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 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic
);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic
);
109 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
112 ktime_t soft
, hard
, now
;
115 if (hrtimer_active(period_timer
))
118 now
= hrtimer_cb_get_time(period_timer
);
119 hrtimer_forward(period_timer
, now
, period
);
121 soft
= hrtimer_get_softexpires(period_timer
);
122 hard
= hrtimer_get_expires(period_timer
);
123 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
124 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
125 HRTIMER_MODE_ABS_PINNED
, 0);
129 DEFINE_MUTEX(sched_domains_mutex
);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
132 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
134 void update_rq_clock(struct rq
*rq
)
138 if (rq
->skip_clock_update
> 0)
141 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
143 update_rq_clock_task(rq
, delta
);
147 * Debugging: various feature bits
150 #define SCHED_FEAT(name, enabled) \
151 (1UL << __SCHED_FEAT_##name) * enabled |
153 const_debug
unsigned int sysctl_sched_features
=
154 #include "features.h"
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled) \
163 static const char * const sched_feat_names
[] = {
164 #include "features.h"
169 static int sched_feat_show(struct seq_file
*m
, void *v
)
173 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
174 if (!(sysctl_sched_features
& (1UL << i
)))
176 seq_printf(m
, "%s ", sched_feat_names
[i
]);
183 #ifdef HAVE_JUMP_LABEL
185 #define jump_label_key__true STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
188 #define SCHED_FEAT(name, enabled) \
189 jump_label_key__##enabled ,
191 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
192 #include "features.h"
197 static void sched_feat_disable(int i
)
199 if (static_key_enabled(&sched_feat_keys
[i
]))
200 static_key_slow_dec(&sched_feat_keys
[i
]);
203 static void sched_feat_enable(int i
)
205 if (!static_key_enabled(&sched_feat_keys
[i
]))
206 static_key_slow_inc(&sched_feat_keys
[i
]);
209 static void sched_feat_disable(int i
) { };
210 static void sched_feat_enable(int i
) { };
211 #endif /* HAVE_JUMP_LABEL */
213 static int sched_feat_set(char *cmp
)
218 if (strncmp(cmp
, "NO_", 3) == 0) {
223 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
224 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
226 sysctl_sched_features
&= ~(1UL << i
);
227 sched_feat_disable(i
);
229 sysctl_sched_features
|= (1UL << i
);
230 sched_feat_enable(i
);
240 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
241 size_t cnt
, loff_t
*ppos
)
250 if (copy_from_user(&buf
, ubuf
, cnt
))
256 i
= sched_feat_set(cmp
);
257 if (i
== __SCHED_FEAT_NR
)
265 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
267 return single_open(filp
, sched_feat_show
, NULL
);
270 static const struct file_operations sched_feat_fops
= {
271 .open
= sched_feat_open
,
272 .write
= sched_feat_write
,
275 .release
= single_release
,
278 static __init
int sched_init_debug(void)
280 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
285 late_initcall(sched_init_debug
);
286 #endif /* CONFIG_SCHED_DEBUG */
289 * Number of tasks to iterate in a single balance run.
290 * Limited because this is done with IRQs disabled.
292 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
295 * period over which we average the RT time consumption, measured
300 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
303 * period over which we measure -rt task cpu usage in us.
306 unsigned int sysctl_sched_rt_period
= 1000000;
308 __read_mostly
int scheduler_running
;
311 * part of the period that we allow rt tasks to run in us.
314 int sysctl_sched_rt_runtime
= 950000;
317 * __task_rq_lock - lock the rq @p resides on.
319 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
324 lockdep_assert_held(&p
->pi_lock
);
328 raw_spin_lock(&rq
->lock
);
329 if (likely(rq
== task_rq(p
)))
331 raw_spin_unlock(&rq
->lock
);
336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
338 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
339 __acquires(p
->pi_lock
)
345 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
347 raw_spin_lock(&rq
->lock
);
348 if (likely(rq
== task_rq(p
)))
350 raw_spin_unlock(&rq
->lock
);
351 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
355 static void __task_rq_unlock(struct rq
*rq
)
358 raw_spin_unlock(&rq
->lock
);
362 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
364 __releases(p
->pi_lock
)
366 raw_spin_unlock(&rq
->lock
);
367 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
371 * this_rq_lock - lock this runqueue and disable interrupts.
373 static struct rq
*this_rq_lock(void)
380 raw_spin_lock(&rq
->lock
);
385 #ifdef CONFIG_SCHED_HRTICK
387 * Use HR-timers to deliver accurate preemption points.
390 static void hrtick_clear(struct rq
*rq
)
392 if (hrtimer_active(&rq
->hrtick_timer
))
393 hrtimer_cancel(&rq
->hrtick_timer
);
397 * High-resolution timer tick.
398 * Runs from hardirq context with interrupts disabled.
400 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
402 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
404 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
406 raw_spin_lock(&rq
->lock
);
408 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
409 raw_spin_unlock(&rq
->lock
);
411 return HRTIMER_NORESTART
;
416 static int __hrtick_restart(struct rq
*rq
)
418 struct hrtimer
*timer
= &rq
->hrtick_timer
;
419 ktime_t time
= hrtimer_get_softexpires(timer
);
421 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
425 * called from hardirq (IPI) context
427 static void __hrtick_start(void *arg
)
431 raw_spin_lock(&rq
->lock
);
432 __hrtick_restart(rq
);
433 rq
->hrtick_csd_pending
= 0;
434 raw_spin_unlock(&rq
->lock
);
438 * Called to set the hrtick timer state.
440 * called with rq->lock held and irqs disabled
442 void hrtick_start(struct rq
*rq
, u64 delay
)
444 struct hrtimer
*timer
= &rq
->hrtick_timer
;
445 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
447 hrtimer_set_expires(timer
, time
);
449 if (rq
== this_rq()) {
450 __hrtick_restart(rq
);
451 } else if (!rq
->hrtick_csd_pending
) {
452 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
453 rq
->hrtick_csd_pending
= 1;
458 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
460 int cpu
= (int)(long)hcpu
;
463 case CPU_UP_CANCELED
:
464 case CPU_UP_CANCELED_FROZEN
:
465 case CPU_DOWN_PREPARE
:
466 case CPU_DOWN_PREPARE_FROZEN
:
468 case CPU_DEAD_FROZEN
:
469 hrtick_clear(cpu_rq(cpu
));
476 static __init
void init_hrtick(void)
478 hotcpu_notifier(hotplug_hrtick
, 0);
482 * Called to set the hrtick timer state.
484 * called with rq->lock held and irqs disabled
486 void hrtick_start(struct rq
*rq
, u64 delay
)
488 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
489 HRTIMER_MODE_REL_PINNED
, 0);
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SMP */
497 static void init_rq_hrtick(struct rq
*rq
)
500 rq
->hrtick_csd_pending
= 0;
502 rq
->hrtick_csd
.flags
= 0;
503 rq
->hrtick_csd
.func
= __hrtick_start
;
504 rq
->hrtick_csd
.info
= rq
;
507 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
508 rq
->hrtick_timer
.function
= hrtick
;
510 #else /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq
*rq
)
515 static inline void init_rq_hrtick(struct rq
*rq
)
519 static inline void init_hrtick(void)
522 #endif /* CONFIG_SCHED_HRTICK */
525 * resched_task - mark a task 'to be rescheduled now'.
527 * On UP this means the setting of the need_resched flag, on SMP it
528 * might also involve a cross-CPU call to trigger the scheduler on
531 void resched_task(struct task_struct
*p
)
535 lockdep_assert_held(&task_rq(p
)->lock
);
537 if (test_tsk_need_resched(p
))
540 set_tsk_need_resched(p
);
543 if (cpu
== smp_processor_id()) {
544 set_preempt_need_resched();
548 /* NEED_RESCHED must be visible before we test polling */
550 if (!tsk_is_polling(p
))
551 smp_send_reschedule(cpu
);
554 void resched_cpu(int cpu
)
556 struct rq
*rq
= cpu_rq(cpu
);
559 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
561 resched_task(cpu_curr(cpu
));
562 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
566 #ifdef CONFIG_NO_HZ_COMMON
568 * In the semi idle case, use the nearest busy cpu for migrating timers
569 * from an idle cpu. This is good for power-savings.
571 * We don't do similar optimization for completely idle system, as
572 * selecting an idle cpu will add more delays to the timers than intended
573 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
575 int get_nohz_timer_target(int pinned
)
577 int cpu
= smp_processor_id();
579 struct sched_domain
*sd
;
581 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
585 for_each_domain(cpu
, sd
) {
586 for_each_cpu(i
, sched_domain_span(sd
)) {
598 * When add_timer_on() enqueues a timer into the timer wheel of an
599 * idle CPU then this timer might expire before the next timer event
600 * which is scheduled to wake up that CPU. In case of a completely
601 * idle system the next event might even be infinite time into the
602 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
603 * leaves the inner idle loop so the newly added timer is taken into
604 * account when the CPU goes back to idle and evaluates the timer
605 * wheel for the next timer event.
607 static void wake_up_idle_cpu(int cpu
)
609 struct rq
*rq
= cpu_rq(cpu
);
611 if (cpu
== smp_processor_id())
615 * This is safe, as this function is called with the timer
616 * wheel base lock of (cpu) held. When the CPU is on the way
617 * to idle and has not yet set rq->curr to idle then it will
618 * be serialized on the timer wheel base lock and take the new
619 * timer into account automatically.
621 if (rq
->curr
!= rq
->idle
)
625 * We can set TIF_RESCHED on the idle task of the other CPU
626 * lockless. The worst case is that the other CPU runs the
627 * idle task through an additional NOOP schedule()
629 set_tsk_need_resched(rq
->idle
);
631 /* NEED_RESCHED must be visible before we test polling */
633 if (!tsk_is_polling(rq
->idle
))
634 smp_send_reschedule(cpu
);
637 static bool wake_up_full_nohz_cpu(int cpu
)
639 if (tick_nohz_full_cpu(cpu
)) {
640 if (cpu
!= smp_processor_id() ||
641 tick_nohz_tick_stopped())
642 smp_send_reschedule(cpu
);
649 void wake_up_nohz_cpu(int cpu
)
651 if (!wake_up_full_nohz_cpu(cpu
))
652 wake_up_idle_cpu(cpu
);
655 static inline bool got_nohz_idle_kick(void)
657 int cpu
= smp_processor_id();
659 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
662 if (idle_cpu(cpu
) && !need_resched())
666 * We can't run Idle Load Balance on this CPU for this time so we
667 * cancel it and clear NOHZ_BALANCE_KICK
669 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
673 #else /* CONFIG_NO_HZ_COMMON */
675 static inline bool got_nohz_idle_kick(void)
680 #endif /* CONFIG_NO_HZ_COMMON */
682 #ifdef CONFIG_NO_HZ_FULL
683 bool sched_can_stop_tick(void)
689 /* Make sure rq->nr_running update is visible after the IPI */
692 /* More than one running task need preemption */
693 if (rq
->nr_running
> 1)
698 #endif /* CONFIG_NO_HZ_FULL */
700 void sched_avg_update(struct rq
*rq
)
702 s64 period
= sched_avg_period();
704 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
706 * Inline assembly required to prevent the compiler
707 * optimising this loop into a divmod call.
708 * See __iter_div_u64_rem() for another example of this.
710 asm("" : "+rm" (rq
->age_stamp
));
711 rq
->age_stamp
+= period
;
716 #endif /* CONFIG_SMP */
718 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
719 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
721 * Iterate task_group tree rooted at *from, calling @down when first entering a
722 * node and @up when leaving it for the final time.
724 * Caller must hold rcu_lock or sufficient equivalent.
726 int walk_tg_tree_from(struct task_group
*from
,
727 tg_visitor down
, tg_visitor up
, void *data
)
729 struct task_group
*parent
, *child
;
735 ret
= (*down
)(parent
, data
);
738 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
745 ret
= (*up
)(parent
, data
);
746 if (ret
|| parent
== from
)
750 parent
= parent
->parent
;
757 int tg_nop(struct task_group
*tg
, void *data
)
763 static void set_load_weight(struct task_struct
*p
)
765 int prio
= p
->static_prio
- MAX_RT_PRIO
;
766 struct load_weight
*load
= &p
->se
.load
;
769 * SCHED_IDLE tasks get minimal weight:
771 if (p
->policy
== SCHED_IDLE
) {
772 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
773 load
->inv_weight
= WMULT_IDLEPRIO
;
777 load
->weight
= scale_load(prio_to_weight
[prio
]);
778 load
->inv_weight
= prio_to_wmult
[prio
];
781 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
784 sched_info_queued(rq
, p
);
785 p
->sched_class
->enqueue_task(rq
, p
, flags
);
788 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
791 sched_info_dequeued(rq
, p
);
792 p
->sched_class
->dequeue_task(rq
, p
, flags
);
795 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
797 if (task_contributes_to_load(p
))
798 rq
->nr_uninterruptible
--;
800 enqueue_task(rq
, p
, flags
);
803 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
805 if (task_contributes_to_load(p
))
806 rq
->nr_uninterruptible
++;
808 dequeue_task(rq
, p
, flags
);
811 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
814 * In theory, the compile should just see 0 here, and optimize out the call
815 * to sched_rt_avg_update. But I don't trust it...
817 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
818 s64 steal
= 0, irq_delta
= 0;
820 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
821 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
824 * Since irq_time is only updated on {soft,}irq_exit, we might run into
825 * this case when a previous update_rq_clock() happened inside a
828 * When this happens, we stop ->clock_task and only update the
829 * prev_irq_time stamp to account for the part that fit, so that a next
830 * update will consume the rest. This ensures ->clock_task is
833 * It does however cause some slight miss-attribution of {soft,}irq
834 * time, a more accurate solution would be to update the irq_time using
835 * the current rq->clock timestamp, except that would require using
838 if (irq_delta
> delta
)
841 rq
->prev_irq_time
+= irq_delta
;
844 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
845 if (static_key_false((¶virt_steal_rq_enabled
))) {
846 steal
= paravirt_steal_clock(cpu_of(rq
));
847 steal
-= rq
->prev_steal_time_rq
;
849 if (unlikely(steal
> delta
))
852 rq
->prev_steal_time_rq
+= steal
;
857 rq
->clock_task
+= delta
;
859 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
860 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
861 sched_rt_avg_update(rq
, irq_delta
+ steal
);
865 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
867 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
868 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
872 * Make it appear like a SCHED_FIFO task, its something
873 * userspace knows about and won't get confused about.
875 * Also, it will make PI more or less work without too
876 * much confusion -- but then, stop work should not
877 * rely on PI working anyway.
879 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
881 stop
->sched_class
= &stop_sched_class
;
884 cpu_rq(cpu
)->stop
= stop
;
888 * Reset it back to a normal scheduling class so that
889 * it can die in pieces.
891 old_stop
->sched_class
= &rt_sched_class
;
896 * __normal_prio - return the priority that is based on the static prio
898 static inline int __normal_prio(struct task_struct
*p
)
900 return p
->static_prio
;
904 * Calculate the expected normal priority: i.e. priority
905 * without taking RT-inheritance into account. Might be
906 * boosted by interactivity modifiers. Changes upon fork,
907 * setprio syscalls, and whenever the interactivity
908 * estimator recalculates.
910 static inline int normal_prio(struct task_struct
*p
)
914 if (task_has_dl_policy(p
))
915 prio
= MAX_DL_PRIO
-1;
916 else if (task_has_rt_policy(p
))
917 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
919 prio
= __normal_prio(p
);
924 * Calculate the current priority, i.e. the priority
925 * taken into account by the scheduler. This value might
926 * be boosted by RT tasks, or might be boosted by
927 * interactivity modifiers. Will be RT if the task got
928 * RT-boosted. If not then it returns p->normal_prio.
930 static int effective_prio(struct task_struct
*p
)
932 p
->normal_prio
= normal_prio(p
);
934 * If we are RT tasks or we were boosted to RT priority,
935 * keep the priority unchanged. Otherwise, update priority
936 * to the normal priority:
938 if (!rt_prio(p
->prio
))
939 return p
->normal_prio
;
944 * task_curr - is this task currently executing on a CPU?
945 * @p: the task in question.
947 * Return: 1 if the task is currently executing. 0 otherwise.
949 inline int task_curr(const struct task_struct
*p
)
951 return cpu_curr(task_cpu(p
)) == p
;
954 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
955 const struct sched_class
*prev_class
,
958 if (prev_class
!= p
->sched_class
) {
959 if (prev_class
->switched_from
)
960 prev_class
->switched_from(rq
, p
);
961 p
->sched_class
->switched_to(rq
, p
);
962 } else if (oldprio
!= p
->prio
|| dl_task(p
))
963 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
966 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
968 const struct sched_class
*class;
970 if (p
->sched_class
== rq
->curr
->sched_class
) {
971 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
973 for_each_class(class) {
974 if (class == rq
->curr
->sched_class
)
976 if (class == p
->sched_class
) {
977 resched_task(rq
->curr
);
984 * A queue event has occurred, and we're going to schedule. In
985 * this case, we can save a useless back to back clock update.
987 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
988 rq
->skip_clock_update
= 1;
992 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
994 #ifdef CONFIG_SCHED_DEBUG
996 * We should never call set_task_cpu() on a blocked task,
997 * ttwu() will sort out the placement.
999 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1000 !(task_preempt_count(p
) & PREEMPT_ACTIVE
));
1002 #ifdef CONFIG_LOCKDEP
1004 * The caller should hold either p->pi_lock or rq->lock, when changing
1005 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1007 * sched_move_task() holds both and thus holding either pins the cgroup,
1010 * Furthermore, all task_rq users should acquire both locks, see
1013 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1014 lockdep_is_held(&task_rq(p
)->lock
)));
1018 trace_sched_migrate_task(p
, new_cpu
);
1020 if (task_cpu(p
) != new_cpu
) {
1021 if (p
->sched_class
->migrate_task_rq
)
1022 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1023 p
->se
.nr_migrations
++;
1024 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1027 __set_task_cpu(p
, new_cpu
);
1030 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1033 struct rq
*src_rq
, *dst_rq
;
1035 src_rq
= task_rq(p
);
1036 dst_rq
= cpu_rq(cpu
);
1038 deactivate_task(src_rq
, p
, 0);
1039 set_task_cpu(p
, cpu
);
1040 activate_task(dst_rq
, p
, 0);
1041 check_preempt_curr(dst_rq
, p
, 0);
1044 * Task isn't running anymore; make it appear like we migrated
1045 * it before it went to sleep. This means on wakeup we make the
1046 * previous cpu our targer instead of where it really is.
1052 struct migration_swap_arg
{
1053 struct task_struct
*src_task
, *dst_task
;
1054 int src_cpu
, dst_cpu
;
1057 static int migrate_swap_stop(void *data
)
1059 struct migration_swap_arg
*arg
= data
;
1060 struct rq
*src_rq
, *dst_rq
;
1063 src_rq
= cpu_rq(arg
->src_cpu
);
1064 dst_rq
= cpu_rq(arg
->dst_cpu
);
1066 double_raw_lock(&arg
->src_task
->pi_lock
,
1067 &arg
->dst_task
->pi_lock
);
1068 double_rq_lock(src_rq
, dst_rq
);
1069 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1072 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1075 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1078 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1081 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1082 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1087 double_rq_unlock(src_rq
, dst_rq
);
1088 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1089 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1095 * Cross migrate two tasks
1097 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1099 struct migration_swap_arg arg
;
1102 arg
= (struct migration_swap_arg
){
1104 .src_cpu
= task_cpu(cur
),
1106 .dst_cpu
= task_cpu(p
),
1109 if (arg
.src_cpu
== arg
.dst_cpu
)
1113 * These three tests are all lockless; this is OK since all of them
1114 * will be re-checked with proper locks held further down the line.
1116 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1119 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1122 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1125 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1126 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1132 struct migration_arg
{
1133 struct task_struct
*task
;
1137 static int migration_cpu_stop(void *data
);
1140 * wait_task_inactive - wait for a thread to unschedule.
1142 * If @match_state is nonzero, it's the @p->state value just checked and
1143 * not expected to change. If it changes, i.e. @p might have woken up,
1144 * then return zero. When we succeed in waiting for @p to be off its CPU,
1145 * we return a positive number (its total switch count). If a second call
1146 * a short while later returns the same number, the caller can be sure that
1147 * @p has remained unscheduled the whole time.
1149 * The caller must ensure that the task *will* unschedule sometime soon,
1150 * else this function might spin for a *long* time. This function can't
1151 * be called with interrupts off, or it may introduce deadlock with
1152 * smp_call_function() if an IPI is sent by the same process we are
1153 * waiting to become inactive.
1155 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1157 unsigned long flags
;
1164 * We do the initial early heuristics without holding
1165 * any task-queue locks at all. We'll only try to get
1166 * the runqueue lock when things look like they will
1172 * If the task is actively running on another CPU
1173 * still, just relax and busy-wait without holding
1176 * NOTE! Since we don't hold any locks, it's not
1177 * even sure that "rq" stays as the right runqueue!
1178 * But we don't care, since "task_running()" will
1179 * return false if the runqueue has changed and p
1180 * is actually now running somewhere else!
1182 while (task_running(rq
, p
)) {
1183 if (match_state
&& unlikely(p
->state
!= match_state
))
1189 * Ok, time to look more closely! We need the rq
1190 * lock now, to be *sure*. If we're wrong, we'll
1191 * just go back and repeat.
1193 rq
= task_rq_lock(p
, &flags
);
1194 trace_sched_wait_task(p
);
1195 running
= task_running(rq
, p
);
1198 if (!match_state
|| p
->state
== match_state
)
1199 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1200 task_rq_unlock(rq
, p
, &flags
);
1203 * If it changed from the expected state, bail out now.
1205 if (unlikely(!ncsw
))
1209 * Was it really running after all now that we
1210 * checked with the proper locks actually held?
1212 * Oops. Go back and try again..
1214 if (unlikely(running
)) {
1220 * It's not enough that it's not actively running,
1221 * it must be off the runqueue _entirely_, and not
1224 * So if it was still runnable (but just not actively
1225 * running right now), it's preempted, and we should
1226 * yield - it could be a while.
1228 if (unlikely(on_rq
)) {
1229 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1231 set_current_state(TASK_UNINTERRUPTIBLE
);
1232 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1237 * Ahh, all good. It wasn't running, and it wasn't
1238 * runnable, which means that it will never become
1239 * running in the future either. We're all done!
1248 * kick_process - kick a running thread to enter/exit the kernel
1249 * @p: the to-be-kicked thread
1251 * Cause a process which is running on another CPU to enter
1252 * kernel-mode, without any delay. (to get signals handled.)
1254 * NOTE: this function doesn't have to take the runqueue lock,
1255 * because all it wants to ensure is that the remote task enters
1256 * the kernel. If the IPI races and the task has been migrated
1257 * to another CPU then no harm is done and the purpose has been
1260 void kick_process(struct task_struct
*p
)
1266 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1267 smp_send_reschedule(cpu
);
1270 EXPORT_SYMBOL_GPL(kick_process
);
1271 #endif /* CONFIG_SMP */
1275 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1277 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1279 int nid
= cpu_to_node(cpu
);
1280 const struct cpumask
*nodemask
= NULL
;
1281 enum { cpuset
, possible
, fail
} state
= cpuset
;
1285 * If the node that the cpu is on has been offlined, cpu_to_node()
1286 * will return -1. There is no cpu on the node, and we should
1287 * select the cpu on the other node.
1290 nodemask
= cpumask_of_node(nid
);
1292 /* Look for allowed, online CPU in same node. */
1293 for_each_cpu(dest_cpu
, nodemask
) {
1294 if (!cpu_online(dest_cpu
))
1296 if (!cpu_active(dest_cpu
))
1298 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1304 /* Any allowed, online CPU? */
1305 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1306 if (!cpu_online(dest_cpu
))
1308 if (!cpu_active(dest_cpu
))
1315 /* No more Mr. Nice Guy. */
1316 cpuset_cpus_allowed_fallback(p
);
1321 do_set_cpus_allowed(p
, cpu_possible_mask
);
1332 if (state
!= cpuset
) {
1334 * Don't tell them about moving exiting tasks or
1335 * kernel threads (both mm NULL), since they never
1338 if (p
->mm
&& printk_ratelimit()) {
1339 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1340 task_pid_nr(p
), p
->comm
, cpu
);
1348 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1351 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1353 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1356 * In order not to call set_task_cpu() on a blocking task we need
1357 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1360 * Since this is common to all placement strategies, this lives here.
1362 * [ this allows ->select_task() to simply return task_cpu(p) and
1363 * not worry about this generic constraint ]
1365 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1367 cpu
= select_fallback_rq(task_cpu(p
), p
);
1372 static void update_avg(u64
*avg
, u64 sample
)
1374 s64 diff
= sample
- *avg
;
1380 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1382 #ifdef CONFIG_SCHEDSTATS
1383 struct rq
*rq
= this_rq();
1386 int this_cpu
= smp_processor_id();
1388 if (cpu
== this_cpu
) {
1389 schedstat_inc(rq
, ttwu_local
);
1390 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1392 struct sched_domain
*sd
;
1394 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1396 for_each_domain(this_cpu
, sd
) {
1397 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1398 schedstat_inc(sd
, ttwu_wake_remote
);
1405 if (wake_flags
& WF_MIGRATED
)
1406 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1408 #endif /* CONFIG_SMP */
1410 schedstat_inc(rq
, ttwu_count
);
1411 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1413 if (wake_flags
& WF_SYNC
)
1414 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1416 #endif /* CONFIG_SCHEDSTATS */
1419 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1421 activate_task(rq
, p
, en_flags
);
1424 /* if a worker is waking up, notify workqueue */
1425 if (p
->flags
& PF_WQ_WORKER
)
1426 wq_worker_waking_up(p
, cpu_of(rq
));
1430 * Mark the task runnable and perform wakeup-preemption.
1433 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1435 check_preempt_curr(rq
, p
, wake_flags
);
1436 trace_sched_wakeup(p
, true);
1438 p
->state
= TASK_RUNNING
;
1440 if (p
->sched_class
->task_woken
)
1441 p
->sched_class
->task_woken(rq
, p
);
1443 if (rq
->idle_stamp
) {
1444 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1445 u64 max
= 2*rq
->max_idle_balance_cost
;
1447 update_avg(&rq
->avg_idle
, delta
);
1449 if (rq
->avg_idle
> max
)
1458 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1461 if (p
->sched_contributes_to_load
)
1462 rq
->nr_uninterruptible
--;
1465 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1466 ttwu_do_wakeup(rq
, p
, wake_flags
);
1470 * Called in case the task @p isn't fully descheduled from its runqueue,
1471 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1472 * since all we need to do is flip p->state to TASK_RUNNING, since
1473 * the task is still ->on_rq.
1475 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1480 rq
= __task_rq_lock(p
);
1482 /* check_preempt_curr() may use rq clock */
1483 update_rq_clock(rq
);
1484 ttwu_do_wakeup(rq
, p
, wake_flags
);
1487 __task_rq_unlock(rq
);
1493 static void sched_ttwu_pending(void)
1495 struct rq
*rq
= this_rq();
1496 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1497 struct task_struct
*p
;
1499 raw_spin_lock(&rq
->lock
);
1502 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1503 llist
= llist_next(llist
);
1504 ttwu_do_activate(rq
, p
, 0);
1507 raw_spin_unlock(&rq
->lock
);
1510 void scheduler_ipi(void)
1513 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1514 * TIF_NEED_RESCHED remotely (for the first time) will also send
1517 preempt_fold_need_resched();
1519 if (llist_empty(&this_rq()->wake_list
)
1520 && !tick_nohz_full_cpu(smp_processor_id())
1521 && !got_nohz_idle_kick())
1525 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1526 * traditionally all their work was done from the interrupt return
1527 * path. Now that we actually do some work, we need to make sure
1530 * Some archs already do call them, luckily irq_enter/exit nest
1533 * Arguably we should visit all archs and update all handlers,
1534 * however a fair share of IPIs are still resched only so this would
1535 * somewhat pessimize the simple resched case.
1538 tick_nohz_full_check();
1539 sched_ttwu_pending();
1542 * Check if someone kicked us for doing the nohz idle load balance.
1544 if (unlikely(got_nohz_idle_kick())) {
1545 this_rq()->idle_balance
= 1;
1546 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1551 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1553 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1554 smp_send_reschedule(cpu
);
1557 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1559 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1561 #endif /* CONFIG_SMP */
1563 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1565 struct rq
*rq
= cpu_rq(cpu
);
1567 #if defined(CONFIG_SMP)
1568 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1569 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1570 ttwu_queue_remote(p
, cpu
);
1575 raw_spin_lock(&rq
->lock
);
1576 ttwu_do_activate(rq
, p
, 0);
1577 raw_spin_unlock(&rq
->lock
);
1581 * try_to_wake_up - wake up a thread
1582 * @p: the thread to be awakened
1583 * @state: the mask of task states that can be woken
1584 * @wake_flags: wake modifier flags (WF_*)
1586 * Put it on the run-queue if it's not already there. The "current"
1587 * thread is always on the run-queue (except when the actual
1588 * re-schedule is in progress), and as such you're allowed to do
1589 * the simpler "current->state = TASK_RUNNING" to mark yourself
1590 * runnable without the overhead of this.
1592 * Return: %true if @p was woken up, %false if it was already running.
1593 * or @state didn't match @p's state.
1596 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1598 unsigned long flags
;
1599 int cpu
, success
= 0;
1602 * If we are going to wake up a thread waiting for CONDITION we
1603 * need to ensure that CONDITION=1 done by the caller can not be
1604 * reordered with p->state check below. This pairs with mb() in
1605 * set_current_state() the waiting thread does.
1607 smp_mb__before_spinlock();
1608 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1609 if (!(p
->state
& state
))
1612 success
= 1; /* we're going to change ->state */
1615 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1620 * If the owning (remote) cpu is still in the middle of schedule() with
1621 * this task as prev, wait until its done referencing the task.
1626 * Pairs with the smp_wmb() in finish_lock_switch().
1630 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1631 p
->state
= TASK_WAKING
;
1633 if (p
->sched_class
->task_waking
)
1634 p
->sched_class
->task_waking(p
);
1636 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1637 if (task_cpu(p
) != cpu
) {
1638 wake_flags
|= WF_MIGRATED
;
1639 set_task_cpu(p
, cpu
);
1641 #endif /* CONFIG_SMP */
1645 ttwu_stat(p
, cpu
, wake_flags
);
1647 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1653 * try_to_wake_up_local - try to wake up a local task with rq lock held
1654 * @p: the thread to be awakened
1656 * Put @p on the run-queue if it's not already there. The caller must
1657 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1660 static void try_to_wake_up_local(struct task_struct
*p
)
1662 struct rq
*rq
= task_rq(p
);
1664 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1665 WARN_ON_ONCE(p
== current
))
1668 lockdep_assert_held(&rq
->lock
);
1670 if (!raw_spin_trylock(&p
->pi_lock
)) {
1671 raw_spin_unlock(&rq
->lock
);
1672 raw_spin_lock(&p
->pi_lock
);
1673 raw_spin_lock(&rq
->lock
);
1676 if (!(p
->state
& TASK_NORMAL
))
1680 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1682 ttwu_do_wakeup(rq
, p
, 0);
1683 ttwu_stat(p
, smp_processor_id(), 0);
1685 raw_spin_unlock(&p
->pi_lock
);
1689 * wake_up_process - Wake up a specific process
1690 * @p: The process to be woken up.
1692 * Attempt to wake up the nominated process and move it to the set of runnable
1695 * Return: 1 if the process was woken up, 0 if it was already running.
1697 * It may be assumed that this function implies a write memory barrier before
1698 * changing the task state if and only if any tasks are woken up.
1700 int wake_up_process(struct task_struct
*p
)
1702 WARN_ON(task_is_stopped_or_traced(p
));
1703 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1705 EXPORT_SYMBOL(wake_up_process
);
1707 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1709 return try_to_wake_up(p
, state
, 0);
1713 * Perform scheduler related setup for a newly forked process p.
1714 * p is forked by current.
1716 * __sched_fork() is basic setup used by init_idle() too:
1718 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1723 p
->se
.exec_start
= 0;
1724 p
->se
.sum_exec_runtime
= 0;
1725 p
->se
.prev_sum_exec_runtime
= 0;
1726 p
->se
.nr_migrations
= 0;
1728 INIT_LIST_HEAD(&p
->se
.group_node
);
1730 #ifdef CONFIG_SCHEDSTATS
1731 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1734 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1735 hrtimer_init(&p
->dl
.dl_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1736 p
->dl
.dl_runtime
= p
->dl
.runtime
= 0;
1737 p
->dl
.dl_deadline
= p
->dl
.deadline
= 0;
1738 p
->dl
.dl_period
= 0;
1741 INIT_LIST_HEAD(&p
->rt
.run_list
);
1743 #ifdef CONFIG_PREEMPT_NOTIFIERS
1744 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1747 #ifdef CONFIG_NUMA_BALANCING
1748 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1749 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1750 p
->mm
->numa_scan_seq
= 0;
1753 if (clone_flags
& CLONE_VM
)
1754 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1756 p
->numa_preferred_nid
= -1;
1758 p
->node_stamp
= 0ULL;
1759 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1760 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1761 p
->numa_work
.next
= &p
->numa_work
;
1762 p
->numa_faults_memory
= NULL
;
1763 p
->numa_faults_buffer_memory
= NULL
;
1764 p
->last_task_numa_placement
= 0;
1765 p
->last_sum_exec_runtime
= 0;
1767 INIT_LIST_HEAD(&p
->numa_entry
);
1768 p
->numa_group
= NULL
;
1769 #endif /* CONFIG_NUMA_BALANCING */
1772 #ifdef CONFIG_NUMA_BALANCING
1773 #ifdef CONFIG_SCHED_DEBUG
1774 void set_numabalancing_state(bool enabled
)
1777 sched_feat_set("NUMA");
1779 sched_feat_set("NO_NUMA");
1782 __read_mostly
bool numabalancing_enabled
;
1784 void set_numabalancing_state(bool enabled
)
1786 numabalancing_enabled
= enabled
;
1788 #endif /* CONFIG_SCHED_DEBUG */
1790 #ifdef CONFIG_PROC_SYSCTL
1791 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1792 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1796 int state
= numabalancing_enabled
;
1798 if (write
&& !capable(CAP_SYS_ADMIN
))
1803 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1807 set_numabalancing_state(state
);
1814 * fork()/clone()-time setup:
1816 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1818 unsigned long flags
;
1819 int cpu
= get_cpu();
1821 __sched_fork(clone_flags
, p
);
1823 * We mark the process as running here. This guarantees that
1824 * nobody will actually run it, and a signal or other external
1825 * event cannot wake it up and insert it on the runqueue either.
1827 p
->state
= TASK_RUNNING
;
1830 * Make sure we do not leak PI boosting priority to the child.
1832 p
->prio
= current
->normal_prio
;
1835 * Revert to default priority/policy on fork if requested.
1837 if (unlikely(p
->sched_reset_on_fork
)) {
1838 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1839 p
->policy
= SCHED_NORMAL
;
1840 p
->static_prio
= NICE_TO_PRIO(0);
1842 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1843 p
->static_prio
= NICE_TO_PRIO(0);
1845 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1849 * We don't need the reset flag anymore after the fork. It has
1850 * fulfilled its duty:
1852 p
->sched_reset_on_fork
= 0;
1855 if (dl_prio(p
->prio
)) {
1858 } else if (rt_prio(p
->prio
)) {
1859 p
->sched_class
= &rt_sched_class
;
1861 p
->sched_class
= &fair_sched_class
;
1864 if (p
->sched_class
->task_fork
)
1865 p
->sched_class
->task_fork(p
);
1868 * The child is not yet in the pid-hash so no cgroup attach races,
1869 * and the cgroup is pinned to this child due to cgroup_fork()
1870 * is ran before sched_fork().
1872 * Silence PROVE_RCU.
1874 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1875 set_task_cpu(p
, cpu
);
1876 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1878 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1879 if (likely(sched_info_on()))
1880 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1882 #if defined(CONFIG_SMP)
1885 init_task_preempt_count(p
);
1887 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1888 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1895 unsigned long to_ratio(u64 period
, u64 runtime
)
1897 if (runtime
== RUNTIME_INF
)
1901 * Doing this here saves a lot of checks in all
1902 * the calling paths, and returning zero seems
1903 * safe for them anyway.
1908 return div64_u64(runtime
<< 20, period
);
1912 inline struct dl_bw
*dl_bw_of(int i
)
1914 return &cpu_rq(i
)->rd
->dl_bw
;
1917 static inline int dl_bw_cpus(int i
)
1919 struct root_domain
*rd
= cpu_rq(i
)->rd
;
1922 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
1928 inline struct dl_bw
*dl_bw_of(int i
)
1930 return &cpu_rq(i
)->dl
.dl_bw
;
1933 static inline int dl_bw_cpus(int i
)
1940 void __dl_clear(struct dl_bw
*dl_b
, u64 tsk_bw
)
1942 dl_b
->total_bw
-= tsk_bw
;
1946 void __dl_add(struct dl_bw
*dl_b
, u64 tsk_bw
)
1948 dl_b
->total_bw
+= tsk_bw
;
1952 bool __dl_overflow(struct dl_bw
*dl_b
, int cpus
, u64 old_bw
, u64 new_bw
)
1954 return dl_b
->bw
!= -1 &&
1955 dl_b
->bw
* cpus
< dl_b
->total_bw
- old_bw
+ new_bw
;
1959 * We must be sure that accepting a new task (or allowing changing the
1960 * parameters of an existing one) is consistent with the bandwidth
1961 * constraints. If yes, this function also accordingly updates the currently
1962 * allocated bandwidth to reflect the new situation.
1964 * This function is called while holding p's rq->lock.
1966 static int dl_overflow(struct task_struct
*p
, int policy
,
1967 const struct sched_attr
*attr
)
1970 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1971 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
1972 u64 runtime
= attr
->sched_runtime
;
1973 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
1976 if (new_bw
== p
->dl
.dl_bw
)
1980 * Either if a task, enters, leave, or stays -deadline but changes
1981 * its parameters, we may need to update accordingly the total
1982 * allocated bandwidth of the container.
1984 raw_spin_lock(&dl_b
->lock
);
1985 cpus
= dl_bw_cpus(task_cpu(p
));
1986 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
1987 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
1988 __dl_add(dl_b
, new_bw
);
1990 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
1991 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
1992 __dl_clear(dl_b
, p
->dl
.dl_bw
);
1993 __dl_add(dl_b
, new_bw
);
1995 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
1996 __dl_clear(dl_b
, p
->dl
.dl_bw
);
1999 raw_spin_unlock(&dl_b
->lock
);
2004 extern void init_dl_bw(struct dl_bw
*dl_b
);
2007 * wake_up_new_task - wake up a newly created task for the first time.
2009 * This function will do some initial scheduler statistics housekeeping
2010 * that must be done for every newly created context, then puts the task
2011 * on the runqueue and wakes it.
2013 void wake_up_new_task(struct task_struct
*p
)
2015 unsigned long flags
;
2018 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2021 * Fork balancing, do it here and not earlier because:
2022 * - cpus_allowed can change in the fork path
2023 * - any previously selected cpu might disappear through hotplug
2025 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2028 /* Initialize new task's runnable average */
2029 init_task_runnable_average(p
);
2030 rq
= __task_rq_lock(p
);
2031 activate_task(rq
, p
, 0);
2033 trace_sched_wakeup_new(p
, true);
2034 check_preempt_curr(rq
, p
, WF_FORK
);
2036 if (p
->sched_class
->task_woken
)
2037 p
->sched_class
->task_woken(rq
, p
);
2039 task_rq_unlock(rq
, p
, &flags
);
2042 #ifdef CONFIG_PREEMPT_NOTIFIERS
2045 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2046 * @notifier: notifier struct to register
2048 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2050 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2052 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2055 * preempt_notifier_unregister - no longer interested in preemption notifications
2056 * @notifier: notifier struct to unregister
2058 * This is safe to call from within a preemption notifier.
2060 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2062 hlist_del(¬ifier
->link
);
2064 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2066 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2068 struct preempt_notifier
*notifier
;
2070 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2071 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2075 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2076 struct task_struct
*next
)
2078 struct preempt_notifier
*notifier
;
2080 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2081 notifier
->ops
->sched_out(notifier
, next
);
2084 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2086 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2091 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2092 struct task_struct
*next
)
2096 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2099 * prepare_task_switch - prepare to switch tasks
2100 * @rq: the runqueue preparing to switch
2101 * @prev: the current task that is being switched out
2102 * @next: the task we are going to switch to.
2104 * This is called with the rq lock held and interrupts off. It must
2105 * be paired with a subsequent finish_task_switch after the context
2108 * prepare_task_switch sets up locking and calls architecture specific
2112 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2113 struct task_struct
*next
)
2115 trace_sched_switch(prev
, next
);
2116 sched_info_switch(rq
, prev
, next
);
2117 perf_event_task_sched_out(prev
, next
);
2118 fire_sched_out_preempt_notifiers(prev
, next
);
2119 prepare_lock_switch(rq
, next
);
2120 prepare_arch_switch(next
);
2124 * finish_task_switch - clean up after a task-switch
2125 * @rq: runqueue associated with task-switch
2126 * @prev: the thread we just switched away from.
2128 * finish_task_switch must be called after the context switch, paired
2129 * with a prepare_task_switch call before the context switch.
2130 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2131 * and do any other architecture-specific cleanup actions.
2133 * Note that we may have delayed dropping an mm in context_switch(). If
2134 * so, we finish that here outside of the runqueue lock. (Doing it
2135 * with the lock held can cause deadlocks; see schedule() for
2138 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2139 __releases(rq
->lock
)
2141 struct mm_struct
*mm
= rq
->prev_mm
;
2147 * A task struct has one reference for the use as "current".
2148 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2149 * schedule one last time. The schedule call will never return, and
2150 * the scheduled task must drop that reference.
2151 * The test for TASK_DEAD must occur while the runqueue locks are
2152 * still held, otherwise prev could be scheduled on another cpu, die
2153 * there before we look at prev->state, and then the reference would
2155 * Manfred Spraul <manfred@colorfullife.com>
2157 prev_state
= prev
->state
;
2158 vtime_task_switch(prev
);
2159 finish_arch_switch(prev
);
2160 perf_event_task_sched_in(prev
, current
);
2161 finish_lock_switch(rq
, prev
);
2162 finish_arch_post_lock_switch();
2164 fire_sched_in_preempt_notifiers(current
);
2167 if (unlikely(prev_state
== TASK_DEAD
)) {
2168 if (prev
->sched_class
->task_dead
)
2169 prev
->sched_class
->task_dead(prev
);
2172 * Remove function-return probe instances associated with this
2173 * task and put them back on the free list.
2175 kprobe_flush_task(prev
);
2176 put_task_struct(prev
);
2179 tick_nohz_task_switch(current
);
2184 /* rq->lock is NOT held, but preemption is disabled */
2185 static inline void post_schedule(struct rq
*rq
)
2187 if (rq
->post_schedule
) {
2188 unsigned long flags
;
2190 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2191 if (rq
->curr
->sched_class
->post_schedule
)
2192 rq
->curr
->sched_class
->post_schedule(rq
);
2193 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2195 rq
->post_schedule
= 0;
2201 static inline void post_schedule(struct rq
*rq
)
2208 * schedule_tail - first thing a freshly forked thread must call.
2209 * @prev: the thread we just switched away from.
2211 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2212 __releases(rq
->lock
)
2214 struct rq
*rq
= this_rq();
2216 finish_task_switch(rq
, prev
);
2219 * FIXME: do we need to worry about rq being invalidated by the
2224 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2225 /* In this case, finish_task_switch does not reenable preemption */
2228 if (current
->set_child_tid
)
2229 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2233 * context_switch - switch to the new MM and the new
2234 * thread's register state.
2237 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2238 struct task_struct
*next
)
2240 struct mm_struct
*mm
, *oldmm
;
2242 prepare_task_switch(rq
, prev
, next
);
2245 oldmm
= prev
->active_mm
;
2247 * For paravirt, this is coupled with an exit in switch_to to
2248 * combine the page table reload and the switch backend into
2251 arch_start_context_switch(prev
);
2254 next
->active_mm
= oldmm
;
2255 atomic_inc(&oldmm
->mm_count
);
2256 enter_lazy_tlb(oldmm
, next
);
2258 switch_mm(oldmm
, mm
, next
);
2261 prev
->active_mm
= NULL
;
2262 rq
->prev_mm
= oldmm
;
2265 * Since the runqueue lock will be released by the next
2266 * task (which is an invalid locking op but in the case
2267 * of the scheduler it's an obvious special-case), so we
2268 * do an early lockdep release here:
2270 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2271 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2274 context_tracking_task_switch(prev
, next
);
2275 /* Here we just switch the register state and the stack. */
2276 switch_to(prev
, next
, prev
);
2280 * this_rq must be evaluated again because prev may have moved
2281 * CPUs since it called schedule(), thus the 'rq' on its stack
2282 * frame will be invalid.
2284 finish_task_switch(this_rq(), prev
);
2288 * nr_running and nr_context_switches:
2290 * externally visible scheduler statistics: current number of runnable
2291 * threads, total number of context switches performed since bootup.
2293 unsigned long nr_running(void)
2295 unsigned long i
, sum
= 0;
2297 for_each_online_cpu(i
)
2298 sum
+= cpu_rq(i
)->nr_running
;
2303 unsigned long long nr_context_switches(void)
2306 unsigned long long sum
= 0;
2308 for_each_possible_cpu(i
)
2309 sum
+= cpu_rq(i
)->nr_switches
;
2314 unsigned long nr_iowait(void)
2316 unsigned long i
, sum
= 0;
2318 for_each_possible_cpu(i
)
2319 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2324 unsigned long nr_iowait_cpu(int cpu
)
2326 struct rq
*this = cpu_rq(cpu
);
2327 return atomic_read(&this->nr_iowait
);
2333 * sched_exec - execve() is a valuable balancing opportunity, because at
2334 * this point the task has the smallest effective memory and cache footprint.
2336 void sched_exec(void)
2338 struct task_struct
*p
= current
;
2339 unsigned long flags
;
2342 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2343 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2344 if (dest_cpu
== smp_processor_id())
2347 if (likely(cpu_active(dest_cpu
))) {
2348 struct migration_arg arg
= { p
, dest_cpu
};
2350 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2351 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2355 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2360 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2361 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2363 EXPORT_PER_CPU_SYMBOL(kstat
);
2364 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2367 * Return any ns on the sched_clock that have not yet been accounted in
2368 * @p in case that task is currently running.
2370 * Called with task_rq_lock() held on @rq.
2372 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2376 if (task_current(rq
, p
)) {
2377 update_rq_clock(rq
);
2378 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2386 unsigned long long task_delta_exec(struct task_struct
*p
)
2388 unsigned long flags
;
2392 rq
= task_rq_lock(p
, &flags
);
2393 ns
= do_task_delta_exec(p
, rq
);
2394 task_rq_unlock(rq
, p
, &flags
);
2400 * Return accounted runtime for the task.
2401 * In case the task is currently running, return the runtime plus current's
2402 * pending runtime that have not been accounted yet.
2404 unsigned long long task_sched_runtime(struct task_struct
*p
)
2406 unsigned long flags
;
2410 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2412 * 64-bit doesn't need locks to atomically read a 64bit value.
2413 * So we have a optimization chance when the task's delta_exec is 0.
2414 * Reading ->on_cpu is racy, but this is ok.
2416 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2417 * If we race with it entering cpu, unaccounted time is 0. This is
2418 * indistinguishable from the read occurring a few cycles earlier.
2421 return p
->se
.sum_exec_runtime
;
2424 rq
= task_rq_lock(p
, &flags
);
2425 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2426 task_rq_unlock(rq
, p
, &flags
);
2432 * This function gets called by the timer code, with HZ frequency.
2433 * We call it with interrupts disabled.
2435 void scheduler_tick(void)
2437 int cpu
= smp_processor_id();
2438 struct rq
*rq
= cpu_rq(cpu
);
2439 struct task_struct
*curr
= rq
->curr
;
2443 raw_spin_lock(&rq
->lock
);
2444 update_rq_clock(rq
);
2445 curr
->sched_class
->task_tick(rq
, curr
, 0);
2446 update_cpu_load_active(rq
);
2447 raw_spin_unlock(&rq
->lock
);
2449 perf_event_task_tick();
2452 rq
->idle_balance
= idle_cpu(cpu
);
2453 trigger_load_balance(rq
);
2455 rq_last_tick_reset(rq
);
2458 #ifdef CONFIG_NO_HZ_FULL
2460 * scheduler_tick_max_deferment
2462 * Keep at least one tick per second when a single
2463 * active task is running because the scheduler doesn't
2464 * yet completely support full dynticks environment.
2466 * This makes sure that uptime, CFS vruntime, load
2467 * balancing, etc... continue to move forward, even
2468 * with a very low granularity.
2470 * Return: Maximum deferment in nanoseconds.
2472 u64
scheduler_tick_max_deferment(void)
2474 struct rq
*rq
= this_rq();
2475 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2477 next
= rq
->last_sched_tick
+ HZ
;
2479 if (time_before_eq(next
, now
))
2482 return jiffies_to_nsecs(next
- now
);
2486 notrace
unsigned long get_parent_ip(unsigned long addr
)
2488 if (in_lock_functions(addr
)) {
2489 addr
= CALLER_ADDR2
;
2490 if (in_lock_functions(addr
))
2491 addr
= CALLER_ADDR3
;
2496 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2497 defined(CONFIG_PREEMPT_TRACER))
2499 void __kprobes
preempt_count_add(int val
)
2501 #ifdef CONFIG_DEBUG_PREEMPT
2505 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2508 __preempt_count_add(val
);
2509 #ifdef CONFIG_DEBUG_PREEMPT
2511 * Spinlock count overflowing soon?
2513 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2516 if (preempt_count() == val
) {
2517 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2518 #ifdef CONFIG_DEBUG_PREEMPT
2519 current
->preempt_disable_ip
= ip
;
2521 trace_preempt_off(CALLER_ADDR0
, ip
);
2524 EXPORT_SYMBOL(preempt_count_add
);
2526 void __kprobes
preempt_count_sub(int val
)
2528 #ifdef CONFIG_DEBUG_PREEMPT
2532 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2535 * Is the spinlock portion underflowing?
2537 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2538 !(preempt_count() & PREEMPT_MASK
)))
2542 if (preempt_count() == val
)
2543 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2544 __preempt_count_sub(val
);
2546 EXPORT_SYMBOL(preempt_count_sub
);
2551 * Print scheduling while atomic bug:
2553 static noinline
void __schedule_bug(struct task_struct
*prev
)
2555 if (oops_in_progress
)
2558 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2559 prev
->comm
, prev
->pid
, preempt_count());
2561 debug_show_held_locks(prev
);
2563 if (irqs_disabled())
2564 print_irqtrace_events(prev
);
2565 #ifdef CONFIG_DEBUG_PREEMPT
2566 if (in_atomic_preempt_off()) {
2567 pr_err("Preemption disabled at:");
2568 print_ip_sym(current
->preempt_disable_ip
);
2573 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2577 * Various schedule()-time debugging checks and statistics:
2579 static inline void schedule_debug(struct task_struct
*prev
)
2582 * Test if we are atomic. Since do_exit() needs to call into
2583 * schedule() atomically, we ignore that path. Otherwise whine
2584 * if we are scheduling when we should not.
2586 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2587 __schedule_bug(prev
);
2590 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2592 schedstat_inc(this_rq(), sched_count
);
2596 * Pick up the highest-prio task:
2598 static inline struct task_struct
*
2599 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2601 const struct sched_class
*class = &fair_sched_class
;
2602 struct task_struct
*p
;
2605 * Optimization: we know that if all tasks are in
2606 * the fair class we can call that function directly:
2608 if (likely(prev
->sched_class
== class &&
2609 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2610 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2611 if (likely(p
&& p
!= RETRY_TASK
))
2616 for_each_class(class) {
2617 p
= class->pick_next_task(rq
, prev
);
2619 if (unlikely(p
== RETRY_TASK
))
2625 BUG(); /* the idle class will always have a runnable task */
2629 * __schedule() is the main scheduler function.
2631 * The main means of driving the scheduler and thus entering this function are:
2633 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2635 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2636 * paths. For example, see arch/x86/entry_64.S.
2638 * To drive preemption between tasks, the scheduler sets the flag in timer
2639 * interrupt handler scheduler_tick().
2641 * 3. Wakeups don't really cause entry into schedule(). They add a
2642 * task to the run-queue and that's it.
2644 * Now, if the new task added to the run-queue preempts the current
2645 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2646 * called on the nearest possible occasion:
2648 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2650 * - in syscall or exception context, at the next outmost
2651 * preempt_enable(). (this might be as soon as the wake_up()'s
2654 * - in IRQ context, return from interrupt-handler to
2655 * preemptible context
2657 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2660 * - cond_resched() call
2661 * - explicit schedule() call
2662 * - return from syscall or exception to user-space
2663 * - return from interrupt-handler to user-space
2665 static void __sched
__schedule(void)
2667 struct task_struct
*prev
, *next
;
2668 unsigned long *switch_count
;
2674 cpu
= smp_processor_id();
2676 rcu_note_context_switch(cpu
);
2679 schedule_debug(prev
);
2681 if (sched_feat(HRTICK
))
2685 * Make sure that signal_pending_state()->signal_pending() below
2686 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2687 * done by the caller to avoid the race with signal_wake_up().
2689 smp_mb__before_spinlock();
2690 raw_spin_lock_irq(&rq
->lock
);
2692 switch_count
= &prev
->nivcsw
;
2693 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2694 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2695 prev
->state
= TASK_RUNNING
;
2697 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2701 * If a worker went to sleep, notify and ask workqueue
2702 * whether it wants to wake up a task to maintain
2705 if (prev
->flags
& PF_WQ_WORKER
) {
2706 struct task_struct
*to_wakeup
;
2708 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2710 try_to_wake_up_local(to_wakeup
);
2713 switch_count
= &prev
->nvcsw
;
2716 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2717 update_rq_clock(rq
);
2719 next
= pick_next_task(rq
, prev
);
2720 clear_tsk_need_resched(prev
);
2721 clear_preempt_need_resched();
2722 rq
->skip_clock_update
= 0;
2724 if (likely(prev
!= next
)) {
2729 context_switch(rq
, prev
, next
); /* unlocks the rq */
2731 * The context switch have flipped the stack from under us
2732 * and restored the local variables which were saved when
2733 * this task called schedule() in the past. prev == current
2734 * is still correct, but it can be moved to another cpu/rq.
2736 cpu
= smp_processor_id();
2739 raw_spin_unlock_irq(&rq
->lock
);
2743 sched_preempt_enable_no_resched();
2748 static inline void sched_submit_work(struct task_struct
*tsk
)
2750 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2753 * If we are going to sleep and we have plugged IO queued,
2754 * make sure to submit it to avoid deadlocks.
2756 if (blk_needs_flush_plug(tsk
))
2757 blk_schedule_flush_plug(tsk
);
2760 asmlinkage
void __sched
schedule(void)
2762 struct task_struct
*tsk
= current
;
2764 sched_submit_work(tsk
);
2767 EXPORT_SYMBOL(schedule
);
2769 #ifdef CONFIG_CONTEXT_TRACKING
2770 asmlinkage
void __sched
schedule_user(void)
2773 * If we come here after a random call to set_need_resched(),
2774 * or we have been woken up remotely but the IPI has not yet arrived,
2775 * we haven't yet exited the RCU idle mode. Do it here manually until
2776 * we find a better solution.
2785 * schedule_preempt_disabled - called with preemption disabled
2787 * Returns with preemption disabled. Note: preempt_count must be 1
2789 void __sched
schedule_preempt_disabled(void)
2791 sched_preempt_enable_no_resched();
2796 #ifdef CONFIG_PREEMPT
2798 * this is the entry point to schedule() from in-kernel preemption
2799 * off of preempt_enable. Kernel preemptions off return from interrupt
2800 * occur there and call schedule directly.
2802 asmlinkage
void __sched notrace
preempt_schedule(void)
2805 * If there is a non-zero preempt_count or interrupts are disabled,
2806 * we do not want to preempt the current task. Just return..
2808 if (likely(!preemptible()))
2812 __preempt_count_add(PREEMPT_ACTIVE
);
2814 __preempt_count_sub(PREEMPT_ACTIVE
);
2817 * Check again in case we missed a preemption opportunity
2818 * between schedule and now.
2821 } while (need_resched());
2823 EXPORT_SYMBOL(preempt_schedule
);
2824 #endif /* CONFIG_PREEMPT */
2827 * this is the entry point to schedule() from kernel preemption
2828 * off of irq context.
2829 * Note, that this is called and return with irqs disabled. This will
2830 * protect us against recursive calling from irq.
2832 asmlinkage
void __sched
preempt_schedule_irq(void)
2834 enum ctx_state prev_state
;
2836 /* Catch callers which need to be fixed */
2837 BUG_ON(preempt_count() || !irqs_disabled());
2839 prev_state
= exception_enter();
2842 __preempt_count_add(PREEMPT_ACTIVE
);
2845 local_irq_disable();
2846 __preempt_count_sub(PREEMPT_ACTIVE
);
2849 * Check again in case we missed a preemption opportunity
2850 * between schedule and now.
2853 } while (need_resched());
2855 exception_exit(prev_state
);
2858 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2861 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2863 EXPORT_SYMBOL(default_wake_function
);
2865 #ifdef CONFIG_RT_MUTEXES
2868 * rt_mutex_setprio - set the current priority of a task
2870 * @prio: prio value (kernel-internal form)
2872 * This function changes the 'effective' priority of a task. It does
2873 * not touch ->normal_prio like __setscheduler().
2875 * Used by the rt_mutex code to implement priority inheritance
2876 * logic. Call site only calls if the priority of the task changed.
2878 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2880 int oldprio
, on_rq
, running
, enqueue_flag
= 0;
2882 const struct sched_class
*prev_class
;
2884 BUG_ON(prio
> MAX_PRIO
);
2886 rq
= __task_rq_lock(p
);
2889 * Idle task boosting is a nono in general. There is one
2890 * exception, when PREEMPT_RT and NOHZ is active:
2892 * The idle task calls get_next_timer_interrupt() and holds
2893 * the timer wheel base->lock on the CPU and another CPU wants
2894 * to access the timer (probably to cancel it). We can safely
2895 * ignore the boosting request, as the idle CPU runs this code
2896 * with interrupts disabled and will complete the lock
2897 * protected section without being interrupted. So there is no
2898 * real need to boost.
2900 if (unlikely(p
== rq
->idle
)) {
2901 WARN_ON(p
!= rq
->curr
);
2902 WARN_ON(p
->pi_blocked_on
);
2906 trace_sched_pi_setprio(p
, prio
);
2907 p
->pi_top_task
= rt_mutex_get_top_task(p
);
2909 prev_class
= p
->sched_class
;
2911 running
= task_current(rq
, p
);
2913 dequeue_task(rq
, p
, 0);
2915 p
->sched_class
->put_prev_task(rq
, p
);
2918 * Boosting condition are:
2919 * 1. -rt task is running and holds mutex A
2920 * --> -dl task blocks on mutex A
2922 * 2. -dl task is running and holds mutex A
2923 * --> -dl task blocks on mutex A and could preempt the
2926 if (dl_prio(prio
)) {
2927 if (!dl_prio(p
->normal_prio
) || (p
->pi_top_task
&&
2928 dl_entity_preempt(&p
->pi_top_task
->dl
, &p
->dl
))) {
2929 p
->dl
.dl_boosted
= 1;
2930 p
->dl
.dl_throttled
= 0;
2931 enqueue_flag
= ENQUEUE_REPLENISH
;
2933 p
->dl
.dl_boosted
= 0;
2934 p
->sched_class
= &dl_sched_class
;
2935 } else if (rt_prio(prio
)) {
2936 if (dl_prio(oldprio
))
2937 p
->dl
.dl_boosted
= 0;
2939 enqueue_flag
= ENQUEUE_HEAD
;
2940 p
->sched_class
= &rt_sched_class
;
2942 if (dl_prio(oldprio
))
2943 p
->dl
.dl_boosted
= 0;
2944 p
->sched_class
= &fair_sched_class
;
2950 p
->sched_class
->set_curr_task(rq
);
2952 enqueue_task(rq
, p
, enqueue_flag
);
2954 check_class_changed(rq
, p
, prev_class
, oldprio
);
2956 __task_rq_unlock(rq
);
2960 void set_user_nice(struct task_struct
*p
, long nice
)
2962 int old_prio
, delta
, on_rq
;
2963 unsigned long flags
;
2966 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
2969 * We have to be careful, if called from sys_setpriority(),
2970 * the task might be in the middle of scheduling on another CPU.
2972 rq
= task_rq_lock(p
, &flags
);
2974 * The RT priorities are set via sched_setscheduler(), but we still
2975 * allow the 'normal' nice value to be set - but as expected
2976 * it wont have any effect on scheduling until the task is
2977 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2979 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2980 p
->static_prio
= NICE_TO_PRIO(nice
);
2985 dequeue_task(rq
, p
, 0);
2987 p
->static_prio
= NICE_TO_PRIO(nice
);
2990 p
->prio
= effective_prio(p
);
2991 delta
= p
->prio
- old_prio
;
2994 enqueue_task(rq
, p
, 0);
2996 * If the task increased its priority or is running and
2997 * lowered its priority, then reschedule its CPU:
2999 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3000 resched_task(rq
->curr
);
3003 task_rq_unlock(rq
, p
, &flags
);
3005 EXPORT_SYMBOL(set_user_nice
);
3008 * can_nice - check if a task can reduce its nice value
3012 int can_nice(const struct task_struct
*p
, const int nice
)
3014 /* convert nice value [19,-20] to rlimit style value [1,40] */
3015 int nice_rlim
= 20 - nice
;
3017 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3018 capable(CAP_SYS_NICE
));
3021 #ifdef __ARCH_WANT_SYS_NICE
3024 * sys_nice - change the priority of the current process.
3025 * @increment: priority increment
3027 * sys_setpriority is a more generic, but much slower function that
3028 * does similar things.
3030 SYSCALL_DEFINE1(nice
, int, increment
)
3035 * Setpriority might change our priority at the same moment.
3036 * We don't have to worry. Conceptually one call occurs first
3037 * and we have a single winner.
3039 if (increment
< -40)
3044 nice
= task_nice(current
) + increment
;
3045 if (nice
< MIN_NICE
)
3047 if (nice
> MAX_NICE
)
3050 if (increment
< 0 && !can_nice(current
, nice
))
3053 retval
= security_task_setnice(current
, nice
);
3057 set_user_nice(current
, nice
);
3064 * task_prio - return the priority value of a given task.
3065 * @p: the task in question.
3067 * Return: The priority value as seen by users in /proc.
3068 * RT tasks are offset by -200. Normal tasks are centered
3069 * around 0, value goes from -16 to +15.
3071 int task_prio(const struct task_struct
*p
)
3073 return p
->prio
- MAX_RT_PRIO
;
3077 * idle_cpu - is a given cpu idle currently?
3078 * @cpu: the processor in question.
3080 * Return: 1 if the CPU is currently idle. 0 otherwise.
3082 int idle_cpu(int cpu
)
3084 struct rq
*rq
= cpu_rq(cpu
);
3086 if (rq
->curr
!= rq
->idle
)
3093 if (!llist_empty(&rq
->wake_list
))
3101 * idle_task - return the idle task for a given cpu.
3102 * @cpu: the processor in question.
3104 * Return: The idle task for the cpu @cpu.
3106 struct task_struct
*idle_task(int cpu
)
3108 return cpu_rq(cpu
)->idle
;
3112 * find_process_by_pid - find a process with a matching PID value.
3113 * @pid: the pid in question.
3115 * The task of @pid, if found. %NULL otherwise.
3117 static struct task_struct
*find_process_by_pid(pid_t pid
)
3119 return pid
? find_task_by_vpid(pid
) : current
;
3123 * This function initializes the sched_dl_entity of a newly becoming
3124 * SCHED_DEADLINE task.
3126 * Only the static values are considered here, the actual runtime and the
3127 * absolute deadline will be properly calculated when the task is enqueued
3128 * for the first time with its new policy.
3131 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3133 struct sched_dl_entity
*dl_se
= &p
->dl
;
3135 init_dl_task_timer(dl_se
);
3136 dl_se
->dl_runtime
= attr
->sched_runtime
;
3137 dl_se
->dl_deadline
= attr
->sched_deadline
;
3138 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3139 dl_se
->flags
= attr
->sched_flags
;
3140 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3141 dl_se
->dl_throttled
= 0;
3145 static void __setscheduler_params(struct task_struct
*p
,
3146 const struct sched_attr
*attr
)
3148 int policy
= attr
->sched_policy
;
3150 if (policy
== -1) /* setparam */
3155 if (dl_policy(policy
))
3156 __setparam_dl(p
, attr
);
3157 else if (fair_policy(policy
))
3158 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3161 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3162 * !rt_policy. Always setting this ensures that things like
3163 * getparam()/getattr() don't report silly values for !rt tasks.
3165 p
->rt_priority
= attr
->sched_priority
;
3166 p
->normal_prio
= normal_prio(p
);
3170 /* Actually do priority change: must hold pi & rq lock. */
3171 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3172 const struct sched_attr
*attr
)
3174 __setscheduler_params(p
, attr
);
3177 * If we get here, there was no pi waiters boosting the
3178 * task. It is safe to use the normal prio.
3180 p
->prio
= normal_prio(p
);
3182 if (dl_prio(p
->prio
))
3183 p
->sched_class
= &dl_sched_class
;
3184 else if (rt_prio(p
->prio
))
3185 p
->sched_class
= &rt_sched_class
;
3187 p
->sched_class
= &fair_sched_class
;
3191 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3193 struct sched_dl_entity
*dl_se
= &p
->dl
;
3195 attr
->sched_priority
= p
->rt_priority
;
3196 attr
->sched_runtime
= dl_se
->dl_runtime
;
3197 attr
->sched_deadline
= dl_se
->dl_deadline
;
3198 attr
->sched_period
= dl_se
->dl_period
;
3199 attr
->sched_flags
= dl_se
->flags
;
3203 * This function validates the new parameters of a -deadline task.
3204 * We ask for the deadline not being zero, and greater or equal
3205 * than the runtime, as well as the period of being zero or
3206 * greater than deadline. Furthermore, we have to be sure that
3207 * user parameters are above the internal resolution (1us); we
3208 * check sched_runtime only since it is always the smaller one.
3211 __checkparam_dl(const struct sched_attr
*attr
)
3213 return attr
&& attr
->sched_deadline
!= 0 &&
3214 (attr
->sched_period
== 0 ||
3215 (s64
)(attr
->sched_period
- attr
->sched_deadline
) >= 0) &&
3216 (s64
)(attr
->sched_deadline
- attr
->sched_runtime
) >= 0 &&
3217 attr
->sched_runtime
>= (2 << (DL_SCALE
- 1));
3221 * check the target process has a UID that matches the current process's
3223 static bool check_same_owner(struct task_struct
*p
)
3225 const struct cred
*cred
= current_cred(), *pcred
;
3229 pcred
= __task_cred(p
);
3230 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3231 uid_eq(cred
->euid
, pcred
->uid
));
3236 static int __sched_setscheduler(struct task_struct
*p
,
3237 const struct sched_attr
*attr
,
3240 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3241 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3242 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3243 int policy
= attr
->sched_policy
;
3244 unsigned long flags
;
3245 const struct sched_class
*prev_class
;
3249 /* may grab non-irq protected spin_locks */
3250 BUG_ON(in_interrupt());
3252 /* double check policy once rq lock held */
3254 reset_on_fork
= p
->sched_reset_on_fork
;
3255 policy
= oldpolicy
= p
->policy
;
3257 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3259 if (policy
!= SCHED_DEADLINE
&&
3260 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3261 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3262 policy
!= SCHED_IDLE
)
3266 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3270 * Valid priorities for SCHED_FIFO and SCHED_RR are
3271 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3272 * SCHED_BATCH and SCHED_IDLE is 0.
3274 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3275 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3277 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3278 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3282 * Allow unprivileged RT tasks to decrease priority:
3284 if (user
&& !capable(CAP_SYS_NICE
)) {
3285 if (fair_policy(policy
)) {
3286 if (attr
->sched_nice
< task_nice(p
) &&
3287 !can_nice(p
, attr
->sched_nice
))
3291 if (rt_policy(policy
)) {
3292 unsigned long rlim_rtprio
=
3293 task_rlimit(p
, RLIMIT_RTPRIO
);
3295 /* can't set/change the rt policy */
3296 if (policy
!= p
->policy
&& !rlim_rtprio
)
3299 /* can't increase priority */
3300 if (attr
->sched_priority
> p
->rt_priority
&&
3301 attr
->sched_priority
> rlim_rtprio
)
3306 * Can't set/change SCHED_DEADLINE policy at all for now
3307 * (safest behavior); in the future we would like to allow
3308 * unprivileged DL tasks to increase their relative deadline
3309 * or reduce their runtime (both ways reducing utilization)
3311 if (dl_policy(policy
))
3315 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3316 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3318 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3319 if (!can_nice(p
, task_nice(p
)))
3323 /* can't change other user's priorities */
3324 if (!check_same_owner(p
))
3327 /* Normal users shall not reset the sched_reset_on_fork flag */
3328 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3333 retval
= security_task_setscheduler(p
);
3339 * make sure no PI-waiters arrive (or leave) while we are
3340 * changing the priority of the task:
3342 * To be able to change p->policy safely, the appropriate
3343 * runqueue lock must be held.
3345 rq
= task_rq_lock(p
, &flags
);
3348 * Changing the policy of the stop threads its a very bad idea
3350 if (p
== rq
->stop
) {
3351 task_rq_unlock(rq
, p
, &flags
);
3356 * If not changing anything there's no need to proceed further,
3357 * but store a possible modification of reset_on_fork.
3359 if (unlikely(policy
== p
->policy
)) {
3360 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3362 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3364 if (dl_policy(policy
))
3367 p
->sched_reset_on_fork
= reset_on_fork
;
3368 task_rq_unlock(rq
, p
, &flags
);
3374 #ifdef CONFIG_RT_GROUP_SCHED
3376 * Do not allow realtime tasks into groups that have no runtime
3379 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3380 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3381 !task_group_is_autogroup(task_group(p
))) {
3382 task_rq_unlock(rq
, p
, &flags
);
3387 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3388 cpumask_t
*span
= rq
->rd
->span
;
3391 * Don't allow tasks with an affinity mask smaller than
3392 * the entire root_domain to become SCHED_DEADLINE. We
3393 * will also fail if there's no bandwidth available.
3395 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3396 rq
->rd
->dl_bw
.bw
== 0) {
3397 task_rq_unlock(rq
, p
, &flags
);
3404 /* recheck policy now with rq lock held */
3405 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3406 policy
= oldpolicy
= -1;
3407 task_rq_unlock(rq
, p
, &flags
);
3412 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3413 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3416 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3417 task_rq_unlock(rq
, p
, &flags
);
3421 p
->sched_reset_on_fork
= reset_on_fork
;
3425 * Special case for priority boosted tasks.
3427 * If the new priority is lower or equal (user space view)
3428 * than the current (boosted) priority, we just store the new
3429 * normal parameters and do not touch the scheduler class and
3430 * the runqueue. This will be done when the task deboost
3433 if (rt_mutex_check_prio(p
, newprio
)) {
3434 __setscheduler_params(p
, attr
);
3435 task_rq_unlock(rq
, p
, &flags
);
3440 running
= task_current(rq
, p
);
3442 dequeue_task(rq
, p
, 0);
3444 p
->sched_class
->put_prev_task(rq
, p
);
3446 prev_class
= p
->sched_class
;
3447 __setscheduler(rq
, p
, attr
);
3450 p
->sched_class
->set_curr_task(rq
);
3453 * We enqueue to tail when the priority of a task is
3454 * increased (user space view).
3456 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3459 check_class_changed(rq
, p
, prev_class
, oldprio
);
3460 task_rq_unlock(rq
, p
, &flags
);
3462 rt_mutex_adjust_pi(p
);
3467 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3468 const struct sched_param
*param
, bool check
)
3470 struct sched_attr attr
= {
3471 .sched_policy
= policy
,
3472 .sched_priority
= param
->sched_priority
,
3473 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3477 * Fixup the legacy SCHED_RESET_ON_FORK hack
3479 if (policy
& SCHED_RESET_ON_FORK
) {
3480 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3481 policy
&= ~SCHED_RESET_ON_FORK
;
3482 attr
.sched_policy
= policy
;
3485 return __sched_setscheduler(p
, &attr
, check
);
3488 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3489 * @p: the task in question.
3490 * @policy: new policy.
3491 * @param: structure containing the new RT priority.
3493 * Return: 0 on success. An error code otherwise.
3495 * NOTE that the task may be already dead.
3497 int sched_setscheduler(struct task_struct
*p
, int policy
,
3498 const struct sched_param
*param
)
3500 return _sched_setscheduler(p
, policy
, param
, true);
3502 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3504 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3506 return __sched_setscheduler(p
, attr
, true);
3508 EXPORT_SYMBOL_GPL(sched_setattr
);
3511 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3512 * @p: the task in question.
3513 * @policy: new policy.
3514 * @param: structure containing the new RT priority.
3516 * Just like sched_setscheduler, only don't bother checking if the
3517 * current context has permission. For example, this is needed in
3518 * stop_machine(): we create temporary high priority worker threads,
3519 * but our caller might not have that capability.
3521 * Return: 0 on success. An error code otherwise.
3523 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3524 const struct sched_param
*param
)
3526 return _sched_setscheduler(p
, policy
, param
, false);
3530 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3532 struct sched_param lparam
;
3533 struct task_struct
*p
;
3536 if (!param
|| pid
< 0)
3538 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3543 p
= find_process_by_pid(pid
);
3545 retval
= sched_setscheduler(p
, policy
, &lparam
);
3552 * Mimics kernel/events/core.c perf_copy_attr().
3554 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3555 struct sched_attr
*attr
)
3560 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3564 * zero the full structure, so that a short copy will be nice.
3566 memset(attr
, 0, sizeof(*attr
));
3568 ret
= get_user(size
, &uattr
->size
);
3572 if (size
> PAGE_SIZE
) /* silly large */
3575 if (!size
) /* abi compat */
3576 size
= SCHED_ATTR_SIZE_VER0
;
3578 if (size
< SCHED_ATTR_SIZE_VER0
)
3582 * If we're handed a bigger struct than we know of,
3583 * ensure all the unknown bits are 0 - i.e. new
3584 * user-space does not rely on any kernel feature
3585 * extensions we dont know about yet.
3587 if (size
> sizeof(*attr
)) {
3588 unsigned char __user
*addr
;
3589 unsigned char __user
*end
;
3592 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3593 end
= (void __user
*)uattr
+ size
;
3595 for (; addr
< end
; addr
++) {
3596 ret
= get_user(val
, addr
);
3602 size
= sizeof(*attr
);
3605 ret
= copy_from_user(attr
, uattr
, size
);
3610 * XXX: do we want to be lenient like existing syscalls; or do we want
3611 * to be strict and return an error on out-of-bounds values?
3613 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3619 put_user(sizeof(*attr
), &uattr
->size
);
3625 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3626 * @pid: the pid in question.
3627 * @policy: new policy.
3628 * @param: structure containing the new RT priority.
3630 * Return: 0 on success. An error code otherwise.
3632 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3633 struct sched_param __user
*, param
)
3635 /* negative values for policy are not valid */
3639 return do_sched_setscheduler(pid
, policy
, param
);
3643 * sys_sched_setparam - set/change the RT priority of a thread
3644 * @pid: the pid in question.
3645 * @param: structure containing the new RT priority.
3647 * Return: 0 on success. An error code otherwise.
3649 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3651 return do_sched_setscheduler(pid
, -1, param
);
3655 * sys_sched_setattr - same as above, but with extended sched_attr
3656 * @pid: the pid in question.
3657 * @uattr: structure containing the extended parameters.
3659 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3660 unsigned int, flags
)
3662 struct sched_attr attr
;
3663 struct task_struct
*p
;
3666 if (!uattr
|| pid
< 0 || flags
)
3669 if (sched_copy_attr(uattr
, &attr
))
3674 p
= find_process_by_pid(pid
);
3676 retval
= sched_setattr(p
, &attr
);
3683 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3684 * @pid: the pid in question.
3686 * Return: On success, the policy of the thread. Otherwise, a negative error
3689 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3691 struct task_struct
*p
;
3699 p
= find_process_by_pid(pid
);
3701 retval
= security_task_getscheduler(p
);
3704 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3711 * sys_sched_getparam - get the RT priority of a thread
3712 * @pid: the pid in question.
3713 * @param: structure containing the RT priority.
3715 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3718 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3720 struct sched_param lp
;
3721 struct task_struct
*p
;
3724 if (!param
|| pid
< 0)
3728 p
= find_process_by_pid(pid
);
3733 retval
= security_task_getscheduler(p
);
3737 if (task_has_dl_policy(p
)) {
3741 lp
.sched_priority
= p
->rt_priority
;
3745 * This one might sleep, we cannot do it with a spinlock held ...
3747 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3756 static int sched_read_attr(struct sched_attr __user
*uattr
,
3757 struct sched_attr
*attr
,
3762 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3766 * If we're handed a smaller struct than we know of,
3767 * ensure all the unknown bits are 0 - i.e. old
3768 * user-space does not get uncomplete information.
3770 if (usize
< sizeof(*attr
)) {
3771 unsigned char *addr
;
3774 addr
= (void *)attr
+ usize
;
3775 end
= (void *)attr
+ sizeof(*attr
);
3777 for (; addr
< end
; addr
++) {
3785 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3798 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3799 * @pid: the pid in question.
3800 * @uattr: structure containing the extended parameters.
3801 * @size: sizeof(attr) for fwd/bwd comp.
3803 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3804 unsigned int, size
, unsigned int, flags
)
3806 struct sched_attr attr
= {
3807 .size
= sizeof(struct sched_attr
),
3809 struct task_struct
*p
;
3812 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3813 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3817 p
= find_process_by_pid(pid
);
3822 retval
= security_task_getscheduler(p
);
3826 attr
.sched_policy
= p
->policy
;
3827 if (p
->sched_reset_on_fork
)
3828 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3829 if (task_has_dl_policy(p
))
3830 __getparam_dl(p
, &attr
);
3831 else if (task_has_rt_policy(p
))
3832 attr
.sched_priority
= p
->rt_priority
;
3834 attr
.sched_nice
= task_nice(p
);
3838 retval
= sched_read_attr(uattr
, &attr
, size
);
3846 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3848 cpumask_var_t cpus_allowed
, new_mask
;
3849 struct task_struct
*p
;
3854 p
= find_process_by_pid(pid
);
3860 /* Prevent p going away */
3864 if (p
->flags
& PF_NO_SETAFFINITY
) {
3868 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3872 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3874 goto out_free_cpus_allowed
;
3877 if (!check_same_owner(p
)) {
3879 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3886 retval
= security_task_setscheduler(p
);
3891 cpuset_cpus_allowed(p
, cpus_allowed
);
3892 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3895 * Since bandwidth control happens on root_domain basis,
3896 * if admission test is enabled, we only admit -deadline
3897 * tasks allowed to run on all the CPUs in the task's
3901 if (task_has_dl_policy(p
)) {
3902 const struct cpumask
*span
= task_rq(p
)->rd
->span
;
3904 if (dl_bandwidth_enabled() && !cpumask_subset(span
, new_mask
)) {
3911 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3914 cpuset_cpus_allowed(p
, cpus_allowed
);
3915 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3917 * We must have raced with a concurrent cpuset
3918 * update. Just reset the cpus_allowed to the
3919 * cpuset's cpus_allowed
3921 cpumask_copy(new_mask
, cpus_allowed
);
3926 free_cpumask_var(new_mask
);
3927 out_free_cpus_allowed
:
3928 free_cpumask_var(cpus_allowed
);
3934 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3935 struct cpumask
*new_mask
)
3937 if (len
< cpumask_size())
3938 cpumask_clear(new_mask
);
3939 else if (len
> cpumask_size())
3940 len
= cpumask_size();
3942 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3946 * sys_sched_setaffinity - set the cpu affinity of a process
3947 * @pid: pid of the process
3948 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3949 * @user_mask_ptr: user-space pointer to the new cpu mask
3951 * Return: 0 on success. An error code otherwise.
3953 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3954 unsigned long __user
*, user_mask_ptr
)
3956 cpumask_var_t new_mask
;
3959 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3962 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3964 retval
= sched_setaffinity(pid
, new_mask
);
3965 free_cpumask_var(new_mask
);
3969 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
3971 struct task_struct
*p
;
3972 unsigned long flags
;
3978 p
= find_process_by_pid(pid
);
3982 retval
= security_task_getscheduler(p
);
3986 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3987 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
3988 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3997 * sys_sched_getaffinity - get the cpu affinity of a process
3998 * @pid: pid of the process
3999 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4000 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4002 * Return: 0 on success. An error code otherwise.
4004 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4005 unsigned long __user
*, user_mask_ptr
)
4010 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4012 if (len
& (sizeof(unsigned long)-1))
4015 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4018 ret
= sched_getaffinity(pid
, mask
);
4020 size_t retlen
= min_t(size_t, len
, cpumask_size());
4022 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4027 free_cpumask_var(mask
);
4033 * sys_sched_yield - yield the current processor to other threads.
4035 * This function yields the current CPU to other tasks. If there are no
4036 * other threads running on this CPU then this function will return.
4040 SYSCALL_DEFINE0(sched_yield
)
4042 struct rq
*rq
= this_rq_lock();
4044 schedstat_inc(rq
, yld_count
);
4045 current
->sched_class
->yield_task(rq
);
4048 * Since we are going to call schedule() anyway, there's
4049 * no need to preempt or enable interrupts:
4051 __release(rq
->lock
);
4052 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4053 do_raw_spin_unlock(&rq
->lock
);
4054 sched_preempt_enable_no_resched();
4061 static void __cond_resched(void)
4063 __preempt_count_add(PREEMPT_ACTIVE
);
4065 __preempt_count_sub(PREEMPT_ACTIVE
);
4068 int __sched
_cond_resched(void)
4070 if (should_resched()) {
4076 EXPORT_SYMBOL(_cond_resched
);
4079 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4080 * call schedule, and on return reacquire the lock.
4082 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4083 * operations here to prevent schedule() from being called twice (once via
4084 * spin_unlock(), once by hand).
4086 int __cond_resched_lock(spinlock_t
*lock
)
4088 int resched
= should_resched();
4091 lockdep_assert_held(lock
);
4093 if (spin_needbreak(lock
) || resched
) {
4104 EXPORT_SYMBOL(__cond_resched_lock
);
4106 int __sched
__cond_resched_softirq(void)
4108 BUG_ON(!in_softirq());
4110 if (should_resched()) {
4118 EXPORT_SYMBOL(__cond_resched_softirq
);
4121 * yield - yield the current processor to other threads.
4123 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4125 * The scheduler is at all times free to pick the calling task as the most
4126 * eligible task to run, if removing the yield() call from your code breaks
4127 * it, its already broken.
4129 * Typical broken usage is:
4134 * where one assumes that yield() will let 'the other' process run that will
4135 * make event true. If the current task is a SCHED_FIFO task that will never
4136 * happen. Never use yield() as a progress guarantee!!
4138 * If you want to use yield() to wait for something, use wait_event().
4139 * If you want to use yield() to be 'nice' for others, use cond_resched().
4140 * If you still want to use yield(), do not!
4142 void __sched
yield(void)
4144 set_current_state(TASK_RUNNING
);
4147 EXPORT_SYMBOL(yield
);
4150 * yield_to - yield the current processor to another thread in
4151 * your thread group, or accelerate that thread toward the
4152 * processor it's on.
4154 * @preempt: whether task preemption is allowed or not
4156 * It's the caller's job to ensure that the target task struct
4157 * can't go away on us before we can do any checks.
4160 * true (>0) if we indeed boosted the target task.
4161 * false (0) if we failed to boost the target.
4162 * -ESRCH if there's no task to yield to.
4164 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4166 struct task_struct
*curr
= current
;
4167 struct rq
*rq
, *p_rq
;
4168 unsigned long flags
;
4171 local_irq_save(flags
);
4177 * If we're the only runnable task on the rq and target rq also
4178 * has only one task, there's absolutely no point in yielding.
4180 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4185 double_rq_lock(rq
, p_rq
);
4186 if (task_rq(p
) != p_rq
) {
4187 double_rq_unlock(rq
, p_rq
);
4191 if (!curr
->sched_class
->yield_to_task
)
4194 if (curr
->sched_class
!= p
->sched_class
)
4197 if (task_running(p_rq
, p
) || p
->state
)
4200 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4202 schedstat_inc(rq
, yld_count
);
4204 * Make p's CPU reschedule; pick_next_entity takes care of
4207 if (preempt
&& rq
!= p_rq
)
4208 resched_task(p_rq
->curr
);
4212 double_rq_unlock(rq
, p_rq
);
4214 local_irq_restore(flags
);
4221 EXPORT_SYMBOL_GPL(yield_to
);
4224 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4225 * that process accounting knows that this is a task in IO wait state.
4227 void __sched
io_schedule(void)
4229 struct rq
*rq
= raw_rq();
4231 delayacct_blkio_start();
4232 atomic_inc(&rq
->nr_iowait
);
4233 blk_flush_plug(current
);
4234 current
->in_iowait
= 1;
4236 current
->in_iowait
= 0;
4237 atomic_dec(&rq
->nr_iowait
);
4238 delayacct_blkio_end();
4240 EXPORT_SYMBOL(io_schedule
);
4242 long __sched
io_schedule_timeout(long timeout
)
4244 struct rq
*rq
= raw_rq();
4247 delayacct_blkio_start();
4248 atomic_inc(&rq
->nr_iowait
);
4249 blk_flush_plug(current
);
4250 current
->in_iowait
= 1;
4251 ret
= schedule_timeout(timeout
);
4252 current
->in_iowait
= 0;
4253 atomic_dec(&rq
->nr_iowait
);
4254 delayacct_blkio_end();
4259 * sys_sched_get_priority_max - return maximum RT priority.
4260 * @policy: scheduling class.
4262 * Return: On success, this syscall returns the maximum
4263 * rt_priority that can be used by a given scheduling class.
4264 * On failure, a negative error code is returned.
4266 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4273 ret
= MAX_USER_RT_PRIO
-1;
4275 case SCHED_DEADLINE
:
4286 * sys_sched_get_priority_min - return minimum RT priority.
4287 * @policy: scheduling class.
4289 * Return: On success, this syscall returns the minimum
4290 * rt_priority that can be used by a given scheduling class.
4291 * On failure, a negative error code is returned.
4293 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4302 case SCHED_DEADLINE
:
4312 * sys_sched_rr_get_interval - return the default timeslice of a process.
4313 * @pid: pid of the process.
4314 * @interval: userspace pointer to the timeslice value.
4316 * this syscall writes the default timeslice value of a given process
4317 * into the user-space timespec buffer. A value of '0' means infinity.
4319 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4322 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4323 struct timespec __user
*, interval
)
4325 struct task_struct
*p
;
4326 unsigned int time_slice
;
4327 unsigned long flags
;
4337 p
= find_process_by_pid(pid
);
4341 retval
= security_task_getscheduler(p
);
4345 rq
= task_rq_lock(p
, &flags
);
4347 if (p
->sched_class
->get_rr_interval
)
4348 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4349 task_rq_unlock(rq
, p
, &flags
);
4352 jiffies_to_timespec(time_slice
, &t
);
4353 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4361 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4363 void sched_show_task(struct task_struct
*p
)
4365 unsigned long free
= 0;
4369 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4370 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4371 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4372 #if BITS_PER_LONG == 32
4373 if (state
== TASK_RUNNING
)
4374 printk(KERN_CONT
" running ");
4376 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4378 if (state
== TASK_RUNNING
)
4379 printk(KERN_CONT
" running task ");
4381 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4383 #ifdef CONFIG_DEBUG_STACK_USAGE
4384 free
= stack_not_used(p
);
4387 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4389 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4390 task_pid_nr(p
), ppid
,
4391 (unsigned long)task_thread_info(p
)->flags
);
4393 print_worker_info(KERN_INFO
, p
);
4394 show_stack(p
, NULL
);
4397 void show_state_filter(unsigned long state_filter
)
4399 struct task_struct
*g
, *p
;
4401 #if BITS_PER_LONG == 32
4403 " task PC stack pid father\n");
4406 " task PC stack pid father\n");
4409 do_each_thread(g
, p
) {
4411 * reset the NMI-timeout, listing all files on a slow
4412 * console might take a lot of time:
4414 touch_nmi_watchdog();
4415 if (!state_filter
|| (p
->state
& state_filter
))
4417 } while_each_thread(g
, p
);
4419 touch_all_softlockup_watchdogs();
4421 #ifdef CONFIG_SCHED_DEBUG
4422 sysrq_sched_debug_show();
4426 * Only show locks if all tasks are dumped:
4429 debug_show_all_locks();
4432 void init_idle_bootup_task(struct task_struct
*idle
)
4434 idle
->sched_class
= &idle_sched_class
;
4438 * init_idle - set up an idle thread for a given CPU
4439 * @idle: task in question
4440 * @cpu: cpu the idle task belongs to
4442 * NOTE: this function does not set the idle thread's NEED_RESCHED
4443 * flag, to make booting more robust.
4445 void init_idle(struct task_struct
*idle
, int cpu
)
4447 struct rq
*rq
= cpu_rq(cpu
);
4448 unsigned long flags
;
4450 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4452 __sched_fork(0, idle
);
4453 idle
->state
= TASK_RUNNING
;
4454 idle
->se
.exec_start
= sched_clock();
4456 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4458 * We're having a chicken and egg problem, even though we are
4459 * holding rq->lock, the cpu isn't yet set to this cpu so the
4460 * lockdep check in task_group() will fail.
4462 * Similar case to sched_fork(). / Alternatively we could
4463 * use task_rq_lock() here and obtain the other rq->lock.
4468 __set_task_cpu(idle
, cpu
);
4471 rq
->curr
= rq
->idle
= idle
;
4473 #if defined(CONFIG_SMP)
4476 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4478 /* Set the preempt count _outside_ the spinlocks! */
4479 init_idle_preempt_count(idle
, cpu
);
4482 * The idle tasks have their own, simple scheduling class:
4484 idle
->sched_class
= &idle_sched_class
;
4485 ftrace_graph_init_idle_task(idle
, cpu
);
4486 vtime_init_idle(idle
, cpu
);
4487 #if defined(CONFIG_SMP)
4488 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4493 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4495 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4496 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4498 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4499 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4503 * This is how migration works:
4505 * 1) we invoke migration_cpu_stop() on the target CPU using
4507 * 2) stopper starts to run (implicitly forcing the migrated thread
4509 * 3) it checks whether the migrated task is still in the wrong runqueue.
4510 * 4) if it's in the wrong runqueue then the migration thread removes
4511 * it and puts it into the right queue.
4512 * 5) stopper completes and stop_one_cpu() returns and the migration
4517 * Change a given task's CPU affinity. Migrate the thread to a
4518 * proper CPU and schedule it away if the CPU it's executing on
4519 * is removed from the allowed bitmask.
4521 * NOTE: the caller must have a valid reference to the task, the
4522 * task must not exit() & deallocate itself prematurely. The
4523 * call is not atomic; no spinlocks may be held.
4525 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4527 unsigned long flags
;
4529 unsigned int dest_cpu
;
4532 rq
= task_rq_lock(p
, &flags
);
4534 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4537 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4542 do_set_cpus_allowed(p
, new_mask
);
4544 /* Can the task run on the task's current CPU? If so, we're done */
4545 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4548 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4550 struct migration_arg arg
= { p
, dest_cpu
};
4551 /* Need help from migration thread: drop lock and wait. */
4552 task_rq_unlock(rq
, p
, &flags
);
4553 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4554 tlb_migrate_finish(p
->mm
);
4558 task_rq_unlock(rq
, p
, &flags
);
4562 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4565 * Move (not current) task off this cpu, onto dest cpu. We're doing
4566 * this because either it can't run here any more (set_cpus_allowed()
4567 * away from this CPU, or CPU going down), or because we're
4568 * attempting to rebalance this task on exec (sched_exec).
4570 * So we race with normal scheduler movements, but that's OK, as long
4571 * as the task is no longer on this CPU.
4573 * Returns non-zero if task was successfully migrated.
4575 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4577 struct rq
*rq_dest
, *rq_src
;
4580 if (unlikely(!cpu_active(dest_cpu
)))
4583 rq_src
= cpu_rq(src_cpu
);
4584 rq_dest
= cpu_rq(dest_cpu
);
4586 raw_spin_lock(&p
->pi_lock
);
4587 double_rq_lock(rq_src
, rq_dest
);
4588 /* Already moved. */
4589 if (task_cpu(p
) != src_cpu
)
4591 /* Affinity changed (again). */
4592 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4596 * If we're not on a rq, the next wake-up will ensure we're
4600 dequeue_task(rq_src
, p
, 0);
4601 set_task_cpu(p
, dest_cpu
);
4602 enqueue_task(rq_dest
, p
, 0);
4603 check_preempt_curr(rq_dest
, p
, 0);
4608 double_rq_unlock(rq_src
, rq_dest
);
4609 raw_spin_unlock(&p
->pi_lock
);
4613 #ifdef CONFIG_NUMA_BALANCING
4614 /* Migrate current task p to target_cpu */
4615 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4617 struct migration_arg arg
= { p
, target_cpu
};
4618 int curr_cpu
= task_cpu(p
);
4620 if (curr_cpu
== target_cpu
)
4623 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4626 /* TODO: This is not properly updating schedstats */
4628 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4629 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4633 * Requeue a task on a given node and accurately track the number of NUMA
4634 * tasks on the runqueues
4636 void sched_setnuma(struct task_struct
*p
, int nid
)
4639 unsigned long flags
;
4640 bool on_rq
, running
;
4642 rq
= task_rq_lock(p
, &flags
);
4644 running
= task_current(rq
, p
);
4647 dequeue_task(rq
, p
, 0);
4649 p
->sched_class
->put_prev_task(rq
, p
);
4651 p
->numa_preferred_nid
= nid
;
4654 p
->sched_class
->set_curr_task(rq
);
4656 enqueue_task(rq
, p
, 0);
4657 task_rq_unlock(rq
, p
, &flags
);
4662 * migration_cpu_stop - this will be executed by a highprio stopper thread
4663 * and performs thread migration by bumping thread off CPU then
4664 * 'pushing' onto another runqueue.
4666 static int migration_cpu_stop(void *data
)
4668 struct migration_arg
*arg
= data
;
4671 * The original target cpu might have gone down and we might
4672 * be on another cpu but it doesn't matter.
4674 local_irq_disable();
4675 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4680 #ifdef CONFIG_HOTPLUG_CPU
4683 * Ensures that the idle task is using init_mm right before its cpu goes
4686 void idle_task_exit(void)
4688 struct mm_struct
*mm
= current
->active_mm
;
4690 BUG_ON(cpu_online(smp_processor_id()));
4692 if (mm
!= &init_mm
) {
4693 switch_mm(mm
, &init_mm
, current
);
4694 finish_arch_post_lock_switch();
4700 * Since this CPU is going 'away' for a while, fold any nr_active delta
4701 * we might have. Assumes we're called after migrate_tasks() so that the
4702 * nr_active count is stable.
4704 * Also see the comment "Global load-average calculations".
4706 static void calc_load_migrate(struct rq
*rq
)
4708 long delta
= calc_load_fold_active(rq
);
4710 atomic_long_add(delta
, &calc_load_tasks
);
4713 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4717 static const struct sched_class fake_sched_class
= {
4718 .put_prev_task
= put_prev_task_fake
,
4721 static struct task_struct fake_task
= {
4723 * Avoid pull_{rt,dl}_task()
4725 .prio
= MAX_PRIO
+ 1,
4726 .sched_class
= &fake_sched_class
,
4730 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4731 * try_to_wake_up()->select_task_rq().
4733 * Called with rq->lock held even though we'er in stop_machine() and
4734 * there's no concurrency possible, we hold the required locks anyway
4735 * because of lock validation efforts.
4737 static void migrate_tasks(unsigned int dead_cpu
)
4739 struct rq
*rq
= cpu_rq(dead_cpu
);
4740 struct task_struct
*next
, *stop
= rq
->stop
;
4744 * Fudge the rq selection such that the below task selection loop
4745 * doesn't get stuck on the currently eligible stop task.
4747 * We're currently inside stop_machine() and the rq is either stuck
4748 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4749 * either way we should never end up calling schedule() until we're
4755 * put_prev_task() and pick_next_task() sched
4756 * class method both need to have an up-to-date
4757 * value of rq->clock[_task]
4759 update_rq_clock(rq
);
4763 * There's this thread running, bail when that's the only
4766 if (rq
->nr_running
== 1)
4769 next
= pick_next_task(rq
, &fake_task
);
4771 next
->sched_class
->put_prev_task(rq
, next
);
4773 /* Find suitable destination for @next, with force if needed. */
4774 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4775 raw_spin_unlock(&rq
->lock
);
4777 __migrate_task(next
, dead_cpu
, dest_cpu
);
4779 raw_spin_lock(&rq
->lock
);
4785 #endif /* CONFIG_HOTPLUG_CPU */
4787 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4789 static struct ctl_table sd_ctl_dir
[] = {
4791 .procname
= "sched_domain",
4797 static struct ctl_table sd_ctl_root
[] = {
4799 .procname
= "kernel",
4801 .child
= sd_ctl_dir
,
4806 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4808 struct ctl_table
*entry
=
4809 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4814 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4816 struct ctl_table
*entry
;
4819 * In the intermediate directories, both the child directory and
4820 * procname are dynamically allocated and could fail but the mode
4821 * will always be set. In the lowest directory the names are
4822 * static strings and all have proc handlers.
4824 for (entry
= *tablep
; entry
->mode
; entry
++) {
4826 sd_free_ctl_entry(&entry
->child
);
4827 if (entry
->proc_handler
== NULL
)
4828 kfree(entry
->procname
);
4835 static int min_load_idx
= 0;
4836 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4839 set_table_entry(struct ctl_table
*entry
,
4840 const char *procname
, void *data
, int maxlen
,
4841 umode_t mode
, proc_handler
*proc_handler
,
4844 entry
->procname
= procname
;
4846 entry
->maxlen
= maxlen
;
4848 entry
->proc_handler
= proc_handler
;
4851 entry
->extra1
= &min_load_idx
;
4852 entry
->extra2
= &max_load_idx
;
4856 static struct ctl_table
*
4857 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4859 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
4864 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4865 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4866 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4867 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4868 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4869 sizeof(int), 0644, proc_dointvec_minmax
, true);
4870 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4871 sizeof(int), 0644, proc_dointvec_minmax
, true);
4872 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4873 sizeof(int), 0644, proc_dointvec_minmax
, true);
4874 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4875 sizeof(int), 0644, proc_dointvec_minmax
, true);
4876 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4877 sizeof(int), 0644, proc_dointvec_minmax
, true);
4878 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4879 sizeof(int), 0644, proc_dointvec_minmax
, false);
4880 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4881 sizeof(int), 0644, proc_dointvec_minmax
, false);
4882 set_table_entry(&table
[9], "cache_nice_tries",
4883 &sd
->cache_nice_tries
,
4884 sizeof(int), 0644, proc_dointvec_minmax
, false);
4885 set_table_entry(&table
[10], "flags", &sd
->flags
,
4886 sizeof(int), 0644, proc_dointvec_minmax
, false);
4887 set_table_entry(&table
[11], "max_newidle_lb_cost",
4888 &sd
->max_newidle_lb_cost
,
4889 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4890 set_table_entry(&table
[12], "name", sd
->name
,
4891 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4892 /* &table[13] is terminator */
4897 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4899 struct ctl_table
*entry
, *table
;
4900 struct sched_domain
*sd
;
4901 int domain_num
= 0, i
;
4904 for_each_domain(cpu
, sd
)
4906 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4911 for_each_domain(cpu
, sd
) {
4912 snprintf(buf
, 32, "domain%d", i
);
4913 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4915 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4922 static struct ctl_table_header
*sd_sysctl_header
;
4923 static void register_sched_domain_sysctl(void)
4925 int i
, cpu_num
= num_possible_cpus();
4926 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4929 WARN_ON(sd_ctl_dir
[0].child
);
4930 sd_ctl_dir
[0].child
= entry
;
4935 for_each_possible_cpu(i
) {
4936 snprintf(buf
, 32, "cpu%d", i
);
4937 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4939 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4943 WARN_ON(sd_sysctl_header
);
4944 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4947 /* may be called multiple times per register */
4948 static void unregister_sched_domain_sysctl(void)
4950 if (sd_sysctl_header
)
4951 unregister_sysctl_table(sd_sysctl_header
);
4952 sd_sysctl_header
= NULL
;
4953 if (sd_ctl_dir
[0].child
)
4954 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4957 static void register_sched_domain_sysctl(void)
4960 static void unregister_sched_domain_sysctl(void)
4965 static void set_rq_online(struct rq
*rq
)
4968 const struct sched_class
*class;
4970 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
4973 for_each_class(class) {
4974 if (class->rq_online
)
4975 class->rq_online(rq
);
4980 static void set_rq_offline(struct rq
*rq
)
4983 const struct sched_class
*class;
4985 for_each_class(class) {
4986 if (class->rq_offline
)
4987 class->rq_offline(rq
);
4990 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
4996 * migration_call - callback that gets triggered when a CPU is added.
4997 * Here we can start up the necessary migration thread for the new CPU.
5000 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5002 int cpu
= (long)hcpu
;
5003 unsigned long flags
;
5004 struct rq
*rq
= cpu_rq(cpu
);
5006 switch (action
& ~CPU_TASKS_FROZEN
) {
5008 case CPU_UP_PREPARE
:
5009 rq
->calc_load_update
= calc_load_update
;
5013 /* Update our root-domain */
5014 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5016 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5020 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5023 #ifdef CONFIG_HOTPLUG_CPU
5025 sched_ttwu_pending();
5026 /* Update our root-domain */
5027 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5029 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5033 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5034 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5038 calc_load_migrate(rq
);
5043 update_max_interval();
5049 * Register at high priority so that task migration (migrate_all_tasks)
5050 * happens before everything else. This has to be lower priority than
5051 * the notifier in the perf_event subsystem, though.
5053 static struct notifier_block migration_notifier
= {
5054 .notifier_call
= migration_call
,
5055 .priority
= CPU_PRI_MIGRATION
,
5058 static int sched_cpu_active(struct notifier_block
*nfb
,
5059 unsigned long action
, void *hcpu
)
5061 switch (action
& ~CPU_TASKS_FROZEN
) {
5063 case CPU_DOWN_FAILED
:
5064 set_cpu_active((long)hcpu
, true);
5071 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5072 unsigned long action
, void *hcpu
)
5074 unsigned long flags
;
5075 long cpu
= (long)hcpu
;
5077 switch (action
& ~CPU_TASKS_FROZEN
) {
5078 case CPU_DOWN_PREPARE
:
5079 set_cpu_active(cpu
, false);
5081 /* explicitly allow suspend */
5082 if (!(action
& CPU_TASKS_FROZEN
)) {
5083 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
5087 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5088 cpus
= dl_bw_cpus(cpu
);
5089 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5090 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5093 return notifier_from_errno(-EBUSY
);
5101 static int __init
migration_init(void)
5103 void *cpu
= (void *)(long)smp_processor_id();
5106 /* Initialize migration for the boot CPU */
5107 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5108 BUG_ON(err
== NOTIFY_BAD
);
5109 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5110 register_cpu_notifier(&migration_notifier
);
5112 /* Register cpu active notifiers */
5113 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5114 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5118 early_initcall(migration_init
);
5123 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5125 #ifdef CONFIG_SCHED_DEBUG
5127 static __read_mostly
int sched_debug_enabled
;
5129 static int __init
sched_debug_setup(char *str
)
5131 sched_debug_enabled
= 1;
5135 early_param("sched_debug", sched_debug_setup
);
5137 static inline bool sched_debug(void)
5139 return sched_debug_enabled
;
5142 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5143 struct cpumask
*groupmask
)
5145 struct sched_group
*group
= sd
->groups
;
5148 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5149 cpumask_clear(groupmask
);
5151 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5153 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5154 printk("does not load-balance\n");
5156 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5161 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5163 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5164 printk(KERN_ERR
"ERROR: domain->span does not contain "
5167 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5168 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5172 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5176 printk(KERN_ERR
"ERROR: group is NULL\n");
5181 * Even though we initialize ->power to something semi-sane,
5182 * we leave power_orig unset. This allows us to detect if
5183 * domain iteration is still funny without causing /0 traps.
5185 if (!group
->sgp
->power_orig
) {
5186 printk(KERN_CONT
"\n");
5187 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5192 if (!cpumask_weight(sched_group_cpus(group
))) {
5193 printk(KERN_CONT
"\n");
5194 printk(KERN_ERR
"ERROR: empty group\n");
5198 if (!(sd
->flags
& SD_OVERLAP
) &&
5199 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5200 printk(KERN_CONT
"\n");
5201 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5205 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5207 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5209 printk(KERN_CONT
" %s", str
);
5210 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5211 printk(KERN_CONT
" (cpu_power = %d)",
5215 group
= group
->next
;
5216 } while (group
!= sd
->groups
);
5217 printk(KERN_CONT
"\n");
5219 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5220 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5223 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5224 printk(KERN_ERR
"ERROR: parent span is not a superset "
5225 "of domain->span\n");
5229 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5233 if (!sched_debug_enabled
)
5237 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5241 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5244 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5252 #else /* !CONFIG_SCHED_DEBUG */
5253 # define sched_domain_debug(sd, cpu) do { } while (0)
5254 static inline bool sched_debug(void)
5258 #endif /* CONFIG_SCHED_DEBUG */
5260 static int sd_degenerate(struct sched_domain
*sd
)
5262 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5265 /* Following flags need at least 2 groups */
5266 if (sd
->flags
& (SD_LOAD_BALANCE
|
5267 SD_BALANCE_NEWIDLE
|
5271 SD_SHARE_PKG_RESOURCES
)) {
5272 if (sd
->groups
!= sd
->groups
->next
)
5276 /* Following flags don't use groups */
5277 if (sd
->flags
& (SD_WAKE_AFFINE
))
5284 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5286 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5288 if (sd_degenerate(parent
))
5291 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5294 /* Flags needing groups don't count if only 1 group in parent */
5295 if (parent
->groups
== parent
->groups
->next
) {
5296 pflags
&= ~(SD_LOAD_BALANCE
|
5297 SD_BALANCE_NEWIDLE
|
5301 SD_SHARE_PKG_RESOURCES
|
5303 if (nr_node_ids
== 1)
5304 pflags
&= ~SD_SERIALIZE
;
5306 if (~cflags
& pflags
)
5312 static void free_rootdomain(struct rcu_head
*rcu
)
5314 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5316 cpupri_cleanup(&rd
->cpupri
);
5317 cpudl_cleanup(&rd
->cpudl
);
5318 free_cpumask_var(rd
->dlo_mask
);
5319 free_cpumask_var(rd
->rto_mask
);
5320 free_cpumask_var(rd
->online
);
5321 free_cpumask_var(rd
->span
);
5325 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5327 struct root_domain
*old_rd
= NULL
;
5328 unsigned long flags
;
5330 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5335 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5338 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5341 * If we dont want to free the old_rd yet then
5342 * set old_rd to NULL to skip the freeing later
5345 if (!atomic_dec_and_test(&old_rd
->refcount
))
5349 atomic_inc(&rd
->refcount
);
5352 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5353 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5356 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5359 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5362 static int init_rootdomain(struct root_domain
*rd
)
5364 memset(rd
, 0, sizeof(*rd
));
5366 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5368 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5370 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5372 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5375 init_dl_bw(&rd
->dl_bw
);
5376 if (cpudl_init(&rd
->cpudl
) != 0)
5379 if (cpupri_init(&rd
->cpupri
) != 0)
5384 free_cpumask_var(rd
->rto_mask
);
5386 free_cpumask_var(rd
->dlo_mask
);
5388 free_cpumask_var(rd
->online
);
5390 free_cpumask_var(rd
->span
);
5396 * By default the system creates a single root-domain with all cpus as
5397 * members (mimicking the global state we have today).
5399 struct root_domain def_root_domain
;
5401 static void init_defrootdomain(void)
5403 init_rootdomain(&def_root_domain
);
5405 atomic_set(&def_root_domain
.refcount
, 1);
5408 static struct root_domain
*alloc_rootdomain(void)
5410 struct root_domain
*rd
;
5412 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5416 if (init_rootdomain(rd
) != 0) {
5424 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5426 struct sched_group
*tmp
, *first
;
5435 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5440 } while (sg
!= first
);
5443 static void free_sched_domain(struct rcu_head
*rcu
)
5445 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5448 * If its an overlapping domain it has private groups, iterate and
5451 if (sd
->flags
& SD_OVERLAP
) {
5452 free_sched_groups(sd
->groups
, 1);
5453 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5454 kfree(sd
->groups
->sgp
);
5460 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5462 call_rcu(&sd
->rcu
, free_sched_domain
);
5465 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5467 for (; sd
; sd
= sd
->parent
)
5468 destroy_sched_domain(sd
, cpu
);
5472 * Keep a special pointer to the highest sched_domain that has
5473 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5474 * allows us to avoid some pointer chasing select_idle_sibling().
5476 * Also keep a unique ID per domain (we use the first cpu number in
5477 * the cpumask of the domain), this allows us to quickly tell if
5478 * two cpus are in the same cache domain, see cpus_share_cache().
5480 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5481 DEFINE_PER_CPU(int, sd_llc_size
);
5482 DEFINE_PER_CPU(int, sd_llc_id
);
5483 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5484 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5485 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5487 static void update_top_cache_domain(int cpu
)
5489 struct sched_domain
*sd
;
5490 struct sched_domain
*busy_sd
= NULL
;
5494 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5496 id
= cpumask_first(sched_domain_span(sd
));
5497 size
= cpumask_weight(sched_domain_span(sd
));
5498 busy_sd
= sd
->parent
; /* sd_busy */
5500 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5502 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5503 per_cpu(sd_llc_size
, cpu
) = size
;
5504 per_cpu(sd_llc_id
, cpu
) = id
;
5506 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5507 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5509 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5510 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5514 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5515 * hold the hotplug lock.
5518 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5520 struct rq
*rq
= cpu_rq(cpu
);
5521 struct sched_domain
*tmp
;
5523 /* Remove the sched domains which do not contribute to scheduling. */
5524 for (tmp
= sd
; tmp
; ) {
5525 struct sched_domain
*parent
= tmp
->parent
;
5529 if (sd_parent_degenerate(tmp
, parent
)) {
5530 tmp
->parent
= parent
->parent
;
5532 parent
->parent
->child
= tmp
;
5534 * Transfer SD_PREFER_SIBLING down in case of a
5535 * degenerate parent; the spans match for this
5536 * so the property transfers.
5538 if (parent
->flags
& SD_PREFER_SIBLING
)
5539 tmp
->flags
|= SD_PREFER_SIBLING
;
5540 destroy_sched_domain(parent
, cpu
);
5545 if (sd
&& sd_degenerate(sd
)) {
5548 destroy_sched_domain(tmp
, cpu
);
5553 sched_domain_debug(sd
, cpu
);
5555 rq_attach_root(rq
, rd
);
5557 rcu_assign_pointer(rq
->sd
, sd
);
5558 destroy_sched_domains(tmp
, cpu
);
5560 update_top_cache_domain(cpu
);
5563 /* cpus with isolated domains */
5564 static cpumask_var_t cpu_isolated_map
;
5566 /* Setup the mask of cpus configured for isolated domains */
5567 static int __init
isolated_cpu_setup(char *str
)
5569 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5570 cpulist_parse(str
, cpu_isolated_map
);
5574 __setup("isolcpus=", isolated_cpu_setup
);
5576 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5578 return cpumask_of_node(cpu_to_node(cpu
));
5582 struct sched_domain
**__percpu sd
;
5583 struct sched_group
**__percpu sg
;
5584 struct sched_group_power
**__percpu sgp
;
5588 struct sched_domain
** __percpu sd
;
5589 struct root_domain
*rd
;
5599 struct sched_domain_topology_level
;
5601 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5602 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5604 #define SDTL_OVERLAP 0x01
5606 struct sched_domain_topology_level
{
5607 sched_domain_init_f init
;
5608 sched_domain_mask_f mask
;
5611 struct sd_data data
;
5615 * Build an iteration mask that can exclude certain CPUs from the upwards
5618 * Asymmetric node setups can result in situations where the domain tree is of
5619 * unequal depth, make sure to skip domains that already cover the entire
5622 * In that case build_sched_domains() will have terminated the iteration early
5623 * and our sibling sd spans will be empty. Domains should always include the
5624 * cpu they're built on, so check that.
5627 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5629 const struct cpumask
*span
= sched_domain_span(sd
);
5630 struct sd_data
*sdd
= sd
->private;
5631 struct sched_domain
*sibling
;
5634 for_each_cpu(i
, span
) {
5635 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5636 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5639 cpumask_set_cpu(i
, sched_group_mask(sg
));
5644 * Return the canonical balance cpu for this group, this is the first cpu
5645 * of this group that's also in the iteration mask.
5647 int group_balance_cpu(struct sched_group
*sg
)
5649 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5653 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5655 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5656 const struct cpumask
*span
= sched_domain_span(sd
);
5657 struct cpumask
*covered
= sched_domains_tmpmask
;
5658 struct sd_data
*sdd
= sd
->private;
5659 struct sched_domain
*child
;
5662 cpumask_clear(covered
);
5664 for_each_cpu(i
, span
) {
5665 struct cpumask
*sg_span
;
5667 if (cpumask_test_cpu(i
, covered
))
5670 child
= *per_cpu_ptr(sdd
->sd
, i
);
5672 /* See the comment near build_group_mask(). */
5673 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5676 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5677 GFP_KERNEL
, cpu_to_node(cpu
));
5682 sg_span
= sched_group_cpus(sg
);
5684 child
= child
->child
;
5685 cpumask_copy(sg_span
, sched_domain_span(child
));
5687 cpumask_set_cpu(i
, sg_span
);
5689 cpumask_or(covered
, covered
, sg_span
);
5691 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5692 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5693 build_group_mask(sd
, sg
);
5696 * Initialize sgp->power such that even if we mess up the
5697 * domains and no possible iteration will get us here, we won't
5700 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5701 sg
->sgp
->power_orig
= sg
->sgp
->power
;
5704 * Make sure the first group of this domain contains the
5705 * canonical balance cpu. Otherwise the sched_domain iteration
5706 * breaks. See update_sg_lb_stats().
5708 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5709 group_balance_cpu(sg
) == cpu
)
5719 sd
->groups
= groups
;
5724 free_sched_groups(first
, 0);
5729 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5731 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5732 struct sched_domain
*child
= sd
->child
;
5735 cpu
= cpumask_first(sched_domain_span(child
));
5738 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5739 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5740 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5747 * build_sched_groups will build a circular linked list of the groups
5748 * covered by the given span, and will set each group's ->cpumask correctly,
5749 * and ->cpu_power to 0.
5751 * Assumes the sched_domain tree is fully constructed
5754 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5756 struct sched_group
*first
= NULL
, *last
= NULL
;
5757 struct sd_data
*sdd
= sd
->private;
5758 const struct cpumask
*span
= sched_domain_span(sd
);
5759 struct cpumask
*covered
;
5762 get_group(cpu
, sdd
, &sd
->groups
);
5763 atomic_inc(&sd
->groups
->ref
);
5765 if (cpu
!= cpumask_first(span
))
5768 lockdep_assert_held(&sched_domains_mutex
);
5769 covered
= sched_domains_tmpmask
;
5771 cpumask_clear(covered
);
5773 for_each_cpu(i
, span
) {
5774 struct sched_group
*sg
;
5777 if (cpumask_test_cpu(i
, covered
))
5780 group
= get_group(i
, sdd
, &sg
);
5781 cpumask_clear(sched_group_cpus(sg
));
5783 cpumask_setall(sched_group_mask(sg
));
5785 for_each_cpu(j
, span
) {
5786 if (get_group(j
, sdd
, NULL
) != group
)
5789 cpumask_set_cpu(j
, covered
);
5790 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5805 * Initialize sched groups cpu_power.
5807 * cpu_power indicates the capacity of sched group, which is used while
5808 * distributing the load between different sched groups in a sched domain.
5809 * Typically cpu_power for all the groups in a sched domain will be same unless
5810 * there are asymmetries in the topology. If there are asymmetries, group
5811 * having more cpu_power will pickup more load compared to the group having
5814 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5816 struct sched_group
*sg
= sd
->groups
;
5821 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5823 } while (sg
!= sd
->groups
);
5825 if (cpu
!= group_balance_cpu(sg
))
5828 update_group_power(sd
, cpu
);
5829 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5832 int __weak
arch_sd_sibling_asym_packing(void)
5834 return 0*SD_ASYM_PACKING
;
5838 * Initializers for schedule domains
5839 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5842 #ifdef CONFIG_SCHED_DEBUG
5843 # define SD_INIT_NAME(sd, type) sd->name = #type
5845 # define SD_INIT_NAME(sd, type) do { } while (0)
5848 #define SD_INIT_FUNC(type) \
5849 static noinline struct sched_domain * \
5850 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5852 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5853 *sd = SD_##type##_INIT; \
5854 SD_INIT_NAME(sd, type); \
5855 sd->private = &tl->data; \
5860 #ifdef CONFIG_SCHED_SMT
5861 SD_INIT_FUNC(SIBLING
)
5863 #ifdef CONFIG_SCHED_MC
5866 #ifdef CONFIG_SCHED_BOOK
5870 static int default_relax_domain_level
= -1;
5871 int sched_domain_level_max
;
5873 static int __init
setup_relax_domain_level(char *str
)
5875 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5876 pr_warn("Unable to set relax_domain_level\n");
5880 __setup("relax_domain_level=", setup_relax_domain_level
);
5882 static void set_domain_attribute(struct sched_domain
*sd
,
5883 struct sched_domain_attr
*attr
)
5887 if (!attr
|| attr
->relax_domain_level
< 0) {
5888 if (default_relax_domain_level
< 0)
5891 request
= default_relax_domain_level
;
5893 request
= attr
->relax_domain_level
;
5894 if (request
< sd
->level
) {
5895 /* turn off idle balance on this domain */
5896 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5898 /* turn on idle balance on this domain */
5899 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5903 static void __sdt_free(const struct cpumask
*cpu_map
);
5904 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5906 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5907 const struct cpumask
*cpu_map
)
5911 if (!atomic_read(&d
->rd
->refcount
))
5912 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5914 free_percpu(d
->sd
); /* fall through */
5916 __sdt_free(cpu_map
); /* fall through */
5922 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5923 const struct cpumask
*cpu_map
)
5925 memset(d
, 0, sizeof(*d
));
5927 if (__sdt_alloc(cpu_map
))
5928 return sa_sd_storage
;
5929 d
->sd
= alloc_percpu(struct sched_domain
*);
5931 return sa_sd_storage
;
5932 d
->rd
= alloc_rootdomain();
5935 return sa_rootdomain
;
5939 * NULL the sd_data elements we've used to build the sched_domain and
5940 * sched_group structure so that the subsequent __free_domain_allocs()
5941 * will not free the data we're using.
5943 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5945 struct sd_data
*sdd
= sd
->private;
5947 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5948 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5950 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5951 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5953 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5954 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5957 #ifdef CONFIG_SCHED_SMT
5958 static const struct cpumask
*cpu_smt_mask(int cpu
)
5960 return topology_thread_cpumask(cpu
);
5965 * Topology list, bottom-up.
5967 static struct sched_domain_topology_level default_topology
[] = {
5968 #ifdef CONFIG_SCHED_SMT
5969 { sd_init_SIBLING
, cpu_smt_mask
, },
5971 #ifdef CONFIG_SCHED_MC
5972 { sd_init_MC
, cpu_coregroup_mask
, },
5974 #ifdef CONFIG_SCHED_BOOK
5975 { sd_init_BOOK
, cpu_book_mask
, },
5977 { sd_init_CPU
, cpu_cpu_mask
, },
5981 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5983 #define for_each_sd_topology(tl) \
5984 for (tl = sched_domain_topology; tl->init; tl++)
5988 static int sched_domains_numa_levels
;
5989 static int *sched_domains_numa_distance
;
5990 static struct cpumask
***sched_domains_numa_masks
;
5991 static int sched_domains_curr_level
;
5993 static inline int sd_local_flags(int level
)
5995 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5998 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6001 static struct sched_domain
*
6002 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6004 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6005 int level
= tl
->numa_level
;
6006 int sd_weight
= cpumask_weight(
6007 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6009 *sd
= (struct sched_domain
){
6010 .min_interval
= sd_weight
,
6011 .max_interval
= 2*sd_weight
,
6013 .imbalance_pct
= 125,
6014 .cache_nice_tries
= 2,
6021 .flags
= 1*SD_LOAD_BALANCE
6022 | 1*SD_BALANCE_NEWIDLE
6027 | 0*SD_SHARE_CPUPOWER
6028 | 0*SD_SHARE_PKG_RESOURCES
6030 | 0*SD_PREFER_SIBLING
6032 | sd_local_flags(level
)
6034 .last_balance
= jiffies
,
6035 .balance_interval
= sd_weight
,
6037 SD_INIT_NAME(sd
, NUMA
);
6038 sd
->private = &tl
->data
;
6041 * Ugly hack to pass state to sd_numa_mask()...
6043 sched_domains_curr_level
= tl
->numa_level
;
6048 static const struct cpumask
*sd_numa_mask(int cpu
)
6050 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6053 static void sched_numa_warn(const char *str
)
6055 static int done
= false;
6063 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6065 for (i
= 0; i
< nr_node_ids
; i
++) {
6066 printk(KERN_WARNING
" ");
6067 for (j
= 0; j
< nr_node_ids
; j
++)
6068 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6069 printk(KERN_CONT
"\n");
6071 printk(KERN_WARNING
"\n");
6074 static bool find_numa_distance(int distance
)
6078 if (distance
== node_distance(0, 0))
6081 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6082 if (sched_domains_numa_distance
[i
] == distance
)
6089 static void sched_init_numa(void)
6091 int next_distance
, curr_distance
= node_distance(0, 0);
6092 struct sched_domain_topology_level
*tl
;
6096 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6097 if (!sched_domains_numa_distance
)
6101 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6102 * unique distances in the node_distance() table.
6104 * Assumes node_distance(0,j) includes all distances in
6105 * node_distance(i,j) in order to avoid cubic time.
6107 next_distance
= curr_distance
;
6108 for (i
= 0; i
< nr_node_ids
; i
++) {
6109 for (j
= 0; j
< nr_node_ids
; j
++) {
6110 for (k
= 0; k
< nr_node_ids
; k
++) {
6111 int distance
= node_distance(i
, k
);
6113 if (distance
> curr_distance
&&
6114 (distance
< next_distance
||
6115 next_distance
== curr_distance
))
6116 next_distance
= distance
;
6119 * While not a strong assumption it would be nice to know
6120 * about cases where if node A is connected to B, B is not
6121 * equally connected to A.
6123 if (sched_debug() && node_distance(k
, i
) != distance
)
6124 sched_numa_warn("Node-distance not symmetric");
6126 if (sched_debug() && i
&& !find_numa_distance(distance
))
6127 sched_numa_warn("Node-0 not representative");
6129 if (next_distance
!= curr_distance
) {
6130 sched_domains_numa_distance
[level
++] = next_distance
;
6131 sched_domains_numa_levels
= level
;
6132 curr_distance
= next_distance
;
6137 * In case of sched_debug() we verify the above assumption.
6143 * 'level' contains the number of unique distances, excluding the
6144 * identity distance node_distance(i,i).
6146 * The sched_domains_numa_distance[] array includes the actual distance
6151 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6152 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6153 * the array will contain less then 'level' members. This could be
6154 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6155 * in other functions.
6157 * We reset it to 'level' at the end of this function.
6159 sched_domains_numa_levels
= 0;
6161 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6162 if (!sched_domains_numa_masks
)
6166 * Now for each level, construct a mask per node which contains all
6167 * cpus of nodes that are that many hops away from us.
6169 for (i
= 0; i
< level
; i
++) {
6170 sched_domains_numa_masks
[i
] =
6171 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6172 if (!sched_domains_numa_masks
[i
])
6175 for (j
= 0; j
< nr_node_ids
; j
++) {
6176 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6180 sched_domains_numa_masks
[i
][j
] = mask
;
6182 for (k
= 0; k
< nr_node_ids
; k
++) {
6183 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6186 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6191 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6192 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6197 * Copy the default topology bits..
6199 for (i
= 0; default_topology
[i
].init
; i
++)
6200 tl
[i
] = default_topology
[i
];
6203 * .. and append 'j' levels of NUMA goodness.
6205 for (j
= 0; j
< level
; i
++, j
++) {
6206 tl
[i
] = (struct sched_domain_topology_level
){
6207 .init
= sd_numa_init
,
6208 .mask
= sd_numa_mask
,
6209 .flags
= SDTL_OVERLAP
,
6214 sched_domain_topology
= tl
;
6216 sched_domains_numa_levels
= level
;
6219 static void sched_domains_numa_masks_set(int cpu
)
6222 int node
= cpu_to_node(cpu
);
6224 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6225 for (j
= 0; j
< nr_node_ids
; j
++) {
6226 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6227 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6232 static void sched_domains_numa_masks_clear(int cpu
)
6235 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6236 for (j
= 0; j
< nr_node_ids
; j
++)
6237 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6242 * Update sched_domains_numa_masks[level][node] array when new cpus
6245 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6246 unsigned long action
,
6249 int cpu
= (long)hcpu
;
6251 switch (action
& ~CPU_TASKS_FROZEN
) {
6253 sched_domains_numa_masks_set(cpu
);
6257 sched_domains_numa_masks_clear(cpu
);
6267 static inline void sched_init_numa(void)
6271 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6272 unsigned long action
,
6277 #endif /* CONFIG_NUMA */
6279 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6281 struct sched_domain_topology_level
*tl
;
6284 for_each_sd_topology(tl
) {
6285 struct sd_data
*sdd
= &tl
->data
;
6287 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6291 sdd
->sg
= alloc_percpu(struct sched_group
*);
6295 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6299 for_each_cpu(j
, cpu_map
) {
6300 struct sched_domain
*sd
;
6301 struct sched_group
*sg
;
6302 struct sched_group_power
*sgp
;
6304 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6305 GFP_KERNEL
, cpu_to_node(j
));
6309 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6311 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6312 GFP_KERNEL
, cpu_to_node(j
));
6318 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6320 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6321 GFP_KERNEL
, cpu_to_node(j
));
6325 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6332 static void __sdt_free(const struct cpumask
*cpu_map
)
6334 struct sched_domain_topology_level
*tl
;
6337 for_each_sd_topology(tl
) {
6338 struct sd_data
*sdd
= &tl
->data
;
6340 for_each_cpu(j
, cpu_map
) {
6341 struct sched_domain
*sd
;
6344 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6345 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6346 free_sched_groups(sd
->groups
, 0);
6347 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6351 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6353 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6355 free_percpu(sdd
->sd
);
6357 free_percpu(sdd
->sg
);
6359 free_percpu(sdd
->sgp
);
6364 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6365 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6366 struct sched_domain
*child
, int cpu
)
6368 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6372 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6374 sd
->level
= child
->level
+ 1;
6375 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6379 set_domain_attribute(sd
, attr
);
6385 * Build sched domains for a given set of cpus and attach the sched domains
6386 * to the individual cpus
6388 static int build_sched_domains(const struct cpumask
*cpu_map
,
6389 struct sched_domain_attr
*attr
)
6391 enum s_alloc alloc_state
;
6392 struct sched_domain
*sd
;
6394 int i
, ret
= -ENOMEM
;
6396 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6397 if (alloc_state
!= sa_rootdomain
)
6400 /* Set up domains for cpus specified by the cpu_map. */
6401 for_each_cpu(i
, cpu_map
) {
6402 struct sched_domain_topology_level
*tl
;
6405 for_each_sd_topology(tl
) {
6406 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6407 if (tl
== sched_domain_topology
)
6408 *per_cpu_ptr(d
.sd
, i
) = sd
;
6409 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6410 sd
->flags
|= SD_OVERLAP
;
6411 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6416 /* Build the groups for the domains */
6417 for_each_cpu(i
, cpu_map
) {
6418 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6419 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6420 if (sd
->flags
& SD_OVERLAP
) {
6421 if (build_overlap_sched_groups(sd
, i
))
6424 if (build_sched_groups(sd
, i
))
6430 /* Calculate CPU power for physical packages and nodes */
6431 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6432 if (!cpumask_test_cpu(i
, cpu_map
))
6435 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6436 claim_allocations(i
, sd
);
6437 init_sched_groups_power(i
, sd
);
6441 /* Attach the domains */
6443 for_each_cpu(i
, cpu_map
) {
6444 sd
= *per_cpu_ptr(d
.sd
, i
);
6445 cpu_attach_domain(sd
, d
.rd
, i
);
6451 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6455 static cpumask_var_t
*doms_cur
; /* current sched domains */
6456 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6457 static struct sched_domain_attr
*dattr_cur
;
6458 /* attribues of custom domains in 'doms_cur' */
6461 * Special case: If a kmalloc of a doms_cur partition (array of
6462 * cpumask) fails, then fallback to a single sched domain,
6463 * as determined by the single cpumask fallback_doms.
6465 static cpumask_var_t fallback_doms
;
6468 * arch_update_cpu_topology lets virtualized architectures update the
6469 * cpu core maps. It is supposed to return 1 if the topology changed
6470 * or 0 if it stayed the same.
6472 int __weak
arch_update_cpu_topology(void)
6477 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6480 cpumask_var_t
*doms
;
6482 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6485 for (i
= 0; i
< ndoms
; i
++) {
6486 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6487 free_sched_domains(doms
, i
);
6494 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6497 for (i
= 0; i
< ndoms
; i
++)
6498 free_cpumask_var(doms
[i
]);
6503 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6504 * For now this just excludes isolated cpus, but could be used to
6505 * exclude other special cases in the future.
6507 static int init_sched_domains(const struct cpumask
*cpu_map
)
6511 arch_update_cpu_topology();
6513 doms_cur
= alloc_sched_domains(ndoms_cur
);
6515 doms_cur
= &fallback_doms
;
6516 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6517 err
= build_sched_domains(doms_cur
[0], NULL
);
6518 register_sched_domain_sysctl();
6524 * Detach sched domains from a group of cpus specified in cpu_map
6525 * These cpus will now be attached to the NULL domain
6527 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6532 for_each_cpu(i
, cpu_map
)
6533 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6537 /* handle null as "default" */
6538 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6539 struct sched_domain_attr
*new, int idx_new
)
6541 struct sched_domain_attr tmp
;
6548 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6549 new ? (new + idx_new
) : &tmp
,
6550 sizeof(struct sched_domain_attr
));
6554 * Partition sched domains as specified by the 'ndoms_new'
6555 * cpumasks in the array doms_new[] of cpumasks. This compares
6556 * doms_new[] to the current sched domain partitioning, doms_cur[].
6557 * It destroys each deleted domain and builds each new domain.
6559 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6560 * The masks don't intersect (don't overlap.) We should setup one
6561 * sched domain for each mask. CPUs not in any of the cpumasks will
6562 * not be load balanced. If the same cpumask appears both in the
6563 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6566 * The passed in 'doms_new' should be allocated using
6567 * alloc_sched_domains. This routine takes ownership of it and will
6568 * free_sched_domains it when done with it. If the caller failed the
6569 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6570 * and partition_sched_domains() will fallback to the single partition
6571 * 'fallback_doms', it also forces the domains to be rebuilt.
6573 * If doms_new == NULL it will be replaced with cpu_online_mask.
6574 * ndoms_new == 0 is a special case for destroying existing domains,
6575 * and it will not create the default domain.
6577 * Call with hotplug lock held
6579 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6580 struct sched_domain_attr
*dattr_new
)
6585 mutex_lock(&sched_domains_mutex
);
6587 /* always unregister in case we don't destroy any domains */
6588 unregister_sched_domain_sysctl();
6590 /* Let architecture update cpu core mappings. */
6591 new_topology
= arch_update_cpu_topology();
6593 n
= doms_new
? ndoms_new
: 0;
6595 /* Destroy deleted domains */
6596 for (i
= 0; i
< ndoms_cur
; i
++) {
6597 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6598 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6599 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6602 /* no match - a current sched domain not in new doms_new[] */
6603 detach_destroy_domains(doms_cur
[i
]);
6609 if (doms_new
== NULL
) {
6611 doms_new
= &fallback_doms
;
6612 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6613 WARN_ON_ONCE(dattr_new
);
6616 /* Build new domains */
6617 for (i
= 0; i
< ndoms_new
; i
++) {
6618 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6619 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6620 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6623 /* no match - add a new doms_new */
6624 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6629 /* Remember the new sched domains */
6630 if (doms_cur
!= &fallback_doms
)
6631 free_sched_domains(doms_cur
, ndoms_cur
);
6632 kfree(dattr_cur
); /* kfree(NULL) is safe */
6633 doms_cur
= doms_new
;
6634 dattr_cur
= dattr_new
;
6635 ndoms_cur
= ndoms_new
;
6637 register_sched_domain_sysctl();
6639 mutex_unlock(&sched_domains_mutex
);
6642 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6645 * Update cpusets according to cpu_active mask. If cpusets are
6646 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6647 * around partition_sched_domains().
6649 * If we come here as part of a suspend/resume, don't touch cpusets because we
6650 * want to restore it back to its original state upon resume anyway.
6652 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6656 case CPU_ONLINE_FROZEN
:
6657 case CPU_DOWN_FAILED_FROZEN
:
6660 * num_cpus_frozen tracks how many CPUs are involved in suspend
6661 * resume sequence. As long as this is not the last online
6662 * operation in the resume sequence, just build a single sched
6663 * domain, ignoring cpusets.
6666 if (likely(num_cpus_frozen
)) {
6667 partition_sched_domains(1, NULL
, NULL
);
6672 * This is the last CPU online operation. So fall through and
6673 * restore the original sched domains by considering the
6674 * cpuset configurations.
6678 case CPU_DOWN_FAILED
:
6679 cpuset_update_active_cpus(true);
6687 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6691 case CPU_DOWN_PREPARE
:
6692 cpuset_update_active_cpus(false);
6694 case CPU_DOWN_PREPARE_FROZEN
:
6696 partition_sched_domains(1, NULL
, NULL
);
6704 void __init
sched_init_smp(void)
6706 cpumask_var_t non_isolated_cpus
;
6708 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6709 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6714 * There's no userspace yet to cause hotplug operations; hence all the
6715 * cpu masks are stable and all blatant races in the below code cannot
6718 mutex_lock(&sched_domains_mutex
);
6719 init_sched_domains(cpu_active_mask
);
6720 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6721 if (cpumask_empty(non_isolated_cpus
))
6722 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6723 mutex_unlock(&sched_domains_mutex
);
6725 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6726 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6727 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6731 /* Move init over to a non-isolated CPU */
6732 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6734 sched_init_granularity();
6735 free_cpumask_var(non_isolated_cpus
);
6737 init_sched_rt_class();
6738 init_sched_dl_class();
6741 void __init
sched_init_smp(void)
6743 sched_init_granularity();
6745 #endif /* CONFIG_SMP */
6747 const_debug
unsigned int sysctl_timer_migration
= 1;
6749 int in_sched_functions(unsigned long addr
)
6751 return in_lock_functions(addr
) ||
6752 (addr
>= (unsigned long)__sched_text_start
6753 && addr
< (unsigned long)__sched_text_end
);
6756 #ifdef CONFIG_CGROUP_SCHED
6758 * Default task group.
6759 * Every task in system belongs to this group at bootup.
6761 struct task_group root_task_group
;
6762 LIST_HEAD(task_groups
);
6765 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6767 void __init
sched_init(void)
6770 unsigned long alloc_size
= 0, ptr
;
6772 #ifdef CONFIG_FAIR_GROUP_SCHED
6773 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6775 #ifdef CONFIG_RT_GROUP_SCHED
6776 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6778 #ifdef CONFIG_CPUMASK_OFFSTACK
6779 alloc_size
+= num_possible_cpus() * cpumask_size();
6782 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6784 #ifdef CONFIG_FAIR_GROUP_SCHED
6785 root_task_group
.se
= (struct sched_entity
**)ptr
;
6786 ptr
+= nr_cpu_ids
* sizeof(void **);
6788 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6789 ptr
+= nr_cpu_ids
* sizeof(void **);
6791 #endif /* CONFIG_FAIR_GROUP_SCHED */
6792 #ifdef CONFIG_RT_GROUP_SCHED
6793 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6794 ptr
+= nr_cpu_ids
* sizeof(void **);
6796 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6797 ptr
+= nr_cpu_ids
* sizeof(void **);
6799 #endif /* CONFIG_RT_GROUP_SCHED */
6800 #ifdef CONFIG_CPUMASK_OFFSTACK
6801 for_each_possible_cpu(i
) {
6802 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6803 ptr
+= cpumask_size();
6805 #endif /* CONFIG_CPUMASK_OFFSTACK */
6808 init_rt_bandwidth(&def_rt_bandwidth
,
6809 global_rt_period(), global_rt_runtime());
6810 init_dl_bandwidth(&def_dl_bandwidth
,
6811 global_rt_period(), global_rt_runtime());
6814 init_defrootdomain();
6817 #ifdef CONFIG_RT_GROUP_SCHED
6818 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6819 global_rt_period(), global_rt_runtime());
6820 #endif /* CONFIG_RT_GROUP_SCHED */
6822 #ifdef CONFIG_CGROUP_SCHED
6823 list_add(&root_task_group
.list
, &task_groups
);
6824 INIT_LIST_HEAD(&root_task_group
.children
);
6825 INIT_LIST_HEAD(&root_task_group
.siblings
);
6826 autogroup_init(&init_task
);
6828 #endif /* CONFIG_CGROUP_SCHED */
6830 for_each_possible_cpu(i
) {
6834 raw_spin_lock_init(&rq
->lock
);
6836 rq
->calc_load_active
= 0;
6837 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6838 init_cfs_rq(&rq
->cfs
);
6839 init_rt_rq(&rq
->rt
, rq
);
6840 init_dl_rq(&rq
->dl
, rq
);
6841 #ifdef CONFIG_FAIR_GROUP_SCHED
6842 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6843 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6845 * How much cpu bandwidth does root_task_group get?
6847 * In case of task-groups formed thr' the cgroup filesystem, it
6848 * gets 100% of the cpu resources in the system. This overall
6849 * system cpu resource is divided among the tasks of
6850 * root_task_group and its child task-groups in a fair manner,
6851 * based on each entity's (task or task-group's) weight
6852 * (se->load.weight).
6854 * In other words, if root_task_group has 10 tasks of weight
6855 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6856 * then A0's share of the cpu resource is:
6858 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6860 * We achieve this by letting root_task_group's tasks sit
6861 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6863 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6864 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6865 #endif /* CONFIG_FAIR_GROUP_SCHED */
6867 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6868 #ifdef CONFIG_RT_GROUP_SCHED
6869 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6872 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6873 rq
->cpu_load
[j
] = 0;
6875 rq
->last_load_update_tick
= jiffies
;
6880 rq
->cpu_power
= SCHED_POWER_SCALE
;
6881 rq
->post_schedule
= 0;
6882 rq
->active_balance
= 0;
6883 rq
->next_balance
= jiffies
;
6888 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6889 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6891 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6893 rq_attach_root(rq
, &def_root_domain
);
6894 #ifdef CONFIG_NO_HZ_COMMON
6897 #ifdef CONFIG_NO_HZ_FULL
6898 rq
->last_sched_tick
= 0;
6902 atomic_set(&rq
->nr_iowait
, 0);
6905 set_load_weight(&init_task
);
6907 #ifdef CONFIG_PREEMPT_NOTIFIERS
6908 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6912 * The boot idle thread does lazy MMU switching as well:
6914 atomic_inc(&init_mm
.mm_count
);
6915 enter_lazy_tlb(&init_mm
, current
);
6918 * Make us the idle thread. Technically, schedule() should not be
6919 * called from this thread, however somewhere below it might be,
6920 * but because we are the idle thread, we just pick up running again
6921 * when this runqueue becomes "idle".
6923 init_idle(current
, smp_processor_id());
6925 calc_load_update
= jiffies
+ LOAD_FREQ
;
6928 * During early bootup we pretend to be a normal task:
6930 current
->sched_class
= &fair_sched_class
;
6933 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6934 /* May be allocated at isolcpus cmdline parse time */
6935 if (cpu_isolated_map
== NULL
)
6936 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6937 idle_thread_set_boot_cpu();
6939 init_sched_fair_class();
6941 scheduler_running
= 1;
6944 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6945 static inline int preempt_count_equals(int preempt_offset
)
6947 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6949 return (nested
== preempt_offset
);
6952 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6954 static unsigned long prev_jiffy
; /* ratelimiting */
6956 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6957 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6958 !is_idle_task(current
)) ||
6959 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6961 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6963 prev_jiffy
= jiffies
;
6966 "BUG: sleeping function called from invalid context at %s:%d\n",
6969 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6970 in_atomic(), irqs_disabled(),
6971 current
->pid
, current
->comm
);
6973 debug_show_held_locks(current
);
6974 if (irqs_disabled())
6975 print_irqtrace_events(current
);
6976 #ifdef CONFIG_DEBUG_PREEMPT
6977 if (!preempt_count_equals(preempt_offset
)) {
6978 pr_err("Preemption disabled at:");
6979 print_ip_sym(current
->preempt_disable_ip
);
6985 EXPORT_SYMBOL(__might_sleep
);
6988 #ifdef CONFIG_MAGIC_SYSRQ
6989 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6991 const struct sched_class
*prev_class
= p
->sched_class
;
6992 struct sched_attr attr
= {
6993 .sched_policy
= SCHED_NORMAL
,
6995 int old_prio
= p
->prio
;
7000 dequeue_task(rq
, p
, 0);
7001 __setscheduler(rq
, p
, &attr
);
7003 enqueue_task(rq
, p
, 0);
7004 resched_task(rq
->curr
);
7007 check_class_changed(rq
, p
, prev_class
, old_prio
);
7010 void normalize_rt_tasks(void)
7012 struct task_struct
*g
, *p
;
7013 unsigned long flags
;
7016 read_lock_irqsave(&tasklist_lock
, flags
);
7017 do_each_thread(g
, p
) {
7019 * Only normalize user tasks:
7024 p
->se
.exec_start
= 0;
7025 #ifdef CONFIG_SCHEDSTATS
7026 p
->se
.statistics
.wait_start
= 0;
7027 p
->se
.statistics
.sleep_start
= 0;
7028 p
->se
.statistics
.block_start
= 0;
7031 if (!dl_task(p
) && !rt_task(p
)) {
7033 * Renice negative nice level userspace
7036 if (task_nice(p
) < 0 && p
->mm
)
7037 set_user_nice(p
, 0);
7041 raw_spin_lock(&p
->pi_lock
);
7042 rq
= __task_rq_lock(p
);
7044 normalize_task(rq
, p
);
7046 __task_rq_unlock(rq
);
7047 raw_spin_unlock(&p
->pi_lock
);
7048 } while_each_thread(g
, p
);
7050 read_unlock_irqrestore(&tasklist_lock
, flags
);
7053 #endif /* CONFIG_MAGIC_SYSRQ */
7055 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7057 * These functions are only useful for the IA64 MCA handling, or kdb.
7059 * They can only be called when the whole system has been
7060 * stopped - every CPU needs to be quiescent, and no scheduling
7061 * activity can take place. Using them for anything else would
7062 * be a serious bug, and as a result, they aren't even visible
7063 * under any other configuration.
7067 * curr_task - return the current task for a given cpu.
7068 * @cpu: the processor in question.
7070 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7072 * Return: The current task for @cpu.
7074 struct task_struct
*curr_task(int cpu
)
7076 return cpu_curr(cpu
);
7079 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7083 * set_curr_task - set the current task for a given cpu.
7084 * @cpu: the processor in question.
7085 * @p: the task pointer to set.
7087 * Description: This function must only be used when non-maskable interrupts
7088 * are serviced on a separate stack. It allows the architecture to switch the
7089 * notion of the current task on a cpu in a non-blocking manner. This function
7090 * must be called with all CPU's synchronized, and interrupts disabled, the
7091 * and caller must save the original value of the current task (see
7092 * curr_task() above) and restore that value before reenabling interrupts and
7093 * re-starting the system.
7095 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7097 void set_curr_task(int cpu
, struct task_struct
*p
)
7104 #ifdef CONFIG_CGROUP_SCHED
7105 /* task_group_lock serializes the addition/removal of task groups */
7106 static DEFINE_SPINLOCK(task_group_lock
);
7108 static void free_sched_group(struct task_group
*tg
)
7110 free_fair_sched_group(tg
);
7111 free_rt_sched_group(tg
);
7116 /* allocate runqueue etc for a new task group */
7117 struct task_group
*sched_create_group(struct task_group
*parent
)
7119 struct task_group
*tg
;
7121 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7123 return ERR_PTR(-ENOMEM
);
7125 if (!alloc_fair_sched_group(tg
, parent
))
7128 if (!alloc_rt_sched_group(tg
, parent
))
7134 free_sched_group(tg
);
7135 return ERR_PTR(-ENOMEM
);
7138 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7140 unsigned long flags
;
7142 spin_lock_irqsave(&task_group_lock
, flags
);
7143 list_add_rcu(&tg
->list
, &task_groups
);
7145 WARN_ON(!parent
); /* root should already exist */
7147 tg
->parent
= parent
;
7148 INIT_LIST_HEAD(&tg
->children
);
7149 list_add_rcu(&tg
->siblings
, &parent
->children
);
7150 spin_unlock_irqrestore(&task_group_lock
, flags
);
7153 /* rcu callback to free various structures associated with a task group */
7154 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7156 /* now it should be safe to free those cfs_rqs */
7157 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7160 /* Destroy runqueue etc associated with a task group */
7161 void sched_destroy_group(struct task_group
*tg
)
7163 /* wait for possible concurrent references to cfs_rqs complete */
7164 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7167 void sched_offline_group(struct task_group
*tg
)
7169 unsigned long flags
;
7172 /* end participation in shares distribution */
7173 for_each_possible_cpu(i
)
7174 unregister_fair_sched_group(tg
, i
);
7176 spin_lock_irqsave(&task_group_lock
, flags
);
7177 list_del_rcu(&tg
->list
);
7178 list_del_rcu(&tg
->siblings
);
7179 spin_unlock_irqrestore(&task_group_lock
, flags
);
7182 /* change task's runqueue when it moves between groups.
7183 * The caller of this function should have put the task in its new group
7184 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7185 * reflect its new group.
7187 void sched_move_task(struct task_struct
*tsk
)
7189 struct task_group
*tg
;
7191 unsigned long flags
;
7194 rq
= task_rq_lock(tsk
, &flags
);
7196 running
= task_current(rq
, tsk
);
7200 dequeue_task(rq
, tsk
, 0);
7201 if (unlikely(running
))
7202 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7204 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
,
7205 lockdep_is_held(&tsk
->sighand
->siglock
)),
7206 struct task_group
, css
);
7207 tg
= autogroup_task_group(tsk
, tg
);
7208 tsk
->sched_task_group
= tg
;
7210 #ifdef CONFIG_FAIR_GROUP_SCHED
7211 if (tsk
->sched_class
->task_move_group
)
7212 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7215 set_task_rq(tsk
, task_cpu(tsk
));
7217 if (unlikely(running
))
7218 tsk
->sched_class
->set_curr_task(rq
);
7220 enqueue_task(rq
, tsk
, 0);
7222 task_rq_unlock(rq
, tsk
, &flags
);
7224 #endif /* CONFIG_CGROUP_SCHED */
7226 #ifdef CONFIG_RT_GROUP_SCHED
7228 * Ensure that the real time constraints are schedulable.
7230 static DEFINE_MUTEX(rt_constraints_mutex
);
7232 /* Must be called with tasklist_lock held */
7233 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7235 struct task_struct
*g
, *p
;
7237 do_each_thread(g
, p
) {
7238 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7240 } while_each_thread(g
, p
);
7245 struct rt_schedulable_data
{
7246 struct task_group
*tg
;
7251 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7253 struct rt_schedulable_data
*d
= data
;
7254 struct task_group
*child
;
7255 unsigned long total
, sum
= 0;
7256 u64 period
, runtime
;
7258 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7259 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7262 period
= d
->rt_period
;
7263 runtime
= d
->rt_runtime
;
7267 * Cannot have more runtime than the period.
7269 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7273 * Ensure we don't starve existing RT tasks.
7275 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7278 total
= to_ratio(period
, runtime
);
7281 * Nobody can have more than the global setting allows.
7283 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7287 * The sum of our children's runtime should not exceed our own.
7289 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7290 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7291 runtime
= child
->rt_bandwidth
.rt_runtime
;
7293 if (child
== d
->tg
) {
7294 period
= d
->rt_period
;
7295 runtime
= d
->rt_runtime
;
7298 sum
+= to_ratio(period
, runtime
);
7307 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7311 struct rt_schedulable_data data
= {
7313 .rt_period
= period
,
7314 .rt_runtime
= runtime
,
7318 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7324 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7325 u64 rt_period
, u64 rt_runtime
)
7329 mutex_lock(&rt_constraints_mutex
);
7330 read_lock(&tasklist_lock
);
7331 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7335 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7336 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7337 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7339 for_each_possible_cpu(i
) {
7340 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7342 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7343 rt_rq
->rt_runtime
= rt_runtime
;
7344 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7346 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7348 read_unlock(&tasklist_lock
);
7349 mutex_unlock(&rt_constraints_mutex
);
7354 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7356 u64 rt_runtime
, rt_period
;
7358 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7359 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7360 if (rt_runtime_us
< 0)
7361 rt_runtime
= RUNTIME_INF
;
7363 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7366 static long sched_group_rt_runtime(struct task_group
*tg
)
7370 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7373 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7374 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7375 return rt_runtime_us
;
7378 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7380 u64 rt_runtime
, rt_period
;
7382 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7383 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7388 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7391 static long sched_group_rt_period(struct task_group
*tg
)
7395 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7396 do_div(rt_period_us
, NSEC_PER_USEC
);
7397 return rt_period_us
;
7399 #endif /* CONFIG_RT_GROUP_SCHED */
7401 #ifdef CONFIG_RT_GROUP_SCHED
7402 static int sched_rt_global_constraints(void)
7406 mutex_lock(&rt_constraints_mutex
);
7407 read_lock(&tasklist_lock
);
7408 ret
= __rt_schedulable(NULL
, 0, 0);
7409 read_unlock(&tasklist_lock
);
7410 mutex_unlock(&rt_constraints_mutex
);
7415 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7417 /* Don't accept realtime tasks when there is no way for them to run */
7418 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7424 #else /* !CONFIG_RT_GROUP_SCHED */
7425 static int sched_rt_global_constraints(void)
7427 unsigned long flags
;
7430 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7431 for_each_possible_cpu(i
) {
7432 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7434 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7435 rt_rq
->rt_runtime
= global_rt_runtime();
7436 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7438 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7442 #endif /* CONFIG_RT_GROUP_SCHED */
7444 static int sched_dl_global_constraints(void)
7446 u64 runtime
= global_rt_runtime();
7447 u64 period
= global_rt_period();
7448 u64 new_bw
= to_ratio(period
, runtime
);
7450 unsigned long flags
;
7453 * Here we want to check the bandwidth not being set to some
7454 * value smaller than the currently allocated bandwidth in
7455 * any of the root_domains.
7457 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7458 * cycling on root_domains... Discussion on different/better
7459 * solutions is welcome!
7461 for_each_possible_cpu(cpu
) {
7462 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7464 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7465 if (new_bw
< dl_b
->total_bw
)
7467 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7476 static void sched_dl_do_global(void)
7480 unsigned long flags
;
7482 def_dl_bandwidth
.dl_period
= global_rt_period();
7483 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7485 if (global_rt_runtime() != RUNTIME_INF
)
7486 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7489 * FIXME: As above...
7491 for_each_possible_cpu(cpu
) {
7492 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7494 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7496 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7500 static int sched_rt_global_validate(void)
7502 if (sysctl_sched_rt_period
<= 0)
7505 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7506 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7512 static void sched_rt_do_global(void)
7514 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7515 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7518 int sched_rt_handler(struct ctl_table
*table
, int write
,
7519 void __user
*buffer
, size_t *lenp
,
7522 int old_period
, old_runtime
;
7523 static DEFINE_MUTEX(mutex
);
7527 old_period
= sysctl_sched_rt_period
;
7528 old_runtime
= sysctl_sched_rt_runtime
;
7530 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7532 if (!ret
&& write
) {
7533 ret
= sched_rt_global_validate();
7537 ret
= sched_rt_global_constraints();
7541 ret
= sched_dl_global_constraints();
7545 sched_rt_do_global();
7546 sched_dl_do_global();
7550 sysctl_sched_rt_period
= old_period
;
7551 sysctl_sched_rt_runtime
= old_runtime
;
7553 mutex_unlock(&mutex
);
7558 int sched_rr_handler(struct ctl_table
*table
, int write
,
7559 void __user
*buffer
, size_t *lenp
,
7563 static DEFINE_MUTEX(mutex
);
7566 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7567 /* make sure that internally we keep jiffies */
7568 /* also, writing zero resets timeslice to default */
7569 if (!ret
&& write
) {
7570 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7571 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7573 mutex_unlock(&mutex
);
7577 #ifdef CONFIG_CGROUP_SCHED
7579 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7581 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7584 static struct cgroup_subsys_state
*
7585 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7587 struct task_group
*parent
= css_tg(parent_css
);
7588 struct task_group
*tg
;
7591 /* This is early initialization for the top cgroup */
7592 return &root_task_group
.css
;
7595 tg
= sched_create_group(parent
);
7597 return ERR_PTR(-ENOMEM
);
7602 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7604 struct task_group
*tg
= css_tg(css
);
7605 struct task_group
*parent
= css_tg(css_parent(css
));
7608 sched_online_group(tg
, parent
);
7612 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7614 struct task_group
*tg
= css_tg(css
);
7616 sched_destroy_group(tg
);
7619 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7621 struct task_group
*tg
= css_tg(css
);
7623 sched_offline_group(tg
);
7626 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7627 struct cgroup_taskset
*tset
)
7629 struct task_struct
*task
;
7631 cgroup_taskset_for_each(task
, tset
) {
7632 #ifdef CONFIG_RT_GROUP_SCHED
7633 if (!sched_rt_can_attach(css_tg(css
), task
))
7636 /* We don't support RT-tasks being in separate groups */
7637 if (task
->sched_class
!= &fair_sched_class
)
7644 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7645 struct cgroup_taskset
*tset
)
7647 struct task_struct
*task
;
7649 cgroup_taskset_for_each(task
, tset
)
7650 sched_move_task(task
);
7653 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7654 struct cgroup_subsys_state
*old_css
,
7655 struct task_struct
*task
)
7658 * cgroup_exit() is called in the copy_process() failure path.
7659 * Ignore this case since the task hasn't ran yet, this avoids
7660 * trying to poke a half freed task state from generic code.
7662 if (!(task
->flags
& PF_EXITING
))
7665 sched_move_task(task
);
7668 #ifdef CONFIG_FAIR_GROUP_SCHED
7669 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7670 struct cftype
*cftype
, u64 shareval
)
7672 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7675 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7678 struct task_group
*tg
= css_tg(css
);
7680 return (u64
) scale_load_down(tg
->shares
);
7683 #ifdef CONFIG_CFS_BANDWIDTH
7684 static DEFINE_MUTEX(cfs_constraints_mutex
);
7686 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7687 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7689 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7691 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7693 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7694 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7696 if (tg
== &root_task_group
)
7700 * Ensure we have at some amount of bandwidth every period. This is
7701 * to prevent reaching a state of large arrears when throttled via
7702 * entity_tick() resulting in prolonged exit starvation.
7704 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7708 * Likewise, bound things on the otherside by preventing insane quota
7709 * periods. This also allows us to normalize in computing quota
7712 if (period
> max_cfs_quota_period
)
7715 mutex_lock(&cfs_constraints_mutex
);
7716 ret
= __cfs_schedulable(tg
, period
, quota
);
7720 runtime_enabled
= quota
!= RUNTIME_INF
;
7721 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7723 * If we need to toggle cfs_bandwidth_used, off->on must occur
7724 * before making related changes, and on->off must occur afterwards
7726 if (runtime_enabled
&& !runtime_was_enabled
)
7727 cfs_bandwidth_usage_inc();
7728 raw_spin_lock_irq(&cfs_b
->lock
);
7729 cfs_b
->period
= ns_to_ktime(period
);
7730 cfs_b
->quota
= quota
;
7732 __refill_cfs_bandwidth_runtime(cfs_b
);
7733 /* restart the period timer (if active) to handle new period expiry */
7734 if (runtime_enabled
&& cfs_b
->timer_active
) {
7735 /* force a reprogram */
7736 cfs_b
->timer_active
= 0;
7737 __start_cfs_bandwidth(cfs_b
);
7739 raw_spin_unlock_irq(&cfs_b
->lock
);
7741 for_each_possible_cpu(i
) {
7742 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7743 struct rq
*rq
= cfs_rq
->rq
;
7745 raw_spin_lock_irq(&rq
->lock
);
7746 cfs_rq
->runtime_enabled
= runtime_enabled
;
7747 cfs_rq
->runtime_remaining
= 0;
7749 if (cfs_rq
->throttled
)
7750 unthrottle_cfs_rq(cfs_rq
);
7751 raw_spin_unlock_irq(&rq
->lock
);
7753 if (runtime_was_enabled
&& !runtime_enabled
)
7754 cfs_bandwidth_usage_dec();
7756 mutex_unlock(&cfs_constraints_mutex
);
7761 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7765 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7766 if (cfs_quota_us
< 0)
7767 quota
= RUNTIME_INF
;
7769 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7771 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7774 long tg_get_cfs_quota(struct task_group
*tg
)
7778 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7781 quota_us
= tg
->cfs_bandwidth
.quota
;
7782 do_div(quota_us
, NSEC_PER_USEC
);
7787 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7791 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7792 quota
= tg
->cfs_bandwidth
.quota
;
7794 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7797 long tg_get_cfs_period(struct task_group
*tg
)
7801 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7802 do_div(cfs_period_us
, NSEC_PER_USEC
);
7804 return cfs_period_us
;
7807 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7810 return tg_get_cfs_quota(css_tg(css
));
7813 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7814 struct cftype
*cftype
, s64 cfs_quota_us
)
7816 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7819 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7822 return tg_get_cfs_period(css_tg(css
));
7825 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7826 struct cftype
*cftype
, u64 cfs_period_us
)
7828 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7831 struct cfs_schedulable_data
{
7832 struct task_group
*tg
;
7837 * normalize group quota/period to be quota/max_period
7838 * note: units are usecs
7840 static u64
normalize_cfs_quota(struct task_group
*tg
,
7841 struct cfs_schedulable_data
*d
)
7849 period
= tg_get_cfs_period(tg
);
7850 quota
= tg_get_cfs_quota(tg
);
7853 /* note: these should typically be equivalent */
7854 if (quota
== RUNTIME_INF
|| quota
== -1)
7857 return to_ratio(period
, quota
);
7860 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7862 struct cfs_schedulable_data
*d
= data
;
7863 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7864 s64 quota
= 0, parent_quota
= -1;
7867 quota
= RUNTIME_INF
;
7869 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7871 quota
= normalize_cfs_quota(tg
, d
);
7872 parent_quota
= parent_b
->hierarchal_quota
;
7875 * ensure max(child_quota) <= parent_quota, inherit when no
7878 if (quota
== RUNTIME_INF
)
7879 quota
= parent_quota
;
7880 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7883 cfs_b
->hierarchal_quota
= quota
;
7888 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7891 struct cfs_schedulable_data data
= {
7897 if (quota
!= RUNTIME_INF
) {
7898 do_div(data
.period
, NSEC_PER_USEC
);
7899 do_div(data
.quota
, NSEC_PER_USEC
);
7903 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7909 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7911 struct task_group
*tg
= css_tg(seq_css(sf
));
7912 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7914 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7915 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7916 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7920 #endif /* CONFIG_CFS_BANDWIDTH */
7921 #endif /* CONFIG_FAIR_GROUP_SCHED */
7923 #ifdef CONFIG_RT_GROUP_SCHED
7924 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7925 struct cftype
*cft
, s64 val
)
7927 return sched_group_set_rt_runtime(css_tg(css
), val
);
7930 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7933 return sched_group_rt_runtime(css_tg(css
));
7936 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7937 struct cftype
*cftype
, u64 rt_period_us
)
7939 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7942 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7945 return sched_group_rt_period(css_tg(css
));
7947 #endif /* CONFIG_RT_GROUP_SCHED */
7949 static struct cftype cpu_files
[] = {
7950 #ifdef CONFIG_FAIR_GROUP_SCHED
7953 .read_u64
= cpu_shares_read_u64
,
7954 .write_u64
= cpu_shares_write_u64
,
7957 #ifdef CONFIG_CFS_BANDWIDTH
7959 .name
= "cfs_quota_us",
7960 .read_s64
= cpu_cfs_quota_read_s64
,
7961 .write_s64
= cpu_cfs_quota_write_s64
,
7964 .name
= "cfs_period_us",
7965 .read_u64
= cpu_cfs_period_read_u64
,
7966 .write_u64
= cpu_cfs_period_write_u64
,
7970 .seq_show
= cpu_stats_show
,
7973 #ifdef CONFIG_RT_GROUP_SCHED
7975 .name
= "rt_runtime_us",
7976 .read_s64
= cpu_rt_runtime_read
,
7977 .write_s64
= cpu_rt_runtime_write
,
7980 .name
= "rt_period_us",
7981 .read_u64
= cpu_rt_period_read_uint
,
7982 .write_u64
= cpu_rt_period_write_uint
,
7988 struct cgroup_subsys cpu_cgrp_subsys
= {
7989 .css_alloc
= cpu_cgroup_css_alloc
,
7990 .css_free
= cpu_cgroup_css_free
,
7991 .css_online
= cpu_cgroup_css_online
,
7992 .css_offline
= cpu_cgroup_css_offline
,
7993 .can_attach
= cpu_cgroup_can_attach
,
7994 .attach
= cpu_cgroup_attach
,
7995 .exit
= cpu_cgroup_exit
,
7996 .base_cftypes
= cpu_files
,
8000 #endif /* CONFIG_CGROUP_SCHED */
8002 void dump_cpu_task(int cpu
)
8004 pr_info("Task dump for CPU %d:\n", cpu
);
8005 sched_show_task(cpu_curr(cpu
));