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
) - 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_sched("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
) - 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 /* check_preempt_curr() may use rq clock */
1381 update_rq_clock(rq
);
1382 ttwu_do_wakeup(rq
, p
, wake_flags
);
1385 __task_rq_unlock(rq
);
1391 static void sched_ttwu_pending(void)
1393 struct rq
*rq
= this_rq();
1394 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1395 struct task_struct
*p
;
1397 raw_spin_lock(&rq
->lock
);
1400 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1401 llist
= llist_next(llist
);
1402 ttwu_do_activate(rq
, p
, 0);
1405 raw_spin_unlock(&rq
->lock
);
1408 void scheduler_ipi(void)
1410 if (llist_empty(&this_rq()->wake_list
)
1411 && !tick_nohz_full_cpu(smp_processor_id())
1412 && !got_nohz_idle_kick())
1416 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1417 * traditionally all their work was done from the interrupt return
1418 * path. Now that we actually do some work, we need to make sure
1421 * Some archs already do call them, luckily irq_enter/exit nest
1424 * Arguably we should visit all archs and update all handlers,
1425 * however a fair share of IPIs are still resched only so this would
1426 * somewhat pessimize the simple resched case.
1429 tick_nohz_full_check();
1430 sched_ttwu_pending();
1433 * Check if someone kicked us for doing the nohz idle load balance.
1435 if (unlikely(got_nohz_idle_kick())) {
1436 this_rq()->idle_balance
= 1;
1437 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1442 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1444 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1445 smp_send_reschedule(cpu
);
1448 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1450 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1452 #endif /* CONFIG_SMP */
1454 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1456 struct rq
*rq
= cpu_rq(cpu
);
1458 #if defined(CONFIG_SMP)
1459 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1460 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1461 ttwu_queue_remote(p
, cpu
);
1466 raw_spin_lock(&rq
->lock
);
1467 ttwu_do_activate(rq
, p
, 0);
1468 raw_spin_unlock(&rq
->lock
);
1472 * try_to_wake_up - wake up a thread
1473 * @p: the thread to be awakened
1474 * @state: the mask of task states that can be woken
1475 * @wake_flags: wake modifier flags (WF_*)
1477 * Put it on the run-queue if it's not already there. The "current"
1478 * thread is always on the run-queue (except when the actual
1479 * re-schedule is in progress), and as such you're allowed to do
1480 * the simpler "current->state = TASK_RUNNING" to mark yourself
1481 * runnable without the overhead of this.
1483 * Returns %true if @p was woken up, %false if it was already running
1484 * or @state didn't match @p's state.
1487 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1489 unsigned long flags
;
1490 int cpu
, success
= 0;
1493 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1494 if (!(p
->state
& state
))
1497 success
= 1; /* we're going to change ->state */
1500 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1505 * If the owning (remote) cpu is still in the middle of schedule() with
1506 * this task as prev, wait until its done referencing the task.
1511 * Pairs with the smp_wmb() in finish_lock_switch().
1515 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1516 p
->state
= TASK_WAKING
;
1518 if (p
->sched_class
->task_waking
)
1519 p
->sched_class
->task_waking(p
);
1521 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1522 if (task_cpu(p
) != cpu
) {
1523 wake_flags
|= WF_MIGRATED
;
1524 set_task_cpu(p
, cpu
);
1526 #endif /* CONFIG_SMP */
1530 ttwu_stat(p
, cpu
, wake_flags
);
1532 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1538 * try_to_wake_up_local - try to wake up a local task with rq lock held
1539 * @p: the thread to be awakened
1541 * Put @p on the run-queue if it's not already there. The caller must
1542 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1545 static void try_to_wake_up_local(struct task_struct
*p
)
1547 struct rq
*rq
= task_rq(p
);
1549 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1550 WARN_ON_ONCE(p
== current
))
1553 lockdep_assert_held(&rq
->lock
);
1555 if (!raw_spin_trylock(&p
->pi_lock
)) {
1556 raw_spin_unlock(&rq
->lock
);
1557 raw_spin_lock(&p
->pi_lock
);
1558 raw_spin_lock(&rq
->lock
);
1561 if (!(p
->state
& TASK_NORMAL
))
1565 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1567 ttwu_do_wakeup(rq
, p
, 0);
1568 ttwu_stat(p
, smp_processor_id(), 0);
1570 raw_spin_unlock(&p
->pi_lock
);
1574 * wake_up_process - Wake up a specific process
1575 * @p: The process to be woken up.
1577 * Attempt to wake up the nominated process and move it to the set of runnable
1578 * processes. Returns 1 if the process was woken up, 0 if it was already
1581 * It may be assumed that this function implies a write memory barrier before
1582 * changing the task state if and only if any tasks are woken up.
1584 int wake_up_process(struct task_struct
*p
)
1586 WARN_ON(task_is_stopped_or_traced(p
));
1587 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1589 EXPORT_SYMBOL(wake_up_process
);
1591 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1593 return try_to_wake_up(p
, state
, 0);
1597 * Perform scheduler related setup for a newly forked process p.
1598 * p is forked by current.
1600 * __sched_fork() is basic setup used by init_idle() too:
1602 static void __sched_fork(struct task_struct
*p
)
1607 p
->se
.exec_start
= 0;
1608 p
->se
.sum_exec_runtime
= 0;
1609 p
->se
.prev_sum_exec_runtime
= 0;
1610 p
->se
.nr_migrations
= 0;
1612 INIT_LIST_HEAD(&p
->se
.group_node
);
1614 #ifdef CONFIG_SCHEDSTATS
1615 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1618 INIT_LIST_HEAD(&p
->rt
.run_list
);
1620 #ifdef CONFIG_PREEMPT_NOTIFIERS
1621 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1624 #ifdef CONFIG_NUMA_BALANCING
1625 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1626 p
->mm
->numa_next_scan
= jiffies
;
1627 p
->mm
->numa_next_reset
= jiffies
;
1628 p
->mm
->numa_scan_seq
= 0;
1631 p
->node_stamp
= 0ULL;
1632 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1633 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1634 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1635 p
->numa_work
.next
= &p
->numa_work
;
1636 #endif /* CONFIG_NUMA_BALANCING */
1639 #ifdef CONFIG_NUMA_BALANCING
1640 #ifdef CONFIG_SCHED_DEBUG
1641 void set_numabalancing_state(bool enabled
)
1644 sched_feat_set("NUMA");
1646 sched_feat_set("NO_NUMA");
1649 __read_mostly
bool numabalancing_enabled
;
1651 void set_numabalancing_state(bool enabled
)
1653 numabalancing_enabled
= enabled
;
1655 #endif /* CONFIG_SCHED_DEBUG */
1656 #endif /* CONFIG_NUMA_BALANCING */
1659 * fork()/clone()-time setup:
1661 void sched_fork(struct task_struct
*p
)
1663 unsigned long flags
;
1664 int cpu
= get_cpu();
1668 * We mark the process as running here. This guarantees that
1669 * nobody will actually run it, and a signal or other external
1670 * event cannot wake it up and insert it on the runqueue either.
1672 p
->state
= TASK_RUNNING
;
1675 * Make sure we do not leak PI boosting priority to the child.
1677 p
->prio
= current
->normal_prio
;
1680 * Revert to default priority/policy on fork if requested.
1682 if (unlikely(p
->sched_reset_on_fork
)) {
1683 if (task_has_rt_policy(p
)) {
1684 p
->policy
= SCHED_NORMAL
;
1685 p
->static_prio
= NICE_TO_PRIO(0);
1687 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1688 p
->static_prio
= NICE_TO_PRIO(0);
1690 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1694 * We don't need the reset flag anymore after the fork. It has
1695 * fulfilled its duty:
1697 p
->sched_reset_on_fork
= 0;
1700 if (!rt_prio(p
->prio
))
1701 p
->sched_class
= &fair_sched_class
;
1703 if (p
->sched_class
->task_fork
)
1704 p
->sched_class
->task_fork(p
);
1707 * The child is not yet in the pid-hash so no cgroup attach races,
1708 * and the cgroup is pinned to this child due to cgroup_fork()
1709 * is ran before sched_fork().
1711 * Silence PROVE_RCU.
1713 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1714 set_task_cpu(p
, cpu
);
1715 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1717 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1718 if (likely(sched_info_on()))
1719 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1721 #if defined(CONFIG_SMP)
1724 #ifdef CONFIG_PREEMPT_COUNT
1725 /* Want to start with kernel preemption disabled. */
1726 task_thread_info(p
)->preempt_count
= 1;
1729 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1736 * wake_up_new_task - wake up a newly created task for the first time.
1738 * This function will do some initial scheduler statistics housekeeping
1739 * that must be done for every newly created context, then puts the task
1740 * on the runqueue and wakes it.
1742 void wake_up_new_task(struct task_struct
*p
)
1744 unsigned long flags
;
1747 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1750 * Fork balancing, do it here and not earlier because:
1751 * - cpus_allowed can change in the fork path
1752 * - any previously selected cpu might disappear through hotplug
1754 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1757 /* Initialize new task's runnable average */
1758 init_task_runnable_average(p
);
1759 rq
= __task_rq_lock(p
);
1760 activate_task(rq
, p
, 0);
1762 trace_sched_wakeup_new(p
, true);
1763 check_preempt_curr(rq
, p
, WF_FORK
);
1765 if (p
->sched_class
->task_woken
)
1766 p
->sched_class
->task_woken(rq
, p
);
1768 task_rq_unlock(rq
, p
, &flags
);
1771 #ifdef CONFIG_PREEMPT_NOTIFIERS
1774 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1775 * @notifier: notifier struct to register
1777 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1779 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1781 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1784 * preempt_notifier_unregister - no longer interested in preemption notifications
1785 * @notifier: notifier struct to unregister
1787 * This is safe to call from within a preemption notifier.
1789 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1791 hlist_del(¬ifier
->link
);
1793 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1795 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1797 struct preempt_notifier
*notifier
;
1799 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1800 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1804 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1805 struct task_struct
*next
)
1807 struct preempt_notifier
*notifier
;
1809 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1810 notifier
->ops
->sched_out(notifier
, next
);
1813 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1815 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1820 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1821 struct task_struct
*next
)
1825 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1828 * prepare_task_switch - prepare to switch tasks
1829 * @rq: the runqueue preparing to switch
1830 * @prev: the current task that is being switched out
1831 * @next: the task we are going to switch to.
1833 * This is called with the rq lock held and interrupts off. It must
1834 * be paired with a subsequent finish_task_switch after the context
1837 * prepare_task_switch sets up locking and calls architecture specific
1841 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1842 struct task_struct
*next
)
1844 trace_sched_switch(prev
, next
);
1845 sched_info_switch(prev
, next
);
1846 perf_event_task_sched_out(prev
, next
);
1847 fire_sched_out_preempt_notifiers(prev
, next
);
1848 prepare_lock_switch(rq
, next
);
1849 prepare_arch_switch(next
);
1853 * finish_task_switch - clean up after a task-switch
1854 * @rq: runqueue associated with task-switch
1855 * @prev: the thread we just switched away from.
1857 * finish_task_switch must be called after the context switch, paired
1858 * with a prepare_task_switch call before the context switch.
1859 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1860 * and do any other architecture-specific cleanup actions.
1862 * Note that we may have delayed dropping an mm in context_switch(). If
1863 * so, we finish that here outside of the runqueue lock. (Doing it
1864 * with the lock held can cause deadlocks; see schedule() for
1867 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1868 __releases(rq
->lock
)
1870 struct mm_struct
*mm
= rq
->prev_mm
;
1876 * A task struct has one reference for the use as "current".
1877 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1878 * schedule one last time. The schedule call will never return, and
1879 * the scheduled task must drop that reference.
1880 * The test for TASK_DEAD must occur while the runqueue locks are
1881 * still held, otherwise prev could be scheduled on another cpu, die
1882 * there before we look at prev->state, and then the reference would
1884 * Manfred Spraul <manfred@colorfullife.com>
1886 prev_state
= prev
->state
;
1887 vtime_task_switch(prev
);
1888 finish_arch_switch(prev
);
1889 perf_event_task_sched_in(prev
, current
);
1890 finish_lock_switch(rq
, prev
);
1891 finish_arch_post_lock_switch();
1893 fire_sched_in_preempt_notifiers(current
);
1896 if (unlikely(prev_state
== TASK_DEAD
)) {
1898 * Remove function-return probe instances associated with this
1899 * task and put them back on the free list.
1901 kprobe_flush_task(prev
);
1902 put_task_struct(prev
);
1905 tick_nohz_task_switch(current
);
1910 /* assumes rq->lock is held */
1911 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1913 if (prev
->sched_class
->pre_schedule
)
1914 prev
->sched_class
->pre_schedule(rq
, prev
);
1917 /* rq->lock is NOT held, but preemption is disabled */
1918 static inline void post_schedule(struct rq
*rq
)
1920 if (rq
->post_schedule
) {
1921 unsigned long flags
;
1923 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1924 if (rq
->curr
->sched_class
->post_schedule
)
1925 rq
->curr
->sched_class
->post_schedule(rq
);
1926 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1928 rq
->post_schedule
= 0;
1934 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1938 static inline void post_schedule(struct rq
*rq
)
1945 * schedule_tail - first thing a freshly forked thread must call.
1946 * @prev: the thread we just switched away from.
1948 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1949 __releases(rq
->lock
)
1951 struct rq
*rq
= this_rq();
1953 finish_task_switch(rq
, prev
);
1956 * FIXME: do we need to worry about rq being invalidated by the
1961 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1962 /* In this case, finish_task_switch does not reenable preemption */
1965 if (current
->set_child_tid
)
1966 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1970 * context_switch - switch to the new MM and the new
1971 * thread's register state.
1974 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1975 struct task_struct
*next
)
1977 struct mm_struct
*mm
, *oldmm
;
1979 prepare_task_switch(rq
, prev
, next
);
1982 oldmm
= prev
->active_mm
;
1984 * For paravirt, this is coupled with an exit in switch_to to
1985 * combine the page table reload and the switch backend into
1988 arch_start_context_switch(prev
);
1991 next
->active_mm
= oldmm
;
1992 atomic_inc(&oldmm
->mm_count
);
1993 enter_lazy_tlb(oldmm
, next
);
1995 switch_mm(oldmm
, mm
, next
);
1998 prev
->active_mm
= NULL
;
1999 rq
->prev_mm
= oldmm
;
2002 * Since the runqueue lock will be released by the next
2003 * task (which is an invalid locking op but in the case
2004 * of the scheduler it's an obvious special-case), so we
2005 * do an early lockdep release here:
2007 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2008 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2011 context_tracking_task_switch(prev
, next
);
2012 /* Here we just switch the register state and the stack. */
2013 switch_to(prev
, next
, prev
);
2017 * this_rq must be evaluated again because prev may have moved
2018 * CPUs since it called schedule(), thus the 'rq' on its stack
2019 * frame will be invalid.
2021 finish_task_switch(this_rq(), prev
);
2025 * nr_running and nr_context_switches:
2027 * externally visible scheduler statistics: current number of runnable
2028 * threads, total number of context switches performed since bootup.
2030 unsigned long nr_running(void)
2032 unsigned long i
, sum
= 0;
2034 for_each_online_cpu(i
)
2035 sum
+= cpu_rq(i
)->nr_running
;
2040 unsigned long long nr_context_switches(void)
2043 unsigned long long sum
= 0;
2045 for_each_possible_cpu(i
)
2046 sum
+= cpu_rq(i
)->nr_switches
;
2051 unsigned long nr_iowait(void)
2053 unsigned long i
, sum
= 0;
2055 for_each_possible_cpu(i
)
2056 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2061 unsigned long nr_iowait_cpu(int cpu
)
2063 struct rq
*this = cpu_rq(cpu
);
2064 return atomic_read(&this->nr_iowait
);
2070 * sched_exec - execve() is a valuable balancing opportunity, because at
2071 * this point the task has the smallest effective memory and cache footprint.
2073 void sched_exec(void)
2075 struct task_struct
*p
= current
;
2076 unsigned long flags
;
2079 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2080 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2081 if (dest_cpu
== smp_processor_id())
2084 if (likely(cpu_active(dest_cpu
))) {
2085 struct migration_arg arg
= { p
, dest_cpu
};
2087 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2088 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2092 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2097 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2098 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2100 EXPORT_PER_CPU_SYMBOL(kstat
);
2101 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2104 * Return any ns on the sched_clock that have not yet been accounted in
2105 * @p in case that task is currently running.
2107 * Called with task_rq_lock() held on @rq.
2109 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2113 if (task_current(rq
, p
)) {
2114 update_rq_clock(rq
);
2115 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2123 unsigned long long task_delta_exec(struct task_struct
*p
)
2125 unsigned long flags
;
2129 rq
= task_rq_lock(p
, &flags
);
2130 ns
= do_task_delta_exec(p
, rq
);
2131 task_rq_unlock(rq
, p
, &flags
);
2137 * Return accounted runtime for the task.
2138 * In case the task is currently running, return the runtime plus current's
2139 * pending runtime that have not been accounted yet.
2141 unsigned long long task_sched_runtime(struct task_struct
*p
)
2143 unsigned long flags
;
2147 rq
= task_rq_lock(p
, &flags
);
2148 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2149 task_rq_unlock(rq
, p
, &flags
);
2155 * This function gets called by the timer code, with HZ frequency.
2156 * We call it with interrupts disabled.
2158 void scheduler_tick(void)
2160 int cpu
= smp_processor_id();
2161 struct rq
*rq
= cpu_rq(cpu
);
2162 struct task_struct
*curr
= rq
->curr
;
2166 raw_spin_lock(&rq
->lock
);
2167 update_rq_clock(rq
);
2168 curr
->sched_class
->task_tick(rq
, curr
, 0);
2169 update_cpu_load_active(rq
);
2170 raw_spin_unlock(&rq
->lock
);
2172 perf_event_task_tick();
2175 rq
->idle_balance
= idle_cpu(cpu
);
2176 trigger_load_balance(rq
, cpu
);
2178 rq_last_tick_reset(rq
);
2181 #ifdef CONFIG_NO_HZ_FULL
2183 * scheduler_tick_max_deferment
2185 * Keep at least one tick per second when a single
2186 * active task is running because the scheduler doesn't
2187 * yet completely support full dynticks environment.
2189 * This makes sure that uptime, CFS vruntime, load
2190 * balancing, etc... continue to move forward, even
2191 * with a very low granularity.
2193 u64
scheduler_tick_max_deferment(void)
2195 struct rq
*rq
= this_rq();
2196 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2198 next
= rq
->last_sched_tick
+ HZ
;
2200 if (time_before_eq(next
, now
))
2203 return jiffies_to_usecs(next
- now
) * NSEC_PER_USEC
;
2207 notrace
unsigned long get_parent_ip(unsigned long addr
)
2209 if (in_lock_functions(addr
)) {
2210 addr
= CALLER_ADDR2
;
2211 if (in_lock_functions(addr
))
2212 addr
= CALLER_ADDR3
;
2217 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2218 defined(CONFIG_PREEMPT_TRACER))
2220 void __kprobes
add_preempt_count(int val
)
2222 #ifdef CONFIG_DEBUG_PREEMPT
2226 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2229 preempt_count() += val
;
2230 #ifdef CONFIG_DEBUG_PREEMPT
2232 * Spinlock count overflowing soon?
2234 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2237 if (preempt_count() == val
)
2238 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2240 EXPORT_SYMBOL(add_preempt_count
);
2242 void __kprobes
sub_preempt_count(int val
)
2244 #ifdef CONFIG_DEBUG_PREEMPT
2248 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2251 * Is the spinlock portion underflowing?
2253 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2254 !(preempt_count() & PREEMPT_MASK
)))
2258 if (preempt_count() == val
)
2259 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2260 preempt_count() -= val
;
2262 EXPORT_SYMBOL(sub_preempt_count
);
2267 * Print scheduling while atomic bug:
2269 static noinline
void __schedule_bug(struct task_struct
*prev
)
2271 if (oops_in_progress
)
2274 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2275 prev
->comm
, prev
->pid
, preempt_count());
2277 debug_show_held_locks(prev
);
2279 if (irqs_disabled())
2280 print_irqtrace_events(prev
);
2282 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2286 * Various schedule()-time debugging checks and statistics:
2288 static inline void schedule_debug(struct task_struct
*prev
)
2291 * Test if we are atomic. Since do_exit() needs to call into
2292 * schedule() atomically, we ignore that path for now.
2293 * Otherwise, whine if we are scheduling when we should not be.
2295 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2296 __schedule_bug(prev
);
2299 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2301 schedstat_inc(this_rq(), sched_count
);
2304 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2306 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2307 update_rq_clock(rq
);
2308 prev
->sched_class
->put_prev_task(rq
, prev
);
2312 * Pick up the highest-prio task:
2314 static inline struct task_struct
*
2315 pick_next_task(struct rq
*rq
)
2317 const struct sched_class
*class;
2318 struct task_struct
*p
;
2321 * Optimization: we know that if all tasks are in
2322 * the fair class we can call that function directly:
2324 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2325 p
= fair_sched_class
.pick_next_task(rq
);
2330 for_each_class(class) {
2331 p
= class->pick_next_task(rq
);
2336 BUG(); /* the idle class will always have a runnable task */
2340 * __schedule() is the main scheduler function.
2342 * The main means of driving the scheduler and thus entering this function are:
2344 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2346 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2347 * paths. For example, see arch/x86/entry_64.S.
2349 * To drive preemption between tasks, the scheduler sets the flag in timer
2350 * interrupt handler scheduler_tick().
2352 * 3. Wakeups don't really cause entry into schedule(). They add a
2353 * task to the run-queue and that's it.
2355 * Now, if the new task added to the run-queue preempts the current
2356 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2357 * called on the nearest possible occasion:
2359 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2361 * - in syscall or exception context, at the next outmost
2362 * preempt_enable(). (this might be as soon as the wake_up()'s
2365 * - in IRQ context, return from interrupt-handler to
2366 * preemptible context
2368 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2371 * - cond_resched() call
2372 * - explicit schedule() call
2373 * - return from syscall or exception to user-space
2374 * - return from interrupt-handler to user-space
2376 static void __sched
__schedule(void)
2378 struct task_struct
*prev
, *next
;
2379 unsigned long *switch_count
;
2385 cpu
= smp_processor_id();
2387 rcu_note_context_switch(cpu
);
2390 schedule_debug(prev
);
2392 if (sched_feat(HRTICK
))
2395 raw_spin_lock_irq(&rq
->lock
);
2397 switch_count
= &prev
->nivcsw
;
2398 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2399 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2400 prev
->state
= TASK_RUNNING
;
2402 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2406 * If a worker went to sleep, notify and ask workqueue
2407 * whether it wants to wake up a task to maintain
2410 if (prev
->flags
& PF_WQ_WORKER
) {
2411 struct task_struct
*to_wakeup
;
2413 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2415 try_to_wake_up_local(to_wakeup
);
2418 switch_count
= &prev
->nvcsw
;
2421 pre_schedule(rq
, prev
);
2423 if (unlikely(!rq
->nr_running
))
2424 idle_balance(cpu
, rq
);
2426 put_prev_task(rq
, prev
);
2427 next
= pick_next_task(rq
);
2428 clear_tsk_need_resched(prev
);
2429 rq
->skip_clock_update
= 0;
2431 if (likely(prev
!= next
)) {
2436 context_switch(rq
, prev
, next
); /* unlocks the rq */
2438 * The context switch have flipped the stack from under us
2439 * and restored the local variables which were saved when
2440 * this task called schedule() in the past. prev == current
2441 * is still correct, but it can be moved to another cpu/rq.
2443 cpu
= smp_processor_id();
2446 raw_spin_unlock_irq(&rq
->lock
);
2450 sched_preempt_enable_no_resched();
2455 static inline void sched_submit_work(struct task_struct
*tsk
)
2457 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2460 * If we are going to sleep and we have plugged IO queued,
2461 * make sure to submit it to avoid deadlocks.
2463 if (blk_needs_flush_plug(tsk
))
2464 blk_schedule_flush_plug(tsk
);
2467 asmlinkage
void __sched
schedule(void)
2469 struct task_struct
*tsk
= current
;
2471 sched_submit_work(tsk
);
2474 EXPORT_SYMBOL(schedule
);
2476 #ifdef CONFIG_CONTEXT_TRACKING
2477 asmlinkage
void __sched
schedule_user(void)
2480 * If we come here after a random call to set_need_resched(),
2481 * or we have been woken up remotely but the IPI has not yet arrived,
2482 * we haven't yet exited the RCU idle mode. Do it here manually until
2483 * we find a better solution.
2492 * schedule_preempt_disabled - called with preemption disabled
2494 * Returns with preemption disabled. Note: preempt_count must be 1
2496 void __sched
schedule_preempt_disabled(void)
2498 sched_preempt_enable_no_resched();
2503 #ifdef CONFIG_PREEMPT
2505 * this is the entry point to schedule() from in-kernel preemption
2506 * off of preempt_enable. Kernel preemptions off return from interrupt
2507 * occur there and call schedule directly.
2509 asmlinkage
void __sched notrace
preempt_schedule(void)
2511 struct thread_info
*ti
= current_thread_info();
2514 * If there is a non-zero preempt_count or interrupts are disabled,
2515 * we do not want to preempt the current task. Just return..
2517 if (likely(ti
->preempt_count
|| irqs_disabled()))
2521 add_preempt_count_notrace(PREEMPT_ACTIVE
);
2523 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
2526 * Check again in case we missed a preemption opportunity
2527 * between schedule and now.
2530 } while (need_resched());
2532 EXPORT_SYMBOL(preempt_schedule
);
2535 * this is the entry point to schedule() from kernel preemption
2536 * off of irq context.
2537 * Note, that this is called and return with irqs disabled. This will
2538 * protect us against recursive calling from irq.
2540 asmlinkage
void __sched
preempt_schedule_irq(void)
2542 struct thread_info
*ti
= current_thread_info();
2543 enum ctx_state prev_state
;
2545 /* Catch callers which need to be fixed */
2546 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
2548 prev_state
= exception_enter();
2551 add_preempt_count(PREEMPT_ACTIVE
);
2554 local_irq_disable();
2555 sub_preempt_count(PREEMPT_ACTIVE
);
2558 * Check again in case we missed a preemption opportunity
2559 * between schedule and now.
2562 } while (need_resched());
2564 exception_exit(prev_state
);
2567 #endif /* CONFIG_PREEMPT */
2569 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2572 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2574 EXPORT_SYMBOL(default_wake_function
);
2577 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2578 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2579 * number) then we wake all the non-exclusive tasks and one exclusive task.
2581 * There are circumstances in which we can try to wake a task which has already
2582 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2583 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2585 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
2586 int nr_exclusive
, int wake_flags
, void *key
)
2588 wait_queue_t
*curr
, *next
;
2590 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
2591 unsigned flags
= curr
->flags
;
2593 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
2594 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
2600 * __wake_up - wake up threads blocked on a waitqueue.
2602 * @mode: which threads
2603 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2604 * @key: is directly passed to the wakeup function
2606 * It may be assumed that this function implies a write memory barrier before
2607 * changing the task state if and only if any tasks are woken up.
2609 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
2610 int nr_exclusive
, void *key
)
2612 unsigned long flags
;
2614 spin_lock_irqsave(&q
->lock
, flags
);
2615 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
2616 spin_unlock_irqrestore(&q
->lock
, flags
);
2618 EXPORT_SYMBOL(__wake_up
);
2621 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2623 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
2625 __wake_up_common(q
, mode
, nr
, 0, NULL
);
2627 EXPORT_SYMBOL_GPL(__wake_up_locked
);
2629 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
2631 __wake_up_common(q
, mode
, 1, 0, key
);
2633 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
2636 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2638 * @mode: which threads
2639 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2640 * @key: opaque value to be passed to wakeup targets
2642 * The sync wakeup differs that the waker knows that it will schedule
2643 * away soon, so while the target thread will be woken up, it will not
2644 * be migrated to another CPU - ie. the two threads are 'synchronized'
2645 * with each other. This can prevent needless bouncing between CPUs.
2647 * On UP it can prevent extra preemption.
2649 * It may be assumed that this function implies a write memory barrier before
2650 * changing the task state if and only if any tasks are woken up.
2652 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
2653 int nr_exclusive
, void *key
)
2655 unsigned long flags
;
2656 int wake_flags
= WF_SYNC
;
2661 if (unlikely(!nr_exclusive
))
2664 spin_lock_irqsave(&q
->lock
, flags
);
2665 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
2666 spin_unlock_irqrestore(&q
->lock
, flags
);
2668 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
2671 * __wake_up_sync - see __wake_up_sync_key()
2673 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
2675 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
2677 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
2680 * complete: - signals a single thread waiting on this completion
2681 * @x: holds the state of this particular completion
2683 * This will wake up a single thread waiting on this completion. Threads will be
2684 * awakened in the same order in which they were queued.
2686 * See also complete_all(), wait_for_completion() and related routines.
2688 * It may be assumed that this function implies a write memory barrier before
2689 * changing the task state if and only if any tasks are woken up.
2691 void complete(struct completion
*x
)
2693 unsigned long flags
;
2695 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2697 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
2698 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2700 EXPORT_SYMBOL(complete
);
2703 * complete_all: - signals all threads waiting on this completion
2704 * @x: holds the state of this particular completion
2706 * This will wake up all threads waiting on this particular completion event.
2708 * It may be assumed that this function implies a write memory barrier before
2709 * changing the task state if and only if any tasks are woken up.
2711 void complete_all(struct completion
*x
)
2713 unsigned long flags
;
2715 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2716 x
->done
+= UINT_MAX
/2;
2717 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
2718 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2720 EXPORT_SYMBOL(complete_all
);
2722 static inline long __sched
2723 do_wait_for_common(struct completion
*x
,
2724 long (*action
)(long), long timeout
, int state
)
2727 DECLARE_WAITQUEUE(wait
, current
);
2729 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
2731 if (signal_pending_state(state
, current
)) {
2732 timeout
= -ERESTARTSYS
;
2735 __set_current_state(state
);
2736 spin_unlock_irq(&x
->wait
.lock
);
2737 timeout
= action(timeout
);
2738 spin_lock_irq(&x
->wait
.lock
);
2739 } while (!x
->done
&& timeout
);
2740 __remove_wait_queue(&x
->wait
, &wait
);
2745 return timeout
?: 1;
2748 static inline long __sched
2749 __wait_for_common(struct completion
*x
,
2750 long (*action
)(long), long timeout
, int state
)
2754 spin_lock_irq(&x
->wait
.lock
);
2755 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
2756 spin_unlock_irq(&x
->wait
.lock
);
2761 wait_for_common(struct completion
*x
, long timeout
, int state
)
2763 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
2767 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
2769 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
2773 * wait_for_completion: - waits for completion of a task
2774 * @x: holds the state of this particular completion
2776 * This waits to be signaled for completion of a specific task. It is NOT
2777 * interruptible and there is no timeout.
2779 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2780 * and interrupt capability. Also see complete().
2782 void __sched
wait_for_completion(struct completion
*x
)
2784 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
2786 EXPORT_SYMBOL(wait_for_completion
);
2789 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2790 * @x: holds the state of this particular completion
2791 * @timeout: timeout value in jiffies
2793 * This waits for either a completion of a specific task to be signaled or for a
2794 * specified timeout to expire. The timeout is in jiffies. It is not
2797 * The return value is 0 if timed out, and positive (at least 1, or number of
2798 * jiffies left till timeout) if completed.
2800 unsigned long __sched
2801 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
2803 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
2805 EXPORT_SYMBOL(wait_for_completion_timeout
);
2808 * wait_for_completion_io: - waits for completion of a task
2809 * @x: holds the state of this particular completion
2811 * This waits to be signaled for completion of a specific task. It is NOT
2812 * interruptible and there is no timeout. The caller is accounted as waiting
2815 void __sched
wait_for_completion_io(struct completion
*x
)
2817 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
2819 EXPORT_SYMBOL(wait_for_completion_io
);
2822 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2823 * @x: holds the state of this particular completion
2824 * @timeout: timeout value in jiffies
2826 * This waits for either a completion of a specific task to be signaled or for a
2827 * specified timeout to expire. The timeout is in jiffies. It is not
2828 * interruptible. The caller is accounted as waiting for IO.
2830 * The return value is 0 if timed out, and positive (at least 1, or number of
2831 * jiffies left till timeout) if completed.
2833 unsigned long __sched
2834 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
2836 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
2838 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
2841 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2842 * @x: holds the state of this particular completion
2844 * This waits for completion of a specific task to be signaled. It is
2847 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2849 int __sched
wait_for_completion_interruptible(struct completion
*x
)
2851 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
2852 if (t
== -ERESTARTSYS
)
2856 EXPORT_SYMBOL(wait_for_completion_interruptible
);
2859 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2860 * @x: holds the state of this particular completion
2861 * @timeout: timeout value in jiffies
2863 * This waits for either a completion of a specific task to be signaled or for a
2864 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2866 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2867 * positive (at least 1, or number of jiffies left till timeout) if completed.
2870 wait_for_completion_interruptible_timeout(struct completion
*x
,
2871 unsigned long timeout
)
2873 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
2875 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
2878 * wait_for_completion_killable: - waits for completion of a task (killable)
2879 * @x: holds the state of this particular completion
2881 * This waits to be signaled for completion of a specific task. It can be
2882 * interrupted by a kill signal.
2884 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2886 int __sched
wait_for_completion_killable(struct completion
*x
)
2888 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
2889 if (t
== -ERESTARTSYS
)
2893 EXPORT_SYMBOL(wait_for_completion_killable
);
2896 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2897 * @x: holds the state of this particular completion
2898 * @timeout: timeout value in jiffies
2900 * This waits for either a completion of a specific task to be
2901 * signaled or for a specified timeout to expire. It can be
2902 * interrupted by a kill signal. The timeout is in jiffies.
2904 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2905 * positive (at least 1, or number of jiffies left till timeout) if completed.
2908 wait_for_completion_killable_timeout(struct completion
*x
,
2909 unsigned long timeout
)
2911 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
2913 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
2916 * try_wait_for_completion - try to decrement a completion without blocking
2917 * @x: completion structure
2919 * Returns: 0 if a decrement cannot be done without blocking
2920 * 1 if a decrement succeeded.
2922 * If a completion is being used as a counting completion,
2923 * attempt to decrement the counter without blocking. This
2924 * enables us to avoid waiting if the resource the completion
2925 * is protecting is not available.
2927 bool try_wait_for_completion(struct completion
*x
)
2929 unsigned long flags
;
2932 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2937 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2940 EXPORT_SYMBOL(try_wait_for_completion
);
2943 * completion_done - Test to see if a completion has any waiters
2944 * @x: completion structure
2946 * Returns: 0 if there are waiters (wait_for_completion() in progress)
2947 * 1 if there are no waiters.
2950 bool completion_done(struct completion
*x
)
2952 unsigned long flags
;
2955 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2958 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2961 EXPORT_SYMBOL(completion_done
);
2964 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
2966 unsigned long flags
;
2969 init_waitqueue_entry(&wait
, current
);
2971 __set_current_state(state
);
2973 spin_lock_irqsave(&q
->lock
, flags
);
2974 __add_wait_queue(q
, &wait
);
2975 spin_unlock(&q
->lock
);
2976 timeout
= schedule_timeout(timeout
);
2977 spin_lock_irq(&q
->lock
);
2978 __remove_wait_queue(q
, &wait
);
2979 spin_unlock_irqrestore(&q
->lock
, flags
);
2984 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
2986 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2988 EXPORT_SYMBOL(interruptible_sleep_on
);
2991 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
2993 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
2995 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
2997 void __sched
sleep_on(wait_queue_head_t
*q
)
2999 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3001 EXPORT_SYMBOL(sleep_on
);
3003 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3005 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3007 EXPORT_SYMBOL(sleep_on_timeout
);
3009 #ifdef CONFIG_RT_MUTEXES
3012 * rt_mutex_setprio - set the current priority of a task
3014 * @prio: prio value (kernel-internal form)
3016 * This function changes the 'effective' priority of a task. It does
3017 * not touch ->normal_prio like __setscheduler().
3019 * Used by the rt_mutex code to implement priority inheritance logic.
3021 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3023 int oldprio
, on_rq
, running
;
3025 const struct sched_class
*prev_class
;
3027 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3029 rq
= __task_rq_lock(p
);
3032 * Idle task boosting is a nono in general. There is one
3033 * exception, when PREEMPT_RT and NOHZ is active:
3035 * The idle task calls get_next_timer_interrupt() and holds
3036 * the timer wheel base->lock on the CPU and another CPU wants
3037 * to access the timer (probably to cancel it). We can safely
3038 * ignore the boosting request, as the idle CPU runs this code
3039 * with interrupts disabled and will complete the lock
3040 * protected section without being interrupted. So there is no
3041 * real need to boost.
3043 if (unlikely(p
== rq
->idle
)) {
3044 WARN_ON(p
!= rq
->curr
);
3045 WARN_ON(p
->pi_blocked_on
);
3049 trace_sched_pi_setprio(p
, prio
);
3051 prev_class
= p
->sched_class
;
3053 running
= task_current(rq
, p
);
3055 dequeue_task(rq
, p
, 0);
3057 p
->sched_class
->put_prev_task(rq
, p
);
3060 p
->sched_class
= &rt_sched_class
;
3062 p
->sched_class
= &fair_sched_class
;
3067 p
->sched_class
->set_curr_task(rq
);
3069 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3071 check_class_changed(rq
, p
, prev_class
, oldprio
);
3073 __task_rq_unlock(rq
);
3076 void set_user_nice(struct task_struct
*p
, long nice
)
3078 int old_prio
, delta
, on_rq
;
3079 unsigned long flags
;
3082 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3085 * We have to be careful, if called from sys_setpriority(),
3086 * the task might be in the middle of scheduling on another CPU.
3088 rq
= task_rq_lock(p
, &flags
);
3090 * The RT priorities are set via sched_setscheduler(), but we still
3091 * allow the 'normal' nice value to be set - but as expected
3092 * it wont have any effect on scheduling until the task is
3093 * SCHED_FIFO/SCHED_RR:
3095 if (task_has_rt_policy(p
)) {
3096 p
->static_prio
= NICE_TO_PRIO(nice
);
3101 dequeue_task(rq
, p
, 0);
3103 p
->static_prio
= NICE_TO_PRIO(nice
);
3106 p
->prio
= effective_prio(p
);
3107 delta
= p
->prio
- old_prio
;
3110 enqueue_task(rq
, p
, 0);
3112 * If the task increased its priority or is running and
3113 * lowered its priority, then reschedule its CPU:
3115 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3116 resched_task(rq
->curr
);
3119 task_rq_unlock(rq
, p
, &flags
);
3121 EXPORT_SYMBOL(set_user_nice
);
3124 * can_nice - check if a task can reduce its nice value
3128 int can_nice(const struct task_struct
*p
, const int nice
)
3130 /* convert nice value [19,-20] to rlimit style value [1,40] */
3131 int nice_rlim
= 20 - nice
;
3133 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3134 capable(CAP_SYS_NICE
));
3137 #ifdef __ARCH_WANT_SYS_NICE
3140 * sys_nice - change the priority of the current process.
3141 * @increment: priority increment
3143 * sys_setpriority is a more generic, but much slower function that
3144 * does similar things.
3146 SYSCALL_DEFINE1(nice
, int, increment
)
3151 * Setpriority might change our priority at the same moment.
3152 * We don't have to worry. Conceptually one call occurs first
3153 * and we have a single winner.
3155 if (increment
< -40)
3160 nice
= TASK_NICE(current
) + increment
;
3166 if (increment
< 0 && !can_nice(current
, nice
))
3169 retval
= security_task_setnice(current
, nice
);
3173 set_user_nice(current
, nice
);
3180 * task_prio - return the priority value of a given task.
3181 * @p: the task in question.
3183 * This is the priority value as seen by users in /proc.
3184 * RT tasks are offset by -200. Normal tasks are centered
3185 * around 0, value goes from -16 to +15.
3187 int task_prio(const struct task_struct
*p
)
3189 return p
->prio
- MAX_RT_PRIO
;
3193 * task_nice - return the nice value of a given task.
3194 * @p: the task in question.
3196 int task_nice(const struct task_struct
*p
)
3198 return TASK_NICE(p
);
3200 EXPORT_SYMBOL(task_nice
);
3203 * idle_cpu - is a given cpu idle currently?
3204 * @cpu: the processor in question.
3206 int idle_cpu(int cpu
)
3208 struct rq
*rq
= cpu_rq(cpu
);
3210 if (rq
->curr
!= rq
->idle
)
3217 if (!llist_empty(&rq
->wake_list
))
3225 * idle_task - return the idle task for a given cpu.
3226 * @cpu: the processor in question.
3228 struct task_struct
*idle_task(int cpu
)
3230 return cpu_rq(cpu
)->idle
;
3234 * find_process_by_pid - find a process with a matching PID value.
3235 * @pid: the pid in question.
3237 static struct task_struct
*find_process_by_pid(pid_t pid
)
3239 return pid
? find_task_by_vpid(pid
) : current
;
3242 /* Actually do priority change: must hold rq lock. */
3244 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3247 p
->rt_priority
= prio
;
3248 p
->normal_prio
= normal_prio(p
);
3249 /* we are holding p->pi_lock already */
3250 p
->prio
= rt_mutex_getprio(p
);
3251 if (rt_prio(p
->prio
))
3252 p
->sched_class
= &rt_sched_class
;
3254 p
->sched_class
= &fair_sched_class
;
3259 * check the target process has a UID that matches the current process's
3261 static bool check_same_owner(struct task_struct
*p
)
3263 const struct cred
*cred
= current_cred(), *pcred
;
3267 pcred
= __task_cred(p
);
3268 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3269 uid_eq(cred
->euid
, pcred
->uid
));
3274 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3275 const struct sched_param
*param
, bool user
)
3277 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3278 unsigned long flags
;
3279 const struct sched_class
*prev_class
;
3283 /* may grab non-irq protected spin_locks */
3284 BUG_ON(in_interrupt());
3286 /* double check policy once rq lock held */
3288 reset_on_fork
= p
->sched_reset_on_fork
;
3289 policy
= oldpolicy
= p
->policy
;
3291 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3292 policy
&= ~SCHED_RESET_ON_FORK
;
3294 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3295 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3296 policy
!= SCHED_IDLE
)
3301 * Valid priorities for SCHED_FIFO and SCHED_RR are
3302 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3303 * SCHED_BATCH and SCHED_IDLE is 0.
3305 if (param
->sched_priority
< 0 ||
3306 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3307 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3309 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3313 * Allow unprivileged RT tasks to decrease priority:
3315 if (user
&& !capable(CAP_SYS_NICE
)) {
3316 if (rt_policy(policy
)) {
3317 unsigned long rlim_rtprio
=
3318 task_rlimit(p
, RLIMIT_RTPRIO
);
3320 /* can't set/change the rt policy */
3321 if (policy
!= p
->policy
&& !rlim_rtprio
)
3324 /* can't increase priority */
3325 if (param
->sched_priority
> p
->rt_priority
&&
3326 param
->sched_priority
> rlim_rtprio
)
3331 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3332 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3334 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3335 if (!can_nice(p
, TASK_NICE(p
)))
3339 /* can't change other user's priorities */
3340 if (!check_same_owner(p
))
3343 /* Normal users shall not reset the sched_reset_on_fork flag */
3344 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3349 retval
= security_task_setscheduler(p
);
3355 * make sure no PI-waiters arrive (or leave) while we are
3356 * changing the priority of the task:
3358 * To be able to change p->policy safely, the appropriate
3359 * runqueue lock must be held.
3361 rq
= task_rq_lock(p
, &flags
);
3364 * Changing the policy of the stop threads its a very bad idea
3366 if (p
== rq
->stop
) {
3367 task_rq_unlock(rq
, p
, &flags
);
3372 * If not changing anything there's no need to proceed further:
3374 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3375 param
->sched_priority
== p
->rt_priority
))) {
3376 task_rq_unlock(rq
, p
, &flags
);
3380 #ifdef CONFIG_RT_GROUP_SCHED
3383 * Do not allow realtime tasks into groups that have no runtime
3386 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3387 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3388 !task_group_is_autogroup(task_group(p
))) {
3389 task_rq_unlock(rq
, p
, &flags
);
3395 /* recheck policy now with rq lock held */
3396 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3397 policy
= oldpolicy
= -1;
3398 task_rq_unlock(rq
, p
, &flags
);
3402 running
= task_current(rq
, p
);
3404 dequeue_task(rq
, p
, 0);
3406 p
->sched_class
->put_prev_task(rq
, p
);
3408 p
->sched_reset_on_fork
= reset_on_fork
;
3411 prev_class
= p
->sched_class
;
3412 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3415 p
->sched_class
->set_curr_task(rq
);
3417 enqueue_task(rq
, p
, 0);
3419 check_class_changed(rq
, p
, prev_class
, oldprio
);
3420 task_rq_unlock(rq
, p
, &flags
);
3422 rt_mutex_adjust_pi(p
);
3428 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3429 * @p: the task in question.
3430 * @policy: new policy.
3431 * @param: structure containing the new RT priority.
3433 * NOTE that the task may be already dead.
3435 int sched_setscheduler(struct task_struct
*p
, int policy
,
3436 const struct sched_param
*param
)
3438 return __sched_setscheduler(p
, policy
, param
, true);
3440 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3443 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3444 * @p: the task in question.
3445 * @policy: new policy.
3446 * @param: structure containing the new RT priority.
3448 * Just like sched_setscheduler, only don't bother checking if the
3449 * current context has permission. For example, this is needed in
3450 * stop_machine(): we create temporary high priority worker threads,
3451 * but our caller might not have that capability.
3453 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3454 const struct sched_param
*param
)
3456 return __sched_setscheduler(p
, policy
, param
, false);
3460 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3462 struct sched_param lparam
;
3463 struct task_struct
*p
;
3466 if (!param
|| pid
< 0)
3468 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3473 p
= find_process_by_pid(pid
);
3475 retval
= sched_setscheduler(p
, policy
, &lparam
);
3482 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3483 * @pid: the pid in question.
3484 * @policy: new policy.
3485 * @param: structure containing the new RT priority.
3487 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3488 struct sched_param __user
*, param
)
3490 /* negative values for policy are not valid */
3494 return do_sched_setscheduler(pid
, policy
, param
);
3498 * sys_sched_setparam - set/change the RT priority of a thread
3499 * @pid: the pid in question.
3500 * @param: structure containing the new RT priority.
3502 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3504 return do_sched_setscheduler(pid
, -1, param
);
3508 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3509 * @pid: the pid in question.
3511 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3513 struct task_struct
*p
;
3521 p
= find_process_by_pid(pid
);
3523 retval
= security_task_getscheduler(p
);
3526 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3533 * sys_sched_getparam - get the RT priority of a thread
3534 * @pid: the pid in question.
3535 * @param: structure containing the RT priority.
3537 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3539 struct sched_param lp
;
3540 struct task_struct
*p
;
3543 if (!param
|| pid
< 0)
3547 p
= find_process_by_pid(pid
);
3552 retval
= security_task_getscheduler(p
);
3556 lp
.sched_priority
= p
->rt_priority
;
3560 * This one might sleep, we cannot do it with a spinlock held ...
3562 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3571 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3573 cpumask_var_t cpus_allowed
, new_mask
;
3574 struct task_struct
*p
;
3580 p
= find_process_by_pid(pid
);
3587 /* Prevent p going away */
3591 if (p
->flags
& PF_NO_SETAFFINITY
) {
3595 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3599 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3601 goto out_free_cpus_allowed
;
3604 if (!check_same_owner(p
)) {
3606 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3613 retval
= security_task_setscheduler(p
);
3617 cpuset_cpus_allowed(p
, cpus_allowed
);
3618 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3620 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3623 cpuset_cpus_allowed(p
, cpus_allowed
);
3624 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3626 * We must have raced with a concurrent cpuset
3627 * update. Just reset the cpus_allowed to the
3628 * cpuset's cpus_allowed
3630 cpumask_copy(new_mask
, cpus_allowed
);
3635 free_cpumask_var(new_mask
);
3636 out_free_cpus_allowed
:
3637 free_cpumask_var(cpus_allowed
);
3644 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3645 struct cpumask
*new_mask
)
3647 if (len
< cpumask_size())
3648 cpumask_clear(new_mask
);
3649 else if (len
> cpumask_size())
3650 len
= cpumask_size();
3652 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3656 * sys_sched_setaffinity - set the cpu affinity of a process
3657 * @pid: pid of the process
3658 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3659 * @user_mask_ptr: user-space pointer to the new cpu mask
3661 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3662 unsigned long __user
*, user_mask_ptr
)
3664 cpumask_var_t new_mask
;
3667 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3670 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3672 retval
= sched_setaffinity(pid
, new_mask
);
3673 free_cpumask_var(new_mask
);
3677 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
3679 struct task_struct
*p
;
3680 unsigned long flags
;
3687 p
= find_process_by_pid(pid
);
3691 retval
= security_task_getscheduler(p
);
3695 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3696 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
3697 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3707 * sys_sched_getaffinity - get the cpu affinity of a process
3708 * @pid: pid of the process
3709 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3710 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3712 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
3713 unsigned long __user
*, user_mask_ptr
)
3718 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
3720 if (len
& (sizeof(unsigned long)-1))
3723 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
3726 ret
= sched_getaffinity(pid
, mask
);
3728 size_t retlen
= min_t(size_t, len
, cpumask_size());
3730 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
3735 free_cpumask_var(mask
);
3741 * sys_sched_yield - yield the current processor to other threads.
3743 * This function yields the current CPU to other tasks. If there are no
3744 * other threads running on this CPU then this function will return.
3746 SYSCALL_DEFINE0(sched_yield
)
3748 struct rq
*rq
= this_rq_lock();
3750 schedstat_inc(rq
, yld_count
);
3751 current
->sched_class
->yield_task(rq
);
3754 * Since we are going to call schedule() anyway, there's
3755 * no need to preempt or enable interrupts:
3757 __release(rq
->lock
);
3758 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3759 do_raw_spin_unlock(&rq
->lock
);
3760 sched_preempt_enable_no_resched();
3767 static inline int should_resched(void)
3769 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
3772 static void __cond_resched(void)
3774 add_preempt_count(PREEMPT_ACTIVE
);
3776 sub_preempt_count(PREEMPT_ACTIVE
);
3779 int __sched
_cond_resched(void)
3781 if (should_resched()) {
3787 EXPORT_SYMBOL(_cond_resched
);
3790 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3791 * call schedule, and on return reacquire the lock.
3793 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3794 * operations here to prevent schedule() from being called twice (once via
3795 * spin_unlock(), once by hand).
3797 int __cond_resched_lock(spinlock_t
*lock
)
3799 int resched
= should_resched();
3802 lockdep_assert_held(lock
);
3804 if (spin_needbreak(lock
) || resched
) {
3815 EXPORT_SYMBOL(__cond_resched_lock
);
3817 int __sched
__cond_resched_softirq(void)
3819 BUG_ON(!in_softirq());
3821 if (should_resched()) {
3829 EXPORT_SYMBOL(__cond_resched_softirq
);
3832 * yield - yield the current processor to other threads.
3834 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3836 * The scheduler is at all times free to pick the calling task as the most
3837 * eligible task to run, if removing the yield() call from your code breaks
3838 * it, its already broken.
3840 * Typical broken usage is:
3845 * where one assumes that yield() will let 'the other' process run that will
3846 * make event true. If the current task is a SCHED_FIFO task that will never
3847 * happen. Never use yield() as a progress guarantee!!
3849 * If you want to use yield() to wait for something, use wait_event().
3850 * If you want to use yield() to be 'nice' for others, use cond_resched().
3851 * If you still want to use yield(), do not!
3853 void __sched
yield(void)
3855 set_current_state(TASK_RUNNING
);
3858 EXPORT_SYMBOL(yield
);
3861 * yield_to - yield the current processor to another thread in
3862 * your thread group, or accelerate that thread toward the
3863 * processor it's on.
3865 * @preempt: whether task preemption is allowed or not
3867 * It's the caller's job to ensure that the target task struct
3868 * can't go away on us before we can do any checks.
3871 * true (>0) if we indeed boosted the target task.
3872 * false (0) if we failed to boost the target.
3873 * -ESRCH if there's no task to yield to.
3875 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
3877 struct task_struct
*curr
= current
;
3878 struct rq
*rq
, *p_rq
;
3879 unsigned long flags
;
3882 local_irq_save(flags
);
3888 * If we're the only runnable task on the rq and target rq also
3889 * has only one task, there's absolutely no point in yielding.
3891 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
3896 double_rq_lock(rq
, p_rq
);
3897 while (task_rq(p
) != p_rq
) {
3898 double_rq_unlock(rq
, p_rq
);
3902 if (!curr
->sched_class
->yield_to_task
)
3905 if (curr
->sched_class
!= p
->sched_class
)
3908 if (task_running(p_rq
, p
) || p
->state
)
3911 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
3913 schedstat_inc(rq
, yld_count
);
3915 * Make p's CPU reschedule; pick_next_entity takes care of
3918 if (preempt
&& rq
!= p_rq
)
3919 resched_task(p_rq
->curr
);
3923 double_rq_unlock(rq
, p_rq
);
3925 local_irq_restore(flags
);
3932 EXPORT_SYMBOL_GPL(yield_to
);
3935 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3936 * that process accounting knows that this is a task in IO wait state.
3938 void __sched
io_schedule(void)
3940 struct rq
*rq
= raw_rq();
3942 delayacct_blkio_start();
3943 atomic_inc(&rq
->nr_iowait
);
3944 blk_flush_plug(current
);
3945 current
->in_iowait
= 1;
3947 current
->in_iowait
= 0;
3948 atomic_dec(&rq
->nr_iowait
);
3949 delayacct_blkio_end();
3951 EXPORT_SYMBOL(io_schedule
);
3953 long __sched
io_schedule_timeout(long timeout
)
3955 struct rq
*rq
= raw_rq();
3958 delayacct_blkio_start();
3959 atomic_inc(&rq
->nr_iowait
);
3960 blk_flush_plug(current
);
3961 current
->in_iowait
= 1;
3962 ret
= schedule_timeout(timeout
);
3963 current
->in_iowait
= 0;
3964 atomic_dec(&rq
->nr_iowait
);
3965 delayacct_blkio_end();
3970 * sys_sched_get_priority_max - return maximum RT priority.
3971 * @policy: scheduling class.
3973 * this syscall returns the maximum rt_priority that can be used
3974 * by a given scheduling class.
3976 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
3983 ret
= MAX_USER_RT_PRIO
-1;
3995 * sys_sched_get_priority_min - return minimum RT priority.
3996 * @policy: scheduling class.
3998 * this syscall returns the minimum rt_priority that can be used
3999 * by a given scheduling class.
4001 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4019 * sys_sched_rr_get_interval - return the default timeslice of a process.
4020 * @pid: pid of the process.
4021 * @interval: userspace pointer to the timeslice value.
4023 * this syscall writes the default timeslice value of a given process
4024 * into the user-space timespec buffer. A value of '0' means infinity.
4026 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4027 struct timespec __user
*, interval
)
4029 struct task_struct
*p
;
4030 unsigned int time_slice
;
4031 unsigned long flags
;
4041 p
= find_process_by_pid(pid
);
4045 retval
= security_task_getscheduler(p
);
4049 rq
= task_rq_lock(p
, &flags
);
4050 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4051 task_rq_unlock(rq
, p
, &flags
);
4054 jiffies_to_timespec(time_slice
, &t
);
4055 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4063 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4065 void sched_show_task(struct task_struct
*p
)
4067 unsigned long free
= 0;
4071 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4072 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4073 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4074 #if BITS_PER_LONG == 32
4075 if (state
== TASK_RUNNING
)
4076 printk(KERN_CONT
" running ");
4078 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4080 if (state
== TASK_RUNNING
)
4081 printk(KERN_CONT
" running task ");
4083 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4085 #ifdef CONFIG_DEBUG_STACK_USAGE
4086 free
= stack_not_used(p
);
4089 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4091 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4092 task_pid_nr(p
), ppid
,
4093 (unsigned long)task_thread_info(p
)->flags
);
4095 print_worker_info(KERN_INFO
, p
);
4096 show_stack(p
, NULL
);
4099 void show_state_filter(unsigned long state_filter
)
4101 struct task_struct
*g
, *p
;
4103 #if BITS_PER_LONG == 32
4105 " task PC stack pid father\n");
4108 " task PC stack pid father\n");
4111 do_each_thread(g
, p
) {
4113 * reset the NMI-timeout, listing all files on a slow
4114 * console might take a lot of time:
4116 touch_nmi_watchdog();
4117 if (!state_filter
|| (p
->state
& state_filter
))
4119 } while_each_thread(g
, p
);
4121 touch_all_softlockup_watchdogs();
4123 #ifdef CONFIG_SCHED_DEBUG
4124 sysrq_sched_debug_show();
4128 * Only show locks if all tasks are dumped:
4131 debug_show_all_locks();
4134 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
4136 idle
->sched_class
= &idle_sched_class
;
4140 * init_idle - set up an idle thread for a given CPU
4141 * @idle: task in question
4142 * @cpu: cpu the idle task belongs to
4144 * NOTE: this function does not set the idle thread's NEED_RESCHED
4145 * flag, to make booting more robust.
4147 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4149 struct rq
*rq
= cpu_rq(cpu
);
4150 unsigned long flags
;
4152 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4155 idle
->state
= TASK_RUNNING
;
4156 idle
->se
.exec_start
= sched_clock();
4158 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4160 * We're having a chicken and egg problem, even though we are
4161 * holding rq->lock, the cpu isn't yet set to this cpu so the
4162 * lockdep check in task_group() will fail.
4164 * Similar case to sched_fork(). / Alternatively we could
4165 * use task_rq_lock() here and obtain the other rq->lock.
4170 __set_task_cpu(idle
, cpu
);
4173 rq
->curr
= rq
->idle
= idle
;
4174 #if defined(CONFIG_SMP)
4177 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4179 /* Set the preempt count _outside_ the spinlocks! */
4180 task_thread_info(idle
)->preempt_count
= 0;
4183 * The idle tasks have their own, simple scheduling class:
4185 idle
->sched_class
= &idle_sched_class
;
4186 ftrace_graph_init_idle_task(idle
, cpu
);
4187 vtime_init_idle(idle
, cpu
);
4188 #if defined(CONFIG_SMP)
4189 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4194 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4196 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4197 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4199 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4200 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4204 * This is how migration works:
4206 * 1) we invoke migration_cpu_stop() on the target CPU using
4208 * 2) stopper starts to run (implicitly forcing the migrated thread
4210 * 3) it checks whether the migrated task is still in the wrong runqueue.
4211 * 4) if it's in the wrong runqueue then the migration thread removes
4212 * it and puts it into the right queue.
4213 * 5) stopper completes and stop_one_cpu() returns and the migration
4218 * Change a given task's CPU affinity. Migrate the thread to a
4219 * proper CPU and schedule it away if the CPU it's executing on
4220 * is removed from the allowed bitmask.
4222 * NOTE: the caller must have a valid reference to the task, the
4223 * task must not exit() & deallocate itself prematurely. The
4224 * call is not atomic; no spinlocks may be held.
4226 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4228 unsigned long flags
;
4230 unsigned int dest_cpu
;
4233 rq
= task_rq_lock(p
, &flags
);
4235 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4238 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4243 do_set_cpus_allowed(p
, new_mask
);
4245 /* Can the task run on the task's current CPU? If so, we're done */
4246 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4249 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4251 struct migration_arg arg
= { p
, dest_cpu
};
4252 /* Need help from migration thread: drop lock and wait. */
4253 task_rq_unlock(rq
, p
, &flags
);
4254 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4255 tlb_migrate_finish(p
->mm
);
4259 task_rq_unlock(rq
, p
, &flags
);
4263 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4266 * Move (not current) task off this cpu, onto dest cpu. We're doing
4267 * this because either it can't run here any more (set_cpus_allowed()
4268 * away from this CPU, or CPU going down), or because we're
4269 * attempting to rebalance this task on exec (sched_exec).
4271 * So we race with normal scheduler movements, but that's OK, as long
4272 * as the task is no longer on this CPU.
4274 * Returns non-zero if task was successfully migrated.
4276 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4278 struct rq
*rq_dest
, *rq_src
;
4281 if (unlikely(!cpu_active(dest_cpu
)))
4284 rq_src
= cpu_rq(src_cpu
);
4285 rq_dest
= cpu_rq(dest_cpu
);
4287 raw_spin_lock(&p
->pi_lock
);
4288 double_rq_lock(rq_src
, rq_dest
);
4289 /* Already moved. */
4290 if (task_cpu(p
) != src_cpu
)
4292 /* Affinity changed (again). */
4293 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4297 * If we're not on a rq, the next wake-up will ensure we're
4301 dequeue_task(rq_src
, p
, 0);
4302 set_task_cpu(p
, dest_cpu
);
4303 enqueue_task(rq_dest
, p
, 0);
4304 check_preempt_curr(rq_dest
, p
, 0);
4309 double_rq_unlock(rq_src
, rq_dest
);
4310 raw_spin_unlock(&p
->pi_lock
);
4315 * migration_cpu_stop - this will be executed by a highprio stopper thread
4316 * and performs thread migration by bumping thread off CPU then
4317 * 'pushing' onto another runqueue.
4319 static int migration_cpu_stop(void *data
)
4321 struct migration_arg
*arg
= data
;
4324 * The original target cpu might have gone down and we might
4325 * be on another cpu but it doesn't matter.
4327 local_irq_disable();
4328 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4333 #ifdef CONFIG_HOTPLUG_CPU
4336 * Ensures that the idle task is using init_mm right before its cpu goes
4339 void idle_task_exit(void)
4341 struct mm_struct
*mm
= current
->active_mm
;
4343 BUG_ON(cpu_online(smp_processor_id()));
4346 switch_mm(mm
, &init_mm
, current
);
4351 * Since this CPU is going 'away' for a while, fold any nr_active delta
4352 * we might have. Assumes we're called after migrate_tasks() so that the
4353 * nr_active count is stable.
4355 * Also see the comment "Global load-average calculations".
4357 static void calc_load_migrate(struct rq
*rq
)
4359 long delta
= calc_load_fold_active(rq
);
4361 atomic_long_add(delta
, &calc_load_tasks
);
4365 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4366 * try_to_wake_up()->select_task_rq().
4368 * Called with rq->lock held even though we'er in stop_machine() and
4369 * there's no concurrency possible, we hold the required locks anyway
4370 * because of lock validation efforts.
4372 static void migrate_tasks(unsigned int dead_cpu
)
4374 struct rq
*rq
= cpu_rq(dead_cpu
);
4375 struct task_struct
*next
, *stop
= rq
->stop
;
4379 * Fudge the rq selection such that the below task selection loop
4380 * doesn't get stuck on the currently eligible stop task.
4382 * We're currently inside stop_machine() and the rq is either stuck
4383 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4384 * either way we should never end up calling schedule() until we're
4390 * put_prev_task() and pick_next_task() sched
4391 * class method both need to have an up-to-date
4392 * value of rq->clock[_task]
4394 update_rq_clock(rq
);
4398 * There's this thread running, bail when that's the only
4401 if (rq
->nr_running
== 1)
4404 next
= pick_next_task(rq
);
4406 next
->sched_class
->put_prev_task(rq
, next
);
4408 /* Find suitable destination for @next, with force if needed. */
4409 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4410 raw_spin_unlock(&rq
->lock
);
4412 __migrate_task(next
, dead_cpu
, dest_cpu
);
4414 raw_spin_lock(&rq
->lock
);
4420 #endif /* CONFIG_HOTPLUG_CPU */
4422 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4424 static struct ctl_table sd_ctl_dir
[] = {
4426 .procname
= "sched_domain",
4432 static struct ctl_table sd_ctl_root
[] = {
4434 .procname
= "kernel",
4436 .child
= sd_ctl_dir
,
4441 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4443 struct ctl_table
*entry
=
4444 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4449 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4451 struct ctl_table
*entry
;
4454 * In the intermediate directories, both the child directory and
4455 * procname are dynamically allocated and could fail but the mode
4456 * will always be set. In the lowest directory the names are
4457 * static strings and all have proc handlers.
4459 for (entry
= *tablep
; entry
->mode
; entry
++) {
4461 sd_free_ctl_entry(&entry
->child
);
4462 if (entry
->proc_handler
== NULL
)
4463 kfree(entry
->procname
);
4470 static int min_load_idx
= 0;
4471 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4474 set_table_entry(struct ctl_table
*entry
,
4475 const char *procname
, void *data
, int maxlen
,
4476 umode_t mode
, proc_handler
*proc_handler
,
4479 entry
->procname
= procname
;
4481 entry
->maxlen
= maxlen
;
4483 entry
->proc_handler
= proc_handler
;
4486 entry
->extra1
= &min_load_idx
;
4487 entry
->extra2
= &max_load_idx
;
4491 static struct ctl_table
*
4492 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4494 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4499 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4500 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4501 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4502 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4503 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4504 sizeof(int), 0644, proc_dointvec_minmax
, true);
4505 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4506 sizeof(int), 0644, proc_dointvec_minmax
, true);
4507 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4508 sizeof(int), 0644, proc_dointvec_minmax
, true);
4509 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4510 sizeof(int), 0644, proc_dointvec_minmax
, true);
4511 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4512 sizeof(int), 0644, proc_dointvec_minmax
, true);
4513 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4514 sizeof(int), 0644, proc_dointvec_minmax
, false);
4515 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4516 sizeof(int), 0644, proc_dointvec_minmax
, false);
4517 set_table_entry(&table
[9], "cache_nice_tries",
4518 &sd
->cache_nice_tries
,
4519 sizeof(int), 0644, proc_dointvec_minmax
, false);
4520 set_table_entry(&table
[10], "flags", &sd
->flags
,
4521 sizeof(int), 0644, proc_dointvec_minmax
, false);
4522 set_table_entry(&table
[11], "name", sd
->name
,
4523 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4524 /* &table[12] is terminator */
4529 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4531 struct ctl_table
*entry
, *table
;
4532 struct sched_domain
*sd
;
4533 int domain_num
= 0, i
;
4536 for_each_domain(cpu
, sd
)
4538 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4543 for_each_domain(cpu
, sd
) {
4544 snprintf(buf
, 32, "domain%d", i
);
4545 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4547 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4554 static struct ctl_table_header
*sd_sysctl_header
;
4555 static void register_sched_domain_sysctl(void)
4557 int i
, cpu_num
= num_possible_cpus();
4558 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4561 WARN_ON(sd_ctl_dir
[0].child
);
4562 sd_ctl_dir
[0].child
= entry
;
4567 for_each_possible_cpu(i
) {
4568 snprintf(buf
, 32, "cpu%d", i
);
4569 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4571 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4575 WARN_ON(sd_sysctl_header
);
4576 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4579 /* may be called multiple times per register */
4580 static void unregister_sched_domain_sysctl(void)
4582 if (sd_sysctl_header
)
4583 unregister_sysctl_table(sd_sysctl_header
);
4584 sd_sysctl_header
= NULL
;
4585 if (sd_ctl_dir
[0].child
)
4586 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4589 static void register_sched_domain_sysctl(void)
4592 static void unregister_sched_domain_sysctl(void)
4597 static void set_rq_online(struct rq
*rq
)
4600 const struct sched_class
*class;
4602 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
4605 for_each_class(class) {
4606 if (class->rq_online
)
4607 class->rq_online(rq
);
4612 static void set_rq_offline(struct rq
*rq
)
4615 const struct sched_class
*class;
4617 for_each_class(class) {
4618 if (class->rq_offline
)
4619 class->rq_offline(rq
);
4622 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
4628 * migration_call - callback that gets triggered when a CPU is added.
4629 * Here we can start up the necessary migration thread for the new CPU.
4631 static int __cpuinit
4632 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
4634 int cpu
= (long)hcpu
;
4635 unsigned long flags
;
4636 struct rq
*rq
= cpu_rq(cpu
);
4638 switch (action
& ~CPU_TASKS_FROZEN
) {
4640 case CPU_UP_PREPARE
:
4641 rq
->calc_load_update
= calc_load_update
;
4645 /* Update our root-domain */
4646 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4648 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
4652 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4655 #ifdef CONFIG_HOTPLUG_CPU
4657 sched_ttwu_pending();
4658 /* Update our root-domain */
4659 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4661 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
4665 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
4666 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4670 calc_load_migrate(rq
);
4675 update_max_interval();
4681 * Register at high priority so that task migration (migrate_all_tasks)
4682 * happens before everything else. This has to be lower priority than
4683 * the notifier in the perf_event subsystem, though.
4685 static struct notifier_block __cpuinitdata migration_notifier
= {
4686 .notifier_call
= migration_call
,
4687 .priority
= CPU_PRI_MIGRATION
,
4690 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
4691 unsigned long action
, void *hcpu
)
4693 switch (action
& ~CPU_TASKS_FROZEN
) {
4695 case CPU_DOWN_FAILED
:
4696 set_cpu_active((long)hcpu
, true);
4703 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
4704 unsigned long action
, void *hcpu
)
4706 switch (action
& ~CPU_TASKS_FROZEN
) {
4707 case CPU_DOWN_PREPARE
:
4708 set_cpu_active((long)hcpu
, false);
4715 static int __init
migration_init(void)
4717 void *cpu
= (void *)(long)smp_processor_id();
4720 /* Initialize migration for the boot CPU */
4721 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4722 BUG_ON(err
== NOTIFY_BAD
);
4723 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4724 register_cpu_notifier(&migration_notifier
);
4726 /* Register cpu active notifiers */
4727 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
4728 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
4732 early_initcall(migration_init
);
4737 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
4739 #ifdef CONFIG_SCHED_DEBUG
4741 static __read_mostly
int sched_debug_enabled
;
4743 static int __init
sched_debug_setup(char *str
)
4745 sched_debug_enabled
= 1;
4749 early_param("sched_debug", sched_debug_setup
);
4751 static inline bool sched_debug(void)
4753 return sched_debug_enabled
;
4756 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
4757 struct cpumask
*groupmask
)
4759 struct sched_group
*group
= sd
->groups
;
4762 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
4763 cpumask_clear(groupmask
);
4765 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
4767 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4768 printk("does not load-balance\n");
4770 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
4775 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
4777 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
4778 printk(KERN_ERR
"ERROR: domain->span does not contain "
4781 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
4782 printk(KERN_ERR
"ERROR: domain->groups does not contain"
4786 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
4790 printk(KERN_ERR
"ERROR: group is NULL\n");
4795 * Even though we initialize ->power to something semi-sane,
4796 * we leave power_orig unset. This allows us to detect if
4797 * domain iteration is still funny without causing /0 traps.
4799 if (!group
->sgp
->power_orig
) {
4800 printk(KERN_CONT
"\n");
4801 printk(KERN_ERR
"ERROR: domain->cpu_power not "
4806 if (!cpumask_weight(sched_group_cpus(group
))) {
4807 printk(KERN_CONT
"\n");
4808 printk(KERN_ERR
"ERROR: empty group\n");
4812 if (!(sd
->flags
& SD_OVERLAP
) &&
4813 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
4814 printk(KERN_CONT
"\n");
4815 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4819 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
4821 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
4823 printk(KERN_CONT
" %s", str
);
4824 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
4825 printk(KERN_CONT
" (cpu_power = %d)",
4829 group
= group
->next
;
4830 } while (group
!= sd
->groups
);
4831 printk(KERN_CONT
"\n");
4833 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
4834 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4837 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
4838 printk(KERN_ERR
"ERROR: parent span is not a superset "
4839 "of domain->span\n");
4843 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4847 if (!sched_debug_enabled
)
4851 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4855 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4858 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
4866 #else /* !CONFIG_SCHED_DEBUG */
4867 # define sched_domain_debug(sd, cpu) do { } while (0)
4868 static inline bool sched_debug(void)
4872 #endif /* CONFIG_SCHED_DEBUG */
4874 static int sd_degenerate(struct sched_domain
*sd
)
4876 if (cpumask_weight(sched_domain_span(sd
)) == 1)
4879 /* Following flags need at least 2 groups */
4880 if (sd
->flags
& (SD_LOAD_BALANCE
|
4881 SD_BALANCE_NEWIDLE
|
4885 SD_SHARE_PKG_RESOURCES
)) {
4886 if (sd
->groups
!= sd
->groups
->next
)
4890 /* Following flags don't use groups */
4891 if (sd
->flags
& (SD_WAKE_AFFINE
))
4898 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
4900 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4902 if (sd_degenerate(parent
))
4905 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
4908 /* Flags needing groups don't count if only 1 group in parent */
4909 if (parent
->groups
== parent
->groups
->next
) {
4910 pflags
&= ~(SD_LOAD_BALANCE
|
4911 SD_BALANCE_NEWIDLE
|
4915 SD_SHARE_PKG_RESOURCES
);
4916 if (nr_node_ids
== 1)
4917 pflags
&= ~SD_SERIALIZE
;
4919 if (~cflags
& pflags
)
4925 static void free_rootdomain(struct rcu_head
*rcu
)
4927 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
4929 cpupri_cleanup(&rd
->cpupri
);
4930 free_cpumask_var(rd
->rto_mask
);
4931 free_cpumask_var(rd
->online
);
4932 free_cpumask_var(rd
->span
);
4936 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
4938 struct root_domain
*old_rd
= NULL
;
4939 unsigned long flags
;
4941 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4946 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
4949 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
4952 * If we dont want to free the old_rt yet then
4953 * set old_rd to NULL to skip the freeing later
4956 if (!atomic_dec_and_test(&old_rd
->refcount
))
4960 atomic_inc(&rd
->refcount
);
4963 cpumask_set_cpu(rq
->cpu
, rd
->span
);
4964 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
4967 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4970 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
4973 static int init_rootdomain(struct root_domain
*rd
)
4975 memset(rd
, 0, sizeof(*rd
));
4977 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
4979 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
4981 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
4984 if (cpupri_init(&rd
->cpupri
) != 0)
4989 free_cpumask_var(rd
->rto_mask
);
4991 free_cpumask_var(rd
->online
);
4993 free_cpumask_var(rd
->span
);
4999 * By default the system creates a single root-domain with all cpus as
5000 * members (mimicking the global state we have today).
5002 struct root_domain def_root_domain
;
5004 static void init_defrootdomain(void)
5006 init_rootdomain(&def_root_domain
);
5008 atomic_set(&def_root_domain
.refcount
, 1);
5011 static struct root_domain
*alloc_rootdomain(void)
5013 struct root_domain
*rd
;
5015 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5019 if (init_rootdomain(rd
) != 0) {
5027 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5029 struct sched_group
*tmp
, *first
;
5038 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5043 } while (sg
!= first
);
5046 static void free_sched_domain(struct rcu_head
*rcu
)
5048 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5051 * If its an overlapping domain it has private groups, iterate and
5054 if (sd
->flags
& SD_OVERLAP
) {
5055 free_sched_groups(sd
->groups
, 1);
5056 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5057 kfree(sd
->groups
->sgp
);
5063 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5065 call_rcu(&sd
->rcu
, free_sched_domain
);
5068 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5070 for (; sd
; sd
= sd
->parent
)
5071 destroy_sched_domain(sd
, cpu
);
5075 * Keep a special pointer to the highest sched_domain that has
5076 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5077 * allows us to avoid some pointer chasing select_idle_sibling().
5079 * Also keep a unique ID per domain (we use the first cpu number in
5080 * the cpumask of the domain), this allows us to quickly tell if
5081 * two cpus are in the same cache domain, see cpus_share_cache().
5083 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5084 DEFINE_PER_CPU(int, sd_llc_id
);
5086 static void update_top_cache_domain(int cpu
)
5088 struct sched_domain
*sd
;
5091 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5093 id
= cpumask_first(sched_domain_span(sd
));
5095 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5096 per_cpu(sd_llc_id
, cpu
) = id
;
5100 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5101 * hold the hotplug lock.
5104 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5106 struct rq
*rq
= cpu_rq(cpu
);
5107 struct sched_domain
*tmp
;
5109 /* Remove the sched domains which do not contribute to scheduling. */
5110 for (tmp
= sd
; tmp
; ) {
5111 struct sched_domain
*parent
= tmp
->parent
;
5115 if (sd_parent_degenerate(tmp
, parent
)) {
5116 tmp
->parent
= parent
->parent
;
5118 parent
->parent
->child
= tmp
;
5119 destroy_sched_domain(parent
, cpu
);
5124 if (sd
&& sd_degenerate(sd
)) {
5127 destroy_sched_domain(tmp
, cpu
);
5132 sched_domain_debug(sd
, cpu
);
5134 rq_attach_root(rq
, rd
);
5136 rcu_assign_pointer(rq
->sd
, sd
);
5137 destroy_sched_domains(tmp
, cpu
);
5139 update_top_cache_domain(cpu
);
5142 /* cpus with isolated domains */
5143 static cpumask_var_t cpu_isolated_map
;
5145 /* Setup the mask of cpus configured for isolated domains */
5146 static int __init
isolated_cpu_setup(char *str
)
5148 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5149 cpulist_parse(str
, cpu_isolated_map
);
5153 __setup("isolcpus=", isolated_cpu_setup
);
5155 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5157 return cpumask_of_node(cpu_to_node(cpu
));
5161 struct sched_domain
**__percpu sd
;
5162 struct sched_group
**__percpu sg
;
5163 struct sched_group_power
**__percpu sgp
;
5167 struct sched_domain
** __percpu sd
;
5168 struct root_domain
*rd
;
5178 struct sched_domain_topology_level
;
5180 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5181 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5183 #define SDTL_OVERLAP 0x01
5185 struct sched_domain_topology_level
{
5186 sched_domain_init_f init
;
5187 sched_domain_mask_f mask
;
5190 struct sd_data data
;
5194 * Build an iteration mask that can exclude certain CPUs from the upwards
5197 * Asymmetric node setups can result in situations where the domain tree is of
5198 * unequal depth, make sure to skip domains that already cover the entire
5201 * In that case build_sched_domains() will have terminated the iteration early
5202 * and our sibling sd spans will be empty. Domains should always include the
5203 * cpu they're built on, so check that.
5206 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5208 const struct cpumask
*span
= sched_domain_span(sd
);
5209 struct sd_data
*sdd
= sd
->private;
5210 struct sched_domain
*sibling
;
5213 for_each_cpu(i
, span
) {
5214 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5215 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5218 cpumask_set_cpu(i
, sched_group_mask(sg
));
5223 * Return the canonical balance cpu for this group, this is the first cpu
5224 * of this group that's also in the iteration mask.
5226 int group_balance_cpu(struct sched_group
*sg
)
5228 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5232 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5234 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5235 const struct cpumask
*span
= sched_domain_span(sd
);
5236 struct cpumask
*covered
= sched_domains_tmpmask
;
5237 struct sd_data
*sdd
= sd
->private;
5238 struct sched_domain
*child
;
5241 cpumask_clear(covered
);
5243 for_each_cpu(i
, span
) {
5244 struct cpumask
*sg_span
;
5246 if (cpumask_test_cpu(i
, covered
))
5249 child
= *per_cpu_ptr(sdd
->sd
, i
);
5251 /* See the comment near build_group_mask(). */
5252 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5255 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5256 GFP_KERNEL
, cpu_to_node(cpu
));
5261 sg_span
= sched_group_cpus(sg
);
5263 child
= child
->child
;
5264 cpumask_copy(sg_span
, sched_domain_span(child
));
5266 cpumask_set_cpu(i
, sg_span
);
5268 cpumask_or(covered
, covered
, sg_span
);
5270 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5271 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5272 build_group_mask(sd
, sg
);
5275 * Initialize sgp->power such that even if we mess up the
5276 * domains and no possible iteration will get us here, we won't
5279 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5282 * Make sure the first group of this domain contains the
5283 * canonical balance cpu. Otherwise the sched_domain iteration
5284 * breaks. See update_sg_lb_stats().
5286 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5287 group_balance_cpu(sg
) == cpu
)
5297 sd
->groups
= groups
;
5302 free_sched_groups(first
, 0);
5307 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5309 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5310 struct sched_domain
*child
= sd
->child
;
5313 cpu
= cpumask_first(sched_domain_span(child
));
5316 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5317 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5318 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5325 * build_sched_groups will build a circular linked list of the groups
5326 * covered by the given span, and will set each group's ->cpumask correctly,
5327 * and ->cpu_power to 0.
5329 * Assumes the sched_domain tree is fully constructed
5332 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5334 struct sched_group
*first
= NULL
, *last
= NULL
;
5335 struct sd_data
*sdd
= sd
->private;
5336 const struct cpumask
*span
= sched_domain_span(sd
);
5337 struct cpumask
*covered
;
5340 get_group(cpu
, sdd
, &sd
->groups
);
5341 atomic_inc(&sd
->groups
->ref
);
5343 if (cpu
!= cpumask_first(span
))
5346 lockdep_assert_held(&sched_domains_mutex
);
5347 covered
= sched_domains_tmpmask
;
5349 cpumask_clear(covered
);
5351 for_each_cpu(i
, span
) {
5352 struct sched_group
*sg
;
5355 if (cpumask_test_cpu(i
, covered
))
5358 group
= get_group(i
, sdd
, &sg
);
5359 cpumask_clear(sched_group_cpus(sg
));
5361 cpumask_setall(sched_group_mask(sg
));
5363 for_each_cpu(j
, span
) {
5364 if (get_group(j
, sdd
, NULL
) != group
)
5367 cpumask_set_cpu(j
, covered
);
5368 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5383 * Initialize sched groups cpu_power.
5385 * cpu_power indicates the capacity of sched group, which is used while
5386 * distributing the load between different sched groups in a sched domain.
5387 * Typically cpu_power for all the groups in a sched domain will be same unless
5388 * there are asymmetries in the topology. If there are asymmetries, group
5389 * having more cpu_power will pickup more load compared to the group having
5392 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5394 struct sched_group
*sg
= sd
->groups
;
5399 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5401 } while (sg
!= sd
->groups
);
5403 if (cpu
!= group_balance_cpu(sg
))
5406 update_group_power(sd
, cpu
);
5407 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5410 int __weak
arch_sd_sibling_asym_packing(void)
5412 return 0*SD_ASYM_PACKING
;
5416 * Initializers for schedule domains
5417 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5420 #ifdef CONFIG_SCHED_DEBUG
5421 # define SD_INIT_NAME(sd, type) sd->name = #type
5423 # define SD_INIT_NAME(sd, type) do { } while (0)
5426 #define SD_INIT_FUNC(type) \
5427 static noinline struct sched_domain * \
5428 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5430 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5431 *sd = SD_##type##_INIT; \
5432 SD_INIT_NAME(sd, type); \
5433 sd->private = &tl->data; \
5438 #ifdef CONFIG_SCHED_SMT
5439 SD_INIT_FUNC(SIBLING
)
5441 #ifdef CONFIG_SCHED_MC
5444 #ifdef CONFIG_SCHED_BOOK
5448 static int default_relax_domain_level
= -1;
5449 int sched_domain_level_max
;
5451 static int __init
setup_relax_domain_level(char *str
)
5453 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5454 pr_warn("Unable to set relax_domain_level\n");
5458 __setup("relax_domain_level=", setup_relax_domain_level
);
5460 static void set_domain_attribute(struct sched_domain
*sd
,
5461 struct sched_domain_attr
*attr
)
5465 if (!attr
|| attr
->relax_domain_level
< 0) {
5466 if (default_relax_domain_level
< 0)
5469 request
= default_relax_domain_level
;
5471 request
= attr
->relax_domain_level
;
5472 if (request
< sd
->level
) {
5473 /* turn off idle balance on this domain */
5474 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5476 /* turn on idle balance on this domain */
5477 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5481 static void __sdt_free(const struct cpumask
*cpu_map
);
5482 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5484 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5485 const struct cpumask
*cpu_map
)
5489 if (!atomic_read(&d
->rd
->refcount
))
5490 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5492 free_percpu(d
->sd
); /* fall through */
5494 __sdt_free(cpu_map
); /* fall through */
5500 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5501 const struct cpumask
*cpu_map
)
5503 memset(d
, 0, sizeof(*d
));
5505 if (__sdt_alloc(cpu_map
))
5506 return sa_sd_storage
;
5507 d
->sd
= alloc_percpu(struct sched_domain
*);
5509 return sa_sd_storage
;
5510 d
->rd
= alloc_rootdomain();
5513 return sa_rootdomain
;
5517 * NULL the sd_data elements we've used to build the sched_domain and
5518 * sched_group structure so that the subsequent __free_domain_allocs()
5519 * will not free the data we're using.
5521 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5523 struct sd_data
*sdd
= sd
->private;
5525 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5526 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5528 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5529 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5531 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5532 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5535 #ifdef CONFIG_SCHED_SMT
5536 static const struct cpumask
*cpu_smt_mask(int cpu
)
5538 return topology_thread_cpumask(cpu
);
5543 * Topology list, bottom-up.
5545 static struct sched_domain_topology_level default_topology
[] = {
5546 #ifdef CONFIG_SCHED_SMT
5547 { sd_init_SIBLING
, cpu_smt_mask
, },
5549 #ifdef CONFIG_SCHED_MC
5550 { sd_init_MC
, cpu_coregroup_mask
, },
5552 #ifdef CONFIG_SCHED_BOOK
5553 { sd_init_BOOK
, cpu_book_mask
, },
5555 { sd_init_CPU
, cpu_cpu_mask
, },
5559 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5561 #define for_each_sd_topology(tl) \
5562 for (tl = sched_domain_topology; tl->init; tl++)
5566 static int sched_domains_numa_levels
;
5567 static int *sched_domains_numa_distance
;
5568 static struct cpumask
***sched_domains_numa_masks
;
5569 static int sched_domains_curr_level
;
5571 static inline int sd_local_flags(int level
)
5573 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5576 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
5579 static struct sched_domain
*
5580 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
5582 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5583 int level
= tl
->numa_level
;
5584 int sd_weight
= cpumask_weight(
5585 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
5587 *sd
= (struct sched_domain
){
5588 .min_interval
= sd_weight
,
5589 .max_interval
= 2*sd_weight
,
5591 .imbalance_pct
= 125,
5592 .cache_nice_tries
= 2,
5599 .flags
= 1*SD_LOAD_BALANCE
5600 | 1*SD_BALANCE_NEWIDLE
5605 | 0*SD_SHARE_CPUPOWER
5606 | 0*SD_SHARE_PKG_RESOURCES
5608 | 0*SD_PREFER_SIBLING
5609 | sd_local_flags(level
)
5611 .last_balance
= jiffies
,
5612 .balance_interval
= sd_weight
,
5614 SD_INIT_NAME(sd
, NUMA
);
5615 sd
->private = &tl
->data
;
5618 * Ugly hack to pass state to sd_numa_mask()...
5620 sched_domains_curr_level
= tl
->numa_level
;
5625 static const struct cpumask
*sd_numa_mask(int cpu
)
5627 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
5630 static void sched_numa_warn(const char *str
)
5632 static int done
= false;
5640 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
5642 for (i
= 0; i
< nr_node_ids
; i
++) {
5643 printk(KERN_WARNING
" ");
5644 for (j
= 0; j
< nr_node_ids
; j
++)
5645 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
5646 printk(KERN_CONT
"\n");
5648 printk(KERN_WARNING
"\n");
5651 static bool find_numa_distance(int distance
)
5655 if (distance
== node_distance(0, 0))
5658 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5659 if (sched_domains_numa_distance
[i
] == distance
)
5666 static void sched_init_numa(void)
5668 int next_distance
, curr_distance
= node_distance(0, 0);
5669 struct sched_domain_topology_level
*tl
;
5673 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
5674 if (!sched_domains_numa_distance
)
5678 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5679 * unique distances in the node_distance() table.
5681 * Assumes node_distance(0,j) includes all distances in
5682 * node_distance(i,j) in order to avoid cubic time.
5684 next_distance
= curr_distance
;
5685 for (i
= 0; i
< nr_node_ids
; i
++) {
5686 for (j
= 0; j
< nr_node_ids
; j
++) {
5687 for (k
= 0; k
< nr_node_ids
; k
++) {
5688 int distance
= node_distance(i
, k
);
5690 if (distance
> curr_distance
&&
5691 (distance
< next_distance
||
5692 next_distance
== curr_distance
))
5693 next_distance
= distance
;
5696 * While not a strong assumption it would be nice to know
5697 * about cases where if node A is connected to B, B is not
5698 * equally connected to A.
5700 if (sched_debug() && node_distance(k
, i
) != distance
)
5701 sched_numa_warn("Node-distance not symmetric");
5703 if (sched_debug() && i
&& !find_numa_distance(distance
))
5704 sched_numa_warn("Node-0 not representative");
5706 if (next_distance
!= curr_distance
) {
5707 sched_domains_numa_distance
[level
++] = next_distance
;
5708 sched_domains_numa_levels
= level
;
5709 curr_distance
= next_distance
;
5714 * In case of sched_debug() we verify the above assumption.
5720 * 'level' contains the number of unique distances, excluding the
5721 * identity distance node_distance(i,i).
5723 * The sched_domains_numa_distance[] array includes the actual distance
5728 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5729 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5730 * the array will contain less then 'level' members. This could be
5731 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5732 * in other functions.
5734 * We reset it to 'level' at the end of this function.
5736 sched_domains_numa_levels
= 0;
5738 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
5739 if (!sched_domains_numa_masks
)
5743 * Now for each level, construct a mask per node which contains all
5744 * cpus of nodes that are that many hops away from us.
5746 for (i
= 0; i
< level
; i
++) {
5747 sched_domains_numa_masks
[i
] =
5748 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
5749 if (!sched_domains_numa_masks
[i
])
5752 for (j
= 0; j
< nr_node_ids
; j
++) {
5753 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
5757 sched_domains_numa_masks
[i
][j
] = mask
;
5759 for (k
= 0; k
< nr_node_ids
; k
++) {
5760 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
5763 cpumask_or(mask
, mask
, cpumask_of_node(k
));
5768 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
5769 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
5774 * Copy the default topology bits..
5776 for (i
= 0; default_topology
[i
].init
; i
++)
5777 tl
[i
] = default_topology
[i
];
5780 * .. and append 'j' levels of NUMA goodness.
5782 for (j
= 0; j
< level
; i
++, j
++) {
5783 tl
[i
] = (struct sched_domain_topology_level
){
5784 .init
= sd_numa_init
,
5785 .mask
= sd_numa_mask
,
5786 .flags
= SDTL_OVERLAP
,
5791 sched_domain_topology
= tl
;
5793 sched_domains_numa_levels
= level
;
5796 static void sched_domains_numa_masks_set(int cpu
)
5799 int node
= cpu_to_node(cpu
);
5801 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5802 for (j
= 0; j
< nr_node_ids
; j
++) {
5803 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
5804 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
5809 static void sched_domains_numa_masks_clear(int cpu
)
5812 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5813 for (j
= 0; j
< nr_node_ids
; j
++)
5814 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
5819 * Update sched_domains_numa_masks[level][node] array when new cpus
5822 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
5823 unsigned long action
,
5826 int cpu
= (long)hcpu
;
5828 switch (action
& ~CPU_TASKS_FROZEN
) {
5830 sched_domains_numa_masks_set(cpu
);
5834 sched_domains_numa_masks_clear(cpu
);
5844 static inline void sched_init_numa(void)
5848 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
5849 unsigned long action
,
5854 #endif /* CONFIG_NUMA */
5856 static int __sdt_alloc(const struct cpumask
*cpu_map
)
5858 struct sched_domain_topology_level
*tl
;
5861 for_each_sd_topology(tl
) {
5862 struct sd_data
*sdd
= &tl
->data
;
5864 sdd
->sd
= alloc_percpu(struct sched_domain
*);
5868 sdd
->sg
= alloc_percpu(struct sched_group
*);
5872 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
5876 for_each_cpu(j
, cpu_map
) {
5877 struct sched_domain
*sd
;
5878 struct sched_group
*sg
;
5879 struct sched_group_power
*sgp
;
5881 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
5882 GFP_KERNEL
, cpu_to_node(j
));
5886 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
5888 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5889 GFP_KERNEL
, cpu_to_node(j
));
5895 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
5897 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
5898 GFP_KERNEL
, cpu_to_node(j
));
5902 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
5909 static void __sdt_free(const struct cpumask
*cpu_map
)
5911 struct sched_domain_topology_level
*tl
;
5914 for_each_sd_topology(tl
) {
5915 struct sd_data
*sdd
= &tl
->data
;
5917 for_each_cpu(j
, cpu_map
) {
5918 struct sched_domain
*sd
;
5921 sd
= *per_cpu_ptr(sdd
->sd
, j
);
5922 if (sd
&& (sd
->flags
& SD_OVERLAP
))
5923 free_sched_groups(sd
->groups
, 0);
5924 kfree(*per_cpu_ptr(sdd
->sd
, j
));
5928 kfree(*per_cpu_ptr(sdd
->sg
, j
));
5930 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
5932 free_percpu(sdd
->sd
);
5934 free_percpu(sdd
->sg
);
5936 free_percpu(sdd
->sgp
);
5941 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
5942 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
5943 struct sched_domain
*child
, int cpu
)
5945 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
5949 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
5951 sd
->level
= child
->level
+ 1;
5952 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
5956 set_domain_attribute(sd
, attr
);
5962 * Build sched domains for a given set of cpus and attach the sched domains
5963 * to the individual cpus
5965 static int build_sched_domains(const struct cpumask
*cpu_map
,
5966 struct sched_domain_attr
*attr
)
5968 enum s_alloc alloc_state
;
5969 struct sched_domain
*sd
;
5971 int i
, ret
= -ENOMEM
;
5973 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
5974 if (alloc_state
!= sa_rootdomain
)
5977 /* Set up domains for cpus specified by the cpu_map. */
5978 for_each_cpu(i
, cpu_map
) {
5979 struct sched_domain_topology_level
*tl
;
5982 for_each_sd_topology(tl
) {
5983 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
5984 if (tl
== sched_domain_topology
)
5985 *per_cpu_ptr(d
.sd
, i
) = sd
;
5986 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
5987 sd
->flags
|= SD_OVERLAP
;
5988 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
5993 /* Build the groups for the domains */
5994 for_each_cpu(i
, cpu_map
) {
5995 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
5996 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
5997 if (sd
->flags
& SD_OVERLAP
) {
5998 if (build_overlap_sched_groups(sd
, i
))
6001 if (build_sched_groups(sd
, i
))
6007 /* Calculate CPU power for physical packages and nodes */
6008 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6009 if (!cpumask_test_cpu(i
, cpu_map
))
6012 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6013 claim_allocations(i
, sd
);
6014 init_sched_groups_power(i
, sd
);
6018 /* Attach the domains */
6020 for_each_cpu(i
, cpu_map
) {
6021 sd
= *per_cpu_ptr(d
.sd
, i
);
6022 cpu_attach_domain(sd
, d
.rd
, i
);
6028 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6032 static cpumask_var_t
*doms_cur
; /* current sched domains */
6033 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6034 static struct sched_domain_attr
*dattr_cur
;
6035 /* attribues of custom domains in 'doms_cur' */
6038 * Special case: If a kmalloc of a doms_cur partition (array of
6039 * cpumask) fails, then fallback to a single sched domain,
6040 * as determined by the single cpumask fallback_doms.
6042 static cpumask_var_t fallback_doms
;
6045 * arch_update_cpu_topology lets virtualized architectures update the
6046 * cpu core maps. It is supposed to return 1 if the topology changed
6047 * or 0 if it stayed the same.
6049 int __attribute__((weak
)) arch_update_cpu_topology(void)
6054 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6057 cpumask_var_t
*doms
;
6059 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6062 for (i
= 0; i
< ndoms
; i
++) {
6063 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6064 free_sched_domains(doms
, i
);
6071 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6074 for (i
= 0; i
< ndoms
; i
++)
6075 free_cpumask_var(doms
[i
]);
6080 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6081 * For now this just excludes isolated cpus, but could be used to
6082 * exclude other special cases in the future.
6084 static int init_sched_domains(const struct cpumask
*cpu_map
)
6088 arch_update_cpu_topology();
6090 doms_cur
= alloc_sched_domains(ndoms_cur
);
6092 doms_cur
= &fallback_doms
;
6093 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6094 err
= build_sched_domains(doms_cur
[0], NULL
);
6095 register_sched_domain_sysctl();
6101 * Detach sched domains from a group of cpus specified in cpu_map
6102 * These cpus will now be attached to the NULL domain
6104 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6109 for_each_cpu(i
, cpu_map
)
6110 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6114 /* handle null as "default" */
6115 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6116 struct sched_domain_attr
*new, int idx_new
)
6118 struct sched_domain_attr tmp
;
6125 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6126 new ? (new + idx_new
) : &tmp
,
6127 sizeof(struct sched_domain_attr
));
6131 * Partition sched domains as specified by the 'ndoms_new'
6132 * cpumasks in the array doms_new[] of cpumasks. This compares
6133 * doms_new[] to the current sched domain partitioning, doms_cur[].
6134 * It destroys each deleted domain and builds each new domain.
6136 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6137 * The masks don't intersect (don't overlap.) We should setup one
6138 * sched domain for each mask. CPUs not in any of the cpumasks will
6139 * not be load balanced. If the same cpumask appears both in the
6140 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6143 * The passed in 'doms_new' should be allocated using
6144 * alloc_sched_domains. This routine takes ownership of it and will
6145 * free_sched_domains it when done with it. If the caller failed the
6146 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6147 * and partition_sched_domains() will fallback to the single partition
6148 * 'fallback_doms', it also forces the domains to be rebuilt.
6150 * If doms_new == NULL it will be replaced with cpu_online_mask.
6151 * ndoms_new == 0 is a special case for destroying existing domains,
6152 * and it will not create the default domain.
6154 * Call with hotplug lock held
6156 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6157 struct sched_domain_attr
*dattr_new
)
6162 mutex_lock(&sched_domains_mutex
);
6164 /* always unregister in case we don't destroy any domains */
6165 unregister_sched_domain_sysctl();
6167 /* Let architecture update cpu core mappings. */
6168 new_topology
= arch_update_cpu_topology();
6170 n
= doms_new
? ndoms_new
: 0;
6172 /* Destroy deleted domains */
6173 for (i
= 0; i
< ndoms_cur
; i
++) {
6174 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6175 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6176 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6179 /* no match - a current sched domain not in new doms_new[] */
6180 detach_destroy_domains(doms_cur
[i
]);
6185 if (doms_new
== NULL
) {
6187 doms_new
= &fallback_doms
;
6188 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6189 WARN_ON_ONCE(dattr_new
);
6192 /* Build new domains */
6193 for (i
= 0; i
< ndoms_new
; i
++) {
6194 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
6195 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6196 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6199 /* no match - add a new doms_new */
6200 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6205 /* Remember the new sched domains */
6206 if (doms_cur
!= &fallback_doms
)
6207 free_sched_domains(doms_cur
, ndoms_cur
);
6208 kfree(dattr_cur
); /* kfree(NULL) is safe */
6209 doms_cur
= doms_new
;
6210 dattr_cur
= dattr_new
;
6211 ndoms_cur
= ndoms_new
;
6213 register_sched_domain_sysctl();
6215 mutex_unlock(&sched_domains_mutex
);
6218 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6221 * Update cpusets according to cpu_active mask. If cpusets are
6222 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6223 * around partition_sched_domains().
6225 * If we come here as part of a suspend/resume, don't touch cpusets because we
6226 * want to restore it back to its original state upon resume anyway.
6228 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6232 case CPU_ONLINE_FROZEN
:
6233 case CPU_DOWN_FAILED_FROZEN
:
6236 * num_cpus_frozen tracks how many CPUs are involved in suspend
6237 * resume sequence. As long as this is not the last online
6238 * operation in the resume sequence, just build a single sched
6239 * domain, ignoring cpusets.
6242 if (likely(num_cpus_frozen
)) {
6243 partition_sched_domains(1, NULL
, NULL
);
6248 * This is the last CPU online operation. So fall through and
6249 * restore the original sched domains by considering the
6250 * cpuset configurations.
6254 case CPU_DOWN_FAILED
:
6255 cpuset_update_active_cpus(true);
6263 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6267 case CPU_DOWN_PREPARE
:
6268 cpuset_update_active_cpus(false);
6270 case CPU_DOWN_PREPARE_FROZEN
:
6272 partition_sched_domains(1, NULL
, NULL
);
6280 void __init
sched_init_smp(void)
6282 cpumask_var_t non_isolated_cpus
;
6284 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6285 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6290 mutex_lock(&sched_domains_mutex
);
6291 init_sched_domains(cpu_active_mask
);
6292 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6293 if (cpumask_empty(non_isolated_cpus
))
6294 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6295 mutex_unlock(&sched_domains_mutex
);
6298 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6299 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6300 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6304 /* Move init over to a non-isolated CPU */
6305 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6307 sched_init_granularity();
6308 free_cpumask_var(non_isolated_cpus
);
6310 init_sched_rt_class();
6313 void __init
sched_init_smp(void)
6315 sched_init_granularity();
6317 #endif /* CONFIG_SMP */
6319 const_debug
unsigned int sysctl_timer_migration
= 1;
6321 int in_sched_functions(unsigned long addr
)
6323 return in_lock_functions(addr
) ||
6324 (addr
>= (unsigned long)__sched_text_start
6325 && addr
< (unsigned long)__sched_text_end
);
6328 #ifdef CONFIG_CGROUP_SCHED
6330 * Default task group.
6331 * Every task in system belongs to this group at bootup.
6333 struct task_group root_task_group
;
6334 LIST_HEAD(task_groups
);
6337 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6339 void __init
sched_init(void)
6342 unsigned long alloc_size
= 0, ptr
;
6344 #ifdef CONFIG_FAIR_GROUP_SCHED
6345 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6347 #ifdef CONFIG_RT_GROUP_SCHED
6348 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6350 #ifdef CONFIG_CPUMASK_OFFSTACK
6351 alloc_size
+= num_possible_cpus() * cpumask_size();
6354 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6356 #ifdef CONFIG_FAIR_GROUP_SCHED
6357 root_task_group
.se
= (struct sched_entity
**)ptr
;
6358 ptr
+= nr_cpu_ids
* sizeof(void **);
6360 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6361 ptr
+= nr_cpu_ids
* sizeof(void **);
6363 #endif /* CONFIG_FAIR_GROUP_SCHED */
6364 #ifdef CONFIG_RT_GROUP_SCHED
6365 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6366 ptr
+= nr_cpu_ids
* sizeof(void **);
6368 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6369 ptr
+= nr_cpu_ids
* sizeof(void **);
6371 #endif /* CONFIG_RT_GROUP_SCHED */
6372 #ifdef CONFIG_CPUMASK_OFFSTACK
6373 for_each_possible_cpu(i
) {
6374 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6375 ptr
+= cpumask_size();
6377 #endif /* CONFIG_CPUMASK_OFFSTACK */
6381 init_defrootdomain();
6384 init_rt_bandwidth(&def_rt_bandwidth
,
6385 global_rt_period(), global_rt_runtime());
6387 #ifdef CONFIG_RT_GROUP_SCHED
6388 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6389 global_rt_period(), global_rt_runtime());
6390 #endif /* CONFIG_RT_GROUP_SCHED */
6392 #ifdef CONFIG_CGROUP_SCHED
6393 list_add(&root_task_group
.list
, &task_groups
);
6394 INIT_LIST_HEAD(&root_task_group
.children
);
6395 INIT_LIST_HEAD(&root_task_group
.siblings
);
6396 autogroup_init(&init_task
);
6398 #endif /* CONFIG_CGROUP_SCHED */
6400 for_each_possible_cpu(i
) {
6404 raw_spin_lock_init(&rq
->lock
);
6406 rq
->calc_load_active
= 0;
6407 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6408 init_cfs_rq(&rq
->cfs
);
6409 init_rt_rq(&rq
->rt
, rq
);
6410 #ifdef CONFIG_FAIR_GROUP_SCHED
6411 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6412 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6414 * How much cpu bandwidth does root_task_group get?
6416 * In case of task-groups formed thr' the cgroup filesystem, it
6417 * gets 100% of the cpu resources in the system. This overall
6418 * system cpu resource is divided among the tasks of
6419 * root_task_group and its child task-groups in a fair manner,
6420 * based on each entity's (task or task-group's) weight
6421 * (se->load.weight).
6423 * In other words, if root_task_group has 10 tasks of weight
6424 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6425 * then A0's share of the cpu resource is:
6427 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6429 * We achieve this by letting root_task_group's tasks sit
6430 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6432 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6433 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6434 #endif /* CONFIG_FAIR_GROUP_SCHED */
6436 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6437 #ifdef CONFIG_RT_GROUP_SCHED
6438 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6439 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6442 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6443 rq
->cpu_load
[j
] = 0;
6445 rq
->last_load_update_tick
= jiffies
;
6450 rq
->cpu_power
= SCHED_POWER_SCALE
;
6451 rq
->post_schedule
= 0;
6452 rq
->active_balance
= 0;
6453 rq
->next_balance
= jiffies
;
6458 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6460 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6462 rq_attach_root(rq
, &def_root_domain
);
6463 #ifdef CONFIG_NO_HZ_COMMON
6466 #ifdef CONFIG_NO_HZ_FULL
6467 rq
->last_sched_tick
= 0;
6471 atomic_set(&rq
->nr_iowait
, 0);
6474 set_load_weight(&init_task
);
6476 #ifdef CONFIG_PREEMPT_NOTIFIERS
6477 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6480 #ifdef CONFIG_RT_MUTEXES
6481 plist_head_init(&init_task
.pi_waiters
);
6485 * The boot idle thread does lazy MMU switching as well:
6487 atomic_inc(&init_mm
.mm_count
);
6488 enter_lazy_tlb(&init_mm
, current
);
6491 * Make us the idle thread. Technically, schedule() should not be
6492 * called from this thread, however somewhere below it might be,
6493 * but because we are the idle thread, we just pick up running again
6494 * when this runqueue becomes "idle".
6496 init_idle(current
, smp_processor_id());
6498 calc_load_update
= jiffies
+ LOAD_FREQ
;
6501 * During early bootup we pretend to be a normal task:
6503 current
->sched_class
= &fair_sched_class
;
6506 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6507 /* May be allocated at isolcpus cmdline parse time */
6508 if (cpu_isolated_map
== NULL
)
6509 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6510 idle_thread_set_boot_cpu();
6512 init_sched_fair_class();
6514 scheduler_running
= 1;
6517 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6518 static inline int preempt_count_equals(int preempt_offset
)
6520 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6522 return (nested
== preempt_offset
);
6525 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6527 static unsigned long prev_jiffy
; /* ratelimiting */
6529 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6530 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6531 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6533 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6535 prev_jiffy
= jiffies
;
6538 "BUG: sleeping function called from invalid context at %s:%d\n",
6541 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6542 in_atomic(), irqs_disabled(),
6543 current
->pid
, current
->comm
);
6545 debug_show_held_locks(current
);
6546 if (irqs_disabled())
6547 print_irqtrace_events(current
);
6550 EXPORT_SYMBOL(__might_sleep
);
6553 #ifdef CONFIG_MAGIC_SYSRQ
6554 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6556 const struct sched_class
*prev_class
= p
->sched_class
;
6557 int old_prio
= p
->prio
;
6562 dequeue_task(rq
, p
, 0);
6563 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6565 enqueue_task(rq
, p
, 0);
6566 resched_task(rq
->curr
);
6569 check_class_changed(rq
, p
, prev_class
, old_prio
);
6572 void normalize_rt_tasks(void)
6574 struct task_struct
*g
, *p
;
6575 unsigned long flags
;
6578 read_lock_irqsave(&tasklist_lock
, flags
);
6579 do_each_thread(g
, p
) {
6581 * Only normalize user tasks:
6586 p
->se
.exec_start
= 0;
6587 #ifdef CONFIG_SCHEDSTATS
6588 p
->se
.statistics
.wait_start
= 0;
6589 p
->se
.statistics
.sleep_start
= 0;
6590 p
->se
.statistics
.block_start
= 0;
6595 * Renice negative nice level userspace
6598 if (TASK_NICE(p
) < 0 && p
->mm
)
6599 set_user_nice(p
, 0);
6603 raw_spin_lock(&p
->pi_lock
);
6604 rq
= __task_rq_lock(p
);
6606 normalize_task(rq
, p
);
6608 __task_rq_unlock(rq
);
6609 raw_spin_unlock(&p
->pi_lock
);
6610 } while_each_thread(g
, p
);
6612 read_unlock_irqrestore(&tasklist_lock
, flags
);
6615 #endif /* CONFIG_MAGIC_SYSRQ */
6617 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6619 * These functions are only useful for the IA64 MCA handling, or kdb.
6621 * They can only be called when the whole system has been
6622 * stopped - every CPU needs to be quiescent, and no scheduling
6623 * activity can take place. Using them for anything else would
6624 * be a serious bug, and as a result, they aren't even visible
6625 * under any other configuration.
6629 * curr_task - return the current task for a given cpu.
6630 * @cpu: the processor in question.
6632 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6634 struct task_struct
*curr_task(int cpu
)
6636 return cpu_curr(cpu
);
6639 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6643 * set_curr_task - set the current task for a given cpu.
6644 * @cpu: the processor in question.
6645 * @p: the task pointer to set.
6647 * Description: This function must only be used when non-maskable interrupts
6648 * are serviced on a separate stack. It allows the architecture to switch the
6649 * notion of the current task on a cpu in a non-blocking manner. This function
6650 * must be called with all CPU's synchronized, and interrupts disabled, the
6651 * and caller must save the original value of the current task (see
6652 * curr_task() above) and restore that value before reenabling interrupts and
6653 * re-starting the system.
6655 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6657 void set_curr_task(int cpu
, struct task_struct
*p
)
6664 #ifdef CONFIG_CGROUP_SCHED
6665 /* task_group_lock serializes the addition/removal of task groups */
6666 static DEFINE_SPINLOCK(task_group_lock
);
6668 static void free_sched_group(struct task_group
*tg
)
6670 free_fair_sched_group(tg
);
6671 free_rt_sched_group(tg
);
6676 /* allocate runqueue etc for a new task group */
6677 struct task_group
*sched_create_group(struct task_group
*parent
)
6679 struct task_group
*tg
;
6681 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6683 return ERR_PTR(-ENOMEM
);
6685 if (!alloc_fair_sched_group(tg
, parent
))
6688 if (!alloc_rt_sched_group(tg
, parent
))
6694 free_sched_group(tg
);
6695 return ERR_PTR(-ENOMEM
);
6698 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6700 unsigned long flags
;
6702 spin_lock_irqsave(&task_group_lock
, flags
);
6703 list_add_rcu(&tg
->list
, &task_groups
);
6705 WARN_ON(!parent
); /* root should already exist */
6707 tg
->parent
= parent
;
6708 INIT_LIST_HEAD(&tg
->children
);
6709 list_add_rcu(&tg
->siblings
, &parent
->children
);
6710 spin_unlock_irqrestore(&task_group_lock
, flags
);
6713 /* rcu callback to free various structures associated with a task group */
6714 static void free_sched_group_rcu(struct rcu_head
*rhp
)
6716 /* now it should be safe to free those cfs_rqs */
6717 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
6720 /* Destroy runqueue etc associated with a task group */
6721 void sched_destroy_group(struct task_group
*tg
)
6723 /* wait for possible concurrent references to cfs_rqs complete */
6724 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
6727 void sched_offline_group(struct task_group
*tg
)
6729 unsigned long flags
;
6732 /* end participation in shares distribution */
6733 for_each_possible_cpu(i
)
6734 unregister_fair_sched_group(tg
, i
);
6736 spin_lock_irqsave(&task_group_lock
, flags
);
6737 list_del_rcu(&tg
->list
);
6738 list_del_rcu(&tg
->siblings
);
6739 spin_unlock_irqrestore(&task_group_lock
, flags
);
6742 /* change task's runqueue when it moves between groups.
6743 * The caller of this function should have put the task in its new group
6744 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6745 * reflect its new group.
6747 void sched_move_task(struct task_struct
*tsk
)
6749 struct task_group
*tg
;
6751 unsigned long flags
;
6754 rq
= task_rq_lock(tsk
, &flags
);
6756 running
= task_current(rq
, tsk
);
6760 dequeue_task(rq
, tsk
, 0);
6761 if (unlikely(running
))
6762 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6764 tg
= container_of(task_subsys_state_check(tsk
, cpu_cgroup_subsys_id
,
6765 lockdep_is_held(&tsk
->sighand
->siglock
)),
6766 struct task_group
, css
);
6767 tg
= autogroup_task_group(tsk
, tg
);
6768 tsk
->sched_task_group
= tg
;
6770 #ifdef CONFIG_FAIR_GROUP_SCHED
6771 if (tsk
->sched_class
->task_move_group
)
6772 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
6775 set_task_rq(tsk
, task_cpu(tsk
));
6777 if (unlikely(running
))
6778 tsk
->sched_class
->set_curr_task(rq
);
6780 enqueue_task(rq
, tsk
, 0);
6782 task_rq_unlock(rq
, tsk
, &flags
);
6784 #endif /* CONFIG_CGROUP_SCHED */
6786 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6787 static unsigned long to_ratio(u64 period
, u64 runtime
)
6789 if (runtime
== RUNTIME_INF
)
6792 return div64_u64(runtime
<< 20, period
);
6796 #ifdef CONFIG_RT_GROUP_SCHED
6798 * Ensure that the real time constraints are schedulable.
6800 static DEFINE_MUTEX(rt_constraints_mutex
);
6802 /* Must be called with tasklist_lock held */
6803 static inline int tg_has_rt_tasks(struct task_group
*tg
)
6805 struct task_struct
*g
, *p
;
6807 do_each_thread(g
, p
) {
6808 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
6810 } while_each_thread(g
, p
);
6815 struct rt_schedulable_data
{
6816 struct task_group
*tg
;
6821 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
6823 struct rt_schedulable_data
*d
= data
;
6824 struct task_group
*child
;
6825 unsigned long total
, sum
= 0;
6826 u64 period
, runtime
;
6828 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6829 runtime
= tg
->rt_bandwidth
.rt_runtime
;
6832 period
= d
->rt_period
;
6833 runtime
= d
->rt_runtime
;
6837 * Cannot have more runtime than the period.
6839 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
6843 * Ensure we don't starve existing RT tasks.
6845 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
6848 total
= to_ratio(period
, runtime
);
6851 * Nobody can have more than the global setting allows.
6853 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
6857 * The sum of our children's runtime should not exceed our own.
6859 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
6860 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
6861 runtime
= child
->rt_bandwidth
.rt_runtime
;
6863 if (child
== d
->tg
) {
6864 period
= d
->rt_period
;
6865 runtime
= d
->rt_runtime
;
6868 sum
+= to_ratio(period
, runtime
);
6877 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
6881 struct rt_schedulable_data data
= {
6883 .rt_period
= period
,
6884 .rt_runtime
= runtime
,
6888 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
6894 static int tg_set_rt_bandwidth(struct task_group
*tg
,
6895 u64 rt_period
, u64 rt_runtime
)
6899 mutex_lock(&rt_constraints_mutex
);
6900 read_lock(&tasklist_lock
);
6901 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
6905 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6906 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
6907 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
6909 for_each_possible_cpu(i
) {
6910 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
6912 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6913 rt_rq
->rt_runtime
= rt_runtime
;
6914 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6916 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6918 read_unlock(&tasklist_lock
);
6919 mutex_unlock(&rt_constraints_mutex
);
6924 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
6926 u64 rt_runtime
, rt_period
;
6928 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6929 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
6930 if (rt_runtime_us
< 0)
6931 rt_runtime
= RUNTIME_INF
;
6933 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6936 static long sched_group_rt_runtime(struct task_group
*tg
)
6940 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
6943 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
6944 do_div(rt_runtime_us
, NSEC_PER_USEC
);
6945 return rt_runtime_us
;
6948 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
6950 u64 rt_runtime
, rt_period
;
6952 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
6953 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
6958 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6961 static long sched_group_rt_period(struct task_group
*tg
)
6965 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6966 do_div(rt_period_us
, NSEC_PER_USEC
);
6967 return rt_period_us
;
6970 static int sched_rt_global_constraints(void)
6972 u64 runtime
, period
;
6975 if (sysctl_sched_rt_period
<= 0)
6978 runtime
= global_rt_runtime();
6979 period
= global_rt_period();
6982 * Sanity check on the sysctl variables.
6984 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
6987 mutex_lock(&rt_constraints_mutex
);
6988 read_lock(&tasklist_lock
);
6989 ret
= __rt_schedulable(NULL
, 0, 0);
6990 read_unlock(&tasklist_lock
);
6991 mutex_unlock(&rt_constraints_mutex
);
6996 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
6998 /* Don't accept realtime tasks when there is no way for them to run */
6999 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7005 #else /* !CONFIG_RT_GROUP_SCHED */
7006 static int sched_rt_global_constraints(void)
7008 unsigned long flags
;
7011 if (sysctl_sched_rt_period
<= 0)
7015 * There's always some RT tasks in the root group
7016 * -- migration, kstopmachine etc..
7018 if (sysctl_sched_rt_runtime
== 0)
7021 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7022 for_each_possible_cpu(i
) {
7023 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7025 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7026 rt_rq
->rt_runtime
= global_rt_runtime();
7027 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7029 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7033 #endif /* CONFIG_RT_GROUP_SCHED */
7035 int sched_rr_handler(struct ctl_table
*table
, int write
,
7036 void __user
*buffer
, size_t *lenp
,
7040 static DEFINE_MUTEX(mutex
);
7043 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7044 /* make sure that internally we keep jiffies */
7045 /* also, writing zero resets timeslice to default */
7046 if (!ret
&& write
) {
7047 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7048 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7050 mutex_unlock(&mutex
);
7054 int sched_rt_handler(struct ctl_table
*table
, int write
,
7055 void __user
*buffer
, size_t *lenp
,
7059 int old_period
, old_runtime
;
7060 static DEFINE_MUTEX(mutex
);
7063 old_period
= sysctl_sched_rt_period
;
7064 old_runtime
= sysctl_sched_rt_runtime
;
7066 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7068 if (!ret
&& write
) {
7069 ret
= sched_rt_global_constraints();
7071 sysctl_sched_rt_period
= old_period
;
7072 sysctl_sched_rt_runtime
= old_runtime
;
7074 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7075 def_rt_bandwidth
.rt_period
=
7076 ns_to_ktime(global_rt_period());
7079 mutex_unlock(&mutex
);
7084 #ifdef CONFIG_CGROUP_SCHED
7086 /* return corresponding task_group object of a cgroup */
7087 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
7089 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
7090 struct task_group
, css
);
7093 static struct cgroup_subsys_state
*cpu_cgroup_css_alloc(struct cgroup
*cgrp
)
7095 struct task_group
*tg
, *parent
;
7097 if (!cgrp
->parent
) {
7098 /* This is early initialization for the top cgroup */
7099 return &root_task_group
.css
;
7102 parent
= cgroup_tg(cgrp
->parent
);
7103 tg
= sched_create_group(parent
);
7105 return ERR_PTR(-ENOMEM
);
7110 static int cpu_cgroup_css_online(struct cgroup
*cgrp
)
7112 struct task_group
*tg
= cgroup_tg(cgrp
);
7113 struct task_group
*parent
;
7118 parent
= cgroup_tg(cgrp
->parent
);
7119 sched_online_group(tg
, parent
);
7123 static void cpu_cgroup_css_free(struct cgroup
*cgrp
)
7125 struct task_group
*tg
= cgroup_tg(cgrp
);
7127 sched_destroy_group(tg
);
7130 static void cpu_cgroup_css_offline(struct cgroup
*cgrp
)
7132 struct task_group
*tg
= cgroup_tg(cgrp
);
7134 sched_offline_group(tg
);
7137 static int cpu_cgroup_can_attach(struct cgroup
*cgrp
,
7138 struct cgroup_taskset
*tset
)
7140 struct task_struct
*task
;
7142 cgroup_taskset_for_each(task
, cgrp
, tset
) {
7143 #ifdef CONFIG_RT_GROUP_SCHED
7144 if (!sched_rt_can_attach(cgroup_tg(cgrp
), task
))
7147 /* We don't support RT-tasks being in separate groups */
7148 if (task
->sched_class
!= &fair_sched_class
)
7155 static void cpu_cgroup_attach(struct cgroup
*cgrp
,
7156 struct cgroup_taskset
*tset
)
7158 struct task_struct
*task
;
7160 cgroup_taskset_for_each(task
, cgrp
, tset
)
7161 sched_move_task(task
);
7165 cpu_cgroup_exit(struct cgroup
*cgrp
, struct cgroup
*old_cgrp
,
7166 struct task_struct
*task
)
7169 * cgroup_exit() is called in the copy_process() failure path.
7170 * Ignore this case since the task hasn't ran yet, this avoids
7171 * trying to poke a half freed task state from generic code.
7173 if (!(task
->flags
& PF_EXITING
))
7176 sched_move_task(task
);
7179 #ifdef CONFIG_FAIR_GROUP_SCHED
7180 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7183 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
7186 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7188 struct task_group
*tg
= cgroup_tg(cgrp
);
7190 return (u64
) scale_load_down(tg
->shares
);
7193 #ifdef CONFIG_CFS_BANDWIDTH
7194 static DEFINE_MUTEX(cfs_constraints_mutex
);
7196 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7197 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7199 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7201 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7203 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7204 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7206 if (tg
== &root_task_group
)
7210 * Ensure we have at some amount of bandwidth every period. This is
7211 * to prevent reaching a state of large arrears when throttled via
7212 * entity_tick() resulting in prolonged exit starvation.
7214 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7218 * Likewise, bound things on the otherside by preventing insane quota
7219 * periods. This also allows us to normalize in computing quota
7222 if (period
> max_cfs_quota_period
)
7225 mutex_lock(&cfs_constraints_mutex
);
7226 ret
= __cfs_schedulable(tg
, period
, quota
);
7230 runtime_enabled
= quota
!= RUNTIME_INF
;
7231 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7232 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7233 raw_spin_lock_irq(&cfs_b
->lock
);
7234 cfs_b
->period
= ns_to_ktime(period
);
7235 cfs_b
->quota
= quota
;
7237 __refill_cfs_bandwidth_runtime(cfs_b
);
7238 /* restart the period timer (if active) to handle new period expiry */
7239 if (runtime_enabled
&& cfs_b
->timer_active
) {
7240 /* force a reprogram */
7241 cfs_b
->timer_active
= 0;
7242 __start_cfs_bandwidth(cfs_b
);
7244 raw_spin_unlock_irq(&cfs_b
->lock
);
7246 for_each_possible_cpu(i
) {
7247 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7248 struct rq
*rq
= cfs_rq
->rq
;
7250 raw_spin_lock_irq(&rq
->lock
);
7251 cfs_rq
->runtime_enabled
= runtime_enabled
;
7252 cfs_rq
->runtime_remaining
= 0;
7254 if (cfs_rq
->throttled
)
7255 unthrottle_cfs_rq(cfs_rq
);
7256 raw_spin_unlock_irq(&rq
->lock
);
7259 mutex_unlock(&cfs_constraints_mutex
);
7264 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7268 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7269 if (cfs_quota_us
< 0)
7270 quota
= RUNTIME_INF
;
7272 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7274 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7277 long tg_get_cfs_quota(struct task_group
*tg
)
7281 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7284 quota_us
= tg
->cfs_bandwidth
.quota
;
7285 do_div(quota_us
, NSEC_PER_USEC
);
7290 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7294 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7295 quota
= tg
->cfs_bandwidth
.quota
;
7297 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7300 long tg_get_cfs_period(struct task_group
*tg
)
7304 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7305 do_div(cfs_period_us
, NSEC_PER_USEC
);
7307 return cfs_period_us
;
7310 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
7312 return tg_get_cfs_quota(cgroup_tg(cgrp
));
7315 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7318 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
7321 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
7323 return tg_get_cfs_period(cgroup_tg(cgrp
));
7326 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
7329 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
7332 struct cfs_schedulable_data
{
7333 struct task_group
*tg
;
7338 * normalize group quota/period to be quota/max_period
7339 * note: units are usecs
7341 static u64
normalize_cfs_quota(struct task_group
*tg
,
7342 struct cfs_schedulable_data
*d
)
7350 period
= tg_get_cfs_period(tg
);
7351 quota
= tg_get_cfs_quota(tg
);
7354 /* note: these should typically be equivalent */
7355 if (quota
== RUNTIME_INF
|| quota
== -1)
7358 return to_ratio(period
, quota
);
7361 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7363 struct cfs_schedulable_data
*d
= data
;
7364 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7365 s64 quota
= 0, parent_quota
= -1;
7368 quota
= RUNTIME_INF
;
7370 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7372 quota
= normalize_cfs_quota(tg
, d
);
7373 parent_quota
= parent_b
->hierarchal_quota
;
7376 * ensure max(child_quota) <= parent_quota, inherit when no
7379 if (quota
== RUNTIME_INF
)
7380 quota
= parent_quota
;
7381 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7384 cfs_b
->hierarchal_quota
= quota
;
7389 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7392 struct cfs_schedulable_data data
= {
7398 if (quota
!= RUNTIME_INF
) {
7399 do_div(data
.period
, NSEC_PER_USEC
);
7400 do_div(data
.quota
, NSEC_PER_USEC
);
7404 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7410 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
7411 struct cgroup_map_cb
*cb
)
7413 struct task_group
*tg
= cgroup_tg(cgrp
);
7414 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7416 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7417 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7418 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7422 #endif /* CONFIG_CFS_BANDWIDTH */
7423 #endif /* CONFIG_FAIR_GROUP_SCHED */
7425 #ifdef CONFIG_RT_GROUP_SCHED
7426 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
7429 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
7432 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
7434 return sched_group_rt_runtime(cgroup_tg(cgrp
));
7437 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
7440 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
7443 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
7445 return sched_group_rt_period(cgroup_tg(cgrp
));
7447 #endif /* CONFIG_RT_GROUP_SCHED */
7449 static struct cftype cpu_files
[] = {
7450 #ifdef CONFIG_FAIR_GROUP_SCHED
7453 .read_u64
= cpu_shares_read_u64
,
7454 .write_u64
= cpu_shares_write_u64
,
7457 #ifdef CONFIG_CFS_BANDWIDTH
7459 .name
= "cfs_quota_us",
7460 .read_s64
= cpu_cfs_quota_read_s64
,
7461 .write_s64
= cpu_cfs_quota_write_s64
,
7464 .name
= "cfs_period_us",
7465 .read_u64
= cpu_cfs_period_read_u64
,
7466 .write_u64
= cpu_cfs_period_write_u64
,
7470 .read_map
= cpu_stats_show
,
7473 #ifdef CONFIG_RT_GROUP_SCHED
7475 .name
= "rt_runtime_us",
7476 .read_s64
= cpu_rt_runtime_read
,
7477 .write_s64
= cpu_rt_runtime_write
,
7480 .name
= "rt_period_us",
7481 .read_u64
= cpu_rt_period_read_uint
,
7482 .write_u64
= cpu_rt_period_write_uint
,
7488 struct cgroup_subsys cpu_cgroup_subsys
= {
7490 .css_alloc
= cpu_cgroup_css_alloc
,
7491 .css_free
= cpu_cgroup_css_free
,
7492 .css_online
= cpu_cgroup_css_online
,
7493 .css_offline
= cpu_cgroup_css_offline
,
7494 .can_attach
= cpu_cgroup_can_attach
,
7495 .attach
= cpu_cgroup_attach
,
7496 .exit
= cpu_cgroup_exit
,
7497 .subsys_id
= cpu_cgroup_subsys_id
,
7498 .base_cftypes
= cpu_files
,
7502 #endif /* CONFIG_CGROUP_SCHED */
7504 void dump_cpu_task(int cpu
)
7506 pr_info("Task dump for CPU %d:\n", cpu
);
7507 sched_show_task(cpu_curr(cpu
));