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 accounted runtime for the task.
2479 * In case the task is currently running, return the runtime plus current's
2480 * pending runtime that have not been accounted yet.
2482 unsigned long long task_sched_runtime(struct task_struct
*p
)
2484 unsigned long flags
;
2488 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2490 * 64-bit doesn't need locks to atomically read a 64bit value.
2491 * So we have a optimization chance when the task's delta_exec is 0.
2492 * Reading ->on_cpu is racy, but this is ok.
2494 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2495 * If we race with it entering cpu, unaccounted time is 0. This is
2496 * indistinguishable from the read occurring a few cycles earlier.
2497 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2498 * been accounted, so we're correct here as well.
2500 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2501 return p
->se
.sum_exec_runtime
;
2504 rq
= task_rq_lock(p
, &flags
);
2506 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2507 * project cycles that may never be accounted to this
2508 * thread, breaking clock_gettime().
2510 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2511 update_rq_clock(rq
);
2512 p
->sched_class
->update_curr(rq
);
2514 ns
= p
->se
.sum_exec_runtime
;
2515 task_rq_unlock(rq
, p
, &flags
);
2521 * This function gets called by the timer code, with HZ frequency.
2522 * We call it with interrupts disabled.
2524 void scheduler_tick(void)
2526 int cpu
= smp_processor_id();
2527 struct rq
*rq
= cpu_rq(cpu
);
2528 struct task_struct
*curr
= rq
->curr
;
2532 raw_spin_lock(&rq
->lock
);
2533 update_rq_clock(rq
);
2534 curr
->sched_class
->task_tick(rq
, curr
, 0);
2535 update_cpu_load_active(rq
);
2536 raw_spin_unlock(&rq
->lock
);
2538 perf_event_task_tick();
2541 rq
->idle_balance
= idle_cpu(cpu
);
2542 trigger_load_balance(rq
);
2544 rq_last_tick_reset(rq
);
2547 #ifdef CONFIG_NO_HZ_FULL
2549 * scheduler_tick_max_deferment
2551 * Keep at least one tick per second when a single
2552 * active task is running because the scheduler doesn't
2553 * yet completely support full dynticks environment.
2555 * This makes sure that uptime, CFS vruntime, load
2556 * balancing, etc... continue to move forward, even
2557 * with a very low granularity.
2559 * Return: Maximum deferment in nanoseconds.
2561 u64
scheduler_tick_max_deferment(void)
2563 struct rq
*rq
= this_rq();
2564 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2566 next
= rq
->last_sched_tick
+ HZ
;
2568 if (time_before_eq(next
, now
))
2571 return jiffies_to_nsecs(next
- now
);
2575 notrace
unsigned long get_parent_ip(unsigned long addr
)
2577 if (in_lock_functions(addr
)) {
2578 addr
= CALLER_ADDR2
;
2579 if (in_lock_functions(addr
))
2580 addr
= CALLER_ADDR3
;
2585 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2586 defined(CONFIG_PREEMPT_TRACER))
2588 void preempt_count_add(int val
)
2590 #ifdef CONFIG_DEBUG_PREEMPT
2594 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2597 __preempt_count_add(val
);
2598 #ifdef CONFIG_DEBUG_PREEMPT
2600 * Spinlock count overflowing soon?
2602 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2605 if (preempt_count() == val
) {
2606 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2607 #ifdef CONFIG_DEBUG_PREEMPT
2608 current
->preempt_disable_ip
= ip
;
2610 trace_preempt_off(CALLER_ADDR0
, ip
);
2613 EXPORT_SYMBOL(preempt_count_add
);
2614 NOKPROBE_SYMBOL(preempt_count_add
);
2616 void preempt_count_sub(int val
)
2618 #ifdef CONFIG_DEBUG_PREEMPT
2622 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2625 * Is the spinlock portion underflowing?
2627 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2628 !(preempt_count() & PREEMPT_MASK
)))
2632 if (preempt_count() == val
)
2633 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2634 __preempt_count_sub(val
);
2636 EXPORT_SYMBOL(preempt_count_sub
);
2637 NOKPROBE_SYMBOL(preempt_count_sub
);
2642 * Print scheduling while atomic bug:
2644 static noinline
void __schedule_bug(struct task_struct
*prev
)
2646 if (oops_in_progress
)
2649 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2650 prev
->comm
, prev
->pid
, preempt_count());
2652 debug_show_held_locks(prev
);
2654 if (irqs_disabled())
2655 print_irqtrace_events(prev
);
2656 #ifdef CONFIG_DEBUG_PREEMPT
2657 if (in_atomic_preempt_off()) {
2658 pr_err("Preemption disabled at:");
2659 print_ip_sym(current
->preempt_disable_ip
);
2664 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2668 * Various schedule()-time debugging checks and statistics:
2670 static inline void schedule_debug(struct task_struct
*prev
)
2672 #ifdef CONFIG_SCHED_STACK_END_CHECK
2673 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2676 * Test if we are atomic. Since do_exit() needs to call into
2677 * schedule() atomically, we ignore that path. Otherwise whine
2678 * if we are scheduling when we should not.
2680 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2681 __schedule_bug(prev
);
2684 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2686 schedstat_inc(this_rq(), sched_count
);
2690 * Pick up the highest-prio task:
2692 static inline struct task_struct
*
2693 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2695 const struct sched_class
*class = &fair_sched_class
;
2696 struct task_struct
*p
;
2699 * Optimization: we know that if all tasks are in
2700 * the fair class we can call that function directly:
2702 if (likely(prev
->sched_class
== class &&
2703 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2704 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2705 if (unlikely(p
== RETRY_TASK
))
2708 /* assumes fair_sched_class->next == idle_sched_class */
2710 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2716 for_each_class(class) {
2717 p
= class->pick_next_task(rq
, prev
);
2719 if (unlikely(p
== RETRY_TASK
))
2725 BUG(); /* the idle class will always have a runnable task */
2729 * __schedule() is the main scheduler function.
2731 * The main means of driving the scheduler and thus entering this function are:
2733 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2735 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2736 * paths. For example, see arch/x86/entry_64.S.
2738 * To drive preemption between tasks, the scheduler sets the flag in timer
2739 * interrupt handler scheduler_tick().
2741 * 3. Wakeups don't really cause entry into schedule(). They add a
2742 * task to the run-queue and that's it.
2744 * Now, if the new task added to the run-queue preempts the current
2745 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2746 * called on the nearest possible occasion:
2748 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2750 * - in syscall or exception context, at the next outmost
2751 * preempt_enable(). (this might be as soon as the wake_up()'s
2754 * - in IRQ context, return from interrupt-handler to
2755 * preemptible context
2757 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2760 * - cond_resched() call
2761 * - explicit schedule() call
2762 * - return from syscall or exception to user-space
2763 * - return from interrupt-handler to user-space
2765 static void __sched
__schedule(void)
2767 struct task_struct
*prev
, *next
;
2768 unsigned long *switch_count
;
2774 cpu
= smp_processor_id();
2776 rcu_note_context_switch(cpu
);
2779 schedule_debug(prev
);
2781 if (sched_feat(HRTICK
))
2785 * Make sure that signal_pending_state()->signal_pending() below
2786 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2787 * done by the caller to avoid the race with signal_wake_up().
2789 smp_mb__before_spinlock();
2790 raw_spin_lock_irq(&rq
->lock
);
2792 switch_count
= &prev
->nivcsw
;
2793 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2794 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2795 prev
->state
= TASK_RUNNING
;
2797 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2801 * If a worker went to sleep, notify and ask workqueue
2802 * whether it wants to wake up a task to maintain
2805 if (prev
->flags
& PF_WQ_WORKER
) {
2806 struct task_struct
*to_wakeup
;
2808 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2810 try_to_wake_up_local(to_wakeup
);
2813 switch_count
= &prev
->nvcsw
;
2816 if (task_on_rq_queued(prev
) || rq
->skip_clock_update
< 0)
2817 update_rq_clock(rq
);
2819 next
= pick_next_task(rq
, prev
);
2820 clear_tsk_need_resched(prev
);
2821 clear_preempt_need_resched();
2822 rq
->skip_clock_update
= 0;
2824 if (likely(prev
!= next
)) {
2829 context_switch(rq
, prev
, next
); /* unlocks the rq */
2831 * The context switch have flipped the stack from under us
2832 * and restored the local variables which were saved when
2833 * this task called schedule() in the past. prev == current
2834 * is still correct, but it can be moved to another cpu/rq.
2836 cpu
= smp_processor_id();
2839 raw_spin_unlock_irq(&rq
->lock
);
2843 sched_preempt_enable_no_resched();
2848 static inline void sched_submit_work(struct task_struct
*tsk
)
2850 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2853 * If we are going to sleep and we have plugged IO queued,
2854 * make sure to submit it to avoid deadlocks.
2856 if (blk_needs_flush_plug(tsk
))
2857 blk_schedule_flush_plug(tsk
);
2860 asmlinkage __visible
void __sched
schedule(void)
2862 struct task_struct
*tsk
= current
;
2864 sched_submit_work(tsk
);
2867 EXPORT_SYMBOL(schedule
);
2869 #ifdef CONFIG_CONTEXT_TRACKING
2870 asmlinkage __visible
void __sched
schedule_user(void)
2873 * If we come here after a random call to set_need_resched(),
2874 * or we have been woken up remotely but the IPI has not yet arrived,
2875 * we haven't yet exited the RCU idle mode. Do it here manually until
2876 * we find a better solution.
2885 * schedule_preempt_disabled - called with preemption disabled
2887 * Returns with preemption disabled. Note: preempt_count must be 1
2889 void __sched
schedule_preempt_disabled(void)
2891 sched_preempt_enable_no_resched();
2896 #ifdef CONFIG_PREEMPT
2898 * this is the entry point to schedule() from in-kernel preemption
2899 * off of preempt_enable. Kernel preemptions off return from interrupt
2900 * occur there and call schedule directly.
2902 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2905 * If there is a non-zero preempt_count or interrupts are disabled,
2906 * we do not want to preempt the current task. Just return..
2908 if (likely(!preemptible()))
2912 __preempt_count_add(PREEMPT_ACTIVE
);
2914 __preempt_count_sub(PREEMPT_ACTIVE
);
2917 * Check again in case we missed a preemption opportunity
2918 * between schedule and now.
2921 } while (need_resched());
2923 NOKPROBE_SYMBOL(preempt_schedule
);
2924 EXPORT_SYMBOL(preempt_schedule
);
2926 #ifdef CONFIG_CONTEXT_TRACKING
2928 * preempt_schedule_context - preempt_schedule called by tracing
2930 * The tracing infrastructure uses preempt_enable_notrace to prevent
2931 * recursion and tracing preempt enabling caused by the tracing
2932 * infrastructure itself. But as tracing can happen in areas coming
2933 * from userspace or just about to enter userspace, a preempt enable
2934 * can occur before user_exit() is called. This will cause the scheduler
2935 * to be called when the system is still in usermode.
2937 * To prevent this, the preempt_enable_notrace will use this function
2938 * instead of preempt_schedule() to exit user context if needed before
2939 * calling the scheduler.
2941 asmlinkage __visible
void __sched notrace
preempt_schedule_context(void)
2943 enum ctx_state prev_ctx
;
2945 if (likely(!preemptible()))
2949 __preempt_count_add(PREEMPT_ACTIVE
);
2951 * Needs preempt disabled in case user_exit() is traced
2952 * and the tracer calls preempt_enable_notrace() causing
2953 * an infinite recursion.
2955 prev_ctx
= exception_enter();
2957 exception_exit(prev_ctx
);
2959 __preempt_count_sub(PREEMPT_ACTIVE
);
2961 } while (need_resched());
2963 EXPORT_SYMBOL_GPL(preempt_schedule_context
);
2964 #endif /* CONFIG_CONTEXT_TRACKING */
2966 #endif /* CONFIG_PREEMPT */
2969 * this is the entry point to schedule() from kernel preemption
2970 * off of irq context.
2971 * Note, that this is called and return with irqs disabled. This will
2972 * protect us against recursive calling from irq.
2974 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2976 enum ctx_state prev_state
;
2978 /* Catch callers which need to be fixed */
2979 BUG_ON(preempt_count() || !irqs_disabled());
2981 prev_state
= exception_enter();
2984 __preempt_count_add(PREEMPT_ACTIVE
);
2987 local_irq_disable();
2988 __preempt_count_sub(PREEMPT_ACTIVE
);
2991 * Check again in case we missed a preemption opportunity
2992 * between schedule and now.
2995 } while (need_resched());
2997 exception_exit(prev_state
);
3000 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3003 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3005 EXPORT_SYMBOL(default_wake_function
);
3007 #ifdef CONFIG_RT_MUTEXES
3010 * rt_mutex_setprio - set the current priority of a task
3012 * @prio: prio value (kernel-internal form)
3014 * This function changes the 'effective' priority of a task. It does
3015 * not touch ->normal_prio like __setscheduler().
3017 * Used by the rt_mutex code to implement priority inheritance
3018 * logic. Call site only calls if the priority of the task changed.
3020 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3022 int oldprio
, queued
, running
, enqueue_flag
= 0;
3024 const struct sched_class
*prev_class
;
3026 BUG_ON(prio
> MAX_PRIO
);
3028 rq
= __task_rq_lock(p
);
3031 * Idle task boosting is a nono in general. There is one
3032 * exception, when PREEMPT_RT and NOHZ is active:
3034 * The idle task calls get_next_timer_interrupt() and holds
3035 * the timer wheel base->lock on the CPU and another CPU wants
3036 * to access the timer (probably to cancel it). We can safely
3037 * ignore the boosting request, as the idle CPU runs this code
3038 * with interrupts disabled and will complete the lock
3039 * protected section without being interrupted. So there is no
3040 * real need to boost.
3042 if (unlikely(p
== rq
->idle
)) {
3043 WARN_ON(p
!= rq
->curr
);
3044 WARN_ON(p
->pi_blocked_on
);
3048 trace_sched_pi_setprio(p
, prio
);
3050 prev_class
= p
->sched_class
;
3051 queued
= task_on_rq_queued(p
);
3052 running
= task_current(rq
, p
);
3054 dequeue_task(rq
, p
, 0);
3056 put_prev_task(rq
, p
);
3059 * Boosting condition are:
3060 * 1. -rt task is running and holds mutex A
3061 * --> -dl task blocks on mutex A
3063 * 2. -dl task is running and holds mutex A
3064 * --> -dl task blocks on mutex A and could preempt the
3067 if (dl_prio(prio
)) {
3068 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3069 if (!dl_prio(p
->normal_prio
) ||
3070 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3071 p
->dl
.dl_boosted
= 1;
3072 p
->dl
.dl_throttled
= 0;
3073 enqueue_flag
= ENQUEUE_REPLENISH
;
3075 p
->dl
.dl_boosted
= 0;
3076 p
->sched_class
= &dl_sched_class
;
3077 } else if (rt_prio(prio
)) {
3078 if (dl_prio(oldprio
))
3079 p
->dl
.dl_boosted
= 0;
3081 enqueue_flag
= ENQUEUE_HEAD
;
3082 p
->sched_class
= &rt_sched_class
;
3084 if (dl_prio(oldprio
))
3085 p
->dl
.dl_boosted
= 0;
3086 p
->sched_class
= &fair_sched_class
;
3092 p
->sched_class
->set_curr_task(rq
);
3094 enqueue_task(rq
, p
, enqueue_flag
);
3096 check_class_changed(rq
, p
, prev_class
, oldprio
);
3098 __task_rq_unlock(rq
);
3102 void set_user_nice(struct task_struct
*p
, long nice
)
3104 int old_prio
, delta
, queued
;
3105 unsigned long flags
;
3108 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3111 * We have to be careful, if called from sys_setpriority(),
3112 * the task might be in the middle of scheduling on another CPU.
3114 rq
= task_rq_lock(p
, &flags
);
3116 * The RT priorities are set via sched_setscheduler(), but we still
3117 * allow the 'normal' nice value to be set - but as expected
3118 * it wont have any effect on scheduling until the task is
3119 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3121 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3122 p
->static_prio
= NICE_TO_PRIO(nice
);
3125 queued
= task_on_rq_queued(p
);
3127 dequeue_task(rq
, p
, 0);
3129 p
->static_prio
= NICE_TO_PRIO(nice
);
3132 p
->prio
= effective_prio(p
);
3133 delta
= p
->prio
- old_prio
;
3136 enqueue_task(rq
, p
, 0);
3138 * If the task increased its priority or is running and
3139 * lowered its priority, then reschedule its CPU:
3141 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3145 task_rq_unlock(rq
, p
, &flags
);
3147 EXPORT_SYMBOL(set_user_nice
);
3150 * can_nice - check if a task can reduce its nice value
3154 int can_nice(const struct task_struct
*p
, const int nice
)
3156 /* convert nice value [19,-20] to rlimit style value [1,40] */
3157 int nice_rlim
= nice_to_rlimit(nice
);
3159 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3160 capable(CAP_SYS_NICE
));
3163 #ifdef __ARCH_WANT_SYS_NICE
3166 * sys_nice - change the priority of the current process.
3167 * @increment: priority increment
3169 * sys_setpriority is a more generic, but much slower function that
3170 * does similar things.
3172 SYSCALL_DEFINE1(nice
, int, increment
)
3177 * Setpriority might change our priority at the same moment.
3178 * We don't have to worry. Conceptually one call occurs first
3179 * and we have a single winner.
3181 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3182 nice
= task_nice(current
) + increment
;
3184 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3185 if (increment
< 0 && !can_nice(current
, nice
))
3188 retval
= security_task_setnice(current
, nice
);
3192 set_user_nice(current
, nice
);
3199 * task_prio - return the priority value of a given task.
3200 * @p: the task in question.
3202 * Return: The priority value as seen by users in /proc.
3203 * RT tasks are offset by -200. Normal tasks are centered
3204 * around 0, value goes from -16 to +15.
3206 int task_prio(const struct task_struct
*p
)
3208 return p
->prio
- MAX_RT_PRIO
;
3212 * idle_cpu - is a given cpu idle currently?
3213 * @cpu: the processor in question.
3215 * Return: 1 if the CPU is currently idle. 0 otherwise.
3217 int idle_cpu(int cpu
)
3219 struct rq
*rq
= cpu_rq(cpu
);
3221 if (rq
->curr
!= rq
->idle
)
3228 if (!llist_empty(&rq
->wake_list
))
3236 * idle_task - return the idle task for a given cpu.
3237 * @cpu: the processor in question.
3239 * Return: The idle task for the cpu @cpu.
3241 struct task_struct
*idle_task(int cpu
)
3243 return cpu_rq(cpu
)->idle
;
3247 * find_process_by_pid - find a process with a matching PID value.
3248 * @pid: the pid in question.
3250 * The task of @pid, if found. %NULL otherwise.
3252 static struct task_struct
*find_process_by_pid(pid_t pid
)
3254 return pid
? find_task_by_vpid(pid
) : current
;
3258 * This function initializes the sched_dl_entity of a newly becoming
3259 * SCHED_DEADLINE task.
3261 * Only the static values are considered here, the actual runtime and the
3262 * absolute deadline will be properly calculated when the task is enqueued
3263 * for the first time with its new policy.
3266 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3268 struct sched_dl_entity
*dl_se
= &p
->dl
;
3270 init_dl_task_timer(dl_se
);
3271 dl_se
->dl_runtime
= attr
->sched_runtime
;
3272 dl_se
->dl_deadline
= attr
->sched_deadline
;
3273 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3274 dl_se
->flags
= attr
->sched_flags
;
3275 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3276 dl_se
->dl_throttled
= 0;
3278 dl_se
->dl_yielded
= 0;
3282 * sched_setparam() passes in -1 for its policy, to let the functions
3283 * it calls know not to change it.
3285 #define SETPARAM_POLICY -1
3287 static void __setscheduler_params(struct task_struct
*p
,
3288 const struct sched_attr
*attr
)
3290 int policy
= attr
->sched_policy
;
3292 if (policy
== SETPARAM_POLICY
)
3297 if (dl_policy(policy
))
3298 __setparam_dl(p
, attr
);
3299 else if (fair_policy(policy
))
3300 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3303 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3304 * !rt_policy. Always setting this ensures that things like
3305 * getparam()/getattr() don't report silly values for !rt tasks.
3307 p
->rt_priority
= attr
->sched_priority
;
3308 p
->normal_prio
= normal_prio(p
);
3312 /* Actually do priority change: must hold pi & rq lock. */
3313 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3314 const struct sched_attr
*attr
)
3316 __setscheduler_params(p
, attr
);
3319 * If we get here, there was no pi waiters boosting the
3320 * task. It is safe to use the normal prio.
3322 p
->prio
= normal_prio(p
);
3324 if (dl_prio(p
->prio
))
3325 p
->sched_class
= &dl_sched_class
;
3326 else if (rt_prio(p
->prio
))
3327 p
->sched_class
= &rt_sched_class
;
3329 p
->sched_class
= &fair_sched_class
;
3333 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3335 struct sched_dl_entity
*dl_se
= &p
->dl
;
3337 attr
->sched_priority
= p
->rt_priority
;
3338 attr
->sched_runtime
= dl_se
->dl_runtime
;
3339 attr
->sched_deadline
= dl_se
->dl_deadline
;
3340 attr
->sched_period
= dl_se
->dl_period
;
3341 attr
->sched_flags
= dl_se
->flags
;
3345 * This function validates the new parameters of a -deadline task.
3346 * We ask for the deadline not being zero, and greater or equal
3347 * than the runtime, as well as the period of being zero or
3348 * greater than deadline. Furthermore, we have to be sure that
3349 * user parameters are above the internal resolution of 1us (we
3350 * check sched_runtime only since it is always the smaller one) and
3351 * below 2^63 ns (we have to check both sched_deadline and
3352 * sched_period, as the latter can be zero).
3355 __checkparam_dl(const struct sched_attr
*attr
)
3358 if (attr
->sched_deadline
== 0)
3362 * Since we truncate DL_SCALE bits, make sure we're at least
3365 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3369 * Since we use the MSB for wrap-around and sign issues, make
3370 * sure it's not set (mind that period can be equal to zero).
3372 if (attr
->sched_deadline
& (1ULL << 63) ||
3373 attr
->sched_period
& (1ULL << 63))
3376 /* runtime <= deadline <= period (if period != 0) */
3377 if ((attr
->sched_period
!= 0 &&
3378 attr
->sched_period
< attr
->sched_deadline
) ||
3379 attr
->sched_deadline
< attr
->sched_runtime
)
3386 * check the target process has a UID that matches the current process's
3388 static bool check_same_owner(struct task_struct
*p
)
3390 const struct cred
*cred
= current_cred(), *pcred
;
3394 pcred
= __task_cred(p
);
3395 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3396 uid_eq(cred
->euid
, pcred
->uid
));
3401 static int __sched_setscheduler(struct task_struct
*p
,
3402 const struct sched_attr
*attr
,
3405 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3406 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3407 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3408 int policy
= attr
->sched_policy
;
3409 unsigned long flags
;
3410 const struct sched_class
*prev_class
;
3414 /* may grab non-irq protected spin_locks */
3415 BUG_ON(in_interrupt());
3417 /* double check policy once rq lock held */
3419 reset_on_fork
= p
->sched_reset_on_fork
;
3420 policy
= oldpolicy
= p
->policy
;
3422 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3424 if (policy
!= SCHED_DEADLINE
&&
3425 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3426 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3427 policy
!= SCHED_IDLE
)
3431 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3435 * Valid priorities for SCHED_FIFO and SCHED_RR are
3436 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3437 * SCHED_BATCH and SCHED_IDLE is 0.
3439 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3440 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3442 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3443 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3447 * Allow unprivileged RT tasks to decrease priority:
3449 if (user
&& !capable(CAP_SYS_NICE
)) {
3450 if (fair_policy(policy
)) {
3451 if (attr
->sched_nice
< task_nice(p
) &&
3452 !can_nice(p
, attr
->sched_nice
))
3456 if (rt_policy(policy
)) {
3457 unsigned long rlim_rtprio
=
3458 task_rlimit(p
, RLIMIT_RTPRIO
);
3460 /* can't set/change the rt policy */
3461 if (policy
!= p
->policy
&& !rlim_rtprio
)
3464 /* can't increase priority */
3465 if (attr
->sched_priority
> p
->rt_priority
&&
3466 attr
->sched_priority
> rlim_rtprio
)
3471 * Can't set/change SCHED_DEADLINE policy at all for now
3472 * (safest behavior); in the future we would like to allow
3473 * unprivileged DL tasks to increase their relative deadline
3474 * or reduce their runtime (both ways reducing utilization)
3476 if (dl_policy(policy
))
3480 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3481 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3483 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3484 if (!can_nice(p
, task_nice(p
)))
3488 /* can't change other user's priorities */
3489 if (!check_same_owner(p
))
3492 /* Normal users shall not reset the sched_reset_on_fork flag */
3493 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3498 retval
= security_task_setscheduler(p
);
3504 * make sure no PI-waiters arrive (or leave) while we are
3505 * changing the priority of the task:
3507 * To be able to change p->policy safely, the appropriate
3508 * runqueue lock must be held.
3510 rq
= task_rq_lock(p
, &flags
);
3513 * Changing the policy of the stop threads its a very bad idea
3515 if (p
== rq
->stop
) {
3516 task_rq_unlock(rq
, p
, &flags
);
3521 * If not changing anything there's no need to proceed further,
3522 * but store a possible modification of reset_on_fork.
3524 if (unlikely(policy
== p
->policy
)) {
3525 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3527 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3529 if (dl_policy(policy
))
3532 p
->sched_reset_on_fork
= reset_on_fork
;
3533 task_rq_unlock(rq
, p
, &flags
);
3539 #ifdef CONFIG_RT_GROUP_SCHED
3541 * Do not allow realtime tasks into groups that have no runtime
3544 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3545 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3546 !task_group_is_autogroup(task_group(p
))) {
3547 task_rq_unlock(rq
, p
, &flags
);
3552 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3553 cpumask_t
*span
= rq
->rd
->span
;
3556 * Don't allow tasks with an affinity mask smaller than
3557 * the entire root_domain to become SCHED_DEADLINE. We
3558 * will also fail if there's no bandwidth available.
3560 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3561 rq
->rd
->dl_bw
.bw
== 0) {
3562 task_rq_unlock(rq
, p
, &flags
);
3569 /* recheck policy now with rq lock held */
3570 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3571 policy
= oldpolicy
= -1;
3572 task_rq_unlock(rq
, p
, &flags
);
3577 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3578 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3581 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3582 task_rq_unlock(rq
, p
, &flags
);
3586 p
->sched_reset_on_fork
= reset_on_fork
;
3590 * Special case for priority boosted tasks.
3592 * If the new priority is lower or equal (user space view)
3593 * than the current (boosted) priority, we just store the new
3594 * normal parameters and do not touch the scheduler class and
3595 * the runqueue. This will be done when the task deboost
3598 if (rt_mutex_check_prio(p
, newprio
)) {
3599 __setscheduler_params(p
, attr
);
3600 task_rq_unlock(rq
, p
, &flags
);
3604 queued
= task_on_rq_queued(p
);
3605 running
= task_current(rq
, p
);
3607 dequeue_task(rq
, p
, 0);
3609 put_prev_task(rq
, p
);
3611 prev_class
= p
->sched_class
;
3612 __setscheduler(rq
, p
, attr
);
3615 p
->sched_class
->set_curr_task(rq
);
3618 * We enqueue to tail when the priority of a task is
3619 * increased (user space view).
3621 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3624 check_class_changed(rq
, p
, prev_class
, oldprio
);
3625 task_rq_unlock(rq
, p
, &flags
);
3627 rt_mutex_adjust_pi(p
);
3632 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3633 const struct sched_param
*param
, bool check
)
3635 struct sched_attr attr
= {
3636 .sched_policy
= policy
,
3637 .sched_priority
= param
->sched_priority
,
3638 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3641 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3642 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3643 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3644 policy
&= ~SCHED_RESET_ON_FORK
;
3645 attr
.sched_policy
= policy
;
3648 return __sched_setscheduler(p
, &attr
, check
);
3651 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3652 * @p: the task in question.
3653 * @policy: new policy.
3654 * @param: structure containing the new RT priority.
3656 * Return: 0 on success. An error code otherwise.
3658 * NOTE that the task may be already dead.
3660 int sched_setscheduler(struct task_struct
*p
, int policy
,
3661 const struct sched_param
*param
)
3663 return _sched_setscheduler(p
, policy
, param
, true);
3665 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3667 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3669 return __sched_setscheduler(p
, attr
, true);
3671 EXPORT_SYMBOL_GPL(sched_setattr
);
3674 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3675 * @p: the task in question.
3676 * @policy: new policy.
3677 * @param: structure containing the new RT priority.
3679 * Just like sched_setscheduler, only don't bother checking if the
3680 * current context has permission. For example, this is needed in
3681 * stop_machine(): we create temporary high priority worker threads,
3682 * but our caller might not have that capability.
3684 * Return: 0 on success. An error code otherwise.
3686 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3687 const struct sched_param
*param
)
3689 return _sched_setscheduler(p
, policy
, param
, false);
3693 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3695 struct sched_param lparam
;
3696 struct task_struct
*p
;
3699 if (!param
|| pid
< 0)
3701 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3706 p
= find_process_by_pid(pid
);
3708 retval
= sched_setscheduler(p
, policy
, &lparam
);
3715 * Mimics kernel/events/core.c perf_copy_attr().
3717 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3718 struct sched_attr
*attr
)
3723 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3727 * zero the full structure, so that a short copy will be nice.
3729 memset(attr
, 0, sizeof(*attr
));
3731 ret
= get_user(size
, &uattr
->size
);
3735 if (size
> PAGE_SIZE
) /* silly large */
3738 if (!size
) /* abi compat */
3739 size
= SCHED_ATTR_SIZE_VER0
;
3741 if (size
< SCHED_ATTR_SIZE_VER0
)
3745 * If we're handed a bigger struct than we know of,
3746 * ensure all the unknown bits are 0 - i.e. new
3747 * user-space does not rely on any kernel feature
3748 * extensions we dont know about yet.
3750 if (size
> sizeof(*attr
)) {
3751 unsigned char __user
*addr
;
3752 unsigned char __user
*end
;
3755 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3756 end
= (void __user
*)uattr
+ size
;
3758 for (; addr
< end
; addr
++) {
3759 ret
= get_user(val
, addr
);
3765 size
= sizeof(*attr
);
3768 ret
= copy_from_user(attr
, uattr
, size
);
3773 * XXX: do we want to be lenient like existing syscalls; or do we want
3774 * to be strict and return an error on out-of-bounds values?
3776 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3781 put_user(sizeof(*attr
), &uattr
->size
);
3786 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3787 * @pid: the pid in question.
3788 * @policy: new policy.
3789 * @param: structure containing the new RT priority.
3791 * Return: 0 on success. An error code otherwise.
3793 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3794 struct sched_param __user
*, param
)
3796 /* negative values for policy are not valid */
3800 return do_sched_setscheduler(pid
, policy
, param
);
3804 * sys_sched_setparam - set/change the RT priority of a thread
3805 * @pid: the pid in question.
3806 * @param: structure containing the new RT priority.
3808 * Return: 0 on success. An error code otherwise.
3810 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3812 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
3816 * sys_sched_setattr - same as above, but with extended sched_attr
3817 * @pid: the pid in question.
3818 * @uattr: structure containing the extended parameters.
3819 * @flags: for future extension.
3821 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3822 unsigned int, flags
)
3824 struct sched_attr attr
;
3825 struct task_struct
*p
;
3828 if (!uattr
|| pid
< 0 || flags
)
3831 retval
= sched_copy_attr(uattr
, &attr
);
3835 if ((int)attr
.sched_policy
< 0)
3840 p
= find_process_by_pid(pid
);
3842 retval
= sched_setattr(p
, &attr
);
3849 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3850 * @pid: the pid in question.
3852 * Return: On success, the policy of the thread. Otherwise, a negative error
3855 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3857 struct task_struct
*p
;
3865 p
= find_process_by_pid(pid
);
3867 retval
= security_task_getscheduler(p
);
3870 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3877 * sys_sched_getparam - get the RT priority of a thread
3878 * @pid: the pid in question.
3879 * @param: structure containing the RT priority.
3881 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3884 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3886 struct sched_param lp
= { .sched_priority
= 0 };
3887 struct task_struct
*p
;
3890 if (!param
|| pid
< 0)
3894 p
= find_process_by_pid(pid
);
3899 retval
= security_task_getscheduler(p
);
3903 if (task_has_rt_policy(p
))
3904 lp
.sched_priority
= p
->rt_priority
;
3908 * This one might sleep, we cannot do it with a spinlock held ...
3910 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3919 static int sched_read_attr(struct sched_attr __user
*uattr
,
3920 struct sched_attr
*attr
,
3925 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3929 * If we're handed a smaller struct than we know of,
3930 * ensure all the unknown bits are 0 - i.e. old
3931 * user-space does not get uncomplete information.
3933 if (usize
< sizeof(*attr
)) {
3934 unsigned char *addr
;
3937 addr
= (void *)attr
+ usize
;
3938 end
= (void *)attr
+ sizeof(*attr
);
3940 for (; addr
< end
; addr
++) {
3948 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3956 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3957 * @pid: the pid in question.
3958 * @uattr: structure containing the extended parameters.
3959 * @size: sizeof(attr) for fwd/bwd comp.
3960 * @flags: for future extension.
3962 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3963 unsigned int, size
, unsigned int, flags
)
3965 struct sched_attr attr
= {
3966 .size
= sizeof(struct sched_attr
),
3968 struct task_struct
*p
;
3971 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3972 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3976 p
= find_process_by_pid(pid
);
3981 retval
= security_task_getscheduler(p
);
3985 attr
.sched_policy
= p
->policy
;
3986 if (p
->sched_reset_on_fork
)
3987 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3988 if (task_has_dl_policy(p
))
3989 __getparam_dl(p
, &attr
);
3990 else if (task_has_rt_policy(p
))
3991 attr
.sched_priority
= p
->rt_priority
;
3993 attr
.sched_nice
= task_nice(p
);
3997 retval
= sched_read_attr(uattr
, &attr
, size
);
4005 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4007 cpumask_var_t cpus_allowed
, new_mask
;
4008 struct task_struct
*p
;
4013 p
= find_process_by_pid(pid
);
4019 /* Prevent p going away */
4023 if (p
->flags
& PF_NO_SETAFFINITY
) {
4027 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4031 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4033 goto out_free_cpus_allowed
;
4036 if (!check_same_owner(p
)) {
4038 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4040 goto out_free_new_mask
;
4045 retval
= security_task_setscheduler(p
);
4047 goto out_free_new_mask
;
4050 cpuset_cpus_allowed(p
, cpus_allowed
);
4051 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4054 * Since bandwidth control happens on root_domain basis,
4055 * if admission test is enabled, we only admit -deadline
4056 * tasks allowed to run on all the CPUs in the task's
4060 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4062 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4065 goto out_free_new_mask
;
4071 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4074 cpuset_cpus_allowed(p
, cpus_allowed
);
4075 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4077 * We must have raced with a concurrent cpuset
4078 * update. Just reset the cpus_allowed to the
4079 * cpuset's cpus_allowed
4081 cpumask_copy(new_mask
, cpus_allowed
);
4086 free_cpumask_var(new_mask
);
4087 out_free_cpus_allowed
:
4088 free_cpumask_var(cpus_allowed
);
4094 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4095 struct cpumask
*new_mask
)
4097 if (len
< cpumask_size())
4098 cpumask_clear(new_mask
);
4099 else if (len
> cpumask_size())
4100 len
= cpumask_size();
4102 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4106 * sys_sched_setaffinity - set the cpu affinity of a process
4107 * @pid: pid of the process
4108 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4109 * @user_mask_ptr: user-space pointer to the new cpu mask
4111 * Return: 0 on success. An error code otherwise.
4113 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4114 unsigned long __user
*, user_mask_ptr
)
4116 cpumask_var_t new_mask
;
4119 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4122 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4124 retval
= sched_setaffinity(pid
, new_mask
);
4125 free_cpumask_var(new_mask
);
4129 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4131 struct task_struct
*p
;
4132 unsigned long flags
;
4138 p
= find_process_by_pid(pid
);
4142 retval
= security_task_getscheduler(p
);
4146 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4147 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4148 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4157 * sys_sched_getaffinity - get the cpu affinity of a process
4158 * @pid: pid of the process
4159 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4160 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4162 * Return: 0 on success. An error code otherwise.
4164 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4165 unsigned long __user
*, user_mask_ptr
)
4170 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4172 if (len
& (sizeof(unsigned long)-1))
4175 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4178 ret
= sched_getaffinity(pid
, mask
);
4180 size_t retlen
= min_t(size_t, len
, cpumask_size());
4182 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4187 free_cpumask_var(mask
);
4193 * sys_sched_yield - yield the current processor to other threads.
4195 * This function yields the current CPU to other tasks. If there are no
4196 * other threads running on this CPU then this function will return.
4200 SYSCALL_DEFINE0(sched_yield
)
4202 struct rq
*rq
= this_rq_lock();
4204 schedstat_inc(rq
, yld_count
);
4205 current
->sched_class
->yield_task(rq
);
4208 * Since we are going to call schedule() anyway, there's
4209 * no need to preempt or enable interrupts:
4211 __release(rq
->lock
);
4212 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4213 do_raw_spin_unlock(&rq
->lock
);
4214 sched_preempt_enable_no_resched();
4221 static void __cond_resched(void)
4223 __preempt_count_add(PREEMPT_ACTIVE
);
4225 __preempt_count_sub(PREEMPT_ACTIVE
);
4228 int __sched
_cond_resched(void)
4230 if (should_resched()) {
4236 EXPORT_SYMBOL(_cond_resched
);
4239 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4240 * call schedule, and on return reacquire the lock.
4242 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4243 * operations here to prevent schedule() from being called twice (once via
4244 * spin_unlock(), once by hand).
4246 int __cond_resched_lock(spinlock_t
*lock
)
4248 int resched
= should_resched();
4251 lockdep_assert_held(lock
);
4253 if (spin_needbreak(lock
) || resched
) {
4264 EXPORT_SYMBOL(__cond_resched_lock
);
4266 int __sched
__cond_resched_softirq(void)
4268 BUG_ON(!in_softirq());
4270 if (should_resched()) {
4278 EXPORT_SYMBOL(__cond_resched_softirq
);
4281 * yield - yield the current processor to other threads.
4283 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4285 * The scheduler is at all times free to pick the calling task as the most
4286 * eligible task to run, if removing the yield() call from your code breaks
4287 * it, its already broken.
4289 * Typical broken usage is:
4294 * where one assumes that yield() will let 'the other' process run that will
4295 * make event true. If the current task is a SCHED_FIFO task that will never
4296 * happen. Never use yield() as a progress guarantee!!
4298 * If you want to use yield() to wait for something, use wait_event().
4299 * If you want to use yield() to be 'nice' for others, use cond_resched().
4300 * If you still want to use yield(), do not!
4302 void __sched
yield(void)
4304 set_current_state(TASK_RUNNING
);
4307 EXPORT_SYMBOL(yield
);
4310 * yield_to - yield the current processor to another thread in
4311 * your thread group, or accelerate that thread toward the
4312 * processor it's on.
4314 * @preempt: whether task preemption is allowed or not
4316 * It's the caller's job to ensure that the target task struct
4317 * can't go away on us before we can do any checks.
4320 * true (>0) if we indeed boosted the target task.
4321 * false (0) if we failed to boost the target.
4322 * -ESRCH if there's no task to yield to.
4324 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4326 struct task_struct
*curr
= current
;
4327 struct rq
*rq
, *p_rq
;
4328 unsigned long flags
;
4331 local_irq_save(flags
);
4337 * If we're the only runnable task on the rq and target rq also
4338 * has only one task, there's absolutely no point in yielding.
4340 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4345 double_rq_lock(rq
, p_rq
);
4346 if (task_rq(p
) != p_rq
) {
4347 double_rq_unlock(rq
, p_rq
);
4351 if (!curr
->sched_class
->yield_to_task
)
4354 if (curr
->sched_class
!= p
->sched_class
)
4357 if (task_running(p_rq
, p
) || p
->state
)
4360 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4362 schedstat_inc(rq
, yld_count
);
4364 * Make p's CPU reschedule; pick_next_entity takes care of
4367 if (preempt
&& rq
!= p_rq
)
4372 double_rq_unlock(rq
, p_rq
);
4374 local_irq_restore(flags
);
4381 EXPORT_SYMBOL_GPL(yield_to
);
4384 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4385 * that process accounting knows that this is a task in IO wait state.
4387 void __sched
io_schedule(void)
4389 struct rq
*rq
= raw_rq();
4391 delayacct_blkio_start();
4392 atomic_inc(&rq
->nr_iowait
);
4393 blk_flush_plug(current
);
4394 current
->in_iowait
= 1;
4396 current
->in_iowait
= 0;
4397 atomic_dec(&rq
->nr_iowait
);
4398 delayacct_blkio_end();
4400 EXPORT_SYMBOL(io_schedule
);
4402 long __sched
io_schedule_timeout(long timeout
)
4404 struct rq
*rq
= raw_rq();
4407 delayacct_blkio_start();
4408 atomic_inc(&rq
->nr_iowait
);
4409 blk_flush_plug(current
);
4410 current
->in_iowait
= 1;
4411 ret
= schedule_timeout(timeout
);
4412 current
->in_iowait
= 0;
4413 atomic_dec(&rq
->nr_iowait
);
4414 delayacct_blkio_end();
4419 * sys_sched_get_priority_max - return maximum RT priority.
4420 * @policy: scheduling class.
4422 * Return: On success, this syscall returns the maximum
4423 * rt_priority that can be used by a given scheduling class.
4424 * On failure, a negative error code is returned.
4426 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4433 ret
= MAX_USER_RT_PRIO
-1;
4435 case SCHED_DEADLINE
:
4446 * sys_sched_get_priority_min - return minimum RT priority.
4447 * @policy: scheduling class.
4449 * Return: On success, this syscall returns the minimum
4450 * rt_priority that can be used by a given scheduling class.
4451 * On failure, a negative error code is returned.
4453 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4462 case SCHED_DEADLINE
:
4472 * sys_sched_rr_get_interval - return the default timeslice of a process.
4473 * @pid: pid of the process.
4474 * @interval: userspace pointer to the timeslice value.
4476 * this syscall writes the default timeslice value of a given process
4477 * into the user-space timespec buffer. A value of '0' means infinity.
4479 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4482 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4483 struct timespec __user
*, interval
)
4485 struct task_struct
*p
;
4486 unsigned int time_slice
;
4487 unsigned long flags
;
4497 p
= find_process_by_pid(pid
);
4501 retval
= security_task_getscheduler(p
);
4505 rq
= task_rq_lock(p
, &flags
);
4507 if (p
->sched_class
->get_rr_interval
)
4508 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4509 task_rq_unlock(rq
, p
, &flags
);
4512 jiffies_to_timespec(time_slice
, &t
);
4513 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4521 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4523 void sched_show_task(struct task_struct
*p
)
4525 unsigned long free
= 0;
4529 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4530 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4531 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4532 #if BITS_PER_LONG == 32
4533 if (state
== TASK_RUNNING
)
4534 printk(KERN_CONT
" running ");
4536 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4538 if (state
== TASK_RUNNING
)
4539 printk(KERN_CONT
" running task ");
4541 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4543 #ifdef CONFIG_DEBUG_STACK_USAGE
4544 free
= stack_not_used(p
);
4547 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4549 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4550 task_pid_nr(p
), ppid
,
4551 (unsigned long)task_thread_info(p
)->flags
);
4553 print_worker_info(KERN_INFO
, p
);
4554 show_stack(p
, NULL
);
4557 void show_state_filter(unsigned long state_filter
)
4559 struct task_struct
*g
, *p
;
4561 #if BITS_PER_LONG == 32
4563 " task PC stack pid father\n");
4566 " task PC stack pid father\n");
4569 for_each_process_thread(g
, p
) {
4571 * reset the NMI-timeout, listing all files on a slow
4572 * console might take a lot of time:
4574 touch_nmi_watchdog();
4575 if (!state_filter
|| (p
->state
& state_filter
))
4579 touch_all_softlockup_watchdogs();
4581 #ifdef CONFIG_SCHED_DEBUG
4582 sysrq_sched_debug_show();
4586 * Only show locks if all tasks are dumped:
4589 debug_show_all_locks();
4592 void init_idle_bootup_task(struct task_struct
*idle
)
4594 idle
->sched_class
= &idle_sched_class
;
4598 * init_idle - set up an idle thread for a given CPU
4599 * @idle: task in question
4600 * @cpu: cpu the idle task belongs to
4602 * NOTE: this function does not set the idle thread's NEED_RESCHED
4603 * flag, to make booting more robust.
4605 void init_idle(struct task_struct
*idle
, int cpu
)
4607 struct rq
*rq
= cpu_rq(cpu
);
4608 unsigned long flags
;
4610 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4612 __sched_fork(0, idle
);
4613 idle
->state
= TASK_RUNNING
;
4614 idle
->se
.exec_start
= sched_clock();
4616 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4618 * We're having a chicken and egg problem, even though we are
4619 * holding rq->lock, the cpu isn't yet set to this cpu so the
4620 * lockdep check in task_group() will fail.
4622 * Similar case to sched_fork(). / Alternatively we could
4623 * use task_rq_lock() here and obtain the other rq->lock.
4628 __set_task_cpu(idle
, cpu
);
4631 rq
->curr
= rq
->idle
= idle
;
4632 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4633 #if defined(CONFIG_SMP)
4636 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4638 /* Set the preempt count _outside_ the spinlocks! */
4639 init_idle_preempt_count(idle
, cpu
);
4642 * The idle tasks have their own, simple scheduling class:
4644 idle
->sched_class
= &idle_sched_class
;
4645 ftrace_graph_init_idle_task(idle
, cpu
);
4646 vtime_init_idle(idle
, cpu
);
4647 #if defined(CONFIG_SMP)
4648 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4654 * move_queued_task - move a queued task to new rq.
4656 * Returns (locked) new rq. Old rq's lock is released.
4658 static struct rq
*move_queued_task(struct task_struct
*p
, int new_cpu
)
4660 struct rq
*rq
= task_rq(p
);
4662 lockdep_assert_held(&rq
->lock
);
4664 dequeue_task(rq
, p
, 0);
4665 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
4666 set_task_cpu(p
, new_cpu
);
4667 raw_spin_unlock(&rq
->lock
);
4669 rq
= cpu_rq(new_cpu
);
4671 raw_spin_lock(&rq
->lock
);
4672 BUG_ON(task_cpu(p
) != new_cpu
);
4673 p
->on_rq
= TASK_ON_RQ_QUEUED
;
4674 enqueue_task(rq
, p
, 0);
4675 check_preempt_curr(rq
, p
, 0);
4680 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4682 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4683 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4685 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4686 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4690 * This is how migration works:
4692 * 1) we invoke migration_cpu_stop() on the target CPU using
4694 * 2) stopper starts to run (implicitly forcing the migrated thread
4696 * 3) it checks whether the migrated task is still in the wrong runqueue.
4697 * 4) if it's in the wrong runqueue then the migration thread removes
4698 * it and puts it into the right queue.
4699 * 5) stopper completes and stop_one_cpu() returns and the migration
4704 * Change a given task's CPU affinity. Migrate the thread to a
4705 * proper CPU and schedule it away if the CPU it's executing on
4706 * is removed from the allowed bitmask.
4708 * NOTE: the caller must have a valid reference to the task, the
4709 * task must not exit() & deallocate itself prematurely. The
4710 * call is not atomic; no spinlocks may be held.
4712 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4714 unsigned long flags
;
4716 unsigned int dest_cpu
;
4719 rq
= task_rq_lock(p
, &flags
);
4721 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4724 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4729 do_set_cpus_allowed(p
, new_mask
);
4731 /* Can the task run on the task's current CPU? If so, we're done */
4732 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4735 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4736 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
4737 struct migration_arg arg
= { p
, dest_cpu
};
4738 /* Need help from migration thread: drop lock and wait. */
4739 task_rq_unlock(rq
, p
, &flags
);
4740 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4741 tlb_migrate_finish(p
->mm
);
4743 } else if (task_on_rq_queued(p
))
4744 rq
= move_queued_task(p
, dest_cpu
);
4746 task_rq_unlock(rq
, p
, &flags
);
4750 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4753 * Move (not current) task off this cpu, onto dest cpu. We're doing
4754 * this because either it can't run here any more (set_cpus_allowed()
4755 * away from this CPU, or CPU going down), or because we're
4756 * attempting to rebalance this task on exec (sched_exec).
4758 * So we race with normal scheduler movements, but that's OK, as long
4759 * as the task is no longer on this CPU.
4761 * Returns non-zero if task was successfully migrated.
4763 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4768 if (unlikely(!cpu_active(dest_cpu
)))
4771 rq
= cpu_rq(src_cpu
);
4773 raw_spin_lock(&p
->pi_lock
);
4774 raw_spin_lock(&rq
->lock
);
4775 /* Already moved. */
4776 if (task_cpu(p
) != src_cpu
)
4779 /* Affinity changed (again). */
4780 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4784 * If we're not on a rq, the next wake-up will ensure we're
4787 if (task_on_rq_queued(p
))
4788 rq
= move_queued_task(p
, dest_cpu
);
4792 raw_spin_unlock(&rq
->lock
);
4793 raw_spin_unlock(&p
->pi_lock
);
4797 #ifdef CONFIG_NUMA_BALANCING
4798 /* Migrate current task p to target_cpu */
4799 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4801 struct migration_arg arg
= { p
, target_cpu
};
4802 int curr_cpu
= task_cpu(p
);
4804 if (curr_cpu
== target_cpu
)
4807 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4810 /* TODO: This is not properly updating schedstats */
4812 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4813 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4817 * Requeue a task on a given node and accurately track the number of NUMA
4818 * tasks on the runqueues
4820 void sched_setnuma(struct task_struct
*p
, int nid
)
4823 unsigned long flags
;
4824 bool queued
, running
;
4826 rq
= task_rq_lock(p
, &flags
);
4827 queued
= task_on_rq_queued(p
);
4828 running
= task_current(rq
, p
);
4831 dequeue_task(rq
, p
, 0);
4833 put_prev_task(rq
, p
);
4835 p
->numa_preferred_nid
= nid
;
4838 p
->sched_class
->set_curr_task(rq
);
4840 enqueue_task(rq
, p
, 0);
4841 task_rq_unlock(rq
, p
, &flags
);
4846 * migration_cpu_stop - this will be executed by a highprio stopper thread
4847 * and performs thread migration by bumping thread off CPU then
4848 * 'pushing' onto another runqueue.
4850 static int migration_cpu_stop(void *data
)
4852 struct migration_arg
*arg
= data
;
4855 * The original target cpu might have gone down and we might
4856 * be on another cpu but it doesn't matter.
4858 local_irq_disable();
4860 * We need to explicitly wake pending tasks before running
4861 * __migrate_task() such that we will not miss enforcing cpus_allowed
4862 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4864 sched_ttwu_pending();
4865 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4870 #ifdef CONFIG_HOTPLUG_CPU
4873 * Ensures that the idle task is using init_mm right before its cpu goes
4876 void idle_task_exit(void)
4878 struct mm_struct
*mm
= current
->active_mm
;
4880 BUG_ON(cpu_online(smp_processor_id()));
4882 if (mm
!= &init_mm
) {
4883 switch_mm(mm
, &init_mm
, current
);
4884 finish_arch_post_lock_switch();
4890 * Since this CPU is going 'away' for a while, fold any nr_active delta
4891 * we might have. Assumes we're called after migrate_tasks() so that the
4892 * nr_active count is stable.
4894 * Also see the comment "Global load-average calculations".
4896 static void calc_load_migrate(struct rq
*rq
)
4898 long delta
= calc_load_fold_active(rq
);
4900 atomic_long_add(delta
, &calc_load_tasks
);
4903 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4907 static const struct sched_class fake_sched_class
= {
4908 .put_prev_task
= put_prev_task_fake
,
4911 static struct task_struct fake_task
= {
4913 * Avoid pull_{rt,dl}_task()
4915 .prio
= MAX_PRIO
+ 1,
4916 .sched_class
= &fake_sched_class
,
4920 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4921 * try_to_wake_up()->select_task_rq().
4923 * Called with rq->lock held even though we'er in stop_machine() and
4924 * there's no concurrency possible, we hold the required locks anyway
4925 * because of lock validation efforts.
4927 static void migrate_tasks(unsigned int dead_cpu
)
4929 struct rq
*rq
= cpu_rq(dead_cpu
);
4930 struct task_struct
*next
, *stop
= rq
->stop
;
4934 * Fudge the rq selection such that the below task selection loop
4935 * doesn't get stuck on the currently eligible stop task.
4937 * We're currently inside stop_machine() and the rq is either stuck
4938 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4939 * either way we should never end up calling schedule() until we're
4945 * put_prev_task() and pick_next_task() sched
4946 * class method both need to have an up-to-date
4947 * value of rq->clock[_task]
4949 update_rq_clock(rq
);
4953 * There's this thread running, bail when that's the only
4956 if (rq
->nr_running
== 1)
4959 next
= pick_next_task(rq
, &fake_task
);
4961 next
->sched_class
->put_prev_task(rq
, next
);
4963 /* Find suitable destination for @next, with force if needed. */
4964 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4965 raw_spin_unlock(&rq
->lock
);
4967 __migrate_task(next
, dead_cpu
, dest_cpu
);
4969 raw_spin_lock(&rq
->lock
);
4975 #endif /* CONFIG_HOTPLUG_CPU */
4977 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4979 static struct ctl_table sd_ctl_dir
[] = {
4981 .procname
= "sched_domain",
4987 static struct ctl_table sd_ctl_root
[] = {
4989 .procname
= "kernel",
4991 .child
= sd_ctl_dir
,
4996 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4998 struct ctl_table
*entry
=
4999 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5004 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5006 struct ctl_table
*entry
;
5009 * In the intermediate directories, both the child directory and
5010 * procname are dynamically allocated and could fail but the mode
5011 * will always be set. In the lowest directory the names are
5012 * static strings and all have proc handlers.
5014 for (entry
= *tablep
; entry
->mode
; entry
++) {
5016 sd_free_ctl_entry(&entry
->child
);
5017 if (entry
->proc_handler
== NULL
)
5018 kfree(entry
->procname
);
5025 static int min_load_idx
= 0;
5026 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5029 set_table_entry(struct ctl_table
*entry
,
5030 const char *procname
, void *data
, int maxlen
,
5031 umode_t mode
, proc_handler
*proc_handler
,
5034 entry
->procname
= procname
;
5036 entry
->maxlen
= maxlen
;
5038 entry
->proc_handler
= proc_handler
;
5041 entry
->extra1
= &min_load_idx
;
5042 entry
->extra2
= &max_load_idx
;
5046 static struct ctl_table
*
5047 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5049 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5054 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5055 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5056 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5057 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5058 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5059 sizeof(int), 0644, proc_dointvec_minmax
, true);
5060 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5061 sizeof(int), 0644, proc_dointvec_minmax
, true);
5062 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5063 sizeof(int), 0644, proc_dointvec_minmax
, true);
5064 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5065 sizeof(int), 0644, proc_dointvec_minmax
, true);
5066 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5067 sizeof(int), 0644, proc_dointvec_minmax
, true);
5068 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5069 sizeof(int), 0644, proc_dointvec_minmax
, false);
5070 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5071 sizeof(int), 0644, proc_dointvec_minmax
, false);
5072 set_table_entry(&table
[9], "cache_nice_tries",
5073 &sd
->cache_nice_tries
,
5074 sizeof(int), 0644, proc_dointvec_minmax
, false);
5075 set_table_entry(&table
[10], "flags", &sd
->flags
,
5076 sizeof(int), 0644, proc_dointvec_minmax
, false);
5077 set_table_entry(&table
[11], "max_newidle_lb_cost",
5078 &sd
->max_newidle_lb_cost
,
5079 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5080 set_table_entry(&table
[12], "name", sd
->name
,
5081 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5082 /* &table[13] is terminator */
5087 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5089 struct ctl_table
*entry
, *table
;
5090 struct sched_domain
*sd
;
5091 int domain_num
= 0, i
;
5094 for_each_domain(cpu
, sd
)
5096 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5101 for_each_domain(cpu
, sd
) {
5102 snprintf(buf
, 32, "domain%d", i
);
5103 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5105 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5112 static struct ctl_table_header
*sd_sysctl_header
;
5113 static void register_sched_domain_sysctl(void)
5115 int i
, cpu_num
= num_possible_cpus();
5116 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5119 WARN_ON(sd_ctl_dir
[0].child
);
5120 sd_ctl_dir
[0].child
= entry
;
5125 for_each_possible_cpu(i
) {
5126 snprintf(buf
, 32, "cpu%d", i
);
5127 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5129 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5133 WARN_ON(sd_sysctl_header
);
5134 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5137 /* may be called multiple times per register */
5138 static void unregister_sched_domain_sysctl(void)
5140 if (sd_sysctl_header
)
5141 unregister_sysctl_table(sd_sysctl_header
);
5142 sd_sysctl_header
= NULL
;
5143 if (sd_ctl_dir
[0].child
)
5144 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5147 static void register_sched_domain_sysctl(void)
5150 static void unregister_sched_domain_sysctl(void)
5155 static void set_rq_online(struct rq
*rq
)
5158 const struct sched_class
*class;
5160 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5163 for_each_class(class) {
5164 if (class->rq_online
)
5165 class->rq_online(rq
);
5170 static void set_rq_offline(struct rq
*rq
)
5173 const struct sched_class
*class;
5175 for_each_class(class) {
5176 if (class->rq_offline
)
5177 class->rq_offline(rq
);
5180 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5186 * migration_call - callback that gets triggered when a CPU is added.
5187 * Here we can start up the necessary migration thread for the new CPU.
5190 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5192 int cpu
= (long)hcpu
;
5193 unsigned long flags
;
5194 struct rq
*rq
= cpu_rq(cpu
);
5196 switch (action
& ~CPU_TASKS_FROZEN
) {
5198 case CPU_UP_PREPARE
:
5199 rq
->calc_load_update
= calc_load_update
;
5203 /* Update our root-domain */
5204 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5206 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5210 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5213 #ifdef CONFIG_HOTPLUG_CPU
5215 sched_ttwu_pending();
5216 /* Update our root-domain */
5217 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5219 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5223 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5224 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5228 calc_load_migrate(rq
);
5233 update_max_interval();
5239 * Register at high priority so that task migration (migrate_all_tasks)
5240 * happens before everything else. This has to be lower priority than
5241 * the notifier in the perf_event subsystem, though.
5243 static struct notifier_block migration_notifier
= {
5244 .notifier_call
= migration_call
,
5245 .priority
= CPU_PRI_MIGRATION
,
5248 static void __cpuinit
set_cpu_rq_start_time(void)
5250 int cpu
= smp_processor_id();
5251 struct rq
*rq
= cpu_rq(cpu
);
5252 rq
->age_stamp
= sched_clock_cpu(cpu
);
5255 static int sched_cpu_active(struct notifier_block
*nfb
,
5256 unsigned long action
, void *hcpu
)
5258 switch (action
& ~CPU_TASKS_FROZEN
) {
5260 set_cpu_rq_start_time();
5262 case CPU_DOWN_FAILED
:
5263 set_cpu_active((long)hcpu
, true);
5270 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5271 unsigned long action
, void *hcpu
)
5273 unsigned long flags
;
5274 long cpu
= (long)hcpu
;
5277 switch (action
& ~CPU_TASKS_FROZEN
) {
5278 case CPU_DOWN_PREPARE
:
5279 set_cpu_active(cpu
, false);
5281 /* explicitly allow suspend */
5282 if (!(action
& CPU_TASKS_FROZEN
)) {
5286 rcu_read_lock_sched();
5287 dl_b
= dl_bw_of(cpu
);
5289 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5290 cpus
= dl_bw_cpus(cpu
);
5291 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5292 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5294 rcu_read_unlock_sched();
5297 return notifier_from_errno(-EBUSY
);
5305 static int __init
migration_init(void)
5307 void *cpu
= (void *)(long)smp_processor_id();
5310 /* Initialize migration for the boot CPU */
5311 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5312 BUG_ON(err
== NOTIFY_BAD
);
5313 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5314 register_cpu_notifier(&migration_notifier
);
5316 /* Register cpu active notifiers */
5317 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5318 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5322 early_initcall(migration_init
);
5327 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5329 #ifdef CONFIG_SCHED_DEBUG
5331 static __read_mostly
int sched_debug_enabled
;
5333 static int __init
sched_debug_setup(char *str
)
5335 sched_debug_enabled
= 1;
5339 early_param("sched_debug", sched_debug_setup
);
5341 static inline bool sched_debug(void)
5343 return sched_debug_enabled
;
5346 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5347 struct cpumask
*groupmask
)
5349 struct sched_group
*group
= sd
->groups
;
5352 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5353 cpumask_clear(groupmask
);
5355 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5357 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5358 printk("does not load-balance\n");
5360 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5365 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5367 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5368 printk(KERN_ERR
"ERROR: domain->span does not contain "
5371 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5372 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5376 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5380 printk(KERN_ERR
"ERROR: group is NULL\n");
5385 * Even though we initialize ->capacity to something semi-sane,
5386 * we leave capacity_orig unset. This allows us to detect if
5387 * domain iteration is still funny without causing /0 traps.
5389 if (!group
->sgc
->capacity_orig
) {
5390 printk(KERN_CONT
"\n");
5391 printk(KERN_ERR
"ERROR: domain->cpu_capacity not set\n");
5395 if (!cpumask_weight(sched_group_cpus(group
))) {
5396 printk(KERN_CONT
"\n");
5397 printk(KERN_ERR
"ERROR: empty group\n");
5401 if (!(sd
->flags
& SD_OVERLAP
) &&
5402 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5403 printk(KERN_CONT
"\n");
5404 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5408 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5410 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5412 printk(KERN_CONT
" %s", str
);
5413 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5414 printk(KERN_CONT
" (cpu_capacity = %d)",
5415 group
->sgc
->capacity
);
5418 group
= group
->next
;
5419 } while (group
!= sd
->groups
);
5420 printk(KERN_CONT
"\n");
5422 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5423 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5426 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5427 printk(KERN_ERR
"ERROR: parent span is not a superset "
5428 "of domain->span\n");
5432 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5436 if (!sched_debug_enabled
)
5440 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5444 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5447 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5455 #else /* !CONFIG_SCHED_DEBUG */
5456 # define sched_domain_debug(sd, cpu) do { } while (0)
5457 static inline bool sched_debug(void)
5461 #endif /* CONFIG_SCHED_DEBUG */
5463 static int sd_degenerate(struct sched_domain
*sd
)
5465 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5468 /* Following flags need at least 2 groups */
5469 if (sd
->flags
& (SD_LOAD_BALANCE
|
5470 SD_BALANCE_NEWIDLE
|
5473 SD_SHARE_CPUCAPACITY
|
5474 SD_SHARE_PKG_RESOURCES
|
5475 SD_SHARE_POWERDOMAIN
)) {
5476 if (sd
->groups
!= sd
->groups
->next
)
5480 /* Following flags don't use groups */
5481 if (sd
->flags
& (SD_WAKE_AFFINE
))
5488 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5490 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5492 if (sd_degenerate(parent
))
5495 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5498 /* Flags needing groups don't count if only 1 group in parent */
5499 if (parent
->groups
== parent
->groups
->next
) {
5500 pflags
&= ~(SD_LOAD_BALANCE
|
5501 SD_BALANCE_NEWIDLE
|
5504 SD_SHARE_CPUCAPACITY
|
5505 SD_SHARE_PKG_RESOURCES
|
5507 SD_SHARE_POWERDOMAIN
);
5508 if (nr_node_ids
== 1)
5509 pflags
&= ~SD_SERIALIZE
;
5511 if (~cflags
& pflags
)
5517 static void free_rootdomain(struct rcu_head
*rcu
)
5519 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5521 cpupri_cleanup(&rd
->cpupri
);
5522 cpudl_cleanup(&rd
->cpudl
);
5523 free_cpumask_var(rd
->dlo_mask
);
5524 free_cpumask_var(rd
->rto_mask
);
5525 free_cpumask_var(rd
->online
);
5526 free_cpumask_var(rd
->span
);
5530 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5532 struct root_domain
*old_rd
= NULL
;
5533 unsigned long flags
;
5535 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5540 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5543 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5546 * If we dont want to free the old_rd yet then
5547 * set old_rd to NULL to skip the freeing later
5550 if (!atomic_dec_and_test(&old_rd
->refcount
))
5554 atomic_inc(&rd
->refcount
);
5557 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5558 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5561 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5564 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5567 static int init_rootdomain(struct root_domain
*rd
)
5569 memset(rd
, 0, sizeof(*rd
));
5571 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5573 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5575 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5577 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5580 init_dl_bw(&rd
->dl_bw
);
5581 if (cpudl_init(&rd
->cpudl
) != 0)
5584 if (cpupri_init(&rd
->cpupri
) != 0)
5589 free_cpumask_var(rd
->rto_mask
);
5591 free_cpumask_var(rd
->dlo_mask
);
5593 free_cpumask_var(rd
->online
);
5595 free_cpumask_var(rd
->span
);
5601 * By default the system creates a single root-domain with all cpus as
5602 * members (mimicking the global state we have today).
5604 struct root_domain def_root_domain
;
5606 static void init_defrootdomain(void)
5608 init_rootdomain(&def_root_domain
);
5610 atomic_set(&def_root_domain
.refcount
, 1);
5613 static struct root_domain
*alloc_rootdomain(void)
5615 struct root_domain
*rd
;
5617 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5621 if (init_rootdomain(rd
) != 0) {
5629 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5631 struct sched_group
*tmp
, *first
;
5640 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5645 } while (sg
!= first
);
5648 static void free_sched_domain(struct rcu_head
*rcu
)
5650 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5653 * If its an overlapping domain it has private groups, iterate and
5656 if (sd
->flags
& SD_OVERLAP
) {
5657 free_sched_groups(sd
->groups
, 1);
5658 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5659 kfree(sd
->groups
->sgc
);
5665 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5667 call_rcu(&sd
->rcu
, free_sched_domain
);
5670 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5672 for (; sd
; sd
= sd
->parent
)
5673 destroy_sched_domain(sd
, cpu
);
5677 * Keep a special pointer to the highest sched_domain that has
5678 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5679 * allows us to avoid some pointer chasing select_idle_sibling().
5681 * Also keep a unique ID per domain (we use the first cpu number in
5682 * the cpumask of the domain), this allows us to quickly tell if
5683 * two cpus are in the same cache domain, see cpus_share_cache().
5685 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5686 DEFINE_PER_CPU(int, sd_llc_size
);
5687 DEFINE_PER_CPU(int, sd_llc_id
);
5688 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5689 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5690 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5692 static void update_top_cache_domain(int cpu
)
5694 struct sched_domain
*sd
;
5695 struct sched_domain
*busy_sd
= NULL
;
5699 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5701 id
= cpumask_first(sched_domain_span(sd
));
5702 size
= cpumask_weight(sched_domain_span(sd
));
5703 busy_sd
= sd
->parent
; /* sd_busy */
5705 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5707 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5708 per_cpu(sd_llc_size
, cpu
) = size
;
5709 per_cpu(sd_llc_id
, cpu
) = id
;
5711 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5712 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5714 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5715 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5719 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5720 * hold the hotplug lock.
5723 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5725 struct rq
*rq
= cpu_rq(cpu
);
5726 struct sched_domain
*tmp
;
5728 /* Remove the sched domains which do not contribute to scheduling. */
5729 for (tmp
= sd
; tmp
; ) {
5730 struct sched_domain
*parent
= tmp
->parent
;
5734 if (sd_parent_degenerate(tmp
, parent
)) {
5735 tmp
->parent
= parent
->parent
;
5737 parent
->parent
->child
= tmp
;
5739 * Transfer SD_PREFER_SIBLING down in case of a
5740 * degenerate parent; the spans match for this
5741 * so the property transfers.
5743 if (parent
->flags
& SD_PREFER_SIBLING
)
5744 tmp
->flags
|= SD_PREFER_SIBLING
;
5745 destroy_sched_domain(parent
, cpu
);
5750 if (sd
&& sd_degenerate(sd
)) {
5753 destroy_sched_domain(tmp
, cpu
);
5758 sched_domain_debug(sd
, cpu
);
5760 rq_attach_root(rq
, rd
);
5762 rcu_assign_pointer(rq
->sd
, sd
);
5763 destroy_sched_domains(tmp
, cpu
);
5765 update_top_cache_domain(cpu
);
5768 /* cpus with isolated domains */
5769 static cpumask_var_t cpu_isolated_map
;
5771 /* Setup the mask of cpus configured for isolated domains */
5772 static int __init
isolated_cpu_setup(char *str
)
5774 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5775 cpulist_parse(str
, cpu_isolated_map
);
5779 __setup("isolcpus=", isolated_cpu_setup
);
5782 struct sched_domain
** __percpu sd
;
5783 struct root_domain
*rd
;
5794 * Build an iteration mask that can exclude certain CPUs from the upwards
5797 * Asymmetric node setups can result in situations where the domain tree is of
5798 * unequal depth, make sure to skip domains that already cover the entire
5801 * In that case build_sched_domains() will have terminated the iteration early
5802 * and our sibling sd spans will be empty. Domains should always include the
5803 * cpu they're built on, so check that.
5806 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5808 const struct cpumask
*span
= sched_domain_span(sd
);
5809 struct sd_data
*sdd
= sd
->private;
5810 struct sched_domain
*sibling
;
5813 for_each_cpu(i
, span
) {
5814 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5815 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5818 cpumask_set_cpu(i
, sched_group_mask(sg
));
5823 * Return the canonical balance cpu for this group, this is the first cpu
5824 * of this group that's also in the iteration mask.
5826 int group_balance_cpu(struct sched_group
*sg
)
5828 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5832 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5834 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5835 const struct cpumask
*span
= sched_domain_span(sd
);
5836 struct cpumask
*covered
= sched_domains_tmpmask
;
5837 struct sd_data
*sdd
= sd
->private;
5838 struct sched_domain
*sibling
;
5841 cpumask_clear(covered
);
5843 for_each_cpu(i
, span
) {
5844 struct cpumask
*sg_span
;
5846 if (cpumask_test_cpu(i
, covered
))
5849 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5851 /* See the comment near build_group_mask(). */
5852 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5855 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5856 GFP_KERNEL
, cpu_to_node(cpu
));
5861 sg_span
= sched_group_cpus(sg
);
5863 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
5865 cpumask_set_cpu(i
, sg_span
);
5867 cpumask_or(covered
, covered
, sg_span
);
5869 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
5870 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
5871 build_group_mask(sd
, sg
);
5874 * Initialize sgc->capacity such that even if we mess up the
5875 * domains and no possible iteration will get us here, we won't
5878 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
5879 sg
->sgc
->capacity_orig
= sg
->sgc
->capacity
;
5882 * Make sure the first group of this domain contains the
5883 * canonical balance cpu. Otherwise the sched_domain iteration
5884 * breaks. See update_sg_lb_stats().
5886 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5887 group_balance_cpu(sg
) == cpu
)
5897 sd
->groups
= groups
;
5902 free_sched_groups(first
, 0);
5907 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5909 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5910 struct sched_domain
*child
= sd
->child
;
5913 cpu
= cpumask_first(sched_domain_span(child
));
5916 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5917 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
5918 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
5925 * build_sched_groups will build a circular linked list of the groups
5926 * covered by the given span, and will set each group's ->cpumask correctly,
5927 * and ->cpu_capacity to 0.
5929 * Assumes the sched_domain tree is fully constructed
5932 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5934 struct sched_group
*first
= NULL
, *last
= NULL
;
5935 struct sd_data
*sdd
= sd
->private;
5936 const struct cpumask
*span
= sched_domain_span(sd
);
5937 struct cpumask
*covered
;
5940 get_group(cpu
, sdd
, &sd
->groups
);
5941 atomic_inc(&sd
->groups
->ref
);
5943 if (cpu
!= cpumask_first(span
))
5946 lockdep_assert_held(&sched_domains_mutex
);
5947 covered
= sched_domains_tmpmask
;
5949 cpumask_clear(covered
);
5951 for_each_cpu(i
, span
) {
5952 struct sched_group
*sg
;
5955 if (cpumask_test_cpu(i
, covered
))
5958 group
= get_group(i
, sdd
, &sg
);
5959 cpumask_setall(sched_group_mask(sg
));
5961 for_each_cpu(j
, span
) {
5962 if (get_group(j
, sdd
, NULL
) != group
)
5965 cpumask_set_cpu(j
, covered
);
5966 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5981 * Initialize sched groups cpu_capacity.
5983 * cpu_capacity indicates the capacity of sched group, which is used while
5984 * distributing the load between different sched groups in a sched domain.
5985 * Typically cpu_capacity for all the groups in a sched domain will be same
5986 * unless there are asymmetries in the topology. If there are asymmetries,
5987 * group having more cpu_capacity will pickup more load compared to the
5988 * group having less cpu_capacity.
5990 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
5992 struct sched_group
*sg
= sd
->groups
;
5997 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5999 } while (sg
!= sd
->groups
);
6001 if (cpu
!= group_balance_cpu(sg
))
6004 update_group_capacity(sd
, cpu
);
6005 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6009 * Initializers for schedule domains
6010 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6013 static int default_relax_domain_level
= -1;
6014 int sched_domain_level_max
;
6016 static int __init
setup_relax_domain_level(char *str
)
6018 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6019 pr_warn("Unable to set relax_domain_level\n");
6023 __setup("relax_domain_level=", setup_relax_domain_level
);
6025 static void set_domain_attribute(struct sched_domain
*sd
,
6026 struct sched_domain_attr
*attr
)
6030 if (!attr
|| attr
->relax_domain_level
< 0) {
6031 if (default_relax_domain_level
< 0)
6034 request
= default_relax_domain_level
;
6036 request
= attr
->relax_domain_level
;
6037 if (request
< sd
->level
) {
6038 /* turn off idle balance on this domain */
6039 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6041 /* turn on idle balance on this domain */
6042 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6046 static void __sdt_free(const struct cpumask
*cpu_map
);
6047 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6049 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6050 const struct cpumask
*cpu_map
)
6054 if (!atomic_read(&d
->rd
->refcount
))
6055 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6057 free_percpu(d
->sd
); /* fall through */
6059 __sdt_free(cpu_map
); /* fall through */
6065 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6066 const struct cpumask
*cpu_map
)
6068 memset(d
, 0, sizeof(*d
));
6070 if (__sdt_alloc(cpu_map
))
6071 return sa_sd_storage
;
6072 d
->sd
= alloc_percpu(struct sched_domain
*);
6074 return sa_sd_storage
;
6075 d
->rd
= alloc_rootdomain();
6078 return sa_rootdomain
;
6082 * NULL the sd_data elements we've used to build the sched_domain and
6083 * sched_group structure so that the subsequent __free_domain_allocs()
6084 * will not free the data we're using.
6086 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6088 struct sd_data
*sdd
= sd
->private;
6090 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6091 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6093 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6094 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6096 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6097 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6101 static int sched_domains_numa_levels
;
6102 static int *sched_domains_numa_distance
;
6103 static struct cpumask
***sched_domains_numa_masks
;
6104 static int sched_domains_curr_level
;
6108 * SD_flags allowed in topology descriptions.
6110 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6111 * SD_SHARE_PKG_RESOURCES - describes shared caches
6112 * SD_NUMA - describes NUMA topologies
6113 * SD_SHARE_POWERDOMAIN - describes shared power domain
6116 * SD_ASYM_PACKING - describes SMT quirks
6118 #define TOPOLOGY_SD_FLAGS \
6119 (SD_SHARE_CPUCAPACITY | \
6120 SD_SHARE_PKG_RESOURCES | \
6123 SD_SHARE_POWERDOMAIN)
6125 static struct sched_domain
*
6126 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6128 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6129 int sd_weight
, sd_flags
= 0;
6133 * Ugly hack to pass state to sd_numa_mask()...
6135 sched_domains_curr_level
= tl
->numa_level
;
6138 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6141 sd_flags
= (*tl
->sd_flags
)();
6142 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6143 "wrong sd_flags in topology description\n"))
6144 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6146 *sd
= (struct sched_domain
){
6147 .min_interval
= sd_weight
,
6148 .max_interval
= 2*sd_weight
,
6150 .imbalance_pct
= 125,
6152 .cache_nice_tries
= 0,
6159 .flags
= 1*SD_LOAD_BALANCE
6160 | 1*SD_BALANCE_NEWIDLE
6165 | 0*SD_SHARE_CPUCAPACITY
6166 | 0*SD_SHARE_PKG_RESOURCES
6168 | 0*SD_PREFER_SIBLING
6173 .last_balance
= jiffies
,
6174 .balance_interval
= sd_weight
,
6176 .max_newidle_lb_cost
= 0,
6177 .next_decay_max_lb_cost
= jiffies
,
6178 #ifdef CONFIG_SCHED_DEBUG
6184 * Convert topological properties into behaviour.
6187 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6188 sd
->imbalance_pct
= 110;
6189 sd
->smt_gain
= 1178; /* ~15% */
6191 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6192 sd
->imbalance_pct
= 117;
6193 sd
->cache_nice_tries
= 1;
6197 } else if (sd
->flags
& SD_NUMA
) {
6198 sd
->cache_nice_tries
= 2;
6202 sd
->flags
|= SD_SERIALIZE
;
6203 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6204 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6211 sd
->flags
|= SD_PREFER_SIBLING
;
6212 sd
->cache_nice_tries
= 1;
6217 sd
->private = &tl
->data
;
6223 * Topology list, bottom-up.
6225 static struct sched_domain_topology_level default_topology
[] = {
6226 #ifdef CONFIG_SCHED_SMT
6227 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6229 #ifdef CONFIG_SCHED_MC
6230 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6232 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6236 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6238 #define for_each_sd_topology(tl) \
6239 for (tl = sched_domain_topology; tl->mask; tl++)
6241 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6243 sched_domain_topology
= tl
;
6248 static const struct cpumask
*sd_numa_mask(int cpu
)
6250 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6253 static void sched_numa_warn(const char *str
)
6255 static int done
= false;
6263 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6265 for (i
= 0; i
< nr_node_ids
; i
++) {
6266 printk(KERN_WARNING
" ");
6267 for (j
= 0; j
< nr_node_ids
; j
++)
6268 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6269 printk(KERN_CONT
"\n");
6271 printk(KERN_WARNING
"\n");
6274 static bool find_numa_distance(int distance
)
6278 if (distance
== node_distance(0, 0))
6281 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6282 if (sched_domains_numa_distance
[i
] == distance
)
6289 static void sched_init_numa(void)
6291 int next_distance
, curr_distance
= node_distance(0, 0);
6292 struct sched_domain_topology_level
*tl
;
6296 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6297 if (!sched_domains_numa_distance
)
6301 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6302 * unique distances in the node_distance() table.
6304 * Assumes node_distance(0,j) includes all distances in
6305 * node_distance(i,j) in order to avoid cubic time.
6307 next_distance
= curr_distance
;
6308 for (i
= 0; i
< nr_node_ids
; i
++) {
6309 for (j
= 0; j
< nr_node_ids
; j
++) {
6310 for (k
= 0; k
< nr_node_ids
; k
++) {
6311 int distance
= node_distance(i
, k
);
6313 if (distance
> curr_distance
&&
6314 (distance
< next_distance
||
6315 next_distance
== curr_distance
))
6316 next_distance
= distance
;
6319 * While not a strong assumption it would be nice to know
6320 * about cases where if node A is connected to B, B is not
6321 * equally connected to A.
6323 if (sched_debug() && node_distance(k
, i
) != distance
)
6324 sched_numa_warn("Node-distance not symmetric");
6326 if (sched_debug() && i
&& !find_numa_distance(distance
))
6327 sched_numa_warn("Node-0 not representative");
6329 if (next_distance
!= curr_distance
) {
6330 sched_domains_numa_distance
[level
++] = next_distance
;
6331 sched_domains_numa_levels
= level
;
6332 curr_distance
= next_distance
;
6337 * In case of sched_debug() we verify the above assumption.
6347 * 'level' contains the number of unique distances, excluding the
6348 * identity distance node_distance(i,i).
6350 * The sched_domains_numa_distance[] array includes the actual distance
6355 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6356 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6357 * the array will contain less then 'level' members. This could be
6358 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6359 * in other functions.
6361 * We reset it to 'level' at the end of this function.
6363 sched_domains_numa_levels
= 0;
6365 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6366 if (!sched_domains_numa_masks
)
6370 * Now for each level, construct a mask per node which contains all
6371 * cpus of nodes that are that many hops away from us.
6373 for (i
= 0; i
< level
; i
++) {
6374 sched_domains_numa_masks
[i
] =
6375 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6376 if (!sched_domains_numa_masks
[i
])
6379 for (j
= 0; j
< nr_node_ids
; j
++) {
6380 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6384 sched_domains_numa_masks
[i
][j
] = mask
;
6386 for (k
= 0; k
< nr_node_ids
; k
++) {
6387 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6390 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6395 /* Compute default topology size */
6396 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6398 tl
= kzalloc((i
+ level
+ 1) *
6399 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6404 * Copy the default topology bits..
6406 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6407 tl
[i
] = sched_domain_topology
[i
];
6410 * .. and append 'j' levels of NUMA goodness.
6412 for (j
= 0; j
< level
; i
++, j
++) {
6413 tl
[i
] = (struct sched_domain_topology_level
){
6414 .mask
= sd_numa_mask
,
6415 .sd_flags
= cpu_numa_flags
,
6416 .flags
= SDTL_OVERLAP
,
6422 sched_domain_topology
= tl
;
6424 sched_domains_numa_levels
= level
;
6427 static void sched_domains_numa_masks_set(int cpu
)
6430 int node
= cpu_to_node(cpu
);
6432 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6433 for (j
= 0; j
< nr_node_ids
; j
++) {
6434 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6435 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6440 static void sched_domains_numa_masks_clear(int cpu
)
6443 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6444 for (j
= 0; j
< nr_node_ids
; j
++)
6445 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6450 * Update sched_domains_numa_masks[level][node] array when new cpus
6453 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6454 unsigned long action
,
6457 int cpu
= (long)hcpu
;
6459 switch (action
& ~CPU_TASKS_FROZEN
) {
6461 sched_domains_numa_masks_set(cpu
);
6465 sched_domains_numa_masks_clear(cpu
);
6475 static inline void sched_init_numa(void)
6479 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6480 unsigned long action
,
6485 #endif /* CONFIG_NUMA */
6487 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6489 struct sched_domain_topology_level
*tl
;
6492 for_each_sd_topology(tl
) {
6493 struct sd_data
*sdd
= &tl
->data
;
6495 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6499 sdd
->sg
= alloc_percpu(struct sched_group
*);
6503 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6507 for_each_cpu(j
, cpu_map
) {
6508 struct sched_domain
*sd
;
6509 struct sched_group
*sg
;
6510 struct sched_group_capacity
*sgc
;
6512 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6513 GFP_KERNEL
, cpu_to_node(j
));
6517 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6519 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6520 GFP_KERNEL
, cpu_to_node(j
));
6526 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6528 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6529 GFP_KERNEL
, cpu_to_node(j
));
6533 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6540 static void __sdt_free(const struct cpumask
*cpu_map
)
6542 struct sched_domain_topology_level
*tl
;
6545 for_each_sd_topology(tl
) {
6546 struct sd_data
*sdd
= &tl
->data
;
6548 for_each_cpu(j
, cpu_map
) {
6549 struct sched_domain
*sd
;
6552 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6553 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6554 free_sched_groups(sd
->groups
, 0);
6555 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6559 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6561 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6563 free_percpu(sdd
->sd
);
6565 free_percpu(sdd
->sg
);
6567 free_percpu(sdd
->sgc
);
6572 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6573 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6574 struct sched_domain
*child
, int cpu
)
6576 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6580 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6582 sd
->level
= child
->level
+ 1;
6583 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6587 if (!cpumask_subset(sched_domain_span(child
),
6588 sched_domain_span(sd
))) {
6589 pr_err("BUG: arch topology borken\n");
6590 #ifdef CONFIG_SCHED_DEBUG
6591 pr_err(" the %s domain not a subset of the %s domain\n",
6592 child
->name
, sd
->name
);
6594 /* Fixup, ensure @sd has at least @child cpus. */
6595 cpumask_or(sched_domain_span(sd
),
6596 sched_domain_span(sd
),
6597 sched_domain_span(child
));
6601 set_domain_attribute(sd
, attr
);
6607 * Build sched domains for a given set of cpus and attach the sched domains
6608 * to the individual cpus
6610 static int build_sched_domains(const struct cpumask
*cpu_map
,
6611 struct sched_domain_attr
*attr
)
6613 enum s_alloc alloc_state
;
6614 struct sched_domain
*sd
;
6616 int i
, ret
= -ENOMEM
;
6618 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6619 if (alloc_state
!= sa_rootdomain
)
6622 /* Set up domains for cpus specified by the cpu_map. */
6623 for_each_cpu(i
, cpu_map
) {
6624 struct sched_domain_topology_level
*tl
;
6627 for_each_sd_topology(tl
) {
6628 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6629 if (tl
== sched_domain_topology
)
6630 *per_cpu_ptr(d
.sd
, i
) = sd
;
6631 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6632 sd
->flags
|= SD_OVERLAP
;
6633 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6638 /* Build the groups for the domains */
6639 for_each_cpu(i
, cpu_map
) {
6640 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6641 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6642 if (sd
->flags
& SD_OVERLAP
) {
6643 if (build_overlap_sched_groups(sd
, i
))
6646 if (build_sched_groups(sd
, i
))
6652 /* Calculate CPU capacity for physical packages and nodes */
6653 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6654 if (!cpumask_test_cpu(i
, cpu_map
))
6657 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6658 claim_allocations(i
, sd
);
6659 init_sched_groups_capacity(i
, sd
);
6663 /* Attach the domains */
6665 for_each_cpu(i
, cpu_map
) {
6666 sd
= *per_cpu_ptr(d
.sd
, i
);
6667 cpu_attach_domain(sd
, d
.rd
, i
);
6673 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6677 static cpumask_var_t
*doms_cur
; /* current sched domains */
6678 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6679 static struct sched_domain_attr
*dattr_cur
;
6680 /* attribues of custom domains in 'doms_cur' */
6683 * Special case: If a kmalloc of a doms_cur partition (array of
6684 * cpumask) fails, then fallback to a single sched domain,
6685 * as determined by the single cpumask fallback_doms.
6687 static cpumask_var_t fallback_doms
;
6690 * arch_update_cpu_topology lets virtualized architectures update the
6691 * cpu core maps. It is supposed to return 1 if the topology changed
6692 * or 0 if it stayed the same.
6694 int __weak
arch_update_cpu_topology(void)
6699 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6702 cpumask_var_t
*doms
;
6704 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6707 for (i
= 0; i
< ndoms
; i
++) {
6708 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6709 free_sched_domains(doms
, i
);
6716 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6719 for (i
= 0; i
< ndoms
; i
++)
6720 free_cpumask_var(doms
[i
]);
6725 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6726 * For now this just excludes isolated cpus, but could be used to
6727 * exclude other special cases in the future.
6729 static int init_sched_domains(const struct cpumask
*cpu_map
)
6733 arch_update_cpu_topology();
6735 doms_cur
= alloc_sched_domains(ndoms_cur
);
6737 doms_cur
= &fallback_doms
;
6738 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6739 err
= build_sched_domains(doms_cur
[0], NULL
);
6740 register_sched_domain_sysctl();
6746 * Detach sched domains from a group of cpus specified in cpu_map
6747 * These cpus will now be attached to the NULL domain
6749 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6754 for_each_cpu(i
, cpu_map
)
6755 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6759 /* handle null as "default" */
6760 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6761 struct sched_domain_attr
*new, int idx_new
)
6763 struct sched_domain_attr tmp
;
6770 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6771 new ? (new + idx_new
) : &tmp
,
6772 sizeof(struct sched_domain_attr
));
6776 * Partition sched domains as specified by the 'ndoms_new'
6777 * cpumasks in the array doms_new[] of cpumasks. This compares
6778 * doms_new[] to the current sched domain partitioning, doms_cur[].
6779 * It destroys each deleted domain and builds each new domain.
6781 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6782 * The masks don't intersect (don't overlap.) We should setup one
6783 * sched domain for each mask. CPUs not in any of the cpumasks will
6784 * not be load balanced. If the same cpumask appears both in the
6785 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6788 * The passed in 'doms_new' should be allocated using
6789 * alloc_sched_domains. This routine takes ownership of it and will
6790 * free_sched_domains it when done with it. If the caller failed the
6791 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6792 * and partition_sched_domains() will fallback to the single partition
6793 * 'fallback_doms', it also forces the domains to be rebuilt.
6795 * If doms_new == NULL it will be replaced with cpu_online_mask.
6796 * ndoms_new == 0 is a special case for destroying existing domains,
6797 * and it will not create the default domain.
6799 * Call with hotplug lock held
6801 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6802 struct sched_domain_attr
*dattr_new
)
6807 mutex_lock(&sched_domains_mutex
);
6809 /* always unregister in case we don't destroy any domains */
6810 unregister_sched_domain_sysctl();
6812 /* Let architecture update cpu core mappings. */
6813 new_topology
= arch_update_cpu_topology();
6815 n
= doms_new
? ndoms_new
: 0;
6817 /* Destroy deleted domains */
6818 for (i
= 0; i
< ndoms_cur
; i
++) {
6819 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6820 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6821 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6824 /* no match - a current sched domain not in new doms_new[] */
6825 detach_destroy_domains(doms_cur
[i
]);
6831 if (doms_new
== NULL
) {
6833 doms_new
= &fallback_doms
;
6834 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6835 WARN_ON_ONCE(dattr_new
);
6838 /* Build new domains */
6839 for (i
= 0; i
< ndoms_new
; i
++) {
6840 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6841 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6842 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6845 /* no match - add a new doms_new */
6846 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6851 /* Remember the new sched domains */
6852 if (doms_cur
!= &fallback_doms
)
6853 free_sched_domains(doms_cur
, ndoms_cur
);
6854 kfree(dattr_cur
); /* kfree(NULL) is safe */
6855 doms_cur
= doms_new
;
6856 dattr_cur
= dattr_new
;
6857 ndoms_cur
= ndoms_new
;
6859 register_sched_domain_sysctl();
6861 mutex_unlock(&sched_domains_mutex
);
6864 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6867 * Update cpusets according to cpu_active mask. If cpusets are
6868 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6869 * around partition_sched_domains().
6871 * If we come here as part of a suspend/resume, don't touch cpusets because we
6872 * want to restore it back to its original state upon resume anyway.
6874 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6878 case CPU_ONLINE_FROZEN
:
6879 case CPU_DOWN_FAILED_FROZEN
:
6882 * num_cpus_frozen tracks how many CPUs are involved in suspend
6883 * resume sequence. As long as this is not the last online
6884 * operation in the resume sequence, just build a single sched
6885 * domain, ignoring cpusets.
6888 if (likely(num_cpus_frozen
)) {
6889 partition_sched_domains(1, NULL
, NULL
);
6894 * This is the last CPU online operation. So fall through and
6895 * restore the original sched domains by considering the
6896 * cpuset configurations.
6900 case CPU_DOWN_FAILED
:
6901 cpuset_update_active_cpus(true);
6909 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6913 case CPU_DOWN_PREPARE
:
6914 cpuset_update_active_cpus(false);
6916 case CPU_DOWN_PREPARE_FROZEN
:
6918 partition_sched_domains(1, NULL
, NULL
);
6926 void __init
sched_init_smp(void)
6928 cpumask_var_t non_isolated_cpus
;
6930 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6931 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6936 * There's no userspace yet to cause hotplug operations; hence all the
6937 * cpu masks are stable and all blatant races in the below code cannot
6940 mutex_lock(&sched_domains_mutex
);
6941 init_sched_domains(cpu_active_mask
);
6942 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6943 if (cpumask_empty(non_isolated_cpus
))
6944 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6945 mutex_unlock(&sched_domains_mutex
);
6947 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6948 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6949 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6953 /* Move init over to a non-isolated CPU */
6954 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6956 sched_init_granularity();
6957 free_cpumask_var(non_isolated_cpus
);
6959 init_sched_rt_class();
6960 init_sched_dl_class();
6963 void __init
sched_init_smp(void)
6965 sched_init_granularity();
6967 #endif /* CONFIG_SMP */
6969 const_debug
unsigned int sysctl_timer_migration
= 1;
6971 int in_sched_functions(unsigned long addr
)
6973 return in_lock_functions(addr
) ||
6974 (addr
>= (unsigned long)__sched_text_start
6975 && addr
< (unsigned long)__sched_text_end
);
6978 #ifdef CONFIG_CGROUP_SCHED
6980 * Default task group.
6981 * Every task in system belongs to this group at bootup.
6983 struct task_group root_task_group
;
6984 LIST_HEAD(task_groups
);
6987 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6989 void __init
sched_init(void)
6992 unsigned long alloc_size
= 0, ptr
;
6994 #ifdef CONFIG_FAIR_GROUP_SCHED
6995 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6997 #ifdef CONFIG_RT_GROUP_SCHED
6998 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7000 #ifdef CONFIG_CPUMASK_OFFSTACK
7001 alloc_size
+= num_possible_cpus() * cpumask_size();
7004 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7006 #ifdef CONFIG_FAIR_GROUP_SCHED
7007 root_task_group
.se
= (struct sched_entity
**)ptr
;
7008 ptr
+= nr_cpu_ids
* sizeof(void **);
7010 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7011 ptr
+= nr_cpu_ids
* sizeof(void **);
7013 #endif /* CONFIG_FAIR_GROUP_SCHED */
7014 #ifdef CONFIG_RT_GROUP_SCHED
7015 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7016 ptr
+= nr_cpu_ids
* sizeof(void **);
7018 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7019 ptr
+= nr_cpu_ids
* sizeof(void **);
7021 #endif /* CONFIG_RT_GROUP_SCHED */
7022 #ifdef CONFIG_CPUMASK_OFFSTACK
7023 for_each_possible_cpu(i
) {
7024 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
7025 ptr
+= cpumask_size();
7027 #endif /* CONFIG_CPUMASK_OFFSTACK */
7030 init_rt_bandwidth(&def_rt_bandwidth
,
7031 global_rt_period(), global_rt_runtime());
7032 init_dl_bandwidth(&def_dl_bandwidth
,
7033 global_rt_period(), global_rt_runtime());
7036 init_defrootdomain();
7039 #ifdef CONFIG_RT_GROUP_SCHED
7040 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7041 global_rt_period(), global_rt_runtime());
7042 #endif /* CONFIG_RT_GROUP_SCHED */
7044 #ifdef CONFIG_CGROUP_SCHED
7045 list_add(&root_task_group
.list
, &task_groups
);
7046 INIT_LIST_HEAD(&root_task_group
.children
);
7047 INIT_LIST_HEAD(&root_task_group
.siblings
);
7048 autogroup_init(&init_task
);
7050 #endif /* CONFIG_CGROUP_SCHED */
7052 for_each_possible_cpu(i
) {
7056 raw_spin_lock_init(&rq
->lock
);
7058 rq
->calc_load_active
= 0;
7059 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7060 init_cfs_rq(&rq
->cfs
);
7061 init_rt_rq(&rq
->rt
, rq
);
7062 init_dl_rq(&rq
->dl
, rq
);
7063 #ifdef CONFIG_FAIR_GROUP_SCHED
7064 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7065 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7067 * How much cpu bandwidth does root_task_group get?
7069 * In case of task-groups formed thr' the cgroup filesystem, it
7070 * gets 100% of the cpu resources in the system. This overall
7071 * system cpu resource is divided among the tasks of
7072 * root_task_group and its child task-groups in a fair manner,
7073 * based on each entity's (task or task-group's) weight
7074 * (se->load.weight).
7076 * In other words, if root_task_group has 10 tasks of weight
7077 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7078 * then A0's share of the cpu resource is:
7080 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7082 * We achieve this by letting root_task_group's tasks sit
7083 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7085 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7086 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7087 #endif /* CONFIG_FAIR_GROUP_SCHED */
7089 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7090 #ifdef CONFIG_RT_GROUP_SCHED
7091 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7094 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7095 rq
->cpu_load
[j
] = 0;
7097 rq
->last_load_update_tick
= jiffies
;
7102 rq
->cpu_capacity
= SCHED_CAPACITY_SCALE
;
7103 rq
->post_schedule
= 0;
7104 rq
->active_balance
= 0;
7105 rq
->next_balance
= jiffies
;
7110 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7111 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7113 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7115 rq_attach_root(rq
, &def_root_domain
);
7116 #ifdef CONFIG_NO_HZ_COMMON
7119 #ifdef CONFIG_NO_HZ_FULL
7120 rq
->last_sched_tick
= 0;
7124 atomic_set(&rq
->nr_iowait
, 0);
7127 set_load_weight(&init_task
);
7129 #ifdef CONFIG_PREEMPT_NOTIFIERS
7130 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7134 * The boot idle thread does lazy MMU switching as well:
7136 atomic_inc(&init_mm
.mm_count
);
7137 enter_lazy_tlb(&init_mm
, current
);
7140 * Make us the idle thread. Technically, schedule() should not be
7141 * called from this thread, however somewhere below it might be,
7142 * but because we are the idle thread, we just pick up running again
7143 * when this runqueue becomes "idle".
7145 init_idle(current
, smp_processor_id());
7147 calc_load_update
= jiffies
+ LOAD_FREQ
;
7150 * During early bootup we pretend to be a normal task:
7152 current
->sched_class
= &fair_sched_class
;
7155 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7156 /* May be allocated at isolcpus cmdline parse time */
7157 if (cpu_isolated_map
== NULL
)
7158 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7159 idle_thread_set_boot_cpu();
7160 set_cpu_rq_start_time();
7162 init_sched_fair_class();
7164 scheduler_running
= 1;
7167 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7168 static inline int preempt_count_equals(int preempt_offset
)
7170 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7172 return (nested
== preempt_offset
);
7175 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7177 static unsigned long prev_jiffy
; /* ratelimiting */
7179 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7180 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7181 !is_idle_task(current
)) ||
7182 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7184 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7186 prev_jiffy
= jiffies
;
7189 "BUG: sleeping function called from invalid context at %s:%d\n",
7192 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7193 in_atomic(), irqs_disabled(),
7194 current
->pid
, current
->comm
);
7196 debug_show_held_locks(current
);
7197 if (irqs_disabled())
7198 print_irqtrace_events(current
);
7199 #ifdef CONFIG_DEBUG_PREEMPT
7200 if (!preempt_count_equals(preempt_offset
)) {
7201 pr_err("Preemption disabled at:");
7202 print_ip_sym(current
->preempt_disable_ip
);
7208 EXPORT_SYMBOL(__might_sleep
);
7211 #ifdef CONFIG_MAGIC_SYSRQ
7212 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7214 const struct sched_class
*prev_class
= p
->sched_class
;
7215 struct sched_attr attr
= {
7216 .sched_policy
= SCHED_NORMAL
,
7218 int old_prio
= p
->prio
;
7221 queued
= task_on_rq_queued(p
);
7223 dequeue_task(rq
, p
, 0);
7224 __setscheduler(rq
, p
, &attr
);
7226 enqueue_task(rq
, p
, 0);
7230 check_class_changed(rq
, p
, prev_class
, old_prio
);
7233 void normalize_rt_tasks(void)
7235 struct task_struct
*g
, *p
;
7236 unsigned long flags
;
7239 read_lock(&tasklist_lock
);
7240 for_each_process_thread(g
, p
) {
7242 * Only normalize user tasks:
7244 if (p
->flags
& PF_KTHREAD
)
7247 p
->se
.exec_start
= 0;
7248 #ifdef CONFIG_SCHEDSTATS
7249 p
->se
.statistics
.wait_start
= 0;
7250 p
->se
.statistics
.sleep_start
= 0;
7251 p
->se
.statistics
.block_start
= 0;
7254 if (!dl_task(p
) && !rt_task(p
)) {
7256 * Renice negative nice level userspace
7259 if (task_nice(p
) < 0)
7260 set_user_nice(p
, 0);
7264 rq
= task_rq_lock(p
, &flags
);
7265 normalize_task(rq
, p
);
7266 task_rq_unlock(rq
, p
, &flags
);
7268 read_unlock(&tasklist_lock
);
7271 #endif /* CONFIG_MAGIC_SYSRQ */
7273 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7275 * These functions are only useful for the IA64 MCA handling, or kdb.
7277 * They can only be called when the whole system has been
7278 * stopped - every CPU needs to be quiescent, and no scheduling
7279 * activity can take place. Using them for anything else would
7280 * be a serious bug, and as a result, they aren't even visible
7281 * under any other configuration.
7285 * curr_task - return the current task for a given cpu.
7286 * @cpu: the processor in question.
7288 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7290 * Return: The current task for @cpu.
7292 struct task_struct
*curr_task(int cpu
)
7294 return cpu_curr(cpu
);
7297 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7301 * set_curr_task - set the current task for a given cpu.
7302 * @cpu: the processor in question.
7303 * @p: the task pointer to set.
7305 * Description: This function must only be used when non-maskable interrupts
7306 * are serviced on a separate stack. It allows the architecture to switch the
7307 * notion of the current task on a cpu in a non-blocking manner. This function
7308 * must be called with all CPU's synchronized, and interrupts disabled, the
7309 * and caller must save the original value of the current task (see
7310 * curr_task() above) and restore that value before reenabling interrupts and
7311 * re-starting the system.
7313 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7315 void set_curr_task(int cpu
, struct task_struct
*p
)
7322 #ifdef CONFIG_CGROUP_SCHED
7323 /* task_group_lock serializes the addition/removal of task groups */
7324 static DEFINE_SPINLOCK(task_group_lock
);
7326 static void free_sched_group(struct task_group
*tg
)
7328 free_fair_sched_group(tg
);
7329 free_rt_sched_group(tg
);
7334 /* allocate runqueue etc for a new task group */
7335 struct task_group
*sched_create_group(struct task_group
*parent
)
7337 struct task_group
*tg
;
7339 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7341 return ERR_PTR(-ENOMEM
);
7343 if (!alloc_fair_sched_group(tg
, parent
))
7346 if (!alloc_rt_sched_group(tg
, parent
))
7352 free_sched_group(tg
);
7353 return ERR_PTR(-ENOMEM
);
7356 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7358 unsigned long flags
;
7360 spin_lock_irqsave(&task_group_lock
, flags
);
7361 list_add_rcu(&tg
->list
, &task_groups
);
7363 WARN_ON(!parent
); /* root should already exist */
7365 tg
->parent
= parent
;
7366 INIT_LIST_HEAD(&tg
->children
);
7367 list_add_rcu(&tg
->siblings
, &parent
->children
);
7368 spin_unlock_irqrestore(&task_group_lock
, flags
);
7371 /* rcu callback to free various structures associated with a task group */
7372 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7374 /* now it should be safe to free those cfs_rqs */
7375 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7378 /* Destroy runqueue etc associated with a task group */
7379 void sched_destroy_group(struct task_group
*tg
)
7381 /* wait for possible concurrent references to cfs_rqs complete */
7382 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7385 void sched_offline_group(struct task_group
*tg
)
7387 unsigned long flags
;
7390 /* end participation in shares distribution */
7391 for_each_possible_cpu(i
)
7392 unregister_fair_sched_group(tg
, i
);
7394 spin_lock_irqsave(&task_group_lock
, flags
);
7395 list_del_rcu(&tg
->list
);
7396 list_del_rcu(&tg
->siblings
);
7397 spin_unlock_irqrestore(&task_group_lock
, flags
);
7400 /* change task's runqueue when it moves between groups.
7401 * The caller of this function should have put the task in its new group
7402 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7403 * reflect its new group.
7405 void sched_move_task(struct task_struct
*tsk
)
7407 struct task_group
*tg
;
7408 int queued
, running
;
7409 unsigned long flags
;
7412 rq
= task_rq_lock(tsk
, &flags
);
7414 running
= task_current(rq
, tsk
);
7415 queued
= task_on_rq_queued(tsk
);
7418 dequeue_task(rq
, tsk
, 0);
7419 if (unlikely(running
))
7420 put_prev_task(rq
, tsk
);
7423 * All callers are synchronized by task_rq_lock(); we do not use RCU
7424 * which is pointless here. Thus, we pass "true" to task_css_check()
7425 * to prevent lockdep warnings.
7427 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7428 struct task_group
, css
);
7429 tg
= autogroup_task_group(tsk
, tg
);
7430 tsk
->sched_task_group
= tg
;
7432 #ifdef CONFIG_FAIR_GROUP_SCHED
7433 if (tsk
->sched_class
->task_move_group
)
7434 tsk
->sched_class
->task_move_group(tsk
, queued
);
7437 set_task_rq(tsk
, task_cpu(tsk
));
7439 if (unlikely(running
))
7440 tsk
->sched_class
->set_curr_task(rq
);
7442 enqueue_task(rq
, tsk
, 0);
7444 task_rq_unlock(rq
, tsk
, &flags
);
7446 #endif /* CONFIG_CGROUP_SCHED */
7448 #ifdef CONFIG_RT_GROUP_SCHED
7450 * Ensure that the real time constraints are schedulable.
7452 static DEFINE_MUTEX(rt_constraints_mutex
);
7454 /* Must be called with tasklist_lock held */
7455 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7457 struct task_struct
*g
, *p
;
7459 for_each_process_thread(g
, p
) {
7460 if (rt_task(p
) && task_group(p
) == tg
)
7467 struct rt_schedulable_data
{
7468 struct task_group
*tg
;
7473 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7475 struct rt_schedulable_data
*d
= data
;
7476 struct task_group
*child
;
7477 unsigned long total
, sum
= 0;
7478 u64 period
, runtime
;
7480 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7481 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7484 period
= d
->rt_period
;
7485 runtime
= d
->rt_runtime
;
7489 * Cannot have more runtime than the period.
7491 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7495 * Ensure we don't starve existing RT tasks.
7497 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7500 total
= to_ratio(period
, runtime
);
7503 * Nobody can have more than the global setting allows.
7505 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7509 * The sum of our children's runtime should not exceed our own.
7511 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7512 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7513 runtime
= child
->rt_bandwidth
.rt_runtime
;
7515 if (child
== d
->tg
) {
7516 period
= d
->rt_period
;
7517 runtime
= d
->rt_runtime
;
7520 sum
+= to_ratio(period
, runtime
);
7529 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7533 struct rt_schedulable_data data
= {
7535 .rt_period
= period
,
7536 .rt_runtime
= runtime
,
7540 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7546 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7547 u64 rt_period
, u64 rt_runtime
)
7551 mutex_lock(&rt_constraints_mutex
);
7552 read_lock(&tasklist_lock
);
7553 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7557 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7558 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7559 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7561 for_each_possible_cpu(i
) {
7562 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7564 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7565 rt_rq
->rt_runtime
= rt_runtime
;
7566 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7568 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7570 read_unlock(&tasklist_lock
);
7571 mutex_unlock(&rt_constraints_mutex
);
7576 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7578 u64 rt_runtime
, rt_period
;
7580 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7581 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7582 if (rt_runtime_us
< 0)
7583 rt_runtime
= RUNTIME_INF
;
7585 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7588 static long sched_group_rt_runtime(struct task_group
*tg
)
7592 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7595 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7596 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7597 return rt_runtime_us
;
7600 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7602 u64 rt_runtime
, rt_period
;
7604 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7605 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7610 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7613 static long sched_group_rt_period(struct task_group
*tg
)
7617 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7618 do_div(rt_period_us
, NSEC_PER_USEC
);
7619 return rt_period_us
;
7621 #endif /* CONFIG_RT_GROUP_SCHED */
7623 #ifdef CONFIG_RT_GROUP_SCHED
7624 static int sched_rt_global_constraints(void)
7628 mutex_lock(&rt_constraints_mutex
);
7629 read_lock(&tasklist_lock
);
7630 ret
= __rt_schedulable(NULL
, 0, 0);
7631 read_unlock(&tasklist_lock
);
7632 mutex_unlock(&rt_constraints_mutex
);
7637 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7639 /* Don't accept realtime tasks when there is no way for them to run */
7640 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7646 #else /* !CONFIG_RT_GROUP_SCHED */
7647 static int sched_rt_global_constraints(void)
7649 unsigned long flags
;
7652 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7653 for_each_possible_cpu(i
) {
7654 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7656 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7657 rt_rq
->rt_runtime
= global_rt_runtime();
7658 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7660 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7664 #endif /* CONFIG_RT_GROUP_SCHED */
7666 static int sched_dl_global_constraints(void)
7668 u64 runtime
= global_rt_runtime();
7669 u64 period
= global_rt_period();
7670 u64 new_bw
= to_ratio(period
, runtime
);
7673 unsigned long flags
;
7676 * Here we want to check the bandwidth not being set to some
7677 * value smaller than the currently allocated bandwidth in
7678 * any of the root_domains.
7680 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7681 * cycling on root_domains... Discussion on different/better
7682 * solutions is welcome!
7684 for_each_possible_cpu(cpu
) {
7685 rcu_read_lock_sched();
7686 dl_b
= dl_bw_of(cpu
);
7688 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7689 if (new_bw
< dl_b
->total_bw
)
7691 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7693 rcu_read_unlock_sched();
7702 static void sched_dl_do_global(void)
7707 unsigned long flags
;
7709 def_dl_bandwidth
.dl_period
= global_rt_period();
7710 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7712 if (global_rt_runtime() != RUNTIME_INF
)
7713 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7716 * FIXME: As above...
7718 for_each_possible_cpu(cpu
) {
7719 rcu_read_lock_sched();
7720 dl_b
= dl_bw_of(cpu
);
7722 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7724 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7726 rcu_read_unlock_sched();
7730 static int sched_rt_global_validate(void)
7732 if (sysctl_sched_rt_period
<= 0)
7735 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7736 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7742 static void sched_rt_do_global(void)
7744 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7745 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7748 int sched_rt_handler(struct ctl_table
*table
, int write
,
7749 void __user
*buffer
, size_t *lenp
,
7752 int old_period
, old_runtime
;
7753 static DEFINE_MUTEX(mutex
);
7757 old_period
= sysctl_sched_rt_period
;
7758 old_runtime
= sysctl_sched_rt_runtime
;
7760 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7762 if (!ret
&& write
) {
7763 ret
= sched_rt_global_validate();
7767 ret
= sched_rt_global_constraints();
7771 ret
= sched_dl_global_constraints();
7775 sched_rt_do_global();
7776 sched_dl_do_global();
7780 sysctl_sched_rt_period
= old_period
;
7781 sysctl_sched_rt_runtime
= old_runtime
;
7783 mutex_unlock(&mutex
);
7788 int sched_rr_handler(struct ctl_table
*table
, int write
,
7789 void __user
*buffer
, size_t *lenp
,
7793 static DEFINE_MUTEX(mutex
);
7796 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7797 /* make sure that internally we keep jiffies */
7798 /* also, writing zero resets timeslice to default */
7799 if (!ret
&& write
) {
7800 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7801 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7803 mutex_unlock(&mutex
);
7807 #ifdef CONFIG_CGROUP_SCHED
7809 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7811 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7814 static struct cgroup_subsys_state
*
7815 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7817 struct task_group
*parent
= css_tg(parent_css
);
7818 struct task_group
*tg
;
7821 /* This is early initialization for the top cgroup */
7822 return &root_task_group
.css
;
7825 tg
= sched_create_group(parent
);
7827 return ERR_PTR(-ENOMEM
);
7832 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7834 struct task_group
*tg
= css_tg(css
);
7835 struct task_group
*parent
= css_tg(css
->parent
);
7838 sched_online_group(tg
, parent
);
7842 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7844 struct task_group
*tg
= css_tg(css
);
7846 sched_destroy_group(tg
);
7849 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7851 struct task_group
*tg
= css_tg(css
);
7853 sched_offline_group(tg
);
7856 static void cpu_cgroup_fork(struct task_struct
*task
)
7858 sched_move_task(task
);
7861 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7862 struct cgroup_taskset
*tset
)
7864 struct task_struct
*task
;
7866 cgroup_taskset_for_each(task
, tset
) {
7867 #ifdef CONFIG_RT_GROUP_SCHED
7868 if (!sched_rt_can_attach(css_tg(css
), task
))
7871 /* We don't support RT-tasks being in separate groups */
7872 if (task
->sched_class
!= &fair_sched_class
)
7879 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7880 struct cgroup_taskset
*tset
)
7882 struct task_struct
*task
;
7884 cgroup_taskset_for_each(task
, tset
)
7885 sched_move_task(task
);
7888 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7889 struct cgroup_subsys_state
*old_css
,
7890 struct task_struct
*task
)
7893 * cgroup_exit() is called in the copy_process() failure path.
7894 * Ignore this case since the task hasn't ran yet, this avoids
7895 * trying to poke a half freed task state from generic code.
7897 if (!(task
->flags
& PF_EXITING
))
7900 sched_move_task(task
);
7903 #ifdef CONFIG_FAIR_GROUP_SCHED
7904 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7905 struct cftype
*cftype
, u64 shareval
)
7907 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7910 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7913 struct task_group
*tg
= css_tg(css
);
7915 return (u64
) scale_load_down(tg
->shares
);
7918 #ifdef CONFIG_CFS_BANDWIDTH
7919 static DEFINE_MUTEX(cfs_constraints_mutex
);
7921 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7922 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7924 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7926 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7928 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7929 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7931 if (tg
== &root_task_group
)
7935 * Ensure we have at some amount of bandwidth every period. This is
7936 * to prevent reaching a state of large arrears when throttled via
7937 * entity_tick() resulting in prolonged exit starvation.
7939 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7943 * Likewise, bound things on the otherside by preventing insane quota
7944 * periods. This also allows us to normalize in computing quota
7947 if (period
> max_cfs_quota_period
)
7951 * Prevent race between setting of cfs_rq->runtime_enabled and
7952 * unthrottle_offline_cfs_rqs().
7955 mutex_lock(&cfs_constraints_mutex
);
7956 ret
= __cfs_schedulable(tg
, period
, quota
);
7960 runtime_enabled
= quota
!= RUNTIME_INF
;
7961 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7963 * If we need to toggle cfs_bandwidth_used, off->on must occur
7964 * before making related changes, and on->off must occur afterwards
7966 if (runtime_enabled
&& !runtime_was_enabled
)
7967 cfs_bandwidth_usage_inc();
7968 raw_spin_lock_irq(&cfs_b
->lock
);
7969 cfs_b
->period
= ns_to_ktime(period
);
7970 cfs_b
->quota
= quota
;
7972 __refill_cfs_bandwidth_runtime(cfs_b
);
7973 /* restart the period timer (if active) to handle new period expiry */
7974 if (runtime_enabled
&& cfs_b
->timer_active
) {
7975 /* force a reprogram */
7976 __start_cfs_bandwidth(cfs_b
, true);
7978 raw_spin_unlock_irq(&cfs_b
->lock
);
7980 for_each_online_cpu(i
) {
7981 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7982 struct rq
*rq
= cfs_rq
->rq
;
7984 raw_spin_lock_irq(&rq
->lock
);
7985 cfs_rq
->runtime_enabled
= runtime_enabled
;
7986 cfs_rq
->runtime_remaining
= 0;
7988 if (cfs_rq
->throttled
)
7989 unthrottle_cfs_rq(cfs_rq
);
7990 raw_spin_unlock_irq(&rq
->lock
);
7992 if (runtime_was_enabled
&& !runtime_enabled
)
7993 cfs_bandwidth_usage_dec();
7995 mutex_unlock(&cfs_constraints_mutex
);
8001 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8005 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8006 if (cfs_quota_us
< 0)
8007 quota
= RUNTIME_INF
;
8009 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8011 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8014 long tg_get_cfs_quota(struct task_group
*tg
)
8018 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8021 quota_us
= tg
->cfs_bandwidth
.quota
;
8022 do_div(quota_us
, NSEC_PER_USEC
);
8027 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8031 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8032 quota
= tg
->cfs_bandwidth
.quota
;
8034 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8037 long tg_get_cfs_period(struct task_group
*tg
)
8041 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8042 do_div(cfs_period_us
, NSEC_PER_USEC
);
8044 return cfs_period_us
;
8047 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8050 return tg_get_cfs_quota(css_tg(css
));
8053 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8054 struct cftype
*cftype
, s64 cfs_quota_us
)
8056 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8059 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8062 return tg_get_cfs_period(css_tg(css
));
8065 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8066 struct cftype
*cftype
, u64 cfs_period_us
)
8068 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8071 struct cfs_schedulable_data
{
8072 struct task_group
*tg
;
8077 * normalize group quota/period to be quota/max_period
8078 * note: units are usecs
8080 static u64
normalize_cfs_quota(struct task_group
*tg
,
8081 struct cfs_schedulable_data
*d
)
8089 period
= tg_get_cfs_period(tg
);
8090 quota
= tg_get_cfs_quota(tg
);
8093 /* note: these should typically be equivalent */
8094 if (quota
== RUNTIME_INF
|| quota
== -1)
8097 return to_ratio(period
, quota
);
8100 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8102 struct cfs_schedulable_data
*d
= data
;
8103 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8104 s64 quota
= 0, parent_quota
= -1;
8107 quota
= RUNTIME_INF
;
8109 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8111 quota
= normalize_cfs_quota(tg
, d
);
8112 parent_quota
= parent_b
->hierarchical_quota
;
8115 * ensure max(child_quota) <= parent_quota, inherit when no
8118 if (quota
== RUNTIME_INF
)
8119 quota
= parent_quota
;
8120 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8123 cfs_b
->hierarchical_quota
= quota
;
8128 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8131 struct cfs_schedulable_data data
= {
8137 if (quota
!= RUNTIME_INF
) {
8138 do_div(data
.period
, NSEC_PER_USEC
);
8139 do_div(data
.quota
, NSEC_PER_USEC
);
8143 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8149 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8151 struct task_group
*tg
= css_tg(seq_css(sf
));
8152 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8154 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8155 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8156 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8160 #endif /* CONFIG_CFS_BANDWIDTH */
8161 #endif /* CONFIG_FAIR_GROUP_SCHED */
8163 #ifdef CONFIG_RT_GROUP_SCHED
8164 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8165 struct cftype
*cft
, s64 val
)
8167 return sched_group_set_rt_runtime(css_tg(css
), val
);
8170 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8173 return sched_group_rt_runtime(css_tg(css
));
8176 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8177 struct cftype
*cftype
, u64 rt_period_us
)
8179 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8182 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8185 return sched_group_rt_period(css_tg(css
));
8187 #endif /* CONFIG_RT_GROUP_SCHED */
8189 static struct cftype cpu_files
[] = {
8190 #ifdef CONFIG_FAIR_GROUP_SCHED
8193 .read_u64
= cpu_shares_read_u64
,
8194 .write_u64
= cpu_shares_write_u64
,
8197 #ifdef CONFIG_CFS_BANDWIDTH
8199 .name
= "cfs_quota_us",
8200 .read_s64
= cpu_cfs_quota_read_s64
,
8201 .write_s64
= cpu_cfs_quota_write_s64
,
8204 .name
= "cfs_period_us",
8205 .read_u64
= cpu_cfs_period_read_u64
,
8206 .write_u64
= cpu_cfs_period_write_u64
,
8210 .seq_show
= cpu_stats_show
,
8213 #ifdef CONFIG_RT_GROUP_SCHED
8215 .name
= "rt_runtime_us",
8216 .read_s64
= cpu_rt_runtime_read
,
8217 .write_s64
= cpu_rt_runtime_write
,
8220 .name
= "rt_period_us",
8221 .read_u64
= cpu_rt_period_read_uint
,
8222 .write_u64
= cpu_rt_period_write_uint
,
8228 struct cgroup_subsys cpu_cgrp_subsys
= {
8229 .css_alloc
= cpu_cgroup_css_alloc
,
8230 .css_free
= cpu_cgroup_css_free
,
8231 .css_online
= cpu_cgroup_css_online
,
8232 .css_offline
= cpu_cgroup_css_offline
,
8233 .fork
= cpu_cgroup_fork
,
8234 .can_attach
= cpu_cgroup_can_attach
,
8235 .attach
= cpu_cgroup_attach
,
8236 .exit
= cpu_cgroup_exit
,
8237 .legacy_cftypes
= cpu_files
,
8241 #endif /* CONFIG_CGROUP_SCHED */
8243 void dump_cpu_task(int cpu
)
8245 pr_info("Task dump for CPU %d:\n", cpu
);
8246 sched_show_task(cpu_curr(cpu
));