */
5474 static int sd_degenerate(struct sched_domain
*sd
)
5476 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5479 /* Following flags need at least 2 groups */
5480 if (sd
->flags
& (SD_LOAD_BALANCE
|
5481 SD_BALANCE_NEWIDLE
|
5485 SD_SHARE_PKG_RESOURCES
)) {
5486 if (sd
->groups
!= sd
->groups
->next
)
5490 /* Following flags don't use groups */
5491 if (sd
->flags
& (SD_WAKE_AFFINE
))
5498 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5500 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5502 if (sd_degenerate(parent
))
5505 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5508 /* Flags needing groups don't count if only 1 group in parent */
5509 if (parent
->groups
== parent
->groups
->next
) {
5510 pflags
&= ~(SD_LOAD_BALANCE
|
5511 SD_BALANCE_NEWIDLE
|
5515 SD_SHARE_PKG_RESOURCES
);
5516 if (nr_node_ids
== 1)
5517 pflags
&= ~SD_SERIALIZE
;
5519 if (~cflags
& pflags
)
5525 static void free_rootdomain(struct rcu_head
*rcu
)
5527 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5529 cpupri_cleanup(&rd
->cpupri
);
5530 free_cpumask_var(rd
->rto_mask
);
5531 free_cpumask_var(rd
->online
);
5532 free_cpumask_var(rd
->span
);
5536 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5538 struct root_domain
*old_rd
= NULL
;
5539 unsigned long flags
;
5541 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5546 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5549 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5552 * If we dont want to free the old_rt yet then
5553 * set old_rd to NULL to skip the freeing later
5556 if (!atomic_dec_and_test(&old_rd
->refcount
))
5560 atomic_inc(&rd
->refcount
);
5563 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5564 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5567 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5570 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5573 static int init_rootdomain(struct root_domain
*rd
)
5575 memset(rd
, 0, sizeof(*rd
));
5577 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5579 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5581 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5584 if (cpupri_init(&rd
->cpupri
) != 0)
5589 free_cpumask_var(rd
->rto_mask
);
5591 free_cpumask_var(rd
->online
);
5593 free_cpumask_var(rd
->span
);
5599 * By default the system creates a single root-domain with all cpus as
5600 * members (mimicking the global state we have today).
5602 struct root_domain def_root_domain
;
5604 static void init_defrootdomain(void)
5606 init_rootdomain(&def_root_domain
);
5608 atomic_set(&def_root_domain
.refcount
, 1);
5611 static struct root_domain
*alloc_rootdomain(void)
5613 struct root_domain
*rd
;
5615 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5619 if (init_rootdomain(rd
) != 0) {
5627 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5629 struct sched_group
*tmp
, *first
;
5638 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5643 } while (sg
!= first
);
5646 static void free_sched_domain(struct rcu_head
*rcu
)
5648 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5651 * If its an overlapping domain it has private groups, iterate and
5654 if (sd
->flags
& SD_OVERLAP
) {
5655 free_sched_groups(sd
->groups
, 1);
5656 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5657 kfree(sd
->groups
->sgp
);
5663 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5665 call_rcu(&sd
->rcu
, free_sched_domain
);
5668 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5670 for (; sd
; sd
= sd
->parent
)
5671 destroy_sched_domain(sd
, cpu
);
5675 * Keep a special pointer to the highest sched_domain that has
5676 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5677 * allows us to avoid some pointer chasing select_idle_sibling().
5679 * Also keep a unique ID per domain (we use the first cpu number in
5680 * the cpumask of the domain), this allows us to quickly tell if
5681 * two cpus are in the same cache domain, see cpus_share_cache().
5683 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5684 DEFINE_PER_CPU(int, sd_llc_id
);
5686 static void update_top_cache_domain(int cpu
)
5688 struct sched_domain
*sd
;
5691 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5693 id
= cpumask_first(sched_domain_span(sd
));
5695 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5696 per_cpu(sd_llc_id
, cpu
) = id
;
5700 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5701 * hold the hotplug lock.
5704 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5706 struct rq
*rq
= cpu_rq(cpu
);
5707 struct sched_domain
*tmp
;
5709 /* Remove the sched domains which do not contribute to scheduling. */
5710 for (tmp
= sd
; tmp
; ) {
5711 struct sched_domain
*parent
= tmp
->parent
;
5715 if (sd_parent_degenerate(tmp
, parent
)) {
5716 tmp
->parent
= parent
->parent
;
5718 parent
->parent
->child
= tmp
;
5719 destroy_sched_domain(parent
, cpu
);
5724 if (sd
&& sd_degenerate(sd
)) {
5727 destroy_sched_domain(tmp
, cpu
);
5732 sched_domain_debug(sd
, cpu
);
5734 rq_attach_root(rq
, rd
);
5736 rcu_assign_pointer(rq
->sd
, sd
);
5737 destroy_sched_domains(tmp
, cpu
);
5739 update_top_cache_domain(cpu
);
5742 /* cpus with isolated domains */
5743 static cpumask_var_t cpu_isolated_map
;
5745 /* Setup the mask of cpus configured for isolated domains */
5746 static int __init
isolated_cpu_setup(char *str
)
5748 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5749 cpulist_parse(str
, cpu_isolated_map
);
5753 __setup("isolcpus=", isolated_cpu_setup
);
5755 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5757 return cpumask_of_node(cpu_to_node(cpu
));
5761 struct sched_domain
**__percpu sd
;
5762 struct sched_group
**__percpu sg
;
5763 struct sched_group_power
**__percpu sgp
;
5767 struct sched_domain
** __percpu sd
;
5768 struct root_domain
*rd
;
5778 struct sched_domain_topology_level
;
5780 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5781 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5783 #define SDTL_OVERLAP 0x01
5785 struct sched_domain_topology_level
{
5786 sched_domain_init_f init
;
5787 sched_domain_mask_f mask
;
5790 struct sd_data data
;
5794 * Build an iteration mask that can exclude certain CPUs from the upwards
5797 * Asymmetric node setups can result in situations where the domain tree is of
5798 * unequal depth, make sure to skip domains that already cover the entire
5801 * In that case build_sched_domains() will have terminated the iteration early
5802 * and our sibling sd spans will be empty. Domains should always include the
5803 * cpu they're built on, so check that.
5806 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5808 const struct cpumask
*span
= sched_domain_span(sd
);
5809 struct sd_data
*sdd
= sd
->private;
5810 struct sched_domain
*sibling
;
5813 for_each_cpu(i
, span
) {
5814 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5815 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5818 cpumask_set_cpu(i
, sched_group_mask(sg
));
5823 * Return the canonical balance cpu for this group, this is the first cpu
5824 * of this group that's also in the iteration mask.
5826 int group_balance_cpu(struct sched_group
*sg
)
5828 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5832 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5834 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5835 const struct cpumask
*span
= sched_domain_span(sd
);
5836 struct cpumask
*covered
= sched_domains_tmpmask
;
5837 struct sd_data
*sdd
= sd
->private;
5838 struct sched_domain
*child
;
5841 cpumask_clear(covered
);
5843 for_each_cpu(i
, span
) {
5844 struct cpumask
*sg_span
;
5846 if (cpumask_test_cpu(i
, covered
))
5849 child
= *per_cpu_ptr(sdd
->sd
, i
);
5851 /* See the comment near build_group_mask(). */
5852 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5855 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5856 GFP_KERNEL
, cpu_to_node(cpu
));
5861 sg_span
= sched_group_cpus(sg
);
5863 child
= child
->child
;
5864 cpumask_copy(sg_span
, sched_domain_span(child
));
5866 cpumask_set_cpu(i
, sg_span
);
5868 cpumask_or(covered
, covered
, sg_span
);
5870 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5871 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5872 build_group_mask(sd
, sg
);
5875 * Initialize sgp->power such that even if we mess up the
5876 * domains and no possible iteration will get us here, we won't
5879 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5882 * Make sure the first group of this domain contains the
5883 * canonical balance cpu. Otherwise the sched_domain iteration
5884 * breaks. See update_sg_lb_stats().
5886 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5887 group_balance_cpu(sg
) == cpu
)
5897 sd
->groups
= groups
;
5902 free_sched_groups(first
, 0);
5907 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5909 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5910 struct sched_domain
*child
= sd
->child
;
5913 cpu
= cpumask_first(sched_domain_span(child
));
5916 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5917 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5918 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5925 * build_sched_groups will build a circular linked list of the groups
5926 * covered by the given span, and will set each group's ->cpumask correctly,
5927 * and ->cpu_power to 0.
5929 * Assumes the sched_domain tree is fully constructed
5932 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5934 struct sched_group
*first
= NULL
, *last
= NULL
;
5935 struct sd_data
*sdd
= sd
->private;
5936 const struct cpumask
*span
= sched_domain_span(sd
);
5937 struct cpumask
*covered
;
5940 get_group(cpu
, sdd
, &sd
->groups
);
5941 atomic_inc(&sd
->groups
->ref
);
5943 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
5946 lockdep_assert_held(&sched_domains_mutex
);
5947 covered
= sched_domains_tmpmask
;
5949 cpumask_clear(covered
);
5951 for_each_cpu(i
, span
) {
5952 struct sched_group
*sg
;
5953 int group
= get_group(i
, sdd
, &sg
);
5956 if (cpumask_test_cpu(i
, covered
))
5959 cpumask_clear(sched_group_cpus(sg
));
5961 cpumask_setall(sched_group_mask(sg
));
5963 for_each_cpu(j
, span
) {
5964 if (get_group(j
, sdd
, NULL
) != group
)
5967 cpumask_set_cpu(j
, covered
);
5968 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5983 * Initialize sched groups cpu_power.
5985 * cpu_power indicates the capacity of sched group, which is used while
5986 * distributing the load between different sched groups in a sched domain.
5987 * Typically cpu_power for all the groups in a sched domain will be same unless
5988 * there are asymmetries in the topology. If there are asymmetries, group
5989 * having more cpu_power will pickup more load compared to the group having
5992 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5994 struct sched_group
*sg
= sd
->groups
;
5996 WARN_ON(!sd
|| !sg
);
5999 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6001 } while (sg
!= sd
->groups
);
6003 if (cpu
!= group_balance_cpu(sg
))
6006 update_group_power(sd
, cpu
);
6007 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
6010 int __weak
arch_sd_sibling_asym_packing(void)
6012 return 0*SD_ASYM_PACKING
;
6016 * Initializers for schedule domains
6017 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6020 #ifdef CONFIG_SCHED_DEBUG
6021 # define SD_INIT_NAME(sd, type) sd->name = #type
6023 # define SD_INIT_NAME(sd, type) do { } while (0)
6026 #define SD_INIT_FUNC(type) \
6027 static noinline struct sched_domain * \
6028 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6030 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6031 *sd = SD_##type##_INIT; \
6032 SD_INIT_NAME(sd, type); \
6033 sd->private = &tl->data; \
6038 #ifdef CONFIG_SCHED_SMT
6039 SD_INIT_FUNC(SIBLING
)
6041 #ifdef CONFIG_SCHED_MC
6044 #ifdef CONFIG_SCHED_BOOK
6048 static int default_relax_domain_level
= -1;
6049 int sched_domain_level_max
;
6051 static int __init
setup_relax_domain_level(char *str
)
6053 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6054 pr_warn("Unable to set relax_domain_level\n");
6058 __setup("relax_domain_level=", setup_relax_domain_level
);
6060 static void set_domain_attribute(struct sched_domain
*sd
,
6061 struct sched_domain_attr
*attr
)
6065 if (!attr
|| attr
->relax_domain_level
< 0) {
6066 if (default_relax_domain_level
< 0)
6069 request
= default_relax_domain_level
;
6071 request
= attr
->relax_domain_level
;
6072 if (request
< sd
->level
) {
6073 /* turn off idle balance on this domain */
6074 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6076 /* turn on idle balance on this domain */
6077 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6081 static void __sdt_free(const struct cpumask
*cpu_map
);
6082 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6084 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6085 const struct cpumask
*cpu_map
)
6089 if (!atomic_read(&d
->rd
->refcount
))
6090 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6092 free_percpu(d
->sd
); /* fall through */
6094 __sdt_free(cpu_map
); /* fall through */
6100 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6101 const struct cpumask
*cpu_map
)
6103 memset(d
, 0, sizeof(*d
));
6105 if (__sdt_alloc(cpu_map
))
6106 return sa_sd_storage
;
6107 d
->sd
= alloc_percpu(struct sched_domain
*);
6109 return sa_sd_storage
;
6110 d
->rd
= alloc_rootdomain();
6113 return sa_rootdomain
;
6117 * NULL the sd_data elements we've used to build the sched_domain and
6118 * sched_group structure so that the subsequent __free_domain_allocs()
6119 * will not free the data we're using.
6121 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6123 struct sd_data
*sdd
= sd
->private;
6125 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6126 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6128 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6129 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6131 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
6132 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
6135 #ifdef CONFIG_SCHED_SMT
6136 static const struct cpumask
*cpu_smt_mask(int cpu
)
6138 return topology_thread_cpumask(cpu
);
6143 * Topology list, bottom-up.
6145 static struct sched_domain_topology_level default_topology
[] = {
6146 #ifdef CONFIG_SCHED_SMT
6147 { sd_init_SIBLING
, cpu_smt_mask
, },
6149 #ifdef CONFIG_SCHED_MC
6150 { sd_init_MC
, cpu_coregroup_mask
, },
6152 #ifdef CONFIG_SCHED_BOOK
6153 { sd_init_BOOK
, cpu_book_mask
, },
6155 { sd_init_CPU
, cpu_cpu_mask
, },
6159 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6163 static int sched_domains_numa_levels
;
6164 static int *sched_domains_numa_distance
;
6165 static struct cpumask
***sched_domains_numa_masks
;
6166 static int sched_domains_curr_level
;
6168 static inline int sd_local_flags(int level
)
6170 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6173 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6176 static struct sched_domain
*
6177 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6179 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6180 int level
= tl
->numa_level
;
6181 int sd_weight
= cpumask_weight(
6182 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6184 *sd
= (struct sched_domain
){
6185 .min_interval
= sd_weight
,
6186 .max_interval
= 2*sd_weight
,
6188 .imbalance_pct
= 125,
6189 .cache_nice_tries
= 2,
6196 .flags
= 1*SD_LOAD_BALANCE
6197 | 1*SD_BALANCE_NEWIDLE
6202 | 0*SD_SHARE_CPUPOWER
6203 | 0*SD_SHARE_PKG_RESOURCES
6205 | 0*SD_PREFER_SIBLING
6206 | sd_local_flags(level
)
6208 .last_balance
= jiffies
,
6209 .balance_interval
= sd_weight
,
6211 SD_INIT_NAME(sd
, NUMA
);
6212 sd
->private = &tl
->data
;
6215 * Ugly hack to pass state to sd_numa_mask()...
6217 sched_domains_curr_level
= tl
->numa_level
;
6222 static const struct cpumask
*sd_numa_mask(int cpu
)
6224 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6227 static void sched_numa_warn(const char *str
)
6229 static int done
= false;
6237 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6239 for (i
= 0; i
< nr_node_ids
; i
++) {
6240 printk(KERN_WARNING
" ");
6241 for (j
= 0; j
< nr_node_ids
; j
++)
6242 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6243 printk(KERN_CONT
"\n");
6245 printk(KERN_WARNING
"\n");
6248 static bool find_numa_distance(int distance
)
6252 if (distance
== node_distance(0, 0))
6255 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6256 if (sched_domains_numa_distance
[i
] == distance
)
6263 static void sched_init_numa(void)
6265 int next_distance
, curr_distance
= node_distance(0, 0);
6266 struct sched_domain_topology_level
*tl
;
6270 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6271 if (!sched_domains_numa_distance
)
6275 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6276 * unique distances in the node_distance() table.
6278 * Assumes node_distance(0,j) includes all distances in
6279 * node_distance(i,j) in order to avoid cubic time.
6281 next_distance
= curr_distance
;
6282 for (i
= 0; i
< nr_node_ids
; i
++) {
6283 for (j
= 0; j
< nr_node_ids
; j
++) {
6284 for (k
= 0; k
< nr_node_ids
; k
++) {
6285 int distance
= node_distance(i
, k
);
6287 if (distance
> curr_distance
&&
6288 (distance
< next_distance
||
6289 next_distance
== curr_distance
))
6290 next_distance
= distance
;
6293 * While not a strong assumption it would be nice to know
6294 * about cases where if node A is connected to B, B is not
6295 * equally connected to A.
6297 if (sched_debug() && node_distance(k
, i
) != distance
)
6298 sched_numa_warn("Node-distance not symmetric");
6300 if (sched_debug() && i
&& !find_numa_distance(distance
))
6301 sched_numa_warn("Node-0 not representative");
6303 if (next_distance
!= curr_distance
) {
6304 sched_domains_numa_distance
[level
++] = next_distance
;
6305 sched_domains_numa_levels
= level
;
6306 curr_distance
= next_distance
;
6311 * In case of sched_debug() we verify the above assumption.
6317 * 'level' contains the number of unique distances, excluding the
6318 * identity distance node_distance(i,i).
6320 * The sched_domains_numa_distance[] array includes the actual distance
6325 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6326 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6327 * the array will contain less then 'level' members. This could be
6328 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6329 * in other functions.
6331 * We reset it to 'level' at the end of this function.
6333 sched_domains_numa_levels
= 0;
6335 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6336 if (!sched_domains_numa_masks
)
6340 * Now for each level, construct a mask per node which contains all
6341 * cpus of nodes that are that many hops away from us.
6343 for (i
= 0; i
< level
; i
++) {
6344 sched_domains_numa_masks
[i
] =
6345 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6346 if (!sched_domains_numa_masks
[i
])
6349 for (j
= 0; j
< nr_node_ids
; j
++) {
6350 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6354 sched_domains_numa_masks
[i
][j
] = mask
;
6356 for (k
= 0; k
< nr_node_ids
; k
++) {
6357 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6360 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6365 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6366 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6371 * Copy the default topology bits..
6373 for (i
= 0; default_topology
[i
].init
; i
++)
6374 tl
[i
] = default_topology
[i
];
6377 * .. and append 'j' levels of NUMA goodness.
6379 for (j
= 0; j
< level
; i
++, j
++) {
6380 tl
[i
] = (struct sched_domain_topology_level
){
6381 .init
= sd_numa_init
,
6382 .mask
= sd_numa_mask
,
6383 .flags
= SDTL_OVERLAP
,
6388 sched_domain_topology
= tl
;
6390 sched_domains_numa_levels
= level
;
6393 static void sched_domains_numa_masks_set(int cpu
)
6396 int node
= cpu_to_node(cpu
);
6398 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6399 for (j
= 0; j
< nr_node_ids
; j
++) {
6400 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6401 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6406 static void sched_domains_numa_masks_clear(int cpu
)
6409 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6410 for (j
= 0; j
< nr_node_ids
; j
++)
6411 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6416 * Update sched_domains_numa_masks[level][node] array when new cpus
6419 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6420 unsigned long action
,
6423 int cpu
= (long)hcpu
;
6425 switch (action
& ~CPU_TASKS_FROZEN
) {
6427 sched_domains_numa_masks_set(cpu
);
6431 sched_domains_numa_masks_clear(cpu
);
6441 static inline void sched_init_numa(void)
6445 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6446 unsigned long action
,
6451 #endif /* CONFIG_NUMA */
6453 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6455 struct sched_domain_topology_level
*tl
;
6458 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6459 struct sd_data
*sdd
= &tl
->data
;
6461 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6465 sdd
->sg
= alloc_percpu(struct sched_group
*);
6469 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6473 for_each_cpu(j
, cpu_map
) {
6474 struct sched_domain
*sd
;
6475 struct sched_group
*sg
;
6476 struct sched_group_power
*sgp
;
6478 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6479 GFP_KERNEL
, cpu_to_node(j
));
6483 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6485 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6486 GFP_KERNEL
, cpu_to_node(j
));
6492 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6494 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6495 GFP_KERNEL
, cpu_to_node(j
));
6499 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6506 static void __sdt_free(const struct cpumask
*cpu_map
)
6508 struct sched_domain_topology_level
*tl
;
6511 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6512 struct sd_data
*sdd
= &tl
->data
;
6514 for_each_cpu(j
, cpu_map
) {
6515 struct sched_domain
*sd
;
6518 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6519 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6520 free_sched_groups(sd
->groups
, 0);
6521 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6525 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6527 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6529 free_percpu(sdd
->sd
);
6531 free_percpu(sdd
->sg
);
6533 free_percpu(sdd
->sgp
);
6538 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6539 struct s_data
*d
, const struct cpumask
*cpu_map
,
6540 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
6543 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6547 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6549 sd
->level
= child
->level
+ 1;
6550 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6554 set_domain_attribute(sd
, attr
);
6560 * Build sched domains for a given set of cpus and attach the sched domains
6561 * to the individual cpus
6563 static int build_sched_domains(const struct cpumask
*cpu_map
,
6564 struct sched_domain_attr
*attr
)
6566 enum s_alloc alloc_state
= sa_none
;
6567 struct sched_domain
*sd
;
6569 int i
, ret
= -ENOMEM
;
6571 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6572 if (alloc_state
!= sa_rootdomain
)
6575 /* Set up domains for cpus specified by the cpu_map. */
6576 for_each_cpu(i
, cpu_map
) {
6577 struct sched_domain_topology_level
*tl
;
6580 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
6581 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
6582 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6583 sd
->flags
|= SD_OVERLAP
;
6584 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6591 *per_cpu_ptr(d
.sd
, i
) = sd
;
6594 /* Build the groups for the domains */
6595 for_each_cpu(i
, cpu_map
) {
6596 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6597 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6598 if (sd
->flags
& SD_OVERLAP
) {
6599 if (build_overlap_sched_groups(sd
, i
))
6602 if (build_sched_groups(sd
, i
))
6608 /* Calculate CPU power for physical packages and nodes */
6609 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6610 if (!cpumask_test_cpu(i
, cpu_map
))
6613 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6614 claim_allocations(i
, sd
);
6615 init_sched_groups_power(i
, sd
);
6619 /* Attach the domains */
6621 for_each_cpu(i
, cpu_map
) {
6622 sd
= *per_cpu_ptr(d
.sd
, i
);
6623 cpu_attach_domain(sd
, d
.rd
, i
);
6629 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6633 static cpumask_var_t
*doms_cur
; /* current sched domains */
6634 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6635 static struct sched_domain_attr
*dattr_cur
;
6636 /* attribues of custom domains in 'doms_cur' */
6639 * Special case: If a kmalloc of a doms_cur partition (array of
6640 * cpumask) fails, then fallback to a single sched domain,
6641 * as determined by the single cpumask fallback_doms.
6643 static cpumask_var_t fallback_doms
;
6646 * arch_update_cpu_topology lets virtualized architectures update the
6647 * cpu core maps. It is supposed to return 1 if the topology changed
6648 * or 0 if it stayed the same.
6650 int __attribute__((weak
)) arch_update_cpu_topology(void)
6655 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6658 cpumask_var_t
*doms
;
6660 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6663 for (i
= 0; i
< ndoms
; i
++) {
6664 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6665 free_sched_domains(doms
, i
);
6672 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6675 for (i
= 0; i
< ndoms
; i
++)
6676 free_cpumask_var(doms
[i
]);
6681 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6682 * For now this just excludes isolated cpus, but could be used to
6683 * exclude other special cases in the future.
6685 static int init_sched_domains(const struct cpumask
*cpu_map
)
6689 arch_update_cpu_topology();
6691 doms_cur
= alloc_sched_domains(ndoms_cur
);
6693 doms_cur
= &fallback_doms
;
6694 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6695 err
= build_sched_domains(doms_cur
[0], NULL
);
6696 register_sched_domain_sysctl();
6702 * Detach sched domains from a group of cpus specified in cpu_map
6703 * These cpus will now be attached to the NULL domain
6705 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6710 for_each_cpu(i
, cpu_map
)
6711 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6715 /* handle null as "default" */
6716 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6717 struct sched_domain_attr
*new, int idx_new
)
6719 struct sched_domain_attr tmp
;
6726 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6727 new ? (new + idx_new
) : &tmp
,
6728 sizeof(struct sched_domain_attr
));
6732 * Partition sched domains as specified by the 'ndoms_new'
6733 * cpumasks in the array doms_new[] of cpumasks. This compares
6734 * doms_new[] to the current sched domain partitioning, doms_cur[].
6735 * It destroys each deleted domain and builds each new domain.
6737 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6738 * The masks don't intersect (don't overlap.) We should setup one
6739 * sched domain for each mask. CPUs not in any of the cpumasks will
6740 * not be load balanced. If the same cpumask appears both in the
6741 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6744 * The passed in 'doms_new' should be allocated using
6745 * alloc_sched_domains. This routine takes ownership of it and will
6746 * free_sched_domains it when done with it. If the caller failed the
6747 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6748 * and partition_sched_domains() will fallback to the single partition
6749 * 'fallback_doms', it also forces the domains to be rebuilt.
6751 * If doms_new == NULL it will be replaced with cpu_online_mask.
6752 * ndoms_new == 0 is a special case for destroying existing domains,
6753 * and it will not create the default domain.
6755 * Call with hotplug lock held
6757 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6758 struct sched_domain_attr
*dattr_new
)
6763 mutex_lock(&sched_domains_mutex
);
6765 /* always unregister in case we don't destroy any domains */
6766 unregister_sched_domain_sysctl();
6768 /* Let architecture update cpu core mappings. */
6769 new_topology
= arch_update_cpu_topology();
6771 n
= doms_new
? ndoms_new
: 0;
6773 /* Destroy deleted domains */
6774 for (i
= 0; i
< ndoms_cur
; i
++) {
6775 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6776 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6777 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6780 /* no match - a current sched domain not in new doms_new[] */
6781 detach_destroy_domains(doms_cur
[i
]);
6786 if (doms_new
== NULL
) {
6788 doms_new
= &fallback_doms
;
6789 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6790 WARN_ON_ONCE(dattr_new
);
6793 /* Build new domains */
6794 for (i
= 0; i
< ndoms_new
; i
++) {
6795 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6796 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6797 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6800 /* no match - add a new doms_new */
6801 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6806 /* Remember the new sched domains */
6807 if (doms_cur
!= &fallback_doms
)
6808 free_sched_domains(doms_cur
, ndoms_cur
);
6809 kfree(dattr_cur
); /* kfree(NULL) is safe */
6810 doms_cur
= doms_new
;
6811 dattr_cur
= dattr_new
;
6812 ndoms_cur
= ndoms_new
;
6814 register_sched_domain_sysctl();
6816 mutex_unlock(&sched_domains_mutex
);
6819 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6822 * Update cpusets according to cpu_active mask. If cpusets are
6823 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6824 * around partition_sched_domains().
6826 * If we come here as part of a suspend/resume, don't touch cpusets because we
6827 * want to restore it back to its original state upon resume anyway.
6829 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6833 case CPU_ONLINE_FROZEN
:
6834 case CPU_DOWN_FAILED_FROZEN
:
6837 * num_cpus_frozen tracks how many CPUs are involved in suspend
6838 * resume sequence. As long as this is not the last online
6839 * operation in the resume sequence, just build a single sched
6840 * domain, ignoring cpusets.
6843 if (likely(num_cpus_frozen
)) {
6844 partition_sched_domains(1, NULL
, NULL
);
6849 * This is the last CPU online operation. So fall through and
6850 * restore the original sched domains by considering the
6851 * cpuset configurations.
6855 case CPU_DOWN_FAILED
:
6856 cpuset_update_active_cpus(true);
6864 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6868 case CPU_DOWN_PREPARE
:
6869 cpuset_update_active_cpus(false);
6871 case CPU_DOWN_PREPARE_FROZEN
:
6873 partition_sched_domains(1, NULL
, NULL
);
6881 void __init
sched_init_smp(void)
6883 cpumask_var_t non_isolated_cpus
;
6885 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6886 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6891 mutex_lock(&sched_domains_mutex
);
6892 init_sched_domains(cpu_active_mask
);
6893 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6894 if (cpumask_empty(non_isolated_cpus
))
6895 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6896 mutex_unlock(&sched_domains_mutex
);
6899 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6900 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6901 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6903 /* RT runtime code needs to handle some hotplug events */
6904 hotcpu_notifier(update_runtime
, 0);
6908 /* Move init over to a non-isolated CPU */
6909 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6911 sched_init_granularity();
6912 free_cpumask_var(non_isolated_cpus
);
6914 init_sched_rt_class();
6917 void __init
sched_init_smp(void)
6919 sched_init_granularity();
6921 #endif /* CONFIG_SMP */
6923 const_debug
unsigned int sysctl_timer_migration
= 1;
6925 int in_sched_functions(unsigned long addr
)
6927 return in_lock_functions(addr
) ||
6928 (addr
>= (unsigned long)__sched_text_start
6929 && addr
< (unsigned long)__sched_text_end
);
6932 #ifdef CONFIG_CGROUP_SCHED
6934 * Default task group.
6935 * Every task in system belongs to this group at bootup.
6937 struct task_group root_task_group
;
6938 LIST_HEAD(task_groups
);
6941 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6943 void __init
sched_init(void)
6946 unsigned long alloc_size
= 0, ptr
;
6948 #ifdef CONFIG_FAIR_GROUP_SCHED
6949 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6951 #ifdef CONFIG_RT_GROUP_SCHED
6952 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6954 #ifdef CONFIG_CPUMASK_OFFSTACK
6955 alloc_size
+= num_possible_cpus() * cpumask_size();
6958 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6960 #ifdef CONFIG_FAIR_GROUP_SCHED
6961 root_task_group
.se
= (struct sched_entity
**)ptr
;
6962 ptr
+= nr_cpu_ids
* sizeof(void **);
6964 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6965 ptr
+= nr_cpu_ids
* sizeof(void **);
6967 #endif /* CONFIG_FAIR_GROUP_SCHED */
6968 #ifdef CONFIG_RT_GROUP_SCHED
6969 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6970 ptr
+= nr_cpu_ids
* sizeof(void **);
6972 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6973 ptr
+= nr_cpu_ids
* sizeof(void **);
6975 #endif /* CONFIG_RT_GROUP_SCHED */
6976 #ifdef CONFIG_CPUMASK_OFFSTACK
6977 for_each_possible_cpu(i
) {
6978 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6979 ptr
+= cpumask_size();
6981 #endif /* CONFIG_CPUMASK_OFFSTACK */
6985 init_defrootdomain();
6988 init_rt_bandwidth(&def_rt_bandwidth
,
6989 global_rt_period(), global_rt_runtime());
6991 #ifdef CONFIG_RT_GROUP_SCHED
6992 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6993 global_rt_period(), global_rt_runtime());
6994 #endif /* CONFIG_RT_GROUP_SCHED */
6996 #ifdef CONFIG_CGROUP_SCHED
6997 list_add(&root_task_group
.list
, &task_groups
);
6998 INIT_LIST_HEAD(&root_task_group
.children
);
6999 INIT_LIST_HEAD(&root_task_group
.siblings
);
7000 autogroup_init(&init_task
);
7002 #endif /* CONFIG_CGROUP_SCHED */
7004 for_each_possible_cpu(i
) {
7008 raw_spin_lock_init(&rq
->lock
);
7010 rq
->calc_load_active
= 0;
7011 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7012 init_cfs_rq(&rq
->cfs
);
7013 init_rt_rq(&rq
->rt
, rq
);
7014 #ifdef CONFIG_FAIR_GROUP_SCHED
7015 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7016 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7018 * How much cpu bandwidth does root_task_group get?
7020 * In case of task-groups formed thr' the cgroup filesystem, it
7021 * gets 100% of the cpu resources in the system. This overall
7022 * system cpu resource is divided among the tasks of
7023 * root_task_group and its child task-groups in a fair manner,
7024 * based on each entity's (task or task-group's) weight
7025 * (se->load.weight).
7027 * In other words, if root_task_group has 10 tasks of weight
7028 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7029 * then A0's share of the cpu resource is:
7031 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7033 * We achieve this by letting root_task_group's tasks sit
7034 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7036 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7037 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7038 #endif /* CONFIG_FAIR_GROUP_SCHED */
7040 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7041 #ifdef CONFIG_RT_GROUP_SCHED
7042 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7043 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7046 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7047 rq
->cpu_load
[j
] = 0;
7049 rq
->last_load_update_tick
= jiffies
;
7054 rq
->cpu_power
= SCHED_POWER_SCALE
;
7055 rq
->post_schedule
= 0;
7056 rq
->active_balance
= 0;
7057 rq
->next_balance
= jiffies
;
7062 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7064 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7066 rq_attach_root(rq
, &def_root_domain
);
7067 #ifdef CONFIG_NO_HZ_COMMON
7070 #ifdef CONFIG_NO_HZ_FULL
7071 rq
->last_sched_tick
= 0;
7075 atomic_set(&rq
->nr_iowait
, 0);
7078 set_load_weight(&init_task
);
7080 #ifdef CONFIG_PREEMPT_NOTIFIERS
7081 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7084 #ifdef CONFIG_RT_MUTEXES
7085 plist_head_init(&init_task
.pi_waiters
);
7089 * The boot idle thread does lazy MMU switching as well:
7091 atomic_inc(&init_mm
.mm_count
);
7092 enter_lazy_tlb(&init_mm
, current
);
7095 * Make us the idle thread. Technically, schedule() should not be
7096 * called from this thread, however somewhere below it might be,
7097 * but because we are the idle thread, we just pick up running again
7098 * when this runqueue becomes "idle".
7100 init_idle(current
, smp_processor_id());
7102 calc_load_update
= jiffies
+ LOAD_FREQ
;
7105 * During early bootup we pretend to be a normal task:
7107 current
->sched_class
= &fair_sched_class
;
7110 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7111 /* May be allocated at isolcpus cmdline parse time */
7112 if (cpu_isolated_map
== NULL
)
7113 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7114 idle_thread_set_boot_cpu();
7116 init_sched_fair_class();
7118 scheduler_running
= 1;
7121 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7122 static inline int preempt_count_equals(int preempt_offset
)
7124 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7126 return (nested
== preempt_offset
);
7129 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7131 static unsigned long prev_jiffy
; /* ratelimiting */
7133 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7134 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7135 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7137 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7139 prev_jiffy
= jiffies
;
7142 "BUG: sleeping function called from invalid context at %s:%d\n",
7145 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7146 in_atomic(), irqs_disabled(),
7147 current
->pid
, current
->comm
);
7149 debug_show_held_locks(current
);
7150 if (irqs_disabled())
7151 print_irqtrace_events(current
);
7154 EXPORT_SYMBOL(__might_sleep
);
7157 #ifdef CONFIG_MAGIC_SYSRQ
7158 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7160 const struct sched_class
*prev_class
= p
->sched_class
;
7161 int old_prio
= p
->prio
;
7166 dequeue_task(rq
, p
, 0);
7167 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7169 enqueue_task(rq
, p
, 0);
7170 resched_task(rq
->curr
);
7173 check_class_changed(rq
, p
, prev_class
, old_prio
);
7176 void normalize_rt_tasks(void)
7178 struct task_struct
*g
, *p
;
7179 unsigned long flags
;
7182 read_lock_irqsave(&tasklist_lock
, flags
);
7183 do_each_thread(g
, p
) {
7185 * Only normalize user tasks:
7190 p
->se
.exec_start
= 0;
7191 #ifdef CONFIG_SCHEDSTATS
7192 p
->se
.statistics
.wait_start
= 0;
7193 p
->se
.statistics
.sleep_start
= 0;
7194 p
->se
.statistics
.block_start
= 0;
7199 * Renice negative nice level userspace
7202 if (TASK_NICE(p
) < 0 && p
->mm
)
7203 set_user_nice(p
, 0);
7207 raw_spin_lock(&p
->pi_lock
);
7208 rq
= __task_rq_lock(p
);
7210 normalize_task(rq
, p
);
7212 __task_rq_unlock(rq
);
7213 raw_spin_unlock(&p
->pi_lock
);
7214 } while_each_thread(g
, p
);
7216 read_unlock_irqrestore(&tasklist_lock
, flags
);
7219 #endif /* CONFIG_MAGIC_SYSRQ */
7221 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7223 * These functions are only useful for the IA64 MCA handling, or kdb.
7225 * They can only be called when the whole system has been
7226 * stopped - every CPU needs to be quiescent, and no scheduling
7227 * activity can take place. Using them for anything else would
7228 * be a serious bug, and as a result, they aren't even visible
7229 * under any other configuration.
7233 * curr_task - return the current task for a given cpu.
7234 * @cpu: the processor in question.
7236 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7238 struct task_struct
*curr_task(int cpu
)
7240 return cpu_curr(cpu
);
7243 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7247 * set_curr_task - set the current task for a given cpu.
7248 * @cpu: the processor in question.
7249 * @p: the task pointer to set.
7251 * Description: This function must only be used when non-maskable interrupts
7252 * are serviced on a separate stack. It allows the architecture to switch the
7253 * notion of the current task on a cpu in a non-blocking manner. This function
7254 * must be called with all CPU's synchronized, and interrupts disabled, the
7255 * and caller must save the original value of the current task (see
7256 * curr_task() above) and restore that value before reenabling interrupts and
7257 * re-starting the system.
7259 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7261 void set_curr_task(int cpu
, struct task_struct
*p
)
7268 #ifdef CONFIG_CGROUP_SCHED
7269 /* task_group_lock serializes the addition/removal of task groups */
7270 static DEFINE_SPINLOCK(task_group_lock
);
7272 static void free_sched_group(struct task_group
*tg
)
7274 free_fair_sched_group(tg
);
7275 free_rt_sched_group(tg
);
7280 /* allocate runqueue etc for a new task group */
7281 struct task_group
*sched_create_group(struct task_group
*parent
)
7283 struct task_group
*tg
;
7285 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7287 return ERR_PTR(-ENOMEM
);
7289 if (!alloc_fair_sched_group(tg
, parent
))
7292 if (!alloc_rt_sched_group(tg
, parent
))
7298 free_sched_group(tg
);
7299 return ERR_PTR(-ENOMEM
);
7302 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7304 unsigned long flags
;
7306 spin_lock_irqsave(&task_group_lock
, flags
);
7307 list_add_rcu(&tg
->list
, &task_groups
);
7309 WARN_ON(!parent
); /* root should already exist */
7311 tg
->parent
= parent
;
7312 INIT_LIST_HEAD(&tg
->children
);
7313 list_add_rcu(&tg
->siblings
, &parent
->children
);
7314 spin_unlock_irqrestore(&task_group_lock
, flags
);
7317 /* rcu callback to free various structures associated with a task group */
7318 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7320 /* now it should be safe to free those cfs_rqs */
7321 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7324 /* Destroy runqueue etc associated with a task group */
7325 void sched_destroy_group(struct task_group
*tg
)
7327 /* wait for possible concurrent references to cfs_rqs complete */
7328 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7331 void sched_offline_group(struct task_group
*tg
)
7333 unsigned long flags
;
7336 /* end participation in shares distribution */
7337 for_each_possible_cpu(i
)
7338 unregister_fair_sched_group(tg
, i
);
7340 spin_lock_irqsave(&task_group_lock
, flags
);
7341 list_del_rcu(&tg
->list
);
7342 list_del_rcu(&tg
->siblings
);
7343 spin_unlock_irqrestore(&task_group_lock
, flags
);
7346 /* change task's runqueue when it moves between groups.
7347 * The caller of this function should have put the task in its new group
7348 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7349 * reflect its new group.
7351 void sched_move_task(struct task_struct
*tsk
)
7353 struct task_group
*tg
;
7355 unsigned long flags
;
7358 rq
= task_rq_lock(tsk
, &flags
);
7360 running
= task_current(rq
, tsk
);
7364 dequeue_task(rq
, tsk
, 0);
7365 if (unlikely(running
))
7366 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7368 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
7369 lockdep_is_held(&tsk
->sighand
->siglock
)),
7370 struct task_group
, css
);
7371 tg
= autogroup_task_group(tsk
, tg
);
7372 tsk
->sched_task_group
= tg
;
7374 #ifdef CONFIG_FAIR_GROUP_SCHED
7375 if (tsk
->sched_class
->task_move_group
)
7376 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7379 set_task_rq(tsk
, task_cpu(tsk
));
7381 if (unlikely(running
))
7382 tsk
->sched_class
->set_curr_task(rq
);
7384 enqueue_task(rq
, tsk
, 0);
7386 task_rq_unlock(rq
, tsk
, &flags
);
7388 #endif /* CONFIG_CGROUP_SCHED */
7390 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7391 static unsigned long to_ratio(u64 period
, u64 runtime
)
7393 if (runtime
== RUNTIME_INF
)
7396 return div64_u64(runtime
<< 20, period
);
7400 #ifdef CONFIG_RT_GROUP_SCHED
7402 * Ensure that the real time constraints are schedulable.
7404 static DEFINE_MUTEX(rt_constraints_mutex
);
7406 /* Must be called with tasklist_lock held */
7407 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7409 struct task_struct
*g
, *p
;
7411 do_each_thread(g
, p
) {
7412 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7414 } while_each_thread(g
, p
);
7419 struct rt_schedulable_data
{
7420 struct task_group
*tg
;
7425 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7427 struct rt_schedulable_data
*d
= data
;
7428 struct task_group
*child
;
7429 unsigned long total
, sum
= 0;
7430 u64 period
, runtime
;
7432 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7433 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7436 period
= d
->rt_period
;
7437 runtime
= d
->rt_runtime
;
7441 * Cannot have more runtime than the period.
7443 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7447 * Ensure we don't starve existing RT tasks.
7449 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7452 total
= to_ratio(period
, runtime
);
7455 * Nobody can have more than the global setting allows.
7457 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7461 * The sum of our children's runtime should not exceed our own.
7463 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7464 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7465 runtime
= child
->rt_bandwidth
.rt_runtime
;
7467 if (child
== d
->tg
) {
7468 period
= d
->rt_period
;
7469 runtime
= d
->rt_runtime
;
7472 sum
+= to_ratio(period
, runtime
);
7481 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7485 struct rt_schedulable_data data
= {
7487 .rt_period
= period
,
7488 .rt_runtime
= runtime
,
7492 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7498 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7499 u64 rt_period
, u64 rt_runtime
)
7503 mutex_lock(&rt_constraints_mutex
);
7504 read_lock(&tasklist_lock
);
7505 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7509 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7510 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7511 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7513 for_each_possible_cpu(i
) {
7514 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7516 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7517 rt_rq
->rt_runtime
= rt_runtime
;
7518 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7520 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7522 read_unlock(&tasklist_lock
);
7523 mutex_unlock(&rt_constraints_mutex
);
7528 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7530 u64 rt_runtime
, rt_period
;
7532 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7533 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7534 if (rt_runtime_us
< 0)
7535 rt_runtime
= RUNTIME_INF
;
7537 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7540 static long sched_group_rt_runtime(struct task_group
*tg
)
7544 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7547 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7548 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7549 return rt_runtime_us
;
7552 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7554 u64 rt_runtime
, rt_period
;
7556 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7557 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7562 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7565 static long sched_group_rt_period(struct task_group
*tg
)
7569 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7570 do_div(rt_period_us
, NSEC_PER_USEC
);
7571 return rt_period_us
;
7574 static int sched_rt_global_constraints(void)
7576 u64 runtime
, period
;
7579 if (sysctl_sched_rt_period
<= 0)
7582 runtime
= global_rt_runtime();
7583 period
= global_rt_period();
7586 * Sanity check on the sysctl variables.
7588 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7591 mutex_lock(&rt_constraints_mutex
);
7592 read_lock(&tasklist_lock
);
7593 ret
= __rt_schedulable(NULL
, 0, 0);
7594 read_unlock(&tasklist_lock
);
7595 mutex_unlock(&rt_constraints_mutex
);
7600 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7602 /* Don't accept realtime tasks when there is no way for them to run */
7603 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7609 #else /* !CONFIG_RT_GROUP_SCHED */
7610 static int sched_rt_global_constraints(void)
7612 unsigned long flags
;
7615 if (sysctl_sched_rt_period
<= 0)
7619 * There's always some RT tasks in the root group
7620 * -- migration, kstopmachine etc..
7622 if (sysctl_sched_rt_runtime
== 0)
7625 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7626 for_each_possible_cpu(i
) {
7627 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7629 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7630 rt_rq
->rt_runtime
= global_rt_runtime();
7631 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7633 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7637 #endif /* CONFIG_RT_GROUP_SCHED */
7639 int sched_rr_handler(struct ctl_table
*table
, int write
,
7640 void __user
*buffer
, size_t *lenp
,
7644 static DEFINE_MUTEX(mutex
);
7647 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7648 /* make sure that internally we keep jiffies */
7649 /* also, writing zero resets timeslice to default */
7650 if (!ret
&& write
) {
7651 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7652 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7654 mutex_unlock(&mutex
);
7658 int sched_rt_handler(struct ctl_table
*table
, int write
,
7659 void __user
*buffer
, size_t *lenp
,
7663 int old_period
, old_runtime
;
7664 static DEFINE_MUTEX(mutex
);
7667 old_period
= sysctl_sched_rt_period
;
7668 old_runtime
= sysctl_sched_rt_runtime
;
7670 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7672 if (!ret
&& write
) {
7673 ret
= sched_rt_global_constraints();
7675 sysctl_sched_rt_period
= old_period
;
7676 sysctl_sched_rt_runtime
= old_runtime
;
7678 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7679 def_rt_bandwidth
.rt_period
=
7680 ns_to_ktime(global_rt_period());
7683 mutex_unlock(&mutex
);
7688 #ifdef CONFIG_CGROUP_SCHED
7690 /* return corresponding task_group object of a cgroup */
7691 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7693 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7694 struct task_group
, css
);
7697 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7699 struct task_group
*tg
, *parent
;
7701 if (!cgrp
->parent
) {
7702 /* This is early initialization for the top cgroup */
7703 return &root_task_group
.css
;
7706 parent
= cgroup_tg(cgrp
->parent
);
7707 tg
= sched_create_group(parent
);
7709 return ERR_PTR(-ENOMEM
);
7714 static int cpu_cgroup_css_online(struct cgroup
*cgrp
)
7716 struct task_group
*tg
= cgroup_tg(cgrp
);
7717 struct task_group
*parent
;
7722 parent
= cgroup_tg(cgrp
->parent
);
7723 sched_online_group(tg
, parent
);
7727 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7729 struct task_group
*tg
= cgroup_tg(cgrp
);
7731 sched_destroy_group(tg
);
7734 static void cpu_cgroup_css_offline(struct cgroup
*cgrp
)
7736 struct task_group
*tg
= cgroup_tg(cgrp
);
7738 sched_offline_group(tg
);
7741 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7742 struct cgroup_taskset
*tset
)
7744 struct task_struct
*task
;
7746 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7747 #ifdef CONFIG_RT_GROUP_SCHED
7748 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7751 /* We don't support RT-tasks being in separate groups */
7752 if (task
->sched_class
!= &fair_sched_class
)
7759 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7760 struct cgroup_taskset
*tset
)
7762 struct task_struct
*task
;
7764 cgroup_taskset_for_each(task
, cgrp
, tset
)
7765 sched_move_task(task
);
7769 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7770 struct task_struct
*task
)
7773 * cgroup_exit() is called in the copy_process() failure path.
7774 * Ignore this case since the task hasn't ran yet, this avoids
7775 * trying to poke a half freed task state from generic code.
7777 if (!(task
->flags
& PF_EXITING
))
7780 sched_move_task(task
);
7783 #ifdef CONFIG_FAIR_GROUP_SCHED
7784 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7787 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7790 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7792 struct task_group
*tg
= cgroup_tg(cgrp
);
7794 return (u64
) scale_load_down(tg
->shares
);
7797 #ifdef CONFIG_CFS_BANDWIDTH
7798 static DEFINE_MUTEX(cfs_constraints_mutex
);
7800 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7801 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7803 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7805 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7807 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7808 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7810 if (tg
== &root_task_group
)
7814 * Ensure we have at some amount of bandwidth every period. This is
7815 * to prevent reaching a state of large arrears when throttled via
7816 * entity_tick() resulting in prolonged exit starvation.
7818 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7822 * Likewise, bound things on the otherside by preventing insane quota
7823 * periods. This also allows us to normalize in computing quota
7826 if (period
> max_cfs_quota_period
)
7829 mutex_lock(&cfs_constraints_mutex
);
7830 ret
= __cfs_schedulable(tg
, period
, quota
);
7834 runtime_enabled
= quota
!= RUNTIME_INF
;
7835 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7837 * If we need to toggle cfs_bandwidth_used, off->on must occur
7838 * before making related changes, and on->off must occur afterwards
7840 if (runtime_enabled
&& !runtime_was_enabled
)
7841 cfs_bandwidth_usage_inc();
7842 raw_spin_lock_irq(&cfs_b
->lock
);
7843 cfs_b
->period
= ns_to_ktime(period
);
7844 cfs_b
->quota
= quota
;
7846 __refill_cfs_bandwidth_runtime(cfs_b
);
7847 /* restart the period timer (if active) to handle new period expiry */
7848 if (runtime_enabled
&& cfs_b
->timer_active
) {
7849 /* force a reprogram */
7850 cfs_b
->timer_active
= 0;
7851 __start_cfs_bandwidth(cfs_b
);
7853 raw_spin_unlock_irq(&cfs_b
->lock
);
7855 for_each_possible_cpu(i
) {
7856 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7857 struct rq
*rq
= cfs_rq
->rq
;
7859 raw_spin_lock_irq(&rq
->lock
);
7860 cfs_rq
->runtime_enabled
= runtime_enabled
;
7861 cfs_rq
->runtime_remaining
= 0;
7863 if (cfs_rq
->throttled
)
7864 unthrottle_cfs_rq(cfs_rq
);
7865 raw_spin_unlock_irq(&rq
->lock
);
7867 if (runtime_was_enabled
&& !runtime_enabled
)
7868 cfs_bandwidth_usage_dec();
7870 mutex_unlock(&cfs_constraints_mutex
);
7875 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7879 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7880 if (cfs_quota_us
< 0)
7881 quota
= RUNTIME_INF
;
7883 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7885 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7888 long tg_get_cfs_quota(struct task_group
*tg
)
7892 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7895 quota_us
= tg
->cfs_bandwidth
.quota
;
7896 do_div(quota_us
, NSEC_PER_USEC
);
7901 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7905 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7906 quota
= tg
->cfs_bandwidth
.quota
;
7908 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7911 long tg_get_cfs_period(struct task_group
*tg
)
7915 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7916 do_div(cfs_period_us
, NSEC_PER_USEC
);
7918 return cfs_period_us
;
7921 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7923 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7926 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7929 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7932 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7934 return tg_get_cfs_period(cgroup_tg(cgrp
));
7937 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7940 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7943 struct cfs_schedulable_data
{
7944 struct task_group
*tg
;
7949 * normalize group quota/period to be quota/max_period
7950 * note: units are usecs
7952 static u64
normalize_cfs_quota(struct task_group
*tg
,
7953 struct cfs_schedulable_data
*d
)
7961 period
= tg_get_cfs_period(tg
);
7962 quota
= tg_get_cfs_quota(tg
);
7965 /* note: these should typically be equivalent */
7966 if (quota
== RUNTIME_INF
|| quota
== -1)
7969 return to_ratio(period
, quota
);
7972 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7974 struct cfs_schedulable_data
*d
= data
;
7975 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7976 s64 quota
= 0, parent_quota
= -1;
7979 quota
= RUNTIME_INF
;
7981 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7983 quota
= normalize_cfs_quota(tg
, d
);
7984 parent_quota
= parent_b
->hierarchal_quota
;
7987 * ensure max(child_quota) <= parent_quota, inherit when no
7990 if (quota
== RUNTIME_INF
)
7991 quota
= parent_quota
;
7992 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7995 cfs_b
->hierarchal_quota
= quota
;
8000 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8003 struct cfs_schedulable_data data
= {
8009 if (quota
!= RUNTIME_INF
) {
8010 do_div(data
.period
, NSEC_PER_USEC
);
8011 do_div(data
.quota
, NSEC_PER_USEC
);
8015 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8021 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8022 struct cgroup_map_cb
*cb
)
8024 struct task_group
*tg
= cgroup_tg(cgrp
);
8025 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8027 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
8028 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
8029 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
8033 #endif /* CONFIG_CFS_BANDWIDTH */
8034 #endif /* CONFIG_FAIR_GROUP_SCHED */
8036 #ifdef CONFIG_RT_GROUP_SCHED
8037 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8040 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8043 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8045 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8048 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8051 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8054 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8056 return sched_group_rt_period(cgroup_tg(cgrp
));
8058 #endif /* CONFIG_RT_GROUP_SCHED */
8060 static struct cftype cpu_files
[] = {
8061 #ifdef CONFIG_FAIR_GROUP_SCHED
8064 .read_u64
= cpu_shares_read_u64
,
8065 .write_u64
= cpu_shares_write_u64
,
8068 #ifdef CONFIG_CFS_BANDWIDTH
8070 .name
= "cfs_quota_us",
8071 .read_s64
= cpu_cfs_quota_read_s64
,
8072 .write_s64
= cpu_cfs_quota_write_s64
,
8075 .name
= "cfs_period_us",
8076 .read_u64
= cpu_cfs_period_read_u64
,
8077 .write_u64
= cpu_cfs_period_write_u64
,
8081 .read_map
= cpu_stats_show
,
8084 #ifdef CONFIG_RT_GROUP_SCHED
8086 .name
= "rt_runtime_us",
8087 .read_s64
= cpu_rt_runtime_read
,
8088 .write_s64
= cpu_rt_runtime_write
,
8091 .name
= "rt_period_us",
8092 .read_u64
= cpu_rt_period_read_uint
,
8093 .write_u64
= cpu_rt_period_write_uint
,
8099 struct cgroup_subsys cpu_cgroup_subsys
= {
8101 .css_alloc
= cpu_cgroup_css_alloc
,
8102 .css_free
= cpu_cgroup_css_free
,
8103 .css_online
= cpu_cgroup_css_online
,
8104 .css_offline
= cpu_cgroup_css_offline
,
8105 .can_attach
= cpu_cgroup_can_attach
,
8106 .attach
= cpu_cgroup_attach
,
8107 .exit
= cpu_cgroup_exit
,
8108 .subsys_id
= cpu_cgroup_subsys_id
,
8109 .base_cftypes
= cpu_files
,
8113 #endif /* CONFIG_CGROUP_SCHED */
8115 void dump_cpu_task(int cpu
)
8117 pr_info("Task dump for CPU %d:\n", cpu
);
8118 sched_show_task(cpu_curr(cpu
));