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
;
1012 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
1014 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1015 const struct sched_class
*prev_class
,
1018 if (prev_class
!= p
->sched_class
) {
1019 if (prev_class
->switched_from
)
1020 prev_class
->switched_from(rq
, p
);
1021 /* Possble rq->lock 'hole'. */
1022 p
->sched_class
->switched_to(rq
, p
);
1023 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1024 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1027 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1029 const struct sched_class
*class;
1031 if (p
->sched_class
== rq
->curr
->sched_class
) {
1032 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1034 for_each_class(class) {
1035 if (class == rq
->curr
->sched_class
)
1037 if (class == p
->sched_class
) {
1045 * A queue event has occurred, and we're going to schedule. In
1046 * this case, we can save a useless back to back clock update.
1048 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1049 rq
->skip_clock_update
= 1;
1053 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1055 #ifdef CONFIG_SCHED_DEBUG
1057 * We should never call set_task_cpu() on a blocked task,
1058 * ttwu() will sort out the placement.
1060 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1063 #ifdef CONFIG_LOCKDEP
1065 * The caller should hold either p->pi_lock or rq->lock, when changing
1066 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1068 * sched_move_task() holds both and thus holding either pins the cgroup,
1071 * Furthermore, all task_rq users should acquire both locks, see
1074 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1075 lockdep_is_held(&task_rq(p
)->lock
)));
1079 trace_sched_migrate_task(p
, new_cpu
);
1081 if (task_cpu(p
) != new_cpu
) {
1082 if (p
->sched_class
->migrate_task_rq
)
1083 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1084 p
->se
.nr_migrations
++;
1085 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1088 __set_task_cpu(p
, new_cpu
);
1091 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1093 if (task_on_rq_queued(p
)) {
1094 struct rq
*src_rq
, *dst_rq
;
1096 src_rq
= task_rq(p
);
1097 dst_rq
= cpu_rq(cpu
);
1099 deactivate_task(src_rq
, p
, 0);
1100 set_task_cpu(p
, cpu
);
1101 activate_task(dst_rq
, p
, 0);
1102 check_preempt_curr(dst_rq
, p
, 0);
1105 * Task isn't running anymore; make it appear like we migrated
1106 * it before it went to sleep. This means on wakeup we make the
1107 * previous cpu our targer instead of where it really is.
1113 struct migration_swap_arg
{
1114 struct task_struct
*src_task
, *dst_task
;
1115 int src_cpu
, dst_cpu
;
1118 static int migrate_swap_stop(void *data
)
1120 struct migration_swap_arg
*arg
= data
;
1121 struct rq
*src_rq
, *dst_rq
;
1124 src_rq
= cpu_rq(arg
->src_cpu
);
1125 dst_rq
= cpu_rq(arg
->dst_cpu
);
1127 double_raw_lock(&arg
->src_task
->pi_lock
,
1128 &arg
->dst_task
->pi_lock
);
1129 double_rq_lock(src_rq
, dst_rq
);
1130 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1133 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1136 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1139 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1142 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1143 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1148 double_rq_unlock(src_rq
, dst_rq
);
1149 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1150 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1156 * Cross migrate two tasks
1158 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1160 struct migration_swap_arg arg
;
1163 arg
= (struct migration_swap_arg
){
1165 .src_cpu
= task_cpu(cur
),
1167 .dst_cpu
= task_cpu(p
),
1170 if (arg
.src_cpu
== arg
.dst_cpu
)
1174 * These three tests are all lockless; this is OK since all of them
1175 * will be re-checked with proper locks held further down the line.
1177 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1180 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1183 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1186 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1187 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1193 struct migration_arg
{
1194 struct task_struct
*task
;
1198 static int migration_cpu_stop(void *data
);
1201 * wait_task_inactive - wait for a thread to unschedule.
1203 * If @match_state is nonzero, it's the @p->state value just checked and
1204 * not expected to change. If it changes, i.e. @p might have woken up,
1205 * then return zero. When we succeed in waiting for @p to be off its CPU,
1206 * we return a positive number (its total switch count). If a second call
1207 * a short while later returns the same number, the caller can be sure that
1208 * @p has remained unscheduled the whole time.
1210 * The caller must ensure that the task *will* unschedule sometime soon,
1211 * else this function might spin for a *long* time. This function can't
1212 * be called with interrupts off, or it may introduce deadlock with
1213 * smp_call_function() if an IPI is sent by the same process we are
1214 * waiting to become inactive.
1216 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1218 unsigned long flags
;
1219 int running
, queued
;
1225 * We do the initial early heuristics without holding
1226 * any task-queue locks at all. We'll only try to get
1227 * the runqueue lock when things look like they will
1233 * If the task is actively running on another CPU
1234 * still, just relax and busy-wait without holding
1237 * NOTE! Since we don't hold any locks, it's not
1238 * even sure that "rq" stays as the right runqueue!
1239 * But we don't care, since "task_running()" will
1240 * return false if the runqueue has changed and p
1241 * is actually now running somewhere else!
1243 while (task_running(rq
, p
)) {
1244 if (match_state
&& unlikely(p
->state
!= match_state
))
1250 * Ok, time to look more closely! We need the rq
1251 * lock now, to be *sure*. If we're wrong, we'll
1252 * just go back and repeat.
1254 rq
= task_rq_lock(p
, &flags
);
1255 trace_sched_wait_task(p
);
1256 running
= task_running(rq
, p
);
1257 queued
= task_on_rq_queued(p
);
1259 if (!match_state
|| p
->state
== match_state
)
1260 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1261 task_rq_unlock(rq
, p
, &flags
);
1264 * If it changed from the expected state, bail out now.
1266 if (unlikely(!ncsw
))
1270 * Was it really running after all now that we
1271 * checked with the proper locks actually held?
1273 * Oops. Go back and try again..
1275 if (unlikely(running
)) {
1281 * It's not enough that it's not actively running,
1282 * it must be off the runqueue _entirely_, and not
1285 * So if it was still runnable (but just not actively
1286 * running right now), it's preempted, and we should
1287 * yield - it could be a while.
1289 if (unlikely(queued
)) {
1290 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1292 set_current_state(TASK_UNINTERRUPTIBLE
);
1293 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1298 * Ahh, all good. It wasn't running, and it wasn't
1299 * runnable, which means that it will never become
1300 * running in the future either. We're all done!
1309 * kick_process - kick a running thread to enter/exit the kernel
1310 * @p: the to-be-kicked thread
1312 * Cause a process which is running on another CPU to enter
1313 * kernel-mode, without any delay. (to get signals handled.)
1315 * NOTE: this function doesn't have to take the runqueue lock,
1316 * because all it wants to ensure is that the remote task enters
1317 * the kernel. If the IPI races and the task has been migrated
1318 * to another CPU then no harm is done and the purpose has been
1321 void kick_process(struct task_struct
*p
)
1327 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1328 smp_send_reschedule(cpu
);
1331 EXPORT_SYMBOL_GPL(kick_process
);
1332 #endif /* CONFIG_SMP */
1336 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1338 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1340 int nid
= cpu_to_node(cpu
);
1341 const struct cpumask
*nodemask
= NULL
;
1342 enum { cpuset
, possible
, fail
} state
= cpuset
;
1346 * If the node that the cpu is on has been offlined, cpu_to_node()
1347 * will return -1. There is no cpu on the node, and we should
1348 * select the cpu on the other node.
1351 nodemask
= cpumask_of_node(nid
);
1353 /* Look for allowed, online CPU in same node. */
1354 for_each_cpu(dest_cpu
, nodemask
) {
1355 if (!cpu_online(dest_cpu
))
1357 if (!cpu_active(dest_cpu
))
1359 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1365 /* Any allowed, online CPU? */
1366 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1367 if (!cpu_online(dest_cpu
))
1369 if (!cpu_active(dest_cpu
))
1376 /* No more Mr. Nice Guy. */
1377 cpuset_cpus_allowed_fallback(p
);
1382 do_set_cpus_allowed(p
, cpu_possible_mask
);
1393 if (state
!= cpuset
) {
1395 * Don't tell them about moving exiting tasks or
1396 * kernel threads (both mm NULL), since they never
1399 if (p
->mm
&& printk_ratelimit()) {
1400 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1401 task_pid_nr(p
), p
->comm
, cpu
);
1409 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1412 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1414 if (p
->nr_cpus_allowed
> 1)
1415 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1418 * In order not to call set_task_cpu() on a blocking task we need
1419 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1422 * Since this is common to all placement strategies, this lives here.
1424 * [ this allows ->select_task() to simply return task_cpu(p) and
1425 * not worry about this generic constraint ]
1427 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1429 cpu
= select_fallback_rq(task_cpu(p
), p
);
1434 static void update_avg(u64
*avg
, u64 sample
)
1436 s64 diff
= sample
- *avg
;
1442 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1444 #ifdef CONFIG_SCHEDSTATS
1445 struct rq
*rq
= this_rq();
1448 int this_cpu
= smp_processor_id();
1450 if (cpu
== this_cpu
) {
1451 schedstat_inc(rq
, ttwu_local
);
1452 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1454 struct sched_domain
*sd
;
1456 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1458 for_each_domain(this_cpu
, sd
) {
1459 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1460 schedstat_inc(sd
, ttwu_wake_remote
);
1467 if (wake_flags
& WF_MIGRATED
)
1468 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1470 #endif /* CONFIG_SMP */
1472 schedstat_inc(rq
, ttwu_count
);
1473 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1475 if (wake_flags
& WF_SYNC
)
1476 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1478 #endif /* CONFIG_SCHEDSTATS */
1481 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1483 activate_task(rq
, p
, en_flags
);
1484 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1486 /* if a worker is waking up, notify workqueue */
1487 if (p
->flags
& PF_WQ_WORKER
)
1488 wq_worker_waking_up(p
, cpu_of(rq
));
1492 * Mark the task runnable and perform wakeup-preemption.
1495 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1497 check_preempt_curr(rq
, p
, wake_flags
);
1498 trace_sched_wakeup(p
, true);
1500 p
->state
= TASK_RUNNING
;
1502 if (p
->sched_class
->task_woken
)
1503 p
->sched_class
->task_woken(rq
, p
);
1505 if (rq
->idle_stamp
) {
1506 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1507 u64 max
= 2*rq
->max_idle_balance_cost
;
1509 update_avg(&rq
->avg_idle
, delta
);
1511 if (rq
->avg_idle
> max
)
1520 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1523 if (p
->sched_contributes_to_load
)
1524 rq
->nr_uninterruptible
--;
1527 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1528 ttwu_do_wakeup(rq
, p
, wake_flags
);
1532 * Called in case the task @p isn't fully descheduled from its runqueue,
1533 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1534 * since all we need to do is flip p->state to TASK_RUNNING, since
1535 * the task is still ->on_rq.
1537 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1542 rq
= __task_rq_lock(p
);
1543 if (task_on_rq_queued(p
)) {
1544 /* check_preempt_curr() may use rq clock */
1545 update_rq_clock(rq
);
1546 ttwu_do_wakeup(rq
, p
, wake_flags
);
1549 __task_rq_unlock(rq
);
1555 void sched_ttwu_pending(void)
1557 struct rq
*rq
= this_rq();
1558 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1559 struct task_struct
*p
;
1560 unsigned long flags
;
1565 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1568 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1569 llist
= llist_next(llist
);
1570 ttwu_do_activate(rq
, p
, 0);
1573 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1576 void scheduler_ipi(void)
1579 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1580 * TIF_NEED_RESCHED remotely (for the first time) will also send
1583 preempt_fold_need_resched();
1585 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1589 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1590 * traditionally all their work was done from the interrupt return
1591 * path. Now that we actually do some work, we need to make sure
1594 * Some archs already do call them, luckily irq_enter/exit nest
1597 * Arguably we should visit all archs and update all handlers,
1598 * however a fair share of IPIs are still resched only so this would
1599 * somewhat pessimize the simple resched case.
1602 sched_ttwu_pending();
1605 * Check if someone kicked us for doing the nohz idle load balance.
1607 if (unlikely(got_nohz_idle_kick())) {
1608 this_rq()->idle_balance
= 1;
1609 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1614 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1616 struct rq
*rq
= cpu_rq(cpu
);
1618 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1619 if (!set_nr_if_polling(rq
->idle
))
1620 smp_send_reschedule(cpu
);
1622 trace_sched_wake_idle_without_ipi(cpu
);
1626 void wake_up_if_idle(int cpu
)
1628 struct rq
*rq
= cpu_rq(cpu
);
1629 unsigned long flags
;
1631 if (!is_idle_task(rq
->curr
))
1634 if (set_nr_if_polling(rq
->idle
)) {
1635 trace_sched_wake_idle_without_ipi(cpu
);
1637 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1638 if (is_idle_task(rq
->curr
))
1639 smp_send_reschedule(cpu
);
1640 /* Else cpu is not in idle, do nothing here */
1641 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1645 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1647 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1649 #endif /* CONFIG_SMP */
1651 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1653 struct rq
*rq
= cpu_rq(cpu
);
1655 #if defined(CONFIG_SMP)
1656 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1657 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1658 ttwu_queue_remote(p
, cpu
);
1663 raw_spin_lock(&rq
->lock
);
1664 ttwu_do_activate(rq
, p
, 0);
1665 raw_spin_unlock(&rq
->lock
);
1669 * try_to_wake_up - wake up a thread
1670 * @p: the thread to be awakened
1671 * @state: the mask of task states that can be woken
1672 * @wake_flags: wake modifier flags (WF_*)
1674 * Put it on the run-queue if it's not already there. The "current"
1675 * thread is always on the run-queue (except when the actual
1676 * re-schedule is in progress), and as such you're allowed to do
1677 * the simpler "current->state = TASK_RUNNING" to mark yourself
1678 * runnable without the overhead of this.
1680 * Return: %true if @p was woken up, %false if it was already running.
1681 * or @state didn't match @p's state.
1684 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1686 unsigned long flags
;
1687 int cpu
, success
= 0;
1690 * If we are going to wake up a thread waiting for CONDITION we
1691 * need to ensure that CONDITION=1 done by the caller can not be
1692 * reordered with p->state check below. This pairs with mb() in
1693 * set_current_state() the waiting thread does.
1695 smp_mb__before_spinlock();
1696 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1697 if (!(p
->state
& state
))
1700 success
= 1; /* we're going to change ->state */
1703 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1708 * If the owning (remote) cpu is still in the middle of schedule() with
1709 * this task as prev, wait until its done referencing the task.
1714 * Pairs with the smp_wmb() in finish_lock_switch().
1718 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1719 p
->state
= TASK_WAKING
;
1721 if (p
->sched_class
->task_waking
)
1722 p
->sched_class
->task_waking(p
);
1724 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1725 if (task_cpu(p
) != cpu
) {
1726 wake_flags
|= WF_MIGRATED
;
1727 set_task_cpu(p
, cpu
);
1729 #endif /* CONFIG_SMP */
1733 ttwu_stat(p
, cpu
, wake_flags
);
1735 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1741 * try_to_wake_up_local - try to wake up a local task with rq lock held
1742 * @p: the thread to be awakened
1744 * Put @p on the run-queue if it's not already there. The caller must
1745 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1748 static void try_to_wake_up_local(struct task_struct
*p
)
1750 struct rq
*rq
= task_rq(p
);
1752 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1753 WARN_ON_ONCE(p
== current
))
1756 lockdep_assert_held(&rq
->lock
);
1758 if (!raw_spin_trylock(&p
->pi_lock
)) {
1759 raw_spin_unlock(&rq
->lock
);
1760 raw_spin_lock(&p
->pi_lock
);
1761 raw_spin_lock(&rq
->lock
);
1764 if (!(p
->state
& TASK_NORMAL
))
1767 if (!task_on_rq_queued(p
))
1768 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1770 ttwu_do_wakeup(rq
, p
, 0);
1771 ttwu_stat(p
, smp_processor_id(), 0);
1773 raw_spin_unlock(&p
->pi_lock
);
1777 * wake_up_process - Wake up a specific process
1778 * @p: The process to be woken up.
1780 * Attempt to wake up the nominated process and move it to the set of runnable
1783 * Return: 1 if the process was woken up, 0 if it was already running.
1785 * It may be assumed that this function implies a write memory barrier before
1786 * changing the task state if and only if any tasks are woken up.
1788 int wake_up_process(struct task_struct
*p
)
1790 WARN_ON(task_is_stopped_or_traced(p
));
1791 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1793 EXPORT_SYMBOL(wake_up_process
);
1795 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1797 return try_to_wake_up(p
, state
, 0);
1801 * This function clears the sched_dl_entity static params.
1803 void __dl_clear_params(struct task_struct
*p
)
1805 struct sched_dl_entity
*dl_se
= &p
->dl
;
1807 dl_se
->dl_runtime
= 0;
1808 dl_se
->dl_deadline
= 0;
1809 dl_se
->dl_period
= 0;
1815 * Perform scheduler related setup for a newly forked process p.
1816 * p is forked by current.
1818 * __sched_fork() is basic setup used by init_idle() too:
1820 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1825 p
->se
.exec_start
= 0;
1826 p
->se
.sum_exec_runtime
= 0;
1827 p
->se
.prev_sum_exec_runtime
= 0;
1828 p
->se
.nr_migrations
= 0;
1830 INIT_LIST_HEAD(&p
->se
.group_node
);
1832 #ifdef CONFIG_SCHEDSTATS
1833 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1836 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1837 hrtimer_init(&p
->dl
.dl_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1838 __dl_clear_params(p
);
1840 INIT_LIST_HEAD(&p
->rt
.run_list
);
1842 #ifdef CONFIG_PREEMPT_NOTIFIERS
1843 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1846 #ifdef CONFIG_NUMA_BALANCING
1847 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1848 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1849 p
->mm
->numa_scan_seq
= 0;
1852 if (clone_flags
& CLONE_VM
)
1853 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1855 p
->numa_preferred_nid
= -1;
1857 p
->node_stamp
= 0ULL;
1858 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1859 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1860 p
->numa_work
.next
= &p
->numa_work
;
1861 p
->numa_faults
= NULL
;
1862 p
->last_task_numa_placement
= 0;
1863 p
->last_sum_exec_runtime
= 0;
1865 p
->numa_group
= NULL
;
1866 #endif /* CONFIG_NUMA_BALANCING */
1869 #ifdef CONFIG_NUMA_BALANCING
1870 #ifdef CONFIG_SCHED_DEBUG
1871 void set_numabalancing_state(bool enabled
)
1874 sched_feat_set("NUMA");
1876 sched_feat_set("NO_NUMA");
1879 __read_mostly
bool numabalancing_enabled
;
1881 void set_numabalancing_state(bool enabled
)
1883 numabalancing_enabled
= enabled
;
1885 #endif /* CONFIG_SCHED_DEBUG */
1887 #ifdef CONFIG_PROC_SYSCTL
1888 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1889 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1893 int state
= numabalancing_enabled
;
1895 if (write
&& !capable(CAP_SYS_ADMIN
))
1900 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1904 set_numabalancing_state(state
);
1911 * fork()/clone()-time setup:
1913 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1915 unsigned long flags
;
1916 int cpu
= get_cpu();
1918 __sched_fork(clone_flags
, p
);
1920 * We mark the process as running here. This guarantees that
1921 * nobody will actually run it, and a signal or other external
1922 * event cannot wake it up and insert it on the runqueue either.
1924 p
->state
= TASK_RUNNING
;
1927 * Make sure we do not leak PI boosting priority to the child.
1929 p
->prio
= current
->normal_prio
;
1932 * Revert to default priority/policy on fork if requested.
1934 if (unlikely(p
->sched_reset_on_fork
)) {
1935 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1936 p
->policy
= SCHED_NORMAL
;
1937 p
->static_prio
= NICE_TO_PRIO(0);
1939 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1940 p
->static_prio
= NICE_TO_PRIO(0);
1942 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1946 * We don't need the reset flag anymore after the fork. It has
1947 * fulfilled its duty:
1949 p
->sched_reset_on_fork
= 0;
1952 if (dl_prio(p
->prio
)) {
1955 } else if (rt_prio(p
->prio
)) {
1956 p
->sched_class
= &rt_sched_class
;
1958 p
->sched_class
= &fair_sched_class
;
1961 if (p
->sched_class
->task_fork
)
1962 p
->sched_class
->task_fork(p
);
1965 * The child is not yet in the pid-hash so no cgroup attach races,
1966 * and the cgroup is pinned to this child due to cgroup_fork()
1967 * is ran before sched_fork().
1969 * Silence PROVE_RCU.
1971 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1972 set_task_cpu(p
, cpu
);
1973 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1975 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1976 if (likely(sched_info_on()))
1977 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1979 #if defined(CONFIG_SMP)
1982 init_task_preempt_count(p
);
1984 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1985 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1992 unsigned long to_ratio(u64 period
, u64 runtime
)
1994 if (runtime
== RUNTIME_INF
)
1998 * Doing this here saves a lot of checks in all
1999 * the calling paths, and returning zero seems
2000 * safe for them anyway.
2005 return div64_u64(runtime
<< 20, period
);
2009 inline struct dl_bw
*dl_bw_of(int i
)
2011 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2012 "sched RCU must be held");
2013 return &cpu_rq(i
)->rd
->dl_bw
;
2016 static inline int dl_bw_cpus(int i
)
2018 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2021 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2022 "sched RCU must be held");
2023 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2029 inline struct dl_bw
*dl_bw_of(int i
)
2031 return &cpu_rq(i
)->dl
.dl_bw
;
2034 static inline int dl_bw_cpus(int i
)
2041 * We must be sure that accepting a new task (or allowing changing the
2042 * parameters of an existing one) is consistent with the bandwidth
2043 * constraints. If yes, this function also accordingly updates the currently
2044 * allocated bandwidth to reflect the new situation.
2046 * This function is called while holding p's rq->lock.
2048 static int dl_overflow(struct task_struct
*p
, int policy
,
2049 const struct sched_attr
*attr
)
2052 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2053 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2054 u64 runtime
= attr
->sched_runtime
;
2055 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2058 if (new_bw
== p
->dl
.dl_bw
)
2062 * Either if a task, enters, leave, or stays -deadline but changes
2063 * its parameters, we may need to update accordingly the total
2064 * allocated bandwidth of the container.
2066 raw_spin_lock(&dl_b
->lock
);
2067 cpus
= dl_bw_cpus(task_cpu(p
));
2068 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2069 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2070 __dl_add(dl_b
, new_bw
);
2072 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2073 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2074 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2075 __dl_add(dl_b
, new_bw
);
2077 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2078 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2081 raw_spin_unlock(&dl_b
->lock
);
2086 extern void init_dl_bw(struct dl_bw
*dl_b
);
2089 * wake_up_new_task - wake up a newly created task for the first time.
2091 * This function will do some initial scheduler statistics housekeeping
2092 * that must be done for every newly created context, then puts the task
2093 * on the runqueue and wakes it.
2095 void wake_up_new_task(struct task_struct
*p
)
2097 unsigned long flags
;
2100 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2103 * Fork balancing, do it here and not earlier because:
2104 * - cpus_allowed can change in the fork path
2105 * - any previously selected cpu might disappear through hotplug
2107 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2110 /* Initialize new task's runnable average */
2111 init_task_runnable_average(p
);
2112 rq
= __task_rq_lock(p
);
2113 activate_task(rq
, p
, 0);
2114 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2115 trace_sched_wakeup_new(p
, true);
2116 check_preempt_curr(rq
, p
, WF_FORK
);
2118 if (p
->sched_class
->task_woken
)
2119 p
->sched_class
->task_woken(rq
, p
);
2121 task_rq_unlock(rq
, p
, &flags
);
2124 #ifdef CONFIG_PREEMPT_NOTIFIERS
2127 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2128 * @notifier: notifier struct to register
2130 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2132 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2134 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2137 * preempt_notifier_unregister - no longer interested in preemption notifications
2138 * @notifier: notifier struct to unregister
2140 * This is safe to call from within a preemption notifier.
2142 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2144 hlist_del(¬ifier
->link
);
2146 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2148 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2150 struct preempt_notifier
*notifier
;
2152 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2153 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2157 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2158 struct task_struct
*next
)
2160 struct preempt_notifier
*notifier
;
2162 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2163 notifier
->ops
->sched_out(notifier
, next
);
2166 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2168 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2173 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2174 struct task_struct
*next
)
2178 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2181 * prepare_task_switch - prepare to switch tasks
2182 * @rq: the runqueue preparing to switch
2183 * @prev: the current task that is being switched out
2184 * @next: the task we are going to switch to.
2186 * This is called with the rq lock held and interrupts off. It must
2187 * be paired with a subsequent finish_task_switch after the context
2190 * prepare_task_switch sets up locking and calls architecture specific
2194 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2195 struct task_struct
*next
)
2197 trace_sched_switch(prev
, next
);
2198 sched_info_switch(rq
, prev
, next
);
2199 perf_event_task_sched_out(prev
, next
);
2200 fire_sched_out_preempt_notifiers(prev
, next
);
2201 prepare_lock_switch(rq
, next
);
2202 prepare_arch_switch(next
);
2206 * finish_task_switch - clean up after a task-switch
2207 * @prev: the thread we just switched away from.
2209 * finish_task_switch must be called after the context switch, paired
2210 * with a prepare_task_switch call before the context switch.
2211 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2212 * and do any other architecture-specific cleanup actions.
2214 * Note that we may have delayed dropping an mm in context_switch(). If
2215 * so, we finish that here outside of the runqueue lock. (Doing it
2216 * with the lock held can cause deadlocks; see schedule() for
2219 * The context switch have flipped the stack from under us and restored the
2220 * local variables which were saved when this task called schedule() in the
2221 * past. prev == current is still correct but we need to recalculate this_rq
2222 * because prev may have moved to another CPU.
2224 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2225 __releases(rq
->lock
)
2227 struct rq
*rq
= this_rq();
2228 struct mm_struct
*mm
= rq
->prev_mm
;
2234 * A task struct has one reference for the use as "current".
2235 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2236 * schedule one last time. The schedule call will never return, and
2237 * the scheduled task must drop that reference.
2238 * The test for TASK_DEAD must occur while the runqueue locks are
2239 * still held, otherwise prev could be scheduled on another cpu, die
2240 * there before we look at prev->state, and then the reference would
2242 * Manfred Spraul <manfred@colorfullife.com>
2244 prev_state
= prev
->state
;
2245 vtime_task_switch(prev
);
2246 finish_arch_switch(prev
);
2247 perf_event_task_sched_in(prev
, current
);
2248 finish_lock_switch(rq
, prev
);
2249 finish_arch_post_lock_switch();
2251 fire_sched_in_preempt_notifiers(current
);
2254 if (unlikely(prev_state
== TASK_DEAD
)) {
2255 if (prev
->sched_class
->task_dead
)
2256 prev
->sched_class
->task_dead(prev
);
2259 * Remove function-return probe instances associated with this
2260 * task and put them back on the free list.
2262 kprobe_flush_task(prev
);
2263 put_task_struct(prev
);
2266 tick_nohz_task_switch(current
);
2272 /* rq->lock is NOT held, but preemption is disabled */
2273 static inline void post_schedule(struct rq
*rq
)
2275 if (rq
->post_schedule
) {
2276 unsigned long flags
;
2278 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2279 if (rq
->curr
->sched_class
->post_schedule
)
2280 rq
->curr
->sched_class
->post_schedule(rq
);
2281 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2283 rq
->post_schedule
= 0;
2289 static inline void post_schedule(struct rq
*rq
)
2296 * schedule_tail - first thing a freshly forked thread must call.
2297 * @prev: the thread we just switched away from.
2299 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2300 __releases(rq
->lock
)
2304 /* finish_task_switch() drops rq->lock and enables preemtion */
2306 rq
= finish_task_switch(prev
);
2310 if (current
->set_child_tid
)
2311 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2315 * context_switch - switch to the new MM and the new thread's register state.
2317 static inline struct rq
*
2318 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2319 struct task_struct
*next
)
2321 struct mm_struct
*mm
, *oldmm
;
2323 prepare_task_switch(rq
, prev
, next
);
2326 oldmm
= prev
->active_mm
;
2328 * For paravirt, this is coupled with an exit in switch_to to
2329 * combine the page table reload and the switch backend into
2332 arch_start_context_switch(prev
);
2335 next
->active_mm
= oldmm
;
2336 atomic_inc(&oldmm
->mm_count
);
2337 enter_lazy_tlb(oldmm
, next
);
2339 switch_mm(oldmm
, mm
, next
);
2342 prev
->active_mm
= NULL
;
2343 rq
->prev_mm
= oldmm
;
2346 * Since the runqueue lock will be released by the next
2347 * task (which is an invalid locking op but in the case
2348 * of the scheduler it's an obvious special-case), so we
2349 * do an early lockdep release here:
2351 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2353 context_tracking_task_switch(prev
, next
);
2354 /* Here we just switch the register state and the stack. */
2355 switch_to(prev
, next
, prev
);
2358 return finish_task_switch(prev
);
2362 * nr_running and nr_context_switches:
2364 * externally visible scheduler statistics: current number of runnable
2365 * threads, total number of context switches performed since bootup.
2367 unsigned long nr_running(void)
2369 unsigned long i
, sum
= 0;
2371 for_each_online_cpu(i
)
2372 sum
+= cpu_rq(i
)->nr_running
;
2378 * Check if only the current task is running on the cpu.
2380 bool single_task_running(void)
2382 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2387 EXPORT_SYMBOL(single_task_running
);
2389 unsigned long long nr_context_switches(void)
2392 unsigned long long sum
= 0;
2394 for_each_possible_cpu(i
)
2395 sum
+= cpu_rq(i
)->nr_switches
;
2400 unsigned long nr_iowait(void)
2402 unsigned long i
, sum
= 0;
2404 for_each_possible_cpu(i
)
2405 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2410 unsigned long nr_iowait_cpu(int cpu
)
2412 struct rq
*this = cpu_rq(cpu
);
2413 return atomic_read(&this->nr_iowait
);
2416 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2418 struct rq
*this = this_rq();
2419 *nr_waiters
= atomic_read(&this->nr_iowait
);
2420 *load
= this->cpu_load
[0];
2426 * sched_exec - execve() is a valuable balancing opportunity, because at
2427 * this point the task has the smallest effective memory and cache footprint.
2429 void sched_exec(void)
2431 struct task_struct
*p
= current
;
2432 unsigned long flags
;
2435 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2436 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2437 if (dest_cpu
== smp_processor_id())
2440 if (likely(cpu_active(dest_cpu
))) {
2441 struct migration_arg arg
= { p
, dest_cpu
};
2443 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2444 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2448 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2453 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2454 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2456 EXPORT_PER_CPU_SYMBOL(kstat
);
2457 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2460 * Return accounted runtime for the task.
2461 * In case the task is currently running, return the runtime plus current's
2462 * pending runtime that have not been accounted yet.
2464 unsigned long long task_sched_runtime(struct task_struct
*p
)
2466 unsigned long flags
;
2470 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2472 * 64-bit doesn't need locks to atomically read a 64bit value.
2473 * So we have a optimization chance when the task's delta_exec is 0.
2474 * Reading ->on_cpu is racy, but this is ok.
2476 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2477 * If we race with it entering cpu, unaccounted time is 0. This is
2478 * indistinguishable from the read occurring a few cycles earlier.
2479 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2480 * been accounted, so we're correct here as well.
2482 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2483 return p
->se
.sum_exec_runtime
;
2486 rq
= task_rq_lock(p
, &flags
);
2488 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2489 * project cycles that may never be accounted to this
2490 * thread, breaking clock_gettime().
2492 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2493 update_rq_clock(rq
);
2494 p
->sched_class
->update_curr(rq
);
2496 ns
= p
->se
.sum_exec_runtime
;
2497 task_rq_unlock(rq
, p
, &flags
);
2503 * This function gets called by the timer code, with HZ frequency.
2504 * We call it with interrupts disabled.
2506 void scheduler_tick(void)
2508 int cpu
= smp_processor_id();
2509 struct rq
*rq
= cpu_rq(cpu
);
2510 struct task_struct
*curr
= rq
->curr
;
2514 raw_spin_lock(&rq
->lock
);
2515 update_rq_clock(rq
);
2516 curr
->sched_class
->task_tick(rq
, curr
, 0);
2517 update_cpu_load_active(rq
);
2518 raw_spin_unlock(&rq
->lock
);
2520 perf_event_task_tick();
2523 rq
->idle_balance
= idle_cpu(cpu
);
2524 trigger_load_balance(rq
);
2526 rq_last_tick_reset(rq
);
2529 #ifdef CONFIG_NO_HZ_FULL
2531 * scheduler_tick_max_deferment
2533 * Keep at least one tick per second when a single
2534 * active task is running because the scheduler doesn't
2535 * yet completely support full dynticks environment.
2537 * This makes sure that uptime, CFS vruntime, load
2538 * balancing, etc... continue to move forward, even
2539 * with a very low granularity.
2541 * Return: Maximum deferment in nanoseconds.
2543 u64
scheduler_tick_max_deferment(void)
2545 struct rq
*rq
= this_rq();
2546 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2548 next
= rq
->last_sched_tick
+ HZ
;
2550 if (time_before_eq(next
, now
))
2553 return jiffies_to_nsecs(next
- now
);
2557 notrace
unsigned long get_parent_ip(unsigned long addr
)
2559 if (in_lock_functions(addr
)) {
2560 addr
= CALLER_ADDR2
;
2561 if (in_lock_functions(addr
))
2562 addr
= CALLER_ADDR3
;
2567 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2568 defined(CONFIG_PREEMPT_TRACER))
2570 void preempt_count_add(int val
)
2572 #ifdef CONFIG_DEBUG_PREEMPT
2576 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2579 __preempt_count_add(val
);
2580 #ifdef CONFIG_DEBUG_PREEMPT
2582 * Spinlock count overflowing soon?
2584 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2587 if (preempt_count() == val
) {
2588 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2589 #ifdef CONFIG_DEBUG_PREEMPT
2590 current
->preempt_disable_ip
= ip
;
2592 trace_preempt_off(CALLER_ADDR0
, ip
);
2595 EXPORT_SYMBOL(preempt_count_add
);
2596 NOKPROBE_SYMBOL(preempt_count_add
);
2598 void preempt_count_sub(int val
)
2600 #ifdef CONFIG_DEBUG_PREEMPT
2604 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2607 * Is the spinlock portion underflowing?
2609 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2610 !(preempt_count() & PREEMPT_MASK
)))
2614 if (preempt_count() == val
)
2615 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2616 __preempt_count_sub(val
);
2618 EXPORT_SYMBOL(preempt_count_sub
);
2619 NOKPROBE_SYMBOL(preempt_count_sub
);
2624 * Print scheduling while atomic bug:
2626 static noinline
void __schedule_bug(struct task_struct
*prev
)
2628 if (oops_in_progress
)
2631 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2632 prev
->comm
, prev
->pid
, preempt_count());
2634 debug_show_held_locks(prev
);
2636 if (irqs_disabled())
2637 print_irqtrace_events(prev
);
2638 #ifdef CONFIG_DEBUG_PREEMPT
2639 if (in_atomic_preempt_off()) {
2640 pr_err("Preemption disabled at:");
2641 print_ip_sym(current
->preempt_disable_ip
);
2646 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2650 * Various schedule()-time debugging checks and statistics:
2652 static inline void schedule_debug(struct task_struct
*prev
)
2654 #ifdef CONFIG_SCHED_STACK_END_CHECK
2655 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2658 * Test if we are atomic. Since do_exit() needs to call into
2659 * schedule() atomically, we ignore that path. Otherwise whine
2660 * if we are scheduling when we should not.
2662 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2663 __schedule_bug(prev
);
2666 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2668 schedstat_inc(this_rq(), sched_count
);
2672 * Pick up the highest-prio task:
2674 static inline struct task_struct
*
2675 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2677 const struct sched_class
*class = &fair_sched_class
;
2678 struct task_struct
*p
;
2681 * Optimization: we know that if all tasks are in
2682 * the fair class we can call that function directly:
2684 if (likely(prev
->sched_class
== class &&
2685 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2686 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2687 if (unlikely(p
== RETRY_TASK
))
2690 /* assumes fair_sched_class->next == idle_sched_class */
2692 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2698 for_each_class(class) {
2699 p
= class->pick_next_task(rq
, prev
);
2701 if (unlikely(p
== RETRY_TASK
))
2707 BUG(); /* the idle class will always have a runnable task */
2711 * __schedule() is the main scheduler function.
2713 * The main means of driving the scheduler and thus entering this function are:
2715 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2717 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2718 * paths. For example, see arch/x86/entry_64.S.
2720 * To drive preemption between tasks, the scheduler sets the flag in timer
2721 * interrupt handler scheduler_tick().
2723 * 3. Wakeups don't really cause entry into schedule(). They add a
2724 * task to the run-queue and that's it.
2726 * Now, if the new task added to the run-queue preempts the current
2727 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2728 * called on the nearest possible occasion:
2730 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2732 * - in syscall or exception context, at the next outmost
2733 * preempt_enable(). (this might be as soon as the wake_up()'s
2736 * - in IRQ context, return from interrupt-handler to
2737 * preemptible context
2739 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2742 * - cond_resched() call
2743 * - explicit schedule() call
2744 * - return from syscall or exception to user-space
2745 * - return from interrupt-handler to user-space
2747 static void __sched
__schedule(void)
2749 struct task_struct
*prev
, *next
;
2750 unsigned long *switch_count
;
2756 cpu
= smp_processor_id();
2758 rcu_note_context_switch(cpu
);
2761 schedule_debug(prev
);
2763 if (sched_feat(HRTICK
))
2767 * Make sure that signal_pending_state()->signal_pending() below
2768 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2769 * done by the caller to avoid the race with signal_wake_up().
2771 smp_mb__before_spinlock();
2772 raw_spin_lock_irq(&rq
->lock
);
2774 switch_count
= &prev
->nivcsw
;
2775 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2776 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2777 prev
->state
= TASK_RUNNING
;
2779 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2783 * If a worker went to sleep, notify and ask workqueue
2784 * whether it wants to wake up a task to maintain
2787 if (prev
->flags
& PF_WQ_WORKER
) {
2788 struct task_struct
*to_wakeup
;
2790 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2792 try_to_wake_up_local(to_wakeup
);
2795 switch_count
= &prev
->nvcsw
;
2798 if (task_on_rq_queued(prev
) || rq
->skip_clock_update
< 0)
2799 update_rq_clock(rq
);
2801 next
= pick_next_task(rq
, prev
);
2802 clear_tsk_need_resched(prev
);
2803 clear_preempt_need_resched();
2804 rq
->skip_clock_update
= 0;
2806 if (likely(prev
!= next
)) {
2811 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
2814 raw_spin_unlock_irq(&rq
->lock
);
2818 sched_preempt_enable_no_resched();
2823 static inline void sched_submit_work(struct task_struct
*tsk
)
2825 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2828 * If we are going to sleep and we have plugged IO queued,
2829 * make sure to submit it to avoid deadlocks.
2831 if (blk_needs_flush_plug(tsk
))
2832 blk_schedule_flush_plug(tsk
);
2835 asmlinkage __visible
void __sched
schedule(void)
2837 struct task_struct
*tsk
= current
;
2839 sched_submit_work(tsk
);
2842 EXPORT_SYMBOL(schedule
);
2844 #ifdef CONFIG_CONTEXT_TRACKING
2845 asmlinkage __visible
void __sched
schedule_user(void)
2848 * If we come here after a random call to set_need_resched(),
2849 * or we have been woken up remotely but the IPI has not yet arrived,
2850 * we haven't yet exited the RCU idle mode. Do it here manually until
2851 * we find a better solution.
2860 * schedule_preempt_disabled - called with preemption disabled
2862 * Returns with preemption disabled. Note: preempt_count must be 1
2864 void __sched
schedule_preempt_disabled(void)
2866 sched_preempt_enable_no_resched();
2871 #ifdef CONFIG_PREEMPT
2873 * this is the entry point to schedule() from in-kernel preemption
2874 * off of preempt_enable. Kernel preemptions off return from interrupt
2875 * occur there and call schedule directly.
2877 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2880 * If there is a non-zero preempt_count or interrupts are disabled,
2881 * we do not want to preempt the current task. Just return..
2883 if (likely(!preemptible()))
2887 __preempt_count_add(PREEMPT_ACTIVE
);
2889 __preempt_count_sub(PREEMPT_ACTIVE
);
2892 * Check again in case we missed a preemption opportunity
2893 * between schedule and now.
2896 } while (need_resched());
2898 NOKPROBE_SYMBOL(preempt_schedule
);
2899 EXPORT_SYMBOL(preempt_schedule
);
2901 #ifdef CONFIG_CONTEXT_TRACKING
2903 * preempt_schedule_context - preempt_schedule called by tracing
2905 * The tracing infrastructure uses preempt_enable_notrace to prevent
2906 * recursion and tracing preempt enabling caused by the tracing
2907 * infrastructure itself. But as tracing can happen in areas coming
2908 * from userspace or just about to enter userspace, a preempt enable
2909 * can occur before user_exit() is called. This will cause the scheduler
2910 * to be called when the system is still in usermode.
2912 * To prevent this, the preempt_enable_notrace will use this function
2913 * instead of preempt_schedule() to exit user context if needed before
2914 * calling the scheduler.
2916 asmlinkage __visible
void __sched notrace
preempt_schedule_context(void)
2918 enum ctx_state prev_ctx
;
2920 if (likely(!preemptible()))
2924 __preempt_count_add(PREEMPT_ACTIVE
);
2926 * Needs preempt disabled in case user_exit() is traced
2927 * and the tracer calls preempt_enable_notrace() causing
2928 * an infinite recursion.
2930 prev_ctx
= exception_enter();
2932 exception_exit(prev_ctx
);
2934 __preempt_count_sub(PREEMPT_ACTIVE
);
2936 } while (need_resched());
2938 EXPORT_SYMBOL_GPL(preempt_schedule_context
);
2939 #endif /* CONFIG_CONTEXT_TRACKING */
2941 #endif /* CONFIG_PREEMPT */
2944 * this is the entry point to schedule() from kernel preemption
2945 * off of irq context.
2946 * Note, that this is called and return with irqs disabled. This will
2947 * protect us against recursive calling from irq.
2949 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2951 enum ctx_state prev_state
;
2953 /* Catch callers which need to be fixed */
2954 BUG_ON(preempt_count() || !irqs_disabled());
2956 prev_state
= exception_enter();
2959 __preempt_count_add(PREEMPT_ACTIVE
);
2962 local_irq_disable();
2963 __preempt_count_sub(PREEMPT_ACTIVE
);
2966 * Check again in case we missed a preemption opportunity
2967 * between schedule and now.
2970 } while (need_resched());
2972 exception_exit(prev_state
);
2975 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2978 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2980 EXPORT_SYMBOL(default_wake_function
);
2982 #ifdef CONFIG_RT_MUTEXES
2985 * rt_mutex_setprio - set the current priority of a task
2987 * @prio: prio value (kernel-internal form)
2989 * This function changes the 'effective' priority of a task. It does
2990 * not touch ->normal_prio like __setscheduler().
2992 * Used by the rt_mutex code to implement priority inheritance
2993 * logic. Call site only calls if the priority of the task changed.
2995 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2997 int oldprio
, queued
, running
, enqueue_flag
= 0;
2999 const struct sched_class
*prev_class
;
3001 BUG_ON(prio
> MAX_PRIO
);
3003 rq
= __task_rq_lock(p
);
3006 * Idle task boosting is a nono in general. There is one
3007 * exception, when PREEMPT_RT and NOHZ is active:
3009 * The idle task calls get_next_timer_interrupt() and holds
3010 * the timer wheel base->lock on the CPU and another CPU wants
3011 * to access the timer (probably to cancel it). We can safely
3012 * ignore the boosting request, as the idle CPU runs this code
3013 * with interrupts disabled and will complete the lock
3014 * protected section without being interrupted. So there is no
3015 * real need to boost.
3017 if (unlikely(p
== rq
->idle
)) {
3018 WARN_ON(p
!= rq
->curr
);
3019 WARN_ON(p
->pi_blocked_on
);
3023 trace_sched_pi_setprio(p
, prio
);
3025 prev_class
= p
->sched_class
;
3026 queued
= task_on_rq_queued(p
);
3027 running
= task_current(rq
, p
);
3029 dequeue_task(rq
, p
, 0);
3031 put_prev_task(rq
, p
);
3034 * Boosting condition are:
3035 * 1. -rt task is running and holds mutex A
3036 * --> -dl task blocks on mutex A
3038 * 2. -dl task is running and holds mutex A
3039 * --> -dl task blocks on mutex A and could preempt the
3042 if (dl_prio(prio
)) {
3043 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3044 if (!dl_prio(p
->normal_prio
) ||
3045 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3046 p
->dl
.dl_boosted
= 1;
3047 p
->dl
.dl_throttled
= 0;
3048 enqueue_flag
= ENQUEUE_REPLENISH
;
3050 p
->dl
.dl_boosted
= 0;
3051 p
->sched_class
= &dl_sched_class
;
3052 } else if (rt_prio(prio
)) {
3053 if (dl_prio(oldprio
))
3054 p
->dl
.dl_boosted
= 0;
3056 enqueue_flag
= ENQUEUE_HEAD
;
3057 p
->sched_class
= &rt_sched_class
;
3059 if (dl_prio(oldprio
))
3060 p
->dl
.dl_boosted
= 0;
3061 p
->sched_class
= &fair_sched_class
;
3067 p
->sched_class
->set_curr_task(rq
);
3069 enqueue_task(rq
, p
, enqueue_flag
);
3071 check_class_changed(rq
, p
, prev_class
, oldprio
);
3073 __task_rq_unlock(rq
);
3077 void set_user_nice(struct task_struct
*p
, long nice
)
3079 int old_prio
, delta
, queued
;
3080 unsigned long flags
;
3083 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3086 * We have to be careful, if called from sys_setpriority(),
3087 * the task might be in the middle of scheduling on another CPU.
3089 rq
= task_rq_lock(p
, &flags
);
3091 * The RT priorities are set via sched_setscheduler(), but we still
3092 * allow the 'normal' nice value to be set - but as expected
3093 * it wont have any effect on scheduling until the task is
3094 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3096 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3097 p
->static_prio
= NICE_TO_PRIO(nice
);
3100 queued
= task_on_rq_queued(p
);
3102 dequeue_task(rq
, p
, 0);
3104 p
->static_prio
= NICE_TO_PRIO(nice
);
3107 p
->prio
= effective_prio(p
);
3108 delta
= p
->prio
- old_prio
;
3111 enqueue_task(rq
, p
, 0);
3113 * If the task increased its priority or is running and
3114 * lowered its priority, then reschedule its CPU:
3116 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3120 task_rq_unlock(rq
, p
, &flags
);
3122 EXPORT_SYMBOL(set_user_nice
);
3125 * can_nice - check if a task can reduce its nice value
3129 int can_nice(const struct task_struct
*p
, const int nice
)
3131 /* convert nice value [19,-20] to rlimit style value [1,40] */
3132 int nice_rlim
= nice_to_rlimit(nice
);
3134 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3135 capable(CAP_SYS_NICE
));
3138 #ifdef __ARCH_WANT_SYS_NICE
3141 * sys_nice - change the priority of the current process.
3142 * @increment: priority increment
3144 * sys_setpriority is a more generic, but much slower function that
3145 * does similar things.
3147 SYSCALL_DEFINE1(nice
, int, increment
)
3152 * Setpriority might change our priority at the same moment.
3153 * We don't have to worry. Conceptually one call occurs first
3154 * and we have a single winner.
3156 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3157 nice
= task_nice(current
) + increment
;
3159 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3160 if (increment
< 0 && !can_nice(current
, nice
))
3163 retval
= security_task_setnice(current
, nice
);
3167 set_user_nice(current
, nice
);
3174 * task_prio - return the priority value of a given task.
3175 * @p: the task in question.
3177 * Return: The priority value as seen by users in /proc.
3178 * RT tasks are offset by -200. Normal tasks are centered
3179 * around 0, value goes from -16 to +15.
3181 int task_prio(const struct task_struct
*p
)
3183 return p
->prio
- MAX_RT_PRIO
;
3187 * idle_cpu - is a given cpu idle currently?
3188 * @cpu: the processor in question.
3190 * Return: 1 if the CPU is currently idle. 0 otherwise.
3192 int idle_cpu(int cpu
)
3194 struct rq
*rq
= cpu_rq(cpu
);
3196 if (rq
->curr
!= rq
->idle
)
3203 if (!llist_empty(&rq
->wake_list
))
3211 * idle_task - return the idle task for a given cpu.
3212 * @cpu: the processor in question.
3214 * Return: The idle task for the cpu @cpu.
3216 struct task_struct
*idle_task(int cpu
)
3218 return cpu_rq(cpu
)->idle
;
3222 * find_process_by_pid - find a process with a matching PID value.
3223 * @pid: the pid in question.
3225 * The task of @pid, if found. %NULL otherwise.
3227 static struct task_struct
*find_process_by_pid(pid_t pid
)
3229 return pid
? find_task_by_vpid(pid
) : current
;
3233 * This function initializes the sched_dl_entity of a newly becoming
3234 * SCHED_DEADLINE task.
3236 * Only the static values are considered here, the actual runtime and the
3237 * absolute deadline will be properly calculated when the task is enqueued
3238 * for the first time with its new policy.
3241 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3243 struct sched_dl_entity
*dl_se
= &p
->dl
;
3245 init_dl_task_timer(dl_se
);
3246 dl_se
->dl_runtime
= attr
->sched_runtime
;
3247 dl_se
->dl_deadline
= attr
->sched_deadline
;
3248 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3249 dl_se
->flags
= attr
->sched_flags
;
3250 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3251 dl_se
->dl_throttled
= 0;
3253 dl_se
->dl_yielded
= 0;
3257 * sched_setparam() passes in -1 for its policy, to let the functions
3258 * it calls know not to change it.
3260 #define SETPARAM_POLICY -1
3262 static void __setscheduler_params(struct task_struct
*p
,
3263 const struct sched_attr
*attr
)
3265 int policy
= attr
->sched_policy
;
3267 if (policy
== SETPARAM_POLICY
)
3272 if (dl_policy(policy
))
3273 __setparam_dl(p
, attr
);
3274 else if (fair_policy(policy
))
3275 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3278 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3279 * !rt_policy. Always setting this ensures that things like
3280 * getparam()/getattr() don't report silly values for !rt tasks.
3282 p
->rt_priority
= attr
->sched_priority
;
3283 p
->normal_prio
= normal_prio(p
);
3287 /* Actually do priority change: must hold pi & rq lock. */
3288 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3289 const struct sched_attr
*attr
)
3291 __setscheduler_params(p
, attr
);
3294 * If we get here, there was no pi waiters boosting the
3295 * task. It is safe to use the normal prio.
3297 p
->prio
= normal_prio(p
);
3299 if (dl_prio(p
->prio
))
3300 p
->sched_class
= &dl_sched_class
;
3301 else if (rt_prio(p
->prio
))
3302 p
->sched_class
= &rt_sched_class
;
3304 p
->sched_class
= &fair_sched_class
;
3308 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3310 struct sched_dl_entity
*dl_se
= &p
->dl
;
3312 attr
->sched_priority
= p
->rt_priority
;
3313 attr
->sched_runtime
= dl_se
->dl_runtime
;
3314 attr
->sched_deadline
= dl_se
->dl_deadline
;
3315 attr
->sched_period
= dl_se
->dl_period
;
3316 attr
->sched_flags
= dl_se
->flags
;
3320 * This function validates the new parameters of a -deadline task.
3321 * We ask for the deadline not being zero, and greater or equal
3322 * than the runtime, as well as the period of being zero or
3323 * greater than deadline. Furthermore, we have to be sure that
3324 * user parameters are above the internal resolution of 1us (we
3325 * check sched_runtime only since it is always the smaller one) and
3326 * below 2^63 ns (we have to check both sched_deadline and
3327 * sched_period, as the latter can be zero).
3330 __checkparam_dl(const struct sched_attr
*attr
)
3333 if (attr
->sched_deadline
== 0)
3337 * Since we truncate DL_SCALE bits, make sure we're at least
3340 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3344 * Since we use the MSB for wrap-around and sign issues, make
3345 * sure it's not set (mind that period can be equal to zero).
3347 if (attr
->sched_deadline
& (1ULL << 63) ||
3348 attr
->sched_period
& (1ULL << 63))
3351 /* runtime <= deadline <= period (if period != 0) */
3352 if ((attr
->sched_period
!= 0 &&
3353 attr
->sched_period
< attr
->sched_deadline
) ||
3354 attr
->sched_deadline
< attr
->sched_runtime
)
3361 * check the target process has a UID that matches the current process's
3363 static bool check_same_owner(struct task_struct
*p
)
3365 const struct cred
*cred
= current_cred(), *pcred
;
3369 pcred
= __task_cred(p
);
3370 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3371 uid_eq(cred
->euid
, pcred
->uid
));
3376 static int __sched_setscheduler(struct task_struct
*p
,
3377 const struct sched_attr
*attr
,
3380 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3381 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3382 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3383 int policy
= attr
->sched_policy
;
3384 unsigned long flags
;
3385 const struct sched_class
*prev_class
;
3389 /* may grab non-irq protected spin_locks */
3390 BUG_ON(in_interrupt());
3392 /* double check policy once rq lock held */
3394 reset_on_fork
= p
->sched_reset_on_fork
;
3395 policy
= oldpolicy
= p
->policy
;
3397 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3399 if (policy
!= SCHED_DEADLINE
&&
3400 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3401 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3402 policy
!= SCHED_IDLE
)
3406 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3410 * Valid priorities for SCHED_FIFO and SCHED_RR are
3411 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3412 * SCHED_BATCH and SCHED_IDLE is 0.
3414 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3415 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3417 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3418 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3422 * Allow unprivileged RT tasks to decrease priority:
3424 if (user
&& !capable(CAP_SYS_NICE
)) {
3425 if (fair_policy(policy
)) {
3426 if (attr
->sched_nice
< task_nice(p
) &&
3427 !can_nice(p
, attr
->sched_nice
))
3431 if (rt_policy(policy
)) {
3432 unsigned long rlim_rtprio
=
3433 task_rlimit(p
, RLIMIT_RTPRIO
);
3435 /* can't set/change the rt policy */
3436 if (policy
!= p
->policy
&& !rlim_rtprio
)
3439 /* can't increase priority */
3440 if (attr
->sched_priority
> p
->rt_priority
&&
3441 attr
->sched_priority
> rlim_rtprio
)
3446 * Can't set/change SCHED_DEADLINE policy at all for now
3447 * (safest behavior); in the future we would like to allow
3448 * unprivileged DL tasks to increase their relative deadline
3449 * or reduce their runtime (both ways reducing utilization)
3451 if (dl_policy(policy
))
3455 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3456 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3458 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3459 if (!can_nice(p
, task_nice(p
)))
3463 /* can't change other user's priorities */
3464 if (!check_same_owner(p
))
3467 /* Normal users shall not reset the sched_reset_on_fork flag */
3468 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3473 retval
= security_task_setscheduler(p
);
3479 * make sure no PI-waiters arrive (or leave) while we are
3480 * changing the priority of the task:
3482 * To be able to change p->policy safely, the appropriate
3483 * runqueue lock must be held.
3485 rq
= task_rq_lock(p
, &flags
);
3488 * Changing the policy of the stop threads its a very bad idea
3490 if (p
== rq
->stop
) {
3491 task_rq_unlock(rq
, p
, &flags
);
3496 * If not changing anything there's no need to proceed further,
3497 * but store a possible modification of reset_on_fork.
3499 if (unlikely(policy
== p
->policy
)) {
3500 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3502 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3504 if (dl_policy(policy
))
3507 p
->sched_reset_on_fork
= reset_on_fork
;
3508 task_rq_unlock(rq
, p
, &flags
);
3514 #ifdef CONFIG_RT_GROUP_SCHED
3516 * Do not allow realtime tasks into groups that have no runtime
3519 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3520 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3521 !task_group_is_autogroup(task_group(p
))) {
3522 task_rq_unlock(rq
, p
, &flags
);
3527 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3528 cpumask_t
*span
= rq
->rd
->span
;
3531 * Don't allow tasks with an affinity mask smaller than
3532 * the entire root_domain to become SCHED_DEADLINE. We
3533 * will also fail if there's no bandwidth available.
3535 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3536 rq
->rd
->dl_bw
.bw
== 0) {
3537 task_rq_unlock(rq
, p
, &flags
);
3544 /* recheck policy now with rq lock held */
3545 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3546 policy
= oldpolicy
= -1;
3547 task_rq_unlock(rq
, p
, &flags
);
3552 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3553 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3556 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3557 task_rq_unlock(rq
, p
, &flags
);
3561 p
->sched_reset_on_fork
= reset_on_fork
;
3565 * Special case for priority boosted tasks.
3567 * If the new priority is lower or equal (user space view)
3568 * than the current (boosted) priority, we just store the new
3569 * normal parameters and do not touch the scheduler class and
3570 * the runqueue. This will be done when the task deboost
3573 if (rt_mutex_check_prio(p
, newprio
)) {
3574 __setscheduler_params(p
, attr
);
3575 task_rq_unlock(rq
, p
, &flags
);
3579 queued
= task_on_rq_queued(p
);
3580 running
= task_current(rq
, p
);
3582 dequeue_task(rq
, p
, 0);
3584 put_prev_task(rq
, p
);
3586 prev_class
= p
->sched_class
;
3587 __setscheduler(rq
, p
, attr
);
3590 p
->sched_class
->set_curr_task(rq
);
3593 * We enqueue to tail when the priority of a task is
3594 * increased (user space view).
3596 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3599 check_class_changed(rq
, p
, prev_class
, oldprio
);
3600 task_rq_unlock(rq
, p
, &flags
);
3602 rt_mutex_adjust_pi(p
);
3607 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3608 const struct sched_param
*param
, bool check
)
3610 struct sched_attr attr
= {
3611 .sched_policy
= policy
,
3612 .sched_priority
= param
->sched_priority
,
3613 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3616 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3617 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3618 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3619 policy
&= ~SCHED_RESET_ON_FORK
;
3620 attr
.sched_policy
= policy
;
3623 return __sched_setscheduler(p
, &attr
, check
);
3626 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3627 * @p: the task in question.
3628 * @policy: new policy.
3629 * @param: structure containing the new RT priority.
3631 * Return: 0 on success. An error code otherwise.
3633 * NOTE that the task may be already dead.
3635 int sched_setscheduler(struct task_struct
*p
, int policy
,
3636 const struct sched_param
*param
)
3638 return _sched_setscheduler(p
, policy
, param
, true);
3640 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3642 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3644 return __sched_setscheduler(p
, attr
, true);
3646 EXPORT_SYMBOL_GPL(sched_setattr
);
3649 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3650 * @p: the task in question.
3651 * @policy: new policy.
3652 * @param: structure containing the new RT priority.
3654 * Just like sched_setscheduler, only don't bother checking if the
3655 * current context has permission. For example, this is needed in
3656 * stop_machine(): we create temporary high priority worker threads,
3657 * but our caller might not have that capability.
3659 * Return: 0 on success. An error code otherwise.
3661 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3662 const struct sched_param
*param
)
3664 return _sched_setscheduler(p
, policy
, param
, false);
3668 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3670 struct sched_param lparam
;
3671 struct task_struct
*p
;
3674 if (!param
|| pid
< 0)
3676 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3681 p
= find_process_by_pid(pid
);
3683 retval
= sched_setscheduler(p
, policy
, &lparam
);
3690 * Mimics kernel/events/core.c perf_copy_attr().
3692 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3693 struct sched_attr
*attr
)
3698 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3702 * zero the full structure, so that a short copy will be nice.
3704 memset(attr
, 0, sizeof(*attr
));
3706 ret
= get_user(size
, &uattr
->size
);
3710 if (size
> PAGE_SIZE
) /* silly large */
3713 if (!size
) /* abi compat */
3714 size
= SCHED_ATTR_SIZE_VER0
;
3716 if (size
< SCHED_ATTR_SIZE_VER0
)
3720 * If we're handed a bigger struct than we know of,
3721 * ensure all the unknown bits are 0 - i.e. new
3722 * user-space does not rely on any kernel feature
3723 * extensions we dont know about yet.
3725 if (size
> sizeof(*attr
)) {
3726 unsigned char __user
*addr
;
3727 unsigned char __user
*end
;
3730 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3731 end
= (void __user
*)uattr
+ size
;
3733 for (; addr
< end
; addr
++) {
3734 ret
= get_user(val
, addr
);
3740 size
= sizeof(*attr
);
3743 ret
= copy_from_user(attr
, uattr
, size
);
3748 * XXX: do we want to be lenient like existing syscalls; or do we want
3749 * to be strict and return an error on out-of-bounds values?
3751 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3756 put_user(sizeof(*attr
), &uattr
->size
);
3761 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3762 * @pid: the pid in question.
3763 * @policy: new policy.
3764 * @param: structure containing the new RT priority.
3766 * Return: 0 on success. An error code otherwise.
3768 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3769 struct sched_param __user
*, param
)
3771 /* negative values for policy are not valid */
3775 return do_sched_setscheduler(pid
, policy
, param
);
3779 * sys_sched_setparam - set/change the RT priority of a thread
3780 * @pid: the pid in question.
3781 * @param: structure containing the new RT priority.
3783 * Return: 0 on success. An error code otherwise.
3785 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3787 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
3791 * sys_sched_setattr - same as above, but with extended sched_attr
3792 * @pid: the pid in question.
3793 * @uattr: structure containing the extended parameters.
3794 * @flags: for future extension.
3796 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3797 unsigned int, flags
)
3799 struct sched_attr attr
;
3800 struct task_struct
*p
;
3803 if (!uattr
|| pid
< 0 || flags
)
3806 retval
= sched_copy_attr(uattr
, &attr
);
3810 if ((int)attr
.sched_policy
< 0)
3815 p
= find_process_by_pid(pid
);
3817 retval
= sched_setattr(p
, &attr
);
3824 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3825 * @pid: the pid in question.
3827 * Return: On success, the policy of the thread. Otherwise, a negative error
3830 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3832 struct task_struct
*p
;
3840 p
= find_process_by_pid(pid
);
3842 retval
= security_task_getscheduler(p
);
3845 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3852 * sys_sched_getparam - get the RT priority of a thread
3853 * @pid: the pid in question.
3854 * @param: structure containing the RT priority.
3856 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3859 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3861 struct sched_param lp
= { .sched_priority
= 0 };
3862 struct task_struct
*p
;
3865 if (!param
|| pid
< 0)
3869 p
= find_process_by_pid(pid
);
3874 retval
= security_task_getscheduler(p
);
3878 if (task_has_rt_policy(p
))
3879 lp
.sched_priority
= p
->rt_priority
;
3883 * This one might sleep, we cannot do it with a spinlock held ...
3885 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3894 static int sched_read_attr(struct sched_attr __user
*uattr
,
3895 struct sched_attr
*attr
,
3900 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3904 * If we're handed a smaller struct than we know of,
3905 * ensure all the unknown bits are 0 - i.e. old
3906 * user-space does not get uncomplete information.
3908 if (usize
< sizeof(*attr
)) {
3909 unsigned char *addr
;
3912 addr
= (void *)attr
+ usize
;
3913 end
= (void *)attr
+ sizeof(*attr
);
3915 for (; addr
< end
; addr
++) {
3923 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3931 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3932 * @pid: the pid in question.
3933 * @uattr: structure containing the extended parameters.
3934 * @size: sizeof(attr) for fwd/bwd comp.
3935 * @flags: for future extension.
3937 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3938 unsigned int, size
, unsigned int, flags
)
3940 struct sched_attr attr
= {
3941 .size
= sizeof(struct sched_attr
),
3943 struct task_struct
*p
;
3946 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3947 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3951 p
= find_process_by_pid(pid
);
3956 retval
= security_task_getscheduler(p
);
3960 attr
.sched_policy
= p
->policy
;
3961 if (p
->sched_reset_on_fork
)
3962 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3963 if (task_has_dl_policy(p
))
3964 __getparam_dl(p
, &attr
);
3965 else if (task_has_rt_policy(p
))
3966 attr
.sched_priority
= p
->rt_priority
;
3968 attr
.sched_nice
= task_nice(p
);
3972 retval
= sched_read_attr(uattr
, &attr
, size
);
3980 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3982 cpumask_var_t cpus_allowed
, new_mask
;
3983 struct task_struct
*p
;
3988 p
= find_process_by_pid(pid
);
3994 /* Prevent p going away */
3998 if (p
->flags
& PF_NO_SETAFFINITY
) {
4002 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4006 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4008 goto out_free_cpus_allowed
;
4011 if (!check_same_owner(p
)) {
4013 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4015 goto out_free_new_mask
;
4020 retval
= security_task_setscheduler(p
);
4022 goto out_free_new_mask
;
4025 cpuset_cpus_allowed(p
, cpus_allowed
);
4026 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4029 * Since bandwidth control happens on root_domain basis,
4030 * if admission test is enabled, we only admit -deadline
4031 * tasks allowed to run on all the CPUs in the task's
4035 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4037 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4040 goto out_free_new_mask
;
4046 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4049 cpuset_cpus_allowed(p
, cpus_allowed
);
4050 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4052 * We must have raced with a concurrent cpuset
4053 * update. Just reset the cpus_allowed to the
4054 * cpuset's cpus_allowed
4056 cpumask_copy(new_mask
, cpus_allowed
);
4061 free_cpumask_var(new_mask
);
4062 out_free_cpus_allowed
:
4063 free_cpumask_var(cpus_allowed
);
4069 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4070 struct cpumask
*new_mask
)
4072 if (len
< cpumask_size())
4073 cpumask_clear(new_mask
);
4074 else if (len
> cpumask_size())
4075 len
= cpumask_size();
4077 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4081 * sys_sched_setaffinity - set the cpu affinity of a process
4082 * @pid: pid of the process
4083 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4084 * @user_mask_ptr: user-space pointer to the new cpu mask
4086 * Return: 0 on success. An error code otherwise.
4088 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4089 unsigned long __user
*, user_mask_ptr
)
4091 cpumask_var_t new_mask
;
4094 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4097 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4099 retval
= sched_setaffinity(pid
, new_mask
);
4100 free_cpumask_var(new_mask
);
4104 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4106 struct task_struct
*p
;
4107 unsigned long flags
;
4113 p
= find_process_by_pid(pid
);
4117 retval
= security_task_getscheduler(p
);
4121 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4122 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4123 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4132 * sys_sched_getaffinity - get the cpu affinity of a process
4133 * @pid: pid of the process
4134 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4135 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4137 * Return: 0 on success. An error code otherwise.
4139 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4140 unsigned long __user
*, user_mask_ptr
)
4145 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4147 if (len
& (sizeof(unsigned long)-1))
4150 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4153 ret
= sched_getaffinity(pid
, mask
);
4155 size_t retlen
= min_t(size_t, len
, cpumask_size());
4157 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4162 free_cpumask_var(mask
);
4168 * sys_sched_yield - yield the current processor to other threads.
4170 * This function yields the current CPU to other tasks. If there are no
4171 * other threads running on this CPU then this function will return.
4175 SYSCALL_DEFINE0(sched_yield
)
4177 struct rq
*rq
= this_rq_lock();
4179 schedstat_inc(rq
, yld_count
);
4180 current
->sched_class
->yield_task(rq
);
4183 * Since we are going to call schedule() anyway, there's
4184 * no need to preempt or enable interrupts:
4186 __release(rq
->lock
);
4187 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4188 do_raw_spin_unlock(&rq
->lock
);
4189 sched_preempt_enable_no_resched();
4196 static void __cond_resched(void)
4198 __preempt_count_add(PREEMPT_ACTIVE
);
4200 __preempt_count_sub(PREEMPT_ACTIVE
);
4203 int __sched
_cond_resched(void)
4205 if (should_resched()) {
4211 EXPORT_SYMBOL(_cond_resched
);
4214 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4215 * call schedule, and on return reacquire the lock.
4217 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4218 * operations here to prevent schedule() from being called twice (once via
4219 * spin_unlock(), once by hand).
4221 int __cond_resched_lock(spinlock_t
*lock
)
4223 int resched
= should_resched();
4226 lockdep_assert_held(lock
);
4228 if (spin_needbreak(lock
) || resched
) {
4239 EXPORT_SYMBOL(__cond_resched_lock
);
4241 int __sched
__cond_resched_softirq(void)
4243 BUG_ON(!in_softirq());
4245 if (should_resched()) {
4253 EXPORT_SYMBOL(__cond_resched_softirq
);
4256 * yield - yield the current processor to other threads.
4258 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4260 * The scheduler is at all times free to pick the calling task as the most
4261 * eligible task to run, if removing the yield() call from your code breaks
4262 * it, its already broken.
4264 * Typical broken usage is:
4269 * where one assumes that yield() will let 'the other' process run that will
4270 * make event true. If the current task is a SCHED_FIFO task that will never
4271 * happen. Never use yield() as a progress guarantee!!
4273 * If you want to use yield() to wait for something, use wait_event().
4274 * If you want to use yield() to be 'nice' for others, use cond_resched().
4275 * If you still want to use yield(), do not!
4277 void __sched
yield(void)
4279 set_current_state(TASK_RUNNING
);
4282 EXPORT_SYMBOL(yield
);
4285 * yield_to - yield the current processor to another thread in
4286 * your thread group, or accelerate that thread toward the
4287 * processor it's on.
4289 * @preempt: whether task preemption is allowed or not
4291 * It's the caller's job to ensure that the target task struct
4292 * can't go away on us before we can do any checks.
4295 * true (>0) if we indeed boosted the target task.
4296 * false (0) if we failed to boost the target.
4297 * -ESRCH if there's no task to yield to.
4299 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4301 struct task_struct
*curr
= current
;
4302 struct rq
*rq
, *p_rq
;
4303 unsigned long flags
;
4306 local_irq_save(flags
);
4312 * If we're the only runnable task on the rq and target rq also
4313 * has only one task, there's absolutely no point in yielding.
4315 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4320 double_rq_lock(rq
, p_rq
);
4321 if (task_rq(p
) != p_rq
) {
4322 double_rq_unlock(rq
, p_rq
);
4326 if (!curr
->sched_class
->yield_to_task
)
4329 if (curr
->sched_class
!= p
->sched_class
)
4332 if (task_running(p_rq
, p
) || p
->state
)
4335 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4337 schedstat_inc(rq
, yld_count
);
4339 * Make p's CPU reschedule; pick_next_entity takes care of
4342 if (preempt
&& rq
!= p_rq
)
4347 double_rq_unlock(rq
, p_rq
);
4349 local_irq_restore(flags
);
4356 EXPORT_SYMBOL_GPL(yield_to
);
4359 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4360 * that process accounting knows that this is a task in IO wait state.
4362 void __sched
io_schedule(void)
4364 struct rq
*rq
= raw_rq();
4366 delayacct_blkio_start();
4367 atomic_inc(&rq
->nr_iowait
);
4368 blk_flush_plug(current
);
4369 current
->in_iowait
= 1;
4371 current
->in_iowait
= 0;
4372 atomic_dec(&rq
->nr_iowait
);
4373 delayacct_blkio_end();
4375 EXPORT_SYMBOL(io_schedule
);
4377 long __sched
io_schedule_timeout(long timeout
)
4379 struct rq
*rq
= raw_rq();
4382 delayacct_blkio_start();
4383 atomic_inc(&rq
->nr_iowait
);
4384 blk_flush_plug(current
);
4385 current
->in_iowait
= 1;
4386 ret
= schedule_timeout(timeout
);
4387 current
->in_iowait
= 0;
4388 atomic_dec(&rq
->nr_iowait
);
4389 delayacct_blkio_end();
4394 * sys_sched_get_priority_max - return maximum RT priority.
4395 * @policy: scheduling class.
4397 * Return: On success, this syscall returns the maximum
4398 * rt_priority that can be used by a given scheduling class.
4399 * On failure, a negative error code is returned.
4401 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4408 ret
= MAX_USER_RT_PRIO
-1;
4410 case SCHED_DEADLINE
:
4421 * sys_sched_get_priority_min - return minimum RT priority.
4422 * @policy: scheduling class.
4424 * Return: On success, this syscall returns the minimum
4425 * rt_priority that can be used by a given scheduling class.
4426 * On failure, a negative error code is returned.
4428 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4437 case SCHED_DEADLINE
:
4447 * sys_sched_rr_get_interval - return the default timeslice of a process.
4448 * @pid: pid of the process.
4449 * @interval: userspace pointer to the timeslice value.
4451 * this syscall writes the default timeslice value of a given process
4452 * into the user-space timespec buffer. A value of '0' means infinity.
4454 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4457 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4458 struct timespec __user
*, interval
)
4460 struct task_struct
*p
;
4461 unsigned int time_slice
;
4462 unsigned long flags
;
4472 p
= find_process_by_pid(pid
);
4476 retval
= security_task_getscheduler(p
);
4480 rq
= task_rq_lock(p
, &flags
);
4482 if (p
->sched_class
->get_rr_interval
)
4483 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4484 task_rq_unlock(rq
, p
, &flags
);
4487 jiffies_to_timespec(time_slice
, &t
);
4488 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4496 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4498 void sched_show_task(struct task_struct
*p
)
4500 unsigned long free
= 0;
4504 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4505 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4506 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4507 #if BITS_PER_LONG == 32
4508 if (state
== TASK_RUNNING
)
4509 printk(KERN_CONT
" running ");
4511 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4513 if (state
== TASK_RUNNING
)
4514 printk(KERN_CONT
" running task ");
4516 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4518 #ifdef CONFIG_DEBUG_STACK_USAGE
4519 free
= stack_not_used(p
);
4522 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4524 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4525 task_pid_nr(p
), ppid
,
4526 (unsigned long)task_thread_info(p
)->flags
);
4528 print_worker_info(KERN_INFO
, p
);
4529 show_stack(p
, NULL
);
4532 void show_state_filter(unsigned long state_filter
)
4534 struct task_struct
*g
, *p
;
4536 #if BITS_PER_LONG == 32
4538 " task PC stack pid father\n");
4541 " task PC stack pid father\n");
4544 for_each_process_thread(g
, p
) {
4546 * reset the NMI-timeout, listing all files on a slow
4547 * console might take a lot of time:
4549 touch_nmi_watchdog();
4550 if (!state_filter
|| (p
->state
& state_filter
))
4554 touch_all_softlockup_watchdogs();
4556 #ifdef CONFIG_SCHED_DEBUG
4557 sysrq_sched_debug_show();
4561 * Only show locks if all tasks are dumped:
4564 debug_show_all_locks();
4567 void init_idle_bootup_task(struct task_struct
*idle
)
4569 idle
->sched_class
= &idle_sched_class
;
4573 * init_idle - set up an idle thread for a given CPU
4574 * @idle: task in question
4575 * @cpu: cpu the idle task belongs to
4577 * NOTE: this function does not set the idle thread's NEED_RESCHED
4578 * flag, to make booting more robust.
4580 void init_idle(struct task_struct
*idle
, int cpu
)
4582 struct rq
*rq
= cpu_rq(cpu
);
4583 unsigned long flags
;
4585 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4587 __sched_fork(0, idle
);
4588 idle
->state
= TASK_RUNNING
;
4589 idle
->se
.exec_start
= sched_clock();
4591 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4593 * We're having a chicken and egg problem, even though we are
4594 * holding rq->lock, the cpu isn't yet set to this cpu so the
4595 * lockdep check in task_group() will fail.
4597 * Similar case to sched_fork(). / Alternatively we could
4598 * use task_rq_lock() here and obtain the other rq->lock.
4603 __set_task_cpu(idle
, cpu
);
4606 rq
->curr
= rq
->idle
= idle
;
4607 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4608 #if defined(CONFIG_SMP)
4611 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4613 /* Set the preempt count _outside_ the spinlocks! */
4614 init_idle_preempt_count(idle
, cpu
);
4617 * The idle tasks have their own, simple scheduling class:
4619 idle
->sched_class
= &idle_sched_class
;
4620 ftrace_graph_init_idle_task(idle
, cpu
);
4621 vtime_init_idle(idle
, cpu
);
4622 #if defined(CONFIG_SMP)
4623 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4627 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4628 const struct cpumask
*trial
)
4630 int ret
= 1, trial_cpus
;
4631 struct dl_bw
*cur_dl_b
;
4632 unsigned long flags
;
4634 rcu_read_lock_sched();
4635 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4636 trial_cpus
= cpumask_weight(trial
);
4638 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4639 if (cur_dl_b
->bw
!= -1 &&
4640 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
4642 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
4643 rcu_read_unlock_sched();
4648 int task_can_attach(struct task_struct
*p
,
4649 const struct cpumask
*cs_cpus_allowed
)
4654 * Kthreads which disallow setaffinity shouldn't be moved
4655 * to a new cpuset; we don't want to change their cpu
4656 * affinity and isolating such threads by their set of
4657 * allowed nodes is unnecessary. Thus, cpusets are not
4658 * applicable for such threads. This prevents checking for
4659 * success of set_cpus_allowed_ptr() on all attached tasks
4660 * before cpus_allowed may be changed.
4662 if (p
->flags
& PF_NO_SETAFFINITY
) {
4668 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
4670 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
4675 unsigned long flags
;
4677 rcu_read_lock_sched();
4678 dl_b
= dl_bw_of(dest_cpu
);
4679 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
4680 cpus
= dl_bw_cpus(dest_cpu
);
4681 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
4686 * We reserve space for this task in the destination
4687 * root_domain, as we can't fail after this point.
4688 * We will free resources in the source root_domain
4689 * later on (see set_cpus_allowed_dl()).
4691 __dl_add(dl_b
, p
->dl
.dl_bw
);
4693 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
4694 rcu_read_unlock_sched();
4704 * move_queued_task - move a queued task to new rq.
4706 * Returns (locked) new rq. Old rq's lock is released.
4708 static struct rq
*move_queued_task(struct task_struct
*p
, int new_cpu
)
4710 struct rq
*rq
= task_rq(p
);
4712 lockdep_assert_held(&rq
->lock
);
4714 dequeue_task(rq
, p
, 0);
4715 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
4716 set_task_cpu(p
, new_cpu
);
4717 raw_spin_unlock(&rq
->lock
);
4719 rq
= cpu_rq(new_cpu
);
4721 raw_spin_lock(&rq
->lock
);
4722 BUG_ON(task_cpu(p
) != new_cpu
);
4723 p
->on_rq
= TASK_ON_RQ_QUEUED
;
4724 enqueue_task(rq
, p
, 0);
4725 check_preempt_curr(rq
, p
, 0);
4730 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4732 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4733 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4735 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4736 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4740 * This is how migration works:
4742 * 1) we invoke migration_cpu_stop() on the target CPU using
4744 * 2) stopper starts to run (implicitly forcing the migrated thread
4746 * 3) it checks whether the migrated task is still in the wrong runqueue.
4747 * 4) if it's in the wrong runqueue then the migration thread removes
4748 * it and puts it into the right queue.
4749 * 5) stopper completes and stop_one_cpu() returns and the migration
4754 * Change a given task's CPU affinity. Migrate the thread to a
4755 * proper CPU and schedule it away if the CPU it's executing on
4756 * is removed from the allowed bitmask.
4758 * NOTE: the caller must have a valid reference to the task, the
4759 * task must not exit() & deallocate itself prematurely. The
4760 * call is not atomic; no spinlocks may be held.
4762 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4764 unsigned long flags
;
4766 unsigned int dest_cpu
;
4769 rq
= task_rq_lock(p
, &flags
);
4771 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4774 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4779 do_set_cpus_allowed(p
, new_mask
);
4781 /* Can the task run on the task's current CPU? If so, we're done */
4782 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4785 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4786 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
4787 struct migration_arg arg
= { p
, dest_cpu
};
4788 /* Need help from migration thread: drop lock and wait. */
4789 task_rq_unlock(rq
, p
, &flags
);
4790 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4791 tlb_migrate_finish(p
->mm
);
4793 } else if (task_on_rq_queued(p
))
4794 rq
= move_queued_task(p
, dest_cpu
);
4796 task_rq_unlock(rq
, p
, &flags
);
4800 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4803 * Move (not current) task off this cpu, onto dest cpu. We're doing
4804 * this because either it can't run here any more (set_cpus_allowed()
4805 * away from this CPU, or CPU going down), or because we're
4806 * attempting to rebalance this task on exec (sched_exec).
4808 * So we race with normal scheduler movements, but that's OK, as long
4809 * as the task is no longer on this CPU.
4811 * Returns non-zero if task was successfully migrated.
4813 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4818 if (unlikely(!cpu_active(dest_cpu
)))
4821 rq
= cpu_rq(src_cpu
);
4823 raw_spin_lock(&p
->pi_lock
);
4824 raw_spin_lock(&rq
->lock
);
4825 /* Already moved. */
4826 if (task_cpu(p
) != src_cpu
)
4829 /* Affinity changed (again). */
4830 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4834 * If we're not on a rq, the next wake-up will ensure we're
4837 if (task_on_rq_queued(p
))
4838 rq
= move_queued_task(p
, dest_cpu
);
4842 raw_spin_unlock(&rq
->lock
);
4843 raw_spin_unlock(&p
->pi_lock
);
4847 #ifdef CONFIG_NUMA_BALANCING
4848 /* Migrate current task p to target_cpu */
4849 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4851 struct migration_arg arg
= { p
, target_cpu
};
4852 int curr_cpu
= task_cpu(p
);
4854 if (curr_cpu
== target_cpu
)
4857 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4860 /* TODO: This is not properly updating schedstats */
4862 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4863 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4867 * Requeue a task on a given node and accurately track the number of NUMA
4868 * tasks on the runqueues
4870 void sched_setnuma(struct task_struct
*p
, int nid
)
4873 unsigned long flags
;
4874 bool queued
, running
;
4876 rq
= task_rq_lock(p
, &flags
);
4877 queued
= task_on_rq_queued(p
);
4878 running
= task_current(rq
, p
);
4881 dequeue_task(rq
, p
, 0);
4883 put_prev_task(rq
, p
);
4885 p
->numa_preferred_nid
= nid
;
4888 p
->sched_class
->set_curr_task(rq
);
4890 enqueue_task(rq
, p
, 0);
4891 task_rq_unlock(rq
, p
, &flags
);
4896 * migration_cpu_stop - this will be executed by a highprio stopper thread
4897 * and performs thread migration by bumping thread off CPU then
4898 * 'pushing' onto another runqueue.
4900 static int migration_cpu_stop(void *data
)
4902 struct migration_arg
*arg
= data
;
4905 * The original target cpu might have gone down and we might
4906 * be on another cpu but it doesn't matter.
4908 local_irq_disable();
4910 * We need to explicitly wake pending tasks before running
4911 * __migrate_task() such that we will not miss enforcing cpus_allowed
4912 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4914 sched_ttwu_pending();
4915 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4920 #ifdef CONFIG_HOTPLUG_CPU
4923 * Ensures that the idle task is using init_mm right before its cpu goes
4926 void idle_task_exit(void)
4928 struct mm_struct
*mm
= current
->active_mm
;
4930 BUG_ON(cpu_online(smp_processor_id()));
4932 if (mm
!= &init_mm
) {
4933 switch_mm(mm
, &init_mm
, current
);
4934 finish_arch_post_lock_switch();
4940 * Since this CPU is going 'away' for a while, fold any nr_active delta
4941 * we might have. Assumes we're called after migrate_tasks() so that the
4942 * nr_active count is stable.
4944 * Also see the comment "Global load-average calculations".
4946 static void calc_load_migrate(struct rq
*rq
)
4948 long delta
= calc_load_fold_active(rq
);
4950 atomic_long_add(delta
, &calc_load_tasks
);
4953 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4957 static const struct sched_class fake_sched_class
= {
4958 .put_prev_task
= put_prev_task_fake
,
4961 static struct task_struct fake_task
= {
4963 * Avoid pull_{rt,dl}_task()
4965 .prio
= MAX_PRIO
+ 1,
4966 .sched_class
= &fake_sched_class
,
4970 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4971 * try_to_wake_up()->select_task_rq().
4973 * Called with rq->lock held even though we'er in stop_machine() and
4974 * there's no concurrency possible, we hold the required locks anyway
4975 * because of lock validation efforts.
4977 static void migrate_tasks(unsigned int dead_cpu
)
4979 struct rq
*rq
= cpu_rq(dead_cpu
);
4980 struct task_struct
*next
, *stop
= rq
->stop
;
4984 * Fudge the rq selection such that the below task selection loop
4985 * doesn't get stuck on the currently eligible stop task.
4987 * We're currently inside stop_machine() and the rq is either stuck
4988 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4989 * either way we should never end up calling schedule() until we're
4995 * put_prev_task() and pick_next_task() sched
4996 * class method both need to have an up-to-date
4997 * value of rq->clock[_task]
4999 update_rq_clock(rq
);
5003 * There's this thread running, bail when that's the only
5006 if (rq
->nr_running
== 1)
5009 next
= pick_next_task(rq
, &fake_task
);
5011 next
->sched_class
->put_prev_task(rq
, next
);
5013 /* Find suitable destination for @next, with force if needed. */
5014 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5015 raw_spin_unlock(&rq
->lock
);
5017 __migrate_task(next
, dead_cpu
, dest_cpu
);
5019 raw_spin_lock(&rq
->lock
);
5025 #endif /* CONFIG_HOTPLUG_CPU */
5027 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5029 static struct ctl_table sd_ctl_dir
[] = {
5031 .procname
= "sched_domain",
5037 static struct ctl_table sd_ctl_root
[] = {
5039 .procname
= "kernel",
5041 .child
= sd_ctl_dir
,
5046 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5048 struct ctl_table
*entry
=
5049 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5054 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5056 struct ctl_table
*entry
;
5059 * In the intermediate directories, both the child directory and
5060 * procname are dynamically allocated and could fail but the mode
5061 * will always be set. In the lowest directory the names are
5062 * static strings and all have proc handlers.
5064 for (entry
= *tablep
; entry
->mode
; entry
++) {
5066 sd_free_ctl_entry(&entry
->child
);
5067 if (entry
->proc_handler
== NULL
)
5068 kfree(entry
->procname
);
5075 static int min_load_idx
= 0;
5076 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5079 set_table_entry(struct ctl_table
*entry
,
5080 const char *procname
, void *data
, int maxlen
,
5081 umode_t mode
, proc_handler
*proc_handler
,
5084 entry
->procname
= procname
;
5086 entry
->maxlen
= maxlen
;
5088 entry
->proc_handler
= proc_handler
;
5091 entry
->extra1
= &min_load_idx
;
5092 entry
->extra2
= &max_load_idx
;
5096 static struct ctl_table
*
5097 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5099 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5104 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5105 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5106 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5107 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5108 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5109 sizeof(int), 0644, proc_dointvec_minmax
, true);
5110 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5111 sizeof(int), 0644, proc_dointvec_minmax
, true);
5112 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5113 sizeof(int), 0644, proc_dointvec_minmax
, true);
5114 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5115 sizeof(int), 0644, proc_dointvec_minmax
, true);
5116 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5117 sizeof(int), 0644, proc_dointvec_minmax
, true);
5118 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5119 sizeof(int), 0644, proc_dointvec_minmax
, false);
5120 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5121 sizeof(int), 0644, proc_dointvec_minmax
, false);
5122 set_table_entry(&table
[9], "cache_nice_tries",
5123 &sd
->cache_nice_tries
,
5124 sizeof(int), 0644, proc_dointvec_minmax
, false);
5125 set_table_entry(&table
[10], "flags", &sd
->flags
,
5126 sizeof(int), 0644, proc_dointvec_minmax
, false);
5127 set_table_entry(&table
[11], "max_newidle_lb_cost",
5128 &sd
->max_newidle_lb_cost
,
5129 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5130 set_table_entry(&table
[12], "name", sd
->name
,
5131 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5132 /* &table[13] is terminator */
5137 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5139 struct ctl_table
*entry
, *table
;
5140 struct sched_domain
*sd
;
5141 int domain_num
= 0, i
;
5144 for_each_domain(cpu
, sd
)
5146 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5151 for_each_domain(cpu
, sd
) {
5152 snprintf(buf
, 32, "domain%d", i
);
5153 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5155 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5162 static struct ctl_table_header
*sd_sysctl_header
;
5163 static void register_sched_domain_sysctl(void)
5165 int i
, cpu_num
= num_possible_cpus();
5166 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5169 WARN_ON(sd_ctl_dir
[0].child
);
5170 sd_ctl_dir
[0].child
= entry
;
5175 for_each_possible_cpu(i
) {
5176 snprintf(buf
, 32, "cpu%d", i
);
5177 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5179 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5183 WARN_ON(sd_sysctl_header
);
5184 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5187 /* may be called multiple times per register */
5188 static void unregister_sched_domain_sysctl(void)
5190 if (sd_sysctl_header
)
5191 unregister_sysctl_table(sd_sysctl_header
);
5192 sd_sysctl_header
= NULL
;
5193 if (sd_ctl_dir
[0].child
)
5194 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5197 static void register_sched_domain_sysctl(void)
5200 static void unregister_sched_domain_sysctl(void)
5205 static void set_rq_online(struct rq
*rq
)
5208 const struct sched_class
*class;
5210 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5213 for_each_class(class) {
5214 if (class->rq_online
)
5215 class->rq_online(rq
);
5220 static void set_rq_offline(struct rq
*rq
)
5223 const struct sched_class
*class;
5225 for_each_class(class) {
5226 if (class->rq_offline
)
5227 class->rq_offline(rq
);
5230 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5236 * migration_call - callback that gets triggered when a CPU is added.
5237 * Here we can start up the necessary migration thread for the new CPU.
5240 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5242 int cpu
= (long)hcpu
;
5243 unsigned long flags
;
5244 struct rq
*rq
= cpu_rq(cpu
);
5246 switch (action
& ~CPU_TASKS_FROZEN
) {
5248 case CPU_UP_PREPARE
:
5249 rq
->calc_load_update
= calc_load_update
;
5253 /* Update our root-domain */
5254 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5256 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5260 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5263 #ifdef CONFIG_HOTPLUG_CPU
5265 sched_ttwu_pending();
5266 /* Update our root-domain */
5267 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5269 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5273 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5274 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5278 calc_load_migrate(rq
);
5283 update_max_interval();
5289 * Register at high priority so that task migration (migrate_all_tasks)
5290 * happens before everything else. This has to be lower priority than
5291 * the notifier in the perf_event subsystem, though.
5293 static struct notifier_block migration_notifier
= {
5294 .notifier_call
= migration_call
,
5295 .priority
= CPU_PRI_MIGRATION
,
5298 static void __cpuinit
set_cpu_rq_start_time(void)
5300 int cpu
= smp_processor_id();
5301 struct rq
*rq
= cpu_rq(cpu
);
5302 rq
->age_stamp
= sched_clock_cpu(cpu
);
5305 static int sched_cpu_active(struct notifier_block
*nfb
,
5306 unsigned long action
, void *hcpu
)
5308 switch (action
& ~CPU_TASKS_FROZEN
) {
5310 set_cpu_rq_start_time();
5312 case CPU_DOWN_FAILED
:
5313 set_cpu_active((long)hcpu
, true);
5320 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5321 unsigned long action
, void *hcpu
)
5323 unsigned long flags
;
5324 long cpu
= (long)hcpu
;
5327 switch (action
& ~CPU_TASKS_FROZEN
) {
5328 case CPU_DOWN_PREPARE
:
5329 set_cpu_active(cpu
, false);
5331 /* explicitly allow suspend */
5332 if (!(action
& CPU_TASKS_FROZEN
)) {
5336 rcu_read_lock_sched();
5337 dl_b
= dl_bw_of(cpu
);
5339 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5340 cpus
= dl_bw_cpus(cpu
);
5341 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5342 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5344 rcu_read_unlock_sched();
5347 return notifier_from_errno(-EBUSY
);
5355 static int __init
migration_init(void)
5357 void *cpu
= (void *)(long)smp_processor_id();
5360 /* Initialize migration for the boot CPU */
5361 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5362 BUG_ON(err
== NOTIFY_BAD
);
5363 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5364 register_cpu_notifier(&migration_notifier
);
5366 /* Register cpu active notifiers */
5367 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5368 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5372 early_initcall(migration_init
);
5377 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5379 #ifdef CONFIG_SCHED_DEBUG
5381 static __read_mostly
int sched_debug_enabled
;
5383 static int __init
sched_debug_setup(char *str
)
5385 sched_debug_enabled
= 1;
5389 early_param("sched_debug", sched_debug_setup
);
5391 static inline bool sched_debug(void)
5393 return sched_debug_enabled
;
5396 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5397 struct cpumask
*groupmask
)
5399 struct sched_group
*group
= sd
->groups
;
5402 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5403 cpumask_clear(groupmask
);
5405 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5407 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5408 printk("does not load-balance\n");
5410 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5415 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5417 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5418 printk(KERN_ERR
"ERROR: domain->span does not contain "
5421 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5422 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5426 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5430 printk(KERN_ERR
"ERROR: group is NULL\n");
5435 * Even though we initialize ->capacity to something semi-sane,
5436 * we leave capacity_orig unset. This allows us to detect if
5437 * domain iteration is still funny without causing /0 traps.
5439 if (!group
->sgc
->capacity_orig
) {
5440 printk(KERN_CONT
"\n");
5441 printk(KERN_ERR
"ERROR: domain->cpu_capacity not set\n");
5445 if (!cpumask_weight(sched_group_cpus(group
))) {
5446 printk(KERN_CONT
"\n");
5447 printk(KERN_ERR
"ERROR: empty group\n");
5451 if (!(sd
->flags
& SD_OVERLAP
) &&
5452 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5453 printk(KERN_CONT
"\n");
5454 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5458 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5460 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5462 printk(KERN_CONT
" %s", str
);
5463 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5464 printk(KERN_CONT
" (cpu_capacity = %d)",
5465 group
->sgc
->capacity
);
5468 group
= group
->next
;
5469 } while (group
!= sd
->groups
);
5470 printk(KERN_CONT
"\n");
5472 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5473 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5476 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5477 printk(KERN_ERR
"ERROR: parent span is not a superset "
5478 "of domain->span\n");
5482 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5486 if (!sched_debug_enabled
)
5490 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5494 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5497 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5505 #else /* !CONFIG_SCHED_DEBUG */
5506 # define sched_domain_debug(sd, cpu) do { } while (0)
5507 static inline bool sched_debug(void)
5511 #endif /* CONFIG_SCHED_DEBUG */
5513 static int sd_degenerate(struct sched_domain
*sd
)
5515 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5518 /* Following flags need at least 2 groups */
5519 if (sd
->flags
& (SD_LOAD_BALANCE
|
5520 SD_BALANCE_NEWIDLE
|
5523 SD_SHARE_CPUCAPACITY
|
5524 SD_SHARE_PKG_RESOURCES
|
5525 SD_SHARE_POWERDOMAIN
)) {
5526 if (sd
->groups
!= sd
->groups
->next
)
5530 /* Following flags don't use groups */
5531 if (sd
->flags
& (SD_WAKE_AFFINE
))
5538 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5540 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5542 if (sd_degenerate(parent
))
5545 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5548 /* Flags needing groups don't count if only 1 group in parent */
5549 if (parent
->groups
== parent
->groups
->next
) {
5550 pflags
&= ~(SD_LOAD_BALANCE
|
5551 SD_BALANCE_NEWIDLE
|
5554 SD_SHARE_CPUCAPACITY
|
5555 SD_SHARE_PKG_RESOURCES
|
5557 SD_SHARE_POWERDOMAIN
);
5558 if (nr_node_ids
== 1)
5559 pflags
&= ~SD_SERIALIZE
;
5561 if (~cflags
& pflags
)
5567 static void free_rootdomain(struct rcu_head
*rcu
)
5569 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5571 cpupri_cleanup(&rd
->cpupri
);
5572 cpudl_cleanup(&rd
->cpudl
);
5573 free_cpumask_var(rd
->dlo_mask
);
5574 free_cpumask_var(rd
->rto_mask
);
5575 free_cpumask_var(rd
->online
);
5576 free_cpumask_var(rd
->span
);
5580 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5582 struct root_domain
*old_rd
= NULL
;
5583 unsigned long flags
;
5585 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5590 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5593 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5596 * If we dont want to free the old_rd yet then
5597 * set old_rd to NULL to skip the freeing later
5600 if (!atomic_dec_and_test(&old_rd
->refcount
))
5604 atomic_inc(&rd
->refcount
);
5607 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5608 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5611 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5614 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5617 static int init_rootdomain(struct root_domain
*rd
)
5619 memset(rd
, 0, sizeof(*rd
));
5621 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5623 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5625 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5627 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5630 init_dl_bw(&rd
->dl_bw
);
5631 if (cpudl_init(&rd
->cpudl
) != 0)
5634 if (cpupri_init(&rd
->cpupri
) != 0)
5639 free_cpumask_var(rd
->rto_mask
);
5641 free_cpumask_var(rd
->dlo_mask
);
5643 free_cpumask_var(rd
->online
);
5645 free_cpumask_var(rd
->span
);
5651 * By default the system creates a single root-domain with all cpus as
5652 * members (mimicking the global state we have today).
5654 struct root_domain def_root_domain
;
5656 static void init_defrootdomain(void)
5658 init_rootdomain(&def_root_domain
);
5660 atomic_set(&def_root_domain
.refcount
, 1);
5663 static struct root_domain
*alloc_rootdomain(void)
5665 struct root_domain
*rd
;
5667 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5671 if (init_rootdomain(rd
) != 0) {
5679 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5681 struct sched_group
*tmp
, *first
;
5690 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5695 } while (sg
!= first
);
5698 static void free_sched_domain(struct rcu_head
*rcu
)
5700 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5703 * If its an overlapping domain it has private groups, iterate and
5706 if (sd
->flags
& SD_OVERLAP
) {
5707 free_sched_groups(sd
->groups
, 1);
5708 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5709 kfree(sd
->groups
->sgc
);
5715 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5717 call_rcu(&sd
->rcu
, free_sched_domain
);
5720 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5722 for (; sd
; sd
= sd
->parent
)
5723 destroy_sched_domain(sd
, cpu
);
5727 * Keep a special pointer to the highest sched_domain that has
5728 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5729 * allows us to avoid some pointer chasing select_idle_sibling().
5731 * Also keep a unique ID per domain (we use the first cpu number in
5732 * the cpumask of the domain), this allows us to quickly tell if
5733 * two cpus are in the same cache domain, see cpus_share_cache().
5735 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5736 DEFINE_PER_CPU(int, sd_llc_size
);
5737 DEFINE_PER_CPU(int, sd_llc_id
);
5738 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5739 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5740 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5742 static void update_top_cache_domain(int cpu
)
5744 struct sched_domain
*sd
;
5745 struct sched_domain
*busy_sd
= NULL
;
5749 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5751 id
= cpumask_first(sched_domain_span(sd
));
5752 size
= cpumask_weight(sched_domain_span(sd
));
5753 busy_sd
= sd
->parent
; /* sd_busy */
5755 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5757 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5758 per_cpu(sd_llc_size
, cpu
) = size
;
5759 per_cpu(sd_llc_id
, cpu
) = id
;
5761 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5762 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5764 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5765 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5769 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5770 * hold the hotplug lock.
5773 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5775 struct rq
*rq
= cpu_rq(cpu
);
5776 struct sched_domain
*tmp
;
5778 /* Remove the sched domains which do not contribute to scheduling. */
5779 for (tmp
= sd
; tmp
; ) {
5780 struct sched_domain
*parent
= tmp
->parent
;
5784 if (sd_parent_degenerate(tmp
, parent
)) {
5785 tmp
->parent
= parent
->parent
;
5787 parent
->parent
->child
= tmp
;
5789 * Transfer SD_PREFER_SIBLING down in case of a
5790 * degenerate parent; the spans match for this
5791 * so the property transfers.
5793 if (parent
->flags
& SD_PREFER_SIBLING
)
5794 tmp
->flags
|= SD_PREFER_SIBLING
;
5795 destroy_sched_domain(parent
, cpu
);
5800 if (sd
&& sd_degenerate(sd
)) {
5803 destroy_sched_domain(tmp
, cpu
);
5808 sched_domain_debug(sd
, cpu
);
5810 rq_attach_root(rq
, rd
);
5812 rcu_assign_pointer(rq
->sd
, sd
);
5813 destroy_sched_domains(tmp
, cpu
);
5815 update_top_cache_domain(cpu
);
5818 /* cpus with isolated domains */
5819 static cpumask_var_t cpu_isolated_map
;
5821 /* Setup the mask of cpus configured for isolated domains */
5822 static int __init
isolated_cpu_setup(char *str
)
5824 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5825 cpulist_parse(str
, cpu_isolated_map
);
5829 __setup("isolcpus=", isolated_cpu_setup
);
5832 struct sched_domain
** __percpu sd
;
5833 struct root_domain
*rd
;
5844 * Build an iteration mask that can exclude certain CPUs from the upwards
5847 * Asymmetric node setups can result in situations where the domain tree is of
5848 * unequal depth, make sure to skip domains that already cover the entire
5851 * In that case build_sched_domains() will have terminated the iteration early
5852 * and our sibling sd spans will be empty. Domains should always include the
5853 * cpu they're built on, so check that.
5856 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5858 const struct cpumask
*span
= sched_domain_span(sd
);
5859 struct sd_data
*sdd
= sd
->private;
5860 struct sched_domain
*sibling
;
5863 for_each_cpu(i
, span
) {
5864 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5865 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5868 cpumask_set_cpu(i
, sched_group_mask(sg
));
5873 * Return the canonical balance cpu for this group, this is the first cpu
5874 * of this group that's also in the iteration mask.
5876 int group_balance_cpu(struct sched_group
*sg
)
5878 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5882 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5884 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5885 const struct cpumask
*span
= sched_domain_span(sd
);
5886 struct cpumask
*covered
= sched_domains_tmpmask
;
5887 struct sd_data
*sdd
= sd
->private;
5888 struct sched_domain
*sibling
;
5891 cpumask_clear(covered
);
5893 for_each_cpu(i
, span
) {
5894 struct cpumask
*sg_span
;
5896 if (cpumask_test_cpu(i
, covered
))
5899 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5901 /* See the comment near build_group_mask(). */
5902 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5905 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5906 GFP_KERNEL
, cpu_to_node(cpu
));
5911 sg_span
= sched_group_cpus(sg
);
5913 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
5915 cpumask_set_cpu(i
, sg_span
);
5917 cpumask_or(covered
, covered
, sg_span
);
5919 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
5920 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
5921 build_group_mask(sd
, sg
);
5924 * Initialize sgc->capacity such that even if we mess up the
5925 * domains and no possible iteration will get us here, we won't
5928 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
5929 sg
->sgc
->capacity_orig
= sg
->sgc
->capacity
;
5932 * Make sure the first group of this domain contains the
5933 * canonical balance cpu. Otherwise the sched_domain iteration
5934 * breaks. See update_sg_lb_stats().
5936 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5937 group_balance_cpu(sg
) == cpu
)
5947 sd
->groups
= groups
;
5952 free_sched_groups(first
, 0);
5957 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5959 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5960 struct sched_domain
*child
= sd
->child
;
5963 cpu
= cpumask_first(sched_domain_span(child
));
5966 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5967 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
5968 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
5975 * build_sched_groups will build a circular linked list of the groups
5976 * covered by the given span, and will set each group's ->cpumask correctly,
5977 * and ->cpu_capacity to 0.
5979 * Assumes the sched_domain tree is fully constructed
5982 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5984 struct sched_group
*first
= NULL
, *last
= NULL
;
5985 struct sd_data
*sdd
= sd
->private;
5986 const struct cpumask
*span
= sched_domain_span(sd
);
5987 struct cpumask
*covered
;
5990 get_group(cpu
, sdd
, &sd
->groups
);
5991 atomic_inc(&sd
->groups
->ref
);
5993 if (cpu
!= cpumask_first(span
))
5996 lockdep_assert_held(&sched_domains_mutex
);
5997 covered
= sched_domains_tmpmask
;
5999 cpumask_clear(covered
);
6001 for_each_cpu(i
, span
) {
6002 struct sched_group
*sg
;
6005 if (cpumask_test_cpu(i
, covered
))
6008 group
= get_group(i
, sdd
, &sg
);
6009 cpumask_setall(sched_group_mask(sg
));
6011 for_each_cpu(j
, span
) {
6012 if (get_group(j
, sdd
, NULL
) != group
)
6015 cpumask_set_cpu(j
, covered
);
6016 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6031 * Initialize sched groups cpu_capacity.
6033 * cpu_capacity indicates the capacity of sched group, which is used while
6034 * distributing the load between different sched groups in a sched domain.
6035 * Typically cpu_capacity for all the groups in a sched domain will be same
6036 * unless there are asymmetries in the topology. If there are asymmetries,
6037 * group having more cpu_capacity will pickup more load compared to the
6038 * group having less cpu_capacity.
6040 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6042 struct sched_group
*sg
= sd
->groups
;
6047 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6049 } while (sg
!= sd
->groups
);
6051 if (cpu
!= group_balance_cpu(sg
))
6054 update_group_capacity(sd
, cpu
);
6055 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6059 * Initializers for schedule domains
6060 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6063 static int default_relax_domain_level
= -1;
6064 int sched_domain_level_max
;
6066 static int __init
setup_relax_domain_level(char *str
)
6068 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6069 pr_warn("Unable to set relax_domain_level\n");
6073 __setup("relax_domain_level=", setup_relax_domain_level
);
6075 static void set_domain_attribute(struct sched_domain
*sd
,
6076 struct sched_domain_attr
*attr
)
6080 if (!attr
|| attr
->relax_domain_level
< 0) {
6081 if (default_relax_domain_level
< 0)
6084 request
= default_relax_domain_level
;
6086 request
= attr
->relax_domain_level
;
6087 if (request
< sd
->level
) {
6088 /* turn off idle balance on this domain */
6089 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6091 /* turn on idle balance on this domain */
6092 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6096 static void __sdt_free(const struct cpumask
*cpu_map
);
6097 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6099 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6100 const struct cpumask
*cpu_map
)
6104 if (!atomic_read(&d
->rd
->refcount
))
6105 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6107 free_percpu(d
->sd
); /* fall through */
6109 __sdt_free(cpu_map
); /* fall through */
6115 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6116 const struct cpumask
*cpu_map
)
6118 memset(d
, 0, sizeof(*d
));
6120 if (__sdt_alloc(cpu_map
))
6121 return sa_sd_storage
;
6122 d
->sd
= alloc_percpu(struct sched_domain
*);
6124 return sa_sd_storage
;
6125 d
->rd
= alloc_rootdomain();
6128 return sa_rootdomain
;
6132 * NULL the sd_data elements we've used to build the sched_domain and
6133 * sched_group structure so that the subsequent __free_domain_allocs()
6134 * will not free the data we're using.
6136 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6138 struct sd_data
*sdd
= sd
->private;
6140 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6141 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6143 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6144 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6146 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6147 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6151 static int sched_domains_numa_levels
;
6152 enum numa_topology_type sched_numa_topology_type
;
6153 static int *sched_domains_numa_distance
;
6154 int sched_max_numa_distance
;
6155 static struct cpumask
***sched_domains_numa_masks
;
6156 static int sched_domains_curr_level
;
6160 * SD_flags allowed in topology descriptions.
6162 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6163 * SD_SHARE_PKG_RESOURCES - describes shared caches
6164 * SD_NUMA - describes NUMA topologies
6165 * SD_SHARE_POWERDOMAIN - describes shared power domain
6168 * SD_ASYM_PACKING - describes SMT quirks
6170 #define TOPOLOGY_SD_FLAGS \
6171 (SD_SHARE_CPUCAPACITY | \
6172 SD_SHARE_PKG_RESOURCES | \
6175 SD_SHARE_POWERDOMAIN)
6177 static struct sched_domain
*
6178 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6180 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6181 int sd_weight
, sd_flags
= 0;
6185 * Ugly hack to pass state to sd_numa_mask()...
6187 sched_domains_curr_level
= tl
->numa_level
;
6190 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6193 sd_flags
= (*tl
->sd_flags
)();
6194 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6195 "wrong sd_flags in topology description\n"))
6196 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6198 *sd
= (struct sched_domain
){
6199 .min_interval
= sd_weight
,
6200 .max_interval
= 2*sd_weight
,
6202 .imbalance_pct
= 125,
6204 .cache_nice_tries
= 0,
6211 .flags
= 1*SD_LOAD_BALANCE
6212 | 1*SD_BALANCE_NEWIDLE
6217 | 0*SD_SHARE_CPUCAPACITY
6218 | 0*SD_SHARE_PKG_RESOURCES
6220 | 0*SD_PREFER_SIBLING
6225 .last_balance
= jiffies
,
6226 .balance_interval
= sd_weight
,
6228 .max_newidle_lb_cost
= 0,
6229 .next_decay_max_lb_cost
= jiffies
,
6230 #ifdef CONFIG_SCHED_DEBUG
6236 * Convert topological properties into behaviour.
6239 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6240 sd
->imbalance_pct
= 110;
6241 sd
->smt_gain
= 1178; /* ~15% */
6243 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6244 sd
->imbalance_pct
= 117;
6245 sd
->cache_nice_tries
= 1;
6249 } else if (sd
->flags
& SD_NUMA
) {
6250 sd
->cache_nice_tries
= 2;
6254 sd
->flags
|= SD_SERIALIZE
;
6255 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6256 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6263 sd
->flags
|= SD_PREFER_SIBLING
;
6264 sd
->cache_nice_tries
= 1;
6269 sd
->private = &tl
->data
;
6275 * Topology list, bottom-up.
6277 static struct sched_domain_topology_level default_topology
[] = {
6278 #ifdef CONFIG_SCHED_SMT
6279 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6281 #ifdef CONFIG_SCHED_MC
6282 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6284 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6288 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6290 #define for_each_sd_topology(tl) \
6291 for (tl = sched_domain_topology; tl->mask; tl++)
6293 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6295 sched_domain_topology
= tl
;
6300 static const struct cpumask
*sd_numa_mask(int cpu
)
6302 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6305 static void sched_numa_warn(const char *str
)
6307 static int done
= false;
6315 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6317 for (i
= 0; i
< nr_node_ids
; i
++) {
6318 printk(KERN_WARNING
" ");
6319 for (j
= 0; j
< nr_node_ids
; j
++)
6320 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6321 printk(KERN_CONT
"\n");
6323 printk(KERN_WARNING
"\n");
6326 bool find_numa_distance(int distance
)
6330 if (distance
== node_distance(0, 0))
6333 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6334 if (sched_domains_numa_distance
[i
] == distance
)
6342 * A system can have three types of NUMA topology:
6343 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6344 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6345 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6347 * The difference between a glueless mesh topology and a backplane
6348 * topology lies in whether communication between not directly
6349 * connected nodes goes through intermediary nodes (where programs
6350 * could run), or through backplane controllers. This affects
6351 * placement of programs.
6353 * The type of topology can be discerned with the following tests:
6354 * - If the maximum distance between any nodes is 1 hop, the system
6355 * is directly connected.
6356 * - If for two nodes A and B, located N > 1 hops away from each other,
6357 * there is an intermediary node C, which is < N hops away from both
6358 * nodes A and B, the system is a glueless mesh.
6360 static void init_numa_topology_type(void)
6364 n
= sched_max_numa_distance
;
6367 sched_numa_topology_type
= NUMA_DIRECT
;
6369 for_each_online_node(a
) {
6370 for_each_online_node(b
) {
6371 /* Find two nodes furthest removed from each other. */
6372 if (node_distance(a
, b
) < n
)
6375 /* Is there an intermediary node between a and b? */
6376 for_each_online_node(c
) {
6377 if (node_distance(a
, c
) < n
&&
6378 node_distance(b
, c
) < n
) {
6379 sched_numa_topology_type
=
6385 sched_numa_topology_type
= NUMA_BACKPLANE
;
6391 static void sched_init_numa(void)
6393 int next_distance
, curr_distance
= node_distance(0, 0);
6394 struct sched_domain_topology_level
*tl
;
6398 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6399 if (!sched_domains_numa_distance
)
6403 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6404 * unique distances in the node_distance() table.
6406 * Assumes node_distance(0,j) includes all distances in
6407 * node_distance(i,j) in order to avoid cubic time.
6409 next_distance
= curr_distance
;
6410 for (i
= 0; i
< nr_node_ids
; i
++) {
6411 for (j
= 0; j
< nr_node_ids
; j
++) {
6412 for (k
= 0; k
< nr_node_ids
; k
++) {
6413 int distance
= node_distance(i
, k
);
6415 if (distance
> curr_distance
&&
6416 (distance
< next_distance
||
6417 next_distance
== curr_distance
))
6418 next_distance
= distance
;
6421 * While not a strong assumption it would be nice to know
6422 * about cases where if node A is connected to B, B is not
6423 * equally connected to A.
6425 if (sched_debug() && node_distance(k
, i
) != distance
)
6426 sched_numa_warn("Node-distance not symmetric");
6428 if (sched_debug() && i
&& !find_numa_distance(distance
))
6429 sched_numa_warn("Node-0 not representative");
6431 if (next_distance
!= curr_distance
) {
6432 sched_domains_numa_distance
[level
++] = next_distance
;
6433 sched_domains_numa_levels
= level
;
6434 curr_distance
= next_distance
;
6439 * In case of sched_debug() we verify the above assumption.
6449 * 'level' contains the number of unique distances, excluding the
6450 * identity distance node_distance(i,i).
6452 * The sched_domains_numa_distance[] array includes the actual distance
6457 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6458 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6459 * the array will contain less then 'level' members. This could be
6460 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6461 * in other functions.
6463 * We reset it to 'level' at the end of this function.
6465 sched_domains_numa_levels
= 0;
6467 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6468 if (!sched_domains_numa_masks
)
6472 * Now for each level, construct a mask per node which contains all
6473 * cpus of nodes that are that many hops away from us.
6475 for (i
= 0; i
< level
; i
++) {
6476 sched_domains_numa_masks
[i
] =
6477 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6478 if (!sched_domains_numa_masks
[i
])
6481 for (j
= 0; j
< nr_node_ids
; j
++) {
6482 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6486 sched_domains_numa_masks
[i
][j
] = mask
;
6488 for (k
= 0; k
< nr_node_ids
; k
++) {
6489 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6492 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6497 /* Compute default topology size */
6498 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6500 tl
= kzalloc((i
+ level
+ 1) *
6501 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6506 * Copy the default topology bits..
6508 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6509 tl
[i
] = sched_domain_topology
[i
];
6512 * .. and append 'j' levels of NUMA goodness.
6514 for (j
= 0; j
< level
; i
++, j
++) {
6515 tl
[i
] = (struct sched_domain_topology_level
){
6516 .mask
= sd_numa_mask
,
6517 .sd_flags
= cpu_numa_flags
,
6518 .flags
= SDTL_OVERLAP
,
6524 sched_domain_topology
= tl
;
6526 sched_domains_numa_levels
= level
;
6527 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6529 init_numa_topology_type();
6532 static void sched_domains_numa_masks_set(int cpu
)
6535 int node
= cpu_to_node(cpu
);
6537 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6538 for (j
= 0; j
< nr_node_ids
; j
++) {
6539 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6540 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6545 static void sched_domains_numa_masks_clear(int cpu
)
6548 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6549 for (j
= 0; j
< nr_node_ids
; j
++)
6550 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6555 * Update sched_domains_numa_masks[level][node] array when new cpus
6558 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6559 unsigned long action
,
6562 int cpu
= (long)hcpu
;
6564 switch (action
& ~CPU_TASKS_FROZEN
) {
6566 sched_domains_numa_masks_set(cpu
);
6570 sched_domains_numa_masks_clear(cpu
);
6580 static inline void sched_init_numa(void)
6584 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6585 unsigned long action
,
6590 #endif /* CONFIG_NUMA */
6592 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6594 struct sched_domain_topology_level
*tl
;
6597 for_each_sd_topology(tl
) {
6598 struct sd_data
*sdd
= &tl
->data
;
6600 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6604 sdd
->sg
= alloc_percpu(struct sched_group
*);
6608 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6612 for_each_cpu(j
, cpu_map
) {
6613 struct sched_domain
*sd
;
6614 struct sched_group
*sg
;
6615 struct sched_group_capacity
*sgc
;
6617 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6618 GFP_KERNEL
, cpu_to_node(j
));
6622 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6624 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6625 GFP_KERNEL
, cpu_to_node(j
));
6631 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6633 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6634 GFP_KERNEL
, cpu_to_node(j
));
6638 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6645 static void __sdt_free(const struct cpumask
*cpu_map
)
6647 struct sched_domain_topology_level
*tl
;
6650 for_each_sd_topology(tl
) {
6651 struct sd_data
*sdd
= &tl
->data
;
6653 for_each_cpu(j
, cpu_map
) {
6654 struct sched_domain
*sd
;
6657 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6658 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6659 free_sched_groups(sd
->groups
, 0);
6660 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6664 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6666 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6668 free_percpu(sdd
->sd
);
6670 free_percpu(sdd
->sg
);
6672 free_percpu(sdd
->sgc
);
6677 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6678 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6679 struct sched_domain
*child
, int cpu
)
6681 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6685 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6687 sd
->level
= child
->level
+ 1;
6688 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6692 if (!cpumask_subset(sched_domain_span(child
),
6693 sched_domain_span(sd
))) {
6694 pr_err("BUG: arch topology borken\n");
6695 #ifdef CONFIG_SCHED_DEBUG
6696 pr_err(" the %s domain not a subset of the %s domain\n",
6697 child
->name
, sd
->name
);
6699 /* Fixup, ensure @sd has at least @child cpus. */
6700 cpumask_or(sched_domain_span(sd
),
6701 sched_domain_span(sd
),
6702 sched_domain_span(child
));
6706 set_domain_attribute(sd
, attr
);
6712 * Build sched domains for a given set of cpus and attach the sched domains
6713 * to the individual cpus
6715 static int build_sched_domains(const struct cpumask
*cpu_map
,
6716 struct sched_domain_attr
*attr
)
6718 enum s_alloc alloc_state
;
6719 struct sched_domain
*sd
;
6721 int i
, ret
= -ENOMEM
;
6723 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6724 if (alloc_state
!= sa_rootdomain
)
6727 /* Set up domains for cpus specified by the cpu_map. */
6728 for_each_cpu(i
, cpu_map
) {
6729 struct sched_domain_topology_level
*tl
;
6732 for_each_sd_topology(tl
) {
6733 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6734 if (tl
== sched_domain_topology
)
6735 *per_cpu_ptr(d
.sd
, i
) = sd
;
6736 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6737 sd
->flags
|= SD_OVERLAP
;
6738 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6743 /* Build the groups for the domains */
6744 for_each_cpu(i
, cpu_map
) {
6745 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6746 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6747 if (sd
->flags
& SD_OVERLAP
) {
6748 if (build_overlap_sched_groups(sd
, i
))
6751 if (build_sched_groups(sd
, i
))
6757 /* Calculate CPU capacity for physical packages and nodes */
6758 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6759 if (!cpumask_test_cpu(i
, cpu_map
))
6762 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6763 claim_allocations(i
, sd
);
6764 init_sched_groups_capacity(i
, sd
);
6768 /* Attach the domains */
6770 for_each_cpu(i
, cpu_map
) {
6771 sd
= *per_cpu_ptr(d
.sd
, i
);
6772 cpu_attach_domain(sd
, d
.rd
, i
);
6778 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6782 static cpumask_var_t
*doms_cur
; /* current sched domains */
6783 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6784 static struct sched_domain_attr
*dattr_cur
;
6785 /* attribues of custom domains in 'doms_cur' */
6788 * Special case: If a kmalloc of a doms_cur partition (array of
6789 * cpumask) fails, then fallback to a single sched domain,
6790 * as determined by the single cpumask fallback_doms.
6792 static cpumask_var_t fallback_doms
;
6795 * arch_update_cpu_topology lets virtualized architectures update the
6796 * cpu core maps. It is supposed to return 1 if the topology changed
6797 * or 0 if it stayed the same.
6799 int __weak
arch_update_cpu_topology(void)
6804 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6807 cpumask_var_t
*doms
;
6809 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6812 for (i
= 0; i
< ndoms
; i
++) {
6813 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6814 free_sched_domains(doms
, i
);
6821 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6824 for (i
= 0; i
< ndoms
; i
++)
6825 free_cpumask_var(doms
[i
]);
6830 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6831 * For now this just excludes isolated cpus, but could be used to
6832 * exclude other special cases in the future.
6834 static int init_sched_domains(const struct cpumask
*cpu_map
)
6838 arch_update_cpu_topology();
6840 doms_cur
= alloc_sched_domains(ndoms_cur
);
6842 doms_cur
= &fallback_doms
;
6843 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6844 err
= build_sched_domains(doms_cur
[0], NULL
);
6845 register_sched_domain_sysctl();
6851 * Detach sched domains from a group of cpus specified in cpu_map
6852 * These cpus will now be attached to the NULL domain
6854 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6859 for_each_cpu(i
, cpu_map
)
6860 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6864 /* handle null as "default" */
6865 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6866 struct sched_domain_attr
*new, int idx_new
)
6868 struct sched_domain_attr tmp
;
6875 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6876 new ? (new + idx_new
) : &tmp
,
6877 sizeof(struct sched_domain_attr
));
6881 * Partition sched domains as specified by the 'ndoms_new'
6882 * cpumasks in the array doms_new[] of cpumasks. This compares
6883 * doms_new[] to the current sched domain partitioning, doms_cur[].
6884 * It destroys each deleted domain and builds each new domain.
6886 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6887 * The masks don't intersect (don't overlap.) We should setup one
6888 * sched domain for each mask. CPUs not in any of the cpumasks will
6889 * not be load balanced. If the same cpumask appears both in the
6890 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6893 * The passed in 'doms_new' should be allocated using
6894 * alloc_sched_domains. This routine takes ownership of it and will
6895 * free_sched_domains it when done with it. If the caller failed the
6896 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6897 * and partition_sched_domains() will fallback to the single partition
6898 * 'fallback_doms', it also forces the domains to be rebuilt.
6900 * If doms_new == NULL it will be replaced with cpu_online_mask.
6901 * ndoms_new == 0 is a special case for destroying existing domains,
6902 * and it will not create the default domain.
6904 * Call with hotplug lock held
6906 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6907 struct sched_domain_attr
*dattr_new
)
6912 mutex_lock(&sched_domains_mutex
);
6914 /* always unregister in case we don't destroy any domains */
6915 unregister_sched_domain_sysctl();
6917 /* Let architecture update cpu core mappings. */
6918 new_topology
= arch_update_cpu_topology();
6920 n
= doms_new
? ndoms_new
: 0;
6922 /* Destroy deleted domains */
6923 for (i
= 0; i
< ndoms_cur
; i
++) {
6924 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6925 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6926 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6929 /* no match - a current sched domain not in new doms_new[] */
6930 detach_destroy_domains(doms_cur
[i
]);
6936 if (doms_new
== NULL
) {
6938 doms_new
= &fallback_doms
;
6939 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6940 WARN_ON_ONCE(dattr_new
);
6943 /* Build new domains */
6944 for (i
= 0; i
< ndoms_new
; i
++) {
6945 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6946 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6947 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6950 /* no match - add a new doms_new */
6951 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6956 /* Remember the new sched domains */
6957 if (doms_cur
!= &fallback_doms
)
6958 free_sched_domains(doms_cur
, ndoms_cur
);
6959 kfree(dattr_cur
); /* kfree(NULL) is safe */
6960 doms_cur
= doms_new
;
6961 dattr_cur
= dattr_new
;
6962 ndoms_cur
= ndoms_new
;
6964 register_sched_domain_sysctl();
6966 mutex_unlock(&sched_domains_mutex
);
6969 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6972 * Update cpusets according to cpu_active mask. If cpusets are
6973 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6974 * around partition_sched_domains().
6976 * If we come here as part of a suspend/resume, don't touch cpusets because we
6977 * want to restore it back to its original state upon resume anyway.
6979 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6983 case CPU_ONLINE_FROZEN
:
6984 case CPU_DOWN_FAILED_FROZEN
:
6987 * num_cpus_frozen tracks how many CPUs are involved in suspend
6988 * resume sequence. As long as this is not the last online
6989 * operation in the resume sequence, just build a single sched
6990 * domain, ignoring cpusets.
6993 if (likely(num_cpus_frozen
)) {
6994 partition_sched_domains(1, NULL
, NULL
);
6999 * This is the last CPU online operation. So fall through and
7000 * restore the original sched domains by considering the
7001 * cpuset configurations.
7005 case CPU_DOWN_FAILED
:
7006 cpuset_update_active_cpus(true);
7014 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7018 case CPU_DOWN_PREPARE
:
7019 cpuset_update_active_cpus(false);
7021 case CPU_DOWN_PREPARE_FROZEN
:
7023 partition_sched_domains(1, NULL
, NULL
);
7031 void __init
sched_init_smp(void)
7033 cpumask_var_t non_isolated_cpus
;
7035 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7036 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7041 * There's no userspace yet to cause hotplug operations; hence all the
7042 * cpu masks are stable and all blatant races in the below code cannot
7045 mutex_lock(&sched_domains_mutex
);
7046 init_sched_domains(cpu_active_mask
);
7047 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7048 if (cpumask_empty(non_isolated_cpus
))
7049 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7050 mutex_unlock(&sched_domains_mutex
);
7052 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7053 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7054 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7058 /* Move init over to a non-isolated CPU */
7059 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7061 sched_init_granularity();
7062 free_cpumask_var(non_isolated_cpus
);
7064 init_sched_rt_class();
7065 init_sched_dl_class();
7068 void __init
sched_init_smp(void)
7070 sched_init_granularity();
7072 #endif /* CONFIG_SMP */
7074 const_debug
unsigned int sysctl_timer_migration
= 1;
7076 int in_sched_functions(unsigned long addr
)
7078 return in_lock_functions(addr
) ||
7079 (addr
>= (unsigned long)__sched_text_start
7080 && addr
< (unsigned long)__sched_text_end
);
7083 #ifdef CONFIG_CGROUP_SCHED
7085 * Default task group.
7086 * Every task in system belongs to this group at bootup.
7088 struct task_group root_task_group
;
7089 LIST_HEAD(task_groups
);
7092 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7094 void __init
sched_init(void)
7097 unsigned long alloc_size
= 0, ptr
;
7099 #ifdef CONFIG_FAIR_GROUP_SCHED
7100 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7102 #ifdef CONFIG_RT_GROUP_SCHED
7103 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7105 #ifdef CONFIG_CPUMASK_OFFSTACK
7106 alloc_size
+= num_possible_cpus() * cpumask_size();
7109 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7111 #ifdef CONFIG_FAIR_GROUP_SCHED
7112 root_task_group
.se
= (struct sched_entity
**)ptr
;
7113 ptr
+= nr_cpu_ids
* sizeof(void **);
7115 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7116 ptr
+= nr_cpu_ids
* sizeof(void **);
7118 #endif /* CONFIG_FAIR_GROUP_SCHED */
7119 #ifdef CONFIG_RT_GROUP_SCHED
7120 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7121 ptr
+= nr_cpu_ids
* sizeof(void **);
7123 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7124 ptr
+= nr_cpu_ids
* sizeof(void **);
7126 #endif /* CONFIG_RT_GROUP_SCHED */
7127 #ifdef CONFIG_CPUMASK_OFFSTACK
7128 for_each_possible_cpu(i
) {
7129 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
7130 ptr
+= cpumask_size();
7132 #endif /* CONFIG_CPUMASK_OFFSTACK */
7135 init_rt_bandwidth(&def_rt_bandwidth
,
7136 global_rt_period(), global_rt_runtime());
7137 init_dl_bandwidth(&def_dl_bandwidth
,
7138 global_rt_period(), global_rt_runtime());
7141 init_defrootdomain();
7144 #ifdef CONFIG_RT_GROUP_SCHED
7145 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7146 global_rt_period(), global_rt_runtime());
7147 #endif /* CONFIG_RT_GROUP_SCHED */
7149 #ifdef CONFIG_CGROUP_SCHED
7150 list_add(&root_task_group
.list
, &task_groups
);
7151 INIT_LIST_HEAD(&root_task_group
.children
);
7152 INIT_LIST_HEAD(&root_task_group
.siblings
);
7153 autogroup_init(&init_task
);
7155 #endif /* CONFIG_CGROUP_SCHED */
7157 for_each_possible_cpu(i
) {
7161 raw_spin_lock_init(&rq
->lock
);
7163 rq
->calc_load_active
= 0;
7164 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7165 init_cfs_rq(&rq
->cfs
);
7166 init_rt_rq(&rq
->rt
, rq
);
7167 init_dl_rq(&rq
->dl
, rq
);
7168 #ifdef CONFIG_FAIR_GROUP_SCHED
7169 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7170 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7172 * How much cpu bandwidth does root_task_group get?
7174 * In case of task-groups formed thr' the cgroup filesystem, it
7175 * gets 100% of the cpu resources in the system. This overall
7176 * system cpu resource is divided among the tasks of
7177 * root_task_group and its child task-groups in a fair manner,
7178 * based on each entity's (task or task-group's) weight
7179 * (se->load.weight).
7181 * In other words, if root_task_group has 10 tasks of weight
7182 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7183 * then A0's share of the cpu resource is:
7185 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7187 * We achieve this by letting root_task_group's tasks sit
7188 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7190 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7191 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7192 #endif /* CONFIG_FAIR_GROUP_SCHED */
7194 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7195 #ifdef CONFIG_RT_GROUP_SCHED
7196 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7199 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7200 rq
->cpu_load
[j
] = 0;
7202 rq
->last_load_update_tick
= jiffies
;
7207 rq
->cpu_capacity
= SCHED_CAPACITY_SCALE
;
7208 rq
->post_schedule
= 0;
7209 rq
->active_balance
= 0;
7210 rq
->next_balance
= jiffies
;
7215 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7216 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7218 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7220 rq_attach_root(rq
, &def_root_domain
);
7221 #ifdef CONFIG_NO_HZ_COMMON
7224 #ifdef CONFIG_NO_HZ_FULL
7225 rq
->last_sched_tick
= 0;
7229 atomic_set(&rq
->nr_iowait
, 0);
7232 set_load_weight(&init_task
);
7234 #ifdef CONFIG_PREEMPT_NOTIFIERS
7235 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7239 * The boot idle thread does lazy MMU switching as well:
7241 atomic_inc(&init_mm
.mm_count
);
7242 enter_lazy_tlb(&init_mm
, current
);
7245 * Make us the idle thread. Technically, schedule() should not be
7246 * called from this thread, however somewhere below it might be,
7247 * but because we are the idle thread, we just pick up running again
7248 * when this runqueue becomes "idle".
7250 init_idle(current
, smp_processor_id());
7252 calc_load_update
= jiffies
+ LOAD_FREQ
;
7255 * During early bootup we pretend to be a normal task:
7257 current
->sched_class
= &fair_sched_class
;
7260 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7261 /* May be allocated at isolcpus cmdline parse time */
7262 if (cpu_isolated_map
== NULL
)
7263 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7264 idle_thread_set_boot_cpu();
7265 set_cpu_rq_start_time();
7267 init_sched_fair_class();
7269 scheduler_running
= 1;
7272 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7273 static inline int preempt_count_equals(int preempt_offset
)
7275 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7277 return (nested
== preempt_offset
);
7280 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7283 * Blocking primitives will set (and therefore destroy) current->state,
7284 * since we will exit with TASK_RUNNING make sure we enter with it,
7285 * otherwise we will destroy state.
7287 if (WARN_ONCE(current
->state
!= TASK_RUNNING
,
7288 "do not call blocking ops when !TASK_RUNNING; "
7289 "state=%lx set at [<%p>] %pS\n",
7291 (void *)current
->task_state_change
,
7292 (void *)current
->task_state_change
))
7293 __set_current_state(TASK_RUNNING
);
7295 ___might_sleep(file
, line
, preempt_offset
);
7297 EXPORT_SYMBOL(__might_sleep
);
7299 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7301 static unsigned long prev_jiffy
; /* ratelimiting */
7303 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7304 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7305 !is_idle_task(current
)) ||
7306 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7308 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7310 prev_jiffy
= jiffies
;
7313 "BUG: sleeping function called from invalid context at %s:%d\n",
7316 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7317 in_atomic(), irqs_disabled(),
7318 current
->pid
, current
->comm
);
7320 debug_show_held_locks(current
);
7321 if (irqs_disabled())
7322 print_irqtrace_events(current
);
7323 #ifdef CONFIG_DEBUG_PREEMPT
7324 if (!preempt_count_equals(preempt_offset
)) {
7325 pr_err("Preemption disabled at:");
7326 print_ip_sym(current
->preempt_disable_ip
);
7332 EXPORT_SYMBOL(___might_sleep
);
7335 #ifdef CONFIG_MAGIC_SYSRQ
7336 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7338 const struct sched_class
*prev_class
= p
->sched_class
;
7339 struct sched_attr attr
= {
7340 .sched_policy
= SCHED_NORMAL
,
7342 int old_prio
= p
->prio
;
7345 queued
= task_on_rq_queued(p
);
7347 dequeue_task(rq
, p
, 0);
7348 __setscheduler(rq
, p
, &attr
);
7350 enqueue_task(rq
, p
, 0);
7354 check_class_changed(rq
, p
, prev_class
, old_prio
);
7357 void normalize_rt_tasks(void)
7359 struct task_struct
*g
, *p
;
7360 unsigned long flags
;
7363 read_lock(&tasklist_lock
);
7364 for_each_process_thread(g
, p
) {
7366 * Only normalize user tasks:
7368 if (p
->flags
& PF_KTHREAD
)
7371 p
->se
.exec_start
= 0;
7372 #ifdef CONFIG_SCHEDSTATS
7373 p
->se
.statistics
.wait_start
= 0;
7374 p
->se
.statistics
.sleep_start
= 0;
7375 p
->se
.statistics
.block_start
= 0;
7378 if (!dl_task(p
) && !rt_task(p
)) {
7380 * Renice negative nice level userspace
7383 if (task_nice(p
) < 0)
7384 set_user_nice(p
, 0);
7388 rq
= task_rq_lock(p
, &flags
);
7389 normalize_task(rq
, p
);
7390 task_rq_unlock(rq
, p
, &flags
);
7392 read_unlock(&tasklist_lock
);
7395 #endif /* CONFIG_MAGIC_SYSRQ */
7397 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7399 * These functions are only useful for the IA64 MCA handling, or kdb.
7401 * They can only be called when the whole system has been
7402 * stopped - every CPU needs to be quiescent, and no scheduling
7403 * activity can take place. Using them for anything else would
7404 * be a serious bug, and as a result, they aren't even visible
7405 * under any other configuration.
7409 * curr_task - return the current task for a given cpu.
7410 * @cpu: the processor in question.
7412 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7414 * Return: The current task for @cpu.
7416 struct task_struct
*curr_task(int cpu
)
7418 return cpu_curr(cpu
);
7421 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7425 * set_curr_task - set the current task for a given cpu.
7426 * @cpu: the processor in question.
7427 * @p: the task pointer to set.
7429 * Description: This function must only be used when non-maskable interrupts
7430 * are serviced on a separate stack. It allows the architecture to switch the
7431 * notion of the current task on a cpu in a non-blocking manner. This function
7432 * must be called with all CPU's synchronized, and interrupts disabled, the
7433 * and caller must save the original value of the current task (see
7434 * curr_task() above) and restore that value before reenabling interrupts and
7435 * re-starting the system.
7437 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7439 void set_curr_task(int cpu
, struct task_struct
*p
)
7446 #ifdef CONFIG_CGROUP_SCHED
7447 /* task_group_lock serializes the addition/removal of task groups */
7448 static DEFINE_SPINLOCK(task_group_lock
);
7450 static void free_sched_group(struct task_group
*tg
)
7452 free_fair_sched_group(tg
);
7453 free_rt_sched_group(tg
);
7458 /* allocate runqueue etc for a new task group */
7459 struct task_group
*sched_create_group(struct task_group
*parent
)
7461 struct task_group
*tg
;
7463 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7465 return ERR_PTR(-ENOMEM
);
7467 if (!alloc_fair_sched_group(tg
, parent
))
7470 if (!alloc_rt_sched_group(tg
, parent
))
7476 free_sched_group(tg
);
7477 return ERR_PTR(-ENOMEM
);
7480 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7482 unsigned long flags
;
7484 spin_lock_irqsave(&task_group_lock
, flags
);
7485 list_add_rcu(&tg
->list
, &task_groups
);
7487 WARN_ON(!parent
); /* root should already exist */
7489 tg
->parent
= parent
;
7490 INIT_LIST_HEAD(&tg
->children
);
7491 list_add_rcu(&tg
->siblings
, &parent
->children
);
7492 spin_unlock_irqrestore(&task_group_lock
, flags
);
7495 /* rcu callback to free various structures associated with a task group */
7496 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7498 /* now it should be safe to free those cfs_rqs */
7499 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7502 /* Destroy runqueue etc associated with a task group */
7503 void sched_destroy_group(struct task_group
*tg
)
7505 /* wait for possible concurrent references to cfs_rqs complete */
7506 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7509 void sched_offline_group(struct task_group
*tg
)
7511 unsigned long flags
;
7514 /* end participation in shares distribution */
7515 for_each_possible_cpu(i
)
7516 unregister_fair_sched_group(tg
, i
);
7518 spin_lock_irqsave(&task_group_lock
, flags
);
7519 list_del_rcu(&tg
->list
);
7520 list_del_rcu(&tg
->siblings
);
7521 spin_unlock_irqrestore(&task_group_lock
, flags
);
7524 /* change task's runqueue when it moves between groups.
7525 * The caller of this function should have put the task in its new group
7526 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7527 * reflect its new group.
7529 void sched_move_task(struct task_struct
*tsk
)
7531 struct task_group
*tg
;
7532 int queued
, running
;
7533 unsigned long flags
;
7536 rq
= task_rq_lock(tsk
, &flags
);
7538 running
= task_current(rq
, tsk
);
7539 queued
= task_on_rq_queued(tsk
);
7542 dequeue_task(rq
, tsk
, 0);
7543 if (unlikely(running
))
7544 put_prev_task(rq
, tsk
);
7547 * All callers are synchronized by task_rq_lock(); we do not use RCU
7548 * which is pointless here. Thus, we pass "true" to task_css_check()
7549 * to prevent lockdep warnings.
7551 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7552 struct task_group
, css
);
7553 tg
= autogroup_task_group(tsk
, tg
);
7554 tsk
->sched_task_group
= tg
;
7556 #ifdef CONFIG_FAIR_GROUP_SCHED
7557 if (tsk
->sched_class
->task_move_group
)
7558 tsk
->sched_class
->task_move_group(tsk
, queued
);
7561 set_task_rq(tsk
, task_cpu(tsk
));
7563 if (unlikely(running
))
7564 tsk
->sched_class
->set_curr_task(rq
);
7566 enqueue_task(rq
, tsk
, 0);
7568 task_rq_unlock(rq
, tsk
, &flags
);
7570 #endif /* CONFIG_CGROUP_SCHED */
7572 #ifdef CONFIG_RT_GROUP_SCHED
7574 * Ensure that the real time constraints are schedulable.
7576 static DEFINE_MUTEX(rt_constraints_mutex
);
7578 /* Must be called with tasklist_lock held */
7579 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7581 struct task_struct
*g
, *p
;
7583 for_each_process_thread(g
, p
) {
7584 if (rt_task(p
) && task_group(p
) == tg
)
7591 struct rt_schedulable_data
{
7592 struct task_group
*tg
;
7597 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7599 struct rt_schedulable_data
*d
= data
;
7600 struct task_group
*child
;
7601 unsigned long total
, sum
= 0;
7602 u64 period
, runtime
;
7604 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7605 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7608 period
= d
->rt_period
;
7609 runtime
= d
->rt_runtime
;
7613 * Cannot have more runtime than the period.
7615 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7619 * Ensure we don't starve existing RT tasks.
7621 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7624 total
= to_ratio(period
, runtime
);
7627 * Nobody can have more than the global setting allows.
7629 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7633 * The sum of our children's runtime should not exceed our own.
7635 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7636 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7637 runtime
= child
->rt_bandwidth
.rt_runtime
;
7639 if (child
== d
->tg
) {
7640 period
= d
->rt_period
;
7641 runtime
= d
->rt_runtime
;
7644 sum
+= to_ratio(period
, runtime
);
7653 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7657 struct rt_schedulable_data data
= {
7659 .rt_period
= period
,
7660 .rt_runtime
= runtime
,
7664 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7670 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7671 u64 rt_period
, u64 rt_runtime
)
7675 mutex_lock(&rt_constraints_mutex
);
7676 read_lock(&tasklist_lock
);
7677 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7681 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7682 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7683 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7685 for_each_possible_cpu(i
) {
7686 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7688 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7689 rt_rq
->rt_runtime
= rt_runtime
;
7690 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7692 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7694 read_unlock(&tasklist_lock
);
7695 mutex_unlock(&rt_constraints_mutex
);
7700 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7702 u64 rt_runtime
, rt_period
;
7704 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7705 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7706 if (rt_runtime_us
< 0)
7707 rt_runtime
= RUNTIME_INF
;
7709 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7712 static long sched_group_rt_runtime(struct task_group
*tg
)
7716 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7719 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7720 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7721 return rt_runtime_us
;
7724 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7726 u64 rt_runtime
, rt_period
;
7728 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7729 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7734 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7737 static long sched_group_rt_period(struct task_group
*tg
)
7741 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7742 do_div(rt_period_us
, NSEC_PER_USEC
);
7743 return rt_period_us
;
7745 #endif /* CONFIG_RT_GROUP_SCHED */
7747 #ifdef CONFIG_RT_GROUP_SCHED
7748 static int sched_rt_global_constraints(void)
7752 mutex_lock(&rt_constraints_mutex
);
7753 read_lock(&tasklist_lock
);
7754 ret
= __rt_schedulable(NULL
, 0, 0);
7755 read_unlock(&tasklist_lock
);
7756 mutex_unlock(&rt_constraints_mutex
);
7761 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7763 /* Don't accept realtime tasks when there is no way for them to run */
7764 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7770 #else /* !CONFIG_RT_GROUP_SCHED */
7771 static int sched_rt_global_constraints(void)
7773 unsigned long flags
;
7776 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7777 for_each_possible_cpu(i
) {
7778 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7780 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7781 rt_rq
->rt_runtime
= global_rt_runtime();
7782 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7784 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7788 #endif /* CONFIG_RT_GROUP_SCHED */
7790 static int sched_dl_global_constraints(void)
7792 u64 runtime
= global_rt_runtime();
7793 u64 period
= global_rt_period();
7794 u64 new_bw
= to_ratio(period
, runtime
);
7797 unsigned long flags
;
7800 * Here we want to check the bandwidth not being set to some
7801 * value smaller than the currently allocated bandwidth in
7802 * any of the root_domains.
7804 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7805 * cycling on root_domains... Discussion on different/better
7806 * solutions is welcome!
7808 for_each_possible_cpu(cpu
) {
7809 rcu_read_lock_sched();
7810 dl_b
= dl_bw_of(cpu
);
7812 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7813 if (new_bw
< dl_b
->total_bw
)
7815 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7817 rcu_read_unlock_sched();
7826 static void sched_dl_do_global(void)
7831 unsigned long flags
;
7833 def_dl_bandwidth
.dl_period
= global_rt_period();
7834 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7836 if (global_rt_runtime() != RUNTIME_INF
)
7837 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7840 * FIXME: As above...
7842 for_each_possible_cpu(cpu
) {
7843 rcu_read_lock_sched();
7844 dl_b
= dl_bw_of(cpu
);
7846 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7848 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7850 rcu_read_unlock_sched();
7854 static int sched_rt_global_validate(void)
7856 if (sysctl_sched_rt_period
<= 0)
7859 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7860 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7866 static void sched_rt_do_global(void)
7868 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7869 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7872 int sched_rt_handler(struct ctl_table
*table
, int write
,
7873 void __user
*buffer
, size_t *lenp
,
7876 int old_period
, old_runtime
;
7877 static DEFINE_MUTEX(mutex
);
7881 old_period
= sysctl_sched_rt_period
;
7882 old_runtime
= sysctl_sched_rt_runtime
;
7884 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7886 if (!ret
&& write
) {
7887 ret
= sched_rt_global_validate();
7891 ret
= sched_rt_global_constraints();
7895 ret
= sched_dl_global_constraints();
7899 sched_rt_do_global();
7900 sched_dl_do_global();
7904 sysctl_sched_rt_period
= old_period
;
7905 sysctl_sched_rt_runtime
= old_runtime
;
7907 mutex_unlock(&mutex
);
7912 int sched_rr_handler(struct ctl_table
*table
, int write
,
7913 void __user
*buffer
, size_t *lenp
,
7917 static DEFINE_MUTEX(mutex
);
7920 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7921 /* make sure that internally we keep jiffies */
7922 /* also, writing zero resets timeslice to default */
7923 if (!ret
&& write
) {
7924 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7925 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7927 mutex_unlock(&mutex
);
7931 #ifdef CONFIG_CGROUP_SCHED
7933 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7935 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7938 static struct cgroup_subsys_state
*
7939 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7941 struct task_group
*parent
= css_tg(parent_css
);
7942 struct task_group
*tg
;
7945 /* This is early initialization for the top cgroup */
7946 return &root_task_group
.css
;
7949 tg
= sched_create_group(parent
);
7951 return ERR_PTR(-ENOMEM
);
7956 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7958 struct task_group
*tg
= css_tg(css
);
7959 struct task_group
*parent
= css_tg(css
->parent
);
7962 sched_online_group(tg
, parent
);
7966 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7968 struct task_group
*tg
= css_tg(css
);
7970 sched_destroy_group(tg
);
7973 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7975 struct task_group
*tg
= css_tg(css
);
7977 sched_offline_group(tg
);
7980 static void cpu_cgroup_fork(struct task_struct
*task
)
7982 sched_move_task(task
);
7985 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7986 struct cgroup_taskset
*tset
)
7988 struct task_struct
*task
;
7990 cgroup_taskset_for_each(task
, tset
) {
7991 #ifdef CONFIG_RT_GROUP_SCHED
7992 if (!sched_rt_can_attach(css_tg(css
), task
))
7995 /* We don't support RT-tasks being in separate groups */
7996 if (task
->sched_class
!= &fair_sched_class
)
8003 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8004 struct cgroup_taskset
*tset
)
8006 struct task_struct
*task
;
8008 cgroup_taskset_for_each(task
, tset
)
8009 sched_move_task(task
);
8012 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8013 struct cgroup_subsys_state
*old_css
,
8014 struct task_struct
*task
)
8017 * cgroup_exit() is called in the copy_process() failure path.
8018 * Ignore this case since the task hasn't ran yet, this avoids
8019 * trying to poke a half freed task state from generic code.
8021 if (!(task
->flags
& PF_EXITING
))
8024 sched_move_task(task
);
8027 #ifdef CONFIG_FAIR_GROUP_SCHED
8028 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8029 struct cftype
*cftype
, u64 shareval
)
8031 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8034 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8037 struct task_group
*tg
= css_tg(css
);
8039 return (u64
) scale_load_down(tg
->shares
);
8042 #ifdef CONFIG_CFS_BANDWIDTH
8043 static DEFINE_MUTEX(cfs_constraints_mutex
);
8045 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8046 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8048 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8050 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8052 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8053 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8055 if (tg
== &root_task_group
)
8059 * Ensure we have at some amount of bandwidth every period. This is
8060 * to prevent reaching a state of large arrears when throttled via
8061 * entity_tick() resulting in prolonged exit starvation.
8063 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8067 * Likewise, bound things on the otherside by preventing insane quota
8068 * periods. This also allows us to normalize in computing quota
8071 if (period
> max_cfs_quota_period
)
8075 * Prevent race between setting of cfs_rq->runtime_enabled and
8076 * unthrottle_offline_cfs_rqs().
8079 mutex_lock(&cfs_constraints_mutex
);
8080 ret
= __cfs_schedulable(tg
, period
, quota
);
8084 runtime_enabled
= quota
!= RUNTIME_INF
;
8085 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8087 * If we need to toggle cfs_bandwidth_used, off->on must occur
8088 * before making related changes, and on->off must occur afterwards
8090 if (runtime_enabled
&& !runtime_was_enabled
)
8091 cfs_bandwidth_usage_inc();
8092 raw_spin_lock_irq(&cfs_b
->lock
);
8093 cfs_b
->period
= ns_to_ktime(period
);
8094 cfs_b
->quota
= quota
;
8096 __refill_cfs_bandwidth_runtime(cfs_b
);
8097 /* restart the period timer (if active) to handle new period expiry */
8098 if (runtime_enabled
&& cfs_b
->timer_active
) {
8099 /* force a reprogram */
8100 __start_cfs_bandwidth(cfs_b
, true);
8102 raw_spin_unlock_irq(&cfs_b
->lock
);
8104 for_each_online_cpu(i
) {
8105 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8106 struct rq
*rq
= cfs_rq
->rq
;
8108 raw_spin_lock_irq(&rq
->lock
);
8109 cfs_rq
->runtime_enabled
= runtime_enabled
;
8110 cfs_rq
->runtime_remaining
= 0;
8112 if (cfs_rq
->throttled
)
8113 unthrottle_cfs_rq(cfs_rq
);
8114 raw_spin_unlock_irq(&rq
->lock
);
8116 if (runtime_was_enabled
&& !runtime_enabled
)
8117 cfs_bandwidth_usage_dec();
8119 mutex_unlock(&cfs_constraints_mutex
);
8125 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8129 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8130 if (cfs_quota_us
< 0)
8131 quota
= RUNTIME_INF
;
8133 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8135 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8138 long tg_get_cfs_quota(struct task_group
*tg
)
8142 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8145 quota_us
= tg
->cfs_bandwidth
.quota
;
8146 do_div(quota_us
, NSEC_PER_USEC
);
8151 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8155 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8156 quota
= tg
->cfs_bandwidth
.quota
;
8158 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8161 long tg_get_cfs_period(struct task_group
*tg
)
8165 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8166 do_div(cfs_period_us
, NSEC_PER_USEC
);
8168 return cfs_period_us
;
8171 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8174 return tg_get_cfs_quota(css_tg(css
));
8177 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8178 struct cftype
*cftype
, s64 cfs_quota_us
)
8180 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8183 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8186 return tg_get_cfs_period(css_tg(css
));
8189 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8190 struct cftype
*cftype
, u64 cfs_period_us
)
8192 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8195 struct cfs_schedulable_data
{
8196 struct task_group
*tg
;
8201 * normalize group quota/period to be quota/max_period
8202 * note: units are usecs
8204 static u64
normalize_cfs_quota(struct task_group
*tg
,
8205 struct cfs_schedulable_data
*d
)
8213 period
= tg_get_cfs_period(tg
);
8214 quota
= tg_get_cfs_quota(tg
);
8217 /* note: these should typically be equivalent */
8218 if (quota
== RUNTIME_INF
|| quota
== -1)
8221 return to_ratio(period
, quota
);
8224 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8226 struct cfs_schedulable_data
*d
= data
;
8227 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8228 s64 quota
= 0, parent_quota
= -1;
8231 quota
= RUNTIME_INF
;
8233 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8235 quota
= normalize_cfs_quota(tg
, d
);
8236 parent_quota
= parent_b
->hierarchical_quota
;
8239 * ensure max(child_quota) <= parent_quota, inherit when no
8242 if (quota
== RUNTIME_INF
)
8243 quota
= parent_quota
;
8244 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8247 cfs_b
->hierarchical_quota
= quota
;
8252 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8255 struct cfs_schedulable_data data
= {
8261 if (quota
!= RUNTIME_INF
) {
8262 do_div(data
.period
, NSEC_PER_USEC
);
8263 do_div(data
.quota
, NSEC_PER_USEC
);
8267 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8273 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8275 struct task_group
*tg
= css_tg(seq_css(sf
));
8276 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8278 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8279 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8280 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8284 #endif /* CONFIG_CFS_BANDWIDTH */
8285 #endif /* CONFIG_FAIR_GROUP_SCHED */
8287 #ifdef CONFIG_RT_GROUP_SCHED
8288 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8289 struct cftype
*cft
, s64 val
)
8291 return sched_group_set_rt_runtime(css_tg(css
), val
);
8294 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8297 return sched_group_rt_runtime(css_tg(css
));
8300 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8301 struct cftype
*cftype
, u64 rt_period_us
)
8303 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8306 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8309 return sched_group_rt_period(css_tg(css
));
8311 #endif /* CONFIG_RT_GROUP_SCHED */
8313 static struct cftype cpu_files
[] = {
8314 #ifdef CONFIG_FAIR_GROUP_SCHED
8317 .read_u64
= cpu_shares_read_u64
,
8318 .write_u64
= cpu_shares_write_u64
,
8321 #ifdef CONFIG_CFS_BANDWIDTH
8323 .name
= "cfs_quota_us",
8324 .read_s64
= cpu_cfs_quota_read_s64
,
8325 .write_s64
= cpu_cfs_quota_write_s64
,
8328 .name
= "cfs_period_us",
8329 .read_u64
= cpu_cfs_period_read_u64
,
8330 .write_u64
= cpu_cfs_period_write_u64
,
8334 .seq_show
= cpu_stats_show
,
8337 #ifdef CONFIG_RT_GROUP_SCHED
8339 .name
= "rt_runtime_us",
8340 .read_s64
= cpu_rt_runtime_read
,
8341 .write_s64
= cpu_rt_runtime_write
,
8344 .name
= "rt_period_us",
8345 .read_u64
= cpu_rt_period_read_uint
,
8346 .write_u64
= cpu_rt_period_write_uint
,
8352 struct cgroup_subsys cpu_cgrp_subsys
= {
8353 .css_alloc
= cpu_cgroup_css_alloc
,
8354 .css_free
= cpu_cgroup_css_free
,
8355 .css_online
= cpu_cgroup_css_online
,
8356 .css_offline
= cpu_cgroup_css_offline
,
8357 .fork
= cpu_cgroup_fork
,
8358 .can_attach
= cpu_cgroup_can_attach
,
8359 .attach
= cpu_cgroup_attach
,
8360 .exit
= cpu_cgroup_exit
,
8361 .legacy_cftypes
= cpu_files
,
8365 #endif /* CONFIG_CGROUP_SCHED */
8367 void dump_cpu_task(int cpu
)
8369 pr_info("Task dump for CPU %d:\n", cpu
);
8370 sched_show_task(cpu_curr(cpu
));