4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
96 ktime_t soft
, hard
, now
;
99 if (hrtimer_active(period_timer
))
102 now
= hrtimer_cb_get_time(period_timer
);
103 hrtimer_forward(period_timer
, now
, period
);
105 soft
= hrtimer_get_softexpires(period_timer
);
106 hard
= hrtimer_get_expires(period_timer
);
107 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
108 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
109 HRTIMER_MODE_ABS_PINNED
, 0);
113 DEFINE_MUTEX(sched_domains_mutex
);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
116 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
118 void update_rq_clock(struct rq
*rq
)
122 if (rq
->skip_clock_update
> 0)
125 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
129 update_rq_clock_task(rq
, delta
);
133 * Debugging: various feature bits
136 #define SCHED_FEAT(name, enabled) \
137 (1UL << __SCHED_FEAT_##name) * enabled |
139 const_debug
unsigned int sysctl_sched_features
=
140 #include "features.h"
145 #ifdef CONFIG_SCHED_DEBUG
146 #define SCHED_FEAT(name, enabled) \
149 static const char * const sched_feat_names
[] = {
150 #include "features.h"
155 static int sched_feat_show(struct seq_file
*m
, void *v
)
159 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
160 if (!(sysctl_sched_features
& (1UL << i
)))
162 seq_printf(m
, "%s ", sched_feat_names
[i
]);
169 #ifdef HAVE_JUMP_LABEL
171 #define jump_label_key__true STATIC_KEY_INIT_TRUE
172 #define jump_label_key__false STATIC_KEY_INIT_FALSE
174 #define SCHED_FEAT(name, enabled) \
175 jump_label_key__##enabled ,
177 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
178 #include "features.h"
183 static void sched_feat_disable(int i
)
185 if (static_key_enabled(&sched_feat_keys
[i
]))
186 static_key_slow_dec(&sched_feat_keys
[i
]);
189 static void sched_feat_enable(int i
)
191 if (!static_key_enabled(&sched_feat_keys
[i
]))
192 static_key_slow_inc(&sched_feat_keys
[i
]);
195 static void sched_feat_disable(int i
) { };
196 static void sched_feat_enable(int i
) { };
197 #endif /* HAVE_JUMP_LABEL */
199 static int sched_feat_set(char *cmp
)
204 if (strncmp(cmp
, "NO_", 3) == 0) {
209 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
210 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
212 sysctl_sched_features
&= ~(1UL << i
);
213 sched_feat_disable(i
);
215 sysctl_sched_features
|= (1UL << i
);
216 sched_feat_enable(i
);
226 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
227 size_t cnt
, loff_t
*ppos
)
237 if (copy_from_user(&buf
, ubuf
, cnt
))
243 /* Ensure the static_key remains in a consistent state */
244 inode
= file_inode(filp
);
245 mutex_lock(&inode
->i_mutex
);
246 i
= sched_feat_set(cmp
);
247 mutex_unlock(&inode
->i_mutex
);
248 if (i
== __SCHED_FEAT_NR
)
256 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
258 return single_open(filp
, sched_feat_show
, NULL
);
261 static const struct file_operations sched_feat_fops
= {
262 .open
= sched_feat_open
,
263 .write
= sched_feat_write
,
266 .release
= single_release
,
269 static __init
int sched_init_debug(void)
271 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
276 late_initcall(sched_init_debug
);
277 #endif /* CONFIG_SCHED_DEBUG */
280 * Number of tasks to iterate in a single balance run.
281 * Limited because this is done with IRQs disabled.
283 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
286 * period over which we average the RT time consumption, measured
291 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
294 * period over which we measure -rt task cpu usage in us.
297 unsigned int sysctl_sched_rt_period
= 1000000;
299 __read_mostly
int scheduler_running
;
302 * part of the period that we allow rt tasks to run in us.
305 int sysctl_sched_rt_runtime
= 950000;
308 * __task_rq_lock - lock the rq @p resides on.
310 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
315 lockdep_assert_held(&p
->pi_lock
);
319 raw_spin_lock(&rq
->lock
);
320 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
)))
322 raw_spin_unlock(&rq
->lock
);
324 while (unlikely(task_on_rq_migrating(p
)))
330 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
332 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
333 __acquires(p
->pi_lock
)
339 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
341 raw_spin_lock(&rq
->lock
);
342 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
)))
344 raw_spin_unlock(&rq
->lock
);
345 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
347 while (unlikely(task_on_rq_migrating(p
)))
352 static void __task_rq_unlock(struct rq
*rq
)
355 raw_spin_unlock(&rq
->lock
);
359 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
361 __releases(p
->pi_lock
)
363 raw_spin_unlock(&rq
->lock
);
364 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
368 * this_rq_lock - lock this runqueue and disable interrupts.
370 static struct rq
*this_rq_lock(void)
377 raw_spin_lock(&rq
->lock
);
382 #ifdef CONFIG_SCHED_HRTICK
384 * Use HR-timers to deliver accurate preemption points.
387 static void hrtick_clear(struct rq
*rq
)
389 if (hrtimer_active(&rq
->hrtick_timer
))
390 hrtimer_cancel(&rq
->hrtick_timer
);
394 * High-resolution timer tick.
395 * Runs from hardirq context with interrupts disabled.
397 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
399 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
401 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
403 raw_spin_lock(&rq
->lock
);
405 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
406 raw_spin_unlock(&rq
->lock
);
408 return HRTIMER_NORESTART
;
413 static int __hrtick_restart(struct rq
*rq
)
415 struct hrtimer
*timer
= &rq
->hrtick_timer
;
416 ktime_t time
= hrtimer_get_softexpires(timer
);
418 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
422 * called from hardirq (IPI) context
424 static void __hrtick_start(void *arg
)
428 raw_spin_lock(&rq
->lock
);
429 __hrtick_restart(rq
);
430 rq
->hrtick_csd_pending
= 0;
431 raw_spin_unlock(&rq
->lock
);
435 * Called to set the hrtick timer state.
437 * called with rq->lock held and irqs disabled
439 void hrtick_start(struct rq
*rq
, u64 delay
)
441 struct hrtimer
*timer
= &rq
->hrtick_timer
;
446 * Don't schedule slices shorter than 10000ns, that just
447 * doesn't make sense and can cause timer DoS.
449 delta
= max_t(s64
, delay
, 10000LL);
450 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
452 hrtimer_set_expires(timer
, time
);
454 if (rq
== this_rq()) {
455 __hrtick_restart(rq
);
456 } else if (!rq
->hrtick_csd_pending
) {
457 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
458 rq
->hrtick_csd_pending
= 1;
463 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
465 int cpu
= (int)(long)hcpu
;
468 case CPU_UP_CANCELED
:
469 case CPU_UP_CANCELED_FROZEN
:
470 case CPU_DOWN_PREPARE
:
471 case CPU_DOWN_PREPARE_FROZEN
:
473 case CPU_DEAD_FROZEN
:
474 hrtick_clear(cpu_rq(cpu
));
481 static __init
void init_hrtick(void)
483 hotcpu_notifier(hotplug_hrtick
, 0);
487 * Called to set the hrtick timer state.
489 * called with rq->lock held and irqs disabled
491 void hrtick_start(struct rq
*rq
, u64 delay
)
493 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
494 HRTIMER_MODE_REL_PINNED
, 0);
497 static inline void init_hrtick(void)
500 #endif /* CONFIG_SMP */
502 static void init_rq_hrtick(struct rq
*rq
)
505 rq
->hrtick_csd_pending
= 0;
507 rq
->hrtick_csd
.flags
= 0;
508 rq
->hrtick_csd
.func
= __hrtick_start
;
509 rq
->hrtick_csd
.info
= rq
;
512 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
513 rq
->hrtick_timer
.function
= hrtick
;
515 #else /* CONFIG_SCHED_HRTICK */
516 static inline void hrtick_clear(struct rq
*rq
)
520 static inline void init_rq_hrtick(struct rq
*rq
)
524 static inline void init_hrtick(void)
527 #endif /* CONFIG_SCHED_HRTICK */
530 * cmpxchg based fetch_or, macro so it works for different integer types
532 #define fetch_or(ptr, val) \
533 ({ typeof(*(ptr)) __old, __val = *(ptr); \
535 __old = cmpxchg((ptr), __val, __val | (val)); \
536 if (__old == __val) \
543 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
545 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
546 * this avoids any races wrt polling state changes and thereby avoids
549 static bool set_nr_and_not_polling(struct task_struct
*p
)
551 struct thread_info
*ti
= task_thread_info(p
);
552 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
556 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
558 * If this returns true, then the idle task promises to call
559 * sched_ttwu_pending() and reschedule soon.
561 static bool set_nr_if_polling(struct task_struct
*p
)
563 struct thread_info
*ti
= task_thread_info(p
);
564 typeof(ti
->flags
) old
, val
= ACCESS_ONCE(ti
->flags
);
567 if (!(val
& _TIF_POLLING_NRFLAG
))
569 if (val
& _TIF_NEED_RESCHED
)
571 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
580 static bool set_nr_and_not_polling(struct task_struct
*p
)
582 set_tsk_need_resched(p
);
587 static bool set_nr_if_polling(struct task_struct
*p
)
595 * resched_curr - mark rq's current task 'to be rescheduled now'.
597 * On UP this means the setting of the need_resched flag, on SMP it
598 * might also involve a cross-CPU call to trigger the scheduler on
601 void resched_curr(struct rq
*rq
)
603 struct task_struct
*curr
= rq
->curr
;
606 lockdep_assert_held(&rq
->lock
);
608 if (test_tsk_need_resched(curr
))
613 if (cpu
== smp_processor_id()) {
614 set_tsk_need_resched(curr
);
615 set_preempt_need_resched();
619 if (set_nr_and_not_polling(curr
))
620 smp_send_reschedule(cpu
);
622 trace_sched_wake_idle_without_ipi(cpu
);
625 void resched_cpu(int cpu
)
627 struct rq
*rq
= cpu_rq(cpu
);
630 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
633 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
637 #ifdef CONFIG_NO_HZ_COMMON
639 * In the semi idle case, use the nearest busy cpu for migrating timers
640 * from an idle cpu. This is good for power-savings.
642 * We don't do similar optimization for completely idle system, as
643 * selecting an idle cpu will add more delays to the timers than intended
644 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
646 int get_nohz_timer_target(int pinned
)
648 int cpu
= smp_processor_id();
650 struct sched_domain
*sd
;
652 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
656 for_each_domain(cpu
, sd
) {
657 for_each_cpu(i
, sched_domain_span(sd
)) {
669 * When add_timer_on() enqueues a timer into the timer wheel of an
670 * idle CPU then this timer might expire before the next timer event
671 * which is scheduled to wake up that CPU. In case of a completely
672 * idle system the next event might even be infinite time into the
673 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
674 * leaves the inner idle loop so the newly added timer is taken into
675 * account when the CPU goes back to idle and evaluates the timer
676 * wheel for the next timer event.
678 static void wake_up_idle_cpu(int cpu
)
680 struct rq
*rq
= cpu_rq(cpu
);
682 if (cpu
== smp_processor_id())
685 if (set_nr_and_not_polling(rq
->idle
))
686 smp_send_reschedule(cpu
);
688 trace_sched_wake_idle_without_ipi(cpu
);
691 static bool wake_up_full_nohz_cpu(int cpu
)
694 * We just need the target to call irq_exit() and re-evaluate
695 * the next tick. The nohz full kick at least implies that.
696 * If needed we can still optimize that later with an
699 if (tick_nohz_full_cpu(cpu
)) {
700 if (cpu
!= smp_processor_id() ||
701 tick_nohz_tick_stopped())
702 tick_nohz_full_kick_cpu(cpu
);
709 void wake_up_nohz_cpu(int cpu
)
711 if (!wake_up_full_nohz_cpu(cpu
))
712 wake_up_idle_cpu(cpu
);
715 static inline bool got_nohz_idle_kick(void)
717 int cpu
= smp_processor_id();
719 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
722 if (idle_cpu(cpu
) && !need_resched())
726 * We can't run Idle Load Balance on this CPU for this time so we
727 * cancel it and clear NOHZ_BALANCE_KICK
729 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
733 #else /* CONFIG_NO_HZ_COMMON */
735 static inline bool got_nohz_idle_kick(void)
740 #endif /* CONFIG_NO_HZ_COMMON */
742 #ifdef CONFIG_NO_HZ_FULL
743 bool sched_can_stop_tick(void)
746 * More than one running task need preemption.
747 * nr_running update is assumed to be visible
748 * after IPI is sent from wakers.
750 if (this_rq()->nr_running
> 1)
755 #endif /* CONFIG_NO_HZ_FULL */
757 void sched_avg_update(struct rq
*rq
)
759 s64 period
= sched_avg_period();
761 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
763 * Inline assembly required to prevent the compiler
764 * optimising this loop into a divmod call.
765 * See __iter_div_u64_rem() for another example of this.
767 asm("" : "+rm" (rq
->age_stamp
));
768 rq
->age_stamp
+= period
;
773 #endif /* CONFIG_SMP */
775 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
776 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
778 * Iterate task_group tree rooted at *from, calling @down when first entering a
779 * node and @up when leaving it for the final time.
781 * Caller must hold rcu_lock or sufficient equivalent.
783 int walk_tg_tree_from(struct task_group
*from
,
784 tg_visitor down
, tg_visitor up
, void *data
)
786 struct task_group
*parent
, *child
;
792 ret
= (*down
)(parent
, data
);
795 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
802 ret
= (*up
)(parent
, data
);
803 if (ret
|| parent
== from
)
807 parent
= parent
->parent
;
814 int tg_nop(struct task_group
*tg
, void *data
)
820 static void set_load_weight(struct task_struct
*p
)
822 int prio
= p
->static_prio
- MAX_RT_PRIO
;
823 struct load_weight
*load
= &p
->se
.load
;
826 * SCHED_IDLE tasks get minimal weight:
828 if (p
->policy
== SCHED_IDLE
) {
829 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
830 load
->inv_weight
= WMULT_IDLEPRIO
;
834 load
->weight
= scale_load(prio_to_weight
[prio
]);
835 load
->inv_weight
= prio_to_wmult
[prio
];
838 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
841 sched_info_queued(rq
, p
);
842 p
->sched_class
->enqueue_task(rq
, p
, flags
);
845 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
848 sched_info_dequeued(rq
, p
);
849 p
->sched_class
->dequeue_task(rq
, p
, flags
);
852 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
854 if (task_contributes_to_load(p
))
855 rq
->nr_uninterruptible
--;
857 enqueue_task(rq
, p
, flags
);
860 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
862 if (task_contributes_to_load(p
))
863 rq
->nr_uninterruptible
++;
865 dequeue_task(rq
, p
, flags
);
868 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
871 * In theory, the compile should just see 0 here, and optimize out the call
872 * to sched_rt_avg_update. But I don't trust it...
874 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
875 s64 steal
= 0, irq_delta
= 0;
877 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
878 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
881 * Since irq_time is only updated on {soft,}irq_exit, we might run into
882 * this case when a previous update_rq_clock() happened inside a
885 * When this happens, we stop ->clock_task and only update the
886 * prev_irq_time stamp to account for the part that fit, so that a next
887 * update will consume the rest. This ensures ->clock_task is
890 * It does however cause some slight miss-attribution of {soft,}irq
891 * time, a more accurate solution would be to update the irq_time using
892 * the current rq->clock timestamp, except that would require using
895 if (irq_delta
> delta
)
898 rq
->prev_irq_time
+= irq_delta
;
901 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
902 if (static_key_false((¶virt_steal_rq_enabled
))) {
903 steal
= paravirt_steal_clock(cpu_of(rq
));
904 steal
-= rq
->prev_steal_time_rq
;
906 if (unlikely(steal
> delta
))
909 rq
->prev_steal_time_rq
+= steal
;
914 rq
->clock_task
+= delta
;
916 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
917 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
918 sched_rt_avg_update(rq
, irq_delta
+ steal
);
922 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
924 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
925 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
929 * Make it appear like a SCHED_FIFO task, its something
930 * userspace knows about and won't get confused about.
932 * Also, it will make PI more or less work without too
933 * much confusion -- but then, stop work should not
934 * rely on PI working anyway.
936 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
938 stop
->sched_class
= &stop_sched_class
;
941 cpu_rq(cpu
)->stop
= stop
;
945 * Reset it back to a normal scheduling class so that
946 * it can die in pieces.
948 old_stop
->sched_class
= &rt_sched_class
;
953 * __normal_prio - return the priority that is based on the static prio
955 static inline int __normal_prio(struct task_struct
*p
)
957 return p
->static_prio
;
961 * Calculate the expected normal priority: i.e. priority
962 * without taking RT-inheritance into account. Might be
963 * boosted by interactivity modifiers. Changes upon fork,
964 * setprio syscalls, and whenever the interactivity
965 * estimator recalculates.
967 static inline int normal_prio(struct task_struct
*p
)
971 if (task_has_dl_policy(p
))
972 prio
= MAX_DL_PRIO
-1;
973 else if (task_has_rt_policy(p
))
974 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
976 prio
= __normal_prio(p
);
981 * Calculate the current priority, i.e. the priority
982 * taken into account by the scheduler. This value might
983 * be boosted by RT tasks, or might be boosted by
984 * interactivity modifiers. Will be RT if the task got
985 * RT-boosted. If not then it returns p->normal_prio.
987 static int effective_prio(struct task_struct
*p
)
989 p
->normal_prio
= normal_prio(p
);
991 * If we are RT tasks or we were boosted to RT priority,
992 * keep the priority unchanged. Otherwise, update priority
993 * to the normal priority:
995 if (!rt_prio(p
->prio
))
996 return p
->normal_prio
;
1001 * task_curr - is this task currently executing on a CPU?
1002 * @p: the task in question.
1004 * Return: 1 if the task is currently executing. 0 otherwise.
1006 inline int task_curr(const struct task_struct
*p
)
1008 return cpu_curr(task_cpu(p
)) == p
;
1011 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1012 const struct sched_class
*prev_class
,
1015 if (prev_class
!= p
->sched_class
) {
1016 if (prev_class
->switched_from
)
1017 prev_class
->switched_from(rq
, p
);
1018 p
->sched_class
->switched_to(rq
, p
);
1019 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1020 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1023 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1025 const struct sched_class
*class;
1027 if (p
->sched_class
== rq
->curr
->sched_class
) {
1028 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1030 for_each_class(class) {
1031 if (class == rq
->curr
->sched_class
)
1033 if (class == p
->sched_class
) {
1041 * A queue event has occurred, and we're going to schedule. In
1042 * this case, we can save a useless back to back clock update.
1044 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1045 rq
->skip_clock_update
= 1;
1049 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1051 #ifdef CONFIG_SCHED_DEBUG
1053 * We should never call set_task_cpu() on a blocked task,
1054 * ttwu() will sort out the placement.
1056 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1057 !(task_preempt_count(p
) & PREEMPT_ACTIVE
));
1059 #ifdef CONFIG_LOCKDEP
1061 * The caller should hold either p->pi_lock or rq->lock, when changing
1062 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1064 * sched_move_task() holds both and thus holding either pins the cgroup,
1067 * Furthermore, all task_rq users should acquire both locks, see
1070 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1071 lockdep_is_held(&task_rq(p
)->lock
)));
1075 trace_sched_migrate_task(p
, new_cpu
);
1077 if (task_cpu(p
) != new_cpu
) {
1078 if (p
->sched_class
->migrate_task_rq
)
1079 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1080 p
->se
.nr_migrations
++;
1081 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1084 __set_task_cpu(p
, new_cpu
);
1087 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1089 if (task_on_rq_queued(p
)) {
1090 struct rq
*src_rq
, *dst_rq
;
1092 src_rq
= task_rq(p
);
1093 dst_rq
= cpu_rq(cpu
);
1095 deactivate_task(src_rq
, p
, 0);
1096 set_task_cpu(p
, cpu
);
1097 activate_task(dst_rq
, p
, 0);
1098 check_preempt_curr(dst_rq
, p
, 0);
1101 * Task isn't running anymore; make it appear like we migrated
1102 * it before it went to sleep. This means on wakeup we make the
1103 * previous cpu our targer instead of where it really is.
1109 struct migration_swap_arg
{
1110 struct task_struct
*src_task
, *dst_task
;
1111 int src_cpu
, dst_cpu
;
1114 static int migrate_swap_stop(void *data
)
1116 struct migration_swap_arg
*arg
= data
;
1117 struct rq
*src_rq
, *dst_rq
;
1120 src_rq
= cpu_rq(arg
->src_cpu
);
1121 dst_rq
= cpu_rq(arg
->dst_cpu
);
1123 double_raw_lock(&arg
->src_task
->pi_lock
,
1124 &arg
->dst_task
->pi_lock
);
1125 double_rq_lock(src_rq
, dst_rq
);
1126 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1129 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1132 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1135 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1138 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1139 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1144 double_rq_unlock(src_rq
, dst_rq
);
1145 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1146 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1152 * Cross migrate two tasks
1154 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1156 struct migration_swap_arg arg
;
1159 arg
= (struct migration_swap_arg
){
1161 .src_cpu
= task_cpu(cur
),
1163 .dst_cpu
= task_cpu(p
),
1166 if (arg
.src_cpu
== arg
.dst_cpu
)
1170 * These three tests are all lockless; this is OK since all of them
1171 * will be re-checked with proper locks held further down the line.
1173 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1176 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1179 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1182 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1183 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1189 struct migration_arg
{
1190 struct task_struct
*task
;
1194 static int migration_cpu_stop(void *data
);
1197 * wait_task_inactive - wait for a thread to unschedule.
1199 * If @match_state is nonzero, it's the @p->state value just checked and
1200 * not expected to change. If it changes, i.e. @p might have woken up,
1201 * then return zero. When we succeed in waiting for @p to be off its CPU,
1202 * we return a positive number (its total switch count). If a second call
1203 * a short while later returns the same number, the caller can be sure that
1204 * @p has remained unscheduled the whole time.
1206 * The caller must ensure that the task *will* unschedule sometime soon,
1207 * else this function might spin for a *long* time. This function can't
1208 * be called with interrupts off, or it may introduce deadlock with
1209 * smp_call_function() if an IPI is sent by the same process we are
1210 * waiting to become inactive.
1212 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1214 unsigned long flags
;
1215 int running
, queued
;
1221 * We do the initial early heuristics without holding
1222 * any task-queue locks at all. We'll only try to get
1223 * the runqueue lock when things look like they will
1229 * If the task is actively running on another CPU
1230 * still, just relax and busy-wait without holding
1233 * NOTE! Since we don't hold any locks, it's not
1234 * even sure that "rq" stays as the right runqueue!
1235 * But we don't care, since "task_running()" will
1236 * return false if the runqueue has changed and p
1237 * is actually now running somewhere else!
1239 while (task_running(rq
, p
)) {
1240 if (match_state
&& unlikely(p
->state
!= match_state
))
1246 * Ok, time to look more closely! We need the rq
1247 * lock now, to be *sure*. If we're wrong, we'll
1248 * just go back and repeat.
1250 rq
= task_rq_lock(p
, &flags
);
1251 trace_sched_wait_task(p
);
1252 running
= task_running(rq
, p
);
1253 queued
= task_on_rq_queued(p
);
1255 if (!match_state
|| p
->state
== match_state
)
1256 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1257 task_rq_unlock(rq
, p
, &flags
);
1260 * If it changed from the expected state, bail out now.
1262 if (unlikely(!ncsw
))
1266 * Was it really running after all now that we
1267 * checked with the proper locks actually held?
1269 * Oops. Go back and try again..
1271 if (unlikely(running
)) {
1277 * It's not enough that it's not actively running,
1278 * it must be off the runqueue _entirely_, and not
1281 * So if it was still runnable (but just not actively
1282 * running right now), it's preempted, and we should
1283 * yield - it could be a while.
1285 if (unlikely(queued
)) {
1286 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1288 set_current_state(TASK_UNINTERRUPTIBLE
);
1289 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1294 * Ahh, all good. It wasn't running, and it wasn't
1295 * runnable, which means that it will never become
1296 * running in the future either. We're all done!
1305 * kick_process - kick a running thread to enter/exit the kernel
1306 * @p: the to-be-kicked thread
1308 * Cause a process which is running on another CPU to enter
1309 * kernel-mode, without any delay. (to get signals handled.)
1311 * NOTE: this function doesn't have to take the runqueue lock,
1312 * because all it wants to ensure is that the remote task enters
1313 * the kernel. If the IPI races and the task has been migrated
1314 * to another CPU then no harm is done and the purpose has been
1317 void kick_process(struct task_struct
*p
)
1323 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1324 smp_send_reschedule(cpu
);
1327 EXPORT_SYMBOL_GPL(kick_process
);
1328 #endif /* CONFIG_SMP */
1332 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1334 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1336 int nid
= cpu_to_node(cpu
);
1337 const struct cpumask
*nodemask
= NULL
;
1338 enum { cpuset
, possible
, fail
} state
= cpuset
;
1342 * If the node that the cpu is on has been offlined, cpu_to_node()
1343 * will return -1. There is no cpu on the node, and we should
1344 * select the cpu on the other node.
1347 nodemask
= cpumask_of_node(nid
);
1349 /* Look for allowed, online CPU in same node. */
1350 for_each_cpu(dest_cpu
, nodemask
) {
1351 if (!cpu_online(dest_cpu
))
1353 if (!cpu_active(dest_cpu
))
1355 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1361 /* Any allowed, online CPU? */
1362 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1363 if (!cpu_online(dest_cpu
))
1365 if (!cpu_active(dest_cpu
))
1372 /* No more Mr. Nice Guy. */
1373 cpuset_cpus_allowed_fallback(p
);
1378 do_set_cpus_allowed(p
, cpu_possible_mask
);
1389 if (state
!= cpuset
) {
1391 * Don't tell them about moving exiting tasks or
1392 * kernel threads (both mm NULL), since they never
1395 if (p
->mm
&& printk_ratelimit()) {
1396 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1397 task_pid_nr(p
), p
->comm
, cpu
);
1405 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1408 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1410 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1413 * In order not to call set_task_cpu() on a blocking task we need
1414 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1417 * Since this is common to all placement strategies, this lives here.
1419 * [ this allows ->select_task() to simply return task_cpu(p) and
1420 * not worry about this generic constraint ]
1422 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1424 cpu
= select_fallback_rq(task_cpu(p
), p
);
1429 static void update_avg(u64
*avg
, u64 sample
)
1431 s64 diff
= sample
- *avg
;
1437 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1439 #ifdef CONFIG_SCHEDSTATS
1440 struct rq
*rq
= this_rq();
1443 int this_cpu
= smp_processor_id();
1445 if (cpu
== this_cpu
) {
1446 schedstat_inc(rq
, ttwu_local
);
1447 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1449 struct sched_domain
*sd
;
1451 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1453 for_each_domain(this_cpu
, sd
) {
1454 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1455 schedstat_inc(sd
, ttwu_wake_remote
);
1462 if (wake_flags
& WF_MIGRATED
)
1463 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1465 #endif /* CONFIG_SMP */
1467 schedstat_inc(rq
, ttwu_count
);
1468 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1470 if (wake_flags
& WF_SYNC
)
1471 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1473 #endif /* CONFIG_SCHEDSTATS */
1476 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1478 activate_task(rq
, p
, en_flags
);
1479 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1481 /* if a worker is waking up, notify workqueue */
1482 if (p
->flags
& PF_WQ_WORKER
)
1483 wq_worker_waking_up(p
, cpu_of(rq
));
1487 * Mark the task runnable and perform wakeup-preemption.
1490 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1492 check_preempt_curr(rq
, p
, wake_flags
);
1493 trace_sched_wakeup(p
, true);
1495 p
->state
= TASK_RUNNING
;
1497 if (p
->sched_class
->task_woken
)
1498 p
->sched_class
->task_woken(rq
, p
);
1500 if (rq
->idle_stamp
) {
1501 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1502 u64 max
= 2*rq
->max_idle_balance_cost
;
1504 update_avg(&rq
->avg_idle
, delta
);
1506 if (rq
->avg_idle
> max
)
1515 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1518 if (p
->sched_contributes_to_load
)
1519 rq
->nr_uninterruptible
--;
1522 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1523 ttwu_do_wakeup(rq
, p
, wake_flags
);
1527 * Called in case the task @p isn't fully descheduled from its runqueue,
1528 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1529 * since all we need to do is flip p->state to TASK_RUNNING, since
1530 * the task is still ->on_rq.
1532 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1537 rq
= __task_rq_lock(p
);
1538 if (task_on_rq_queued(p
)) {
1539 /* check_preempt_curr() may use rq clock */
1540 update_rq_clock(rq
);
1541 ttwu_do_wakeup(rq
, p
, wake_flags
);
1544 __task_rq_unlock(rq
);
1550 void sched_ttwu_pending(void)
1552 struct rq
*rq
= this_rq();
1553 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1554 struct task_struct
*p
;
1555 unsigned long flags
;
1560 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1563 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1564 llist
= llist_next(llist
);
1565 ttwu_do_activate(rq
, p
, 0);
1568 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1571 void scheduler_ipi(void)
1574 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1575 * TIF_NEED_RESCHED remotely (for the first time) will also send
1578 preempt_fold_need_resched();
1580 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1584 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1585 * traditionally all their work was done from the interrupt return
1586 * path. Now that we actually do some work, we need to make sure
1589 * Some archs already do call them, luckily irq_enter/exit nest
1592 * Arguably we should visit all archs and update all handlers,
1593 * however a fair share of IPIs are still resched only so this would
1594 * somewhat pessimize the simple resched case.
1597 sched_ttwu_pending();
1600 * Check if someone kicked us for doing the nohz idle load balance.
1602 if (unlikely(got_nohz_idle_kick())) {
1603 this_rq()->idle_balance
= 1;
1604 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1609 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1611 struct rq
*rq
= cpu_rq(cpu
);
1613 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1614 if (!set_nr_if_polling(rq
->idle
))
1615 smp_send_reschedule(cpu
);
1617 trace_sched_wake_idle_without_ipi(cpu
);
1621 void wake_up_if_idle(int cpu
)
1623 struct rq
*rq
= cpu_rq(cpu
);
1624 unsigned long flags
;
1626 if (!is_idle_task(rq
->curr
))
1629 if (set_nr_if_polling(rq
->idle
)) {
1630 trace_sched_wake_idle_without_ipi(cpu
);
1632 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1633 if (is_idle_task(rq
->curr
))
1634 smp_send_reschedule(cpu
);
1635 /* Else cpu is not in idle, do nothing here */
1636 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1640 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1642 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1644 #endif /* CONFIG_SMP */
1646 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1648 struct rq
*rq
= cpu_rq(cpu
);
1650 #if defined(CONFIG_SMP)
1651 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1652 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1653 ttwu_queue_remote(p
, cpu
);
1658 raw_spin_lock(&rq
->lock
);
1659 ttwu_do_activate(rq
, p
, 0);
1660 raw_spin_unlock(&rq
->lock
);
1664 * try_to_wake_up - wake up a thread
1665 * @p: the thread to be awakened
1666 * @state: the mask of task states that can be woken
1667 * @wake_flags: wake modifier flags (WF_*)
1669 * Put it on the run-queue if it's not already there. The "current"
1670 * thread is always on the run-queue (except when the actual
1671 * re-schedule is in progress), and as such you're allowed to do
1672 * the simpler "current->state = TASK_RUNNING" to mark yourself
1673 * runnable without the overhead of this.
1675 * Return: %true if @p was woken up, %false if it was already running.
1676 * or @state didn't match @p's state.
1679 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1681 unsigned long flags
;
1682 int cpu
, success
= 0;
1685 * If we are going to wake up a thread waiting for CONDITION we
1686 * need to ensure that CONDITION=1 done by the caller can not be
1687 * reordered with p->state check below. This pairs with mb() in
1688 * set_current_state() the waiting thread does.
1690 smp_mb__before_spinlock();
1691 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1692 if (!(p
->state
& state
))
1695 success
= 1; /* we're going to change ->state */
1698 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1703 * If the owning (remote) cpu is still in the middle of schedule() with
1704 * this task as prev, wait until its done referencing the task.
1709 * Pairs with the smp_wmb() in finish_lock_switch().
1713 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1714 p
->state
= TASK_WAKING
;
1716 if (p
->sched_class
->task_waking
)
1717 p
->sched_class
->task_waking(p
);
1719 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1720 if (task_cpu(p
) != cpu
) {
1721 wake_flags
|= WF_MIGRATED
;
1722 set_task_cpu(p
, cpu
);
1724 #endif /* CONFIG_SMP */
1728 ttwu_stat(p
, cpu
, wake_flags
);
1730 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1736 * try_to_wake_up_local - try to wake up a local task with rq lock held
1737 * @p: the thread to be awakened
1739 * Put @p on the run-queue if it's not already there. The caller must
1740 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1743 static void try_to_wake_up_local(struct task_struct
*p
)
1745 struct rq
*rq
= task_rq(p
);
1747 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1748 WARN_ON_ONCE(p
== current
))
1751 lockdep_assert_held(&rq
->lock
);
1753 if (!raw_spin_trylock(&p
->pi_lock
)) {
1754 raw_spin_unlock(&rq
->lock
);
1755 raw_spin_lock(&p
->pi_lock
);
1756 raw_spin_lock(&rq
->lock
);
1759 if (!(p
->state
& TASK_NORMAL
))
1762 if (!task_on_rq_queued(p
))
1763 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1765 ttwu_do_wakeup(rq
, p
, 0);
1766 ttwu_stat(p
, smp_processor_id(), 0);
1768 raw_spin_unlock(&p
->pi_lock
);
1772 * wake_up_process - Wake up a specific process
1773 * @p: The process to be woken up.
1775 * Attempt to wake up the nominated process and move it to the set of runnable
1778 * Return: 1 if the process was woken up, 0 if it was already running.
1780 * It may be assumed that this function implies a write memory barrier before
1781 * changing the task state if and only if any tasks are woken up.
1783 int wake_up_process(struct task_struct
*p
)
1785 WARN_ON(task_is_stopped_or_traced(p
));
1786 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1788 EXPORT_SYMBOL(wake_up_process
);
1790 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1792 return try_to_wake_up(p
, state
, 0);
1796 * This function clears the sched_dl_entity static params.
1798 void __dl_clear_params(struct task_struct
*p
)
1800 struct sched_dl_entity
*dl_se
= &p
->dl
;
1802 dl_se
->dl_runtime
= 0;
1803 dl_se
->dl_deadline
= 0;
1804 dl_se
->dl_period
= 0;
1810 * Perform scheduler related setup for a newly forked process p.
1811 * p is forked by current.
1813 * __sched_fork() is basic setup used by init_idle() too:
1815 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1820 p
->se
.exec_start
= 0;
1821 p
->se
.sum_exec_runtime
= 0;
1822 p
->se
.prev_sum_exec_runtime
= 0;
1823 p
->se
.nr_migrations
= 0;
1825 INIT_LIST_HEAD(&p
->se
.group_node
);
1827 #ifdef CONFIG_SCHEDSTATS
1828 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1831 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1832 hrtimer_init(&p
->dl
.dl_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1833 __dl_clear_params(p
);
1835 INIT_LIST_HEAD(&p
->rt
.run_list
);
1837 #ifdef CONFIG_PREEMPT_NOTIFIERS
1838 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1841 #ifdef CONFIG_NUMA_BALANCING
1842 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1843 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1844 p
->mm
->numa_scan_seq
= 0;
1847 if (clone_flags
& CLONE_VM
)
1848 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1850 p
->numa_preferred_nid
= -1;
1852 p
->node_stamp
= 0ULL;
1853 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1854 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1855 p
->numa_work
.next
= &p
->numa_work
;
1856 p
->numa_faults_memory
= NULL
;
1857 p
->numa_faults_buffer_memory
= NULL
;
1858 p
->last_task_numa_placement
= 0;
1859 p
->last_sum_exec_runtime
= 0;
1861 INIT_LIST_HEAD(&p
->numa_entry
);
1862 p
->numa_group
= NULL
;
1863 #endif /* CONFIG_NUMA_BALANCING */
1866 #ifdef CONFIG_NUMA_BALANCING
1867 #ifdef CONFIG_SCHED_DEBUG
1868 void set_numabalancing_state(bool enabled
)
1871 sched_feat_set("NUMA");
1873 sched_feat_set("NO_NUMA");
1876 __read_mostly
bool numabalancing_enabled
;
1878 void set_numabalancing_state(bool enabled
)
1880 numabalancing_enabled
= enabled
;
1882 #endif /* CONFIG_SCHED_DEBUG */
1884 #ifdef CONFIG_PROC_SYSCTL
1885 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1886 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1890 int state
= numabalancing_enabled
;
1892 if (write
&& !capable(CAP_SYS_ADMIN
))
1897 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1901 set_numabalancing_state(state
);
1908 * fork()/clone()-time setup:
1910 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1912 unsigned long flags
;
1913 int cpu
= get_cpu();
1915 __sched_fork(clone_flags
, p
);
1917 * We mark the process as running here. This guarantees that
1918 * nobody will actually run it, and a signal or other external
1919 * event cannot wake it up and insert it on the runqueue either.
1921 p
->state
= TASK_RUNNING
;
1924 * Make sure we do not leak PI boosting priority to the child.
1926 p
->prio
= current
->normal_prio
;
1929 * Revert to default priority/policy on fork if requested.
1931 if (unlikely(p
->sched_reset_on_fork
)) {
1932 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1933 p
->policy
= SCHED_NORMAL
;
1934 p
->static_prio
= NICE_TO_PRIO(0);
1936 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1937 p
->static_prio
= NICE_TO_PRIO(0);
1939 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1943 * We don't need the reset flag anymore after the fork. It has
1944 * fulfilled its duty:
1946 p
->sched_reset_on_fork
= 0;
1949 if (dl_prio(p
->prio
)) {
1952 } else if (rt_prio(p
->prio
)) {
1953 p
->sched_class
= &rt_sched_class
;
1955 p
->sched_class
= &fair_sched_class
;
1958 if (p
->sched_class
->task_fork
)
1959 p
->sched_class
->task_fork(p
);
1962 * The child is not yet in the pid-hash so no cgroup attach races,
1963 * and the cgroup is pinned to this child due to cgroup_fork()
1964 * is ran before sched_fork().
1966 * Silence PROVE_RCU.
1968 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1969 set_task_cpu(p
, cpu
);
1970 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1972 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1973 if (likely(sched_info_on()))
1974 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1976 #if defined(CONFIG_SMP)
1979 init_task_preempt_count(p
);
1981 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1982 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1989 unsigned long to_ratio(u64 period
, u64 runtime
)
1991 if (runtime
== RUNTIME_INF
)
1995 * Doing this here saves a lot of checks in all
1996 * the calling paths, and returning zero seems
1997 * safe for them anyway.
2002 return div64_u64(runtime
<< 20, period
);
2006 inline struct dl_bw
*dl_bw_of(int i
)
2008 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2009 "sched RCU must be held");
2010 return &cpu_rq(i
)->rd
->dl_bw
;
2013 static inline int dl_bw_cpus(int i
)
2015 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2018 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2019 "sched RCU must be held");
2020 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2026 inline struct dl_bw
*dl_bw_of(int i
)
2028 return &cpu_rq(i
)->dl
.dl_bw
;
2031 static inline int dl_bw_cpus(int i
)
2038 void __dl_clear(struct dl_bw
*dl_b
, u64 tsk_bw
)
2040 dl_b
->total_bw
-= tsk_bw
;
2044 void __dl_add(struct dl_bw
*dl_b
, u64 tsk_bw
)
2046 dl_b
->total_bw
+= tsk_bw
;
2050 bool __dl_overflow(struct dl_bw
*dl_b
, int cpus
, u64 old_bw
, u64 new_bw
)
2052 return dl_b
->bw
!= -1 &&
2053 dl_b
->bw
* cpus
< dl_b
->total_bw
- old_bw
+ new_bw
;
2057 * We must be sure that accepting a new task (or allowing changing the
2058 * parameters of an existing one) is consistent with the bandwidth
2059 * constraints. If yes, this function also accordingly updates the currently
2060 * allocated bandwidth to reflect the new situation.
2062 * This function is called while holding p's rq->lock.
2064 static int dl_overflow(struct task_struct
*p
, int policy
,
2065 const struct sched_attr
*attr
)
2068 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2069 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2070 u64 runtime
= attr
->sched_runtime
;
2071 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2074 if (new_bw
== p
->dl
.dl_bw
)
2078 * Either if a task, enters, leave, or stays -deadline but changes
2079 * its parameters, we may need to update accordingly the total
2080 * allocated bandwidth of the container.
2082 raw_spin_lock(&dl_b
->lock
);
2083 cpus
= dl_bw_cpus(task_cpu(p
));
2084 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2085 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2086 __dl_add(dl_b
, new_bw
);
2088 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2089 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2090 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2091 __dl_add(dl_b
, new_bw
);
2093 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2094 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2097 raw_spin_unlock(&dl_b
->lock
);
2102 extern void init_dl_bw(struct dl_bw
*dl_b
);
2105 * wake_up_new_task - wake up a newly created task for the first time.
2107 * This function will do some initial scheduler statistics housekeeping
2108 * that must be done for every newly created context, then puts the task
2109 * on the runqueue and wakes it.
2111 void wake_up_new_task(struct task_struct
*p
)
2113 unsigned long flags
;
2116 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2119 * Fork balancing, do it here and not earlier because:
2120 * - cpus_allowed can change in the fork path
2121 * - any previously selected cpu might disappear through hotplug
2123 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2126 /* Initialize new task's runnable average */
2127 init_task_runnable_average(p
);
2128 rq
= __task_rq_lock(p
);
2129 activate_task(rq
, p
, 0);
2130 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2131 trace_sched_wakeup_new(p
, true);
2132 check_preempt_curr(rq
, p
, WF_FORK
);
2134 if (p
->sched_class
->task_woken
)
2135 p
->sched_class
->task_woken(rq
, p
);
2137 task_rq_unlock(rq
, p
, &flags
);
2140 #ifdef CONFIG_PREEMPT_NOTIFIERS
2143 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2144 * @notifier: notifier struct to register
2146 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2148 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2150 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2153 * preempt_notifier_unregister - no longer interested in preemption notifications
2154 * @notifier: notifier struct to unregister
2156 * This is safe to call from within a preemption notifier.
2158 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2160 hlist_del(¬ifier
->link
);
2162 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2164 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2166 struct preempt_notifier
*notifier
;
2168 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2169 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2173 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2174 struct task_struct
*next
)
2176 struct preempt_notifier
*notifier
;
2178 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2179 notifier
->ops
->sched_out(notifier
, next
);
2182 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2184 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2189 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2190 struct task_struct
*next
)
2194 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2197 * prepare_task_switch - prepare to switch tasks
2198 * @rq: the runqueue preparing to switch
2199 * @prev: the current task that is being switched out
2200 * @next: the task we are going to switch to.
2202 * This is called with the rq lock held and interrupts off. It must
2203 * be paired with a subsequent finish_task_switch after the context
2206 * prepare_task_switch sets up locking and calls architecture specific
2210 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2211 struct task_struct
*next
)
2213 trace_sched_switch(prev
, next
);
2214 sched_info_switch(rq
, prev
, next
);
2215 perf_event_task_sched_out(prev
, next
);
2216 fire_sched_out_preempt_notifiers(prev
, next
);
2217 prepare_lock_switch(rq
, next
);
2218 prepare_arch_switch(next
);
2222 * finish_task_switch - clean up after a task-switch
2223 * @rq: runqueue associated with task-switch
2224 * @prev: the thread we just switched away from.
2226 * finish_task_switch must be called after the context switch, paired
2227 * with a prepare_task_switch call before the context switch.
2228 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2229 * and do any other architecture-specific cleanup actions.
2231 * Note that we may have delayed dropping an mm in context_switch(). If
2232 * so, we finish that here outside of the runqueue lock. (Doing it
2233 * with the lock held can cause deadlocks; see schedule() for
2236 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2237 __releases(rq
->lock
)
2239 struct mm_struct
*mm
= rq
->prev_mm
;
2245 * A task struct has one reference for the use as "current".
2246 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2247 * schedule one last time. The schedule call will never return, and
2248 * the scheduled task must drop that reference.
2249 * The test for TASK_DEAD must occur while the runqueue locks are
2250 * still held, otherwise prev could be scheduled on another cpu, die
2251 * there before we look at prev->state, and then the reference would
2253 * Manfred Spraul <manfred@colorfullife.com>
2255 prev_state
= prev
->state
;
2256 vtime_task_switch(prev
);
2257 finish_arch_switch(prev
);
2258 perf_event_task_sched_in(prev
, current
);
2259 finish_lock_switch(rq
, prev
);
2260 finish_arch_post_lock_switch();
2262 fire_sched_in_preempt_notifiers(current
);
2265 if (unlikely(prev_state
== TASK_DEAD
)) {
2266 if (prev
->sched_class
->task_dead
)
2267 prev
->sched_class
->task_dead(prev
);
2270 * Remove function-return probe instances associated with this
2271 * task and put them back on the free list.
2273 kprobe_flush_task(prev
);
2274 put_task_struct(prev
);
2277 tick_nohz_task_switch(current
);
2282 /* rq->lock is NOT held, but preemption is disabled */
2283 static inline void post_schedule(struct rq
*rq
)
2285 if (rq
->post_schedule
) {
2286 unsigned long flags
;
2288 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2289 if (rq
->curr
->sched_class
->post_schedule
)
2290 rq
->curr
->sched_class
->post_schedule(rq
);
2291 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2293 rq
->post_schedule
= 0;
2299 static inline void post_schedule(struct rq
*rq
)
2306 * schedule_tail - first thing a freshly forked thread must call.
2307 * @prev: the thread we just switched away from.
2309 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2310 __releases(rq
->lock
)
2312 struct rq
*rq
= this_rq();
2314 finish_task_switch(rq
, prev
);
2317 * FIXME: do we need to worry about rq being invalidated by the
2322 if (current
->set_child_tid
)
2323 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2327 * context_switch - switch to the new MM and the new
2328 * thread's register state.
2331 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2332 struct task_struct
*next
)
2334 struct mm_struct
*mm
, *oldmm
;
2336 prepare_task_switch(rq
, prev
, next
);
2339 oldmm
= prev
->active_mm
;
2341 * For paravirt, this is coupled with an exit in switch_to to
2342 * combine the page table reload and the switch backend into
2345 arch_start_context_switch(prev
);
2348 next
->active_mm
= oldmm
;
2349 atomic_inc(&oldmm
->mm_count
);
2350 enter_lazy_tlb(oldmm
, next
);
2352 switch_mm(oldmm
, mm
, next
);
2355 prev
->active_mm
= NULL
;
2356 rq
->prev_mm
= oldmm
;
2359 * Since the runqueue lock will be released by the next
2360 * task (which is an invalid locking op but in the case
2361 * of the scheduler it's an obvious special-case), so we
2362 * do an early lockdep release here:
2364 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2366 context_tracking_task_switch(prev
, next
);
2367 /* Here we just switch the register state and the stack. */
2368 switch_to(prev
, next
, prev
);
2372 * this_rq must be evaluated again because prev may have moved
2373 * CPUs since it called schedule(), thus the 'rq' on its stack
2374 * frame will be invalid.
2376 finish_task_switch(this_rq(), prev
);
2380 * nr_running and nr_context_switches:
2382 * externally visible scheduler statistics: current number of runnable
2383 * threads, total number of context switches performed since bootup.
2385 unsigned long nr_running(void)
2387 unsigned long i
, sum
= 0;
2389 for_each_online_cpu(i
)
2390 sum
+= cpu_rq(i
)->nr_running
;
2396 * Check if only the current task is running on the cpu.
2398 bool single_task_running(void)
2400 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2405 EXPORT_SYMBOL(single_task_running
);
2407 unsigned long long nr_context_switches(void)
2410 unsigned long long sum
= 0;
2412 for_each_possible_cpu(i
)
2413 sum
+= cpu_rq(i
)->nr_switches
;
2418 unsigned long nr_iowait(void)
2420 unsigned long i
, sum
= 0;
2422 for_each_possible_cpu(i
)
2423 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2428 unsigned long nr_iowait_cpu(int cpu
)
2430 struct rq
*this = cpu_rq(cpu
);
2431 return atomic_read(&this->nr_iowait
);
2434 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2436 struct rq
*this = this_rq();
2437 *nr_waiters
= atomic_read(&this->nr_iowait
);
2438 *load
= this->cpu_load
[0];
2444 * sched_exec - execve() is a valuable balancing opportunity, because at
2445 * this point the task has the smallest effective memory and cache footprint.
2447 void sched_exec(void)
2449 struct task_struct
*p
= current
;
2450 unsigned long flags
;
2453 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2454 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2455 if (dest_cpu
== smp_processor_id())
2458 if (likely(cpu_active(dest_cpu
))) {
2459 struct migration_arg arg
= { p
, dest_cpu
};
2461 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2462 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2466 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2471 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2472 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2474 EXPORT_PER_CPU_SYMBOL(kstat
);
2475 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2478 * Return any ns on the sched_clock that have not yet been accounted in
2479 * @p in case that task is currently running.
2481 * Called with task_rq_lock() held on @rq.
2483 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2488 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2489 * project cycles that may never be accounted to this
2490 * thread, breaking clock_gettime().
2492 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2493 update_rq_clock(rq
);
2494 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2502 unsigned long long task_delta_exec(struct task_struct
*p
)
2504 unsigned long flags
;
2508 rq
= task_rq_lock(p
, &flags
);
2509 ns
= do_task_delta_exec(p
, rq
);
2510 task_rq_unlock(rq
, p
, &flags
);
2516 * Return accounted runtime for the task.
2517 * In case the task is currently running, return the runtime plus current's
2518 * pending runtime that have not been accounted yet.
2520 unsigned long long task_sched_runtime(struct task_struct
*p
)
2522 unsigned long flags
;
2526 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2528 * 64-bit doesn't need locks to atomically read a 64bit value.
2529 * So we have a optimization chance when the task's delta_exec is 0.
2530 * Reading ->on_cpu is racy, but this is ok.
2532 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2533 * If we race with it entering cpu, unaccounted time is 0. This is
2534 * indistinguishable from the read occurring a few cycles earlier.
2535 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2536 * been accounted, so we're correct here as well.
2538 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2539 return p
->se
.sum_exec_runtime
;
2542 rq
= task_rq_lock(p
, &flags
);
2543 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2544 task_rq_unlock(rq
, p
, &flags
);
2550 * This function gets called by the timer code, with HZ frequency.
2551 * We call it with interrupts disabled.
2553 void scheduler_tick(void)
2555 int cpu
= smp_processor_id();
2556 struct rq
*rq
= cpu_rq(cpu
);
2557 struct task_struct
*curr
= rq
->curr
;
2561 raw_spin_lock(&rq
->lock
);
2562 update_rq_clock(rq
);
2563 curr
->sched_class
->task_tick(rq
, curr
, 0);
2564 update_cpu_load_active(rq
);
2565 raw_spin_unlock(&rq
->lock
);
2567 perf_event_task_tick();
2570 rq
->idle_balance
= idle_cpu(cpu
);
2571 trigger_load_balance(rq
);
2573 rq_last_tick_reset(rq
);
2576 #ifdef CONFIG_NO_HZ_FULL
2578 * scheduler_tick_max_deferment
2580 * Keep at least one tick per second when a single
2581 * active task is running because the scheduler doesn't
2582 * yet completely support full dynticks environment.
2584 * This makes sure that uptime, CFS vruntime, load
2585 * balancing, etc... continue to move forward, even
2586 * with a very low granularity.
2588 * Return: Maximum deferment in nanoseconds.
2590 u64
scheduler_tick_max_deferment(void)
2592 struct rq
*rq
= this_rq();
2593 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2595 next
= rq
->last_sched_tick
+ HZ
;
2597 if (time_before_eq(next
, now
))
2600 return jiffies_to_nsecs(next
- now
);
2604 notrace
unsigned long get_parent_ip(unsigned long addr
)
2606 if (in_lock_functions(addr
)) {
2607 addr
= CALLER_ADDR2
;
2608 if (in_lock_functions(addr
))
2609 addr
= CALLER_ADDR3
;
2614 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2615 defined(CONFIG_PREEMPT_TRACER))
2617 void preempt_count_add(int val
)
2619 #ifdef CONFIG_DEBUG_PREEMPT
2623 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2626 __preempt_count_add(val
);
2627 #ifdef CONFIG_DEBUG_PREEMPT
2629 * Spinlock count overflowing soon?
2631 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2634 if (preempt_count() == val
) {
2635 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2636 #ifdef CONFIG_DEBUG_PREEMPT
2637 current
->preempt_disable_ip
= ip
;
2639 trace_preempt_off(CALLER_ADDR0
, ip
);
2642 EXPORT_SYMBOL(preempt_count_add
);
2643 NOKPROBE_SYMBOL(preempt_count_add
);
2645 void preempt_count_sub(int val
)
2647 #ifdef CONFIG_DEBUG_PREEMPT
2651 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2654 * Is the spinlock portion underflowing?
2656 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2657 !(preempt_count() & PREEMPT_MASK
)))
2661 if (preempt_count() == val
)
2662 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2663 __preempt_count_sub(val
);
2665 EXPORT_SYMBOL(preempt_count_sub
);
2666 NOKPROBE_SYMBOL(preempt_count_sub
);
2671 * Print scheduling while atomic bug:
2673 static noinline
void __schedule_bug(struct task_struct
*prev
)
2675 if (oops_in_progress
)
2678 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2679 prev
->comm
, prev
->pid
, preempt_count());
2681 debug_show_held_locks(prev
);
2683 if (irqs_disabled())
2684 print_irqtrace_events(prev
);
2685 #ifdef CONFIG_DEBUG_PREEMPT
2686 if (in_atomic_preempt_off()) {
2687 pr_err("Preemption disabled at:");
2688 print_ip_sym(current
->preempt_disable_ip
);
2693 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2697 * Various schedule()-time debugging checks and statistics:
2699 static inline void schedule_debug(struct task_struct
*prev
)
2701 #ifdef CONFIG_SCHED_STACK_END_CHECK
2702 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2705 * Test if we are atomic. Since do_exit() needs to call into
2706 * schedule() atomically, we ignore that path. Otherwise whine
2707 * if we are scheduling when we should not.
2709 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2710 __schedule_bug(prev
);
2713 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2715 schedstat_inc(this_rq(), sched_count
);
2719 * Pick up the highest-prio task:
2721 static inline struct task_struct
*
2722 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2724 const struct sched_class
*class = &fair_sched_class
;
2725 struct task_struct
*p
;
2728 * Optimization: we know that if all tasks are in
2729 * the fair class we can call that function directly:
2731 if (likely(prev
->sched_class
== class &&
2732 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2733 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2734 if (unlikely(p
== RETRY_TASK
))
2737 /* assumes fair_sched_class->next == idle_sched_class */
2739 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2745 for_each_class(class) {
2746 p
= class->pick_next_task(rq
, prev
);
2748 if (unlikely(p
== RETRY_TASK
))
2754 BUG(); /* the idle class will always have a runnable task */
2758 * __schedule() is the main scheduler function.
2760 * The main means of driving the scheduler and thus entering this function are:
2762 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2764 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2765 * paths. For example, see arch/x86/entry_64.S.
2767 * To drive preemption between tasks, the scheduler sets the flag in timer
2768 * interrupt handler scheduler_tick().
2770 * 3. Wakeups don't really cause entry into schedule(). They add a
2771 * task to the run-queue and that's it.
2773 * Now, if the new task added to the run-queue preempts the current
2774 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2775 * called on the nearest possible occasion:
2777 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2779 * - in syscall or exception context, at the next outmost
2780 * preempt_enable(). (this might be as soon as the wake_up()'s
2783 * - in IRQ context, return from interrupt-handler to
2784 * preemptible context
2786 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2789 * - cond_resched() call
2790 * - explicit schedule() call
2791 * - return from syscall or exception to user-space
2792 * - return from interrupt-handler to user-space
2794 static void __sched
__schedule(void)
2796 struct task_struct
*prev
, *next
;
2797 unsigned long *switch_count
;
2803 cpu
= smp_processor_id();
2805 rcu_note_context_switch(cpu
);
2808 schedule_debug(prev
);
2810 if (sched_feat(HRTICK
))
2814 * Make sure that signal_pending_state()->signal_pending() below
2815 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2816 * done by the caller to avoid the race with signal_wake_up().
2818 smp_mb__before_spinlock();
2819 raw_spin_lock_irq(&rq
->lock
);
2821 switch_count
= &prev
->nivcsw
;
2822 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2823 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2824 prev
->state
= TASK_RUNNING
;
2826 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2830 * If a worker went to sleep, notify and ask workqueue
2831 * whether it wants to wake up a task to maintain
2834 if (prev
->flags
& PF_WQ_WORKER
) {
2835 struct task_struct
*to_wakeup
;
2837 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2839 try_to_wake_up_local(to_wakeup
);
2842 switch_count
= &prev
->nvcsw
;
2845 if (task_on_rq_queued(prev
) || rq
->skip_clock_update
< 0)
2846 update_rq_clock(rq
);
2848 next
= pick_next_task(rq
, prev
);
2849 clear_tsk_need_resched(prev
);
2850 clear_preempt_need_resched();
2851 rq
->skip_clock_update
= 0;
2853 if (likely(prev
!= next
)) {
2858 context_switch(rq
, prev
, next
); /* unlocks the rq */
2860 * The context switch have flipped the stack from under us
2861 * and restored the local variables which were saved when
2862 * this task called schedule() in the past. prev == current
2863 * is still correct, but it can be moved to another cpu/rq.
2865 cpu
= smp_processor_id();
2868 raw_spin_unlock_irq(&rq
->lock
);
2872 sched_preempt_enable_no_resched();
2877 static inline void sched_submit_work(struct task_struct
*tsk
)
2879 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2882 * If we are going to sleep and we have plugged IO queued,
2883 * make sure to submit it to avoid deadlocks.
2885 if (blk_needs_flush_plug(tsk
))
2886 blk_schedule_flush_plug(tsk
);
2889 asmlinkage __visible
void __sched
schedule(void)
2891 struct task_struct
*tsk
= current
;
2893 sched_submit_work(tsk
);
2896 EXPORT_SYMBOL(schedule
);
2898 #ifdef CONFIG_CONTEXT_TRACKING
2899 asmlinkage __visible
void __sched
schedule_user(void)
2902 * If we come here after a random call to set_need_resched(),
2903 * or we have been woken up remotely but the IPI has not yet arrived,
2904 * we haven't yet exited the RCU idle mode. Do it here manually until
2905 * we find a better solution.
2914 * schedule_preempt_disabled - called with preemption disabled
2916 * Returns with preemption disabled. Note: preempt_count must be 1
2918 void __sched
schedule_preempt_disabled(void)
2920 sched_preempt_enable_no_resched();
2925 #ifdef CONFIG_PREEMPT
2927 * this is the entry point to schedule() from in-kernel preemption
2928 * off of preempt_enable. Kernel preemptions off return from interrupt
2929 * occur there and call schedule directly.
2931 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2934 * If there is a non-zero preempt_count or interrupts are disabled,
2935 * we do not want to preempt the current task. Just return..
2937 if (likely(!preemptible()))
2941 __preempt_count_add(PREEMPT_ACTIVE
);
2943 __preempt_count_sub(PREEMPT_ACTIVE
);
2946 * Check again in case we missed a preemption opportunity
2947 * between schedule and now.
2950 } while (need_resched());
2952 NOKPROBE_SYMBOL(preempt_schedule
);
2953 EXPORT_SYMBOL(preempt_schedule
);
2954 #endif /* CONFIG_PREEMPT */
2957 * this is the entry point to schedule() from kernel preemption
2958 * off of irq context.
2959 * Note, that this is called and return with irqs disabled. This will
2960 * protect us against recursive calling from irq.
2962 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2964 enum ctx_state prev_state
;
2966 /* Catch callers which need to be fixed */
2967 BUG_ON(preempt_count() || !irqs_disabled());
2969 prev_state
= exception_enter();
2972 __preempt_count_add(PREEMPT_ACTIVE
);
2975 local_irq_disable();
2976 __preempt_count_sub(PREEMPT_ACTIVE
);
2979 * Check again in case we missed a preemption opportunity
2980 * between schedule and now.
2983 } while (need_resched());
2985 exception_exit(prev_state
);
2988 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2991 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2993 EXPORT_SYMBOL(default_wake_function
);
2995 #ifdef CONFIG_RT_MUTEXES
2998 * rt_mutex_setprio - set the current priority of a task
3000 * @prio: prio value (kernel-internal form)
3002 * This function changes the 'effective' priority of a task. It does
3003 * not touch ->normal_prio like __setscheduler().
3005 * Used by the rt_mutex code to implement priority inheritance
3006 * logic. Call site only calls if the priority of the task changed.
3008 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3010 int oldprio
, queued
, running
, enqueue_flag
= 0;
3012 const struct sched_class
*prev_class
;
3014 BUG_ON(prio
> MAX_PRIO
);
3016 rq
= __task_rq_lock(p
);
3019 * Idle task boosting is a nono in general. There is one
3020 * exception, when PREEMPT_RT and NOHZ is active:
3022 * The idle task calls get_next_timer_interrupt() and holds
3023 * the timer wheel base->lock on the CPU and another CPU wants
3024 * to access the timer (probably to cancel it). We can safely
3025 * ignore the boosting request, as the idle CPU runs this code
3026 * with interrupts disabled and will complete the lock
3027 * protected section without being interrupted. So there is no
3028 * real need to boost.
3030 if (unlikely(p
== rq
->idle
)) {
3031 WARN_ON(p
!= rq
->curr
);
3032 WARN_ON(p
->pi_blocked_on
);
3036 trace_sched_pi_setprio(p
, prio
);
3038 prev_class
= p
->sched_class
;
3039 queued
= task_on_rq_queued(p
);
3040 running
= task_current(rq
, p
);
3042 dequeue_task(rq
, p
, 0);
3044 put_prev_task(rq
, p
);
3047 * Boosting condition are:
3048 * 1. -rt task is running and holds mutex A
3049 * --> -dl task blocks on mutex A
3051 * 2. -dl task is running and holds mutex A
3052 * --> -dl task blocks on mutex A and could preempt the
3055 if (dl_prio(prio
)) {
3056 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3057 if (!dl_prio(p
->normal_prio
) ||
3058 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3059 p
->dl
.dl_boosted
= 1;
3060 p
->dl
.dl_throttled
= 0;
3061 enqueue_flag
= ENQUEUE_REPLENISH
;
3063 p
->dl
.dl_boosted
= 0;
3064 p
->sched_class
= &dl_sched_class
;
3065 } else if (rt_prio(prio
)) {
3066 if (dl_prio(oldprio
))
3067 p
->dl
.dl_boosted
= 0;
3069 enqueue_flag
= ENQUEUE_HEAD
;
3070 p
->sched_class
= &rt_sched_class
;
3072 if (dl_prio(oldprio
))
3073 p
->dl
.dl_boosted
= 0;
3074 p
->sched_class
= &fair_sched_class
;
3080 p
->sched_class
->set_curr_task(rq
);
3082 enqueue_task(rq
, p
, enqueue_flag
);
3084 check_class_changed(rq
, p
, prev_class
, oldprio
);
3086 __task_rq_unlock(rq
);
3090 void set_user_nice(struct task_struct
*p
, long nice
)
3092 int old_prio
, delta
, queued
;
3093 unsigned long flags
;
3096 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3099 * We have to be careful, if called from sys_setpriority(),
3100 * the task might be in the middle of scheduling on another CPU.
3102 rq
= task_rq_lock(p
, &flags
);
3104 * The RT priorities are set via sched_setscheduler(), but we still
3105 * allow the 'normal' nice value to be set - but as expected
3106 * it wont have any effect on scheduling until the task is
3107 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3109 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3110 p
->static_prio
= NICE_TO_PRIO(nice
);
3113 queued
= task_on_rq_queued(p
);
3115 dequeue_task(rq
, p
, 0);
3117 p
->static_prio
= NICE_TO_PRIO(nice
);
3120 p
->prio
= effective_prio(p
);
3121 delta
= p
->prio
- old_prio
;
3124 enqueue_task(rq
, p
, 0);
3126 * If the task increased its priority or is running and
3127 * lowered its priority, then reschedule its CPU:
3129 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3133 task_rq_unlock(rq
, p
, &flags
);
3135 EXPORT_SYMBOL(set_user_nice
);
3138 * can_nice - check if a task can reduce its nice value
3142 int can_nice(const struct task_struct
*p
, const int nice
)
3144 /* convert nice value [19,-20] to rlimit style value [1,40] */
3145 int nice_rlim
= nice_to_rlimit(nice
);
3147 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3148 capable(CAP_SYS_NICE
));
3151 #ifdef __ARCH_WANT_SYS_NICE
3154 * sys_nice - change the priority of the current process.
3155 * @increment: priority increment
3157 * sys_setpriority is a more generic, but much slower function that
3158 * does similar things.
3160 SYSCALL_DEFINE1(nice
, int, increment
)
3165 * Setpriority might change our priority at the same moment.
3166 * We don't have to worry. Conceptually one call occurs first
3167 * and we have a single winner.
3169 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3170 nice
= task_nice(current
) + increment
;
3172 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3173 if (increment
< 0 && !can_nice(current
, nice
))
3176 retval
= security_task_setnice(current
, nice
);
3180 set_user_nice(current
, nice
);
3187 * task_prio - return the priority value of a given task.
3188 * @p: the task in question.
3190 * Return: The priority value as seen by users in /proc.
3191 * RT tasks are offset by -200. Normal tasks are centered
3192 * around 0, value goes from -16 to +15.
3194 int task_prio(const struct task_struct
*p
)
3196 return p
->prio
- MAX_RT_PRIO
;
3200 * idle_cpu - is a given cpu idle currently?
3201 * @cpu: the processor in question.
3203 * Return: 1 if the CPU is currently idle. 0 otherwise.
3205 int idle_cpu(int cpu
)
3207 struct rq
*rq
= cpu_rq(cpu
);
3209 if (rq
->curr
!= rq
->idle
)
3216 if (!llist_empty(&rq
->wake_list
))
3224 * idle_task - return the idle task for a given cpu.
3225 * @cpu: the processor in question.
3227 * Return: The idle task for the cpu @cpu.
3229 struct task_struct
*idle_task(int cpu
)
3231 return cpu_rq(cpu
)->idle
;
3235 * find_process_by_pid - find a process with a matching PID value.
3236 * @pid: the pid in question.
3238 * The task of @pid, if found. %NULL otherwise.
3240 static struct task_struct
*find_process_by_pid(pid_t pid
)
3242 return pid
? find_task_by_vpid(pid
) : current
;
3246 * This function initializes the sched_dl_entity of a newly becoming
3247 * SCHED_DEADLINE task.
3249 * Only the static values are considered here, the actual runtime and the
3250 * absolute deadline will be properly calculated when the task is enqueued
3251 * for the first time with its new policy.
3254 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3256 struct sched_dl_entity
*dl_se
= &p
->dl
;
3258 init_dl_task_timer(dl_se
);
3259 dl_se
->dl_runtime
= attr
->sched_runtime
;
3260 dl_se
->dl_deadline
= attr
->sched_deadline
;
3261 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3262 dl_se
->flags
= attr
->sched_flags
;
3263 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3264 dl_se
->dl_throttled
= 0;
3266 dl_se
->dl_yielded
= 0;
3270 * sched_setparam() passes in -1 for its policy, to let the functions
3271 * it calls know not to change it.
3273 #define SETPARAM_POLICY -1
3275 static void __setscheduler_params(struct task_struct
*p
,
3276 const struct sched_attr
*attr
)
3278 int policy
= attr
->sched_policy
;
3280 if (policy
== SETPARAM_POLICY
)
3285 if (dl_policy(policy
))
3286 __setparam_dl(p
, attr
);
3287 else if (fair_policy(policy
))
3288 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3291 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3292 * !rt_policy. Always setting this ensures that things like
3293 * getparam()/getattr() don't report silly values for !rt tasks.
3295 p
->rt_priority
= attr
->sched_priority
;
3296 p
->normal_prio
= normal_prio(p
);
3300 /* Actually do priority change: must hold pi & rq lock. */
3301 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3302 const struct sched_attr
*attr
)
3304 __setscheduler_params(p
, attr
);
3307 * If we get here, there was no pi waiters boosting the
3308 * task. It is safe to use the normal prio.
3310 p
->prio
= normal_prio(p
);
3312 if (dl_prio(p
->prio
))
3313 p
->sched_class
= &dl_sched_class
;
3314 else if (rt_prio(p
->prio
))
3315 p
->sched_class
= &rt_sched_class
;
3317 p
->sched_class
= &fair_sched_class
;
3321 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3323 struct sched_dl_entity
*dl_se
= &p
->dl
;
3325 attr
->sched_priority
= p
->rt_priority
;
3326 attr
->sched_runtime
= dl_se
->dl_runtime
;
3327 attr
->sched_deadline
= dl_se
->dl_deadline
;
3328 attr
->sched_period
= dl_se
->dl_period
;
3329 attr
->sched_flags
= dl_se
->flags
;
3333 * This function validates the new parameters of a -deadline task.
3334 * We ask for the deadline not being zero, and greater or equal
3335 * than the runtime, as well as the period of being zero or
3336 * greater than deadline. Furthermore, we have to be sure that
3337 * user parameters are above the internal resolution of 1us (we
3338 * check sched_runtime only since it is always the smaller one) and
3339 * below 2^63 ns (we have to check both sched_deadline and
3340 * sched_period, as the latter can be zero).
3343 __checkparam_dl(const struct sched_attr
*attr
)
3346 if (attr
->sched_deadline
== 0)
3350 * Since we truncate DL_SCALE bits, make sure we're at least
3353 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3357 * Since we use the MSB for wrap-around and sign issues, make
3358 * sure it's not set (mind that period can be equal to zero).
3360 if (attr
->sched_deadline
& (1ULL << 63) ||
3361 attr
->sched_period
& (1ULL << 63))
3364 /* runtime <= deadline <= period (if period != 0) */
3365 if ((attr
->sched_period
!= 0 &&
3366 attr
->sched_period
< attr
->sched_deadline
) ||
3367 attr
->sched_deadline
< attr
->sched_runtime
)
3374 * check the target process has a UID that matches the current process's
3376 static bool check_same_owner(struct task_struct
*p
)
3378 const struct cred
*cred
= current_cred(), *pcred
;
3382 pcred
= __task_cred(p
);
3383 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3384 uid_eq(cred
->euid
, pcred
->uid
));
3389 static int __sched_setscheduler(struct task_struct
*p
,
3390 const struct sched_attr
*attr
,
3393 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3394 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3395 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3396 int policy
= attr
->sched_policy
;
3397 unsigned long flags
;
3398 const struct sched_class
*prev_class
;
3402 /* may grab non-irq protected spin_locks */
3403 BUG_ON(in_interrupt());
3405 /* double check policy once rq lock held */
3407 reset_on_fork
= p
->sched_reset_on_fork
;
3408 policy
= oldpolicy
= p
->policy
;
3410 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3412 if (policy
!= SCHED_DEADLINE
&&
3413 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3414 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3415 policy
!= SCHED_IDLE
)
3419 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3423 * Valid priorities for SCHED_FIFO and SCHED_RR are
3424 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3425 * SCHED_BATCH and SCHED_IDLE is 0.
3427 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3428 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3430 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3431 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3435 * Allow unprivileged RT tasks to decrease priority:
3437 if (user
&& !capable(CAP_SYS_NICE
)) {
3438 if (fair_policy(policy
)) {
3439 if (attr
->sched_nice
< task_nice(p
) &&
3440 !can_nice(p
, attr
->sched_nice
))
3444 if (rt_policy(policy
)) {
3445 unsigned long rlim_rtprio
=
3446 task_rlimit(p
, RLIMIT_RTPRIO
);
3448 /* can't set/change the rt policy */
3449 if (policy
!= p
->policy
&& !rlim_rtprio
)
3452 /* can't increase priority */
3453 if (attr
->sched_priority
> p
->rt_priority
&&
3454 attr
->sched_priority
> rlim_rtprio
)
3459 * Can't set/change SCHED_DEADLINE policy at all for now
3460 * (safest behavior); in the future we would like to allow
3461 * unprivileged DL tasks to increase their relative deadline
3462 * or reduce their runtime (both ways reducing utilization)
3464 if (dl_policy(policy
))
3468 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3469 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3471 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3472 if (!can_nice(p
, task_nice(p
)))
3476 /* can't change other user's priorities */
3477 if (!check_same_owner(p
))
3480 /* Normal users shall not reset the sched_reset_on_fork flag */
3481 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3486 retval
= security_task_setscheduler(p
);
3492 * make sure no PI-waiters arrive (or leave) while we are
3493 * changing the priority of the task:
3495 * To be able to change p->policy safely, the appropriate
3496 * runqueue lock must be held.
3498 rq
= task_rq_lock(p
, &flags
);
3501 * Changing the policy of the stop threads its a very bad idea
3503 if (p
== rq
->stop
) {
3504 task_rq_unlock(rq
, p
, &flags
);
3509 * If not changing anything there's no need to proceed further,
3510 * but store a possible modification of reset_on_fork.
3512 if (unlikely(policy
== p
->policy
)) {
3513 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3515 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3517 if (dl_policy(policy
))
3520 p
->sched_reset_on_fork
= reset_on_fork
;
3521 task_rq_unlock(rq
, p
, &flags
);
3527 #ifdef CONFIG_RT_GROUP_SCHED
3529 * Do not allow realtime tasks into groups that have no runtime
3532 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3533 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3534 !task_group_is_autogroup(task_group(p
))) {
3535 task_rq_unlock(rq
, p
, &flags
);
3540 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3541 cpumask_t
*span
= rq
->rd
->span
;
3544 * Don't allow tasks with an affinity mask smaller than
3545 * the entire root_domain to become SCHED_DEADLINE. We
3546 * will also fail if there's no bandwidth available.
3548 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3549 rq
->rd
->dl_bw
.bw
== 0) {
3550 task_rq_unlock(rq
, p
, &flags
);
3557 /* recheck policy now with rq lock held */
3558 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3559 policy
= oldpolicy
= -1;
3560 task_rq_unlock(rq
, p
, &flags
);
3565 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3566 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3569 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3570 task_rq_unlock(rq
, p
, &flags
);
3574 p
->sched_reset_on_fork
= reset_on_fork
;
3578 * Special case for priority boosted tasks.
3580 * If the new priority is lower or equal (user space view)
3581 * than the current (boosted) priority, we just store the new
3582 * normal parameters and do not touch the scheduler class and
3583 * the runqueue. This will be done when the task deboost
3586 if (rt_mutex_check_prio(p
, newprio
)) {
3587 __setscheduler_params(p
, attr
);
3588 task_rq_unlock(rq
, p
, &flags
);
3592 queued
= task_on_rq_queued(p
);
3593 running
= task_current(rq
, p
);
3595 dequeue_task(rq
, p
, 0);
3597 put_prev_task(rq
, p
);
3599 prev_class
= p
->sched_class
;
3600 __setscheduler(rq
, p
, attr
);
3603 p
->sched_class
->set_curr_task(rq
);
3606 * We enqueue to tail when the priority of a task is
3607 * increased (user space view).
3609 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3612 check_class_changed(rq
, p
, prev_class
, oldprio
);
3613 task_rq_unlock(rq
, p
, &flags
);
3615 rt_mutex_adjust_pi(p
);
3620 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3621 const struct sched_param
*param
, bool check
)
3623 struct sched_attr attr
= {
3624 .sched_policy
= policy
,
3625 .sched_priority
= param
->sched_priority
,
3626 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3629 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3630 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3631 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3632 policy
&= ~SCHED_RESET_ON_FORK
;
3633 attr
.sched_policy
= policy
;
3636 return __sched_setscheduler(p
, &attr
, check
);
3639 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3640 * @p: the task in question.
3641 * @policy: new policy.
3642 * @param: structure containing the new RT priority.
3644 * Return: 0 on success. An error code otherwise.
3646 * NOTE that the task may be already dead.
3648 int sched_setscheduler(struct task_struct
*p
, int policy
,
3649 const struct sched_param
*param
)
3651 return _sched_setscheduler(p
, policy
, param
, true);
3653 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3655 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3657 return __sched_setscheduler(p
, attr
, true);
3659 EXPORT_SYMBOL_GPL(sched_setattr
);
3662 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3663 * @p: the task in question.
3664 * @policy: new policy.
3665 * @param: structure containing the new RT priority.
3667 * Just like sched_setscheduler, only don't bother checking if the
3668 * current context has permission. For example, this is needed in
3669 * stop_machine(): we create temporary high priority worker threads,
3670 * but our caller might not have that capability.
3672 * Return: 0 on success. An error code otherwise.
3674 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3675 const struct sched_param
*param
)
3677 return _sched_setscheduler(p
, policy
, param
, false);
3681 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3683 struct sched_param lparam
;
3684 struct task_struct
*p
;
3687 if (!param
|| pid
< 0)
3689 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3694 p
= find_process_by_pid(pid
);
3696 retval
= sched_setscheduler(p
, policy
, &lparam
);
3703 * Mimics kernel/events/core.c perf_copy_attr().
3705 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3706 struct sched_attr
*attr
)
3711 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3715 * zero the full structure, so that a short copy will be nice.
3717 memset(attr
, 0, sizeof(*attr
));
3719 ret
= get_user(size
, &uattr
->size
);
3723 if (size
> PAGE_SIZE
) /* silly large */
3726 if (!size
) /* abi compat */
3727 size
= SCHED_ATTR_SIZE_VER0
;
3729 if (size
< SCHED_ATTR_SIZE_VER0
)
3733 * If we're handed a bigger struct than we know of,
3734 * ensure all the unknown bits are 0 - i.e. new
3735 * user-space does not rely on any kernel feature
3736 * extensions we dont know about yet.
3738 if (size
> sizeof(*attr
)) {
3739 unsigned char __user
*addr
;
3740 unsigned char __user
*end
;
3743 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3744 end
= (void __user
*)uattr
+ size
;
3746 for (; addr
< end
; addr
++) {
3747 ret
= get_user(val
, addr
);
3753 size
= sizeof(*attr
);
3756 ret
= copy_from_user(attr
, uattr
, size
);
3761 * XXX: do we want to be lenient like existing syscalls; or do we want
3762 * to be strict and return an error on out-of-bounds values?
3764 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3769 put_user(sizeof(*attr
), &uattr
->size
);
3774 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3775 * @pid: the pid in question.
3776 * @policy: new policy.
3777 * @param: structure containing the new RT priority.
3779 * Return: 0 on success. An error code otherwise.
3781 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3782 struct sched_param __user
*, param
)
3784 /* negative values for policy are not valid */
3788 return do_sched_setscheduler(pid
, policy
, param
);
3792 * sys_sched_setparam - set/change the RT priority of a thread
3793 * @pid: the pid in question.
3794 * @param: structure containing the new RT priority.
3796 * Return: 0 on success. An error code otherwise.
3798 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3800 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
3804 * sys_sched_setattr - same as above, but with extended sched_attr
3805 * @pid: the pid in question.
3806 * @uattr: structure containing the extended parameters.
3807 * @flags: for future extension.
3809 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3810 unsigned int, flags
)
3812 struct sched_attr attr
;
3813 struct task_struct
*p
;
3816 if (!uattr
|| pid
< 0 || flags
)
3819 retval
= sched_copy_attr(uattr
, &attr
);
3823 if ((int)attr
.sched_policy
< 0)
3828 p
= find_process_by_pid(pid
);
3830 retval
= sched_setattr(p
, &attr
);
3837 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3838 * @pid: the pid in question.
3840 * Return: On success, the policy of the thread. Otherwise, a negative error
3843 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3845 struct task_struct
*p
;
3853 p
= find_process_by_pid(pid
);
3855 retval
= security_task_getscheduler(p
);
3858 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3865 * sys_sched_getparam - get the RT priority of a thread
3866 * @pid: the pid in question.
3867 * @param: structure containing the RT priority.
3869 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3872 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3874 struct sched_param lp
= { .sched_priority
= 0 };
3875 struct task_struct
*p
;
3878 if (!param
|| pid
< 0)
3882 p
= find_process_by_pid(pid
);
3887 retval
= security_task_getscheduler(p
);
3891 if (task_has_rt_policy(p
))
3892 lp
.sched_priority
= p
->rt_priority
;
3896 * This one might sleep, we cannot do it with a spinlock held ...
3898 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3907 static int sched_read_attr(struct sched_attr __user
*uattr
,
3908 struct sched_attr
*attr
,
3913 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3917 * If we're handed a smaller struct than we know of,
3918 * ensure all the unknown bits are 0 - i.e. old
3919 * user-space does not get uncomplete information.
3921 if (usize
< sizeof(*attr
)) {
3922 unsigned char *addr
;
3925 addr
= (void *)attr
+ usize
;
3926 end
= (void *)attr
+ sizeof(*attr
);
3928 for (; addr
< end
; addr
++) {
3936 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3944 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3945 * @pid: the pid in question.
3946 * @uattr: structure containing the extended parameters.
3947 * @size: sizeof(attr) for fwd/bwd comp.
3948 * @flags: for future extension.
3950 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3951 unsigned int, size
, unsigned int, flags
)
3953 struct sched_attr attr
= {
3954 .size
= sizeof(struct sched_attr
),
3956 struct task_struct
*p
;
3959 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3960 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3964 p
= find_process_by_pid(pid
);
3969 retval
= security_task_getscheduler(p
);
3973 attr
.sched_policy
= p
->policy
;
3974 if (p
->sched_reset_on_fork
)
3975 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3976 if (task_has_dl_policy(p
))
3977 __getparam_dl(p
, &attr
);
3978 else if (task_has_rt_policy(p
))
3979 attr
.sched_priority
= p
->rt_priority
;
3981 attr
.sched_nice
= task_nice(p
);
3985 retval
= sched_read_attr(uattr
, &attr
, size
);
3993 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3995 cpumask_var_t cpus_allowed
, new_mask
;
3996 struct task_struct
*p
;
4001 p
= find_process_by_pid(pid
);
4007 /* Prevent p going away */
4011 if (p
->flags
& PF_NO_SETAFFINITY
) {
4015 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4019 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4021 goto out_free_cpus_allowed
;
4024 if (!check_same_owner(p
)) {
4026 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4028 goto out_free_new_mask
;
4033 retval
= security_task_setscheduler(p
);
4035 goto out_free_new_mask
;
4038 cpuset_cpus_allowed(p
, cpus_allowed
);
4039 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4042 * Since bandwidth control happens on root_domain basis,
4043 * if admission test is enabled, we only admit -deadline
4044 * tasks allowed to run on all the CPUs in the task's
4048 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4050 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4053 goto out_free_new_mask
;
4059 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4062 cpuset_cpus_allowed(p
, cpus_allowed
);
4063 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4065 * We must have raced with a concurrent cpuset
4066 * update. Just reset the cpus_allowed to the
4067 * cpuset's cpus_allowed
4069 cpumask_copy(new_mask
, cpus_allowed
);
4074 free_cpumask_var(new_mask
);
4075 out_free_cpus_allowed
:
4076 free_cpumask_var(cpus_allowed
);
4082 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4083 struct cpumask
*new_mask
)
4085 if (len
< cpumask_size())
4086 cpumask_clear(new_mask
);
4087 else if (len
> cpumask_size())
4088 len
= cpumask_size();
4090 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4094 * sys_sched_setaffinity - set the cpu affinity of a process
4095 * @pid: pid of the process
4096 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4097 * @user_mask_ptr: user-space pointer to the new cpu mask
4099 * Return: 0 on success. An error code otherwise.
4101 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4102 unsigned long __user
*, user_mask_ptr
)
4104 cpumask_var_t new_mask
;
4107 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4110 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4112 retval
= sched_setaffinity(pid
, new_mask
);
4113 free_cpumask_var(new_mask
);
4117 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4119 struct task_struct
*p
;
4120 unsigned long flags
;
4126 p
= find_process_by_pid(pid
);
4130 retval
= security_task_getscheduler(p
);
4134 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4135 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4136 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4145 * sys_sched_getaffinity - get the cpu affinity of a process
4146 * @pid: pid of the process
4147 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4148 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4150 * Return: 0 on success. An error code otherwise.
4152 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4153 unsigned long __user
*, user_mask_ptr
)
4158 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4160 if (len
& (sizeof(unsigned long)-1))
4163 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4166 ret
= sched_getaffinity(pid
, mask
);
4168 size_t retlen
= min_t(size_t, len
, cpumask_size());
4170 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4175 free_cpumask_var(mask
);
4181 * sys_sched_yield - yield the current processor to other threads.
4183 * This function yields the current CPU to other tasks. If there are no
4184 * other threads running on this CPU then this function will return.
4188 SYSCALL_DEFINE0(sched_yield
)
4190 struct rq
*rq
= this_rq_lock();
4192 schedstat_inc(rq
, yld_count
);
4193 current
->sched_class
->yield_task(rq
);
4196 * Since we are going to call schedule() anyway, there's
4197 * no need to preempt or enable interrupts:
4199 __release(rq
->lock
);
4200 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4201 do_raw_spin_unlock(&rq
->lock
);
4202 sched_preempt_enable_no_resched();
4209 static void __cond_resched(void)
4211 __preempt_count_add(PREEMPT_ACTIVE
);
4213 __preempt_count_sub(PREEMPT_ACTIVE
);
4216 int __sched
_cond_resched(void)
4218 if (should_resched()) {
4224 EXPORT_SYMBOL(_cond_resched
);
4227 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4228 * call schedule, and on return reacquire the lock.
4230 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4231 * operations here to prevent schedule() from being called twice (once via
4232 * spin_unlock(), once by hand).
4234 int __cond_resched_lock(spinlock_t
*lock
)
4236 int resched
= should_resched();
4239 lockdep_assert_held(lock
);
4241 if (spin_needbreak(lock
) || resched
) {
4252 EXPORT_SYMBOL(__cond_resched_lock
);
4254 int __sched
__cond_resched_softirq(void)
4256 BUG_ON(!in_softirq());
4258 if (should_resched()) {
4266 EXPORT_SYMBOL(__cond_resched_softirq
);
4269 * yield - yield the current processor to other threads.
4271 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4273 * The scheduler is at all times free to pick the calling task as the most
4274 * eligible task to run, if removing the yield() call from your code breaks
4275 * it, its already broken.
4277 * Typical broken usage is:
4282 * where one assumes that yield() will let 'the other' process run that will
4283 * make event true. If the current task is a SCHED_FIFO task that will never
4284 * happen. Never use yield() as a progress guarantee!!
4286 * If you want to use yield() to wait for something, use wait_event().
4287 * If you want to use yield() to be 'nice' for others, use cond_resched().
4288 * If you still want to use yield(), do not!
4290 void __sched
yield(void)
4292 set_current_state(TASK_RUNNING
);
4295 EXPORT_SYMBOL(yield
);
4298 * yield_to - yield the current processor to another thread in
4299 * your thread group, or accelerate that thread toward the
4300 * processor it's on.
4302 * @preempt: whether task preemption is allowed or not
4304 * It's the caller's job to ensure that the target task struct
4305 * can't go away on us before we can do any checks.
4308 * true (>0) if we indeed boosted the target task.
4309 * false (0) if we failed to boost the target.
4310 * -ESRCH if there's no task to yield to.
4312 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4314 struct task_struct
*curr
= current
;
4315 struct rq
*rq
, *p_rq
;
4316 unsigned long flags
;
4319 local_irq_save(flags
);
4325 * If we're the only runnable task on the rq and target rq also
4326 * has only one task, there's absolutely no point in yielding.
4328 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4333 double_rq_lock(rq
, p_rq
);
4334 if (task_rq(p
) != p_rq
) {
4335 double_rq_unlock(rq
, p_rq
);
4339 if (!curr
->sched_class
->yield_to_task
)
4342 if (curr
->sched_class
!= p
->sched_class
)
4345 if (task_running(p_rq
, p
) || p
->state
)
4348 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4350 schedstat_inc(rq
, yld_count
);
4352 * Make p's CPU reschedule; pick_next_entity takes care of
4355 if (preempt
&& rq
!= p_rq
)
4360 double_rq_unlock(rq
, p_rq
);
4362 local_irq_restore(flags
);
4369 EXPORT_SYMBOL_GPL(yield_to
);
4372 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4373 * that process accounting knows that this is a task in IO wait state.
4375 void __sched
io_schedule(void)
4377 struct rq
*rq
= raw_rq();
4379 delayacct_blkio_start();
4380 atomic_inc(&rq
->nr_iowait
);
4381 blk_flush_plug(current
);
4382 current
->in_iowait
= 1;
4384 current
->in_iowait
= 0;
4385 atomic_dec(&rq
->nr_iowait
);
4386 delayacct_blkio_end();
4388 EXPORT_SYMBOL(io_schedule
);
4390 long __sched
io_schedule_timeout(long timeout
)
4392 struct rq
*rq
= raw_rq();
4395 delayacct_blkio_start();
4396 atomic_inc(&rq
->nr_iowait
);
4397 blk_flush_plug(current
);
4398 current
->in_iowait
= 1;
4399 ret
= schedule_timeout(timeout
);
4400 current
->in_iowait
= 0;
4401 atomic_dec(&rq
->nr_iowait
);
4402 delayacct_blkio_end();
4407 * sys_sched_get_priority_max - return maximum RT priority.
4408 * @policy: scheduling class.
4410 * Return: On success, this syscall returns the maximum
4411 * rt_priority that can be used by a given scheduling class.
4412 * On failure, a negative error code is returned.
4414 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4421 ret
= MAX_USER_RT_PRIO
-1;
4423 case SCHED_DEADLINE
:
4434 * sys_sched_get_priority_min - return minimum RT priority.
4435 * @policy: scheduling class.
4437 * Return: On success, this syscall returns the minimum
4438 * rt_priority that can be used by a given scheduling class.
4439 * On failure, a negative error code is returned.
4441 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4450 case SCHED_DEADLINE
:
4460 * sys_sched_rr_get_interval - return the default timeslice of a process.
4461 * @pid: pid of the process.
4462 * @interval: userspace pointer to the timeslice value.
4464 * this syscall writes the default timeslice value of a given process
4465 * into the user-space timespec buffer. A value of '0' means infinity.
4467 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4470 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4471 struct timespec __user
*, interval
)
4473 struct task_struct
*p
;
4474 unsigned int time_slice
;
4475 unsigned long flags
;
4485 p
= find_process_by_pid(pid
);
4489 retval
= security_task_getscheduler(p
);
4493 rq
= task_rq_lock(p
, &flags
);
4495 if (p
->sched_class
->get_rr_interval
)
4496 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4497 task_rq_unlock(rq
, p
, &flags
);
4500 jiffies_to_timespec(time_slice
, &t
);
4501 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4509 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4511 void sched_show_task(struct task_struct
*p
)
4513 unsigned long free
= 0;
4517 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4518 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4519 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4520 #if BITS_PER_LONG == 32
4521 if (state
== TASK_RUNNING
)
4522 printk(KERN_CONT
" running ");
4524 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4526 if (state
== TASK_RUNNING
)
4527 printk(KERN_CONT
" running task ");
4529 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4531 #ifdef CONFIG_DEBUG_STACK_USAGE
4532 free
= stack_not_used(p
);
4535 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4537 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4538 task_pid_nr(p
), ppid
,
4539 (unsigned long)task_thread_info(p
)->flags
);
4541 print_worker_info(KERN_INFO
, p
);
4542 show_stack(p
, NULL
);
4545 void show_state_filter(unsigned long state_filter
)
4547 struct task_struct
*g
, *p
;
4549 #if BITS_PER_LONG == 32
4551 " task PC stack pid father\n");
4554 " task PC stack pid father\n");
4557 for_each_process_thread(g
, p
) {
4559 * reset the NMI-timeout, listing all files on a slow
4560 * console might take a lot of time:
4562 touch_nmi_watchdog();
4563 if (!state_filter
|| (p
->state
& state_filter
))
4567 touch_all_softlockup_watchdogs();
4569 #ifdef CONFIG_SCHED_DEBUG
4570 sysrq_sched_debug_show();
4574 * Only show locks if all tasks are dumped:
4577 debug_show_all_locks();
4580 void init_idle_bootup_task(struct task_struct
*idle
)
4582 idle
->sched_class
= &idle_sched_class
;
4586 * init_idle - set up an idle thread for a given CPU
4587 * @idle: task in question
4588 * @cpu: cpu the idle task belongs to
4590 * NOTE: this function does not set the idle thread's NEED_RESCHED
4591 * flag, to make booting more robust.
4593 void init_idle(struct task_struct
*idle
, int cpu
)
4595 struct rq
*rq
= cpu_rq(cpu
);
4596 unsigned long flags
;
4598 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4600 __sched_fork(0, idle
);
4601 idle
->state
= TASK_RUNNING
;
4602 idle
->se
.exec_start
= sched_clock();
4604 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4606 * We're having a chicken and egg problem, even though we are
4607 * holding rq->lock, the cpu isn't yet set to this cpu so the
4608 * lockdep check in task_group() will fail.
4610 * Similar case to sched_fork(). / Alternatively we could
4611 * use task_rq_lock() here and obtain the other rq->lock.
4616 __set_task_cpu(idle
, cpu
);
4619 rq
->curr
= rq
->idle
= idle
;
4620 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4621 #if defined(CONFIG_SMP)
4624 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4626 /* Set the preempt count _outside_ the spinlocks! */
4627 init_idle_preempt_count(idle
, cpu
);
4630 * The idle tasks have their own, simple scheduling class:
4632 idle
->sched_class
= &idle_sched_class
;
4633 ftrace_graph_init_idle_task(idle
, cpu
);
4634 vtime_init_idle(idle
, cpu
);
4635 #if defined(CONFIG_SMP)
4636 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4642 * move_queued_task - move a queued task to new rq.
4644 * Returns (locked) new rq. Old rq's lock is released.
4646 static struct rq
*move_queued_task(struct task_struct
*p
, int new_cpu
)
4648 struct rq
*rq
= task_rq(p
);
4650 lockdep_assert_held(&rq
->lock
);
4652 dequeue_task(rq
, p
, 0);
4653 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
4654 set_task_cpu(p
, new_cpu
);
4655 raw_spin_unlock(&rq
->lock
);
4657 rq
= cpu_rq(new_cpu
);
4659 raw_spin_lock(&rq
->lock
);
4660 BUG_ON(task_cpu(p
) != new_cpu
);
4661 p
->on_rq
= TASK_ON_RQ_QUEUED
;
4662 enqueue_task(rq
, p
, 0);
4663 check_preempt_curr(rq
, p
, 0);
4668 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4670 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4671 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4673 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4674 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4678 * This is how migration works:
4680 * 1) we invoke migration_cpu_stop() on the target CPU using
4682 * 2) stopper starts to run (implicitly forcing the migrated thread
4684 * 3) it checks whether the migrated task is still in the wrong runqueue.
4685 * 4) if it's in the wrong runqueue then the migration thread removes
4686 * it and puts it into the right queue.
4687 * 5) stopper completes and stop_one_cpu() returns and the migration
4692 * Change a given task's CPU affinity. Migrate the thread to a
4693 * proper CPU and schedule it away if the CPU it's executing on
4694 * is removed from the allowed bitmask.
4696 * NOTE: the caller must have a valid reference to the task, the
4697 * task must not exit() & deallocate itself prematurely. The
4698 * call is not atomic; no spinlocks may be held.
4700 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4702 unsigned long flags
;
4704 unsigned int dest_cpu
;
4707 rq
= task_rq_lock(p
, &flags
);
4709 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4712 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4717 do_set_cpus_allowed(p
, new_mask
);
4719 /* Can the task run on the task's current CPU? If so, we're done */
4720 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4723 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4724 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
4725 struct migration_arg arg
= { p
, dest_cpu
};
4726 /* Need help from migration thread: drop lock and wait. */
4727 task_rq_unlock(rq
, p
, &flags
);
4728 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4729 tlb_migrate_finish(p
->mm
);
4731 } else if (task_on_rq_queued(p
))
4732 rq
= move_queued_task(p
, dest_cpu
);
4734 task_rq_unlock(rq
, p
, &flags
);
4738 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4741 * Move (not current) task off this cpu, onto dest cpu. We're doing
4742 * this because either it can't run here any more (set_cpus_allowed()
4743 * away from this CPU, or CPU going down), or because we're
4744 * attempting to rebalance this task on exec (sched_exec).
4746 * So we race with normal scheduler movements, but that's OK, as long
4747 * as the task is no longer on this CPU.
4749 * Returns non-zero if task was successfully migrated.
4751 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4756 if (unlikely(!cpu_active(dest_cpu
)))
4759 rq
= cpu_rq(src_cpu
);
4761 raw_spin_lock(&p
->pi_lock
);
4762 raw_spin_lock(&rq
->lock
);
4763 /* Already moved. */
4764 if (task_cpu(p
) != src_cpu
)
4767 /* Affinity changed (again). */
4768 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4772 * If we're not on a rq, the next wake-up will ensure we're
4775 if (task_on_rq_queued(p
))
4776 rq
= move_queued_task(p
, dest_cpu
);
4780 raw_spin_unlock(&rq
->lock
);
4781 raw_spin_unlock(&p
->pi_lock
);
4785 #ifdef CONFIG_NUMA_BALANCING
4786 /* Migrate current task p to target_cpu */
4787 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4789 struct migration_arg arg
= { p
, target_cpu
};
4790 int curr_cpu
= task_cpu(p
);
4792 if (curr_cpu
== target_cpu
)
4795 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4798 /* TODO: This is not properly updating schedstats */
4800 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4801 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4805 * Requeue a task on a given node and accurately track the number of NUMA
4806 * tasks on the runqueues
4808 void sched_setnuma(struct task_struct
*p
, int nid
)
4811 unsigned long flags
;
4812 bool queued
, running
;
4814 rq
= task_rq_lock(p
, &flags
);
4815 queued
= task_on_rq_queued(p
);
4816 running
= task_current(rq
, p
);
4819 dequeue_task(rq
, p
, 0);
4821 put_prev_task(rq
, p
);
4823 p
->numa_preferred_nid
= nid
;
4826 p
->sched_class
->set_curr_task(rq
);
4828 enqueue_task(rq
, p
, 0);
4829 task_rq_unlock(rq
, p
, &flags
);
4834 * migration_cpu_stop - this will be executed by a highprio stopper thread
4835 * and performs thread migration by bumping thread off CPU then
4836 * 'pushing' onto another runqueue.
4838 static int migration_cpu_stop(void *data
)
4840 struct migration_arg
*arg
= data
;
4843 * The original target cpu might have gone down and we might
4844 * be on another cpu but it doesn't matter.
4846 local_irq_disable();
4848 * We need to explicitly wake pending tasks before running
4849 * __migrate_task() such that we will not miss enforcing cpus_allowed
4850 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4852 sched_ttwu_pending();
4853 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4858 #ifdef CONFIG_HOTPLUG_CPU
4861 * Ensures that the idle task is using init_mm right before its cpu goes
4864 void idle_task_exit(void)
4866 struct mm_struct
*mm
= current
->active_mm
;
4868 BUG_ON(cpu_online(smp_processor_id()));
4870 if (mm
!= &init_mm
) {
4871 switch_mm(mm
, &init_mm
, current
);
4872 finish_arch_post_lock_switch();
4878 * Since this CPU is going 'away' for a while, fold any nr_active delta
4879 * we might have. Assumes we're called after migrate_tasks() so that the
4880 * nr_active count is stable.
4882 * Also see the comment "Global load-average calculations".
4884 static void calc_load_migrate(struct rq
*rq
)
4886 long delta
= calc_load_fold_active(rq
);
4888 atomic_long_add(delta
, &calc_load_tasks
);
4891 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4895 static const struct sched_class fake_sched_class
= {
4896 .put_prev_task
= put_prev_task_fake
,
4899 static struct task_struct fake_task
= {
4901 * Avoid pull_{rt,dl}_task()
4903 .prio
= MAX_PRIO
+ 1,
4904 .sched_class
= &fake_sched_class
,
4908 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4909 * try_to_wake_up()->select_task_rq().
4911 * Called with rq->lock held even though we'er in stop_machine() and
4912 * there's no concurrency possible, we hold the required locks anyway
4913 * because of lock validation efforts.
4915 static void migrate_tasks(unsigned int dead_cpu
)
4917 struct rq
*rq
= cpu_rq(dead_cpu
);
4918 struct task_struct
*next
, *stop
= rq
->stop
;
4922 * Fudge the rq selection such that the below task selection loop
4923 * doesn't get stuck on the currently eligible stop task.
4925 * We're currently inside stop_machine() and the rq is either stuck
4926 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4927 * either way we should never end up calling schedule() until we're
4933 * put_prev_task() and pick_next_task() sched
4934 * class method both need to have an up-to-date
4935 * value of rq->clock[_task]
4937 update_rq_clock(rq
);
4941 * There's this thread running, bail when that's the only
4944 if (rq
->nr_running
== 1)
4947 next
= pick_next_task(rq
, &fake_task
);
4949 next
->sched_class
->put_prev_task(rq
, next
);
4951 /* Find suitable destination for @next, with force if needed. */
4952 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4953 raw_spin_unlock(&rq
->lock
);
4955 __migrate_task(next
, dead_cpu
, dest_cpu
);
4957 raw_spin_lock(&rq
->lock
);
4963 #endif /* CONFIG_HOTPLUG_CPU */
4965 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4967 static struct ctl_table sd_ctl_dir
[] = {
4969 .procname
= "sched_domain",
4975 static struct ctl_table sd_ctl_root
[] = {
4977 .procname
= "kernel",
4979 .child
= sd_ctl_dir
,
4984 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4986 struct ctl_table
*entry
=
4987 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4992 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4994 struct ctl_table
*entry
;
4997 * In the intermediate directories, both the child directory and
4998 * procname are dynamically allocated and could fail but the mode
4999 * will always be set. In the lowest directory the names are
5000 * static strings and all have proc handlers.
5002 for (entry
= *tablep
; entry
->mode
; entry
++) {
5004 sd_free_ctl_entry(&entry
->child
);
5005 if (entry
->proc_handler
== NULL
)
5006 kfree(entry
->procname
);
5013 static int min_load_idx
= 0;
5014 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5017 set_table_entry(struct ctl_table
*entry
,
5018 const char *procname
, void *data
, int maxlen
,
5019 umode_t mode
, proc_handler
*proc_handler
,
5022 entry
->procname
= procname
;
5024 entry
->maxlen
= maxlen
;
5026 entry
->proc_handler
= proc_handler
;
5029 entry
->extra1
= &min_load_idx
;
5030 entry
->extra2
= &max_load_idx
;
5034 static struct ctl_table
*
5035 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5037 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5042 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5043 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5044 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5045 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5046 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5047 sizeof(int), 0644, proc_dointvec_minmax
, true);
5048 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5049 sizeof(int), 0644, proc_dointvec_minmax
, true);
5050 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5051 sizeof(int), 0644, proc_dointvec_minmax
, true);
5052 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5053 sizeof(int), 0644, proc_dointvec_minmax
, true);
5054 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5055 sizeof(int), 0644, proc_dointvec_minmax
, true);
5056 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5057 sizeof(int), 0644, proc_dointvec_minmax
, false);
5058 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5059 sizeof(int), 0644, proc_dointvec_minmax
, false);
5060 set_table_entry(&table
[9], "cache_nice_tries",
5061 &sd
->cache_nice_tries
,
5062 sizeof(int), 0644, proc_dointvec_minmax
, false);
5063 set_table_entry(&table
[10], "flags", &sd
->flags
,
5064 sizeof(int), 0644, proc_dointvec_minmax
, false);
5065 set_table_entry(&table
[11], "max_newidle_lb_cost",
5066 &sd
->max_newidle_lb_cost
,
5067 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5068 set_table_entry(&table
[12], "name", sd
->name
,
5069 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5070 /* &table[13] is terminator */
5075 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5077 struct ctl_table
*entry
, *table
;
5078 struct sched_domain
*sd
;
5079 int domain_num
= 0, i
;
5082 for_each_domain(cpu
, sd
)
5084 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5089 for_each_domain(cpu
, sd
) {
5090 snprintf(buf
, 32, "domain%d", i
);
5091 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5093 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5100 static struct ctl_table_header
*sd_sysctl_header
;
5101 static void register_sched_domain_sysctl(void)
5103 int i
, cpu_num
= num_possible_cpus();
5104 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5107 WARN_ON(sd_ctl_dir
[0].child
);
5108 sd_ctl_dir
[0].child
= entry
;
5113 for_each_possible_cpu(i
) {
5114 snprintf(buf
, 32, "cpu%d", i
);
5115 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5117 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5121 WARN_ON(sd_sysctl_header
);
5122 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5125 /* may be called multiple times per register */
5126 static void unregister_sched_domain_sysctl(void)
5128 if (sd_sysctl_header
)
5129 unregister_sysctl_table(sd_sysctl_header
);
5130 sd_sysctl_header
= NULL
;
5131 if (sd_ctl_dir
[0].child
)
5132 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5135 static void register_sched_domain_sysctl(void)
5138 static void unregister_sched_domain_sysctl(void)
5143 static void set_rq_online(struct rq
*rq
)
5146 const struct sched_class
*class;
5148 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5151 for_each_class(class) {
5152 if (class->rq_online
)
5153 class->rq_online(rq
);
5158 static void set_rq_offline(struct rq
*rq
)
5161 const struct sched_class
*class;
5163 for_each_class(class) {
5164 if (class->rq_offline
)
5165 class->rq_offline(rq
);
5168 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5174 * migration_call - callback that gets triggered when a CPU is added.
5175 * Here we can start up the necessary migration thread for the new CPU.
5178 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5180 int cpu
= (long)hcpu
;
5181 unsigned long flags
;
5182 struct rq
*rq
= cpu_rq(cpu
);
5184 switch (action
& ~CPU_TASKS_FROZEN
) {
5186 case CPU_UP_PREPARE
:
5187 rq
->calc_load_update
= calc_load_update
;
5191 /* Update our root-domain */
5192 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5194 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5198 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5201 #ifdef CONFIG_HOTPLUG_CPU
5203 sched_ttwu_pending();
5204 /* Update our root-domain */
5205 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5207 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5211 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5212 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5216 calc_load_migrate(rq
);
5221 update_max_interval();
5227 * Register at high priority so that task migration (migrate_all_tasks)
5228 * happens before everything else. This has to be lower priority than
5229 * the notifier in the perf_event subsystem, though.
5231 static struct notifier_block migration_notifier
= {
5232 .notifier_call
= migration_call
,
5233 .priority
= CPU_PRI_MIGRATION
,
5236 static void __cpuinit
set_cpu_rq_start_time(void)
5238 int cpu
= smp_processor_id();
5239 struct rq
*rq
= cpu_rq(cpu
);
5240 rq
->age_stamp
= sched_clock_cpu(cpu
);
5243 static int sched_cpu_active(struct notifier_block
*nfb
,
5244 unsigned long action
, void *hcpu
)
5246 switch (action
& ~CPU_TASKS_FROZEN
) {
5248 set_cpu_rq_start_time();
5250 case CPU_DOWN_FAILED
:
5251 set_cpu_active((long)hcpu
, true);
5258 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5259 unsigned long action
, void *hcpu
)
5261 unsigned long flags
;
5262 long cpu
= (long)hcpu
;
5265 switch (action
& ~CPU_TASKS_FROZEN
) {
5266 case CPU_DOWN_PREPARE
:
5267 set_cpu_active(cpu
, false);
5269 /* explicitly allow suspend */
5270 if (!(action
& CPU_TASKS_FROZEN
)) {
5274 rcu_read_lock_sched();
5275 dl_b
= dl_bw_of(cpu
);
5277 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5278 cpus
= dl_bw_cpus(cpu
);
5279 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5280 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5282 rcu_read_unlock_sched();
5285 return notifier_from_errno(-EBUSY
);
5293 static int __init
migration_init(void)
5295 void *cpu
= (void *)(long)smp_processor_id();
5298 /* Initialize migration for the boot CPU */
5299 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5300 BUG_ON(err
== NOTIFY_BAD
);
5301 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5302 register_cpu_notifier(&migration_notifier
);
5304 /* Register cpu active notifiers */
5305 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5306 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5310 early_initcall(migration_init
);
5315 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5317 #ifdef CONFIG_SCHED_DEBUG
5319 static __read_mostly
int sched_debug_enabled
;
5321 static int __init
sched_debug_setup(char *str
)
5323 sched_debug_enabled
= 1;
5327 early_param("sched_debug", sched_debug_setup
);
5329 static inline bool sched_debug(void)
5331 return sched_debug_enabled
;
5334 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5335 struct cpumask
*groupmask
)
5337 struct sched_group
*group
= sd
->groups
;
5340 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5341 cpumask_clear(groupmask
);
5343 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5345 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5346 printk("does not load-balance\n");
5348 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5353 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5355 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5356 printk(KERN_ERR
"ERROR: domain->span does not contain "
5359 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5360 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5364 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5368 printk(KERN_ERR
"ERROR: group is NULL\n");
5373 * Even though we initialize ->capacity to something semi-sane,
5374 * we leave capacity_orig unset. This allows us to detect if
5375 * domain iteration is still funny without causing /0 traps.
5377 if (!group
->sgc
->capacity_orig
) {
5378 printk(KERN_CONT
"\n");
5379 printk(KERN_ERR
"ERROR: domain->cpu_capacity not set\n");
5383 if (!cpumask_weight(sched_group_cpus(group
))) {
5384 printk(KERN_CONT
"\n");
5385 printk(KERN_ERR
"ERROR: empty group\n");
5389 if (!(sd
->flags
& SD_OVERLAP
) &&
5390 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5391 printk(KERN_CONT
"\n");
5392 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5396 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5398 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5400 printk(KERN_CONT
" %s", str
);
5401 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5402 printk(KERN_CONT
" (cpu_capacity = %d)",
5403 group
->sgc
->capacity
);
5406 group
= group
->next
;
5407 } while (group
!= sd
->groups
);
5408 printk(KERN_CONT
"\n");
5410 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5411 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5414 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5415 printk(KERN_ERR
"ERROR: parent span is not a superset "
5416 "of domain->span\n");
5420 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5424 if (!sched_debug_enabled
)
5428 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5432 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5435 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5443 #else /* !CONFIG_SCHED_DEBUG */
5444 # define sched_domain_debug(sd, cpu) do { } while (0)
5445 static inline bool sched_debug(void)
5449 #endif /* CONFIG_SCHED_DEBUG */
5451 static int sd_degenerate(struct sched_domain
*sd
)
5453 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5456 /* Following flags need at least 2 groups */
5457 if (sd
->flags
& (SD_LOAD_BALANCE
|
5458 SD_BALANCE_NEWIDLE
|
5461 SD_SHARE_CPUCAPACITY
|
5462 SD_SHARE_PKG_RESOURCES
|
5463 SD_SHARE_POWERDOMAIN
)) {
5464 if (sd
->groups
!= sd
->groups
->next
)
5468 /* Following flags don't use groups */
5469 if (sd
->flags
& (SD_WAKE_AFFINE
))
5476 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5478 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5480 if (sd_degenerate(parent
))
5483 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5486 /* Flags needing groups don't count if only 1 group in parent */
5487 if (parent
->groups
== parent
->groups
->next
) {
5488 pflags
&= ~(SD_LOAD_BALANCE
|
5489 SD_BALANCE_NEWIDLE
|
5492 SD_SHARE_CPUCAPACITY
|
5493 SD_SHARE_PKG_RESOURCES
|
5495 SD_SHARE_POWERDOMAIN
);
5496 if (nr_node_ids
== 1)
5497 pflags
&= ~SD_SERIALIZE
;
5499 if (~cflags
& pflags
)
5505 static void free_rootdomain(struct rcu_head
*rcu
)
5507 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5509 cpupri_cleanup(&rd
->cpupri
);
5510 cpudl_cleanup(&rd
->cpudl
);
5511 free_cpumask_var(rd
->dlo_mask
);
5512 free_cpumask_var(rd
->rto_mask
);
5513 free_cpumask_var(rd
->online
);
5514 free_cpumask_var(rd
->span
);
5518 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5520 struct root_domain
*old_rd
= NULL
;
5521 unsigned long flags
;
5523 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5528 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5531 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5534 * If we dont want to free the old_rd yet then
5535 * set old_rd to NULL to skip the freeing later
5538 if (!atomic_dec_and_test(&old_rd
->refcount
))
5542 atomic_inc(&rd
->refcount
);
5545 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5546 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5549 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5552 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5555 static int init_rootdomain(struct root_domain
*rd
)
5557 memset(rd
, 0, sizeof(*rd
));
5559 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5561 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5563 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5565 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5568 init_dl_bw(&rd
->dl_bw
);
5569 if (cpudl_init(&rd
->cpudl
) != 0)
5572 if (cpupri_init(&rd
->cpupri
) != 0)
5577 free_cpumask_var(rd
->rto_mask
);
5579 free_cpumask_var(rd
->dlo_mask
);
5581 free_cpumask_var(rd
->online
);
5583 free_cpumask_var(rd
->span
);
5589 * By default the system creates a single root-domain with all cpus as
5590 * members (mimicking the global state we have today).
5592 struct root_domain def_root_domain
;
5594 static void init_defrootdomain(void)
5596 init_rootdomain(&def_root_domain
);
5598 atomic_set(&def_root_domain
.refcount
, 1);
5601 static struct root_domain
*alloc_rootdomain(void)
5603 struct root_domain
*rd
;
5605 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5609 if (init_rootdomain(rd
) != 0) {
5617 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5619 struct sched_group
*tmp
, *first
;
5628 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5633 } while (sg
!= first
);
5636 static void free_sched_domain(struct rcu_head
*rcu
)
5638 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5641 * If its an overlapping domain it has private groups, iterate and
5644 if (sd
->flags
& SD_OVERLAP
) {
5645 free_sched_groups(sd
->groups
, 1);
5646 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5647 kfree(sd
->groups
->sgc
);
5653 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5655 call_rcu(&sd
->rcu
, free_sched_domain
);
5658 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5660 for (; sd
; sd
= sd
->parent
)
5661 destroy_sched_domain(sd
, cpu
);
5665 * Keep a special pointer to the highest sched_domain that has
5666 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5667 * allows us to avoid some pointer chasing select_idle_sibling().
5669 * Also keep a unique ID per domain (we use the first cpu number in
5670 * the cpumask of the domain), this allows us to quickly tell if
5671 * two cpus are in the same cache domain, see cpus_share_cache().
5673 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5674 DEFINE_PER_CPU(int, sd_llc_size
);
5675 DEFINE_PER_CPU(int, sd_llc_id
);
5676 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5677 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5678 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5680 static void update_top_cache_domain(int cpu
)
5682 struct sched_domain
*sd
;
5683 struct sched_domain
*busy_sd
= NULL
;
5687 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5689 id
= cpumask_first(sched_domain_span(sd
));
5690 size
= cpumask_weight(sched_domain_span(sd
));
5691 busy_sd
= sd
->parent
; /* sd_busy */
5693 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5695 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5696 per_cpu(sd_llc_size
, cpu
) = size
;
5697 per_cpu(sd_llc_id
, cpu
) = id
;
5699 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5700 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5702 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5703 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5707 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5708 * hold the hotplug lock.
5711 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5713 struct rq
*rq
= cpu_rq(cpu
);
5714 struct sched_domain
*tmp
;
5716 /* Remove the sched domains which do not contribute to scheduling. */
5717 for (tmp
= sd
; tmp
; ) {
5718 struct sched_domain
*parent
= tmp
->parent
;
5722 if (sd_parent_degenerate(tmp
, parent
)) {
5723 tmp
->parent
= parent
->parent
;
5725 parent
->parent
->child
= tmp
;
5727 * Transfer SD_PREFER_SIBLING down in case of a
5728 * degenerate parent; the spans match for this
5729 * so the property transfers.
5731 if (parent
->flags
& SD_PREFER_SIBLING
)
5732 tmp
->flags
|= SD_PREFER_SIBLING
;
5733 destroy_sched_domain(parent
, cpu
);
5738 if (sd
&& sd_degenerate(sd
)) {
5741 destroy_sched_domain(tmp
, cpu
);
5746 sched_domain_debug(sd
, cpu
);
5748 rq_attach_root(rq
, rd
);
5750 rcu_assign_pointer(rq
->sd
, sd
);
5751 destroy_sched_domains(tmp
, cpu
);
5753 update_top_cache_domain(cpu
);
5756 /* cpus with isolated domains */
5757 static cpumask_var_t cpu_isolated_map
;
5759 /* Setup the mask of cpus configured for isolated domains */
5760 static int __init
isolated_cpu_setup(char *str
)
5762 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5763 cpulist_parse(str
, cpu_isolated_map
);
5767 __setup("isolcpus=", isolated_cpu_setup
);
5770 struct sched_domain
** __percpu sd
;
5771 struct root_domain
*rd
;
5782 * Build an iteration mask that can exclude certain CPUs from the upwards
5785 * Asymmetric node setups can result in situations where the domain tree is of
5786 * unequal depth, make sure to skip domains that already cover the entire
5789 * In that case build_sched_domains() will have terminated the iteration early
5790 * and our sibling sd spans will be empty. Domains should always include the
5791 * cpu they're built on, so check that.
5794 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5796 const struct cpumask
*span
= sched_domain_span(sd
);
5797 struct sd_data
*sdd
= sd
->private;
5798 struct sched_domain
*sibling
;
5801 for_each_cpu(i
, span
) {
5802 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5803 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5806 cpumask_set_cpu(i
, sched_group_mask(sg
));
5811 * Return the canonical balance cpu for this group, this is the first cpu
5812 * of this group that's also in the iteration mask.
5814 int group_balance_cpu(struct sched_group
*sg
)
5816 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5820 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5822 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5823 const struct cpumask
*span
= sched_domain_span(sd
);
5824 struct cpumask
*covered
= sched_domains_tmpmask
;
5825 struct sd_data
*sdd
= sd
->private;
5826 struct sched_domain
*sibling
;
5829 cpumask_clear(covered
);
5831 for_each_cpu(i
, span
) {
5832 struct cpumask
*sg_span
;
5834 if (cpumask_test_cpu(i
, covered
))
5837 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5839 /* See the comment near build_group_mask(). */
5840 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5843 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5844 GFP_KERNEL
, cpu_to_node(cpu
));
5849 sg_span
= sched_group_cpus(sg
);
5851 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
5853 cpumask_set_cpu(i
, sg_span
);
5855 cpumask_or(covered
, covered
, sg_span
);
5857 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
5858 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
5859 build_group_mask(sd
, sg
);
5862 * Initialize sgc->capacity such that even if we mess up the
5863 * domains and no possible iteration will get us here, we won't
5866 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
5867 sg
->sgc
->capacity_orig
= sg
->sgc
->capacity
;
5870 * Make sure the first group of this domain contains the
5871 * canonical balance cpu. Otherwise the sched_domain iteration
5872 * breaks. See update_sg_lb_stats().
5874 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5875 group_balance_cpu(sg
) == cpu
)
5885 sd
->groups
= groups
;
5890 free_sched_groups(first
, 0);
5895 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5897 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5898 struct sched_domain
*child
= sd
->child
;
5901 cpu
= cpumask_first(sched_domain_span(child
));
5904 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5905 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
5906 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
5913 * build_sched_groups will build a circular linked list of the groups
5914 * covered by the given span, and will set each group's ->cpumask correctly,
5915 * and ->cpu_capacity to 0.
5917 * Assumes the sched_domain tree is fully constructed
5920 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5922 struct sched_group
*first
= NULL
, *last
= NULL
;
5923 struct sd_data
*sdd
= sd
->private;
5924 const struct cpumask
*span
= sched_domain_span(sd
);
5925 struct cpumask
*covered
;
5928 get_group(cpu
, sdd
, &sd
->groups
);
5929 atomic_inc(&sd
->groups
->ref
);
5931 if (cpu
!= cpumask_first(span
))
5934 lockdep_assert_held(&sched_domains_mutex
);
5935 covered
= sched_domains_tmpmask
;
5937 cpumask_clear(covered
);
5939 for_each_cpu(i
, span
) {
5940 struct sched_group
*sg
;
5943 if (cpumask_test_cpu(i
, covered
))
5946 group
= get_group(i
, sdd
, &sg
);
5947 cpumask_setall(sched_group_mask(sg
));
5949 for_each_cpu(j
, span
) {
5950 if (get_group(j
, sdd
, NULL
) != group
)
5953 cpumask_set_cpu(j
, covered
);
5954 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5969 * Initialize sched groups cpu_capacity.
5971 * cpu_capacity indicates the capacity of sched group, which is used while
5972 * distributing the load between different sched groups in a sched domain.
5973 * Typically cpu_capacity for all the groups in a sched domain will be same
5974 * unless there are asymmetries in the topology. If there are asymmetries,
5975 * group having more cpu_capacity will pickup more load compared to the
5976 * group having less cpu_capacity.
5978 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
5980 struct sched_group
*sg
= sd
->groups
;
5985 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5987 } while (sg
!= sd
->groups
);
5989 if (cpu
!= group_balance_cpu(sg
))
5992 update_group_capacity(sd
, cpu
);
5993 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
5997 * Initializers for schedule domains
5998 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6001 static int default_relax_domain_level
= -1;
6002 int sched_domain_level_max
;
6004 static int __init
setup_relax_domain_level(char *str
)
6006 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6007 pr_warn("Unable to set relax_domain_level\n");
6011 __setup("relax_domain_level=", setup_relax_domain_level
);
6013 static void set_domain_attribute(struct sched_domain
*sd
,
6014 struct sched_domain_attr
*attr
)
6018 if (!attr
|| attr
->relax_domain_level
< 0) {
6019 if (default_relax_domain_level
< 0)
6022 request
= default_relax_domain_level
;
6024 request
= attr
->relax_domain_level
;
6025 if (request
< sd
->level
) {
6026 /* turn off idle balance on this domain */
6027 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6029 /* turn on idle balance on this domain */
6030 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6034 static void __sdt_free(const struct cpumask
*cpu_map
);
6035 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6037 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6038 const struct cpumask
*cpu_map
)
6042 if (!atomic_read(&d
->rd
->refcount
))
6043 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6045 free_percpu(d
->sd
); /* fall through */
6047 __sdt_free(cpu_map
); /* fall through */
6053 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6054 const struct cpumask
*cpu_map
)
6056 memset(d
, 0, sizeof(*d
));
6058 if (__sdt_alloc(cpu_map
))
6059 return sa_sd_storage
;
6060 d
->sd
= alloc_percpu(struct sched_domain
*);
6062 return sa_sd_storage
;
6063 d
->rd
= alloc_rootdomain();
6066 return sa_rootdomain
;
6070 * NULL the sd_data elements we've used to build the sched_domain and
6071 * sched_group structure so that the subsequent __free_domain_allocs()
6072 * will not free the data we're using.
6074 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6076 struct sd_data
*sdd
= sd
->private;
6078 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6079 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6081 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6082 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6084 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6085 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6089 static int sched_domains_numa_levels
;
6090 static int *sched_domains_numa_distance
;
6091 static struct cpumask
***sched_domains_numa_masks
;
6092 static int sched_domains_curr_level
;
6096 * SD_flags allowed in topology descriptions.
6098 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6099 * SD_SHARE_PKG_RESOURCES - describes shared caches
6100 * SD_NUMA - describes NUMA topologies
6101 * SD_SHARE_POWERDOMAIN - describes shared power domain
6104 * SD_ASYM_PACKING - describes SMT quirks
6106 #define TOPOLOGY_SD_FLAGS \
6107 (SD_SHARE_CPUCAPACITY | \
6108 SD_SHARE_PKG_RESOURCES | \
6111 SD_SHARE_POWERDOMAIN)
6113 static struct sched_domain
*
6114 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6116 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6117 int sd_weight
, sd_flags
= 0;
6121 * Ugly hack to pass state to sd_numa_mask()...
6123 sched_domains_curr_level
= tl
->numa_level
;
6126 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6129 sd_flags
= (*tl
->sd_flags
)();
6130 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6131 "wrong sd_flags in topology description\n"))
6132 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6134 *sd
= (struct sched_domain
){
6135 .min_interval
= sd_weight
,
6136 .max_interval
= 2*sd_weight
,
6138 .imbalance_pct
= 125,
6140 .cache_nice_tries
= 0,
6147 .flags
= 1*SD_LOAD_BALANCE
6148 | 1*SD_BALANCE_NEWIDLE
6153 | 0*SD_SHARE_CPUCAPACITY
6154 | 0*SD_SHARE_PKG_RESOURCES
6156 | 0*SD_PREFER_SIBLING
6161 .last_balance
= jiffies
,
6162 .balance_interval
= sd_weight
,
6164 .max_newidle_lb_cost
= 0,
6165 .next_decay_max_lb_cost
= jiffies
,
6166 #ifdef CONFIG_SCHED_DEBUG
6172 * Convert topological properties into behaviour.
6175 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6176 sd
->imbalance_pct
= 110;
6177 sd
->smt_gain
= 1178; /* ~15% */
6179 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6180 sd
->imbalance_pct
= 117;
6181 sd
->cache_nice_tries
= 1;
6185 } else if (sd
->flags
& SD_NUMA
) {
6186 sd
->cache_nice_tries
= 2;
6190 sd
->flags
|= SD_SERIALIZE
;
6191 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6192 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6199 sd
->flags
|= SD_PREFER_SIBLING
;
6200 sd
->cache_nice_tries
= 1;
6205 sd
->private = &tl
->data
;
6211 * Topology list, bottom-up.
6213 static struct sched_domain_topology_level default_topology
[] = {
6214 #ifdef CONFIG_SCHED_SMT
6215 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6217 #ifdef CONFIG_SCHED_MC
6218 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6220 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6224 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6226 #define for_each_sd_topology(tl) \
6227 for (tl = sched_domain_topology; tl->mask; tl++)
6229 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6231 sched_domain_topology
= tl
;
6236 static const struct cpumask
*sd_numa_mask(int cpu
)
6238 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6241 static void sched_numa_warn(const char *str
)
6243 static int done
= false;
6251 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6253 for (i
= 0; i
< nr_node_ids
; i
++) {
6254 printk(KERN_WARNING
" ");
6255 for (j
= 0; j
< nr_node_ids
; j
++)
6256 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6257 printk(KERN_CONT
"\n");
6259 printk(KERN_WARNING
"\n");
6262 static bool find_numa_distance(int distance
)
6266 if (distance
== node_distance(0, 0))
6269 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6270 if (sched_domains_numa_distance
[i
] == distance
)
6277 static void sched_init_numa(void)
6279 int next_distance
, curr_distance
= node_distance(0, 0);
6280 struct sched_domain_topology_level
*tl
;
6284 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6285 if (!sched_domains_numa_distance
)
6289 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6290 * unique distances in the node_distance() table.
6292 * Assumes node_distance(0,j) includes all distances in
6293 * node_distance(i,j) in order to avoid cubic time.
6295 next_distance
= curr_distance
;
6296 for (i
= 0; i
< nr_node_ids
; i
++) {
6297 for (j
= 0; j
< nr_node_ids
; j
++) {
6298 for (k
= 0; k
< nr_node_ids
; k
++) {
6299 int distance
= node_distance(i
, k
);
6301 if (distance
> curr_distance
&&
6302 (distance
< next_distance
||
6303 next_distance
== curr_distance
))
6304 next_distance
= distance
;
6307 * While not a strong assumption it would be nice to know
6308 * about cases where if node A is connected to B, B is not
6309 * equally connected to A.
6311 if (sched_debug() && node_distance(k
, i
) != distance
)
6312 sched_numa_warn("Node-distance not symmetric");
6314 if (sched_debug() && i
&& !find_numa_distance(distance
))
6315 sched_numa_warn("Node-0 not representative");
6317 if (next_distance
!= curr_distance
) {
6318 sched_domains_numa_distance
[level
++] = next_distance
;
6319 sched_domains_numa_levels
= level
;
6320 curr_distance
= next_distance
;
6325 * In case of sched_debug() we verify the above assumption.
6331 * 'level' contains the number of unique distances, excluding the
6332 * identity distance node_distance(i,i).
6334 * The sched_domains_numa_distance[] array includes the actual distance
6339 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6340 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6341 * the array will contain less then 'level' members. This could be
6342 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6343 * in other functions.
6345 * We reset it to 'level' at the end of this function.
6347 sched_domains_numa_levels
= 0;
6349 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6350 if (!sched_domains_numa_masks
)
6354 * Now for each level, construct a mask per node which contains all
6355 * cpus of nodes that are that many hops away from us.
6357 for (i
= 0; i
< level
; i
++) {
6358 sched_domains_numa_masks
[i
] =
6359 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6360 if (!sched_domains_numa_masks
[i
])
6363 for (j
= 0; j
< nr_node_ids
; j
++) {
6364 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6368 sched_domains_numa_masks
[i
][j
] = mask
;
6370 for (k
= 0; k
< nr_node_ids
; k
++) {
6371 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6374 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6379 /* Compute default topology size */
6380 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6382 tl
= kzalloc((i
+ level
+ 1) *
6383 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6388 * Copy the default topology bits..
6390 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6391 tl
[i
] = sched_domain_topology
[i
];
6394 * .. and append 'j' levels of NUMA goodness.
6396 for (j
= 0; j
< level
; i
++, j
++) {
6397 tl
[i
] = (struct sched_domain_topology_level
){
6398 .mask
= sd_numa_mask
,
6399 .sd_flags
= cpu_numa_flags
,
6400 .flags
= SDTL_OVERLAP
,
6406 sched_domain_topology
= tl
;
6408 sched_domains_numa_levels
= level
;
6411 static void sched_domains_numa_masks_set(int cpu
)
6414 int node
= cpu_to_node(cpu
);
6416 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6417 for (j
= 0; j
< nr_node_ids
; j
++) {
6418 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6419 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6424 static void sched_domains_numa_masks_clear(int cpu
)
6427 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6428 for (j
= 0; j
< nr_node_ids
; j
++)
6429 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6434 * Update sched_domains_numa_masks[level][node] array when new cpus
6437 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6438 unsigned long action
,
6441 int cpu
= (long)hcpu
;
6443 switch (action
& ~CPU_TASKS_FROZEN
) {
6445 sched_domains_numa_masks_set(cpu
);
6449 sched_domains_numa_masks_clear(cpu
);
6459 static inline void sched_init_numa(void)
6463 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6464 unsigned long action
,
6469 #endif /* CONFIG_NUMA */
6471 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6473 struct sched_domain_topology_level
*tl
;
6476 for_each_sd_topology(tl
) {
6477 struct sd_data
*sdd
= &tl
->data
;
6479 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6483 sdd
->sg
= alloc_percpu(struct sched_group
*);
6487 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6491 for_each_cpu(j
, cpu_map
) {
6492 struct sched_domain
*sd
;
6493 struct sched_group
*sg
;
6494 struct sched_group_capacity
*sgc
;
6496 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6497 GFP_KERNEL
, cpu_to_node(j
));
6501 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6503 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6504 GFP_KERNEL
, cpu_to_node(j
));
6510 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6512 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6513 GFP_KERNEL
, cpu_to_node(j
));
6517 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6524 static void __sdt_free(const struct cpumask
*cpu_map
)
6526 struct sched_domain_topology_level
*tl
;
6529 for_each_sd_topology(tl
) {
6530 struct sd_data
*sdd
= &tl
->data
;
6532 for_each_cpu(j
, cpu_map
) {
6533 struct sched_domain
*sd
;
6536 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6537 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6538 free_sched_groups(sd
->groups
, 0);
6539 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6543 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6545 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6547 free_percpu(sdd
->sd
);
6549 free_percpu(sdd
->sg
);
6551 free_percpu(sdd
->sgc
);
6556 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6557 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6558 struct sched_domain
*child
, int cpu
)
6560 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6564 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6566 sd
->level
= child
->level
+ 1;
6567 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6571 if (!cpumask_subset(sched_domain_span(child
),
6572 sched_domain_span(sd
))) {
6573 pr_err("BUG: arch topology borken\n");
6574 #ifdef CONFIG_SCHED_DEBUG
6575 pr_err(" the %s domain not a subset of the %s domain\n",
6576 child
->name
, sd
->name
);
6578 /* Fixup, ensure @sd has at least @child cpus. */
6579 cpumask_or(sched_domain_span(sd
),
6580 sched_domain_span(sd
),
6581 sched_domain_span(child
));
6585 set_domain_attribute(sd
, attr
);
6591 * Build sched domains for a given set of cpus and attach the sched domains
6592 * to the individual cpus
6594 static int build_sched_domains(const struct cpumask
*cpu_map
,
6595 struct sched_domain_attr
*attr
)
6597 enum s_alloc alloc_state
;
6598 struct sched_domain
*sd
;
6600 int i
, ret
= -ENOMEM
;
6602 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6603 if (alloc_state
!= sa_rootdomain
)
6606 /* Set up domains for cpus specified by the cpu_map. */
6607 for_each_cpu(i
, cpu_map
) {
6608 struct sched_domain_topology_level
*tl
;
6611 for_each_sd_topology(tl
) {
6612 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6613 if (tl
== sched_domain_topology
)
6614 *per_cpu_ptr(d
.sd
, i
) = sd
;
6615 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6616 sd
->flags
|= SD_OVERLAP
;
6617 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6622 /* Build the groups for the domains */
6623 for_each_cpu(i
, cpu_map
) {
6624 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6625 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6626 if (sd
->flags
& SD_OVERLAP
) {
6627 if (build_overlap_sched_groups(sd
, i
))
6630 if (build_sched_groups(sd
, i
))
6636 /* Calculate CPU capacity for physical packages and nodes */
6637 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6638 if (!cpumask_test_cpu(i
, cpu_map
))
6641 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6642 claim_allocations(i
, sd
);
6643 init_sched_groups_capacity(i
, sd
);
6647 /* Attach the domains */
6649 for_each_cpu(i
, cpu_map
) {
6650 sd
= *per_cpu_ptr(d
.sd
, i
);
6651 cpu_attach_domain(sd
, d
.rd
, i
);
6657 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6661 static cpumask_var_t
*doms_cur
; /* current sched domains */
6662 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6663 static struct sched_domain_attr
*dattr_cur
;
6664 /* attribues of custom domains in 'doms_cur' */
6667 * Special case: If a kmalloc of a doms_cur partition (array of
6668 * cpumask) fails, then fallback to a single sched domain,
6669 * as determined by the single cpumask fallback_doms.
6671 static cpumask_var_t fallback_doms
;
6674 * arch_update_cpu_topology lets virtualized architectures update the
6675 * cpu core maps. It is supposed to return 1 if the topology changed
6676 * or 0 if it stayed the same.
6678 int __weak
arch_update_cpu_topology(void)
6683 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6686 cpumask_var_t
*doms
;
6688 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6691 for (i
= 0; i
< ndoms
; i
++) {
6692 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6693 free_sched_domains(doms
, i
);
6700 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6703 for (i
= 0; i
< ndoms
; i
++)
6704 free_cpumask_var(doms
[i
]);
6709 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6710 * For now this just excludes isolated cpus, but could be used to
6711 * exclude other special cases in the future.
6713 static int init_sched_domains(const struct cpumask
*cpu_map
)
6717 arch_update_cpu_topology();
6719 doms_cur
= alloc_sched_domains(ndoms_cur
);
6721 doms_cur
= &fallback_doms
;
6722 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6723 err
= build_sched_domains(doms_cur
[0], NULL
);
6724 register_sched_domain_sysctl();
6730 * Detach sched domains from a group of cpus specified in cpu_map
6731 * These cpus will now be attached to the NULL domain
6733 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6738 for_each_cpu(i
, cpu_map
)
6739 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6743 /* handle null as "default" */
6744 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6745 struct sched_domain_attr
*new, int idx_new
)
6747 struct sched_domain_attr tmp
;
6754 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6755 new ? (new + idx_new
) : &tmp
,
6756 sizeof(struct sched_domain_attr
));
6760 * Partition sched domains as specified by the 'ndoms_new'
6761 * cpumasks in the array doms_new[] of cpumasks. This compares
6762 * doms_new[] to the current sched domain partitioning, doms_cur[].
6763 * It destroys each deleted domain and builds each new domain.
6765 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6766 * The masks don't intersect (don't overlap.) We should setup one
6767 * sched domain for each mask. CPUs not in any of the cpumasks will
6768 * not be load balanced. If the same cpumask appears both in the
6769 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6772 * The passed in 'doms_new' should be allocated using
6773 * alloc_sched_domains. This routine takes ownership of it and will
6774 * free_sched_domains it when done with it. If the caller failed the
6775 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6776 * and partition_sched_domains() will fallback to the single partition
6777 * 'fallback_doms', it also forces the domains to be rebuilt.
6779 * If doms_new == NULL it will be replaced with cpu_online_mask.
6780 * ndoms_new == 0 is a special case for destroying existing domains,
6781 * and it will not create the default domain.
6783 * Call with hotplug lock held
6785 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6786 struct sched_domain_attr
*dattr_new
)
6791 mutex_lock(&sched_domains_mutex
);
6793 /* always unregister in case we don't destroy any domains */
6794 unregister_sched_domain_sysctl();
6796 /* Let architecture update cpu core mappings. */
6797 new_topology
= arch_update_cpu_topology();
6799 n
= doms_new
? ndoms_new
: 0;
6801 /* Destroy deleted domains */
6802 for (i
= 0; i
< ndoms_cur
; i
++) {
6803 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6804 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6805 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6808 /* no match - a current sched domain not in new doms_new[] */
6809 detach_destroy_domains(doms_cur
[i
]);
6815 if (doms_new
== NULL
) {
6817 doms_new
= &fallback_doms
;
6818 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6819 WARN_ON_ONCE(dattr_new
);
6822 /* Build new domains */
6823 for (i
= 0; i
< ndoms_new
; i
++) {
6824 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6825 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6826 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6829 /* no match - add a new doms_new */
6830 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6835 /* Remember the new sched domains */
6836 if (doms_cur
!= &fallback_doms
)
6837 free_sched_domains(doms_cur
, ndoms_cur
);
6838 kfree(dattr_cur
); /* kfree(NULL) is safe */
6839 doms_cur
= doms_new
;
6840 dattr_cur
= dattr_new
;
6841 ndoms_cur
= ndoms_new
;
6843 register_sched_domain_sysctl();
6845 mutex_unlock(&sched_domains_mutex
);
6848 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6851 * Update cpusets according to cpu_active mask. If cpusets are
6852 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6853 * around partition_sched_domains().
6855 * If we come here as part of a suspend/resume, don't touch cpusets because we
6856 * want to restore it back to its original state upon resume anyway.
6858 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6862 case CPU_ONLINE_FROZEN
:
6863 case CPU_DOWN_FAILED_FROZEN
:
6866 * num_cpus_frozen tracks how many CPUs are involved in suspend
6867 * resume sequence. As long as this is not the last online
6868 * operation in the resume sequence, just build a single sched
6869 * domain, ignoring cpusets.
6872 if (likely(num_cpus_frozen
)) {
6873 partition_sched_domains(1, NULL
, NULL
);
6878 * This is the last CPU online operation. So fall through and
6879 * restore the original sched domains by considering the
6880 * cpuset configurations.
6884 case CPU_DOWN_FAILED
:
6885 cpuset_update_active_cpus(true);
6893 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6897 case CPU_DOWN_PREPARE
:
6898 cpuset_update_active_cpus(false);
6900 case CPU_DOWN_PREPARE_FROZEN
:
6902 partition_sched_domains(1, NULL
, NULL
);
6910 void __init
sched_init_smp(void)
6912 cpumask_var_t non_isolated_cpus
;
6914 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6915 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6920 * There's no userspace yet to cause hotplug operations; hence all the
6921 * cpu masks are stable and all blatant races in the below code cannot
6924 mutex_lock(&sched_domains_mutex
);
6925 init_sched_domains(cpu_active_mask
);
6926 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6927 if (cpumask_empty(non_isolated_cpus
))
6928 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6929 mutex_unlock(&sched_domains_mutex
);
6931 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6932 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6933 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6937 /* Move init over to a non-isolated CPU */
6938 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6940 sched_init_granularity();
6941 free_cpumask_var(non_isolated_cpus
);
6943 init_sched_rt_class();
6944 init_sched_dl_class();
6947 void __init
sched_init_smp(void)
6949 sched_init_granularity();
6951 #endif /* CONFIG_SMP */
6953 const_debug
unsigned int sysctl_timer_migration
= 1;
6955 int in_sched_functions(unsigned long addr
)
6957 return in_lock_functions(addr
) ||
6958 (addr
>= (unsigned long)__sched_text_start
6959 && addr
< (unsigned long)__sched_text_end
);
6962 #ifdef CONFIG_CGROUP_SCHED
6964 * Default task group.
6965 * Every task in system belongs to this group at bootup.
6967 struct task_group root_task_group
;
6968 LIST_HEAD(task_groups
);
6971 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6973 void __init
sched_init(void)
6976 unsigned long alloc_size
= 0, ptr
;
6978 #ifdef CONFIG_FAIR_GROUP_SCHED
6979 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6981 #ifdef CONFIG_RT_GROUP_SCHED
6982 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6984 #ifdef CONFIG_CPUMASK_OFFSTACK
6985 alloc_size
+= num_possible_cpus() * cpumask_size();
6988 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6990 #ifdef CONFIG_FAIR_GROUP_SCHED
6991 root_task_group
.se
= (struct sched_entity
**)ptr
;
6992 ptr
+= nr_cpu_ids
* sizeof(void **);
6994 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6995 ptr
+= nr_cpu_ids
* sizeof(void **);
6997 #endif /* CONFIG_FAIR_GROUP_SCHED */
6998 #ifdef CONFIG_RT_GROUP_SCHED
6999 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7000 ptr
+= nr_cpu_ids
* sizeof(void **);
7002 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7003 ptr
+= nr_cpu_ids
* sizeof(void **);
7005 #endif /* CONFIG_RT_GROUP_SCHED */
7006 #ifdef CONFIG_CPUMASK_OFFSTACK
7007 for_each_possible_cpu(i
) {
7008 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
7009 ptr
+= cpumask_size();
7011 #endif /* CONFIG_CPUMASK_OFFSTACK */
7014 init_rt_bandwidth(&def_rt_bandwidth
,
7015 global_rt_period(), global_rt_runtime());
7016 init_dl_bandwidth(&def_dl_bandwidth
,
7017 global_rt_period(), global_rt_runtime());
7020 init_defrootdomain();
7023 #ifdef CONFIG_RT_GROUP_SCHED
7024 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7025 global_rt_period(), global_rt_runtime());
7026 #endif /* CONFIG_RT_GROUP_SCHED */
7028 #ifdef CONFIG_CGROUP_SCHED
7029 list_add(&root_task_group
.list
, &task_groups
);
7030 INIT_LIST_HEAD(&root_task_group
.children
);
7031 INIT_LIST_HEAD(&root_task_group
.siblings
);
7032 autogroup_init(&init_task
);
7034 #endif /* CONFIG_CGROUP_SCHED */
7036 for_each_possible_cpu(i
) {
7040 raw_spin_lock_init(&rq
->lock
);
7042 rq
->calc_load_active
= 0;
7043 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7044 init_cfs_rq(&rq
->cfs
);
7045 init_rt_rq(&rq
->rt
, rq
);
7046 init_dl_rq(&rq
->dl
, rq
);
7047 #ifdef CONFIG_FAIR_GROUP_SCHED
7048 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7049 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7051 * How much cpu bandwidth does root_task_group get?
7053 * In case of task-groups formed thr' the cgroup filesystem, it
7054 * gets 100% of the cpu resources in the system. This overall
7055 * system cpu resource is divided among the tasks of
7056 * root_task_group and its child task-groups in a fair manner,
7057 * based on each entity's (task or task-group's) weight
7058 * (se->load.weight).
7060 * In other words, if root_task_group has 10 tasks of weight
7061 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7062 * then A0's share of the cpu resource is:
7064 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7066 * We achieve this by letting root_task_group's tasks sit
7067 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7069 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7070 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7071 #endif /* CONFIG_FAIR_GROUP_SCHED */
7073 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7074 #ifdef CONFIG_RT_GROUP_SCHED
7075 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7078 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7079 rq
->cpu_load
[j
] = 0;
7081 rq
->last_load_update_tick
= jiffies
;
7086 rq
->cpu_capacity
= SCHED_CAPACITY_SCALE
;
7087 rq
->post_schedule
= 0;
7088 rq
->active_balance
= 0;
7089 rq
->next_balance
= jiffies
;
7094 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7095 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7097 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7099 rq_attach_root(rq
, &def_root_domain
);
7100 #ifdef CONFIG_NO_HZ_COMMON
7103 #ifdef CONFIG_NO_HZ_FULL
7104 rq
->last_sched_tick
= 0;
7108 atomic_set(&rq
->nr_iowait
, 0);
7111 set_load_weight(&init_task
);
7113 #ifdef CONFIG_PREEMPT_NOTIFIERS
7114 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7118 * The boot idle thread does lazy MMU switching as well:
7120 atomic_inc(&init_mm
.mm_count
);
7121 enter_lazy_tlb(&init_mm
, current
);
7124 * Make us the idle thread. Technically, schedule() should not be
7125 * called from this thread, however somewhere below it might be,
7126 * but because we are the idle thread, we just pick up running again
7127 * when this runqueue becomes "idle".
7129 init_idle(current
, smp_processor_id());
7131 calc_load_update
= jiffies
+ LOAD_FREQ
;
7134 * During early bootup we pretend to be a normal task:
7136 current
->sched_class
= &fair_sched_class
;
7139 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7140 /* May be allocated at isolcpus cmdline parse time */
7141 if (cpu_isolated_map
== NULL
)
7142 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7143 idle_thread_set_boot_cpu();
7144 set_cpu_rq_start_time();
7146 init_sched_fair_class();
7148 scheduler_running
= 1;
7151 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7152 static inline int preempt_count_equals(int preempt_offset
)
7154 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7156 return (nested
== preempt_offset
);
7159 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7161 static unsigned long prev_jiffy
; /* ratelimiting */
7163 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7164 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7165 !is_idle_task(current
)) ||
7166 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7168 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7170 prev_jiffy
= jiffies
;
7173 "BUG: sleeping function called from invalid context at %s:%d\n",
7176 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7177 in_atomic(), irqs_disabled(),
7178 current
->pid
, current
->comm
);
7180 debug_show_held_locks(current
);
7181 if (irqs_disabled())
7182 print_irqtrace_events(current
);
7183 #ifdef CONFIG_DEBUG_PREEMPT
7184 if (!preempt_count_equals(preempt_offset
)) {
7185 pr_err("Preemption disabled at:");
7186 print_ip_sym(current
->preempt_disable_ip
);
7192 EXPORT_SYMBOL(__might_sleep
);
7195 #ifdef CONFIG_MAGIC_SYSRQ
7196 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7198 const struct sched_class
*prev_class
= p
->sched_class
;
7199 struct sched_attr attr
= {
7200 .sched_policy
= SCHED_NORMAL
,
7202 int old_prio
= p
->prio
;
7205 queued
= task_on_rq_queued(p
);
7207 dequeue_task(rq
, p
, 0);
7208 __setscheduler(rq
, p
, &attr
);
7210 enqueue_task(rq
, p
, 0);
7214 check_class_changed(rq
, p
, prev_class
, old_prio
);
7217 void normalize_rt_tasks(void)
7219 struct task_struct
*g
, *p
;
7220 unsigned long flags
;
7223 read_lock(&tasklist_lock
);
7224 for_each_process_thread(g
, p
) {
7226 * Only normalize user tasks:
7228 if (p
->flags
& PF_KTHREAD
)
7231 p
->se
.exec_start
= 0;
7232 #ifdef CONFIG_SCHEDSTATS
7233 p
->se
.statistics
.wait_start
= 0;
7234 p
->se
.statistics
.sleep_start
= 0;
7235 p
->se
.statistics
.block_start
= 0;
7238 if (!dl_task(p
) && !rt_task(p
)) {
7240 * Renice negative nice level userspace
7243 if (task_nice(p
) < 0)
7244 set_user_nice(p
, 0);
7248 rq
= task_rq_lock(p
, &flags
);
7249 normalize_task(rq
, p
);
7250 task_rq_unlock(rq
, p
, &flags
);
7252 read_unlock(&tasklist_lock
);
7255 #endif /* CONFIG_MAGIC_SYSRQ */
7257 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7259 * These functions are only useful for the IA64 MCA handling, or kdb.
7261 * They can only be called when the whole system has been
7262 * stopped - every CPU needs to be quiescent, and no scheduling
7263 * activity can take place. Using them for anything else would
7264 * be a serious bug, and as a result, they aren't even visible
7265 * under any other configuration.
7269 * curr_task - return the current task for a given cpu.
7270 * @cpu: the processor in question.
7272 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7274 * Return: The current task for @cpu.
7276 struct task_struct
*curr_task(int cpu
)
7278 return cpu_curr(cpu
);
7281 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7285 * set_curr_task - set the current task for a given cpu.
7286 * @cpu: the processor in question.
7287 * @p: the task pointer to set.
7289 * Description: This function must only be used when non-maskable interrupts
7290 * are serviced on a separate stack. It allows the architecture to switch the
7291 * notion of the current task on a cpu in a non-blocking manner. This function
7292 * must be called with all CPU's synchronized, and interrupts disabled, the
7293 * and caller must save the original value of the current task (see
7294 * curr_task() above) and restore that value before reenabling interrupts and
7295 * re-starting the system.
7297 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7299 void set_curr_task(int cpu
, struct task_struct
*p
)
7306 #ifdef CONFIG_CGROUP_SCHED
7307 /* task_group_lock serializes the addition/removal of task groups */
7308 static DEFINE_SPINLOCK(task_group_lock
);
7310 static void free_sched_group(struct task_group
*tg
)
7312 free_fair_sched_group(tg
);
7313 free_rt_sched_group(tg
);
7318 /* allocate runqueue etc for a new task group */
7319 struct task_group
*sched_create_group(struct task_group
*parent
)
7321 struct task_group
*tg
;
7323 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7325 return ERR_PTR(-ENOMEM
);
7327 if (!alloc_fair_sched_group(tg
, parent
))
7330 if (!alloc_rt_sched_group(tg
, parent
))
7336 free_sched_group(tg
);
7337 return ERR_PTR(-ENOMEM
);
7340 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7342 unsigned long flags
;
7344 spin_lock_irqsave(&task_group_lock
, flags
);
7345 list_add_rcu(&tg
->list
, &task_groups
);
7347 WARN_ON(!parent
); /* root should already exist */
7349 tg
->parent
= parent
;
7350 INIT_LIST_HEAD(&tg
->children
);
7351 list_add_rcu(&tg
->siblings
, &parent
->children
);
7352 spin_unlock_irqrestore(&task_group_lock
, flags
);
7355 /* rcu callback to free various structures associated with a task group */
7356 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7358 /* now it should be safe to free those cfs_rqs */
7359 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7362 /* Destroy runqueue etc associated with a task group */
7363 void sched_destroy_group(struct task_group
*tg
)
7365 /* wait for possible concurrent references to cfs_rqs complete */
7366 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7369 void sched_offline_group(struct task_group
*tg
)
7371 unsigned long flags
;
7374 /* end participation in shares distribution */
7375 for_each_possible_cpu(i
)
7376 unregister_fair_sched_group(tg
, i
);
7378 spin_lock_irqsave(&task_group_lock
, flags
);
7379 list_del_rcu(&tg
->list
);
7380 list_del_rcu(&tg
->siblings
);
7381 spin_unlock_irqrestore(&task_group_lock
, flags
);
7384 /* change task's runqueue when it moves between groups.
7385 * The caller of this function should have put the task in its new group
7386 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7387 * reflect its new group.
7389 void sched_move_task(struct task_struct
*tsk
)
7391 struct task_group
*tg
;
7392 int queued
, running
;
7393 unsigned long flags
;
7396 rq
= task_rq_lock(tsk
, &flags
);
7398 running
= task_current(rq
, tsk
);
7399 queued
= task_on_rq_queued(tsk
);
7402 dequeue_task(rq
, tsk
, 0);
7403 if (unlikely(running
))
7404 put_prev_task(rq
, tsk
);
7406 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
,
7407 lockdep_is_held(&tsk
->sighand
->siglock
)),
7408 struct task_group
, css
);
7409 tg
= autogroup_task_group(tsk
, tg
);
7410 tsk
->sched_task_group
= tg
;
7412 #ifdef CONFIG_FAIR_GROUP_SCHED
7413 if (tsk
->sched_class
->task_move_group
)
7414 tsk
->sched_class
->task_move_group(tsk
, queued
);
7417 set_task_rq(tsk
, task_cpu(tsk
));
7419 if (unlikely(running
))
7420 tsk
->sched_class
->set_curr_task(rq
);
7422 enqueue_task(rq
, tsk
, 0);
7424 task_rq_unlock(rq
, tsk
, &flags
);
7426 #endif /* CONFIG_CGROUP_SCHED */
7428 #ifdef CONFIG_RT_GROUP_SCHED
7430 * Ensure that the real time constraints are schedulable.
7432 static DEFINE_MUTEX(rt_constraints_mutex
);
7434 /* Must be called with tasklist_lock held */
7435 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7437 struct task_struct
*g
, *p
;
7439 for_each_process_thread(g
, p
) {
7440 if (rt_task(p
) && task_group(p
) == tg
)
7447 struct rt_schedulable_data
{
7448 struct task_group
*tg
;
7453 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7455 struct rt_schedulable_data
*d
= data
;
7456 struct task_group
*child
;
7457 unsigned long total
, sum
= 0;
7458 u64 period
, runtime
;
7460 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7461 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7464 period
= d
->rt_period
;
7465 runtime
= d
->rt_runtime
;
7469 * Cannot have more runtime than the period.
7471 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7475 * Ensure we don't starve existing RT tasks.
7477 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7480 total
= to_ratio(period
, runtime
);
7483 * Nobody can have more than the global setting allows.
7485 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7489 * The sum of our children's runtime should not exceed our own.
7491 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7492 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7493 runtime
= child
->rt_bandwidth
.rt_runtime
;
7495 if (child
== d
->tg
) {
7496 period
= d
->rt_period
;
7497 runtime
= d
->rt_runtime
;
7500 sum
+= to_ratio(period
, runtime
);
7509 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7513 struct rt_schedulable_data data
= {
7515 .rt_period
= period
,
7516 .rt_runtime
= runtime
,
7520 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7526 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7527 u64 rt_period
, u64 rt_runtime
)
7531 mutex_lock(&rt_constraints_mutex
);
7532 read_lock(&tasklist_lock
);
7533 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7537 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7538 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7539 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7541 for_each_possible_cpu(i
) {
7542 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7544 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7545 rt_rq
->rt_runtime
= rt_runtime
;
7546 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7548 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7550 read_unlock(&tasklist_lock
);
7551 mutex_unlock(&rt_constraints_mutex
);
7556 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7558 u64 rt_runtime
, rt_period
;
7560 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7561 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7562 if (rt_runtime_us
< 0)
7563 rt_runtime
= RUNTIME_INF
;
7565 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7568 static long sched_group_rt_runtime(struct task_group
*tg
)
7572 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7575 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7576 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7577 return rt_runtime_us
;
7580 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7582 u64 rt_runtime
, rt_period
;
7584 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7585 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7590 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7593 static long sched_group_rt_period(struct task_group
*tg
)
7597 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7598 do_div(rt_period_us
, NSEC_PER_USEC
);
7599 return rt_period_us
;
7601 #endif /* CONFIG_RT_GROUP_SCHED */
7603 #ifdef CONFIG_RT_GROUP_SCHED
7604 static int sched_rt_global_constraints(void)
7608 mutex_lock(&rt_constraints_mutex
);
7609 read_lock(&tasklist_lock
);
7610 ret
= __rt_schedulable(NULL
, 0, 0);
7611 read_unlock(&tasklist_lock
);
7612 mutex_unlock(&rt_constraints_mutex
);
7617 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7619 /* Don't accept realtime tasks when there is no way for them to run */
7620 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7626 #else /* !CONFIG_RT_GROUP_SCHED */
7627 static int sched_rt_global_constraints(void)
7629 unsigned long flags
;
7632 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7633 for_each_possible_cpu(i
) {
7634 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7636 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7637 rt_rq
->rt_runtime
= global_rt_runtime();
7638 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7640 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7644 #endif /* CONFIG_RT_GROUP_SCHED */
7646 static int sched_dl_global_constraints(void)
7648 u64 runtime
= global_rt_runtime();
7649 u64 period
= global_rt_period();
7650 u64 new_bw
= to_ratio(period
, runtime
);
7653 unsigned long flags
;
7656 * Here we want to check the bandwidth not being set to some
7657 * value smaller than the currently allocated bandwidth in
7658 * any of the root_domains.
7660 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7661 * cycling on root_domains... Discussion on different/better
7662 * solutions is welcome!
7664 for_each_possible_cpu(cpu
) {
7665 rcu_read_lock_sched();
7666 dl_b
= dl_bw_of(cpu
);
7668 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7669 if (new_bw
< dl_b
->total_bw
)
7671 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7673 rcu_read_unlock_sched();
7682 static void sched_dl_do_global(void)
7687 unsigned long flags
;
7689 def_dl_bandwidth
.dl_period
= global_rt_period();
7690 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7692 if (global_rt_runtime() != RUNTIME_INF
)
7693 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7696 * FIXME: As above...
7698 for_each_possible_cpu(cpu
) {
7699 rcu_read_lock_sched();
7700 dl_b
= dl_bw_of(cpu
);
7702 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7704 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7706 rcu_read_unlock_sched();
7710 static int sched_rt_global_validate(void)
7712 if (sysctl_sched_rt_period
<= 0)
7715 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7716 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7722 static void sched_rt_do_global(void)
7724 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7725 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7728 int sched_rt_handler(struct ctl_table
*table
, int write
,
7729 void __user
*buffer
, size_t *lenp
,
7732 int old_period
, old_runtime
;
7733 static DEFINE_MUTEX(mutex
);
7737 old_period
= sysctl_sched_rt_period
;
7738 old_runtime
= sysctl_sched_rt_runtime
;
7740 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7742 if (!ret
&& write
) {
7743 ret
= sched_rt_global_validate();
7747 ret
= sched_rt_global_constraints();
7751 ret
= sched_dl_global_constraints();
7755 sched_rt_do_global();
7756 sched_dl_do_global();
7760 sysctl_sched_rt_period
= old_period
;
7761 sysctl_sched_rt_runtime
= old_runtime
;
7763 mutex_unlock(&mutex
);
7768 int sched_rr_handler(struct ctl_table
*table
, int write
,
7769 void __user
*buffer
, size_t *lenp
,
7773 static DEFINE_MUTEX(mutex
);
7776 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7777 /* make sure that internally we keep jiffies */
7778 /* also, writing zero resets timeslice to default */
7779 if (!ret
&& write
) {
7780 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7781 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7783 mutex_unlock(&mutex
);
7787 #ifdef CONFIG_CGROUP_SCHED
7789 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7791 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7794 static struct cgroup_subsys_state
*
7795 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7797 struct task_group
*parent
= css_tg(parent_css
);
7798 struct task_group
*tg
;
7801 /* This is early initialization for the top cgroup */
7802 return &root_task_group
.css
;
7805 tg
= sched_create_group(parent
);
7807 return ERR_PTR(-ENOMEM
);
7812 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7814 struct task_group
*tg
= css_tg(css
);
7815 struct task_group
*parent
= css_tg(css
->parent
);
7818 sched_online_group(tg
, parent
);
7822 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7824 struct task_group
*tg
= css_tg(css
);
7826 sched_destroy_group(tg
);
7829 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7831 struct task_group
*tg
= css_tg(css
);
7833 sched_offline_group(tg
);
7836 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7837 struct cgroup_taskset
*tset
)
7839 struct task_struct
*task
;
7841 cgroup_taskset_for_each(task
, tset
) {
7842 #ifdef CONFIG_RT_GROUP_SCHED
7843 if (!sched_rt_can_attach(css_tg(css
), task
))
7846 /* We don't support RT-tasks being in separate groups */
7847 if (task
->sched_class
!= &fair_sched_class
)
7854 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7855 struct cgroup_taskset
*tset
)
7857 struct task_struct
*task
;
7859 cgroup_taskset_for_each(task
, tset
)
7860 sched_move_task(task
);
7863 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7864 struct cgroup_subsys_state
*old_css
,
7865 struct task_struct
*task
)
7868 * cgroup_exit() is called in the copy_process() failure path.
7869 * Ignore this case since the task hasn't ran yet, this avoids
7870 * trying to poke a half freed task state from generic code.
7872 if (!(task
->flags
& PF_EXITING
))
7875 sched_move_task(task
);
7878 #ifdef CONFIG_FAIR_GROUP_SCHED
7879 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7880 struct cftype
*cftype
, u64 shareval
)
7882 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7885 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7888 struct task_group
*tg
= css_tg(css
);
7890 return (u64
) scale_load_down(tg
->shares
);
7893 #ifdef CONFIG_CFS_BANDWIDTH
7894 static DEFINE_MUTEX(cfs_constraints_mutex
);
7896 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7897 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7899 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7901 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7903 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7904 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7906 if (tg
== &root_task_group
)
7910 * Ensure we have at some amount of bandwidth every period. This is
7911 * to prevent reaching a state of large arrears when throttled via
7912 * entity_tick() resulting in prolonged exit starvation.
7914 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7918 * Likewise, bound things on the otherside by preventing insane quota
7919 * periods. This also allows us to normalize in computing quota
7922 if (period
> max_cfs_quota_period
)
7926 * Prevent race between setting of cfs_rq->runtime_enabled and
7927 * unthrottle_offline_cfs_rqs().
7930 mutex_lock(&cfs_constraints_mutex
);
7931 ret
= __cfs_schedulable(tg
, period
, quota
);
7935 runtime_enabled
= quota
!= RUNTIME_INF
;
7936 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7938 * If we need to toggle cfs_bandwidth_used, off->on must occur
7939 * before making related changes, and on->off must occur afterwards
7941 if (runtime_enabled
&& !runtime_was_enabled
)
7942 cfs_bandwidth_usage_inc();
7943 raw_spin_lock_irq(&cfs_b
->lock
);
7944 cfs_b
->period
= ns_to_ktime(period
);
7945 cfs_b
->quota
= quota
;
7947 __refill_cfs_bandwidth_runtime(cfs_b
);
7948 /* restart the period timer (if active) to handle new period expiry */
7949 if (runtime_enabled
&& cfs_b
->timer_active
) {
7950 /* force a reprogram */
7951 __start_cfs_bandwidth(cfs_b
, true);
7953 raw_spin_unlock_irq(&cfs_b
->lock
);
7955 for_each_online_cpu(i
) {
7956 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7957 struct rq
*rq
= cfs_rq
->rq
;
7959 raw_spin_lock_irq(&rq
->lock
);
7960 cfs_rq
->runtime_enabled
= runtime_enabled
;
7961 cfs_rq
->runtime_remaining
= 0;
7963 if (cfs_rq
->throttled
)
7964 unthrottle_cfs_rq(cfs_rq
);
7965 raw_spin_unlock_irq(&rq
->lock
);
7967 if (runtime_was_enabled
&& !runtime_enabled
)
7968 cfs_bandwidth_usage_dec();
7970 mutex_unlock(&cfs_constraints_mutex
);
7976 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7980 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7981 if (cfs_quota_us
< 0)
7982 quota
= RUNTIME_INF
;
7984 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7986 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7989 long tg_get_cfs_quota(struct task_group
*tg
)
7993 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7996 quota_us
= tg
->cfs_bandwidth
.quota
;
7997 do_div(quota_us
, NSEC_PER_USEC
);
8002 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8006 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8007 quota
= tg
->cfs_bandwidth
.quota
;
8009 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8012 long tg_get_cfs_period(struct task_group
*tg
)
8016 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8017 do_div(cfs_period_us
, NSEC_PER_USEC
);
8019 return cfs_period_us
;
8022 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8025 return tg_get_cfs_quota(css_tg(css
));
8028 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8029 struct cftype
*cftype
, s64 cfs_quota_us
)
8031 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8034 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8037 return tg_get_cfs_period(css_tg(css
));
8040 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8041 struct cftype
*cftype
, u64 cfs_period_us
)
8043 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8046 struct cfs_schedulable_data
{
8047 struct task_group
*tg
;
8052 * normalize group quota/period to be quota/max_period
8053 * note: units are usecs
8055 static u64
normalize_cfs_quota(struct task_group
*tg
,
8056 struct cfs_schedulable_data
*d
)
8064 period
= tg_get_cfs_period(tg
);
8065 quota
= tg_get_cfs_quota(tg
);
8068 /* note: these should typically be equivalent */
8069 if (quota
== RUNTIME_INF
|| quota
== -1)
8072 return to_ratio(period
, quota
);
8075 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8077 struct cfs_schedulable_data
*d
= data
;
8078 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8079 s64 quota
= 0, parent_quota
= -1;
8082 quota
= RUNTIME_INF
;
8084 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8086 quota
= normalize_cfs_quota(tg
, d
);
8087 parent_quota
= parent_b
->hierarchical_quota
;
8090 * ensure max(child_quota) <= parent_quota, inherit when no
8093 if (quota
== RUNTIME_INF
)
8094 quota
= parent_quota
;
8095 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8098 cfs_b
->hierarchical_quota
= quota
;
8103 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8106 struct cfs_schedulable_data data
= {
8112 if (quota
!= RUNTIME_INF
) {
8113 do_div(data
.period
, NSEC_PER_USEC
);
8114 do_div(data
.quota
, NSEC_PER_USEC
);
8118 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8124 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8126 struct task_group
*tg
= css_tg(seq_css(sf
));
8127 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8129 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8130 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8131 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8135 #endif /* CONFIG_CFS_BANDWIDTH */
8136 #endif /* CONFIG_FAIR_GROUP_SCHED */
8138 #ifdef CONFIG_RT_GROUP_SCHED
8139 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8140 struct cftype
*cft
, s64 val
)
8142 return sched_group_set_rt_runtime(css_tg(css
), val
);
8145 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8148 return sched_group_rt_runtime(css_tg(css
));
8151 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8152 struct cftype
*cftype
, u64 rt_period_us
)
8154 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8157 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8160 return sched_group_rt_period(css_tg(css
));
8162 #endif /* CONFIG_RT_GROUP_SCHED */
8164 static struct cftype cpu_files
[] = {
8165 #ifdef CONFIG_FAIR_GROUP_SCHED
8168 .read_u64
= cpu_shares_read_u64
,
8169 .write_u64
= cpu_shares_write_u64
,
8172 #ifdef CONFIG_CFS_BANDWIDTH
8174 .name
= "cfs_quota_us",
8175 .read_s64
= cpu_cfs_quota_read_s64
,
8176 .write_s64
= cpu_cfs_quota_write_s64
,
8179 .name
= "cfs_period_us",
8180 .read_u64
= cpu_cfs_period_read_u64
,
8181 .write_u64
= cpu_cfs_period_write_u64
,
8185 .seq_show
= cpu_stats_show
,
8188 #ifdef CONFIG_RT_GROUP_SCHED
8190 .name
= "rt_runtime_us",
8191 .read_s64
= cpu_rt_runtime_read
,
8192 .write_s64
= cpu_rt_runtime_write
,
8195 .name
= "rt_period_us",
8196 .read_u64
= cpu_rt_period_read_uint
,
8197 .write_u64
= cpu_rt_period_write_uint
,
8203 struct cgroup_subsys cpu_cgrp_subsys
= {
8204 .css_alloc
= cpu_cgroup_css_alloc
,
8205 .css_free
= cpu_cgroup_css_free
,
8206 .css_online
= cpu_cgroup_css_online
,
8207 .css_offline
= cpu_cgroup_css_offline
,
8208 .can_attach
= cpu_cgroup_can_attach
,
8209 .attach
= cpu_cgroup_attach
,
8210 .exit
= cpu_cgroup_exit
,
8211 .legacy_cftypes
= cpu_files
,
8215 #endif /* CONFIG_CGROUP_SCHED */
8217 void dump_cpu_task(int cpu
)
8219 pr_info("Task dump for CPU %d:\n", cpu
);
8220 sched_show_task(cpu_curr(cpu
));