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>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
95 ktime_t soft
, hard
, now
;
98 if (hrtimer_active(period_timer
))
101 now
= hrtimer_cb_get_time(period_timer
);
102 hrtimer_forward(period_timer
, now
, period
);
104 soft
= hrtimer_get_softexpires(period_timer
);
105 hard
= hrtimer_get_expires(period_timer
);
106 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
107 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
108 HRTIMER_MODE_ABS_PINNED
, 0);
112 DEFINE_MUTEX(sched_domains_mutex
);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
115 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
117 void update_rq_clock(struct rq
*rq
)
121 if (rq
->skip_clock_update
> 0)
124 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
126 update_rq_clock_task(rq
, delta
);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug
unsigned int sysctl_sched_features
=
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names
[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp
)
201 if (strncmp(cmp
, "NO_", 3) == 0) {
206 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
207 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
209 sysctl_sched_features
&= ~(1UL << i
);
210 sched_feat_disable(i
);
212 sysctl_sched_features
|= (1UL << i
);
213 sched_feat_enable(i
);
223 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
224 size_t cnt
, loff_t
*ppos
)
233 if (copy_from_user(&buf
, ubuf
, cnt
))
239 i
= sched_feat_set(cmp
);
240 if (i
== __SCHED_FEAT_NR
)
248 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
250 return single_open(filp
, sched_feat_show
, NULL
);
253 static const struct file_operations sched_feat_fops
= {
254 .open
= sched_feat_open
,
255 .write
= sched_feat_write
,
258 .release
= single_release
,
261 static __init
int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
268 late_initcall(sched_init_debug
);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
278 * period over which we average the RT time consumption, measured
283 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period
= 1000000;
291 __read_mostly
int scheduler_running
;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime
= 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
309 lockdep_assert_held(&p
->pi_lock
);
313 raw_spin_lock(&rq
->lock
);
314 if (likely(rq
== task_rq(p
)))
316 raw_spin_unlock(&rq
->lock
);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
324 __acquires(p
->pi_lock
)
330 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
332 raw_spin_lock(&rq
->lock
);
333 if (likely(rq
== task_rq(p
)))
335 raw_spin_unlock(&rq
->lock
);
336 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
340 static void __task_rq_unlock(struct rq
*rq
)
343 raw_spin_unlock(&rq
->lock
);
347 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
349 __releases(p
->pi_lock
)
351 raw_spin_unlock(&rq
->lock
);
352 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq
*this_rq_lock(void)
365 raw_spin_lock(&rq
->lock
);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
375 static void hrtick_clear(struct rq
*rq
)
377 if (hrtimer_active(&rq
->hrtick_timer
))
378 hrtimer_cancel(&rq
->hrtick_timer
);
382 * High-resolution timer tick.
383 * Runs from hardirq context with interrupts disabled.
385 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
387 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
389 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
391 raw_spin_lock(&rq
->lock
);
393 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
394 raw_spin_unlock(&rq
->lock
);
396 return HRTIMER_NORESTART
;
401 static int __hrtick_restart(struct rq
*rq
)
403 struct hrtimer
*timer
= &rq
->hrtick_timer
;
404 ktime_t time
= hrtimer_get_softexpires(timer
);
406 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
410 * called from hardirq (IPI) context
412 static void __hrtick_start(void *arg
)
416 raw_spin_lock(&rq
->lock
);
417 __hrtick_restart(rq
);
418 rq
->hrtick_csd_pending
= 0;
419 raw_spin_unlock(&rq
->lock
);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq
*rq
, u64 delay
)
429 struct hrtimer
*timer
= &rq
->hrtick_timer
;
430 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
432 hrtimer_set_expires(timer
, time
);
434 if (rq
== this_rq()) {
435 __hrtick_restart(rq
);
436 } else if (!rq
->hrtick_csd_pending
) {
437 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
438 rq
->hrtick_csd_pending
= 1;
443 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
445 int cpu
= (int)(long)hcpu
;
448 case CPU_UP_CANCELED
:
449 case CPU_UP_CANCELED_FROZEN
:
450 case CPU_DOWN_PREPARE
:
451 case CPU_DOWN_PREPARE_FROZEN
:
453 case CPU_DEAD_FROZEN
:
454 hrtick_clear(cpu_rq(cpu
));
461 static __init
void init_hrtick(void)
463 hotcpu_notifier(hotplug_hrtick
, 0);
467 * Called to set the hrtick timer state.
469 * called with rq->lock held and irqs disabled
471 void hrtick_start(struct rq
*rq
, u64 delay
)
473 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
474 HRTIMER_MODE_REL_PINNED
, 0);
477 static inline void init_hrtick(void)
480 #endif /* CONFIG_SMP */
482 static void init_rq_hrtick(struct rq
*rq
)
485 rq
->hrtick_csd_pending
= 0;
487 rq
->hrtick_csd
.flags
= 0;
488 rq
->hrtick_csd
.func
= __hrtick_start
;
489 rq
->hrtick_csd
.info
= rq
;
492 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
493 rq
->hrtick_timer
.function
= hrtick
;
495 #else /* CONFIG_SCHED_HRTICK */
496 static inline void hrtick_clear(struct rq
*rq
)
500 static inline void init_rq_hrtick(struct rq
*rq
)
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SCHED_HRTICK */
510 * resched_task - mark a task 'to be rescheduled now'.
512 * On UP this means the setting of the need_resched flag, on SMP it
513 * might also involve a cross-CPU call to trigger the scheduler on
517 void resched_task(struct task_struct
*p
)
521 assert_raw_spin_locked(&task_rq(p
)->lock
);
523 if (test_tsk_need_resched(p
))
526 set_tsk_need_resched(p
);
529 if (cpu
== smp_processor_id())
532 /* NEED_RESCHED must be visible before we test polling */
534 if (!tsk_is_polling(p
))
535 smp_send_reschedule(cpu
);
538 void resched_cpu(int cpu
)
540 struct rq
*rq
= cpu_rq(cpu
);
543 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
545 resched_task(cpu_curr(cpu
));
546 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
549 #ifdef CONFIG_NO_HZ_COMMON
551 * In the semi idle case, use the nearest busy cpu for migrating timers
552 * from an idle cpu. This is good for power-savings.
554 * We don't do similar optimization for completely idle system, as
555 * selecting an idle cpu will add more delays to the timers than intended
556 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 int get_nohz_timer_target(void)
560 int cpu
= smp_processor_id();
562 struct sched_domain
*sd
;
565 for_each_domain(cpu
, sd
) {
566 for_each_cpu(i
, sched_domain_span(sd
)) {
578 * When add_timer_on() enqueues a timer into the timer wheel of an
579 * idle CPU then this timer might expire before the next timer event
580 * which is scheduled to wake up that CPU. In case of a completely
581 * idle system the next event might even be infinite time into the
582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583 * leaves the inner idle loop so the newly added timer is taken into
584 * account when the CPU goes back to idle and evaluates the timer
585 * wheel for the next timer event.
587 static void wake_up_idle_cpu(int cpu
)
589 struct rq
*rq
= cpu_rq(cpu
);
591 if (cpu
== smp_processor_id())
595 * This is safe, as this function is called with the timer
596 * wheel base lock of (cpu) held. When the CPU is on the way
597 * to idle and has not yet set rq->curr to idle then it will
598 * be serialized on the timer wheel base lock and take the new
599 * timer into account automatically.
601 if (rq
->curr
!= rq
->idle
)
605 * We can set TIF_RESCHED on the idle task of the other CPU
606 * lockless. The worst case is that the other CPU runs the
607 * idle task through an additional NOOP schedule()
609 set_tsk_need_resched(rq
->idle
);
611 /* NEED_RESCHED must be visible before we test polling */
613 if (!tsk_is_polling(rq
->idle
))
614 smp_send_reschedule(cpu
);
617 static bool wake_up_full_nohz_cpu(int cpu
)
619 if (tick_nohz_full_cpu(cpu
)) {
620 if (cpu
!= smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 smp_send_reschedule(cpu
);
629 void wake_up_nohz_cpu(int cpu
)
631 if (!wake_up_full_nohz_cpu(cpu
))
632 wake_up_idle_cpu(cpu
);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu
= smp_processor_id();
639 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
642 if (idle_cpu(cpu
) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
669 /* Make sure rq->nr_running update is visible after the IPI */
672 /* More than one running task need preemption */
673 if (rq
->nr_running
> 1)
678 #endif /* CONFIG_NO_HZ_FULL */
680 void sched_avg_update(struct rq
*rq
)
682 s64 period
= sched_avg_period();
684 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
686 * Inline assembly required to prevent the compiler
687 * optimising this loop into a divmod call.
688 * See __iter_div_u64_rem() for another example of this.
690 asm("" : "+rm" (rq
->age_stamp
));
691 rq
->age_stamp
+= period
;
696 #else /* !CONFIG_SMP */
697 void resched_task(struct task_struct
*p
)
699 assert_raw_spin_locked(&task_rq(p
)->lock
);
700 set_tsk_need_resched(p
);
702 #endif /* CONFIG_SMP */
704 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
705 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
707 * Iterate task_group tree rooted at *from, calling @down when first entering a
708 * node and @up when leaving it for the final time.
710 * Caller must hold rcu_lock or sufficient equivalent.
712 int walk_tg_tree_from(struct task_group
*from
,
713 tg_visitor down
, tg_visitor up
, void *data
)
715 struct task_group
*parent
, *child
;
721 ret
= (*down
)(parent
, data
);
724 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
731 ret
= (*up
)(parent
, data
);
732 if (ret
|| parent
== from
)
736 parent
= parent
->parent
;
743 int tg_nop(struct task_group
*tg
, void *data
)
749 static void set_load_weight(struct task_struct
*p
)
751 int prio
= p
->static_prio
- MAX_RT_PRIO
;
752 struct load_weight
*load
= &p
->se
.load
;
755 * SCHED_IDLE tasks get minimal weight:
757 if (p
->policy
== SCHED_IDLE
) {
758 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
759 load
->inv_weight
= WMULT_IDLEPRIO
;
763 load
->weight
= scale_load(prio_to_weight
[prio
]);
764 load
->inv_weight
= prio_to_wmult
[prio
];
767 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
770 sched_info_queued(p
);
771 p
->sched_class
->enqueue_task(rq
, p
, flags
);
774 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
777 sched_info_dequeued(p
);
778 p
->sched_class
->dequeue_task(rq
, p
, flags
);
781 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
783 if (task_contributes_to_load(p
))
784 rq
->nr_uninterruptible
--;
786 enqueue_task(rq
, p
, flags
);
789 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
791 if (task_contributes_to_load(p
))
792 rq
->nr_uninterruptible
++;
794 dequeue_task(rq
, p
, flags
);
797 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
800 * In theory, the compile should just see 0 here, and optimize out the call
801 * to sched_rt_avg_update. But I don't trust it...
803 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
804 s64 steal
= 0, irq_delta
= 0;
806 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
807 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
810 * Since irq_time is only updated on {soft,}irq_exit, we might run into
811 * this case when a previous update_rq_clock() happened inside a
814 * When this happens, we stop ->clock_task and only update the
815 * prev_irq_time stamp to account for the part that fit, so that a next
816 * update will consume the rest. This ensures ->clock_task is
819 * It does however cause some slight miss-attribution of {soft,}irq
820 * time, a more accurate solution would be to update the irq_time using
821 * the current rq->clock timestamp, except that would require using
824 if (irq_delta
> delta
)
827 rq
->prev_irq_time
+= irq_delta
;
830 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
831 if (static_key_false((¶virt_steal_rq_enabled
))) {
834 steal
= paravirt_steal_clock(cpu_of(rq
));
835 steal
-= rq
->prev_steal_time_rq
;
837 if (unlikely(steal
> delta
))
840 st
= steal_ticks(steal
);
841 steal
= st
* TICK_NSEC
;
843 rq
->prev_steal_time_rq
+= steal
;
849 rq
->clock_task
+= delta
;
851 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
852 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
853 sched_rt_avg_update(rq
, irq_delta
+ steal
);
857 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
859 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
860 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
864 * Make it appear like a SCHED_FIFO task, its something
865 * userspace knows about and won't get confused about.
867 * Also, it will make PI more or less work without too
868 * much confusion -- but then, stop work should not
869 * rely on PI working anyway.
871 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
873 stop
->sched_class
= &stop_sched_class
;
876 cpu_rq(cpu
)->stop
= stop
;
880 * Reset it back to a normal scheduling class so that
881 * it can die in pieces.
883 old_stop
->sched_class
= &rt_sched_class
;
888 * __normal_prio - return the priority that is based on the static prio
890 static inline int __normal_prio(struct task_struct
*p
)
892 return p
->static_prio
;
896 * Calculate the expected normal priority: i.e. priority
897 * without taking RT-inheritance into account. Might be
898 * boosted by interactivity modifiers. Changes upon fork,
899 * setprio syscalls, and whenever the interactivity
900 * estimator recalculates.
902 static inline int normal_prio(struct task_struct
*p
)
906 if (task_has_rt_policy(p
))
907 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
909 prio
= __normal_prio(p
);
914 * Calculate the current priority, i.e. the priority
915 * taken into account by the scheduler. This value might
916 * be boosted by RT tasks, or might be boosted by
917 * interactivity modifiers. Will be RT if the task got
918 * RT-boosted. If not then it returns p->normal_prio.
920 static int effective_prio(struct task_struct
*p
)
922 p
->normal_prio
= normal_prio(p
);
924 * If we are RT tasks or we were boosted to RT priority,
925 * keep the priority unchanged. Otherwise, update priority
926 * to the normal priority:
928 if (!rt_prio(p
->prio
))
929 return p
->normal_prio
;
934 * task_curr - is this task currently executing on a CPU?
935 * @p: the task in question.
937 * Return: 1 if the task is currently executing. 0 otherwise.
939 inline int task_curr(const struct task_struct
*p
)
941 return cpu_curr(task_cpu(p
)) == p
;
944 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
945 const struct sched_class
*prev_class
,
948 if (prev_class
!= p
->sched_class
) {
949 if (prev_class
->switched_from
)
950 prev_class
->switched_from(rq
, p
);
951 p
->sched_class
->switched_to(rq
, p
);
952 } else if (oldprio
!= p
->prio
)
953 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
956 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
958 const struct sched_class
*class;
960 if (p
->sched_class
== rq
->curr
->sched_class
) {
961 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
963 for_each_class(class) {
964 if (class == rq
->curr
->sched_class
)
966 if (class == p
->sched_class
) {
967 resched_task(rq
->curr
);
974 * A queue event has occurred, and we're going to schedule. In
975 * this case, we can save a useless back to back clock update.
977 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
978 rq
->skip_clock_update
= 1;
981 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier
);
983 void register_task_migration_notifier(struct notifier_block
*n
)
985 atomic_notifier_chain_register(&task_migration_notifier
, n
);
989 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
991 #ifdef CONFIG_SCHED_DEBUG
993 * We should never call set_task_cpu() on a blocked task,
994 * ttwu() will sort out the placement.
996 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
997 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
999 #ifdef CONFIG_LOCKDEP
1001 * The caller should hold either p->pi_lock or rq->lock, when changing
1002 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1004 * sched_move_task() holds both and thus holding either pins the cgroup,
1007 * Furthermore, all task_rq users should acquire both locks, see
1010 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1011 lockdep_is_held(&task_rq(p
)->lock
)));
1015 trace_sched_migrate_task(p
, new_cpu
);
1017 if (task_cpu(p
) != new_cpu
) {
1018 struct task_migration_notifier tmn
;
1020 if (p
->sched_class
->migrate_task_rq
)
1021 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1022 p
->se
.nr_migrations
++;
1023 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1026 tmn
.from_cpu
= task_cpu(p
);
1027 tmn
.to_cpu
= new_cpu
;
1029 atomic_notifier_call_chain(&task_migration_notifier
, 0, &tmn
);
1032 __set_task_cpu(p
, new_cpu
);
1035 struct migration_arg
{
1036 struct task_struct
*task
;
1040 static int migration_cpu_stop(void *data
);
1043 * wait_task_inactive - wait for a thread to unschedule.
1045 * If @match_state is nonzero, it's the @p->state value just checked and
1046 * not expected to change. If it changes, i.e. @p might have woken up,
1047 * then return zero. When we succeed in waiting for @p to be off its CPU,
1048 * we return a positive number (its total switch count). If a second call
1049 * a short while later returns the same number, the caller can be sure that
1050 * @p has remained unscheduled the whole time.
1052 * The caller must ensure that the task *will* unschedule sometime soon,
1053 * else this function might spin for a *long* time. This function can't
1054 * be called with interrupts off, or it may introduce deadlock with
1055 * smp_call_function() if an IPI is sent by the same process we are
1056 * waiting to become inactive.
1058 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1060 unsigned long flags
;
1067 * We do the initial early heuristics without holding
1068 * any task-queue locks at all. We'll only try to get
1069 * the runqueue lock when things look like they will
1075 * If the task is actively running on another CPU
1076 * still, just relax and busy-wait without holding
1079 * NOTE! Since we don't hold any locks, it's not
1080 * even sure that "rq" stays as the right runqueue!
1081 * But we don't care, since "task_running()" will
1082 * return false if the runqueue has changed and p
1083 * is actually now running somewhere else!
1085 while (task_running(rq
, p
)) {
1086 if (match_state
&& unlikely(p
->state
!= match_state
))
1092 * Ok, time to look more closely! We need the rq
1093 * lock now, to be *sure*. If we're wrong, we'll
1094 * just go back and repeat.
1096 rq
= task_rq_lock(p
, &flags
);
1097 trace_sched_wait_task(p
);
1098 running
= task_running(rq
, p
);
1101 if (!match_state
|| p
->state
== match_state
)
1102 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1103 task_rq_unlock(rq
, p
, &flags
);
1106 * If it changed from the expected state, bail out now.
1108 if (unlikely(!ncsw
))
1112 * Was it really running after all now that we
1113 * checked with the proper locks actually held?
1115 * Oops. Go back and try again..
1117 if (unlikely(running
)) {
1123 * It's not enough that it's not actively running,
1124 * it must be off the runqueue _entirely_, and not
1127 * So if it was still runnable (but just not actively
1128 * running right now), it's preempted, and we should
1129 * yield - it could be a while.
1131 if (unlikely(on_rq
)) {
1132 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1134 set_current_state(TASK_UNINTERRUPTIBLE
);
1135 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1140 * Ahh, all good. It wasn't running, and it wasn't
1141 * runnable, which means that it will never become
1142 * running in the future either. We're all done!
1151 * kick_process - kick a running thread to enter/exit the kernel
1152 * @p: the to-be-kicked thread
1154 * Cause a process which is running on another CPU to enter
1155 * kernel-mode, without any delay. (to get signals handled.)
1157 * NOTE: this function doesn't have to take the runqueue lock,
1158 * because all it wants to ensure is that the remote task enters
1159 * the kernel. If the IPI races and the task has been migrated
1160 * to another CPU then no harm is done and the purpose has been
1163 void kick_process(struct task_struct
*p
)
1169 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1170 smp_send_reschedule(cpu
);
1173 EXPORT_SYMBOL_GPL(kick_process
);
1174 #endif /* CONFIG_SMP */
1178 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1180 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1182 int nid
= cpu_to_node(cpu
);
1183 const struct cpumask
*nodemask
= NULL
;
1184 enum { cpuset
, possible
, fail
} state
= cpuset
;
1188 * If the node that the cpu is on has been offlined, cpu_to_node()
1189 * will return -1. There is no cpu on the node, and we should
1190 * select the cpu on the other node.
1193 nodemask
= cpumask_of_node(nid
);
1195 /* Look for allowed, online CPU in same node. */
1196 for_each_cpu(dest_cpu
, nodemask
) {
1197 if (!cpu_online(dest_cpu
))
1199 if (!cpu_active(dest_cpu
))
1201 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1207 /* Any allowed, online CPU? */
1208 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1209 if (!cpu_online(dest_cpu
))
1211 if (!cpu_active(dest_cpu
))
1218 /* No more Mr. Nice Guy. */
1219 cpuset_cpus_allowed_fallback(p
);
1224 do_set_cpus_allowed(p
, cpu_possible_mask
);
1235 if (state
!= cpuset
) {
1237 * Don't tell them about moving exiting tasks or
1238 * kernel threads (both mm NULL), since they never
1241 if (p
->mm
&& printk_ratelimit()) {
1242 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1243 task_pid_nr(p
), p
->comm
, cpu
);
1251 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1254 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
1256 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
1259 * In order not to call set_task_cpu() on a blocking task we need
1260 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1263 * Since this is common to all placement strategies, this lives here.
1265 * [ this allows ->select_task() to simply return task_cpu(p) and
1266 * not worry about this generic constraint ]
1268 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1270 cpu
= select_fallback_rq(task_cpu(p
), p
);
1275 static void update_avg(u64
*avg
, u64 sample
)
1277 s64 diff
= sample
- *avg
;
1283 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1285 #ifdef CONFIG_SCHEDSTATS
1286 struct rq
*rq
= this_rq();
1289 int this_cpu
= smp_processor_id();
1291 if (cpu
== this_cpu
) {
1292 schedstat_inc(rq
, ttwu_local
);
1293 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1295 struct sched_domain
*sd
;
1297 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1299 for_each_domain(this_cpu
, sd
) {
1300 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1301 schedstat_inc(sd
, ttwu_wake_remote
);
1308 if (wake_flags
& WF_MIGRATED
)
1309 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1311 #endif /* CONFIG_SMP */
1313 schedstat_inc(rq
, ttwu_count
);
1314 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1316 if (wake_flags
& WF_SYNC
)
1317 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1319 #endif /* CONFIG_SCHEDSTATS */
1322 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1324 activate_task(rq
, p
, en_flags
);
1327 /* if a worker is waking up, notify workqueue */
1328 if (p
->flags
& PF_WQ_WORKER
)
1329 wq_worker_waking_up(p
, cpu_of(rq
));
1333 * Mark the task runnable and perform wakeup-preemption.
1336 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1338 check_preempt_curr(rq
, p
, wake_flags
);
1339 trace_sched_wakeup(p
, true);
1341 p
->state
= TASK_RUNNING
;
1343 if (p
->sched_class
->task_woken
)
1344 p
->sched_class
->task_woken(rq
, p
);
1346 if (rq
->idle_stamp
) {
1347 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1348 u64 max
= 2*sysctl_sched_migration_cost
;
1353 update_avg(&rq
->avg_idle
, delta
);
1360 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1363 if (p
->sched_contributes_to_load
)
1364 rq
->nr_uninterruptible
--;
1367 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1368 ttwu_do_wakeup(rq
, p
, wake_flags
);
1372 * Called in case the task @p isn't fully descheduled from its runqueue,
1373 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1374 * since all we need to do is flip p->state to TASK_RUNNING, since
1375 * the task is still ->on_rq.
1377 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1382 rq
= __task_rq_lock(p
);
1384 /* check_preempt_curr() may use rq clock */
1385 update_rq_clock(rq
);
1386 ttwu_do_wakeup(rq
, p
, wake_flags
);
1389 __task_rq_unlock(rq
);
1395 static void sched_ttwu_pending(void)
1397 struct rq
*rq
= this_rq();
1398 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1399 struct task_struct
*p
;
1401 raw_spin_lock(&rq
->lock
);
1404 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1405 llist
= llist_next(llist
);
1406 ttwu_do_activate(rq
, p
, 0);
1409 raw_spin_unlock(&rq
->lock
);
1412 void scheduler_ipi(void)
1414 if (llist_empty(&this_rq()->wake_list
)
1415 && !tick_nohz_full_cpu(smp_processor_id())
1416 && !got_nohz_idle_kick())
1420 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1421 * traditionally all their work was done from the interrupt return
1422 * path. Now that we actually do some work, we need to make sure
1425 * Some archs already do call them, luckily irq_enter/exit nest
1428 * Arguably we should visit all archs and update all handlers,
1429 * however a fair share of IPIs are still resched only so this would
1430 * somewhat pessimize the simple resched case.
1433 tick_nohz_full_check();
1434 sched_ttwu_pending();
1437 * Check if someone kicked us for doing the nohz idle load balance.
1439 if (unlikely(got_nohz_idle_kick())) {
1440 this_rq()->idle_balance
= 1;
1441 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1446 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1448 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1449 smp_send_reschedule(cpu
);
1452 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1454 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1456 #endif /* CONFIG_SMP */
1458 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1460 struct rq
*rq
= cpu_rq(cpu
);
1462 #if defined(CONFIG_SMP)
1463 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1464 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1465 ttwu_queue_remote(p
, cpu
);
1470 raw_spin_lock(&rq
->lock
);
1471 ttwu_do_activate(rq
, p
, 0);
1472 raw_spin_unlock(&rq
->lock
);
1476 * try_to_wake_up - wake up a thread
1477 * @p: the thread to be awakened
1478 * @state: the mask of task states that can be woken
1479 * @wake_flags: wake modifier flags (WF_*)
1481 * Put it on the run-queue if it's not already there. The "current"
1482 * thread is always on the run-queue (except when the actual
1483 * re-schedule is in progress), and as such you're allowed to do
1484 * the simpler "current->state = TASK_RUNNING" to mark yourself
1485 * runnable without the overhead of this.
1487 * Return: %true if @p was woken up, %false if it was already running.
1488 * or @state didn't match @p's state.
1491 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1493 unsigned long flags
;
1494 int cpu
, success
= 0;
1497 * If we are going to wake up a thread waiting for CONDITION we
1498 * need to ensure that CONDITION=1 done by the caller can not be
1499 * reordered with p->state check below. This pairs with mb() in
1500 * set_current_state() the waiting thread does.
1502 smp_mb__before_spinlock();
1503 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1504 if (!(p
->state
& state
))
1507 success
= 1; /* we're going to change ->state */
1510 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1515 * If the owning (remote) cpu is still in the middle of schedule() with
1516 * this task as prev, wait until its done referencing the task.
1521 * Pairs with the smp_wmb() in finish_lock_switch().
1525 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1526 p
->state
= TASK_WAKING
;
1528 if (p
->sched_class
->task_waking
)
1529 p
->sched_class
->task_waking(p
);
1531 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
1532 if (task_cpu(p
) != cpu
) {
1533 wake_flags
|= WF_MIGRATED
;
1534 set_task_cpu(p
, cpu
);
1536 #endif /* CONFIG_SMP */
1540 ttwu_stat(p
, cpu
, wake_flags
);
1542 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1548 * try_to_wake_up_local - try to wake up a local task with rq lock held
1549 * @p: the thread to be awakened
1551 * Put @p on the run-queue if it's not already there. The caller must
1552 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1555 static void try_to_wake_up_local(struct task_struct
*p
)
1557 struct rq
*rq
= task_rq(p
);
1559 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1560 WARN_ON_ONCE(p
== current
))
1563 lockdep_assert_held(&rq
->lock
);
1565 if (!raw_spin_trylock(&p
->pi_lock
)) {
1566 raw_spin_unlock(&rq
->lock
);
1567 raw_spin_lock(&p
->pi_lock
);
1568 raw_spin_lock(&rq
->lock
);
1571 if (!(p
->state
& TASK_NORMAL
))
1575 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1577 ttwu_do_wakeup(rq
, p
, 0);
1578 ttwu_stat(p
, smp_processor_id(), 0);
1580 raw_spin_unlock(&p
->pi_lock
);
1584 * wake_up_process - Wake up a specific process
1585 * @p: The process to be woken up.
1587 * Attempt to wake up the nominated process and move it to the set of runnable
1590 * Return: 1 if the process was woken up, 0 if it was already running.
1592 * It may be assumed that this function implies a write memory barrier before
1593 * changing the task state if and only if any tasks are woken up.
1595 int wake_up_process(struct task_struct
*p
)
1597 WARN_ON(task_is_stopped_or_traced(p
));
1598 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1600 EXPORT_SYMBOL(wake_up_process
);
1602 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1604 return try_to_wake_up(p
, state
, 0);
1608 * Perform scheduler related setup for a newly forked process p.
1609 * p is forked by current.
1611 * __sched_fork() is basic setup used by init_idle() too:
1613 static void __sched_fork(struct task_struct
*p
)
1618 p
->se
.exec_start
= 0;
1619 p
->se
.sum_exec_runtime
= 0;
1620 p
->se
.prev_sum_exec_runtime
= 0;
1621 p
->se
.nr_migrations
= 0;
1623 INIT_LIST_HEAD(&p
->se
.group_node
);
1625 #ifdef CONFIG_SCHEDSTATS
1626 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1629 INIT_LIST_HEAD(&p
->rt
.run_list
);
1631 #ifdef CONFIG_PREEMPT_NOTIFIERS
1632 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1635 #ifdef CONFIG_NUMA_BALANCING
1636 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1637 p
->mm
->numa_next_scan
= jiffies
;
1638 p
->mm
->numa_next_reset
= jiffies
;
1639 p
->mm
->numa_scan_seq
= 0;
1642 p
->node_stamp
= 0ULL;
1643 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1644 p
->numa_migrate_seq
= p
->mm
? p
->mm
->numa_scan_seq
- 1 : 0;
1645 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1646 p
->numa_work
.next
= &p
->numa_work
;
1647 #endif /* CONFIG_NUMA_BALANCING */
1650 #ifdef CONFIG_NUMA_BALANCING
1651 #ifdef CONFIG_SCHED_DEBUG
1652 void set_numabalancing_state(bool enabled
)
1655 sched_feat_set("NUMA");
1657 sched_feat_set("NO_NUMA");
1660 __read_mostly
bool numabalancing_enabled
;
1662 void set_numabalancing_state(bool enabled
)
1664 numabalancing_enabled
= enabled
;
1666 #endif /* CONFIG_SCHED_DEBUG */
1667 #endif /* CONFIG_NUMA_BALANCING */
1670 * fork()/clone()-time setup:
1672 void sched_fork(struct task_struct
*p
)
1674 unsigned long flags
;
1675 int cpu
= get_cpu();
1679 * We mark the process as running here. This guarantees that
1680 * nobody will actually run it, and a signal or other external
1681 * event cannot wake it up and insert it on the runqueue either.
1683 p
->state
= TASK_RUNNING
;
1686 * Make sure we do not leak PI boosting priority to the child.
1688 p
->prio
= current
->normal_prio
;
1691 * Revert to default priority/policy on fork if requested.
1693 if (unlikely(p
->sched_reset_on_fork
)) {
1694 if (task_has_rt_policy(p
)) {
1695 p
->policy
= SCHED_NORMAL
;
1696 p
->static_prio
= NICE_TO_PRIO(0);
1698 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1699 p
->static_prio
= NICE_TO_PRIO(0);
1701 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1705 * We don't need the reset flag anymore after the fork. It has
1706 * fulfilled its duty:
1708 p
->sched_reset_on_fork
= 0;
1711 if (!rt_prio(p
->prio
))
1712 p
->sched_class
= &fair_sched_class
;
1714 if (p
->sched_class
->task_fork
)
1715 p
->sched_class
->task_fork(p
);
1718 * The child is not yet in the pid-hash so no cgroup attach races,
1719 * and the cgroup is pinned to this child due to cgroup_fork()
1720 * is ran before sched_fork().
1722 * Silence PROVE_RCU.
1724 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1725 set_task_cpu(p
, cpu
);
1726 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1728 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1729 if (likely(sched_info_on()))
1730 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1732 #if defined(CONFIG_SMP)
1735 #ifdef CONFIG_PREEMPT_COUNT
1736 /* Want to start with kernel preemption disabled. */
1737 task_thread_info(p
)->preempt_count
= 1;
1740 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1747 * wake_up_new_task - wake up a newly created task for the first time.
1749 * This function will do some initial scheduler statistics housekeeping
1750 * that must be done for every newly created context, then puts the task
1751 * on the runqueue and wakes it.
1753 void wake_up_new_task(struct task_struct
*p
)
1755 unsigned long flags
;
1758 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1761 * Fork balancing, do it here and not earlier because:
1762 * - cpus_allowed can change in the fork path
1763 * - any previously selected cpu might disappear through hotplug
1765 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
1768 /* Initialize new task's runnable average */
1769 init_task_runnable_average(p
);
1770 rq
= __task_rq_lock(p
);
1771 activate_task(rq
, p
, 0);
1773 trace_sched_wakeup_new(p
, true);
1774 check_preempt_curr(rq
, p
, WF_FORK
);
1776 if (p
->sched_class
->task_woken
)
1777 p
->sched_class
->task_woken(rq
, p
);
1779 task_rq_unlock(rq
, p
, &flags
);
1782 #ifdef CONFIG_PREEMPT_NOTIFIERS
1785 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1786 * @notifier: notifier struct to register
1788 void preempt_notifier_register(struct preempt_notifier
*notifier
)
1790 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
1792 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
1795 * preempt_notifier_unregister - no longer interested in preemption notifications
1796 * @notifier: notifier struct to unregister
1798 * This is safe to call from within a preemption notifier.
1800 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
1802 hlist_del(¬ifier
->link
);
1804 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
1806 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1808 struct preempt_notifier
*notifier
;
1810 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1811 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
1815 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1816 struct task_struct
*next
)
1818 struct preempt_notifier
*notifier
;
1820 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
1821 notifier
->ops
->sched_out(notifier
, next
);
1824 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1826 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
1831 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
1832 struct task_struct
*next
)
1836 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1839 * prepare_task_switch - prepare to switch tasks
1840 * @rq: the runqueue preparing to switch
1841 * @prev: the current task that is being switched out
1842 * @next: the task we are going to switch to.
1844 * This is called with the rq lock held and interrupts off. It must
1845 * be paired with a subsequent finish_task_switch after the context
1848 * prepare_task_switch sets up locking and calls architecture specific
1852 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
1853 struct task_struct
*next
)
1855 trace_sched_switch(prev
, next
);
1856 sched_info_switch(prev
, next
);
1857 perf_event_task_sched_out(prev
, next
);
1858 fire_sched_out_preempt_notifiers(prev
, next
);
1859 prepare_lock_switch(rq
, next
);
1860 prepare_arch_switch(next
);
1864 * finish_task_switch - clean up after a task-switch
1865 * @rq: runqueue associated with task-switch
1866 * @prev: the thread we just switched away from.
1868 * finish_task_switch must be called after the context switch, paired
1869 * with a prepare_task_switch call before the context switch.
1870 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1871 * and do any other architecture-specific cleanup actions.
1873 * Note that we may have delayed dropping an mm in context_switch(). If
1874 * so, we finish that here outside of the runqueue lock. (Doing it
1875 * with the lock held can cause deadlocks; see schedule() for
1878 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1879 __releases(rq
->lock
)
1881 struct mm_struct
*mm
= rq
->prev_mm
;
1887 * A task struct has one reference for the use as "current".
1888 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1889 * schedule one last time. The schedule call will never return, and
1890 * the scheduled task must drop that reference.
1891 * The test for TASK_DEAD must occur while the runqueue locks are
1892 * still held, otherwise prev could be scheduled on another cpu, die
1893 * there before we look at prev->state, and then the reference would
1895 * Manfred Spraul <manfred@colorfullife.com>
1897 prev_state
= prev
->state
;
1898 vtime_task_switch(prev
);
1899 finish_arch_switch(prev
);
1900 perf_event_task_sched_in(prev
, current
);
1901 finish_lock_switch(rq
, prev
);
1902 finish_arch_post_lock_switch();
1904 fire_sched_in_preempt_notifiers(current
);
1907 if (unlikely(prev_state
== TASK_DEAD
)) {
1909 * Remove function-return probe instances associated with this
1910 * task and put them back on the free list.
1912 kprobe_flush_task(prev
);
1913 put_task_struct(prev
);
1916 tick_nohz_task_switch(current
);
1921 /* assumes rq->lock is held */
1922 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
1924 if (prev
->sched_class
->pre_schedule
)
1925 prev
->sched_class
->pre_schedule(rq
, prev
);
1928 /* rq->lock is NOT held, but preemption is disabled */
1929 static inline void post_schedule(struct rq
*rq
)
1931 if (rq
->post_schedule
) {
1932 unsigned long flags
;
1934 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1935 if (rq
->curr
->sched_class
->post_schedule
)
1936 rq
->curr
->sched_class
->post_schedule(rq
);
1937 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1939 rq
->post_schedule
= 0;
1945 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
1949 static inline void post_schedule(struct rq
*rq
)
1956 * schedule_tail - first thing a freshly forked thread must call.
1957 * @prev: the thread we just switched away from.
1959 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1960 __releases(rq
->lock
)
1962 struct rq
*rq
= this_rq();
1964 finish_task_switch(rq
, prev
);
1967 * FIXME: do we need to worry about rq being invalidated by the
1972 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1973 /* In this case, finish_task_switch does not reenable preemption */
1976 if (current
->set_child_tid
)
1977 put_user(task_pid_vnr(current
), current
->set_child_tid
);
1981 * context_switch - switch to the new MM and the new
1982 * thread's register state.
1985 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1986 struct task_struct
*next
)
1988 struct mm_struct
*mm
, *oldmm
;
1990 prepare_task_switch(rq
, prev
, next
);
1993 oldmm
= prev
->active_mm
;
1995 * For paravirt, this is coupled with an exit in switch_to to
1996 * combine the page table reload and the switch backend into
1999 arch_start_context_switch(prev
);
2002 next
->active_mm
= oldmm
;
2003 atomic_inc(&oldmm
->mm_count
);
2004 enter_lazy_tlb(oldmm
, next
);
2006 switch_mm(oldmm
, mm
, next
);
2009 prev
->active_mm
= NULL
;
2010 rq
->prev_mm
= oldmm
;
2013 * Since the runqueue lock will be released by the next
2014 * task (which is an invalid locking op but in the case
2015 * of the scheduler it's an obvious special-case), so we
2016 * do an early lockdep release here:
2018 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2019 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2022 context_tracking_task_switch(prev
, next
);
2023 /* Here we just switch the register state and the stack. */
2024 switch_to(prev
, next
, prev
);
2028 * this_rq must be evaluated again because prev may have moved
2029 * CPUs since it called schedule(), thus the 'rq' on its stack
2030 * frame will be invalid.
2032 finish_task_switch(this_rq(), prev
);
2036 * nr_running and nr_context_switches:
2038 * externally visible scheduler statistics: current number of runnable
2039 * threads, total number of context switches performed since bootup.
2041 unsigned long nr_running(void)
2043 unsigned long i
, sum
= 0;
2045 for_each_online_cpu(i
)
2046 sum
+= cpu_rq(i
)->nr_running
;
2051 unsigned long long nr_context_switches(void)
2054 unsigned long long sum
= 0;
2056 for_each_possible_cpu(i
)
2057 sum
+= cpu_rq(i
)->nr_switches
;
2062 unsigned long nr_iowait(void)
2064 unsigned long i
, sum
= 0;
2066 for_each_possible_cpu(i
)
2067 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2072 unsigned long nr_iowait_cpu(int cpu
)
2074 struct rq
*this = cpu_rq(cpu
);
2075 return atomic_read(&this->nr_iowait
);
2081 * sched_exec - execve() is a valuable balancing opportunity, because at
2082 * this point the task has the smallest effective memory and cache footprint.
2084 void sched_exec(void)
2086 struct task_struct
*p
= current
;
2087 unsigned long flags
;
2090 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2091 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
2092 if (dest_cpu
== smp_processor_id())
2095 if (likely(cpu_active(dest_cpu
))) {
2096 struct migration_arg arg
= { p
, dest_cpu
};
2098 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2099 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2103 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2108 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2109 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2111 EXPORT_PER_CPU_SYMBOL(kstat
);
2112 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2115 * Return any ns on the sched_clock that have not yet been accounted in
2116 * @p in case that task is currently running.
2118 * Called with task_rq_lock() held on @rq.
2120 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2124 if (task_current(rq
, p
)) {
2125 update_rq_clock(rq
);
2126 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2134 unsigned long long task_delta_exec(struct task_struct
*p
)
2136 unsigned long flags
;
2140 rq
= task_rq_lock(p
, &flags
);
2141 ns
= do_task_delta_exec(p
, rq
);
2142 task_rq_unlock(rq
, p
, &flags
);
2148 * Return accounted runtime for the task.
2149 * In case the task is currently running, return the runtime plus current's
2150 * pending runtime that have not been accounted yet.
2152 unsigned long long task_sched_runtime(struct task_struct
*p
)
2154 unsigned long flags
;
2158 rq
= task_rq_lock(p
, &flags
);
2159 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2160 task_rq_unlock(rq
, p
, &flags
);
2166 * This function gets called by the timer code, with HZ frequency.
2167 * We call it with interrupts disabled.
2169 void scheduler_tick(void)
2171 int cpu
= smp_processor_id();
2172 struct rq
*rq
= cpu_rq(cpu
);
2173 struct task_struct
*curr
= rq
->curr
;
2177 raw_spin_lock(&rq
->lock
);
2178 update_rq_clock(rq
);
2179 curr
->sched_class
->task_tick(rq
, curr
, 0);
2180 update_cpu_load_active(rq
);
2181 raw_spin_unlock(&rq
->lock
);
2183 perf_event_task_tick();
2186 rq
->idle_balance
= idle_cpu(cpu
);
2187 trigger_load_balance(rq
, cpu
);
2189 rq_last_tick_reset(rq
);
2192 #ifdef CONFIG_NO_HZ_FULL
2194 * scheduler_tick_max_deferment
2196 * Keep at least one tick per second when a single
2197 * active task is running because the scheduler doesn't
2198 * yet completely support full dynticks environment.
2200 * This makes sure that uptime, CFS vruntime, load
2201 * balancing, etc... continue to move forward, even
2202 * with a very low granularity.
2204 * Return: Maximum deferment in nanoseconds.
2206 u64
scheduler_tick_max_deferment(void)
2208 struct rq
*rq
= this_rq();
2209 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2211 next
= rq
->last_sched_tick
+ HZ
;
2213 if (time_before_eq(next
, now
))
2216 return jiffies_to_usecs(next
- now
) * NSEC_PER_USEC
;
2220 notrace
unsigned long get_parent_ip(unsigned long addr
)
2222 if (in_lock_functions(addr
)) {
2223 addr
= CALLER_ADDR2
;
2224 if (in_lock_functions(addr
))
2225 addr
= CALLER_ADDR3
;
2230 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2231 defined(CONFIG_PREEMPT_TRACER))
2233 void __kprobes
add_preempt_count(int val
)
2235 #ifdef CONFIG_DEBUG_PREEMPT
2239 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2242 preempt_count() += val
;
2243 #ifdef CONFIG_DEBUG_PREEMPT
2245 * Spinlock count overflowing soon?
2247 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2250 if (preempt_count() == val
)
2251 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2253 EXPORT_SYMBOL(add_preempt_count
);
2255 void __kprobes
sub_preempt_count(int val
)
2257 #ifdef CONFIG_DEBUG_PREEMPT
2261 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2264 * Is the spinlock portion underflowing?
2266 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2267 !(preempt_count() & PREEMPT_MASK
)))
2271 if (preempt_count() == val
)
2272 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2273 preempt_count() -= val
;
2275 EXPORT_SYMBOL(sub_preempt_count
);
2280 * Print scheduling while atomic bug:
2282 static noinline
void __schedule_bug(struct task_struct
*prev
)
2284 if (oops_in_progress
)
2287 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2288 prev
->comm
, prev
->pid
, preempt_count());
2290 debug_show_held_locks(prev
);
2292 if (irqs_disabled())
2293 print_irqtrace_events(prev
);
2295 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2299 * Various schedule()-time debugging checks and statistics:
2301 static inline void schedule_debug(struct task_struct
*prev
)
2304 * Test if we are atomic. Since do_exit() needs to call into
2305 * schedule() atomically, we ignore that path for now.
2306 * Otherwise, whine if we are scheduling when we should not be.
2308 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
2309 __schedule_bug(prev
);
2312 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2314 schedstat_inc(this_rq(), sched_count
);
2317 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2319 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2320 update_rq_clock(rq
);
2321 prev
->sched_class
->put_prev_task(rq
, prev
);
2325 * Pick up the highest-prio task:
2327 static inline struct task_struct
*
2328 pick_next_task(struct rq
*rq
)
2330 const struct sched_class
*class;
2331 struct task_struct
*p
;
2334 * Optimization: we know that if all tasks are in
2335 * the fair class we can call that function directly:
2337 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2338 p
= fair_sched_class
.pick_next_task(rq
);
2343 for_each_class(class) {
2344 p
= class->pick_next_task(rq
);
2349 BUG(); /* the idle class will always have a runnable task */
2353 * __schedule() is the main scheduler function.
2355 * The main means of driving the scheduler and thus entering this function are:
2357 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2359 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2360 * paths. For example, see arch/x86/entry_64.S.
2362 * To drive preemption between tasks, the scheduler sets the flag in timer
2363 * interrupt handler scheduler_tick().
2365 * 3. Wakeups don't really cause entry into schedule(). They add a
2366 * task to the run-queue and that's it.
2368 * Now, if the new task added to the run-queue preempts the current
2369 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2370 * called on the nearest possible occasion:
2372 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2374 * - in syscall or exception context, at the next outmost
2375 * preempt_enable(). (this might be as soon as the wake_up()'s
2378 * - in IRQ context, return from interrupt-handler to
2379 * preemptible context
2381 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2384 * - cond_resched() call
2385 * - explicit schedule() call
2386 * - return from syscall or exception to user-space
2387 * - return from interrupt-handler to user-space
2389 static void __sched
__schedule(void)
2391 struct task_struct
*prev
, *next
;
2392 unsigned long *switch_count
;
2398 cpu
= smp_processor_id();
2400 rcu_note_context_switch(cpu
);
2403 schedule_debug(prev
);
2405 if (sched_feat(HRTICK
))
2409 * Make sure that signal_pending_state()->signal_pending() below
2410 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2411 * done by the caller to avoid the race with signal_wake_up().
2413 smp_mb__before_spinlock();
2414 raw_spin_lock_irq(&rq
->lock
);
2416 switch_count
= &prev
->nivcsw
;
2417 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2418 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2419 prev
->state
= TASK_RUNNING
;
2421 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2425 * If a worker went to sleep, notify and ask workqueue
2426 * whether it wants to wake up a task to maintain
2429 if (prev
->flags
& PF_WQ_WORKER
) {
2430 struct task_struct
*to_wakeup
;
2432 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2434 try_to_wake_up_local(to_wakeup
);
2437 switch_count
= &prev
->nvcsw
;
2440 pre_schedule(rq
, prev
);
2442 if (unlikely(!rq
->nr_running
))
2443 idle_balance(cpu
, rq
);
2445 put_prev_task(rq
, prev
);
2446 next
= pick_next_task(rq
);
2447 clear_tsk_need_resched(prev
);
2448 rq
->skip_clock_update
= 0;
2450 if (likely(prev
!= next
)) {
2455 context_switch(rq
, prev
, next
); /* unlocks the rq */
2457 * The context switch have flipped the stack from under us
2458 * and restored the local variables which were saved when
2459 * this task called schedule() in the past. prev == current
2460 * is still correct, but it can be moved to another cpu/rq.
2462 cpu
= smp_processor_id();
2465 raw_spin_unlock_irq(&rq
->lock
);
2469 sched_preempt_enable_no_resched();
2474 static inline void sched_submit_work(struct task_struct
*tsk
)
2476 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2479 * If we are going to sleep and we have plugged IO queued,
2480 * make sure to submit it to avoid deadlocks.
2482 if (blk_needs_flush_plug(tsk
))
2483 blk_schedule_flush_plug(tsk
);
2486 asmlinkage
void __sched
schedule(void)
2488 struct task_struct
*tsk
= current
;
2490 sched_submit_work(tsk
);
2493 EXPORT_SYMBOL(schedule
);
2495 #ifdef CONFIG_CONTEXT_TRACKING
2496 asmlinkage
void __sched
schedule_user(void)
2499 * If we come here after a random call to set_need_resched(),
2500 * or we have been woken up remotely but the IPI has not yet arrived,
2501 * we haven't yet exited the RCU idle mode. Do it here manually until
2502 * we find a better solution.
2511 * schedule_preempt_disabled - called with preemption disabled
2513 * Returns with preemption disabled. Note: preempt_count must be 1
2515 void __sched
schedule_preempt_disabled(void)
2517 sched_preempt_enable_no_resched();
2522 #ifdef CONFIG_PREEMPT
2524 * this is the entry point to schedule() from in-kernel preemption
2525 * off of preempt_enable. Kernel preemptions off return from interrupt
2526 * occur there and call schedule directly.
2528 asmlinkage
void __sched notrace
preempt_schedule(void)
2531 * If there is a non-zero preempt_count or interrupts are disabled,
2532 * we do not want to preempt the current task. Just return..
2534 if (likely(!preemptible()))
2538 add_preempt_count_notrace(PREEMPT_ACTIVE
);
2540 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
2543 * Check again in case we missed a preemption opportunity
2544 * between schedule and now.
2547 } while (need_resched());
2549 EXPORT_SYMBOL(preempt_schedule
);
2552 * this is the entry point to schedule() from kernel preemption
2553 * off of irq context.
2554 * Note, that this is called and return with irqs disabled. This will
2555 * protect us against recursive calling from irq.
2557 asmlinkage
void __sched
preempt_schedule_irq(void)
2559 struct thread_info
*ti
= current_thread_info();
2560 enum ctx_state prev_state
;
2562 /* Catch callers which need to be fixed */
2563 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
2565 prev_state
= exception_enter();
2568 add_preempt_count(PREEMPT_ACTIVE
);
2571 local_irq_disable();
2572 sub_preempt_count(PREEMPT_ACTIVE
);
2575 * Check again in case we missed a preemption opportunity
2576 * between schedule and now.
2579 } while (need_resched());
2581 exception_exit(prev_state
);
2584 #endif /* CONFIG_PREEMPT */
2586 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2589 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2591 EXPORT_SYMBOL(default_wake_function
);
2594 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2595 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2596 * number) then we wake all the non-exclusive tasks and one exclusive task.
2598 * There are circumstances in which we can try to wake a task which has already
2599 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2600 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2602 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
2603 int nr_exclusive
, int wake_flags
, void *key
)
2605 wait_queue_t
*curr
, *next
;
2607 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
2608 unsigned flags
= curr
->flags
;
2610 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
2611 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
2617 * __wake_up - wake up threads blocked on a waitqueue.
2619 * @mode: which threads
2620 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2621 * @key: is directly passed to the wakeup function
2623 * It may be assumed that this function implies a write memory barrier before
2624 * changing the task state if and only if any tasks are woken up.
2626 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
2627 int nr_exclusive
, void *key
)
2629 unsigned long flags
;
2631 spin_lock_irqsave(&q
->lock
, flags
);
2632 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
2633 spin_unlock_irqrestore(&q
->lock
, flags
);
2635 EXPORT_SYMBOL(__wake_up
);
2638 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2640 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
, int nr
)
2642 __wake_up_common(q
, mode
, nr
, 0, NULL
);
2644 EXPORT_SYMBOL_GPL(__wake_up_locked
);
2646 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
2648 __wake_up_common(q
, mode
, 1, 0, key
);
2650 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
2653 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2655 * @mode: which threads
2656 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2657 * @key: opaque value to be passed to wakeup targets
2659 * The sync wakeup differs that the waker knows that it will schedule
2660 * away soon, so while the target thread will be woken up, it will not
2661 * be migrated to another CPU - ie. the two threads are 'synchronized'
2662 * with each other. This can prevent needless bouncing between CPUs.
2664 * On UP it can prevent extra preemption.
2666 * It may be assumed that this function implies a write memory barrier before
2667 * changing the task state if and only if any tasks are woken up.
2669 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
2670 int nr_exclusive
, void *key
)
2672 unsigned long flags
;
2673 int wake_flags
= WF_SYNC
;
2678 if (unlikely(nr_exclusive
!= 1))
2681 spin_lock_irqsave(&q
->lock
, flags
);
2682 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
2683 spin_unlock_irqrestore(&q
->lock
, flags
);
2685 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
2688 * __wake_up_sync - see __wake_up_sync_key()
2690 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
2692 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
2694 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
2697 * complete: - signals a single thread waiting on this completion
2698 * @x: holds the state of this particular completion
2700 * This will wake up a single thread waiting on this completion. Threads will be
2701 * awakened in the same order in which they were queued.
2703 * See also complete_all(), wait_for_completion() and related routines.
2705 * It may be assumed that this function implies a write memory barrier before
2706 * changing the task state if and only if any tasks are woken up.
2708 void complete(struct completion
*x
)
2710 unsigned long flags
;
2712 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2714 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
2715 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2717 EXPORT_SYMBOL(complete
);
2720 * complete_all: - signals all threads waiting on this completion
2721 * @x: holds the state of this particular completion
2723 * This will wake up all threads waiting on this particular completion event.
2725 * It may be assumed that this function implies a write memory barrier before
2726 * changing the task state if and only if any tasks are woken up.
2728 void complete_all(struct completion
*x
)
2730 unsigned long flags
;
2732 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2733 x
->done
+= UINT_MAX
/2;
2734 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
2735 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2737 EXPORT_SYMBOL(complete_all
);
2739 static inline long __sched
2740 do_wait_for_common(struct completion
*x
,
2741 long (*action
)(long), long timeout
, int state
)
2744 DECLARE_WAITQUEUE(wait
, current
);
2746 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
2748 if (signal_pending_state(state
, current
)) {
2749 timeout
= -ERESTARTSYS
;
2752 __set_current_state(state
);
2753 spin_unlock_irq(&x
->wait
.lock
);
2754 timeout
= action(timeout
);
2755 spin_lock_irq(&x
->wait
.lock
);
2756 } while (!x
->done
&& timeout
);
2757 __remove_wait_queue(&x
->wait
, &wait
);
2762 return timeout
?: 1;
2765 static inline long __sched
2766 __wait_for_common(struct completion
*x
,
2767 long (*action
)(long), long timeout
, int state
)
2771 spin_lock_irq(&x
->wait
.lock
);
2772 timeout
= do_wait_for_common(x
, action
, timeout
, state
);
2773 spin_unlock_irq(&x
->wait
.lock
);
2778 wait_for_common(struct completion
*x
, long timeout
, int state
)
2780 return __wait_for_common(x
, schedule_timeout
, timeout
, state
);
2784 wait_for_common_io(struct completion
*x
, long timeout
, int state
)
2786 return __wait_for_common(x
, io_schedule_timeout
, timeout
, state
);
2790 * wait_for_completion: - waits for completion of a task
2791 * @x: holds the state of this particular completion
2793 * This waits to be signaled for completion of a specific task. It is NOT
2794 * interruptible and there is no timeout.
2796 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2797 * and interrupt capability. Also see complete().
2799 void __sched
wait_for_completion(struct completion
*x
)
2801 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
2803 EXPORT_SYMBOL(wait_for_completion
);
2806 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2807 * @x: holds the state of this particular completion
2808 * @timeout: timeout value in jiffies
2810 * This waits for either a completion of a specific task to be signaled or for a
2811 * specified timeout to expire. The timeout is in jiffies. It is not
2814 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2815 * till timeout) if completed.
2817 unsigned long __sched
2818 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
2820 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
2822 EXPORT_SYMBOL(wait_for_completion_timeout
);
2825 * wait_for_completion_io: - waits for completion of a task
2826 * @x: holds the state of this particular completion
2828 * This waits to be signaled for completion of a specific task. It is NOT
2829 * interruptible and there is no timeout. The caller is accounted as waiting
2832 void __sched
wait_for_completion_io(struct completion
*x
)
2834 wait_for_common_io(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
2836 EXPORT_SYMBOL(wait_for_completion_io
);
2839 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2840 * @x: holds the state of this particular completion
2841 * @timeout: timeout value in jiffies
2843 * This waits for either a completion of a specific task to be signaled or for a
2844 * specified timeout to expire. The timeout is in jiffies. It is not
2845 * interruptible. The caller is accounted as waiting for IO.
2847 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2848 * till timeout) if completed.
2850 unsigned long __sched
2851 wait_for_completion_io_timeout(struct completion
*x
, unsigned long timeout
)
2853 return wait_for_common_io(x
, timeout
, TASK_UNINTERRUPTIBLE
);
2855 EXPORT_SYMBOL(wait_for_completion_io_timeout
);
2858 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2859 * @x: holds the state of this particular completion
2861 * This waits for completion of a specific task to be signaled. It is
2864 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2866 int __sched
wait_for_completion_interruptible(struct completion
*x
)
2868 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
2869 if (t
== -ERESTARTSYS
)
2873 EXPORT_SYMBOL(wait_for_completion_interruptible
);
2876 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2877 * @x: holds the state of this particular completion
2878 * @timeout: timeout value in jiffies
2880 * This waits for either a completion of a specific task to be signaled or for a
2881 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2883 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2884 * or number of jiffies left till timeout) if completed.
2887 wait_for_completion_interruptible_timeout(struct completion
*x
,
2888 unsigned long timeout
)
2890 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
2892 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
2895 * wait_for_completion_killable: - waits for completion of a task (killable)
2896 * @x: holds the state of this particular completion
2898 * This waits to be signaled for completion of a specific task. It can be
2899 * interrupted by a kill signal.
2901 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2903 int __sched
wait_for_completion_killable(struct completion
*x
)
2905 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
2906 if (t
== -ERESTARTSYS
)
2910 EXPORT_SYMBOL(wait_for_completion_killable
);
2913 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2914 * @x: holds the state of this particular completion
2915 * @timeout: timeout value in jiffies
2917 * This waits for either a completion of a specific task to be
2918 * signaled or for a specified timeout to expire. It can be
2919 * interrupted by a kill signal. The timeout is in jiffies.
2921 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2922 * or number of jiffies left till timeout) if completed.
2925 wait_for_completion_killable_timeout(struct completion
*x
,
2926 unsigned long timeout
)
2928 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
2930 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
2933 * try_wait_for_completion - try to decrement a completion without blocking
2934 * @x: completion structure
2936 * Return: 0 if a decrement cannot be done without blocking
2937 * 1 if a decrement succeeded.
2939 * If a completion is being used as a counting completion,
2940 * attempt to decrement the counter without blocking. This
2941 * enables us to avoid waiting if the resource the completion
2942 * is protecting is not available.
2944 bool try_wait_for_completion(struct completion
*x
)
2946 unsigned long flags
;
2949 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2954 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2957 EXPORT_SYMBOL(try_wait_for_completion
);
2960 * completion_done - Test to see if a completion has any waiters
2961 * @x: completion structure
2963 * Return: 0 if there are waiters (wait_for_completion() in progress)
2964 * 1 if there are no waiters.
2967 bool completion_done(struct completion
*x
)
2969 unsigned long flags
;
2972 spin_lock_irqsave(&x
->wait
.lock
, flags
);
2975 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
2978 EXPORT_SYMBOL(completion_done
);
2981 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
2983 unsigned long flags
;
2986 init_waitqueue_entry(&wait
, current
);
2988 __set_current_state(state
);
2990 spin_lock_irqsave(&q
->lock
, flags
);
2991 __add_wait_queue(q
, &wait
);
2992 spin_unlock(&q
->lock
);
2993 timeout
= schedule_timeout(timeout
);
2994 spin_lock_irq(&q
->lock
);
2995 __remove_wait_queue(q
, &wait
);
2996 spin_unlock_irqrestore(&q
->lock
, flags
);
3001 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3003 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3005 EXPORT_SYMBOL(interruptible_sleep_on
);
3008 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3010 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
3012 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3014 void __sched
sleep_on(wait_queue_head_t
*q
)
3016 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
3018 EXPORT_SYMBOL(sleep_on
);
3020 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3022 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
3024 EXPORT_SYMBOL(sleep_on_timeout
);
3026 #ifdef CONFIG_RT_MUTEXES
3029 * rt_mutex_setprio - set the current priority of a task
3031 * @prio: prio value (kernel-internal form)
3033 * This function changes the 'effective' priority of a task. It does
3034 * not touch ->normal_prio like __setscheduler().
3036 * Used by the rt_mutex code to implement priority inheritance logic.
3038 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3040 int oldprio
, on_rq
, running
;
3042 const struct sched_class
*prev_class
;
3044 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3046 rq
= __task_rq_lock(p
);
3049 * Idle task boosting is a nono in general. There is one
3050 * exception, when PREEMPT_RT and NOHZ is active:
3052 * The idle task calls get_next_timer_interrupt() and holds
3053 * the timer wheel base->lock on the CPU and another CPU wants
3054 * to access the timer (probably to cancel it). We can safely
3055 * ignore the boosting request, as the idle CPU runs this code
3056 * with interrupts disabled and will complete the lock
3057 * protected section without being interrupted. So there is no
3058 * real need to boost.
3060 if (unlikely(p
== rq
->idle
)) {
3061 WARN_ON(p
!= rq
->curr
);
3062 WARN_ON(p
->pi_blocked_on
);
3066 trace_sched_pi_setprio(p
, prio
);
3068 prev_class
= p
->sched_class
;
3070 running
= task_current(rq
, p
);
3072 dequeue_task(rq
, p
, 0);
3074 p
->sched_class
->put_prev_task(rq
, p
);
3077 p
->sched_class
= &rt_sched_class
;
3079 p
->sched_class
= &fair_sched_class
;
3084 p
->sched_class
->set_curr_task(rq
);
3086 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
3088 check_class_changed(rq
, p
, prev_class
, oldprio
);
3090 __task_rq_unlock(rq
);
3093 void set_user_nice(struct task_struct
*p
, long nice
)
3095 int old_prio
, delta
, on_rq
;
3096 unsigned long flags
;
3099 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3102 * We have to be careful, if called from sys_setpriority(),
3103 * the task might be in the middle of scheduling on another CPU.
3105 rq
= task_rq_lock(p
, &flags
);
3107 * The RT priorities are set via sched_setscheduler(), but we still
3108 * allow the 'normal' nice value to be set - but as expected
3109 * it wont have any effect on scheduling until the task is
3110 * SCHED_FIFO/SCHED_RR:
3112 if (task_has_rt_policy(p
)) {
3113 p
->static_prio
= NICE_TO_PRIO(nice
);
3118 dequeue_task(rq
, p
, 0);
3120 p
->static_prio
= NICE_TO_PRIO(nice
);
3123 p
->prio
= effective_prio(p
);
3124 delta
= p
->prio
- old_prio
;
3127 enqueue_task(rq
, p
, 0);
3129 * If the task increased its priority or is running and
3130 * lowered its priority, then reschedule its CPU:
3132 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3133 resched_task(rq
->curr
);
3136 task_rq_unlock(rq
, p
, &flags
);
3138 EXPORT_SYMBOL(set_user_nice
);
3141 * can_nice - check if a task can reduce its nice value
3145 int can_nice(const struct task_struct
*p
, const int nice
)
3147 /* convert nice value [19,-20] to rlimit style value [1,40] */
3148 int nice_rlim
= 20 - nice
;
3150 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3151 capable(CAP_SYS_NICE
));
3154 #ifdef __ARCH_WANT_SYS_NICE
3157 * sys_nice - change the priority of the current process.
3158 * @increment: priority increment
3160 * sys_setpriority is a more generic, but much slower function that
3161 * does similar things.
3163 SYSCALL_DEFINE1(nice
, int, increment
)
3168 * Setpriority might change our priority at the same moment.
3169 * We don't have to worry. Conceptually one call occurs first
3170 * and we have a single winner.
3172 if (increment
< -40)
3177 nice
= TASK_NICE(current
) + increment
;
3183 if (increment
< 0 && !can_nice(current
, nice
))
3186 retval
= security_task_setnice(current
, nice
);
3190 set_user_nice(current
, nice
);
3197 * task_prio - return the priority value of a given task.
3198 * @p: the task in question.
3200 * Return: The priority value as seen by users in /proc.
3201 * RT tasks are offset by -200. Normal tasks are centered
3202 * around 0, value goes from -16 to +15.
3204 int task_prio(const struct task_struct
*p
)
3206 return p
->prio
- MAX_RT_PRIO
;
3210 * task_nice - return the nice value of a given task.
3211 * @p: the task in question.
3213 * Return: The nice value [ -20 ... 0 ... 19 ].
3215 int task_nice(const struct task_struct
*p
)
3217 return TASK_NICE(p
);
3219 EXPORT_SYMBOL(task_nice
);
3222 * idle_cpu - is a given cpu idle currently?
3223 * @cpu: the processor in question.
3225 * Return: 1 if the CPU is currently idle. 0 otherwise.
3227 int idle_cpu(int cpu
)
3229 struct rq
*rq
= cpu_rq(cpu
);
3231 if (rq
->curr
!= rq
->idle
)
3238 if (!llist_empty(&rq
->wake_list
))
3246 * idle_task - return the idle task for a given cpu.
3247 * @cpu: the processor in question.
3249 * Return: The idle task for the cpu @cpu.
3251 struct task_struct
*idle_task(int cpu
)
3253 return cpu_rq(cpu
)->idle
;
3257 * find_process_by_pid - find a process with a matching PID value.
3258 * @pid: the pid in question.
3260 * The task of @pid, if found. %NULL otherwise.
3262 static struct task_struct
*find_process_by_pid(pid_t pid
)
3264 return pid
? find_task_by_vpid(pid
) : current
;
3267 /* Actually do priority change: must hold rq lock. */
3269 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
3272 p
->rt_priority
= prio
;
3273 p
->normal_prio
= normal_prio(p
);
3274 /* we are holding p->pi_lock already */
3275 p
->prio
= rt_mutex_getprio(p
);
3276 if (rt_prio(p
->prio
))
3277 p
->sched_class
= &rt_sched_class
;
3279 p
->sched_class
= &fair_sched_class
;
3284 * check the target process has a UID that matches the current process's
3286 static bool check_same_owner(struct task_struct
*p
)
3288 const struct cred
*cred
= current_cred(), *pcred
;
3292 pcred
= __task_cred(p
);
3293 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3294 uid_eq(cred
->euid
, pcred
->uid
));
3299 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
3300 const struct sched_param
*param
, bool user
)
3302 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3303 unsigned long flags
;
3304 const struct sched_class
*prev_class
;
3308 /* may grab non-irq protected spin_locks */
3309 BUG_ON(in_interrupt());
3311 /* double check policy once rq lock held */
3313 reset_on_fork
= p
->sched_reset_on_fork
;
3314 policy
= oldpolicy
= p
->policy
;
3316 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
3317 policy
&= ~SCHED_RESET_ON_FORK
;
3319 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3320 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3321 policy
!= SCHED_IDLE
)
3326 * Valid priorities for SCHED_FIFO and SCHED_RR are
3327 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3328 * SCHED_BATCH and SCHED_IDLE is 0.
3330 if (param
->sched_priority
< 0 ||
3331 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3332 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
3334 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
3338 * Allow unprivileged RT tasks to decrease priority:
3340 if (user
&& !capable(CAP_SYS_NICE
)) {
3341 if (rt_policy(policy
)) {
3342 unsigned long rlim_rtprio
=
3343 task_rlimit(p
, RLIMIT_RTPRIO
);
3345 /* can't set/change the rt policy */
3346 if (policy
!= p
->policy
&& !rlim_rtprio
)
3349 /* can't increase priority */
3350 if (param
->sched_priority
> p
->rt_priority
&&
3351 param
->sched_priority
> rlim_rtprio
)
3356 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3357 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3359 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3360 if (!can_nice(p
, TASK_NICE(p
)))
3364 /* can't change other user's priorities */
3365 if (!check_same_owner(p
))
3368 /* Normal users shall not reset the sched_reset_on_fork flag */
3369 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3374 retval
= security_task_setscheduler(p
);
3380 * make sure no PI-waiters arrive (or leave) while we are
3381 * changing the priority of the task:
3383 * To be able to change p->policy safely, the appropriate
3384 * runqueue lock must be held.
3386 rq
= task_rq_lock(p
, &flags
);
3389 * Changing the policy of the stop threads its a very bad idea
3391 if (p
== rq
->stop
) {
3392 task_rq_unlock(rq
, p
, &flags
);
3397 * If not changing anything there's no need to proceed further:
3399 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
3400 param
->sched_priority
== p
->rt_priority
))) {
3401 task_rq_unlock(rq
, p
, &flags
);
3405 #ifdef CONFIG_RT_GROUP_SCHED
3408 * Do not allow realtime tasks into groups that have no runtime
3411 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3412 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3413 !task_group_is_autogroup(task_group(p
))) {
3414 task_rq_unlock(rq
, p
, &flags
);
3420 /* recheck policy now with rq lock held */
3421 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3422 policy
= oldpolicy
= -1;
3423 task_rq_unlock(rq
, p
, &flags
);
3427 running
= task_current(rq
, p
);
3429 dequeue_task(rq
, p
, 0);
3431 p
->sched_class
->put_prev_task(rq
, p
);
3433 p
->sched_reset_on_fork
= reset_on_fork
;
3436 prev_class
= p
->sched_class
;
3437 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
3440 p
->sched_class
->set_curr_task(rq
);
3442 enqueue_task(rq
, p
, 0);
3444 check_class_changed(rq
, p
, prev_class
, oldprio
);
3445 task_rq_unlock(rq
, p
, &flags
);
3447 rt_mutex_adjust_pi(p
);
3453 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3454 * @p: the task in question.
3455 * @policy: new policy.
3456 * @param: structure containing the new RT priority.
3458 * Return: 0 on success. An error code otherwise.
3460 * NOTE that the task may be already dead.
3462 int sched_setscheduler(struct task_struct
*p
, int policy
,
3463 const struct sched_param
*param
)
3465 return __sched_setscheduler(p
, policy
, param
, true);
3467 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3470 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3471 * @p: the task in question.
3472 * @policy: new policy.
3473 * @param: structure containing the new RT priority.
3475 * Just like sched_setscheduler, only don't bother checking if the
3476 * current context has permission. For example, this is needed in
3477 * stop_machine(): we create temporary high priority worker threads,
3478 * but our caller might not have that capability.
3480 * Return: 0 on success. An error code otherwise.
3482 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3483 const struct sched_param
*param
)
3485 return __sched_setscheduler(p
, policy
, param
, false);
3489 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3491 struct sched_param lparam
;
3492 struct task_struct
*p
;
3495 if (!param
|| pid
< 0)
3497 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3502 p
= find_process_by_pid(pid
);
3504 retval
= sched_setscheduler(p
, policy
, &lparam
);
3511 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3512 * @pid: the pid in question.
3513 * @policy: new policy.
3514 * @param: structure containing the new RT priority.
3516 * Return: 0 on success. An error code otherwise.
3518 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3519 struct sched_param __user
*, param
)
3521 /* negative values for policy are not valid */
3525 return do_sched_setscheduler(pid
, policy
, param
);
3529 * sys_sched_setparam - set/change the RT priority of a thread
3530 * @pid: the pid in question.
3531 * @param: structure containing the new RT priority.
3533 * Return: 0 on success. An error code otherwise.
3535 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3537 return do_sched_setscheduler(pid
, -1, param
);
3541 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3542 * @pid: the pid in question.
3544 * Return: On success, the policy of the thread. Otherwise, a negative error
3547 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3549 struct task_struct
*p
;
3557 p
= find_process_by_pid(pid
);
3559 retval
= security_task_getscheduler(p
);
3562 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3569 * sys_sched_getparam - get the RT priority of a thread
3570 * @pid: the pid in question.
3571 * @param: structure containing the RT priority.
3573 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3576 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3578 struct sched_param lp
;
3579 struct task_struct
*p
;
3582 if (!param
|| pid
< 0)
3586 p
= find_process_by_pid(pid
);
3591 retval
= security_task_getscheduler(p
);
3595 lp
.sched_priority
= p
->rt_priority
;
3599 * This one might sleep, we cannot do it with a spinlock held ...
3601 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3610 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3612 cpumask_var_t cpus_allowed
, new_mask
;
3613 struct task_struct
*p
;
3619 p
= find_process_by_pid(pid
);
3626 /* Prevent p going away */
3630 if (p
->flags
& PF_NO_SETAFFINITY
) {
3634 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3638 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3640 goto out_free_cpus_allowed
;
3643 if (!check_same_owner(p
)) {
3645 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3652 retval
= security_task_setscheduler(p
);
3656 cpuset_cpus_allowed(p
, cpus_allowed
);
3657 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3659 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3662 cpuset_cpus_allowed(p
, cpus_allowed
);
3663 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3665 * We must have raced with a concurrent cpuset
3666 * update. Just reset the cpus_allowed to the
3667 * cpuset's cpus_allowed
3669 cpumask_copy(new_mask
, cpus_allowed
);
3674 free_cpumask_var(new_mask
);
3675 out_free_cpus_allowed
:
3676 free_cpumask_var(cpus_allowed
);
3683 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3684 struct cpumask
*new_mask
)
3686 if (len
< cpumask_size())
3687 cpumask_clear(new_mask
);
3688 else if (len
> cpumask_size())
3689 len
= cpumask_size();
3691 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3695 * sys_sched_setaffinity - set the cpu affinity of a process
3696 * @pid: pid of the process
3697 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3698 * @user_mask_ptr: user-space pointer to the new cpu mask
3700 * Return: 0 on success. An error code otherwise.
3702 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3703 unsigned long __user
*, user_mask_ptr
)
3705 cpumask_var_t new_mask
;
3708 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3711 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3713 retval
= sched_setaffinity(pid
, new_mask
);
3714 free_cpumask_var(new_mask
);
3718 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
3720 struct task_struct
*p
;
3721 unsigned long flags
;
3728 p
= find_process_by_pid(pid
);
3732 retval
= security_task_getscheduler(p
);
3736 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3737 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
3738 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3748 * sys_sched_getaffinity - get the cpu affinity of a process
3749 * @pid: pid of the process
3750 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3751 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3753 * Return: 0 on success. An error code otherwise.
3755 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
3756 unsigned long __user
*, user_mask_ptr
)
3761 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
3763 if (len
& (sizeof(unsigned long)-1))
3766 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
3769 ret
= sched_getaffinity(pid
, mask
);
3771 size_t retlen
= min_t(size_t, len
, cpumask_size());
3773 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
3778 free_cpumask_var(mask
);
3784 * sys_sched_yield - yield the current processor to other threads.
3786 * This function yields the current CPU to other tasks. If there are no
3787 * other threads running on this CPU then this function will return.
3791 SYSCALL_DEFINE0(sched_yield
)
3793 struct rq
*rq
= this_rq_lock();
3795 schedstat_inc(rq
, yld_count
);
3796 current
->sched_class
->yield_task(rq
);
3799 * Since we are going to call schedule() anyway, there's
3800 * no need to preempt or enable interrupts:
3802 __release(rq
->lock
);
3803 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3804 do_raw_spin_unlock(&rq
->lock
);
3805 sched_preempt_enable_no_resched();
3812 static inline int should_resched(void)
3814 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
3817 static void __cond_resched(void)
3819 add_preempt_count(PREEMPT_ACTIVE
);
3821 sub_preempt_count(PREEMPT_ACTIVE
);
3824 int __sched
_cond_resched(void)
3826 if (should_resched()) {
3832 EXPORT_SYMBOL(_cond_resched
);
3835 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3836 * call schedule, and on return reacquire the lock.
3838 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3839 * operations here to prevent schedule() from being called twice (once via
3840 * spin_unlock(), once by hand).
3842 int __cond_resched_lock(spinlock_t
*lock
)
3844 int resched
= should_resched();
3847 lockdep_assert_held(lock
);
3849 if (spin_needbreak(lock
) || resched
) {
3860 EXPORT_SYMBOL(__cond_resched_lock
);
3862 int __sched
__cond_resched_softirq(void)
3864 BUG_ON(!in_softirq());
3866 if (should_resched()) {
3874 EXPORT_SYMBOL(__cond_resched_softirq
);
3877 * yield - yield the current processor to other threads.
3879 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3881 * The scheduler is at all times free to pick the calling task as the most
3882 * eligible task to run, if removing the yield() call from your code breaks
3883 * it, its already broken.
3885 * Typical broken usage is:
3890 * where one assumes that yield() will let 'the other' process run that will
3891 * make event true. If the current task is a SCHED_FIFO task that will never
3892 * happen. Never use yield() as a progress guarantee!!
3894 * If you want to use yield() to wait for something, use wait_event().
3895 * If you want to use yield() to be 'nice' for others, use cond_resched().
3896 * If you still want to use yield(), do not!
3898 void __sched
yield(void)
3900 set_current_state(TASK_RUNNING
);
3903 EXPORT_SYMBOL(yield
);
3906 * yield_to - yield the current processor to another thread in
3907 * your thread group, or accelerate that thread toward the
3908 * processor it's on.
3910 * @preempt: whether task preemption is allowed or not
3912 * It's the caller's job to ensure that the target task struct
3913 * can't go away on us before we can do any checks.
3916 * true (>0) if we indeed boosted the target task.
3917 * false (0) if we failed to boost the target.
3918 * -ESRCH if there's no task to yield to.
3920 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
3922 struct task_struct
*curr
= current
;
3923 struct rq
*rq
, *p_rq
;
3924 unsigned long flags
;
3927 local_irq_save(flags
);
3933 * If we're the only runnable task on the rq and target rq also
3934 * has only one task, there's absolutely no point in yielding.
3936 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
3941 double_rq_lock(rq
, p_rq
);
3942 while (task_rq(p
) != p_rq
) {
3943 double_rq_unlock(rq
, p_rq
);
3947 if (!curr
->sched_class
->yield_to_task
)
3950 if (curr
->sched_class
!= p
->sched_class
)
3953 if (task_running(p_rq
, p
) || p
->state
)
3956 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
3958 schedstat_inc(rq
, yld_count
);
3960 * Make p's CPU reschedule; pick_next_entity takes care of
3963 if (preempt
&& rq
!= p_rq
)
3964 resched_task(p_rq
->curr
);
3968 double_rq_unlock(rq
, p_rq
);
3970 local_irq_restore(flags
);
3977 EXPORT_SYMBOL_GPL(yield_to
);
3980 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3981 * that process accounting knows that this is a task in IO wait state.
3983 void __sched
io_schedule(void)
3985 struct rq
*rq
= raw_rq();
3987 delayacct_blkio_start();
3988 atomic_inc(&rq
->nr_iowait
);
3989 blk_flush_plug(current
);
3990 current
->in_iowait
= 1;
3992 current
->in_iowait
= 0;
3993 atomic_dec(&rq
->nr_iowait
);
3994 delayacct_blkio_end();
3996 EXPORT_SYMBOL(io_schedule
);
3998 long __sched
io_schedule_timeout(long timeout
)
4000 struct rq
*rq
= raw_rq();
4003 delayacct_blkio_start();
4004 atomic_inc(&rq
->nr_iowait
);
4005 blk_flush_plug(current
);
4006 current
->in_iowait
= 1;
4007 ret
= schedule_timeout(timeout
);
4008 current
->in_iowait
= 0;
4009 atomic_dec(&rq
->nr_iowait
);
4010 delayacct_blkio_end();
4015 * sys_sched_get_priority_max - return maximum RT priority.
4016 * @policy: scheduling class.
4018 * Return: On success, this syscall returns the maximum
4019 * rt_priority that can be used by a given scheduling class.
4020 * On failure, a negative error code is returned.
4022 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4029 ret
= MAX_USER_RT_PRIO
-1;
4041 * sys_sched_get_priority_min - return minimum RT priority.
4042 * @policy: scheduling class.
4044 * Return: On success, this syscall returns the minimum
4045 * rt_priority that can be used by a given scheduling class.
4046 * On failure, a negative error code is returned.
4048 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4066 * sys_sched_rr_get_interval - return the default timeslice of a process.
4067 * @pid: pid of the process.
4068 * @interval: userspace pointer to the timeslice value.
4070 * this syscall writes the default timeslice value of a given process
4071 * into the user-space timespec buffer. A value of '0' means infinity.
4073 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4076 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4077 struct timespec __user
*, interval
)
4079 struct task_struct
*p
;
4080 unsigned int time_slice
;
4081 unsigned long flags
;
4091 p
= find_process_by_pid(pid
);
4095 retval
= security_task_getscheduler(p
);
4099 rq
= task_rq_lock(p
, &flags
);
4100 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4101 task_rq_unlock(rq
, p
, &flags
);
4104 jiffies_to_timespec(time_slice
, &t
);
4105 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4113 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4115 void sched_show_task(struct task_struct
*p
)
4117 unsigned long free
= 0;
4121 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4122 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4123 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4124 #if BITS_PER_LONG == 32
4125 if (state
== TASK_RUNNING
)
4126 printk(KERN_CONT
" running ");
4128 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4130 if (state
== TASK_RUNNING
)
4131 printk(KERN_CONT
" running task ");
4133 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4135 #ifdef CONFIG_DEBUG_STACK_USAGE
4136 free
= stack_not_used(p
);
4139 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4141 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4142 task_pid_nr(p
), ppid
,
4143 (unsigned long)task_thread_info(p
)->flags
);
4145 print_worker_info(KERN_INFO
, p
);
4146 show_stack(p
, NULL
);
4149 void show_state_filter(unsigned long state_filter
)
4151 struct task_struct
*g
, *p
;
4153 #if BITS_PER_LONG == 32
4155 " task PC stack pid father\n");
4158 " task PC stack pid father\n");
4161 do_each_thread(g
, p
) {
4163 * reset the NMI-timeout, listing all files on a slow
4164 * console might take a lot of time:
4166 touch_nmi_watchdog();
4167 if (!state_filter
|| (p
->state
& state_filter
))
4169 } while_each_thread(g
, p
);
4171 touch_all_softlockup_watchdogs();
4173 #ifdef CONFIG_SCHED_DEBUG
4174 sysrq_sched_debug_show();
4178 * Only show locks if all tasks are dumped:
4181 debug_show_all_locks();
4184 void init_idle_bootup_task(struct task_struct
*idle
)
4186 idle
->sched_class
= &idle_sched_class
;
4190 * init_idle - set up an idle thread for a given CPU
4191 * @idle: task in question
4192 * @cpu: cpu the idle task belongs to
4194 * NOTE: this function does not set the idle thread's NEED_RESCHED
4195 * flag, to make booting more robust.
4197 void init_idle(struct task_struct
*idle
, int cpu
)
4199 struct rq
*rq
= cpu_rq(cpu
);
4200 unsigned long flags
;
4202 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4205 idle
->state
= TASK_RUNNING
;
4206 idle
->se
.exec_start
= sched_clock();
4208 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4210 * We're having a chicken and egg problem, even though we are
4211 * holding rq->lock, the cpu isn't yet set to this cpu so the
4212 * lockdep check in task_group() will fail.
4214 * Similar case to sched_fork(). / Alternatively we could
4215 * use task_rq_lock() here and obtain the other rq->lock.
4220 __set_task_cpu(idle
, cpu
);
4223 rq
->curr
= rq
->idle
= idle
;
4224 #if defined(CONFIG_SMP)
4227 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4229 /* Set the preempt count _outside_ the spinlocks! */
4230 task_thread_info(idle
)->preempt_count
= 0;
4233 * The idle tasks have their own, simple scheduling class:
4235 idle
->sched_class
= &idle_sched_class
;
4236 ftrace_graph_init_idle_task(idle
, cpu
);
4237 vtime_init_idle(idle
, cpu
);
4238 #if defined(CONFIG_SMP)
4239 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4244 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4246 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4247 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4249 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4250 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4254 * This is how migration works:
4256 * 1) we invoke migration_cpu_stop() on the target CPU using
4258 * 2) stopper starts to run (implicitly forcing the migrated thread
4260 * 3) it checks whether the migrated task is still in the wrong runqueue.
4261 * 4) if it's in the wrong runqueue then the migration thread removes
4262 * it and puts it into the right queue.
4263 * 5) stopper completes and stop_one_cpu() returns and the migration
4268 * Change a given task's CPU affinity. Migrate the thread to a
4269 * proper CPU and schedule it away if the CPU it's executing on
4270 * is removed from the allowed bitmask.
4272 * NOTE: the caller must have a valid reference to the task, the
4273 * task must not exit() & deallocate itself prematurely. The
4274 * call is not atomic; no spinlocks may be held.
4276 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4278 unsigned long flags
;
4280 unsigned int dest_cpu
;
4283 rq
= task_rq_lock(p
, &flags
);
4285 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4288 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4293 do_set_cpus_allowed(p
, new_mask
);
4295 /* Can the task run on the task's current CPU? If so, we're done */
4296 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4299 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4301 struct migration_arg arg
= { p
, dest_cpu
};
4302 /* Need help from migration thread: drop lock and wait. */
4303 task_rq_unlock(rq
, p
, &flags
);
4304 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4305 tlb_migrate_finish(p
->mm
);
4309 task_rq_unlock(rq
, p
, &flags
);
4313 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4316 * Move (not current) task off this cpu, onto dest cpu. We're doing
4317 * this because either it can't run here any more (set_cpus_allowed()
4318 * away from this CPU, or CPU going down), or because we're
4319 * attempting to rebalance this task on exec (sched_exec).
4321 * So we race with normal scheduler movements, but that's OK, as long
4322 * as the task is no longer on this CPU.
4324 * Returns non-zero if task was successfully migrated.
4326 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4328 struct rq
*rq_dest
, *rq_src
;
4331 if (unlikely(!cpu_active(dest_cpu
)))
4334 rq_src
= cpu_rq(src_cpu
);
4335 rq_dest
= cpu_rq(dest_cpu
);
4337 raw_spin_lock(&p
->pi_lock
);
4338 double_rq_lock(rq_src
, rq_dest
);
4339 /* Already moved. */
4340 if (task_cpu(p
) != src_cpu
)
4342 /* Affinity changed (again). */
4343 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4347 * If we're not on a rq, the next wake-up will ensure we're
4351 dequeue_task(rq_src
, p
, 0);
4352 set_task_cpu(p
, dest_cpu
);
4353 enqueue_task(rq_dest
, p
, 0);
4354 check_preempt_curr(rq_dest
, p
, 0);
4359 double_rq_unlock(rq_src
, rq_dest
);
4360 raw_spin_unlock(&p
->pi_lock
);
4365 * migration_cpu_stop - this will be executed by a highprio stopper thread
4366 * and performs thread migration by bumping thread off CPU then
4367 * 'pushing' onto another runqueue.
4369 static int migration_cpu_stop(void *data
)
4371 struct migration_arg
*arg
= data
;
4374 * The original target cpu might have gone down and we might
4375 * be on another cpu but it doesn't matter.
4377 local_irq_disable();
4378 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4383 #ifdef CONFIG_HOTPLUG_CPU
4386 * Ensures that the idle task is using init_mm right before its cpu goes
4389 void idle_task_exit(void)
4391 struct mm_struct
*mm
= current
->active_mm
;
4393 BUG_ON(cpu_online(smp_processor_id()));
4396 switch_mm(mm
, &init_mm
, current
);
4401 * Since this CPU is going 'away' for a while, fold any nr_active delta
4402 * we might have. Assumes we're called after migrate_tasks() so that the
4403 * nr_active count is stable.
4405 * Also see the comment "Global load-average calculations".
4407 static void calc_load_migrate(struct rq
*rq
)
4409 long delta
= calc_load_fold_active(rq
);
4411 atomic_long_add(delta
, &calc_load_tasks
);
4415 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4416 * try_to_wake_up()->select_task_rq().
4418 * Called with rq->lock held even though we'er in stop_machine() and
4419 * there's no concurrency possible, we hold the required locks anyway
4420 * because of lock validation efforts.
4422 static void migrate_tasks(unsigned int dead_cpu
)
4424 struct rq
*rq
= cpu_rq(dead_cpu
);
4425 struct task_struct
*next
, *stop
= rq
->stop
;
4429 * Fudge the rq selection such that the below task selection loop
4430 * doesn't get stuck on the currently eligible stop task.
4432 * We're currently inside stop_machine() and the rq is either stuck
4433 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4434 * either way we should never end up calling schedule() until we're
4440 * put_prev_task() and pick_next_task() sched
4441 * class method both need to have an up-to-date
4442 * value of rq->clock[_task]
4444 update_rq_clock(rq
);
4448 * There's this thread running, bail when that's the only
4451 if (rq
->nr_running
== 1)
4454 next
= pick_next_task(rq
);
4456 next
->sched_class
->put_prev_task(rq
, next
);
4458 /* Find suitable destination for @next, with force if needed. */
4459 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4460 raw_spin_unlock(&rq
->lock
);
4462 __migrate_task(next
, dead_cpu
, dest_cpu
);
4464 raw_spin_lock(&rq
->lock
);
4470 #endif /* CONFIG_HOTPLUG_CPU */
4472 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4474 static struct ctl_table sd_ctl_dir
[] = {
4476 .procname
= "sched_domain",
4482 static struct ctl_table sd_ctl_root
[] = {
4484 .procname
= "kernel",
4486 .child
= sd_ctl_dir
,
4491 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4493 struct ctl_table
*entry
=
4494 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4499 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4501 struct ctl_table
*entry
;
4504 * In the intermediate directories, both the child directory and
4505 * procname are dynamically allocated and could fail but the mode
4506 * will always be set. In the lowest directory the names are
4507 * static strings and all have proc handlers.
4509 for (entry
= *tablep
; entry
->mode
; entry
++) {
4511 sd_free_ctl_entry(&entry
->child
);
4512 if (entry
->proc_handler
== NULL
)
4513 kfree(entry
->procname
);
4520 static int min_load_idx
= 0;
4521 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4524 set_table_entry(struct ctl_table
*entry
,
4525 const char *procname
, void *data
, int maxlen
,
4526 umode_t mode
, proc_handler
*proc_handler
,
4529 entry
->procname
= procname
;
4531 entry
->maxlen
= maxlen
;
4533 entry
->proc_handler
= proc_handler
;
4536 entry
->extra1
= &min_load_idx
;
4537 entry
->extra2
= &max_load_idx
;
4541 static struct ctl_table
*
4542 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4544 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4549 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4550 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4551 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4552 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4553 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4554 sizeof(int), 0644, proc_dointvec_minmax
, true);
4555 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4556 sizeof(int), 0644, proc_dointvec_minmax
, true);
4557 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4558 sizeof(int), 0644, proc_dointvec_minmax
, true);
4559 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4560 sizeof(int), 0644, proc_dointvec_minmax
, true);
4561 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4562 sizeof(int), 0644, proc_dointvec_minmax
, true);
4563 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4564 sizeof(int), 0644, proc_dointvec_minmax
, false);
4565 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4566 sizeof(int), 0644, proc_dointvec_minmax
, false);
4567 set_table_entry(&table
[9], "cache_nice_tries",
4568 &sd
->cache_nice_tries
,
4569 sizeof(int), 0644, proc_dointvec_minmax
, false);
4570 set_table_entry(&table
[10], "flags", &sd
->flags
,
4571 sizeof(int), 0644, proc_dointvec_minmax
, false);
4572 set_table_entry(&table
[11], "name", sd
->name
,
4573 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4574 /* &table[12] is terminator */
4579 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4581 struct ctl_table
*entry
, *table
;
4582 struct sched_domain
*sd
;
4583 int domain_num
= 0, i
;
4586 for_each_domain(cpu
, sd
)
4588 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4593 for_each_domain(cpu
, sd
) {
4594 snprintf(buf
, 32, "domain%d", i
);
4595 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4597 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4604 static struct ctl_table_header
*sd_sysctl_header
;
4605 static void register_sched_domain_sysctl(void)
4607 int i
, cpu_num
= num_possible_cpus();
4608 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4611 WARN_ON(sd_ctl_dir
[0].child
);
4612 sd_ctl_dir
[0].child
= entry
;
4617 for_each_possible_cpu(i
) {
4618 snprintf(buf
, 32, "cpu%d", i
);
4619 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4621 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4625 WARN_ON(sd_sysctl_header
);
4626 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4629 /* may be called multiple times per register */
4630 static void unregister_sched_domain_sysctl(void)
4632 if (sd_sysctl_header
)
4633 unregister_sysctl_table(sd_sysctl_header
);
4634 sd_sysctl_header
= NULL
;
4635 if (sd_ctl_dir
[0].child
)
4636 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4639 static void register_sched_domain_sysctl(void)
4642 static void unregister_sched_domain_sysctl(void)
4647 static void set_rq_online(struct rq
*rq
)
4650 const struct sched_class
*class;
4652 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
4655 for_each_class(class) {
4656 if (class->rq_online
)
4657 class->rq_online(rq
);
4662 static void set_rq_offline(struct rq
*rq
)
4665 const struct sched_class
*class;
4667 for_each_class(class) {
4668 if (class->rq_offline
)
4669 class->rq_offline(rq
);
4672 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
4678 * migration_call - callback that gets triggered when a CPU is added.
4679 * Here we can start up the necessary migration thread for the new CPU.
4682 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
4684 int cpu
= (long)hcpu
;
4685 unsigned long flags
;
4686 struct rq
*rq
= cpu_rq(cpu
);
4688 switch (action
& ~CPU_TASKS_FROZEN
) {
4690 case CPU_UP_PREPARE
:
4691 rq
->calc_load_update
= calc_load_update
;
4695 /* Update our root-domain */
4696 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4698 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
4702 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4705 #ifdef CONFIG_HOTPLUG_CPU
4707 sched_ttwu_pending();
4708 /* Update our root-domain */
4709 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4711 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
4715 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
4716 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4720 calc_load_migrate(rq
);
4725 update_max_interval();
4731 * Register at high priority so that task migration (migrate_all_tasks)
4732 * happens before everything else. This has to be lower priority than
4733 * the notifier in the perf_event subsystem, though.
4735 static struct notifier_block migration_notifier
= {
4736 .notifier_call
= migration_call
,
4737 .priority
= CPU_PRI_MIGRATION
,
4740 static int sched_cpu_active(struct notifier_block
*nfb
,
4741 unsigned long action
, void *hcpu
)
4743 switch (action
& ~CPU_TASKS_FROZEN
) {
4745 case CPU_DOWN_FAILED
:
4746 set_cpu_active((long)hcpu
, true);
4753 static int sched_cpu_inactive(struct notifier_block
*nfb
,
4754 unsigned long action
, void *hcpu
)
4756 switch (action
& ~CPU_TASKS_FROZEN
) {
4757 case CPU_DOWN_PREPARE
:
4758 set_cpu_active((long)hcpu
, false);
4765 static int __init
migration_init(void)
4767 void *cpu
= (void *)(long)smp_processor_id();
4770 /* Initialize migration for the boot CPU */
4771 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
4772 BUG_ON(err
== NOTIFY_BAD
);
4773 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
4774 register_cpu_notifier(&migration_notifier
);
4776 /* Register cpu active notifiers */
4777 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
4778 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
4782 early_initcall(migration_init
);
4787 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
4789 #ifdef CONFIG_SCHED_DEBUG
4791 static __read_mostly
int sched_debug_enabled
;
4793 static int __init
sched_debug_setup(char *str
)
4795 sched_debug_enabled
= 1;
4799 early_param("sched_debug", sched_debug_setup
);
4801 static inline bool sched_debug(void)
4803 return sched_debug_enabled
;
4806 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
4807 struct cpumask
*groupmask
)
4809 struct sched_group
*group
= sd
->groups
;
4812 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
4813 cpumask_clear(groupmask
);
4815 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
4817 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
4818 printk("does not load-balance\n");
4820 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
4825 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
4827 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
4828 printk(KERN_ERR
"ERROR: domain->span does not contain "
4831 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
4832 printk(KERN_ERR
"ERROR: domain->groups does not contain"
4836 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
4840 printk(KERN_ERR
"ERROR: group is NULL\n");
4845 * Even though we initialize ->power to something semi-sane,
4846 * we leave power_orig unset. This allows us to detect if
4847 * domain iteration is still funny without causing /0 traps.
4849 if (!group
->sgp
->power_orig
) {
4850 printk(KERN_CONT
"\n");
4851 printk(KERN_ERR
"ERROR: domain->cpu_power not "
4856 if (!cpumask_weight(sched_group_cpus(group
))) {
4857 printk(KERN_CONT
"\n");
4858 printk(KERN_ERR
"ERROR: empty group\n");
4862 if (!(sd
->flags
& SD_OVERLAP
) &&
4863 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
4864 printk(KERN_CONT
"\n");
4865 printk(KERN_ERR
"ERROR: repeated CPUs\n");
4869 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
4871 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
4873 printk(KERN_CONT
" %s", str
);
4874 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
4875 printk(KERN_CONT
" (cpu_power = %d)",
4879 group
= group
->next
;
4880 } while (group
!= sd
->groups
);
4881 printk(KERN_CONT
"\n");
4883 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
4884 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
4887 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
4888 printk(KERN_ERR
"ERROR: parent span is not a superset "
4889 "of domain->span\n");
4893 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
4897 if (!sched_debug_enabled
)
4901 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
4905 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
4908 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
4916 #else /* !CONFIG_SCHED_DEBUG */
4917 # define sched_domain_debug(sd, cpu) do { } while (0)
4918 static inline bool sched_debug(void)
4922 #endif /* CONFIG_SCHED_DEBUG */
4924 static int sd_degenerate(struct sched_domain
*sd
)
4926 if (cpumask_weight(sched_domain_span(sd
)) == 1)
4929 /* Following flags need at least 2 groups */
4930 if (sd
->flags
& (SD_LOAD_BALANCE
|
4931 SD_BALANCE_NEWIDLE
|
4935 SD_SHARE_PKG_RESOURCES
)) {
4936 if (sd
->groups
!= sd
->groups
->next
)
4940 /* Following flags don't use groups */
4941 if (sd
->flags
& (SD_WAKE_AFFINE
))
4948 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
4950 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
4952 if (sd_degenerate(parent
))
4955 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
4958 /* Flags needing groups don't count if only 1 group in parent */
4959 if (parent
->groups
== parent
->groups
->next
) {
4960 pflags
&= ~(SD_LOAD_BALANCE
|
4961 SD_BALANCE_NEWIDLE
|
4965 SD_SHARE_PKG_RESOURCES
|
4967 if (nr_node_ids
== 1)
4968 pflags
&= ~SD_SERIALIZE
;
4970 if (~cflags
& pflags
)
4976 static void free_rootdomain(struct rcu_head
*rcu
)
4978 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
4980 cpupri_cleanup(&rd
->cpupri
);
4981 free_cpumask_var(rd
->rto_mask
);
4982 free_cpumask_var(rd
->online
);
4983 free_cpumask_var(rd
->span
);
4987 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
4989 struct root_domain
*old_rd
= NULL
;
4990 unsigned long flags
;
4992 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4997 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5000 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5003 * If we dont want to free the old_rt yet then
5004 * set old_rd to NULL to skip the freeing later
5007 if (!atomic_dec_and_test(&old_rd
->refcount
))
5011 atomic_inc(&rd
->refcount
);
5014 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5015 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5018 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5021 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5024 static int init_rootdomain(struct root_domain
*rd
)
5026 memset(rd
, 0, sizeof(*rd
));
5028 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5030 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5032 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5035 if (cpupri_init(&rd
->cpupri
) != 0)
5040 free_cpumask_var(rd
->rto_mask
);
5042 free_cpumask_var(rd
->online
);
5044 free_cpumask_var(rd
->span
);
5050 * By default the system creates a single root-domain with all cpus as
5051 * members (mimicking the global state we have today).
5053 struct root_domain def_root_domain
;
5055 static void init_defrootdomain(void)
5057 init_rootdomain(&def_root_domain
);
5059 atomic_set(&def_root_domain
.refcount
, 1);
5062 static struct root_domain
*alloc_rootdomain(void)
5064 struct root_domain
*rd
;
5066 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5070 if (init_rootdomain(rd
) != 0) {
5078 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5080 struct sched_group
*tmp
, *first
;
5089 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5094 } while (sg
!= first
);
5097 static void free_sched_domain(struct rcu_head
*rcu
)
5099 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5102 * If its an overlapping domain it has private groups, iterate and
5105 if (sd
->flags
& SD_OVERLAP
) {
5106 free_sched_groups(sd
->groups
, 1);
5107 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5108 kfree(sd
->groups
->sgp
);
5114 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5116 call_rcu(&sd
->rcu
, free_sched_domain
);
5119 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5121 for (; sd
; sd
= sd
->parent
)
5122 destroy_sched_domain(sd
, cpu
);
5126 * Keep a special pointer to the highest sched_domain that has
5127 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5128 * allows us to avoid some pointer chasing select_idle_sibling().
5130 * Also keep a unique ID per domain (we use the first cpu number in
5131 * the cpumask of the domain), this allows us to quickly tell if
5132 * two cpus are in the same cache domain, see cpus_share_cache().
5134 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5135 DEFINE_PER_CPU(int, sd_llc_size
);
5136 DEFINE_PER_CPU(int, sd_llc_id
);
5138 static void update_top_cache_domain(int cpu
)
5140 struct sched_domain
*sd
;
5144 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5146 id
= cpumask_first(sched_domain_span(sd
));
5147 size
= cpumask_weight(sched_domain_span(sd
));
5150 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5151 per_cpu(sd_llc_size
, cpu
) = size
;
5152 per_cpu(sd_llc_id
, cpu
) = id
;
5156 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5157 * hold the hotplug lock.
5160 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5162 struct rq
*rq
= cpu_rq(cpu
);
5163 struct sched_domain
*tmp
;
5165 /* Remove the sched domains which do not contribute to scheduling. */
5166 for (tmp
= sd
; tmp
; ) {
5167 struct sched_domain
*parent
= tmp
->parent
;
5171 if (sd_parent_degenerate(tmp
, parent
)) {
5172 tmp
->parent
= parent
->parent
;
5174 parent
->parent
->child
= tmp
;
5176 * Transfer SD_PREFER_SIBLING down in case of a
5177 * degenerate parent; the spans match for this
5178 * so the property transfers.
5180 if (parent
->flags
& SD_PREFER_SIBLING
)
5181 tmp
->flags
|= SD_PREFER_SIBLING
;
5182 destroy_sched_domain(parent
, cpu
);
5187 if (sd
&& sd_degenerate(sd
)) {
5190 destroy_sched_domain(tmp
, cpu
);
5195 sched_domain_debug(sd
, cpu
);
5197 rq_attach_root(rq
, rd
);
5199 rcu_assign_pointer(rq
->sd
, sd
);
5200 destroy_sched_domains(tmp
, cpu
);
5202 update_top_cache_domain(cpu
);
5205 /* cpus with isolated domains */
5206 static cpumask_var_t cpu_isolated_map
;
5208 /* Setup the mask of cpus configured for isolated domains */
5209 static int __init
isolated_cpu_setup(char *str
)
5211 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5212 cpulist_parse(str
, cpu_isolated_map
);
5216 __setup("isolcpus=", isolated_cpu_setup
);
5218 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5220 return cpumask_of_node(cpu_to_node(cpu
));
5224 struct sched_domain
**__percpu sd
;
5225 struct sched_group
**__percpu sg
;
5226 struct sched_group_power
**__percpu sgp
;
5230 struct sched_domain
** __percpu sd
;
5231 struct root_domain
*rd
;
5241 struct sched_domain_topology_level
;
5243 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5244 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5246 #define SDTL_OVERLAP 0x01
5248 struct sched_domain_topology_level
{
5249 sched_domain_init_f init
;
5250 sched_domain_mask_f mask
;
5253 struct sd_data data
;
5257 * Build an iteration mask that can exclude certain CPUs from the upwards
5260 * Asymmetric node setups can result in situations where the domain tree is of
5261 * unequal depth, make sure to skip domains that already cover the entire
5264 * In that case build_sched_domains() will have terminated the iteration early
5265 * and our sibling sd spans will be empty. Domains should always include the
5266 * cpu they're built on, so check that.
5269 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5271 const struct cpumask
*span
= sched_domain_span(sd
);
5272 struct sd_data
*sdd
= sd
->private;
5273 struct sched_domain
*sibling
;
5276 for_each_cpu(i
, span
) {
5277 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5278 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5281 cpumask_set_cpu(i
, sched_group_mask(sg
));
5286 * Return the canonical balance cpu for this group, this is the first cpu
5287 * of this group that's also in the iteration mask.
5289 int group_balance_cpu(struct sched_group
*sg
)
5291 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5295 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5297 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5298 const struct cpumask
*span
= sched_domain_span(sd
);
5299 struct cpumask
*covered
= sched_domains_tmpmask
;
5300 struct sd_data
*sdd
= sd
->private;
5301 struct sched_domain
*child
;
5304 cpumask_clear(covered
);
5306 for_each_cpu(i
, span
) {
5307 struct cpumask
*sg_span
;
5309 if (cpumask_test_cpu(i
, covered
))
5312 child
= *per_cpu_ptr(sdd
->sd
, i
);
5314 /* See the comment near build_group_mask(). */
5315 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5318 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5319 GFP_KERNEL
, cpu_to_node(cpu
));
5324 sg_span
= sched_group_cpus(sg
);
5326 child
= child
->child
;
5327 cpumask_copy(sg_span
, sched_domain_span(child
));
5329 cpumask_set_cpu(i
, sg_span
);
5331 cpumask_or(covered
, covered
, sg_span
);
5333 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5334 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5335 build_group_mask(sd
, sg
);
5338 * Initialize sgp->power such that even if we mess up the
5339 * domains and no possible iteration will get us here, we won't
5342 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5345 * Make sure the first group of this domain contains the
5346 * canonical balance cpu. Otherwise the sched_domain iteration
5347 * breaks. See update_sg_lb_stats().
5349 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5350 group_balance_cpu(sg
) == cpu
)
5360 sd
->groups
= groups
;
5365 free_sched_groups(first
, 0);
5370 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5372 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5373 struct sched_domain
*child
= sd
->child
;
5376 cpu
= cpumask_first(sched_domain_span(child
));
5379 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5380 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5381 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5388 * build_sched_groups will build a circular linked list of the groups
5389 * covered by the given span, and will set each group's ->cpumask correctly,
5390 * and ->cpu_power to 0.
5392 * Assumes the sched_domain tree is fully constructed
5395 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5397 struct sched_group
*first
= NULL
, *last
= NULL
;
5398 struct sd_data
*sdd
= sd
->private;
5399 const struct cpumask
*span
= sched_domain_span(sd
);
5400 struct cpumask
*covered
;
5403 get_group(cpu
, sdd
, &sd
->groups
);
5404 atomic_inc(&sd
->groups
->ref
);
5406 if (cpu
!= cpumask_first(span
))
5409 lockdep_assert_held(&sched_domains_mutex
);
5410 covered
= sched_domains_tmpmask
;
5412 cpumask_clear(covered
);
5414 for_each_cpu(i
, span
) {
5415 struct sched_group
*sg
;
5418 if (cpumask_test_cpu(i
, covered
))
5421 group
= get_group(i
, sdd
, &sg
);
5422 cpumask_clear(sched_group_cpus(sg
));
5424 cpumask_setall(sched_group_mask(sg
));
5426 for_each_cpu(j
, span
) {
5427 if (get_group(j
, sdd
, NULL
) != group
)
5430 cpumask_set_cpu(j
, covered
);
5431 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5446 * Initialize sched groups cpu_power.
5448 * cpu_power indicates the capacity of sched group, which is used while
5449 * distributing the load between different sched groups in a sched domain.
5450 * Typically cpu_power for all the groups in a sched domain will be same unless
5451 * there are asymmetries in the topology. If there are asymmetries, group
5452 * having more cpu_power will pickup more load compared to the group having
5455 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5457 struct sched_group
*sg
= sd
->groups
;
5462 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5464 } while (sg
!= sd
->groups
);
5466 if (cpu
!= group_balance_cpu(sg
))
5469 update_group_power(sd
, cpu
);
5470 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5473 int __weak
arch_sd_sibling_asym_packing(void)
5475 return 0*SD_ASYM_PACKING
;
5479 * Initializers for schedule domains
5480 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5483 #ifdef CONFIG_SCHED_DEBUG
5484 # define SD_INIT_NAME(sd, type) sd->name = #type
5486 # define SD_INIT_NAME(sd, type) do { } while (0)
5489 #define SD_INIT_FUNC(type) \
5490 static noinline struct sched_domain * \
5491 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5493 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5494 *sd = SD_##type##_INIT; \
5495 SD_INIT_NAME(sd, type); \
5496 sd->private = &tl->data; \
5501 #ifdef CONFIG_SCHED_SMT
5502 SD_INIT_FUNC(SIBLING
)
5504 #ifdef CONFIG_SCHED_MC
5507 #ifdef CONFIG_SCHED_BOOK
5511 static int default_relax_domain_level
= -1;
5512 int sched_domain_level_max
;
5514 static int __init
setup_relax_domain_level(char *str
)
5516 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5517 pr_warn("Unable to set relax_domain_level\n");
5521 __setup("relax_domain_level=", setup_relax_domain_level
);
5523 static void set_domain_attribute(struct sched_domain
*sd
,
5524 struct sched_domain_attr
*attr
)
5528 if (!attr
|| attr
->relax_domain_level
< 0) {
5529 if (default_relax_domain_level
< 0)
5532 request
= default_relax_domain_level
;
5534 request
= attr
->relax_domain_level
;
5535 if (request
< sd
->level
) {
5536 /* turn off idle balance on this domain */
5537 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5539 /* turn on idle balance on this domain */
5540 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5544 static void __sdt_free(const struct cpumask
*cpu_map
);
5545 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5547 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5548 const struct cpumask
*cpu_map
)
5552 if (!atomic_read(&d
->rd
->refcount
))
5553 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5555 free_percpu(d
->sd
); /* fall through */
5557 __sdt_free(cpu_map
); /* fall through */
5563 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5564 const struct cpumask
*cpu_map
)
5566 memset(d
, 0, sizeof(*d
));
5568 if (__sdt_alloc(cpu_map
))
5569 return sa_sd_storage
;
5570 d
->sd
= alloc_percpu(struct sched_domain
*);
5572 return sa_sd_storage
;
5573 d
->rd
= alloc_rootdomain();
5576 return sa_rootdomain
;
5580 * NULL the sd_data elements we've used to build the sched_domain and
5581 * sched_group structure so that the subsequent __free_domain_allocs()
5582 * will not free the data we're using.
5584 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5586 struct sd_data
*sdd
= sd
->private;
5588 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5589 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5591 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5592 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5594 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5595 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5598 #ifdef CONFIG_SCHED_SMT
5599 static const struct cpumask
*cpu_smt_mask(int cpu
)
5601 return topology_thread_cpumask(cpu
);
5606 * Topology list, bottom-up.
5608 static struct sched_domain_topology_level default_topology
[] = {
5609 #ifdef CONFIG_SCHED_SMT
5610 { sd_init_SIBLING
, cpu_smt_mask
, },
5612 #ifdef CONFIG_SCHED_MC
5613 { sd_init_MC
, cpu_coregroup_mask
, },
5615 #ifdef CONFIG_SCHED_BOOK
5616 { sd_init_BOOK
, cpu_book_mask
, },
5618 { sd_init_CPU
, cpu_cpu_mask
, },
5622 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5624 #define for_each_sd_topology(tl) \
5625 for (tl = sched_domain_topology; tl->init; tl++)
5629 static int sched_domains_numa_levels
;
5630 static int *sched_domains_numa_distance
;
5631 static struct cpumask
***sched_domains_numa_masks
;
5632 static int sched_domains_curr_level
;
5634 static inline int sd_local_flags(int level
)
5636 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5639 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
5642 static struct sched_domain
*
5643 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
5645 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5646 int level
= tl
->numa_level
;
5647 int sd_weight
= cpumask_weight(
5648 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
5650 *sd
= (struct sched_domain
){
5651 .min_interval
= sd_weight
,
5652 .max_interval
= 2*sd_weight
,
5654 .imbalance_pct
= 125,
5655 .cache_nice_tries
= 2,
5662 .flags
= 1*SD_LOAD_BALANCE
5663 | 1*SD_BALANCE_NEWIDLE
5668 | 0*SD_SHARE_CPUPOWER
5669 | 0*SD_SHARE_PKG_RESOURCES
5671 | 0*SD_PREFER_SIBLING
5672 | sd_local_flags(level
)
5674 .last_balance
= jiffies
,
5675 .balance_interval
= sd_weight
,
5677 SD_INIT_NAME(sd
, NUMA
);
5678 sd
->private = &tl
->data
;
5681 * Ugly hack to pass state to sd_numa_mask()...
5683 sched_domains_curr_level
= tl
->numa_level
;
5688 static const struct cpumask
*sd_numa_mask(int cpu
)
5690 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
5693 static void sched_numa_warn(const char *str
)
5695 static int done
= false;
5703 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
5705 for (i
= 0; i
< nr_node_ids
; i
++) {
5706 printk(KERN_WARNING
" ");
5707 for (j
= 0; j
< nr_node_ids
; j
++)
5708 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
5709 printk(KERN_CONT
"\n");
5711 printk(KERN_WARNING
"\n");
5714 static bool find_numa_distance(int distance
)
5718 if (distance
== node_distance(0, 0))
5721 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5722 if (sched_domains_numa_distance
[i
] == distance
)
5729 static void sched_init_numa(void)
5731 int next_distance
, curr_distance
= node_distance(0, 0);
5732 struct sched_domain_topology_level
*tl
;
5736 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
5737 if (!sched_domains_numa_distance
)
5741 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5742 * unique distances in the node_distance() table.
5744 * Assumes node_distance(0,j) includes all distances in
5745 * node_distance(i,j) in order to avoid cubic time.
5747 next_distance
= curr_distance
;
5748 for (i
= 0; i
< nr_node_ids
; i
++) {
5749 for (j
= 0; j
< nr_node_ids
; j
++) {
5750 for (k
= 0; k
< nr_node_ids
; k
++) {
5751 int distance
= node_distance(i
, k
);
5753 if (distance
> curr_distance
&&
5754 (distance
< next_distance
||
5755 next_distance
== curr_distance
))
5756 next_distance
= distance
;
5759 * While not a strong assumption it would be nice to know
5760 * about cases where if node A is connected to B, B is not
5761 * equally connected to A.
5763 if (sched_debug() && node_distance(k
, i
) != distance
)
5764 sched_numa_warn("Node-distance not symmetric");
5766 if (sched_debug() && i
&& !find_numa_distance(distance
))
5767 sched_numa_warn("Node-0 not representative");
5769 if (next_distance
!= curr_distance
) {
5770 sched_domains_numa_distance
[level
++] = next_distance
;
5771 sched_domains_numa_levels
= level
;
5772 curr_distance
= next_distance
;
5777 * In case of sched_debug() we verify the above assumption.
5783 * 'level' contains the number of unique distances, excluding the
5784 * identity distance node_distance(i,i).
5786 * The sched_domains_numa_distance[] array includes the actual distance
5791 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5792 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5793 * the array will contain less then 'level' members. This could be
5794 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5795 * in other functions.
5797 * We reset it to 'level' at the end of this function.
5799 sched_domains_numa_levels
= 0;
5801 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
5802 if (!sched_domains_numa_masks
)
5806 * Now for each level, construct a mask per node which contains all
5807 * cpus of nodes that are that many hops away from us.
5809 for (i
= 0; i
< level
; i
++) {
5810 sched_domains_numa_masks
[i
] =
5811 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
5812 if (!sched_domains_numa_masks
[i
])
5815 for (j
= 0; j
< nr_node_ids
; j
++) {
5816 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
5820 sched_domains_numa_masks
[i
][j
] = mask
;
5822 for (k
= 0; k
< nr_node_ids
; k
++) {
5823 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
5826 cpumask_or(mask
, mask
, cpumask_of_node(k
));
5831 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
5832 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
5837 * Copy the default topology bits..
5839 for (i
= 0; default_topology
[i
].init
; i
++)
5840 tl
[i
] = default_topology
[i
];
5843 * .. and append 'j' levels of NUMA goodness.
5845 for (j
= 0; j
< level
; i
++, j
++) {
5846 tl
[i
] = (struct sched_domain_topology_level
){
5847 .init
= sd_numa_init
,
5848 .mask
= sd_numa_mask
,
5849 .flags
= SDTL_OVERLAP
,
5854 sched_domain_topology
= tl
;
5856 sched_domains_numa_levels
= level
;
5859 static void sched_domains_numa_masks_set(int cpu
)
5862 int node
= cpu_to_node(cpu
);
5864 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5865 for (j
= 0; j
< nr_node_ids
; j
++) {
5866 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
5867 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
5872 static void sched_domains_numa_masks_clear(int cpu
)
5875 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
5876 for (j
= 0; j
< nr_node_ids
; j
++)
5877 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
5882 * Update sched_domains_numa_masks[level][node] array when new cpus
5885 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
5886 unsigned long action
,
5889 int cpu
= (long)hcpu
;
5891 switch (action
& ~CPU_TASKS_FROZEN
) {
5893 sched_domains_numa_masks_set(cpu
);
5897 sched_domains_numa_masks_clear(cpu
);
5907 static inline void sched_init_numa(void)
5911 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
5912 unsigned long action
,
5917 #endif /* CONFIG_NUMA */
5919 static int __sdt_alloc(const struct cpumask
*cpu_map
)
5921 struct sched_domain_topology_level
*tl
;
5924 for_each_sd_topology(tl
) {
5925 struct sd_data
*sdd
= &tl
->data
;
5927 sdd
->sd
= alloc_percpu(struct sched_domain
*);
5931 sdd
->sg
= alloc_percpu(struct sched_group
*);
5935 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
5939 for_each_cpu(j
, cpu_map
) {
5940 struct sched_domain
*sd
;
5941 struct sched_group
*sg
;
5942 struct sched_group_power
*sgp
;
5944 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
5945 GFP_KERNEL
, cpu_to_node(j
));
5949 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
5951 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5952 GFP_KERNEL
, cpu_to_node(j
));
5958 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
5960 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
5961 GFP_KERNEL
, cpu_to_node(j
));
5965 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
5972 static void __sdt_free(const struct cpumask
*cpu_map
)
5974 struct sched_domain_topology_level
*tl
;
5977 for_each_sd_topology(tl
) {
5978 struct sd_data
*sdd
= &tl
->data
;
5980 for_each_cpu(j
, cpu_map
) {
5981 struct sched_domain
*sd
;
5984 sd
= *per_cpu_ptr(sdd
->sd
, j
);
5985 if (sd
&& (sd
->flags
& SD_OVERLAP
))
5986 free_sched_groups(sd
->groups
, 0);
5987 kfree(*per_cpu_ptr(sdd
->sd
, j
));
5991 kfree(*per_cpu_ptr(sdd
->sg
, j
));
5993 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
5995 free_percpu(sdd
->sd
);
5997 free_percpu(sdd
->sg
);
5999 free_percpu(sdd
->sgp
);
6004 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6005 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6006 struct sched_domain
*child
, int cpu
)
6008 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6012 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6014 sd
->level
= child
->level
+ 1;
6015 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6019 set_domain_attribute(sd
, attr
);
6025 * Build sched domains for a given set of cpus and attach the sched domains
6026 * to the individual cpus
6028 static int build_sched_domains(const struct cpumask
*cpu_map
,
6029 struct sched_domain_attr
*attr
)
6031 enum s_alloc alloc_state
;
6032 struct sched_domain
*sd
;
6034 int i
, ret
= -ENOMEM
;
6036 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6037 if (alloc_state
!= sa_rootdomain
)
6040 /* Set up domains for cpus specified by the cpu_map. */
6041 for_each_cpu(i
, cpu_map
) {
6042 struct sched_domain_topology_level
*tl
;
6045 for_each_sd_topology(tl
) {
6046 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6047 if (tl
== sched_domain_topology
)
6048 *per_cpu_ptr(d
.sd
, i
) = sd
;
6049 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6050 sd
->flags
|= SD_OVERLAP
;
6051 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6056 /* Build the groups for the domains */
6057 for_each_cpu(i
, cpu_map
) {
6058 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6059 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6060 if (sd
->flags
& SD_OVERLAP
) {
6061 if (build_overlap_sched_groups(sd
, i
))
6064 if (build_sched_groups(sd
, i
))
6070 /* Calculate CPU power for physical packages and nodes */
6071 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6072 if (!cpumask_test_cpu(i
, cpu_map
))
6075 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6076 claim_allocations(i
, sd
);
6077 init_sched_groups_power(i
, sd
);
6081 /* Attach the domains */
6083 for_each_cpu(i
, cpu_map
) {
6084 sd
= *per_cpu_ptr(d
.sd
, i
);
6085 cpu_attach_domain(sd
, d
.rd
, i
);
6091 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6095 static cpumask_var_t
*doms_cur
; /* current sched domains */
6096 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6097 static struct sched_domain_attr
*dattr_cur
;
6098 /* attribues of custom domains in 'doms_cur' */
6101 * Special case: If a kmalloc of a doms_cur partition (array of
6102 * cpumask) fails, then fallback to a single sched domain,
6103 * as determined by the single cpumask fallback_doms.
6105 static cpumask_var_t fallback_doms
;
6108 * arch_update_cpu_topology lets virtualized architectures update the
6109 * cpu core maps. It is supposed to return 1 if the topology changed
6110 * or 0 if it stayed the same.
6112 int __attribute__((weak
)) arch_update_cpu_topology(void)
6117 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6120 cpumask_var_t
*doms
;
6122 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6125 for (i
= 0; i
< ndoms
; i
++) {
6126 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6127 free_sched_domains(doms
, i
);
6134 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6137 for (i
= 0; i
< ndoms
; i
++)
6138 free_cpumask_var(doms
[i
]);
6143 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6144 * For now this just excludes isolated cpus, but could be used to
6145 * exclude other special cases in the future.
6147 static int init_sched_domains(const struct cpumask
*cpu_map
)
6151 arch_update_cpu_topology();
6153 doms_cur
= alloc_sched_domains(ndoms_cur
);
6155 doms_cur
= &fallback_doms
;
6156 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6157 err
= build_sched_domains(doms_cur
[0], NULL
);
6158 register_sched_domain_sysctl();
6164 * Detach sched domains from a group of cpus specified in cpu_map
6165 * These cpus will now be attached to the NULL domain
6167 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6172 for_each_cpu(i
, cpu_map
)
6173 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6177 /* handle null as "default" */
6178 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6179 struct sched_domain_attr
*new, int idx_new
)
6181 struct sched_domain_attr tmp
;
6188 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6189 new ? (new + idx_new
) : &tmp
,
6190 sizeof(struct sched_domain_attr
));
6194 * Partition sched domains as specified by the 'ndoms_new'
6195 * cpumasks in the array doms_new[] of cpumasks. This compares
6196 * doms_new[] to the current sched domain partitioning, doms_cur[].
6197 * It destroys each deleted domain and builds each new domain.
6199 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6200 * The masks don't intersect (don't overlap.) We should setup one
6201 * sched domain for each mask. CPUs not in any of the cpumasks will
6202 * not be load balanced. If the same cpumask appears both in the
6203 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6206 * The passed in 'doms_new' should be allocated using
6207 * alloc_sched_domains. This routine takes ownership of it and will
6208 * free_sched_domains it when done with it. If the caller failed the
6209 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6210 * and partition_sched_domains() will fallback to the single partition
6211 * 'fallback_doms', it also forces the domains to be rebuilt.
6213 * If doms_new == NULL it will be replaced with cpu_online_mask.
6214 * ndoms_new == 0 is a special case for destroying existing domains,
6215 * and it will not create the default domain.
6217 * Call with hotplug lock held
6219 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6220 struct sched_domain_attr
*dattr_new
)
6225 mutex_lock(&sched_domains_mutex
);
6227 /* always unregister in case we don't destroy any domains */
6228 unregister_sched_domain_sysctl();
6230 /* Let architecture update cpu core mappings. */
6231 new_topology
= arch_update_cpu_topology();
6233 n
= doms_new
? ndoms_new
: 0;
6235 /* Destroy deleted domains */
6236 for (i
= 0; i
< ndoms_cur
; i
++) {
6237 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6238 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6239 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6242 /* no match - a current sched domain not in new doms_new[] */
6243 detach_destroy_domains(doms_cur
[i
]);
6249 if (doms_new
== NULL
) {
6251 doms_new
= &fallback_doms
;
6252 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6253 WARN_ON_ONCE(dattr_new
);
6256 /* Build new domains */
6257 for (i
= 0; i
< ndoms_new
; i
++) {
6258 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6259 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6260 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6263 /* no match - add a new doms_new */
6264 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6269 /* Remember the new sched domains */
6270 if (doms_cur
!= &fallback_doms
)
6271 free_sched_domains(doms_cur
, ndoms_cur
);
6272 kfree(dattr_cur
); /* kfree(NULL) is safe */
6273 doms_cur
= doms_new
;
6274 dattr_cur
= dattr_new
;
6275 ndoms_cur
= ndoms_new
;
6277 register_sched_domain_sysctl();
6279 mutex_unlock(&sched_domains_mutex
);
6282 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6285 * Update cpusets according to cpu_active mask. If cpusets are
6286 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6287 * around partition_sched_domains().
6289 * If we come here as part of a suspend/resume, don't touch cpusets because we
6290 * want to restore it back to its original state upon resume anyway.
6292 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6296 case CPU_ONLINE_FROZEN
:
6297 case CPU_DOWN_FAILED_FROZEN
:
6300 * num_cpus_frozen tracks how many CPUs are involved in suspend
6301 * resume sequence. As long as this is not the last online
6302 * operation in the resume sequence, just build a single sched
6303 * domain, ignoring cpusets.
6306 if (likely(num_cpus_frozen
)) {
6307 partition_sched_domains(1, NULL
, NULL
);
6312 * This is the last CPU online operation. So fall through and
6313 * restore the original sched domains by considering the
6314 * cpuset configurations.
6318 case CPU_DOWN_FAILED
:
6319 cpuset_update_active_cpus(true);
6327 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6331 case CPU_DOWN_PREPARE
:
6332 cpuset_update_active_cpus(false);
6334 case CPU_DOWN_PREPARE_FROZEN
:
6336 partition_sched_domains(1, NULL
, NULL
);
6344 void __init
sched_init_smp(void)
6346 cpumask_var_t non_isolated_cpus
;
6348 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6349 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6354 mutex_lock(&sched_domains_mutex
);
6355 init_sched_domains(cpu_active_mask
);
6356 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6357 if (cpumask_empty(non_isolated_cpus
))
6358 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6359 mutex_unlock(&sched_domains_mutex
);
6362 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6363 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6364 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6368 /* Move init over to a non-isolated CPU */
6369 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6371 sched_init_granularity();
6372 free_cpumask_var(non_isolated_cpus
);
6374 init_sched_rt_class();
6377 void __init
sched_init_smp(void)
6379 sched_init_granularity();
6381 #endif /* CONFIG_SMP */
6383 const_debug
unsigned int sysctl_timer_migration
= 1;
6385 int in_sched_functions(unsigned long addr
)
6387 return in_lock_functions(addr
) ||
6388 (addr
>= (unsigned long)__sched_text_start
6389 && addr
< (unsigned long)__sched_text_end
);
6392 #ifdef CONFIG_CGROUP_SCHED
6394 * Default task group.
6395 * Every task in system belongs to this group at bootup.
6397 struct task_group root_task_group
;
6398 LIST_HEAD(task_groups
);
6401 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6403 void __init
sched_init(void)
6406 unsigned long alloc_size
= 0, ptr
;
6408 #ifdef CONFIG_FAIR_GROUP_SCHED
6409 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6411 #ifdef CONFIG_RT_GROUP_SCHED
6412 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6414 #ifdef CONFIG_CPUMASK_OFFSTACK
6415 alloc_size
+= num_possible_cpus() * cpumask_size();
6418 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6420 #ifdef CONFIG_FAIR_GROUP_SCHED
6421 root_task_group
.se
= (struct sched_entity
**)ptr
;
6422 ptr
+= nr_cpu_ids
* sizeof(void **);
6424 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6425 ptr
+= nr_cpu_ids
* sizeof(void **);
6427 #endif /* CONFIG_FAIR_GROUP_SCHED */
6428 #ifdef CONFIG_RT_GROUP_SCHED
6429 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6430 ptr
+= nr_cpu_ids
* sizeof(void **);
6432 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6433 ptr
+= nr_cpu_ids
* sizeof(void **);
6435 #endif /* CONFIG_RT_GROUP_SCHED */
6436 #ifdef CONFIG_CPUMASK_OFFSTACK
6437 for_each_possible_cpu(i
) {
6438 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6439 ptr
+= cpumask_size();
6441 #endif /* CONFIG_CPUMASK_OFFSTACK */
6445 init_defrootdomain();
6448 init_rt_bandwidth(&def_rt_bandwidth
,
6449 global_rt_period(), global_rt_runtime());
6451 #ifdef CONFIG_RT_GROUP_SCHED
6452 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6453 global_rt_period(), global_rt_runtime());
6454 #endif /* CONFIG_RT_GROUP_SCHED */
6456 #ifdef CONFIG_CGROUP_SCHED
6457 list_add(&root_task_group
.list
, &task_groups
);
6458 INIT_LIST_HEAD(&root_task_group
.children
);
6459 INIT_LIST_HEAD(&root_task_group
.siblings
);
6460 autogroup_init(&init_task
);
6462 #endif /* CONFIG_CGROUP_SCHED */
6464 for_each_possible_cpu(i
) {
6468 raw_spin_lock_init(&rq
->lock
);
6470 rq
->calc_load_active
= 0;
6471 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6472 init_cfs_rq(&rq
->cfs
);
6473 init_rt_rq(&rq
->rt
, rq
);
6474 #ifdef CONFIG_FAIR_GROUP_SCHED
6475 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6476 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6478 * How much cpu bandwidth does root_task_group get?
6480 * In case of task-groups formed thr' the cgroup filesystem, it
6481 * gets 100% of the cpu resources in the system. This overall
6482 * system cpu resource is divided among the tasks of
6483 * root_task_group and its child task-groups in a fair manner,
6484 * based on each entity's (task or task-group's) weight
6485 * (se->load.weight).
6487 * In other words, if root_task_group has 10 tasks of weight
6488 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6489 * then A0's share of the cpu resource is:
6491 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6493 * We achieve this by letting root_task_group's tasks sit
6494 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6496 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6497 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6498 #endif /* CONFIG_FAIR_GROUP_SCHED */
6500 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6501 #ifdef CONFIG_RT_GROUP_SCHED
6502 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6503 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6506 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6507 rq
->cpu_load
[j
] = 0;
6509 rq
->last_load_update_tick
= jiffies
;
6514 rq
->cpu_power
= SCHED_POWER_SCALE
;
6515 rq
->post_schedule
= 0;
6516 rq
->active_balance
= 0;
6517 rq
->next_balance
= jiffies
;
6522 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6524 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6526 rq_attach_root(rq
, &def_root_domain
);
6527 #ifdef CONFIG_NO_HZ_COMMON
6530 #ifdef CONFIG_NO_HZ_FULL
6531 rq
->last_sched_tick
= 0;
6535 atomic_set(&rq
->nr_iowait
, 0);
6538 set_load_weight(&init_task
);
6540 #ifdef CONFIG_PREEMPT_NOTIFIERS
6541 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6544 #ifdef CONFIG_RT_MUTEXES
6545 plist_head_init(&init_task
.pi_waiters
);
6549 * The boot idle thread does lazy MMU switching as well:
6551 atomic_inc(&init_mm
.mm_count
);
6552 enter_lazy_tlb(&init_mm
, current
);
6555 * Make us the idle thread. Technically, schedule() should not be
6556 * called from this thread, however somewhere below it might be,
6557 * but because we are the idle thread, we just pick up running again
6558 * when this runqueue becomes "idle".
6560 init_idle(current
, smp_processor_id());
6562 calc_load_update
= jiffies
+ LOAD_FREQ
;
6565 * During early bootup we pretend to be a normal task:
6567 current
->sched_class
= &fair_sched_class
;
6570 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6571 /* May be allocated at isolcpus cmdline parse time */
6572 if (cpu_isolated_map
== NULL
)
6573 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6574 idle_thread_set_boot_cpu();
6576 init_sched_fair_class();
6578 scheduler_running
= 1;
6581 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6582 static inline int preempt_count_equals(int preempt_offset
)
6584 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6586 return (nested
== preempt_offset
);
6589 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6591 static unsigned long prev_jiffy
; /* ratelimiting */
6593 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6594 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6595 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6597 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6599 prev_jiffy
= jiffies
;
6602 "BUG: sleeping function called from invalid context at %s:%d\n",
6605 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6606 in_atomic(), irqs_disabled(),
6607 current
->pid
, current
->comm
);
6609 debug_show_held_locks(current
);
6610 if (irqs_disabled())
6611 print_irqtrace_events(current
);
6614 EXPORT_SYMBOL(__might_sleep
);
6617 #ifdef CONFIG_MAGIC_SYSRQ
6618 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6620 const struct sched_class
*prev_class
= p
->sched_class
;
6621 int old_prio
= p
->prio
;
6626 dequeue_task(rq
, p
, 0);
6627 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
6629 enqueue_task(rq
, p
, 0);
6630 resched_task(rq
->curr
);
6633 check_class_changed(rq
, p
, prev_class
, old_prio
);
6636 void normalize_rt_tasks(void)
6638 struct task_struct
*g
, *p
;
6639 unsigned long flags
;
6642 read_lock_irqsave(&tasklist_lock
, flags
);
6643 do_each_thread(g
, p
) {
6645 * Only normalize user tasks:
6650 p
->se
.exec_start
= 0;
6651 #ifdef CONFIG_SCHEDSTATS
6652 p
->se
.statistics
.wait_start
= 0;
6653 p
->se
.statistics
.sleep_start
= 0;
6654 p
->se
.statistics
.block_start
= 0;
6659 * Renice negative nice level userspace
6662 if (TASK_NICE(p
) < 0 && p
->mm
)
6663 set_user_nice(p
, 0);
6667 raw_spin_lock(&p
->pi_lock
);
6668 rq
= __task_rq_lock(p
);
6670 normalize_task(rq
, p
);
6672 __task_rq_unlock(rq
);
6673 raw_spin_unlock(&p
->pi_lock
);
6674 } while_each_thread(g
, p
);
6676 read_unlock_irqrestore(&tasklist_lock
, flags
);
6679 #endif /* CONFIG_MAGIC_SYSRQ */
6681 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6683 * These functions are only useful for the IA64 MCA handling, or kdb.
6685 * They can only be called when the whole system has been
6686 * stopped - every CPU needs to be quiescent, and no scheduling
6687 * activity can take place. Using them for anything else would
6688 * be a serious bug, and as a result, they aren't even visible
6689 * under any other configuration.
6693 * curr_task - return the current task for a given cpu.
6694 * @cpu: the processor in question.
6696 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6698 * Return: The current task for @cpu.
6700 struct task_struct
*curr_task(int cpu
)
6702 return cpu_curr(cpu
);
6705 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6709 * set_curr_task - set the current task for a given cpu.
6710 * @cpu: the processor in question.
6711 * @p: the task pointer to set.
6713 * Description: This function must only be used when non-maskable interrupts
6714 * are serviced on a separate stack. It allows the architecture to switch the
6715 * notion of the current task on a cpu in a non-blocking manner. This function
6716 * must be called with all CPU's synchronized, and interrupts disabled, the
6717 * and caller must save the original value of the current task (see
6718 * curr_task() above) and restore that value before reenabling interrupts and
6719 * re-starting the system.
6721 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6723 void set_curr_task(int cpu
, struct task_struct
*p
)
6730 #ifdef CONFIG_CGROUP_SCHED
6731 /* task_group_lock serializes the addition/removal of task groups */
6732 static DEFINE_SPINLOCK(task_group_lock
);
6734 static void free_sched_group(struct task_group
*tg
)
6736 free_fair_sched_group(tg
);
6737 free_rt_sched_group(tg
);
6742 /* allocate runqueue etc for a new task group */
6743 struct task_group
*sched_create_group(struct task_group
*parent
)
6745 struct task_group
*tg
;
6747 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
6749 return ERR_PTR(-ENOMEM
);
6751 if (!alloc_fair_sched_group(tg
, parent
))
6754 if (!alloc_rt_sched_group(tg
, parent
))
6760 free_sched_group(tg
);
6761 return ERR_PTR(-ENOMEM
);
6764 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6766 unsigned long flags
;
6768 spin_lock_irqsave(&task_group_lock
, flags
);
6769 list_add_rcu(&tg
->list
, &task_groups
);
6771 WARN_ON(!parent
); /* root should already exist */
6773 tg
->parent
= parent
;
6774 INIT_LIST_HEAD(&tg
->children
);
6775 list_add_rcu(&tg
->siblings
, &parent
->children
);
6776 spin_unlock_irqrestore(&task_group_lock
, flags
);
6779 /* rcu callback to free various structures associated with a task group */
6780 static void free_sched_group_rcu(struct rcu_head
*rhp
)
6782 /* now it should be safe to free those cfs_rqs */
6783 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
6786 /* Destroy runqueue etc associated with a task group */
6787 void sched_destroy_group(struct task_group
*tg
)
6789 /* wait for possible concurrent references to cfs_rqs complete */
6790 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
6793 void sched_offline_group(struct task_group
*tg
)
6795 unsigned long flags
;
6798 /* end participation in shares distribution */
6799 for_each_possible_cpu(i
)
6800 unregister_fair_sched_group(tg
, i
);
6802 spin_lock_irqsave(&task_group_lock
, flags
);
6803 list_del_rcu(&tg
->list
);
6804 list_del_rcu(&tg
->siblings
);
6805 spin_unlock_irqrestore(&task_group_lock
, flags
);
6808 /* change task's runqueue when it moves between groups.
6809 * The caller of this function should have put the task in its new group
6810 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6811 * reflect its new group.
6813 void sched_move_task(struct task_struct
*tsk
)
6815 struct task_group
*tg
;
6817 unsigned long flags
;
6820 rq
= task_rq_lock(tsk
, &flags
);
6822 running
= task_current(rq
, tsk
);
6826 dequeue_task(rq
, tsk
, 0);
6827 if (unlikely(running
))
6828 tsk
->sched_class
->put_prev_task(rq
, tsk
);
6830 tg
= container_of(task_css_check(tsk
, cpu_cgroup_subsys_id
,
6831 lockdep_is_held(&tsk
->sighand
->siglock
)),
6832 struct task_group
, css
);
6833 tg
= autogroup_task_group(tsk
, tg
);
6834 tsk
->sched_task_group
= tg
;
6836 #ifdef CONFIG_FAIR_GROUP_SCHED
6837 if (tsk
->sched_class
->task_move_group
)
6838 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
6841 set_task_rq(tsk
, task_cpu(tsk
));
6843 if (unlikely(running
))
6844 tsk
->sched_class
->set_curr_task(rq
);
6846 enqueue_task(rq
, tsk
, 0);
6848 task_rq_unlock(rq
, tsk
, &flags
);
6850 #endif /* CONFIG_CGROUP_SCHED */
6852 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6853 static unsigned long to_ratio(u64 period
, u64 runtime
)
6855 if (runtime
== RUNTIME_INF
)
6858 return div64_u64(runtime
<< 20, period
);
6862 #ifdef CONFIG_RT_GROUP_SCHED
6864 * Ensure that the real time constraints are schedulable.
6866 static DEFINE_MUTEX(rt_constraints_mutex
);
6868 /* Must be called with tasklist_lock held */
6869 static inline int tg_has_rt_tasks(struct task_group
*tg
)
6871 struct task_struct
*g
, *p
;
6873 do_each_thread(g
, p
) {
6874 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
6876 } while_each_thread(g
, p
);
6881 struct rt_schedulable_data
{
6882 struct task_group
*tg
;
6887 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
6889 struct rt_schedulable_data
*d
= data
;
6890 struct task_group
*child
;
6891 unsigned long total
, sum
= 0;
6892 u64 period
, runtime
;
6894 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6895 runtime
= tg
->rt_bandwidth
.rt_runtime
;
6898 period
= d
->rt_period
;
6899 runtime
= d
->rt_runtime
;
6903 * Cannot have more runtime than the period.
6905 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
6909 * Ensure we don't starve existing RT tasks.
6911 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
6914 total
= to_ratio(period
, runtime
);
6917 * Nobody can have more than the global setting allows.
6919 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
6923 * The sum of our children's runtime should not exceed our own.
6925 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
6926 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
6927 runtime
= child
->rt_bandwidth
.rt_runtime
;
6929 if (child
== d
->tg
) {
6930 period
= d
->rt_period
;
6931 runtime
= d
->rt_runtime
;
6934 sum
+= to_ratio(period
, runtime
);
6943 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
6947 struct rt_schedulable_data data
= {
6949 .rt_period
= period
,
6950 .rt_runtime
= runtime
,
6954 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
6960 static int tg_set_rt_bandwidth(struct task_group
*tg
,
6961 u64 rt_period
, u64 rt_runtime
)
6965 mutex_lock(&rt_constraints_mutex
);
6966 read_lock(&tasklist_lock
);
6967 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
6971 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6972 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
6973 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
6975 for_each_possible_cpu(i
) {
6976 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
6978 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6979 rt_rq
->rt_runtime
= rt_runtime
;
6980 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6982 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6984 read_unlock(&tasklist_lock
);
6985 mutex_unlock(&rt_constraints_mutex
);
6990 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
6992 u64 rt_runtime
, rt_period
;
6994 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6995 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
6996 if (rt_runtime_us
< 0)
6997 rt_runtime
= RUNTIME_INF
;
6999 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7002 static long sched_group_rt_runtime(struct task_group
*tg
)
7006 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7009 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7010 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7011 return rt_runtime_us
;
7014 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7016 u64 rt_runtime
, rt_period
;
7018 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7019 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7024 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7027 static long sched_group_rt_period(struct task_group
*tg
)
7031 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7032 do_div(rt_period_us
, NSEC_PER_USEC
);
7033 return rt_period_us
;
7036 static int sched_rt_global_constraints(void)
7038 u64 runtime
, period
;
7041 if (sysctl_sched_rt_period
<= 0)
7044 runtime
= global_rt_runtime();
7045 period
= global_rt_period();
7048 * Sanity check on the sysctl variables.
7050 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7053 mutex_lock(&rt_constraints_mutex
);
7054 read_lock(&tasklist_lock
);
7055 ret
= __rt_schedulable(NULL
, 0, 0);
7056 read_unlock(&tasklist_lock
);
7057 mutex_unlock(&rt_constraints_mutex
);
7062 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7064 /* Don't accept realtime tasks when there is no way for them to run */
7065 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7071 #else /* !CONFIG_RT_GROUP_SCHED */
7072 static int sched_rt_global_constraints(void)
7074 unsigned long flags
;
7077 if (sysctl_sched_rt_period
<= 0)
7081 * There's always some RT tasks in the root group
7082 * -- migration, kstopmachine etc..
7084 if (sysctl_sched_rt_runtime
== 0)
7087 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7088 for_each_possible_cpu(i
) {
7089 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7091 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7092 rt_rq
->rt_runtime
= global_rt_runtime();
7093 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7095 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7099 #endif /* CONFIG_RT_GROUP_SCHED */
7101 int sched_rr_handler(struct ctl_table
*table
, int write
,
7102 void __user
*buffer
, size_t *lenp
,
7106 static DEFINE_MUTEX(mutex
);
7109 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7110 /* make sure that internally we keep jiffies */
7111 /* also, writing zero resets timeslice to default */
7112 if (!ret
&& write
) {
7113 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7114 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7116 mutex_unlock(&mutex
);
7120 int sched_rt_handler(struct ctl_table
*table
, int write
,
7121 void __user
*buffer
, size_t *lenp
,
7125 int old_period
, old_runtime
;
7126 static DEFINE_MUTEX(mutex
);
7129 old_period
= sysctl_sched_rt_period
;
7130 old_runtime
= sysctl_sched_rt_runtime
;
7132 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7134 if (!ret
&& write
) {
7135 ret
= sched_rt_global_constraints();
7137 sysctl_sched_rt_period
= old_period
;
7138 sysctl_sched_rt_runtime
= old_runtime
;
7140 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7141 def_rt_bandwidth
.rt_period
=
7142 ns_to_ktime(global_rt_period());
7145 mutex_unlock(&mutex
);
7150 #ifdef CONFIG_CGROUP_SCHED
7152 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7154 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7157 static struct cgroup_subsys_state
*
7158 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7160 struct task_group
*parent
= css_tg(parent_css
);
7161 struct task_group
*tg
;
7164 /* This is early initialization for the top cgroup */
7165 return &root_task_group
.css
;
7168 tg
= sched_create_group(parent
);
7170 return ERR_PTR(-ENOMEM
);
7175 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7177 struct task_group
*tg
= css_tg(css
);
7178 struct task_group
*parent
= css_tg(css_parent(css
));
7181 sched_online_group(tg
, parent
);
7185 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7187 struct task_group
*tg
= css_tg(css
);
7189 sched_destroy_group(tg
);
7192 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7194 struct task_group
*tg
= css_tg(css
);
7196 sched_offline_group(tg
);
7199 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7200 struct cgroup_taskset
*tset
)
7202 struct task_struct
*task
;
7204 cgroup_taskset_for_each(task
, css
, tset
) {
7205 #ifdef CONFIG_RT_GROUP_SCHED
7206 if (!sched_rt_can_attach(css_tg(css
), task
))
7209 /* We don't support RT-tasks being in separate groups */
7210 if (task
->sched_class
!= &fair_sched_class
)
7217 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7218 struct cgroup_taskset
*tset
)
7220 struct task_struct
*task
;
7222 cgroup_taskset_for_each(task
, css
, tset
)
7223 sched_move_task(task
);
7226 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7227 struct cgroup_subsys_state
*old_css
,
7228 struct task_struct
*task
)
7231 * cgroup_exit() is called in the copy_process() failure path.
7232 * Ignore this case since the task hasn't ran yet, this avoids
7233 * trying to poke a half freed task state from generic code.
7235 if (!(task
->flags
& PF_EXITING
))
7238 sched_move_task(task
);
7241 #ifdef CONFIG_FAIR_GROUP_SCHED
7242 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7243 struct cftype
*cftype
, u64 shareval
)
7245 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7248 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7251 struct task_group
*tg
= css_tg(css
);
7253 return (u64
) scale_load_down(tg
->shares
);
7256 #ifdef CONFIG_CFS_BANDWIDTH
7257 static DEFINE_MUTEX(cfs_constraints_mutex
);
7259 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7260 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7262 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7264 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7266 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7267 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7269 if (tg
== &root_task_group
)
7273 * Ensure we have at some amount of bandwidth every period. This is
7274 * to prevent reaching a state of large arrears when throttled via
7275 * entity_tick() resulting in prolonged exit starvation.
7277 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7281 * Likewise, bound things on the otherside by preventing insane quota
7282 * periods. This also allows us to normalize in computing quota
7285 if (period
> max_cfs_quota_period
)
7288 mutex_lock(&cfs_constraints_mutex
);
7289 ret
= __cfs_schedulable(tg
, period
, quota
);
7293 runtime_enabled
= quota
!= RUNTIME_INF
;
7294 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7295 account_cfs_bandwidth_used(runtime_enabled
, runtime_was_enabled
);
7296 raw_spin_lock_irq(&cfs_b
->lock
);
7297 cfs_b
->period
= ns_to_ktime(period
);
7298 cfs_b
->quota
= quota
;
7300 __refill_cfs_bandwidth_runtime(cfs_b
);
7301 /* restart the period timer (if active) to handle new period expiry */
7302 if (runtime_enabled
&& cfs_b
->timer_active
) {
7303 /* force a reprogram */
7304 cfs_b
->timer_active
= 0;
7305 __start_cfs_bandwidth(cfs_b
);
7307 raw_spin_unlock_irq(&cfs_b
->lock
);
7309 for_each_possible_cpu(i
) {
7310 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7311 struct rq
*rq
= cfs_rq
->rq
;
7313 raw_spin_lock_irq(&rq
->lock
);
7314 cfs_rq
->runtime_enabled
= runtime_enabled
;
7315 cfs_rq
->runtime_remaining
= 0;
7317 if (cfs_rq
->throttled
)
7318 unthrottle_cfs_rq(cfs_rq
);
7319 raw_spin_unlock_irq(&rq
->lock
);
7322 mutex_unlock(&cfs_constraints_mutex
);
7327 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7331 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7332 if (cfs_quota_us
< 0)
7333 quota
= RUNTIME_INF
;
7335 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7337 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7340 long tg_get_cfs_quota(struct task_group
*tg
)
7344 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7347 quota_us
= tg
->cfs_bandwidth
.quota
;
7348 do_div(quota_us
, NSEC_PER_USEC
);
7353 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7357 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7358 quota
= tg
->cfs_bandwidth
.quota
;
7360 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7363 long tg_get_cfs_period(struct task_group
*tg
)
7367 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7368 do_div(cfs_period_us
, NSEC_PER_USEC
);
7370 return cfs_period_us
;
7373 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7376 return tg_get_cfs_quota(css_tg(css
));
7379 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7380 struct cftype
*cftype
, s64 cfs_quota_us
)
7382 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7385 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7388 return tg_get_cfs_period(css_tg(css
));
7391 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7392 struct cftype
*cftype
, u64 cfs_period_us
)
7394 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7397 struct cfs_schedulable_data
{
7398 struct task_group
*tg
;
7403 * normalize group quota/period to be quota/max_period
7404 * note: units are usecs
7406 static u64
normalize_cfs_quota(struct task_group
*tg
,
7407 struct cfs_schedulable_data
*d
)
7415 period
= tg_get_cfs_period(tg
);
7416 quota
= tg_get_cfs_quota(tg
);
7419 /* note: these should typically be equivalent */
7420 if (quota
== RUNTIME_INF
|| quota
== -1)
7423 return to_ratio(period
, quota
);
7426 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7428 struct cfs_schedulable_data
*d
= data
;
7429 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7430 s64 quota
= 0, parent_quota
= -1;
7433 quota
= RUNTIME_INF
;
7435 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7437 quota
= normalize_cfs_quota(tg
, d
);
7438 parent_quota
= parent_b
->hierarchal_quota
;
7441 * ensure max(child_quota) <= parent_quota, inherit when no
7444 if (quota
== RUNTIME_INF
)
7445 quota
= parent_quota
;
7446 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7449 cfs_b
->hierarchal_quota
= quota
;
7454 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7457 struct cfs_schedulable_data data
= {
7463 if (quota
!= RUNTIME_INF
) {
7464 do_div(data
.period
, NSEC_PER_USEC
);
7465 do_div(data
.quota
, NSEC_PER_USEC
);
7469 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7475 static int cpu_stats_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
7476 struct cgroup_map_cb
*cb
)
7478 struct task_group
*tg
= css_tg(css
);
7479 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7481 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
7482 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
7483 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
7487 #endif /* CONFIG_CFS_BANDWIDTH */
7488 #endif /* CONFIG_FAIR_GROUP_SCHED */
7490 #ifdef CONFIG_RT_GROUP_SCHED
7491 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7492 struct cftype
*cft
, s64 val
)
7494 return sched_group_set_rt_runtime(css_tg(css
), val
);
7497 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7500 return sched_group_rt_runtime(css_tg(css
));
7503 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7504 struct cftype
*cftype
, u64 rt_period_us
)
7506 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7509 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7512 return sched_group_rt_period(css_tg(css
));
7514 #endif /* CONFIG_RT_GROUP_SCHED */
7516 static struct cftype cpu_files
[] = {
7517 #ifdef CONFIG_FAIR_GROUP_SCHED
7520 .read_u64
= cpu_shares_read_u64
,
7521 .write_u64
= cpu_shares_write_u64
,
7524 #ifdef CONFIG_CFS_BANDWIDTH
7526 .name
= "cfs_quota_us",
7527 .read_s64
= cpu_cfs_quota_read_s64
,
7528 .write_s64
= cpu_cfs_quota_write_s64
,
7531 .name
= "cfs_period_us",
7532 .read_u64
= cpu_cfs_period_read_u64
,
7533 .write_u64
= cpu_cfs_period_write_u64
,
7537 .read_map
= cpu_stats_show
,
7540 #ifdef CONFIG_RT_GROUP_SCHED
7542 .name
= "rt_runtime_us",
7543 .read_s64
= cpu_rt_runtime_read
,
7544 .write_s64
= cpu_rt_runtime_write
,
7547 .name
= "rt_period_us",
7548 .read_u64
= cpu_rt_period_read_uint
,
7549 .write_u64
= cpu_rt_period_write_uint
,
7555 struct cgroup_subsys cpu_cgroup_subsys
= {
7557 .css_alloc
= cpu_cgroup_css_alloc
,
7558 .css_free
= cpu_cgroup_css_free
,
7559 .css_online
= cpu_cgroup_css_online
,
7560 .css_offline
= cpu_cgroup_css_offline
,
7561 .can_attach
= cpu_cgroup_can_attach
,
7562 .attach
= cpu_cgroup_attach
,
7563 .exit
= cpu_cgroup_exit
,
7564 .subsys_id
= cpu_cgroup_subsys_id
,
7565 .base_cftypes
= cpu_files
,
7569 #endif /* CONFIG_CGROUP_SCHED */
7571 void dump_cpu_task(int cpu
)
7573 pr_info("Task dump for CPU %d:\n", cpu
);
7574 sched_show_task(cpu_curr(cpu
));