4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
96 ktime_t soft
, hard
, now
;
99 if (hrtimer_active(period_timer
))
102 now
= hrtimer_cb_get_time(period_timer
);
103 hrtimer_forward(period_timer
, now
, period
);
105 soft
= hrtimer_get_softexpires(period_timer
);
106 hard
= hrtimer_get_expires(period_timer
);
107 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
108 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
109 HRTIMER_MODE_ABS_PINNED
, 0);
113 DEFINE_MUTEX(sched_domains_mutex
);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
116 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
118 void update_rq_clock(struct rq
*rq
)
122 if (rq
->skip_clock_update
> 0)
125 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
127 update_rq_clock_task(rq
, delta
);
131 * Debugging: various feature bits
134 #define SCHED_FEAT(name, enabled) \
135 (1UL << __SCHED_FEAT_##name) * enabled |
137 const_debug
unsigned int sysctl_sched_features
=
138 #include "features.h"
143 #ifdef CONFIG_SCHED_DEBUG
144 #define SCHED_FEAT(name, enabled) \
147 static const char * const sched_feat_names
[] = {
148 #include "features.h"
153 static int sched_feat_show(struct seq_file
*m
, void *v
)
157 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
158 if (!(sysctl_sched_features
& (1UL << i
)))
160 seq_printf(m
, "%s ", sched_feat_names
[i
]);
167 #ifdef HAVE_JUMP_LABEL
169 #define jump_label_key__true STATIC_KEY_INIT_TRUE
170 #define jump_label_key__false STATIC_KEY_INIT_FALSE
172 #define SCHED_FEAT(name, enabled) \
173 jump_label_key__##enabled ,
175 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
176 #include "features.h"
181 static void sched_feat_disable(int i
)
183 if (static_key_enabled(&sched_feat_keys
[i
]))
184 static_key_slow_dec(&sched_feat_keys
[i
]);
187 static void sched_feat_enable(int i
)
189 if (!static_key_enabled(&sched_feat_keys
[i
]))
190 static_key_slow_inc(&sched_feat_keys
[i
]);
193 static void sched_feat_disable(int i
) { };
194 static void sched_feat_enable(int i
) { };
195 #endif /* HAVE_JUMP_LABEL */
197 static int sched_feat_set(char *cmp
)
202 if (strncmp(cmp
, "NO_", 3) == 0) {
207 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
208 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
210 sysctl_sched_features
&= ~(1UL << i
);
211 sched_feat_disable(i
);
213 sysctl_sched_features
|= (1UL << i
);
214 sched_feat_enable(i
);
224 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
225 size_t cnt
, loff_t
*ppos
)
234 if (copy_from_user(&buf
, ubuf
, cnt
))
240 i
= sched_feat_set(cmp
);
241 if (i
== __SCHED_FEAT_NR
)
249 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
251 return single_open(filp
, sched_feat_show
, NULL
);
254 static const struct file_operations sched_feat_fops
= {
255 .open
= sched_feat_open
,
256 .write
= sched_feat_write
,
259 .release
= single_release
,
262 static __init
int sched_init_debug(void)
264 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
269 late_initcall(sched_init_debug
);
270 #endif /* CONFIG_SCHED_DEBUG */
273 * Number of tasks to iterate in a single balance run.
274 * Limited because this is done with IRQs disabled.
276 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
279 * period over which we average the RT time consumption, measured
284 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
287 * period over which we measure -rt task cpu usage in us.
290 unsigned int sysctl_sched_rt_period
= 1000000;
292 __read_mostly
int scheduler_running
;
295 * part of the period that we allow rt tasks to run in us.
298 int sysctl_sched_rt_runtime
= 950000;
301 * __task_rq_lock - lock the rq @p resides on.
303 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
308 lockdep_assert_held(&p
->pi_lock
);
312 raw_spin_lock(&rq
->lock
);
313 if (likely(rq
== task_rq(p
)))
315 raw_spin_unlock(&rq
->lock
);
320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
322 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
323 __acquires(p
->pi_lock
)
329 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
331 raw_spin_lock(&rq
->lock
);
332 if (likely(rq
== task_rq(p
)))
334 raw_spin_unlock(&rq
->lock
);
335 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
339 static void __task_rq_unlock(struct rq
*rq
)
342 raw_spin_unlock(&rq
->lock
);
346 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
348 __releases(p
->pi_lock
)
350 raw_spin_unlock(&rq
->lock
);
351 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
355 * this_rq_lock - lock this runqueue and disable interrupts.
357 static struct rq
*this_rq_lock(void)
364 raw_spin_lock(&rq
->lock
);
369 #ifdef CONFIG_SCHED_HRTICK
371 * Use HR-timers to deliver accurate preemption points.
374 static void hrtick_clear(struct rq
*rq
)
376 if (hrtimer_active(&rq
->hrtick_timer
))
377 hrtimer_cancel(&rq
->hrtick_timer
);
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
384 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
386 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
388 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
390 raw_spin_lock(&rq
->lock
);
392 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
393 raw_spin_unlock(&rq
->lock
);
395 return HRTIMER_NORESTART
;
400 static int __hrtick_restart(struct rq
*rq
)
402 struct hrtimer
*timer
= &rq
->hrtick_timer
;
403 ktime_t time
= hrtimer_get_softexpires(timer
);
405 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
409 * called from hardirq (IPI) context
411 static void __hrtick_start(void *arg
)
415 raw_spin_lock(&rq
->lock
);
416 __hrtick_restart(rq
);
417 rq
->hrtick_csd_pending
= 0;
418 raw_spin_unlock(&rq
->lock
);
422 * Called to set the hrtick timer state.
424 * called with rq->lock held and irqs disabled
426 void hrtick_start(struct rq
*rq
, u64 delay
)
428 struct hrtimer
*timer
= &rq
->hrtick_timer
;
429 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
431 hrtimer_set_expires(timer
, time
);
433 if (rq
== this_rq()) {
434 __hrtick_restart(rq
);
435 } else if (!rq
->hrtick_csd_pending
) {
436 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
437 rq
->hrtick_csd_pending
= 1;
442 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
444 int cpu
= (int)(long)hcpu
;
447 case CPU_UP_CANCELED
:
448 case CPU_UP_CANCELED_FROZEN
:
449 case CPU_DOWN_PREPARE
:
450 case CPU_DOWN_PREPARE_FROZEN
:
452 case CPU_DEAD_FROZEN
:
453 hrtick_clear(cpu_rq(cpu
));
460 static __init
void init_hrtick(void)
462 hotcpu_notifier(hotplug_hrtick
, 0);
466 * Called to set the hrtick timer state.
468 * called with rq->lock held and irqs disabled
470 void hrtick_start(struct rq
*rq
, u64 delay
)
472 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
473 HRTIMER_MODE_REL_PINNED
, 0);
476 static inline void init_hrtick(void)
479 #endif /* CONFIG_SMP */
481 static void init_rq_hrtick(struct rq
*rq
)
484 rq
->hrtick_csd_pending
= 0;
486 rq
->hrtick_csd
.flags
= 0;
487 rq
->hrtick_csd
.func
= __hrtick_start
;
488 rq
->hrtick_csd
.info
= rq
;
491 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
492 rq
->hrtick_timer
.function
= hrtick
;
494 #else /* CONFIG_SCHED_HRTICK */
495 static inline void hrtick_clear(struct rq
*rq
)
499 static inline void init_rq_hrtick(struct rq
*rq
)
503 static inline void init_hrtick(void)
506 #endif /* CONFIG_SCHED_HRTICK */
509 * resched_task - mark a task 'to be rescheduled now'.
511 * On UP this means the setting of the need_resched flag, on SMP it
512 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct
*p
)
519 lockdep_assert_held(&task_rq(p
)->lock
);
521 if (test_tsk_need_resched(p
))
524 set_tsk_need_resched(p
);
527 if (cpu
== smp_processor_id()) {
528 set_preempt_need_resched();
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
);
550 #ifdef CONFIG_NO_HZ_COMMON
552 * In the semi idle case, use the nearest busy cpu for migrating timers
553 * from an idle cpu. This is good for power-savings.
555 * We don't do similar optimization for completely idle system, as
556 * selecting an idle cpu will add more delays to the timers than intended
557 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
559 int get_nohz_timer_target(int pinned
)
561 int cpu
= smp_processor_id();
563 struct sched_domain
*sd
;
565 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
569 for_each_domain(cpu
, sd
) {
570 for_each_cpu(i
, sched_domain_span(sd
)) {
582 * When add_timer_on() enqueues a timer into the timer wheel of an
583 * idle CPU then this timer might expire before the next timer event
584 * which is scheduled to wake up that CPU. In case of a completely
585 * idle system the next event might even be infinite time into the
586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587 * leaves the inner idle loop so the newly added timer is taken into
588 * account when the CPU goes back to idle and evaluates the timer
589 * wheel for the next timer event.
591 static void wake_up_idle_cpu(int cpu
)
593 struct rq
*rq
= cpu_rq(cpu
);
595 if (cpu
== smp_processor_id())
599 * This is safe, as this function is called with the timer
600 * wheel base lock of (cpu) held. When the CPU is on the way
601 * to idle and has not yet set rq->curr to idle then it will
602 * be serialized on the timer wheel base lock and take the new
603 * timer into account automatically.
605 if (rq
->curr
!= rq
->idle
)
609 * We can set TIF_RESCHED on the idle task of the other CPU
610 * lockless. The worst case is that the other CPU runs the
611 * idle task through an additional NOOP schedule()
613 set_tsk_need_resched(rq
->idle
);
615 /* NEED_RESCHED must be visible before we test polling */
617 if (!tsk_is_polling(rq
->idle
))
618 smp_send_reschedule(cpu
);
621 static bool wake_up_full_nohz_cpu(int cpu
)
623 if (tick_nohz_full_cpu(cpu
)) {
624 if (cpu
!= smp_processor_id() ||
625 tick_nohz_tick_stopped())
626 smp_send_reschedule(cpu
);
633 void wake_up_nohz_cpu(int cpu
)
635 if (!wake_up_full_nohz_cpu(cpu
))
636 wake_up_idle_cpu(cpu
);
639 static inline bool got_nohz_idle_kick(void)
641 int cpu
= smp_processor_id();
643 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
646 if (idle_cpu(cpu
) && !need_resched())
650 * We can't run Idle Load Balance on this CPU for this time so we
651 * cancel it and clear NOHZ_BALANCE_KICK
653 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
657 #else /* CONFIG_NO_HZ_COMMON */
659 static inline bool got_nohz_idle_kick(void)
664 #endif /* CONFIG_NO_HZ_COMMON */
666 #ifdef CONFIG_NO_HZ_FULL
667 bool sched_can_stop_tick(void)
673 /* Make sure rq->nr_running update is visible after the IPI */
676 /* More than one running task need preemption */
677 if (rq
->nr_running
> 1)
682 #endif /* CONFIG_NO_HZ_FULL */
684 void sched_avg_update(struct rq
*rq
)
686 s64 period
= sched_avg_period();
688 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
690 * Inline assembly required to prevent the compiler
691 * optimising this loop into a divmod call.
692 * See __iter_div_u64_rem() for another example of this.
694 asm("" : "+rm" (rq
->age_stamp
));
695 rq
->age_stamp
+= period
;
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group
*from
,
711 tg_visitor down
, tg_visitor up
, void *data
)
713 struct task_group
*parent
, *child
;
719 ret
= (*down
)(parent
, data
);
722 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
729 ret
= (*up
)(parent
, data
);
730 if (ret
|| parent
== from
)
734 parent
= parent
->parent
;
741 int tg_nop(struct task_group
*tg
, void *data
)
747 static void set_load_weight(struct task_struct
*p
)
749 int prio
= p
->static_prio
- MAX_RT_PRIO
;
750 struct load_weight
*load
= &p
->se
.load
;
753 * SCHED_IDLE tasks get minimal weight:
755 if (p
->policy
== SCHED_IDLE
) {
756 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
757 load
->inv_weight
= WMULT_IDLEPRIO
;
761 load
->weight
= scale_load(prio_to_weight
[prio
]);
762 load
->inv_weight
= prio_to_wmult
[prio
];
765 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
768 sched_info_queued(rq
, p
);
769 p
->sched_class
->enqueue_task(rq
, p
, flags
);
772 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
775 sched_info_dequeued(rq
, p
);
776 p
->sched_class
->dequeue_task(rq
, p
, flags
);
779 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
781 if (task_contributes_to_load(p
))
782 rq
->nr_uninterruptible
--;
784 enqueue_task(rq
, p
, flags
);
787 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
789 if (task_contributes_to_load(p
))
790 rq
->nr_uninterruptible
++;
792 dequeue_task(rq
, p
, flags
);
795 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal
= 0, irq_delta
= 0;
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
822 if (irq_delta
> delta
)
825 rq
->prev_irq_time
+= irq_delta
;
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled
))) {
830 steal
= paravirt_steal_clock(cpu_of(rq
));
831 steal
-= rq
->prev_steal_time_rq
;
833 if (unlikely(steal
> delta
))
836 rq
->prev_steal_time_rq
+= steal
;
841 rq
->clock_task
+= delta
;
843 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
844 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
845 sched_rt_avg_update(rq
, irq_delta
+ steal
);
849 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
851 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
852 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
856 * Make it appear like a SCHED_FIFO task, its something
857 * userspace knows about and won't get confused about.
859 * Also, it will make PI more or less work without too
860 * much confusion -- but then, stop work should not
861 * rely on PI working anyway.
863 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
865 stop
->sched_class
= &stop_sched_class
;
868 cpu_rq(cpu
)->stop
= stop
;
872 * Reset it back to a normal scheduling class so that
873 * it can die in pieces.
875 old_stop
->sched_class
= &rt_sched_class
;
880 * __normal_prio - return the priority that is based on the static prio
882 static inline int __normal_prio(struct task_struct
*p
)
884 return p
->static_prio
;
888 * Calculate the expected normal priority: i.e. priority
889 * without taking RT-inheritance into account. Might be
890 * boosted by interactivity modifiers. Changes upon fork,
891 * setprio syscalls, and whenever the interactivity
892 * estimator recalculates.
894 static inline int normal_prio(struct task_struct
*p
)
898 if (task_has_dl_policy(p
))
899 prio
= MAX_DL_PRIO
-1;
900 else if (task_has_rt_policy(p
))
901 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
903 prio
= __normal_prio(p
);
908 * Calculate the current priority, i.e. the priority
909 * taken into account by the scheduler. This value might
910 * be boosted by RT tasks, or might be boosted by
911 * interactivity modifiers. Will be RT if the task got
912 * RT-boosted. If not then it returns p->normal_prio.
914 static int effective_prio(struct task_struct
*p
)
916 p
->normal_prio
= normal_prio(p
);
918 * If we are RT tasks or we were boosted to RT priority,
919 * keep the priority unchanged. Otherwise, update priority
920 * to the normal priority:
922 if (!rt_prio(p
->prio
))
923 return p
->normal_prio
;
928 * task_curr - is this task currently executing on a CPU?
929 * @p: the task in question.
931 * Return: 1 if the task is currently executing. 0 otherwise.
933 inline int task_curr(const struct task_struct
*p
)
935 return cpu_curr(task_cpu(p
)) == p
;
938 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
939 const struct sched_class
*prev_class
,
942 if (prev_class
!= p
->sched_class
) {
943 if (prev_class
->switched_from
)
944 prev_class
->switched_from(rq
, p
);
945 p
->sched_class
->switched_to(rq
, p
);
946 } else if (oldprio
!= p
->prio
|| dl_task(p
))
947 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
950 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
952 const struct sched_class
*class;
954 if (p
->sched_class
== rq
->curr
->sched_class
) {
955 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
957 for_each_class(class) {
958 if (class == rq
->curr
->sched_class
)
960 if (class == p
->sched_class
) {
961 resched_task(rq
->curr
);
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
972 rq
->skip_clock_update
= 1;
976 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
978 #ifdef CONFIG_SCHED_DEBUG
980 * We should never call set_task_cpu() on a blocked task,
981 * ttwu() will sort out the placement.
983 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
984 !(task_preempt_count(p
) & PREEMPT_ACTIVE
));
986 #ifdef CONFIG_LOCKDEP
988 * The caller should hold either p->pi_lock or rq->lock, when changing
989 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
991 * sched_move_task() holds both and thus holding either pins the cgroup,
994 * Furthermore, all task_rq users should acquire both locks, see
997 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
998 lockdep_is_held(&task_rq(p
)->lock
)));
1002 trace_sched_migrate_task(p
, new_cpu
);
1004 if (task_cpu(p
) != new_cpu
) {
1005 if (p
->sched_class
->migrate_task_rq
)
1006 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1007 p
->se
.nr_migrations
++;
1008 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1011 __set_task_cpu(p
, new_cpu
);
1014 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1017 struct rq
*src_rq
, *dst_rq
;
1019 src_rq
= task_rq(p
);
1020 dst_rq
= cpu_rq(cpu
);
1022 deactivate_task(src_rq
, p
, 0);
1023 set_task_cpu(p
, cpu
);
1024 activate_task(dst_rq
, p
, 0);
1025 check_preempt_curr(dst_rq
, p
, 0);
1028 * Task isn't running anymore; make it appear like we migrated
1029 * it before it went to sleep. This means on wakeup we make the
1030 * previous cpu our targer instead of where it really is.
1036 struct migration_swap_arg
{
1037 struct task_struct
*src_task
, *dst_task
;
1038 int src_cpu
, dst_cpu
;
1041 static int migrate_swap_stop(void *data
)
1043 struct migration_swap_arg
*arg
= data
;
1044 struct rq
*src_rq
, *dst_rq
;
1047 src_rq
= cpu_rq(arg
->src_cpu
);
1048 dst_rq
= cpu_rq(arg
->dst_cpu
);
1050 double_raw_lock(&arg
->src_task
->pi_lock
,
1051 &arg
->dst_task
->pi_lock
);
1052 double_rq_lock(src_rq
, dst_rq
);
1053 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1056 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1059 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1062 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1065 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1066 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1071 double_rq_unlock(src_rq
, dst_rq
);
1072 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1073 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1079 * Cross migrate two tasks
1081 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1083 struct migration_swap_arg arg
;
1086 arg
= (struct migration_swap_arg
){
1088 .src_cpu
= task_cpu(cur
),
1090 .dst_cpu
= task_cpu(p
),
1093 if (arg
.src_cpu
== arg
.dst_cpu
)
1097 * These three tests are all lockless; this is OK since all of them
1098 * will be re-checked with proper locks held further down the line.
1100 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1103 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1106 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1109 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1110 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1116 struct migration_arg
{
1117 struct task_struct
*task
;
1121 static int migration_cpu_stop(void *data
);
1124 * wait_task_inactive - wait for a thread to unschedule.
1126 * If @match_state is nonzero, it's the @p->state value just checked and
1127 * not expected to change. If it changes, i.e. @p might have woken up,
1128 * then return zero. When we succeed in waiting for @p to be off its CPU,
1129 * we return a positive number (its total switch count). If a second call
1130 * a short while later returns the same number, the caller can be sure that
1131 * @p has remained unscheduled the whole time.
1133 * The caller must ensure that the task *will* unschedule sometime soon,
1134 * else this function might spin for a *long* time. This function can't
1135 * be called with interrupts off, or it may introduce deadlock with
1136 * smp_call_function() if an IPI is sent by the same process we are
1137 * waiting to become inactive.
1139 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1141 unsigned long flags
;
1148 * We do the initial early heuristics without holding
1149 * any task-queue locks at all. We'll only try to get
1150 * the runqueue lock when things look like they will
1156 * If the task is actively running on another CPU
1157 * still, just relax and busy-wait without holding
1160 * NOTE! Since we don't hold any locks, it's not
1161 * even sure that "rq" stays as the right runqueue!
1162 * But we don't care, since "task_running()" will
1163 * return false if the runqueue has changed and p
1164 * is actually now running somewhere else!
1166 while (task_running(rq
, p
)) {
1167 if (match_state
&& unlikely(p
->state
!= match_state
))
1173 * Ok, time to look more closely! We need the rq
1174 * lock now, to be *sure*. If we're wrong, we'll
1175 * just go back and repeat.
1177 rq
= task_rq_lock(p
, &flags
);
1178 trace_sched_wait_task(p
);
1179 running
= task_running(rq
, p
);
1182 if (!match_state
|| p
->state
== match_state
)
1183 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1184 task_rq_unlock(rq
, p
, &flags
);
1187 * If it changed from the expected state, bail out now.
1189 if (unlikely(!ncsw
))
1193 * Was it really running after all now that we
1194 * checked with the proper locks actually held?
1196 * Oops. Go back and try again..
1198 if (unlikely(running
)) {
1204 * It's not enough that it's not actively running,
1205 * it must be off the runqueue _entirely_, and not
1208 * So if it was still runnable (but just not actively
1209 * running right now), it's preempted, and we should
1210 * yield - it could be a while.
1212 if (unlikely(on_rq
)) {
1213 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1215 set_current_state(TASK_UNINTERRUPTIBLE
);
1216 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1221 * Ahh, all good. It wasn't running, and it wasn't
1222 * runnable, which means that it will never become
1223 * running in the future either. We're all done!
1232 * kick_process - kick a running thread to enter/exit the kernel
1233 * @p: the to-be-kicked thread
1235 * Cause a process which is running on another CPU to enter
1236 * kernel-mode, without any delay. (to get signals handled.)
1238 * NOTE: this function doesn't have to take the runqueue lock,
1239 * because all it wants to ensure is that the remote task enters
1240 * the kernel. If the IPI races and the task has been migrated
1241 * to another CPU then no harm is done and the purpose has been
1244 void kick_process(struct task_struct
*p
)
1250 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1251 smp_send_reschedule(cpu
);
1254 EXPORT_SYMBOL_GPL(kick_process
);
1255 #endif /* CONFIG_SMP */
1259 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1261 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1263 int nid
= cpu_to_node(cpu
);
1264 const struct cpumask
*nodemask
= NULL
;
1265 enum { cpuset
, possible
, fail
} state
= cpuset
;
1269 * If the node that the cpu is on has been offlined, cpu_to_node()
1270 * will return -1. There is no cpu on the node, and we should
1271 * select the cpu on the other node.
1274 nodemask
= cpumask_of_node(nid
);
1276 /* Look for allowed, online CPU in same node. */
1277 for_each_cpu(dest_cpu
, nodemask
) {
1278 if (!cpu_online(dest_cpu
))
1280 if (!cpu_active(dest_cpu
))
1282 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1288 /* Any allowed, online CPU? */
1289 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1290 if (!cpu_online(dest_cpu
))
1292 if (!cpu_active(dest_cpu
))
1299 /* No more Mr. Nice Guy. */
1300 cpuset_cpus_allowed_fallback(p
);
1305 do_set_cpus_allowed(p
, cpu_possible_mask
);
1316 if (state
!= cpuset
) {
1318 * Don't tell them about moving exiting tasks or
1319 * kernel threads (both mm NULL), since they never
1322 if (p
->mm
&& printk_ratelimit()) {
1323 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1324 task_pid_nr(p
), p
->comm
, cpu
);
1332 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1335 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1337 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1340 * In order not to call set_task_cpu() on a blocking task we need
1341 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1344 * Since this is common to all placement strategies, this lives here.
1346 * [ this allows ->select_task() to simply return task_cpu(p) and
1347 * not worry about this generic constraint ]
1349 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1351 cpu
= select_fallback_rq(task_cpu(p
), p
);
1356 static void update_avg(u64
*avg
, u64 sample
)
1358 s64 diff
= sample
- *avg
;
1364 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1366 #ifdef CONFIG_SCHEDSTATS
1367 struct rq
*rq
= this_rq();
1370 int this_cpu
= smp_processor_id();
1372 if (cpu
== this_cpu
) {
1373 schedstat_inc(rq
, ttwu_local
);
1374 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1376 struct sched_domain
*sd
;
1378 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1380 for_each_domain(this_cpu
, sd
) {
1381 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1382 schedstat_inc(sd
, ttwu_wake_remote
);
1389 if (wake_flags
& WF_MIGRATED
)
1390 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1392 #endif /* CONFIG_SMP */
1394 schedstat_inc(rq
, ttwu_count
);
1395 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1397 if (wake_flags
& WF_SYNC
)
1398 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1400 #endif /* CONFIG_SCHEDSTATS */
1403 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1405 activate_task(rq
, p
, en_flags
);
1408 /* if a worker is waking up, notify workqueue */
1409 if (p
->flags
& PF_WQ_WORKER
)
1410 wq_worker_waking_up(p
, cpu_of(rq
));
1414 * Mark the task runnable and perform wakeup-preemption.
1417 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1419 check_preempt_curr(rq
, p
, wake_flags
);
1420 trace_sched_wakeup(p
, true);
1422 p
->state
= TASK_RUNNING
;
1424 if (p
->sched_class
->task_woken
)
1425 p
->sched_class
->task_woken(rq
, p
);
1427 if (rq
->idle_stamp
) {
1428 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1429 u64 max
= 2*rq
->max_idle_balance_cost
;
1431 update_avg(&rq
->avg_idle
, delta
);
1433 if (rq
->avg_idle
> max
)
1442 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1445 if (p
->sched_contributes_to_load
)
1446 rq
->nr_uninterruptible
--;
1449 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1450 ttwu_do_wakeup(rq
, p
, wake_flags
);
1454 * Called in case the task @p isn't fully descheduled from its runqueue,
1455 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1456 * since all we need to do is flip p->state to TASK_RUNNING, since
1457 * the task is still ->on_rq.
1459 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1464 rq
= __task_rq_lock(p
);
1466 /* check_preempt_curr() may use rq clock */
1467 update_rq_clock(rq
);
1468 ttwu_do_wakeup(rq
, p
, wake_flags
);
1471 __task_rq_unlock(rq
);
1477 static void sched_ttwu_pending(void)
1479 struct rq
*rq
= this_rq();
1480 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1481 struct task_struct
*p
;
1483 raw_spin_lock(&rq
->lock
);
1486 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1487 llist
= llist_next(llist
);
1488 ttwu_do_activate(rq
, p
, 0);
1491 raw_spin_unlock(&rq
->lock
);
1494 void scheduler_ipi(void)
1497 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1498 * TIF_NEED_RESCHED remotely (for the first time) will also send
1501 preempt_fold_need_resched();
1503 if (llist_empty(&this_rq()->wake_list
)
1504 && !tick_nohz_full_cpu(smp_processor_id())
1505 && !got_nohz_idle_kick())
1509 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1510 * traditionally all their work was done from the interrupt return
1511 * path. Now that we actually do some work, we need to make sure
1514 * Some archs already do call them, luckily irq_enter/exit nest
1517 * Arguably we should visit all archs and update all handlers,
1518 * however a fair share of IPIs are still resched only so this would
1519 * somewhat pessimize the simple resched case.
1522 tick_nohz_full_check();
1523 sched_ttwu_pending();
1526 * Check if someone kicked us for doing the nohz idle load balance.
1528 if (unlikely(got_nohz_idle_kick())) {
1529 this_rq()->idle_balance
= 1;
1530 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1535 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1537 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1538 smp_send_reschedule(cpu
);
1541 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1543 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1545 #endif /* CONFIG_SMP */
1547 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1549 struct rq
*rq
= cpu_rq(cpu
);
1551 #if defined(CONFIG_SMP)
1552 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1553 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1554 ttwu_queue_remote(p
, cpu
);
1559 raw_spin_lock(&rq
->lock
);
1560 ttwu_do_activate(rq
, p
, 0);
1561 raw_spin_unlock(&rq
->lock
);
1565 * try_to_wake_up - wake up a thread
1566 * @p: the thread to be awakened
1567 * @state: the mask of task states that can be woken
1568 * @wake_flags: wake modifier flags (WF_*)
1570 * Put it on the run-queue if it's not already there. The "current"
1571 * thread is always on the run-queue (except when the actual
1572 * re-schedule is in progress), and as such you're allowed to do
1573 * the simpler "current->state = TASK_RUNNING" to mark yourself
1574 * runnable without the overhead of this.
1576 * Return: %true if @p was woken up, %false if it was already running.
1577 * or @state didn't match @p's state.
1580 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1582 unsigned long flags
;
1583 int cpu
, success
= 0;
1586 * If we are going to wake up a thread waiting for CONDITION we
1587 * need to ensure that CONDITION=1 done by the caller can not be
1588 * reordered with p->state check below. This pairs with mb() in
1589 * set_current_state() the waiting thread does.
1591 smp_mb__before_spinlock();
1592 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1593 if (!(p
->state
& state
))
1596 success
= 1; /* we're going to change ->state */
1599 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1604 * If the owning (remote) cpu is still in the middle of schedule() with
1605 * this task as prev, wait until its done referencing the task.
1610 * Pairs with the smp_wmb() in finish_lock_switch().
1614 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1615 p
->state
= TASK_WAKING
;
1617 if (p
->sched_class
->task_waking
)
1618 p
->sched_class
->task_waking(p
);
1620 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1621 if (task_cpu(p
) != cpu
) {
1622 wake_flags
|= WF_MIGRATED
;
1623 set_task_cpu(p
, cpu
);
1625 #endif /* CONFIG_SMP */
1629 ttwu_stat(p
, cpu
, wake_flags
);
1631 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1637 * try_to_wake_up_local - try to wake up a local task with rq lock held
1638 * @p: the thread to be awakened
1640 * Put @p on the run-queue if it's not already there. The caller must
1641 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1644 static void try_to_wake_up_local(struct task_struct
*p
)
1646 struct rq
*rq
= task_rq(p
);
1648 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1649 WARN_ON_ONCE(p
== current
))
1652 lockdep_assert_held(&rq
->lock
);
1654 if (!raw_spin_trylock(&p
->pi_lock
)) {
1655 raw_spin_unlock(&rq
->lock
);
1656 raw_spin_lock(&p
->pi_lock
);
1657 raw_spin_lock(&rq
->lock
);
1660 if (!(p
->state
& TASK_NORMAL
))
1664 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1666 ttwu_do_wakeup(rq
, p
, 0);
1667 ttwu_stat(p
, smp_processor_id(), 0);
1669 raw_spin_unlock(&p
->pi_lock
);
1673 * wake_up_process - Wake up a specific process
1674 * @p: The process to be woken up.
1676 * Attempt to wake up the nominated process and move it to the set of runnable
1679 * Return: 1 if the process was woken up, 0 if it was already running.
1681 * It may be assumed that this function implies a write memory barrier before
1682 * changing the task state if and only if any tasks are woken up.
1684 int wake_up_process(struct task_struct
*p
)
1686 WARN_ON(task_is_stopped_or_traced(p
));
1687 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1689 EXPORT_SYMBOL(wake_up_process
);
1691 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1693 return try_to_wake_up(p
, state
, 0);
1697 * Perform scheduler related setup for a newly forked process p.
1698 * p is forked by current.
1700 * __sched_fork() is basic setup used by init_idle() too:
1702 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1707 p
->se
.exec_start
= 0;
1708 p
->se
.sum_exec_runtime
= 0;
1709 p
->se
.prev_sum_exec_runtime
= 0;
1710 p
->se
.nr_migrations
= 0;
1712 INIT_LIST_HEAD(&p
->se
.group_node
);
1714 #ifdef CONFIG_SCHEDSTATS
1715 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1718 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1719 hrtimer_init(&p
->dl
.dl_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1720 p
->dl
.dl_runtime
= p
->dl
.runtime
= 0;
1721 p
->dl
.dl_deadline
= p
->dl
.deadline
= 0;
1722 p
->dl
.dl_period
= 0;
1725 INIT_LIST_HEAD(&p
->rt
.run_list
);
1727 #ifdef CONFIG_PREEMPT_NOTIFIERS
1728 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1731 #ifdef CONFIG_NUMA_BALANCING
1732 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1733 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1734 p
->mm
->numa_scan_seq
= 0;
1737 if (clone_flags
& CLONE_VM
)
1738 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1740 p
->numa_preferred_nid
= -1;
1742 p
->node_stamp
= 0ULL;
1743 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1744 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1745 p
->numa_work
.next
= &p
->numa_work
;
1746 p
->numa_faults_memory
= NULL
;
1747 p
->numa_faults_buffer_memory
= NULL
;
1748 p
->last_task_numa_placement
= 0;
1749 p
->last_sum_exec_runtime
= 0;
1751 INIT_LIST_HEAD(&p
->numa_entry
);
1752 p
->numa_group
= NULL
;
1753 #endif /* CONFIG_NUMA_BALANCING */
1756 #ifdef CONFIG_NUMA_BALANCING
1757 #ifdef CONFIG_SCHED_DEBUG
1758 void set_numabalancing_state(bool enabled
)
1761 sched_feat_set("NUMA");
1763 sched_feat_set("NO_NUMA");
1766 __read_mostly
bool numabalancing_enabled
;
1768 void set_numabalancing_state(bool enabled
)
1770 numabalancing_enabled
= enabled
;
1772 #endif /* CONFIG_SCHED_DEBUG */
1774 #ifdef CONFIG_PROC_SYSCTL
1775 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1776 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1780 int state
= numabalancing_enabled
;
1782 if (write
&& !capable(CAP_SYS_ADMIN
))
1787 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1791 set_numabalancing_state(state
);
1798 * fork()/clone()-time setup:
1800 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1802 unsigned long flags
;
1803 int cpu
= get_cpu();
1805 __sched_fork(clone_flags
, p
);
1807 * We mark the process as running here. This guarantees that
1808 * nobody will actually run it, and a signal or other external
1809 * event cannot wake it up and insert it on the runqueue either.
1811 p
->state
= TASK_RUNNING
;
1814 * Make sure we do not leak PI boosting priority to the child.
1816 p
->prio
= current
->normal_prio
;
1819 * Revert to default priority/policy on fork if requested.
1821 if (unlikely(p
->sched_reset_on_fork
)) {
1822 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1823 p
->policy
= SCHED_NORMAL
;
1824 p
->static_prio
= NICE_TO_PRIO(0);
1826 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1827 p
->static_prio
= NICE_TO_PRIO(0);
1829 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1833 * We don't need the reset flag anymore after the fork. It has
1834 * fulfilled its duty:
1836 p
->sched_reset_on_fork
= 0;
1839 if (dl_prio(p
->prio
)) {
1842 } else if (rt_prio(p
->prio
)) {
1843 p
->sched_class
= &rt_sched_class
;
1845 p
->sched_class
= &fair_sched_class
;
1848 if (p
->sched_class
->task_fork
)
1849 p
->sched_class
->task_fork(p
);
1852 * The child is not yet in the pid-hash so no cgroup attach races,
1853 * and the cgroup is pinned to this child due to cgroup_fork()
1854 * is ran before sched_fork().
1856 * Silence PROVE_RCU.
1858 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1859 set_task_cpu(p
, cpu
);
1860 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1862 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1863 if (likely(sched_info_on()))
1864 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1866 #if defined(CONFIG_SMP)
1869 init_task_preempt_count(p
);
1871 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1872 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1879 unsigned long to_ratio(u64 period
, u64 runtime
)
1881 if (runtime
== RUNTIME_INF
)
1885 * Doing this here saves a lot of checks in all
1886 * the calling paths, and returning zero seems
1887 * safe for them anyway.
1892 return div64_u64(runtime
<< 20, period
);
1896 inline struct dl_bw
*dl_bw_of(int i
)
1898 return &cpu_rq(i
)->rd
->dl_bw
;
1901 static inline int dl_bw_cpus(int i
)
1903 struct root_domain
*rd
= cpu_rq(i
)->rd
;
1906 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
1912 inline struct dl_bw
*dl_bw_of(int i
)
1914 return &cpu_rq(i
)->dl
.dl_bw
;
1917 static inline int dl_bw_cpus(int i
)
1924 void __dl_clear(struct dl_bw
*dl_b
, u64 tsk_bw
)
1926 dl_b
->total_bw
-= tsk_bw
;
1930 void __dl_add(struct dl_bw
*dl_b
, u64 tsk_bw
)
1932 dl_b
->total_bw
+= tsk_bw
;
1936 bool __dl_overflow(struct dl_bw
*dl_b
, int cpus
, u64 old_bw
, u64 new_bw
)
1938 return dl_b
->bw
!= -1 &&
1939 dl_b
->bw
* cpus
< dl_b
->total_bw
- old_bw
+ new_bw
;
1943 * We must be sure that accepting a new task (or allowing changing the
1944 * parameters of an existing one) is consistent with the bandwidth
1945 * constraints. If yes, this function also accordingly updates the currently
1946 * allocated bandwidth to reflect the new situation.
1948 * This function is called while holding p's rq->lock.
1950 static int dl_overflow(struct task_struct
*p
, int policy
,
1951 const struct sched_attr
*attr
)
1954 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1955 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
1956 u64 runtime
= attr
->sched_runtime
;
1957 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
1960 if (new_bw
== p
->dl
.dl_bw
)
1964 * Either if a task, enters, leave, or stays -deadline but changes
1965 * its parameters, we may need to update accordingly the total
1966 * allocated bandwidth of the container.
1968 raw_spin_lock(&dl_b
->lock
);
1969 cpus
= dl_bw_cpus(task_cpu(p
));
1970 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
1971 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
1972 __dl_add(dl_b
, new_bw
);
1974 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
1975 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
1976 __dl_clear(dl_b
, p
->dl
.dl_bw
);
1977 __dl_add(dl_b
, new_bw
);
1979 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
1980 __dl_clear(dl_b
, p
->dl
.dl_bw
);
1983 raw_spin_unlock(&dl_b
->lock
);
1988 extern void init_dl_bw(struct dl_bw
*dl_b
);
1991 * wake_up_new_task - wake up a newly created task for the first time.
1993 * This function will do some initial scheduler statistics housekeeping
1994 * that must be done for every newly created context, then puts the task
1995 * on the runqueue and wakes it.
1997 void wake_up_new_task(struct task_struct
*p
)
1999 unsigned long flags
;
2002 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2005 * Fork balancing, do it here and not earlier because:
2006 * - cpus_allowed can change in the fork path
2007 * - any previously selected cpu might disappear through hotplug
2009 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2012 /* Initialize new task's runnable average */
2013 init_task_runnable_average(p
);
2014 rq
= __task_rq_lock(p
);
2015 activate_task(rq
, p
, 0);
2017 trace_sched_wakeup_new(p
, true);
2018 check_preempt_curr(rq
, p
, WF_FORK
);
2020 if (p
->sched_class
->task_woken
)
2021 p
->sched_class
->task_woken(rq
, p
);
2023 task_rq_unlock(rq
, p
, &flags
);
2026 #ifdef CONFIG_PREEMPT_NOTIFIERS
2029 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2030 * @notifier: notifier struct to register
2032 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2034 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2036 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2039 * preempt_notifier_unregister - no longer interested in preemption notifications
2040 * @notifier: notifier struct to unregister
2042 * This is safe to call from within a preemption notifier.
2044 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2046 hlist_del(¬ifier
->link
);
2048 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2050 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2052 struct preempt_notifier
*notifier
;
2054 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2055 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2059 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2060 struct task_struct
*next
)
2062 struct preempt_notifier
*notifier
;
2064 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2065 notifier
->ops
->sched_out(notifier
, next
);
2068 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2070 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2075 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2076 struct task_struct
*next
)
2080 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2083 * prepare_task_switch - prepare to switch tasks
2084 * @rq: the runqueue preparing to switch
2085 * @prev: the current task that is being switched out
2086 * @next: the task we are going to switch to.
2088 * This is called with the rq lock held and interrupts off. It must
2089 * be paired with a subsequent finish_task_switch after the context
2092 * prepare_task_switch sets up locking and calls architecture specific
2096 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2097 struct task_struct
*next
)
2099 trace_sched_switch(prev
, next
);
2100 sched_info_switch(rq
, prev
, next
);
2101 perf_event_task_sched_out(prev
, next
);
2102 fire_sched_out_preempt_notifiers(prev
, next
);
2103 prepare_lock_switch(rq
, next
);
2104 prepare_arch_switch(next
);
2108 * finish_task_switch - clean up after a task-switch
2109 * @rq: runqueue associated with task-switch
2110 * @prev: the thread we just switched away from.
2112 * finish_task_switch must be called after the context switch, paired
2113 * with a prepare_task_switch call before the context switch.
2114 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2115 * and do any other architecture-specific cleanup actions.
2117 * Note that we may have delayed dropping an mm in context_switch(). If
2118 * so, we finish that here outside of the runqueue lock. (Doing it
2119 * with the lock held can cause deadlocks; see schedule() for
2122 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2123 __releases(rq
->lock
)
2125 struct mm_struct
*mm
= rq
->prev_mm
;
2131 * A task struct has one reference for the use as "current".
2132 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2133 * schedule one last time. The schedule call will never return, and
2134 * the scheduled task must drop that reference.
2135 * The test for TASK_DEAD must occur while the runqueue locks are
2136 * still held, otherwise prev could be scheduled on another cpu, die
2137 * there before we look at prev->state, and then the reference would
2139 * Manfred Spraul <manfred@colorfullife.com>
2141 prev_state
= prev
->state
;
2142 vtime_task_switch(prev
);
2143 finish_arch_switch(prev
);
2144 perf_event_task_sched_in(prev
, current
);
2145 finish_lock_switch(rq
, prev
);
2146 finish_arch_post_lock_switch();
2148 fire_sched_in_preempt_notifiers(current
);
2151 if (unlikely(prev_state
== TASK_DEAD
)) {
2152 if (prev
->sched_class
->task_dead
)
2153 prev
->sched_class
->task_dead(prev
);
2156 * Remove function-return probe instances associated with this
2157 * task and put them back on the free list.
2159 kprobe_flush_task(prev
);
2160 put_task_struct(prev
);
2163 tick_nohz_task_switch(current
);
2168 /* rq->lock is NOT held, but preemption is disabled */
2169 static inline void post_schedule(struct rq
*rq
)
2171 if (rq
->post_schedule
) {
2172 unsigned long flags
;
2174 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2175 if (rq
->curr
->sched_class
->post_schedule
)
2176 rq
->curr
->sched_class
->post_schedule(rq
);
2177 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2179 rq
->post_schedule
= 0;
2185 static inline void post_schedule(struct rq
*rq
)
2192 * schedule_tail - first thing a freshly forked thread must call.
2193 * @prev: the thread we just switched away from.
2195 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2196 __releases(rq
->lock
)
2198 struct rq
*rq
= this_rq();
2200 finish_task_switch(rq
, prev
);
2203 * FIXME: do we need to worry about rq being invalidated by the
2208 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2209 /* In this case, finish_task_switch does not reenable preemption */
2212 if (current
->set_child_tid
)
2213 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2217 * context_switch - switch to the new MM and the new
2218 * thread's register state.
2221 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2222 struct task_struct
*next
)
2224 struct mm_struct
*mm
, *oldmm
;
2226 prepare_task_switch(rq
, prev
, next
);
2229 oldmm
= prev
->active_mm
;
2231 * For paravirt, this is coupled with an exit in switch_to to
2232 * combine the page table reload and the switch backend into
2235 arch_start_context_switch(prev
);
2238 next
->active_mm
= oldmm
;
2239 atomic_inc(&oldmm
->mm_count
);
2240 enter_lazy_tlb(oldmm
, next
);
2242 switch_mm(oldmm
, mm
, next
);
2245 prev
->active_mm
= NULL
;
2246 rq
->prev_mm
= oldmm
;
2249 * Since the runqueue lock will be released by the next
2250 * task (which is an invalid locking op but in the case
2251 * of the scheduler it's an obvious special-case), so we
2252 * do an early lockdep release here:
2254 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2255 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2258 context_tracking_task_switch(prev
, next
);
2259 /* Here we just switch the register state and the stack. */
2260 switch_to(prev
, next
, prev
);
2264 * this_rq must be evaluated again because prev may have moved
2265 * CPUs since it called schedule(), thus the 'rq' on its stack
2266 * frame will be invalid.
2268 finish_task_switch(this_rq(), prev
);
2272 * nr_running and nr_context_switches:
2274 * externally visible scheduler statistics: current number of runnable
2275 * threads, total number of context switches performed since bootup.
2277 unsigned long nr_running(void)
2279 unsigned long i
, sum
= 0;
2281 for_each_online_cpu(i
)
2282 sum
+= cpu_rq(i
)->nr_running
;
2287 unsigned long long nr_context_switches(void)
2290 unsigned long long sum
= 0;
2292 for_each_possible_cpu(i
)
2293 sum
+= cpu_rq(i
)->nr_switches
;
2298 unsigned long nr_iowait(void)
2300 unsigned long i
, sum
= 0;
2302 for_each_possible_cpu(i
)
2303 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2308 unsigned long nr_iowait_cpu(int cpu
)
2310 struct rq
*this = cpu_rq(cpu
);
2311 return atomic_read(&this->nr_iowait
);
2317 * sched_exec - execve() is a valuable balancing opportunity, because at
2318 * this point the task has the smallest effective memory and cache footprint.
2320 void sched_exec(void)
2322 struct task_struct
*p
= current
;
2323 unsigned long flags
;
2326 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2327 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2328 if (dest_cpu
== smp_processor_id())
2331 if (likely(cpu_active(dest_cpu
))) {
2332 struct migration_arg arg
= { p
, dest_cpu
};
2334 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2335 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2339 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2344 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2345 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2347 EXPORT_PER_CPU_SYMBOL(kstat
);
2348 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2351 * Return any ns on the sched_clock that have not yet been accounted in
2352 * @p in case that task is currently running.
2354 * Called with task_rq_lock() held on @rq.
2356 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2360 if (task_current(rq
, p
)) {
2361 update_rq_clock(rq
);
2362 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2370 unsigned long long task_delta_exec(struct task_struct
*p
)
2372 unsigned long flags
;
2376 rq
= task_rq_lock(p
, &flags
);
2377 ns
= do_task_delta_exec(p
, rq
);
2378 task_rq_unlock(rq
, p
, &flags
);
2384 * Return accounted runtime for the task.
2385 * In case the task is currently running, return the runtime plus current's
2386 * pending runtime that have not been accounted yet.
2388 unsigned long long task_sched_runtime(struct task_struct
*p
)
2390 unsigned long flags
;
2394 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2396 * 64-bit doesn't need locks to atomically read a 64bit value.
2397 * So we have a optimization chance when the task's delta_exec is 0.
2398 * Reading ->on_cpu is racy, but this is ok.
2400 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2401 * If we race with it entering cpu, unaccounted time is 0. This is
2402 * indistinguishable from the read occurring a few cycles earlier.
2405 return p
->se
.sum_exec_runtime
;
2408 rq
= task_rq_lock(p
, &flags
);
2409 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2410 task_rq_unlock(rq
, p
, &flags
);
2416 * This function gets called by the timer code, with HZ frequency.
2417 * We call it with interrupts disabled.
2419 void scheduler_tick(void)
2421 int cpu
= smp_processor_id();
2422 struct rq
*rq
= cpu_rq(cpu
);
2423 struct task_struct
*curr
= rq
->curr
;
2427 raw_spin_lock(&rq
->lock
);
2428 update_rq_clock(rq
);
2429 curr
->sched_class
->task_tick(rq
, curr
, 0);
2430 update_cpu_load_active(rq
);
2431 raw_spin_unlock(&rq
->lock
);
2433 perf_event_task_tick();
2436 rq
->idle_balance
= idle_cpu(cpu
);
2437 trigger_load_balance(rq
);
2439 rq_last_tick_reset(rq
);
2442 #ifdef CONFIG_NO_HZ_FULL
2444 * scheduler_tick_max_deferment
2446 * Keep at least one tick per second when a single
2447 * active task is running because the scheduler doesn't
2448 * yet completely support full dynticks environment.
2450 * This makes sure that uptime, CFS vruntime, load
2451 * balancing, etc... continue to move forward, even
2452 * with a very low granularity.
2454 * Return: Maximum deferment in nanoseconds.
2456 u64
scheduler_tick_max_deferment(void)
2458 struct rq
*rq
= this_rq();
2459 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2461 next
= rq
->last_sched_tick
+ HZ
;
2463 if (time_before_eq(next
, now
))
2466 return jiffies_to_nsecs(next
- now
);
2470 notrace
unsigned long get_parent_ip(unsigned long addr
)
2472 if (in_lock_functions(addr
)) {
2473 addr
= CALLER_ADDR2
;
2474 if (in_lock_functions(addr
))
2475 addr
= CALLER_ADDR3
;
2480 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2481 defined(CONFIG_PREEMPT_TRACER))
2483 void __kprobes
preempt_count_add(int val
)
2485 #ifdef CONFIG_DEBUG_PREEMPT
2489 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2492 __preempt_count_add(val
);
2493 #ifdef CONFIG_DEBUG_PREEMPT
2495 * Spinlock count overflowing soon?
2497 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2500 if (preempt_count() == val
) {
2501 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2502 #ifdef CONFIG_DEBUG_PREEMPT
2503 current
->preempt_disable_ip
= ip
;
2505 trace_preempt_off(CALLER_ADDR0
, ip
);
2508 EXPORT_SYMBOL(preempt_count_add
);
2510 void __kprobes
preempt_count_sub(int val
)
2512 #ifdef CONFIG_DEBUG_PREEMPT
2516 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2519 * Is the spinlock portion underflowing?
2521 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2522 !(preempt_count() & PREEMPT_MASK
)))
2526 if (preempt_count() == val
)
2527 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2528 __preempt_count_sub(val
);
2530 EXPORT_SYMBOL(preempt_count_sub
);
2535 * Print scheduling while atomic bug:
2537 static noinline
void __schedule_bug(struct task_struct
*prev
)
2539 if (oops_in_progress
)
2542 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2543 prev
->comm
, prev
->pid
, preempt_count());
2545 debug_show_held_locks(prev
);
2547 if (irqs_disabled())
2548 print_irqtrace_events(prev
);
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550 if (in_atomic_preempt_off()) {
2551 pr_err("Preemption disabled at:");
2552 print_ip_sym(current
->preempt_disable_ip
);
2557 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2561 * Various schedule()-time debugging checks and statistics:
2563 static inline void schedule_debug(struct task_struct
*prev
)
2566 * Test if we are atomic. Since do_exit() needs to call into
2567 * schedule() atomically, we ignore that path. Otherwise whine
2568 * if we are scheduling when we should not.
2570 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2571 __schedule_bug(prev
);
2574 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2576 schedstat_inc(this_rq(), sched_count
);
2580 * Pick up the highest-prio task:
2582 static inline struct task_struct
*
2583 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2585 const struct sched_class
*class = &fair_sched_class
;
2586 struct task_struct
*p
;
2589 * Optimization: we know that if all tasks are in
2590 * the fair class we can call that function directly:
2592 if (likely(prev
->sched_class
== class &&
2593 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2594 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2595 if (unlikely(p
== RETRY_TASK
))
2598 /* assumes fair_sched_class->next == idle_sched_class */
2600 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2606 for_each_class(class) {
2607 p
= class->pick_next_task(rq
, prev
);
2609 if (unlikely(p
== RETRY_TASK
))
2615 BUG(); /* the idle class will always have a runnable task */
2619 * __schedule() is the main scheduler function.
2621 * The main means of driving the scheduler and thus entering this function are:
2623 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2625 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2626 * paths. For example, see arch/x86/entry_64.S.
2628 * To drive preemption between tasks, the scheduler sets the flag in timer
2629 * interrupt handler scheduler_tick().
2631 * 3. Wakeups don't really cause entry into schedule(). They add a
2632 * task to the run-queue and that's it.
2634 * Now, if the new task added to the run-queue preempts the current
2635 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2636 * called on the nearest possible occasion:
2638 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2640 * - in syscall or exception context, at the next outmost
2641 * preempt_enable(). (this might be as soon as the wake_up()'s
2644 * - in IRQ context, return from interrupt-handler to
2645 * preemptible context
2647 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2650 * - cond_resched() call
2651 * - explicit schedule() call
2652 * - return from syscall or exception to user-space
2653 * - return from interrupt-handler to user-space
2655 static void __sched
__schedule(void)
2657 struct task_struct
*prev
, *next
;
2658 unsigned long *switch_count
;
2664 cpu
= smp_processor_id();
2666 rcu_note_context_switch(cpu
);
2669 schedule_debug(prev
);
2671 if (sched_feat(HRTICK
))
2675 * Make sure that signal_pending_state()->signal_pending() below
2676 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2677 * done by the caller to avoid the race with signal_wake_up().
2679 smp_mb__before_spinlock();
2680 raw_spin_lock_irq(&rq
->lock
);
2682 switch_count
= &prev
->nivcsw
;
2683 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2684 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2685 prev
->state
= TASK_RUNNING
;
2687 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2691 * If a worker went to sleep, notify and ask workqueue
2692 * whether it wants to wake up a task to maintain
2695 if (prev
->flags
& PF_WQ_WORKER
) {
2696 struct task_struct
*to_wakeup
;
2698 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2700 try_to_wake_up_local(to_wakeup
);
2703 switch_count
= &prev
->nvcsw
;
2706 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2707 update_rq_clock(rq
);
2709 next
= pick_next_task(rq
, prev
);
2710 clear_tsk_need_resched(prev
);
2711 clear_preempt_need_resched();
2712 rq
->skip_clock_update
= 0;
2714 if (likely(prev
!= next
)) {
2719 context_switch(rq
, prev
, next
); /* unlocks the rq */
2721 * The context switch have flipped the stack from under us
2722 * and restored the local variables which were saved when
2723 * this task called schedule() in the past. prev == current
2724 * is still correct, but it can be moved to another cpu/rq.
2726 cpu
= smp_processor_id();
2729 raw_spin_unlock_irq(&rq
->lock
);
2733 sched_preempt_enable_no_resched();
2738 static inline void sched_submit_work(struct task_struct
*tsk
)
2740 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2743 * If we are going to sleep and we have plugged IO queued,
2744 * make sure to submit it to avoid deadlocks.
2746 if (blk_needs_flush_plug(tsk
))
2747 blk_schedule_flush_plug(tsk
);
2750 asmlinkage
void __sched
schedule(void)
2752 struct task_struct
*tsk
= current
;
2754 sched_submit_work(tsk
);
2757 EXPORT_SYMBOL(schedule
);
2759 #ifdef CONFIG_CONTEXT_TRACKING
2760 asmlinkage
void __sched
schedule_user(void)
2763 * If we come here after a random call to set_need_resched(),
2764 * or we have been woken up remotely but the IPI has not yet arrived,
2765 * we haven't yet exited the RCU idle mode. Do it here manually until
2766 * we find a better solution.
2775 * schedule_preempt_disabled - called with preemption disabled
2777 * Returns with preemption disabled. Note: preempt_count must be 1
2779 void __sched
schedule_preempt_disabled(void)
2781 sched_preempt_enable_no_resched();
2786 #ifdef CONFIG_PREEMPT
2788 * this is the entry point to schedule() from in-kernel preemption
2789 * off of preempt_enable. Kernel preemptions off return from interrupt
2790 * occur there and call schedule directly.
2792 asmlinkage
void __sched notrace
preempt_schedule(void)
2795 * If there is a non-zero preempt_count or interrupts are disabled,
2796 * we do not want to preempt the current task. Just return..
2798 if (likely(!preemptible()))
2802 __preempt_count_add(PREEMPT_ACTIVE
);
2804 __preempt_count_sub(PREEMPT_ACTIVE
);
2807 * Check again in case we missed a preemption opportunity
2808 * between schedule and now.
2811 } while (need_resched());
2813 EXPORT_SYMBOL(preempt_schedule
);
2814 #endif /* CONFIG_PREEMPT */
2817 * this is the entry point to schedule() from kernel preemption
2818 * off of irq context.
2819 * Note, that this is called and return with irqs disabled. This will
2820 * protect us against recursive calling from irq.
2822 asmlinkage
void __sched
preempt_schedule_irq(void)
2824 enum ctx_state prev_state
;
2826 /* Catch callers which need to be fixed */
2827 BUG_ON(preempt_count() || !irqs_disabled());
2829 prev_state
= exception_enter();
2832 __preempt_count_add(PREEMPT_ACTIVE
);
2835 local_irq_disable();
2836 __preempt_count_sub(PREEMPT_ACTIVE
);
2839 * Check again in case we missed a preemption opportunity
2840 * between schedule and now.
2843 } while (need_resched());
2845 exception_exit(prev_state
);
2848 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2851 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2853 EXPORT_SYMBOL(default_wake_function
);
2855 #ifdef CONFIG_RT_MUTEXES
2858 * rt_mutex_setprio - set the current priority of a task
2860 * @prio: prio value (kernel-internal form)
2862 * This function changes the 'effective' priority of a task. It does
2863 * not touch ->normal_prio like __setscheduler().
2865 * Used by the rt_mutex code to implement priority inheritance
2866 * logic. Call site only calls if the priority of the task changed.
2868 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2870 int oldprio
, on_rq
, running
, enqueue_flag
= 0;
2872 const struct sched_class
*prev_class
;
2874 BUG_ON(prio
> MAX_PRIO
);
2876 rq
= __task_rq_lock(p
);
2879 * Idle task boosting is a nono in general. There is one
2880 * exception, when PREEMPT_RT and NOHZ is active:
2882 * The idle task calls get_next_timer_interrupt() and holds
2883 * the timer wheel base->lock on the CPU and another CPU wants
2884 * to access the timer (probably to cancel it). We can safely
2885 * ignore the boosting request, as the idle CPU runs this code
2886 * with interrupts disabled and will complete the lock
2887 * protected section without being interrupted. So there is no
2888 * real need to boost.
2890 if (unlikely(p
== rq
->idle
)) {
2891 WARN_ON(p
!= rq
->curr
);
2892 WARN_ON(p
->pi_blocked_on
);
2896 trace_sched_pi_setprio(p
, prio
);
2897 p
->pi_top_task
= rt_mutex_get_top_task(p
);
2899 prev_class
= p
->sched_class
;
2901 running
= task_current(rq
, p
);
2903 dequeue_task(rq
, p
, 0);
2905 p
->sched_class
->put_prev_task(rq
, p
);
2908 * Boosting condition are:
2909 * 1. -rt task is running and holds mutex A
2910 * --> -dl task blocks on mutex A
2912 * 2. -dl task is running and holds mutex A
2913 * --> -dl task blocks on mutex A and could preempt the
2916 if (dl_prio(prio
)) {
2917 if (!dl_prio(p
->normal_prio
) || (p
->pi_top_task
&&
2918 dl_entity_preempt(&p
->pi_top_task
->dl
, &p
->dl
))) {
2919 p
->dl
.dl_boosted
= 1;
2920 p
->dl
.dl_throttled
= 0;
2921 enqueue_flag
= ENQUEUE_REPLENISH
;
2923 p
->dl
.dl_boosted
= 0;
2924 p
->sched_class
= &dl_sched_class
;
2925 } else if (rt_prio(prio
)) {
2926 if (dl_prio(oldprio
))
2927 p
->dl
.dl_boosted
= 0;
2929 enqueue_flag
= ENQUEUE_HEAD
;
2930 p
->sched_class
= &rt_sched_class
;
2932 if (dl_prio(oldprio
))
2933 p
->dl
.dl_boosted
= 0;
2934 p
->sched_class
= &fair_sched_class
;
2940 p
->sched_class
->set_curr_task(rq
);
2942 enqueue_task(rq
, p
, enqueue_flag
);
2944 check_class_changed(rq
, p
, prev_class
, oldprio
);
2946 __task_rq_unlock(rq
);
2950 void set_user_nice(struct task_struct
*p
, long nice
)
2952 int old_prio
, delta
, on_rq
;
2953 unsigned long flags
;
2956 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
2959 * We have to be careful, if called from sys_setpriority(),
2960 * the task might be in the middle of scheduling on another CPU.
2962 rq
= task_rq_lock(p
, &flags
);
2964 * The RT priorities are set via sched_setscheduler(), but we still
2965 * allow the 'normal' nice value to be set - but as expected
2966 * it wont have any effect on scheduling until the task is
2967 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2969 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2970 p
->static_prio
= NICE_TO_PRIO(nice
);
2975 dequeue_task(rq
, p
, 0);
2977 p
->static_prio
= NICE_TO_PRIO(nice
);
2980 p
->prio
= effective_prio(p
);
2981 delta
= p
->prio
- old_prio
;
2984 enqueue_task(rq
, p
, 0);
2986 * If the task increased its priority or is running and
2987 * lowered its priority, then reschedule its CPU:
2989 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
2990 resched_task(rq
->curr
);
2993 task_rq_unlock(rq
, p
, &flags
);
2995 EXPORT_SYMBOL(set_user_nice
);
2998 * can_nice - check if a task can reduce its nice value
3002 int can_nice(const struct task_struct
*p
, const int nice
)
3004 /* convert nice value [19,-20] to rlimit style value [1,40] */
3005 int nice_rlim
= 20 - nice
;
3007 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3008 capable(CAP_SYS_NICE
));
3011 #ifdef __ARCH_WANT_SYS_NICE
3014 * sys_nice - change the priority of the current process.
3015 * @increment: priority increment
3017 * sys_setpriority is a more generic, but much slower function that
3018 * does similar things.
3020 SYSCALL_DEFINE1(nice
, int, increment
)
3025 * Setpriority might change our priority at the same moment.
3026 * We don't have to worry. Conceptually one call occurs first
3027 * and we have a single winner.
3029 if (increment
< -40)
3034 nice
= task_nice(current
) + increment
;
3035 if (nice
< MIN_NICE
)
3037 if (nice
> MAX_NICE
)
3040 if (increment
< 0 && !can_nice(current
, nice
))
3043 retval
= security_task_setnice(current
, nice
);
3047 set_user_nice(current
, nice
);
3054 * task_prio - return the priority value of a given task.
3055 * @p: the task in question.
3057 * Return: The priority value as seen by users in /proc.
3058 * RT tasks are offset by -200. Normal tasks are centered
3059 * around 0, value goes from -16 to +15.
3061 int task_prio(const struct task_struct
*p
)
3063 return p
->prio
- MAX_RT_PRIO
;
3067 * idle_cpu - is a given cpu idle currently?
3068 * @cpu: the processor in question.
3070 * Return: 1 if the CPU is currently idle. 0 otherwise.
3072 int idle_cpu(int cpu
)
3074 struct rq
*rq
= cpu_rq(cpu
);
3076 if (rq
->curr
!= rq
->idle
)
3083 if (!llist_empty(&rq
->wake_list
))
3091 * idle_task - return the idle task for a given cpu.
3092 * @cpu: the processor in question.
3094 * Return: The idle task for the cpu @cpu.
3096 struct task_struct
*idle_task(int cpu
)
3098 return cpu_rq(cpu
)->idle
;
3102 * find_process_by_pid - find a process with a matching PID value.
3103 * @pid: the pid in question.
3105 * The task of @pid, if found. %NULL otherwise.
3107 static struct task_struct
*find_process_by_pid(pid_t pid
)
3109 return pid
? find_task_by_vpid(pid
) : current
;
3113 * This function initializes the sched_dl_entity of a newly becoming
3114 * SCHED_DEADLINE task.
3116 * Only the static values are considered here, the actual runtime and the
3117 * absolute deadline will be properly calculated when the task is enqueued
3118 * for the first time with its new policy.
3121 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3123 struct sched_dl_entity
*dl_se
= &p
->dl
;
3125 init_dl_task_timer(dl_se
);
3126 dl_se
->dl_runtime
= attr
->sched_runtime
;
3127 dl_se
->dl_deadline
= attr
->sched_deadline
;
3128 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3129 dl_se
->flags
= attr
->sched_flags
;
3130 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3131 dl_se
->dl_throttled
= 0;
3133 dl_se
->dl_yielded
= 0;
3136 static void __setscheduler_params(struct task_struct
*p
,
3137 const struct sched_attr
*attr
)
3139 int policy
= attr
->sched_policy
;
3141 if (policy
== -1) /* setparam */
3146 if (dl_policy(policy
))
3147 __setparam_dl(p
, attr
);
3148 else if (fair_policy(policy
))
3149 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3152 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3153 * !rt_policy. Always setting this ensures that things like
3154 * getparam()/getattr() don't report silly values for !rt tasks.
3156 p
->rt_priority
= attr
->sched_priority
;
3157 p
->normal_prio
= normal_prio(p
);
3161 /* Actually do priority change: must hold pi & rq lock. */
3162 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3163 const struct sched_attr
*attr
)
3165 __setscheduler_params(p
, attr
);
3168 * If we get here, there was no pi waiters boosting the
3169 * task. It is safe to use the normal prio.
3171 p
->prio
= normal_prio(p
);
3173 if (dl_prio(p
->prio
))
3174 p
->sched_class
= &dl_sched_class
;
3175 else if (rt_prio(p
->prio
))
3176 p
->sched_class
= &rt_sched_class
;
3178 p
->sched_class
= &fair_sched_class
;
3182 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3184 struct sched_dl_entity
*dl_se
= &p
->dl
;
3186 attr
->sched_priority
= p
->rt_priority
;
3187 attr
->sched_runtime
= dl_se
->dl_runtime
;
3188 attr
->sched_deadline
= dl_se
->dl_deadline
;
3189 attr
->sched_period
= dl_se
->dl_period
;
3190 attr
->sched_flags
= dl_se
->flags
;
3194 * This function validates the new parameters of a -deadline task.
3195 * We ask for the deadline not being zero, and greater or equal
3196 * than the runtime, as well as the period of being zero or
3197 * greater than deadline. Furthermore, we have to be sure that
3198 * user parameters are above the internal resolution (1us); we
3199 * check sched_runtime only since it is always the smaller one.
3202 __checkparam_dl(const struct sched_attr
*attr
)
3204 return attr
&& attr
->sched_deadline
!= 0 &&
3205 (attr
->sched_period
== 0 ||
3206 (s64
)(attr
->sched_period
- attr
->sched_deadline
) >= 0) &&
3207 (s64
)(attr
->sched_deadline
- attr
->sched_runtime
) >= 0 &&
3208 attr
->sched_runtime
>= (2 << (DL_SCALE
- 1));
3212 * check the target process has a UID that matches the current process's
3214 static bool check_same_owner(struct task_struct
*p
)
3216 const struct cred
*cred
= current_cred(), *pcred
;
3220 pcred
= __task_cred(p
);
3221 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3222 uid_eq(cred
->euid
, pcred
->uid
));
3227 static int __sched_setscheduler(struct task_struct
*p
,
3228 const struct sched_attr
*attr
,
3231 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3232 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3233 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3234 int policy
= attr
->sched_policy
;
3235 unsigned long flags
;
3236 const struct sched_class
*prev_class
;
3240 /* may grab non-irq protected spin_locks */
3241 BUG_ON(in_interrupt());
3243 /* double check policy once rq lock held */
3245 reset_on_fork
= p
->sched_reset_on_fork
;
3246 policy
= oldpolicy
= p
->policy
;
3248 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3250 if (policy
!= SCHED_DEADLINE
&&
3251 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3252 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3253 policy
!= SCHED_IDLE
)
3257 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3261 * Valid priorities for SCHED_FIFO and SCHED_RR are
3262 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3263 * SCHED_BATCH and SCHED_IDLE is 0.
3265 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3266 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3268 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3269 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3273 * Allow unprivileged RT tasks to decrease priority:
3275 if (user
&& !capable(CAP_SYS_NICE
)) {
3276 if (fair_policy(policy
)) {
3277 if (attr
->sched_nice
< task_nice(p
) &&
3278 !can_nice(p
, attr
->sched_nice
))
3282 if (rt_policy(policy
)) {
3283 unsigned long rlim_rtprio
=
3284 task_rlimit(p
, RLIMIT_RTPRIO
);
3286 /* can't set/change the rt policy */
3287 if (policy
!= p
->policy
&& !rlim_rtprio
)
3290 /* can't increase priority */
3291 if (attr
->sched_priority
> p
->rt_priority
&&
3292 attr
->sched_priority
> rlim_rtprio
)
3297 * Can't set/change SCHED_DEADLINE policy at all for now
3298 * (safest behavior); in the future we would like to allow
3299 * unprivileged DL tasks to increase their relative deadline
3300 * or reduce their runtime (both ways reducing utilization)
3302 if (dl_policy(policy
))
3306 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3307 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3309 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3310 if (!can_nice(p
, task_nice(p
)))
3314 /* can't change other user's priorities */
3315 if (!check_same_owner(p
))
3318 /* Normal users shall not reset the sched_reset_on_fork flag */
3319 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3324 retval
= security_task_setscheduler(p
);
3330 * make sure no PI-waiters arrive (or leave) while we are
3331 * changing the priority of the task:
3333 * To be able to change p->policy safely, the appropriate
3334 * runqueue lock must be held.
3336 rq
= task_rq_lock(p
, &flags
);
3339 * Changing the policy of the stop threads its a very bad idea
3341 if (p
== rq
->stop
) {
3342 task_rq_unlock(rq
, p
, &flags
);
3347 * If not changing anything there's no need to proceed further,
3348 * but store a possible modification of reset_on_fork.
3350 if (unlikely(policy
== p
->policy
)) {
3351 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3353 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3355 if (dl_policy(policy
))
3358 p
->sched_reset_on_fork
= reset_on_fork
;
3359 task_rq_unlock(rq
, p
, &flags
);
3365 #ifdef CONFIG_RT_GROUP_SCHED
3367 * Do not allow realtime tasks into groups that have no runtime
3370 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3371 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3372 !task_group_is_autogroup(task_group(p
))) {
3373 task_rq_unlock(rq
, p
, &flags
);
3378 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3379 cpumask_t
*span
= rq
->rd
->span
;
3382 * Don't allow tasks with an affinity mask smaller than
3383 * the entire root_domain to become SCHED_DEADLINE. We
3384 * will also fail if there's no bandwidth available.
3386 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3387 rq
->rd
->dl_bw
.bw
== 0) {
3388 task_rq_unlock(rq
, p
, &flags
);
3395 /* recheck policy now with rq lock held */
3396 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3397 policy
= oldpolicy
= -1;
3398 task_rq_unlock(rq
, p
, &flags
);
3403 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3404 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3407 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3408 task_rq_unlock(rq
, p
, &flags
);
3412 p
->sched_reset_on_fork
= reset_on_fork
;
3416 * Special case for priority boosted tasks.
3418 * If the new priority is lower or equal (user space view)
3419 * than the current (boosted) priority, we just store the new
3420 * normal parameters and do not touch the scheduler class and
3421 * the runqueue. This will be done when the task deboost
3424 if (rt_mutex_check_prio(p
, newprio
)) {
3425 __setscheduler_params(p
, attr
);
3426 task_rq_unlock(rq
, p
, &flags
);
3431 running
= task_current(rq
, p
);
3433 dequeue_task(rq
, p
, 0);
3435 p
->sched_class
->put_prev_task(rq
, p
);
3437 prev_class
= p
->sched_class
;
3438 __setscheduler(rq
, p
, attr
);
3441 p
->sched_class
->set_curr_task(rq
);
3444 * We enqueue to tail when the priority of a task is
3445 * increased (user space view).
3447 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3450 check_class_changed(rq
, p
, prev_class
, oldprio
);
3451 task_rq_unlock(rq
, p
, &flags
);
3453 rt_mutex_adjust_pi(p
);
3458 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3459 const struct sched_param
*param
, bool check
)
3461 struct sched_attr attr
= {
3462 .sched_policy
= policy
,
3463 .sched_priority
= param
->sched_priority
,
3464 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3468 * Fixup the legacy SCHED_RESET_ON_FORK hack
3470 if (policy
& SCHED_RESET_ON_FORK
) {
3471 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3472 policy
&= ~SCHED_RESET_ON_FORK
;
3473 attr
.sched_policy
= policy
;
3476 return __sched_setscheduler(p
, &attr
, check
);
3479 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3480 * @p: the task in question.
3481 * @policy: new policy.
3482 * @param: structure containing the new RT priority.
3484 * Return: 0 on success. An error code otherwise.
3486 * NOTE that the task may be already dead.
3488 int sched_setscheduler(struct task_struct
*p
, int policy
,
3489 const struct sched_param
*param
)
3491 return _sched_setscheduler(p
, policy
, param
, true);
3493 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3495 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3497 return __sched_setscheduler(p
, attr
, true);
3499 EXPORT_SYMBOL_GPL(sched_setattr
);
3502 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3503 * @p: the task in question.
3504 * @policy: new policy.
3505 * @param: structure containing the new RT priority.
3507 * Just like sched_setscheduler, only don't bother checking if the
3508 * current context has permission. For example, this is needed in
3509 * stop_machine(): we create temporary high priority worker threads,
3510 * but our caller might not have that capability.
3512 * Return: 0 on success. An error code otherwise.
3514 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3515 const struct sched_param
*param
)
3517 return _sched_setscheduler(p
, policy
, param
, false);
3521 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3523 struct sched_param lparam
;
3524 struct task_struct
*p
;
3527 if (!param
|| pid
< 0)
3529 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3534 p
= find_process_by_pid(pid
);
3536 retval
= sched_setscheduler(p
, policy
, &lparam
);
3543 * Mimics kernel/events/core.c perf_copy_attr().
3545 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3546 struct sched_attr
*attr
)
3551 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3555 * zero the full structure, so that a short copy will be nice.
3557 memset(attr
, 0, sizeof(*attr
));
3559 ret
= get_user(size
, &uattr
->size
);
3563 if (size
> PAGE_SIZE
) /* silly large */
3566 if (!size
) /* abi compat */
3567 size
= SCHED_ATTR_SIZE_VER0
;
3569 if (size
< SCHED_ATTR_SIZE_VER0
)
3573 * If we're handed a bigger struct than we know of,
3574 * ensure all the unknown bits are 0 - i.e. new
3575 * user-space does not rely on any kernel feature
3576 * extensions we dont know about yet.
3578 if (size
> sizeof(*attr
)) {
3579 unsigned char __user
*addr
;
3580 unsigned char __user
*end
;
3583 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3584 end
= (void __user
*)uattr
+ size
;
3586 for (; addr
< end
; addr
++) {
3587 ret
= get_user(val
, addr
);
3593 size
= sizeof(*attr
);
3596 ret
= copy_from_user(attr
, uattr
, size
);
3601 * XXX: do we want to be lenient like existing syscalls; or do we want
3602 * to be strict and return an error on out-of-bounds values?
3604 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3610 put_user(sizeof(*attr
), &uattr
->size
);
3616 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3617 * @pid: the pid in question.
3618 * @policy: new policy.
3619 * @param: structure containing the new RT priority.
3621 * Return: 0 on success. An error code otherwise.
3623 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3624 struct sched_param __user
*, param
)
3626 /* negative values for policy are not valid */
3630 return do_sched_setscheduler(pid
, policy
, param
);
3634 * sys_sched_setparam - set/change the RT priority of a thread
3635 * @pid: the pid in question.
3636 * @param: structure containing the new RT priority.
3638 * Return: 0 on success. An error code otherwise.
3640 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3642 return do_sched_setscheduler(pid
, -1, param
);
3646 * sys_sched_setattr - same as above, but with extended sched_attr
3647 * @pid: the pid in question.
3648 * @uattr: structure containing the extended parameters.
3649 * @flags: for future extension.
3651 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3652 unsigned int, flags
)
3654 struct sched_attr attr
;
3655 struct task_struct
*p
;
3658 if (!uattr
|| pid
< 0 || flags
)
3661 retval
= sched_copy_attr(uattr
, &attr
);
3665 if (attr
.sched_policy
< 0)
3670 p
= find_process_by_pid(pid
);
3672 retval
= sched_setattr(p
, &attr
);
3679 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3680 * @pid: the pid in question.
3682 * Return: On success, the policy of the thread. Otherwise, a negative error
3685 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3687 struct task_struct
*p
;
3695 p
= find_process_by_pid(pid
);
3697 retval
= security_task_getscheduler(p
);
3700 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3707 * sys_sched_getparam - get the RT priority of a thread
3708 * @pid: the pid in question.
3709 * @param: structure containing the RT priority.
3711 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3714 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3716 struct sched_param lp
= { .sched_priority
= 0 };
3717 struct task_struct
*p
;
3720 if (!param
|| pid
< 0)
3724 p
= find_process_by_pid(pid
);
3729 retval
= security_task_getscheduler(p
);
3733 if (task_has_rt_policy(p
))
3734 lp
.sched_priority
= p
->rt_priority
;
3738 * This one might sleep, we cannot do it with a spinlock held ...
3740 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3749 static int sched_read_attr(struct sched_attr __user
*uattr
,
3750 struct sched_attr
*attr
,
3755 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3759 * If we're handed a smaller struct than we know of,
3760 * ensure all the unknown bits are 0 - i.e. old
3761 * user-space does not get uncomplete information.
3763 if (usize
< sizeof(*attr
)) {
3764 unsigned char *addr
;
3767 addr
= (void *)attr
+ usize
;
3768 end
= (void *)attr
+ sizeof(*attr
);
3770 for (; addr
< end
; addr
++) {
3778 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3791 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3792 * @pid: the pid in question.
3793 * @uattr: structure containing the extended parameters.
3794 * @size: sizeof(attr) for fwd/bwd comp.
3795 * @flags: for future extension.
3797 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3798 unsigned int, size
, unsigned int, flags
)
3800 struct sched_attr attr
= {
3801 .size
= sizeof(struct sched_attr
),
3803 struct task_struct
*p
;
3806 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3807 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3811 p
= find_process_by_pid(pid
);
3816 retval
= security_task_getscheduler(p
);
3820 attr
.sched_policy
= p
->policy
;
3821 if (p
->sched_reset_on_fork
)
3822 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3823 if (task_has_dl_policy(p
))
3824 __getparam_dl(p
, &attr
);
3825 else if (task_has_rt_policy(p
))
3826 attr
.sched_priority
= p
->rt_priority
;
3828 attr
.sched_nice
= task_nice(p
);
3832 retval
= sched_read_attr(uattr
, &attr
, size
);
3840 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3842 cpumask_var_t cpus_allowed
, new_mask
;
3843 struct task_struct
*p
;
3848 p
= find_process_by_pid(pid
);
3854 /* Prevent p going away */
3858 if (p
->flags
& PF_NO_SETAFFINITY
) {
3862 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3866 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3868 goto out_free_cpus_allowed
;
3871 if (!check_same_owner(p
)) {
3873 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3880 retval
= security_task_setscheduler(p
);
3885 cpuset_cpus_allowed(p
, cpus_allowed
);
3886 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3889 * Since bandwidth control happens on root_domain basis,
3890 * if admission test is enabled, we only admit -deadline
3891 * tasks allowed to run on all the CPUs in the task's
3895 if (task_has_dl_policy(p
)) {
3896 const struct cpumask
*span
= task_rq(p
)->rd
->span
;
3898 if (dl_bandwidth_enabled() && !cpumask_subset(span
, new_mask
)) {
3905 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3908 cpuset_cpus_allowed(p
, cpus_allowed
);
3909 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3911 * We must have raced with a concurrent cpuset
3912 * update. Just reset the cpus_allowed to the
3913 * cpuset's cpus_allowed
3915 cpumask_copy(new_mask
, cpus_allowed
);
3920 free_cpumask_var(new_mask
);
3921 out_free_cpus_allowed
:
3922 free_cpumask_var(cpus_allowed
);
3928 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3929 struct cpumask
*new_mask
)
3931 if (len
< cpumask_size())
3932 cpumask_clear(new_mask
);
3933 else if (len
> cpumask_size())
3934 len
= cpumask_size();
3936 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3940 * sys_sched_setaffinity - set the cpu affinity of a process
3941 * @pid: pid of the process
3942 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3943 * @user_mask_ptr: user-space pointer to the new cpu mask
3945 * Return: 0 on success. An error code otherwise.
3947 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3948 unsigned long __user
*, user_mask_ptr
)
3950 cpumask_var_t new_mask
;
3953 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3956 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3958 retval
= sched_setaffinity(pid
, new_mask
);
3959 free_cpumask_var(new_mask
);
3963 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
3965 struct task_struct
*p
;
3966 unsigned long flags
;
3972 p
= find_process_by_pid(pid
);
3976 retval
= security_task_getscheduler(p
);
3980 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3981 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
3982 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3991 * sys_sched_getaffinity - get the cpu affinity of a process
3992 * @pid: pid of the process
3993 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3994 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3996 * Return: 0 on success. An error code otherwise.
3998 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
3999 unsigned long __user
*, user_mask_ptr
)
4004 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4006 if (len
& (sizeof(unsigned long)-1))
4009 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4012 ret
= sched_getaffinity(pid
, mask
);
4014 size_t retlen
= min_t(size_t, len
, cpumask_size());
4016 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4021 free_cpumask_var(mask
);
4027 * sys_sched_yield - yield the current processor to other threads.
4029 * This function yields the current CPU to other tasks. If there are no
4030 * other threads running on this CPU then this function will return.
4034 SYSCALL_DEFINE0(sched_yield
)
4036 struct rq
*rq
= this_rq_lock();
4038 schedstat_inc(rq
, yld_count
);
4039 current
->sched_class
->yield_task(rq
);
4042 * Since we are going to call schedule() anyway, there's
4043 * no need to preempt or enable interrupts:
4045 __release(rq
->lock
);
4046 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4047 do_raw_spin_unlock(&rq
->lock
);
4048 sched_preempt_enable_no_resched();
4055 static void __cond_resched(void)
4057 __preempt_count_add(PREEMPT_ACTIVE
);
4059 __preempt_count_sub(PREEMPT_ACTIVE
);
4062 int __sched
_cond_resched(void)
4064 if (should_resched()) {
4070 EXPORT_SYMBOL(_cond_resched
);
4073 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4074 * call schedule, and on return reacquire the lock.
4076 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4077 * operations here to prevent schedule() from being called twice (once via
4078 * spin_unlock(), once by hand).
4080 int __cond_resched_lock(spinlock_t
*lock
)
4082 int resched
= should_resched();
4085 lockdep_assert_held(lock
);
4087 if (spin_needbreak(lock
) || resched
) {
4098 EXPORT_SYMBOL(__cond_resched_lock
);
4100 int __sched
__cond_resched_softirq(void)
4102 BUG_ON(!in_softirq());
4104 if (should_resched()) {
4112 EXPORT_SYMBOL(__cond_resched_softirq
);
4115 * yield - yield the current processor to other threads.
4117 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4119 * The scheduler is at all times free to pick the calling task as the most
4120 * eligible task to run, if removing the yield() call from your code breaks
4121 * it, its already broken.
4123 * Typical broken usage is:
4128 * where one assumes that yield() will let 'the other' process run that will
4129 * make event true. If the current task is a SCHED_FIFO task that will never
4130 * happen. Never use yield() as a progress guarantee!!
4132 * If you want to use yield() to wait for something, use wait_event().
4133 * If you want to use yield() to be 'nice' for others, use cond_resched().
4134 * If you still want to use yield(), do not!
4136 void __sched
yield(void)
4138 set_current_state(TASK_RUNNING
);
4141 EXPORT_SYMBOL(yield
);
4144 * yield_to - yield the current processor to another thread in
4145 * your thread group, or accelerate that thread toward the
4146 * processor it's on.
4148 * @preempt: whether task preemption is allowed or not
4150 * It's the caller's job to ensure that the target task struct
4151 * can't go away on us before we can do any checks.
4154 * true (>0) if we indeed boosted the target task.
4155 * false (0) if we failed to boost the target.
4156 * -ESRCH if there's no task to yield to.
4158 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4160 struct task_struct
*curr
= current
;
4161 struct rq
*rq
, *p_rq
;
4162 unsigned long flags
;
4165 local_irq_save(flags
);
4171 * If we're the only runnable task on the rq and target rq also
4172 * has only one task, there's absolutely no point in yielding.
4174 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4179 double_rq_lock(rq
, p_rq
);
4180 if (task_rq(p
) != p_rq
) {
4181 double_rq_unlock(rq
, p_rq
);
4185 if (!curr
->sched_class
->yield_to_task
)
4188 if (curr
->sched_class
!= p
->sched_class
)
4191 if (task_running(p_rq
, p
) || p
->state
)
4194 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4196 schedstat_inc(rq
, yld_count
);
4198 * Make p's CPU reschedule; pick_next_entity takes care of
4201 if (preempt
&& rq
!= p_rq
)
4202 resched_task(p_rq
->curr
);
4206 double_rq_unlock(rq
, p_rq
);
4208 local_irq_restore(flags
);
4215 EXPORT_SYMBOL_GPL(yield_to
);
4218 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4219 * that process accounting knows that this is a task in IO wait state.
4221 void __sched
io_schedule(void)
4223 struct rq
*rq
= raw_rq();
4225 delayacct_blkio_start();
4226 atomic_inc(&rq
->nr_iowait
);
4227 blk_flush_plug(current
);
4228 current
->in_iowait
= 1;
4230 current
->in_iowait
= 0;
4231 atomic_dec(&rq
->nr_iowait
);
4232 delayacct_blkio_end();
4234 EXPORT_SYMBOL(io_schedule
);
4236 long __sched
io_schedule_timeout(long timeout
)
4238 struct rq
*rq
= raw_rq();
4241 delayacct_blkio_start();
4242 atomic_inc(&rq
->nr_iowait
);
4243 blk_flush_plug(current
);
4244 current
->in_iowait
= 1;
4245 ret
= schedule_timeout(timeout
);
4246 current
->in_iowait
= 0;
4247 atomic_dec(&rq
->nr_iowait
);
4248 delayacct_blkio_end();
4253 * sys_sched_get_priority_max - return maximum RT priority.
4254 * @policy: scheduling class.
4256 * Return: On success, this syscall returns the maximum
4257 * rt_priority that can be used by a given scheduling class.
4258 * On failure, a negative error code is returned.
4260 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4267 ret
= MAX_USER_RT_PRIO
-1;
4269 case SCHED_DEADLINE
:
4280 * sys_sched_get_priority_min - return minimum RT priority.
4281 * @policy: scheduling class.
4283 * Return: On success, this syscall returns the minimum
4284 * rt_priority that can be used by a given scheduling class.
4285 * On failure, a negative error code is returned.
4287 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4296 case SCHED_DEADLINE
:
4306 * sys_sched_rr_get_interval - return the default timeslice of a process.
4307 * @pid: pid of the process.
4308 * @interval: userspace pointer to the timeslice value.
4310 * this syscall writes the default timeslice value of a given process
4311 * into the user-space timespec buffer. A value of '0' means infinity.
4313 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4316 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4317 struct timespec __user
*, interval
)
4319 struct task_struct
*p
;
4320 unsigned int time_slice
;
4321 unsigned long flags
;
4331 p
= find_process_by_pid(pid
);
4335 retval
= security_task_getscheduler(p
);
4339 rq
= task_rq_lock(p
, &flags
);
4341 if (p
->sched_class
->get_rr_interval
)
4342 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4343 task_rq_unlock(rq
, p
, &flags
);
4346 jiffies_to_timespec(time_slice
, &t
);
4347 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4355 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4357 void sched_show_task(struct task_struct
*p
)
4359 unsigned long free
= 0;
4363 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4364 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4365 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4366 #if BITS_PER_LONG == 32
4367 if (state
== TASK_RUNNING
)
4368 printk(KERN_CONT
" running ");
4370 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4372 if (state
== TASK_RUNNING
)
4373 printk(KERN_CONT
" running task ");
4375 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4377 #ifdef CONFIG_DEBUG_STACK_USAGE
4378 free
= stack_not_used(p
);
4381 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4383 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4384 task_pid_nr(p
), ppid
,
4385 (unsigned long)task_thread_info(p
)->flags
);
4387 print_worker_info(KERN_INFO
, p
);
4388 show_stack(p
, NULL
);
4391 void show_state_filter(unsigned long state_filter
)
4393 struct task_struct
*g
, *p
;
4395 #if BITS_PER_LONG == 32
4397 " task PC stack pid father\n");
4400 " task PC stack pid father\n");
4403 do_each_thread(g
, p
) {
4405 * reset the NMI-timeout, listing all files on a slow
4406 * console might take a lot of time:
4408 touch_nmi_watchdog();
4409 if (!state_filter
|| (p
->state
& state_filter
))
4411 } while_each_thread(g
, p
);
4413 touch_all_softlockup_watchdogs();
4415 #ifdef CONFIG_SCHED_DEBUG
4416 sysrq_sched_debug_show();
4420 * Only show locks if all tasks are dumped:
4423 debug_show_all_locks();
4426 void init_idle_bootup_task(struct task_struct
*idle
)
4428 idle
->sched_class
= &idle_sched_class
;
4432 * init_idle - set up an idle thread for a given CPU
4433 * @idle: task in question
4434 * @cpu: cpu the idle task belongs to
4436 * NOTE: this function does not set the idle thread's NEED_RESCHED
4437 * flag, to make booting more robust.
4439 void init_idle(struct task_struct
*idle
, int cpu
)
4441 struct rq
*rq
= cpu_rq(cpu
);
4442 unsigned long flags
;
4444 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4446 __sched_fork(0, idle
);
4447 idle
->state
= TASK_RUNNING
;
4448 idle
->se
.exec_start
= sched_clock();
4450 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4452 * We're having a chicken and egg problem, even though we are
4453 * holding rq->lock, the cpu isn't yet set to this cpu so the
4454 * lockdep check in task_group() will fail.
4456 * Similar case to sched_fork(). / Alternatively we could
4457 * use task_rq_lock() here and obtain the other rq->lock.
4462 __set_task_cpu(idle
, cpu
);
4465 rq
->curr
= rq
->idle
= idle
;
4467 #if defined(CONFIG_SMP)
4470 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4472 /* Set the preempt count _outside_ the spinlocks! */
4473 init_idle_preempt_count(idle
, cpu
);
4476 * The idle tasks have their own, simple scheduling class:
4478 idle
->sched_class
= &idle_sched_class
;
4479 ftrace_graph_init_idle_task(idle
, cpu
);
4480 vtime_init_idle(idle
, cpu
);
4481 #if defined(CONFIG_SMP)
4482 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4487 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4489 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4490 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4492 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4493 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4497 * This is how migration works:
4499 * 1) we invoke migration_cpu_stop() on the target CPU using
4501 * 2) stopper starts to run (implicitly forcing the migrated thread
4503 * 3) it checks whether the migrated task is still in the wrong runqueue.
4504 * 4) if it's in the wrong runqueue then the migration thread removes
4505 * it and puts it into the right queue.
4506 * 5) stopper completes and stop_one_cpu() returns and the migration
4511 * Change a given task's CPU affinity. Migrate the thread to a
4512 * proper CPU and schedule it away if the CPU it's executing on
4513 * is removed from the allowed bitmask.
4515 * NOTE: the caller must have a valid reference to the task, the
4516 * task must not exit() & deallocate itself prematurely. The
4517 * call is not atomic; no spinlocks may be held.
4519 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4521 unsigned long flags
;
4523 unsigned int dest_cpu
;
4526 rq
= task_rq_lock(p
, &flags
);
4528 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4531 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4536 do_set_cpus_allowed(p
, new_mask
);
4538 /* Can the task run on the task's current CPU? If so, we're done */
4539 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4542 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4544 struct migration_arg arg
= { p
, dest_cpu
};
4545 /* Need help from migration thread: drop lock and wait. */
4546 task_rq_unlock(rq
, p
, &flags
);
4547 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4548 tlb_migrate_finish(p
->mm
);
4552 task_rq_unlock(rq
, p
, &flags
);
4556 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4559 * Move (not current) task off this cpu, onto dest cpu. We're doing
4560 * this because either it can't run here any more (set_cpus_allowed()
4561 * away from this CPU, or CPU going down), or because we're
4562 * attempting to rebalance this task on exec (sched_exec).
4564 * So we race with normal scheduler movements, but that's OK, as long
4565 * as the task is no longer on this CPU.
4567 * Returns non-zero if task was successfully migrated.
4569 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4571 struct rq
*rq_dest
, *rq_src
;
4574 if (unlikely(!cpu_active(dest_cpu
)))
4577 rq_src
= cpu_rq(src_cpu
);
4578 rq_dest
= cpu_rq(dest_cpu
);
4580 raw_spin_lock(&p
->pi_lock
);
4581 double_rq_lock(rq_src
, rq_dest
);
4582 /* Already moved. */
4583 if (task_cpu(p
) != src_cpu
)
4585 /* Affinity changed (again). */
4586 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4590 * If we're not on a rq, the next wake-up will ensure we're
4594 dequeue_task(rq_src
, p
, 0);
4595 set_task_cpu(p
, dest_cpu
);
4596 enqueue_task(rq_dest
, p
, 0);
4597 check_preempt_curr(rq_dest
, p
, 0);
4602 double_rq_unlock(rq_src
, rq_dest
);
4603 raw_spin_unlock(&p
->pi_lock
);
4607 #ifdef CONFIG_NUMA_BALANCING
4608 /* Migrate current task p to target_cpu */
4609 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4611 struct migration_arg arg
= { p
, target_cpu
};
4612 int curr_cpu
= task_cpu(p
);
4614 if (curr_cpu
== target_cpu
)
4617 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4620 /* TODO: This is not properly updating schedstats */
4622 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4623 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4627 * Requeue a task on a given node and accurately track the number of NUMA
4628 * tasks on the runqueues
4630 void sched_setnuma(struct task_struct
*p
, int nid
)
4633 unsigned long flags
;
4634 bool on_rq
, running
;
4636 rq
= task_rq_lock(p
, &flags
);
4638 running
= task_current(rq
, p
);
4641 dequeue_task(rq
, p
, 0);
4643 p
->sched_class
->put_prev_task(rq
, p
);
4645 p
->numa_preferred_nid
= nid
;
4648 p
->sched_class
->set_curr_task(rq
);
4650 enqueue_task(rq
, p
, 0);
4651 task_rq_unlock(rq
, p
, &flags
);
4656 * migration_cpu_stop - this will be executed by a highprio stopper thread
4657 * and performs thread migration by bumping thread off CPU then
4658 * 'pushing' onto another runqueue.
4660 static int migration_cpu_stop(void *data
)
4662 struct migration_arg
*arg
= data
;
4665 * The original target cpu might have gone down and we might
4666 * be on another cpu but it doesn't matter.
4668 local_irq_disable();
4669 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4674 #ifdef CONFIG_HOTPLUG_CPU
4677 * Ensures that the idle task is using init_mm right before its cpu goes
4680 void idle_task_exit(void)
4682 struct mm_struct
*mm
= current
->active_mm
;
4684 BUG_ON(cpu_online(smp_processor_id()));
4686 if (mm
!= &init_mm
) {
4687 switch_mm(mm
, &init_mm
, current
);
4688 finish_arch_post_lock_switch();
4694 * Since this CPU is going 'away' for a while, fold any nr_active delta
4695 * we might have. Assumes we're called after migrate_tasks() so that the
4696 * nr_active count is stable.
4698 * Also see the comment "Global load-average calculations".
4700 static void calc_load_migrate(struct rq
*rq
)
4702 long delta
= calc_load_fold_active(rq
);
4704 atomic_long_add(delta
, &calc_load_tasks
);
4707 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4711 static const struct sched_class fake_sched_class
= {
4712 .put_prev_task
= put_prev_task_fake
,
4715 static struct task_struct fake_task
= {
4717 * Avoid pull_{rt,dl}_task()
4719 .prio
= MAX_PRIO
+ 1,
4720 .sched_class
= &fake_sched_class
,
4724 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4725 * try_to_wake_up()->select_task_rq().
4727 * Called with rq->lock held even though we'er in stop_machine() and
4728 * there's no concurrency possible, we hold the required locks anyway
4729 * because of lock validation efforts.
4731 static void migrate_tasks(unsigned int dead_cpu
)
4733 struct rq
*rq
= cpu_rq(dead_cpu
);
4734 struct task_struct
*next
, *stop
= rq
->stop
;
4738 * Fudge the rq selection such that the below task selection loop
4739 * doesn't get stuck on the currently eligible stop task.
4741 * We're currently inside stop_machine() and the rq is either stuck
4742 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4743 * either way we should never end up calling schedule() until we're
4749 * put_prev_task() and pick_next_task() sched
4750 * class method both need to have an up-to-date
4751 * value of rq->clock[_task]
4753 update_rq_clock(rq
);
4757 * There's this thread running, bail when that's the only
4760 if (rq
->nr_running
== 1)
4763 next
= pick_next_task(rq
, &fake_task
);
4765 next
->sched_class
->put_prev_task(rq
, next
);
4767 /* Find suitable destination for @next, with force if needed. */
4768 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4769 raw_spin_unlock(&rq
->lock
);
4771 __migrate_task(next
, dead_cpu
, dest_cpu
);
4773 raw_spin_lock(&rq
->lock
);
4779 #endif /* CONFIG_HOTPLUG_CPU */
4781 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4783 static struct ctl_table sd_ctl_dir
[] = {
4785 .procname
= "sched_domain",
4791 static struct ctl_table sd_ctl_root
[] = {
4793 .procname
= "kernel",
4795 .child
= sd_ctl_dir
,
4800 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4802 struct ctl_table
*entry
=
4803 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4808 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4810 struct ctl_table
*entry
;
4813 * In the intermediate directories, both the child directory and
4814 * procname are dynamically allocated and could fail but the mode
4815 * will always be set. In the lowest directory the names are
4816 * static strings and all have proc handlers.
4818 for (entry
= *tablep
; entry
->mode
; entry
++) {
4820 sd_free_ctl_entry(&entry
->child
);
4821 if (entry
->proc_handler
== NULL
)
4822 kfree(entry
->procname
);
4829 static int min_load_idx
= 0;
4830 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4833 set_table_entry(struct ctl_table
*entry
,
4834 const char *procname
, void *data
, int maxlen
,
4835 umode_t mode
, proc_handler
*proc_handler
,
4838 entry
->procname
= procname
;
4840 entry
->maxlen
= maxlen
;
4842 entry
->proc_handler
= proc_handler
;
4845 entry
->extra1
= &min_load_idx
;
4846 entry
->extra2
= &max_load_idx
;
4850 static struct ctl_table
*
4851 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4853 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
4858 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4859 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4860 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4861 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4862 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4863 sizeof(int), 0644, proc_dointvec_minmax
, true);
4864 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4865 sizeof(int), 0644, proc_dointvec_minmax
, true);
4866 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4867 sizeof(int), 0644, proc_dointvec_minmax
, true);
4868 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4869 sizeof(int), 0644, proc_dointvec_minmax
, true);
4870 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4871 sizeof(int), 0644, proc_dointvec_minmax
, true);
4872 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4873 sizeof(int), 0644, proc_dointvec_minmax
, false);
4874 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4875 sizeof(int), 0644, proc_dointvec_minmax
, false);
4876 set_table_entry(&table
[9], "cache_nice_tries",
4877 &sd
->cache_nice_tries
,
4878 sizeof(int), 0644, proc_dointvec_minmax
, false);
4879 set_table_entry(&table
[10], "flags", &sd
->flags
,
4880 sizeof(int), 0644, proc_dointvec_minmax
, false);
4881 set_table_entry(&table
[11], "max_newidle_lb_cost",
4882 &sd
->max_newidle_lb_cost
,
4883 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4884 set_table_entry(&table
[12], "name", sd
->name
,
4885 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4886 /* &table[13] is terminator */
4891 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4893 struct ctl_table
*entry
, *table
;
4894 struct sched_domain
*sd
;
4895 int domain_num
= 0, i
;
4898 for_each_domain(cpu
, sd
)
4900 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4905 for_each_domain(cpu
, sd
) {
4906 snprintf(buf
, 32, "domain%d", i
);
4907 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4909 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4916 static struct ctl_table_header
*sd_sysctl_header
;
4917 static void register_sched_domain_sysctl(void)
4919 int i
, cpu_num
= num_possible_cpus();
4920 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4923 WARN_ON(sd_ctl_dir
[0].child
);
4924 sd_ctl_dir
[0].child
= entry
;
4929 for_each_possible_cpu(i
) {
4930 snprintf(buf
, 32, "cpu%d", i
);
4931 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4933 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4937 WARN_ON(sd_sysctl_header
);
4938 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4941 /* may be called multiple times per register */
4942 static void unregister_sched_domain_sysctl(void)
4944 if (sd_sysctl_header
)
4945 unregister_sysctl_table(sd_sysctl_header
);
4946 sd_sysctl_header
= NULL
;
4947 if (sd_ctl_dir
[0].child
)
4948 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4951 static void register_sched_domain_sysctl(void)
4954 static void unregister_sched_domain_sysctl(void)
4959 static void set_rq_online(struct rq
*rq
)
4962 const struct sched_class
*class;
4964 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
4967 for_each_class(class) {
4968 if (class->rq_online
)
4969 class->rq_online(rq
);
4974 static void set_rq_offline(struct rq
*rq
)
4977 const struct sched_class
*class;
4979 for_each_class(class) {
4980 if (class->rq_offline
)
4981 class->rq_offline(rq
);
4984 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
4990 * migration_call - callback that gets triggered when a CPU is added.
4991 * Here we can start up the necessary migration thread for the new CPU.
4994 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
4996 int cpu
= (long)hcpu
;
4997 unsigned long flags
;
4998 struct rq
*rq
= cpu_rq(cpu
);
5000 switch (action
& ~CPU_TASKS_FROZEN
) {
5002 case CPU_UP_PREPARE
:
5003 rq
->calc_load_update
= calc_load_update
;
5007 /* Update our root-domain */
5008 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5010 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5014 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5017 #ifdef CONFIG_HOTPLUG_CPU
5019 sched_ttwu_pending();
5020 /* Update our root-domain */
5021 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5023 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5027 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5028 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5032 calc_load_migrate(rq
);
5037 update_max_interval();
5043 * Register at high priority so that task migration (migrate_all_tasks)
5044 * happens before everything else. This has to be lower priority than
5045 * the notifier in the perf_event subsystem, though.
5047 static struct notifier_block migration_notifier
= {
5048 .notifier_call
= migration_call
,
5049 .priority
= CPU_PRI_MIGRATION
,
5052 static int sched_cpu_active(struct notifier_block
*nfb
,
5053 unsigned long action
, void *hcpu
)
5055 switch (action
& ~CPU_TASKS_FROZEN
) {
5057 case CPU_DOWN_FAILED
:
5058 set_cpu_active((long)hcpu
, true);
5065 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5066 unsigned long action
, void *hcpu
)
5068 unsigned long flags
;
5069 long cpu
= (long)hcpu
;
5071 switch (action
& ~CPU_TASKS_FROZEN
) {
5072 case CPU_DOWN_PREPARE
:
5073 set_cpu_active(cpu
, false);
5075 /* explicitly allow suspend */
5076 if (!(action
& CPU_TASKS_FROZEN
)) {
5077 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
5081 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5082 cpus
= dl_bw_cpus(cpu
);
5083 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5084 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5087 return notifier_from_errno(-EBUSY
);
5095 static int __init
migration_init(void)
5097 void *cpu
= (void *)(long)smp_processor_id();
5100 /* Initialize migration for the boot CPU */
5101 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5102 BUG_ON(err
== NOTIFY_BAD
);
5103 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5104 register_cpu_notifier(&migration_notifier
);
5106 /* Register cpu active notifiers */
5107 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5108 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5112 early_initcall(migration_init
);
5117 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5119 #ifdef CONFIG_SCHED_DEBUG
5121 static __read_mostly
int sched_debug_enabled
;
5123 static int __init
sched_debug_setup(char *str
)
5125 sched_debug_enabled
= 1;
5129 early_param("sched_debug", sched_debug_setup
);
5131 static inline bool sched_debug(void)
5133 return sched_debug_enabled
;
5136 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5137 struct cpumask
*groupmask
)
5139 struct sched_group
*group
= sd
->groups
;
5142 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5143 cpumask_clear(groupmask
);
5145 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5147 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5148 printk("does not load-balance\n");
5150 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5155 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5157 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5158 printk(KERN_ERR
"ERROR: domain->span does not contain "
5161 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5162 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5166 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5170 printk(KERN_ERR
"ERROR: group is NULL\n");
5175 * Even though we initialize ->power to something semi-sane,
5176 * we leave power_orig unset. This allows us to detect if
5177 * domain iteration is still funny without causing /0 traps.
5179 if (!group
->sgp
->power_orig
) {
5180 printk(KERN_CONT
"\n");
5181 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5186 if (!cpumask_weight(sched_group_cpus(group
))) {
5187 printk(KERN_CONT
"\n");
5188 printk(KERN_ERR
"ERROR: empty group\n");
5192 if (!(sd
->flags
& SD_OVERLAP
) &&
5193 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5194 printk(KERN_CONT
"\n");
5195 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5199 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5201 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5203 printk(KERN_CONT
" %s", str
);
5204 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5205 printk(KERN_CONT
" (cpu_power = %d)",
5209 group
= group
->next
;
5210 } while (group
!= sd
->groups
);
5211 printk(KERN_CONT
"\n");
5213 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5214 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5217 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5218 printk(KERN_ERR
"ERROR: parent span is not a superset "
5219 "of domain->span\n");
5223 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5227 if (!sched_debug_enabled
)
5231 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5235 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5238 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5246 #else /* !CONFIG_SCHED_DEBUG */
5247 # define sched_domain_debug(sd, cpu) do { } while (0)
5248 static inline bool sched_debug(void)
5252 #endif /* CONFIG_SCHED_DEBUG */
5254 static int sd_degenerate(struct sched_domain
*sd
)
5256 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5259 /* Following flags need at least 2 groups */
5260 if (sd
->flags
& (SD_LOAD_BALANCE
|
5261 SD_BALANCE_NEWIDLE
|
5265 SD_SHARE_PKG_RESOURCES
)) {
5266 if (sd
->groups
!= sd
->groups
->next
)
5270 /* Following flags don't use groups */
5271 if (sd
->flags
& (SD_WAKE_AFFINE
))
5278 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5280 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5282 if (sd_degenerate(parent
))
5285 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5288 /* Flags needing groups don't count if only 1 group in parent */
5289 if (parent
->groups
== parent
->groups
->next
) {
5290 pflags
&= ~(SD_LOAD_BALANCE
|
5291 SD_BALANCE_NEWIDLE
|
5295 SD_SHARE_PKG_RESOURCES
|
5297 if (nr_node_ids
== 1)
5298 pflags
&= ~SD_SERIALIZE
;
5300 if (~cflags
& pflags
)
5306 static void free_rootdomain(struct rcu_head
*rcu
)
5308 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5310 cpupri_cleanup(&rd
->cpupri
);
5311 cpudl_cleanup(&rd
->cpudl
);
5312 free_cpumask_var(rd
->dlo_mask
);
5313 free_cpumask_var(rd
->rto_mask
);
5314 free_cpumask_var(rd
->online
);
5315 free_cpumask_var(rd
->span
);
5319 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5321 struct root_domain
*old_rd
= NULL
;
5322 unsigned long flags
;
5324 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5329 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5332 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5335 * If we dont want to free the old_rd yet then
5336 * set old_rd to NULL to skip the freeing later
5339 if (!atomic_dec_and_test(&old_rd
->refcount
))
5343 atomic_inc(&rd
->refcount
);
5346 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5347 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5350 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5353 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5356 static int init_rootdomain(struct root_domain
*rd
)
5358 memset(rd
, 0, sizeof(*rd
));
5360 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5362 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5364 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5366 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5369 init_dl_bw(&rd
->dl_bw
);
5370 if (cpudl_init(&rd
->cpudl
) != 0)
5373 if (cpupri_init(&rd
->cpupri
) != 0)
5378 free_cpumask_var(rd
->rto_mask
);
5380 free_cpumask_var(rd
->dlo_mask
);
5382 free_cpumask_var(rd
->online
);
5384 free_cpumask_var(rd
->span
);
5390 * By default the system creates a single root-domain with all cpus as
5391 * members (mimicking the global state we have today).
5393 struct root_domain def_root_domain
;
5395 static void init_defrootdomain(void)
5397 init_rootdomain(&def_root_domain
);
5399 atomic_set(&def_root_domain
.refcount
, 1);
5402 static struct root_domain
*alloc_rootdomain(void)
5404 struct root_domain
*rd
;
5406 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5410 if (init_rootdomain(rd
) != 0) {
5418 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5420 struct sched_group
*tmp
, *first
;
5429 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5434 } while (sg
!= first
);
5437 static void free_sched_domain(struct rcu_head
*rcu
)
5439 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5442 * If its an overlapping domain it has private groups, iterate and
5445 if (sd
->flags
& SD_OVERLAP
) {
5446 free_sched_groups(sd
->groups
, 1);
5447 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5448 kfree(sd
->groups
->sgp
);
5454 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5456 call_rcu(&sd
->rcu
, free_sched_domain
);
5459 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5461 for (; sd
; sd
= sd
->parent
)
5462 destroy_sched_domain(sd
, cpu
);
5466 * Keep a special pointer to the highest sched_domain that has
5467 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5468 * allows us to avoid some pointer chasing select_idle_sibling().
5470 * Also keep a unique ID per domain (we use the first cpu number in
5471 * the cpumask of the domain), this allows us to quickly tell if
5472 * two cpus are in the same cache domain, see cpus_share_cache().
5474 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5475 DEFINE_PER_CPU(int, sd_llc_size
);
5476 DEFINE_PER_CPU(int, sd_llc_id
);
5477 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5478 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5479 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5481 static void update_top_cache_domain(int cpu
)
5483 struct sched_domain
*sd
;
5484 struct sched_domain
*busy_sd
= NULL
;
5488 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5490 id
= cpumask_first(sched_domain_span(sd
));
5491 size
= cpumask_weight(sched_domain_span(sd
));
5492 busy_sd
= sd
->parent
; /* sd_busy */
5494 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5496 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5497 per_cpu(sd_llc_size
, cpu
) = size
;
5498 per_cpu(sd_llc_id
, cpu
) = id
;
5500 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5501 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5503 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5504 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5508 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5509 * hold the hotplug lock.
5512 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5514 struct rq
*rq
= cpu_rq(cpu
);
5515 struct sched_domain
*tmp
;
5517 /* Remove the sched domains which do not contribute to scheduling. */
5518 for (tmp
= sd
; tmp
; ) {
5519 struct sched_domain
*parent
= tmp
->parent
;
5523 if (sd_parent_degenerate(tmp
, parent
)) {
5524 tmp
->parent
= parent
->parent
;
5526 parent
->parent
->child
= tmp
;
5528 * Transfer SD_PREFER_SIBLING down in case of a
5529 * degenerate parent; the spans match for this
5530 * so the property transfers.
5532 if (parent
->flags
& SD_PREFER_SIBLING
)
5533 tmp
->flags
|= SD_PREFER_SIBLING
;
5534 destroy_sched_domain(parent
, cpu
);
5539 if (sd
&& sd_degenerate(sd
)) {
5542 destroy_sched_domain(tmp
, cpu
);
5547 sched_domain_debug(sd
, cpu
);
5549 rq_attach_root(rq
, rd
);
5551 rcu_assign_pointer(rq
->sd
, sd
);
5552 destroy_sched_domains(tmp
, cpu
);
5554 update_top_cache_domain(cpu
);
5557 /* cpus with isolated domains */
5558 static cpumask_var_t cpu_isolated_map
;
5560 /* Setup the mask of cpus configured for isolated domains */
5561 static int __init
isolated_cpu_setup(char *str
)
5563 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5564 cpulist_parse(str
, cpu_isolated_map
);
5568 __setup("isolcpus=", isolated_cpu_setup
);
5570 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5572 return cpumask_of_node(cpu_to_node(cpu
));
5576 struct sched_domain
**__percpu sd
;
5577 struct sched_group
**__percpu sg
;
5578 struct sched_group_power
**__percpu sgp
;
5582 struct sched_domain
** __percpu sd
;
5583 struct root_domain
*rd
;
5593 struct sched_domain_topology_level
;
5595 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5596 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5598 #define SDTL_OVERLAP 0x01
5600 struct sched_domain_topology_level
{
5601 sched_domain_init_f init
;
5602 sched_domain_mask_f mask
;
5605 struct sd_data data
;
5609 * Build an iteration mask that can exclude certain CPUs from the upwards
5612 * Asymmetric node setups can result in situations where the domain tree is of
5613 * unequal depth, make sure to skip domains that already cover the entire
5616 * In that case build_sched_domains() will have terminated the iteration early
5617 * and our sibling sd spans will be empty. Domains should always include the
5618 * cpu they're built on, so check that.
5621 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5623 const struct cpumask
*span
= sched_domain_span(sd
);
5624 struct sd_data
*sdd
= sd
->private;
5625 struct sched_domain
*sibling
;
5628 for_each_cpu(i
, span
) {
5629 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5630 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5633 cpumask_set_cpu(i
, sched_group_mask(sg
));
5638 * Return the canonical balance cpu for this group, this is the first cpu
5639 * of this group that's also in the iteration mask.
5641 int group_balance_cpu(struct sched_group
*sg
)
5643 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5647 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5649 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5650 const struct cpumask
*span
= sched_domain_span(sd
);
5651 struct cpumask
*covered
= sched_domains_tmpmask
;
5652 struct sd_data
*sdd
= sd
->private;
5653 struct sched_domain
*child
;
5656 cpumask_clear(covered
);
5658 for_each_cpu(i
, span
) {
5659 struct cpumask
*sg_span
;
5661 if (cpumask_test_cpu(i
, covered
))
5664 child
= *per_cpu_ptr(sdd
->sd
, i
);
5666 /* See the comment near build_group_mask(). */
5667 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5670 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5671 GFP_KERNEL
, cpu_to_node(cpu
));
5676 sg_span
= sched_group_cpus(sg
);
5678 child
= child
->child
;
5679 cpumask_copy(sg_span
, sched_domain_span(child
));
5681 cpumask_set_cpu(i
, sg_span
);
5683 cpumask_or(covered
, covered
, sg_span
);
5685 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5686 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5687 build_group_mask(sd
, sg
);
5690 * Initialize sgp->power such that even if we mess up the
5691 * domains and no possible iteration will get us here, we won't
5694 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5695 sg
->sgp
->power_orig
= sg
->sgp
->power
;
5698 * Make sure the first group of this domain contains the
5699 * canonical balance cpu. Otherwise the sched_domain iteration
5700 * breaks. See update_sg_lb_stats().
5702 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5703 group_balance_cpu(sg
) == cpu
)
5713 sd
->groups
= groups
;
5718 free_sched_groups(first
, 0);
5723 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5725 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5726 struct sched_domain
*child
= sd
->child
;
5729 cpu
= cpumask_first(sched_domain_span(child
));
5732 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5733 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5734 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5741 * build_sched_groups will build a circular linked list of the groups
5742 * covered by the given span, and will set each group's ->cpumask correctly,
5743 * and ->cpu_power to 0.
5745 * Assumes the sched_domain tree is fully constructed
5748 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5750 struct sched_group
*first
= NULL
, *last
= NULL
;
5751 struct sd_data
*sdd
= sd
->private;
5752 const struct cpumask
*span
= sched_domain_span(sd
);
5753 struct cpumask
*covered
;
5756 get_group(cpu
, sdd
, &sd
->groups
);
5757 atomic_inc(&sd
->groups
->ref
);
5759 if (cpu
!= cpumask_first(span
))
5762 lockdep_assert_held(&sched_domains_mutex
);
5763 covered
= sched_domains_tmpmask
;
5765 cpumask_clear(covered
);
5767 for_each_cpu(i
, span
) {
5768 struct sched_group
*sg
;
5771 if (cpumask_test_cpu(i
, covered
))
5774 group
= get_group(i
, sdd
, &sg
);
5775 cpumask_clear(sched_group_cpus(sg
));
5777 cpumask_setall(sched_group_mask(sg
));
5779 for_each_cpu(j
, span
) {
5780 if (get_group(j
, sdd
, NULL
) != group
)
5783 cpumask_set_cpu(j
, covered
);
5784 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5799 * Initialize sched groups cpu_power.
5801 * cpu_power indicates the capacity of sched group, which is used while
5802 * distributing the load between different sched groups in a sched domain.
5803 * Typically cpu_power for all the groups in a sched domain will be same unless
5804 * there are asymmetries in the topology. If there are asymmetries, group
5805 * having more cpu_power will pickup more load compared to the group having
5808 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5810 struct sched_group
*sg
= sd
->groups
;
5815 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5817 } while (sg
!= sd
->groups
);
5819 if (cpu
!= group_balance_cpu(sg
))
5822 update_group_power(sd
, cpu
);
5823 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5826 int __weak
arch_sd_sibling_asym_packing(void)
5828 return 0*SD_ASYM_PACKING
;
5832 * Initializers for schedule domains
5833 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5836 #ifdef CONFIG_SCHED_DEBUG
5837 # define SD_INIT_NAME(sd, type) sd->name = #type
5839 # define SD_INIT_NAME(sd, type) do { } while (0)
5842 #define SD_INIT_FUNC(type) \
5843 static noinline struct sched_domain * \
5844 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5846 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5847 *sd = SD_##type##_INIT; \
5848 SD_INIT_NAME(sd, type); \
5849 sd->private = &tl->data; \
5854 #ifdef CONFIG_SCHED_SMT
5855 SD_INIT_FUNC(SIBLING
)
5857 #ifdef CONFIG_SCHED_MC
5860 #ifdef CONFIG_SCHED_BOOK
5864 static int default_relax_domain_level
= -1;
5865 int sched_domain_level_max
;
5867 static int __init
setup_relax_domain_level(char *str
)
5869 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5870 pr_warn("Unable to set relax_domain_level\n");
5874 __setup("relax_domain_level=", setup_relax_domain_level
);
5876 static void set_domain_attribute(struct sched_domain
*sd
,
5877 struct sched_domain_attr
*attr
)
5881 if (!attr
|| attr
->relax_domain_level
< 0) {
5882 if (default_relax_domain_level
< 0)
5885 request
= default_relax_domain_level
;
5887 request
= attr
->relax_domain_level
;
5888 if (request
< sd
->level
) {
5889 /* turn off idle balance on this domain */
5890 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5892 /* turn on idle balance on this domain */
5893 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5897 static void __sdt_free(const struct cpumask
*cpu_map
);
5898 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5900 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5901 const struct cpumask
*cpu_map
)
5905 if (!atomic_read(&d
->rd
->refcount
))
5906 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5908 free_percpu(d
->sd
); /* fall through */
5910 __sdt_free(cpu_map
); /* fall through */
5916 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5917 const struct cpumask
*cpu_map
)
5919 memset(d
, 0, sizeof(*d
));
5921 if (__sdt_alloc(cpu_map
))
5922 return sa_sd_storage
;
5923 d
->sd
= alloc_percpu(struct sched_domain
*);
5925 return sa_sd_storage
;
5926 d
->rd
= alloc_rootdomain();
5929 return sa_rootdomain
;
5933 * NULL the sd_data elements we've used to build the sched_domain and
5934 * sched_group structure so that the subsequent __free_domain_allocs()
5935 * will not free the data we're using.
5937 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5939 struct sd_data
*sdd
= sd
->private;
5941 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5942 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5944 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5945 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5947 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5948 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5951 #ifdef CONFIG_SCHED_SMT
5952 static const struct cpumask
*cpu_smt_mask(int cpu
)
5954 return topology_thread_cpumask(cpu
);
5959 * Topology list, bottom-up.
5961 static struct sched_domain_topology_level default_topology
[] = {
5962 #ifdef CONFIG_SCHED_SMT
5963 { sd_init_SIBLING
, cpu_smt_mask
, },
5965 #ifdef CONFIG_SCHED_MC
5966 { sd_init_MC
, cpu_coregroup_mask
, },
5968 #ifdef CONFIG_SCHED_BOOK
5969 { sd_init_BOOK
, cpu_book_mask
, },
5971 { sd_init_CPU
, cpu_cpu_mask
, },
5975 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5977 #define for_each_sd_topology(tl) \
5978 for (tl = sched_domain_topology; tl->init; tl++)
5982 static int sched_domains_numa_levels
;
5983 static int *sched_domains_numa_distance
;
5984 static struct cpumask
***sched_domains_numa_masks
;
5985 static int sched_domains_curr_level
;
5987 static inline int sd_local_flags(int level
)
5989 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5992 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
5995 static struct sched_domain
*
5996 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
5998 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5999 int level
= tl
->numa_level
;
6000 int sd_weight
= cpumask_weight(
6001 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6003 *sd
= (struct sched_domain
){
6004 .min_interval
= sd_weight
,
6005 .max_interval
= 2*sd_weight
,
6007 .imbalance_pct
= 125,
6008 .cache_nice_tries
= 2,
6015 .flags
= 1*SD_LOAD_BALANCE
6016 | 1*SD_BALANCE_NEWIDLE
6021 | 0*SD_SHARE_CPUPOWER
6022 | 0*SD_SHARE_PKG_RESOURCES
6024 | 0*SD_PREFER_SIBLING
6026 | sd_local_flags(level
)
6028 .last_balance
= jiffies
,
6029 .balance_interval
= sd_weight
,
6030 .max_newidle_lb_cost
= 0,
6031 .next_decay_max_lb_cost
= jiffies
,
6033 SD_INIT_NAME(sd
, NUMA
);
6034 sd
->private = &tl
->data
;
6037 * Ugly hack to pass state to sd_numa_mask()...
6039 sched_domains_curr_level
= tl
->numa_level
;
6044 static const struct cpumask
*sd_numa_mask(int cpu
)
6046 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6049 static void sched_numa_warn(const char *str
)
6051 static int done
= false;
6059 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6061 for (i
= 0; i
< nr_node_ids
; i
++) {
6062 printk(KERN_WARNING
" ");
6063 for (j
= 0; j
< nr_node_ids
; j
++)
6064 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6065 printk(KERN_CONT
"\n");
6067 printk(KERN_WARNING
"\n");
6070 static bool find_numa_distance(int distance
)
6074 if (distance
== node_distance(0, 0))
6077 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6078 if (sched_domains_numa_distance
[i
] == distance
)
6085 static void sched_init_numa(void)
6087 int next_distance
, curr_distance
= node_distance(0, 0);
6088 struct sched_domain_topology_level
*tl
;
6092 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6093 if (!sched_domains_numa_distance
)
6097 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6098 * unique distances in the node_distance() table.
6100 * Assumes node_distance(0,j) includes all distances in
6101 * node_distance(i,j) in order to avoid cubic time.
6103 next_distance
= curr_distance
;
6104 for (i
= 0; i
< nr_node_ids
; i
++) {
6105 for (j
= 0; j
< nr_node_ids
; j
++) {
6106 for (k
= 0; k
< nr_node_ids
; k
++) {
6107 int distance
= node_distance(i
, k
);
6109 if (distance
> curr_distance
&&
6110 (distance
< next_distance
||
6111 next_distance
== curr_distance
))
6112 next_distance
= distance
;
6115 * While not a strong assumption it would be nice to know
6116 * about cases where if node A is connected to B, B is not
6117 * equally connected to A.
6119 if (sched_debug() && node_distance(k
, i
) != distance
)
6120 sched_numa_warn("Node-distance not symmetric");
6122 if (sched_debug() && i
&& !find_numa_distance(distance
))
6123 sched_numa_warn("Node-0 not representative");
6125 if (next_distance
!= curr_distance
) {
6126 sched_domains_numa_distance
[level
++] = next_distance
;
6127 sched_domains_numa_levels
= level
;
6128 curr_distance
= next_distance
;
6133 * In case of sched_debug() we verify the above assumption.
6139 * 'level' contains the number of unique distances, excluding the
6140 * identity distance node_distance(i,i).
6142 * The sched_domains_numa_distance[] array includes the actual distance
6147 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6148 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6149 * the array will contain less then 'level' members. This could be
6150 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6151 * in other functions.
6153 * We reset it to 'level' at the end of this function.
6155 sched_domains_numa_levels
= 0;
6157 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6158 if (!sched_domains_numa_masks
)
6162 * Now for each level, construct a mask per node which contains all
6163 * cpus of nodes that are that many hops away from us.
6165 for (i
= 0; i
< level
; i
++) {
6166 sched_domains_numa_masks
[i
] =
6167 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6168 if (!sched_domains_numa_masks
[i
])
6171 for (j
= 0; j
< nr_node_ids
; j
++) {
6172 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6176 sched_domains_numa_masks
[i
][j
] = mask
;
6178 for (k
= 0; k
< nr_node_ids
; k
++) {
6179 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6182 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6187 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6188 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6193 * Copy the default topology bits..
6195 for (i
= 0; default_topology
[i
].init
; i
++)
6196 tl
[i
] = default_topology
[i
];
6199 * .. and append 'j' levels of NUMA goodness.
6201 for (j
= 0; j
< level
; i
++, j
++) {
6202 tl
[i
] = (struct sched_domain_topology_level
){
6203 .init
= sd_numa_init
,
6204 .mask
= sd_numa_mask
,
6205 .flags
= SDTL_OVERLAP
,
6210 sched_domain_topology
= tl
;
6212 sched_domains_numa_levels
= level
;
6215 static void sched_domains_numa_masks_set(int cpu
)
6218 int node
= cpu_to_node(cpu
);
6220 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6221 for (j
= 0; j
< nr_node_ids
; j
++) {
6222 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6223 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6228 static void sched_domains_numa_masks_clear(int cpu
)
6231 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6232 for (j
= 0; j
< nr_node_ids
; j
++)
6233 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6238 * Update sched_domains_numa_masks[level][node] array when new cpus
6241 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6242 unsigned long action
,
6245 int cpu
= (long)hcpu
;
6247 switch (action
& ~CPU_TASKS_FROZEN
) {
6249 sched_domains_numa_masks_set(cpu
);
6253 sched_domains_numa_masks_clear(cpu
);
6263 static inline void sched_init_numa(void)
6267 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6268 unsigned long action
,
6273 #endif /* CONFIG_NUMA */
6275 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6277 struct sched_domain_topology_level
*tl
;
6280 for_each_sd_topology(tl
) {
6281 struct sd_data
*sdd
= &tl
->data
;
6283 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6287 sdd
->sg
= alloc_percpu(struct sched_group
*);
6291 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6295 for_each_cpu(j
, cpu_map
) {
6296 struct sched_domain
*sd
;
6297 struct sched_group
*sg
;
6298 struct sched_group_power
*sgp
;
6300 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6301 GFP_KERNEL
, cpu_to_node(j
));
6305 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6307 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6308 GFP_KERNEL
, cpu_to_node(j
));
6314 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6316 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6317 GFP_KERNEL
, cpu_to_node(j
));
6321 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6328 static void __sdt_free(const struct cpumask
*cpu_map
)
6330 struct sched_domain_topology_level
*tl
;
6333 for_each_sd_topology(tl
) {
6334 struct sd_data
*sdd
= &tl
->data
;
6336 for_each_cpu(j
, cpu_map
) {
6337 struct sched_domain
*sd
;
6340 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6341 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6342 free_sched_groups(sd
->groups
, 0);
6343 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6347 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6349 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6351 free_percpu(sdd
->sd
);
6353 free_percpu(sdd
->sg
);
6355 free_percpu(sdd
->sgp
);
6360 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6361 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6362 struct sched_domain
*child
, int cpu
)
6364 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6368 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6370 sd
->level
= child
->level
+ 1;
6371 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6375 set_domain_attribute(sd
, attr
);
6381 * Build sched domains for a given set of cpus and attach the sched domains
6382 * to the individual cpus
6384 static int build_sched_domains(const struct cpumask
*cpu_map
,
6385 struct sched_domain_attr
*attr
)
6387 enum s_alloc alloc_state
;
6388 struct sched_domain
*sd
;
6390 int i
, ret
= -ENOMEM
;
6392 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6393 if (alloc_state
!= sa_rootdomain
)
6396 /* Set up domains for cpus specified by the cpu_map. */
6397 for_each_cpu(i
, cpu_map
) {
6398 struct sched_domain_topology_level
*tl
;
6401 for_each_sd_topology(tl
) {
6402 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6403 if (tl
== sched_domain_topology
)
6404 *per_cpu_ptr(d
.sd
, i
) = sd
;
6405 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6406 sd
->flags
|= SD_OVERLAP
;
6407 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6412 /* Build the groups for the domains */
6413 for_each_cpu(i
, cpu_map
) {
6414 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6415 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6416 if (sd
->flags
& SD_OVERLAP
) {
6417 if (build_overlap_sched_groups(sd
, i
))
6420 if (build_sched_groups(sd
, i
))
6426 /* Calculate CPU power for physical packages and nodes */
6427 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6428 if (!cpumask_test_cpu(i
, cpu_map
))
6431 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6432 claim_allocations(i
, sd
);
6433 init_sched_groups_power(i
, sd
);
6437 /* Attach the domains */
6439 for_each_cpu(i
, cpu_map
) {
6440 sd
= *per_cpu_ptr(d
.sd
, i
);
6441 cpu_attach_domain(sd
, d
.rd
, i
);
6447 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6451 static cpumask_var_t
*doms_cur
; /* current sched domains */
6452 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6453 static struct sched_domain_attr
*dattr_cur
;
6454 /* attribues of custom domains in 'doms_cur' */
6457 * Special case: If a kmalloc of a doms_cur partition (array of
6458 * cpumask) fails, then fallback to a single sched domain,
6459 * as determined by the single cpumask fallback_doms.
6461 static cpumask_var_t fallback_doms
;
6464 * arch_update_cpu_topology lets virtualized architectures update the
6465 * cpu core maps. It is supposed to return 1 if the topology changed
6466 * or 0 if it stayed the same.
6468 int __weak
arch_update_cpu_topology(void)
6473 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6476 cpumask_var_t
*doms
;
6478 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6481 for (i
= 0; i
< ndoms
; i
++) {
6482 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6483 free_sched_domains(doms
, i
);
6490 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6493 for (i
= 0; i
< ndoms
; i
++)
6494 free_cpumask_var(doms
[i
]);
6499 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6500 * For now this just excludes isolated cpus, but could be used to
6501 * exclude other special cases in the future.
6503 static int init_sched_domains(const struct cpumask
*cpu_map
)
6507 arch_update_cpu_topology();
6509 doms_cur
= alloc_sched_domains(ndoms_cur
);
6511 doms_cur
= &fallback_doms
;
6512 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6513 err
= build_sched_domains(doms_cur
[0], NULL
);
6514 register_sched_domain_sysctl();
6520 * Detach sched domains from a group of cpus specified in cpu_map
6521 * These cpus will now be attached to the NULL domain
6523 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6528 for_each_cpu(i
, cpu_map
)
6529 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6533 /* handle null as "default" */
6534 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6535 struct sched_domain_attr
*new, int idx_new
)
6537 struct sched_domain_attr tmp
;
6544 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6545 new ? (new + idx_new
) : &tmp
,
6546 sizeof(struct sched_domain_attr
));
6550 * Partition sched domains as specified by the 'ndoms_new'
6551 * cpumasks in the array doms_new[] of cpumasks. This compares
6552 * doms_new[] to the current sched domain partitioning, doms_cur[].
6553 * It destroys each deleted domain and builds each new domain.
6555 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6556 * The masks don't intersect (don't overlap.) We should setup one
6557 * sched domain for each mask. CPUs not in any of the cpumasks will
6558 * not be load balanced. If the same cpumask appears both in the
6559 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6562 * The passed in 'doms_new' should be allocated using
6563 * alloc_sched_domains. This routine takes ownership of it and will
6564 * free_sched_domains it when done with it. If the caller failed the
6565 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6566 * and partition_sched_domains() will fallback to the single partition
6567 * 'fallback_doms', it also forces the domains to be rebuilt.
6569 * If doms_new == NULL it will be replaced with cpu_online_mask.
6570 * ndoms_new == 0 is a special case for destroying existing domains,
6571 * and it will not create the default domain.
6573 * Call with hotplug lock held
6575 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6576 struct sched_domain_attr
*dattr_new
)
6581 mutex_lock(&sched_domains_mutex
);
6583 /* always unregister in case we don't destroy any domains */
6584 unregister_sched_domain_sysctl();
6586 /* Let architecture update cpu core mappings. */
6587 new_topology
= arch_update_cpu_topology();
6589 n
= doms_new
? ndoms_new
: 0;
6591 /* Destroy deleted domains */
6592 for (i
= 0; i
< ndoms_cur
; i
++) {
6593 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6594 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6595 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6598 /* no match - a current sched domain not in new doms_new[] */
6599 detach_destroy_domains(doms_cur
[i
]);
6605 if (doms_new
== NULL
) {
6607 doms_new
= &fallback_doms
;
6608 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6609 WARN_ON_ONCE(dattr_new
);
6612 /* Build new domains */
6613 for (i
= 0; i
< ndoms_new
; i
++) {
6614 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6615 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6616 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6619 /* no match - add a new doms_new */
6620 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6625 /* Remember the new sched domains */
6626 if (doms_cur
!= &fallback_doms
)
6627 free_sched_domains(doms_cur
, ndoms_cur
);
6628 kfree(dattr_cur
); /* kfree(NULL) is safe */
6629 doms_cur
= doms_new
;
6630 dattr_cur
= dattr_new
;
6631 ndoms_cur
= ndoms_new
;
6633 register_sched_domain_sysctl();
6635 mutex_unlock(&sched_domains_mutex
);
6638 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6641 * Update cpusets according to cpu_active mask. If cpusets are
6642 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6643 * around partition_sched_domains().
6645 * If we come here as part of a suspend/resume, don't touch cpusets because we
6646 * want to restore it back to its original state upon resume anyway.
6648 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6652 case CPU_ONLINE_FROZEN
:
6653 case CPU_DOWN_FAILED_FROZEN
:
6656 * num_cpus_frozen tracks how many CPUs are involved in suspend
6657 * resume sequence. As long as this is not the last online
6658 * operation in the resume sequence, just build a single sched
6659 * domain, ignoring cpusets.
6662 if (likely(num_cpus_frozen
)) {
6663 partition_sched_domains(1, NULL
, NULL
);
6668 * This is the last CPU online operation. So fall through and
6669 * restore the original sched domains by considering the
6670 * cpuset configurations.
6674 case CPU_DOWN_FAILED
:
6675 cpuset_update_active_cpus(true);
6683 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6687 case CPU_DOWN_PREPARE
:
6688 cpuset_update_active_cpus(false);
6690 case CPU_DOWN_PREPARE_FROZEN
:
6692 partition_sched_domains(1, NULL
, NULL
);
6700 void __init
sched_init_smp(void)
6702 cpumask_var_t non_isolated_cpus
;
6704 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6705 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6710 * There's no userspace yet to cause hotplug operations; hence all the
6711 * cpu masks are stable and all blatant races in the below code cannot
6714 mutex_lock(&sched_domains_mutex
);
6715 init_sched_domains(cpu_active_mask
);
6716 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6717 if (cpumask_empty(non_isolated_cpus
))
6718 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6719 mutex_unlock(&sched_domains_mutex
);
6721 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6722 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6723 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6727 /* Move init over to a non-isolated CPU */
6728 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6730 sched_init_granularity();
6731 free_cpumask_var(non_isolated_cpus
);
6733 init_sched_rt_class();
6734 init_sched_dl_class();
6737 void __init
sched_init_smp(void)
6739 sched_init_granularity();
6741 #endif /* CONFIG_SMP */
6743 const_debug
unsigned int sysctl_timer_migration
= 1;
6745 int in_sched_functions(unsigned long addr
)
6747 return in_lock_functions(addr
) ||
6748 (addr
>= (unsigned long)__sched_text_start
6749 && addr
< (unsigned long)__sched_text_end
);
6752 #ifdef CONFIG_CGROUP_SCHED
6754 * Default task group.
6755 * Every task in system belongs to this group at bootup.
6757 struct task_group root_task_group
;
6758 LIST_HEAD(task_groups
);
6761 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6763 void __init
sched_init(void)
6766 unsigned long alloc_size
= 0, ptr
;
6768 #ifdef CONFIG_FAIR_GROUP_SCHED
6769 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6771 #ifdef CONFIG_RT_GROUP_SCHED
6772 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6774 #ifdef CONFIG_CPUMASK_OFFSTACK
6775 alloc_size
+= num_possible_cpus() * cpumask_size();
6778 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6780 #ifdef CONFIG_FAIR_GROUP_SCHED
6781 root_task_group
.se
= (struct sched_entity
**)ptr
;
6782 ptr
+= nr_cpu_ids
* sizeof(void **);
6784 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6785 ptr
+= nr_cpu_ids
* sizeof(void **);
6787 #endif /* CONFIG_FAIR_GROUP_SCHED */
6788 #ifdef CONFIG_RT_GROUP_SCHED
6789 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6790 ptr
+= nr_cpu_ids
* sizeof(void **);
6792 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6793 ptr
+= nr_cpu_ids
* sizeof(void **);
6795 #endif /* CONFIG_RT_GROUP_SCHED */
6796 #ifdef CONFIG_CPUMASK_OFFSTACK
6797 for_each_possible_cpu(i
) {
6798 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6799 ptr
+= cpumask_size();
6801 #endif /* CONFIG_CPUMASK_OFFSTACK */
6804 init_rt_bandwidth(&def_rt_bandwidth
,
6805 global_rt_period(), global_rt_runtime());
6806 init_dl_bandwidth(&def_dl_bandwidth
,
6807 global_rt_period(), global_rt_runtime());
6810 init_defrootdomain();
6813 #ifdef CONFIG_RT_GROUP_SCHED
6814 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6815 global_rt_period(), global_rt_runtime());
6816 #endif /* CONFIG_RT_GROUP_SCHED */
6818 #ifdef CONFIG_CGROUP_SCHED
6819 list_add(&root_task_group
.list
, &task_groups
);
6820 INIT_LIST_HEAD(&root_task_group
.children
);
6821 INIT_LIST_HEAD(&root_task_group
.siblings
);
6822 autogroup_init(&init_task
);
6824 #endif /* CONFIG_CGROUP_SCHED */
6826 for_each_possible_cpu(i
) {
6830 raw_spin_lock_init(&rq
->lock
);
6832 rq
->calc_load_active
= 0;
6833 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6834 init_cfs_rq(&rq
->cfs
);
6835 init_rt_rq(&rq
->rt
, rq
);
6836 init_dl_rq(&rq
->dl
, rq
);
6837 #ifdef CONFIG_FAIR_GROUP_SCHED
6838 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6839 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6841 * How much cpu bandwidth does root_task_group get?
6843 * In case of task-groups formed thr' the cgroup filesystem, it
6844 * gets 100% of the cpu resources in the system. This overall
6845 * system cpu resource is divided among the tasks of
6846 * root_task_group and its child task-groups in a fair manner,
6847 * based on each entity's (task or task-group's) weight
6848 * (se->load.weight).
6850 * In other words, if root_task_group has 10 tasks of weight
6851 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6852 * then A0's share of the cpu resource is:
6854 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6856 * We achieve this by letting root_task_group's tasks sit
6857 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6859 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6860 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6861 #endif /* CONFIG_FAIR_GROUP_SCHED */
6863 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6864 #ifdef CONFIG_RT_GROUP_SCHED
6865 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6868 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6869 rq
->cpu_load
[j
] = 0;
6871 rq
->last_load_update_tick
= jiffies
;
6876 rq
->cpu_power
= SCHED_POWER_SCALE
;
6877 rq
->post_schedule
= 0;
6878 rq
->active_balance
= 0;
6879 rq
->next_balance
= jiffies
;
6884 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6885 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6887 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6889 rq_attach_root(rq
, &def_root_domain
);
6890 #ifdef CONFIG_NO_HZ_COMMON
6893 #ifdef CONFIG_NO_HZ_FULL
6894 rq
->last_sched_tick
= 0;
6898 atomic_set(&rq
->nr_iowait
, 0);
6901 set_load_weight(&init_task
);
6903 #ifdef CONFIG_PREEMPT_NOTIFIERS
6904 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6908 * The boot idle thread does lazy MMU switching as well:
6910 atomic_inc(&init_mm
.mm_count
);
6911 enter_lazy_tlb(&init_mm
, current
);
6914 * Make us the idle thread. Technically, schedule() should not be
6915 * called from this thread, however somewhere below it might be,
6916 * but because we are the idle thread, we just pick up running again
6917 * when this runqueue becomes "idle".
6919 init_idle(current
, smp_processor_id());
6921 calc_load_update
= jiffies
+ LOAD_FREQ
;
6924 * During early bootup we pretend to be a normal task:
6926 current
->sched_class
= &fair_sched_class
;
6929 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6930 /* May be allocated at isolcpus cmdline parse time */
6931 if (cpu_isolated_map
== NULL
)
6932 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6933 idle_thread_set_boot_cpu();
6935 init_sched_fair_class();
6937 scheduler_running
= 1;
6940 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6941 static inline int preempt_count_equals(int preempt_offset
)
6943 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6945 return (nested
== preempt_offset
);
6948 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6950 static unsigned long prev_jiffy
; /* ratelimiting */
6952 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6953 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6954 !is_idle_task(current
)) ||
6955 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6957 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6959 prev_jiffy
= jiffies
;
6962 "BUG: sleeping function called from invalid context at %s:%d\n",
6965 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6966 in_atomic(), irqs_disabled(),
6967 current
->pid
, current
->comm
);
6969 debug_show_held_locks(current
);
6970 if (irqs_disabled())
6971 print_irqtrace_events(current
);
6972 #ifdef CONFIG_DEBUG_PREEMPT
6973 if (!preempt_count_equals(preempt_offset
)) {
6974 pr_err("Preemption disabled at:");
6975 print_ip_sym(current
->preempt_disable_ip
);
6981 EXPORT_SYMBOL(__might_sleep
);
6984 #ifdef CONFIG_MAGIC_SYSRQ
6985 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6987 const struct sched_class
*prev_class
= p
->sched_class
;
6988 struct sched_attr attr
= {
6989 .sched_policy
= SCHED_NORMAL
,
6991 int old_prio
= p
->prio
;
6996 dequeue_task(rq
, p
, 0);
6997 __setscheduler(rq
, p
, &attr
);
6999 enqueue_task(rq
, p
, 0);
7000 resched_task(rq
->curr
);
7003 check_class_changed(rq
, p
, prev_class
, old_prio
);
7006 void normalize_rt_tasks(void)
7008 struct task_struct
*g
, *p
;
7009 unsigned long flags
;
7012 read_lock_irqsave(&tasklist_lock
, flags
);
7013 do_each_thread(g
, p
) {
7015 * Only normalize user tasks:
7020 p
->se
.exec_start
= 0;
7021 #ifdef CONFIG_SCHEDSTATS
7022 p
->se
.statistics
.wait_start
= 0;
7023 p
->se
.statistics
.sleep_start
= 0;
7024 p
->se
.statistics
.block_start
= 0;
7027 if (!dl_task(p
) && !rt_task(p
)) {
7029 * Renice negative nice level userspace
7032 if (task_nice(p
) < 0 && p
->mm
)
7033 set_user_nice(p
, 0);
7037 raw_spin_lock(&p
->pi_lock
);
7038 rq
= __task_rq_lock(p
);
7040 normalize_task(rq
, p
);
7042 __task_rq_unlock(rq
);
7043 raw_spin_unlock(&p
->pi_lock
);
7044 } while_each_thread(g
, p
);
7046 read_unlock_irqrestore(&tasklist_lock
, flags
);
7049 #endif /* CONFIG_MAGIC_SYSRQ */
7051 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7053 * These functions are only useful for the IA64 MCA handling, or kdb.
7055 * They can only be called when the whole system has been
7056 * stopped - every CPU needs to be quiescent, and no scheduling
7057 * activity can take place. Using them for anything else would
7058 * be a serious bug, and as a result, they aren't even visible
7059 * under any other configuration.
7063 * curr_task - return the current task for a given cpu.
7064 * @cpu: the processor in question.
7066 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7068 * Return: The current task for @cpu.
7070 struct task_struct
*curr_task(int cpu
)
7072 return cpu_curr(cpu
);
7075 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7079 * set_curr_task - set the current task for a given cpu.
7080 * @cpu: the processor in question.
7081 * @p: the task pointer to set.
7083 * Description: This function must only be used when non-maskable interrupts
7084 * are serviced on a separate stack. It allows the architecture to switch the
7085 * notion of the current task on a cpu in a non-blocking manner. This function
7086 * must be called with all CPU's synchronized, and interrupts disabled, the
7087 * and caller must save the original value of the current task (see
7088 * curr_task() above) and restore that value before reenabling interrupts and
7089 * re-starting the system.
7091 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7093 void set_curr_task(int cpu
, struct task_struct
*p
)
7100 #ifdef CONFIG_CGROUP_SCHED
7101 /* task_group_lock serializes the addition/removal of task groups */
7102 static DEFINE_SPINLOCK(task_group_lock
);
7104 static void free_sched_group(struct task_group
*tg
)
7106 free_fair_sched_group(tg
);
7107 free_rt_sched_group(tg
);
7112 /* allocate runqueue etc for a new task group */
7113 struct task_group
*sched_create_group(struct task_group
*parent
)
7115 struct task_group
*tg
;
7117 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7119 return ERR_PTR(-ENOMEM
);
7121 if (!alloc_fair_sched_group(tg
, parent
))
7124 if (!alloc_rt_sched_group(tg
, parent
))
7130 free_sched_group(tg
);
7131 return ERR_PTR(-ENOMEM
);
7134 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7136 unsigned long flags
;
7138 spin_lock_irqsave(&task_group_lock
, flags
);
7139 list_add_rcu(&tg
->list
, &task_groups
);
7141 WARN_ON(!parent
); /* root should already exist */
7143 tg
->parent
= parent
;
7144 INIT_LIST_HEAD(&tg
->children
);
7145 list_add_rcu(&tg
->siblings
, &parent
->children
);
7146 spin_unlock_irqrestore(&task_group_lock
, flags
);
7149 /* rcu callback to free various structures associated with a task group */
7150 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7152 /* now it should be safe to free those cfs_rqs */
7153 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7156 /* Destroy runqueue etc associated with a task group */
7157 void sched_destroy_group(struct task_group
*tg
)
7159 /* wait for possible concurrent references to cfs_rqs complete */
7160 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7163 void sched_offline_group(struct task_group
*tg
)
7165 unsigned long flags
;
7168 /* end participation in shares distribution */
7169 for_each_possible_cpu(i
)
7170 unregister_fair_sched_group(tg
, i
);
7172 spin_lock_irqsave(&task_group_lock
, flags
);
7173 list_del_rcu(&tg
->list
);
7174 list_del_rcu(&tg
->siblings
);
7175 spin_unlock_irqrestore(&task_group_lock
, flags
);
7178 /* change task's runqueue when it moves between groups.
7179 * The caller of this function should have put the task in its new group
7180 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7181 * reflect its new group.
7183 void sched_move_task(struct task_struct
*tsk
)
7185 struct task_group
*tg
;
7187 unsigned long flags
;
7190 rq
= task_rq_lock(tsk
, &flags
);
7192 running
= task_current(rq
, tsk
);
7196 dequeue_task(rq
, tsk
, 0);
7197 if (unlikely(running
))
7198 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7200 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
,
7201 lockdep_is_held(&tsk
->sighand
->siglock
)),
7202 struct task_group
, css
);
7203 tg
= autogroup_task_group(tsk
, tg
);
7204 tsk
->sched_task_group
= tg
;
7206 #ifdef CONFIG_FAIR_GROUP_SCHED
7207 if (tsk
->sched_class
->task_move_group
)
7208 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7211 set_task_rq(tsk
, task_cpu(tsk
));
7213 if (unlikely(running
))
7214 tsk
->sched_class
->set_curr_task(rq
);
7216 enqueue_task(rq
, tsk
, 0);
7218 task_rq_unlock(rq
, tsk
, &flags
);
7220 #endif /* CONFIG_CGROUP_SCHED */
7222 #ifdef CONFIG_RT_GROUP_SCHED
7224 * Ensure that the real time constraints are schedulable.
7226 static DEFINE_MUTEX(rt_constraints_mutex
);
7228 /* Must be called with tasklist_lock held */
7229 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7231 struct task_struct
*g
, *p
;
7233 do_each_thread(g
, p
) {
7234 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7236 } while_each_thread(g
, p
);
7241 struct rt_schedulable_data
{
7242 struct task_group
*tg
;
7247 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7249 struct rt_schedulable_data
*d
= data
;
7250 struct task_group
*child
;
7251 unsigned long total
, sum
= 0;
7252 u64 period
, runtime
;
7254 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7255 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7258 period
= d
->rt_period
;
7259 runtime
= d
->rt_runtime
;
7263 * Cannot have more runtime than the period.
7265 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7269 * Ensure we don't starve existing RT tasks.
7271 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7274 total
= to_ratio(period
, runtime
);
7277 * Nobody can have more than the global setting allows.
7279 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7283 * The sum of our children's runtime should not exceed our own.
7285 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7286 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7287 runtime
= child
->rt_bandwidth
.rt_runtime
;
7289 if (child
== d
->tg
) {
7290 period
= d
->rt_period
;
7291 runtime
= d
->rt_runtime
;
7294 sum
+= to_ratio(period
, runtime
);
7303 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7307 struct rt_schedulable_data data
= {
7309 .rt_period
= period
,
7310 .rt_runtime
= runtime
,
7314 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7320 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7321 u64 rt_period
, u64 rt_runtime
)
7325 mutex_lock(&rt_constraints_mutex
);
7326 read_lock(&tasklist_lock
);
7327 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7331 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7332 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7333 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7335 for_each_possible_cpu(i
) {
7336 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7338 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7339 rt_rq
->rt_runtime
= rt_runtime
;
7340 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7342 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7344 read_unlock(&tasklist_lock
);
7345 mutex_unlock(&rt_constraints_mutex
);
7350 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7352 u64 rt_runtime
, rt_period
;
7354 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7355 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7356 if (rt_runtime_us
< 0)
7357 rt_runtime
= RUNTIME_INF
;
7359 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7362 static long sched_group_rt_runtime(struct task_group
*tg
)
7366 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7369 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7370 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7371 return rt_runtime_us
;
7374 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7376 u64 rt_runtime
, rt_period
;
7378 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7379 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7384 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7387 static long sched_group_rt_period(struct task_group
*tg
)
7391 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7392 do_div(rt_period_us
, NSEC_PER_USEC
);
7393 return rt_period_us
;
7395 #endif /* CONFIG_RT_GROUP_SCHED */
7397 #ifdef CONFIG_RT_GROUP_SCHED
7398 static int sched_rt_global_constraints(void)
7402 mutex_lock(&rt_constraints_mutex
);
7403 read_lock(&tasklist_lock
);
7404 ret
= __rt_schedulable(NULL
, 0, 0);
7405 read_unlock(&tasklist_lock
);
7406 mutex_unlock(&rt_constraints_mutex
);
7411 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7413 /* Don't accept realtime tasks when there is no way for them to run */
7414 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7420 #else /* !CONFIG_RT_GROUP_SCHED */
7421 static int sched_rt_global_constraints(void)
7423 unsigned long flags
;
7426 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7427 for_each_possible_cpu(i
) {
7428 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7430 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7431 rt_rq
->rt_runtime
= global_rt_runtime();
7432 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7434 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7438 #endif /* CONFIG_RT_GROUP_SCHED */
7440 static int sched_dl_global_constraints(void)
7442 u64 runtime
= global_rt_runtime();
7443 u64 period
= global_rt_period();
7444 u64 new_bw
= to_ratio(period
, runtime
);
7446 unsigned long flags
;
7449 * Here we want to check the bandwidth not being set to some
7450 * value smaller than the currently allocated bandwidth in
7451 * any of the root_domains.
7453 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7454 * cycling on root_domains... Discussion on different/better
7455 * solutions is welcome!
7457 for_each_possible_cpu(cpu
) {
7458 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7460 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7461 if (new_bw
< dl_b
->total_bw
)
7463 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7472 static void sched_dl_do_global(void)
7476 unsigned long flags
;
7478 def_dl_bandwidth
.dl_period
= global_rt_period();
7479 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7481 if (global_rt_runtime() != RUNTIME_INF
)
7482 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7485 * FIXME: As above...
7487 for_each_possible_cpu(cpu
) {
7488 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7490 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7492 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7496 static int sched_rt_global_validate(void)
7498 if (sysctl_sched_rt_period
<= 0)
7501 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7502 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7508 static void sched_rt_do_global(void)
7510 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7511 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7514 int sched_rt_handler(struct ctl_table
*table
, int write
,
7515 void __user
*buffer
, size_t *lenp
,
7518 int old_period
, old_runtime
;
7519 static DEFINE_MUTEX(mutex
);
7523 old_period
= sysctl_sched_rt_period
;
7524 old_runtime
= sysctl_sched_rt_runtime
;
7526 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7528 if (!ret
&& write
) {
7529 ret
= sched_rt_global_validate();
7533 ret
= sched_rt_global_constraints();
7537 ret
= sched_dl_global_constraints();
7541 sched_rt_do_global();
7542 sched_dl_do_global();
7546 sysctl_sched_rt_period
= old_period
;
7547 sysctl_sched_rt_runtime
= old_runtime
;
7549 mutex_unlock(&mutex
);
7554 int sched_rr_handler(struct ctl_table
*table
, int write
,
7555 void __user
*buffer
, size_t *lenp
,
7559 static DEFINE_MUTEX(mutex
);
7562 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7563 /* make sure that internally we keep jiffies */
7564 /* also, writing zero resets timeslice to default */
7565 if (!ret
&& write
) {
7566 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7567 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7569 mutex_unlock(&mutex
);
7573 #ifdef CONFIG_CGROUP_SCHED
7575 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7577 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7580 static struct cgroup_subsys_state
*
7581 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7583 struct task_group
*parent
= css_tg(parent_css
);
7584 struct task_group
*tg
;
7587 /* This is early initialization for the top cgroup */
7588 return &root_task_group
.css
;
7591 tg
= sched_create_group(parent
);
7593 return ERR_PTR(-ENOMEM
);
7598 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7600 struct task_group
*tg
= css_tg(css
);
7601 struct task_group
*parent
= css_tg(css_parent(css
));
7604 sched_online_group(tg
, parent
);
7608 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7610 struct task_group
*tg
= css_tg(css
);
7612 sched_destroy_group(tg
);
7615 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7617 struct task_group
*tg
= css_tg(css
);
7619 sched_offline_group(tg
);
7622 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7623 struct cgroup_taskset
*tset
)
7625 struct task_struct
*task
;
7627 cgroup_taskset_for_each(task
, tset
) {
7628 #ifdef CONFIG_RT_GROUP_SCHED
7629 if (!sched_rt_can_attach(css_tg(css
), task
))
7632 /* We don't support RT-tasks being in separate groups */
7633 if (task
->sched_class
!= &fair_sched_class
)
7640 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7641 struct cgroup_taskset
*tset
)
7643 struct task_struct
*task
;
7645 cgroup_taskset_for_each(task
, tset
)
7646 sched_move_task(task
);
7649 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7650 struct cgroup_subsys_state
*old_css
,
7651 struct task_struct
*task
)
7654 * cgroup_exit() is called in the copy_process() failure path.
7655 * Ignore this case since the task hasn't ran yet, this avoids
7656 * trying to poke a half freed task state from generic code.
7658 if (!(task
->flags
& PF_EXITING
))
7661 sched_move_task(task
);
7664 #ifdef CONFIG_FAIR_GROUP_SCHED
7665 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7666 struct cftype
*cftype
, u64 shareval
)
7668 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7671 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7674 struct task_group
*tg
= css_tg(css
);
7676 return (u64
) scale_load_down(tg
->shares
);
7679 #ifdef CONFIG_CFS_BANDWIDTH
7680 static DEFINE_MUTEX(cfs_constraints_mutex
);
7682 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7683 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7685 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7687 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7689 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7690 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7692 if (tg
== &root_task_group
)
7696 * Ensure we have at some amount of bandwidth every period. This is
7697 * to prevent reaching a state of large arrears when throttled via
7698 * entity_tick() resulting in prolonged exit starvation.
7700 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7704 * Likewise, bound things on the otherside by preventing insane quota
7705 * periods. This also allows us to normalize in computing quota
7708 if (period
> max_cfs_quota_period
)
7711 mutex_lock(&cfs_constraints_mutex
);
7712 ret
= __cfs_schedulable(tg
, period
, quota
);
7716 runtime_enabled
= quota
!= RUNTIME_INF
;
7717 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7719 * If we need to toggle cfs_bandwidth_used, off->on must occur
7720 * before making related changes, and on->off must occur afterwards
7722 if (runtime_enabled
&& !runtime_was_enabled
)
7723 cfs_bandwidth_usage_inc();
7724 raw_spin_lock_irq(&cfs_b
->lock
);
7725 cfs_b
->period
= ns_to_ktime(period
);
7726 cfs_b
->quota
= quota
;
7728 __refill_cfs_bandwidth_runtime(cfs_b
);
7729 /* restart the period timer (if active) to handle new period expiry */
7730 if (runtime_enabled
&& cfs_b
->timer_active
) {
7731 /* force a reprogram */
7732 cfs_b
->timer_active
= 0;
7733 __start_cfs_bandwidth(cfs_b
);
7735 raw_spin_unlock_irq(&cfs_b
->lock
);
7737 for_each_possible_cpu(i
) {
7738 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7739 struct rq
*rq
= cfs_rq
->rq
;
7741 raw_spin_lock_irq(&rq
->lock
);
7742 cfs_rq
->runtime_enabled
= runtime_enabled
;
7743 cfs_rq
->runtime_remaining
= 0;
7745 if (cfs_rq
->throttled
)
7746 unthrottle_cfs_rq(cfs_rq
);
7747 raw_spin_unlock_irq(&rq
->lock
);
7749 if (runtime_was_enabled
&& !runtime_enabled
)
7750 cfs_bandwidth_usage_dec();
7752 mutex_unlock(&cfs_constraints_mutex
);
7757 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7761 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7762 if (cfs_quota_us
< 0)
7763 quota
= RUNTIME_INF
;
7765 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7767 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7770 long tg_get_cfs_quota(struct task_group
*tg
)
7774 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7777 quota_us
= tg
->cfs_bandwidth
.quota
;
7778 do_div(quota_us
, NSEC_PER_USEC
);
7783 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7787 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7788 quota
= tg
->cfs_bandwidth
.quota
;
7790 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7793 long tg_get_cfs_period(struct task_group
*tg
)
7797 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7798 do_div(cfs_period_us
, NSEC_PER_USEC
);
7800 return cfs_period_us
;
7803 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7806 return tg_get_cfs_quota(css_tg(css
));
7809 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7810 struct cftype
*cftype
, s64 cfs_quota_us
)
7812 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7815 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7818 return tg_get_cfs_period(css_tg(css
));
7821 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7822 struct cftype
*cftype
, u64 cfs_period_us
)
7824 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7827 struct cfs_schedulable_data
{
7828 struct task_group
*tg
;
7833 * normalize group quota/period to be quota/max_period
7834 * note: units are usecs
7836 static u64
normalize_cfs_quota(struct task_group
*tg
,
7837 struct cfs_schedulable_data
*d
)
7845 period
= tg_get_cfs_period(tg
);
7846 quota
= tg_get_cfs_quota(tg
);
7849 /* note: these should typically be equivalent */
7850 if (quota
== RUNTIME_INF
|| quota
== -1)
7853 return to_ratio(period
, quota
);
7856 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7858 struct cfs_schedulable_data
*d
= data
;
7859 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7860 s64 quota
= 0, parent_quota
= -1;
7863 quota
= RUNTIME_INF
;
7865 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7867 quota
= normalize_cfs_quota(tg
, d
);
7868 parent_quota
= parent_b
->hierarchal_quota
;
7871 * ensure max(child_quota) <= parent_quota, inherit when no
7874 if (quota
== RUNTIME_INF
)
7875 quota
= parent_quota
;
7876 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7879 cfs_b
->hierarchal_quota
= quota
;
7884 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7887 struct cfs_schedulable_data data
= {
7893 if (quota
!= RUNTIME_INF
) {
7894 do_div(data
.period
, NSEC_PER_USEC
);
7895 do_div(data
.quota
, NSEC_PER_USEC
);
7899 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7905 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7907 struct task_group
*tg
= css_tg(seq_css(sf
));
7908 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7910 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7911 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7912 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7916 #endif /* CONFIG_CFS_BANDWIDTH */
7917 #endif /* CONFIG_FAIR_GROUP_SCHED */
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7921 struct cftype
*cft
, s64 val
)
7923 return sched_group_set_rt_runtime(css_tg(css
), val
);
7926 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7929 return sched_group_rt_runtime(css_tg(css
));
7932 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7933 struct cftype
*cftype
, u64 rt_period_us
)
7935 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7938 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7941 return sched_group_rt_period(css_tg(css
));
7943 #endif /* CONFIG_RT_GROUP_SCHED */
7945 static struct cftype cpu_files
[] = {
7946 #ifdef CONFIG_FAIR_GROUP_SCHED
7949 .read_u64
= cpu_shares_read_u64
,
7950 .write_u64
= cpu_shares_write_u64
,
7953 #ifdef CONFIG_CFS_BANDWIDTH
7955 .name
= "cfs_quota_us",
7956 .read_s64
= cpu_cfs_quota_read_s64
,
7957 .write_s64
= cpu_cfs_quota_write_s64
,
7960 .name
= "cfs_period_us",
7961 .read_u64
= cpu_cfs_period_read_u64
,
7962 .write_u64
= cpu_cfs_period_write_u64
,
7966 .seq_show
= cpu_stats_show
,
7969 #ifdef CONFIG_RT_GROUP_SCHED
7971 .name
= "rt_runtime_us",
7972 .read_s64
= cpu_rt_runtime_read
,
7973 .write_s64
= cpu_rt_runtime_write
,
7976 .name
= "rt_period_us",
7977 .read_u64
= cpu_rt_period_read_uint
,
7978 .write_u64
= cpu_rt_period_write_uint
,
7984 struct cgroup_subsys cpu_cgrp_subsys
= {
7985 .css_alloc
= cpu_cgroup_css_alloc
,
7986 .css_free
= cpu_cgroup_css_free
,
7987 .css_online
= cpu_cgroup_css_online
,
7988 .css_offline
= cpu_cgroup_css_offline
,
7989 .can_attach
= cpu_cgroup_can_attach
,
7990 .attach
= cpu_cgroup_attach
,
7991 .exit
= cpu_cgroup_exit
,
7992 .base_cftypes
= cpu_files
,
7996 #endif /* CONFIG_CGROUP_SCHED */
7998 void dump_cpu_task(int cpu
)
8000 pr_info("Task dump for CPU %d:\n", cpu
);
8001 sched_show_task(cpu_curr(cpu
));