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 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic
);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic
);
109 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
112 ktime_t soft
, hard
, now
;
115 if (hrtimer_active(period_timer
))
118 now
= hrtimer_cb_get_time(period_timer
);
119 hrtimer_forward(period_timer
, now
, period
);
121 soft
= hrtimer_get_softexpires(period_timer
);
122 hard
= hrtimer_get_expires(period_timer
);
123 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
124 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
125 HRTIMER_MODE_ABS_PINNED
, 0);
129 DEFINE_MUTEX(sched_domains_mutex
);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
132 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
134 void update_rq_clock(struct rq
*rq
)
138 if (rq
->skip_clock_update
> 0)
141 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
143 update_rq_clock_task(rq
, delta
);
147 * Debugging: various feature bits
150 #define SCHED_FEAT(name, enabled) \
151 (1UL << __SCHED_FEAT_##name) * enabled |
153 const_debug
unsigned int sysctl_sched_features
=
154 #include "features.h"
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled) \
163 static const char * const sched_feat_names
[] = {
164 #include "features.h"
169 static int sched_feat_show(struct seq_file
*m
, void *v
)
173 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
174 if (!(sysctl_sched_features
& (1UL << i
)))
176 seq_printf(m
, "%s ", sched_feat_names
[i
]);
183 #ifdef HAVE_JUMP_LABEL
185 #define jump_label_key__true STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
188 #define SCHED_FEAT(name, enabled) \
189 jump_label_key__##enabled ,
191 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
192 #include "features.h"
197 static void sched_feat_disable(int i
)
199 if (static_key_enabled(&sched_feat_keys
[i
]))
200 static_key_slow_dec(&sched_feat_keys
[i
]);
203 static void sched_feat_enable(int i
)
205 if (!static_key_enabled(&sched_feat_keys
[i
]))
206 static_key_slow_inc(&sched_feat_keys
[i
]);
209 static void sched_feat_disable(int i
) { };
210 static void sched_feat_enable(int i
) { };
211 #endif /* HAVE_JUMP_LABEL */
213 static int sched_feat_set(char *cmp
)
218 if (strncmp(cmp
, "NO_", 3) == 0) {
223 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
224 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
226 sysctl_sched_features
&= ~(1UL << i
);
227 sched_feat_disable(i
);
229 sysctl_sched_features
|= (1UL << i
);
230 sched_feat_enable(i
);
240 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
241 size_t cnt
, loff_t
*ppos
)
250 if (copy_from_user(&buf
, ubuf
, cnt
))
256 i
= sched_feat_set(cmp
);
257 if (i
== __SCHED_FEAT_NR
)
265 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
267 return single_open(filp
, sched_feat_show
, NULL
);
270 static const struct file_operations sched_feat_fops
= {
271 .open
= sched_feat_open
,
272 .write
= sched_feat_write
,
275 .release
= single_release
,
278 static __init
int sched_init_debug(void)
280 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
285 late_initcall(sched_init_debug
);
286 #endif /* CONFIG_SCHED_DEBUG */
289 * Number of tasks to iterate in a single balance run.
290 * Limited because this is done with IRQs disabled.
292 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
295 * period over which we average the RT time consumption, measured
300 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
303 * period over which we measure -rt task cpu usage in us.
306 unsigned int sysctl_sched_rt_period
= 1000000;
308 __read_mostly
int scheduler_running
;
311 * part of the period that we allow rt tasks to run in us.
314 int sysctl_sched_rt_runtime
= 950000;
317 * __task_rq_lock - lock the rq @p resides on.
319 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
324 lockdep_assert_held(&p
->pi_lock
);
328 raw_spin_lock(&rq
->lock
);
329 if (likely(rq
== task_rq(p
)))
331 raw_spin_unlock(&rq
->lock
);
336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
338 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
339 __acquires(p
->pi_lock
)
345 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
347 raw_spin_lock(&rq
->lock
);
348 if (likely(rq
== task_rq(p
)))
350 raw_spin_unlock(&rq
->lock
);
351 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
355 static void __task_rq_unlock(struct rq
*rq
)
358 raw_spin_unlock(&rq
->lock
);
362 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
364 __releases(p
->pi_lock
)
366 raw_spin_unlock(&rq
->lock
);
367 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
371 * this_rq_lock - lock this runqueue and disable interrupts.
373 static struct rq
*this_rq_lock(void)
380 raw_spin_lock(&rq
->lock
);
385 #ifdef CONFIG_SCHED_HRTICK
387 * Use HR-timers to deliver accurate preemption points.
390 static void hrtick_clear(struct rq
*rq
)
392 if (hrtimer_active(&rq
->hrtick_timer
))
393 hrtimer_cancel(&rq
->hrtick_timer
);
397 * High-resolution timer tick.
398 * Runs from hardirq context with interrupts disabled.
400 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
402 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
404 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
406 raw_spin_lock(&rq
->lock
);
408 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
409 raw_spin_unlock(&rq
->lock
);
411 return HRTIMER_NORESTART
;
416 static int __hrtick_restart(struct rq
*rq
)
418 struct hrtimer
*timer
= &rq
->hrtick_timer
;
419 ktime_t time
= hrtimer_get_softexpires(timer
);
421 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
425 * called from hardirq (IPI) context
427 static void __hrtick_start(void *arg
)
431 raw_spin_lock(&rq
->lock
);
432 __hrtick_restart(rq
);
433 rq
->hrtick_csd_pending
= 0;
434 raw_spin_unlock(&rq
->lock
);
438 * Called to set the hrtick timer state.
440 * called with rq->lock held and irqs disabled
442 void hrtick_start(struct rq
*rq
, u64 delay
)
444 struct hrtimer
*timer
= &rq
->hrtick_timer
;
445 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
447 hrtimer_set_expires(timer
, time
);
449 if (rq
== this_rq()) {
450 __hrtick_restart(rq
);
451 } else if (!rq
->hrtick_csd_pending
) {
452 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
453 rq
->hrtick_csd_pending
= 1;
458 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
460 int cpu
= (int)(long)hcpu
;
463 case CPU_UP_CANCELED
:
464 case CPU_UP_CANCELED_FROZEN
:
465 case CPU_DOWN_PREPARE
:
466 case CPU_DOWN_PREPARE_FROZEN
:
468 case CPU_DEAD_FROZEN
:
469 hrtick_clear(cpu_rq(cpu
));
476 static __init
void init_hrtick(void)
478 hotcpu_notifier(hotplug_hrtick
, 0);
482 * Called to set the hrtick timer state.
484 * called with rq->lock held and irqs disabled
486 void hrtick_start(struct rq
*rq
, u64 delay
)
488 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
489 HRTIMER_MODE_REL_PINNED
, 0);
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SMP */
497 static void init_rq_hrtick(struct rq
*rq
)
500 rq
->hrtick_csd_pending
= 0;
502 rq
->hrtick_csd
.flags
= 0;
503 rq
->hrtick_csd
.func
= __hrtick_start
;
504 rq
->hrtick_csd
.info
= rq
;
507 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
508 rq
->hrtick_timer
.function
= hrtick
;
510 #else /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq
*rq
)
515 static inline void init_rq_hrtick(struct rq
*rq
)
519 static inline void init_hrtick(void)
522 #endif /* CONFIG_SCHED_HRTICK */
525 * cmpxchg based fetch_or, macro so it works for different integer types
527 #define fetch_or(ptr, val) \
528 ({ typeof(*(ptr)) __old, __val = *(ptr); \
530 __old = cmpxchg((ptr), __val, __val | (val)); \
531 if (__old == __val) \
538 #ifdef TIF_POLLING_NRFLAG
540 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
541 * this avoids any races wrt polling state changes and thereby avoids
544 static bool set_nr_and_not_polling(struct task_struct
*p
)
546 struct thread_info
*ti
= task_thread_info(p
);
547 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
550 static bool set_nr_and_not_polling(struct task_struct
*p
)
552 set_tsk_need_resched(p
);
558 * resched_task - mark a task 'to be rescheduled now'.
560 * On UP this means the setting of the need_resched flag, on SMP it
561 * might also involve a cross-CPU call to trigger the scheduler on
564 void resched_task(struct task_struct
*p
)
568 lockdep_assert_held(&task_rq(p
)->lock
);
570 if (test_tsk_need_resched(p
))
575 if (cpu
== smp_processor_id()) {
576 set_tsk_need_resched(p
);
577 set_preempt_need_resched();
581 if (set_nr_and_not_polling(p
))
582 smp_send_reschedule(cpu
);
585 void resched_cpu(int cpu
)
587 struct rq
*rq
= cpu_rq(cpu
);
590 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
592 resched_task(cpu_curr(cpu
));
593 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
597 #ifdef CONFIG_NO_HZ_COMMON
599 * In the semi idle case, use the nearest busy cpu for migrating timers
600 * from an idle cpu. This is good for power-savings.
602 * We don't do similar optimization for completely idle system, as
603 * selecting an idle cpu will add more delays to the timers than intended
604 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
606 int get_nohz_timer_target(int pinned
)
608 int cpu
= smp_processor_id();
610 struct sched_domain
*sd
;
612 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
616 for_each_domain(cpu
, sd
) {
617 for_each_cpu(i
, sched_domain_span(sd
)) {
629 * When add_timer_on() enqueues a timer into the timer wheel of an
630 * idle CPU then this timer might expire before the next timer event
631 * which is scheduled to wake up that CPU. In case of a completely
632 * idle system the next event might even be infinite time into the
633 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
634 * leaves the inner idle loop so the newly added timer is taken into
635 * account when the CPU goes back to idle and evaluates the timer
636 * wheel for the next timer event.
638 static void wake_up_idle_cpu(int cpu
)
640 struct rq
*rq
= cpu_rq(cpu
);
642 if (cpu
== smp_processor_id())
646 * This is safe, as this function is called with the timer
647 * wheel base lock of (cpu) held. When the CPU is on the way
648 * to idle and has not yet set rq->curr to idle then it will
649 * be serialized on the timer wheel base lock and take the new
650 * timer into account automatically.
652 if (rq
->curr
!= rq
->idle
)
656 * We can set TIF_RESCHED on the idle task of the other CPU
657 * lockless. The worst case is that the other CPU runs the
658 * idle task through an additional NOOP schedule()
660 set_tsk_need_resched(rq
->idle
);
662 /* NEED_RESCHED must be visible before we test polling */
664 if (!tsk_is_polling(rq
->idle
))
665 smp_send_reschedule(cpu
);
668 static bool wake_up_full_nohz_cpu(int cpu
)
670 if (tick_nohz_full_cpu(cpu
)) {
671 if (cpu
!= smp_processor_id() ||
672 tick_nohz_tick_stopped())
673 smp_send_reschedule(cpu
);
680 void wake_up_nohz_cpu(int cpu
)
682 if (!wake_up_full_nohz_cpu(cpu
))
683 wake_up_idle_cpu(cpu
);
686 static inline bool got_nohz_idle_kick(void)
688 int cpu
= smp_processor_id();
690 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
693 if (idle_cpu(cpu
) && !need_resched())
697 * We can't run Idle Load Balance on this CPU for this time so we
698 * cancel it and clear NOHZ_BALANCE_KICK
700 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
704 #else /* CONFIG_NO_HZ_COMMON */
706 static inline bool got_nohz_idle_kick(void)
711 #endif /* CONFIG_NO_HZ_COMMON */
713 #ifdef CONFIG_NO_HZ_FULL
714 bool sched_can_stop_tick(void)
720 /* Make sure rq->nr_running update is visible after the IPI */
723 /* More than one running task need preemption */
724 if (rq
->nr_running
> 1)
729 #endif /* CONFIG_NO_HZ_FULL */
731 void sched_avg_update(struct rq
*rq
)
733 s64 period
= sched_avg_period();
735 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
737 * Inline assembly required to prevent the compiler
738 * optimising this loop into a divmod call.
739 * See __iter_div_u64_rem() for another example of this.
741 asm("" : "+rm" (rq
->age_stamp
));
742 rq
->age_stamp
+= period
;
747 #endif /* CONFIG_SMP */
749 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
750 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
752 * Iterate task_group tree rooted at *from, calling @down when first entering a
753 * node and @up when leaving it for the final time.
755 * Caller must hold rcu_lock or sufficient equivalent.
757 int walk_tg_tree_from(struct task_group
*from
,
758 tg_visitor down
, tg_visitor up
, void *data
)
760 struct task_group
*parent
, *child
;
766 ret
= (*down
)(parent
, data
);
769 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
776 ret
= (*up
)(parent
, data
);
777 if (ret
|| parent
== from
)
781 parent
= parent
->parent
;
788 int tg_nop(struct task_group
*tg
, void *data
)
794 static void set_load_weight(struct task_struct
*p
)
796 int prio
= p
->static_prio
- MAX_RT_PRIO
;
797 struct load_weight
*load
= &p
->se
.load
;
800 * SCHED_IDLE tasks get minimal weight:
802 if (p
->policy
== SCHED_IDLE
) {
803 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
804 load
->inv_weight
= WMULT_IDLEPRIO
;
808 load
->weight
= scale_load(prio_to_weight
[prio
]);
809 load
->inv_weight
= prio_to_wmult
[prio
];
812 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
815 sched_info_queued(rq
, p
);
816 p
->sched_class
->enqueue_task(rq
, p
, flags
);
819 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
822 sched_info_dequeued(rq
, p
);
823 p
->sched_class
->dequeue_task(rq
, p
, flags
);
826 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
828 if (task_contributes_to_load(p
))
829 rq
->nr_uninterruptible
--;
831 enqueue_task(rq
, p
, flags
);
834 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
836 if (task_contributes_to_load(p
))
837 rq
->nr_uninterruptible
++;
839 dequeue_task(rq
, p
, flags
);
842 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
845 * In theory, the compile should just see 0 here, and optimize out the call
846 * to sched_rt_avg_update. But I don't trust it...
848 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
849 s64 steal
= 0, irq_delta
= 0;
851 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
852 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
855 * Since irq_time is only updated on {soft,}irq_exit, we might run into
856 * this case when a previous update_rq_clock() happened inside a
859 * When this happens, we stop ->clock_task and only update the
860 * prev_irq_time stamp to account for the part that fit, so that a next
861 * update will consume the rest. This ensures ->clock_task is
864 * It does however cause some slight miss-attribution of {soft,}irq
865 * time, a more accurate solution would be to update the irq_time using
866 * the current rq->clock timestamp, except that would require using
869 if (irq_delta
> delta
)
872 rq
->prev_irq_time
+= irq_delta
;
875 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
876 if (static_key_false((¶virt_steal_rq_enabled
))) {
877 steal
= paravirt_steal_clock(cpu_of(rq
));
878 steal
-= rq
->prev_steal_time_rq
;
880 if (unlikely(steal
> delta
))
883 rq
->prev_steal_time_rq
+= steal
;
888 rq
->clock_task
+= delta
;
890 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
891 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
892 sched_rt_avg_update(rq
, irq_delta
+ steal
);
896 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
898 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
899 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
903 * Make it appear like a SCHED_FIFO task, its something
904 * userspace knows about and won't get confused about.
906 * Also, it will make PI more or less work without too
907 * much confusion -- but then, stop work should not
908 * rely on PI working anyway.
910 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
912 stop
->sched_class
= &stop_sched_class
;
915 cpu_rq(cpu
)->stop
= stop
;
919 * Reset it back to a normal scheduling class so that
920 * it can die in pieces.
922 old_stop
->sched_class
= &rt_sched_class
;
927 * __normal_prio - return the priority that is based on the static prio
929 static inline int __normal_prio(struct task_struct
*p
)
931 return p
->static_prio
;
935 * Calculate the expected normal priority: i.e. priority
936 * without taking RT-inheritance into account. Might be
937 * boosted by interactivity modifiers. Changes upon fork,
938 * setprio syscalls, and whenever the interactivity
939 * estimator recalculates.
941 static inline int normal_prio(struct task_struct
*p
)
945 if (task_has_dl_policy(p
))
946 prio
= MAX_DL_PRIO
-1;
947 else if (task_has_rt_policy(p
))
948 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
950 prio
= __normal_prio(p
);
955 * Calculate the current priority, i.e. the priority
956 * taken into account by the scheduler. This value might
957 * be boosted by RT tasks, or might be boosted by
958 * interactivity modifiers. Will be RT if the task got
959 * RT-boosted. If not then it returns p->normal_prio.
961 static int effective_prio(struct task_struct
*p
)
963 p
->normal_prio
= normal_prio(p
);
965 * If we are RT tasks or we were boosted to RT priority,
966 * keep the priority unchanged. Otherwise, update priority
967 * to the normal priority:
969 if (!rt_prio(p
->prio
))
970 return p
->normal_prio
;
975 * task_curr - is this task currently executing on a CPU?
976 * @p: the task in question.
978 * Return: 1 if the task is currently executing. 0 otherwise.
980 inline int task_curr(const struct task_struct
*p
)
982 return cpu_curr(task_cpu(p
)) == p
;
985 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
986 const struct sched_class
*prev_class
,
989 if (prev_class
!= p
->sched_class
) {
990 if (prev_class
->switched_from
)
991 prev_class
->switched_from(rq
, p
);
992 p
->sched_class
->switched_to(rq
, p
);
993 } else if (oldprio
!= p
->prio
|| dl_task(p
))
994 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
997 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
999 const struct sched_class
*class;
1001 if (p
->sched_class
== rq
->curr
->sched_class
) {
1002 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1004 for_each_class(class) {
1005 if (class == rq
->curr
->sched_class
)
1007 if (class == p
->sched_class
) {
1008 resched_task(rq
->curr
);
1015 * A queue event has occurred, and we're going to schedule. In
1016 * this case, we can save a useless back to back clock update.
1018 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
1019 rq
->skip_clock_update
= 1;
1023 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1025 #ifdef CONFIG_SCHED_DEBUG
1027 * We should never call set_task_cpu() on a blocked task,
1028 * ttwu() will sort out the placement.
1030 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1031 !(task_preempt_count(p
) & PREEMPT_ACTIVE
));
1033 #ifdef CONFIG_LOCKDEP
1035 * The caller should hold either p->pi_lock or rq->lock, when changing
1036 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1038 * sched_move_task() holds both and thus holding either pins the cgroup,
1041 * Furthermore, all task_rq users should acquire both locks, see
1044 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1045 lockdep_is_held(&task_rq(p
)->lock
)));
1049 trace_sched_migrate_task(p
, new_cpu
);
1051 if (task_cpu(p
) != new_cpu
) {
1052 if (p
->sched_class
->migrate_task_rq
)
1053 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1054 p
->se
.nr_migrations
++;
1055 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1058 __set_task_cpu(p
, new_cpu
);
1061 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1064 struct rq
*src_rq
, *dst_rq
;
1066 src_rq
= task_rq(p
);
1067 dst_rq
= cpu_rq(cpu
);
1069 deactivate_task(src_rq
, p
, 0);
1070 set_task_cpu(p
, cpu
);
1071 activate_task(dst_rq
, p
, 0);
1072 check_preempt_curr(dst_rq
, p
, 0);
1075 * Task isn't running anymore; make it appear like we migrated
1076 * it before it went to sleep. This means on wakeup we make the
1077 * previous cpu our targer instead of where it really is.
1083 struct migration_swap_arg
{
1084 struct task_struct
*src_task
, *dst_task
;
1085 int src_cpu
, dst_cpu
;
1088 static int migrate_swap_stop(void *data
)
1090 struct migration_swap_arg
*arg
= data
;
1091 struct rq
*src_rq
, *dst_rq
;
1094 src_rq
= cpu_rq(arg
->src_cpu
);
1095 dst_rq
= cpu_rq(arg
->dst_cpu
);
1097 double_raw_lock(&arg
->src_task
->pi_lock
,
1098 &arg
->dst_task
->pi_lock
);
1099 double_rq_lock(src_rq
, dst_rq
);
1100 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1103 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1106 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1109 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1112 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1113 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1118 double_rq_unlock(src_rq
, dst_rq
);
1119 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1120 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1126 * Cross migrate two tasks
1128 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1130 struct migration_swap_arg arg
;
1133 arg
= (struct migration_swap_arg
){
1135 .src_cpu
= task_cpu(cur
),
1137 .dst_cpu
= task_cpu(p
),
1140 if (arg
.src_cpu
== arg
.dst_cpu
)
1144 * These three tests are all lockless; this is OK since all of them
1145 * will be re-checked with proper locks held further down the line.
1147 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1150 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1153 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1156 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1157 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1163 struct migration_arg
{
1164 struct task_struct
*task
;
1168 static int migration_cpu_stop(void *data
);
1171 * wait_task_inactive - wait for a thread to unschedule.
1173 * If @match_state is nonzero, it's the @p->state value just checked and
1174 * not expected to change. If it changes, i.e. @p might have woken up,
1175 * then return zero. When we succeed in waiting for @p to be off its CPU,
1176 * we return a positive number (its total switch count). If a second call
1177 * a short while later returns the same number, the caller can be sure that
1178 * @p has remained unscheduled the whole time.
1180 * The caller must ensure that the task *will* unschedule sometime soon,
1181 * else this function might spin for a *long* time. This function can't
1182 * be called with interrupts off, or it may introduce deadlock with
1183 * smp_call_function() if an IPI is sent by the same process we are
1184 * waiting to become inactive.
1186 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1188 unsigned long flags
;
1195 * We do the initial early heuristics without holding
1196 * any task-queue locks at all. We'll only try to get
1197 * the runqueue lock when things look like they will
1203 * If the task is actively running on another CPU
1204 * still, just relax and busy-wait without holding
1207 * NOTE! Since we don't hold any locks, it's not
1208 * even sure that "rq" stays as the right runqueue!
1209 * But we don't care, since "task_running()" will
1210 * return false if the runqueue has changed and p
1211 * is actually now running somewhere else!
1213 while (task_running(rq
, p
)) {
1214 if (match_state
&& unlikely(p
->state
!= match_state
))
1220 * Ok, time to look more closely! We need the rq
1221 * lock now, to be *sure*. If we're wrong, we'll
1222 * just go back and repeat.
1224 rq
= task_rq_lock(p
, &flags
);
1225 trace_sched_wait_task(p
);
1226 running
= task_running(rq
, p
);
1229 if (!match_state
|| p
->state
== match_state
)
1230 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1231 task_rq_unlock(rq
, p
, &flags
);
1234 * If it changed from the expected state, bail out now.
1236 if (unlikely(!ncsw
))
1240 * Was it really running after all now that we
1241 * checked with the proper locks actually held?
1243 * Oops. Go back and try again..
1245 if (unlikely(running
)) {
1251 * It's not enough that it's not actively running,
1252 * it must be off the runqueue _entirely_, and not
1255 * So if it was still runnable (but just not actively
1256 * running right now), it's preempted, and we should
1257 * yield - it could be a while.
1259 if (unlikely(on_rq
)) {
1260 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1262 set_current_state(TASK_UNINTERRUPTIBLE
);
1263 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1268 * Ahh, all good. It wasn't running, and it wasn't
1269 * runnable, which means that it will never become
1270 * running in the future either. We're all done!
1279 * kick_process - kick a running thread to enter/exit the kernel
1280 * @p: the to-be-kicked thread
1282 * Cause a process which is running on another CPU to enter
1283 * kernel-mode, without any delay. (to get signals handled.)
1285 * NOTE: this function doesn't have to take the runqueue lock,
1286 * because all it wants to ensure is that the remote task enters
1287 * the kernel. If the IPI races and the task has been migrated
1288 * to another CPU then no harm is done and the purpose has been
1291 void kick_process(struct task_struct
*p
)
1297 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1298 smp_send_reschedule(cpu
);
1301 EXPORT_SYMBOL_GPL(kick_process
);
1302 #endif /* CONFIG_SMP */
1306 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1308 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1310 int nid
= cpu_to_node(cpu
);
1311 const struct cpumask
*nodemask
= NULL
;
1312 enum { cpuset
, possible
, fail
} state
= cpuset
;
1316 * If the node that the cpu is on has been offlined, cpu_to_node()
1317 * will return -1. There is no cpu on the node, and we should
1318 * select the cpu on the other node.
1321 nodemask
= cpumask_of_node(nid
);
1323 /* Look for allowed, online CPU in same node. */
1324 for_each_cpu(dest_cpu
, nodemask
) {
1325 if (!cpu_online(dest_cpu
))
1327 if (!cpu_active(dest_cpu
))
1329 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1335 /* Any allowed, online CPU? */
1336 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1337 if (!cpu_online(dest_cpu
))
1339 if (!cpu_active(dest_cpu
))
1346 /* No more Mr. Nice Guy. */
1347 cpuset_cpus_allowed_fallback(p
);
1352 do_set_cpus_allowed(p
, cpu_possible_mask
);
1363 if (state
!= cpuset
) {
1365 * Don't tell them about moving exiting tasks or
1366 * kernel threads (both mm NULL), since they never
1369 if (p
->mm
&& printk_ratelimit()) {
1370 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1371 task_pid_nr(p
), p
->comm
, cpu
);
1379 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1382 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1384 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1387 * In order not to call set_task_cpu() on a blocking task we need
1388 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1391 * Since this is common to all placement strategies, this lives here.
1393 * [ this allows ->select_task() to simply return task_cpu(p) and
1394 * not worry about this generic constraint ]
1396 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1398 cpu
= select_fallback_rq(task_cpu(p
), p
);
1403 static void update_avg(u64
*avg
, u64 sample
)
1405 s64 diff
= sample
- *avg
;
1411 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1413 #ifdef CONFIG_SCHEDSTATS
1414 struct rq
*rq
= this_rq();
1417 int this_cpu
= smp_processor_id();
1419 if (cpu
== this_cpu
) {
1420 schedstat_inc(rq
, ttwu_local
);
1421 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1423 struct sched_domain
*sd
;
1425 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1427 for_each_domain(this_cpu
, sd
) {
1428 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1429 schedstat_inc(sd
, ttwu_wake_remote
);
1436 if (wake_flags
& WF_MIGRATED
)
1437 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1439 #endif /* CONFIG_SMP */
1441 schedstat_inc(rq
, ttwu_count
);
1442 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1444 if (wake_flags
& WF_SYNC
)
1445 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1447 #endif /* CONFIG_SCHEDSTATS */
1450 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1452 activate_task(rq
, p
, en_flags
);
1455 /* if a worker is waking up, notify workqueue */
1456 if (p
->flags
& PF_WQ_WORKER
)
1457 wq_worker_waking_up(p
, cpu_of(rq
));
1461 * Mark the task runnable and perform wakeup-preemption.
1464 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1466 check_preempt_curr(rq
, p
, wake_flags
);
1467 trace_sched_wakeup(p
, true);
1469 p
->state
= TASK_RUNNING
;
1471 if (p
->sched_class
->task_woken
)
1472 p
->sched_class
->task_woken(rq
, p
);
1474 if (rq
->idle_stamp
) {
1475 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1476 u64 max
= 2*rq
->max_idle_balance_cost
;
1478 update_avg(&rq
->avg_idle
, delta
);
1480 if (rq
->avg_idle
> max
)
1489 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1492 if (p
->sched_contributes_to_load
)
1493 rq
->nr_uninterruptible
--;
1496 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1497 ttwu_do_wakeup(rq
, p
, wake_flags
);
1501 * Called in case the task @p isn't fully descheduled from its runqueue,
1502 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1503 * since all we need to do is flip p->state to TASK_RUNNING, since
1504 * the task is still ->on_rq.
1506 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1511 rq
= __task_rq_lock(p
);
1513 /* check_preempt_curr() may use rq clock */
1514 update_rq_clock(rq
);
1515 ttwu_do_wakeup(rq
, p
, wake_flags
);
1518 __task_rq_unlock(rq
);
1524 static void sched_ttwu_pending(void)
1526 struct rq
*rq
= this_rq();
1527 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1528 struct task_struct
*p
;
1530 raw_spin_lock(&rq
->lock
);
1533 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1534 llist
= llist_next(llist
);
1535 ttwu_do_activate(rq
, p
, 0);
1538 raw_spin_unlock(&rq
->lock
);
1541 void scheduler_ipi(void)
1544 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1545 * TIF_NEED_RESCHED remotely (for the first time) will also send
1548 preempt_fold_need_resched();
1550 if (llist_empty(&this_rq()->wake_list
)
1551 && !tick_nohz_full_cpu(smp_processor_id())
1552 && !got_nohz_idle_kick())
1556 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1557 * traditionally all their work was done from the interrupt return
1558 * path. Now that we actually do some work, we need to make sure
1561 * Some archs already do call them, luckily irq_enter/exit nest
1564 * Arguably we should visit all archs and update all handlers,
1565 * however a fair share of IPIs are still resched only so this would
1566 * somewhat pessimize the simple resched case.
1569 tick_nohz_full_check();
1570 sched_ttwu_pending();
1573 * Check if someone kicked us for doing the nohz idle load balance.
1575 if (unlikely(got_nohz_idle_kick())) {
1576 this_rq()->idle_balance
= 1;
1577 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1582 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1584 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1585 smp_send_reschedule(cpu
);
1588 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1590 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1592 #endif /* CONFIG_SMP */
1594 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1596 struct rq
*rq
= cpu_rq(cpu
);
1598 #if defined(CONFIG_SMP)
1599 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1600 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1601 ttwu_queue_remote(p
, cpu
);
1606 raw_spin_lock(&rq
->lock
);
1607 ttwu_do_activate(rq
, p
, 0);
1608 raw_spin_unlock(&rq
->lock
);
1612 * try_to_wake_up - wake up a thread
1613 * @p: the thread to be awakened
1614 * @state: the mask of task states that can be woken
1615 * @wake_flags: wake modifier flags (WF_*)
1617 * Put it on the run-queue if it's not already there. The "current"
1618 * thread is always on the run-queue (except when the actual
1619 * re-schedule is in progress), and as such you're allowed to do
1620 * the simpler "current->state = TASK_RUNNING" to mark yourself
1621 * runnable without the overhead of this.
1623 * Return: %true if @p was woken up, %false if it was already running.
1624 * or @state didn't match @p's state.
1627 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1629 unsigned long flags
;
1630 int cpu
, success
= 0;
1633 * If we are going to wake up a thread waiting for CONDITION we
1634 * need to ensure that CONDITION=1 done by the caller can not be
1635 * reordered with p->state check below. This pairs with mb() in
1636 * set_current_state() the waiting thread does.
1638 smp_mb__before_spinlock();
1639 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1640 if (!(p
->state
& state
))
1643 success
= 1; /* we're going to change ->state */
1646 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1651 * If the owning (remote) cpu is still in the middle of schedule() with
1652 * this task as prev, wait until its done referencing the task.
1657 * Pairs with the smp_wmb() in finish_lock_switch().
1661 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1662 p
->state
= TASK_WAKING
;
1664 if (p
->sched_class
->task_waking
)
1665 p
->sched_class
->task_waking(p
);
1667 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1668 if (task_cpu(p
) != cpu
) {
1669 wake_flags
|= WF_MIGRATED
;
1670 set_task_cpu(p
, cpu
);
1672 #endif /* CONFIG_SMP */
1676 ttwu_stat(p
, cpu
, wake_flags
);
1678 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1684 * try_to_wake_up_local - try to wake up a local task with rq lock held
1685 * @p: the thread to be awakened
1687 * Put @p on the run-queue if it's not already there. The caller must
1688 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1691 static void try_to_wake_up_local(struct task_struct
*p
)
1693 struct rq
*rq
= task_rq(p
);
1695 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1696 WARN_ON_ONCE(p
== current
))
1699 lockdep_assert_held(&rq
->lock
);
1701 if (!raw_spin_trylock(&p
->pi_lock
)) {
1702 raw_spin_unlock(&rq
->lock
);
1703 raw_spin_lock(&p
->pi_lock
);
1704 raw_spin_lock(&rq
->lock
);
1707 if (!(p
->state
& TASK_NORMAL
))
1711 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1713 ttwu_do_wakeup(rq
, p
, 0);
1714 ttwu_stat(p
, smp_processor_id(), 0);
1716 raw_spin_unlock(&p
->pi_lock
);
1720 * wake_up_process - Wake up a specific process
1721 * @p: The process to be woken up.
1723 * Attempt to wake up the nominated process and move it to the set of runnable
1726 * Return: 1 if the process was woken up, 0 if it was already running.
1728 * It may be assumed that this function implies a write memory barrier before
1729 * changing the task state if and only if any tasks are woken up.
1731 int wake_up_process(struct task_struct
*p
)
1733 WARN_ON(task_is_stopped_or_traced(p
));
1734 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1736 EXPORT_SYMBOL(wake_up_process
);
1738 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1740 return try_to_wake_up(p
, state
, 0);
1744 * Perform scheduler related setup for a newly forked process p.
1745 * p is forked by current.
1747 * __sched_fork() is basic setup used by init_idle() too:
1749 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1754 p
->se
.exec_start
= 0;
1755 p
->se
.sum_exec_runtime
= 0;
1756 p
->se
.prev_sum_exec_runtime
= 0;
1757 p
->se
.nr_migrations
= 0;
1759 INIT_LIST_HEAD(&p
->se
.group_node
);
1761 #ifdef CONFIG_SCHEDSTATS
1762 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1765 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1766 hrtimer_init(&p
->dl
.dl_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1767 p
->dl
.dl_runtime
= p
->dl
.runtime
= 0;
1768 p
->dl
.dl_deadline
= p
->dl
.deadline
= 0;
1769 p
->dl
.dl_period
= 0;
1772 INIT_LIST_HEAD(&p
->rt
.run_list
);
1774 #ifdef CONFIG_PREEMPT_NOTIFIERS
1775 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1778 #ifdef CONFIG_NUMA_BALANCING
1779 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1780 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1781 p
->mm
->numa_scan_seq
= 0;
1784 if (clone_flags
& CLONE_VM
)
1785 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1787 p
->numa_preferred_nid
= -1;
1789 p
->node_stamp
= 0ULL;
1790 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1791 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1792 p
->numa_work
.next
= &p
->numa_work
;
1793 p
->numa_faults_memory
= NULL
;
1794 p
->numa_faults_buffer_memory
= NULL
;
1795 p
->last_task_numa_placement
= 0;
1796 p
->last_sum_exec_runtime
= 0;
1798 INIT_LIST_HEAD(&p
->numa_entry
);
1799 p
->numa_group
= NULL
;
1800 #endif /* CONFIG_NUMA_BALANCING */
1803 #ifdef CONFIG_NUMA_BALANCING
1804 #ifdef CONFIG_SCHED_DEBUG
1805 void set_numabalancing_state(bool enabled
)
1808 sched_feat_set("NUMA");
1810 sched_feat_set("NO_NUMA");
1813 __read_mostly
bool numabalancing_enabled
;
1815 void set_numabalancing_state(bool enabled
)
1817 numabalancing_enabled
= enabled
;
1819 #endif /* CONFIG_SCHED_DEBUG */
1821 #ifdef CONFIG_PROC_SYSCTL
1822 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1823 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1827 int state
= numabalancing_enabled
;
1829 if (write
&& !capable(CAP_SYS_ADMIN
))
1834 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1838 set_numabalancing_state(state
);
1845 * fork()/clone()-time setup:
1847 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1849 unsigned long flags
;
1850 int cpu
= get_cpu();
1852 __sched_fork(clone_flags
, p
);
1854 * We mark the process as running here. This guarantees that
1855 * nobody will actually run it, and a signal or other external
1856 * event cannot wake it up and insert it on the runqueue either.
1858 p
->state
= TASK_RUNNING
;
1861 * Make sure we do not leak PI boosting priority to the child.
1863 p
->prio
= current
->normal_prio
;
1866 * Revert to default priority/policy on fork if requested.
1868 if (unlikely(p
->sched_reset_on_fork
)) {
1869 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1870 p
->policy
= SCHED_NORMAL
;
1871 p
->static_prio
= NICE_TO_PRIO(0);
1873 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1874 p
->static_prio
= NICE_TO_PRIO(0);
1876 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1880 * We don't need the reset flag anymore after the fork. It has
1881 * fulfilled its duty:
1883 p
->sched_reset_on_fork
= 0;
1886 if (dl_prio(p
->prio
)) {
1889 } else if (rt_prio(p
->prio
)) {
1890 p
->sched_class
= &rt_sched_class
;
1892 p
->sched_class
= &fair_sched_class
;
1895 if (p
->sched_class
->task_fork
)
1896 p
->sched_class
->task_fork(p
);
1899 * The child is not yet in the pid-hash so no cgroup attach races,
1900 * and the cgroup is pinned to this child due to cgroup_fork()
1901 * is ran before sched_fork().
1903 * Silence PROVE_RCU.
1905 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1906 set_task_cpu(p
, cpu
);
1907 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1909 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1910 if (likely(sched_info_on()))
1911 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1913 #if defined(CONFIG_SMP)
1916 init_task_preempt_count(p
);
1918 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1919 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1926 unsigned long to_ratio(u64 period
, u64 runtime
)
1928 if (runtime
== RUNTIME_INF
)
1932 * Doing this here saves a lot of checks in all
1933 * the calling paths, and returning zero seems
1934 * safe for them anyway.
1939 return div64_u64(runtime
<< 20, period
);
1943 inline struct dl_bw
*dl_bw_of(int i
)
1945 return &cpu_rq(i
)->rd
->dl_bw
;
1948 static inline int dl_bw_cpus(int i
)
1950 struct root_domain
*rd
= cpu_rq(i
)->rd
;
1953 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
1959 inline struct dl_bw
*dl_bw_of(int i
)
1961 return &cpu_rq(i
)->dl
.dl_bw
;
1964 static inline int dl_bw_cpus(int i
)
1971 void __dl_clear(struct dl_bw
*dl_b
, u64 tsk_bw
)
1973 dl_b
->total_bw
-= tsk_bw
;
1977 void __dl_add(struct dl_bw
*dl_b
, u64 tsk_bw
)
1979 dl_b
->total_bw
+= tsk_bw
;
1983 bool __dl_overflow(struct dl_bw
*dl_b
, int cpus
, u64 old_bw
, u64 new_bw
)
1985 return dl_b
->bw
!= -1 &&
1986 dl_b
->bw
* cpus
< dl_b
->total_bw
- old_bw
+ new_bw
;
1990 * We must be sure that accepting a new task (or allowing changing the
1991 * parameters of an existing one) is consistent with the bandwidth
1992 * constraints. If yes, this function also accordingly updates the currently
1993 * allocated bandwidth to reflect the new situation.
1995 * This function is called while holding p's rq->lock.
1997 static int dl_overflow(struct task_struct
*p
, int policy
,
1998 const struct sched_attr
*attr
)
2001 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2002 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2003 u64 runtime
= attr
->sched_runtime
;
2004 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2007 if (new_bw
== p
->dl
.dl_bw
)
2011 * Either if a task, enters, leave, or stays -deadline but changes
2012 * its parameters, we may need to update accordingly the total
2013 * allocated bandwidth of the container.
2015 raw_spin_lock(&dl_b
->lock
);
2016 cpus
= dl_bw_cpus(task_cpu(p
));
2017 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2018 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2019 __dl_add(dl_b
, new_bw
);
2021 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2022 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2023 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2024 __dl_add(dl_b
, new_bw
);
2026 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2027 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2030 raw_spin_unlock(&dl_b
->lock
);
2035 extern void init_dl_bw(struct dl_bw
*dl_b
);
2038 * wake_up_new_task - wake up a newly created task for the first time.
2040 * This function will do some initial scheduler statistics housekeeping
2041 * that must be done for every newly created context, then puts the task
2042 * on the runqueue and wakes it.
2044 void wake_up_new_task(struct task_struct
*p
)
2046 unsigned long flags
;
2049 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2052 * Fork balancing, do it here and not earlier because:
2053 * - cpus_allowed can change in the fork path
2054 * - any previously selected cpu might disappear through hotplug
2056 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2059 /* Initialize new task's runnable average */
2060 init_task_runnable_average(p
);
2061 rq
= __task_rq_lock(p
);
2062 activate_task(rq
, p
, 0);
2064 trace_sched_wakeup_new(p
, true);
2065 check_preempt_curr(rq
, p
, WF_FORK
);
2067 if (p
->sched_class
->task_woken
)
2068 p
->sched_class
->task_woken(rq
, p
);
2070 task_rq_unlock(rq
, p
, &flags
);
2073 #ifdef CONFIG_PREEMPT_NOTIFIERS
2076 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2077 * @notifier: notifier struct to register
2079 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2081 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2083 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2086 * preempt_notifier_unregister - no longer interested in preemption notifications
2087 * @notifier: notifier struct to unregister
2089 * This is safe to call from within a preemption notifier.
2091 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2093 hlist_del(¬ifier
->link
);
2095 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2097 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2099 struct preempt_notifier
*notifier
;
2101 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2102 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2106 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2107 struct task_struct
*next
)
2109 struct preempt_notifier
*notifier
;
2111 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2112 notifier
->ops
->sched_out(notifier
, next
);
2115 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2117 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2122 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2123 struct task_struct
*next
)
2127 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2130 * prepare_task_switch - prepare to switch tasks
2131 * @rq: the runqueue preparing to switch
2132 * @prev: the current task that is being switched out
2133 * @next: the task we are going to switch to.
2135 * This is called with the rq lock held and interrupts off. It must
2136 * be paired with a subsequent finish_task_switch after the context
2139 * prepare_task_switch sets up locking and calls architecture specific
2143 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2144 struct task_struct
*next
)
2146 trace_sched_switch(prev
, next
);
2147 sched_info_switch(rq
, prev
, next
);
2148 perf_event_task_sched_out(prev
, next
);
2149 fire_sched_out_preempt_notifiers(prev
, next
);
2150 prepare_lock_switch(rq
, next
);
2151 prepare_arch_switch(next
);
2155 * finish_task_switch - clean up after a task-switch
2156 * @rq: runqueue associated with task-switch
2157 * @prev: the thread we just switched away from.
2159 * finish_task_switch must be called after the context switch, paired
2160 * with a prepare_task_switch call before the context switch.
2161 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2162 * and do any other architecture-specific cleanup actions.
2164 * Note that we may have delayed dropping an mm in context_switch(). If
2165 * so, we finish that here outside of the runqueue lock. (Doing it
2166 * with the lock held can cause deadlocks; see schedule() for
2169 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2170 __releases(rq
->lock
)
2172 struct mm_struct
*mm
= rq
->prev_mm
;
2178 * A task struct has one reference for the use as "current".
2179 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2180 * schedule one last time. The schedule call will never return, and
2181 * the scheduled task must drop that reference.
2182 * The test for TASK_DEAD must occur while the runqueue locks are
2183 * still held, otherwise prev could be scheduled on another cpu, die
2184 * there before we look at prev->state, and then the reference would
2186 * Manfred Spraul <manfred@colorfullife.com>
2188 prev_state
= prev
->state
;
2189 vtime_task_switch(prev
);
2190 finish_arch_switch(prev
);
2191 perf_event_task_sched_in(prev
, current
);
2192 finish_lock_switch(rq
, prev
);
2193 finish_arch_post_lock_switch();
2195 fire_sched_in_preempt_notifiers(current
);
2198 if (unlikely(prev_state
== TASK_DEAD
)) {
2199 if (prev
->sched_class
->task_dead
)
2200 prev
->sched_class
->task_dead(prev
);
2203 * Remove function-return probe instances associated with this
2204 * task and put them back on the free list.
2206 kprobe_flush_task(prev
);
2207 put_task_struct(prev
);
2210 tick_nohz_task_switch(current
);
2215 /* rq->lock is NOT held, but preemption is disabled */
2216 static inline void post_schedule(struct rq
*rq
)
2218 if (rq
->post_schedule
) {
2219 unsigned long flags
;
2221 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2222 if (rq
->curr
->sched_class
->post_schedule
)
2223 rq
->curr
->sched_class
->post_schedule(rq
);
2224 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2226 rq
->post_schedule
= 0;
2232 static inline void post_schedule(struct rq
*rq
)
2239 * schedule_tail - first thing a freshly forked thread must call.
2240 * @prev: the thread we just switched away from.
2242 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2243 __releases(rq
->lock
)
2245 struct rq
*rq
= this_rq();
2247 finish_task_switch(rq
, prev
);
2250 * FIXME: do we need to worry about rq being invalidated by the
2255 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2256 /* In this case, finish_task_switch does not reenable preemption */
2259 if (current
->set_child_tid
)
2260 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2264 * context_switch - switch to the new MM and the new
2265 * thread's register state.
2268 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2269 struct task_struct
*next
)
2271 struct mm_struct
*mm
, *oldmm
;
2273 prepare_task_switch(rq
, prev
, next
);
2276 oldmm
= prev
->active_mm
;
2278 * For paravirt, this is coupled with an exit in switch_to to
2279 * combine the page table reload and the switch backend into
2282 arch_start_context_switch(prev
);
2285 next
->active_mm
= oldmm
;
2286 atomic_inc(&oldmm
->mm_count
);
2287 enter_lazy_tlb(oldmm
, next
);
2289 switch_mm(oldmm
, mm
, next
);
2292 prev
->active_mm
= NULL
;
2293 rq
->prev_mm
= oldmm
;
2296 * Since the runqueue lock will be released by the next
2297 * task (which is an invalid locking op but in the case
2298 * of the scheduler it's an obvious special-case), so we
2299 * do an early lockdep release here:
2301 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2302 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2305 context_tracking_task_switch(prev
, next
);
2306 /* Here we just switch the register state and the stack. */
2307 switch_to(prev
, next
, prev
);
2311 * this_rq must be evaluated again because prev may have moved
2312 * CPUs since it called schedule(), thus the 'rq' on its stack
2313 * frame will be invalid.
2315 finish_task_switch(this_rq(), prev
);
2319 * nr_running and nr_context_switches:
2321 * externally visible scheduler statistics: current number of runnable
2322 * threads, total number of context switches performed since bootup.
2324 unsigned long nr_running(void)
2326 unsigned long i
, sum
= 0;
2328 for_each_online_cpu(i
)
2329 sum
+= cpu_rq(i
)->nr_running
;
2334 unsigned long long nr_context_switches(void)
2337 unsigned long long sum
= 0;
2339 for_each_possible_cpu(i
)
2340 sum
+= cpu_rq(i
)->nr_switches
;
2345 unsigned long nr_iowait(void)
2347 unsigned long i
, sum
= 0;
2349 for_each_possible_cpu(i
)
2350 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2355 unsigned long nr_iowait_cpu(int cpu
)
2357 struct rq
*this = cpu_rq(cpu
);
2358 return atomic_read(&this->nr_iowait
);
2364 * sched_exec - execve() is a valuable balancing opportunity, because at
2365 * this point the task has the smallest effective memory and cache footprint.
2367 void sched_exec(void)
2369 struct task_struct
*p
= current
;
2370 unsigned long flags
;
2373 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2374 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2375 if (dest_cpu
== smp_processor_id())
2378 if (likely(cpu_active(dest_cpu
))) {
2379 struct migration_arg arg
= { p
, dest_cpu
};
2381 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2382 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2386 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2391 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2392 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2394 EXPORT_PER_CPU_SYMBOL(kstat
);
2395 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2398 * Return any ns on the sched_clock that have not yet been accounted in
2399 * @p in case that task is currently running.
2401 * Called with task_rq_lock() held on @rq.
2403 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2407 if (task_current(rq
, p
)) {
2408 update_rq_clock(rq
);
2409 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2417 unsigned long long task_delta_exec(struct task_struct
*p
)
2419 unsigned long flags
;
2423 rq
= task_rq_lock(p
, &flags
);
2424 ns
= do_task_delta_exec(p
, rq
);
2425 task_rq_unlock(rq
, p
, &flags
);
2431 * Return accounted runtime for the task.
2432 * In case the task is currently running, return the runtime plus current's
2433 * pending runtime that have not been accounted yet.
2435 unsigned long long task_sched_runtime(struct task_struct
*p
)
2437 unsigned long flags
;
2441 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2443 * 64-bit doesn't need locks to atomically read a 64bit value.
2444 * So we have a optimization chance when the task's delta_exec is 0.
2445 * Reading ->on_cpu is racy, but this is ok.
2447 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2448 * If we race with it entering cpu, unaccounted time is 0. This is
2449 * indistinguishable from the read occurring a few cycles earlier.
2452 return p
->se
.sum_exec_runtime
;
2455 rq
= task_rq_lock(p
, &flags
);
2456 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2457 task_rq_unlock(rq
, p
, &flags
);
2463 * This function gets called by the timer code, with HZ frequency.
2464 * We call it with interrupts disabled.
2466 void scheduler_tick(void)
2468 int cpu
= smp_processor_id();
2469 struct rq
*rq
= cpu_rq(cpu
);
2470 struct task_struct
*curr
= rq
->curr
;
2474 raw_spin_lock(&rq
->lock
);
2475 update_rq_clock(rq
);
2476 curr
->sched_class
->task_tick(rq
, curr
, 0);
2477 update_cpu_load_active(rq
);
2478 raw_spin_unlock(&rq
->lock
);
2480 perf_event_task_tick();
2483 rq
->idle_balance
= idle_cpu(cpu
);
2484 trigger_load_balance(rq
);
2486 rq_last_tick_reset(rq
);
2489 #ifdef CONFIG_NO_HZ_FULL
2491 * scheduler_tick_max_deferment
2493 * Keep at least one tick per second when a single
2494 * active task is running because the scheduler doesn't
2495 * yet completely support full dynticks environment.
2497 * This makes sure that uptime, CFS vruntime, load
2498 * balancing, etc... continue to move forward, even
2499 * with a very low granularity.
2501 * Return: Maximum deferment in nanoseconds.
2503 u64
scheduler_tick_max_deferment(void)
2505 struct rq
*rq
= this_rq();
2506 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2508 next
= rq
->last_sched_tick
+ HZ
;
2510 if (time_before_eq(next
, now
))
2513 return jiffies_to_nsecs(next
- now
);
2517 notrace
unsigned long get_parent_ip(unsigned long addr
)
2519 if (in_lock_functions(addr
)) {
2520 addr
= CALLER_ADDR2
;
2521 if (in_lock_functions(addr
))
2522 addr
= CALLER_ADDR3
;
2527 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2528 defined(CONFIG_PREEMPT_TRACER))
2530 void __kprobes
preempt_count_add(int val
)
2532 #ifdef CONFIG_DEBUG_PREEMPT
2536 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2539 __preempt_count_add(val
);
2540 #ifdef CONFIG_DEBUG_PREEMPT
2542 * Spinlock count overflowing soon?
2544 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2547 if (preempt_count() == val
) {
2548 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550 current
->preempt_disable_ip
= ip
;
2552 trace_preempt_off(CALLER_ADDR0
, ip
);
2555 EXPORT_SYMBOL(preempt_count_add
);
2557 void __kprobes
preempt_count_sub(int val
)
2559 #ifdef CONFIG_DEBUG_PREEMPT
2563 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2566 * Is the spinlock portion underflowing?
2568 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2569 !(preempt_count() & PREEMPT_MASK
)))
2573 if (preempt_count() == val
)
2574 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2575 __preempt_count_sub(val
);
2577 EXPORT_SYMBOL(preempt_count_sub
);
2582 * Print scheduling while atomic bug:
2584 static noinline
void __schedule_bug(struct task_struct
*prev
)
2586 if (oops_in_progress
)
2589 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2590 prev
->comm
, prev
->pid
, preempt_count());
2592 debug_show_held_locks(prev
);
2594 if (irqs_disabled())
2595 print_irqtrace_events(prev
);
2596 #ifdef CONFIG_DEBUG_PREEMPT
2597 if (in_atomic_preempt_off()) {
2598 pr_err("Preemption disabled at:");
2599 print_ip_sym(current
->preempt_disable_ip
);
2604 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2608 * Various schedule()-time debugging checks and statistics:
2610 static inline void schedule_debug(struct task_struct
*prev
)
2613 * Test if we are atomic. Since do_exit() needs to call into
2614 * schedule() atomically, we ignore that path. Otherwise whine
2615 * if we are scheduling when we should not.
2617 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2618 __schedule_bug(prev
);
2621 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2623 schedstat_inc(this_rq(), sched_count
);
2627 * Pick up the highest-prio task:
2629 static inline struct task_struct
*
2630 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2632 const struct sched_class
*class = &fair_sched_class
;
2633 struct task_struct
*p
;
2636 * Optimization: we know that if all tasks are in
2637 * the fair class we can call that function directly:
2639 if (likely(prev
->sched_class
== class &&
2640 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2641 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2642 if (unlikely(p
== RETRY_TASK
))
2645 /* assumes fair_sched_class->next == idle_sched_class */
2647 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2653 for_each_class(class) {
2654 p
= class->pick_next_task(rq
, prev
);
2656 if (unlikely(p
== RETRY_TASK
))
2662 BUG(); /* the idle class will always have a runnable task */
2666 * __schedule() is the main scheduler function.
2668 * The main means of driving the scheduler and thus entering this function are:
2670 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2672 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2673 * paths. For example, see arch/x86/entry_64.S.
2675 * To drive preemption between tasks, the scheduler sets the flag in timer
2676 * interrupt handler scheduler_tick().
2678 * 3. Wakeups don't really cause entry into schedule(). They add a
2679 * task to the run-queue and that's it.
2681 * Now, if the new task added to the run-queue preempts the current
2682 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2683 * called on the nearest possible occasion:
2685 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2687 * - in syscall or exception context, at the next outmost
2688 * preempt_enable(). (this might be as soon as the wake_up()'s
2691 * - in IRQ context, return from interrupt-handler to
2692 * preemptible context
2694 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2697 * - cond_resched() call
2698 * - explicit schedule() call
2699 * - return from syscall or exception to user-space
2700 * - return from interrupt-handler to user-space
2702 static void __sched
__schedule(void)
2704 struct task_struct
*prev
, *next
;
2705 unsigned long *switch_count
;
2711 cpu
= smp_processor_id();
2713 rcu_note_context_switch(cpu
);
2716 schedule_debug(prev
);
2718 if (sched_feat(HRTICK
))
2722 * Make sure that signal_pending_state()->signal_pending() below
2723 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2724 * done by the caller to avoid the race with signal_wake_up().
2726 smp_mb__before_spinlock();
2727 raw_spin_lock_irq(&rq
->lock
);
2729 switch_count
= &prev
->nivcsw
;
2730 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2731 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2732 prev
->state
= TASK_RUNNING
;
2734 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2738 * If a worker went to sleep, notify and ask workqueue
2739 * whether it wants to wake up a task to maintain
2742 if (prev
->flags
& PF_WQ_WORKER
) {
2743 struct task_struct
*to_wakeup
;
2745 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2747 try_to_wake_up_local(to_wakeup
);
2750 switch_count
= &prev
->nvcsw
;
2753 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2754 update_rq_clock(rq
);
2756 next
= pick_next_task(rq
, prev
);
2757 clear_tsk_need_resched(prev
);
2758 clear_preempt_need_resched();
2759 rq
->skip_clock_update
= 0;
2761 if (likely(prev
!= next
)) {
2766 context_switch(rq
, prev
, next
); /* unlocks the rq */
2768 * The context switch have flipped the stack from under us
2769 * and restored the local variables which were saved when
2770 * this task called schedule() in the past. prev == current
2771 * is still correct, but it can be moved to another cpu/rq.
2773 cpu
= smp_processor_id();
2776 raw_spin_unlock_irq(&rq
->lock
);
2780 sched_preempt_enable_no_resched();
2785 static inline void sched_submit_work(struct task_struct
*tsk
)
2787 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2790 * If we are going to sleep and we have plugged IO queued,
2791 * make sure to submit it to avoid deadlocks.
2793 if (blk_needs_flush_plug(tsk
))
2794 blk_schedule_flush_plug(tsk
);
2797 asmlinkage __visible
void __sched
schedule(void)
2799 struct task_struct
*tsk
= current
;
2801 sched_submit_work(tsk
);
2804 EXPORT_SYMBOL(schedule
);
2806 #ifdef CONFIG_CONTEXT_TRACKING
2807 asmlinkage __visible
void __sched
schedule_user(void)
2810 * If we come here after a random call to set_need_resched(),
2811 * or we have been woken up remotely but the IPI has not yet arrived,
2812 * we haven't yet exited the RCU idle mode. Do it here manually until
2813 * we find a better solution.
2822 * schedule_preempt_disabled - called with preemption disabled
2824 * Returns with preemption disabled. Note: preempt_count must be 1
2826 void __sched
schedule_preempt_disabled(void)
2828 sched_preempt_enable_no_resched();
2833 #ifdef CONFIG_PREEMPT
2835 * this is the entry point to schedule() from in-kernel preemption
2836 * off of preempt_enable. Kernel preemptions off return from interrupt
2837 * occur there and call schedule directly.
2839 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2842 * If there is a non-zero preempt_count or interrupts are disabled,
2843 * we do not want to preempt the current task. Just return..
2845 if (likely(!preemptible()))
2849 __preempt_count_add(PREEMPT_ACTIVE
);
2851 __preempt_count_sub(PREEMPT_ACTIVE
);
2854 * Check again in case we missed a preemption opportunity
2855 * between schedule and now.
2858 } while (need_resched());
2860 EXPORT_SYMBOL(preempt_schedule
);
2861 #endif /* CONFIG_PREEMPT */
2864 * this is the entry point to schedule() from kernel preemption
2865 * off of irq context.
2866 * Note, that this is called and return with irqs disabled. This will
2867 * protect us against recursive calling from irq.
2869 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2871 enum ctx_state prev_state
;
2873 /* Catch callers which need to be fixed */
2874 BUG_ON(preempt_count() || !irqs_disabled());
2876 prev_state
= exception_enter();
2879 __preempt_count_add(PREEMPT_ACTIVE
);
2882 local_irq_disable();
2883 __preempt_count_sub(PREEMPT_ACTIVE
);
2886 * Check again in case we missed a preemption opportunity
2887 * between schedule and now.
2890 } while (need_resched());
2892 exception_exit(prev_state
);
2895 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2898 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2900 EXPORT_SYMBOL(default_wake_function
);
2902 #ifdef CONFIG_RT_MUTEXES
2905 * rt_mutex_setprio - set the current priority of a task
2907 * @prio: prio value (kernel-internal form)
2909 * This function changes the 'effective' priority of a task. It does
2910 * not touch ->normal_prio like __setscheduler().
2912 * Used by the rt_mutex code to implement priority inheritance
2913 * logic. Call site only calls if the priority of the task changed.
2915 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2917 int oldprio
, on_rq
, running
, enqueue_flag
= 0;
2919 const struct sched_class
*prev_class
;
2921 BUG_ON(prio
> MAX_PRIO
);
2923 rq
= __task_rq_lock(p
);
2926 * Idle task boosting is a nono in general. There is one
2927 * exception, when PREEMPT_RT and NOHZ is active:
2929 * The idle task calls get_next_timer_interrupt() and holds
2930 * the timer wheel base->lock on the CPU and another CPU wants
2931 * to access the timer (probably to cancel it). We can safely
2932 * ignore the boosting request, as the idle CPU runs this code
2933 * with interrupts disabled and will complete the lock
2934 * protected section without being interrupted. So there is no
2935 * real need to boost.
2937 if (unlikely(p
== rq
->idle
)) {
2938 WARN_ON(p
!= rq
->curr
);
2939 WARN_ON(p
->pi_blocked_on
);
2943 trace_sched_pi_setprio(p
, prio
);
2944 p
->pi_top_task
= rt_mutex_get_top_task(p
);
2946 prev_class
= p
->sched_class
;
2948 running
= task_current(rq
, p
);
2950 dequeue_task(rq
, p
, 0);
2952 p
->sched_class
->put_prev_task(rq
, p
);
2955 * Boosting condition are:
2956 * 1. -rt task is running and holds mutex A
2957 * --> -dl task blocks on mutex A
2959 * 2. -dl task is running and holds mutex A
2960 * --> -dl task blocks on mutex A and could preempt the
2963 if (dl_prio(prio
)) {
2964 if (!dl_prio(p
->normal_prio
) || (p
->pi_top_task
&&
2965 dl_entity_preempt(&p
->pi_top_task
->dl
, &p
->dl
))) {
2966 p
->dl
.dl_boosted
= 1;
2967 p
->dl
.dl_throttled
= 0;
2968 enqueue_flag
= ENQUEUE_REPLENISH
;
2970 p
->dl
.dl_boosted
= 0;
2971 p
->sched_class
= &dl_sched_class
;
2972 } else if (rt_prio(prio
)) {
2973 if (dl_prio(oldprio
))
2974 p
->dl
.dl_boosted
= 0;
2976 enqueue_flag
= ENQUEUE_HEAD
;
2977 p
->sched_class
= &rt_sched_class
;
2979 if (dl_prio(oldprio
))
2980 p
->dl
.dl_boosted
= 0;
2981 p
->sched_class
= &fair_sched_class
;
2987 p
->sched_class
->set_curr_task(rq
);
2989 enqueue_task(rq
, p
, enqueue_flag
);
2991 check_class_changed(rq
, p
, prev_class
, oldprio
);
2993 __task_rq_unlock(rq
);
2997 void set_user_nice(struct task_struct
*p
, long nice
)
2999 int old_prio
, delta
, on_rq
;
3000 unsigned long flags
;
3003 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3006 * We have to be careful, if called from sys_setpriority(),
3007 * the task might be in the middle of scheduling on another CPU.
3009 rq
= task_rq_lock(p
, &flags
);
3011 * The RT priorities are set via sched_setscheduler(), but we still
3012 * allow the 'normal' nice value to be set - but as expected
3013 * it wont have any effect on scheduling until the task is
3014 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3016 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3017 p
->static_prio
= NICE_TO_PRIO(nice
);
3022 dequeue_task(rq
, p
, 0);
3024 p
->static_prio
= NICE_TO_PRIO(nice
);
3027 p
->prio
= effective_prio(p
);
3028 delta
= p
->prio
- old_prio
;
3031 enqueue_task(rq
, p
, 0);
3033 * If the task increased its priority or is running and
3034 * lowered its priority, then reschedule its CPU:
3036 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3037 resched_task(rq
->curr
);
3040 task_rq_unlock(rq
, p
, &flags
);
3042 EXPORT_SYMBOL(set_user_nice
);
3045 * can_nice - check if a task can reduce its nice value
3049 int can_nice(const struct task_struct
*p
, const int nice
)
3051 /* convert nice value [19,-20] to rlimit style value [1,40] */
3052 int nice_rlim
= nice_to_rlimit(nice
);
3054 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3055 capable(CAP_SYS_NICE
));
3058 #ifdef __ARCH_WANT_SYS_NICE
3061 * sys_nice - change the priority of the current process.
3062 * @increment: priority increment
3064 * sys_setpriority is a more generic, but much slower function that
3065 * does similar things.
3067 SYSCALL_DEFINE1(nice
, int, increment
)
3072 * Setpriority might change our priority at the same moment.
3073 * We don't have to worry. Conceptually one call occurs first
3074 * and we have a single winner.
3076 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3077 nice
= task_nice(current
) + increment
;
3079 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3080 if (increment
< 0 && !can_nice(current
, nice
))
3083 retval
= security_task_setnice(current
, nice
);
3087 set_user_nice(current
, nice
);
3094 * task_prio - return the priority value of a given task.
3095 * @p: the task in question.
3097 * Return: The priority value as seen by users in /proc.
3098 * RT tasks are offset by -200. Normal tasks are centered
3099 * around 0, value goes from -16 to +15.
3101 int task_prio(const struct task_struct
*p
)
3103 return p
->prio
- MAX_RT_PRIO
;
3107 * idle_cpu - is a given cpu idle currently?
3108 * @cpu: the processor in question.
3110 * Return: 1 if the CPU is currently idle. 0 otherwise.
3112 int idle_cpu(int cpu
)
3114 struct rq
*rq
= cpu_rq(cpu
);
3116 if (rq
->curr
!= rq
->idle
)
3123 if (!llist_empty(&rq
->wake_list
))
3131 * idle_task - return the idle task for a given cpu.
3132 * @cpu: the processor in question.
3134 * Return: The idle task for the cpu @cpu.
3136 struct task_struct
*idle_task(int cpu
)
3138 return cpu_rq(cpu
)->idle
;
3142 * find_process_by_pid - find a process with a matching PID value.
3143 * @pid: the pid in question.
3145 * The task of @pid, if found. %NULL otherwise.
3147 static struct task_struct
*find_process_by_pid(pid_t pid
)
3149 return pid
? find_task_by_vpid(pid
) : current
;
3153 * This function initializes the sched_dl_entity of a newly becoming
3154 * SCHED_DEADLINE task.
3156 * Only the static values are considered here, the actual runtime and the
3157 * absolute deadline will be properly calculated when the task is enqueued
3158 * for the first time with its new policy.
3161 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3163 struct sched_dl_entity
*dl_se
= &p
->dl
;
3165 init_dl_task_timer(dl_se
);
3166 dl_se
->dl_runtime
= attr
->sched_runtime
;
3167 dl_se
->dl_deadline
= attr
->sched_deadline
;
3168 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3169 dl_se
->flags
= attr
->sched_flags
;
3170 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3171 dl_se
->dl_throttled
= 0;
3173 dl_se
->dl_yielded
= 0;
3176 static void __setscheduler_params(struct task_struct
*p
,
3177 const struct sched_attr
*attr
)
3179 int policy
= attr
->sched_policy
;
3181 if (policy
== -1) /* setparam */
3186 if (dl_policy(policy
))
3187 __setparam_dl(p
, attr
);
3188 else if (fair_policy(policy
))
3189 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3192 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3193 * !rt_policy. Always setting this ensures that things like
3194 * getparam()/getattr() don't report silly values for !rt tasks.
3196 p
->rt_priority
= attr
->sched_priority
;
3197 p
->normal_prio
= normal_prio(p
);
3201 /* Actually do priority change: must hold pi & rq lock. */
3202 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3203 const struct sched_attr
*attr
)
3205 __setscheduler_params(p
, attr
);
3208 * If we get here, there was no pi waiters boosting the
3209 * task. It is safe to use the normal prio.
3211 p
->prio
= normal_prio(p
);
3213 if (dl_prio(p
->prio
))
3214 p
->sched_class
= &dl_sched_class
;
3215 else if (rt_prio(p
->prio
))
3216 p
->sched_class
= &rt_sched_class
;
3218 p
->sched_class
= &fair_sched_class
;
3222 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3224 struct sched_dl_entity
*dl_se
= &p
->dl
;
3226 attr
->sched_priority
= p
->rt_priority
;
3227 attr
->sched_runtime
= dl_se
->dl_runtime
;
3228 attr
->sched_deadline
= dl_se
->dl_deadline
;
3229 attr
->sched_period
= dl_se
->dl_period
;
3230 attr
->sched_flags
= dl_se
->flags
;
3234 * This function validates the new parameters of a -deadline task.
3235 * We ask for the deadline not being zero, and greater or equal
3236 * than the runtime, as well as the period of being zero or
3237 * greater than deadline. Furthermore, we have to be sure that
3238 * user parameters are above the internal resolution of 1us (we
3239 * check sched_runtime only since it is always the smaller one) and
3240 * below 2^63 ns (we have to check both sched_deadline and
3241 * sched_period, as the latter can be zero).
3244 __checkparam_dl(const struct sched_attr
*attr
)
3247 if (attr
->sched_deadline
== 0)
3251 * Since we truncate DL_SCALE bits, make sure we're at least
3254 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3258 * Since we use the MSB for wrap-around and sign issues, make
3259 * sure it's not set (mind that period can be equal to zero).
3261 if (attr
->sched_deadline
& (1ULL << 63) ||
3262 attr
->sched_period
& (1ULL << 63))
3265 /* runtime <= deadline <= period (if period != 0) */
3266 if ((attr
->sched_period
!= 0 &&
3267 attr
->sched_period
< attr
->sched_deadline
) ||
3268 attr
->sched_deadline
< attr
->sched_runtime
)
3275 * check the target process has a UID that matches the current process's
3277 static bool check_same_owner(struct task_struct
*p
)
3279 const struct cred
*cred
= current_cred(), *pcred
;
3283 pcred
= __task_cred(p
);
3284 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3285 uid_eq(cred
->euid
, pcred
->uid
));
3290 static int __sched_setscheduler(struct task_struct
*p
,
3291 const struct sched_attr
*attr
,
3294 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3295 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3296 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3297 int policy
= attr
->sched_policy
;
3298 unsigned long flags
;
3299 const struct sched_class
*prev_class
;
3303 /* may grab non-irq protected spin_locks */
3304 BUG_ON(in_interrupt());
3306 /* double check policy once rq lock held */
3308 reset_on_fork
= p
->sched_reset_on_fork
;
3309 policy
= oldpolicy
= p
->policy
;
3311 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3313 if (policy
!= SCHED_DEADLINE
&&
3314 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3315 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3316 policy
!= SCHED_IDLE
)
3320 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3324 * Valid priorities for SCHED_FIFO and SCHED_RR are
3325 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3326 * SCHED_BATCH and SCHED_IDLE is 0.
3328 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3329 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3331 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3332 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3336 * Allow unprivileged RT tasks to decrease priority:
3338 if (user
&& !capable(CAP_SYS_NICE
)) {
3339 if (fair_policy(policy
)) {
3340 if (attr
->sched_nice
< task_nice(p
) &&
3341 !can_nice(p
, attr
->sched_nice
))
3345 if (rt_policy(policy
)) {
3346 unsigned long rlim_rtprio
=
3347 task_rlimit(p
, RLIMIT_RTPRIO
);
3349 /* can't set/change the rt policy */
3350 if (policy
!= p
->policy
&& !rlim_rtprio
)
3353 /* can't increase priority */
3354 if (attr
->sched_priority
> p
->rt_priority
&&
3355 attr
->sched_priority
> rlim_rtprio
)
3360 * Can't set/change SCHED_DEADLINE policy at all for now
3361 * (safest behavior); in the future we would like to allow
3362 * unprivileged DL tasks to increase their relative deadline
3363 * or reduce their runtime (both ways reducing utilization)
3365 if (dl_policy(policy
))
3369 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3370 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3372 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3373 if (!can_nice(p
, task_nice(p
)))
3377 /* can't change other user's priorities */
3378 if (!check_same_owner(p
))
3381 /* Normal users shall not reset the sched_reset_on_fork flag */
3382 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3387 retval
= security_task_setscheduler(p
);
3393 * make sure no PI-waiters arrive (or leave) while we are
3394 * changing the priority of the task:
3396 * To be able to change p->policy safely, the appropriate
3397 * runqueue lock must be held.
3399 rq
= task_rq_lock(p
, &flags
);
3402 * Changing the policy of the stop threads its a very bad idea
3404 if (p
== rq
->stop
) {
3405 task_rq_unlock(rq
, p
, &flags
);
3410 * If not changing anything there's no need to proceed further,
3411 * but store a possible modification of reset_on_fork.
3413 if (unlikely(policy
== p
->policy
)) {
3414 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3416 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3418 if (dl_policy(policy
))
3421 p
->sched_reset_on_fork
= reset_on_fork
;
3422 task_rq_unlock(rq
, p
, &flags
);
3428 #ifdef CONFIG_RT_GROUP_SCHED
3430 * Do not allow realtime tasks into groups that have no runtime
3433 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3434 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3435 !task_group_is_autogroup(task_group(p
))) {
3436 task_rq_unlock(rq
, p
, &flags
);
3441 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3442 cpumask_t
*span
= rq
->rd
->span
;
3445 * Don't allow tasks with an affinity mask smaller than
3446 * the entire root_domain to become SCHED_DEADLINE. We
3447 * will also fail if there's no bandwidth available.
3449 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3450 rq
->rd
->dl_bw
.bw
== 0) {
3451 task_rq_unlock(rq
, p
, &flags
);
3458 /* recheck policy now with rq lock held */
3459 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3460 policy
= oldpolicy
= -1;
3461 task_rq_unlock(rq
, p
, &flags
);
3466 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3467 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3470 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3471 task_rq_unlock(rq
, p
, &flags
);
3475 p
->sched_reset_on_fork
= reset_on_fork
;
3479 * Special case for priority boosted tasks.
3481 * If the new priority is lower or equal (user space view)
3482 * than the current (boosted) priority, we just store the new
3483 * normal parameters and do not touch the scheduler class and
3484 * the runqueue. This will be done when the task deboost
3487 if (rt_mutex_check_prio(p
, newprio
)) {
3488 __setscheduler_params(p
, attr
);
3489 task_rq_unlock(rq
, p
, &flags
);
3494 running
= task_current(rq
, p
);
3496 dequeue_task(rq
, p
, 0);
3498 p
->sched_class
->put_prev_task(rq
, p
);
3500 prev_class
= p
->sched_class
;
3501 __setscheduler(rq
, p
, attr
);
3504 p
->sched_class
->set_curr_task(rq
);
3507 * We enqueue to tail when the priority of a task is
3508 * increased (user space view).
3510 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3513 check_class_changed(rq
, p
, prev_class
, oldprio
);
3514 task_rq_unlock(rq
, p
, &flags
);
3516 rt_mutex_adjust_pi(p
);
3521 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3522 const struct sched_param
*param
, bool check
)
3524 struct sched_attr attr
= {
3525 .sched_policy
= policy
,
3526 .sched_priority
= param
->sched_priority
,
3527 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3531 * Fixup the legacy SCHED_RESET_ON_FORK hack
3533 if (policy
& SCHED_RESET_ON_FORK
) {
3534 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3535 policy
&= ~SCHED_RESET_ON_FORK
;
3536 attr
.sched_policy
= policy
;
3539 return __sched_setscheduler(p
, &attr
, check
);
3542 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3543 * @p: the task in question.
3544 * @policy: new policy.
3545 * @param: structure containing the new RT priority.
3547 * Return: 0 on success. An error code otherwise.
3549 * NOTE that the task may be already dead.
3551 int sched_setscheduler(struct task_struct
*p
, int policy
,
3552 const struct sched_param
*param
)
3554 return _sched_setscheduler(p
, policy
, param
, true);
3556 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3558 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3560 return __sched_setscheduler(p
, attr
, true);
3562 EXPORT_SYMBOL_GPL(sched_setattr
);
3565 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3566 * @p: the task in question.
3567 * @policy: new policy.
3568 * @param: structure containing the new RT priority.
3570 * Just like sched_setscheduler, only don't bother checking if the
3571 * current context has permission. For example, this is needed in
3572 * stop_machine(): we create temporary high priority worker threads,
3573 * but our caller might not have that capability.
3575 * Return: 0 on success. An error code otherwise.
3577 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3578 const struct sched_param
*param
)
3580 return _sched_setscheduler(p
, policy
, param
, false);
3584 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3586 struct sched_param lparam
;
3587 struct task_struct
*p
;
3590 if (!param
|| pid
< 0)
3592 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3597 p
= find_process_by_pid(pid
);
3599 retval
= sched_setscheduler(p
, policy
, &lparam
);
3606 * Mimics kernel/events/core.c perf_copy_attr().
3608 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3609 struct sched_attr
*attr
)
3614 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3618 * zero the full structure, so that a short copy will be nice.
3620 memset(attr
, 0, sizeof(*attr
));
3622 ret
= get_user(size
, &uattr
->size
);
3626 if (size
> PAGE_SIZE
) /* silly large */
3629 if (!size
) /* abi compat */
3630 size
= SCHED_ATTR_SIZE_VER0
;
3632 if (size
< SCHED_ATTR_SIZE_VER0
)
3636 * If we're handed a bigger struct than we know of,
3637 * ensure all the unknown bits are 0 - i.e. new
3638 * user-space does not rely on any kernel feature
3639 * extensions we dont know about yet.
3641 if (size
> sizeof(*attr
)) {
3642 unsigned char __user
*addr
;
3643 unsigned char __user
*end
;
3646 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3647 end
= (void __user
*)uattr
+ size
;
3649 for (; addr
< end
; addr
++) {
3650 ret
= get_user(val
, addr
);
3656 size
= sizeof(*attr
);
3659 ret
= copy_from_user(attr
, uattr
, size
);
3664 * XXX: do we want to be lenient like existing syscalls; or do we want
3665 * to be strict and return an error on out-of-bounds values?
3667 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3672 put_user(sizeof(*attr
), &uattr
->size
);
3677 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3678 * @pid: the pid in question.
3679 * @policy: new policy.
3680 * @param: structure containing the new RT priority.
3682 * Return: 0 on success. An error code otherwise.
3684 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3685 struct sched_param __user
*, param
)
3687 /* negative values for policy are not valid */
3691 return do_sched_setscheduler(pid
, policy
, param
);
3695 * sys_sched_setparam - set/change the RT priority of a thread
3696 * @pid: the pid in question.
3697 * @param: structure containing the new RT priority.
3699 * Return: 0 on success. An error code otherwise.
3701 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3703 return do_sched_setscheduler(pid
, -1, param
);
3707 * sys_sched_setattr - same as above, but with extended sched_attr
3708 * @pid: the pid in question.
3709 * @uattr: structure containing the extended parameters.
3710 * @flags: for future extension.
3712 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3713 unsigned int, flags
)
3715 struct sched_attr attr
;
3716 struct task_struct
*p
;
3719 if (!uattr
|| pid
< 0 || flags
)
3722 retval
= sched_copy_attr(uattr
, &attr
);
3726 if (attr
.sched_policy
< 0)
3731 p
= find_process_by_pid(pid
);
3733 retval
= sched_setattr(p
, &attr
);
3740 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3741 * @pid: the pid in question.
3743 * Return: On success, the policy of the thread. Otherwise, a negative error
3746 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3748 struct task_struct
*p
;
3756 p
= find_process_by_pid(pid
);
3758 retval
= security_task_getscheduler(p
);
3761 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3768 * sys_sched_getparam - get the RT priority of a thread
3769 * @pid: the pid in question.
3770 * @param: structure containing the RT priority.
3772 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3775 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3777 struct sched_param lp
= { .sched_priority
= 0 };
3778 struct task_struct
*p
;
3781 if (!param
|| pid
< 0)
3785 p
= find_process_by_pid(pid
);
3790 retval
= security_task_getscheduler(p
);
3794 if (task_has_rt_policy(p
))
3795 lp
.sched_priority
= p
->rt_priority
;
3799 * This one might sleep, we cannot do it with a spinlock held ...
3801 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3810 static int sched_read_attr(struct sched_attr __user
*uattr
,
3811 struct sched_attr
*attr
,
3816 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3820 * If we're handed a smaller struct than we know of,
3821 * ensure all the unknown bits are 0 - i.e. old
3822 * user-space does not get uncomplete information.
3824 if (usize
< sizeof(*attr
)) {
3825 unsigned char *addr
;
3828 addr
= (void *)attr
+ usize
;
3829 end
= (void *)attr
+ sizeof(*attr
);
3831 for (; addr
< end
; addr
++) {
3839 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3847 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3848 * @pid: the pid in question.
3849 * @uattr: structure containing the extended parameters.
3850 * @size: sizeof(attr) for fwd/bwd comp.
3851 * @flags: for future extension.
3853 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3854 unsigned int, size
, unsigned int, flags
)
3856 struct sched_attr attr
= {
3857 .size
= sizeof(struct sched_attr
),
3859 struct task_struct
*p
;
3862 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3863 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3867 p
= find_process_by_pid(pid
);
3872 retval
= security_task_getscheduler(p
);
3876 attr
.sched_policy
= p
->policy
;
3877 if (p
->sched_reset_on_fork
)
3878 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3879 if (task_has_dl_policy(p
))
3880 __getparam_dl(p
, &attr
);
3881 else if (task_has_rt_policy(p
))
3882 attr
.sched_priority
= p
->rt_priority
;
3884 attr
.sched_nice
= task_nice(p
);
3888 retval
= sched_read_attr(uattr
, &attr
, size
);
3896 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3898 cpumask_var_t cpus_allowed
, new_mask
;
3899 struct task_struct
*p
;
3904 p
= find_process_by_pid(pid
);
3910 /* Prevent p going away */
3914 if (p
->flags
& PF_NO_SETAFFINITY
) {
3918 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3922 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3924 goto out_free_cpus_allowed
;
3927 if (!check_same_owner(p
)) {
3929 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3936 retval
= security_task_setscheduler(p
);
3941 cpuset_cpus_allowed(p
, cpus_allowed
);
3942 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3945 * Since bandwidth control happens on root_domain basis,
3946 * if admission test is enabled, we only admit -deadline
3947 * tasks allowed to run on all the CPUs in the task's
3951 if (task_has_dl_policy(p
)) {
3952 const struct cpumask
*span
= task_rq(p
)->rd
->span
;
3954 if (dl_bandwidth_enabled() && !cpumask_subset(span
, new_mask
)) {
3961 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3964 cpuset_cpus_allowed(p
, cpus_allowed
);
3965 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3967 * We must have raced with a concurrent cpuset
3968 * update. Just reset the cpus_allowed to the
3969 * cpuset's cpus_allowed
3971 cpumask_copy(new_mask
, cpus_allowed
);
3976 free_cpumask_var(new_mask
);
3977 out_free_cpus_allowed
:
3978 free_cpumask_var(cpus_allowed
);
3984 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3985 struct cpumask
*new_mask
)
3987 if (len
< cpumask_size())
3988 cpumask_clear(new_mask
);
3989 else if (len
> cpumask_size())
3990 len
= cpumask_size();
3992 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3996 * sys_sched_setaffinity - set the cpu affinity of a process
3997 * @pid: pid of the process
3998 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3999 * @user_mask_ptr: user-space pointer to the new cpu mask
4001 * Return: 0 on success. An error code otherwise.
4003 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4004 unsigned long __user
*, user_mask_ptr
)
4006 cpumask_var_t new_mask
;
4009 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4012 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4014 retval
= sched_setaffinity(pid
, new_mask
);
4015 free_cpumask_var(new_mask
);
4019 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4021 struct task_struct
*p
;
4022 unsigned long flags
;
4028 p
= find_process_by_pid(pid
);
4032 retval
= security_task_getscheduler(p
);
4036 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4037 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4038 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4047 * sys_sched_getaffinity - get the cpu affinity of a process
4048 * @pid: pid of the process
4049 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4050 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4052 * Return: 0 on success. An error code otherwise.
4054 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4055 unsigned long __user
*, user_mask_ptr
)
4060 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4062 if (len
& (sizeof(unsigned long)-1))
4065 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4068 ret
= sched_getaffinity(pid
, mask
);
4070 size_t retlen
= min_t(size_t, len
, cpumask_size());
4072 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4077 free_cpumask_var(mask
);
4083 * sys_sched_yield - yield the current processor to other threads.
4085 * This function yields the current CPU to other tasks. If there are no
4086 * other threads running on this CPU then this function will return.
4090 SYSCALL_DEFINE0(sched_yield
)
4092 struct rq
*rq
= this_rq_lock();
4094 schedstat_inc(rq
, yld_count
);
4095 current
->sched_class
->yield_task(rq
);
4098 * Since we are going to call schedule() anyway, there's
4099 * no need to preempt or enable interrupts:
4101 __release(rq
->lock
);
4102 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4103 do_raw_spin_unlock(&rq
->lock
);
4104 sched_preempt_enable_no_resched();
4111 static void __cond_resched(void)
4113 __preempt_count_add(PREEMPT_ACTIVE
);
4115 __preempt_count_sub(PREEMPT_ACTIVE
);
4118 int __sched
_cond_resched(void)
4121 if (should_resched()) {
4127 EXPORT_SYMBOL(_cond_resched
);
4130 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4131 * call schedule, and on return reacquire the lock.
4133 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4134 * operations here to prevent schedule() from being called twice (once via
4135 * spin_unlock(), once by hand).
4137 int __cond_resched_lock(spinlock_t
*lock
)
4139 bool need_rcu_resched
= rcu_should_resched();
4140 int resched
= should_resched();
4143 lockdep_assert_held(lock
);
4145 if (spin_needbreak(lock
) || resched
|| need_rcu_resched
) {
4149 else if (unlikely(need_rcu_resched
))
4158 EXPORT_SYMBOL(__cond_resched_lock
);
4160 int __sched
__cond_resched_softirq(void)
4162 BUG_ON(!in_softirq());
4164 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */
4165 if (should_resched()) {
4173 EXPORT_SYMBOL(__cond_resched_softirq
);
4176 * yield - yield the current processor to other threads.
4178 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4180 * The scheduler is at all times free to pick the calling task as the most
4181 * eligible task to run, if removing the yield() call from your code breaks
4182 * it, its already broken.
4184 * Typical broken usage is:
4189 * where one assumes that yield() will let 'the other' process run that will
4190 * make event true. If the current task is a SCHED_FIFO task that will never
4191 * happen. Never use yield() as a progress guarantee!!
4193 * If you want to use yield() to wait for something, use wait_event().
4194 * If you want to use yield() to be 'nice' for others, use cond_resched().
4195 * If you still want to use yield(), do not!
4197 void __sched
yield(void)
4199 set_current_state(TASK_RUNNING
);
4202 EXPORT_SYMBOL(yield
);
4205 * yield_to - yield the current processor to another thread in
4206 * your thread group, or accelerate that thread toward the
4207 * processor it's on.
4209 * @preempt: whether task preemption is allowed or not
4211 * It's the caller's job to ensure that the target task struct
4212 * can't go away on us before we can do any checks.
4215 * true (>0) if we indeed boosted the target task.
4216 * false (0) if we failed to boost the target.
4217 * -ESRCH if there's no task to yield to.
4219 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4221 struct task_struct
*curr
= current
;
4222 struct rq
*rq
, *p_rq
;
4223 unsigned long flags
;
4226 local_irq_save(flags
);
4232 * If we're the only runnable task on the rq and target rq also
4233 * has only one task, there's absolutely no point in yielding.
4235 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4240 double_rq_lock(rq
, p_rq
);
4241 if (task_rq(p
) != p_rq
) {
4242 double_rq_unlock(rq
, p_rq
);
4246 if (!curr
->sched_class
->yield_to_task
)
4249 if (curr
->sched_class
!= p
->sched_class
)
4252 if (task_running(p_rq
, p
) || p
->state
)
4255 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4257 schedstat_inc(rq
, yld_count
);
4259 * Make p's CPU reschedule; pick_next_entity takes care of
4262 if (preempt
&& rq
!= p_rq
)
4263 resched_task(p_rq
->curr
);
4267 double_rq_unlock(rq
, p_rq
);
4269 local_irq_restore(flags
);
4276 EXPORT_SYMBOL_GPL(yield_to
);
4279 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4280 * that process accounting knows that this is a task in IO wait state.
4282 void __sched
io_schedule(void)
4284 struct rq
*rq
= raw_rq();
4286 delayacct_blkio_start();
4287 atomic_inc(&rq
->nr_iowait
);
4288 blk_flush_plug(current
);
4289 current
->in_iowait
= 1;
4291 current
->in_iowait
= 0;
4292 atomic_dec(&rq
->nr_iowait
);
4293 delayacct_blkio_end();
4295 EXPORT_SYMBOL(io_schedule
);
4297 long __sched
io_schedule_timeout(long timeout
)
4299 struct rq
*rq
= raw_rq();
4302 delayacct_blkio_start();
4303 atomic_inc(&rq
->nr_iowait
);
4304 blk_flush_plug(current
);
4305 current
->in_iowait
= 1;
4306 ret
= schedule_timeout(timeout
);
4307 current
->in_iowait
= 0;
4308 atomic_dec(&rq
->nr_iowait
);
4309 delayacct_blkio_end();
4314 * sys_sched_get_priority_max - return maximum RT priority.
4315 * @policy: scheduling class.
4317 * Return: On success, this syscall returns the maximum
4318 * rt_priority that can be used by a given scheduling class.
4319 * On failure, a negative error code is returned.
4321 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4328 ret
= MAX_USER_RT_PRIO
-1;
4330 case SCHED_DEADLINE
:
4341 * sys_sched_get_priority_min - return minimum RT priority.
4342 * @policy: scheduling class.
4344 * Return: On success, this syscall returns the minimum
4345 * rt_priority that can be used by a given scheduling class.
4346 * On failure, a negative error code is returned.
4348 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4357 case SCHED_DEADLINE
:
4367 * sys_sched_rr_get_interval - return the default timeslice of a process.
4368 * @pid: pid of the process.
4369 * @interval: userspace pointer to the timeslice value.
4371 * this syscall writes the default timeslice value of a given process
4372 * into the user-space timespec buffer. A value of '0' means infinity.
4374 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4377 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4378 struct timespec __user
*, interval
)
4380 struct task_struct
*p
;
4381 unsigned int time_slice
;
4382 unsigned long flags
;
4392 p
= find_process_by_pid(pid
);
4396 retval
= security_task_getscheduler(p
);
4400 rq
= task_rq_lock(p
, &flags
);
4402 if (p
->sched_class
->get_rr_interval
)
4403 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4404 task_rq_unlock(rq
, p
, &flags
);
4407 jiffies_to_timespec(time_slice
, &t
);
4408 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4416 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4418 void sched_show_task(struct task_struct
*p
)
4420 unsigned long free
= 0;
4424 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4425 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4426 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4427 #if BITS_PER_LONG == 32
4428 if (state
== TASK_RUNNING
)
4429 printk(KERN_CONT
" running ");
4431 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4433 if (state
== TASK_RUNNING
)
4434 printk(KERN_CONT
" running task ");
4436 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4438 #ifdef CONFIG_DEBUG_STACK_USAGE
4439 free
= stack_not_used(p
);
4442 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4444 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4445 task_pid_nr(p
), ppid
,
4446 (unsigned long)task_thread_info(p
)->flags
);
4448 print_worker_info(KERN_INFO
, p
);
4449 show_stack(p
, NULL
);
4452 void show_state_filter(unsigned long state_filter
)
4454 struct task_struct
*g
, *p
;
4456 #if BITS_PER_LONG == 32
4458 " task PC stack pid father\n");
4461 " task PC stack pid father\n");
4464 do_each_thread(g
, p
) {
4466 * reset the NMI-timeout, listing all files on a slow
4467 * console might take a lot of time:
4469 touch_nmi_watchdog();
4470 if (!state_filter
|| (p
->state
& state_filter
))
4472 } while_each_thread(g
, p
);
4474 touch_all_softlockup_watchdogs();
4476 #ifdef CONFIG_SCHED_DEBUG
4477 sysrq_sched_debug_show();
4481 * Only show locks if all tasks are dumped:
4484 debug_show_all_locks();
4487 void init_idle_bootup_task(struct task_struct
*idle
)
4489 idle
->sched_class
= &idle_sched_class
;
4493 * init_idle - set up an idle thread for a given CPU
4494 * @idle: task in question
4495 * @cpu: cpu the idle task belongs to
4497 * NOTE: this function does not set the idle thread's NEED_RESCHED
4498 * flag, to make booting more robust.
4500 void init_idle(struct task_struct
*idle
, int cpu
)
4502 struct rq
*rq
= cpu_rq(cpu
);
4503 unsigned long flags
;
4505 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4507 __sched_fork(0, idle
);
4508 idle
->state
= TASK_RUNNING
;
4509 idle
->se
.exec_start
= sched_clock();
4511 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4513 * We're having a chicken and egg problem, even though we are
4514 * holding rq->lock, the cpu isn't yet set to this cpu so the
4515 * lockdep check in task_group() will fail.
4517 * Similar case to sched_fork(). / Alternatively we could
4518 * use task_rq_lock() here and obtain the other rq->lock.
4523 __set_task_cpu(idle
, cpu
);
4526 rq
->curr
= rq
->idle
= idle
;
4528 #if defined(CONFIG_SMP)
4531 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4533 /* Set the preempt count _outside_ the spinlocks! */
4534 init_idle_preempt_count(idle
, cpu
);
4537 * The idle tasks have their own, simple scheduling class:
4539 idle
->sched_class
= &idle_sched_class
;
4540 ftrace_graph_init_idle_task(idle
, cpu
);
4541 vtime_init_idle(idle
, cpu
);
4542 #if defined(CONFIG_SMP)
4543 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4548 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4550 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4551 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4553 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4554 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4558 * This is how migration works:
4560 * 1) we invoke migration_cpu_stop() on the target CPU using
4562 * 2) stopper starts to run (implicitly forcing the migrated thread
4564 * 3) it checks whether the migrated task is still in the wrong runqueue.
4565 * 4) if it's in the wrong runqueue then the migration thread removes
4566 * it and puts it into the right queue.
4567 * 5) stopper completes and stop_one_cpu() returns and the migration
4572 * Change a given task's CPU affinity. Migrate the thread to a
4573 * proper CPU and schedule it away if the CPU it's executing on
4574 * is removed from the allowed bitmask.
4576 * NOTE: the caller must have a valid reference to the task, the
4577 * task must not exit() & deallocate itself prematurely. The
4578 * call is not atomic; no spinlocks may be held.
4580 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4582 unsigned long flags
;
4584 unsigned int dest_cpu
;
4587 rq
= task_rq_lock(p
, &flags
);
4589 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4592 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4597 do_set_cpus_allowed(p
, new_mask
);
4599 /* Can the task run on the task's current CPU? If so, we're done */
4600 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4603 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4605 struct migration_arg arg
= { p
, dest_cpu
};
4606 /* Need help from migration thread: drop lock and wait. */
4607 task_rq_unlock(rq
, p
, &flags
);
4608 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4609 tlb_migrate_finish(p
->mm
);
4613 task_rq_unlock(rq
, p
, &flags
);
4617 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4620 * Move (not current) task off this cpu, onto dest cpu. We're doing
4621 * this because either it can't run here any more (set_cpus_allowed()
4622 * away from this CPU, or CPU going down), or because we're
4623 * attempting to rebalance this task on exec (sched_exec).
4625 * So we race with normal scheduler movements, but that's OK, as long
4626 * as the task is no longer on this CPU.
4628 * Returns non-zero if task was successfully migrated.
4630 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4632 struct rq
*rq_dest
, *rq_src
;
4635 if (unlikely(!cpu_active(dest_cpu
)))
4638 rq_src
= cpu_rq(src_cpu
);
4639 rq_dest
= cpu_rq(dest_cpu
);
4641 raw_spin_lock(&p
->pi_lock
);
4642 double_rq_lock(rq_src
, rq_dest
);
4643 /* Already moved. */
4644 if (task_cpu(p
) != src_cpu
)
4646 /* Affinity changed (again). */
4647 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4651 * If we're not on a rq, the next wake-up will ensure we're
4655 dequeue_task(rq_src
, p
, 0);
4656 set_task_cpu(p
, dest_cpu
);
4657 enqueue_task(rq_dest
, p
, 0);
4658 check_preempt_curr(rq_dest
, p
, 0);
4663 double_rq_unlock(rq_src
, rq_dest
);
4664 raw_spin_unlock(&p
->pi_lock
);
4668 #ifdef CONFIG_NUMA_BALANCING
4669 /* Migrate current task p to target_cpu */
4670 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4672 struct migration_arg arg
= { p
, target_cpu
};
4673 int curr_cpu
= task_cpu(p
);
4675 if (curr_cpu
== target_cpu
)
4678 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4681 /* TODO: This is not properly updating schedstats */
4683 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4684 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4688 * Requeue a task on a given node and accurately track the number of NUMA
4689 * tasks on the runqueues
4691 void sched_setnuma(struct task_struct
*p
, int nid
)
4694 unsigned long flags
;
4695 bool on_rq
, running
;
4697 rq
= task_rq_lock(p
, &flags
);
4699 running
= task_current(rq
, p
);
4702 dequeue_task(rq
, p
, 0);
4704 p
->sched_class
->put_prev_task(rq
, p
);
4706 p
->numa_preferred_nid
= nid
;
4709 p
->sched_class
->set_curr_task(rq
);
4711 enqueue_task(rq
, p
, 0);
4712 task_rq_unlock(rq
, p
, &flags
);
4717 * migration_cpu_stop - this will be executed by a highprio stopper thread
4718 * and performs thread migration by bumping thread off CPU then
4719 * 'pushing' onto another runqueue.
4721 static int migration_cpu_stop(void *data
)
4723 struct migration_arg
*arg
= data
;
4726 * The original target cpu might have gone down and we might
4727 * be on another cpu but it doesn't matter.
4729 local_irq_disable();
4730 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4735 #ifdef CONFIG_HOTPLUG_CPU
4738 * Ensures that the idle task is using init_mm right before its cpu goes
4741 void idle_task_exit(void)
4743 struct mm_struct
*mm
= current
->active_mm
;
4745 BUG_ON(cpu_online(smp_processor_id()));
4747 if (mm
!= &init_mm
) {
4748 switch_mm(mm
, &init_mm
, current
);
4749 finish_arch_post_lock_switch();
4755 * Since this CPU is going 'away' for a while, fold any nr_active delta
4756 * we might have. Assumes we're called after migrate_tasks() so that the
4757 * nr_active count is stable.
4759 * Also see the comment "Global load-average calculations".
4761 static void calc_load_migrate(struct rq
*rq
)
4763 long delta
= calc_load_fold_active(rq
);
4765 atomic_long_add(delta
, &calc_load_tasks
);
4768 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4772 static const struct sched_class fake_sched_class
= {
4773 .put_prev_task
= put_prev_task_fake
,
4776 static struct task_struct fake_task
= {
4778 * Avoid pull_{rt,dl}_task()
4780 .prio
= MAX_PRIO
+ 1,
4781 .sched_class
= &fake_sched_class
,
4785 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4786 * try_to_wake_up()->select_task_rq().
4788 * Called with rq->lock held even though we'er in stop_machine() and
4789 * there's no concurrency possible, we hold the required locks anyway
4790 * because of lock validation efforts.
4792 static void migrate_tasks(unsigned int dead_cpu
)
4794 struct rq
*rq
= cpu_rq(dead_cpu
);
4795 struct task_struct
*next
, *stop
= rq
->stop
;
4799 * Fudge the rq selection such that the below task selection loop
4800 * doesn't get stuck on the currently eligible stop task.
4802 * We're currently inside stop_machine() and the rq is either stuck
4803 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4804 * either way we should never end up calling schedule() until we're
4810 * put_prev_task() and pick_next_task() sched
4811 * class method both need to have an up-to-date
4812 * value of rq->clock[_task]
4814 update_rq_clock(rq
);
4818 * There's this thread running, bail when that's the only
4821 if (rq
->nr_running
== 1)
4824 next
= pick_next_task(rq
, &fake_task
);
4826 next
->sched_class
->put_prev_task(rq
, next
);
4828 /* Find suitable destination for @next, with force if needed. */
4829 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4830 raw_spin_unlock(&rq
->lock
);
4832 __migrate_task(next
, dead_cpu
, dest_cpu
);
4834 raw_spin_lock(&rq
->lock
);
4840 #endif /* CONFIG_HOTPLUG_CPU */
4842 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4844 static struct ctl_table sd_ctl_dir
[] = {
4846 .procname
= "sched_domain",
4852 static struct ctl_table sd_ctl_root
[] = {
4854 .procname
= "kernel",
4856 .child
= sd_ctl_dir
,
4861 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4863 struct ctl_table
*entry
=
4864 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4869 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4871 struct ctl_table
*entry
;
4874 * In the intermediate directories, both the child directory and
4875 * procname are dynamically allocated and could fail but the mode
4876 * will always be set. In the lowest directory the names are
4877 * static strings and all have proc handlers.
4879 for (entry
= *tablep
; entry
->mode
; entry
++) {
4881 sd_free_ctl_entry(&entry
->child
);
4882 if (entry
->proc_handler
== NULL
)
4883 kfree(entry
->procname
);
4890 static int min_load_idx
= 0;
4891 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4894 set_table_entry(struct ctl_table
*entry
,
4895 const char *procname
, void *data
, int maxlen
,
4896 umode_t mode
, proc_handler
*proc_handler
,
4899 entry
->procname
= procname
;
4901 entry
->maxlen
= maxlen
;
4903 entry
->proc_handler
= proc_handler
;
4906 entry
->extra1
= &min_load_idx
;
4907 entry
->extra2
= &max_load_idx
;
4911 static struct ctl_table
*
4912 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4914 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
4919 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4920 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4921 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4922 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4923 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4924 sizeof(int), 0644, proc_dointvec_minmax
, true);
4925 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4926 sizeof(int), 0644, proc_dointvec_minmax
, true);
4927 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4928 sizeof(int), 0644, proc_dointvec_minmax
, true);
4929 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4930 sizeof(int), 0644, proc_dointvec_minmax
, true);
4931 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4932 sizeof(int), 0644, proc_dointvec_minmax
, true);
4933 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4934 sizeof(int), 0644, proc_dointvec_minmax
, false);
4935 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4936 sizeof(int), 0644, proc_dointvec_minmax
, false);
4937 set_table_entry(&table
[9], "cache_nice_tries",
4938 &sd
->cache_nice_tries
,
4939 sizeof(int), 0644, proc_dointvec_minmax
, false);
4940 set_table_entry(&table
[10], "flags", &sd
->flags
,
4941 sizeof(int), 0644, proc_dointvec_minmax
, false);
4942 set_table_entry(&table
[11], "max_newidle_lb_cost",
4943 &sd
->max_newidle_lb_cost
,
4944 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4945 set_table_entry(&table
[12], "name", sd
->name
,
4946 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4947 /* &table[13] is terminator */
4952 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4954 struct ctl_table
*entry
, *table
;
4955 struct sched_domain
*sd
;
4956 int domain_num
= 0, i
;
4959 for_each_domain(cpu
, sd
)
4961 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4966 for_each_domain(cpu
, sd
) {
4967 snprintf(buf
, 32, "domain%d", i
);
4968 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4970 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4977 static struct ctl_table_header
*sd_sysctl_header
;
4978 static void register_sched_domain_sysctl(void)
4980 int i
, cpu_num
= num_possible_cpus();
4981 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4984 WARN_ON(sd_ctl_dir
[0].child
);
4985 sd_ctl_dir
[0].child
= entry
;
4990 for_each_possible_cpu(i
) {
4991 snprintf(buf
, 32, "cpu%d", i
);
4992 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4994 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4998 WARN_ON(sd_sysctl_header
);
4999 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5002 /* may be called multiple times per register */
5003 static void unregister_sched_domain_sysctl(void)
5005 if (sd_sysctl_header
)
5006 unregister_sysctl_table(sd_sysctl_header
);
5007 sd_sysctl_header
= NULL
;
5008 if (sd_ctl_dir
[0].child
)
5009 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5012 static void register_sched_domain_sysctl(void)
5015 static void unregister_sched_domain_sysctl(void)
5020 static void set_rq_online(struct rq
*rq
)
5023 const struct sched_class
*class;
5025 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5028 for_each_class(class) {
5029 if (class->rq_online
)
5030 class->rq_online(rq
);
5035 static void set_rq_offline(struct rq
*rq
)
5038 const struct sched_class
*class;
5040 for_each_class(class) {
5041 if (class->rq_offline
)
5042 class->rq_offline(rq
);
5045 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5051 * migration_call - callback that gets triggered when a CPU is added.
5052 * Here we can start up the necessary migration thread for the new CPU.
5055 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5057 int cpu
= (long)hcpu
;
5058 unsigned long flags
;
5059 struct rq
*rq
= cpu_rq(cpu
);
5061 switch (action
& ~CPU_TASKS_FROZEN
) {
5063 case CPU_UP_PREPARE
:
5064 rq
->calc_load_update
= calc_load_update
;
5068 /* Update our root-domain */
5069 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5071 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5075 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5078 #ifdef CONFIG_HOTPLUG_CPU
5080 sched_ttwu_pending();
5081 /* Update our root-domain */
5082 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5084 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5088 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5089 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5093 calc_load_migrate(rq
);
5098 update_max_interval();
5104 * Register at high priority so that task migration (migrate_all_tasks)
5105 * happens before everything else. This has to be lower priority than
5106 * the notifier in the perf_event subsystem, though.
5108 static struct notifier_block migration_notifier
= {
5109 .notifier_call
= migration_call
,
5110 .priority
= CPU_PRI_MIGRATION
,
5113 static void __cpuinit
set_cpu_rq_start_time(void)
5115 int cpu
= smp_processor_id();
5116 struct rq
*rq
= cpu_rq(cpu
);
5117 rq
->age_stamp
= sched_clock_cpu(cpu
);
5120 static int sched_cpu_active(struct notifier_block
*nfb
,
5121 unsigned long action
, void *hcpu
)
5123 switch (action
& ~CPU_TASKS_FROZEN
) {
5125 set_cpu_rq_start_time();
5127 case CPU_DOWN_FAILED
:
5128 set_cpu_active((long)hcpu
, true);
5135 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5136 unsigned long action
, void *hcpu
)
5138 unsigned long flags
;
5139 long cpu
= (long)hcpu
;
5141 switch (action
& ~CPU_TASKS_FROZEN
) {
5142 case CPU_DOWN_PREPARE
:
5143 set_cpu_active(cpu
, false);
5145 /* explicitly allow suspend */
5146 if (!(action
& CPU_TASKS_FROZEN
)) {
5147 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
5151 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5152 cpus
= dl_bw_cpus(cpu
);
5153 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5154 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5157 return notifier_from_errno(-EBUSY
);
5165 static int __init
migration_init(void)
5167 void *cpu
= (void *)(long)smp_processor_id();
5170 /* Initialize migration for the boot CPU */
5171 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5172 BUG_ON(err
== NOTIFY_BAD
);
5173 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5174 register_cpu_notifier(&migration_notifier
);
5176 /* Register cpu active notifiers */
5177 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5178 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5182 early_initcall(migration_init
);
5187 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5189 #ifdef CONFIG_SCHED_DEBUG
5191 static __read_mostly
int sched_debug_enabled
;
5193 static int __init
sched_debug_setup(char *str
)
5195 sched_debug_enabled
= 1;
5199 early_param("sched_debug", sched_debug_setup
);
5201 static inline bool sched_debug(void)
5203 return sched_debug_enabled
;
5206 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5207 struct cpumask
*groupmask
)
5209 struct sched_group
*group
= sd
->groups
;
5212 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5213 cpumask_clear(groupmask
);
5215 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5217 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5218 printk("does not load-balance\n");
5220 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5225 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5227 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5228 printk(KERN_ERR
"ERROR: domain->span does not contain "
5231 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5232 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5236 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5240 printk(KERN_ERR
"ERROR: group is NULL\n");
5245 * Even though we initialize ->power to something semi-sane,
5246 * we leave power_orig unset. This allows us to detect if
5247 * domain iteration is still funny without causing /0 traps.
5249 if (!group
->sgp
->power_orig
) {
5250 printk(KERN_CONT
"\n");
5251 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5256 if (!cpumask_weight(sched_group_cpus(group
))) {
5257 printk(KERN_CONT
"\n");
5258 printk(KERN_ERR
"ERROR: empty group\n");
5262 if (!(sd
->flags
& SD_OVERLAP
) &&
5263 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5264 printk(KERN_CONT
"\n");
5265 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5269 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5271 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5273 printk(KERN_CONT
" %s", str
);
5274 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5275 printk(KERN_CONT
" (cpu_power = %d)",
5279 group
= group
->next
;
5280 } while (group
!= sd
->groups
);
5281 printk(KERN_CONT
"\n");
5283 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5284 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5287 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5288 printk(KERN_ERR
"ERROR: parent span is not a superset "
5289 "of domain->span\n");
5293 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5297 if (!sched_debug_enabled
)
5301 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5305 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5308 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5316 #else /* !CONFIG_SCHED_DEBUG */
5317 # define sched_domain_debug(sd, cpu) do { } while (0)
5318 static inline bool sched_debug(void)
5322 #endif /* CONFIG_SCHED_DEBUG */
5324 static int sd_degenerate(struct sched_domain
*sd
)
5326 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5329 /* Following flags need at least 2 groups */
5330 if (sd
->flags
& (SD_LOAD_BALANCE
|
5331 SD_BALANCE_NEWIDLE
|
5335 SD_SHARE_PKG_RESOURCES
|
5336 SD_SHARE_POWERDOMAIN
)) {
5337 if (sd
->groups
!= sd
->groups
->next
)
5341 /* Following flags don't use groups */
5342 if (sd
->flags
& (SD_WAKE_AFFINE
))
5349 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5351 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5353 if (sd_degenerate(parent
))
5356 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5359 /* Flags needing groups don't count if only 1 group in parent */
5360 if (parent
->groups
== parent
->groups
->next
) {
5361 pflags
&= ~(SD_LOAD_BALANCE
|
5362 SD_BALANCE_NEWIDLE
|
5366 SD_SHARE_PKG_RESOURCES
|
5368 SD_SHARE_POWERDOMAIN
);
5369 if (nr_node_ids
== 1)
5370 pflags
&= ~SD_SERIALIZE
;
5372 if (~cflags
& pflags
)
5378 static void free_rootdomain(struct rcu_head
*rcu
)
5380 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5382 cpupri_cleanup(&rd
->cpupri
);
5383 cpudl_cleanup(&rd
->cpudl
);
5384 free_cpumask_var(rd
->dlo_mask
);
5385 free_cpumask_var(rd
->rto_mask
);
5386 free_cpumask_var(rd
->online
);
5387 free_cpumask_var(rd
->span
);
5391 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5393 struct root_domain
*old_rd
= NULL
;
5394 unsigned long flags
;
5396 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5401 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5404 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5407 * If we dont want to free the old_rd yet then
5408 * set old_rd to NULL to skip the freeing later
5411 if (!atomic_dec_and_test(&old_rd
->refcount
))
5415 atomic_inc(&rd
->refcount
);
5418 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5419 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5422 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5425 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5428 static int init_rootdomain(struct root_domain
*rd
)
5430 memset(rd
, 0, sizeof(*rd
));
5432 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5434 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5436 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5438 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5441 init_dl_bw(&rd
->dl_bw
);
5442 if (cpudl_init(&rd
->cpudl
) != 0)
5445 if (cpupri_init(&rd
->cpupri
) != 0)
5450 free_cpumask_var(rd
->rto_mask
);
5452 free_cpumask_var(rd
->dlo_mask
);
5454 free_cpumask_var(rd
->online
);
5456 free_cpumask_var(rd
->span
);
5462 * By default the system creates a single root-domain with all cpus as
5463 * members (mimicking the global state we have today).
5465 struct root_domain def_root_domain
;
5467 static void init_defrootdomain(void)
5469 init_rootdomain(&def_root_domain
);
5471 atomic_set(&def_root_domain
.refcount
, 1);
5474 static struct root_domain
*alloc_rootdomain(void)
5476 struct root_domain
*rd
;
5478 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5482 if (init_rootdomain(rd
) != 0) {
5490 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5492 struct sched_group
*tmp
, *first
;
5501 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5506 } while (sg
!= first
);
5509 static void free_sched_domain(struct rcu_head
*rcu
)
5511 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5514 * If its an overlapping domain it has private groups, iterate and
5517 if (sd
->flags
& SD_OVERLAP
) {
5518 free_sched_groups(sd
->groups
, 1);
5519 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5520 kfree(sd
->groups
->sgp
);
5526 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5528 call_rcu(&sd
->rcu
, free_sched_domain
);
5531 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5533 for (; sd
; sd
= sd
->parent
)
5534 destroy_sched_domain(sd
, cpu
);
5538 * Keep a special pointer to the highest sched_domain that has
5539 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5540 * allows us to avoid some pointer chasing select_idle_sibling().
5542 * Also keep a unique ID per domain (we use the first cpu number in
5543 * the cpumask of the domain), this allows us to quickly tell if
5544 * two cpus are in the same cache domain, see cpus_share_cache().
5546 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5547 DEFINE_PER_CPU(int, sd_llc_size
);
5548 DEFINE_PER_CPU(int, sd_llc_id
);
5549 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5550 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5551 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5553 static void update_top_cache_domain(int cpu
)
5555 struct sched_domain
*sd
;
5556 struct sched_domain
*busy_sd
= NULL
;
5560 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5562 id
= cpumask_first(sched_domain_span(sd
));
5563 size
= cpumask_weight(sched_domain_span(sd
));
5564 busy_sd
= sd
->parent
; /* sd_busy */
5566 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5568 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5569 per_cpu(sd_llc_size
, cpu
) = size
;
5570 per_cpu(sd_llc_id
, cpu
) = id
;
5572 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5573 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5575 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5576 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5581 * hold the hotplug lock.
5584 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5586 struct rq
*rq
= cpu_rq(cpu
);
5587 struct sched_domain
*tmp
;
5589 /* Remove the sched domains which do not contribute to scheduling. */
5590 for (tmp
= sd
; tmp
; ) {
5591 struct sched_domain
*parent
= tmp
->parent
;
5595 if (sd_parent_degenerate(tmp
, parent
)) {
5596 tmp
->parent
= parent
->parent
;
5598 parent
->parent
->child
= tmp
;
5600 * Transfer SD_PREFER_SIBLING down in case of a
5601 * degenerate parent; the spans match for this
5602 * so the property transfers.
5604 if (parent
->flags
& SD_PREFER_SIBLING
)
5605 tmp
->flags
|= SD_PREFER_SIBLING
;
5606 destroy_sched_domain(parent
, cpu
);
5611 if (sd
&& sd_degenerate(sd
)) {
5614 destroy_sched_domain(tmp
, cpu
);
5619 sched_domain_debug(sd
, cpu
);
5621 rq_attach_root(rq
, rd
);
5623 rcu_assign_pointer(rq
->sd
, sd
);
5624 destroy_sched_domains(tmp
, cpu
);
5626 update_top_cache_domain(cpu
);
5629 /* cpus with isolated domains */
5630 static cpumask_var_t cpu_isolated_map
;
5632 /* Setup the mask of cpus configured for isolated domains */
5633 static int __init
isolated_cpu_setup(char *str
)
5635 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5636 cpulist_parse(str
, cpu_isolated_map
);
5640 __setup("isolcpus=", isolated_cpu_setup
);
5643 struct sched_domain
** __percpu sd
;
5644 struct root_domain
*rd
;
5655 * Build an iteration mask that can exclude certain CPUs from the upwards
5658 * Asymmetric node setups can result in situations where the domain tree is of
5659 * unequal depth, make sure to skip domains that already cover the entire
5662 * In that case build_sched_domains() will have terminated the iteration early
5663 * and our sibling sd spans will be empty. Domains should always include the
5664 * cpu they're built on, so check that.
5667 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5669 const struct cpumask
*span
= sched_domain_span(sd
);
5670 struct sd_data
*sdd
= sd
->private;
5671 struct sched_domain
*sibling
;
5674 for_each_cpu(i
, span
) {
5675 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5676 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5679 cpumask_set_cpu(i
, sched_group_mask(sg
));
5684 * Return the canonical balance cpu for this group, this is the first cpu
5685 * of this group that's also in the iteration mask.
5687 int group_balance_cpu(struct sched_group
*sg
)
5689 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5693 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5695 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5696 const struct cpumask
*span
= sched_domain_span(sd
);
5697 struct cpumask
*covered
= sched_domains_tmpmask
;
5698 struct sd_data
*sdd
= sd
->private;
5699 struct sched_domain
*child
;
5702 cpumask_clear(covered
);
5704 for_each_cpu(i
, span
) {
5705 struct cpumask
*sg_span
;
5707 if (cpumask_test_cpu(i
, covered
))
5710 child
= *per_cpu_ptr(sdd
->sd
, i
);
5712 /* See the comment near build_group_mask(). */
5713 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5716 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5717 GFP_KERNEL
, cpu_to_node(cpu
));
5722 sg_span
= sched_group_cpus(sg
);
5724 child
= child
->child
;
5725 cpumask_copy(sg_span
, sched_domain_span(child
));
5727 cpumask_set_cpu(i
, sg_span
);
5729 cpumask_or(covered
, covered
, sg_span
);
5731 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5732 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5733 build_group_mask(sd
, sg
);
5736 * Initialize sgp->power such that even if we mess up the
5737 * domains and no possible iteration will get us here, we won't
5740 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5741 sg
->sgp
->power_orig
= sg
->sgp
->power
;
5744 * Make sure the first group of this domain contains the
5745 * canonical balance cpu. Otherwise the sched_domain iteration
5746 * breaks. See update_sg_lb_stats().
5748 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5749 group_balance_cpu(sg
) == cpu
)
5759 sd
->groups
= groups
;
5764 free_sched_groups(first
, 0);
5769 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5771 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5772 struct sched_domain
*child
= sd
->child
;
5775 cpu
= cpumask_first(sched_domain_span(child
));
5778 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5779 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5780 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5787 * build_sched_groups will build a circular linked list of the groups
5788 * covered by the given span, and will set each group's ->cpumask correctly,
5789 * and ->cpu_power to 0.
5791 * Assumes the sched_domain tree is fully constructed
5794 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5796 struct sched_group
*first
= NULL
, *last
= NULL
;
5797 struct sd_data
*sdd
= sd
->private;
5798 const struct cpumask
*span
= sched_domain_span(sd
);
5799 struct cpumask
*covered
;
5802 get_group(cpu
, sdd
, &sd
->groups
);
5803 atomic_inc(&sd
->groups
->ref
);
5805 if (cpu
!= cpumask_first(span
))
5808 lockdep_assert_held(&sched_domains_mutex
);
5809 covered
= sched_domains_tmpmask
;
5811 cpumask_clear(covered
);
5813 for_each_cpu(i
, span
) {
5814 struct sched_group
*sg
;
5817 if (cpumask_test_cpu(i
, covered
))
5820 group
= get_group(i
, sdd
, &sg
);
5821 cpumask_setall(sched_group_mask(sg
));
5823 for_each_cpu(j
, span
) {
5824 if (get_group(j
, sdd
, NULL
) != group
)
5827 cpumask_set_cpu(j
, covered
);
5828 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5843 * Initialize sched groups cpu_power.
5845 * cpu_power indicates the capacity of sched group, which is used while
5846 * distributing the load between different sched groups in a sched domain.
5847 * Typically cpu_power for all the groups in a sched domain will be same unless
5848 * there are asymmetries in the topology. If there are asymmetries, group
5849 * having more cpu_power will pickup more load compared to the group having
5852 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5854 struct sched_group
*sg
= sd
->groups
;
5859 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5861 } while (sg
!= sd
->groups
);
5863 if (cpu
!= group_balance_cpu(sg
))
5866 update_group_power(sd
, cpu
);
5867 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5871 * Initializers for schedule domains
5872 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5875 static int default_relax_domain_level
= -1;
5876 int sched_domain_level_max
;
5878 static int __init
setup_relax_domain_level(char *str
)
5880 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5881 pr_warn("Unable to set relax_domain_level\n");
5885 __setup("relax_domain_level=", setup_relax_domain_level
);
5887 static void set_domain_attribute(struct sched_domain
*sd
,
5888 struct sched_domain_attr
*attr
)
5892 if (!attr
|| attr
->relax_domain_level
< 0) {
5893 if (default_relax_domain_level
< 0)
5896 request
= default_relax_domain_level
;
5898 request
= attr
->relax_domain_level
;
5899 if (request
< sd
->level
) {
5900 /* turn off idle balance on this domain */
5901 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5903 /* turn on idle balance on this domain */
5904 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5908 static void __sdt_free(const struct cpumask
*cpu_map
);
5909 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5911 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5912 const struct cpumask
*cpu_map
)
5916 if (!atomic_read(&d
->rd
->refcount
))
5917 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5919 free_percpu(d
->sd
); /* fall through */
5921 __sdt_free(cpu_map
); /* fall through */
5927 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5928 const struct cpumask
*cpu_map
)
5930 memset(d
, 0, sizeof(*d
));
5932 if (__sdt_alloc(cpu_map
))
5933 return sa_sd_storage
;
5934 d
->sd
= alloc_percpu(struct sched_domain
*);
5936 return sa_sd_storage
;
5937 d
->rd
= alloc_rootdomain();
5940 return sa_rootdomain
;
5944 * NULL the sd_data elements we've used to build the sched_domain and
5945 * sched_group structure so that the subsequent __free_domain_allocs()
5946 * will not free the data we're using.
5948 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5950 struct sd_data
*sdd
= sd
->private;
5952 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5953 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5955 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5956 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5958 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5959 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5963 static int sched_domains_numa_levels
;
5964 static int *sched_domains_numa_distance
;
5965 static struct cpumask
***sched_domains_numa_masks
;
5966 static int sched_domains_curr_level
;
5970 * SD_flags allowed in topology descriptions.
5972 * SD_SHARE_CPUPOWER - describes SMT topologies
5973 * SD_SHARE_PKG_RESOURCES - describes shared caches
5974 * SD_NUMA - describes NUMA topologies
5975 * SD_SHARE_POWERDOMAIN - describes shared power domain
5978 * SD_ASYM_PACKING - describes SMT quirks
5980 #define TOPOLOGY_SD_FLAGS \
5981 (SD_SHARE_CPUPOWER | \
5982 SD_SHARE_PKG_RESOURCES | \
5985 SD_SHARE_POWERDOMAIN)
5987 static struct sched_domain
*
5988 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
5990 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5991 int sd_weight
, sd_flags
= 0;
5995 * Ugly hack to pass state to sd_numa_mask()...
5997 sched_domains_curr_level
= tl
->numa_level
;
6000 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6003 sd_flags
= (*tl
->sd_flags
)();
6004 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6005 "wrong sd_flags in topology description\n"))
6006 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6008 *sd
= (struct sched_domain
){
6009 .min_interval
= sd_weight
,
6010 .max_interval
= 2*sd_weight
,
6012 .imbalance_pct
= 125,
6014 .cache_nice_tries
= 0,
6021 .flags
= 1*SD_LOAD_BALANCE
6022 | 1*SD_BALANCE_NEWIDLE
6027 | 0*SD_SHARE_CPUPOWER
6028 | 0*SD_SHARE_PKG_RESOURCES
6030 | 0*SD_PREFER_SIBLING
6035 .last_balance
= jiffies
,
6036 .balance_interval
= sd_weight
,
6038 .max_newidle_lb_cost
= 0,
6039 .next_decay_max_lb_cost
= jiffies
,
6040 #ifdef CONFIG_SCHED_DEBUG
6046 * Convert topological properties into behaviour.
6049 if (sd
->flags
& SD_SHARE_CPUPOWER
) {
6050 sd
->imbalance_pct
= 110;
6051 sd
->smt_gain
= 1178; /* ~15% */
6053 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6054 sd
->imbalance_pct
= 117;
6055 sd
->cache_nice_tries
= 1;
6059 } else if (sd
->flags
& SD_NUMA
) {
6060 sd
->cache_nice_tries
= 2;
6064 sd
->flags
|= SD_SERIALIZE
;
6065 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6066 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6073 sd
->flags
|= SD_PREFER_SIBLING
;
6074 sd
->cache_nice_tries
= 1;
6079 sd
->private = &tl
->data
;
6085 * Topology list, bottom-up.
6087 static struct sched_domain_topology_level default_topology
[] = {
6088 #ifdef CONFIG_SCHED_SMT
6089 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6091 #ifdef CONFIG_SCHED_MC
6092 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6094 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6098 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6100 #define for_each_sd_topology(tl) \
6101 for (tl = sched_domain_topology; tl->mask; tl++)
6103 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6105 sched_domain_topology
= tl
;
6110 static const struct cpumask
*sd_numa_mask(int cpu
)
6112 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6115 static void sched_numa_warn(const char *str
)
6117 static int done
= false;
6125 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6127 for (i
= 0; i
< nr_node_ids
; i
++) {
6128 printk(KERN_WARNING
" ");
6129 for (j
= 0; j
< nr_node_ids
; j
++)
6130 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6131 printk(KERN_CONT
"\n");
6133 printk(KERN_WARNING
"\n");
6136 static bool find_numa_distance(int distance
)
6140 if (distance
== node_distance(0, 0))
6143 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6144 if (sched_domains_numa_distance
[i
] == distance
)
6151 static void sched_init_numa(void)
6153 int next_distance
, curr_distance
= node_distance(0, 0);
6154 struct sched_domain_topology_level
*tl
;
6158 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6159 if (!sched_domains_numa_distance
)
6163 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6164 * unique distances in the node_distance() table.
6166 * Assumes node_distance(0,j) includes all distances in
6167 * node_distance(i,j) in order to avoid cubic time.
6169 next_distance
= curr_distance
;
6170 for (i
= 0; i
< nr_node_ids
; i
++) {
6171 for (j
= 0; j
< nr_node_ids
; j
++) {
6172 for (k
= 0; k
< nr_node_ids
; k
++) {
6173 int distance
= node_distance(i
, k
);
6175 if (distance
> curr_distance
&&
6176 (distance
< next_distance
||
6177 next_distance
== curr_distance
))
6178 next_distance
= distance
;
6181 * While not a strong assumption it would be nice to know
6182 * about cases where if node A is connected to B, B is not
6183 * equally connected to A.
6185 if (sched_debug() && node_distance(k
, i
) != distance
)
6186 sched_numa_warn("Node-distance not symmetric");
6188 if (sched_debug() && i
&& !find_numa_distance(distance
))
6189 sched_numa_warn("Node-0 not representative");
6191 if (next_distance
!= curr_distance
) {
6192 sched_domains_numa_distance
[level
++] = next_distance
;
6193 sched_domains_numa_levels
= level
;
6194 curr_distance
= next_distance
;
6199 * In case of sched_debug() we verify the above assumption.
6205 * 'level' contains the number of unique distances, excluding the
6206 * identity distance node_distance(i,i).
6208 * The sched_domains_numa_distance[] array includes the actual distance
6213 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6214 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6215 * the array will contain less then 'level' members. This could be
6216 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6217 * in other functions.
6219 * We reset it to 'level' at the end of this function.
6221 sched_domains_numa_levels
= 0;
6223 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6224 if (!sched_domains_numa_masks
)
6228 * Now for each level, construct a mask per node which contains all
6229 * cpus of nodes that are that many hops away from us.
6231 for (i
= 0; i
< level
; i
++) {
6232 sched_domains_numa_masks
[i
] =
6233 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6234 if (!sched_domains_numa_masks
[i
])
6237 for (j
= 0; j
< nr_node_ids
; j
++) {
6238 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6242 sched_domains_numa_masks
[i
][j
] = mask
;
6244 for (k
= 0; k
< nr_node_ids
; k
++) {
6245 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6248 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6253 /* Compute default topology size */
6254 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6256 tl
= kzalloc((i
+ level
+ 1) *
6257 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6262 * Copy the default topology bits..
6264 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6265 tl
[i
] = sched_domain_topology
[i
];
6268 * .. and append 'j' levels of NUMA goodness.
6270 for (j
= 0; j
< level
; i
++, j
++) {
6271 tl
[i
] = (struct sched_domain_topology_level
){
6272 .mask
= sd_numa_mask
,
6273 .sd_flags
= cpu_numa_flags
,
6274 .flags
= SDTL_OVERLAP
,
6280 sched_domain_topology
= tl
;
6282 sched_domains_numa_levels
= level
;
6285 static void sched_domains_numa_masks_set(int cpu
)
6288 int node
= cpu_to_node(cpu
);
6290 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6291 for (j
= 0; j
< nr_node_ids
; j
++) {
6292 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6293 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6298 static void sched_domains_numa_masks_clear(int cpu
)
6301 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6302 for (j
= 0; j
< nr_node_ids
; j
++)
6303 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6308 * Update sched_domains_numa_masks[level][node] array when new cpus
6311 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6312 unsigned long action
,
6315 int cpu
= (long)hcpu
;
6317 switch (action
& ~CPU_TASKS_FROZEN
) {
6319 sched_domains_numa_masks_set(cpu
);
6323 sched_domains_numa_masks_clear(cpu
);
6333 static inline void sched_init_numa(void)
6337 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6338 unsigned long action
,
6343 #endif /* CONFIG_NUMA */
6345 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6347 struct sched_domain_topology_level
*tl
;
6350 for_each_sd_topology(tl
) {
6351 struct sd_data
*sdd
= &tl
->data
;
6353 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6357 sdd
->sg
= alloc_percpu(struct sched_group
*);
6361 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6365 for_each_cpu(j
, cpu_map
) {
6366 struct sched_domain
*sd
;
6367 struct sched_group
*sg
;
6368 struct sched_group_power
*sgp
;
6370 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6371 GFP_KERNEL
, cpu_to_node(j
));
6375 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6377 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6378 GFP_KERNEL
, cpu_to_node(j
));
6384 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6386 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6387 GFP_KERNEL
, cpu_to_node(j
));
6391 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6398 static void __sdt_free(const struct cpumask
*cpu_map
)
6400 struct sched_domain_topology_level
*tl
;
6403 for_each_sd_topology(tl
) {
6404 struct sd_data
*sdd
= &tl
->data
;
6406 for_each_cpu(j
, cpu_map
) {
6407 struct sched_domain
*sd
;
6410 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6411 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6412 free_sched_groups(sd
->groups
, 0);
6413 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6417 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6419 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6421 free_percpu(sdd
->sd
);
6423 free_percpu(sdd
->sg
);
6425 free_percpu(sdd
->sgp
);
6430 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6431 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6432 struct sched_domain
*child
, int cpu
)
6434 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6438 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6440 sd
->level
= child
->level
+ 1;
6441 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6445 set_domain_attribute(sd
, attr
);
6451 * Build sched domains for a given set of cpus and attach the sched domains
6452 * to the individual cpus
6454 static int build_sched_domains(const struct cpumask
*cpu_map
,
6455 struct sched_domain_attr
*attr
)
6457 enum s_alloc alloc_state
;
6458 struct sched_domain
*sd
;
6460 int i
, ret
= -ENOMEM
;
6462 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6463 if (alloc_state
!= sa_rootdomain
)
6466 /* Set up domains for cpus specified by the cpu_map. */
6467 for_each_cpu(i
, cpu_map
) {
6468 struct sched_domain_topology_level
*tl
;
6471 for_each_sd_topology(tl
) {
6472 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6473 if (tl
== sched_domain_topology
)
6474 *per_cpu_ptr(d
.sd
, i
) = sd
;
6475 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6476 sd
->flags
|= SD_OVERLAP
;
6477 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6482 /* Build the groups for the domains */
6483 for_each_cpu(i
, cpu_map
) {
6484 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6485 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6486 if (sd
->flags
& SD_OVERLAP
) {
6487 if (build_overlap_sched_groups(sd
, i
))
6490 if (build_sched_groups(sd
, i
))
6496 /* Calculate CPU power for physical packages and nodes */
6497 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6498 if (!cpumask_test_cpu(i
, cpu_map
))
6501 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6502 claim_allocations(i
, sd
);
6503 init_sched_groups_power(i
, sd
);
6507 /* Attach the domains */
6509 for_each_cpu(i
, cpu_map
) {
6510 sd
= *per_cpu_ptr(d
.sd
, i
);
6511 cpu_attach_domain(sd
, d
.rd
, i
);
6517 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6521 static cpumask_var_t
*doms_cur
; /* current sched domains */
6522 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6523 static struct sched_domain_attr
*dattr_cur
;
6524 /* attribues of custom domains in 'doms_cur' */
6527 * Special case: If a kmalloc of a doms_cur partition (array of
6528 * cpumask) fails, then fallback to a single sched domain,
6529 * as determined by the single cpumask fallback_doms.
6531 static cpumask_var_t fallback_doms
;
6534 * arch_update_cpu_topology lets virtualized architectures update the
6535 * cpu core maps. It is supposed to return 1 if the topology changed
6536 * or 0 if it stayed the same.
6538 int __weak
arch_update_cpu_topology(void)
6543 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6546 cpumask_var_t
*doms
;
6548 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6551 for (i
= 0; i
< ndoms
; i
++) {
6552 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6553 free_sched_domains(doms
, i
);
6560 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6563 for (i
= 0; i
< ndoms
; i
++)
6564 free_cpumask_var(doms
[i
]);
6569 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6570 * For now this just excludes isolated cpus, but could be used to
6571 * exclude other special cases in the future.
6573 static int init_sched_domains(const struct cpumask
*cpu_map
)
6577 arch_update_cpu_topology();
6579 doms_cur
= alloc_sched_domains(ndoms_cur
);
6581 doms_cur
= &fallback_doms
;
6582 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6583 err
= build_sched_domains(doms_cur
[0], NULL
);
6584 register_sched_domain_sysctl();
6590 * Detach sched domains from a group of cpus specified in cpu_map
6591 * These cpus will now be attached to the NULL domain
6593 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6598 for_each_cpu(i
, cpu_map
)
6599 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6603 /* handle null as "default" */
6604 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6605 struct sched_domain_attr
*new, int idx_new
)
6607 struct sched_domain_attr tmp
;
6614 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6615 new ? (new + idx_new
) : &tmp
,
6616 sizeof(struct sched_domain_attr
));
6620 * Partition sched domains as specified by the 'ndoms_new'
6621 * cpumasks in the array doms_new[] of cpumasks. This compares
6622 * doms_new[] to the current sched domain partitioning, doms_cur[].
6623 * It destroys each deleted domain and builds each new domain.
6625 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6626 * The masks don't intersect (don't overlap.) We should setup one
6627 * sched domain for each mask. CPUs not in any of the cpumasks will
6628 * not be load balanced. If the same cpumask appears both in the
6629 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6632 * The passed in 'doms_new' should be allocated using
6633 * alloc_sched_domains. This routine takes ownership of it and will
6634 * free_sched_domains it when done with it. If the caller failed the
6635 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6636 * and partition_sched_domains() will fallback to the single partition
6637 * 'fallback_doms', it also forces the domains to be rebuilt.
6639 * If doms_new == NULL it will be replaced with cpu_online_mask.
6640 * ndoms_new == 0 is a special case for destroying existing domains,
6641 * and it will not create the default domain.
6643 * Call with hotplug lock held
6645 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6646 struct sched_domain_attr
*dattr_new
)
6651 mutex_lock(&sched_domains_mutex
);
6653 /* always unregister in case we don't destroy any domains */
6654 unregister_sched_domain_sysctl();
6656 /* Let architecture update cpu core mappings. */
6657 new_topology
= arch_update_cpu_topology();
6659 n
= doms_new
? ndoms_new
: 0;
6661 /* Destroy deleted domains */
6662 for (i
= 0; i
< ndoms_cur
; i
++) {
6663 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6664 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6665 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6668 /* no match - a current sched domain not in new doms_new[] */
6669 detach_destroy_domains(doms_cur
[i
]);
6675 if (doms_new
== NULL
) {
6677 doms_new
= &fallback_doms
;
6678 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6679 WARN_ON_ONCE(dattr_new
);
6682 /* Build new domains */
6683 for (i
= 0; i
< ndoms_new
; i
++) {
6684 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6685 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6686 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6689 /* no match - add a new doms_new */
6690 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6695 /* Remember the new sched domains */
6696 if (doms_cur
!= &fallback_doms
)
6697 free_sched_domains(doms_cur
, ndoms_cur
);
6698 kfree(dattr_cur
); /* kfree(NULL) is safe */
6699 doms_cur
= doms_new
;
6700 dattr_cur
= dattr_new
;
6701 ndoms_cur
= ndoms_new
;
6703 register_sched_domain_sysctl();
6705 mutex_unlock(&sched_domains_mutex
);
6708 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6711 * Update cpusets according to cpu_active mask. If cpusets are
6712 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6713 * around partition_sched_domains().
6715 * If we come here as part of a suspend/resume, don't touch cpusets because we
6716 * want to restore it back to its original state upon resume anyway.
6718 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6722 case CPU_ONLINE_FROZEN
:
6723 case CPU_DOWN_FAILED_FROZEN
:
6726 * num_cpus_frozen tracks how many CPUs are involved in suspend
6727 * resume sequence. As long as this is not the last online
6728 * operation in the resume sequence, just build a single sched
6729 * domain, ignoring cpusets.
6732 if (likely(num_cpus_frozen
)) {
6733 partition_sched_domains(1, NULL
, NULL
);
6738 * This is the last CPU online operation. So fall through and
6739 * restore the original sched domains by considering the
6740 * cpuset configurations.
6744 case CPU_DOWN_FAILED
:
6745 cpuset_update_active_cpus(true);
6753 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6757 case CPU_DOWN_PREPARE
:
6758 cpuset_update_active_cpus(false);
6760 case CPU_DOWN_PREPARE_FROZEN
:
6762 partition_sched_domains(1, NULL
, NULL
);
6770 void __init
sched_init_smp(void)
6772 cpumask_var_t non_isolated_cpus
;
6774 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6775 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6780 * There's no userspace yet to cause hotplug operations; hence all the
6781 * cpu masks are stable and all blatant races in the below code cannot
6784 mutex_lock(&sched_domains_mutex
);
6785 init_sched_domains(cpu_active_mask
);
6786 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6787 if (cpumask_empty(non_isolated_cpus
))
6788 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6789 mutex_unlock(&sched_domains_mutex
);
6791 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6792 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6793 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6797 /* Move init over to a non-isolated CPU */
6798 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6800 sched_init_granularity();
6801 free_cpumask_var(non_isolated_cpus
);
6803 init_sched_rt_class();
6804 init_sched_dl_class();
6807 void __init
sched_init_smp(void)
6809 sched_init_granularity();
6811 #endif /* CONFIG_SMP */
6813 const_debug
unsigned int sysctl_timer_migration
= 1;
6815 int in_sched_functions(unsigned long addr
)
6817 return in_lock_functions(addr
) ||
6818 (addr
>= (unsigned long)__sched_text_start
6819 && addr
< (unsigned long)__sched_text_end
);
6822 #ifdef CONFIG_CGROUP_SCHED
6824 * Default task group.
6825 * Every task in system belongs to this group at bootup.
6827 struct task_group root_task_group
;
6828 LIST_HEAD(task_groups
);
6831 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6833 void __init
sched_init(void)
6836 unsigned long alloc_size
= 0, ptr
;
6838 #ifdef CONFIG_FAIR_GROUP_SCHED
6839 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6841 #ifdef CONFIG_RT_GROUP_SCHED
6842 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6844 #ifdef CONFIG_CPUMASK_OFFSTACK
6845 alloc_size
+= num_possible_cpus() * cpumask_size();
6848 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6850 #ifdef CONFIG_FAIR_GROUP_SCHED
6851 root_task_group
.se
= (struct sched_entity
**)ptr
;
6852 ptr
+= nr_cpu_ids
* sizeof(void **);
6854 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6855 ptr
+= nr_cpu_ids
* sizeof(void **);
6857 #endif /* CONFIG_FAIR_GROUP_SCHED */
6858 #ifdef CONFIG_RT_GROUP_SCHED
6859 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6860 ptr
+= nr_cpu_ids
* sizeof(void **);
6862 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6863 ptr
+= nr_cpu_ids
* sizeof(void **);
6865 #endif /* CONFIG_RT_GROUP_SCHED */
6866 #ifdef CONFIG_CPUMASK_OFFSTACK
6867 for_each_possible_cpu(i
) {
6868 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6869 ptr
+= cpumask_size();
6871 #endif /* CONFIG_CPUMASK_OFFSTACK */
6874 init_rt_bandwidth(&def_rt_bandwidth
,
6875 global_rt_period(), global_rt_runtime());
6876 init_dl_bandwidth(&def_dl_bandwidth
,
6877 global_rt_period(), global_rt_runtime());
6880 init_defrootdomain();
6883 #ifdef CONFIG_RT_GROUP_SCHED
6884 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6885 global_rt_period(), global_rt_runtime());
6886 #endif /* CONFIG_RT_GROUP_SCHED */
6888 #ifdef CONFIG_CGROUP_SCHED
6889 list_add(&root_task_group
.list
, &task_groups
);
6890 INIT_LIST_HEAD(&root_task_group
.children
);
6891 INIT_LIST_HEAD(&root_task_group
.siblings
);
6892 autogroup_init(&init_task
);
6894 #endif /* CONFIG_CGROUP_SCHED */
6896 for_each_possible_cpu(i
) {
6900 raw_spin_lock_init(&rq
->lock
);
6902 rq
->calc_load_active
= 0;
6903 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6904 init_cfs_rq(&rq
->cfs
);
6905 init_rt_rq(&rq
->rt
, rq
);
6906 init_dl_rq(&rq
->dl
, rq
);
6907 #ifdef CONFIG_FAIR_GROUP_SCHED
6908 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6909 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6911 * How much cpu bandwidth does root_task_group get?
6913 * In case of task-groups formed thr' the cgroup filesystem, it
6914 * gets 100% of the cpu resources in the system. This overall
6915 * system cpu resource is divided among the tasks of
6916 * root_task_group and its child task-groups in a fair manner,
6917 * based on each entity's (task or task-group's) weight
6918 * (se->load.weight).
6920 * In other words, if root_task_group has 10 tasks of weight
6921 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6922 * then A0's share of the cpu resource is:
6924 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6926 * We achieve this by letting root_task_group's tasks sit
6927 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6929 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6930 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6931 #endif /* CONFIG_FAIR_GROUP_SCHED */
6933 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6934 #ifdef CONFIG_RT_GROUP_SCHED
6935 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6938 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6939 rq
->cpu_load
[j
] = 0;
6941 rq
->last_load_update_tick
= jiffies
;
6946 rq
->cpu_power
= SCHED_POWER_SCALE
;
6947 rq
->post_schedule
= 0;
6948 rq
->active_balance
= 0;
6949 rq
->next_balance
= jiffies
;
6954 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6955 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6957 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6959 rq_attach_root(rq
, &def_root_domain
);
6960 #ifdef CONFIG_NO_HZ_COMMON
6963 #ifdef CONFIG_NO_HZ_FULL
6964 rq
->last_sched_tick
= 0;
6968 atomic_set(&rq
->nr_iowait
, 0);
6971 set_load_weight(&init_task
);
6973 #ifdef CONFIG_PREEMPT_NOTIFIERS
6974 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6978 * The boot idle thread does lazy MMU switching as well:
6980 atomic_inc(&init_mm
.mm_count
);
6981 enter_lazy_tlb(&init_mm
, current
);
6984 * Make us the idle thread. Technically, schedule() should not be
6985 * called from this thread, however somewhere below it might be,
6986 * but because we are the idle thread, we just pick up running again
6987 * when this runqueue becomes "idle".
6989 init_idle(current
, smp_processor_id());
6991 calc_load_update
= jiffies
+ LOAD_FREQ
;
6994 * During early bootup we pretend to be a normal task:
6996 current
->sched_class
= &fair_sched_class
;
6999 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7000 /* May be allocated at isolcpus cmdline parse time */
7001 if (cpu_isolated_map
== NULL
)
7002 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7003 idle_thread_set_boot_cpu();
7004 set_cpu_rq_start_time();
7006 init_sched_fair_class();
7008 scheduler_running
= 1;
7011 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7012 static inline int preempt_count_equals(int preempt_offset
)
7014 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7016 return (nested
== preempt_offset
);
7019 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7021 static unsigned long prev_jiffy
; /* ratelimiting */
7023 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7024 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7025 !is_idle_task(current
)) ||
7026 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7028 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7030 prev_jiffy
= jiffies
;
7033 "BUG: sleeping function called from invalid context at %s:%d\n",
7036 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7037 in_atomic(), irqs_disabled(),
7038 current
->pid
, current
->comm
);
7040 debug_show_held_locks(current
);
7041 if (irqs_disabled())
7042 print_irqtrace_events(current
);
7043 #ifdef CONFIG_DEBUG_PREEMPT
7044 if (!preempt_count_equals(preempt_offset
)) {
7045 pr_err("Preemption disabled at:");
7046 print_ip_sym(current
->preempt_disable_ip
);
7052 EXPORT_SYMBOL(__might_sleep
);
7055 #ifdef CONFIG_MAGIC_SYSRQ
7056 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7058 const struct sched_class
*prev_class
= p
->sched_class
;
7059 struct sched_attr attr
= {
7060 .sched_policy
= SCHED_NORMAL
,
7062 int old_prio
= p
->prio
;
7067 dequeue_task(rq
, p
, 0);
7068 __setscheduler(rq
, p
, &attr
);
7070 enqueue_task(rq
, p
, 0);
7071 resched_task(rq
->curr
);
7074 check_class_changed(rq
, p
, prev_class
, old_prio
);
7077 void normalize_rt_tasks(void)
7079 struct task_struct
*g
, *p
;
7080 unsigned long flags
;
7083 read_lock_irqsave(&tasklist_lock
, flags
);
7084 do_each_thread(g
, p
) {
7086 * Only normalize user tasks:
7091 p
->se
.exec_start
= 0;
7092 #ifdef CONFIG_SCHEDSTATS
7093 p
->se
.statistics
.wait_start
= 0;
7094 p
->se
.statistics
.sleep_start
= 0;
7095 p
->se
.statistics
.block_start
= 0;
7098 if (!dl_task(p
) && !rt_task(p
)) {
7100 * Renice negative nice level userspace
7103 if (task_nice(p
) < 0 && p
->mm
)
7104 set_user_nice(p
, 0);
7108 raw_spin_lock(&p
->pi_lock
);
7109 rq
= __task_rq_lock(p
);
7111 normalize_task(rq
, p
);
7113 __task_rq_unlock(rq
);
7114 raw_spin_unlock(&p
->pi_lock
);
7115 } while_each_thread(g
, p
);
7117 read_unlock_irqrestore(&tasklist_lock
, flags
);
7120 #endif /* CONFIG_MAGIC_SYSRQ */
7122 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7124 * These functions are only useful for the IA64 MCA handling, or kdb.
7126 * They can only be called when the whole system has been
7127 * stopped - every CPU needs to be quiescent, and no scheduling
7128 * activity can take place. Using them for anything else would
7129 * be a serious bug, and as a result, they aren't even visible
7130 * under any other configuration.
7134 * curr_task - return the current task for a given cpu.
7135 * @cpu: the processor in question.
7137 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7139 * Return: The current task for @cpu.
7141 struct task_struct
*curr_task(int cpu
)
7143 return cpu_curr(cpu
);
7146 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7150 * set_curr_task - set the current task for a given cpu.
7151 * @cpu: the processor in question.
7152 * @p: the task pointer to set.
7154 * Description: This function must only be used when non-maskable interrupts
7155 * are serviced on a separate stack. It allows the architecture to switch the
7156 * notion of the current task on a cpu in a non-blocking manner. This function
7157 * must be called with all CPU's synchronized, and interrupts disabled, the
7158 * and caller must save the original value of the current task (see
7159 * curr_task() above) and restore that value before reenabling interrupts and
7160 * re-starting the system.
7162 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7164 void set_curr_task(int cpu
, struct task_struct
*p
)
7171 #ifdef CONFIG_CGROUP_SCHED
7172 /* task_group_lock serializes the addition/removal of task groups */
7173 static DEFINE_SPINLOCK(task_group_lock
);
7175 static void free_sched_group(struct task_group
*tg
)
7177 free_fair_sched_group(tg
);
7178 free_rt_sched_group(tg
);
7183 /* allocate runqueue etc for a new task group */
7184 struct task_group
*sched_create_group(struct task_group
*parent
)
7186 struct task_group
*tg
;
7188 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7190 return ERR_PTR(-ENOMEM
);
7192 if (!alloc_fair_sched_group(tg
, parent
))
7195 if (!alloc_rt_sched_group(tg
, parent
))
7201 free_sched_group(tg
);
7202 return ERR_PTR(-ENOMEM
);
7205 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7207 unsigned long flags
;
7209 spin_lock_irqsave(&task_group_lock
, flags
);
7210 list_add_rcu(&tg
->list
, &task_groups
);
7212 WARN_ON(!parent
); /* root should already exist */
7214 tg
->parent
= parent
;
7215 INIT_LIST_HEAD(&tg
->children
);
7216 list_add_rcu(&tg
->siblings
, &parent
->children
);
7217 spin_unlock_irqrestore(&task_group_lock
, flags
);
7220 /* rcu callback to free various structures associated with a task group */
7221 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7223 /* now it should be safe to free those cfs_rqs */
7224 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7227 /* Destroy runqueue etc associated with a task group */
7228 void sched_destroy_group(struct task_group
*tg
)
7230 /* wait for possible concurrent references to cfs_rqs complete */
7231 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7234 void sched_offline_group(struct task_group
*tg
)
7236 unsigned long flags
;
7239 /* end participation in shares distribution */
7240 for_each_possible_cpu(i
)
7241 unregister_fair_sched_group(tg
, i
);
7243 spin_lock_irqsave(&task_group_lock
, flags
);
7244 list_del_rcu(&tg
->list
);
7245 list_del_rcu(&tg
->siblings
);
7246 spin_unlock_irqrestore(&task_group_lock
, flags
);
7249 /* change task's runqueue when it moves between groups.
7250 * The caller of this function should have put the task in its new group
7251 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7252 * reflect its new group.
7254 void sched_move_task(struct task_struct
*tsk
)
7256 struct task_group
*tg
;
7258 unsigned long flags
;
7261 rq
= task_rq_lock(tsk
, &flags
);
7263 running
= task_current(rq
, tsk
);
7267 dequeue_task(rq
, tsk
, 0);
7268 if (unlikely(running
))
7269 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7271 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
,
7272 lockdep_is_held(&tsk
->sighand
->siglock
)),
7273 struct task_group
, css
);
7274 tg
= autogroup_task_group(tsk
, tg
);
7275 tsk
->sched_task_group
= tg
;
7277 #ifdef CONFIG_FAIR_GROUP_SCHED
7278 if (tsk
->sched_class
->task_move_group
)
7279 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7282 set_task_rq(tsk
, task_cpu(tsk
));
7284 if (unlikely(running
))
7285 tsk
->sched_class
->set_curr_task(rq
);
7287 enqueue_task(rq
, tsk
, 0);
7289 task_rq_unlock(rq
, tsk
, &flags
);
7291 #endif /* CONFIG_CGROUP_SCHED */
7293 #ifdef CONFIG_RT_GROUP_SCHED
7295 * Ensure that the real time constraints are schedulable.
7297 static DEFINE_MUTEX(rt_constraints_mutex
);
7299 /* Must be called with tasklist_lock held */
7300 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7302 struct task_struct
*g
, *p
;
7304 do_each_thread(g
, p
) {
7305 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7307 } while_each_thread(g
, p
);
7312 struct rt_schedulable_data
{
7313 struct task_group
*tg
;
7318 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7320 struct rt_schedulable_data
*d
= data
;
7321 struct task_group
*child
;
7322 unsigned long total
, sum
= 0;
7323 u64 period
, runtime
;
7325 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7326 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7329 period
= d
->rt_period
;
7330 runtime
= d
->rt_runtime
;
7334 * Cannot have more runtime than the period.
7336 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7340 * Ensure we don't starve existing RT tasks.
7342 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7345 total
= to_ratio(period
, runtime
);
7348 * Nobody can have more than the global setting allows.
7350 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7354 * The sum of our children's runtime should not exceed our own.
7356 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7357 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7358 runtime
= child
->rt_bandwidth
.rt_runtime
;
7360 if (child
== d
->tg
) {
7361 period
= d
->rt_period
;
7362 runtime
= d
->rt_runtime
;
7365 sum
+= to_ratio(period
, runtime
);
7374 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7378 struct rt_schedulable_data data
= {
7380 .rt_period
= period
,
7381 .rt_runtime
= runtime
,
7385 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7391 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7392 u64 rt_period
, u64 rt_runtime
)
7396 mutex_lock(&rt_constraints_mutex
);
7397 read_lock(&tasklist_lock
);
7398 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7402 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7403 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7404 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7406 for_each_possible_cpu(i
) {
7407 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7409 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7410 rt_rq
->rt_runtime
= rt_runtime
;
7411 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7413 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7415 read_unlock(&tasklist_lock
);
7416 mutex_unlock(&rt_constraints_mutex
);
7421 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7423 u64 rt_runtime
, rt_period
;
7425 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7426 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7427 if (rt_runtime_us
< 0)
7428 rt_runtime
= RUNTIME_INF
;
7430 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7433 static long sched_group_rt_runtime(struct task_group
*tg
)
7437 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7440 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7441 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7442 return rt_runtime_us
;
7445 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7447 u64 rt_runtime
, rt_period
;
7449 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7450 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7455 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7458 static long sched_group_rt_period(struct task_group
*tg
)
7462 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7463 do_div(rt_period_us
, NSEC_PER_USEC
);
7464 return rt_period_us
;
7466 #endif /* CONFIG_RT_GROUP_SCHED */
7468 #ifdef CONFIG_RT_GROUP_SCHED
7469 static int sched_rt_global_constraints(void)
7473 mutex_lock(&rt_constraints_mutex
);
7474 read_lock(&tasklist_lock
);
7475 ret
= __rt_schedulable(NULL
, 0, 0);
7476 read_unlock(&tasklist_lock
);
7477 mutex_unlock(&rt_constraints_mutex
);
7482 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7484 /* Don't accept realtime tasks when there is no way for them to run */
7485 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7491 #else /* !CONFIG_RT_GROUP_SCHED */
7492 static int sched_rt_global_constraints(void)
7494 unsigned long flags
;
7497 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7498 for_each_possible_cpu(i
) {
7499 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7501 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7502 rt_rq
->rt_runtime
= global_rt_runtime();
7503 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7505 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7509 #endif /* CONFIG_RT_GROUP_SCHED */
7511 static int sched_dl_global_constraints(void)
7513 u64 runtime
= global_rt_runtime();
7514 u64 period
= global_rt_period();
7515 u64 new_bw
= to_ratio(period
, runtime
);
7517 unsigned long flags
;
7520 * Here we want to check the bandwidth not being set to some
7521 * value smaller than the currently allocated bandwidth in
7522 * any of the root_domains.
7524 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7525 * cycling on root_domains... Discussion on different/better
7526 * solutions is welcome!
7528 for_each_possible_cpu(cpu
) {
7529 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7531 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7532 if (new_bw
< dl_b
->total_bw
)
7534 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7543 static void sched_dl_do_global(void)
7547 unsigned long flags
;
7549 def_dl_bandwidth
.dl_period
= global_rt_period();
7550 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7552 if (global_rt_runtime() != RUNTIME_INF
)
7553 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7556 * FIXME: As above...
7558 for_each_possible_cpu(cpu
) {
7559 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7561 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7563 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7567 static int sched_rt_global_validate(void)
7569 if (sysctl_sched_rt_period
<= 0)
7572 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7573 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7579 static void sched_rt_do_global(void)
7581 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7582 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7585 int sched_rt_handler(struct ctl_table
*table
, int write
,
7586 void __user
*buffer
, size_t *lenp
,
7589 int old_period
, old_runtime
;
7590 static DEFINE_MUTEX(mutex
);
7594 old_period
= sysctl_sched_rt_period
;
7595 old_runtime
= sysctl_sched_rt_runtime
;
7597 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7599 if (!ret
&& write
) {
7600 ret
= sched_rt_global_validate();
7604 ret
= sched_rt_global_constraints();
7608 ret
= sched_dl_global_constraints();
7612 sched_rt_do_global();
7613 sched_dl_do_global();
7617 sysctl_sched_rt_period
= old_period
;
7618 sysctl_sched_rt_runtime
= old_runtime
;
7620 mutex_unlock(&mutex
);
7625 int sched_rr_handler(struct ctl_table
*table
, int write
,
7626 void __user
*buffer
, size_t *lenp
,
7630 static DEFINE_MUTEX(mutex
);
7633 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7634 /* make sure that internally we keep jiffies */
7635 /* also, writing zero resets timeslice to default */
7636 if (!ret
&& write
) {
7637 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7638 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7640 mutex_unlock(&mutex
);
7644 #ifdef CONFIG_CGROUP_SCHED
7646 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7648 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7651 static struct cgroup_subsys_state
*
7652 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7654 struct task_group
*parent
= css_tg(parent_css
);
7655 struct task_group
*tg
;
7658 /* This is early initialization for the top cgroup */
7659 return &root_task_group
.css
;
7662 tg
= sched_create_group(parent
);
7664 return ERR_PTR(-ENOMEM
);
7669 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7671 struct task_group
*tg
= css_tg(css
);
7672 struct task_group
*parent
= css_tg(css_parent(css
));
7675 sched_online_group(tg
, parent
);
7679 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7681 struct task_group
*tg
= css_tg(css
);
7683 sched_destroy_group(tg
);
7686 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7688 struct task_group
*tg
= css_tg(css
);
7690 sched_offline_group(tg
);
7693 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7694 struct cgroup_taskset
*tset
)
7696 struct task_struct
*task
;
7698 cgroup_taskset_for_each(task
, tset
) {
7699 #ifdef CONFIG_RT_GROUP_SCHED
7700 if (!sched_rt_can_attach(css_tg(css
), task
))
7703 /* We don't support RT-tasks being in separate groups */
7704 if (task
->sched_class
!= &fair_sched_class
)
7711 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7712 struct cgroup_taskset
*tset
)
7714 struct task_struct
*task
;
7716 cgroup_taskset_for_each(task
, tset
)
7717 sched_move_task(task
);
7720 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7721 struct cgroup_subsys_state
*old_css
,
7722 struct task_struct
*task
)
7725 * cgroup_exit() is called in the copy_process() failure path.
7726 * Ignore this case since the task hasn't ran yet, this avoids
7727 * trying to poke a half freed task state from generic code.
7729 if (!(task
->flags
& PF_EXITING
))
7732 sched_move_task(task
);
7735 #ifdef CONFIG_FAIR_GROUP_SCHED
7736 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7737 struct cftype
*cftype
, u64 shareval
)
7739 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7742 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7745 struct task_group
*tg
= css_tg(css
);
7747 return (u64
) scale_load_down(tg
->shares
);
7750 #ifdef CONFIG_CFS_BANDWIDTH
7751 static DEFINE_MUTEX(cfs_constraints_mutex
);
7753 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7754 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7756 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7758 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7760 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7761 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7763 if (tg
== &root_task_group
)
7767 * Ensure we have at some amount of bandwidth every period. This is
7768 * to prevent reaching a state of large arrears when throttled via
7769 * entity_tick() resulting in prolonged exit starvation.
7771 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7775 * Likewise, bound things on the otherside by preventing insane quota
7776 * periods. This also allows us to normalize in computing quota
7779 if (period
> max_cfs_quota_period
)
7782 mutex_lock(&cfs_constraints_mutex
);
7783 ret
= __cfs_schedulable(tg
, period
, quota
);
7787 runtime_enabled
= quota
!= RUNTIME_INF
;
7788 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7790 * If we need to toggle cfs_bandwidth_used, off->on must occur
7791 * before making related changes, and on->off must occur afterwards
7793 if (runtime_enabled
&& !runtime_was_enabled
)
7794 cfs_bandwidth_usage_inc();
7795 raw_spin_lock_irq(&cfs_b
->lock
);
7796 cfs_b
->period
= ns_to_ktime(period
);
7797 cfs_b
->quota
= quota
;
7799 __refill_cfs_bandwidth_runtime(cfs_b
);
7800 /* restart the period timer (if active) to handle new period expiry */
7801 if (runtime_enabled
&& cfs_b
->timer_active
) {
7802 /* force a reprogram */
7803 cfs_b
->timer_active
= 0;
7804 __start_cfs_bandwidth(cfs_b
);
7806 raw_spin_unlock_irq(&cfs_b
->lock
);
7808 for_each_possible_cpu(i
) {
7809 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7810 struct rq
*rq
= cfs_rq
->rq
;
7812 raw_spin_lock_irq(&rq
->lock
);
7813 cfs_rq
->runtime_enabled
= runtime_enabled
;
7814 cfs_rq
->runtime_remaining
= 0;
7816 if (cfs_rq
->throttled
)
7817 unthrottle_cfs_rq(cfs_rq
);
7818 raw_spin_unlock_irq(&rq
->lock
);
7820 if (runtime_was_enabled
&& !runtime_enabled
)
7821 cfs_bandwidth_usage_dec();
7823 mutex_unlock(&cfs_constraints_mutex
);
7828 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7832 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7833 if (cfs_quota_us
< 0)
7834 quota
= RUNTIME_INF
;
7836 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7838 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7841 long tg_get_cfs_quota(struct task_group
*tg
)
7845 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7848 quota_us
= tg
->cfs_bandwidth
.quota
;
7849 do_div(quota_us
, NSEC_PER_USEC
);
7854 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7858 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7859 quota
= tg
->cfs_bandwidth
.quota
;
7861 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7864 long tg_get_cfs_period(struct task_group
*tg
)
7868 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7869 do_div(cfs_period_us
, NSEC_PER_USEC
);
7871 return cfs_period_us
;
7874 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7877 return tg_get_cfs_quota(css_tg(css
));
7880 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7881 struct cftype
*cftype
, s64 cfs_quota_us
)
7883 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7886 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7889 return tg_get_cfs_period(css_tg(css
));
7892 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7893 struct cftype
*cftype
, u64 cfs_period_us
)
7895 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7898 struct cfs_schedulable_data
{
7899 struct task_group
*tg
;
7904 * normalize group quota/period to be quota/max_period
7905 * note: units are usecs
7907 static u64
normalize_cfs_quota(struct task_group
*tg
,
7908 struct cfs_schedulable_data
*d
)
7916 period
= tg_get_cfs_period(tg
);
7917 quota
= tg_get_cfs_quota(tg
);
7920 /* note: these should typically be equivalent */
7921 if (quota
== RUNTIME_INF
|| quota
== -1)
7924 return to_ratio(period
, quota
);
7927 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7929 struct cfs_schedulable_data
*d
= data
;
7930 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7931 s64 quota
= 0, parent_quota
= -1;
7934 quota
= RUNTIME_INF
;
7936 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7938 quota
= normalize_cfs_quota(tg
, d
);
7939 parent_quota
= parent_b
->hierarchal_quota
;
7942 * ensure max(child_quota) <= parent_quota, inherit when no
7945 if (quota
== RUNTIME_INF
)
7946 quota
= parent_quota
;
7947 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7950 cfs_b
->hierarchal_quota
= quota
;
7955 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7958 struct cfs_schedulable_data data
= {
7964 if (quota
!= RUNTIME_INF
) {
7965 do_div(data
.period
, NSEC_PER_USEC
);
7966 do_div(data
.quota
, NSEC_PER_USEC
);
7970 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7976 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7978 struct task_group
*tg
= css_tg(seq_css(sf
));
7979 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7981 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7982 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7983 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7987 #endif /* CONFIG_CFS_BANDWIDTH */
7988 #endif /* CONFIG_FAIR_GROUP_SCHED */
7990 #ifdef CONFIG_RT_GROUP_SCHED
7991 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7992 struct cftype
*cft
, s64 val
)
7994 return sched_group_set_rt_runtime(css_tg(css
), val
);
7997 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8000 return sched_group_rt_runtime(css_tg(css
));
8003 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8004 struct cftype
*cftype
, u64 rt_period_us
)
8006 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8009 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8012 return sched_group_rt_period(css_tg(css
));
8014 #endif /* CONFIG_RT_GROUP_SCHED */
8016 static struct cftype cpu_files
[] = {
8017 #ifdef CONFIG_FAIR_GROUP_SCHED
8020 .read_u64
= cpu_shares_read_u64
,
8021 .write_u64
= cpu_shares_write_u64
,
8024 #ifdef CONFIG_CFS_BANDWIDTH
8026 .name
= "cfs_quota_us",
8027 .read_s64
= cpu_cfs_quota_read_s64
,
8028 .write_s64
= cpu_cfs_quota_write_s64
,
8031 .name
= "cfs_period_us",
8032 .read_u64
= cpu_cfs_period_read_u64
,
8033 .write_u64
= cpu_cfs_period_write_u64
,
8037 .seq_show
= cpu_stats_show
,
8040 #ifdef CONFIG_RT_GROUP_SCHED
8042 .name
= "rt_runtime_us",
8043 .read_s64
= cpu_rt_runtime_read
,
8044 .write_s64
= cpu_rt_runtime_write
,
8047 .name
= "rt_period_us",
8048 .read_u64
= cpu_rt_period_read_uint
,
8049 .write_u64
= cpu_rt_period_write_uint
,
8055 struct cgroup_subsys cpu_cgrp_subsys
= {
8056 .css_alloc
= cpu_cgroup_css_alloc
,
8057 .css_free
= cpu_cgroup_css_free
,
8058 .css_online
= cpu_cgroup_css_online
,
8059 .css_offline
= cpu_cgroup_css_offline
,
8060 .can_attach
= cpu_cgroup_can_attach
,
8061 .attach
= cpu_cgroup_attach
,
8062 .exit
= cpu_cgroup_exit
,
8063 .base_cftypes
= cpu_files
,
8067 #endif /* CONFIG_CGROUP_SCHED */
8069 void dump_cpu_task(int cpu
)
8071 pr_info("Task dump for CPU %d:\n", cpu
);
8072 sched_show_task(cpu_curr(cpu
));