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>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
96 ktime_t soft
, hard
, now
;
99 if (hrtimer_active(period_timer
))
102 now
= hrtimer_cb_get_time(period_timer
);
103 hrtimer_forward(period_timer
, now
, period
);
105 soft
= hrtimer_get_softexpires(period_timer
);
106 hard
= hrtimer_get_expires(period_timer
);
107 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
108 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
109 HRTIMER_MODE_ABS_PINNED
, 0);
113 DEFINE_MUTEX(sched_domains_mutex
);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
116 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
118 void update_rq_clock(struct rq
*rq
)
122 lockdep_assert_held(&rq
->lock
);
124 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
127 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
131 update_rq_clock_task(rq
, delta
);
135 * Debugging: various feature bits
138 #define SCHED_FEAT(name, enabled) \
139 (1UL << __SCHED_FEAT_##name) * enabled |
141 const_debug
unsigned int sysctl_sched_features
=
142 #include "features.h"
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled) \
151 static const char * const sched_feat_names
[] = {
152 #include "features.h"
157 static int sched_feat_show(struct seq_file
*m
, void *v
)
161 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
162 if (!(sysctl_sched_features
& (1UL << i
)))
164 seq_printf(m
, "%s ", sched_feat_names
[i
]);
171 #ifdef HAVE_JUMP_LABEL
173 #define jump_label_key__true STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
176 #define SCHED_FEAT(name, enabled) \
177 jump_label_key__##enabled ,
179 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
180 #include "features.h"
185 static void sched_feat_disable(int i
)
187 if (static_key_enabled(&sched_feat_keys
[i
]))
188 static_key_slow_dec(&sched_feat_keys
[i
]);
191 static void sched_feat_enable(int i
)
193 if (!static_key_enabled(&sched_feat_keys
[i
]))
194 static_key_slow_inc(&sched_feat_keys
[i
]);
197 static void sched_feat_disable(int i
) { };
198 static void sched_feat_enable(int i
) { };
199 #endif /* HAVE_JUMP_LABEL */
201 static int sched_feat_set(char *cmp
)
206 if (strncmp(cmp
, "NO_", 3) == 0) {
211 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
212 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
214 sysctl_sched_features
&= ~(1UL << i
);
215 sched_feat_disable(i
);
217 sysctl_sched_features
|= (1UL << i
);
218 sched_feat_enable(i
);
228 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
229 size_t cnt
, loff_t
*ppos
)
239 if (copy_from_user(&buf
, ubuf
, cnt
))
245 /* Ensure the static_key remains in a consistent state */
246 inode
= file_inode(filp
);
247 mutex_lock(&inode
->i_mutex
);
248 i
= sched_feat_set(cmp
);
249 mutex_unlock(&inode
->i_mutex
);
250 if (i
== __SCHED_FEAT_NR
)
258 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
260 return single_open(filp
, sched_feat_show
, NULL
);
263 static const struct file_operations sched_feat_fops
= {
264 .open
= sched_feat_open
,
265 .write
= sched_feat_write
,
268 .release
= single_release
,
271 static __init
int sched_init_debug(void)
273 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
278 late_initcall(sched_init_debug
);
279 #endif /* CONFIG_SCHED_DEBUG */
282 * Number of tasks to iterate in a single balance run.
283 * Limited because this is done with IRQs disabled.
285 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
288 * period over which we average the RT time consumption, measured
293 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
296 * period over which we measure -rt task cpu usage in us.
299 unsigned int sysctl_sched_rt_period
= 1000000;
301 __read_mostly
int scheduler_running
;
304 * part of the period that we allow rt tasks to run in us.
307 int sysctl_sched_rt_runtime
= 950000;
310 * __task_rq_lock - lock the rq @p resides on.
312 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
317 lockdep_assert_held(&p
->pi_lock
);
321 raw_spin_lock(&rq
->lock
);
322 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
)))
324 raw_spin_unlock(&rq
->lock
);
326 while (unlikely(task_on_rq_migrating(p
)))
332 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
334 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
335 __acquires(p
->pi_lock
)
341 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
343 raw_spin_lock(&rq
->lock
);
344 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
)))
346 raw_spin_unlock(&rq
->lock
);
347 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
349 while (unlikely(task_on_rq_migrating(p
)))
354 static void __task_rq_unlock(struct rq
*rq
)
357 raw_spin_unlock(&rq
->lock
);
361 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
363 __releases(p
->pi_lock
)
365 raw_spin_unlock(&rq
->lock
);
366 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
370 * this_rq_lock - lock this runqueue and disable interrupts.
372 static struct rq
*this_rq_lock(void)
379 raw_spin_lock(&rq
->lock
);
384 #ifdef CONFIG_SCHED_HRTICK
386 * Use HR-timers to deliver accurate preemption points.
389 static void hrtick_clear(struct rq
*rq
)
391 if (hrtimer_active(&rq
->hrtick_timer
))
392 hrtimer_cancel(&rq
->hrtick_timer
);
396 * High-resolution timer tick.
397 * Runs from hardirq context with interrupts disabled.
399 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
401 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
403 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
405 raw_spin_lock(&rq
->lock
);
407 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
408 raw_spin_unlock(&rq
->lock
);
410 return HRTIMER_NORESTART
;
415 static int __hrtick_restart(struct rq
*rq
)
417 struct hrtimer
*timer
= &rq
->hrtick_timer
;
418 ktime_t time
= hrtimer_get_softexpires(timer
);
420 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
424 * called from hardirq (IPI) context
426 static void __hrtick_start(void *arg
)
430 raw_spin_lock(&rq
->lock
);
431 __hrtick_restart(rq
);
432 rq
->hrtick_csd_pending
= 0;
433 raw_spin_unlock(&rq
->lock
);
437 * Called to set the hrtick timer state.
439 * called with rq->lock held and irqs disabled
441 void hrtick_start(struct rq
*rq
, u64 delay
)
443 struct hrtimer
*timer
= &rq
->hrtick_timer
;
448 * Don't schedule slices shorter than 10000ns, that just
449 * doesn't make sense and can cause timer DoS.
451 delta
= max_t(s64
, delay
, 10000LL);
452 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
454 hrtimer_set_expires(timer
, time
);
456 if (rq
== this_rq()) {
457 __hrtick_restart(rq
);
458 } else if (!rq
->hrtick_csd_pending
) {
459 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
460 rq
->hrtick_csd_pending
= 1;
465 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
467 int cpu
= (int)(long)hcpu
;
470 case CPU_UP_CANCELED
:
471 case CPU_UP_CANCELED_FROZEN
:
472 case CPU_DOWN_PREPARE
:
473 case CPU_DOWN_PREPARE_FROZEN
:
475 case CPU_DEAD_FROZEN
:
476 hrtick_clear(cpu_rq(cpu
));
483 static __init
void init_hrtick(void)
485 hotcpu_notifier(hotplug_hrtick
, 0);
489 * Called to set the hrtick timer state.
491 * called with rq->lock held and irqs disabled
493 void hrtick_start(struct rq
*rq
, u64 delay
)
496 * Don't schedule slices shorter than 10000ns, that just
497 * doesn't make sense. Rely on vruntime for fairness.
499 delay
= max_t(u64
, delay
, 10000LL);
500 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
501 HRTIMER_MODE_REL_PINNED
, 0);
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SMP */
509 static void init_rq_hrtick(struct rq
*rq
)
512 rq
->hrtick_csd_pending
= 0;
514 rq
->hrtick_csd
.flags
= 0;
515 rq
->hrtick_csd
.func
= __hrtick_start
;
516 rq
->hrtick_csd
.info
= rq
;
519 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
520 rq
->hrtick_timer
.function
= hrtick
;
522 #else /* CONFIG_SCHED_HRTICK */
523 static inline void hrtick_clear(struct rq
*rq
)
527 static inline void init_rq_hrtick(struct rq
*rq
)
531 static inline void init_hrtick(void)
534 #endif /* CONFIG_SCHED_HRTICK */
537 * cmpxchg based fetch_or, macro so it works for different integer types
539 #define fetch_or(ptr, val) \
540 ({ typeof(*(ptr)) __old, __val = *(ptr); \
542 __old = cmpxchg((ptr), __val, __val | (val)); \
543 if (__old == __val) \
550 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
552 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
553 * this avoids any races wrt polling state changes and thereby avoids
556 static bool set_nr_and_not_polling(struct task_struct
*p
)
558 struct thread_info
*ti
= task_thread_info(p
);
559 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
563 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
565 * If this returns true, then the idle task promises to call
566 * sched_ttwu_pending() and reschedule soon.
568 static bool set_nr_if_polling(struct task_struct
*p
)
570 struct thread_info
*ti
= task_thread_info(p
);
571 typeof(ti
->flags
) old
, val
= ACCESS_ONCE(ti
->flags
);
574 if (!(val
& _TIF_POLLING_NRFLAG
))
576 if (val
& _TIF_NEED_RESCHED
)
578 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
587 static bool set_nr_and_not_polling(struct task_struct
*p
)
589 set_tsk_need_resched(p
);
594 static bool set_nr_if_polling(struct task_struct
*p
)
602 * resched_curr - mark rq's current task 'to be rescheduled now'.
604 * On UP this means the setting of the need_resched flag, on SMP it
605 * might also involve a cross-CPU call to trigger the scheduler on
608 void resched_curr(struct rq
*rq
)
610 struct task_struct
*curr
= rq
->curr
;
613 lockdep_assert_held(&rq
->lock
);
615 if (test_tsk_need_resched(curr
))
620 if (cpu
== smp_processor_id()) {
621 set_tsk_need_resched(curr
);
622 set_preempt_need_resched();
626 if (set_nr_and_not_polling(curr
))
627 smp_send_reschedule(cpu
);
629 trace_sched_wake_idle_without_ipi(cpu
);
632 void resched_cpu(int cpu
)
634 struct rq
*rq
= cpu_rq(cpu
);
637 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
640 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
644 #ifdef CONFIG_NO_HZ_COMMON
646 * In the semi idle case, use the nearest busy cpu for migrating timers
647 * from an idle cpu. This is good for power-savings.
649 * We don't do similar optimization for completely idle system, as
650 * selecting an idle cpu will add more delays to the timers than intended
651 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
653 int get_nohz_timer_target(int pinned
)
655 int cpu
= smp_processor_id();
657 struct sched_domain
*sd
;
659 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
663 for_each_domain(cpu
, sd
) {
664 for_each_cpu(i
, sched_domain_span(sd
)) {
676 * When add_timer_on() enqueues a timer into the timer wheel of an
677 * idle CPU then this timer might expire before the next timer event
678 * which is scheduled to wake up that CPU. In case of a completely
679 * idle system the next event might even be infinite time into the
680 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
681 * leaves the inner idle loop so the newly added timer is taken into
682 * account when the CPU goes back to idle and evaluates the timer
683 * wheel for the next timer event.
685 static void wake_up_idle_cpu(int cpu
)
687 struct rq
*rq
= cpu_rq(cpu
);
689 if (cpu
== smp_processor_id())
692 if (set_nr_and_not_polling(rq
->idle
))
693 smp_send_reschedule(cpu
);
695 trace_sched_wake_idle_without_ipi(cpu
);
698 static bool wake_up_full_nohz_cpu(int cpu
)
701 * We just need the target to call irq_exit() and re-evaluate
702 * the next tick. The nohz full kick at least implies that.
703 * If needed we can still optimize that later with an
706 if (tick_nohz_full_cpu(cpu
)) {
707 if (cpu
!= smp_processor_id() ||
708 tick_nohz_tick_stopped())
709 tick_nohz_full_kick_cpu(cpu
);
716 void wake_up_nohz_cpu(int cpu
)
718 if (!wake_up_full_nohz_cpu(cpu
))
719 wake_up_idle_cpu(cpu
);
722 static inline bool got_nohz_idle_kick(void)
724 int cpu
= smp_processor_id();
726 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
729 if (idle_cpu(cpu
) && !need_resched())
733 * We can't run Idle Load Balance on this CPU for this time so we
734 * cancel it and clear NOHZ_BALANCE_KICK
736 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
740 #else /* CONFIG_NO_HZ_COMMON */
742 static inline bool got_nohz_idle_kick(void)
747 #endif /* CONFIG_NO_HZ_COMMON */
749 #ifdef CONFIG_NO_HZ_FULL
750 bool sched_can_stop_tick(void)
753 * More than one running task need preemption.
754 * nr_running update is assumed to be visible
755 * after IPI is sent from wakers.
757 if (this_rq()->nr_running
> 1)
762 #endif /* CONFIG_NO_HZ_FULL */
764 void sched_avg_update(struct rq
*rq
)
766 s64 period
= sched_avg_period();
768 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
770 * Inline assembly required to prevent the compiler
771 * optimising this loop into a divmod call.
772 * See __iter_div_u64_rem() for another example of this.
774 asm("" : "+rm" (rq
->age_stamp
));
775 rq
->age_stamp
+= period
;
780 #endif /* CONFIG_SMP */
782 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
783 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
785 * Iterate task_group tree rooted at *from, calling @down when first entering a
786 * node and @up when leaving it for the final time.
788 * Caller must hold rcu_lock or sufficient equivalent.
790 int walk_tg_tree_from(struct task_group
*from
,
791 tg_visitor down
, tg_visitor up
, void *data
)
793 struct task_group
*parent
, *child
;
799 ret
= (*down
)(parent
, data
);
802 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
809 ret
= (*up
)(parent
, data
);
810 if (ret
|| parent
== from
)
814 parent
= parent
->parent
;
821 int tg_nop(struct task_group
*tg
, void *data
)
827 static void set_load_weight(struct task_struct
*p
)
829 int prio
= p
->static_prio
- MAX_RT_PRIO
;
830 struct load_weight
*load
= &p
->se
.load
;
833 * SCHED_IDLE tasks get minimal weight:
835 if (p
->policy
== SCHED_IDLE
) {
836 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
837 load
->inv_weight
= WMULT_IDLEPRIO
;
841 load
->weight
= scale_load(prio_to_weight
[prio
]);
842 load
->inv_weight
= prio_to_wmult
[prio
];
845 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
848 sched_info_queued(rq
, p
);
849 p
->sched_class
->enqueue_task(rq
, p
, flags
);
852 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
855 sched_info_dequeued(rq
, p
);
856 p
->sched_class
->dequeue_task(rq
, p
, flags
);
859 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
861 if (task_contributes_to_load(p
))
862 rq
->nr_uninterruptible
--;
864 enqueue_task(rq
, p
, flags
);
867 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
869 if (task_contributes_to_load(p
))
870 rq
->nr_uninterruptible
++;
872 dequeue_task(rq
, p
, flags
);
875 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
878 * In theory, the compile should just see 0 here, and optimize out the call
879 * to sched_rt_avg_update. But I don't trust it...
881 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
882 s64 steal
= 0, irq_delta
= 0;
884 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
885 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
888 * Since irq_time is only updated on {soft,}irq_exit, we might run into
889 * this case when a previous update_rq_clock() happened inside a
892 * When this happens, we stop ->clock_task and only update the
893 * prev_irq_time stamp to account for the part that fit, so that a next
894 * update will consume the rest. This ensures ->clock_task is
897 * It does however cause some slight miss-attribution of {soft,}irq
898 * time, a more accurate solution would be to update the irq_time using
899 * the current rq->clock timestamp, except that would require using
902 if (irq_delta
> delta
)
905 rq
->prev_irq_time
+= irq_delta
;
908 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
909 if (static_key_false((¶virt_steal_rq_enabled
))) {
910 steal
= paravirt_steal_clock(cpu_of(rq
));
911 steal
-= rq
->prev_steal_time_rq
;
913 if (unlikely(steal
> delta
))
916 rq
->prev_steal_time_rq
+= steal
;
921 rq
->clock_task
+= delta
;
923 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
924 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
925 sched_rt_avg_update(rq
, irq_delta
+ steal
);
929 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
931 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
932 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
936 * Make it appear like a SCHED_FIFO task, its something
937 * userspace knows about and won't get confused about.
939 * Also, it will make PI more or less work without too
940 * much confusion -- but then, stop work should not
941 * rely on PI working anyway.
943 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
945 stop
->sched_class
= &stop_sched_class
;
948 cpu_rq(cpu
)->stop
= stop
;
952 * Reset it back to a normal scheduling class so that
953 * it can die in pieces.
955 old_stop
->sched_class
= &rt_sched_class
;
960 * __normal_prio - return the priority that is based on the static prio
962 static inline int __normal_prio(struct task_struct
*p
)
964 return p
->static_prio
;
968 * Calculate the expected normal priority: i.e. priority
969 * without taking RT-inheritance into account. Might be
970 * boosted by interactivity modifiers. Changes upon fork,
971 * setprio syscalls, and whenever the interactivity
972 * estimator recalculates.
974 static inline int normal_prio(struct task_struct
*p
)
978 if (task_has_dl_policy(p
))
979 prio
= MAX_DL_PRIO
-1;
980 else if (task_has_rt_policy(p
))
981 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
983 prio
= __normal_prio(p
);
988 * Calculate the current priority, i.e. the priority
989 * taken into account by the scheduler. This value might
990 * be boosted by RT tasks, or might be boosted by
991 * interactivity modifiers. Will be RT if the task got
992 * RT-boosted. If not then it returns p->normal_prio.
994 static int effective_prio(struct task_struct
*p
)
996 p
->normal_prio
= normal_prio(p
);
998 * If we are RT tasks or we were boosted to RT priority,
999 * keep the priority unchanged. Otherwise, update priority
1000 * to the normal priority:
1002 if (!rt_prio(p
->prio
))
1003 return p
->normal_prio
;
1008 * task_curr - is this task currently executing on a CPU?
1009 * @p: the task in question.
1011 * Return: 1 if the task is currently executing. 0 otherwise.
1013 inline int task_curr(const struct task_struct
*p
)
1015 return cpu_curr(task_cpu(p
)) == p
;
1019 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
1021 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1022 const struct sched_class
*prev_class
,
1025 if (prev_class
!= p
->sched_class
) {
1026 if (prev_class
->switched_from
)
1027 prev_class
->switched_from(rq
, p
);
1028 /* Possble rq->lock 'hole'. */
1029 p
->sched_class
->switched_to(rq
, p
);
1030 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1031 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1034 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1036 const struct sched_class
*class;
1038 if (p
->sched_class
== rq
->curr
->sched_class
) {
1039 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1041 for_each_class(class) {
1042 if (class == rq
->curr
->sched_class
)
1044 if (class == p
->sched_class
) {
1052 * A queue event has occurred, and we're going to schedule. In
1053 * this case, we can save a useless back to back clock update.
1055 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1056 rq_clock_skip_update(rq
, true);
1060 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1062 #ifdef CONFIG_SCHED_DEBUG
1064 * We should never call set_task_cpu() on a blocked task,
1065 * ttwu() will sort out the placement.
1067 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1070 #ifdef CONFIG_LOCKDEP
1072 * The caller should hold either p->pi_lock or rq->lock, when changing
1073 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1075 * sched_move_task() holds both and thus holding either pins the cgroup,
1078 * Furthermore, all task_rq users should acquire both locks, see
1081 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1082 lockdep_is_held(&task_rq(p
)->lock
)));
1086 trace_sched_migrate_task(p
, new_cpu
);
1088 if (task_cpu(p
) != new_cpu
) {
1089 if (p
->sched_class
->migrate_task_rq
)
1090 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1091 p
->se
.nr_migrations
++;
1092 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 0);
1095 __set_task_cpu(p
, new_cpu
);
1098 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1100 if (task_on_rq_queued(p
)) {
1101 struct rq
*src_rq
, *dst_rq
;
1103 src_rq
= task_rq(p
);
1104 dst_rq
= cpu_rq(cpu
);
1106 deactivate_task(src_rq
, p
, 0);
1107 set_task_cpu(p
, cpu
);
1108 activate_task(dst_rq
, p
, 0);
1109 check_preempt_curr(dst_rq
, p
, 0);
1112 * Task isn't running anymore; make it appear like we migrated
1113 * it before it went to sleep. This means on wakeup we make the
1114 * previous cpu our targer instead of where it really is.
1120 struct migration_swap_arg
{
1121 struct task_struct
*src_task
, *dst_task
;
1122 int src_cpu
, dst_cpu
;
1125 static int migrate_swap_stop(void *data
)
1127 struct migration_swap_arg
*arg
= data
;
1128 struct rq
*src_rq
, *dst_rq
;
1131 src_rq
= cpu_rq(arg
->src_cpu
);
1132 dst_rq
= cpu_rq(arg
->dst_cpu
);
1134 double_raw_lock(&arg
->src_task
->pi_lock
,
1135 &arg
->dst_task
->pi_lock
);
1136 double_rq_lock(src_rq
, dst_rq
);
1137 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1140 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1143 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1146 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1149 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1150 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1155 double_rq_unlock(src_rq
, dst_rq
);
1156 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1157 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1163 * Cross migrate two tasks
1165 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1167 struct migration_swap_arg arg
;
1170 arg
= (struct migration_swap_arg
){
1172 .src_cpu
= task_cpu(cur
),
1174 .dst_cpu
= task_cpu(p
),
1177 if (arg
.src_cpu
== arg
.dst_cpu
)
1181 * These three tests are all lockless; this is OK since all of them
1182 * will be re-checked with proper locks held further down the line.
1184 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1187 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1190 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1193 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1194 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1200 struct migration_arg
{
1201 struct task_struct
*task
;
1205 static int migration_cpu_stop(void *data
);
1208 * wait_task_inactive - wait for a thread to unschedule.
1210 * If @match_state is nonzero, it's the @p->state value just checked and
1211 * not expected to change. If it changes, i.e. @p might have woken up,
1212 * then return zero. When we succeed in waiting for @p to be off its CPU,
1213 * we return a positive number (its total switch count). If a second call
1214 * a short while later returns the same number, the caller can be sure that
1215 * @p has remained unscheduled the whole time.
1217 * The caller must ensure that the task *will* unschedule sometime soon,
1218 * else this function might spin for a *long* time. This function can't
1219 * be called with interrupts off, or it may introduce deadlock with
1220 * smp_call_function() if an IPI is sent by the same process we are
1221 * waiting to become inactive.
1223 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1225 unsigned long flags
;
1226 int running
, queued
;
1232 * We do the initial early heuristics without holding
1233 * any task-queue locks at all. We'll only try to get
1234 * the runqueue lock when things look like they will
1240 * If the task is actively running on another CPU
1241 * still, just relax and busy-wait without holding
1244 * NOTE! Since we don't hold any locks, it's not
1245 * even sure that "rq" stays as the right runqueue!
1246 * But we don't care, since "task_running()" will
1247 * return false if the runqueue has changed and p
1248 * is actually now running somewhere else!
1250 while (task_running(rq
, p
)) {
1251 if (match_state
&& unlikely(p
->state
!= match_state
))
1257 * Ok, time to look more closely! We need the rq
1258 * lock now, to be *sure*. If we're wrong, we'll
1259 * just go back and repeat.
1261 rq
= task_rq_lock(p
, &flags
);
1262 trace_sched_wait_task(p
);
1263 running
= task_running(rq
, p
);
1264 queued
= task_on_rq_queued(p
);
1266 if (!match_state
|| p
->state
== match_state
)
1267 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1268 task_rq_unlock(rq
, p
, &flags
);
1271 * If it changed from the expected state, bail out now.
1273 if (unlikely(!ncsw
))
1277 * Was it really running after all now that we
1278 * checked with the proper locks actually held?
1280 * Oops. Go back and try again..
1282 if (unlikely(running
)) {
1288 * It's not enough that it's not actively running,
1289 * it must be off the runqueue _entirely_, and not
1292 * So if it was still runnable (but just not actively
1293 * running right now), it's preempted, and we should
1294 * yield - it could be a while.
1296 if (unlikely(queued
)) {
1297 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1299 set_current_state(TASK_UNINTERRUPTIBLE
);
1300 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1305 * Ahh, all good. It wasn't running, and it wasn't
1306 * runnable, which means that it will never become
1307 * running in the future either. We're all done!
1316 * kick_process - kick a running thread to enter/exit the kernel
1317 * @p: the to-be-kicked thread
1319 * Cause a process which is running on another CPU to enter
1320 * kernel-mode, without any delay. (to get signals handled.)
1322 * NOTE: this function doesn't have to take the runqueue lock,
1323 * because all it wants to ensure is that the remote task enters
1324 * the kernel. If the IPI races and the task has been migrated
1325 * to another CPU then no harm is done and the purpose has been
1328 void kick_process(struct task_struct
*p
)
1334 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1335 smp_send_reschedule(cpu
);
1338 EXPORT_SYMBOL_GPL(kick_process
);
1339 #endif /* CONFIG_SMP */
1343 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1345 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1347 int nid
= cpu_to_node(cpu
);
1348 const struct cpumask
*nodemask
= NULL
;
1349 enum { cpuset
, possible
, fail
} state
= cpuset
;
1353 * If the node that the cpu is on has been offlined, cpu_to_node()
1354 * will return -1. There is no cpu on the node, and we should
1355 * select the cpu on the other node.
1358 nodemask
= cpumask_of_node(nid
);
1360 /* Look for allowed, online CPU in same node. */
1361 for_each_cpu(dest_cpu
, nodemask
) {
1362 if (!cpu_online(dest_cpu
))
1364 if (!cpu_active(dest_cpu
))
1366 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1372 /* Any allowed, online CPU? */
1373 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1374 if (!cpu_online(dest_cpu
))
1376 if (!cpu_active(dest_cpu
))
1383 /* No more Mr. Nice Guy. */
1384 cpuset_cpus_allowed_fallback(p
);
1389 do_set_cpus_allowed(p
, cpu_possible_mask
);
1400 if (state
!= cpuset
) {
1402 * Don't tell them about moving exiting tasks or
1403 * kernel threads (both mm NULL), since they never
1406 if (p
->mm
&& printk_ratelimit()) {
1407 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1408 task_pid_nr(p
), p
->comm
, cpu
);
1416 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1419 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1421 if (p
->nr_cpus_allowed
> 1)
1422 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1425 * In order not to call set_task_cpu() on a blocking task we need
1426 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1429 * Since this is common to all placement strategies, this lives here.
1431 * [ this allows ->select_task() to simply return task_cpu(p) and
1432 * not worry about this generic constraint ]
1434 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1436 cpu
= select_fallback_rq(task_cpu(p
), p
);
1441 static void update_avg(u64
*avg
, u64 sample
)
1443 s64 diff
= sample
- *avg
;
1449 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1451 #ifdef CONFIG_SCHEDSTATS
1452 struct rq
*rq
= this_rq();
1455 int this_cpu
= smp_processor_id();
1457 if (cpu
== this_cpu
) {
1458 schedstat_inc(rq
, ttwu_local
);
1459 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1461 struct sched_domain
*sd
;
1463 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1465 for_each_domain(this_cpu
, sd
) {
1466 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1467 schedstat_inc(sd
, ttwu_wake_remote
);
1474 if (wake_flags
& WF_MIGRATED
)
1475 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1477 #endif /* CONFIG_SMP */
1479 schedstat_inc(rq
, ttwu_count
);
1480 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1482 if (wake_flags
& WF_SYNC
)
1483 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1485 #endif /* CONFIG_SCHEDSTATS */
1488 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1490 activate_task(rq
, p
, en_flags
);
1491 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1493 /* if a worker is waking up, notify workqueue */
1494 if (p
->flags
& PF_WQ_WORKER
)
1495 wq_worker_waking_up(p
, cpu_of(rq
));
1499 * Mark the task runnable and perform wakeup-preemption.
1502 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1504 check_preempt_curr(rq
, p
, wake_flags
);
1505 trace_sched_wakeup(p
, true);
1507 p
->state
= TASK_RUNNING
;
1509 if (p
->sched_class
->task_woken
)
1510 p
->sched_class
->task_woken(rq
, p
);
1512 if (rq
->idle_stamp
) {
1513 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1514 u64 max
= 2*rq
->max_idle_balance_cost
;
1516 update_avg(&rq
->avg_idle
, delta
);
1518 if (rq
->avg_idle
> max
)
1527 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1530 if (p
->sched_contributes_to_load
)
1531 rq
->nr_uninterruptible
--;
1534 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1535 ttwu_do_wakeup(rq
, p
, wake_flags
);
1539 * Called in case the task @p isn't fully descheduled from its runqueue,
1540 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1541 * since all we need to do is flip p->state to TASK_RUNNING, since
1542 * the task is still ->on_rq.
1544 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1549 rq
= __task_rq_lock(p
);
1550 if (task_on_rq_queued(p
)) {
1551 /* check_preempt_curr() may use rq clock */
1552 update_rq_clock(rq
);
1553 ttwu_do_wakeup(rq
, p
, wake_flags
);
1556 __task_rq_unlock(rq
);
1562 void sched_ttwu_pending(void)
1564 struct rq
*rq
= this_rq();
1565 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1566 struct task_struct
*p
;
1567 unsigned long flags
;
1572 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1575 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1576 llist
= llist_next(llist
);
1577 ttwu_do_activate(rq
, p
, 0);
1580 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1583 void scheduler_ipi(void)
1586 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1587 * TIF_NEED_RESCHED remotely (for the first time) will also send
1590 preempt_fold_need_resched();
1592 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1596 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1597 * traditionally all their work was done from the interrupt return
1598 * path. Now that we actually do some work, we need to make sure
1601 * Some archs already do call them, luckily irq_enter/exit nest
1604 * Arguably we should visit all archs and update all handlers,
1605 * however a fair share of IPIs are still resched only so this would
1606 * somewhat pessimize the simple resched case.
1609 sched_ttwu_pending();
1612 * Check if someone kicked us for doing the nohz idle load balance.
1614 if (unlikely(got_nohz_idle_kick())) {
1615 this_rq()->idle_balance
= 1;
1616 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1621 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1623 struct rq
*rq
= cpu_rq(cpu
);
1625 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1626 if (!set_nr_if_polling(rq
->idle
))
1627 smp_send_reschedule(cpu
);
1629 trace_sched_wake_idle_without_ipi(cpu
);
1633 void wake_up_if_idle(int cpu
)
1635 struct rq
*rq
= cpu_rq(cpu
);
1636 unsigned long flags
;
1640 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1643 if (set_nr_if_polling(rq
->idle
)) {
1644 trace_sched_wake_idle_without_ipi(cpu
);
1646 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1647 if (is_idle_task(rq
->curr
))
1648 smp_send_reschedule(cpu
);
1649 /* Else cpu is not in idle, do nothing here */
1650 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1657 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1659 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1661 #endif /* CONFIG_SMP */
1663 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1665 struct rq
*rq
= cpu_rq(cpu
);
1667 #if defined(CONFIG_SMP)
1668 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1669 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1670 ttwu_queue_remote(p
, cpu
);
1675 raw_spin_lock(&rq
->lock
);
1676 ttwu_do_activate(rq
, p
, 0);
1677 raw_spin_unlock(&rq
->lock
);
1681 * try_to_wake_up - wake up a thread
1682 * @p: the thread to be awakened
1683 * @state: the mask of task states that can be woken
1684 * @wake_flags: wake modifier flags (WF_*)
1686 * Put it on the run-queue if it's not already there. The "current"
1687 * thread is always on the run-queue (except when the actual
1688 * re-schedule is in progress), and as such you're allowed to do
1689 * the simpler "current->state = TASK_RUNNING" to mark yourself
1690 * runnable without the overhead of this.
1692 * Return: %true if @p was woken up, %false if it was already running.
1693 * or @state didn't match @p's state.
1696 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1698 unsigned long flags
;
1699 int cpu
, success
= 0;
1702 * If we are going to wake up a thread waiting for CONDITION we
1703 * need to ensure that CONDITION=1 done by the caller can not be
1704 * reordered with p->state check below. This pairs with mb() in
1705 * set_current_state() the waiting thread does.
1707 smp_mb__before_spinlock();
1708 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1709 if (!(p
->state
& state
))
1712 success
= 1; /* we're going to change ->state */
1715 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1720 * If the owning (remote) cpu is still in the middle of schedule() with
1721 * this task as prev, wait until its done referencing the task.
1726 * Pairs with the smp_wmb() in finish_lock_switch().
1730 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1731 p
->state
= TASK_WAKING
;
1733 if (p
->sched_class
->task_waking
)
1734 p
->sched_class
->task_waking(p
);
1736 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1737 if (task_cpu(p
) != cpu
) {
1738 wake_flags
|= WF_MIGRATED
;
1739 set_task_cpu(p
, cpu
);
1741 #endif /* CONFIG_SMP */
1745 ttwu_stat(p
, cpu
, wake_flags
);
1747 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1753 * try_to_wake_up_local - try to wake up a local task with rq lock held
1754 * @p: the thread to be awakened
1756 * Put @p on the run-queue if it's not already there. The caller must
1757 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1760 static void try_to_wake_up_local(struct task_struct
*p
)
1762 struct rq
*rq
= task_rq(p
);
1764 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1765 WARN_ON_ONCE(p
== current
))
1768 lockdep_assert_held(&rq
->lock
);
1770 if (!raw_spin_trylock(&p
->pi_lock
)) {
1771 raw_spin_unlock(&rq
->lock
);
1772 raw_spin_lock(&p
->pi_lock
);
1773 raw_spin_lock(&rq
->lock
);
1776 if (!(p
->state
& TASK_NORMAL
))
1779 if (!task_on_rq_queued(p
))
1780 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1782 ttwu_do_wakeup(rq
, p
, 0);
1783 ttwu_stat(p
, smp_processor_id(), 0);
1785 raw_spin_unlock(&p
->pi_lock
);
1789 * wake_up_process - Wake up a specific process
1790 * @p: The process to be woken up.
1792 * Attempt to wake up the nominated process and move it to the set of runnable
1795 * Return: 1 if the process was woken up, 0 if it was already running.
1797 * It may be assumed that this function implies a write memory barrier before
1798 * changing the task state if and only if any tasks are woken up.
1800 int wake_up_process(struct task_struct
*p
)
1802 WARN_ON(task_is_stopped_or_traced(p
));
1803 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1805 EXPORT_SYMBOL(wake_up_process
);
1807 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1809 return try_to_wake_up(p
, state
, 0);
1813 * This function clears the sched_dl_entity static params.
1815 void __dl_clear_params(struct task_struct
*p
)
1817 struct sched_dl_entity
*dl_se
= &p
->dl
;
1819 dl_se
->dl_runtime
= 0;
1820 dl_se
->dl_deadline
= 0;
1821 dl_se
->dl_period
= 0;
1825 dl_se
->dl_throttled
= 0;
1827 dl_se
->dl_yielded
= 0;
1831 * Perform scheduler related setup for a newly forked process p.
1832 * p is forked by current.
1834 * __sched_fork() is basic setup used by init_idle() too:
1836 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1841 p
->se
.exec_start
= 0;
1842 p
->se
.sum_exec_runtime
= 0;
1843 p
->se
.prev_sum_exec_runtime
= 0;
1844 p
->se
.nr_migrations
= 0;
1847 p
->se
.avg
.decay_count
= 0;
1849 INIT_LIST_HEAD(&p
->se
.group_node
);
1851 #ifdef CONFIG_SCHEDSTATS
1852 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1855 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1856 init_dl_task_timer(&p
->dl
);
1857 __dl_clear_params(p
);
1859 INIT_LIST_HEAD(&p
->rt
.run_list
);
1861 #ifdef CONFIG_PREEMPT_NOTIFIERS
1862 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1865 #ifdef CONFIG_NUMA_BALANCING
1866 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1867 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1868 p
->mm
->numa_scan_seq
= 0;
1871 if (clone_flags
& CLONE_VM
)
1872 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1874 p
->numa_preferred_nid
= -1;
1876 p
->node_stamp
= 0ULL;
1877 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1878 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1879 p
->numa_work
.next
= &p
->numa_work
;
1880 p
->numa_faults
= NULL
;
1881 p
->last_task_numa_placement
= 0;
1882 p
->last_sum_exec_runtime
= 0;
1884 p
->numa_group
= NULL
;
1885 #endif /* CONFIG_NUMA_BALANCING */
1888 #ifdef CONFIG_NUMA_BALANCING
1889 #ifdef CONFIG_SCHED_DEBUG
1890 void set_numabalancing_state(bool enabled
)
1893 sched_feat_set("NUMA");
1895 sched_feat_set("NO_NUMA");
1898 __read_mostly
bool numabalancing_enabled
;
1900 void set_numabalancing_state(bool enabled
)
1902 numabalancing_enabled
= enabled
;
1904 #endif /* CONFIG_SCHED_DEBUG */
1906 #ifdef CONFIG_PROC_SYSCTL
1907 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1908 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1912 int state
= numabalancing_enabled
;
1914 if (write
&& !capable(CAP_SYS_ADMIN
))
1919 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1923 set_numabalancing_state(state
);
1930 * fork()/clone()-time setup:
1932 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1934 unsigned long flags
;
1935 int cpu
= get_cpu();
1937 __sched_fork(clone_flags
, p
);
1939 * We mark the process as running here. This guarantees that
1940 * nobody will actually run it, and a signal or other external
1941 * event cannot wake it up and insert it on the runqueue either.
1943 p
->state
= TASK_RUNNING
;
1946 * Make sure we do not leak PI boosting priority to the child.
1948 p
->prio
= current
->normal_prio
;
1951 * Revert to default priority/policy on fork if requested.
1953 if (unlikely(p
->sched_reset_on_fork
)) {
1954 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1955 p
->policy
= SCHED_NORMAL
;
1956 p
->static_prio
= NICE_TO_PRIO(0);
1958 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1959 p
->static_prio
= NICE_TO_PRIO(0);
1961 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1965 * We don't need the reset flag anymore after the fork. It has
1966 * fulfilled its duty:
1968 p
->sched_reset_on_fork
= 0;
1971 if (dl_prio(p
->prio
)) {
1974 } else if (rt_prio(p
->prio
)) {
1975 p
->sched_class
= &rt_sched_class
;
1977 p
->sched_class
= &fair_sched_class
;
1980 if (p
->sched_class
->task_fork
)
1981 p
->sched_class
->task_fork(p
);
1984 * The child is not yet in the pid-hash so no cgroup attach races,
1985 * and the cgroup is pinned to this child due to cgroup_fork()
1986 * is ran before sched_fork().
1988 * Silence PROVE_RCU.
1990 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1991 set_task_cpu(p
, cpu
);
1992 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1994 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1995 if (likely(sched_info_on()))
1996 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1998 #if defined(CONFIG_SMP)
2001 init_task_preempt_count(p
);
2003 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2004 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2011 unsigned long to_ratio(u64 period
, u64 runtime
)
2013 if (runtime
== RUNTIME_INF
)
2017 * Doing this here saves a lot of checks in all
2018 * the calling paths, and returning zero seems
2019 * safe for them anyway.
2024 return div64_u64(runtime
<< 20, period
);
2028 inline struct dl_bw
*dl_bw_of(int i
)
2030 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2031 "sched RCU must be held");
2032 return &cpu_rq(i
)->rd
->dl_bw
;
2035 static inline int dl_bw_cpus(int i
)
2037 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2040 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2041 "sched RCU must be held");
2042 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2048 inline struct dl_bw
*dl_bw_of(int i
)
2050 return &cpu_rq(i
)->dl
.dl_bw
;
2053 static inline int dl_bw_cpus(int i
)
2060 * We must be sure that accepting a new task (or allowing changing the
2061 * parameters of an existing one) is consistent with the bandwidth
2062 * constraints. If yes, this function also accordingly updates the currently
2063 * allocated bandwidth to reflect the new situation.
2065 * This function is called while holding p's rq->lock.
2067 * XXX we should delay bw change until the task's 0-lag point, see
2070 static int dl_overflow(struct task_struct
*p
, int policy
,
2071 const struct sched_attr
*attr
)
2074 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2075 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2076 u64 runtime
= attr
->sched_runtime
;
2077 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2080 if (new_bw
== p
->dl
.dl_bw
)
2084 * Either if a task, enters, leave, or stays -deadline but changes
2085 * its parameters, we may need to update accordingly the total
2086 * allocated bandwidth of the container.
2088 raw_spin_lock(&dl_b
->lock
);
2089 cpus
= dl_bw_cpus(task_cpu(p
));
2090 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2091 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2092 __dl_add(dl_b
, new_bw
);
2094 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2095 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2096 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2097 __dl_add(dl_b
, new_bw
);
2099 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2100 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2103 raw_spin_unlock(&dl_b
->lock
);
2108 extern void init_dl_bw(struct dl_bw
*dl_b
);
2111 * wake_up_new_task - wake up a newly created task for the first time.
2113 * This function will do some initial scheduler statistics housekeeping
2114 * that must be done for every newly created context, then puts the task
2115 * on the runqueue and wakes it.
2117 void wake_up_new_task(struct task_struct
*p
)
2119 unsigned long flags
;
2122 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2125 * Fork balancing, do it here and not earlier because:
2126 * - cpus_allowed can change in the fork path
2127 * - any previously selected cpu might disappear through hotplug
2129 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2132 /* Initialize new task's runnable average */
2133 init_task_runnable_average(p
);
2134 rq
= __task_rq_lock(p
);
2135 activate_task(rq
, p
, 0);
2136 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2137 trace_sched_wakeup_new(p
, true);
2138 check_preempt_curr(rq
, p
, WF_FORK
);
2140 if (p
->sched_class
->task_woken
)
2141 p
->sched_class
->task_woken(rq
, p
);
2143 task_rq_unlock(rq
, p
, &flags
);
2146 #ifdef CONFIG_PREEMPT_NOTIFIERS
2149 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2150 * @notifier: notifier struct to register
2152 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2154 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2156 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2159 * preempt_notifier_unregister - no longer interested in preemption notifications
2160 * @notifier: notifier struct to unregister
2162 * This is safe to call from within a preemption notifier.
2164 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2166 hlist_del(¬ifier
->link
);
2168 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2170 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2172 struct preempt_notifier
*notifier
;
2174 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2175 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2179 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2180 struct task_struct
*next
)
2182 struct preempt_notifier
*notifier
;
2184 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2185 notifier
->ops
->sched_out(notifier
, next
);
2188 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2190 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2195 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2196 struct task_struct
*next
)
2200 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2203 * prepare_task_switch - prepare to switch tasks
2204 * @rq: the runqueue preparing to switch
2205 * @prev: the current task that is being switched out
2206 * @next: the task we are going to switch to.
2208 * This is called with the rq lock held and interrupts off. It must
2209 * be paired with a subsequent finish_task_switch after the context
2212 * prepare_task_switch sets up locking and calls architecture specific
2216 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2217 struct task_struct
*next
)
2219 trace_sched_switch(prev
, next
);
2220 sched_info_switch(rq
, prev
, next
);
2221 perf_event_task_sched_out(prev
, next
);
2222 fire_sched_out_preempt_notifiers(prev
, next
);
2223 prepare_lock_switch(rq
, next
);
2224 prepare_arch_switch(next
);
2228 * finish_task_switch - clean up after a task-switch
2229 * @prev: the thread we just switched away from.
2231 * finish_task_switch must be called after the context switch, paired
2232 * with a prepare_task_switch call before the context switch.
2233 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2234 * and do any other architecture-specific cleanup actions.
2236 * Note that we may have delayed dropping an mm in context_switch(). If
2237 * so, we finish that here outside of the runqueue lock. (Doing it
2238 * with the lock held can cause deadlocks; see schedule() for
2241 * The context switch have flipped the stack from under us and restored the
2242 * local variables which were saved when this task called schedule() in the
2243 * past. prev == current is still correct but we need to recalculate this_rq
2244 * because prev may have moved to another CPU.
2246 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2247 __releases(rq
->lock
)
2249 struct rq
*rq
= this_rq();
2250 struct mm_struct
*mm
= rq
->prev_mm
;
2256 * A task struct has one reference for the use as "current".
2257 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2258 * schedule one last time. The schedule call will never return, and
2259 * the scheduled task must drop that reference.
2260 * The test for TASK_DEAD must occur while the runqueue locks are
2261 * still held, otherwise prev could be scheduled on another cpu, die
2262 * there before we look at prev->state, and then the reference would
2264 * Manfred Spraul <manfred@colorfullife.com>
2266 prev_state
= prev
->state
;
2267 vtime_task_switch(prev
);
2268 finish_arch_switch(prev
);
2269 perf_event_task_sched_in(prev
, current
);
2270 finish_lock_switch(rq
, prev
);
2271 finish_arch_post_lock_switch();
2273 fire_sched_in_preempt_notifiers(current
);
2276 if (unlikely(prev_state
== TASK_DEAD
)) {
2277 if (prev
->sched_class
->task_dead
)
2278 prev
->sched_class
->task_dead(prev
);
2281 * Remove function-return probe instances associated with this
2282 * task and put them back on the free list.
2284 kprobe_flush_task(prev
);
2285 put_task_struct(prev
);
2288 tick_nohz_task_switch(current
);
2294 /* rq->lock is NOT held, but preemption is disabled */
2295 static inline void post_schedule(struct rq
*rq
)
2297 if (rq
->post_schedule
) {
2298 unsigned long flags
;
2300 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2301 if (rq
->curr
->sched_class
->post_schedule
)
2302 rq
->curr
->sched_class
->post_schedule(rq
);
2303 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2305 rq
->post_schedule
= 0;
2311 static inline void post_schedule(struct rq
*rq
)
2318 * schedule_tail - first thing a freshly forked thread must call.
2319 * @prev: the thread we just switched away from.
2321 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2322 __releases(rq
->lock
)
2326 /* finish_task_switch() drops rq->lock and enables preemtion */
2328 rq
= finish_task_switch(prev
);
2332 if (current
->set_child_tid
)
2333 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2337 * context_switch - switch to the new MM and the new thread's register state.
2339 static inline struct rq
*
2340 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2341 struct task_struct
*next
)
2343 struct mm_struct
*mm
, *oldmm
;
2345 prepare_task_switch(rq
, prev
, next
);
2348 oldmm
= prev
->active_mm
;
2350 * For paravirt, this is coupled with an exit in switch_to to
2351 * combine the page table reload and the switch backend into
2354 arch_start_context_switch(prev
);
2357 next
->active_mm
= oldmm
;
2358 atomic_inc(&oldmm
->mm_count
);
2359 enter_lazy_tlb(oldmm
, next
);
2361 switch_mm(oldmm
, mm
, next
);
2364 prev
->active_mm
= NULL
;
2365 rq
->prev_mm
= oldmm
;
2368 * Since the runqueue lock will be released by the next
2369 * task (which is an invalid locking op but in the case
2370 * of the scheduler it's an obvious special-case), so we
2371 * do an early lockdep release here:
2373 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2375 context_tracking_task_switch(prev
, next
);
2376 /* Here we just switch the register state and the stack. */
2377 switch_to(prev
, next
, prev
);
2380 return finish_task_switch(prev
);
2384 * nr_running and nr_context_switches:
2386 * externally visible scheduler statistics: current number of runnable
2387 * threads, total number of context switches performed since bootup.
2389 unsigned long nr_running(void)
2391 unsigned long i
, sum
= 0;
2393 for_each_online_cpu(i
)
2394 sum
+= cpu_rq(i
)->nr_running
;
2400 * Check if only the current task is running on the cpu.
2402 bool single_task_running(void)
2404 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2409 EXPORT_SYMBOL(single_task_running
);
2411 unsigned long long nr_context_switches(void)
2414 unsigned long long sum
= 0;
2416 for_each_possible_cpu(i
)
2417 sum
+= cpu_rq(i
)->nr_switches
;
2422 unsigned long nr_iowait(void)
2424 unsigned long i
, sum
= 0;
2426 for_each_possible_cpu(i
)
2427 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2432 unsigned long nr_iowait_cpu(int cpu
)
2434 struct rq
*this = cpu_rq(cpu
);
2435 return atomic_read(&this->nr_iowait
);
2438 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2440 struct rq
*this = this_rq();
2441 *nr_waiters
= atomic_read(&this->nr_iowait
);
2442 *load
= this->cpu_load
[0];
2448 * sched_exec - execve() is a valuable balancing opportunity, because at
2449 * this point the task has the smallest effective memory and cache footprint.
2451 void sched_exec(void)
2453 struct task_struct
*p
= current
;
2454 unsigned long flags
;
2457 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2458 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2459 if (dest_cpu
== smp_processor_id())
2462 if (likely(cpu_active(dest_cpu
))) {
2463 struct migration_arg arg
= { p
, dest_cpu
};
2465 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2466 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2470 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2475 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2476 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2478 EXPORT_PER_CPU_SYMBOL(kstat
);
2479 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2482 * Return accounted runtime for the task.
2483 * In case the task is currently running, return the runtime plus current's
2484 * pending runtime that have not been accounted yet.
2486 unsigned long long task_sched_runtime(struct task_struct
*p
)
2488 unsigned long flags
;
2492 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2494 * 64-bit doesn't need locks to atomically read a 64bit value.
2495 * So we have a optimization chance when the task's delta_exec is 0.
2496 * Reading ->on_cpu is racy, but this is ok.
2498 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2499 * If we race with it entering cpu, unaccounted time is 0. This is
2500 * indistinguishable from the read occurring a few cycles earlier.
2501 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2502 * been accounted, so we're correct here as well.
2504 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2505 return p
->se
.sum_exec_runtime
;
2508 rq
= task_rq_lock(p
, &flags
);
2510 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2511 * project cycles that may never be accounted to this
2512 * thread, breaking clock_gettime().
2514 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2515 update_rq_clock(rq
);
2516 p
->sched_class
->update_curr(rq
);
2518 ns
= p
->se
.sum_exec_runtime
;
2519 task_rq_unlock(rq
, p
, &flags
);
2525 * This function gets called by the timer code, with HZ frequency.
2526 * We call it with interrupts disabled.
2528 void scheduler_tick(void)
2530 int cpu
= smp_processor_id();
2531 struct rq
*rq
= cpu_rq(cpu
);
2532 struct task_struct
*curr
= rq
->curr
;
2536 raw_spin_lock(&rq
->lock
);
2537 update_rq_clock(rq
);
2538 curr
->sched_class
->task_tick(rq
, curr
, 0);
2539 update_cpu_load_active(rq
);
2540 raw_spin_unlock(&rq
->lock
);
2542 perf_event_task_tick();
2545 rq
->idle_balance
= idle_cpu(cpu
);
2546 trigger_load_balance(rq
);
2548 rq_last_tick_reset(rq
);
2551 #ifdef CONFIG_NO_HZ_FULL
2553 * scheduler_tick_max_deferment
2555 * Keep at least one tick per second when a single
2556 * active task is running because the scheduler doesn't
2557 * yet completely support full dynticks environment.
2559 * This makes sure that uptime, CFS vruntime, load
2560 * balancing, etc... continue to move forward, even
2561 * with a very low granularity.
2563 * Return: Maximum deferment in nanoseconds.
2565 u64
scheduler_tick_max_deferment(void)
2567 struct rq
*rq
= this_rq();
2568 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2570 next
= rq
->last_sched_tick
+ HZ
;
2572 if (time_before_eq(next
, now
))
2575 return jiffies_to_nsecs(next
- now
);
2579 notrace
unsigned long get_parent_ip(unsigned long addr
)
2581 if (in_lock_functions(addr
)) {
2582 addr
= CALLER_ADDR2
;
2583 if (in_lock_functions(addr
))
2584 addr
= CALLER_ADDR3
;
2589 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2590 defined(CONFIG_PREEMPT_TRACER))
2592 void preempt_count_add(int val
)
2594 #ifdef CONFIG_DEBUG_PREEMPT
2598 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2601 __preempt_count_add(val
);
2602 #ifdef CONFIG_DEBUG_PREEMPT
2604 * Spinlock count overflowing soon?
2606 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2609 if (preempt_count() == val
) {
2610 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2611 #ifdef CONFIG_DEBUG_PREEMPT
2612 current
->preempt_disable_ip
= ip
;
2614 trace_preempt_off(CALLER_ADDR0
, ip
);
2617 EXPORT_SYMBOL(preempt_count_add
);
2618 NOKPROBE_SYMBOL(preempt_count_add
);
2620 void preempt_count_sub(int val
)
2622 #ifdef CONFIG_DEBUG_PREEMPT
2626 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2629 * Is the spinlock portion underflowing?
2631 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2632 !(preempt_count() & PREEMPT_MASK
)))
2636 if (preempt_count() == val
)
2637 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2638 __preempt_count_sub(val
);
2640 EXPORT_SYMBOL(preempt_count_sub
);
2641 NOKPROBE_SYMBOL(preempt_count_sub
);
2646 * Print scheduling while atomic bug:
2648 static noinline
void __schedule_bug(struct task_struct
*prev
)
2650 if (oops_in_progress
)
2653 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2654 prev
->comm
, prev
->pid
, preempt_count());
2656 debug_show_held_locks(prev
);
2658 if (irqs_disabled())
2659 print_irqtrace_events(prev
);
2660 #ifdef CONFIG_DEBUG_PREEMPT
2661 if (in_atomic_preempt_off()) {
2662 pr_err("Preemption disabled at:");
2663 print_ip_sym(current
->preempt_disable_ip
);
2668 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2672 * Various schedule()-time debugging checks and statistics:
2674 static inline void schedule_debug(struct task_struct
*prev
)
2676 #ifdef CONFIG_SCHED_STACK_END_CHECK
2677 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2680 * Test if we are atomic. Since do_exit() needs to call into
2681 * schedule() atomically, we ignore that path. Otherwise whine
2682 * if we are scheduling when we should not.
2684 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2685 __schedule_bug(prev
);
2688 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2690 schedstat_inc(this_rq(), sched_count
);
2694 * Pick up the highest-prio task:
2696 static inline struct task_struct
*
2697 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2699 const struct sched_class
*class = &fair_sched_class
;
2700 struct task_struct
*p
;
2703 * Optimization: we know that if all tasks are in
2704 * the fair class we can call that function directly:
2706 if (likely(prev
->sched_class
== class &&
2707 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2708 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2709 if (unlikely(p
== RETRY_TASK
))
2712 /* assumes fair_sched_class->next == idle_sched_class */
2714 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2720 for_each_class(class) {
2721 p
= class->pick_next_task(rq
, prev
);
2723 if (unlikely(p
== RETRY_TASK
))
2729 BUG(); /* the idle class will always have a runnable task */
2733 * __schedule() is the main scheduler function.
2735 * The main means of driving the scheduler and thus entering this function are:
2737 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2739 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2740 * paths. For example, see arch/x86/entry_64.S.
2742 * To drive preemption between tasks, the scheduler sets the flag in timer
2743 * interrupt handler scheduler_tick().
2745 * 3. Wakeups don't really cause entry into schedule(). They add a
2746 * task to the run-queue and that's it.
2748 * Now, if the new task added to the run-queue preempts the current
2749 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2750 * called on the nearest possible occasion:
2752 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2754 * - in syscall or exception context, at the next outmost
2755 * preempt_enable(). (this might be as soon as the wake_up()'s
2758 * - in IRQ context, return from interrupt-handler to
2759 * preemptible context
2761 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2764 * - cond_resched() call
2765 * - explicit schedule() call
2766 * - return from syscall or exception to user-space
2767 * - return from interrupt-handler to user-space
2769 * WARNING: all callers must re-check need_resched() afterward and reschedule
2770 * accordingly in case an event triggered the need for rescheduling (such as
2771 * an interrupt waking up a task) while preemption was disabled in __schedule().
2773 static void __sched
__schedule(void)
2775 struct task_struct
*prev
, *next
;
2776 unsigned long *switch_count
;
2781 cpu
= smp_processor_id();
2783 rcu_note_context_switch();
2786 schedule_debug(prev
);
2788 if (sched_feat(HRTICK
))
2792 * Make sure that signal_pending_state()->signal_pending() below
2793 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2794 * done by the caller to avoid the race with signal_wake_up().
2796 smp_mb__before_spinlock();
2797 raw_spin_lock_irq(&rq
->lock
);
2799 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
2801 switch_count
= &prev
->nivcsw
;
2802 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2803 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2804 prev
->state
= TASK_RUNNING
;
2806 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2810 * If a worker went to sleep, notify and ask workqueue
2811 * whether it wants to wake up a task to maintain
2814 if (prev
->flags
& PF_WQ_WORKER
) {
2815 struct task_struct
*to_wakeup
;
2817 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2819 try_to_wake_up_local(to_wakeup
);
2822 switch_count
= &prev
->nvcsw
;
2825 if (task_on_rq_queued(prev
))
2826 update_rq_clock(rq
);
2828 next
= pick_next_task(rq
, prev
);
2829 clear_tsk_need_resched(prev
);
2830 clear_preempt_need_resched();
2831 rq
->clock_skip_update
= 0;
2833 if (likely(prev
!= next
)) {
2838 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
2841 raw_spin_unlock_irq(&rq
->lock
);
2845 sched_preempt_enable_no_resched();
2848 static inline void sched_submit_work(struct task_struct
*tsk
)
2850 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2853 * If we are going to sleep and we have plugged IO queued,
2854 * make sure to submit it to avoid deadlocks.
2856 if (blk_needs_flush_plug(tsk
))
2857 blk_schedule_flush_plug(tsk
);
2860 asmlinkage __visible
void __sched
schedule(void)
2862 struct task_struct
*tsk
= current
;
2864 sched_submit_work(tsk
);
2867 } while (need_resched());
2869 EXPORT_SYMBOL(schedule
);
2871 #ifdef CONFIG_CONTEXT_TRACKING
2872 asmlinkage __visible
void __sched
schedule_user(void)
2875 * If we come here after a random call to set_need_resched(),
2876 * or we have been woken up remotely but the IPI has not yet arrived,
2877 * we haven't yet exited the RCU idle mode. Do it here manually until
2878 * we find a better solution.
2880 * NB: There are buggy callers of this function. Ideally we
2881 * should warn if prev_state != IN_USER, but that will trigger
2882 * too frequently to make sense yet.
2884 enum ctx_state prev_state
= exception_enter();
2886 exception_exit(prev_state
);
2891 * schedule_preempt_disabled - called with preemption disabled
2893 * Returns with preemption disabled. Note: preempt_count must be 1
2895 void __sched
schedule_preempt_disabled(void)
2897 sched_preempt_enable_no_resched();
2902 static void preempt_schedule_common(void)
2905 __preempt_count_add(PREEMPT_ACTIVE
);
2907 __preempt_count_sub(PREEMPT_ACTIVE
);
2910 * Check again in case we missed a preemption opportunity
2911 * between schedule and now.
2914 } while (need_resched());
2917 #ifdef CONFIG_PREEMPT
2919 * this is the entry point to schedule() from in-kernel preemption
2920 * off of preempt_enable. Kernel preemptions off return from interrupt
2921 * occur there and call schedule directly.
2923 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2926 * If there is a non-zero preempt_count or interrupts are disabled,
2927 * we do not want to preempt the current task. Just return..
2929 if (likely(!preemptible()))
2932 preempt_schedule_common();
2934 NOKPROBE_SYMBOL(preempt_schedule
);
2935 EXPORT_SYMBOL(preempt_schedule
);
2937 #ifdef CONFIG_CONTEXT_TRACKING
2939 * preempt_schedule_context - preempt_schedule called by tracing
2941 * The tracing infrastructure uses preempt_enable_notrace to prevent
2942 * recursion and tracing preempt enabling caused by the tracing
2943 * infrastructure itself. But as tracing can happen in areas coming
2944 * from userspace or just about to enter userspace, a preempt enable
2945 * can occur before user_exit() is called. This will cause the scheduler
2946 * to be called when the system is still in usermode.
2948 * To prevent this, the preempt_enable_notrace will use this function
2949 * instead of preempt_schedule() to exit user context if needed before
2950 * calling the scheduler.
2952 asmlinkage __visible
void __sched notrace
preempt_schedule_context(void)
2954 enum ctx_state prev_ctx
;
2956 if (likely(!preemptible()))
2960 __preempt_count_add(PREEMPT_ACTIVE
);
2962 * Needs preempt disabled in case user_exit() is traced
2963 * and the tracer calls preempt_enable_notrace() causing
2964 * an infinite recursion.
2966 prev_ctx
= exception_enter();
2968 exception_exit(prev_ctx
);
2970 __preempt_count_sub(PREEMPT_ACTIVE
);
2972 } while (need_resched());
2974 EXPORT_SYMBOL_GPL(preempt_schedule_context
);
2975 #endif /* CONFIG_CONTEXT_TRACKING */
2977 #endif /* CONFIG_PREEMPT */
2980 * this is the entry point to schedule() from kernel preemption
2981 * off of irq context.
2982 * Note, that this is called and return with irqs disabled. This will
2983 * protect us against recursive calling from irq.
2985 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2987 enum ctx_state prev_state
;
2989 /* Catch callers which need to be fixed */
2990 BUG_ON(preempt_count() || !irqs_disabled());
2992 prev_state
= exception_enter();
2995 __preempt_count_add(PREEMPT_ACTIVE
);
2998 local_irq_disable();
2999 __preempt_count_sub(PREEMPT_ACTIVE
);
3002 * Check again in case we missed a preemption opportunity
3003 * between schedule and now.
3006 } while (need_resched());
3008 exception_exit(prev_state
);
3011 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3014 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3016 EXPORT_SYMBOL(default_wake_function
);
3018 #ifdef CONFIG_RT_MUTEXES
3021 * rt_mutex_setprio - set the current priority of a task
3023 * @prio: prio value (kernel-internal form)
3025 * This function changes the 'effective' priority of a task. It does
3026 * not touch ->normal_prio like __setscheduler().
3028 * Used by the rt_mutex code to implement priority inheritance
3029 * logic. Call site only calls if the priority of the task changed.
3031 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3033 int oldprio
, queued
, running
, enqueue_flag
= 0;
3035 const struct sched_class
*prev_class
;
3037 BUG_ON(prio
> MAX_PRIO
);
3039 rq
= __task_rq_lock(p
);
3042 * Idle task boosting is a nono in general. There is one
3043 * exception, when PREEMPT_RT and NOHZ is active:
3045 * The idle task calls get_next_timer_interrupt() and holds
3046 * the timer wheel base->lock on the CPU and another CPU wants
3047 * to access the timer (probably to cancel it). We can safely
3048 * ignore the boosting request, as the idle CPU runs this code
3049 * with interrupts disabled and will complete the lock
3050 * protected section without being interrupted. So there is no
3051 * real need to boost.
3053 if (unlikely(p
== rq
->idle
)) {
3054 WARN_ON(p
!= rq
->curr
);
3055 WARN_ON(p
->pi_blocked_on
);
3059 trace_sched_pi_setprio(p
, prio
);
3061 prev_class
= p
->sched_class
;
3062 queued
= task_on_rq_queued(p
);
3063 running
= task_current(rq
, p
);
3065 dequeue_task(rq
, p
, 0);
3067 put_prev_task(rq
, p
);
3070 * Boosting condition are:
3071 * 1. -rt task is running and holds mutex A
3072 * --> -dl task blocks on mutex A
3074 * 2. -dl task is running and holds mutex A
3075 * --> -dl task blocks on mutex A and could preempt the
3078 if (dl_prio(prio
)) {
3079 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3080 if (!dl_prio(p
->normal_prio
) ||
3081 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3082 p
->dl
.dl_boosted
= 1;
3083 p
->dl
.dl_throttled
= 0;
3084 enqueue_flag
= ENQUEUE_REPLENISH
;
3086 p
->dl
.dl_boosted
= 0;
3087 p
->sched_class
= &dl_sched_class
;
3088 } else if (rt_prio(prio
)) {
3089 if (dl_prio(oldprio
))
3090 p
->dl
.dl_boosted
= 0;
3092 enqueue_flag
= ENQUEUE_HEAD
;
3093 p
->sched_class
= &rt_sched_class
;
3095 if (dl_prio(oldprio
))
3096 p
->dl
.dl_boosted
= 0;
3097 p
->sched_class
= &fair_sched_class
;
3103 p
->sched_class
->set_curr_task(rq
);
3105 enqueue_task(rq
, p
, enqueue_flag
);
3107 check_class_changed(rq
, p
, prev_class
, oldprio
);
3109 __task_rq_unlock(rq
);
3113 void set_user_nice(struct task_struct
*p
, long nice
)
3115 int old_prio
, delta
, queued
;
3116 unsigned long flags
;
3119 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3122 * We have to be careful, if called from sys_setpriority(),
3123 * the task might be in the middle of scheduling on another CPU.
3125 rq
= task_rq_lock(p
, &flags
);
3127 * The RT priorities are set via sched_setscheduler(), but we still
3128 * allow the 'normal' nice value to be set - but as expected
3129 * it wont have any effect on scheduling until the task is
3130 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3132 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3133 p
->static_prio
= NICE_TO_PRIO(nice
);
3136 queued
= task_on_rq_queued(p
);
3138 dequeue_task(rq
, p
, 0);
3140 p
->static_prio
= NICE_TO_PRIO(nice
);
3143 p
->prio
= effective_prio(p
);
3144 delta
= p
->prio
- old_prio
;
3147 enqueue_task(rq
, p
, 0);
3149 * If the task increased its priority or is running and
3150 * lowered its priority, then reschedule its CPU:
3152 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3156 task_rq_unlock(rq
, p
, &flags
);
3158 EXPORT_SYMBOL(set_user_nice
);
3161 * can_nice - check if a task can reduce its nice value
3165 int can_nice(const struct task_struct
*p
, const int nice
)
3167 /* convert nice value [19,-20] to rlimit style value [1,40] */
3168 int nice_rlim
= nice_to_rlimit(nice
);
3170 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3171 capable(CAP_SYS_NICE
));
3174 #ifdef __ARCH_WANT_SYS_NICE
3177 * sys_nice - change the priority of the current process.
3178 * @increment: priority increment
3180 * sys_setpriority is a more generic, but much slower function that
3181 * does similar things.
3183 SYSCALL_DEFINE1(nice
, int, increment
)
3188 * Setpriority might change our priority at the same moment.
3189 * We don't have to worry. Conceptually one call occurs first
3190 * and we have a single winner.
3192 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3193 nice
= task_nice(current
) + increment
;
3195 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3196 if (increment
< 0 && !can_nice(current
, nice
))
3199 retval
= security_task_setnice(current
, nice
);
3203 set_user_nice(current
, nice
);
3210 * task_prio - return the priority value of a given task.
3211 * @p: the task in question.
3213 * Return: The priority value as seen by users in /proc.
3214 * RT tasks are offset by -200. Normal tasks are centered
3215 * around 0, value goes from -16 to +15.
3217 int task_prio(const struct task_struct
*p
)
3219 return p
->prio
- MAX_RT_PRIO
;
3223 * idle_cpu - is a given cpu idle currently?
3224 * @cpu: the processor in question.
3226 * Return: 1 if the CPU is currently idle. 0 otherwise.
3228 int idle_cpu(int cpu
)
3230 struct rq
*rq
= cpu_rq(cpu
);
3232 if (rq
->curr
!= rq
->idle
)
3239 if (!llist_empty(&rq
->wake_list
))
3247 * idle_task - return the idle task for a given cpu.
3248 * @cpu: the processor in question.
3250 * Return: The idle task for the cpu @cpu.
3252 struct task_struct
*idle_task(int cpu
)
3254 return cpu_rq(cpu
)->idle
;
3258 * find_process_by_pid - find a process with a matching PID value.
3259 * @pid: the pid in question.
3261 * The task of @pid, if found. %NULL otherwise.
3263 static struct task_struct
*find_process_by_pid(pid_t pid
)
3265 return pid
? find_task_by_vpid(pid
) : current
;
3269 * This function initializes the sched_dl_entity of a newly becoming
3270 * SCHED_DEADLINE task.
3272 * Only the static values are considered here, the actual runtime and the
3273 * absolute deadline will be properly calculated when the task is enqueued
3274 * for the first time with its new policy.
3277 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3279 struct sched_dl_entity
*dl_se
= &p
->dl
;
3281 dl_se
->dl_runtime
= attr
->sched_runtime
;
3282 dl_se
->dl_deadline
= attr
->sched_deadline
;
3283 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3284 dl_se
->flags
= attr
->sched_flags
;
3285 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3288 * Changing the parameters of a task is 'tricky' and we're not doing
3289 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3291 * What we SHOULD do is delay the bandwidth release until the 0-lag
3292 * point. This would include retaining the task_struct until that time
3293 * and change dl_overflow() to not immediately decrement the current
3296 * Instead we retain the current runtime/deadline and let the new
3297 * parameters take effect after the current reservation period lapses.
3298 * This is safe (albeit pessimistic) because the 0-lag point is always
3299 * before the current scheduling deadline.
3301 * We can still have temporary overloads because we do not delay the
3302 * change in bandwidth until that time; so admission control is
3303 * not on the safe side. It does however guarantee tasks will never
3304 * consume more than promised.
3309 * sched_setparam() passes in -1 for its policy, to let the functions
3310 * it calls know not to change it.
3312 #define SETPARAM_POLICY -1
3314 static void __setscheduler_params(struct task_struct
*p
,
3315 const struct sched_attr
*attr
)
3317 int policy
= attr
->sched_policy
;
3319 if (policy
== SETPARAM_POLICY
)
3324 if (dl_policy(policy
))
3325 __setparam_dl(p
, attr
);
3326 else if (fair_policy(policy
))
3327 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3330 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3331 * !rt_policy. Always setting this ensures that things like
3332 * getparam()/getattr() don't report silly values for !rt tasks.
3334 p
->rt_priority
= attr
->sched_priority
;
3335 p
->normal_prio
= normal_prio(p
);
3339 /* Actually do priority change: must hold pi & rq lock. */
3340 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3341 const struct sched_attr
*attr
)
3343 __setscheduler_params(p
, attr
);
3346 * If we get here, there was no pi waiters boosting the
3347 * task. It is safe to use the normal prio.
3349 p
->prio
= normal_prio(p
);
3351 if (dl_prio(p
->prio
))
3352 p
->sched_class
= &dl_sched_class
;
3353 else if (rt_prio(p
->prio
))
3354 p
->sched_class
= &rt_sched_class
;
3356 p
->sched_class
= &fair_sched_class
;
3360 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3362 struct sched_dl_entity
*dl_se
= &p
->dl
;
3364 attr
->sched_priority
= p
->rt_priority
;
3365 attr
->sched_runtime
= dl_se
->dl_runtime
;
3366 attr
->sched_deadline
= dl_se
->dl_deadline
;
3367 attr
->sched_period
= dl_se
->dl_period
;
3368 attr
->sched_flags
= dl_se
->flags
;
3372 * This function validates the new parameters of a -deadline task.
3373 * We ask for the deadline not being zero, and greater or equal
3374 * than the runtime, as well as the period of being zero or
3375 * greater than deadline. Furthermore, we have to be sure that
3376 * user parameters are above the internal resolution of 1us (we
3377 * check sched_runtime only since it is always the smaller one) and
3378 * below 2^63 ns (we have to check both sched_deadline and
3379 * sched_period, as the latter can be zero).
3382 __checkparam_dl(const struct sched_attr
*attr
)
3385 if (attr
->sched_deadline
== 0)
3389 * Since we truncate DL_SCALE bits, make sure we're at least
3392 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3396 * Since we use the MSB for wrap-around and sign issues, make
3397 * sure it's not set (mind that period can be equal to zero).
3399 if (attr
->sched_deadline
& (1ULL << 63) ||
3400 attr
->sched_period
& (1ULL << 63))
3403 /* runtime <= deadline <= period (if period != 0) */
3404 if ((attr
->sched_period
!= 0 &&
3405 attr
->sched_period
< attr
->sched_deadline
) ||
3406 attr
->sched_deadline
< attr
->sched_runtime
)
3413 * check the target process has a UID that matches the current process's
3415 static bool check_same_owner(struct task_struct
*p
)
3417 const struct cred
*cred
= current_cred(), *pcred
;
3421 pcred
= __task_cred(p
);
3422 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3423 uid_eq(cred
->euid
, pcred
->uid
));
3428 static bool dl_param_changed(struct task_struct
*p
,
3429 const struct sched_attr
*attr
)
3431 struct sched_dl_entity
*dl_se
= &p
->dl
;
3433 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3434 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3435 dl_se
->dl_period
!= attr
->sched_period
||
3436 dl_se
->flags
!= attr
->sched_flags
)
3442 static int __sched_setscheduler(struct task_struct
*p
,
3443 const struct sched_attr
*attr
,
3446 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3447 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3448 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3449 int policy
= attr
->sched_policy
;
3450 unsigned long flags
;
3451 const struct sched_class
*prev_class
;
3455 /* may grab non-irq protected spin_locks */
3456 BUG_ON(in_interrupt());
3458 /* double check policy once rq lock held */
3460 reset_on_fork
= p
->sched_reset_on_fork
;
3461 policy
= oldpolicy
= p
->policy
;
3463 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3465 if (policy
!= SCHED_DEADLINE
&&
3466 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3467 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3468 policy
!= SCHED_IDLE
)
3472 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3476 * Valid priorities for SCHED_FIFO and SCHED_RR are
3477 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3478 * SCHED_BATCH and SCHED_IDLE is 0.
3480 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3481 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3483 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3484 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3488 * Allow unprivileged RT tasks to decrease priority:
3490 if (user
&& !capable(CAP_SYS_NICE
)) {
3491 if (fair_policy(policy
)) {
3492 if (attr
->sched_nice
< task_nice(p
) &&
3493 !can_nice(p
, attr
->sched_nice
))
3497 if (rt_policy(policy
)) {
3498 unsigned long rlim_rtprio
=
3499 task_rlimit(p
, RLIMIT_RTPRIO
);
3501 /* can't set/change the rt policy */
3502 if (policy
!= p
->policy
&& !rlim_rtprio
)
3505 /* can't increase priority */
3506 if (attr
->sched_priority
> p
->rt_priority
&&
3507 attr
->sched_priority
> rlim_rtprio
)
3512 * Can't set/change SCHED_DEADLINE policy at all for now
3513 * (safest behavior); in the future we would like to allow
3514 * unprivileged DL tasks to increase their relative deadline
3515 * or reduce their runtime (both ways reducing utilization)
3517 if (dl_policy(policy
))
3521 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3522 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3524 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3525 if (!can_nice(p
, task_nice(p
)))
3529 /* can't change other user's priorities */
3530 if (!check_same_owner(p
))
3533 /* Normal users shall not reset the sched_reset_on_fork flag */
3534 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3539 retval
= security_task_setscheduler(p
);
3545 * make sure no PI-waiters arrive (or leave) while we are
3546 * changing the priority of the task:
3548 * To be able to change p->policy safely, the appropriate
3549 * runqueue lock must be held.
3551 rq
= task_rq_lock(p
, &flags
);
3554 * Changing the policy of the stop threads its a very bad idea
3556 if (p
== rq
->stop
) {
3557 task_rq_unlock(rq
, p
, &flags
);
3562 * If not changing anything there's no need to proceed further,
3563 * but store a possible modification of reset_on_fork.
3565 if (unlikely(policy
== p
->policy
)) {
3566 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3568 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3570 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3573 p
->sched_reset_on_fork
= reset_on_fork
;
3574 task_rq_unlock(rq
, p
, &flags
);
3580 #ifdef CONFIG_RT_GROUP_SCHED
3582 * Do not allow realtime tasks into groups that have no runtime
3585 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3586 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3587 !task_group_is_autogroup(task_group(p
))) {
3588 task_rq_unlock(rq
, p
, &flags
);
3593 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3594 cpumask_t
*span
= rq
->rd
->span
;
3597 * Don't allow tasks with an affinity mask smaller than
3598 * the entire root_domain to become SCHED_DEADLINE. We
3599 * will also fail if there's no bandwidth available.
3601 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3602 rq
->rd
->dl_bw
.bw
== 0) {
3603 task_rq_unlock(rq
, p
, &flags
);
3610 /* recheck policy now with rq lock held */
3611 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3612 policy
= oldpolicy
= -1;
3613 task_rq_unlock(rq
, p
, &flags
);
3618 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3619 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3622 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3623 task_rq_unlock(rq
, p
, &flags
);
3627 p
->sched_reset_on_fork
= reset_on_fork
;
3631 * Special case for priority boosted tasks.
3633 * If the new priority is lower or equal (user space view)
3634 * than the current (boosted) priority, we just store the new
3635 * normal parameters and do not touch the scheduler class and
3636 * the runqueue. This will be done when the task deboost
3639 if (rt_mutex_check_prio(p
, newprio
)) {
3640 __setscheduler_params(p
, attr
);
3641 task_rq_unlock(rq
, p
, &flags
);
3645 queued
= task_on_rq_queued(p
);
3646 running
= task_current(rq
, p
);
3648 dequeue_task(rq
, p
, 0);
3650 put_prev_task(rq
, p
);
3652 prev_class
= p
->sched_class
;
3653 __setscheduler(rq
, p
, attr
);
3656 p
->sched_class
->set_curr_task(rq
);
3659 * We enqueue to tail when the priority of a task is
3660 * increased (user space view).
3662 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3665 check_class_changed(rq
, p
, prev_class
, oldprio
);
3666 task_rq_unlock(rq
, p
, &flags
);
3668 rt_mutex_adjust_pi(p
);
3673 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3674 const struct sched_param
*param
, bool check
)
3676 struct sched_attr attr
= {
3677 .sched_policy
= policy
,
3678 .sched_priority
= param
->sched_priority
,
3679 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3682 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3683 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3684 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3685 policy
&= ~SCHED_RESET_ON_FORK
;
3686 attr
.sched_policy
= policy
;
3689 return __sched_setscheduler(p
, &attr
, check
);
3692 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3693 * @p: the task in question.
3694 * @policy: new policy.
3695 * @param: structure containing the new RT priority.
3697 * Return: 0 on success. An error code otherwise.
3699 * NOTE that the task may be already dead.
3701 int sched_setscheduler(struct task_struct
*p
, int policy
,
3702 const struct sched_param
*param
)
3704 return _sched_setscheduler(p
, policy
, param
, true);
3706 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3708 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3710 return __sched_setscheduler(p
, attr
, true);
3712 EXPORT_SYMBOL_GPL(sched_setattr
);
3715 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3716 * @p: the task in question.
3717 * @policy: new policy.
3718 * @param: structure containing the new RT priority.
3720 * Just like sched_setscheduler, only don't bother checking if the
3721 * current context has permission. For example, this is needed in
3722 * stop_machine(): we create temporary high priority worker threads,
3723 * but our caller might not have that capability.
3725 * Return: 0 on success. An error code otherwise.
3727 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3728 const struct sched_param
*param
)
3730 return _sched_setscheduler(p
, policy
, param
, false);
3734 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3736 struct sched_param lparam
;
3737 struct task_struct
*p
;
3740 if (!param
|| pid
< 0)
3742 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3747 p
= find_process_by_pid(pid
);
3749 retval
= sched_setscheduler(p
, policy
, &lparam
);
3756 * Mimics kernel/events/core.c perf_copy_attr().
3758 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3759 struct sched_attr
*attr
)
3764 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3768 * zero the full structure, so that a short copy will be nice.
3770 memset(attr
, 0, sizeof(*attr
));
3772 ret
= get_user(size
, &uattr
->size
);
3776 if (size
> PAGE_SIZE
) /* silly large */
3779 if (!size
) /* abi compat */
3780 size
= SCHED_ATTR_SIZE_VER0
;
3782 if (size
< SCHED_ATTR_SIZE_VER0
)
3786 * If we're handed a bigger struct than we know of,
3787 * ensure all the unknown bits are 0 - i.e. new
3788 * user-space does not rely on any kernel feature
3789 * extensions we dont know about yet.
3791 if (size
> sizeof(*attr
)) {
3792 unsigned char __user
*addr
;
3793 unsigned char __user
*end
;
3796 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3797 end
= (void __user
*)uattr
+ size
;
3799 for (; addr
< end
; addr
++) {
3800 ret
= get_user(val
, addr
);
3806 size
= sizeof(*attr
);
3809 ret
= copy_from_user(attr
, uattr
, size
);
3814 * XXX: do we want to be lenient like existing syscalls; or do we want
3815 * to be strict and return an error on out-of-bounds values?
3817 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3822 put_user(sizeof(*attr
), &uattr
->size
);
3827 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3828 * @pid: the pid in question.
3829 * @policy: new policy.
3830 * @param: structure containing the new RT priority.
3832 * Return: 0 on success. An error code otherwise.
3834 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3835 struct sched_param __user
*, param
)
3837 /* negative values for policy are not valid */
3841 return do_sched_setscheduler(pid
, policy
, param
);
3845 * sys_sched_setparam - set/change the RT priority of a thread
3846 * @pid: the pid in question.
3847 * @param: structure containing the new RT priority.
3849 * Return: 0 on success. An error code otherwise.
3851 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3853 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
3857 * sys_sched_setattr - same as above, but with extended sched_attr
3858 * @pid: the pid in question.
3859 * @uattr: structure containing the extended parameters.
3860 * @flags: for future extension.
3862 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3863 unsigned int, flags
)
3865 struct sched_attr attr
;
3866 struct task_struct
*p
;
3869 if (!uattr
|| pid
< 0 || flags
)
3872 retval
= sched_copy_attr(uattr
, &attr
);
3876 if ((int)attr
.sched_policy
< 0)
3881 p
= find_process_by_pid(pid
);
3883 retval
= sched_setattr(p
, &attr
);
3890 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3891 * @pid: the pid in question.
3893 * Return: On success, the policy of the thread. Otherwise, a negative error
3896 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3898 struct task_struct
*p
;
3906 p
= find_process_by_pid(pid
);
3908 retval
= security_task_getscheduler(p
);
3911 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3918 * sys_sched_getparam - get the RT priority of a thread
3919 * @pid: the pid in question.
3920 * @param: structure containing the RT priority.
3922 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3925 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3927 struct sched_param lp
= { .sched_priority
= 0 };
3928 struct task_struct
*p
;
3931 if (!param
|| pid
< 0)
3935 p
= find_process_by_pid(pid
);
3940 retval
= security_task_getscheduler(p
);
3944 if (task_has_rt_policy(p
))
3945 lp
.sched_priority
= p
->rt_priority
;
3949 * This one might sleep, we cannot do it with a spinlock held ...
3951 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3960 static int sched_read_attr(struct sched_attr __user
*uattr
,
3961 struct sched_attr
*attr
,
3966 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3970 * If we're handed a smaller struct than we know of,
3971 * ensure all the unknown bits are 0 - i.e. old
3972 * user-space does not get uncomplete information.
3974 if (usize
< sizeof(*attr
)) {
3975 unsigned char *addr
;
3978 addr
= (void *)attr
+ usize
;
3979 end
= (void *)attr
+ sizeof(*attr
);
3981 for (; addr
< end
; addr
++) {
3989 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3997 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3998 * @pid: the pid in question.
3999 * @uattr: structure containing the extended parameters.
4000 * @size: sizeof(attr) for fwd/bwd comp.
4001 * @flags: for future extension.
4003 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4004 unsigned int, size
, unsigned int, flags
)
4006 struct sched_attr attr
= {
4007 .size
= sizeof(struct sched_attr
),
4009 struct task_struct
*p
;
4012 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4013 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4017 p
= find_process_by_pid(pid
);
4022 retval
= security_task_getscheduler(p
);
4026 attr
.sched_policy
= p
->policy
;
4027 if (p
->sched_reset_on_fork
)
4028 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4029 if (task_has_dl_policy(p
))
4030 __getparam_dl(p
, &attr
);
4031 else if (task_has_rt_policy(p
))
4032 attr
.sched_priority
= p
->rt_priority
;
4034 attr
.sched_nice
= task_nice(p
);
4038 retval
= sched_read_attr(uattr
, &attr
, size
);
4046 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4048 cpumask_var_t cpus_allowed
, new_mask
;
4049 struct task_struct
*p
;
4054 p
= find_process_by_pid(pid
);
4060 /* Prevent p going away */
4064 if (p
->flags
& PF_NO_SETAFFINITY
) {
4068 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4072 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4074 goto out_free_cpus_allowed
;
4077 if (!check_same_owner(p
)) {
4079 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4081 goto out_free_new_mask
;
4086 retval
= security_task_setscheduler(p
);
4088 goto out_free_new_mask
;
4091 cpuset_cpus_allowed(p
, cpus_allowed
);
4092 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4095 * Since bandwidth control happens on root_domain basis,
4096 * if admission test is enabled, we only admit -deadline
4097 * tasks allowed to run on all the CPUs in the task's
4101 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4103 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4106 goto out_free_new_mask
;
4112 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4115 cpuset_cpus_allowed(p
, cpus_allowed
);
4116 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4118 * We must have raced with a concurrent cpuset
4119 * update. Just reset the cpus_allowed to the
4120 * cpuset's cpus_allowed
4122 cpumask_copy(new_mask
, cpus_allowed
);
4127 free_cpumask_var(new_mask
);
4128 out_free_cpus_allowed
:
4129 free_cpumask_var(cpus_allowed
);
4135 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4136 struct cpumask
*new_mask
)
4138 if (len
< cpumask_size())
4139 cpumask_clear(new_mask
);
4140 else if (len
> cpumask_size())
4141 len
= cpumask_size();
4143 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4147 * sys_sched_setaffinity - set the cpu affinity of a process
4148 * @pid: pid of the process
4149 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4150 * @user_mask_ptr: user-space pointer to the new cpu mask
4152 * Return: 0 on success. An error code otherwise.
4154 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4155 unsigned long __user
*, user_mask_ptr
)
4157 cpumask_var_t new_mask
;
4160 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4163 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4165 retval
= sched_setaffinity(pid
, new_mask
);
4166 free_cpumask_var(new_mask
);
4170 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4172 struct task_struct
*p
;
4173 unsigned long flags
;
4179 p
= find_process_by_pid(pid
);
4183 retval
= security_task_getscheduler(p
);
4187 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4188 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4189 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4198 * sys_sched_getaffinity - get the cpu affinity of a process
4199 * @pid: pid of the process
4200 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4201 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4203 * Return: 0 on success. An error code otherwise.
4205 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4206 unsigned long __user
*, user_mask_ptr
)
4211 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4213 if (len
& (sizeof(unsigned long)-1))
4216 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4219 ret
= sched_getaffinity(pid
, mask
);
4221 size_t retlen
= min_t(size_t, len
, cpumask_size());
4223 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4228 free_cpumask_var(mask
);
4234 * sys_sched_yield - yield the current processor to other threads.
4236 * This function yields the current CPU to other tasks. If there are no
4237 * other threads running on this CPU then this function will return.
4241 SYSCALL_DEFINE0(sched_yield
)
4243 struct rq
*rq
= this_rq_lock();
4245 schedstat_inc(rq
, yld_count
);
4246 current
->sched_class
->yield_task(rq
);
4249 * Since we are going to call schedule() anyway, there's
4250 * no need to preempt or enable interrupts:
4252 __release(rq
->lock
);
4253 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4254 do_raw_spin_unlock(&rq
->lock
);
4255 sched_preempt_enable_no_resched();
4262 int __sched
_cond_resched(void)
4264 if (should_resched()) {
4265 preempt_schedule_common();
4270 EXPORT_SYMBOL(_cond_resched
);
4273 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4274 * call schedule, and on return reacquire the lock.
4276 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4277 * operations here to prevent schedule() from being called twice (once via
4278 * spin_unlock(), once by hand).
4280 int __cond_resched_lock(spinlock_t
*lock
)
4282 int resched
= should_resched();
4285 lockdep_assert_held(lock
);
4287 if (spin_needbreak(lock
) || resched
) {
4290 preempt_schedule_common();
4298 EXPORT_SYMBOL(__cond_resched_lock
);
4300 int __sched
__cond_resched_softirq(void)
4302 BUG_ON(!in_softirq());
4304 if (should_resched()) {
4306 preempt_schedule_common();
4312 EXPORT_SYMBOL(__cond_resched_softirq
);
4315 * yield - yield the current processor to other threads.
4317 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4319 * The scheduler is at all times free to pick the calling task as the most
4320 * eligible task to run, if removing the yield() call from your code breaks
4321 * it, its already broken.
4323 * Typical broken usage is:
4328 * where one assumes that yield() will let 'the other' process run that will
4329 * make event true. If the current task is a SCHED_FIFO task that will never
4330 * happen. Never use yield() as a progress guarantee!!
4332 * If you want to use yield() to wait for something, use wait_event().
4333 * If you want to use yield() to be 'nice' for others, use cond_resched().
4334 * If you still want to use yield(), do not!
4336 void __sched
yield(void)
4338 set_current_state(TASK_RUNNING
);
4341 EXPORT_SYMBOL(yield
);
4344 * yield_to - yield the current processor to another thread in
4345 * your thread group, or accelerate that thread toward the
4346 * processor it's on.
4348 * @preempt: whether task preemption is allowed or not
4350 * It's the caller's job to ensure that the target task struct
4351 * can't go away on us before we can do any checks.
4354 * true (>0) if we indeed boosted the target task.
4355 * false (0) if we failed to boost the target.
4356 * -ESRCH if there's no task to yield to.
4358 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4360 struct task_struct
*curr
= current
;
4361 struct rq
*rq
, *p_rq
;
4362 unsigned long flags
;
4365 local_irq_save(flags
);
4371 * If we're the only runnable task on the rq and target rq also
4372 * has only one task, there's absolutely no point in yielding.
4374 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4379 double_rq_lock(rq
, p_rq
);
4380 if (task_rq(p
) != p_rq
) {
4381 double_rq_unlock(rq
, p_rq
);
4385 if (!curr
->sched_class
->yield_to_task
)
4388 if (curr
->sched_class
!= p
->sched_class
)
4391 if (task_running(p_rq
, p
) || p
->state
)
4394 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4396 schedstat_inc(rq
, yld_count
);
4398 * Make p's CPU reschedule; pick_next_entity takes care of
4401 if (preempt
&& rq
!= p_rq
)
4406 double_rq_unlock(rq
, p_rq
);
4408 local_irq_restore(flags
);
4415 EXPORT_SYMBOL_GPL(yield_to
);
4418 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4419 * that process accounting knows that this is a task in IO wait state.
4421 void __sched
io_schedule(void)
4423 struct rq
*rq
= raw_rq();
4425 delayacct_blkio_start();
4426 atomic_inc(&rq
->nr_iowait
);
4427 blk_flush_plug(current
);
4428 current
->in_iowait
= 1;
4430 current
->in_iowait
= 0;
4431 atomic_dec(&rq
->nr_iowait
);
4432 delayacct_blkio_end();
4434 EXPORT_SYMBOL(io_schedule
);
4436 long __sched
io_schedule_timeout(long timeout
)
4438 struct rq
*rq
= raw_rq();
4441 delayacct_blkio_start();
4442 atomic_inc(&rq
->nr_iowait
);
4443 blk_flush_plug(current
);
4444 current
->in_iowait
= 1;
4445 ret
= schedule_timeout(timeout
);
4446 current
->in_iowait
= 0;
4447 atomic_dec(&rq
->nr_iowait
);
4448 delayacct_blkio_end();
4453 * sys_sched_get_priority_max - return maximum RT priority.
4454 * @policy: scheduling class.
4456 * Return: On success, this syscall returns the maximum
4457 * rt_priority that can be used by a given scheduling class.
4458 * On failure, a negative error code is returned.
4460 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4467 ret
= MAX_USER_RT_PRIO
-1;
4469 case SCHED_DEADLINE
:
4480 * sys_sched_get_priority_min - return minimum RT priority.
4481 * @policy: scheduling class.
4483 * Return: On success, this syscall returns the minimum
4484 * rt_priority that can be used by a given scheduling class.
4485 * On failure, a negative error code is returned.
4487 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4496 case SCHED_DEADLINE
:
4506 * sys_sched_rr_get_interval - return the default timeslice of a process.
4507 * @pid: pid of the process.
4508 * @interval: userspace pointer to the timeslice value.
4510 * this syscall writes the default timeslice value of a given process
4511 * into the user-space timespec buffer. A value of '0' means infinity.
4513 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4516 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4517 struct timespec __user
*, interval
)
4519 struct task_struct
*p
;
4520 unsigned int time_slice
;
4521 unsigned long flags
;
4531 p
= find_process_by_pid(pid
);
4535 retval
= security_task_getscheduler(p
);
4539 rq
= task_rq_lock(p
, &flags
);
4541 if (p
->sched_class
->get_rr_interval
)
4542 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4543 task_rq_unlock(rq
, p
, &flags
);
4546 jiffies_to_timespec(time_slice
, &t
);
4547 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4555 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4557 void sched_show_task(struct task_struct
*p
)
4559 unsigned long free
= 0;
4561 unsigned long state
= p
->state
;
4564 state
= __ffs(state
) + 1;
4565 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4566 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4567 #if BITS_PER_LONG == 32
4568 if (state
== TASK_RUNNING
)
4569 printk(KERN_CONT
" running ");
4571 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4573 if (state
== TASK_RUNNING
)
4574 printk(KERN_CONT
" running task ");
4576 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4578 #ifdef CONFIG_DEBUG_STACK_USAGE
4579 free
= stack_not_used(p
);
4584 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4586 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4587 task_pid_nr(p
), ppid
,
4588 (unsigned long)task_thread_info(p
)->flags
);
4590 print_worker_info(KERN_INFO
, p
);
4591 show_stack(p
, NULL
);
4594 void show_state_filter(unsigned long state_filter
)
4596 struct task_struct
*g
, *p
;
4598 #if BITS_PER_LONG == 32
4600 " task PC stack pid father\n");
4603 " task PC stack pid father\n");
4606 for_each_process_thread(g
, p
) {
4608 * reset the NMI-timeout, listing all files on a slow
4609 * console might take a lot of time:
4611 touch_nmi_watchdog();
4612 if (!state_filter
|| (p
->state
& state_filter
))
4616 touch_all_softlockup_watchdogs();
4618 #ifdef CONFIG_SCHED_DEBUG
4619 sysrq_sched_debug_show();
4623 * Only show locks if all tasks are dumped:
4626 debug_show_all_locks();
4629 void init_idle_bootup_task(struct task_struct
*idle
)
4631 idle
->sched_class
= &idle_sched_class
;
4635 * init_idle - set up an idle thread for a given CPU
4636 * @idle: task in question
4637 * @cpu: cpu the idle task belongs to
4639 * NOTE: this function does not set the idle thread's NEED_RESCHED
4640 * flag, to make booting more robust.
4642 void init_idle(struct task_struct
*idle
, int cpu
)
4644 struct rq
*rq
= cpu_rq(cpu
);
4645 unsigned long flags
;
4647 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4649 __sched_fork(0, idle
);
4650 idle
->state
= TASK_RUNNING
;
4651 idle
->se
.exec_start
= sched_clock();
4653 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4655 * We're having a chicken and egg problem, even though we are
4656 * holding rq->lock, the cpu isn't yet set to this cpu so the
4657 * lockdep check in task_group() will fail.
4659 * Similar case to sched_fork(). / Alternatively we could
4660 * use task_rq_lock() here and obtain the other rq->lock.
4665 __set_task_cpu(idle
, cpu
);
4668 rq
->curr
= rq
->idle
= idle
;
4669 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4670 #if defined(CONFIG_SMP)
4673 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4675 /* Set the preempt count _outside_ the spinlocks! */
4676 init_idle_preempt_count(idle
, cpu
);
4679 * The idle tasks have their own, simple scheduling class:
4681 idle
->sched_class
= &idle_sched_class
;
4682 ftrace_graph_init_idle_task(idle
, cpu
);
4683 vtime_init_idle(idle
, cpu
);
4684 #if defined(CONFIG_SMP)
4685 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4689 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4690 const struct cpumask
*trial
)
4692 int ret
= 1, trial_cpus
;
4693 struct dl_bw
*cur_dl_b
;
4694 unsigned long flags
;
4696 if (!cpumask_weight(cur
))
4699 rcu_read_lock_sched();
4700 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4701 trial_cpus
= cpumask_weight(trial
);
4703 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4704 if (cur_dl_b
->bw
!= -1 &&
4705 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
4707 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
4708 rcu_read_unlock_sched();
4713 int task_can_attach(struct task_struct
*p
,
4714 const struct cpumask
*cs_cpus_allowed
)
4719 * Kthreads which disallow setaffinity shouldn't be moved
4720 * to a new cpuset; we don't want to change their cpu
4721 * affinity and isolating such threads by their set of
4722 * allowed nodes is unnecessary. Thus, cpusets are not
4723 * applicable for such threads. This prevents checking for
4724 * success of set_cpus_allowed_ptr() on all attached tasks
4725 * before cpus_allowed may be changed.
4727 if (p
->flags
& PF_NO_SETAFFINITY
) {
4733 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
4735 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
4740 unsigned long flags
;
4742 rcu_read_lock_sched();
4743 dl_b
= dl_bw_of(dest_cpu
);
4744 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
4745 cpus
= dl_bw_cpus(dest_cpu
);
4746 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
4751 * We reserve space for this task in the destination
4752 * root_domain, as we can't fail after this point.
4753 * We will free resources in the source root_domain
4754 * later on (see set_cpus_allowed_dl()).
4756 __dl_add(dl_b
, p
->dl
.dl_bw
);
4758 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
4759 rcu_read_unlock_sched();
4769 * move_queued_task - move a queued task to new rq.
4771 * Returns (locked) new rq. Old rq's lock is released.
4773 static struct rq
*move_queued_task(struct task_struct
*p
, int new_cpu
)
4775 struct rq
*rq
= task_rq(p
);
4777 lockdep_assert_held(&rq
->lock
);
4779 dequeue_task(rq
, p
, 0);
4780 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
4781 set_task_cpu(p
, new_cpu
);
4782 raw_spin_unlock(&rq
->lock
);
4784 rq
= cpu_rq(new_cpu
);
4786 raw_spin_lock(&rq
->lock
);
4787 BUG_ON(task_cpu(p
) != new_cpu
);
4788 p
->on_rq
= TASK_ON_RQ_QUEUED
;
4789 enqueue_task(rq
, p
, 0);
4790 check_preempt_curr(rq
, p
, 0);
4795 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4797 if (p
->sched_class
->set_cpus_allowed
)
4798 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4800 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4801 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4805 * This is how migration works:
4807 * 1) we invoke migration_cpu_stop() on the target CPU using
4809 * 2) stopper starts to run (implicitly forcing the migrated thread
4811 * 3) it checks whether the migrated task is still in the wrong runqueue.
4812 * 4) if it's in the wrong runqueue then the migration thread removes
4813 * it and puts it into the right queue.
4814 * 5) stopper completes and stop_one_cpu() returns and the migration
4819 * Change a given task's CPU affinity. Migrate the thread to a
4820 * proper CPU and schedule it away if the CPU it's executing on
4821 * is removed from the allowed bitmask.
4823 * NOTE: the caller must have a valid reference to the task, the
4824 * task must not exit() & deallocate itself prematurely. The
4825 * call is not atomic; no spinlocks may be held.
4827 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4829 unsigned long flags
;
4831 unsigned int dest_cpu
;
4834 rq
= task_rq_lock(p
, &flags
);
4836 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4839 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4844 do_set_cpus_allowed(p
, new_mask
);
4846 /* Can the task run on the task's current CPU? If so, we're done */
4847 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4850 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4851 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
4852 struct migration_arg arg
= { p
, dest_cpu
};
4853 /* Need help from migration thread: drop lock and wait. */
4854 task_rq_unlock(rq
, p
, &flags
);
4855 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4856 tlb_migrate_finish(p
->mm
);
4858 } else if (task_on_rq_queued(p
))
4859 rq
= move_queued_task(p
, dest_cpu
);
4861 task_rq_unlock(rq
, p
, &flags
);
4865 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4868 * Move (not current) task off this cpu, onto dest cpu. We're doing
4869 * this because either it can't run here any more (set_cpus_allowed()
4870 * away from this CPU, or CPU going down), or because we're
4871 * attempting to rebalance this task on exec (sched_exec).
4873 * So we race with normal scheduler movements, but that's OK, as long
4874 * as the task is no longer on this CPU.
4876 * Returns non-zero if task was successfully migrated.
4878 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4883 if (unlikely(!cpu_active(dest_cpu
)))
4886 rq
= cpu_rq(src_cpu
);
4888 raw_spin_lock(&p
->pi_lock
);
4889 raw_spin_lock(&rq
->lock
);
4890 /* Already moved. */
4891 if (task_cpu(p
) != src_cpu
)
4894 /* Affinity changed (again). */
4895 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4899 * If we're not on a rq, the next wake-up will ensure we're
4902 if (task_on_rq_queued(p
))
4903 rq
= move_queued_task(p
, dest_cpu
);
4907 raw_spin_unlock(&rq
->lock
);
4908 raw_spin_unlock(&p
->pi_lock
);
4912 #ifdef CONFIG_NUMA_BALANCING
4913 /* Migrate current task p to target_cpu */
4914 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4916 struct migration_arg arg
= { p
, target_cpu
};
4917 int curr_cpu
= task_cpu(p
);
4919 if (curr_cpu
== target_cpu
)
4922 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4925 /* TODO: This is not properly updating schedstats */
4927 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4928 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4932 * Requeue a task on a given node and accurately track the number of NUMA
4933 * tasks on the runqueues
4935 void sched_setnuma(struct task_struct
*p
, int nid
)
4938 unsigned long flags
;
4939 bool queued
, running
;
4941 rq
= task_rq_lock(p
, &flags
);
4942 queued
= task_on_rq_queued(p
);
4943 running
= task_current(rq
, p
);
4946 dequeue_task(rq
, p
, 0);
4948 put_prev_task(rq
, p
);
4950 p
->numa_preferred_nid
= nid
;
4953 p
->sched_class
->set_curr_task(rq
);
4955 enqueue_task(rq
, p
, 0);
4956 task_rq_unlock(rq
, p
, &flags
);
4961 * migration_cpu_stop - this will be executed by a highprio stopper thread
4962 * and performs thread migration by bumping thread off CPU then
4963 * 'pushing' onto another runqueue.
4965 static int migration_cpu_stop(void *data
)
4967 struct migration_arg
*arg
= data
;
4970 * The original target cpu might have gone down and we might
4971 * be on another cpu but it doesn't matter.
4973 local_irq_disable();
4975 * We need to explicitly wake pending tasks before running
4976 * __migrate_task() such that we will not miss enforcing cpus_allowed
4977 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4979 sched_ttwu_pending();
4980 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4985 #ifdef CONFIG_HOTPLUG_CPU
4988 * Ensures that the idle task is using init_mm right before its cpu goes
4991 void idle_task_exit(void)
4993 struct mm_struct
*mm
= current
->active_mm
;
4995 BUG_ON(cpu_online(smp_processor_id()));
4997 if (mm
!= &init_mm
) {
4998 switch_mm(mm
, &init_mm
, current
);
4999 finish_arch_post_lock_switch();
5005 * Since this CPU is going 'away' for a while, fold any nr_active delta
5006 * we might have. Assumes we're called after migrate_tasks() so that the
5007 * nr_active count is stable.
5009 * Also see the comment "Global load-average calculations".
5011 static void calc_load_migrate(struct rq
*rq
)
5013 long delta
= calc_load_fold_active(rq
);
5015 atomic_long_add(delta
, &calc_load_tasks
);
5018 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5022 static const struct sched_class fake_sched_class
= {
5023 .put_prev_task
= put_prev_task_fake
,
5026 static struct task_struct fake_task
= {
5028 * Avoid pull_{rt,dl}_task()
5030 .prio
= MAX_PRIO
+ 1,
5031 .sched_class
= &fake_sched_class
,
5035 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5036 * try_to_wake_up()->select_task_rq().
5038 * Called with rq->lock held even though we'er in stop_machine() and
5039 * there's no concurrency possible, we hold the required locks anyway
5040 * because of lock validation efforts.
5042 static void migrate_tasks(unsigned int dead_cpu
)
5044 struct rq
*rq
= cpu_rq(dead_cpu
);
5045 struct task_struct
*next
, *stop
= rq
->stop
;
5049 * Fudge the rq selection such that the below task selection loop
5050 * doesn't get stuck on the currently eligible stop task.
5052 * We're currently inside stop_machine() and the rq is either stuck
5053 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5054 * either way we should never end up calling schedule() until we're
5060 * put_prev_task() and pick_next_task() sched
5061 * class method both need to have an up-to-date
5062 * value of rq->clock[_task]
5064 update_rq_clock(rq
);
5068 * There's this thread running, bail when that's the only
5071 if (rq
->nr_running
== 1)
5074 next
= pick_next_task(rq
, &fake_task
);
5076 next
->sched_class
->put_prev_task(rq
, next
);
5078 /* Find suitable destination for @next, with force if needed. */
5079 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5080 raw_spin_unlock(&rq
->lock
);
5082 __migrate_task(next
, dead_cpu
, dest_cpu
);
5084 raw_spin_lock(&rq
->lock
);
5090 #endif /* CONFIG_HOTPLUG_CPU */
5092 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5094 static struct ctl_table sd_ctl_dir
[] = {
5096 .procname
= "sched_domain",
5102 static struct ctl_table sd_ctl_root
[] = {
5104 .procname
= "kernel",
5106 .child
= sd_ctl_dir
,
5111 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5113 struct ctl_table
*entry
=
5114 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5119 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5121 struct ctl_table
*entry
;
5124 * In the intermediate directories, both the child directory and
5125 * procname are dynamically allocated and could fail but the mode
5126 * will always be set. In the lowest directory the names are
5127 * static strings and all have proc handlers.
5129 for (entry
= *tablep
; entry
->mode
; entry
++) {
5131 sd_free_ctl_entry(&entry
->child
);
5132 if (entry
->proc_handler
== NULL
)
5133 kfree(entry
->procname
);
5140 static int min_load_idx
= 0;
5141 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5144 set_table_entry(struct ctl_table
*entry
,
5145 const char *procname
, void *data
, int maxlen
,
5146 umode_t mode
, proc_handler
*proc_handler
,
5149 entry
->procname
= procname
;
5151 entry
->maxlen
= maxlen
;
5153 entry
->proc_handler
= proc_handler
;
5156 entry
->extra1
= &min_load_idx
;
5157 entry
->extra2
= &max_load_idx
;
5161 static struct ctl_table
*
5162 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5164 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5169 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5170 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5171 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5172 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5173 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5174 sizeof(int), 0644, proc_dointvec_minmax
, true);
5175 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5176 sizeof(int), 0644, proc_dointvec_minmax
, true);
5177 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5178 sizeof(int), 0644, proc_dointvec_minmax
, true);
5179 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5180 sizeof(int), 0644, proc_dointvec_minmax
, true);
5181 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5182 sizeof(int), 0644, proc_dointvec_minmax
, true);
5183 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5184 sizeof(int), 0644, proc_dointvec_minmax
, false);
5185 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5186 sizeof(int), 0644, proc_dointvec_minmax
, false);
5187 set_table_entry(&table
[9], "cache_nice_tries",
5188 &sd
->cache_nice_tries
,
5189 sizeof(int), 0644, proc_dointvec_minmax
, false);
5190 set_table_entry(&table
[10], "flags", &sd
->flags
,
5191 sizeof(int), 0644, proc_dointvec_minmax
, false);
5192 set_table_entry(&table
[11], "max_newidle_lb_cost",
5193 &sd
->max_newidle_lb_cost
,
5194 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5195 set_table_entry(&table
[12], "name", sd
->name
,
5196 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5197 /* &table[13] is terminator */
5202 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5204 struct ctl_table
*entry
, *table
;
5205 struct sched_domain
*sd
;
5206 int domain_num
= 0, i
;
5209 for_each_domain(cpu
, sd
)
5211 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5216 for_each_domain(cpu
, sd
) {
5217 snprintf(buf
, 32, "domain%d", i
);
5218 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5220 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5227 static struct ctl_table_header
*sd_sysctl_header
;
5228 static void register_sched_domain_sysctl(void)
5230 int i
, cpu_num
= num_possible_cpus();
5231 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5234 WARN_ON(sd_ctl_dir
[0].child
);
5235 sd_ctl_dir
[0].child
= entry
;
5240 for_each_possible_cpu(i
) {
5241 snprintf(buf
, 32, "cpu%d", i
);
5242 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5244 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5248 WARN_ON(sd_sysctl_header
);
5249 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5252 /* may be called multiple times per register */
5253 static void unregister_sched_domain_sysctl(void)
5255 if (sd_sysctl_header
)
5256 unregister_sysctl_table(sd_sysctl_header
);
5257 sd_sysctl_header
= NULL
;
5258 if (sd_ctl_dir
[0].child
)
5259 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5262 static void register_sched_domain_sysctl(void)
5265 static void unregister_sched_domain_sysctl(void)
5270 static void set_rq_online(struct rq
*rq
)
5273 const struct sched_class
*class;
5275 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5278 for_each_class(class) {
5279 if (class->rq_online
)
5280 class->rq_online(rq
);
5285 static void set_rq_offline(struct rq
*rq
)
5288 const struct sched_class
*class;
5290 for_each_class(class) {
5291 if (class->rq_offline
)
5292 class->rq_offline(rq
);
5295 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5301 * migration_call - callback that gets triggered when a CPU is added.
5302 * Here we can start up the necessary migration thread for the new CPU.
5305 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5307 int cpu
= (long)hcpu
;
5308 unsigned long flags
;
5309 struct rq
*rq
= cpu_rq(cpu
);
5311 switch (action
& ~CPU_TASKS_FROZEN
) {
5313 case CPU_UP_PREPARE
:
5314 rq
->calc_load_update
= calc_load_update
;
5318 /* Update our root-domain */
5319 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5321 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5325 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5328 #ifdef CONFIG_HOTPLUG_CPU
5330 sched_ttwu_pending();
5331 /* Update our root-domain */
5332 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5334 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5338 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5339 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5343 calc_load_migrate(rq
);
5348 update_max_interval();
5354 * Register at high priority so that task migration (migrate_all_tasks)
5355 * happens before everything else. This has to be lower priority than
5356 * the notifier in the perf_event subsystem, though.
5358 static struct notifier_block migration_notifier
= {
5359 .notifier_call
= migration_call
,
5360 .priority
= CPU_PRI_MIGRATION
,
5363 static void __cpuinit
set_cpu_rq_start_time(void)
5365 int cpu
= smp_processor_id();
5366 struct rq
*rq
= cpu_rq(cpu
);
5367 rq
->age_stamp
= sched_clock_cpu(cpu
);
5370 static int sched_cpu_active(struct notifier_block
*nfb
,
5371 unsigned long action
, void *hcpu
)
5373 switch (action
& ~CPU_TASKS_FROZEN
) {
5375 set_cpu_rq_start_time();
5377 case CPU_DOWN_FAILED
:
5378 set_cpu_active((long)hcpu
, true);
5385 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5386 unsigned long action
, void *hcpu
)
5388 unsigned long flags
;
5389 long cpu
= (long)hcpu
;
5392 switch (action
& ~CPU_TASKS_FROZEN
) {
5393 case CPU_DOWN_PREPARE
:
5394 set_cpu_active(cpu
, false);
5396 /* explicitly allow suspend */
5397 if (!(action
& CPU_TASKS_FROZEN
)) {
5401 rcu_read_lock_sched();
5402 dl_b
= dl_bw_of(cpu
);
5404 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5405 cpus
= dl_bw_cpus(cpu
);
5406 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5407 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5409 rcu_read_unlock_sched();
5412 return notifier_from_errno(-EBUSY
);
5420 static int __init
migration_init(void)
5422 void *cpu
= (void *)(long)smp_processor_id();
5425 /* Initialize migration for the boot CPU */
5426 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5427 BUG_ON(err
== NOTIFY_BAD
);
5428 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5429 register_cpu_notifier(&migration_notifier
);
5431 /* Register cpu active notifiers */
5432 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5433 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5437 early_initcall(migration_init
);
5442 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5444 #ifdef CONFIG_SCHED_DEBUG
5446 static __read_mostly
int sched_debug_enabled
;
5448 static int __init
sched_debug_setup(char *str
)
5450 sched_debug_enabled
= 1;
5454 early_param("sched_debug", sched_debug_setup
);
5456 static inline bool sched_debug(void)
5458 return sched_debug_enabled
;
5461 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5462 struct cpumask
*groupmask
)
5464 struct sched_group
*group
= sd
->groups
;
5466 cpumask_clear(groupmask
);
5468 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5470 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5471 printk("does not load-balance\n");
5473 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5478 printk(KERN_CONT
"span %*pbl level %s\n",
5479 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5481 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5482 printk(KERN_ERR
"ERROR: domain->span does not contain "
5485 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5486 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5490 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5494 printk(KERN_ERR
"ERROR: group is NULL\n");
5499 * Even though we initialize ->capacity to something semi-sane,
5500 * we leave capacity_orig unset. This allows us to detect if
5501 * domain iteration is still funny without causing /0 traps.
5503 if (!group
->sgc
->capacity_orig
) {
5504 printk(KERN_CONT
"\n");
5505 printk(KERN_ERR
"ERROR: domain->cpu_capacity not set\n");
5509 if (!cpumask_weight(sched_group_cpus(group
))) {
5510 printk(KERN_CONT
"\n");
5511 printk(KERN_ERR
"ERROR: empty group\n");
5515 if (!(sd
->flags
& SD_OVERLAP
) &&
5516 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5517 printk(KERN_CONT
"\n");
5518 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5522 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5524 printk(KERN_CONT
" %*pbl",
5525 cpumask_pr_args(sched_group_cpus(group
)));
5526 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5527 printk(KERN_CONT
" (cpu_capacity = %d)",
5528 group
->sgc
->capacity
);
5531 group
= group
->next
;
5532 } while (group
!= sd
->groups
);
5533 printk(KERN_CONT
"\n");
5535 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5536 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5539 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5540 printk(KERN_ERR
"ERROR: parent span is not a superset "
5541 "of domain->span\n");
5545 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5549 if (!sched_debug_enabled
)
5553 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5557 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5560 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5568 #else /* !CONFIG_SCHED_DEBUG */
5569 # define sched_domain_debug(sd, cpu) do { } while (0)
5570 static inline bool sched_debug(void)
5574 #endif /* CONFIG_SCHED_DEBUG */
5576 static int sd_degenerate(struct sched_domain
*sd
)
5578 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5581 /* Following flags need at least 2 groups */
5582 if (sd
->flags
& (SD_LOAD_BALANCE
|
5583 SD_BALANCE_NEWIDLE
|
5586 SD_SHARE_CPUCAPACITY
|
5587 SD_SHARE_PKG_RESOURCES
|
5588 SD_SHARE_POWERDOMAIN
)) {
5589 if (sd
->groups
!= sd
->groups
->next
)
5593 /* Following flags don't use groups */
5594 if (sd
->flags
& (SD_WAKE_AFFINE
))
5601 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5603 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5605 if (sd_degenerate(parent
))
5608 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5611 /* Flags needing groups don't count if only 1 group in parent */
5612 if (parent
->groups
== parent
->groups
->next
) {
5613 pflags
&= ~(SD_LOAD_BALANCE
|
5614 SD_BALANCE_NEWIDLE
|
5617 SD_SHARE_CPUCAPACITY
|
5618 SD_SHARE_PKG_RESOURCES
|
5620 SD_SHARE_POWERDOMAIN
);
5621 if (nr_node_ids
== 1)
5622 pflags
&= ~SD_SERIALIZE
;
5624 if (~cflags
& pflags
)
5630 static void free_rootdomain(struct rcu_head
*rcu
)
5632 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5634 cpupri_cleanup(&rd
->cpupri
);
5635 cpudl_cleanup(&rd
->cpudl
);
5636 free_cpumask_var(rd
->dlo_mask
);
5637 free_cpumask_var(rd
->rto_mask
);
5638 free_cpumask_var(rd
->online
);
5639 free_cpumask_var(rd
->span
);
5643 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5645 struct root_domain
*old_rd
= NULL
;
5646 unsigned long flags
;
5648 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5653 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5656 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5659 * If we dont want to free the old_rd yet then
5660 * set old_rd to NULL to skip the freeing later
5663 if (!atomic_dec_and_test(&old_rd
->refcount
))
5667 atomic_inc(&rd
->refcount
);
5670 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5671 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5674 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5677 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5680 static int init_rootdomain(struct root_domain
*rd
)
5682 memset(rd
, 0, sizeof(*rd
));
5684 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5686 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5688 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5690 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5693 init_dl_bw(&rd
->dl_bw
);
5694 if (cpudl_init(&rd
->cpudl
) != 0)
5697 if (cpupri_init(&rd
->cpupri
) != 0)
5702 free_cpumask_var(rd
->rto_mask
);
5704 free_cpumask_var(rd
->dlo_mask
);
5706 free_cpumask_var(rd
->online
);
5708 free_cpumask_var(rd
->span
);
5714 * By default the system creates a single root-domain with all cpus as
5715 * members (mimicking the global state we have today).
5717 struct root_domain def_root_domain
;
5719 static void init_defrootdomain(void)
5721 init_rootdomain(&def_root_domain
);
5723 atomic_set(&def_root_domain
.refcount
, 1);
5726 static struct root_domain
*alloc_rootdomain(void)
5728 struct root_domain
*rd
;
5730 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5734 if (init_rootdomain(rd
) != 0) {
5742 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5744 struct sched_group
*tmp
, *first
;
5753 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5758 } while (sg
!= first
);
5761 static void free_sched_domain(struct rcu_head
*rcu
)
5763 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5766 * If its an overlapping domain it has private groups, iterate and
5769 if (sd
->flags
& SD_OVERLAP
) {
5770 free_sched_groups(sd
->groups
, 1);
5771 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5772 kfree(sd
->groups
->sgc
);
5778 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5780 call_rcu(&sd
->rcu
, free_sched_domain
);
5783 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5785 for (; sd
; sd
= sd
->parent
)
5786 destroy_sched_domain(sd
, cpu
);
5790 * Keep a special pointer to the highest sched_domain that has
5791 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5792 * allows us to avoid some pointer chasing select_idle_sibling().
5794 * Also keep a unique ID per domain (we use the first cpu number in
5795 * the cpumask of the domain), this allows us to quickly tell if
5796 * two cpus are in the same cache domain, see cpus_share_cache().
5798 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5799 DEFINE_PER_CPU(int, sd_llc_size
);
5800 DEFINE_PER_CPU(int, sd_llc_id
);
5801 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5802 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5803 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5805 static void update_top_cache_domain(int cpu
)
5807 struct sched_domain
*sd
;
5808 struct sched_domain
*busy_sd
= NULL
;
5812 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5814 id
= cpumask_first(sched_domain_span(sd
));
5815 size
= cpumask_weight(sched_domain_span(sd
));
5816 busy_sd
= sd
->parent
; /* sd_busy */
5818 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5820 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5821 per_cpu(sd_llc_size
, cpu
) = size
;
5822 per_cpu(sd_llc_id
, cpu
) = id
;
5824 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5825 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5827 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5828 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5832 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5833 * hold the hotplug lock.
5836 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5838 struct rq
*rq
= cpu_rq(cpu
);
5839 struct sched_domain
*tmp
;
5841 /* Remove the sched domains which do not contribute to scheduling. */
5842 for (tmp
= sd
; tmp
; ) {
5843 struct sched_domain
*parent
= tmp
->parent
;
5847 if (sd_parent_degenerate(tmp
, parent
)) {
5848 tmp
->parent
= parent
->parent
;
5850 parent
->parent
->child
= tmp
;
5852 * Transfer SD_PREFER_SIBLING down in case of a
5853 * degenerate parent; the spans match for this
5854 * so the property transfers.
5856 if (parent
->flags
& SD_PREFER_SIBLING
)
5857 tmp
->flags
|= SD_PREFER_SIBLING
;
5858 destroy_sched_domain(parent
, cpu
);
5863 if (sd
&& sd_degenerate(sd
)) {
5866 destroy_sched_domain(tmp
, cpu
);
5871 sched_domain_debug(sd
, cpu
);
5873 rq_attach_root(rq
, rd
);
5875 rcu_assign_pointer(rq
->sd
, sd
);
5876 destroy_sched_domains(tmp
, cpu
);
5878 update_top_cache_domain(cpu
);
5881 /* cpus with isolated domains */
5882 static cpumask_var_t cpu_isolated_map
;
5884 /* Setup the mask of cpus configured for isolated domains */
5885 static int __init
isolated_cpu_setup(char *str
)
5887 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5888 cpulist_parse(str
, cpu_isolated_map
);
5892 __setup("isolcpus=", isolated_cpu_setup
);
5895 struct sched_domain
** __percpu sd
;
5896 struct root_domain
*rd
;
5907 * Build an iteration mask that can exclude certain CPUs from the upwards
5910 * Asymmetric node setups can result in situations where the domain tree is of
5911 * unequal depth, make sure to skip domains that already cover the entire
5914 * In that case build_sched_domains() will have terminated the iteration early
5915 * and our sibling sd spans will be empty. Domains should always include the
5916 * cpu they're built on, so check that.
5919 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5921 const struct cpumask
*span
= sched_domain_span(sd
);
5922 struct sd_data
*sdd
= sd
->private;
5923 struct sched_domain
*sibling
;
5926 for_each_cpu(i
, span
) {
5927 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5928 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5931 cpumask_set_cpu(i
, sched_group_mask(sg
));
5936 * Return the canonical balance cpu for this group, this is the first cpu
5937 * of this group that's also in the iteration mask.
5939 int group_balance_cpu(struct sched_group
*sg
)
5941 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5945 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5947 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5948 const struct cpumask
*span
= sched_domain_span(sd
);
5949 struct cpumask
*covered
= sched_domains_tmpmask
;
5950 struct sd_data
*sdd
= sd
->private;
5951 struct sched_domain
*sibling
;
5954 cpumask_clear(covered
);
5956 for_each_cpu(i
, span
) {
5957 struct cpumask
*sg_span
;
5959 if (cpumask_test_cpu(i
, covered
))
5962 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5964 /* See the comment near build_group_mask(). */
5965 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5968 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5969 GFP_KERNEL
, cpu_to_node(cpu
));
5974 sg_span
= sched_group_cpus(sg
);
5976 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
5978 cpumask_set_cpu(i
, sg_span
);
5980 cpumask_or(covered
, covered
, sg_span
);
5982 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
5983 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
5984 build_group_mask(sd
, sg
);
5987 * Initialize sgc->capacity such that even if we mess up the
5988 * domains and no possible iteration will get us here, we won't
5991 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
5992 sg
->sgc
->capacity_orig
= sg
->sgc
->capacity
;
5995 * Make sure the first group of this domain contains the
5996 * canonical balance cpu. Otherwise the sched_domain iteration
5997 * breaks. See update_sg_lb_stats().
5999 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6000 group_balance_cpu(sg
) == cpu
)
6010 sd
->groups
= groups
;
6015 free_sched_groups(first
, 0);
6020 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6022 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6023 struct sched_domain
*child
= sd
->child
;
6026 cpu
= cpumask_first(sched_domain_span(child
));
6029 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6030 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6031 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6038 * build_sched_groups will build a circular linked list of the groups
6039 * covered by the given span, and will set each group's ->cpumask correctly,
6040 * and ->cpu_capacity to 0.
6042 * Assumes the sched_domain tree is fully constructed
6045 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6047 struct sched_group
*first
= NULL
, *last
= NULL
;
6048 struct sd_data
*sdd
= sd
->private;
6049 const struct cpumask
*span
= sched_domain_span(sd
);
6050 struct cpumask
*covered
;
6053 get_group(cpu
, sdd
, &sd
->groups
);
6054 atomic_inc(&sd
->groups
->ref
);
6056 if (cpu
!= cpumask_first(span
))
6059 lockdep_assert_held(&sched_domains_mutex
);
6060 covered
= sched_domains_tmpmask
;
6062 cpumask_clear(covered
);
6064 for_each_cpu(i
, span
) {
6065 struct sched_group
*sg
;
6068 if (cpumask_test_cpu(i
, covered
))
6071 group
= get_group(i
, sdd
, &sg
);
6072 cpumask_setall(sched_group_mask(sg
));
6074 for_each_cpu(j
, span
) {
6075 if (get_group(j
, sdd
, NULL
) != group
)
6078 cpumask_set_cpu(j
, covered
);
6079 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6094 * Initialize sched groups cpu_capacity.
6096 * cpu_capacity indicates the capacity of sched group, which is used while
6097 * distributing the load between different sched groups in a sched domain.
6098 * Typically cpu_capacity for all the groups in a sched domain will be same
6099 * unless there are asymmetries in the topology. If there are asymmetries,
6100 * group having more cpu_capacity will pickup more load compared to the
6101 * group having less cpu_capacity.
6103 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6105 struct sched_group
*sg
= sd
->groups
;
6110 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6112 } while (sg
!= sd
->groups
);
6114 if (cpu
!= group_balance_cpu(sg
))
6117 update_group_capacity(sd
, cpu
);
6118 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6122 * Initializers for schedule domains
6123 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6126 static int default_relax_domain_level
= -1;
6127 int sched_domain_level_max
;
6129 static int __init
setup_relax_domain_level(char *str
)
6131 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6132 pr_warn("Unable to set relax_domain_level\n");
6136 __setup("relax_domain_level=", setup_relax_domain_level
);
6138 static void set_domain_attribute(struct sched_domain
*sd
,
6139 struct sched_domain_attr
*attr
)
6143 if (!attr
|| attr
->relax_domain_level
< 0) {
6144 if (default_relax_domain_level
< 0)
6147 request
= default_relax_domain_level
;
6149 request
= attr
->relax_domain_level
;
6150 if (request
< sd
->level
) {
6151 /* turn off idle balance on this domain */
6152 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6154 /* turn on idle balance on this domain */
6155 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6159 static void __sdt_free(const struct cpumask
*cpu_map
);
6160 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6162 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6163 const struct cpumask
*cpu_map
)
6167 if (!atomic_read(&d
->rd
->refcount
))
6168 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6170 free_percpu(d
->sd
); /* fall through */
6172 __sdt_free(cpu_map
); /* fall through */
6178 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6179 const struct cpumask
*cpu_map
)
6181 memset(d
, 0, sizeof(*d
));
6183 if (__sdt_alloc(cpu_map
))
6184 return sa_sd_storage
;
6185 d
->sd
= alloc_percpu(struct sched_domain
*);
6187 return sa_sd_storage
;
6188 d
->rd
= alloc_rootdomain();
6191 return sa_rootdomain
;
6195 * NULL the sd_data elements we've used to build the sched_domain and
6196 * sched_group structure so that the subsequent __free_domain_allocs()
6197 * will not free the data we're using.
6199 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6201 struct sd_data
*sdd
= sd
->private;
6203 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6204 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6206 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6207 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6209 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6210 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6214 static int sched_domains_numa_levels
;
6215 enum numa_topology_type sched_numa_topology_type
;
6216 static int *sched_domains_numa_distance
;
6217 int sched_max_numa_distance
;
6218 static struct cpumask
***sched_domains_numa_masks
;
6219 static int sched_domains_curr_level
;
6223 * SD_flags allowed in topology descriptions.
6225 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6226 * SD_SHARE_PKG_RESOURCES - describes shared caches
6227 * SD_NUMA - describes NUMA topologies
6228 * SD_SHARE_POWERDOMAIN - describes shared power domain
6231 * SD_ASYM_PACKING - describes SMT quirks
6233 #define TOPOLOGY_SD_FLAGS \
6234 (SD_SHARE_CPUCAPACITY | \
6235 SD_SHARE_PKG_RESOURCES | \
6238 SD_SHARE_POWERDOMAIN)
6240 static struct sched_domain
*
6241 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6243 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6244 int sd_weight
, sd_flags
= 0;
6248 * Ugly hack to pass state to sd_numa_mask()...
6250 sched_domains_curr_level
= tl
->numa_level
;
6253 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6256 sd_flags
= (*tl
->sd_flags
)();
6257 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6258 "wrong sd_flags in topology description\n"))
6259 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6261 *sd
= (struct sched_domain
){
6262 .min_interval
= sd_weight
,
6263 .max_interval
= 2*sd_weight
,
6265 .imbalance_pct
= 125,
6267 .cache_nice_tries
= 0,
6274 .flags
= 1*SD_LOAD_BALANCE
6275 | 1*SD_BALANCE_NEWIDLE
6280 | 0*SD_SHARE_CPUCAPACITY
6281 | 0*SD_SHARE_PKG_RESOURCES
6283 | 0*SD_PREFER_SIBLING
6288 .last_balance
= jiffies
,
6289 .balance_interval
= sd_weight
,
6291 .max_newidle_lb_cost
= 0,
6292 .next_decay_max_lb_cost
= jiffies
,
6293 #ifdef CONFIG_SCHED_DEBUG
6299 * Convert topological properties into behaviour.
6302 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6303 sd
->imbalance_pct
= 110;
6304 sd
->smt_gain
= 1178; /* ~15% */
6306 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6307 sd
->imbalance_pct
= 117;
6308 sd
->cache_nice_tries
= 1;
6312 } else if (sd
->flags
& SD_NUMA
) {
6313 sd
->cache_nice_tries
= 2;
6317 sd
->flags
|= SD_SERIALIZE
;
6318 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6319 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6326 sd
->flags
|= SD_PREFER_SIBLING
;
6327 sd
->cache_nice_tries
= 1;
6332 sd
->private = &tl
->data
;
6338 * Topology list, bottom-up.
6340 static struct sched_domain_topology_level default_topology
[] = {
6341 #ifdef CONFIG_SCHED_SMT
6342 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6344 #ifdef CONFIG_SCHED_MC
6345 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6347 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6351 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6353 #define for_each_sd_topology(tl) \
6354 for (tl = sched_domain_topology; tl->mask; tl++)
6356 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6358 sched_domain_topology
= tl
;
6363 static const struct cpumask
*sd_numa_mask(int cpu
)
6365 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6368 static void sched_numa_warn(const char *str
)
6370 static int done
= false;
6378 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6380 for (i
= 0; i
< nr_node_ids
; i
++) {
6381 printk(KERN_WARNING
" ");
6382 for (j
= 0; j
< nr_node_ids
; j
++)
6383 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6384 printk(KERN_CONT
"\n");
6386 printk(KERN_WARNING
"\n");
6389 bool find_numa_distance(int distance
)
6393 if (distance
== node_distance(0, 0))
6396 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6397 if (sched_domains_numa_distance
[i
] == distance
)
6405 * A system can have three types of NUMA topology:
6406 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6407 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6408 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6410 * The difference between a glueless mesh topology and a backplane
6411 * topology lies in whether communication between not directly
6412 * connected nodes goes through intermediary nodes (where programs
6413 * could run), or through backplane controllers. This affects
6414 * placement of programs.
6416 * The type of topology can be discerned with the following tests:
6417 * - If the maximum distance between any nodes is 1 hop, the system
6418 * is directly connected.
6419 * - If for two nodes A and B, located N > 1 hops away from each other,
6420 * there is an intermediary node C, which is < N hops away from both
6421 * nodes A and B, the system is a glueless mesh.
6423 static void init_numa_topology_type(void)
6427 n
= sched_max_numa_distance
;
6430 sched_numa_topology_type
= NUMA_DIRECT
;
6432 for_each_online_node(a
) {
6433 for_each_online_node(b
) {
6434 /* Find two nodes furthest removed from each other. */
6435 if (node_distance(a
, b
) < n
)
6438 /* Is there an intermediary node between a and b? */
6439 for_each_online_node(c
) {
6440 if (node_distance(a
, c
) < n
&&
6441 node_distance(b
, c
) < n
) {
6442 sched_numa_topology_type
=
6448 sched_numa_topology_type
= NUMA_BACKPLANE
;
6454 static void sched_init_numa(void)
6456 int next_distance
, curr_distance
= node_distance(0, 0);
6457 struct sched_domain_topology_level
*tl
;
6461 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6462 if (!sched_domains_numa_distance
)
6466 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6467 * unique distances in the node_distance() table.
6469 * Assumes node_distance(0,j) includes all distances in
6470 * node_distance(i,j) in order to avoid cubic time.
6472 next_distance
= curr_distance
;
6473 for (i
= 0; i
< nr_node_ids
; i
++) {
6474 for (j
= 0; j
< nr_node_ids
; j
++) {
6475 for (k
= 0; k
< nr_node_ids
; k
++) {
6476 int distance
= node_distance(i
, k
);
6478 if (distance
> curr_distance
&&
6479 (distance
< next_distance
||
6480 next_distance
== curr_distance
))
6481 next_distance
= distance
;
6484 * While not a strong assumption it would be nice to know
6485 * about cases where if node A is connected to B, B is not
6486 * equally connected to A.
6488 if (sched_debug() && node_distance(k
, i
) != distance
)
6489 sched_numa_warn("Node-distance not symmetric");
6491 if (sched_debug() && i
&& !find_numa_distance(distance
))
6492 sched_numa_warn("Node-0 not representative");
6494 if (next_distance
!= curr_distance
) {
6495 sched_domains_numa_distance
[level
++] = next_distance
;
6496 sched_domains_numa_levels
= level
;
6497 curr_distance
= next_distance
;
6502 * In case of sched_debug() we verify the above assumption.
6512 * 'level' contains the number of unique distances, excluding the
6513 * identity distance node_distance(i,i).
6515 * The sched_domains_numa_distance[] array includes the actual distance
6520 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6521 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6522 * the array will contain less then 'level' members. This could be
6523 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6524 * in other functions.
6526 * We reset it to 'level' at the end of this function.
6528 sched_domains_numa_levels
= 0;
6530 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6531 if (!sched_domains_numa_masks
)
6535 * Now for each level, construct a mask per node which contains all
6536 * cpus of nodes that are that many hops away from us.
6538 for (i
= 0; i
< level
; i
++) {
6539 sched_domains_numa_masks
[i
] =
6540 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6541 if (!sched_domains_numa_masks
[i
])
6544 for (j
= 0; j
< nr_node_ids
; j
++) {
6545 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6549 sched_domains_numa_masks
[i
][j
] = mask
;
6551 for (k
= 0; k
< nr_node_ids
; k
++) {
6552 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6555 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6560 /* Compute default topology size */
6561 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6563 tl
= kzalloc((i
+ level
+ 1) *
6564 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6569 * Copy the default topology bits..
6571 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6572 tl
[i
] = sched_domain_topology
[i
];
6575 * .. and append 'j' levels of NUMA goodness.
6577 for (j
= 0; j
< level
; i
++, j
++) {
6578 tl
[i
] = (struct sched_domain_topology_level
){
6579 .mask
= sd_numa_mask
,
6580 .sd_flags
= cpu_numa_flags
,
6581 .flags
= SDTL_OVERLAP
,
6587 sched_domain_topology
= tl
;
6589 sched_domains_numa_levels
= level
;
6590 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6592 init_numa_topology_type();
6595 static void sched_domains_numa_masks_set(int cpu
)
6598 int node
= cpu_to_node(cpu
);
6600 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6601 for (j
= 0; j
< nr_node_ids
; j
++) {
6602 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6603 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6608 static void sched_domains_numa_masks_clear(int cpu
)
6611 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6612 for (j
= 0; j
< nr_node_ids
; j
++)
6613 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6618 * Update sched_domains_numa_masks[level][node] array when new cpus
6621 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6622 unsigned long action
,
6625 int cpu
= (long)hcpu
;
6627 switch (action
& ~CPU_TASKS_FROZEN
) {
6629 sched_domains_numa_masks_set(cpu
);
6633 sched_domains_numa_masks_clear(cpu
);
6643 static inline void sched_init_numa(void)
6647 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6648 unsigned long action
,
6653 #endif /* CONFIG_NUMA */
6655 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6657 struct sched_domain_topology_level
*tl
;
6660 for_each_sd_topology(tl
) {
6661 struct sd_data
*sdd
= &tl
->data
;
6663 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6667 sdd
->sg
= alloc_percpu(struct sched_group
*);
6671 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6675 for_each_cpu(j
, cpu_map
) {
6676 struct sched_domain
*sd
;
6677 struct sched_group
*sg
;
6678 struct sched_group_capacity
*sgc
;
6680 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6681 GFP_KERNEL
, cpu_to_node(j
));
6685 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6687 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6688 GFP_KERNEL
, cpu_to_node(j
));
6694 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6696 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6697 GFP_KERNEL
, cpu_to_node(j
));
6701 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6708 static void __sdt_free(const struct cpumask
*cpu_map
)
6710 struct sched_domain_topology_level
*tl
;
6713 for_each_sd_topology(tl
) {
6714 struct sd_data
*sdd
= &tl
->data
;
6716 for_each_cpu(j
, cpu_map
) {
6717 struct sched_domain
*sd
;
6720 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6721 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6722 free_sched_groups(sd
->groups
, 0);
6723 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6727 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6729 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6731 free_percpu(sdd
->sd
);
6733 free_percpu(sdd
->sg
);
6735 free_percpu(sdd
->sgc
);
6740 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6741 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6742 struct sched_domain
*child
, int cpu
)
6744 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6748 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6750 sd
->level
= child
->level
+ 1;
6751 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6755 if (!cpumask_subset(sched_domain_span(child
),
6756 sched_domain_span(sd
))) {
6757 pr_err("BUG: arch topology borken\n");
6758 #ifdef CONFIG_SCHED_DEBUG
6759 pr_err(" the %s domain not a subset of the %s domain\n",
6760 child
->name
, sd
->name
);
6762 /* Fixup, ensure @sd has at least @child cpus. */
6763 cpumask_or(sched_domain_span(sd
),
6764 sched_domain_span(sd
),
6765 sched_domain_span(child
));
6769 set_domain_attribute(sd
, attr
);
6775 * Build sched domains for a given set of cpus and attach the sched domains
6776 * to the individual cpus
6778 static int build_sched_domains(const struct cpumask
*cpu_map
,
6779 struct sched_domain_attr
*attr
)
6781 enum s_alloc alloc_state
;
6782 struct sched_domain
*sd
;
6784 int i
, ret
= -ENOMEM
;
6786 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6787 if (alloc_state
!= sa_rootdomain
)
6790 /* Set up domains for cpus specified by the cpu_map. */
6791 for_each_cpu(i
, cpu_map
) {
6792 struct sched_domain_topology_level
*tl
;
6795 for_each_sd_topology(tl
) {
6796 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6797 if (tl
== sched_domain_topology
)
6798 *per_cpu_ptr(d
.sd
, i
) = sd
;
6799 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6800 sd
->flags
|= SD_OVERLAP
;
6801 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6806 /* Build the groups for the domains */
6807 for_each_cpu(i
, cpu_map
) {
6808 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6809 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6810 if (sd
->flags
& SD_OVERLAP
) {
6811 if (build_overlap_sched_groups(sd
, i
))
6814 if (build_sched_groups(sd
, i
))
6820 /* Calculate CPU capacity for physical packages and nodes */
6821 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6822 if (!cpumask_test_cpu(i
, cpu_map
))
6825 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6826 claim_allocations(i
, sd
);
6827 init_sched_groups_capacity(i
, sd
);
6831 /* Attach the domains */
6833 for_each_cpu(i
, cpu_map
) {
6834 sd
= *per_cpu_ptr(d
.sd
, i
);
6835 cpu_attach_domain(sd
, d
.rd
, i
);
6841 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6845 static cpumask_var_t
*doms_cur
; /* current sched domains */
6846 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6847 static struct sched_domain_attr
*dattr_cur
;
6848 /* attribues of custom domains in 'doms_cur' */
6851 * Special case: If a kmalloc of a doms_cur partition (array of
6852 * cpumask) fails, then fallback to a single sched domain,
6853 * as determined by the single cpumask fallback_doms.
6855 static cpumask_var_t fallback_doms
;
6858 * arch_update_cpu_topology lets virtualized architectures update the
6859 * cpu core maps. It is supposed to return 1 if the topology changed
6860 * or 0 if it stayed the same.
6862 int __weak
arch_update_cpu_topology(void)
6867 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6870 cpumask_var_t
*doms
;
6872 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6875 for (i
= 0; i
< ndoms
; i
++) {
6876 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6877 free_sched_domains(doms
, i
);
6884 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6887 for (i
= 0; i
< ndoms
; i
++)
6888 free_cpumask_var(doms
[i
]);
6893 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6894 * For now this just excludes isolated cpus, but could be used to
6895 * exclude other special cases in the future.
6897 static int init_sched_domains(const struct cpumask
*cpu_map
)
6901 arch_update_cpu_topology();
6903 doms_cur
= alloc_sched_domains(ndoms_cur
);
6905 doms_cur
= &fallback_doms
;
6906 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6907 err
= build_sched_domains(doms_cur
[0], NULL
);
6908 register_sched_domain_sysctl();
6914 * Detach sched domains from a group of cpus specified in cpu_map
6915 * These cpus will now be attached to the NULL domain
6917 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6922 for_each_cpu(i
, cpu_map
)
6923 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6927 /* handle null as "default" */
6928 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6929 struct sched_domain_attr
*new, int idx_new
)
6931 struct sched_domain_attr tmp
;
6938 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6939 new ? (new + idx_new
) : &tmp
,
6940 sizeof(struct sched_domain_attr
));
6944 * Partition sched domains as specified by the 'ndoms_new'
6945 * cpumasks in the array doms_new[] of cpumasks. This compares
6946 * doms_new[] to the current sched domain partitioning, doms_cur[].
6947 * It destroys each deleted domain and builds each new domain.
6949 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6950 * The masks don't intersect (don't overlap.) We should setup one
6951 * sched domain for each mask. CPUs not in any of the cpumasks will
6952 * not be load balanced. If the same cpumask appears both in the
6953 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6956 * The passed in 'doms_new' should be allocated using
6957 * alloc_sched_domains. This routine takes ownership of it and will
6958 * free_sched_domains it when done with it. If the caller failed the
6959 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6960 * and partition_sched_domains() will fallback to the single partition
6961 * 'fallback_doms', it also forces the domains to be rebuilt.
6963 * If doms_new == NULL it will be replaced with cpu_online_mask.
6964 * ndoms_new == 0 is a special case for destroying existing domains,
6965 * and it will not create the default domain.
6967 * Call with hotplug lock held
6969 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6970 struct sched_domain_attr
*dattr_new
)
6975 mutex_lock(&sched_domains_mutex
);
6977 /* always unregister in case we don't destroy any domains */
6978 unregister_sched_domain_sysctl();
6980 /* Let architecture update cpu core mappings. */
6981 new_topology
= arch_update_cpu_topology();
6983 n
= doms_new
? ndoms_new
: 0;
6985 /* Destroy deleted domains */
6986 for (i
= 0; i
< ndoms_cur
; i
++) {
6987 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6988 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6989 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6992 /* no match - a current sched domain not in new doms_new[] */
6993 detach_destroy_domains(doms_cur
[i
]);
6999 if (doms_new
== NULL
) {
7001 doms_new
= &fallback_doms
;
7002 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7003 WARN_ON_ONCE(dattr_new
);
7006 /* Build new domains */
7007 for (i
= 0; i
< ndoms_new
; i
++) {
7008 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7009 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7010 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7013 /* no match - add a new doms_new */
7014 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7019 /* Remember the new sched domains */
7020 if (doms_cur
!= &fallback_doms
)
7021 free_sched_domains(doms_cur
, ndoms_cur
);
7022 kfree(dattr_cur
); /* kfree(NULL) is safe */
7023 doms_cur
= doms_new
;
7024 dattr_cur
= dattr_new
;
7025 ndoms_cur
= ndoms_new
;
7027 register_sched_domain_sysctl();
7029 mutex_unlock(&sched_domains_mutex
);
7032 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7035 * Update cpusets according to cpu_active mask. If cpusets are
7036 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7037 * around partition_sched_domains().
7039 * If we come here as part of a suspend/resume, don't touch cpusets because we
7040 * want to restore it back to its original state upon resume anyway.
7042 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7046 case CPU_ONLINE_FROZEN
:
7047 case CPU_DOWN_FAILED_FROZEN
:
7050 * num_cpus_frozen tracks how many CPUs are involved in suspend
7051 * resume sequence. As long as this is not the last online
7052 * operation in the resume sequence, just build a single sched
7053 * domain, ignoring cpusets.
7056 if (likely(num_cpus_frozen
)) {
7057 partition_sched_domains(1, NULL
, NULL
);
7062 * This is the last CPU online operation. So fall through and
7063 * restore the original sched domains by considering the
7064 * cpuset configurations.
7068 case CPU_DOWN_FAILED
:
7069 cpuset_update_active_cpus(true);
7077 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7081 case CPU_DOWN_PREPARE
:
7082 cpuset_update_active_cpus(false);
7084 case CPU_DOWN_PREPARE_FROZEN
:
7086 partition_sched_domains(1, NULL
, NULL
);
7094 void __init
sched_init_smp(void)
7096 cpumask_var_t non_isolated_cpus
;
7098 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7099 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7104 * There's no userspace yet to cause hotplug operations; hence all the
7105 * cpu masks are stable and all blatant races in the below code cannot
7108 mutex_lock(&sched_domains_mutex
);
7109 init_sched_domains(cpu_active_mask
);
7110 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7111 if (cpumask_empty(non_isolated_cpus
))
7112 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7113 mutex_unlock(&sched_domains_mutex
);
7115 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7116 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7117 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7121 /* Move init over to a non-isolated CPU */
7122 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7124 sched_init_granularity();
7125 free_cpumask_var(non_isolated_cpus
);
7127 init_sched_rt_class();
7128 init_sched_dl_class();
7131 void __init
sched_init_smp(void)
7133 sched_init_granularity();
7135 #endif /* CONFIG_SMP */
7137 const_debug
unsigned int sysctl_timer_migration
= 1;
7139 int in_sched_functions(unsigned long addr
)
7141 return in_lock_functions(addr
) ||
7142 (addr
>= (unsigned long)__sched_text_start
7143 && addr
< (unsigned long)__sched_text_end
);
7146 #ifdef CONFIG_CGROUP_SCHED
7148 * Default task group.
7149 * Every task in system belongs to this group at bootup.
7151 struct task_group root_task_group
;
7152 LIST_HEAD(task_groups
);
7155 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7157 void __init
sched_init(void)
7160 unsigned long alloc_size
= 0, ptr
;
7162 #ifdef CONFIG_FAIR_GROUP_SCHED
7163 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7165 #ifdef CONFIG_RT_GROUP_SCHED
7166 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7169 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7171 #ifdef CONFIG_FAIR_GROUP_SCHED
7172 root_task_group
.se
= (struct sched_entity
**)ptr
;
7173 ptr
+= nr_cpu_ids
* sizeof(void **);
7175 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7176 ptr
+= nr_cpu_ids
* sizeof(void **);
7178 #endif /* CONFIG_FAIR_GROUP_SCHED */
7179 #ifdef CONFIG_RT_GROUP_SCHED
7180 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7181 ptr
+= nr_cpu_ids
* sizeof(void **);
7183 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7184 ptr
+= nr_cpu_ids
* sizeof(void **);
7186 #endif /* CONFIG_RT_GROUP_SCHED */
7188 #ifdef CONFIG_CPUMASK_OFFSTACK
7189 for_each_possible_cpu(i
) {
7190 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7191 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7193 #endif /* CONFIG_CPUMASK_OFFSTACK */
7195 init_rt_bandwidth(&def_rt_bandwidth
,
7196 global_rt_period(), global_rt_runtime());
7197 init_dl_bandwidth(&def_dl_bandwidth
,
7198 global_rt_period(), global_rt_runtime());
7201 init_defrootdomain();
7204 #ifdef CONFIG_RT_GROUP_SCHED
7205 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7206 global_rt_period(), global_rt_runtime());
7207 #endif /* CONFIG_RT_GROUP_SCHED */
7209 #ifdef CONFIG_CGROUP_SCHED
7210 list_add(&root_task_group
.list
, &task_groups
);
7211 INIT_LIST_HEAD(&root_task_group
.children
);
7212 INIT_LIST_HEAD(&root_task_group
.siblings
);
7213 autogroup_init(&init_task
);
7215 #endif /* CONFIG_CGROUP_SCHED */
7217 for_each_possible_cpu(i
) {
7221 raw_spin_lock_init(&rq
->lock
);
7223 rq
->calc_load_active
= 0;
7224 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7225 init_cfs_rq(&rq
->cfs
);
7226 init_rt_rq(&rq
->rt
, rq
);
7227 init_dl_rq(&rq
->dl
, rq
);
7228 #ifdef CONFIG_FAIR_GROUP_SCHED
7229 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7230 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7232 * How much cpu bandwidth does root_task_group get?
7234 * In case of task-groups formed thr' the cgroup filesystem, it
7235 * gets 100% of the cpu resources in the system. This overall
7236 * system cpu resource is divided among the tasks of
7237 * root_task_group and its child task-groups in a fair manner,
7238 * based on each entity's (task or task-group's) weight
7239 * (se->load.weight).
7241 * In other words, if root_task_group has 10 tasks of weight
7242 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7243 * then A0's share of the cpu resource is:
7245 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7247 * We achieve this by letting root_task_group's tasks sit
7248 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7250 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7251 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7252 #endif /* CONFIG_FAIR_GROUP_SCHED */
7254 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7255 #ifdef CONFIG_RT_GROUP_SCHED
7256 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7259 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7260 rq
->cpu_load
[j
] = 0;
7262 rq
->last_load_update_tick
= jiffies
;
7267 rq
->cpu_capacity
= SCHED_CAPACITY_SCALE
;
7268 rq
->post_schedule
= 0;
7269 rq
->active_balance
= 0;
7270 rq
->next_balance
= jiffies
;
7275 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7276 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7278 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7280 rq_attach_root(rq
, &def_root_domain
);
7281 #ifdef CONFIG_NO_HZ_COMMON
7284 #ifdef CONFIG_NO_HZ_FULL
7285 rq
->last_sched_tick
= 0;
7289 atomic_set(&rq
->nr_iowait
, 0);
7292 set_load_weight(&init_task
);
7294 #ifdef CONFIG_PREEMPT_NOTIFIERS
7295 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7299 * The boot idle thread does lazy MMU switching as well:
7301 atomic_inc(&init_mm
.mm_count
);
7302 enter_lazy_tlb(&init_mm
, current
);
7305 * During early bootup we pretend to be a normal task:
7307 current
->sched_class
= &fair_sched_class
;
7310 * Make us the idle thread. Technically, schedule() should not be
7311 * called from this thread, however somewhere below it might be,
7312 * but because we are the idle thread, we just pick up running again
7313 * when this runqueue becomes "idle".
7315 init_idle(current
, smp_processor_id());
7317 calc_load_update
= jiffies
+ LOAD_FREQ
;
7320 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7321 /* May be allocated at isolcpus cmdline parse time */
7322 if (cpu_isolated_map
== NULL
)
7323 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7324 idle_thread_set_boot_cpu();
7325 set_cpu_rq_start_time();
7327 init_sched_fair_class();
7329 scheduler_running
= 1;
7332 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7333 static inline int preempt_count_equals(int preempt_offset
)
7335 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7337 return (nested
== preempt_offset
);
7340 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7343 * Blocking primitives will set (and therefore destroy) current->state,
7344 * since we will exit with TASK_RUNNING make sure we enter with it,
7345 * otherwise we will destroy state.
7347 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7348 "do not call blocking ops when !TASK_RUNNING; "
7349 "state=%lx set at [<%p>] %pS\n",
7351 (void *)current
->task_state_change
,
7352 (void *)current
->task_state_change
);
7354 ___might_sleep(file
, line
, preempt_offset
);
7356 EXPORT_SYMBOL(__might_sleep
);
7358 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7360 static unsigned long prev_jiffy
; /* ratelimiting */
7362 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7363 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7364 !is_idle_task(current
)) ||
7365 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7367 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7369 prev_jiffy
= jiffies
;
7372 "BUG: sleeping function called from invalid context at %s:%d\n",
7375 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7376 in_atomic(), irqs_disabled(),
7377 current
->pid
, current
->comm
);
7379 if (task_stack_end_corrupted(current
))
7380 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7382 debug_show_held_locks(current
);
7383 if (irqs_disabled())
7384 print_irqtrace_events(current
);
7385 #ifdef CONFIG_DEBUG_PREEMPT
7386 if (!preempt_count_equals(preempt_offset
)) {
7387 pr_err("Preemption disabled at:");
7388 print_ip_sym(current
->preempt_disable_ip
);
7394 EXPORT_SYMBOL(___might_sleep
);
7397 #ifdef CONFIG_MAGIC_SYSRQ
7398 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7400 const struct sched_class
*prev_class
= p
->sched_class
;
7401 struct sched_attr attr
= {
7402 .sched_policy
= SCHED_NORMAL
,
7404 int old_prio
= p
->prio
;
7407 queued
= task_on_rq_queued(p
);
7409 dequeue_task(rq
, p
, 0);
7410 __setscheduler(rq
, p
, &attr
);
7412 enqueue_task(rq
, p
, 0);
7416 check_class_changed(rq
, p
, prev_class
, old_prio
);
7419 void normalize_rt_tasks(void)
7421 struct task_struct
*g
, *p
;
7422 unsigned long flags
;
7425 read_lock(&tasklist_lock
);
7426 for_each_process_thread(g
, p
) {
7428 * Only normalize user tasks:
7430 if (p
->flags
& PF_KTHREAD
)
7433 p
->se
.exec_start
= 0;
7434 #ifdef CONFIG_SCHEDSTATS
7435 p
->se
.statistics
.wait_start
= 0;
7436 p
->se
.statistics
.sleep_start
= 0;
7437 p
->se
.statistics
.block_start
= 0;
7440 if (!dl_task(p
) && !rt_task(p
)) {
7442 * Renice negative nice level userspace
7445 if (task_nice(p
) < 0)
7446 set_user_nice(p
, 0);
7450 rq
= task_rq_lock(p
, &flags
);
7451 normalize_task(rq
, p
);
7452 task_rq_unlock(rq
, p
, &flags
);
7454 read_unlock(&tasklist_lock
);
7457 #endif /* CONFIG_MAGIC_SYSRQ */
7459 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7461 * These functions are only useful for the IA64 MCA handling, or kdb.
7463 * They can only be called when the whole system has been
7464 * stopped - every CPU needs to be quiescent, and no scheduling
7465 * activity can take place. Using them for anything else would
7466 * be a serious bug, and as a result, they aren't even visible
7467 * under any other configuration.
7471 * curr_task - return the current task for a given cpu.
7472 * @cpu: the processor in question.
7474 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7476 * Return: The current task for @cpu.
7478 struct task_struct
*curr_task(int cpu
)
7480 return cpu_curr(cpu
);
7483 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7487 * set_curr_task - set the current task for a given cpu.
7488 * @cpu: the processor in question.
7489 * @p: the task pointer to set.
7491 * Description: This function must only be used when non-maskable interrupts
7492 * are serviced on a separate stack. It allows the architecture to switch the
7493 * notion of the current task on a cpu in a non-blocking manner. This function
7494 * must be called with all CPU's synchronized, and interrupts disabled, the
7495 * and caller must save the original value of the current task (see
7496 * curr_task() above) and restore that value before reenabling interrupts and
7497 * re-starting the system.
7499 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7501 void set_curr_task(int cpu
, struct task_struct
*p
)
7508 #ifdef CONFIG_CGROUP_SCHED
7509 /* task_group_lock serializes the addition/removal of task groups */
7510 static DEFINE_SPINLOCK(task_group_lock
);
7512 static void free_sched_group(struct task_group
*tg
)
7514 free_fair_sched_group(tg
);
7515 free_rt_sched_group(tg
);
7520 /* allocate runqueue etc for a new task group */
7521 struct task_group
*sched_create_group(struct task_group
*parent
)
7523 struct task_group
*tg
;
7525 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7527 return ERR_PTR(-ENOMEM
);
7529 if (!alloc_fair_sched_group(tg
, parent
))
7532 if (!alloc_rt_sched_group(tg
, parent
))
7538 free_sched_group(tg
);
7539 return ERR_PTR(-ENOMEM
);
7542 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7544 unsigned long flags
;
7546 spin_lock_irqsave(&task_group_lock
, flags
);
7547 list_add_rcu(&tg
->list
, &task_groups
);
7549 WARN_ON(!parent
); /* root should already exist */
7551 tg
->parent
= parent
;
7552 INIT_LIST_HEAD(&tg
->children
);
7553 list_add_rcu(&tg
->siblings
, &parent
->children
);
7554 spin_unlock_irqrestore(&task_group_lock
, flags
);
7557 /* rcu callback to free various structures associated with a task group */
7558 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7560 /* now it should be safe to free those cfs_rqs */
7561 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7564 /* Destroy runqueue etc associated with a task group */
7565 void sched_destroy_group(struct task_group
*tg
)
7567 /* wait for possible concurrent references to cfs_rqs complete */
7568 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7571 void sched_offline_group(struct task_group
*tg
)
7573 unsigned long flags
;
7576 /* end participation in shares distribution */
7577 for_each_possible_cpu(i
)
7578 unregister_fair_sched_group(tg
, i
);
7580 spin_lock_irqsave(&task_group_lock
, flags
);
7581 list_del_rcu(&tg
->list
);
7582 list_del_rcu(&tg
->siblings
);
7583 spin_unlock_irqrestore(&task_group_lock
, flags
);
7586 /* change task's runqueue when it moves between groups.
7587 * The caller of this function should have put the task in its new group
7588 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7589 * reflect its new group.
7591 void sched_move_task(struct task_struct
*tsk
)
7593 struct task_group
*tg
;
7594 int queued
, running
;
7595 unsigned long flags
;
7598 rq
= task_rq_lock(tsk
, &flags
);
7600 running
= task_current(rq
, tsk
);
7601 queued
= task_on_rq_queued(tsk
);
7604 dequeue_task(rq
, tsk
, 0);
7605 if (unlikely(running
))
7606 put_prev_task(rq
, tsk
);
7609 * All callers are synchronized by task_rq_lock(); we do not use RCU
7610 * which is pointless here. Thus, we pass "true" to task_css_check()
7611 * to prevent lockdep warnings.
7613 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7614 struct task_group
, css
);
7615 tg
= autogroup_task_group(tsk
, tg
);
7616 tsk
->sched_task_group
= tg
;
7618 #ifdef CONFIG_FAIR_GROUP_SCHED
7619 if (tsk
->sched_class
->task_move_group
)
7620 tsk
->sched_class
->task_move_group(tsk
, queued
);
7623 set_task_rq(tsk
, task_cpu(tsk
));
7625 if (unlikely(running
))
7626 tsk
->sched_class
->set_curr_task(rq
);
7628 enqueue_task(rq
, tsk
, 0);
7630 task_rq_unlock(rq
, tsk
, &flags
);
7632 #endif /* CONFIG_CGROUP_SCHED */
7634 #ifdef CONFIG_RT_GROUP_SCHED
7636 * Ensure that the real time constraints are schedulable.
7638 static DEFINE_MUTEX(rt_constraints_mutex
);
7640 /* Must be called with tasklist_lock held */
7641 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7643 struct task_struct
*g
, *p
;
7645 for_each_process_thread(g
, p
) {
7646 if (rt_task(p
) && task_group(p
) == tg
)
7653 struct rt_schedulable_data
{
7654 struct task_group
*tg
;
7659 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7661 struct rt_schedulable_data
*d
= data
;
7662 struct task_group
*child
;
7663 unsigned long total
, sum
= 0;
7664 u64 period
, runtime
;
7666 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7667 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7670 period
= d
->rt_period
;
7671 runtime
= d
->rt_runtime
;
7675 * Cannot have more runtime than the period.
7677 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7681 * Ensure we don't starve existing RT tasks.
7683 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7686 total
= to_ratio(period
, runtime
);
7689 * Nobody can have more than the global setting allows.
7691 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7695 * The sum of our children's runtime should not exceed our own.
7697 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7698 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7699 runtime
= child
->rt_bandwidth
.rt_runtime
;
7701 if (child
== d
->tg
) {
7702 period
= d
->rt_period
;
7703 runtime
= d
->rt_runtime
;
7706 sum
+= to_ratio(period
, runtime
);
7715 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7719 struct rt_schedulable_data data
= {
7721 .rt_period
= period
,
7722 .rt_runtime
= runtime
,
7726 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7732 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7733 u64 rt_period
, u64 rt_runtime
)
7737 mutex_lock(&rt_constraints_mutex
);
7738 read_lock(&tasklist_lock
);
7739 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7743 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7744 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7745 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7747 for_each_possible_cpu(i
) {
7748 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7750 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7751 rt_rq
->rt_runtime
= rt_runtime
;
7752 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7754 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7756 read_unlock(&tasklist_lock
);
7757 mutex_unlock(&rt_constraints_mutex
);
7762 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7764 u64 rt_runtime
, rt_period
;
7766 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7767 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7768 if (rt_runtime_us
< 0)
7769 rt_runtime
= RUNTIME_INF
;
7771 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7774 static long sched_group_rt_runtime(struct task_group
*tg
)
7778 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7781 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7782 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7783 return rt_runtime_us
;
7786 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7788 u64 rt_runtime
, rt_period
;
7790 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7791 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7796 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7799 static long sched_group_rt_period(struct task_group
*tg
)
7803 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7804 do_div(rt_period_us
, NSEC_PER_USEC
);
7805 return rt_period_us
;
7807 #endif /* CONFIG_RT_GROUP_SCHED */
7809 #ifdef CONFIG_RT_GROUP_SCHED
7810 static int sched_rt_global_constraints(void)
7814 mutex_lock(&rt_constraints_mutex
);
7815 read_lock(&tasklist_lock
);
7816 ret
= __rt_schedulable(NULL
, 0, 0);
7817 read_unlock(&tasklist_lock
);
7818 mutex_unlock(&rt_constraints_mutex
);
7823 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7825 /* Don't accept realtime tasks when there is no way for them to run */
7826 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7832 #else /* !CONFIG_RT_GROUP_SCHED */
7833 static int sched_rt_global_constraints(void)
7835 unsigned long flags
;
7838 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7839 for_each_possible_cpu(i
) {
7840 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7842 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7843 rt_rq
->rt_runtime
= global_rt_runtime();
7844 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7846 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7850 #endif /* CONFIG_RT_GROUP_SCHED */
7852 static int sched_dl_global_constraints(void)
7854 u64 runtime
= global_rt_runtime();
7855 u64 period
= global_rt_period();
7856 u64 new_bw
= to_ratio(period
, runtime
);
7859 unsigned long flags
;
7862 * Here we want to check the bandwidth not being set to some
7863 * value smaller than the currently allocated bandwidth in
7864 * any of the root_domains.
7866 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7867 * cycling on root_domains... Discussion on different/better
7868 * solutions is welcome!
7870 for_each_possible_cpu(cpu
) {
7871 rcu_read_lock_sched();
7872 dl_b
= dl_bw_of(cpu
);
7874 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7875 if (new_bw
< dl_b
->total_bw
)
7877 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7879 rcu_read_unlock_sched();
7888 static void sched_dl_do_global(void)
7893 unsigned long flags
;
7895 def_dl_bandwidth
.dl_period
= global_rt_period();
7896 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7898 if (global_rt_runtime() != RUNTIME_INF
)
7899 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7902 * FIXME: As above...
7904 for_each_possible_cpu(cpu
) {
7905 rcu_read_lock_sched();
7906 dl_b
= dl_bw_of(cpu
);
7908 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7910 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7912 rcu_read_unlock_sched();
7916 static int sched_rt_global_validate(void)
7918 if (sysctl_sched_rt_period
<= 0)
7921 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7922 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7928 static void sched_rt_do_global(void)
7930 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7931 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7934 int sched_rt_handler(struct ctl_table
*table
, int write
,
7935 void __user
*buffer
, size_t *lenp
,
7938 int old_period
, old_runtime
;
7939 static DEFINE_MUTEX(mutex
);
7943 old_period
= sysctl_sched_rt_period
;
7944 old_runtime
= sysctl_sched_rt_runtime
;
7946 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7948 if (!ret
&& write
) {
7949 ret
= sched_rt_global_validate();
7953 ret
= sched_rt_global_constraints();
7957 ret
= sched_dl_global_constraints();
7961 sched_rt_do_global();
7962 sched_dl_do_global();
7966 sysctl_sched_rt_period
= old_period
;
7967 sysctl_sched_rt_runtime
= old_runtime
;
7969 mutex_unlock(&mutex
);
7974 int sched_rr_handler(struct ctl_table
*table
, int write
,
7975 void __user
*buffer
, size_t *lenp
,
7979 static DEFINE_MUTEX(mutex
);
7982 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7983 /* make sure that internally we keep jiffies */
7984 /* also, writing zero resets timeslice to default */
7985 if (!ret
&& write
) {
7986 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7987 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7989 mutex_unlock(&mutex
);
7993 #ifdef CONFIG_CGROUP_SCHED
7995 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7997 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8000 static struct cgroup_subsys_state
*
8001 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8003 struct task_group
*parent
= css_tg(parent_css
);
8004 struct task_group
*tg
;
8007 /* This is early initialization for the top cgroup */
8008 return &root_task_group
.css
;
8011 tg
= sched_create_group(parent
);
8013 return ERR_PTR(-ENOMEM
);
8018 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8020 struct task_group
*tg
= css_tg(css
);
8021 struct task_group
*parent
= css_tg(css
->parent
);
8024 sched_online_group(tg
, parent
);
8028 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8030 struct task_group
*tg
= css_tg(css
);
8032 sched_destroy_group(tg
);
8035 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8037 struct task_group
*tg
= css_tg(css
);
8039 sched_offline_group(tg
);
8042 static void cpu_cgroup_fork(struct task_struct
*task
)
8044 sched_move_task(task
);
8047 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
8048 struct cgroup_taskset
*tset
)
8050 struct task_struct
*task
;
8052 cgroup_taskset_for_each(task
, tset
) {
8053 #ifdef CONFIG_RT_GROUP_SCHED
8054 if (!sched_rt_can_attach(css_tg(css
), task
))
8057 /* We don't support RT-tasks being in separate groups */
8058 if (task
->sched_class
!= &fair_sched_class
)
8065 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8066 struct cgroup_taskset
*tset
)
8068 struct task_struct
*task
;
8070 cgroup_taskset_for_each(task
, tset
)
8071 sched_move_task(task
);
8074 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8075 struct cgroup_subsys_state
*old_css
,
8076 struct task_struct
*task
)
8079 * cgroup_exit() is called in the copy_process() failure path.
8080 * Ignore this case since the task hasn't ran yet, this avoids
8081 * trying to poke a half freed task state from generic code.
8083 if (!(task
->flags
& PF_EXITING
))
8086 sched_move_task(task
);
8089 #ifdef CONFIG_FAIR_GROUP_SCHED
8090 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8091 struct cftype
*cftype
, u64 shareval
)
8093 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8096 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8099 struct task_group
*tg
= css_tg(css
);
8101 return (u64
) scale_load_down(tg
->shares
);
8104 #ifdef CONFIG_CFS_BANDWIDTH
8105 static DEFINE_MUTEX(cfs_constraints_mutex
);
8107 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8108 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8110 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8112 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8114 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8115 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8117 if (tg
== &root_task_group
)
8121 * Ensure we have at some amount of bandwidth every period. This is
8122 * to prevent reaching a state of large arrears when throttled via
8123 * entity_tick() resulting in prolonged exit starvation.
8125 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8129 * Likewise, bound things on the otherside by preventing insane quota
8130 * periods. This also allows us to normalize in computing quota
8133 if (period
> max_cfs_quota_period
)
8137 * Prevent race between setting of cfs_rq->runtime_enabled and
8138 * unthrottle_offline_cfs_rqs().
8141 mutex_lock(&cfs_constraints_mutex
);
8142 ret
= __cfs_schedulable(tg
, period
, quota
);
8146 runtime_enabled
= quota
!= RUNTIME_INF
;
8147 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8149 * If we need to toggle cfs_bandwidth_used, off->on must occur
8150 * before making related changes, and on->off must occur afterwards
8152 if (runtime_enabled
&& !runtime_was_enabled
)
8153 cfs_bandwidth_usage_inc();
8154 raw_spin_lock_irq(&cfs_b
->lock
);
8155 cfs_b
->period
= ns_to_ktime(period
);
8156 cfs_b
->quota
= quota
;
8158 __refill_cfs_bandwidth_runtime(cfs_b
);
8159 /* restart the period timer (if active) to handle new period expiry */
8160 if (runtime_enabled
&& cfs_b
->timer_active
) {
8161 /* force a reprogram */
8162 __start_cfs_bandwidth(cfs_b
, true);
8164 raw_spin_unlock_irq(&cfs_b
->lock
);
8166 for_each_online_cpu(i
) {
8167 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8168 struct rq
*rq
= cfs_rq
->rq
;
8170 raw_spin_lock_irq(&rq
->lock
);
8171 cfs_rq
->runtime_enabled
= runtime_enabled
;
8172 cfs_rq
->runtime_remaining
= 0;
8174 if (cfs_rq
->throttled
)
8175 unthrottle_cfs_rq(cfs_rq
);
8176 raw_spin_unlock_irq(&rq
->lock
);
8178 if (runtime_was_enabled
&& !runtime_enabled
)
8179 cfs_bandwidth_usage_dec();
8181 mutex_unlock(&cfs_constraints_mutex
);
8187 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8191 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8192 if (cfs_quota_us
< 0)
8193 quota
= RUNTIME_INF
;
8195 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8197 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8200 long tg_get_cfs_quota(struct task_group
*tg
)
8204 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8207 quota_us
= tg
->cfs_bandwidth
.quota
;
8208 do_div(quota_us
, NSEC_PER_USEC
);
8213 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8217 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8218 quota
= tg
->cfs_bandwidth
.quota
;
8220 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8223 long tg_get_cfs_period(struct task_group
*tg
)
8227 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8228 do_div(cfs_period_us
, NSEC_PER_USEC
);
8230 return cfs_period_us
;
8233 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8236 return tg_get_cfs_quota(css_tg(css
));
8239 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8240 struct cftype
*cftype
, s64 cfs_quota_us
)
8242 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8245 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8248 return tg_get_cfs_period(css_tg(css
));
8251 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8252 struct cftype
*cftype
, u64 cfs_period_us
)
8254 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8257 struct cfs_schedulable_data
{
8258 struct task_group
*tg
;
8263 * normalize group quota/period to be quota/max_period
8264 * note: units are usecs
8266 static u64
normalize_cfs_quota(struct task_group
*tg
,
8267 struct cfs_schedulable_data
*d
)
8275 period
= tg_get_cfs_period(tg
);
8276 quota
= tg_get_cfs_quota(tg
);
8279 /* note: these should typically be equivalent */
8280 if (quota
== RUNTIME_INF
|| quota
== -1)
8283 return to_ratio(period
, quota
);
8286 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8288 struct cfs_schedulable_data
*d
= data
;
8289 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8290 s64 quota
= 0, parent_quota
= -1;
8293 quota
= RUNTIME_INF
;
8295 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8297 quota
= normalize_cfs_quota(tg
, d
);
8298 parent_quota
= parent_b
->hierarchical_quota
;
8301 * ensure max(child_quota) <= parent_quota, inherit when no
8304 if (quota
== RUNTIME_INF
)
8305 quota
= parent_quota
;
8306 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8309 cfs_b
->hierarchical_quota
= quota
;
8314 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8317 struct cfs_schedulable_data data
= {
8323 if (quota
!= RUNTIME_INF
) {
8324 do_div(data
.period
, NSEC_PER_USEC
);
8325 do_div(data
.quota
, NSEC_PER_USEC
);
8329 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8335 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8337 struct task_group
*tg
= css_tg(seq_css(sf
));
8338 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8340 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8341 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8342 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8346 #endif /* CONFIG_CFS_BANDWIDTH */
8347 #endif /* CONFIG_FAIR_GROUP_SCHED */
8349 #ifdef CONFIG_RT_GROUP_SCHED
8350 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8351 struct cftype
*cft
, s64 val
)
8353 return sched_group_set_rt_runtime(css_tg(css
), val
);
8356 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8359 return sched_group_rt_runtime(css_tg(css
));
8362 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8363 struct cftype
*cftype
, u64 rt_period_us
)
8365 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8368 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8371 return sched_group_rt_period(css_tg(css
));
8373 #endif /* CONFIG_RT_GROUP_SCHED */
8375 static struct cftype cpu_files
[] = {
8376 #ifdef CONFIG_FAIR_GROUP_SCHED
8379 .read_u64
= cpu_shares_read_u64
,
8380 .write_u64
= cpu_shares_write_u64
,
8383 #ifdef CONFIG_CFS_BANDWIDTH
8385 .name
= "cfs_quota_us",
8386 .read_s64
= cpu_cfs_quota_read_s64
,
8387 .write_s64
= cpu_cfs_quota_write_s64
,
8390 .name
= "cfs_period_us",
8391 .read_u64
= cpu_cfs_period_read_u64
,
8392 .write_u64
= cpu_cfs_period_write_u64
,
8396 .seq_show
= cpu_stats_show
,
8399 #ifdef CONFIG_RT_GROUP_SCHED
8401 .name
= "rt_runtime_us",
8402 .read_s64
= cpu_rt_runtime_read
,
8403 .write_s64
= cpu_rt_runtime_write
,
8406 .name
= "rt_period_us",
8407 .read_u64
= cpu_rt_period_read_uint
,
8408 .write_u64
= cpu_rt_period_write_uint
,
8414 struct cgroup_subsys cpu_cgrp_subsys
= {
8415 .css_alloc
= cpu_cgroup_css_alloc
,
8416 .css_free
= cpu_cgroup_css_free
,
8417 .css_online
= cpu_cgroup_css_online
,
8418 .css_offline
= cpu_cgroup_css_offline
,
8419 .fork
= cpu_cgroup_fork
,
8420 .can_attach
= cpu_cgroup_can_attach
,
8421 .attach
= cpu_cgroup_attach
,
8422 .exit
= cpu_cgroup_exit
,
8423 .legacy_cftypes
= cpu_files
,
8427 #endif /* CONFIG_CGROUP_SCHED */
8429 void dump_cpu_task(int cpu
)
8431 pr_info("Task dump for CPU %d:\n", cpu
);
8432 sched_show_task(cpu_curr(cpu
));