4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
95 ktime_t soft
, hard
, now
;
98 if (hrtimer_active(period_timer
))
101 now
= hrtimer_cb_get_time(period_timer
);
102 hrtimer_forward(period_timer
, now
, period
);
104 soft
= hrtimer_get_softexpires(period_timer
);
105 hard
= hrtimer_get_expires(period_timer
);
106 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
107 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
108 HRTIMER_MODE_ABS_PINNED
, 0);
112 DEFINE_MUTEX(sched_domains_mutex
);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
115 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
117 void update_rq_clock(struct rq
*rq
)
121 if (rq
->skip_clock_update
> 0)
124 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
126 update_rq_clock_task(rq
, delta
);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug
unsigned int sysctl_sched_features
=
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names
[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file
*m
, void *v
)
156 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
157 if (!(sysctl_sched_features
& (1UL << i
)))
159 seq_printf(m
, "%s ", sched_feat_names
[i
]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
175 #include "features.h"
180 static void sched_feat_disable(int i
)
182 if (static_key_enabled(&sched_feat_keys
[i
]))
183 static_key_slow_dec(&sched_feat_keys
[i
]);
186 static void sched_feat_enable(int i
)
188 if (!static_key_enabled(&sched_feat_keys
[i
]))
189 static_key_slow_inc(&sched_feat_keys
[i
]);
192 static void sched_feat_disable(int i
) { };
193 static void sched_feat_enable(int i
) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp
)
201 if (strncmp(cmp
, "NO_", 3) == 0) {
206 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
207 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
209 sysctl_sched_features
&= ~(1UL << i
);
210 sched_feat_disable(i
);
212 sysctl_sched_features
|= (1UL << i
);
213 sched_feat_enable(i
);
223 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
224 size_t cnt
, loff_t
*ppos
)
233 if (copy_from_user(&buf
, ubuf
, cnt
))
239 i
= sched_feat_set(cmp
);
240 if (i
== __SCHED_FEAT_NR
)
248 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
250 return single_open(filp
, sched_feat_show
, NULL
);
253 static const struct file_operations sched_feat_fops
= {
254 .open
= sched_feat_open
,
255 .write
= sched_feat_write
,
258 .release
= single_release
,
261 static __init
int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
268 late_initcall(sched_init_debug
);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
278 * period over which we average the RT time consumption, measured
283 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period
= 1000000;
291 __read_mostly
int scheduler_running
;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime
= 950000;
300 * __task_rq_lock - lock the rq @p resides on.
302 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
307 lockdep_assert_held(&p
->pi_lock
);
311 raw_spin_lock(&rq
->lock
);
312 if (likely(rq
== task_rq(p
)))
314 raw_spin_unlock(&rq
->lock
);
319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
322 __acquires(p
->pi_lock
)
328 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
330 raw_spin_lock(&rq
->lock
);
331 if (likely(rq
== task_rq(p
)))
333 raw_spin_unlock(&rq
->lock
);
334 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
338 static void __task_rq_unlock(struct rq
*rq
)
341 raw_spin_unlock(&rq
->lock
);
345 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
347 __releases(p
->pi_lock
)
349 raw_spin_unlock(&rq
->lock
);
350 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
354 * this_rq_lock - lock this runqueue and disable interrupts.
356 static struct rq
*this_rq_lock(void)
363 raw_spin_lock(&rq
->lock
);
368 #ifdef CONFIG_SCHED_HRTICK
370 * Use HR-timers to deliver accurate preemption points.
373 static void hrtick_clear(struct rq
*rq
)
375 if (hrtimer_active(&rq
->hrtick_timer
))
376 hrtimer_cancel(&rq
->hrtick_timer
);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
385 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
387 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
389 raw_spin_lock(&rq
->lock
);
391 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
392 raw_spin_unlock(&rq
->lock
);
394 return HRTIMER_NORESTART
;
399 static int __hrtick_restart(struct rq
*rq
)
401 struct hrtimer
*timer
= &rq
->hrtick_timer
;
402 ktime_t time
= hrtimer_get_softexpires(timer
);
404 return __hrtimer_start_range_ns(timer
, time
, 0, HRTIMER_MODE_ABS_PINNED
, 0);
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg
)
414 raw_spin_lock(&rq
->lock
);
415 __hrtick_restart(rq
);
416 rq
->hrtick_csd_pending
= 0;
417 raw_spin_unlock(&rq
->lock
);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq
*rq
, u64 delay
)
427 struct hrtimer
*timer
= &rq
->hrtick_timer
;
428 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
430 hrtimer_set_expires(timer
, time
);
432 if (rq
== this_rq()) {
433 __hrtick_restart(rq
);
434 } else if (!rq
->hrtick_csd_pending
) {
435 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
436 rq
->hrtick_csd_pending
= 1;
441 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
443 int cpu
= (int)(long)hcpu
;
446 case CPU_UP_CANCELED
:
447 case CPU_UP_CANCELED_FROZEN
:
448 case CPU_DOWN_PREPARE
:
449 case CPU_DOWN_PREPARE_FROZEN
:
451 case CPU_DEAD_FROZEN
:
452 hrtick_clear(cpu_rq(cpu
));
459 static __init
void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick
, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq
*rq
, u64 delay
)
471 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
472 HRTIMER_MODE_REL_PINNED
, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq
*rq
)
483 rq
->hrtick_csd_pending
= 0;
485 rq
->hrtick_csd
.flags
= 0;
486 rq
->hrtick_csd
.func
= __hrtick_start
;
487 rq
->hrtick_csd
.info
= rq
;
490 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
491 rq
->hrtick_timer
.function
= hrtick
;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq
*rq
)
498 static inline void init_rq_hrtick(struct rq
*rq
)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
514 void resched_task(struct task_struct
*p
)
518 lockdep_assert_held(&task_rq(p
)->lock
);
520 if (test_tsk_need_resched(p
))
523 set_tsk_need_resched(p
);
526 if (cpu
== smp_processor_id()) {
527 set_preempt_need_resched();
531 /* NEED_RESCHED must be visible before we test polling */
533 if (!tsk_is_polling(p
))
534 smp_send_reschedule(cpu
);
537 void resched_cpu(int cpu
)
539 struct rq
*rq
= cpu_rq(cpu
);
542 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
544 resched_task(cpu_curr(cpu
));
545 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
549 #ifdef CONFIG_NO_HZ_COMMON
551 * In the semi idle case, use the nearest busy cpu for migrating timers
552 * from an idle cpu. This is good for power-savings.
554 * We don't do similar optimization for completely idle system, as
555 * selecting an idle cpu will add more delays to the timers than intended
556 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 int get_nohz_timer_target(void)
560 int cpu
= smp_processor_id();
562 struct sched_domain
*sd
;
565 for_each_domain(cpu
, sd
) {
566 for_each_cpu(i
, sched_domain_span(sd
)) {
578 * When add_timer_on() enqueues a timer into the timer wheel of an
579 * idle CPU then this timer might expire before the next timer event
580 * which is scheduled to wake up that CPU. In case of a completely
581 * idle system the next event might even be infinite time into the
582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583 * leaves the inner idle loop so the newly added timer is taken into
584 * account when the CPU goes back to idle and evaluates the timer
585 * wheel for the next timer event.
587 static void wake_up_idle_cpu(int cpu
)
589 struct rq
*rq
= cpu_rq(cpu
);
591 if (cpu
== smp_processor_id())
595 * This is safe, as this function is called with the timer
596 * wheel base lock of (cpu) held. When the CPU is on the way
597 * to idle and has not yet set rq->curr to idle then it will
598 * be serialized on the timer wheel base lock and take the new
599 * timer into account automatically.
601 if (rq
->curr
!= rq
->idle
)
605 * We can set TIF_RESCHED on the idle task of the other CPU
606 * lockless. The worst case is that the other CPU runs the
607 * idle task through an additional NOOP schedule()
609 set_tsk_need_resched(rq
->idle
);
611 /* NEED_RESCHED must be visible before we test polling */
613 if (!tsk_is_polling(rq
->idle
))
614 smp_send_reschedule(cpu
);
617 static bool wake_up_full_nohz_cpu(int cpu
)
619 if (tick_nohz_full_cpu(cpu
)) {
620 if (cpu
!= smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 smp_send_reschedule(cpu
);
629 void wake_up_nohz_cpu(int cpu
)
631 if (!wake_up_full_nohz_cpu(cpu
))
632 wake_up_idle_cpu(cpu
);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu
= smp_processor_id();
639 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
642 if (idle_cpu(cpu
) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
669 /* Make sure rq->nr_running update is visible after the IPI */
672 /* More than one running task need preemption */
673 if (rq
->nr_running
> 1)
678 #endif /* CONFIG_NO_HZ_FULL */
680 void sched_avg_update(struct rq
*rq
)
682 s64 period
= sched_avg_period();
684 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
686 * Inline assembly required to prevent the compiler
687 * optimising this loop into a divmod call.
688 * See __iter_div_u64_rem() for another example of this.
690 asm("" : "+rm" (rq
->age_stamp
));
691 rq
->age_stamp
+= period
;
696 #endif /* CONFIG_SMP */
698 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
701 * Iterate task_group tree rooted at *from, calling @down when first entering a
702 * node and @up when leaving it for the final time.
704 * Caller must hold rcu_lock or sufficient equivalent.
706 int walk_tg_tree_from(struct task_group
*from
,
707 tg_visitor down
, tg_visitor up
, void *data
)
709 struct task_group
*parent
, *child
;
715 ret
= (*down
)(parent
, data
);
718 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
725 ret
= (*up
)(parent
, data
);
726 if (ret
|| parent
== from
)
730 parent
= parent
->parent
;
737 int tg_nop(struct task_group
*tg
, void *data
)
743 static void set_load_weight(struct task_struct
*p
)
745 int prio
= p
->static_prio
- MAX_RT_PRIO
;
746 struct load_weight
*load
= &p
->se
.load
;
749 * SCHED_IDLE tasks get minimal weight:
751 if (p
->policy
== SCHED_IDLE
) {
752 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
753 load
->inv_weight
= WMULT_IDLEPRIO
;
757 load
->weight
= scale_load(prio_to_weight
[prio
]);
758 load
->inv_weight
= prio_to_wmult
[prio
];
761 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
764 sched_info_queued(rq
, p
);
765 p
->sched_class
->enqueue_task(rq
, p
, flags
);
768 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
771 sched_info_dequeued(rq
, p
);
772 p
->sched_class
->dequeue_task(rq
, p
, flags
);
775 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
777 if (task_contributes_to_load(p
))
778 rq
->nr_uninterruptible
--;
780 enqueue_task(rq
, p
, flags
);
783 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
785 if (task_contributes_to_load(p
))
786 rq
->nr_uninterruptible
++;
788 dequeue_task(rq
, p
, flags
);
791 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
794 * In theory, the compile should just see 0 here, and optimize out the call
795 * to sched_rt_avg_update. But I don't trust it...
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798 s64 steal
= 0, irq_delta
= 0;
800 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
801 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
804 * Since irq_time is only updated on {soft,}irq_exit, we might run into
805 * this case when a previous update_rq_clock() happened inside a
808 * When this happens, we stop ->clock_task and only update the
809 * prev_irq_time stamp to account for the part that fit, so that a next
810 * update will consume the rest. This ensures ->clock_task is
813 * It does however cause some slight miss-attribution of {soft,}irq
814 * time, a more accurate solution would be to update the irq_time using
815 * the current rq->clock timestamp, except that would require using
818 if (irq_delta
> delta
)
821 rq
->prev_irq_time
+= irq_delta
;
824 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
825 if (static_key_false((¶virt_steal_rq_enabled
))) {
828 steal
= paravirt_steal_clock(cpu_of(rq
));
829 steal
-= rq
->prev_steal_time_rq
;
831 if (unlikely(steal
> delta
))
834 st
= steal_ticks(steal
);
835 steal
= st
* TICK_NSEC
;
837 rq
->prev_steal_time_rq
+= steal
;
843 rq
->clock_task
+= delta
;
845 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
846 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
847 sched_rt_avg_update(rq
, irq_delta
+ steal
);
851 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
853 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
854 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
858 * Make it appear like a SCHED_FIFO task, its something
859 * userspace knows about and won't get confused about.
861 * Also, it will make PI more or less work without too
862 * much confusion -- but then, stop work should not
863 * rely on PI working anyway.
865 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
867 stop
->sched_class
= &stop_sched_class
;
870 cpu_rq(cpu
)->stop
= stop
;
874 * Reset it back to a normal scheduling class so that
875 * it can die in pieces.
877 old_stop
->sched_class
= &rt_sched_class
;
882 * __normal_prio - return the priority that is based on the static prio
884 static inline int __normal_prio(struct task_struct
*p
)
886 return p
->static_prio
;
890 * Calculate the expected normal priority: i.e. priority
891 * without taking RT-inheritance into account. Might be
892 * boosted by interactivity modifiers. Changes upon fork,
893 * setprio syscalls, and whenever the interactivity
894 * estimator recalculates.
896 static inline int normal_prio(struct task_struct
*p
)
900 if (task_has_dl_policy(p
))
901 prio
= MAX_DL_PRIO
-1;
902 else if (task_has_rt_policy(p
))
903 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
905 prio
= __normal_prio(p
);
910 * Calculate the current priority, i.e. the priority
911 * taken into account by the scheduler. This value might
912 * be boosted by RT tasks, or might be boosted by
913 * interactivity modifiers. Will be RT if the task got
914 * RT-boosted. If not then it returns p->normal_prio.
916 static int effective_prio(struct task_struct
*p
)
918 p
->normal_prio
= normal_prio(p
);
920 * If we are RT tasks or we were boosted to RT priority,
921 * keep the priority unchanged. Otherwise, update priority
922 * to the normal priority:
924 if (!rt_prio(p
->prio
))
925 return p
->normal_prio
;
930 * task_curr - is this task currently executing on a CPU?
931 * @p: the task in question.
933 * Return: 1 if the task is currently executing. 0 otherwise.
935 inline int task_curr(const struct task_struct
*p
)
937 return cpu_curr(task_cpu(p
)) == p
;
940 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
941 const struct sched_class
*prev_class
,
944 if (prev_class
!= p
->sched_class
) {
945 if (prev_class
->switched_from
)
946 prev_class
->switched_from(rq
, p
);
947 p
->sched_class
->switched_to(rq
, p
);
948 } else if (oldprio
!= p
->prio
|| dl_task(p
))
949 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
952 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
954 const struct sched_class
*class;
956 if (p
->sched_class
== rq
->curr
->sched_class
) {
957 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
959 for_each_class(class) {
960 if (class == rq
->curr
->sched_class
)
962 if (class == p
->sched_class
) {
963 resched_task(rq
->curr
);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
974 rq
->skip_clock_update
= 1;
978 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
986 !(task_preempt_count(p
) & PREEMPT_ACTIVE
));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1000 lockdep_is_held(&task_rq(p
)->lock
)));
1004 trace_sched_migrate_task(p
, new_cpu
);
1006 if (task_cpu(p
) != new_cpu
) {
1007 if (p
->sched_class
->migrate_task_rq
)
1008 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1009 p
->se
.nr_migrations
++;
1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
1013 __set_task_cpu(p
, new_cpu
);
1016 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1019 struct rq
*src_rq
, *dst_rq
;
1021 src_rq
= task_rq(p
);
1022 dst_rq
= cpu_rq(cpu
);
1024 deactivate_task(src_rq
, p
, 0);
1025 set_task_cpu(p
, cpu
);
1026 activate_task(dst_rq
, p
, 0);
1027 check_preempt_curr(dst_rq
, p
, 0);
1030 * Task isn't running anymore; make it appear like we migrated
1031 * it before it went to sleep. This means on wakeup we make the
1032 * previous cpu our targer instead of where it really is.
1038 struct migration_swap_arg
{
1039 struct task_struct
*src_task
, *dst_task
;
1040 int src_cpu
, dst_cpu
;
1043 static int migrate_swap_stop(void *data
)
1045 struct migration_swap_arg
*arg
= data
;
1046 struct rq
*src_rq
, *dst_rq
;
1049 src_rq
= cpu_rq(arg
->src_cpu
);
1050 dst_rq
= cpu_rq(arg
->dst_cpu
);
1052 double_raw_lock(&arg
->src_task
->pi_lock
,
1053 &arg
->dst_task
->pi_lock
);
1054 double_rq_lock(src_rq
, dst_rq
);
1055 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1058 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1061 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1064 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1067 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1068 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1073 double_rq_unlock(src_rq
, dst_rq
);
1074 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1075 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1081 * Cross migrate two tasks
1083 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1085 struct migration_swap_arg arg
;
1088 arg
= (struct migration_swap_arg
){
1090 .src_cpu
= task_cpu(cur
),
1092 .dst_cpu
= task_cpu(p
),
1095 if (arg
.src_cpu
== arg
.dst_cpu
)
1099 * These three tests are all lockless; this is OK since all of them
1100 * will be re-checked with proper locks held further down the line.
1102 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1105 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1108 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1111 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1112 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1118 struct migration_arg
{
1119 struct task_struct
*task
;
1123 static int migration_cpu_stop(void *data
);
1126 * wait_task_inactive - wait for a thread to unschedule.
1128 * If @match_state is nonzero, it's the @p->state value just checked and
1129 * not expected to change. If it changes, i.e. @p might have woken up,
1130 * then return zero. When we succeed in waiting for @p to be off its CPU,
1131 * we return a positive number (its total switch count). If a second call
1132 * a short while later returns the same number, the caller can be sure that
1133 * @p has remained unscheduled the whole time.
1135 * The caller must ensure that the task *will* unschedule sometime soon,
1136 * else this function might spin for a *long* time. This function can't
1137 * be called with interrupts off, or it may introduce deadlock with
1138 * smp_call_function() if an IPI is sent by the same process we are
1139 * waiting to become inactive.
1141 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1143 unsigned long flags
;
1150 * We do the initial early heuristics without holding
1151 * any task-queue locks at all. We'll only try to get
1152 * the runqueue lock when things look like they will
1158 * If the task is actively running on another CPU
1159 * still, just relax and busy-wait without holding
1162 * NOTE! Since we don't hold any locks, it's not
1163 * even sure that "rq" stays as the right runqueue!
1164 * But we don't care, since "task_running()" will
1165 * return false if the runqueue has changed and p
1166 * is actually now running somewhere else!
1168 while (task_running(rq
, p
)) {
1169 if (match_state
&& unlikely(p
->state
!= match_state
))
1175 * Ok, time to look more closely! We need the rq
1176 * lock now, to be *sure*. If we're wrong, we'll
1177 * just go back and repeat.
1179 rq
= task_rq_lock(p
, &flags
);
1180 trace_sched_wait_task(p
);
1181 running
= task_running(rq
, p
);
1184 if (!match_state
|| p
->state
== match_state
)
1185 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1186 task_rq_unlock(rq
, p
, &flags
);
1189 * If it changed from the expected state, bail out now.
1191 if (unlikely(!ncsw
))
1195 * Was it really running after all now that we
1196 * checked with the proper locks actually held?
1198 * Oops. Go back and try again..
1200 if (unlikely(running
)) {
1206 * It's not enough that it's not actively running,
1207 * it must be off the runqueue _entirely_, and not
1210 * So if it was still runnable (but just not actively
1211 * running right now), it's preempted, and we should
1212 * yield - it could be a while.
1214 if (unlikely(on_rq
)) {
1215 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1217 set_current_state(TASK_UNINTERRUPTIBLE
);
1218 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1223 * Ahh, all good. It wasn't running, and it wasn't
1224 * runnable, which means that it will never become
1225 * running in the future either. We're all done!
1234 * kick_process - kick a running thread to enter/exit the kernel
1235 * @p: the to-be-kicked thread
1237 * Cause a process which is running on another CPU to enter
1238 * kernel-mode, without any delay. (to get signals handled.)
1240 * NOTE: this function doesn't have to take the runqueue lock,
1241 * because all it wants to ensure is that the remote task enters
1242 * the kernel. If the IPI races and the task has been migrated
1243 * to another CPU then no harm is done and the purpose has been
1246 void kick_process(struct task_struct
*p
)
1252 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1253 smp_send_reschedule(cpu
);
1256 EXPORT_SYMBOL_GPL(kick_process
);
1257 #endif /* CONFIG_SMP */
1261 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1265 int nid
= cpu_to_node(cpu
);
1266 const struct cpumask
*nodemask
= NULL
;
1267 enum { cpuset
, possible
, fail
} state
= cpuset
;
1271 * If the node that the cpu is on has been offlined, cpu_to_node()
1272 * will return -1. There is no cpu on the node, and we should
1273 * select the cpu on the other node.
1276 nodemask
= cpumask_of_node(nid
);
1278 /* Look for allowed, online CPU in same node. */
1279 for_each_cpu(dest_cpu
, nodemask
) {
1280 if (!cpu_online(dest_cpu
))
1282 if (!cpu_active(dest_cpu
))
1284 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1290 /* Any allowed, online CPU? */
1291 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1292 if (!cpu_online(dest_cpu
))
1294 if (!cpu_active(dest_cpu
))
1301 /* No more Mr. Nice Guy. */
1302 cpuset_cpus_allowed_fallback(p
);
1307 do_set_cpus_allowed(p
, cpu_possible_mask
);
1318 if (state
!= cpuset
) {
1320 * Don't tell them about moving exiting tasks or
1321 * kernel threads (both mm NULL), since they never
1324 if (p
->mm
&& printk_ratelimit()) {
1325 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1326 task_pid_nr(p
), p
->comm
, cpu
);
1334 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1337 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1339 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1342 * In order not to call set_task_cpu() on a blocking task we need
1343 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1346 * Since this is common to all placement strategies, this lives here.
1348 * [ this allows ->select_task() to simply return task_cpu(p) and
1349 * not worry about this generic constraint ]
1351 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1353 cpu
= select_fallback_rq(task_cpu(p
), p
);
1358 static void update_avg(u64
*avg
, u64 sample
)
1360 s64 diff
= sample
- *avg
;
1366 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1368 #ifdef CONFIG_SCHEDSTATS
1369 struct rq
*rq
= this_rq();
1372 int this_cpu
= smp_processor_id();
1374 if (cpu
== this_cpu
) {
1375 schedstat_inc(rq
, ttwu_local
);
1376 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1378 struct sched_domain
*sd
;
1380 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1382 for_each_domain(this_cpu
, sd
) {
1383 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1384 schedstat_inc(sd
, ttwu_wake_remote
);
1391 if (wake_flags
& WF_MIGRATED
)
1392 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1394 #endif /* CONFIG_SMP */
1396 schedstat_inc(rq
, ttwu_count
);
1397 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1399 if (wake_flags
& WF_SYNC
)
1400 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1402 #endif /* CONFIG_SCHEDSTATS */
1405 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1407 activate_task(rq
, p
, en_flags
);
1410 /* if a worker is waking up, notify workqueue */
1411 if (p
->flags
& PF_WQ_WORKER
)
1412 wq_worker_waking_up(p
, cpu_of(rq
));
1416 * Mark the task runnable and perform wakeup-preemption.
1419 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1421 check_preempt_curr(rq
, p
, wake_flags
);
1422 trace_sched_wakeup(p
, true);
1424 p
->state
= TASK_RUNNING
;
1426 if (p
->sched_class
->task_woken
)
1427 p
->sched_class
->task_woken(rq
, p
);
1429 if (rq
->idle_stamp
) {
1430 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1431 u64 max
= 2*rq
->max_idle_balance_cost
;
1433 update_avg(&rq
->avg_idle
, delta
);
1435 if (rq
->avg_idle
> max
)
1444 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1447 if (p
->sched_contributes_to_load
)
1448 rq
->nr_uninterruptible
--;
1451 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1452 ttwu_do_wakeup(rq
, p
, wake_flags
);
1456 * Called in case the task @p isn't fully descheduled from its runqueue,
1457 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1458 * since all we need to do is flip p->state to TASK_RUNNING, since
1459 * the task is still ->on_rq.
1461 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1466 rq
= __task_rq_lock(p
);
1468 /* check_preempt_curr() may use rq clock */
1469 update_rq_clock(rq
);
1470 ttwu_do_wakeup(rq
, p
, wake_flags
);
1473 __task_rq_unlock(rq
);
1479 static void sched_ttwu_pending(void)
1481 struct rq
*rq
= this_rq();
1482 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1483 struct task_struct
*p
;
1485 raw_spin_lock(&rq
->lock
);
1488 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1489 llist
= llist_next(llist
);
1490 ttwu_do_activate(rq
, p
, 0);
1493 raw_spin_unlock(&rq
->lock
);
1496 void scheduler_ipi(void)
1499 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1500 * TIF_NEED_RESCHED remotely (for the first time) will also send
1503 preempt_fold_need_resched();
1505 if (llist_empty(&this_rq()->wake_list
)
1506 && !tick_nohz_full_cpu(smp_processor_id())
1507 && !got_nohz_idle_kick())
1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1512 * traditionally all their work was done from the interrupt return
1513 * path. Now that we actually do some work, we need to make sure
1516 * Some archs already do call them, luckily irq_enter/exit nest
1519 * Arguably we should visit all archs and update all handlers,
1520 * however a fair share of IPIs are still resched only so this would
1521 * somewhat pessimize the simple resched case.
1524 tick_nohz_full_check();
1525 sched_ttwu_pending();
1528 * Check if someone kicked us for doing the nohz idle load balance.
1530 if (unlikely(got_nohz_idle_kick())) {
1531 this_rq()->idle_balance
= 1;
1532 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1537 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1539 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
1540 smp_send_reschedule(cpu
);
1543 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1545 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1547 #endif /* CONFIG_SMP */
1549 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1551 struct rq
*rq
= cpu_rq(cpu
);
1553 #if defined(CONFIG_SMP)
1554 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1555 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1556 ttwu_queue_remote(p
, cpu
);
1561 raw_spin_lock(&rq
->lock
);
1562 ttwu_do_activate(rq
, p
, 0);
1563 raw_spin_unlock(&rq
->lock
);
1567 * try_to_wake_up - wake up a thread
1568 * @p: the thread to be awakened
1569 * @state: the mask of task states that can be woken
1570 * @wake_flags: wake modifier flags (WF_*)
1572 * Put it on the run-queue if it's not already there. The "current"
1573 * thread is always on the run-queue (except when the actual
1574 * re-schedule is in progress), and as such you're allowed to do
1575 * the simpler "current->state = TASK_RUNNING" to mark yourself
1576 * runnable without the overhead of this.
1578 * Return: %true if @p was woken up, %false if it was already running.
1579 * or @state didn't match @p's state.
1582 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1584 unsigned long flags
;
1585 int cpu
, success
= 0;
1588 * If we are going to wake up a thread waiting for CONDITION we
1589 * need to ensure that CONDITION=1 done by the caller can not be
1590 * reordered with p->state check below. This pairs with mb() in
1591 * set_current_state() the waiting thread does.
1593 smp_mb__before_spinlock();
1594 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1595 if (!(p
->state
& state
))
1598 success
= 1; /* we're going to change ->state */
1601 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1606 * If the owning (remote) cpu is still in the middle of schedule() with
1607 * this task as prev, wait until its done referencing the task.
1612 * Pairs with the smp_wmb() in finish_lock_switch().
1616 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1617 p
->state
= TASK_WAKING
;
1619 if (p
->sched_class
->task_waking
)
1620 p
->sched_class
->task_waking(p
);
1622 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1623 if (task_cpu(p
) != cpu
) {
1624 wake_flags
|= WF_MIGRATED
;
1625 set_task_cpu(p
, cpu
);
1627 #endif /* CONFIG_SMP */
1631 ttwu_stat(p
, cpu
, wake_flags
);
1633 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1639 * try_to_wake_up_local - try to wake up a local task with rq lock held
1640 * @p: the thread to be awakened
1642 * Put @p on the run-queue if it's not already there. The caller must
1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1646 static void try_to_wake_up_local(struct task_struct
*p
)
1648 struct rq
*rq
= task_rq(p
);
1650 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1651 WARN_ON_ONCE(p
== current
))
1654 lockdep_assert_held(&rq
->lock
);
1656 if (!raw_spin_trylock(&p
->pi_lock
)) {
1657 raw_spin_unlock(&rq
->lock
);
1658 raw_spin_lock(&p
->pi_lock
);
1659 raw_spin_lock(&rq
->lock
);
1662 if (!(p
->state
& TASK_NORMAL
))
1666 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1668 ttwu_do_wakeup(rq
, p
, 0);
1669 ttwu_stat(p
, smp_processor_id(), 0);
1671 raw_spin_unlock(&p
->pi_lock
);
1675 * wake_up_process - Wake up a specific process
1676 * @p: The process to be woken up.
1678 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * Return: 1 if the process was woken up, 0 if it was already running.
1683 * It may be assumed that this function implies a write memory barrier before
1684 * changing the task state if and only if any tasks are woken up.
1686 int wake_up_process(struct task_struct
*p
)
1688 WARN_ON(task_is_stopped_or_traced(p
));
1689 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1691 EXPORT_SYMBOL(wake_up_process
);
1693 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1695 return try_to_wake_up(p
, state
, 0);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 * __sched_fork() is basic setup used by init_idle() too:
1704 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1709 p
->se
.exec_start
= 0;
1710 p
->se
.sum_exec_runtime
= 0;
1711 p
->se
.prev_sum_exec_runtime
= 0;
1712 p
->se
.nr_migrations
= 0;
1714 INIT_LIST_HEAD(&p
->se
.group_node
);
1716 #ifdef CONFIG_SCHEDSTATS
1717 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1720 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1721 hrtimer_init(&p
->dl
.dl_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1722 p
->dl
.dl_runtime
= p
->dl
.runtime
= 0;
1723 p
->dl
.dl_deadline
= p
->dl
.deadline
= 0;
1724 p
->dl
.dl_period
= 0;
1727 INIT_LIST_HEAD(&p
->rt
.run_list
);
1729 #ifdef CONFIG_PREEMPT_NOTIFIERS
1730 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1733 #ifdef CONFIG_NUMA_BALANCING
1734 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1735 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1736 p
->mm
->numa_scan_seq
= 0;
1739 if (clone_flags
& CLONE_VM
)
1740 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1742 p
->numa_preferred_nid
= -1;
1744 p
->node_stamp
= 0ULL;
1745 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1746 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1747 p
->numa_work
.next
= &p
->numa_work
;
1748 p
->numa_faults_memory
= NULL
;
1749 p
->numa_faults_buffer_memory
= NULL
;
1750 p
->last_task_numa_placement
= 0;
1751 p
->last_sum_exec_runtime
= 0;
1753 INIT_LIST_HEAD(&p
->numa_entry
);
1754 p
->numa_group
= NULL
;
1755 #endif /* CONFIG_NUMA_BALANCING */
1758 #ifdef CONFIG_NUMA_BALANCING
1759 #ifdef CONFIG_SCHED_DEBUG
1760 void set_numabalancing_state(bool enabled
)
1763 sched_feat_set("NUMA");
1765 sched_feat_set("NO_NUMA");
1768 __read_mostly
bool numabalancing_enabled
;
1770 void set_numabalancing_state(bool enabled
)
1772 numabalancing_enabled
= enabled
;
1774 #endif /* CONFIG_SCHED_DEBUG */
1776 #ifdef CONFIG_PROC_SYSCTL
1777 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1778 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1782 int state
= numabalancing_enabled
;
1784 if (write
&& !capable(CAP_SYS_ADMIN
))
1789 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1793 set_numabalancing_state(state
);
1800 * fork()/clone()-time setup:
1802 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1804 unsigned long flags
;
1805 int cpu
= get_cpu();
1807 __sched_fork(clone_flags
, p
);
1809 * We mark the process as running here. This guarantees that
1810 * nobody will actually run it, and a signal or other external
1811 * event cannot wake it up and insert it on the runqueue either.
1813 p
->state
= TASK_RUNNING
;
1816 * Make sure we do not leak PI boosting priority to the child.
1818 p
->prio
= current
->normal_prio
;
1821 * Revert to default priority/policy on fork if requested.
1823 if (unlikely(p
->sched_reset_on_fork
)) {
1824 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1825 p
->policy
= SCHED_NORMAL
;
1826 p
->static_prio
= NICE_TO_PRIO(0);
1828 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1829 p
->static_prio
= NICE_TO_PRIO(0);
1831 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1835 * We don't need the reset flag anymore after the fork. It has
1836 * fulfilled its duty:
1838 p
->sched_reset_on_fork
= 0;
1841 if (dl_prio(p
->prio
)) {
1844 } else if (rt_prio(p
->prio
)) {
1845 p
->sched_class
= &rt_sched_class
;
1847 p
->sched_class
= &fair_sched_class
;
1850 if (p
->sched_class
->task_fork
)
1851 p
->sched_class
->task_fork(p
);
1854 * The child is not yet in the pid-hash so no cgroup attach races,
1855 * and the cgroup is pinned to this child due to cgroup_fork()
1856 * is ran before sched_fork().
1858 * Silence PROVE_RCU.
1860 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1861 set_task_cpu(p
, cpu
);
1862 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1864 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1865 if (likely(sched_info_on()))
1866 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1868 #if defined(CONFIG_SMP)
1871 init_task_preempt_count(p
);
1873 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1874 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1881 unsigned long to_ratio(u64 period
, u64 runtime
)
1883 if (runtime
== RUNTIME_INF
)
1887 * Doing this here saves a lot of checks in all
1888 * the calling paths, and returning zero seems
1889 * safe for them anyway.
1894 return div64_u64(runtime
<< 20, period
);
1898 inline struct dl_bw
*dl_bw_of(int i
)
1900 return &cpu_rq(i
)->rd
->dl_bw
;
1903 static inline int dl_bw_cpus(int i
)
1905 struct root_domain
*rd
= cpu_rq(i
)->rd
;
1908 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
1914 inline struct dl_bw
*dl_bw_of(int i
)
1916 return &cpu_rq(i
)->dl
.dl_bw
;
1919 static inline int dl_bw_cpus(int i
)
1926 void __dl_clear(struct dl_bw
*dl_b
, u64 tsk_bw
)
1928 dl_b
->total_bw
-= tsk_bw
;
1932 void __dl_add(struct dl_bw
*dl_b
, u64 tsk_bw
)
1934 dl_b
->total_bw
+= tsk_bw
;
1938 bool __dl_overflow(struct dl_bw
*dl_b
, int cpus
, u64 old_bw
, u64 new_bw
)
1940 return dl_b
->bw
!= -1 &&
1941 dl_b
->bw
* cpus
< dl_b
->total_bw
- old_bw
+ new_bw
;
1945 * We must be sure that accepting a new task (or allowing changing the
1946 * parameters of an existing one) is consistent with the bandwidth
1947 * constraints. If yes, this function also accordingly updates the currently
1948 * allocated bandwidth to reflect the new situation.
1950 * This function is called while holding p's rq->lock.
1952 static int dl_overflow(struct task_struct
*p
, int policy
,
1953 const struct sched_attr
*attr
)
1956 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1957 u64 period
= attr
->sched_period
;
1958 u64 runtime
= attr
->sched_runtime
;
1959 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
1962 if (new_bw
== p
->dl
.dl_bw
)
1966 * Either if a task, enters, leave, or stays -deadline but changes
1967 * its parameters, we may need to update accordingly the total
1968 * allocated bandwidth of the container.
1970 raw_spin_lock(&dl_b
->lock
);
1971 cpus
= dl_bw_cpus(task_cpu(p
));
1972 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
1973 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
1974 __dl_add(dl_b
, new_bw
);
1976 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
1977 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
1978 __dl_clear(dl_b
, p
->dl
.dl_bw
);
1979 __dl_add(dl_b
, new_bw
);
1981 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
1982 __dl_clear(dl_b
, p
->dl
.dl_bw
);
1985 raw_spin_unlock(&dl_b
->lock
);
1990 extern void init_dl_bw(struct dl_bw
*dl_b
);
1993 * wake_up_new_task - wake up a newly created task for the first time.
1995 * This function will do some initial scheduler statistics housekeeping
1996 * that must be done for every newly created context, then puts the task
1997 * on the runqueue and wakes it.
1999 void wake_up_new_task(struct task_struct
*p
)
2001 unsigned long flags
;
2004 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2007 * Fork balancing, do it here and not earlier because:
2008 * - cpus_allowed can change in the fork path
2009 * - any previously selected cpu might disappear through hotplug
2011 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2014 /* Initialize new task's runnable average */
2015 init_task_runnable_average(p
);
2016 rq
= __task_rq_lock(p
);
2017 activate_task(rq
, p
, 0);
2019 trace_sched_wakeup_new(p
, true);
2020 check_preempt_curr(rq
, p
, WF_FORK
);
2022 if (p
->sched_class
->task_woken
)
2023 p
->sched_class
->task_woken(rq
, p
);
2025 task_rq_unlock(rq
, p
, &flags
);
2028 #ifdef CONFIG_PREEMPT_NOTIFIERS
2031 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2032 * @notifier: notifier struct to register
2034 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2036 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2038 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2041 * preempt_notifier_unregister - no longer interested in preemption notifications
2042 * @notifier: notifier struct to unregister
2044 * This is safe to call from within a preemption notifier.
2046 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2048 hlist_del(¬ifier
->link
);
2050 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2052 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2054 struct preempt_notifier
*notifier
;
2056 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2057 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2061 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2062 struct task_struct
*next
)
2064 struct preempt_notifier
*notifier
;
2066 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2067 notifier
->ops
->sched_out(notifier
, next
);
2070 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2072 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2077 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2078 struct task_struct
*next
)
2082 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2085 * prepare_task_switch - prepare to switch tasks
2086 * @rq: the runqueue preparing to switch
2087 * @prev: the current task that is being switched out
2088 * @next: the task we are going to switch to.
2090 * This is called with the rq lock held and interrupts off. It must
2091 * be paired with a subsequent finish_task_switch after the context
2094 * prepare_task_switch sets up locking and calls architecture specific
2098 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2099 struct task_struct
*next
)
2101 trace_sched_switch(prev
, next
);
2102 sched_info_switch(rq
, prev
, next
);
2103 perf_event_task_sched_out(prev
, next
);
2104 fire_sched_out_preempt_notifiers(prev
, next
);
2105 prepare_lock_switch(rq
, next
);
2106 prepare_arch_switch(next
);
2110 * finish_task_switch - clean up after a task-switch
2111 * @rq: runqueue associated with task-switch
2112 * @prev: the thread we just switched away from.
2114 * finish_task_switch must be called after the context switch, paired
2115 * with a prepare_task_switch call before the context switch.
2116 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2117 * and do any other architecture-specific cleanup actions.
2119 * Note that we may have delayed dropping an mm in context_switch(). If
2120 * so, we finish that here outside of the runqueue lock. (Doing it
2121 * with the lock held can cause deadlocks; see schedule() for
2124 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2125 __releases(rq
->lock
)
2127 struct mm_struct
*mm
= rq
->prev_mm
;
2133 * A task struct has one reference for the use as "current".
2134 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2135 * schedule one last time. The schedule call will never return, and
2136 * the scheduled task must drop that reference.
2137 * The test for TASK_DEAD must occur while the runqueue locks are
2138 * still held, otherwise prev could be scheduled on another cpu, die
2139 * there before we look at prev->state, and then the reference would
2141 * Manfred Spraul <manfred@colorfullife.com>
2143 prev_state
= prev
->state
;
2144 vtime_task_switch(prev
);
2145 finish_arch_switch(prev
);
2146 perf_event_task_sched_in(prev
, current
);
2147 finish_lock_switch(rq
, prev
);
2148 finish_arch_post_lock_switch();
2150 fire_sched_in_preempt_notifiers(current
);
2153 if (unlikely(prev_state
== TASK_DEAD
)) {
2154 task_numa_free(prev
);
2156 if (prev
->sched_class
->task_dead
)
2157 prev
->sched_class
->task_dead(prev
);
2160 * Remove function-return probe instances associated with this
2161 * task and put them back on the free list.
2163 kprobe_flush_task(prev
);
2164 put_task_struct(prev
);
2167 tick_nohz_task_switch(current
);
2172 /* assumes rq->lock is held */
2173 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2175 if (prev
->sched_class
->pre_schedule
)
2176 prev
->sched_class
->pre_schedule(rq
, prev
);
2179 /* rq->lock is NOT held, but preemption is disabled */
2180 static inline void post_schedule(struct rq
*rq
)
2182 if (rq
->post_schedule
) {
2183 unsigned long flags
;
2185 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2186 if (rq
->curr
->sched_class
->post_schedule
)
2187 rq
->curr
->sched_class
->post_schedule(rq
);
2188 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2190 rq
->post_schedule
= 0;
2196 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2200 static inline void post_schedule(struct rq
*rq
)
2207 * schedule_tail - first thing a freshly forked thread must call.
2208 * @prev: the thread we just switched away from.
2210 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2211 __releases(rq
->lock
)
2213 struct rq
*rq
= this_rq();
2215 finish_task_switch(rq
, prev
);
2218 * FIXME: do we need to worry about rq being invalidated by the
2223 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2224 /* In this case, finish_task_switch does not reenable preemption */
2227 if (current
->set_child_tid
)
2228 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2232 * context_switch - switch to the new MM and the new
2233 * thread's register state.
2236 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2237 struct task_struct
*next
)
2239 struct mm_struct
*mm
, *oldmm
;
2241 prepare_task_switch(rq
, prev
, next
);
2244 oldmm
= prev
->active_mm
;
2246 * For paravirt, this is coupled with an exit in switch_to to
2247 * combine the page table reload and the switch backend into
2250 arch_start_context_switch(prev
);
2253 next
->active_mm
= oldmm
;
2254 atomic_inc(&oldmm
->mm_count
);
2255 enter_lazy_tlb(oldmm
, next
);
2257 switch_mm(oldmm
, mm
, next
);
2260 prev
->active_mm
= NULL
;
2261 rq
->prev_mm
= oldmm
;
2264 * Since the runqueue lock will be released by the next
2265 * task (which is an invalid locking op but in the case
2266 * of the scheduler it's an obvious special-case), so we
2267 * do an early lockdep release here:
2269 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2270 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2273 context_tracking_task_switch(prev
, next
);
2274 /* Here we just switch the register state and the stack. */
2275 switch_to(prev
, next
, prev
);
2279 * this_rq must be evaluated again because prev may have moved
2280 * CPUs since it called schedule(), thus the 'rq' on its stack
2281 * frame will be invalid.
2283 finish_task_switch(this_rq(), prev
);
2287 * nr_running and nr_context_switches:
2289 * externally visible scheduler statistics: current number of runnable
2290 * threads, total number of context switches performed since bootup.
2292 unsigned long nr_running(void)
2294 unsigned long i
, sum
= 0;
2296 for_each_online_cpu(i
)
2297 sum
+= cpu_rq(i
)->nr_running
;
2302 unsigned long long nr_context_switches(void)
2305 unsigned long long sum
= 0;
2307 for_each_possible_cpu(i
)
2308 sum
+= cpu_rq(i
)->nr_switches
;
2313 unsigned long nr_iowait(void)
2315 unsigned long i
, sum
= 0;
2317 for_each_possible_cpu(i
)
2318 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2323 unsigned long nr_iowait_cpu(int cpu
)
2325 struct rq
*this = cpu_rq(cpu
);
2326 return atomic_read(&this->nr_iowait
);
2332 * sched_exec - execve() is a valuable balancing opportunity, because at
2333 * this point the task has the smallest effective memory and cache footprint.
2335 void sched_exec(void)
2337 struct task_struct
*p
= current
;
2338 unsigned long flags
;
2341 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2342 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2343 if (dest_cpu
== smp_processor_id())
2346 if (likely(cpu_active(dest_cpu
))) {
2347 struct migration_arg arg
= { p
, dest_cpu
};
2349 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2350 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2354 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2359 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2360 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2362 EXPORT_PER_CPU_SYMBOL(kstat
);
2363 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2366 * Return any ns on the sched_clock that have not yet been accounted in
2367 * @p in case that task is currently running.
2369 * Called with task_rq_lock() held on @rq.
2371 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2375 if (task_current(rq
, p
)) {
2376 update_rq_clock(rq
);
2377 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2385 unsigned long long task_delta_exec(struct task_struct
*p
)
2387 unsigned long flags
;
2391 rq
= task_rq_lock(p
, &flags
);
2392 ns
= do_task_delta_exec(p
, rq
);
2393 task_rq_unlock(rq
, p
, &flags
);
2399 * Return accounted runtime for the task.
2400 * In case the task is currently running, return the runtime plus current's
2401 * pending runtime that have not been accounted yet.
2403 unsigned long long task_sched_runtime(struct task_struct
*p
)
2405 unsigned long flags
;
2409 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2411 * 64-bit doesn't need locks to atomically read a 64bit value.
2412 * So we have a optimization chance when the task's delta_exec is 0.
2413 * Reading ->on_cpu is racy, but this is ok.
2415 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2416 * If we race with it entering cpu, unaccounted time is 0. This is
2417 * indistinguishable from the read occurring a few cycles earlier.
2420 return p
->se
.sum_exec_runtime
;
2423 rq
= task_rq_lock(p
, &flags
);
2424 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2425 task_rq_unlock(rq
, p
, &flags
);
2431 * This function gets called by the timer code, with HZ frequency.
2432 * We call it with interrupts disabled.
2434 void scheduler_tick(void)
2436 int cpu
= smp_processor_id();
2437 struct rq
*rq
= cpu_rq(cpu
);
2438 struct task_struct
*curr
= rq
->curr
;
2442 raw_spin_lock(&rq
->lock
);
2443 update_rq_clock(rq
);
2444 curr
->sched_class
->task_tick(rq
, curr
, 0);
2445 update_cpu_load_active(rq
);
2446 raw_spin_unlock(&rq
->lock
);
2448 perf_event_task_tick();
2451 rq
->idle_balance
= idle_cpu(cpu
);
2452 trigger_load_balance(rq
);
2454 rq_last_tick_reset(rq
);
2457 #ifdef CONFIG_NO_HZ_FULL
2459 * scheduler_tick_max_deferment
2461 * Keep at least one tick per second when a single
2462 * active task is running because the scheduler doesn't
2463 * yet completely support full dynticks environment.
2465 * This makes sure that uptime, CFS vruntime, load
2466 * balancing, etc... continue to move forward, even
2467 * with a very low granularity.
2469 * Return: Maximum deferment in nanoseconds.
2471 u64
scheduler_tick_max_deferment(void)
2473 struct rq
*rq
= this_rq();
2474 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2476 next
= rq
->last_sched_tick
+ HZ
;
2478 if (time_before_eq(next
, now
))
2481 return jiffies_to_nsecs(next
- now
);
2485 notrace
unsigned long get_parent_ip(unsigned long addr
)
2487 if (in_lock_functions(addr
)) {
2488 addr
= CALLER_ADDR2
;
2489 if (in_lock_functions(addr
))
2490 addr
= CALLER_ADDR3
;
2495 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2496 defined(CONFIG_PREEMPT_TRACER))
2498 void __kprobes
preempt_count_add(int val
)
2500 #ifdef CONFIG_DEBUG_PREEMPT
2504 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2507 __preempt_count_add(val
);
2508 #ifdef CONFIG_DEBUG_PREEMPT
2510 * Spinlock count overflowing soon?
2512 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2515 if (preempt_count() == val
)
2516 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2518 EXPORT_SYMBOL(preempt_count_add
);
2520 void __kprobes
preempt_count_sub(int val
)
2522 #ifdef CONFIG_DEBUG_PREEMPT
2526 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2529 * Is the spinlock portion underflowing?
2531 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2532 !(preempt_count() & PREEMPT_MASK
)))
2536 if (preempt_count() == val
)
2537 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2538 __preempt_count_sub(val
);
2540 EXPORT_SYMBOL(preempt_count_sub
);
2545 * Print scheduling while atomic bug:
2547 static noinline
void __schedule_bug(struct task_struct
*prev
)
2549 if (oops_in_progress
)
2552 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2553 prev
->comm
, prev
->pid
, preempt_count());
2555 debug_show_held_locks(prev
);
2557 if (irqs_disabled())
2558 print_irqtrace_events(prev
);
2560 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2564 * Various schedule()-time debugging checks and statistics:
2566 static inline void schedule_debug(struct task_struct
*prev
)
2569 * Test if we are atomic. Since do_exit() needs to call into
2570 * schedule() atomically, we ignore that path. Otherwise whine
2571 * if we are scheduling when we should not.
2573 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2574 __schedule_bug(prev
);
2577 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2579 schedstat_inc(this_rq(), sched_count
);
2582 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
2584 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2585 update_rq_clock(rq
);
2586 prev
->sched_class
->put_prev_task(rq
, prev
);
2590 * Pick up the highest-prio task:
2592 static inline struct task_struct
*
2593 pick_next_task(struct rq
*rq
)
2595 const struct sched_class
*class;
2596 struct task_struct
*p
;
2599 * Optimization: we know that if all tasks are in
2600 * the fair class we can call that function directly:
2602 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2603 p
= fair_sched_class
.pick_next_task(rq
);
2608 for_each_class(class) {
2609 p
= class->pick_next_task(rq
);
2614 BUG(); /* the idle class will always have a runnable task */
2618 * __schedule() is the main scheduler function.
2620 * The main means of driving the scheduler and thus entering this function are:
2622 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2624 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2625 * paths. For example, see arch/x86/entry_64.S.
2627 * To drive preemption between tasks, the scheduler sets the flag in timer
2628 * interrupt handler scheduler_tick().
2630 * 3. Wakeups don't really cause entry into schedule(). They add a
2631 * task to the run-queue and that's it.
2633 * Now, if the new task added to the run-queue preempts the current
2634 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2635 * called on the nearest possible occasion:
2637 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2639 * - in syscall or exception context, at the next outmost
2640 * preempt_enable(). (this might be as soon as the wake_up()'s
2643 * - in IRQ context, return from interrupt-handler to
2644 * preemptible context
2646 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2649 * - cond_resched() call
2650 * - explicit schedule() call
2651 * - return from syscall or exception to user-space
2652 * - return from interrupt-handler to user-space
2654 static void __sched
__schedule(void)
2656 struct task_struct
*prev
, *next
;
2657 unsigned long *switch_count
;
2663 cpu
= smp_processor_id();
2665 rcu_note_context_switch(cpu
);
2668 schedule_debug(prev
);
2670 if (sched_feat(HRTICK
))
2674 * Make sure that signal_pending_state()->signal_pending() below
2675 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2676 * done by the caller to avoid the race with signal_wake_up().
2678 smp_mb__before_spinlock();
2679 raw_spin_lock_irq(&rq
->lock
);
2681 switch_count
= &prev
->nivcsw
;
2682 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2683 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2684 prev
->state
= TASK_RUNNING
;
2686 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2690 * If a worker went to sleep, notify and ask workqueue
2691 * whether it wants to wake up a task to maintain
2694 if (prev
->flags
& PF_WQ_WORKER
) {
2695 struct task_struct
*to_wakeup
;
2697 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2699 try_to_wake_up_local(to_wakeup
);
2702 switch_count
= &prev
->nvcsw
;
2705 pre_schedule(rq
, prev
);
2707 if (unlikely(!rq
->nr_running
))
2708 idle_balance(cpu
, rq
);
2710 put_prev_task(rq
, prev
);
2711 next
= pick_next_task(rq
);
2712 clear_tsk_need_resched(prev
);
2713 clear_preempt_need_resched();
2714 rq
->skip_clock_update
= 0;
2716 if (likely(prev
!= next
)) {
2721 context_switch(rq
, prev
, next
); /* unlocks the rq */
2723 * The context switch have flipped the stack from under us
2724 * and restored the local variables which were saved when
2725 * this task called schedule() in the past. prev == current
2726 * is still correct, but it can be moved to another cpu/rq.
2728 cpu
= smp_processor_id();
2731 raw_spin_unlock_irq(&rq
->lock
);
2735 sched_preempt_enable_no_resched();
2740 static inline void sched_submit_work(struct task_struct
*tsk
)
2742 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2745 * If we are going to sleep and we have plugged IO queued,
2746 * make sure to submit it to avoid deadlocks.
2748 if (blk_needs_flush_plug(tsk
))
2749 blk_schedule_flush_plug(tsk
);
2752 asmlinkage
void __sched
schedule(void)
2754 struct task_struct
*tsk
= current
;
2756 sched_submit_work(tsk
);
2759 EXPORT_SYMBOL(schedule
);
2761 #ifdef CONFIG_CONTEXT_TRACKING
2762 asmlinkage
void __sched
schedule_user(void)
2765 * If we come here after a random call to set_need_resched(),
2766 * or we have been woken up remotely but the IPI has not yet arrived,
2767 * we haven't yet exited the RCU idle mode. Do it here manually until
2768 * we find a better solution.
2777 * schedule_preempt_disabled - called with preemption disabled
2779 * Returns with preemption disabled. Note: preempt_count must be 1
2781 void __sched
schedule_preempt_disabled(void)
2783 sched_preempt_enable_no_resched();
2788 #ifdef CONFIG_PREEMPT
2790 * this is the entry point to schedule() from in-kernel preemption
2791 * off of preempt_enable. Kernel preemptions off return from interrupt
2792 * occur there and call schedule directly.
2794 asmlinkage
void __sched notrace
preempt_schedule(void)
2797 * If there is a non-zero preempt_count or interrupts are disabled,
2798 * we do not want to preempt the current task. Just return..
2800 if (likely(!preemptible()))
2804 __preempt_count_add(PREEMPT_ACTIVE
);
2806 __preempt_count_sub(PREEMPT_ACTIVE
);
2809 * Check again in case we missed a preemption opportunity
2810 * between schedule and now.
2813 } while (need_resched());
2815 EXPORT_SYMBOL(preempt_schedule
);
2816 #endif /* CONFIG_PREEMPT */
2819 * this is the entry point to schedule() from kernel preemption
2820 * off of irq context.
2821 * Note, that this is called and return with irqs disabled. This will
2822 * protect us against recursive calling from irq.
2824 asmlinkage
void __sched
preempt_schedule_irq(void)
2826 enum ctx_state prev_state
;
2828 /* Catch callers which need to be fixed */
2829 BUG_ON(preempt_count() || !irqs_disabled());
2831 prev_state
= exception_enter();
2834 __preempt_count_add(PREEMPT_ACTIVE
);
2837 local_irq_disable();
2838 __preempt_count_sub(PREEMPT_ACTIVE
);
2841 * Check again in case we missed a preemption opportunity
2842 * between schedule and now.
2845 } while (need_resched());
2847 exception_exit(prev_state
);
2850 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2853 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2855 EXPORT_SYMBOL(default_wake_function
);
2858 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
2860 unsigned long flags
;
2863 init_waitqueue_entry(&wait
, current
);
2865 __set_current_state(state
);
2867 spin_lock_irqsave(&q
->lock
, flags
);
2868 __add_wait_queue(q
, &wait
);
2869 spin_unlock(&q
->lock
);
2870 timeout
= schedule_timeout(timeout
);
2871 spin_lock_irq(&q
->lock
);
2872 __remove_wait_queue(q
, &wait
);
2873 spin_unlock_irqrestore(&q
->lock
, flags
);
2878 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
2880 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2882 EXPORT_SYMBOL(interruptible_sleep_on
);
2885 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
2887 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
2889 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
2891 void __sched
sleep_on(wait_queue_head_t
*q
)
2893 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2895 EXPORT_SYMBOL(sleep_on
);
2897 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
2899 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
2901 EXPORT_SYMBOL(sleep_on_timeout
);
2903 #ifdef CONFIG_RT_MUTEXES
2906 * rt_mutex_setprio - set the current priority of a task
2908 * @prio: prio value (kernel-internal form)
2910 * This function changes the 'effective' priority of a task. It does
2911 * not touch ->normal_prio like __setscheduler().
2913 * Used by the rt_mutex code to implement priority inheritance logic.
2915 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2917 int oldprio
, on_rq
, running
, enqueue_flag
= 0;
2919 const struct sched_class
*prev_class
;
2921 BUG_ON(prio
> MAX_PRIO
);
2923 rq
= __task_rq_lock(p
);
2926 * Idle task boosting is a nono in general. There is one
2927 * exception, when PREEMPT_RT and NOHZ is active:
2929 * The idle task calls get_next_timer_interrupt() and holds
2930 * the timer wheel base->lock on the CPU and another CPU wants
2931 * to access the timer (probably to cancel it). We can safely
2932 * ignore the boosting request, as the idle CPU runs this code
2933 * with interrupts disabled and will complete the lock
2934 * protected section without being interrupted. So there is no
2935 * real need to boost.
2937 if (unlikely(p
== rq
->idle
)) {
2938 WARN_ON(p
!= rq
->curr
);
2939 WARN_ON(p
->pi_blocked_on
);
2943 trace_sched_pi_setprio(p
, prio
);
2944 p
->pi_top_task
= rt_mutex_get_top_task(p
);
2946 prev_class
= p
->sched_class
;
2948 running
= task_current(rq
, p
);
2950 dequeue_task(rq
, p
, 0);
2952 p
->sched_class
->put_prev_task(rq
, p
);
2955 * Boosting condition are:
2956 * 1. -rt task is running and holds mutex A
2957 * --> -dl task blocks on mutex A
2959 * 2. -dl task is running and holds mutex A
2960 * --> -dl task blocks on mutex A and could preempt the
2963 if (dl_prio(prio
)) {
2964 if (!dl_prio(p
->normal_prio
) || (p
->pi_top_task
&&
2965 dl_entity_preempt(&p
->pi_top_task
->dl
, &p
->dl
))) {
2966 p
->dl
.dl_boosted
= 1;
2967 p
->dl
.dl_throttled
= 0;
2968 enqueue_flag
= ENQUEUE_REPLENISH
;
2970 p
->dl
.dl_boosted
= 0;
2971 p
->sched_class
= &dl_sched_class
;
2972 } else if (rt_prio(prio
)) {
2973 if (dl_prio(oldprio
))
2974 p
->dl
.dl_boosted
= 0;
2976 enqueue_flag
= ENQUEUE_HEAD
;
2977 p
->sched_class
= &rt_sched_class
;
2979 if (dl_prio(oldprio
))
2980 p
->dl
.dl_boosted
= 0;
2981 p
->sched_class
= &fair_sched_class
;
2987 p
->sched_class
->set_curr_task(rq
);
2989 enqueue_task(rq
, p
, enqueue_flag
);
2991 check_class_changed(rq
, p
, prev_class
, oldprio
);
2993 __task_rq_unlock(rq
);
2997 void set_user_nice(struct task_struct
*p
, long nice
)
2999 int old_prio
, delta
, on_rq
;
3000 unsigned long flags
;
3003 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
3006 * We have to be careful, if called from sys_setpriority(),
3007 * the task might be in the middle of scheduling on another CPU.
3009 rq
= task_rq_lock(p
, &flags
);
3011 * The RT priorities are set via sched_setscheduler(), but we still
3012 * allow the 'normal' nice value to be set - but as expected
3013 * it wont have any effect on scheduling until the task is
3014 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3016 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3017 p
->static_prio
= NICE_TO_PRIO(nice
);
3022 dequeue_task(rq
, p
, 0);
3024 p
->static_prio
= NICE_TO_PRIO(nice
);
3027 p
->prio
= effective_prio(p
);
3028 delta
= p
->prio
- old_prio
;
3031 enqueue_task(rq
, p
, 0);
3033 * If the task increased its priority or is running and
3034 * lowered its priority, then reschedule its CPU:
3036 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3037 resched_task(rq
->curr
);
3040 task_rq_unlock(rq
, p
, &flags
);
3042 EXPORT_SYMBOL(set_user_nice
);
3045 * can_nice - check if a task can reduce its nice value
3049 int can_nice(const struct task_struct
*p
, const int nice
)
3051 /* convert nice value [19,-20] to rlimit style value [1,40] */
3052 int nice_rlim
= 20 - nice
;
3054 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3055 capable(CAP_SYS_NICE
));
3058 #ifdef __ARCH_WANT_SYS_NICE
3061 * sys_nice - change the priority of the current process.
3062 * @increment: priority increment
3064 * sys_setpriority is a more generic, but much slower function that
3065 * does similar things.
3067 SYSCALL_DEFINE1(nice
, int, increment
)
3072 * Setpriority might change our priority at the same moment.
3073 * We don't have to worry. Conceptually one call occurs first
3074 * and we have a single winner.
3076 if (increment
< -40)
3081 nice
= TASK_NICE(current
) + increment
;
3087 if (increment
< 0 && !can_nice(current
, nice
))
3090 retval
= security_task_setnice(current
, nice
);
3094 set_user_nice(current
, nice
);
3101 * task_prio - return the priority value of a given task.
3102 * @p: the task in question.
3104 * Return: The priority value as seen by users in /proc.
3105 * RT tasks are offset by -200. Normal tasks are centered
3106 * around 0, value goes from -16 to +15.
3108 int task_prio(const struct task_struct
*p
)
3110 return p
->prio
- MAX_RT_PRIO
;
3114 * task_nice - return the nice value of a given task.
3115 * @p: the task in question.
3117 * Return: The nice value [ -20 ... 0 ... 19 ].
3119 int task_nice(const struct task_struct
*p
)
3121 return TASK_NICE(p
);
3123 EXPORT_SYMBOL(task_nice
);
3126 * idle_cpu - is a given cpu idle currently?
3127 * @cpu: the processor in question.
3129 * Return: 1 if the CPU is currently idle. 0 otherwise.
3131 int idle_cpu(int cpu
)
3133 struct rq
*rq
= cpu_rq(cpu
);
3135 if (rq
->curr
!= rq
->idle
)
3142 if (!llist_empty(&rq
->wake_list
))
3150 * idle_task - return the idle task for a given cpu.
3151 * @cpu: the processor in question.
3153 * Return: The idle task for the cpu @cpu.
3155 struct task_struct
*idle_task(int cpu
)
3157 return cpu_rq(cpu
)->idle
;
3161 * find_process_by_pid - find a process with a matching PID value.
3162 * @pid: the pid in question.
3164 * The task of @pid, if found. %NULL otherwise.
3166 static struct task_struct
*find_process_by_pid(pid_t pid
)
3168 return pid
? find_task_by_vpid(pid
) : current
;
3172 * This function initializes the sched_dl_entity of a newly becoming
3173 * SCHED_DEADLINE task.
3175 * Only the static values are considered here, the actual runtime and the
3176 * absolute deadline will be properly calculated when the task is enqueued
3177 * for the first time with its new policy.
3180 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3182 struct sched_dl_entity
*dl_se
= &p
->dl
;
3184 init_dl_task_timer(dl_se
);
3185 dl_se
->dl_runtime
= attr
->sched_runtime
;
3186 dl_se
->dl_deadline
= attr
->sched_deadline
;
3187 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3188 dl_se
->flags
= attr
->sched_flags
;
3189 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3190 dl_se
->dl_throttled
= 0;
3194 /* Actually do priority change: must hold pi & rq lock. */
3195 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3196 const struct sched_attr
*attr
)
3198 int policy
= attr
->sched_policy
;
3200 if (policy
== -1) /* setparam */
3205 if (dl_policy(policy
))
3206 __setparam_dl(p
, attr
);
3207 else if (fair_policy(policy
))
3208 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3211 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3212 * !rt_policy. Always setting this ensures that things like
3213 * getparam()/getattr() don't report silly values for !rt tasks.
3215 p
->rt_priority
= attr
->sched_priority
;
3217 p
->normal_prio
= normal_prio(p
);
3218 p
->prio
= rt_mutex_getprio(p
);
3220 if (dl_prio(p
->prio
))
3221 p
->sched_class
= &dl_sched_class
;
3222 else if (rt_prio(p
->prio
))
3223 p
->sched_class
= &rt_sched_class
;
3225 p
->sched_class
= &fair_sched_class
;
3231 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3233 struct sched_dl_entity
*dl_se
= &p
->dl
;
3235 attr
->sched_priority
= p
->rt_priority
;
3236 attr
->sched_runtime
= dl_se
->dl_runtime
;
3237 attr
->sched_deadline
= dl_se
->dl_deadline
;
3238 attr
->sched_period
= dl_se
->dl_period
;
3239 attr
->sched_flags
= dl_se
->flags
;
3243 * This function validates the new parameters of a -deadline task.
3244 * We ask for the deadline not being zero, and greater or equal
3245 * than the runtime, as well as the period of being zero or
3246 * greater than deadline. Furthermore, we have to be sure that
3247 * user parameters are above the internal resolution (1us); we
3248 * check sched_runtime only since it is always the smaller one.
3251 __checkparam_dl(const struct sched_attr
*attr
)
3253 return attr
&& attr
->sched_deadline
!= 0 &&
3254 (attr
->sched_period
== 0 ||
3255 (s64
)(attr
->sched_period
- attr
->sched_deadline
) >= 0) &&
3256 (s64
)(attr
->sched_deadline
- attr
->sched_runtime
) >= 0 &&
3257 attr
->sched_runtime
>= (2 << (DL_SCALE
- 1));
3261 * check the target process has a UID that matches the current process's
3263 static bool check_same_owner(struct task_struct
*p
)
3265 const struct cred
*cred
= current_cred(), *pcred
;
3269 pcred
= __task_cred(p
);
3270 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3271 uid_eq(cred
->euid
, pcred
->uid
));
3276 static int __sched_setscheduler(struct task_struct
*p
,
3277 const struct sched_attr
*attr
,
3280 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3281 int policy
= attr
->sched_policy
;
3282 unsigned long flags
;
3283 const struct sched_class
*prev_class
;
3287 /* may grab non-irq protected spin_locks */
3288 BUG_ON(in_interrupt());
3290 /* double check policy once rq lock held */
3292 reset_on_fork
= p
->sched_reset_on_fork
;
3293 policy
= oldpolicy
= p
->policy
;
3295 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3297 if (policy
!= SCHED_DEADLINE
&&
3298 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3299 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3300 policy
!= SCHED_IDLE
)
3304 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3308 * Valid priorities for SCHED_FIFO and SCHED_RR are
3309 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3310 * SCHED_BATCH and SCHED_IDLE is 0.
3312 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3313 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3315 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3316 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3320 * Allow unprivileged RT tasks to decrease priority:
3322 if (user
&& !capable(CAP_SYS_NICE
)) {
3323 if (fair_policy(policy
)) {
3324 if (attr
->sched_nice
< TASK_NICE(p
) &&
3325 !can_nice(p
, attr
->sched_nice
))
3329 if (rt_policy(policy
)) {
3330 unsigned long rlim_rtprio
=
3331 task_rlimit(p
, RLIMIT_RTPRIO
);
3333 /* can't set/change the rt policy */
3334 if (policy
!= p
->policy
&& !rlim_rtprio
)
3337 /* can't increase priority */
3338 if (attr
->sched_priority
> p
->rt_priority
&&
3339 attr
->sched_priority
> rlim_rtprio
)
3344 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3345 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3347 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3348 if (!can_nice(p
, TASK_NICE(p
)))
3352 /* can't change other user's priorities */
3353 if (!check_same_owner(p
))
3356 /* Normal users shall not reset the sched_reset_on_fork flag */
3357 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3362 retval
= security_task_setscheduler(p
);
3368 * make sure no PI-waiters arrive (or leave) while we are
3369 * changing the priority of the task:
3371 * To be able to change p->policy safely, the appropriate
3372 * runqueue lock must be held.
3374 rq
= task_rq_lock(p
, &flags
);
3377 * Changing the policy of the stop threads its a very bad idea
3379 if (p
== rq
->stop
) {
3380 task_rq_unlock(rq
, p
, &flags
);
3385 * If not changing anything there's no need to proceed further:
3387 if (unlikely(policy
== p
->policy
)) {
3388 if (fair_policy(policy
) && attr
->sched_nice
!= TASK_NICE(p
))
3390 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3392 if (dl_policy(policy
))
3395 task_rq_unlock(rq
, p
, &flags
);
3401 #ifdef CONFIG_RT_GROUP_SCHED
3403 * Do not allow realtime tasks into groups that have no runtime
3406 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3407 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3408 !task_group_is_autogroup(task_group(p
))) {
3409 task_rq_unlock(rq
, p
, &flags
);
3414 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3415 cpumask_t
*span
= rq
->rd
->span
;
3418 * Don't allow tasks with an affinity mask smaller than
3419 * the entire root_domain to become SCHED_DEADLINE. We
3420 * will also fail if there's no bandwidth available.
3422 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3423 rq
->rd
->dl_bw
.bw
== 0) {
3424 task_rq_unlock(rq
, p
, &flags
);
3431 /* recheck policy now with rq lock held */
3432 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3433 policy
= oldpolicy
= -1;
3434 task_rq_unlock(rq
, p
, &flags
);
3439 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3440 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3443 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3444 task_rq_unlock(rq
, p
, &flags
);
3449 running
= task_current(rq
, p
);
3451 dequeue_task(rq
, p
, 0);
3453 p
->sched_class
->put_prev_task(rq
, p
);
3455 p
->sched_reset_on_fork
= reset_on_fork
;
3458 prev_class
= p
->sched_class
;
3459 __setscheduler(rq
, p
, attr
);
3462 p
->sched_class
->set_curr_task(rq
);
3464 enqueue_task(rq
, p
, 0);
3466 check_class_changed(rq
, p
, prev_class
, oldprio
);
3467 task_rq_unlock(rq
, p
, &flags
);
3469 rt_mutex_adjust_pi(p
);
3474 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3475 const struct sched_param
*param
, bool check
)
3477 struct sched_attr attr
= {
3478 .sched_policy
= policy
,
3479 .sched_priority
= param
->sched_priority
,
3480 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3484 * Fixup the legacy SCHED_RESET_ON_FORK hack
3486 if (policy
& SCHED_RESET_ON_FORK
) {
3487 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3488 policy
&= ~SCHED_RESET_ON_FORK
;
3489 attr
.sched_policy
= policy
;
3492 return __sched_setscheduler(p
, &attr
, check
);
3495 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3496 * @p: the task in question.
3497 * @policy: new policy.
3498 * @param: structure containing the new RT priority.
3500 * Return: 0 on success. An error code otherwise.
3502 * NOTE that the task may be already dead.
3504 int sched_setscheduler(struct task_struct
*p
, int policy
,
3505 const struct sched_param
*param
)
3507 return _sched_setscheduler(p
, policy
, param
, true);
3509 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3511 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3513 return __sched_setscheduler(p
, attr
, true);
3515 EXPORT_SYMBOL_GPL(sched_setattr
);
3518 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3519 * @p: the task in question.
3520 * @policy: new policy.
3521 * @param: structure containing the new RT priority.
3523 * Just like sched_setscheduler, only don't bother checking if the
3524 * current context has permission. For example, this is needed in
3525 * stop_machine(): we create temporary high priority worker threads,
3526 * but our caller might not have that capability.
3528 * Return: 0 on success. An error code otherwise.
3530 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3531 const struct sched_param
*param
)
3533 return _sched_setscheduler(p
, policy
, param
, false);
3537 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3539 struct sched_param lparam
;
3540 struct task_struct
*p
;
3543 if (!param
|| pid
< 0)
3545 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3550 p
= find_process_by_pid(pid
);
3552 retval
= sched_setscheduler(p
, policy
, &lparam
);
3559 * Mimics kernel/events/core.c perf_copy_attr().
3561 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3562 struct sched_attr
*attr
)
3567 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3571 * zero the full structure, so that a short copy will be nice.
3573 memset(attr
, 0, sizeof(*attr
));
3575 ret
= get_user(size
, &uattr
->size
);
3579 if (size
> PAGE_SIZE
) /* silly large */
3582 if (!size
) /* abi compat */
3583 size
= SCHED_ATTR_SIZE_VER0
;
3585 if (size
< SCHED_ATTR_SIZE_VER0
)
3589 * If we're handed a bigger struct than we know of,
3590 * ensure all the unknown bits are 0 - i.e. new
3591 * user-space does not rely on any kernel feature
3592 * extensions we dont know about yet.
3594 if (size
> sizeof(*attr
)) {
3595 unsigned char __user
*addr
;
3596 unsigned char __user
*end
;
3599 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3600 end
= (void __user
*)uattr
+ size
;
3602 for (; addr
< end
; addr
++) {
3603 ret
= get_user(val
, addr
);
3609 size
= sizeof(*attr
);
3612 ret
= copy_from_user(attr
, uattr
, size
);
3617 * XXX: do we want to be lenient like existing syscalls; or do we want
3618 * to be strict and return an error on out-of-bounds values?
3620 attr
->sched_nice
= clamp(attr
->sched_nice
, -20, 19);
3626 put_user(sizeof(*attr
), &uattr
->size
);
3632 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3633 * @pid: the pid in question.
3634 * @policy: new policy.
3635 * @param: structure containing the new RT priority.
3637 * Return: 0 on success. An error code otherwise.
3639 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3640 struct sched_param __user
*, param
)
3642 /* negative values for policy are not valid */
3646 return do_sched_setscheduler(pid
, policy
, param
);
3650 * sys_sched_setparam - set/change the RT priority of a thread
3651 * @pid: the pid in question.
3652 * @param: structure containing the new RT priority.
3654 * Return: 0 on success. An error code otherwise.
3656 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3658 return do_sched_setscheduler(pid
, -1, param
);
3662 * sys_sched_setattr - same as above, but with extended sched_attr
3663 * @pid: the pid in question.
3664 * @uattr: structure containing the extended parameters.
3666 SYSCALL_DEFINE2(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
)
3668 struct sched_attr attr
;
3669 struct task_struct
*p
;
3672 if (!uattr
|| pid
< 0)
3675 if (sched_copy_attr(uattr
, &attr
))
3680 p
= find_process_by_pid(pid
);
3682 retval
= sched_setattr(p
, &attr
);
3689 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3690 * @pid: the pid in question.
3692 * Return: On success, the policy of the thread. Otherwise, a negative error
3695 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3697 struct task_struct
*p
;
3705 p
= find_process_by_pid(pid
);
3707 retval
= security_task_getscheduler(p
);
3710 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3717 * sys_sched_getparam - get the RT priority of a thread
3718 * @pid: the pid in question.
3719 * @param: structure containing the RT priority.
3721 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3724 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3726 struct sched_param lp
;
3727 struct task_struct
*p
;
3730 if (!param
|| pid
< 0)
3734 p
= find_process_by_pid(pid
);
3739 retval
= security_task_getscheduler(p
);
3743 if (task_has_dl_policy(p
)) {
3747 lp
.sched_priority
= p
->rt_priority
;
3751 * This one might sleep, we cannot do it with a spinlock held ...
3753 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3762 static int sched_read_attr(struct sched_attr __user
*uattr
,
3763 struct sched_attr
*attr
,
3768 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3772 * If we're handed a smaller struct than we know of,
3773 * ensure all the unknown bits are 0 - i.e. old
3774 * user-space does not get uncomplete information.
3776 if (usize
< sizeof(*attr
)) {
3777 unsigned char *addr
;
3780 addr
= (void *)attr
+ usize
;
3781 end
= (void *)attr
+ sizeof(*attr
);
3783 for (; addr
< end
; addr
++) {
3791 ret
= copy_to_user(uattr
, attr
, usize
);
3804 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3805 * @pid: the pid in question.
3806 * @uattr: structure containing the extended parameters.
3807 * @size: sizeof(attr) for fwd/bwd comp.
3809 SYSCALL_DEFINE3(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3812 struct sched_attr attr
= {
3813 .size
= sizeof(struct sched_attr
),
3815 struct task_struct
*p
;
3818 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3819 size
< SCHED_ATTR_SIZE_VER0
)
3823 p
= find_process_by_pid(pid
);
3828 retval
= security_task_getscheduler(p
);
3832 attr
.sched_policy
= p
->policy
;
3833 if (p
->sched_reset_on_fork
)
3834 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3835 if (task_has_dl_policy(p
))
3836 __getparam_dl(p
, &attr
);
3837 else if (task_has_rt_policy(p
))
3838 attr
.sched_priority
= p
->rt_priority
;
3840 attr
.sched_nice
= TASK_NICE(p
);
3844 retval
= sched_read_attr(uattr
, &attr
, size
);
3852 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3854 cpumask_var_t cpus_allowed
, new_mask
;
3855 struct task_struct
*p
;
3860 p
= find_process_by_pid(pid
);
3866 /* Prevent p going away */
3870 if (p
->flags
& PF_NO_SETAFFINITY
) {
3874 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3878 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3880 goto out_free_cpus_allowed
;
3883 if (!check_same_owner(p
)) {
3885 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3892 retval
= security_task_setscheduler(p
);
3897 cpuset_cpus_allowed(p
, cpus_allowed
);
3898 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3901 * Since bandwidth control happens on root_domain basis,
3902 * if admission test is enabled, we only admit -deadline
3903 * tasks allowed to run on all the CPUs in the task's
3907 if (task_has_dl_policy(p
)) {
3908 const struct cpumask
*span
= task_rq(p
)->rd
->span
;
3910 if (dl_bandwidth_enabled() && !cpumask_subset(span
, new_mask
)) {
3917 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3920 cpuset_cpus_allowed(p
, cpus_allowed
);
3921 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3923 * We must have raced with a concurrent cpuset
3924 * update. Just reset the cpus_allowed to the
3925 * cpuset's cpus_allowed
3927 cpumask_copy(new_mask
, cpus_allowed
);
3932 free_cpumask_var(new_mask
);
3933 out_free_cpus_allowed
:
3934 free_cpumask_var(cpus_allowed
);
3940 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3941 struct cpumask
*new_mask
)
3943 if (len
< cpumask_size())
3944 cpumask_clear(new_mask
);
3945 else if (len
> cpumask_size())
3946 len
= cpumask_size();
3948 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3952 * sys_sched_setaffinity - set the cpu affinity of a process
3953 * @pid: pid of the process
3954 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3955 * @user_mask_ptr: user-space pointer to the new cpu mask
3957 * Return: 0 on success. An error code otherwise.
3959 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3960 unsigned long __user
*, user_mask_ptr
)
3962 cpumask_var_t new_mask
;
3965 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3968 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3970 retval
= sched_setaffinity(pid
, new_mask
);
3971 free_cpumask_var(new_mask
);
3975 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
3977 struct task_struct
*p
;
3978 unsigned long flags
;
3984 p
= find_process_by_pid(pid
);
3988 retval
= security_task_getscheduler(p
);
3992 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3993 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
3994 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4003 * sys_sched_getaffinity - get the cpu affinity of a process
4004 * @pid: pid of the process
4005 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4006 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4008 * Return: 0 on success. An error code otherwise.
4010 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4011 unsigned long __user
*, user_mask_ptr
)
4016 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4018 if (len
& (sizeof(unsigned long)-1))
4021 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4024 ret
= sched_getaffinity(pid
, mask
);
4026 size_t retlen
= min_t(size_t, len
, cpumask_size());
4028 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4033 free_cpumask_var(mask
);
4039 * sys_sched_yield - yield the current processor to other threads.
4041 * This function yields the current CPU to other tasks. If there are no
4042 * other threads running on this CPU then this function will return.
4046 SYSCALL_DEFINE0(sched_yield
)
4048 struct rq
*rq
= this_rq_lock();
4050 schedstat_inc(rq
, yld_count
);
4051 current
->sched_class
->yield_task(rq
);
4054 * Since we are going to call schedule() anyway, there's
4055 * no need to preempt or enable interrupts:
4057 __release(rq
->lock
);
4058 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4059 do_raw_spin_unlock(&rq
->lock
);
4060 sched_preempt_enable_no_resched();
4067 static void __cond_resched(void)
4069 __preempt_count_add(PREEMPT_ACTIVE
);
4071 __preempt_count_sub(PREEMPT_ACTIVE
);
4074 int __sched
_cond_resched(void)
4076 if (should_resched()) {
4082 EXPORT_SYMBOL(_cond_resched
);
4085 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4086 * call schedule, and on return reacquire the lock.
4088 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4089 * operations here to prevent schedule() from being called twice (once via
4090 * spin_unlock(), once by hand).
4092 int __cond_resched_lock(spinlock_t
*lock
)
4094 int resched
= should_resched();
4097 lockdep_assert_held(lock
);
4099 if (spin_needbreak(lock
) || resched
) {
4110 EXPORT_SYMBOL(__cond_resched_lock
);
4112 int __sched
__cond_resched_softirq(void)
4114 BUG_ON(!in_softirq());
4116 if (should_resched()) {
4124 EXPORT_SYMBOL(__cond_resched_softirq
);
4127 * yield - yield the current processor to other threads.
4129 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4131 * The scheduler is at all times free to pick the calling task as the most
4132 * eligible task to run, if removing the yield() call from your code breaks
4133 * it, its already broken.
4135 * Typical broken usage is:
4140 * where one assumes that yield() will let 'the other' process run that will
4141 * make event true. If the current task is a SCHED_FIFO task that will never
4142 * happen. Never use yield() as a progress guarantee!!
4144 * If you want to use yield() to wait for something, use wait_event().
4145 * If you want to use yield() to be 'nice' for others, use cond_resched().
4146 * If you still want to use yield(), do not!
4148 void __sched
yield(void)
4150 set_current_state(TASK_RUNNING
);
4153 EXPORT_SYMBOL(yield
);
4156 * yield_to - yield the current processor to another thread in
4157 * your thread group, or accelerate that thread toward the
4158 * processor it's on.
4160 * @preempt: whether task preemption is allowed or not
4162 * It's the caller's job to ensure that the target task struct
4163 * can't go away on us before we can do any checks.
4166 * true (>0) if we indeed boosted the target task.
4167 * false (0) if we failed to boost the target.
4168 * -ESRCH if there's no task to yield to.
4170 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4172 struct task_struct
*curr
= current
;
4173 struct rq
*rq
, *p_rq
;
4174 unsigned long flags
;
4177 local_irq_save(flags
);
4183 * If we're the only runnable task on the rq and target rq also
4184 * has only one task, there's absolutely no point in yielding.
4186 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4191 double_rq_lock(rq
, p_rq
);
4192 if (task_rq(p
) != p_rq
) {
4193 double_rq_unlock(rq
, p_rq
);
4197 if (!curr
->sched_class
->yield_to_task
)
4200 if (curr
->sched_class
!= p
->sched_class
)
4203 if (task_running(p_rq
, p
) || p
->state
)
4206 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4208 schedstat_inc(rq
, yld_count
);
4210 * Make p's CPU reschedule; pick_next_entity takes care of
4213 if (preempt
&& rq
!= p_rq
)
4214 resched_task(p_rq
->curr
);
4218 double_rq_unlock(rq
, p_rq
);
4220 local_irq_restore(flags
);
4227 EXPORT_SYMBOL_GPL(yield_to
);
4230 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4231 * that process accounting knows that this is a task in IO wait state.
4233 void __sched
io_schedule(void)
4235 struct rq
*rq
= raw_rq();
4237 delayacct_blkio_start();
4238 atomic_inc(&rq
->nr_iowait
);
4239 blk_flush_plug(current
);
4240 current
->in_iowait
= 1;
4242 current
->in_iowait
= 0;
4243 atomic_dec(&rq
->nr_iowait
);
4244 delayacct_blkio_end();
4246 EXPORT_SYMBOL(io_schedule
);
4248 long __sched
io_schedule_timeout(long timeout
)
4250 struct rq
*rq
= raw_rq();
4253 delayacct_blkio_start();
4254 atomic_inc(&rq
->nr_iowait
);
4255 blk_flush_plug(current
);
4256 current
->in_iowait
= 1;
4257 ret
= schedule_timeout(timeout
);
4258 current
->in_iowait
= 0;
4259 atomic_dec(&rq
->nr_iowait
);
4260 delayacct_blkio_end();
4265 * sys_sched_get_priority_max - return maximum RT priority.
4266 * @policy: scheduling class.
4268 * Return: On success, this syscall returns the maximum
4269 * rt_priority that can be used by a given scheduling class.
4270 * On failure, a negative error code is returned.
4272 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4279 ret
= MAX_USER_RT_PRIO
-1;
4281 case SCHED_DEADLINE
:
4292 * sys_sched_get_priority_min - return minimum RT priority.
4293 * @policy: scheduling class.
4295 * Return: On success, this syscall returns the minimum
4296 * rt_priority that can be used by a given scheduling class.
4297 * On failure, a negative error code is returned.
4299 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4308 case SCHED_DEADLINE
:
4318 * sys_sched_rr_get_interval - return the default timeslice of a process.
4319 * @pid: pid of the process.
4320 * @interval: userspace pointer to the timeslice value.
4322 * this syscall writes the default timeslice value of a given process
4323 * into the user-space timespec buffer. A value of '0' means infinity.
4325 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4328 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4329 struct timespec __user
*, interval
)
4331 struct task_struct
*p
;
4332 unsigned int time_slice
;
4333 unsigned long flags
;
4343 p
= find_process_by_pid(pid
);
4347 retval
= security_task_getscheduler(p
);
4351 rq
= task_rq_lock(p
, &flags
);
4353 if (p
->sched_class
->get_rr_interval
)
4354 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4355 task_rq_unlock(rq
, p
, &flags
);
4358 jiffies_to_timespec(time_slice
, &t
);
4359 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4367 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4369 void sched_show_task(struct task_struct
*p
)
4371 unsigned long free
= 0;
4375 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4376 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4377 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4378 #if BITS_PER_LONG == 32
4379 if (state
== TASK_RUNNING
)
4380 printk(KERN_CONT
" running ");
4382 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4384 if (state
== TASK_RUNNING
)
4385 printk(KERN_CONT
" running task ");
4387 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4389 #ifdef CONFIG_DEBUG_STACK_USAGE
4390 free
= stack_not_used(p
);
4393 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4395 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4396 task_pid_nr(p
), ppid
,
4397 (unsigned long)task_thread_info(p
)->flags
);
4399 print_worker_info(KERN_INFO
, p
);
4400 show_stack(p
, NULL
);
4403 void show_state_filter(unsigned long state_filter
)
4405 struct task_struct
*g
, *p
;
4407 #if BITS_PER_LONG == 32
4409 " task PC stack pid father\n");
4412 " task PC stack pid father\n");
4415 do_each_thread(g
, p
) {
4417 * reset the NMI-timeout, listing all files on a slow
4418 * console might take a lot of time:
4420 touch_nmi_watchdog();
4421 if (!state_filter
|| (p
->state
& state_filter
))
4423 } while_each_thread(g
, p
);
4425 touch_all_softlockup_watchdogs();
4427 #ifdef CONFIG_SCHED_DEBUG
4428 sysrq_sched_debug_show();
4432 * Only show locks if all tasks are dumped:
4435 debug_show_all_locks();
4438 void init_idle_bootup_task(struct task_struct
*idle
)
4440 idle
->sched_class
= &idle_sched_class
;
4444 * init_idle - set up an idle thread for a given CPU
4445 * @idle: task in question
4446 * @cpu: cpu the idle task belongs to
4448 * NOTE: this function does not set the idle thread's NEED_RESCHED
4449 * flag, to make booting more robust.
4451 void init_idle(struct task_struct
*idle
, int cpu
)
4453 struct rq
*rq
= cpu_rq(cpu
);
4454 unsigned long flags
;
4456 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4458 __sched_fork(0, idle
);
4459 idle
->state
= TASK_RUNNING
;
4460 idle
->se
.exec_start
= sched_clock();
4462 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4464 * We're having a chicken and egg problem, even though we are
4465 * holding rq->lock, the cpu isn't yet set to this cpu so the
4466 * lockdep check in task_group() will fail.
4468 * Similar case to sched_fork(). / Alternatively we could
4469 * use task_rq_lock() here and obtain the other rq->lock.
4474 __set_task_cpu(idle
, cpu
);
4477 rq
->curr
= rq
->idle
= idle
;
4478 #if defined(CONFIG_SMP)
4481 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4483 /* Set the preempt count _outside_ the spinlocks! */
4484 init_idle_preempt_count(idle
, cpu
);
4487 * The idle tasks have their own, simple scheduling class:
4489 idle
->sched_class
= &idle_sched_class
;
4490 ftrace_graph_init_idle_task(idle
, cpu
);
4491 vtime_init_idle(idle
, cpu
);
4492 #if defined(CONFIG_SMP)
4493 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4498 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4500 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4501 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4503 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4504 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4508 * This is how migration works:
4510 * 1) we invoke migration_cpu_stop() on the target CPU using
4512 * 2) stopper starts to run (implicitly forcing the migrated thread
4514 * 3) it checks whether the migrated task is still in the wrong runqueue.
4515 * 4) if it's in the wrong runqueue then the migration thread removes
4516 * it and puts it into the right queue.
4517 * 5) stopper completes and stop_one_cpu() returns and the migration
4522 * Change a given task's CPU affinity. Migrate the thread to a
4523 * proper CPU and schedule it away if the CPU it's executing on
4524 * is removed from the allowed bitmask.
4526 * NOTE: the caller must have a valid reference to the task, the
4527 * task must not exit() & deallocate itself prematurely. The
4528 * call is not atomic; no spinlocks may be held.
4530 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4532 unsigned long flags
;
4534 unsigned int dest_cpu
;
4537 rq
= task_rq_lock(p
, &flags
);
4539 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4542 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4547 do_set_cpus_allowed(p
, new_mask
);
4549 /* Can the task run on the task's current CPU? If so, we're done */
4550 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4553 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4555 struct migration_arg arg
= { p
, dest_cpu
};
4556 /* Need help from migration thread: drop lock and wait. */
4557 task_rq_unlock(rq
, p
, &flags
);
4558 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4559 tlb_migrate_finish(p
->mm
);
4563 task_rq_unlock(rq
, p
, &flags
);
4567 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4570 * Move (not current) task off this cpu, onto dest cpu. We're doing
4571 * this because either it can't run here any more (set_cpus_allowed()
4572 * away from this CPU, or CPU going down), or because we're
4573 * attempting to rebalance this task on exec (sched_exec).
4575 * So we race with normal scheduler movements, but that's OK, as long
4576 * as the task is no longer on this CPU.
4578 * Returns non-zero if task was successfully migrated.
4580 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4582 struct rq
*rq_dest
, *rq_src
;
4585 if (unlikely(!cpu_active(dest_cpu
)))
4588 rq_src
= cpu_rq(src_cpu
);
4589 rq_dest
= cpu_rq(dest_cpu
);
4591 raw_spin_lock(&p
->pi_lock
);
4592 double_rq_lock(rq_src
, rq_dest
);
4593 /* Already moved. */
4594 if (task_cpu(p
) != src_cpu
)
4596 /* Affinity changed (again). */
4597 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4601 * If we're not on a rq, the next wake-up will ensure we're
4605 dequeue_task(rq_src
, p
, 0);
4606 set_task_cpu(p
, dest_cpu
);
4607 enqueue_task(rq_dest
, p
, 0);
4608 check_preempt_curr(rq_dest
, p
, 0);
4613 double_rq_unlock(rq_src
, rq_dest
);
4614 raw_spin_unlock(&p
->pi_lock
);
4618 #ifdef CONFIG_NUMA_BALANCING
4619 /* Migrate current task p to target_cpu */
4620 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4622 struct migration_arg arg
= { p
, target_cpu
};
4623 int curr_cpu
= task_cpu(p
);
4625 if (curr_cpu
== target_cpu
)
4628 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4631 /* TODO: This is not properly updating schedstats */
4633 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4634 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4638 * Requeue a task on a given node and accurately track the number of NUMA
4639 * tasks on the runqueues
4641 void sched_setnuma(struct task_struct
*p
, int nid
)
4644 unsigned long flags
;
4645 bool on_rq
, running
;
4647 rq
= task_rq_lock(p
, &flags
);
4649 running
= task_current(rq
, p
);
4652 dequeue_task(rq
, p
, 0);
4654 p
->sched_class
->put_prev_task(rq
, p
);
4656 p
->numa_preferred_nid
= nid
;
4659 p
->sched_class
->set_curr_task(rq
);
4661 enqueue_task(rq
, p
, 0);
4662 task_rq_unlock(rq
, p
, &flags
);
4667 * migration_cpu_stop - this will be executed by a highprio stopper thread
4668 * and performs thread migration by bumping thread off CPU then
4669 * 'pushing' onto another runqueue.
4671 static int migration_cpu_stop(void *data
)
4673 struct migration_arg
*arg
= data
;
4676 * The original target cpu might have gone down and we might
4677 * be on another cpu but it doesn't matter.
4679 local_irq_disable();
4680 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4685 #ifdef CONFIG_HOTPLUG_CPU
4688 * Ensures that the idle task is using init_mm right before its cpu goes
4691 void idle_task_exit(void)
4693 struct mm_struct
*mm
= current
->active_mm
;
4695 BUG_ON(cpu_online(smp_processor_id()));
4698 switch_mm(mm
, &init_mm
, current
);
4703 * Since this CPU is going 'away' for a while, fold any nr_active delta
4704 * we might have. Assumes we're called after migrate_tasks() so that the
4705 * nr_active count is stable.
4707 * Also see the comment "Global load-average calculations".
4709 static void calc_load_migrate(struct rq
*rq
)
4711 long delta
= calc_load_fold_active(rq
);
4713 atomic_long_add(delta
, &calc_load_tasks
);
4717 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4718 * try_to_wake_up()->select_task_rq().
4720 * Called with rq->lock held even though we'er in stop_machine() and
4721 * there's no concurrency possible, we hold the required locks anyway
4722 * because of lock validation efforts.
4724 static void migrate_tasks(unsigned int dead_cpu
)
4726 struct rq
*rq
= cpu_rq(dead_cpu
);
4727 struct task_struct
*next
, *stop
= rq
->stop
;
4731 * Fudge the rq selection such that the below task selection loop
4732 * doesn't get stuck on the currently eligible stop task.
4734 * We're currently inside stop_machine() and the rq is either stuck
4735 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4736 * either way we should never end up calling schedule() until we're
4742 * put_prev_task() and pick_next_task() sched
4743 * class method both need to have an up-to-date
4744 * value of rq->clock[_task]
4746 update_rq_clock(rq
);
4750 * There's this thread running, bail when that's the only
4753 if (rq
->nr_running
== 1)
4756 next
= pick_next_task(rq
);
4758 next
->sched_class
->put_prev_task(rq
, next
);
4760 /* Find suitable destination for @next, with force if needed. */
4761 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4762 raw_spin_unlock(&rq
->lock
);
4764 __migrate_task(next
, dead_cpu
, dest_cpu
);
4766 raw_spin_lock(&rq
->lock
);
4772 #endif /* CONFIG_HOTPLUG_CPU */
4774 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4776 static struct ctl_table sd_ctl_dir
[] = {
4778 .procname
= "sched_domain",
4784 static struct ctl_table sd_ctl_root
[] = {
4786 .procname
= "kernel",
4788 .child
= sd_ctl_dir
,
4793 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4795 struct ctl_table
*entry
=
4796 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4801 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4803 struct ctl_table
*entry
;
4806 * In the intermediate directories, both the child directory and
4807 * procname are dynamically allocated and could fail but the mode
4808 * will always be set. In the lowest directory the names are
4809 * static strings and all have proc handlers.
4811 for (entry
= *tablep
; entry
->mode
; entry
++) {
4813 sd_free_ctl_entry(&entry
->child
);
4814 if (entry
->proc_handler
== NULL
)
4815 kfree(entry
->procname
);
4822 static int min_load_idx
= 0;
4823 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4826 set_table_entry(struct ctl_table
*entry
,
4827 const char *procname
, void *data
, int maxlen
,
4828 umode_t mode
, proc_handler
*proc_handler
,
4831 entry
->procname
= procname
;
4833 entry
->maxlen
= maxlen
;
4835 entry
->proc_handler
= proc_handler
;
4838 entry
->extra1
= &min_load_idx
;
4839 entry
->extra2
= &max_load_idx
;
4843 static struct ctl_table
*
4844 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4846 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
4851 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4852 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4853 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4854 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4855 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4856 sizeof(int), 0644, proc_dointvec_minmax
, true);
4857 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4858 sizeof(int), 0644, proc_dointvec_minmax
, true);
4859 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4860 sizeof(int), 0644, proc_dointvec_minmax
, true);
4861 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4862 sizeof(int), 0644, proc_dointvec_minmax
, true);
4863 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4864 sizeof(int), 0644, proc_dointvec_minmax
, true);
4865 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4866 sizeof(int), 0644, proc_dointvec_minmax
, false);
4867 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4868 sizeof(int), 0644, proc_dointvec_minmax
, false);
4869 set_table_entry(&table
[9], "cache_nice_tries",
4870 &sd
->cache_nice_tries
,
4871 sizeof(int), 0644, proc_dointvec_minmax
, false);
4872 set_table_entry(&table
[10], "flags", &sd
->flags
,
4873 sizeof(int), 0644, proc_dointvec_minmax
, false);
4874 set_table_entry(&table
[11], "name", sd
->name
,
4875 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4876 /* &table[12] is terminator */
4881 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4883 struct ctl_table
*entry
, *table
;
4884 struct sched_domain
*sd
;
4885 int domain_num
= 0, i
;
4888 for_each_domain(cpu
, sd
)
4890 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4895 for_each_domain(cpu
, sd
) {
4896 snprintf(buf
, 32, "domain%d", i
);
4897 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4899 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4906 static struct ctl_table_header
*sd_sysctl_header
;
4907 static void register_sched_domain_sysctl(void)
4909 int i
, cpu_num
= num_possible_cpus();
4910 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4913 WARN_ON(sd_ctl_dir
[0].child
);
4914 sd_ctl_dir
[0].child
= entry
;
4919 for_each_possible_cpu(i
) {
4920 snprintf(buf
, 32, "cpu%d", i
);
4921 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4923 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4927 WARN_ON(sd_sysctl_header
);
4928 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4931 /* may be called multiple times per register */
4932 static void unregister_sched_domain_sysctl(void)
4934 if (sd_sysctl_header
)
4935 unregister_sysctl_table(sd_sysctl_header
);
4936 sd_sysctl_header
= NULL
;
4937 if (sd_ctl_dir
[0].child
)
4938 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4941 static void register_sched_domain_sysctl(void)
4944 static void unregister_sched_domain_sysctl(void)
4949 static void set_rq_online(struct rq
*rq
)
4952 const struct sched_class
*class;
4954 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
4957 for_each_class(class) {
4958 if (class->rq_online
)
4959 class->rq_online(rq
);
4964 static void set_rq_offline(struct rq
*rq
)
4967 const struct sched_class
*class;
4969 for_each_class(class) {
4970 if (class->rq_offline
)
4971 class->rq_offline(rq
);
4974 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
4980 * migration_call - callback that gets triggered when a CPU is added.
4981 * Here we can start up the necessary migration thread for the new CPU.
4984 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
4986 int cpu
= (long)hcpu
;
4987 unsigned long flags
;
4988 struct rq
*rq
= cpu_rq(cpu
);
4990 switch (action
& ~CPU_TASKS_FROZEN
) {
4992 case CPU_UP_PREPARE
:
4993 rq
->calc_load_update
= calc_load_update
;
4997 /* Update our root-domain */
4998 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5000 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5004 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5007 #ifdef CONFIG_HOTPLUG_CPU
5009 sched_ttwu_pending();
5010 /* Update our root-domain */
5011 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5013 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5017 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5018 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5022 calc_load_migrate(rq
);
5027 update_max_interval();
5033 * Register at high priority so that task migration (migrate_all_tasks)
5034 * happens before everything else. This has to be lower priority than
5035 * the notifier in the perf_event subsystem, though.
5037 static struct notifier_block migration_notifier
= {
5038 .notifier_call
= migration_call
,
5039 .priority
= CPU_PRI_MIGRATION
,
5042 static int sched_cpu_active(struct notifier_block
*nfb
,
5043 unsigned long action
, void *hcpu
)
5045 switch (action
& ~CPU_TASKS_FROZEN
) {
5047 case CPU_DOWN_FAILED
:
5048 set_cpu_active((long)hcpu
, true);
5055 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5056 unsigned long action
, void *hcpu
)
5058 unsigned long flags
;
5059 long cpu
= (long)hcpu
;
5061 switch (action
& ~CPU_TASKS_FROZEN
) {
5062 case CPU_DOWN_PREPARE
:
5063 set_cpu_active(cpu
, false);
5065 /* explicitly allow suspend */
5066 if (!(action
& CPU_TASKS_FROZEN
)) {
5067 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
5071 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5072 cpus
= dl_bw_cpus(cpu
);
5073 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5074 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5077 return notifier_from_errno(-EBUSY
);
5085 static int __init
migration_init(void)
5087 void *cpu
= (void *)(long)smp_processor_id();
5090 /* Initialize migration for the boot CPU */
5091 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5092 BUG_ON(err
== NOTIFY_BAD
);
5093 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5094 register_cpu_notifier(&migration_notifier
);
5096 /* Register cpu active notifiers */
5097 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5098 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5102 early_initcall(migration_init
);
5107 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5109 #ifdef CONFIG_SCHED_DEBUG
5111 static __read_mostly
int sched_debug_enabled
;
5113 static int __init
sched_debug_setup(char *str
)
5115 sched_debug_enabled
= 1;
5119 early_param("sched_debug", sched_debug_setup
);
5121 static inline bool sched_debug(void)
5123 return sched_debug_enabled
;
5126 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5127 struct cpumask
*groupmask
)
5129 struct sched_group
*group
= sd
->groups
;
5132 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5133 cpumask_clear(groupmask
);
5135 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5137 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5138 printk("does not load-balance\n");
5140 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5145 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5147 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5148 printk(KERN_ERR
"ERROR: domain->span does not contain "
5151 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5152 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5156 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5160 printk(KERN_ERR
"ERROR: group is NULL\n");
5165 * Even though we initialize ->power to something semi-sane,
5166 * we leave power_orig unset. This allows us to detect if
5167 * domain iteration is still funny without causing /0 traps.
5169 if (!group
->sgp
->power_orig
) {
5170 printk(KERN_CONT
"\n");
5171 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5176 if (!cpumask_weight(sched_group_cpus(group
))) {
5177 printk(KERN_CONT
"\n");
5178 printk(KERN_ERR
"ERROR: empty group\n");
5182 if (!(sd
->flags
& SD_OVERLAP
) &&
5183 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5184 printk(KERN_CONT
"\n");
5185 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5189 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5191 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5193 printk(KERN_CONT
" %s", str
);
5194 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5195 printk(KERN_CONT
" (cpu_power = %d)",
5199 group
= group
->next
;
5200 } while (group
!= sd
->groups
);
5201 printk(KERN_CONT
"\n");
5203 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5204 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5207 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5208 printk(KERN_ERR
"ERROR: parent span is not a superset "
5209 "of domain->span\n");
5213 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5217 if (!sched_debug_enabled
)
5221 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5225 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5228 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5236 #else /* !CONFIG_SCHED_DEBUG */
5237 # define sched_domain_debug(sd, cpu) do { } while (0)
5238 static inline bool sched_debug(void)
5242 #endif /* CONFIG_SCHED_DEBUG */
5244 static int sd_degenerate(struct sched_domain
*sd
)
5246 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5249 /* Following flags need at least 2 groups */
5250 if (sd
->flags
& (SD_LOAD_BALANCE
|
5251 SD_BALANCE_NEWIDLE
|
5255 SD_SHARE_PKG_RESOURCES
)) {
5256 if (sd
->groups
!= sd
->groups
->next
)
5260 /* Following flags don't use groups */
5261 if (sd
->flags
& (SD_WAKE_AFFINE
))
5268 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5270 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5272 if (sd_degenerate(parent
))
5275 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5278 /* Flags needing groups don't count if only 1 group in parent */
5279 if (parent
->groups
== parent
->groups
->next
) {
5280 pflags
&= ~(SD_LOAD_BALANCE
|
5281 SD_BALANCE_NEWIDLE
|
5285 SD_SHARE_PKG_RESOURCES
|
5287 if (nr_node_ids
== 1)
5288 pflags
&= ~SD_SERIALIZE
;
5290 if (~cflags
& pflags
)
5296 static void free_rootdomain(struct rcu_head
*rcu
)
5298 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5300 cpupri_cleanup(&rd
->cpupri
);
5301 cpudl_cleanup(&rd
->cpudl
);
5302 free_cpumask_var(rd
->dlo_mask
);
5303 free_cpumask_var(rd
->rto_mask
);
5304 free_cpumask_var(rd
->online
);
5305 free_cpumask_var(rd
->span
);
5309 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5311 struct root_domain
*old_rd
= NULL
;
5312 unsigned long flags
;
5314 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5319 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5322 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5325 * If we dont want to free the old_rd yet then
5326 * set old_rd to NULL to skip the freeing later
5329 if (!atomic_dec_and_test(&old_rd
->refcount
))
5333 atomic_inc(&rd
->refcount
);
5336 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5337 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5340 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5343 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5346 static int init_rootdomain(struct root_domain
*rd
)
5348 memset(rd
, 0, sizeof(*rd
));
5350 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5352 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5354 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5356 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5359 init_dl_bw(&rd
->dl_bw
);
5360 if (cpudl_init(&rd
->cpudl
) != 0)
5363 if (cpupri_init(&rd
->cpupri
) != 0)
5368 free_cpumask_var(rd
->rto_mask
);
5370 free_cpumask_var(rd
->dlo_mask
);
5372 free_cpumask_var(rd
->online
);
5374 free_cpumask_var(rd
->span
);
5380 * By default the system creates a single root-domain with all cpus as
5381 * members (mimicking the global state we have today).
5383 struct root_domain def_root_domain
;
5385 static void init_defrootdomain(void)
5387 init_rootdomain(&def_root_domain
);
5389 atomic_set(&def_root_domain
.refcount
, 1);
5392 static struct root_domain
*alloc_rootdomain(void)
5394 struct root_domain
*rd
;
5396 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5400 if (init_rootdomain(rd
) != 0) {
5408 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5410 struct sched_group
*tmp
, *first
;
5419 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5424 } while (sg
!= first
);
5427 static void free_sched_domain(struct rcu_head
*rcu
)
5429 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5432 * If its an overlapping domain it has private groups, iterate and
5435 if (sd
->flags
& SD_OVERLAP
) {
5436 free_sched_groups(sd
->groups
, 1);
5437 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5438 kfree(sd
->groups
->sgp
);
5444 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5446 call_rcu(&sd
->rcu
, free_sched_domain
);
5449 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5451 for (; sd
; sd
= sd
->parent
)
5452 destroy_sched_domain(sd
, cpu
);
5456 * Keep a special pointer to the highest sched_domain that has
5457 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5458 * allows us to avoid some pointer chasing select_idle_sibling().
5460 * Also keep a unique ID per domain (we use the first cpu number in
5461 * the cpumask of the domain), this allows us to quickly tell if
5462 * two cpus are in the same cache domain, see cpus_share_cache().
5464 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5465 DEFINE_PER_CPU(int, sd_llc_size
);
5466 DEFINE_PER_CPU(int, sd_llc_id
);
5467 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5468 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5469 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5471 static void update_top_cache_domain(int cpu
)
5473 struct sched_domain
*sd
;
5474 struct sched_domain
*busy_sd
= NULL
;
5478 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5480 id
= cpumask_first(sched_domain_span(sd
));
5481 size
= cpumask_weight(sched_domain_span(sd
));
5482 busy_sd
= sd
->parent
; /* sd_busy */
5484 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5486 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5487 per_cpu(sd_llc_size
, cpu
) = size
;
5488 per_cpu(sd_llc_id
, cpu
) = id
;
5490 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5491 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5493 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5494 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5498 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5499 * hold the hotplug lock.
5502 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5504 struct rq
*rq
= cpu_rq(cpu
);
5505 struct sched_domain
*tmp
;
5507 /* Remove the sched domains which do not contribute to scheduling. */
5508 for (tmp
= sd
; tmp
; ) {
5509 struct sched_domain
*parent
= tmp
->parent
;
5513 if (sd_parent_degenerate(tmp
, parent
)) {
5514 tmp
->parent
= parent
->parent
;
5516 parent
->parent
->child
= tmp
;
5518 * Transfer SD_PREFER_SIBLING down in case of a
5519 * degenerate parent; the spans match for this
5520 * so the property transfers.
5522 if (parent
->flags
& SD_PREFER_SIBLING
)
5523 tmp
->flags
|= SD_PREFER_SIBLING
;
5524 destroy_sched_domain(parent
, cpu
);
5529 if (sd
&& sd_degenerate(sd
)) {
5532 destroy_sched_domain(tmp
, cpu
);
5537 sched_domain_debug(sd
, cpu
);
5539 rq_attach_root(rq
, rd
);
5541 rcu_assign_pointer(rq
->sd
, sd
);
5542 destroy_sched_domains(tmp
, cpu
);
5544 update_top_cache_domain(cpu
);
5547 /* cpus with isolated domains */
5548 static cpumask_var_t cpu_isolated_map
;
5550 /* Setup the mask of cpus configured for isolated domains */
5551 static int __init
isolated_cpu_setup(char *str
)
5553 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5554 cpulist_parse(str
, cpu_isolated_map
);
5558 __setup("isolcpus=", isolated_cpu_setup
);
5560 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5562 return cpumask_of_node(cpu_to_node(cpu
));
5566 struct sched_domain
**__percpu sd
;
5567 struct sched_group
**__percpu sg
;
5568 struct sched_group_power
**__percpu sgp
;
5572 struct sched_domain
** __percpu sd
;
5573 struct root_domain
*rd
;
5583 struct sched_domain_topology_level
;
5585 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5586 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5588 #define SDTL_OVERLAP 0x01
5590 struct sched_domain_topology_level
{
5591 sched_domain_init_f init
;
5592 sched_domain_mask_f mask
;
5595 struct sd_data data
;
5599 * Build an iteration mask that can exclude certain CPUs from the upwards
5602 * Asymmetric node setups can result in situations where the domain tree is of
5603 * unequal depth, make sure to skip domains that already cover the entire
5606 * In that case build_sched_domains() will have terminated the iteration early
5607 * and our sibling sd spans will be empty. Domains should always include the
5608 * cpu they're built on, so check that.
5611 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5613 const struct cpumask
*span
= sched_domain_span(sd
);
5614 struct sd_data
*sdd
= sd
->private;
5615 struct sched_domain
*sibling
;
5618 for_each_cpu(i
, span
) {
5619 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5620 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5623 cpumask_set_cpu(i
, sched_group_mask(sg
));
5628 * Return the canonical balance cpu for this group, this is the first cpu
5629 * of this group that's also in the iteration mask.
5631 int group_balance_cpu(struct sched_group
*sg
)
5633 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5637 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5639 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5640 const struct cpumask
*span
= sched_domain_span(sd
);
5641 struct cpumask
*covered
= sched_domains_tmpmask
;
5642 struct sd_data
*sdd
= sd
->private;
5643 struct sched_domain
*child
;
5646 cpumask_clear(covered
);
5648 for_each_cpu(i
, span
) {
5649 struct cpumask
*sg_span
;
5651 if (cpumask_test_cpu(i
, covered
))
5654 child
= *per_cpu_ptr(sdd
->sd
, i
);
5656 /* See the comment near build_group_mask(). */
5657 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5660 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5661 GFP_KERNEL
, cpu_to_node(cpu
));
5666 sg_span
= sched_group_cpus(sg
);
5668 child
= child
->child
;
5669 cpumask_copy(sg_span
, sched_domain_span(child
));
5671 cpumask_set_cpu(i
, sg_span
);
5673 cpumask_or(covered
, covered
, sg_span
);
5675 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5676 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5677 build_group_mask(sd
, sg
);
5680 * Initialize sgp->power such that even if we mess up the
5681 * domains and no possible iteration will get us here, we won't
5684 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5685 sg
->sgp
->power_orig
= sg
->sgp
->power
;
5688 * Make sure the first group of this domain contains the
5689 * canonical balance cpu. Otherwise the sched_domain iteration
5690 * breaks. See update_sg_lb_stats().
5692 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5693 group_balance_cpu(sg
) == cpu
)
5703 sd
->groups
= groups
;
5708 free_sched_groups(first
, 0);
5713 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5715 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5716 struct sched_domain
*child
= sd
->child
;
5719 cpu
= cpumask_first(sched_domain_span(child
));
5722 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5723 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5724 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5731 * build_sched_groups will build a circular linked list of the groups
5732 * covered by the given span, and will set each group's ->cpumask correctly,
5733 * and ->cpu_power to 0.
5735 * Assumes the sched_domain tree is fully constructed
5738 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5740 struct sched_group
*first
= NULL
, *last
= NULL
;
5741 struct sd_data
*sdd
= sd
->private;
5742 const struct cpumask
*span
= sched_domain_span(sd
);
5743 struct cpumask
*covered
;
5746 get_group(cpu
, sdd
, &sd
->groups
);
5747 atomic_inc(&sd
->groups
->ref
);
5749 if (cpu
!= cpumask_first(span
))
5752 lockdep_assert_held(&sched_domains_mutex
);
5753 covered
= sched_domains_tmpmask
;
5755 cpumask_clear(covered
);
5757 for_each_cpu(i
, span
) {
5758 struct sched_group
*sg
;
5761 if (cpumask_test_cpu(i
, covered
))
5764 group
= get_group(i
, sdd
, &sg
);
5765 cpumask_clear(sched_group_cpus(sg
));
5767 cpumask_setall(sched_group_mask(sg
));
5769 for_each_cpu(j
, span
) {
5770 if (get_group(j
, sdd
, NULL
) != group
)
5773 cpumask_set_cpu(j
, covered
);
5774 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5789 * Initialize sched groups cpu_power.
5791 * cpu_power indicates the capacity of sched group, which is used while
5792 * distributing the load between different sched groups in a sched domain.
5793 * Typically cpu_power for all the groups in a sched domain will be same unless
5794 * there are asymmetries in the topology. If there are asymmetries, group
5795 * having more cpu_power will pickup more load compared to the group having
5798 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5800 struct sched_group
*sg
= sd
->groups
;
5805 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5807 } while (sg
!= sd
->groups
);
5809 if (cpu
!= group_balance_cpu(sg
))
5812 update_group_power(sd
, cpu
);
5813 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5816 int __weak
arch_sd_sibling_asym_packing(void)
5818 return 0*SD_ASYM_PACKING
;
5822 * Initializers for schedule domains
5823 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5826 #ifdef CONFIG_SCHED_DEBUG
5827 # define SD_INIT_NAME(sd, type) sd->name = #type
5829 # define SD_INIT_NAME(sd, type) do { } while (0)
5832 #define SD_INIT_FUNC(type) \
5833 static noinline struct sched_domain * \
5834 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5836 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5837 *sd = SD_##type##_INIT; \
5838 SD_INIT_NAME(sd, type); \
5839 sd->private = &tl->data; \
5844 #ifdef CONFIG_SCHED_SMT
5845 SD_INIT_FUNC(SIBLING
)
5847 #ifdef CONFIG_SCHED_MC
5850 #ifdef CONFIG_SCHED_BOOK
5854 static int default_relax_domain_level
= -1;
5855 int sched_domain_level_max
;
5857 static int __init
setup_relax_domain_level(char *str
)
5859 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5860 pr_warn("Unable to set relax_domain_level\n");
5864 __setup("relax_domain_level=", setup_relax_domain_level
);
5866 static void set_domain_attribute(struct sched_domain
*sd
,
5867 struct sched_domain_attr
*attr
)
5871 if (!attr
|| attr
->relax_domain_level
< 0) {
5872 if (default_relax_domain_level
< 0)
5875 request
= default_relax_domain_level
;
5877 request
= attr
->relax_domain_level
;
5878 if (request
< sd
->level
) {
5879 /* turn off idle balance on this domain */
5880 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5882 /* turn on idle balance on this domain */
5883 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5887 static void __sdt_free(const struct cpumask
*cpu_map
);
5888 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5890 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5891 const struct cpumask
*cpu_map
)
5895 if (!atomic_read(&d
->rd
->refcount
))
5896 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5898 free_percpu(d
->sd
); /* fall through */
5900 __sdt_free(cpu_map
); /* fall through */
5906 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5907 const struct cpumask
*cpu_map
)
5909 memset(d
, 0, sizeof(*d
));
5911 if (__sdt_alloc(cpu_map
))
5912 return sa_sd_storage
;
5913 d
->sd
= alloc_percpu(struct sched_domain
*);
5915 return sa_sd_storage
;
5916 d
->rd
= alloc_rootdomain();
5919 return sa_rootdomain
;
5923 * NULL the sd_data elements we've used to build the sched_domain and
5924 * sched_group structure so that the subsequent __free_domain_allocs()
5925 * will not free the data we're using.
5927 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5929 struct sd_data
*sdd
= sd
->private;
5931 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5932 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5934 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5935 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5937 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5938 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5941 #ifdef CONFIG_SCHED_SMT
5942 static const struct cpumask
*cpu_smt_mask(int cpu
)
5944 return topology_thread_cpumask(cpu
);
5949 * Topology list, bottom-up.
5951 static struct sched_domain_topology_level default_topology
[] = {
5952 #ifdef CONFIG_SCHED_SMT
5953 { sd_init_SIBLING
, cpu_smt_mask
, },
5955 #ifdef CONFIG_SCHED_MC
5956 { sd_init_MC
, cpu_coregroup_mask
, },
5958 #ifdef CONFIG_SCHED_BOOK
5959 { sd_init_BOOK
, cpu_book_mask
, },
5961 { sd_init_CPU
, cpu_cpu_mask
, },
5965 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
5967 #define for_each_sd_topology(tl) \
5968 for (tl = sched_domain_topology; tl->init; tl++)
5972 static int sched_domains_numa_levels
;
5973 static int *sched_domains_numa_distance
;
5974 static struct cpumask
***sched_domains_numa_masks
;
5975 static int sched_domains_curr_level
;
5977 static inline int sd_local_flags(int level
)
5979 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
5982 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
5985 static struct sched_domain
*
5986 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
5988 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
5989 int level
= tl
->numa_level
;
5990 int sd_weight
= cpumask_weight(
5991 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
5993 *sd
= (struct sched_domain
){
5994 .min_interval
= sd_weight
,
5995 .max_interval
= 2*sd_weight
,
5997 .imbalance_pct
= 125,
5998 .cache_nice_tries
= 2,
6005 .flags
= 1*SD_LOAD_BALANCE
6006 | 1*SD_BALANCE_NEWIDLE
6011 | 0*SD_SHARE_CPUPOWER
6012 | 0*SD_SHARE_PKG_RESOURCES
6014 | 0*SD_PREFER_SIBLING
6016 | sd_local_flags(level
)
6018 .last_balance
= jiffies
,
6019 .balance_interval
= sd_weight
,
6021 SD_INIT_NAME(sd
, NUMA
);
6022 sd
->private = &tl
->data
;
6025 * Ugly hack to pass state to sd_numa_mask()...
6027 sched_domains_curr_level
= tl
->numa_level
;
6032 static const struct cpumask
*sd_numa_mask(int cpu
)
6034 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6037 static void sched_numa_warn(const char *str
)
6039 static int done
= false;
6047 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6049 for (i
= 0; i
< nr_node_ids
; i
++) {
6050 printk(KERN_WARNING
" ");
6051 for (j
= 0; j
< nr_node_ids
; j
++)
6052 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6053 printk(KERN_CONT
"\n");
6055 printk(KERN_WARNING
"\n");
6058 static bool find_numa_distance(int distance
)
6062 if (distance
== node_distance(0, 0))
6065 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6066 if (sched_domains_numa_distance
[i
] == distance
)
6073 static void sched_init_numa(void)
6075 int next_distance
, curr_distance
= node_distance(0, 0);
6076 struct sched_domain_topology_level
*tl
;
6080 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6081 if (!sched_domains_numa_distance
)
6085 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6086 * unique distances in the node_distance() table.
6088 * Assumes node_distance(0,j) includes all distances in
6089 * node_distance(i,j) in order to avoid cubic time.
6091 next_distance
= curr_distance
;
6092 for (i
= 0; i
< nr_node_ids
; i
++) {
6093 for (j
= 0; j
< nr_node_ids
; j
++) {
6094 for (k
= 0; k
< nr_node_ids
; k
++) {
6095 int distance
= node_distance(i
, k
);
6097 if (distance
> curr_distance
&&
6098 (distance
< next_distance
||
6099 next_distance
== curr_distance
))
6100 next_distance
= distance
;
6103 * While not a strong assumption it would be nice to know
6104 * about cases where if node A is connected to B, B is not
6105 * equally connected to A.
6107 if (sched_debug() && node_distance(k
, i
) != distance
)
6108 sched_numa_warn("Node-distance not symmetric");
6110 if (sched_debug() && i
&& !find_numa_distance(distance
))
6111 sched_numa_warn("Node-0 not representative");
6113 if (next_distance
!= curr_distance
) {
6114 sched_domains_numa_distance
[level
++] = next_distance
;
6115 sched_domains_numa_levels
= level
;
6116 curr_distance
= next_distance
;
6121 * In case of sched_debug() we verify the above assumption.
6127 * 'level' contains the number of unique distances, excluding the
6128 * identity distance node_distance(i,i).
6130 * The sched_domains_numa_distance[] array includes the actual distance
6135 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6136 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6137 * the array will contain less then 'level' members. This could be
6138 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6139 * in other functions.
6141 * We reset it to 'level' at the end of this function.
6143 sched_domains_numa_levels
= 0;
6145 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6146 if (!sched_domains_numa_masks
)
6150 * Now for each level, construct a mask per node which contains all
6151 * cpus of nodes that are that many hops away from us.
6153 for (i
= 0; i
< level
; i
++) {
6154 sched_domains_numa_masks
[i
] =
6155 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6156 if (!sched_domains_numa_masks
[i
])
6159 for (j
= 0; j
< nr_node_ids
; j
++) {
6160 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6164 sched_domains_numa_masks
[i
][j
] = mask
;
6166 for (k
= 0; k
< nr_node_ids
; k
++) {
6167 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6170 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6175 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6176 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6181 * Copy the default topology bits..
6183 for (i
= 0; default_topology
[i
].init
; i
++)
6184 tl
[i
] = default_topology
[i
];
6187 * .. and append 'j' levels of NUMA goodness.
6189 for (j
= 0; j
< level
; i
++, j
++) {
6190 tl
[i
] = (struct sched_domain_topology_level
){
6191 .init
= sd_numa_init
,
6192 .mask
= sd_numa_mask
,
6193 .flags
= SDTL_OVERLAP
,
6198 sched_domain_topology
= tl
;
6200 sched_domains_numa_levels
= level
;
6203 static void sched_domains_numa_masks_set(int cpu
)
6206 int node
= cpu_to_node(cpu
);
6208 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6209 for (j
= 0; j
< nr_node_ids
; j
++) {
6210 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6211 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6216 static void sched_domains_numa_masks_clear(int cpu
)
6219 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6220 for (j
= 0; j
< nr_node_ids
; j
++)
6221 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6226 * Update sched_domains_numa_masks[level][node] array when new cpus
6229 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6230 unsigned long action
,
6233 int cpu
= (long)hcpu
;
6235 switch (action
& ~CPU_TASKS_FROZEN
) {
6237 sched_domains_numa_masks_set(cpu
);
6241 sched_domains_numa_masks_clear(cpu
);
6251 static inline void sched_init_numa(void)
6255 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6256 unsigned long action
,
6261 #endif /* CONFIG_NUMA */
6263 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6265 struct sched_domain_topology_level
*tl
;
6268 for_each_sd_topology(tl
) {
6269 struct sd_data
*sdd
= &tl
->data
;
6271 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6275 sdd
->sg
= alloc_percpu(struct sched_group
*);
6279 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6283 for_each_cpu(j
, cpu_map
) {
6284 struct sched_domain
*sd
;
6285 struct sched_group
*sg
;
6286 struct sched_group_power
*sgp
;
6288 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6289 GFP_KERNEL
, cpu_to_node(j
));
6293 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6295 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6296 GFP_KERNEL
, cpu_to_node(j
));
6302 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6304 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6305 GFP_KERNEL
, cpu_to_node(j
));
6309 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6316 static void __sdt_free(const struct cpumask
*cpu_map
)
6318 struct sched_domain_topology_level
*tl
;
6321 for_each_sd_topology(tl
) {
6322 struct sd_data
*sdd
= &tl
->data
;
6324 for_each_cpu(j
, cpu_map
) {
6325 struct sched_domain
*sd
;
6328 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6329 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6330 free_sched_groups(sd
->groups
, 0);
6331 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6335 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6337 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6339 free_percpu(sdd
->sd
);
6341 free_percpu(sdd
->sg
);
6343 free_percpu(sdd
->sgp
);
6348 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6349 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6350 struct sched_domain
*child
, int cpu
)
6352 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6356 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6358 sd
->level
= child
->level
+ 1;
6359 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6363 set_domain_attribute(sd
, attr
);
6369 * Build sched domains for a given set of cpus and attach the sched domains
6370 * to the individual cpus
6372 static int build_sched_domains(const struct cpumask
*cpu_map
,
6373 struct sched_domain_attr
*attr
)
6375 enum s_alloc alloc_state
;
6376 struct sched_domain
*sd
;
6378 int i
, ret
= -ENOMEM
;
6380 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6381 if (alloc_state
!= sa_rootdomain
)
6384 /* Set up domains for cpus specified by the cpu_map. */
6385 for_each_cpu(i
, cpu_map
) {
6386 struct sched_domain_topology_level
*tl
;
6389 for_each_sd_topology(tl
) {
6390 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6391 if (tl
== sched_domain_topology
)
6392 *per_cpu_ptr(d
.sd
, i
) = sd
;
6393 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6394 sd
->flags
|= SD_OVERLAP
;
6395 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6400 /* Build the groups for the domains */
6401 for_each_cpu(i
, cpu_map
) {
6402 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6403 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6404 if (sd
->flags
& SD_OVERLAP
) {
6405 if (build_overlap_sched_groups(sd
, i
))
6408 if (build_sched_groups(sd
, i
))
6414 /* Calculate CPU power for physical packages and nodes */
6415 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6416 if (!cpumask_test_cpu(i
, cpu_map
))
6419 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6420 claim_allocations(i
, sd
);
6421 init_sched_groups_power(i
, sd
);
6425 /* Attach the domains */
6427 for_each_cpu(i
, cpu_map
) {
6428 sd
= *per_cpu_ptr(d
.sd
, i
);
6429 cpu_attach_domain(sd
, d
.rd
, i
);
6435 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6439 static cpumask_var_t
*doms_cur
; /* current sched domains */
6440 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6441 static struct sched_domain_attr
*dattr_cur
;
6442 /* attribues of custom domains in 'doms_cur' */
6445 * Special case: If a kmalloc of a doms_cur partition (array of
6446 * cpumask) fails, then fallback to a single sched domain,
6447 * as determined by the single cpumask fallback_doms.
6449 static cpumask_var_t fallback_doms
;
6452 * arch_update_cpu_topology lets virtualized architectures update the
6453 * cpu core maps. It is supposed to return 1 if the topology changed
6454 * or 0 if it stayed the same.
6456 int __attribute__((weak
)) arch_update_cpu_topology(void)
6461 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6464 cpumask_var_t
*doms
;
6466 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6469 for (i
= 0; i
< ndoms
; i
++) {
6470 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6471 free_sched_domains(doms
, i
);
6478 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6481 for (i
= 0; i
< ndoms
; i
++)
6482 free_cpumask_var(doms
[i
]);
6487 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6488 * For now this just excludes isolated cpus, but could be used to
6489 * exclude other special cases in the future.
6491 static int init_sched_domains(const struct cpumask
*cpu_map
)
6495 arch_update_cpu_topology();
6497 doms_cur
= alloc_sched_domains(ndoms_cur
);
6499 doms_cur
= &fallback_doms
;
6500 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6501 err
= build_sched_domains(doms_cur
[0], NULL
);
6502 register_sched_domain_sysctl();
6508 * Detach sched domains from a group of cpus specified in cpu_map
6509 * These cpus will now be attached to the NULL domain
6511 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6516 for_each_cpu(i
, cpu_map
)
6517 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6521 /* handle null as "default" */
6522 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6523 struct sched_domain_attr
*new, int idx_new
)
6525 struct sched_domain_attr tmp
;
6532 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6533 new ? (new + idx_new
) : &tmp
,
6534 sizeof(struct sched_domain_attr
));
6538 * Partition sched domains as specified by the 'ndoms_new'
6539 * cpumasks in the array doms_new[] of cpumasks. This compares
6540 * doms_new[] to the current sched domain partitioning, doms_cur[].
6541 * It destroys each deleted domain and builds each new domain.
6543 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6544 * The masks don't intersect (don't overlap.) We should setup one
6545 * sched domain for each mask. CPUs not in any of the cpumasks will
6546 * not be load balanced. If the same cpumask appears both in the
6547 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6550 * The passed in 'doms_new' should be allocated using
6551 * alloc_sched_domains. This routine takes ownership of it and will
6552 * free_sched_domains it when done with it. If the caller failed the
6553 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6554 * and partition_sched_domains() will fallback to the single partition
6555 * 'fallback_doms', it also forces the domains to be rebuilt.
6557 * If doms_new == NULL it will be replaced with cpu_online_mask.
6558 * ndoms_new == 0 is a special case for destroying existing domains,
6559 * and it will not create the default domain.
6561 * Call with hotplug lock held
6563 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6564 struct sched_domain_attr
*dattr_new
)
6569 mutex_lock(&sched_domains_mutex
);
6571 /* always unregister in case we don't destroy any domains */
6572 unregister_sched_domain_sysctl();
6574 /* Let architecture update cpu core mappings. */
6575 new_topology
= arch_update_cpu_topology();
6577 n
= doms_new
? ndoms_new
: 0;
6579 /* Destroy deleted domains */
6580 for (i
= 0; i
< ndoms_cur
; i
++) {
6581 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6582 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6583 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6586 /* no match - a current sched domain not in new doms_new[] */
6587 detach_destroy_domains(doms_cur
[i
]);
6593 if (doms_new
== NULL
) {
6595 doms_new
= &fallback_doms
;
6596 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6597 WARN_ON_ONCE(dattr_new
);
6600 /* Build new domains */
6601 for (i
= 0; i
< ndoms_new
; i
++) {
6602 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6603 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6604 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6607 /* no match - add a new doms_new */
6608 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6613 /* Remember the new sched domains */
6614 if (doms_cur
!= &fallback_doms
)
6615 free_sched_domains(doms_cur
, ndoms_cur
);
6616 kfree(dattr_cur
); /* kfree(NULL) is safe */
6617 doms_cur
= doms_new
;
6618 dattr_cur
= dattr_new
;
6619 ndoms_cur
= ndoms_new
;
6621 register_sched_domain_sysctl();
6623 mutex_unlock(&sched_domains_mutex
);
6626 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6629 * Update cpusets according to cpu_active mask. If cpusets are
6630 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6631 * around partition_sched_domains().
6633 * If we come here as part of a suspend/resume, don't touch cpusets because we
6634 * want to restore it back to its original state upon resume anyway.
6636 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6640 case CPU_ONLINE_FROZEN
:
6641 case CPU_DOWN_FAILED_FROZEN
:
6644 * num_cpus_frozen tracks how many CPUs are involved in suspend
6645 * resume sequence. As long as this is not the last online
6646 * operation in the resume sequence, just build a single sched
6647 * domain, ignoring cpusets.
6650 if (likely(num_cpus_frozen
)) {
6651 partition_sched_domains(1, NULL
, NULL
);
6656 * This is the last CPU online operation. So fall through and
6657 * restore the original sched domains by considering the
6658 * cpuset configurations.
6662 case CPU_DOWN_FAILED
:
6663 cpuset_update_active_cpus(true);
6671 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6675 case CPU_DOWN_PREPARE
:
6676 cpuset_update_active_cpus(false);
6678 case CPU_DOWN_PREPARE_FROZEN
:
6680 partition_sched_domains(1, NULL
, NULL
);
6688 void __init
sched_init_smp(void)
6690 cpumask_var_t non_isolated_cpus
;
6692 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6693 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6698 * There's no userspace yet to cause hotplug operations; hence all the
6699 * cpu masks are stable and all blatant races in the below code cannot
6702 mutex_lock(&sched_domains_mutex
);
6703 init_sched_domains(cpu_active_mask
);
6704 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6705 if (cpumask_empty(non_isolated_cpus
))
6706 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6707 mutex_unlock(&sched_domains_mutex
);
6709 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6710 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6711 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6715 /* Move init over to a non-isolated CPU */
6716 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6718 sched_init_granularity();
6719 free_cpumask_var(non_isolated_cpus
);
6721 init_sched_rt_class();
6722 init_sched_dl_class();
6725 void __init
sched_init_smp(void)
6727 sched_init_granularity();
6729 #endif /* CONFIG_SMP */
6731 const_debug
unsigned int sysctl_timer_migration
= 1;
6733 int in_sched_functions(unsigned long addr
)
6735 return in_lock_functions(addr
) ||
6736 (addr
>= (unsigned long)__sched_text_start
6737 && addr
< (unsigned long)__sched_text_end
);
6740 #ifdef CONFIG_CGROUP_SCHED
6742 * Default task group.
6743 * Every task in system belongs to this group at bootup.
6745 struct task_group root_task_group
;
6746 LIST_HEAD(task_groups
);
6749 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6751 void __init
sched_init(void)
6754 unsigned long alloc_size
= 0, ptr
;
6756 #ifdef CONFIG_FAIR_GROUP_SCHED
6757 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6759 #ifdef CONFIG_RT_GROUP_SCHED
6760 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6762 #ifdef CONFIG_CPUMASK_OFFSTACK
6763 alloc_size
+= num_possible_cpus() * cpumask_size();
6766 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6768 #ifdef CONFIG_FAIR_GROUP_SCHED
6769 root_task_group
.se
= (struct sched_entity
**)ptr
;
6770 ptr
+= nr_cpu_ids
* sizeof(void **);
6772 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6773 ptr
+= nr_cpu_ids
* sizeof(void **);
6775 #endif /* CONFIG_FAIR_GROUP_SCHED */
6776 #ifdef CONFIG_RT_GROUP_SCHED
6777 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6778 ptr
+= nr_cpu_ids
* sizeof(void **);
6780 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6781 ptr
+= nr_cpu_ids
* sizeof(void **);
6783 #endif /* CONFIG_RT_GROUP_SCHED */
6784 #ifdef CONFIG_CPUMASK_OFFSTACK
6785 for_each_possible_cpu(i
) {
6786 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6787 ptr
+= cpumask_size();
6789 #endif /* CONFIG_CPUMASK_OFFSTACK */
6792 init_rt_bandwidth(&def_rt_bandwidth
,
6793 global_rt_period(), global_rt_runtime());
6794 init_dl_bandwidth(&def_dl_bandwidth
,
6795 global_rt_period(), global_rt_runtime());
6798 init_defrootdomain();
6801 #ifdef CONFIG_RT_GROUP_SCHED
6802 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6803 global_rt_period(), global_rt_runtime());
6804 #endif /* CONFIG_RT_GROUP_SCHED */
6806 #ifdef CONFIG_CGROUP_SCHED
6807 list_add(&root_task_group
.list
, &task_groups
);
6808 INIT_LIST_HEAD(&root_task_group
.children
);
6809 INIT_LIST_HEAD(&root_task_group
.siblings
);
6810 autogroup_init(&init_task
);
6812 #endif /* CONFIG_CGROUP_SCHED */
6814 for_each_possible_cpu(i
) {
6818 raw_spin_lock_init(&rq
->lock
);
6820 rq
->calc_load_active
= 0;
6821 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6822 init_cfs_rq(&rq
->cfs
);
6823 init_rt_rq(&rq
->rt
, rq
);
6824 init_dl_rq(&rq
->dl
, rq
);
6825 #ifdef CONFIG_FAIR_GROUP_SCHED
6826 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6827 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6829 * How much cpu bandwidth does root_task_group get?
6831 * In case of task-groups formed thr' the cgroup filesystem, it
6832 * gets 100% of the cpu resources in the system. This overall
6833 * system cpu resource is divided among the tasks of
6834 * root_task_group and its child task-groups in a fair manner,
6835 * based on each entity's (task or task-group's) weight
6836 * (se->load.weight).
6838 * In other words, if root_task_group has 10 tasks of weight
6839 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6840 * then A0's share of the cpu resource is:
6842 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6844 * We achieve this by letting root_task_group's tasks sit
6845 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6847 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6848 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6849 #endif /* CONFIG_FAIR_GROUP_SCHED */
6851 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6852 #ifdef CONFIG_RT_GROUP_SCHED
6853 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
6854 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6857 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6858 rq
->cpu_load
[j
] = 0;
6860 rq
->last_load_update_tick
= jiffies
;
6865 rq
->cpu_power
= SCHED_POWER_SCALE
;
6866 rq
->post_schedule
= 0;
6867 rq
->active_balance
= 0;
6868 rq
->next_balance
= jiffies
;
6873 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6874 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6876 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6878 rq_attach_root(rq
, &def_root_domain
);
6879 #ifdef CONFIG_NO_HZ_COMMON
6882 #ifdef CONFIG_NO_HZ_FULL
6883 rq
->last_sched_tick
= 0;
6887 atomic_set(&rq
->nr_iowait
, 0);
6890 set_load_weight(&init_task
);
6892 #ifdef CONFIG_PREEMPT_NOTIFIERS
6893 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6897 * The boot idle thread does lazy MMU switching as well:
6899 atomic_inc(&init_mm
.mm_count
);
6900 enter_lazy_tlb(&init_mm
, current
);
6903 * Make us the idle thread. Technically, schedule() should not be
6904 * called from this thread, however somewhere below it might be,
6905 * but because we are the idle thread, we just pick up running again
6906 * when this runqueue becomes "idle".
6908 init_idle(current
, smp_processor_id());
6910 calc_load_update
= jiffies
+ LOAD_FREQ
;
6913 * During early bootup we pretend to be a normal task:
6915 current
->sched_class
= &fair_sched_class
;
6918 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6919 /* May be allocated at isolcpus cmdline parse time */
6920 if (cpu_isolated_map
== NULL
)
6921 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6922 idle_thread_set_boot_cpu();
6924 init_sched_fair_class();
6926 scheduler_running
= 1;
6929 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6930 static inline int preempt_count_equals(int preempt_offset
)
6932 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6934 return (nested
== preempt_offset
);
6937 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6939 static unsigned long prev_jiffy
; /* ratelimiting */
6941 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6942 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
6943 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6945 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6947 prev_jiffy
= jiffies
;
6950 "BUG: sleeping function called from invalid context at %s:%d\n",
6953 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6954 in_atomic(), irqs_disabled(),
6955 current
->pid
, current
->comm
);
6957 debug_show_held_locks(current
);
6958 if (irqs_disabled())
6959 print_irqtrace_events(current
);
6962 EXPORT_SYMBOL(__might_sleep
);
6965 #ifdef CONFIG_MAGIC_SYSRQ
6966 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
6968 const struct sched_class
*prev_class
= p
->sched_class
;
6969 struct sched_attr attr
= {
6970 .sched_policy
= SCHED_NORMAL
,
6972 int old_prio
= p
->prio
;
6977 dequeue_task(rq
, p
, 0);
6978 __setscheduler(rq
, p
, &attr
);
6980 enqueue_task(rq
, p
, 0);
6981 resched_task(rq
->curr
);
6984 check_class_changed(rq
, p
, prev_class
, old_prio
);
6987 void normalize_rt_tasks(void)
6989 struct task_struct
*g
, *p
;
6990 unsigned long flags
;
6993 read_lock_irqsave(&tasklist_lock
, flags
);
6994 do_each_thread(g
, p
) {
6996 * Only normalize user tasks:
7001 p
->se
.exec_start
= 0;
7002 #ifdef CONFIG_SCHEDSTATS
7003 p
->se
.statistics
.wait_start
= 0;
7004 p
->se
.statistics
.sleep_start
= 0;
7005 p
->se
.statistics
.block_start
= 0;
7008 if (!dl_task(p
) && !rt_task(p
)) {
7010 * Renice negative nice level userspace
7013 if (TASK_NICE(p
) < 0 && p
->mm
)
7014 set_user_nice(p
, 0);
7018 raw_spin_lock(&p
->pi_lock
);
7019 rq
= __task_rq_lock(p
);
7021 normalize_task(rq
, p
);
7023 __task_rq_unlock(rq
);
7024 raw_spin_unlock(&p
->pi_lock
);
7025 } while_each_thread(g
, p
);
7027 read_unlock_irqrestore(&tasklist_lock
, flags
);
7030 #endif /* CONFIG_MAGIC_SYSRQ */
7032 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7034 * These functions are only useful for the IA64 MCA handling, or kdb.
7036 * They can only be called when the whole system has been
7037 * stopped - every CPU needs to be quiescent, and no scheduling
7038 * activity can take place. Using them for anything else would
7039 * be a serious bug, and as a result, they aren't even visible
7040 * under any other configuration.
7044 * curr_task - return the current task for a given cpu.
7045 * @cpu: the processor in question.
7047 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7049 * Return: The current task for @cpu.
7051 struct task_struct
*curr_task(int cpu
)
7053 return cpu_curr(cpu
);
7056 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7060 * set_curr_task - set the current task for a given cpu.
7061 * @cpu: the processor in question.
7062 * @p: the task pointer to set.
7064 * Description: This function must only be used when non-maskable interrupts
7065 * are serviced on a separate stack. It allows the architecture to switch the
7066 * notion of the current task on a cpu in a non-blocking manner. This function
7067 * must be called with all CPU's synchronized, and interrupts disabled, the
7068 * and caller must save the original value of the current task (see
7069 * curr_task() above) and restore that value before reenabling interrupts and
7070 * re-starting the system.
7072 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7074 void set_curr_task(int cpu
, struct task_struct
*p
)
7081 #ifdef CONFIG_CGROUP_SCHED
7082 /* task_group_lock serializes the addition/removal of task groups */
7083 static DEFINE_SPINLOCK(task_group_lock
);
7085 static void free_sched_group(struct task_group
*tg
)
7087 free_fair_sched_group(tg
);
7088 free_rt_sched_group(tg
);
7093 /* allocate runqueue etc for a new task group */
7094 struct task_group
*sched_create_group(struct task_group
*parent
)
7096 struct task_group
*tg
;
7098 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7100 return ERR_PTR(-ENOMEM
);
7102 if (!alloc_fair_sched_group(tg
, parent
))
7105 if (!alloc_rt_sched_group(tg
, parent
))
7111 free_sched_group(tg
);
7112 return ERR_PTR(-ENOMEM
);
7115 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7117 unsigned long flags
;
7119 spin_lock_irqsave(&task_group_lock
, flags
);
7120 list_add_rcu(&tg
->list
, &task_groups
);
7122 WARN_ON(!parent
); /* root should already exist */
7124 tg
->parent
= parent
;
7125 INIT_LIST_HEAD(&tg
->children
);
7126 list_add_rcu(&tg
->siblings
, &parent
->children
);
7127 spin_unlock_irqrestore(&task_group_lock
, flags
);
7130 /* rcu callback to free various structures associated with a task group */
7131 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7133 /* now it should be safe to free those cfs_rqs */
7134 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7137 /* Destroy runqueue etc associated with a task group */
7138 void sched_destroy_group(struct task_group
*tg
)
7140 /* wait for possible concurrent references to cfs_rqs complete */
7141 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7144 void sched_offline_group(struct task_group
*tg
)
7146 unsigned long flags
;
7149 /* end participation in shares distribution */
7150 for_each_possible_cpu(i
)
7151 unregister_fair_sched_group(tg
, i
);
7153 spin_lock_irqsave(&task_group_lock
, flags
);
7154 list_del_rcu(&tg
->list
);
7155 list_del_rcu(&tg
->siblings
);
7156 spin_unlock_irqrestore(&task_group_lock
, flags
);
7159 /* change task's runqueue when it moves between groups.
7160 * The caller of this function should have put the task in its new group
7161 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7162 * reflect its new group.
7164 void sched_move_task(struct task_struct
*tsk
)
7166 struct task_group
*tg
;
7168 unsigned long flags
;
7171 rq
= task_rq_lock(tsk
, &flags
);
7173 running
= task_current(rq
, tsk
);
7177 dequeue_task(rq
, tsk
, 0);
7178 if (unlikely(running
))
7179 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7181 tg
= container_of(task_css_check(tsk
, cpu_cgroup_subsys_id
,
7182 lockdep_is_held(&tsk
->sighand
->siglock
)),
7183 struct task_group
, css
);
7184 tg
= autogroup_task_group(tsk
, tg
);
7185 tsk
->sched_task_group
= tg
;
7187 #ifdef CONFIG_FAIR_GROUP_SCHED
7188 if (tsk
->sched_class
->task_move_group
)
7189 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7192 set_task_rq(tsk
, task_cpu(tsk
));
7194 if (unlikely(running
))
7195 tsk
->sched_class
->set_curr_task(rq
);
7197 enqueue_task(rq
, tsk
, 0);
7199 task_rq_unlock(rq
, tsk
, &flags
);
7201 #endif /* CONFIG_CGROUP_SCHED */
7203 #ifdef CONFIG_RT_GROUP_SCHED
7205 * Ensure that the real time constraints are schedulable.
7207 static DEFINE_MUTEX(rt_constraints_mutex
);
7209 /* Must be called with tasklist_lock held */
7210 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7212 struct task_struct
*g
, *p
;
7214 do_each_thread(g
, p
) {
7215 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7217 } while_each_thread(g
, p
);
7222 struct rt_schedulable_data
{
7223 struct task_group
*tg
;
7228 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7230 struct rt_schedulable_data
*d
= data
;
7231 struct task_group
*child
;
7232 unsigned long total
, sum
= 0;
7233 u64 period
, runtime
;
7235 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7236 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7239 period
= d
->rt_period
;
7240 runtime
= d
->rt_runtime
;
7244 * Cannot have more runtime than the period.
7246 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7250 * Ensure we don't starve existing RT tasks.
7252 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7255 total
= to_ratio(period
, runtime
);
7258 * Nobody can have more than the global setting allows.
7260 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7264 * The sum of our children's runtime should not exceed our own.
7266 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7267 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7268 runtime
= child
->rt_bandwidth
.rt_runtime
;
7270 if (child
== d
->tg
) {
7271 period
= d
->rt_period
;
7272 runtime
= d
->rt_runtime
;
7275 sum
+= to_ratio(period
, runtime
);
7284 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7288 struct rt_schedulable_data data
= {
7290 .rt_period
= period
,
7291 .rt_runtime
= runtime
,
7295 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7301 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7302 u64 rt_period
, u64 rt_runtime
)
7306 mutex_lock(&rt_constraints_mutex
);
7307 read_lock(&tasklist_lock
);
7308 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7312 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7313 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7314 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7316 for_each_possible_cpu(i
) {
7317 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7319 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7320 rt_rq
->rt_runtime
= rt_runtime
;
7321 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7323 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7325 read_unlock(&tasklist_lock
);
7326 mutex_unlock(&rt_constraints_mutex
);
7331 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7333 u64 rt_runtime
, rt_period
;
7335 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7336 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7337 if (rt_runtime_us
< 0)
7338 rt_runtime
= RUNTIME_INF
;
7340 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7343 static long sched_group_rt_runtime(struct task_group
*tg
)
7347 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7350 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7351 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7352 return rt_runtime_us
;
7355 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7357 u64 rt_runtime
, rt_period
;
7359 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7360 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7365 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7368 static long sched_group_rt_period(struct task_group
*tg
)
7372 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7373 do_div(rt_period_us
, NSEC_PER_USEC
);
7374 return rt_period_us
;
7376 #endif /* CONFIG_RT_GROUP_SCHED */
7378 #ifdef CONFIG_RT_GROUP_SCHED
7379 static int sched_rt_global_constraints(void)
7383 mutex_lock(&rt_constraints_mutex
);
7384 read_lock(&tasklist_lock
);
7385 ret
= __rt_schedulable(NULL
, 0, 0);
7386 read_unlock(&tasklist_lock
);
7387 mutex_unlock(&rt_constraints_mutex
);
7392 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7394 /* Don't accept realtime tasks when there is no way for them to run */
7395 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7401 #else /* !CONFIG_RT_GROUP_SCHED */
7402 static int sched_rt_global_constraints(void)
7404 unsigned long flags
;
7407 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7408 for_each_possible_cpu(i
) {
7409 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7411 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7412 rt_rq
->rt_runtime
= global_rt_runtime();
7413 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7415 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7419 #endif /* CONFIG_RT_GROUP_SCHED */
7421 static int sched_dl_global_constraints(void)
7423 u64 runtime
= global_rt_runtime();
7424 u64 period
= global_rt_period();
7425 u64 new_bw
= to_ratio(period
, runtime
);
7429 * Here we want to check the bandwidth not being set to some
7430 * value smaller than the currently allocated bandwidth in
7431 * any of the root_domains.
7433 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7434 * cycling on root_domains... Discussion on different/better
7435 * solutions is welcome!
7437 for_each_possible_cpu(cpu
) {
7438 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7440 raw_spin_lock(&dl_b
->lock
);
7441 if (new_bw
< dl_b
->total_bw
)
7443 raw_spin_unlock(&dl_b
->lock
);
7452 static void sched_dl_do_global(void)
7457 def_dl_bandwidth
.dl_period
= global_rt_period();
7458 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7460 if (global_rt_runtime() != RUNTIME_INF
)
7461 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7464 * FIXME: As above...
7466 for_each_possible_cpu(cpu
) {
7467 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7469 raw_spin_lock(&dl_b
->lock
);
7471 raw_spin_unlock(&dl_b
->lock
);
7475 static int sched_rt_global_validate(void)
7477 if (sysctl_sched_rt_period
<= 0)
7480 if (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
)
7486 static void sched_rt_do_global(void)
7488 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7489 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7492 int sched_rt_handler(struct ctl_table
*table
, int write
,
7493 void __user
*buffer
, size_t *lenp
,
7496 int old_period
, old_runtime
;
7497 static DEFINE_MUTEX(mutex
);
7501 old_period
= sysctl_sched_rt_period
;
7502 old_runtime
= sysctl_sched_rt_runtime
;
7504 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7506 if (!ret
&& write
) {
7507 ret
= sched_rt_global_validate();
7511 ret
= sched_rt_global_constraints();
7515 ret
= sched_dl_global_constraints();
7519 sched_rt_do_global();
7520 sched_dl_do_global();
7524 sysctl_sched_rt_period
= old_period
;
7525 sysctl_sched_rt_runtime
= old_runtime
;
7527 mutex_unlock(&mutex
);
7532 int sched_rr_handler(struct ctl_table
*table
, int write
,
7533 void __user
*buffer
, size_t *lenp
,
7537 static DEFINE_MUTEX(mutex
);
7540 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7541 /* make sure that internally we keep jiffies */
7542 /* also, writing zero resets timeslice to default */
7543 if (!ret
&& write
) {
7544 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7545 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7547 mutex_unlock(&mutex
);
7551 #ifdef CONFIG_CGROUP_SCHED
7553 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7555 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7558 static struct cgroup_subsys_state
*
7559 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7561 struct task_group
*parent
= css_tg(parent_css
);
7562 struct task_group
*tg
;
7565 /* This is early initialization for the top cgroup */
7566 return &root_task_group
.css
;
7569 tg
= sched_create_group(parent
);
7571 return ERR_PTR(-ENOMEM
);
7576 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7578 struct task_group
*tg
= css_tg(css
);
7579 struct task_group
*parent
= css_tg(css_parent(css
));
7582 sched_online_group(tg
, parent
);
7586 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7588 struct task_group
*tg
= css_tg(css
);
7590 sched_destroy_group(tg
);
7593 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7595 struct task_group
*tg
= css_tg(css
);
7597 sched_offline_group(tg
);
7600 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7601 struct cgroup_taskset
*tset
)
7603 struct task_struct
*task
;
7605 cgroup_taskset_for_each(task
, css
, tset
) {
7606 #ifdef CONFIG_RT_GROUP_SCHED
7607 if (!sched_rt_can_attach(css_tg(css
), task
))
7610 /* We don't support RT-tasks being in separate groups */
7611 if (task
->sched_class
!= &fair_sched_class
)
7618 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7619 struct cgroup_taskset
*tset
)
7621 struct task_struct
*task
;
7623 cgroup_taskset_for_each(task
, css
, tset
)
7624 sched_move_task(task
);
7627 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7628 struct cgroup_subsys_state
*old_css
,
7629 struct task_struct
*task
)
7632 * cgroup_exit() is called in the copy_process() failure path.
7633 * Ignore this case since the task hasn't ran yet, this avoids
7634 * trying to poke a half freed task state from generic code.
7636 if (!(task
->flags
& PF_EXITING
))
7639 sched_move_task(task
);
7642 #ifdef CONFIG_FAIR_GROUP_SCHED
7643 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7644 struct cftype
*cftype
, u64 shareval
)
7646 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7649 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7652 struct task_group
*tg
= css_tg(css
);
7654 return (u64
) scale_load_down(tg
->shares
);
7657 #ifdef CONFIG_CFS_BANDWIDTH
7658 static DEFINE_MUTEX(cfs_constraints_mutex
);
7660 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7661 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7663 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7665 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7667 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7668 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7670 if (tg
== &root_task_group
)
7674 * Ensure we have at some amount of bandwidth every period. This is
7675 * to prevent reaching a state of large arrears when throttled via
7676 * entity_tick() resulting in prolonged exit starvation.
7678 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7682 * Likewise, bound things on the otherside by preventing insane quota
7683 * periods. This also allows us to normalize in computing quota
7686 if (period
> max_cfs_quota_period
)
7689 mutex_lock(&cfs_constraints_mutex
);
7690 ret
= __cfs_schedulable(tg
, period
, quota
);
7694 runtime_enabled
= quota
!= RUNTIME_INF
;
7695 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7697 * If we need to toggle cfs_bandwidth_used, off->on must occur
7698 * before making related changes, and on->off must occur afterwards
7700 if (runtime_enabled
&& !runtime_was_enabled
)
7701 cfs_bandwidth_usage_inc();
7702 raw_spin_lock_irq(&cfs_b
->lock
);
7703 cfs_b
->period
= ns_to_ktime(period
);
7704 cfs_b
->quota
= quota
;
7706 __refill_cfs_bandwidth_runtime(cfs_b
);
7707 /* restart the period timer (if active) to handle new period expiry */
7708 if (runtime_enabled
&& cfs_b
->timer_active
) {
7709 /* force a reprogram */
7710 cfs_b
->timer_active
= 0;
7711 __start_cfs_bandwidth(cfs_b
);
7713 raw_spin_unlock_irq(&cfs_b
->lock
);
7715 for_each_possible_cpu(i
) {
7716 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7717 struct rq
*rq
= cfs_rq
->rq
;
7719 raw_spin_lock_irq(&rq
->lock
);
7720 cfs_rq
->runtime_enabled
= runtime_enabled
;
7721 cfs_rq
->runtime_remaining
= 0;
7723 if (cfs_rq
->throttled
)
7724 unthrottle_cfs_rq(cfs_rq
);
7725 raw_spin_unlock_irq(&rq
->lock
);
7727 if (runtime_was_enabled
&& !runtime_enabled
)
7728 cfs_bandwidth_usage_dec();
7730 mutex_unlock(&cfs_constraints_mutex
);
7735 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7739 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7740 if (cfs_quota_us
< 0)
7741 quota
= RUNTIME_INF
;
7743 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7745 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7748 long tg_get_cfs_quota(struct task_group
*tg
)
7752 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7755 quota_us
= tg
->cfs_bandwidth
.quota
;
7756 do_div(quota_us
, NSEC_PER_USEC
);
7761 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7765 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7766 quota
= tg
->cfs_bandwidth
.quota
;
7768 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7771 long tg_get_cfs_period(struct task_group
*tg
)
7775 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7776 do_div(cfs_period_us
, NSEC_PER_USEC
);
7778 return cfs_period_us
;
7781 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7784 return tg_get_cfs_quota(css_tg(css
));
7787 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7788 struct cftype
*cftype
, s64 cfs_quota_us
)
7790 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7793 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7796 return tg_get_cfs_period(css_tg(css
));
7799 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7800 struct cftype
*cftype
, u64 cfs_period_us
)
7802 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7805 struct cfs_schedulable_data
{
7806 struct task_group
*tg
;
7811 * normalize group quota/period to be quota/max_period
7812 * note: units are usecs
7814 static u64
normalize_cfs_quota(struct task_group
*tg
,
7815 struct cfs_schedulable_data
*d
)
7823 period
= tg_get_cfs_period(tg
);
7824 quota
= tg_get_cfs_quota(tg
);
7827 /* note: these should typically be equivalent */
7828 if (quota
== RUNTIME_INF
|| quota
== -1)
7831 return to_ratio(period
, quota
);
7834 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7836 struct cfs_schedulable_data
*d
= data
;
7837 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7838 s64 quota
= 0, parent_quota
= -1;
7841 quota
= RUNTIME_INF
;
7843 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7845 quota
= normalize_cfs_quota(tg
, d
);
7846 parent_quota
= parent_b
->hierarchal_quota
;
7849 * ensure max(child_quota) <= parent_quota, inherit when no
7852 if (quota
== RUNTIME_INF
)
7853 quota
= parent_quota
;
7854 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7857 cfs_b
->hierarchal_quota
= quota
;
7862 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7865 struct cfs_schedulable_data data
= {
7871 if (quota
!= RUNTIME_INF
) {
7872 do_div(data
.period
, NSEC_PER_USEC
);
7873 do_div(data
.quota
, NSEC_PER_USEC
);
7877 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7883 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7885 struct task_group
*tg
= css_tg(seq_css(sf
));
7886 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7888 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7889 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7890 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7894 #endif /* CONFIG_CFS_BANDWIDTH */
7895 #endif /* CONFIG_FAIR_GROUP_SCHED */
7897 #ifdef CONFIG_RT_GROUP_SCHED
7898 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7899 struct cftype
*cft
, s64 val
)
7901 return sched_group_set_rt_runtime(css_tg(css
), val
);
7904 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7907 return sched_group_rt_runtime(css_tg(css
));
7910 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7911 struct cftype
*cftype
, u64 rt_period_us
)
7913 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7916 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7919 return sched_group_rt_period(css_tg(css
));
7921 #endif /* CONFIG_RT_GROUP_SCHED */
7923 static struct cftype cpu_files
[] = {
7924 #ifdef CONFIG_FAIR_GROUP_SCHED
7927 .read_u64
= cpu_shares_read_u64
,
7928 .write_u64
= cpu_shares_write_u64
,
7931 #ifdef CONFIG_CFS_BANDWIDTH
7933 .name
= "cfs_quota_us",
7934 .read_s64
= cpu_cfs_quota_read_s64
,
7935 .write_s64
= cpu_cfs_quota_write_s64
,
7938 .name
= "cfs_period_us",
7939 .read_u64
= cpu_cfs_period_read_u64
,
7940 .write_u64
= cpu_cfs_period_write_u64
,
7944 .seq_show
= cpu_stats_show
,
7947 #ifdef CONFIG_RT_GROUP_SCHED
7949 .name
= "rt_runtime_us",
7950 .read_s64
= cpu_rt_runtime_read
,
7951 .write_s64
= cpu_rt_runtime_write
,
7954 .name
= "rt_period_us",
7955 .read_u64
= cpu_rt_period_read_uint
,
7956 .write_u64
= cpu_rt_period_write_uint
,
7962 struct cgroup_subsys cpu_cgroup_subsys
= {
7964 .css_alloc
= cpu_cgroup_css_alloc
,
7965 .css_free
= cpu_cgroup_css_free
,
7966 .css_online
= cpu_cgroup_css_online
,
7967 .css_offline
= cpu_cgroup_css_offline
,
7968 .can_attach
= cpu_cgroup_can_attach
,
7969 .attach
= cpu_cgroup_attach
,
7970 .exit
= cpu_cgroup_exit
,
7971 .subsys_id
= cpu_cgroup_subsys_id
,
7972 .base_cftypes
= cpu_files
,
7976 #endif /* CONFIG_CGROUP_SCHED */
7978 void dump_cpu_task(int cpu
)
7980 pr_info("Task dump for CPU %d:\n", cpu
);
7981 sched_show_task(cpu_curr(cpu
));