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
?: attr
->sched_deadline
;
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 if (prev
->sched_class
->task_dead
)
2155 prev
->sched_class
->task_dead(prev
);
2158 * Remove function-return probe instances associated with this
2159 * task and put them back on the free list.
2161 kprobe_flush_task(prev
);
2162 put_task_struct(prev
);
2165 tick_nohz_task_switch(current
);
2170 /* rq->lock is NOT held, but preemption is disabled */
2171 static inline void post_schedule(struct rq
*rq
)
2173 if (rq
->post_schedule
) {
2174 unsigned long flags
;
2176 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2177 if (rq
->curr
->sched_class
->post_schedule
)
2178 rq
->curr
->sched_class
->post_schedule(rq
);
2179 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2181 rq
->post_schedule
= 0;
2187 static inline void post_schedule(struct rq
*rq
)
2194 * schedule_tail - first thing a freshly forked thread must call.
2195 * @prev: the thread we just switched away from.
2197 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2198 __releases(rq
->lock
)
2200 struct rq
*rq
= this_rq();
2202 finish_task_switch(rq
, prev
);
2205 * FIXME: do we need to worry about rq being invalidated by the
2210 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2211 /* In this case, finish_task_switch does not reenable preemption */
2214 if (current
->set_child_tid
)
2215 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2219 * context_switch - switch to the new MM and the new
2220 * thread's register state.
2223 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2224 struct task_struct
*next
)
2226 struct mm_struct
*mm
, *oldmm
;
2228 prepare_task_switch(rq
, prev
, next
);
2231 oldmm
= prev
->active_mm
;
2233 * For paravirt, this is coupled with an exit in switch_to to
2234 * combine the page table reload and the switch backend into
2237 arch_start_context_switch(prev
);
2240 next
->active_mm
= oldmm
;
2241 atomic_inc(&oldmm
->mm_count
);
2242 enter_lazy_tlb(oldmm
, next
);
2244 switch_mm(oldmm
, mm
, next
);
2247 prev
->active_mm
= NULL
;
2248 rq
->prev_mm
= oldmm
;
2251 * Since the runqueue lock will be released by the next
2252 * task (which is an invalid locking op but in the case
2253 * of the scheduler it's an obvious special-case), so we
2254 * do an early lockdep release here:
2256 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2257 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2260 context_tracking_task_switch(prev
, next
);
2261 /* Here we just switch the register state and the stack. */
2262 switch_to(prev
, next
, prev
);
2266 * this_rq must be evaluated again because prev may have moved
2267 * CPUs since it called schedule(), thus the 'rq' on its stack
2268 * frame will be invalid.
2270 finish_task_switch(this_rq(), prev
);
2274 * nr_running and nr_context_switches:
2276 * externally visible scheduler statistics: current number of runnable
2277 * threads, total number of context switches performed since bootup.
2279 unsigned long nr_running(void)
2281 unsigned long i
, sum
= 0;
2283 for_each_online_cpu(i
)
2284 sum
+= cpu_rq(i
)->nr_running
;
2289 unsigned long long nr_context_switches(void)
2292 unsigned long long sum
= 0;
2294 for_each_possible_cpu(i
)
2295 sum
+= cpu_rq(i
)->nr_switches
;
2300 unsigned long nr_iowait(void)
2302 unsigned long i
, sum
= 0;
2304 for_each_possible_cpu(i
)
2305 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2310 unsigned long nr_iowait_cpu(int cpu
)
2312 struct rq
*this = cpu_rq(cpu
);
2313 return atomic_read(&this->nr_iowait
);
2319 * sched_exec - execve() is a valuable balancing opportunity, because at
2320 * this point the task has the smallest effective memory and cache footprint.
2322 void sched_exec(void)
2324 struct task_struct
*p
= current
;
2325 unsigned long flags
;
2328 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2329 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2330 if (dest_cpu
== smp_processor_id())
2333 if (likely(cpu_active(dest_cpu
))) {
2334 struct migration_arg arg
= { p
, dest_cpu
};
2336 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2337 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2341 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2346 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2347 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2349 EXPORT_PER_CPU_SYMBOL(kstat
);
2350 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2353 * Return any ns on the sched_clock that have not yet been accounted in
2354 * @p in case that task is currently running.
2356 * Called with task_rq_lock() held on @rq.
2358 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
2362 if (task_current(rq
, p
)) {
2363 update_rq_clock(rq
);
2364 ns
= rq_clock_task(rq
) - p
->se
.exec_start
;
2372 unsigned long long task_delta_exec(struct task_struct
*p
)
2374 unsigned long flags
;
2378 rq
= task_rq_lock(p
, &flags
);
2379 ns
= do_task_delta_exec(p
, rq
);
2380 task_rq_unlock(rq
, p
, &flags
);
2386 * Return accounted runtime for the task.
2387 * In case the task is currently running, return the runtime plus current's
2388 * pending runtime that have not been accounted yet.
2390 unsigned long long task_sched_runtime(struct task_struct
*p
)
2392 unsigned long flags
;
2396 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2398 * 64-bit doesn't need locks to atomically read a 64bit value.
2399 * So we have a optimization chance when the task's delta_exec is 0.
2400 * Reading ->on_cpu is racy, but this is ok.
2402 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2403 * If we race with it entering cpu, unaccounted time is 0. This is
2404 * indistinguishable from the read occurring a few cycles earlier.
2407 return p
->se
.sum_exec_runtime
;
2410 rq
= task_rq_lock(p
, &flags
);
2411 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
2412 task_rq_unlock(rq
, p
, &flags
);
2418 * This function gets called by the timer code, with HZ frequency.
2419 * We call it with interrupts disabled.
2421 void scheduler_tick(void)
2423 int cpu
= smp_processor_id();
2424 struct rq
*rq
= cpu_rq(cpu
);
2425 struct task_struct
*curr
= rq
->curr
;
2429 raw_spin_lock(&rq
->lock
);
2430 update_rq_clock(rq
);
2431 curr
->sched_class
->task_tick(rq
, curr
, 0);
2432 update_cpu_load_active(rq
);
2433 raw_spin_unlock(&rq
->lock
);
2435 perf_event_task_tick();
2438 rq
->idle_balance
= idle_cpu(cpu
);
2439 trigger_load_balance(rq
);
2441 rq_last_tick_reset(rq
);
2444 #ifdef CONFIG_NO_HZ_FULL
2446 * scheduler_tick_max_deferment
2448 * Keep at least one tick per second when a single
2449 * active task is running because the scheduler doesn't
2450 * yet completely support full dynticks environment.
2452 * This makes sure that uptime, CFS vruntime, load
2453 * balancing, etc... continue to move forward, even
2454 * with a very low granularity.
2456 * Return: Maximum deferment in nanoseconds.
2458 u64
scheduler_tick_max_deferment(void)
2460 struct rq
*rq
= this_rq();
2461 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2463 next
= rq
->last_sched_tick
+ HZ
;
2465 if (time_before_eq(next
, now
))
2468 return jiffies_to_nsecs(next
- now
);
2472 notrace
unsigned long get_parent_ip(unsigned long addr
)
2474 if (in_lock_functions(addr
)) {
2475 addr
= CALLER_ADDR2
;
2476 if (in_lock_functions(addr
))
2477 addr
= CALLER_ADDR3
;
2482 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2483 defined(CONFIG_PREEMPT_TRACER))
2485 void __kprobes
preempt_count_add(int val
)
2487 #ifdef CONFIG_DEBUG_PREEMPT
2491 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2494 __preempt_count_add(val
);
2495 #ifdef CONFIG_DEBUG_PREEMPT
2497 * Spinlock count overflowing soon?
2499 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2502 if (preempt_count() == val
) {
2503 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2504 #ifdef CONFIG_DEBUG_PREEMPT
2505 current
->preempt_disable_ip
= ip
;
2507 trace_preempt_off(CALLER_ADDR0
, ip
);
2510 EXPORT_SYMBOL(preempt_count_add
);
2512 void __kprobes
preempt_count_sub(int val
)
2514 #ifdef CONFIG_DEBUG_PREEMPT
2518 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2521 * Is the spinlock portion underflowing?
2523 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2524 !(preempt_count() & PREEMPT_MASK
)))
2528 if (preempt_count() == val
)
2529 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2530 __preempt_count_sub(val
);
2532 EXPORT_SYMBOL(preempt_count_sub
);
2537 * Print scheduling while atomic bug:
2539 static noinline
void __schedule_bug(struct task_struct
*prev
)
2541 if (oops_in_progress
)
2544 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2545 prev
->comm
, prev
->pid
, preempt_count());
2547 debug_show_held_locks(prev
);
2549 if (irqs_disabled())
2550 print_irqtrace_events(prev
);
2551 #ifdef CONFIG_DEBUG_PREEMPT
2552 if (in_atomic_preempt_off()) {
2553 pr_err("Preemption disabled at:");
2554 print_ip_sym(current
->preempt_disable_ip
);
2559 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2563 * Various schedule()-time debugging checks and statistics:
2565 static inline void schedule_debug(struct task_struct
*prev
)
2568 * Test if we are atomic. Since do_exit() needs to call into
2569 * schedule() atomically, we ignore that path. Otherwise whine
2570 * if we are scheduling when we should not.
2572 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2573 __schedule_bug(prev
);
2576 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2578 schedstat_inc(this_rq(), sched_count
);
2582 * Pick up the highest-prio task:
2584 static inline struct task_struct
*
2585 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2587 const struct sched_class
*class = &fair_sched_class
;
2588 struct task_struct
*p
;
2591 * Optimization: we know that if all tasks are in
2592 * the fair class we can call that function directly:
2594 if (likely(prev
->sched_class
== class &&
2595 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2596 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2597 if (likely(p
&& p
!= RETRY_TASK
))
2602 for_each_class(class) {
2603 p
= class->pick_next_task(rq
, prev
);
2605 if (unlikely(p
== RETRY_TASK
))
2611 BUG(); /* the idle class will always have a runnable task */
2615 * __schedule() is the main scheduler function.
2617 * The main means of driving the scheduler and thus entering this function are:
2619 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2621 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2622 * paths. For example, see arch/x86/entry_64.S.
2624 * To drive preemption between tasks, the scheduler sets the flag in timer
2625 * interrupt handler scheduler_tick().
2627 * 3. Wakeups don't really cause entry into schedule(). They add a
2628 * task to the run-queue and that's it.
2630 * Now, if the new task added to the run-queue preempts the current
2631 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2632 * called on the nearest possible occasion:
2634 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2636 * - in syscall or exception context, at the next outmost
2637 * preempt_enable(). (this might be as soon as the wake_up()'s
2640 * - in IRQ context, return from interrupt-handler to
2641 * preemptible context
2643 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2646 * - cond_resched() call
2647 * - explicit schedule() call
2648 * - return from syscall or exception to user-space
2649 * - return from interrupt-handler to user-space
2651 static void __sched
__schedule(void)
2653 struct task_struct
*prev
, *next
;
2654 unsigned long *switch_count
;
2660 cpu
= smp_processor_id();
2662 rcu_note_context_switch(cpu
);
2665 schedule_debug(prev
);
2667 if (sched_feat(HRTICK
))
2671 * Make sure that signal_pending_state()->signal_pending() below
2672 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2673 * done by the caller to avoid the race with signal_wake_up().
2675 smp_mb__before_spinlock();
2676 raw_spin_lock_irq(&rq
->lock
);
2678 switch_count
= &prev
->nivcsw
;
2679 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2680 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2681 prev
->state
= TASK_RUNNING
;
2683 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2687 * If a worker went to sleep, notify and ask workqueue
2688 * whether it wants to wake up a task to maintain
2691 if (prev
->flags
& PF_WQ_WORKER
) {
2692 struct task_struct
*to_wakeup
;
2694 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2696 try_to_wake_up_local(to_wakeup
);
2699 switch_count
= &prev
->nvcsw
;
2702 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
2703 update_rq_clock(rq
);
2705 next
= pick_next_task(rq
, prev
);
2706 clear_tsk_need_resched(prev
);
2707 clear_preempt_need_resched();
2708 rq
->skip_clock_update
= 0;
2710 if (likely(prev
!= next
)) {
2715 context_switch(rq
, prev
, next
); /* unlocks the rq */
2717 * The context switch have flipped the stack from under us
2718 * and restored the local variables which were saved when
2719 * this task called schedule() in the past. prev == current
2720 * is still correct, but it can be moved to another cpu/rq.
2722 cpu
= smp_processor_id();
2725 raw_spin_unlock_irq(&rq
->lock
);
2729 sched_preempt_enable_no_resched();
2734 static inline void sched_submit_work(struct task_struct
*tsk
)
2736 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2739 * If we are going to sleep and we have plugged IO queued,
2740 * make sure to submit it to avoid deadlocks.
2742 if (blk_needs_flush_plug(tsk
))
2743 blk_schedule_flush_plug(tsk
);
2746 asmlinkage
void __sched
schedule(void)
2748 struct task_struct
*tsk
= current
;
2750 sched_submit_work(tsk
);
2753 EXPORT_SYMBOL(schedule
);
2755 #ifdef CONFIG_CONTEXT_TRACKING
2756 asmlinkage
void __sched
schedule_user(void)
2759 * If we come here after a random call to set_need_resched(),
2760 * or we have been woken up remotely but the IPI has not yet arrived,
2761 * we haven't yet exited the RCU idle mode. Do it here manually until
2762 * we find a better solution.
2771 * schedule_preempt_disabled - called with preemption disabled
2773 * Returns with preemption disabled. Note: preempt_count must be 1
2775 void __sched
schedule_preempt_disabled(void)
2777 sched_preempt_enable_no_resched();
2782 #ifdef CONFIG_PREEMPT
2784 * this is the entry point to schedule() from in-kernel preemption
2785 * off of preempt_enable. Kernel preemptions off return from interrupt
2786 * occur there and call schedule directly.
2788 asmlinkage
void __sched notrace
preempt_schedule(void)
2791 * If there is a non-zero preempt_count or interrupts are disabled,
2792 * we do not want to preempt the current task. Just return..
2794 if (likely(!preemptible()))
2798 __preempt_count_add(PREEMPT_ACTIVE
);
2800 __preempt_count_sub(PREEMPT_ACTIVE
);
2803 * Check again in case we missed a preemption opportunity
2804 * between schedule and now.
2807 } while (need_resched());
2809 EXPORT_SYMBOL(preempt_schedule
);
2810 #endif /* CONFIG_PREEMPT */
2813 * this is the entry point to schedule() from kernel preemption
2814 * off of irq context.
2815 * Note, that this is called and return with irqs disabled. This will
2816 * protect us against recursive calling from irq.
2818 asmlinkage
void __sched
preempt_schedule_irq(void)
2820 enum ctx_state prev_state
;
2822 /* Catch callers which need to be fixed */
2823 BUG_ON(preempt_count() || !irqs_disabled());
2825 prev_state
= exception_enter();
2828 __preempt_count_add(PREEMPT_ACTIVE
);
2831 local_irq_disable();
2832 __preempt_count_sub(PREEMPT_ACTIVE
);
2835 * Check again in case we missed a preemption opportunity
2836 * between schedule and now.
2839 } while (need_resched());
2841 exception_exit(prev_state
);
2844 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2847 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2849 EXPORT_SYMBOL(default_wake_function
);
2852 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
2854 unsigned long flags
;
2857 init_waitqueue_entry(&wait
, current
);
2859 __set_current_state(state
);
2861 spin_lock_irqsave(&q
->lock
, flags
);
2862 __add_wait_queue(q
, &wait
);
2863 spin_unlock(&q
->lock
);
2864 timeout
= schedule_timeout(timeout
);
2865 spin_lock_irq(&q
->lock
);
2866 __remove_wait_queue(q
, &wait
);
2867 spin_unlock_irqrestore(&q
->lock
, flags
);
2872 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
2874 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2876 EXPORT_SYMBOL(interruptible_sleep_on
);
2879 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
2881 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
2883 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
2885 void __sched
sleep_on(wait_queue_head_t
*q
)
2887 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2889 EXPORT_SYMBOL(sleep_on
);
2891 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
2893 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
2895 EXPORT_SYMBOL(sleep_on_timeout
);
2897 #ifdef CONFIG_RT_MUTEXES
2900 * rt_mutex_setprio - set the current priority of a task
2902 * @prio: prio value (kernel-internal form)
2904 * This function changes the 'effective' priority of a task. It does
2905 * not touch ->normal_prio like __setscheduler().
2907 * Used by the rt_mutex code to implement priority inheritance
2908 * logic. Call site only calls if the priority of the task changed.
2910 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2912 int oldprio
, on_rq
, running
, enqueue_flag
= 0;
2914 const struct sched_class
*prev_class
;
2916 BUG_ON(prio
> MAX_PRIO
);
2918 rq
= __task_rq_lock(p
);
2921 * Idle task boosting is a nono in general. There is one
2922 * exception, when PREEMPT_RT and NOHZ is active:
2924 * The idle task calls get_next_timer_interrupt() and holds
2925 * the timer wheel base->lock on the CPU and another CPU wants
2926 * to access the timer (probably to cancel it). We can safely
2927 * ignore the boosting request, as the idle CPU runs this code
2928 * with interrupts disabled and will complete the lock
2929 * protected section without being interrupted. So there is no
2930 * real need to boost.
2932 if (unlikely(p
== rq
->idle
)) {
2933 WARN_ON(p
!= rq
->curr
);
2934 WARN_ON(p
->pi_blocked_on
);
2938 trace_sched_pi_setprio(p
, prio
);
2939 p
->pi_top_task
= rt_mutex_get_top_task(p
);
2941 prev_class
= p
->sched_class
;
2943 running
= task_current(rq
, p
);
2945 dequeue_task(rq
, p
, 0);
2947 p
->sched_class
->put_prev_task(rq
, p
);
2950 * Boosting condition are:
2951 * 1. -rt task is running and holds mutex A
2952 * --> -dl task blocks on mutex A
2954 * 2. -dl task is running and holds mutex A
2955 * --> -dl task blocks on mutex A and could preempt the
2958 if (dl_prio(prio
)) {
2959 if (!dl_prio(p
->normal_prio
) || (p
->pi_top_task
&&
2960 dl_entity_preempt(&p
->pi_top_task
->dl
, &p
->dl
))) {
2961 p
->dl
.dl_boosted
= 1;
2962 p
->dl
.dl_throttled
= 0;
2963 enqueue_flag
= ENQUEUE_REPLENISH
;
2965 p
->dl
.dl_boosted
= 0;
2966 p
->sched_class
= &dl_sched_class
;
2967 } else if (rt_prio(prio
)) {
2968 if (dl_prio(oldprio
))
2969 p
->dl
.dl_boosted
= 0;
2971 enqueue_flag
= ENQUEUE_HEAD
;
2972 p
->sched_class
= &rt_sched_class
;
2974 if (dl_prio(oldprio
))
2975 p
->dl
.dl_boosted
= 0;
2976 p
->sched_class
= &fair_sched_class
;
2982 p
->sched_class
->set_curr_task(rq
);
2984 enqueue_task(rq
, p
, enqueue_flag
);
2986 check_class_changed(rq
, p
, prev_class
, oldprio
);
2988 __task_rq_unlock(rq
);
2992 void set_user_nice(struct task_struct
*p
, long nice
)
2994 int old_prio
, delta
, on_rq
;
2995 unsigned long flags
;
2998 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3001 * We have to be careful, if called from sys_setpriority(),
3002 * the task might be in the middle of scheduling on another CPU.
3004 rq
= task_rq_lock(p
, &flags
);
3006 * The RT priorities are set via sched_setscheduler(), but we still
3007 * allow the 'normal' nice value to be set - but as expected
3008 * it wont have any effect on scheduling until the task is
3009 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3011 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3012 p
->static_prio
= NICE_TO_PRIO(nice
);
3017 dequeue_task(rq
, p
, 0);
3019 p
->static_prio
= NICE_TO_PRIO(nice
);
3022 p
->prio
= effective_prio(p
);
3023 delta
= p
->prio
- old_prio
;
3026 enqueue_task(rq
, p
, 0);
3028 * If the task increased its priority or is running and
3029 * lowered its priority, then reschedule its CPU:
3031 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3032 resched_task(rq
->curr
);
3035 task_rq_unlock(rq
, p
, &flags
);
3037 EXPORT_SYMBOL(set_user_nice
);
3040 * can_nice - check if a task can reduce its nice value
3044 int can_nice(const struct task_struct
*p
, const int nice
)
3046 /* convert nice value [19,-20] to rlimit style value [1,40] */
3047 int nice_rlim
= 20 - nice
;
3049 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3050 capable(CAP_SYS_NICE
));
3053 #ifdef __ARCH_WANT_SYS_NICE
3056 * sys_nice - change the priority of the current process.
3057 * @increment: priority increment
3059 * sys_setpriority is a more generic, but much slower function that
3060 * does similar things.
3062 SYSCALL_DEFINE1(nice
, int, increment
)
3067 * Setpriority might change our priority at the same moment.
3068 * We don't have to worry. Conceptually one call occurs first
3069 * and we have a single winner.
3071 if (increment
< -40)
3076 nice
= task_nice(current
) + increment
;
3077 if (nice
< MIN_NICE
)
3079 if (nice
> MAX_NICE
)
3082 if (increment
< 0 && !can_nice(current
, nice
))
3085 retval
= security_task_setnice(current
, nice
);
3089 set_user_nice(current
, nice
);
3096 * task_prio - return the priority value of a given task.
3097 * @p: the task in question.
3099 * Return: The priority value as seen by users in /proc.
3100 * RT tasks are offset by -200. Normal tasks are centered
3101 * around 0, value goes from -16 to +15.
3103 int task_prio(const struct task_struct
*p
)
3105 return p
->prio
- MAX_RT_PRIO
;
3109 * idle_cpu - is a given cpu idle currently?
3110 * @cpu: the processor in question.
3112 * Return: 1 if the CPU is currently idle. 0 otherwise.
3114 int idle_cpu(int cpu
)
3116 struct rq
*rq
= cpu_rq(cpu
);
3118 if (rq
->curr
!= rq
->idle
)
3125 if (!llist_empty(&rq
->wake_list
))
3133 * idle_task - return the idle task for a given cpu.
3134 * @cpu: the processor in question.
3136 * Return: The idle task for the cpu @cpu.
3138 struct task_struct
*idle_task(int cpu
)
3140 return cpu_rq(cpu
)->idle
;
3144 * find_process_by_pid - find a process with a matching PID value.
3145 * @pid: the pid in question.
3147 * The task of @pid, if found. %NULL otherwise.
3149 static struct task_struct
*find_process_by_pid(pid_t pid
)
3151 return pid
? find_task_by_vpid(pid
) : current
;
3155 * This function initializes the sched_dl_entity of a newly becoming
3156 * SCHED_DEADLINE task.
3158 * Only the static values are considered here, the actual runtime and the
3159 * absolute deadline will be properly calculated when the task is enqueued
3160 * for the first time with its new policy.
3163 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3165 struct sched_dl_entity
*dl_se
= &p
->dl
;
3167 init_dl_task_timer(dl_se
);
3168 dl_se
->dl_runtime
= attr
->sched_runtime
;
3169 dl_se
->dl_deadline
= attr
->sched_deadline
;
3170 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3171 dl_se
->flags
= attr
->sched_flags
;
3172 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3173 dl_se
->dl_throttled
= 0;
3177 static void __setscheduler_params(struct task_struct
*p
,
3178 const struct sched_attr
*attr
)
3180 int policy
= attr
->sched_policy
;
3182 if (policy
== -1) /* setparam */
3187 if (dl_policy(policy
))
3188 __setparam_dl(p
, attr
);
3189 else if (fair_policy(policy
))
3190 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3193 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3194 * !rt_policy. Always setting this ensures that things like
3195 * getparam()/getattr() don't report silly values for !rt tasks.
3197 p
->rt_priority
= attr
->sched_priority
;
3198 p
->normal_prio
= normal_prio(p
);
3202 /* Actually do priority change: must hold pi & rq lock. */
3203 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3204 const struct sched_attr
*attr
)
3206 __setscheduler_params(p
, attr
);
3209 * If we get here, there was no pi waiters boosting the
3210 * task. It is safe to use the normal prio.
3212 p
->prio
= normal_prio(p
);
3214 if (dl_prio(p
->prio
))
3215 p
->sched_class
= &dl_sched_class
;
3216 else if (rt_prio(p
->prio
))
3217 p
->sched_class
= &rt_sched_class
;
3219 p
->sched_class
= &fair_sched_class
;
3223 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3225 struct sched_dl_entity
*dl_se
= &p
->dl
;
3227 attr
->sched_priority
= p
->rt_priority
;
3228 attr
->sched_runtime
= dl_se
->dl_runtime
;
3229 attr
->sched_deadline
= dl_se
->dl_deadline
;
3230 attr
->sched_period
= dl_se
->dl_period
;
3231 attr
->sched_flags
= dl_se
->flags
;
3235 * This function validates the new parameters of a -deadline task.
3236 * We ask for the deadline not being zero, and greater or equal
3237 * than the runtime, as well as the period of being zero or
3238 * greater than deadline. Furthermore, we have to be sure that
3239 * user parameters are above the internal resolution (1us); we
3240 * check sched_runtime only since it is always the smaller one.
3243 __checkparam_dl(const struct sched_attr
*attr
)
3245 return attr
&& attr
->sched_deadline
!= 0 &&
3246 (attr
->sched_period
== 0 ||
3247 (s64
)(attr
->sched_period
- attr
->sched_deadline
) >= 0) &&
3248 (s64
)(attr
->sched_deadline
- attr
->sched_runtime
) >= 0 &&
3249 attr
->sched_runtime
>= (2 << (DL_SCALE
- 1));
3253 * check the target process has a UID that matches the current process's
3255 static bool check_same_owner(struct task_struct
*p
)
3257 const struct cred
*cred
= current_cred(), *pcred
;
3261 pcred
= __task_cred(p
);
3262 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3263 uid_eq(cred
->euid
, pcred
->uid
));
3268 static int __sched_setscheduler(struct task_struct
*p
,
3269 const struct sched_attr
*attr
,
3272 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3273 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3274 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
3275 int policy
= attr
->sched_policy
;
3276 unsigned long flags
;
3277 const struct sched_class
*prev_class
;
3281 /* may grab non-irq protected spin_locks */
3282 BUG_ON(in_interrupt());
3284 /* double check policy once rq lock held */
3286 reset_on_fork
= p
->sched_reset_on_fork
;
3287 policy
= oldpolicy
= p
->policy
;
3289 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3291 if (policy
!= SCHED_DEADLINE
&&
3292 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3293 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3294 policy
!= SCHED_IDLE
)
3298 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3302 * Valid priorities for SCHED_FIFO and SCHED_RR are
3303 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3304 * SCHED_BATCH and SCHED_IDLE is 0.
3306 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3307 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3309 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3310 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3314 * Allow unprivileged RT tasks to decrease priority:
3316 if (user
&& !capable(CAP_SYS_NICE
)) {
3317 if (fair_policy(policy
)) {
3318 if (attr
->sched_nice
< task_nice(p
) &&
3319 !can_nice(p
, attr
->sched_nice
))
3323 if (rt_policy(policy
)) {
3324 unsigned long rlim_rtprio
=
3325 task_rlimit(p
, RLIMIT_RTPRIO
);
3327 /* can't set/change the rt policy */
3328 if (policy
!= p
->policy
&& !rlim_rtprio
)
3331 /* can't increase priority */
3332 if (attr
->sched_priority
> p
->rt_priority
&&
3333 attr
->sched_priority
> rlim_rtprio
)
3338 * Can't set/change SCHED_DEADLINE policy at all for now
3339 * (safest behavior); in the future we would like to allow
3340 * unprivileged DL tasks to increase their relative deadline
3341 * or reduce their runtime (both ways reducing utilization)
3343 if (dl_policy(policy
))
3347 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3348 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3350 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3351 if (!can_nice(p
, task_nice(p
)))
3355 /* can't change other user's priorities */
3356 if (!check_same_owner(p
))
3359 /* Normal users shall not reset the sched_reset_on_fork flag */
3360 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3365 retval
= security_task_setscheduler(p
);
3371 * make sure no PI-waiters arrive (or leave) while we are
3372 * changing the priority of the task:
3374 * To be able to change p->policy safely, the appropriate
3375 * runqueue lock must be held.
3377 rq
= task_rq_lock(p
, &flags
);
3380 * Changing the policy of the stop threads its a very bad idea
3382 if (p
== rq
->stop
) {
3383 task_rq_unlock(rq
, p
, &flags
);
3388 * If not changing anything there's no need to proceed further,
3389 * but store a possible modification of reset_on_fork.
3391 if (unlikely(policy
== p
->policy
)) {
3392 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3394 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3396 if (dl_policy(policy
))
3399 p
->sched_reset_on_fork
= reset_on_fork
;
3400 task_rq_unlock(rq
, p
, &flags
);
3406 #ifdef CONFIG_RT_GROUP_SCHED
3408 * Do not allow realtime tasks into groups that have no runtime
3411 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3412 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3413 !task_group_is_autogroup(task_group(p
))) {
3414 task_rq_unlock(rq
, p
, &flags
);
3419 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3420 cpumask_t
*span
= rq
->rd
->span
;
3423 * Don't allow tasks with an affinity mask smaller than
3424 * the entire root_domain to become SCHED_DEADLINE. We
3425 * will also fail if there's no bandwidth available.
3427 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3428 rq
->rd
->dl_bw
.bw
== 0) {
3429 task_rq_unlock(rq
, p
, &flags
);
3436 /* recheck policy now with rq lock held */
3437 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3438 policy
= oldpolicy
= -1;
3439 task_rq_unlock(rq
, p
, &flags
);
3444 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3445 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3448 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3449 task_rq_unlock(rq
, p
, &flags
);
3453 p
->sched_reset_on_fork
= reset_on_fork
;
3457 * Special case for priority boosted tasks.
3459 * If the new priority is lower or equal (user space view)
3460 * than the current (boosted) priority, we just store the new
3461 * normal parameters and do not touch the scheduler class and
3462 * the runqueue. This will be done when the task deboost
3465 if (rt_mutex_check_prio(p
, newprio
)) {
3466 __setscheduler_params(p
, attr
);
3467 task_rq_unlock(rq
, p
, &flags
);
3472 running
= task_current(rq
, p
);
3474 dequeue_task(rq
, p
, 0);
3476 p
->sched_class
->put_prev_task(rq
, p
);
3478 prev_class
= p
->sched_class
;
3479 __setscheduler(rq
, p
, attr
);
3482 p
->sched_class
->set_curr_task(rq
);
3485 * We enqueue to tail when the priority of a task is
3486 * increased (user space view).
3488 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3491 check_class_changed(rq
, p
, prev_class
, oldprio
);
3492 task_rq_unlock(rq
, p
, &flags
);
3494 rt_mutex_adjust_pi(p
);
3499 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3500 const struct sched_param
*param
, bool check
)
3502 struct sched_attr attr
= {
3503 .sched_policy
= policy
,
3504 .sched_priority
= param
->sched_priority
,
3505 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3509 * Fixup the legacy SCHED_RESET_ON_FORK hack
3511 if (policy
& SCHED_RESET_ON_FORK
) {
3512 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3513 policy
&= ~SCHED_RESET_ON_FORK
;
3514 attr
.sched_policy
= policy
;
3517 return __sched_setscheduler(p
, &attr
, check
);
3520 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3521 * @p: the task in question.
3522 * @policy: new policy.
3523 * @param: structure containing the new RT priority.
3525 * Return: 0 on success. An error code otherwise.
3527 * NOTE that the task may be already dead.
3529 int sched_setscheduler(struct task_struct
*p
, int policy
,
3530 const struct sched_param
*param
)
3532 return _sched_setscheduler(p
, policy
, param
, true);
3534 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3536 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3538 return __sched_setscheduler(p
, attr
, true);
3540 EXPORT_SYMBOL_GPL(sched_setattr
);
3543 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3544 * @p: the task in question.
3545 * @policy: new policy.
3546 * @param: structure containing the new RT priority.
3548 * Just like sched_setscheduler, only don't bother checking if the
3549 * current context has permission. For example, this is needed in
3550 * stop_machine(): we create temporary high priority worker threads,
3551 * but our caller might not have that capability.
3553 * Return: 0 on success. An error code otherwise.
3555 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3556 const struct sched_param
*param
)
3558 return _sched_setscheduler(p
, policy
, param
, false);
3562 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3564 struct sched_param lparam
;
3565 struct task_struct
*p
;
3568 if (!param
|| pid
< 0)
3570 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3575 p
= find_process_by_pid(pid
);
3577 retval
= sched_setscheduler(p
, policy
, &lparam
);
3584 * Mimics kernel/events/core.c perf_copy_attr().
3586 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3587 struct sched_attr
*attr
)
3592 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3596 * zero the full structure, so that a short copy will be nice.
3598 memset(attr
, 0, sizeof(*attr
));
3600 ret
= get_user(size
, &uattr
->size
);
3604 if (size
> PAGE_SIZE
) /* silly large */
3607 if (!size
) /* abi compat */
3608 size
= SCHED_ATTR_SIZE_VER0
;
3610 if (size
< SCHED_ATTR_SIZE_VER0
)
3614 * If we're handed a bigger struct than we know of,
3615 * ensure all the unknown bits are 0 - i.e. new
3616 * user-space does not rely on any kernel feature
3617 * extensions we dont know about yet.
3619 if (size
> sizeof(*attr
)) {
3620 unsigned char __user
*addr
;
3621 unsigned char __user
*end
;
3624 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3625 end
= (void __user
*)uattr
+ size
;
3627 for (; addr
< end
; addr
++) {
3628 ret
= get_user(val
, addr
);
3634 size
= sizeof(*attr
);
3637 ret
= copy_from_user(attr
, uattr
, size
);
3642 * XXX: do we want to be lenient like existing syscalls; or do we want
3643 * to be strict and return an error on out-of-bounds values?
3645 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3651 put_user(sizeof(*attr
), &uattr
->size
);
3657 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3658 * @pid: the pid in question.
3659 * @policy: new policy.
3660 * @param: structure containing the new RT priority.
3662 * Return: 0 on success. An error code otherwise.
3664 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3665 struct sched_param __user
*, param
)
3667 /* negative values for policy are not valid */
3671 return do_sched_setscheduler(pid
, policy
, param
);
3675 * sys_sched_setparam - set/change the RT priority of a thread
3676 * @pid: the pid in question.
3677 * @param: structure containing the new RT priority.
3679 * Return: 0 on success. An error code otherwise.
3681 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3683 return do_sched_setscheduler(pid
, -1, param
);
3687 * sys_sched_setattr - same as above, but with extended sched_attr
3688 * @pid: the pid in question.
3689 * @uattr: structure containing the extended parameters.
3691 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3692 unsigned int, flags
)
3694 struct sched_attr attr
;
3695 struct task_struct
*p
;
3698 if (!uattr
|| pid
< 0 || flags
)
3701 if (sched_copy_attr(uattr
, &attr
))
3706 p
= find_process_by_pid(pid
);
3708 retval
= sched_setattr(p
, &attr
);
3715 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3716 * @pid: the pid in question.
3718 * Return: On success, the policy of the thread. Otherwise, a negative error
3721 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3723 struct task_struct
*p
;
3731 p
= find_process_by_pid(pid
);
3733 retval
= security_task_getscheduler(p
);
3736 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3743 * sys_sched_getparam - get the RT priority of a thread
3744 * @pid: the pid in question.
3745 * @param: structure containing the RT priority.
3747 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3750 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3752 struct sched_param lp
;
3753 struct task_struct
*p
;
3756 if (!param
|| pid
< 0)
3760 p
= find_process_by_pid(pid
);
3765 retval
= security_task_getscheduler(p
);
3769 if (task_has_dl_policy(p
)) {
3773 lp
.sched_priority
= p
->rt_priority
;
3777 * This one might sleep, we cannot do it with a spinlock held ...
3779 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3788 static int sched_read_attr(struct sched_attr __user
*uattr
,
3789 struct sched_attr
*attr
,
3794 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3798 * If we're handed a smaller struct than we know of,
3799 * ensure all the unknown bits are 0 - i.e. old
3800 * user-space does not get uncomplete information.
3802 if (usize
< sizeof(*attr
)) {
3803 unsigned char *addr
;
3806 addr
= (void *)attr
+ usize
;
3807 end
= (void *)attr
+ sizeof(*attr
);
3809 for (; addr
< end
; addr
++) {
3817 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3830 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3831 * @pid: the pid in question.
3832 * @uattr: structure containing the extended parameters.
3833 * @size: sizeof(attr) for fwd/bwd comp.
3835 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3836 unsigned int, size
, unsigned int, flags
)
3838 struct sched_attr attr
= {
3839 .size
= sizeof(struct sched_attr
),
3841 struct task_struct
*p
;
3844 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3845 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3849 p
= find_process_by_pid(pid
);
3854 retval
= security_task_getscheduler(p
);
3858 attr
.sched_policy
= p
->policy
;
3859 if (p
->sched_reset_on_fork
)
3860 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3861 if (task_has_dl_policy(p
))
3862 __getparam_dl(p
, &attr
);
3863 else if (task_has_rt_policy(p
))
3864 attr
.sched_priority
= p
->rt_priority
;
3866 attr
.sched_nice
= task_nice(p
);
3870 retval
= sched_read_attr(uattr
, &attr
, size
);
3878 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3880 cpumask_var_t cpus_allowed
, new_mask
;
3881 struct task_struct
*p
;
3886 p
= find_process_by_pid(pid
);
3892 /* Prevent p going away */
3896 if (p
->flags
& PF_NO_SETAFFINITY
) {
3900 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
3904 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
3906 goto out_free_cpus_allowed
;
3909 if (!check_same_owner(p
)) {
3911 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
3918 retval
= security_task_setscheduler(p
);
3923 cpuset_cpus_allowed(p
, cpus_allowed
);
3924 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
3927 * Since bandwidth control happens on root_domain basis,
3928 * if admission test is enabled, we only admit -deadline
3929 * tasks allowed to run on all the CPUs in the task's
3933 if (task_has_dl_policy(p
)) {
3934 const struct cpumask
*span
= task_rq(p
)->rd
->span
;
3936 if (dl_bandwidth_enabled() && !cpumask_subset(span
, new_mask
)) {
3943 retval
= set_cpus_allowed_ptr(p
, new_mask
);
3946 cpuset_cpus_allowed(p
, cpus_allowed
);
3947 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
3949 * We must have raced with a concurrent cpuset
3950 * update. Just reset the cpus_allowed to the
3951 * cpuset's cpus_allowed
3953 cpumask_copy(new_mask
, cpus_allowed
);
3958 free_cpumask_var(new_mask
);
3959 out_free_cpus_allowed
:
3960 free_cpumask_var(cpus_allowed
);
3966 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
3967 struct cpumask
*new_mask
)
3969 if (len
< cpumask_size())
3970 cpumask_clear(new_mask
);
3971 else if (len
> cpumask_size())
3972 len
= cpumask_size();
3974 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
3978 * sys_sched_setaffinity - set the cpu affinity of a process
3979 * @pid: pid of the process
3980 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3981 * @user_mask_ptr: user-space pointer to the new cpu mask
3983 * Return: 0 on success. An error code otherwise.
3985 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
3986 unsigned long __user
*, user_mask_ptr
)
3988 cpumask_var_t new_mask
;
3991 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3994 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
3996 retval
= sched_setaffinity(pid
, new_mask
);
3997 free_cpumask_var(new_mask
);
4001 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4003 struct task_struct
*p
;
4004 unsigned long flags
;
4010 p
= find_process_by_pid(pid
);
4014 retval
= security_task_getscheduler(p
);
4018 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4019 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4020 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4029 * sys_sched_getaffinity - get the cpu affinity of a process
4030 * @pid: pid of the process
4031 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4032 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4034 * Return: 0 on success. An error code otherwise.
4036 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4037 unsigned long __user
*, user_mask_ptr
)
4042 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4044 if (len
& (sizeof(unsigned long)-1))
4047 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4050 ret
= sched_getaffinity(pid
, mask
);
4052 size_t retlen
= min_t(size_t, len
, cpumask_size());
4054 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4059 free_cpumask_var(mask
);
4065 * sys_sched_yield - yield the current processor to other threads.
4067 * This function yields the current CPU to other tasks. If there are no
4068 * other threads running on this CPU then this function will return.
4072 SYSCALL_DEFINE0(sched_yield
)
4074 struct rq
*rq
= this_rq_lock();
4076 schedstat_inc(rq
, yld_count
);
4077 current
->sched_class
->yield_task(rq
);
4080 * Since we are going to call schedule() anyway, there's
4081 * no need to preempt or enable interrupts:
4083 __release(rq
->lock
);
4084 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4085 do_raw_spin_unlock(&rq
->lock
);
4086 sched_preempt_enable_no_resched();
4093 static void __cond_resched(void)
4095 __preempt_count_add(PREEMPT_ACTIVE
);
4097 __preempt_count_sub(PREEMPT_ACTIVE
);
4100 int __sched
_cond_resched(void)
4102 if (should_resched()) {
4108 EXPORT_SYMBOL(_cond_resched
);
4111 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4112 * call schedule, and on return reacquire the lock.
4114 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4115 * operations here to prevent schedule() from being called twice (once via
4116 * spin_unlock(), once by hand).
4118 int __cond_resched_lock(spinlock_t
*lock
)
4120 int resched
= should_resched();
4123 lockdep_assert_held(lock
);
4125 if (spin_needbreak(lock
) || resched
) {
4136 EXPORT_SYMBOL(__cond_resched_lock
);
4138 int __sched
__cond_resched_softirq(void)
4140 BUG_ON(!in_softirq());
4142 if (should_resched()) {
4150 EXPORT_SYMBOL(__cond_resched_softirq
);
4153 * yield - yield the current processor to other threads.
4155 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4157 * The scheduler is at all times free to pick the calling task as the most
4158 * eligible task to run, if removing the yield() call from your code breaks
4159 * it, its already broken.
4161 * Typical broken usage is:
4166 * where one assumes that yield() will let 'the other' process run that will
4167 * make event true. If the current task is a SCHED_FIFO task that will never
4168 * happen. Never use yield() as a progress guarantee!!
4170 * If you want to use yield() to wait for something, use wait_event().
4171 * If you want to use yield() to be 'nice' for others, use cond_resched().
4172 * If you still want to use yield(), do not!
4174 void __sched
yield(void)
4176 set_current_state(TASK_RUNNING
);
4179 EXPORT_SYMBOL(yield
);
4182 * yield_to - yield the current processor to another thread in
4183 * your thread group, or accelerate that thread toward the
4184 * processor it's on.
4186 * @preempt: whether task preemption is allowed or not
4188 * It's the caller's job to ensure that the target task struct
4189 * can't go away on us before we can do any checks.
4192 * true (>0) if we indeed boosted the target task.
4193 * false (0) if we failed to boost the target.
4194 * -ESRCH if there's no task to yield to.
4196 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
4198 struct task_struct
*curr
= current
;
4199 struct rq
*rq
, *p_rq
;
4200 unsigned long flags
;
4203 local_irq_save(flags
);
4209 * If we're the only runnable task on the rq and target rq also
4210 * has only one task, there's absolutely no point in yielding.
4212 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4217 double_rq_lock(rq
, p_rq
);
4218 if (task_rq(p
) != p_rq
) {
4219 double_rq_unlock(rq
, p_rq
);
4223 if (!curr
->sched_class
->yield_to_task
)
4226 if (curr
->sched_class
!= p
->sched_class
)
4229 if (task_running(p_rq
, p
) || p
->state
)
4232 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4234 schedstat_inc(rq
, yld_count
);
4236 * Make p's CPU reschedule; pick_next_entity takes care of
4239 if (preempt
&& rq
!= p_rq
)
4240 resched_task(p_rq
->curr
);
4244 double_rq_unlock(rq
, p_rq
);
4246 local_irq_restore(flags
);
4253 EXPORT_SYMBOL_GPL(yield_to
);
4256 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4257 * that process accounting knows that this is a task in IO wait state.
4259 void __sched
io_schedule(void)
4261 struct rq
*rq
= raw_rq();
4263 delayacct_blkio_start();
4264 atomic_inc(&rq
->nr_iowait
);
4265 blk_flush_plug(current
);
4266 current
->in_iowait
= 1;
4268 current
->in_iowait
= 0;
4269 atomic_dec(&rq
->nr_iowait
);
4270 delayacct_blkio_end();
4272 EXPORT_SYMBOL(io_schedule
);
4274 long __sched
io_schedule_timeout(long timeout
)
4276 struct rq
*rq
= raw_rq();
4279 delayacct_blkio_start();
4280 atomic_inc(&rq
->nr_iowait
);
4281 blk_flush_plug(current
);
4282 current
->in_iowait
= 1;
4283 ret
= schedule_timeout(timeout
);
4284 current
->in_iowait
= 0;
4285 atomic_dec(&rq
->nr_iowait
);
4286 delayacct_blkio_end();
4291 * sys_sched_get_priority_max - return maximum RT priority.
4292 * @policy: scheduling class.
4294 * Return: On success, this syscall returns the maximum
4295 * rt_priority that can be used by a given scheduling class.
4296 * On failure, a negative error code is returned.
4298 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4305 ret
= MAX_USER_RT_PRIO
-1;
4307 case SCHED_DEADLINE
:
4318 * sys_sched_get_priority_min - return minimum RT priority.
4319 * @policy: scheduling class.
4321 * Return: On success, this syscall returns the minimum
4322 * rt_priority that can be used by a given scheduling class.
4323 * On failure, a negative error code is returned.
4325 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4334 case SCHED_DEADLINE
:
4344 * sys_sched_rr_get_interval - return the default timeslice of a process.
4345 * @pid: pid of the process.
4346 * @interval: userspace pointer to the timeslice value.
4348 * this syscall writes the default timeslice value of a given process
4349 * into the user-space timespec buffer. A value of '0' means infinity.
4351 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4354 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4355 struct timespec __user
*, interval
)
4357 struct task_struct
*p
;
4358 unsigned int time_slice
;
4359 unsigned long flags
;
4369 p
= find_process_by_pid(pid
);
4373 retval
= security_task_getscheduler(p
);
4377 rq
= task_rq_lock(p
, &flags
);
4379 if (p
->sched_class
->get_rr_interval
)
4380 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4381 task_rq_unlock(rq
, p
, &flags
);
4384 jiffies_to_timespec(time_slice
, &t
);
4385 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4393 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4395 void sched_show_task(struct task_struct
*p
)
4397 unsigned long free
= 0;
4401 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4402 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4403 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4404 #if BITS_PER_LONG == 32
4405 if (state
== TASK_RUNNING
)
4406 printk(KERN_CONT
" running ");
4408 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4410 if (state
== TASK_RUNNING
)
4411 printk(KERN_CONT
" running task ");
4413 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4415 #ifdef CONFIG_DEBUG_STACK_USAGE
4416 free
= stack_not_used(p
);
4419 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4421 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4422 task_pid_nr(p
), ppid
,
4423 (unsigned long)task_thread_info(p
)->flags
);
4425 print_worker_info(KERN_INFO
, p
);
4426 show_stack(p
, NULL
);
4429 void show_state_filter(unsigned long state_filter
)
4431 struct task_struct
*g
, *p
;
4433 #if BITS_PER_LONG == 32
4435 " task PC stack pid father\n");
4438 " task PC stack pid father\n");
4441 do_each_thread(g
, p
) {
4443 * reset the NMI-timeout, listing all files on a slow
4444 * console might take a lot of time:
4446 touch_nmi_watchdog();
4447 if (!state_filter
|| (p
->state
& state_filter
))
4449 } while_each_thread(g
, p
);
4451 touch_all_softlockup_watchdogs();
4453 #ifdef CONFIG_SCHED_DEBUG
4454 sysrq_sched_debug_show();
4458 * Only show locks if all tasks are dumped:
4461 debug_show_all_locks();
4464 void init_idle_bootup_task(struct task_struct
*idle
)
4466 idle
->sched_class
= &idle_sched_class
;
4470 * init_idle - set up an idle thread for a given CPU
4471 * @idle: task in question
4472 * @cpu: cpu the idle task belongs to
4474 * NOTE: this function does not set the idle thread's NEED_RESCHED
4475 * flag, to make booting more robust.
4477 void init_idle(struct task_struct
*idle
, int cpu
)
4479 struct rq
*rq
= cpu_rq(cpu
);
4480 unsigned long flags
;
4482 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4484 __sched_fork(0, idle
);
4485 idle
->state
= TASK_RUNNING
;
4486 idle
->se
.exec_start
= sched_clock();
4488 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4490 * We're having a chicken and egg problem, even though we are
4491 * holding rq->lock, the cpu isn't yet set to this cpu so the
4492 * lockdep check in task_group() will fail.
4494 * Similar case to sched_fork(). / Alternatively we could
4495 * use task_rq_lock() here and obtain the other rq->lock.
4500 __set_task_cpu(idle
, cpu
);
4503 rq
->curr
= rq
->idle
= idle
;
4505 #if defined(CONFIG_SMP)
4508 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4510 /* Set the preempt count _outside_ the spinlocks! */
4511 init_idle_preempt_count(idle
, cpu
);
4514 * The idle tasks have their own, simple scheduling class:
4516 idle
->sched_class
= &idle_sched_class
;
4517 ftrace_graph_init_idle_task(idle
, cpu
);
4518 vtime_init_idle(idle
, cpu
);
4519 #if defined(CONFIG_SMP)
4520 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4525 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4527 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
4528 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4530 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4531 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4535 * This is how migration works:
4537 * 1) we invoke migration_cpu_stop() on the target CPU using
4539 * 2) stopper starts to run (implicitly forcing the migrated thread
4541 * 3) it checks whether the migrated task is still in the wrong runqueue.
4542 * 4) if it's in the wrong runqueue then the migration thread removes
4543 * it and puts it into the right queue.
4544 * 5) stopper completes and stop_one_cpu() returns and the migration
4549 * Change a given task's CPU affinity. Migrate the thread to a
4550 * proper CPU and schedule it away if the CPU it's executing on
4551 * is removed from the allowed bitmask.
4553 * NOTE: the caller must have a valid reference to the task, the
4554 * task must not exit() & deallocate itself prematurely. The
4555 * call is not atomic; no spinlocks may be held.
4557 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4559 unsigned long flags
;
4561 unsigned int dest_cpu
;
4564 rq
= task_rq_lock(p
, &flags
);
4566 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4569 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4574 do_set_cpus_allowed(p
, new_mask
);
4576 /* Can the task run on the task's current CPU? If so, we're done */
4577 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4580 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4582 struct migration_arg arg
= { p
, dest_cpu
};
4583 /* Need help from migration thread: drop lock and wait. */
4584 task_rq_unlock(rq
, p
, &flags
);
4585 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4586 tlb_migrate_finish(p
->mm
);
4590 task_rq_unlock(rq
, p
, &flags
);
4594 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4597 * Move (not current) task off this cpu, onto dest cpu. We're doing
4598 * this because either it can't run here any more (set_cpus_allowed()
4599 * away from this CPU, or CPU going down), or because we're
4600 * attempting to rebalance this task on exec (sched_exec).
4602 * So we race with normal scheduler movements, but that's OK, as long
4603 * as the task is no longer on this CPU.
4605 * Returns non-zero if task was successfully migrated.
4607 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4609 struct rq
*rq_dest
, *rq_src
;
4612 if (unlikely(!cpu_active(dest_cpu
)))
4615 rq_src
= cpu_rq(src_cpu
);
4616 rq_dest
= cpu_rq(dest_cpu
);
4618 raw_spin_lock(&p
->pi_lock
);
4619 double_rq_lock(rq_src
, rq_dest
);
4620 /* Already moved. */
4621 if (task_cpu(p
) != src_cpu
)
4623 /* Affinity changed (again). */
4624 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4628 * If we're not on a rq, the next wake-up will ensure we're
4632 dequeue_task(rq_src
, p
, 0);
4633 set_task_cpu(p
, dest_cpu
);
4634 enqueue_task(rq_dest
, p
, 0);
4635 check_preempt_curr(rq_dest
, p
, 0);
4640 double_rq_unlock(rq_src
, rq_dest
);
4641 raw_spin_unlock(&p
->pi_lock
);
4645 #ifdef CONFIG_NUMA_BALANCING
4646 /* Migrate current task p to target_cpu */
4647 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4649 struct migration_arg arg
= { p
, target_cpu
};
4650 int curr_cpu
= task_cpu(p
);
4652 if (curr_cpu
== target_cpu
)
4655 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4658 /* TODO: This is not properly updating schedstats */
4660 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4661 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4665 * Requeue a task on a given node and accurately track the number of NUMA
4666 * tasks on the runqueues
4668 void sched_setnuma(struct task_struct
*p
, int nid
)
4671 unsigned long flags
;
4672 bool on_rq
, running
;
4674 rq
= task_rq_lock(p
, &flags
);
4676 running
= task_current(rq
, p
);
4679 dequeue_task(rq
, p
, 0);
4681 p
->sched_class
->put_prev_task(rq
, p
);
4683 p
->numa_preferred_nid
= nid
;
4686 p
->sched_class
->set_curr_task(rq
);
4688 enqueue_task(rq
, p
, 0);
4689 task_rq_unlock(rq
, p
, &flags
);
4694 * migration_cpu_stop - this will be executed by a highprio stopper thread
4695 * and performs thread migration by bumping thread off CPU then
4696 * 'pushing' onto another runqueue.
4698 static int migration_cpu_stop(void *data
)
4700 struct migration_arg
*arg
= data
;
4703 * The original target cpu might have gone down and we might
4704 * be on another cpu but it doesn't matter.
4706 local_irq_disable();
4707 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4712 #ifdef CONFIG_HOTPLUG_CPU
4715 * Ensures that the idle task is using init_mm right before its cpu goes
4718 void idle_task_exit(void)
4720 struct mm_struct
*mm
= current
->active_mm
;
4722 BUG_ON(cpu_online(smp_processor_id()));
4724 if (mm
!= &init_mm
) {
4725 switch_mm(mm
, &init_mm
, current
);
4726 finish_arch_post_lock_switch();
4732 * Since this CPU is going 'away' for a while, fold any nr_active delta
4733 * we might have. Assumes we're called after migrate_tasks() so that the
4734 * nr_active count is stable.
4736 * Also see the comment "Global load-average calculations".
4738 static void calc_load_migrate(struct rq
*rq
)
4740 long delta
= calc_load_fold_active(rq
);
4742 atomic_long_add(delta
, &calc_load_tasks
);
4745 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4749 static const struct sched_class fake_sched_class
= {
4750 .put_prev_task
= put_prev_task_fake
,
4753 static struct task_struct fake_task
= {
4755 * Avoid pull_{rt,dl}_task()
4757 .prio
= MAX_PRIO
+ 1,
4758 .sched_class
= &fake_sched_class
,
4762 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4763 * try_to_wake_up()->select_task_rq().
4765 * Called with rq->lock held even though we'er in stop_machine() and
4766 * there's no concurrency possible, we hold the required locks anyway
4767 * because of lock validation efforts.
4769 static void migrate_tasks(unsigned int dead_cpu
)
4771 struct rq
*rq
= cpu_rq(dead_cpu
);
4772 struct task_struct
*next
, *stop
= rq
->stop
;
4776 * Fudge the rq selection such that the below task selection loop
4777 * doesn't get stuck on the currently eligible stop task.
4779 * We're currently inside stop_machine() and the rq is either stuck
4780 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4781 * either way we should never end up calling schedule() until we're
4787 * put_prev_task() and pick_next_task() sched
4788 * class method both need to have an up-to-date
4789 * value of rq->clock[_task]
4791 update_rq_clock(rq
);
4795 * There's this thread running, bail when that's the only
4798 if (rq
->nr_running
== 1)
4801 next
= pick_next_task(rq
, &fake_task
);
4803 next
->sched_class
->put_prev_task(rq
, next
);
4805 /* Find suitable destination for @next, with force if needed. */
4806 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
4807 raw_spin_unlock(&rq
->lock
);
4809 __migrate_task(next
, dead_cpu
, dest_cpu
);
4811 raw_spin_lock(&rq
->lock
);
4817 #endif /* CONFIG_HOTPLUG_CPU */
4819 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4821 static struct ctl_table sd_ctl_dir
[] = {
4823 .procname
= "sched_domain",
4829 static struct ctl_table sd_ctl_root
[] = {
4831 .procname
= "kernel",
4833 .child
= sd_ctl_dir
,
4838 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
4840 struct ctl_table
*entry
=
4841 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
4846 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
4848 struct ctl_table
*entry
;
4851 * In the intermediate directories, both the child directory and
4852 * procname are dynamically allocated and could fail but the mode
4853 * will always be set. In the lowest directory the names are
4854 * static strings and all have proc handlers.
4856 for (entry
= *tablep
; entry
->mode
; entry
++) {
4858 sd_free_ctl_entry(&entry
->child
);
4859 if (entry
->proc_handler
== NULL
)
4860 kfree(entry
->procname
);
4867 static int min_load_idx
= 0;
4868 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
4871 set_table_entry(struct ctl_table
*entry
,
4872 const char *procname
, void *data
, int maxlen
,
4873 umode_t mode
, proc_handler
*proc_handler
,
4876 entry
->procname
= procname
;
4878 entry
->maxlen
= maxlen
;
4880 entry
->proc_handler
= proc_handler
;
4883 entry
->extra1
= &min_load_idx
;
4884 entry
->extra2
= &max_load_idx
;
4888 static struct ctl_table
*
4889 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
4891 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
4896 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
4897 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4898 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
4899 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4900 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
4901 sizeof(int), 0644, proc_dointvec_minmax
, true);
4902 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
4903 sizeof(int), 0644, proc_dointvec_minmax
, true);
4904 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
4905 sizeof(int), 0644, proc_dointvec_minmax
, true);
4906 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
4907 sizeof(int), 0644, proc_dointvec_minmax
, true);
4908 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
4909 sizeof(int), 0644, proc_dointvec_minmax
, true);
4910 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
4911 sizeof(int), 0644, proc_dointvec_minmax
, false);
4912 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
4913 sizeof(int), 0644, proc_dointvec_minmax
, false);
4914 set_table_entry(&table
[9], "cache_nice_tries",
4915 &sd
->cache_nice_tries
,
4916 sizeof(int), 0644, proc_dointvec_minmax
, false);
4917 set_table_entry(&table
[10], "flags", &sd
->flags
,
4918 sizeof(int), 0644, proc_dointvec_minmax
, false);
4919 set_table_entry(&table
[11], "max_newidle_lb_cost",
4920 &sd
->max_newidle_lb_cost
,
4921 sizeof(long), 0644, proc_doulongvec_minmax
, false);
4922 set_table_entry(&table
[12], "name", sd
->name
,
4923 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
4924 /* &table[13] is terminator */
4929 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
4931 struct ctl_table
*entry
, *table
;
4932 struct sched_domain
*sd
;
4933 int domain_num
= 0, i
;
4936 for_each_domain(cpu
, sd
)
4938 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
4943 for_each_domain(cpu
, sd
) {
4944 snprintf(buf
, 32, "domain%d", i
);
4945 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4947 entry
->child
= sd_alloc_ctl_domain_table(sd
);
4954 static struct ctl_table_header
*sd_sysctl_header
;
4955 static void register_sched_domain_sysctl(void)
4957 int i
, cpu_num
= num_possible_cpus();
4958 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
4961 WARN_ON(sd_ctl_dir
[0].child
);
4962 sd_ctl_dir
[0].child
= entry
;
4967 for_each_possible_cpu(i
) {
4968 snprintf(buf
, 32, "cpu%d", i
);
4969 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
4971 entry
->child
= sd_alloc_ctl_cpu_table(i
);
4975 WARN_ON(sd_sysctl_header
);
4976 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
4979 /* may be called multiple times per register */
4980 static void unregister_sched_domain_sysctl(void)
4982 if (sd_sysctl_header
)
4983 unregister_sysctl_table(sd_sysctl_header
);
4984 sd_sysctl_header
= NULL
;
4985 if (sd_ctl_dir
[0].child
)
4986 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
4989 static void register_sched_domain_sysctl(void)
4992 static void unregister_sched_domain_sysctl(void)
4997 static void set_rq_online(struct rq
*rq
)
5000 const struct sched_class
*class;
5002 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5005 for_each_class(class) {
5006 if (class->rq_online
)
5007 class->rq_online(rq
);
5012 static void set_rq_offline(struct rq
*rq
)
5015 const struct sched_class
*class;
5017 for_each_class(class) {
5018 if (class->rq_offline
)
5019 class->rq_offline(rq
);
5022 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5028 * migration_call - callback that gets triggered when a CPU is added.
5029 * Here we can start up the necessary migration thread for the new CPU.
5032 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5034 int cpu
= (long)hcpu
;
5035 unsigned long flags
;
5036 struct rq
*rq
= cpu_rq(cpu
);
5038 switch (action
& ~CPU_TASKS_FROZEN
) {
5040 case CPU_UP_PREPARE
:
5041 rq
->calc_load_update
= calc_load_update
;
5045 /* Update our root-domain */
5046 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5048 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5052 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5055 #ifdef CONFIG_HOTPLUG_CPU
5057 sched_ttwu_pending();
5058 /* Update our root-domain */
5059 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5061 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5065 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5066 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5070 calc_load_migrate(rq
);
5075 update_max_interval();
5081 * Register at high priority so that task migration (migrate_all_tasks)
5082 * happens before everything else. This has to be lower priority than
5083 * the notifier in the perf_event subsystem, though.
5085 static struct notifier_block migration_notifier
= {
5086 .notifier_call
= migration_call
,
5087 .priority
= CPU_PRI_MIGRATION
,
5090 static int sched_cpu_active(struct notifier_block
*nfb
,
5091 unsigned long action
, void *hcpu
)
5093 switch (action
& ~CPU_TASKS_FROZEN
) {
5095 case CPU_DOWN_FAILED
:
5096 set_cpu_active((long)hcpu
, true);
5103 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5104 unsigned long action
, void *hcpu
)
5106 unsigned long flags
;
5107 long cpu
= (long)hcpu
;
5109 switch (action
& ~CPU_TASKS_FROZEN
) {
5110 case CPU_DOWN_PREPARE
:
5111 set_cpu_active(cpu
, false);
5113 /* explicitly allow suspend */
5114 if (!(action
& CPU_TASKS_FROZEN
)) {
5115 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
5119 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5120 cpus
= dl_bw_cpus(cpu
);
5121 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5122 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5125 return notifier_from_errno(-EBUSY
);
5133 static int __init
migration_init(void)
5135 void *cpu
= (void *)(long)smp_processor_id();
5138 /* Initialize migration for the boot CPU */
5139 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5140 BUG_ON(err
== NOTIFY_BAD
);
5141 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5142 register_cpu_notifier(&migration_notifier
);
5144 /* Register cpu active notifiers */
5145 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5146 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5150 early_initcall(migration_init
);
5155 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5157 #ifdef CONFIG_SCHED_DEBUG
5159 static __read_mostly
int sched_debug_enabled
;
5161 static int __init
sched_debug_setup(char *str
)
5163 sched_debug_enabled
= 1;
5167 early_param("sched_debug", sched_debug_setup
);
5169 static inline bool sched_debug(void)
5171 return sched_debug_enabled
;
5174 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5175 struct cpumask
*groupmask
)
5177 struct sched_group
*group
= sd
->groups
;
5180 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5181 cpumask_clear(groupmask
);
5183 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5185 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5186 printk("does not load-balance\n");
5188 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5193 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5195 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5196 printk(KERN_ERR
"ERROR: domain->span does not contain "
5199 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5200 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5204 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5208 printk(KERN_ERR
"ERROR: group is NULL\n");
5213 * Even though we initialize ->power to something semi-sane,
5214 * we leave power_orig unset. This allows us to detect if
5215 * domain iteration is still funny without causing /0 traps.
5217 if (!group
->sgp
->power_orig
) {
5218 printk(KERN_CONT
"\n");
5219 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5224 if (!cpumask_weight(sched_group_cpus(group
))) {
5225 printk(KERN_CONT
"\n");
5226 printk(KERN_ERR
"ERROR: empty group\n");
5230 if (!(sd
->flags
& SD_OVERLAP
) &&
5231 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5232 printk(KERN_CONT
"\n");
5233 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5237 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5239 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5241 printk(KERN_CONT
" %s", str
);
5242 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
5243 printk(KERN_CONT
" (cpu_power = %d)",
5247 group
= group
->next
;
5248 } while (group
!= sd
->groups
);
5249 printk(KERN_CONT
"\n");
5251 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5252 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5255 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5256 printk(KERN_ERR
"ERROR: parent span is not a superset "
5257 "of domain->span\n");
5261 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5265 if (!sched_debug_enabled
)
5269 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5273 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5276 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5284 #else /* !CONFIG_SCHED_DEBUG */
5285 # define sched_domain_debug(sd, cpu) do { } while (0)
5286 static inline bool sched_debug(void)
5290 #endif /* CONFIG_SCHED_DEBUG */
5292 static int sd_degenerate(struct sched_domain
*sd
)
5294 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5297 /* Following flags need at least 2 groups */
5298 if (sd
->flags
& (SD_LOAD_BALANCE
|
5299 SD_BALANCE_NEWIDLE
|
5303 SD_SHARE_PKG_RESOURCES
)) {
5304 if (sd
->groups
!= sd
->groups
->next
)
5308 /* Following flags don't use groups */
5309 if (sd
->flags
& (SD_WAKE_AFFINE
))
5316 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5318 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5320 if (sd_degenerate(parent
))
5323 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5326 /* Flags needing groups don't count if only 1 group in parent */
5327 if (parent
->groups
== parent
->groups
->next
) {
5328 pflags
&= ~(SD_LOAD_BALANCE
|
5329 SD_BALANCE_NEWIDLE
|
5333 SD_SHARE_PKG_RESOURCES
|
5335 if (nr_node_ids
== 1)
5336 pflags
&= ~SD_SERIALIZE
;
5338 if (~cflags
& pflags
)
5344 static void free_rootdomain(struct rcu_head
*rcu
)
5346 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5348 cpupri_cleanup(&rd
->cpupri
);
5349 cpudl_cleanup(&rd
->cpudl
);
5350 free_cpumask_var(rd
->dlo_mask
);
5351 free_cpumask_var(rd
->rto_mask
);
5352 free_cpumask_var(rd
->online
);
5353 free_cpumask_var(rd
->span
);
5357 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5359 struct root_domain
*old_rd
= NULL
;
5360 unsigned long flags
;
5362 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5367 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5370 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5373 * If we dont want to free the old_rd yet then
5374 * set old_rd to NULL to skip the freeing later
5377 if (!atomic_dec_and_test(&old_rd
->refcount
))
5381 atomic_inc(&rd
->refcount
);
5384 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5385 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5388 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5391 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5394 static int init_rootdomain(struct root_domain
*rd
)
5396 memset(rd
, 0, sizeof(*rd
));
5398 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5400 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5402 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5404 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5407 init_dl_bw(&rd
->dl_bw
);
5408 if (cpudl_init(&rd
->cpudl
) != 0)
5411 if (cpupri_init(&rd
->cpupri
) != 0)
5416 free_cpumask_var(rd
->rto_mask
);
5418 free_cpumask_var(rd
->dlo_mask
);
5420 free_cpumask_var(rd
->online
);
5422 free_cpumask_var(rd
->span
);
5428 * By default the system creates a single root-domain with all cpus as
5429 * members (mimicking the global state we have today).
5431 struct root_domain def_root_domain
;
5433 static void init_defrootdomain(void)
5435 init_rootdomain(&def_root_domain
);
5437 atomic_set(&def_root_domain
.refcount
, 1);
5440 static struct root_domain
*alloc_rootdomain(void)
5442 struct root_domain
*rd
;
5444 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5448 if (init_rootdomain(rd
) != 0) {
5456 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
5458 struct sched_group
*tmp
, *first
;
5467 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
5472 } while (sg
!= first
);
5475 static void free_sched_domain(struct rcu_head
*rcu
)
5477 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5480 * If its an overlapping domain it has private groups, iterate and
5483 if (sd
->flags
& SD_OVERLAP
) {
5484 free_sched_groups(sd
->groups
, 1);
5485 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5486 kfree(sd
->groups
->sgp
);
5492 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5494 call_rcu(&sd
->rcu
, free_sched_domain
);
5497 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5499 for (; sd
; sd
= sd
->parent
)
5500 destroy_sched_domain(sd
, cpu
);
5504 * Keep a special pointer to the highest sched_domain that has
5505 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5506 * allows us to avoid some pointer chasing select_idle_sibling().
5508 * Also keep a unique ID per domain (we use the first cpu number in
5509 * the cpumask of the domain), this allows us to quickly tell if
5510 * two cpus are in the same cache domain, see cpus_share_cache().
5512 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5513 DEFINE_PER_CPU(int, sd_llc_size
);
5514 DEFINE_PER_CPU(int, sd_llc_id
);
5515 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5516 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5517 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5519 static void update_top_cache_domain(int cpu
)
5521 struct sched_domain
*sd
;
5522 struct sched_domain
*busy_sd
= NULL
;
5526 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5528 id
= cpumask_first(sched_domain_span(sd
));
5529 size
= cpumask_weight(sched_domain_span(sd
));
5530 busy_sd
= sd
->parent
; /* sd_busy */
5532 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5534 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5535 per_cpu(sd_llc_size
, cpu
) = size
;
5536 per_cpu(sd_llc_id
, cpu
) = id
;
5538 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5539 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5541 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5542 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5546 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5547 * hold the hotplug lock.
5550 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5552 struct rq
*rq
= cpu_rq(cpu
);
5553 struct sched_domain
*tmp
;
5555 /* Remove the sched domains which do not contribute to scheduling. */
5556 for (tmp
= sd
; tmp
; ) {
5557 struct sched_domain
*parent
= tmp
->parent
;
5561 if (sd_parent_degenerate(tmp
, parent
)) {
5562 tmp
->parent
= parent
->parent
;
5564 parent
->parent
->child
= tmp
;
5566 * Transfer SD_PREFER_SIBLING down in case of a
5567 * degenerate parent; the spans match for this
5568 * so the property transfers.
5570 if (parent
->flags
& SD_PREFER_SIBLING
)
5571 tmp
->flags
|= SD_PREFER_SIBLING
;
5572 destroy_sched_domain(parent
, cpu
);
5577 if (sd
&& sd_degenerate(sd
)) {
5580 destroy_sched_domain(tmp
, cpu
);
5585 sched_domain_debug(sd
, cpu
);
5587 rq_attach_root(rq
, rd
);
5589 rcu_assign_pointer(rq
->sd
, sd
);
5590 destroy_sched_domains(tmp
, cpu
);
5592 update_top_cache_domain(cpu
);
5595 /* cpus with isolated domains */
5596 static cpumask_var_t cpu_isolated_map
;
5598 /* Setup the mask of cpus configured for isolated domains */
5599 static int __init
isolated_cpu_setup(char *str
)
5601 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5602 cpulist_parse(str
, cpu_isolated_map
);
5606 __setup("isolcpus=", isolated_cpu_setup
);
5608 static const struct cpumask
*cpu_cpu_mask(int cpu
)
5610 return cpumask_of_node(cpu_to_node(cpu
));
5614 struct sched_domain
**__percpu sd
;
5615 struct sched_group
**__percpu sg
;
5616 struct sched_group_power
**__percpu sgp
;
5620 struct sched_domain
** __percpu sd
;
5621 struct root_domain
*rd
;
5631 struct sched_domain_topology_level
;
5633 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
5634 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
5636 #define SDTL_OVERLAP 0x01
5638 struct sched_domain_topology_level
{
5639 sched_domain_init_f init
;
5640 sched_domain_mask_f mask
;
5643 struct sd_data data
;
5647 * Build an iteration mask that can exclude certain CPUs from the upwards
5650 * Asymmetric node setups can result in situations where the domain tree is of
5651 * unequal depth, make sure to skip domains that already cover the entire
5654 * In that case build_sched_domains() will have terminated the iteration early
5655 * and our sibling sd spans will be empty. Domains should always include the
5656 * cpu they're built on, so check that.
5659 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5661 const struct cpumask
*span
= sched_domain_span(sd
);
5662 struct sd_data
*sdd
= sd
->private;
5663 struct sched_domain
*sibling
;
5666 for_each_cpu(i
, span
) {
5667 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5668 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5671 cpumask_set_cpu(i
, sched_group_mask(sg
));
5676 * Return the canonical balance cpu for this group, this is the first cpu
5677 * of this group that's also in the iteration mask.
5679 int group_balance_cpu(struct sched_group
*sg
)
5681 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5685 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5687 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5688 const struct cpumask
*span
= sched_domain_span(sd
);
5689 struct cpumask
*covered
= sched_domains_tmpmask
;
5690 struct sd_data
*sdd
= sd
->private;
5691 struct sched_domain
*child
;
5694 cpumask_clear(covered
);
5696 for_each_cpu(i
, span
) {
5697 struct cpumask
*sg_span
;
5699 if (cpumask_test_cpu(i
, covered
))
5702 child
= *per_cpu_ptr(sdd
->sd
, i
);
5704 /* See the comment near build_group_mask(). */
5705 if (!cpumask_test_cpu(i
, sched_domain_span(child
)))
5708 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5709 GFP_KERNEL
, cpu_to_node(cpu
));
5714 sg_span
= sched_group_cpus(sg
);
5716 child
= child
->child
;
5717 cpumask_copy(sg_span
, sched_domain_span(child
));
5719 cpumask_set_cpu(i
, sg_span
);
5721 cpumask_or(covered
, covered
, sg_span
);
5723 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, i
);
5724 if (atomic_inc_return(&sg
->sgp
->ref
) == 1)
5725 build_group_mask(sd
, sg
);
5728 * Initialize sgp->power such that even if we mess up the
5729 * domains and no possible iteration will get us here, we won't
5732 sg
->sgp
->power
= SCHED_POWER_SCALE
* cpumask_weight(sg_span
);
5733 sg
->sgp
->power_orig
= sg
->sgp
->power
;
5736 * Make sure the first group of this domain contains the
5737 * canonical balance cpu. Otherwise the sched_domain iteration
5738 * breaks. See update_sg_lb_stats().
5740 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5741 group_balance_cpu(sg
) == cpu
)
5751 sd
->groups
= groups
;
5756 free_sched_groups(first
, 0);
5761 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5763 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5764 struct sched_domain
*child
= sd
->child
;
5767 cpu
= cpumask_first(sched_domain_span(child
));
5770 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5771 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
5772 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
5779 * build_sched_groups will build a circular linked list of the groups
5780 * covered by the given span, and will set each group's ->cpumask correctly,
5781 * and ->cpu_power to 0.
5783 * Assumes the sched_domain tree is fully constructed
5786 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5788 struct sched_group
*first
= NULL
, *last
= NULL
;
5789 struct sd_data
*sdd
= sd
->private;
5790 const struct cpumask
*span
= sched_domain_span(sd
);
5791 struct cpumask
*covered
;
5794 get_group(cpu
, sdd
, &sd
->groups
);
5795 atomic_inc(&sd
->groups
->ref
);
5797 if (cpu
!= cpumask_first(span
))
5800 lockdep_assert_held(&sched_domains_mutex
);
5801 covered
= sched_domains_tmpmask
;
5803 cpumask_clear(covered
);
5805 for_each_cpu(i
, span
) {
5806 struct sched_group
*sg
;
5809 if (cpumask_test_cpu(i
, covered
))
5812 group
= get_group(i
, sdd
, &sg
);
5813 cpumask_clear(sched_group_cpus(sg
));
5815 cpumask_setall(sched_group_mask(sg
));
5817 for_each_cpu(j
, span
) {
5818 if (get_group(j
, sdd
, NULL
) != group
)
5821 cpumask_set_cpu(j
, covered
);
5822 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5837 * Initialize sched groups cpu_power.
5839 * cpu_power indicates the capacity of sched group, which is used while
5840 * distributing the load between different sched groups in a sched domain.
5841 * Typically cpu_power for all the groups in a sched domain will be same unless
5842 * there are asymmetries in the topology. If there are asymmetries, group
5843 * having more cpu_power will pickup more load compared to the group having
5846 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
5848 struct sched_group
*sg
= sd
->groups
;
5853 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
5855 } while (sg
!= sd
->groups
);
5857 if (cpu
!= group_balance_cpu(sg
))
5860 update_group_power(sd
, cpu
);
5861 atomic_set(&sg
->sgp
->nr_busy_cpus
, sg
->group_weight
);
5864 int __weak
arch_sd_sibling_asym_packing(void)
5866 return 0*SD_ASYM_PACKING
;
5870 * Initializers for schedule domains
5871 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5874 #ifdef CONFIG_SCHED_DEBUG
5875 # define SD_INIT_NAME(sd, type) sd->name = #type
5877 # define SD_INIT_NAME(sd, type) do { } while (0)
5880 #define SD_INIT_FUNC(type) \
5881 static noinline struct sched_domain * \
5882 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5884 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5885 *sd = SD_##type##_INIT; \
5886 SD_INIT_NAME(sd, type); \
5887 sd->private = &tl->data; \
5892 #ifdef CONFIG_SCHED_SMT
5893 SD_INIT_FUNC(SIBLING
)
5895 #ifdef CONFIG_SCHED_MC
5898 #ifdef CONFIG_SCHED_BOOK
5902 static int default_relax_domain_level
= -1;
5903 int sched_domain_level_max
;
5905 static int __init
setup_relax_domain_level(char *str
)
5907 if (kstrtoint(str
, 0, &default_relax_domain_level
))
5908 pr_warn("Unable to set relax_domain_level\n");
5912 __setup("relax_domain_level=", setup_relax_domain_level
);
5914 static void set_domain_attribute(struct sched_domain
*sd
,
5915 struct sched_domain_attr
*attr
)
5919 if (!attr
|| attr
->relax_domain_level
< 0) {
5920 if (default_relax_domain_level
< 0)
5923 request
= default_relax_domain_level
;
5925 request
= attr
->relax_domain_level
;
5926 if (request
< sd
->level
) {
5927 /* turn off idle balance on this domain */
5928 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5930 /* turn on idle balance on this domain */
5931 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
5935 static void __sdt_free(const struct cpumask
*cpu_map
);
5936 static int __sdt_alloc(const struct cpumask
*cpu_map
);
5938 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
5939 const struct cpumask
*cpu_map
)
5943 if (!atomic_read(&d
->rd
->refcount
))
5944 free_rootdomain(&d
->rd
->rcu
); /* fall through */
5946 free_percpu(d
->sd
); /* fall through */
5948 __sdt_free(cpu_map
); /* fall through */
5954 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
5955 const struct cpumask
*cpu_map
)
5957 memset(d
, 0, sizeof(*d
));
5959 if (__sdt_alloc(cpu_map
))
5960 return sa_sd_storage
;
5961 d
->sd
= alloc_percpu(struct sched_domain
*);
5963 return sa_sd_storage
;
5964 d
->rd
= alloc_rootdomain();
5967 return sa_rootdomain
;
5971 * NULL the sd_data elements we've used to build the sched_domain and
5972 * sched_group structure so that the subsequent __free_domain_allocs()
5973 * will not free the data we're using.
5975 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
5977 struct sd_data
*sdd
= sd
->private;
5979 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
5980 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
5982 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
5983 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
5985 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
5986 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
5989 #ifdef CONFIG_SCHED_SMT
5990 static const struct cpumask
*cpu_smt_mask(int cpu
)
5992 return topology_thread_cpumask(cpu
);
5997 * Topology list, bottom-up.
5999 static struct sched_domain_topology_level default_topology
[] = {
6000 #ifdef CONFIG_SCHED_SMT
6001 { sd_init_SIBLING
, cpu_smt_mask
, },
6003 #ifdef CONFIG_SCHED_MC
6004 { sd_init_MC
, cpu_coregroup_mask
, },
6006 #ifdef CONFIG_SCHED_BOOK
6007 { sd_init_BOOK
, cpu_book_mask
, },
6009 { sd_init_CPU
, cpu_cpu_mask
, },
6013 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6015 #define for_each_sd_topology(tl) \
6016 for (tl = sched_domain_topology; tl->init; tl++)
6020 static int sched_domains_numa_levels
;
6021 static int *sched_domains_numa_distance
;
6022 static struct cpumask
***sched_domains_numa_masks
;
6023 static int sched_domains_curr_level
;
6025 static inline int sd_local_flags(int level
)
6027 if (sched_domains_numa_distance
[level
] > RECLAIM_DISTANCE
)
6030 return SD_BALANCE_EXEC
| SD_BALANCE_FORK
| SD_WAKE_AFFINE
;
6033 static struct sched_domain
*
6034 sd_numa_init(struct sched_domain_topology_level
*tl
, int cpu
)
6036 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6037 int level
= tl
->numa_level
;
6038 int sd_weight
= cpumask_weight(
6039 sched_domains_numa_masks
[level
][cpu_to_node(cpu
)]);
6041 *sd
= (struct sched_domain
){
6042 .min_interval
= sd_weight
,
6043 .max_interval
= 2*sd_weight
,
6045 .imbalance_pct
= 125,
6046 .cache_nice_tries
= 2,
6053 .flags
= 1*SD_LOAD_BALANCE
6054 | 1*SD_BALANCE_NEWIDLE
6059 | 0*SD_SHARE_CPUPOWER
6060 | 0*SD_SHARE_PKG_RESOURCES
6062 | 0*SD_PREFER_SIBLING
6064 | sd_local_flags(level
)
6066 .last_balance
= jiffies
,
6067 .balance_interval
= sd_weight
,
6069 SD_INIT_NAME(sd
, NUMA
);
6070 sd
->private = &tl
->data
;
6073 * Ugly hack to pass state to sd_numa_mask()...
6075 sched_domains_curr_level
= tl
->numa_level
;
6080 static const struct cpumask
*sd_numa_mask(int cpu
)
6082 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6085 static void sched_numa_warn(const char *str
)
6087 static int done
= false;
6095 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6097 for (i
= 0; i
< nr_node_ids
; i
++) {
6098 printk(KERN_WARNING
" ");
6099 for (j
= 0; j
< nr_node_ids
; j
++)
6100 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6101 printk(KERN_CONT
"\n");
6103 printk(KERN_WARNING
"\n");
6106 static bool find_numa_distance(int distance
)
6110 if (distance
== node_distance(0, 0))
6113 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6114 if (sched_domains_numa_distance
[i
] == distance
)
6121 static void sched_init_numa(void)
6123 int next_distance
, curr_distance
= node_distance(0, 0);
6124 struct sched_domain_topology_level
*tl
;
6128 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6129 if (!sched_domains_numa_distance
)
6133 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6134 * unique distances in the node_distance() table.
6136 * Assumes node_distance(0,j) includes all distances in
6137 * node_distance(i,j) in order to avoid cubic time.
6139 next_distance
= curr_distance
;
6140 for (i
= 0; i
< nr_node_ids
; i
++) {
6141 for (j
= 0; j
< nr_node_ids
; j
++) {
6142 for (k
= 0; k
< nr_node_ids
; k
++) {
6143 int distance
= node_distance(i
, k
);
6145 if (distance
> curr_distance
&&
6146 (distance
< next_distance
||
6147 next_distance
== curr_distance
))
6148 next_distance
= distance
;
6151 * While not a strong assumption it would be nice to know
6152 * about cases where if node A is connected to B, B is not
6153 * equally connected to A.
6155 if (sched_debug() && node_distance(k
, i
) != distance
)
6156 sched_numa_warn("Node-distance not symmetric");
6158 if (sched_debug() && i
&& !find_numa_distance(distance
))
6159 sched_numa_warn("Node-0 not representative");
6161 if (next_distance
!= curr_distance
) {
6162 sched_domains_numa_distance
[level
++] = next_distance
;
6163 sched_domains_numa_levels
= level
;
6164 curr_distance
= next_distance
;
6169 * In case of sched_debug() we verify the above assumption.
6175 * 'level' contains the number of unique distances, excluding the
6176 * identity distance node_distance(i,i).
6178 * The sched_domains_numa_distance[] array includes the actual distance
6183 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6184 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6185 * the array will contain less then 'level' members. This could be
6186 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6187 * in other functions.
6189 * We reset it to 'level' at the end of this function.
6191 sched_domains_numa_levels
= 0;
6193 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6194 if (!sched_domains_numa_masks
)
6198 * Now for each level, construct a mask per node which contains all
6199 * cpus of nodes that are that many hops away from us.
6201 for (i
= 0; i
< level
; i
++) {
6202 sched_domains_numa_masks
[i
] =
6203 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6204 if (!sched_domains_numa_masks
[i
])
6207 for (j
= 0; j
< nr_node_ids
; j
++) {
6208 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6212 sched_domains_numa_masks
[i
][j
] = mask
;
6214 for (k
= 0; k
< nr_node_ids
; k
++) {
6215 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6218 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6223 tl
= kzalloc((ARRAY_SIZE(default_topology
) + level
) *
6224 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6229 * Copy the default topology bits..
6231 for (i
= 0; default_topology
[i
].init
; i
++)
6232 tl
[i
] = default_topology
[i
];
6235 * .. and append 'j' levels of NUMA goodness.
6237 for (j
= 0; j
< level
; i
++, j
++) {
6238 tl
[i
] = (struct sched_domain_topology_level
){
6239 .init
= sd_numa_init
,
6240 .mask
= sd_numa_mask
,
6241 .flags
= SDTL_OVERLAP
,
6246 sched_domain_topology
= tl
;
6248 sched_domains_numa_levels
= level
;
6251 static void sched_domains_numa_masks_set(int cpu
)
6254 int node
= cpu_to_node(cpu
);
6256 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6257 for (j
= 0; j
< nr_node_ids
; j
++) {
6258 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6259 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6264 static void sched_domains_numa_masks_clear(int cpu
)
6267 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6268 for (j
= 0; j
< nr_node_ids
; j
++)
6269 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6274 * Update sched_domains_numa_masks[level][node] array when new cpus
6277 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6278 unsigned long action
,
6281 int cpu
= (long)hcpu
;
6283 switch (action
& ~CPU_TASKS_FROZEN
) {
6285 sched_domains_numa_masks_set(cpu
);
6289 sched_domains_numa_masks_clear(cpu
);
6299 static inline void sched_init_numa(void)
6303 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6304 unsigned long action
,
6309 #endif /* CONFIG_NUMA */
6311 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6313 struct sched_domain_topology_level
*tl
;
6316 for_each_sd_topology(tl
) {
6317 struct sd_data
*sdd
= &tl
->data
;
6319 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6323 sdd
->sg
= alloc_percpu(struct sched_group
*);
6327 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
6331 for_each_cpu(j
, cpu_map
) {
6332 struct sched_domain
*sd
;
6333 struct sched_group
*sg
;
6334 struct sched_group_power
*sgp
;
6336 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6337 GFP_KERNEL
, cpu_to_node(j
));
6341 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6343 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6344 GFP_KERNEL
, cpu_to_node(j
));
6350 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6352 sgp
= kzalloc_node(sizeof(struct sched_group_power
) + cpumask_size(),
6353 GFP_KERNEL
, cpu_to_node(j
));
6357 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
6364 static void __sdt_free(const struct cpumask
*cpu_map
)
6366 struct sched_domain_topology_level
*tl
;
6369 for_each_sd_topology(tl
) {
6370 struct sd_data
*sdd
= &tl
->data
;
6372 for_each_cpu(j
, cpu_map
) {
6373 struct sched_domain
*sd
;
6376 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6377 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6378 free_sched_groups(sd
->groups
, 0);
6379 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6383 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6385 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
6387 free_percpu(sdd
->sd
);
6389 free_percpu(sdd
->sg
);
6391 free_percpu(sdd
->sgp
);
6396 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6397 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6398 struct sched_domain
*child
, int cpu
)
6400 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
6404 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6406 sd
->level
= child
->level
+ 1;
6407 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6411 set_domain_attribute(sd
, attr
);
6417 * Build sched domains for a given set of cpus and attach the sched domains
6418 * to the individual cpus
6420 static int build_sched_domains(const struct cpumask
*cpu_map
,
6421 struct sched_domain_attr
*attr
)
6423 enum s_alloc alloc_state
;
6424 struct sched_domain
*sd
;
6426 int i
, ret
= -ENOMEM
;
6428 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6429 if (alloc_state
!= sa_rootdomain
)
6432 /* Set up domains for cpus specified by the cpu_map. */
6433 for_each_cpu(i
, cpu_map
) {
6434 struct sched_domain_topology_level
*tl
;
6437 for_each_sd_topology(tl
) {
6438 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6439 if (tl
== sched_domain_topology
)
6440 *per_cpu_ptr(d
.sd
, i
) = sd
;
6441 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6442 sd
->flags
|= SD_OVERLAP
;
6443 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6448 /* Build the groups for the domains */
6449 for_each_cpu(i
, cpu_map
) {
6450 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6451 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6452 if (sd
->flags
& SD_OVERLAP
) {
6453 if (build_overlap_sched_groups(sd
, i
))
6456 if (build_sched_groups(sd
, i
))
6462 /* Calculate CPU power for physical packages and nodes */
6463 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6464 if (!cpumask_test_cpu(i
, cpu_map
))
6467 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6468 claim_allocations(i
, sd
);
6469 init_sched_groups_power(i
, sd
);
6473 /* Attach the domains */
6475 for_each_cpu(i
, cpu_map
) {
6476 sd
= *per_cpu_ptr(d
.sd
, i
);
6477 cpu_attach_domain(sd
, d
.rd
, i
);
6483 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6487 static cpumask_var_t
*doms_cur
; /* current sched domains */
6488 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6489 static struct sched_domain_attr
*dattr_cur
;
6490 /* attribues of custom domains in 'doms_cur' */
6493 * Special case: If a kmalloc of a doms_cur partition (array of
6494 * cpumask) fails, then fallback to a single sched domain,
6495 * as determined by the single cpumask fallback_doms.
6497 static cpumask_var_t fallback_doms
;
6500 * arch_update_cpu_topology lets virtualized architectures update the
6501 * cpu core maps. It is supposed to return 1 if the topology changed
6502 * or 0 if it stayed the same.
6504 int __attribute__((weak
)) arch_update_cpu_topology(void)
6509 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6512 cpumask_var_t
*doms
;
6514 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6517 for (i
= 0; i
< ndoms
; i
++) {
6518 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6519 free_sched_domains(doms
, i
);
6526 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6529 for (i
= 0; i
< ndoms
; i
++)
6530 free_cpumask_var(doms
[i
]);
6535 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6536 * For now this just excludes isolated cpus, but could be used to
6537 * exclude other special cases in the future.
6539 static int init_sched_domains(const struct cpumask
*cpu_map
)
6543 arch_update_cpu_topology();
6545 doms_cur
= alloc_sched_domains(ndoms_cur
);
6547 doms_cur
= &fallback_doms
;
6548 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6549 err
= build_sched_domains(doms_cur
[0], NULL
);
6550 register_sched_domain_sysctl();
6556 * Detach sched domains from a group of cpus specified in cpu_map
6557 * These cpus will now be attached to the NULL domain
6559 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6564 for_each_cpu(i
, cpu_map
)
6565 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6569 /* handle null as "default" */
6570 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6571 struct sched_domain_attr
*new, int idx_new
)
6573 struct sched_domain_attr tmp
;
6580 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6581 new ? (new + idx_new
) : &tmp
,
6582 sizeof(struct sched_domain_attr
));
6586 * Partition sched domains as specified by the 'ndoms_new'
6587 * cpumasks in the array doms_new[] of cpumasks. This compares
6588 * doms_new[] to the current sched domain partitioning, doms_cur[].
6589 * It destroys each deleted domain and builds each new domain.
6591 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6592 * The masks don't intersect (don't overlap.) We should setup one
6593 * sched domain for each mask. CPUs not in any of the cpumasks will
6594 * not be load balanced. If the same cpumask appears both in the
6595 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6598 * The passed in 'doms_new' should be allocated using
6599 * alloc_sched_domains. This routine takes ownership of it and will
6600 * free_sched_domains it when done with it. If the caller failed the
6601 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6602 * and partition_sched_domains() will fallback to the single partition
6603 * 'fallback_doms', it also forces the domains to be rebuilt.
6605 * If doms_new == NULL it will be replaced with cpu_online_mask.
6606 * ndoms_new == 0 is a special case for destroying existing domains,
6607 * and it will not create the default domain.
6609 * Call with hotplug lock held
6611 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6612 struct sched_domain_attr
*dattr_new
)
6617 mutex_lock(&sched_domains_mutex
);
6619 /* always unregister in case we don't destroy any domains */
6620 unregister_sched_domain_sysctl();
6622 /* Let architecture update cpu core mappings. */
6623 new_topology
= arch_update_cpu_topology();
6625 n
= doms_new
? ndoms_new
: 0;
6627 /* Destroy deleted domains */
6628 for (i
= 0; i
< ndoms_cur
; i
++) {
6629 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6630 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6631 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6634 /* no match - a current sched domain not in new doms_new[] */
6635 detach_destroy_domains(doms_cur
[i
]);
6641 if (doms_new
== NULL
) {
6643 doms_new
= &fallback_doms
;
6644 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6645 WARN_ON_ONCE(dattr_new
);
6648 /* Build new domains */
6649 for (i
= 0; i
< ndoms_new
; i
++) {
6650 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6651 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6652 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6655 /* no match - add a new doms_new */
6656 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6661 /* Remember the new sched domains */
6662 if (doms_cur
!= &fallback_doms
)
6663 free_sched_domains(doms_cur
, ndoms_cur
);
6664 kfree(dattr_cur
); /* kfree(NULL) is safe */
6665 doms_cur
= doms_new
;
6666 dattr_cur
= dattr_new
;
6667 ndoms_cur
= ndoms_new
;
6669 register_sched_domain_sysctl();
6671 mutex_unlock(&sched_domains_mutex
);
6674 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6677 * Update cpusets according to cpu_active mask. If cpusets are
6678 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6679 * around partition_sched_domains().
6681 * If we come here as part of a suspend/resume, don't touch cpusets because we
6682 * want to restore it back to its original state upon resume anyway.
6684 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6688 case CPU_ONLINE_FROZEN
:
6689 case CPU_DOWN_FAILED_FROZEN
:
6692 * num_cpus_frozen tracks how many CPUs are involved in suspend
6693 * resume sequence. As long as this is not the last online
6694 * operation in the resume sequence, just build a single sched
6695 * domain, ignoring cpusets.
6698 if (likely(num_cpus_frozen
)) {
6699 partition_sched_domains(1, NULL
, NULL
);
6704 * This is the last CPU online operation. So fall through and
6705 * restore the original sched domains by considering the
6706 * cpuset configurations.
6710 case CPU_DOWN_FAILED
:
6711 cpuset_update_active_cpus(true);
6719 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6723 case CPU_DOWN_PREPARE
:
6724 cpuset_update_active_cpus(false);
6726 case CPU_DOWN_PREPARE_FROZEN
:
6728 partition_sched_domains(1, NULL
, NULL
);
6736 void __init
sched_init_smp(void)
6738 cpumask_var_t non_isolated_cpus
;
6740 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
6741 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
6746 * There's no userspace yet to cause hotplug operations; hence all the
6747 * cpu masks are stable and all blatant races in the below code cannot
6750 mutex_lock(&sched_domains_mutex
);
6751 init_sched_domains(cpu_active_mask
);
6752 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
6753 if (cpumask_empty(non_isolated_cpus
))
6754 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
6755 mutex_unlock(&sched_domains_mutex
);
6757 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
6758 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
6759 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
6763 /* Move init over to a non-isolated CPU */
6764 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
6766 sched_init_granularity();
6767 free_cpumask_var(non_isolated_cpus
);
6769 init_sched_rt_class();
6770 init_sched_dl_class();
6773 void __init
sched_init_smp(void)
6775 sched_init_granularity();
6777 #endif /* CONFIG_SMP */
6779 const_debug
unsigned int sysctl_timer_migration
= 1;
6781 int in_sched_functions(unsigned long addr
)
6783 return in_lock_functions(addr
) ||
6784 (addr
>= (unsigned long)__sched_text_start
6785 && addr
< (unsigned long)__sched_text_end
);
6788 #ifdef CONFIG_CGROUP_SCHED
6790 * Default task group.
6791 * Every task in system belongs to this group at bootup.
6793 struct task_group root_task_group
;
6794 LIST_HEAD(task_groups
);
6797 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6799 void __init
sched_init(void)
6802 unsigned long alloc_size
= 0, ptr
;
6804 #ifdef CONFIG_FAIR_GROUP_SCHED
6805 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6807 #ifdef CONFIG_RT_GROUP_SCHED
6808 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
6810 #ifdef CONFIG_CPUMASK_OFFSTACK
6811 alloc_size
+= num_possible_cpus() * cpumask_size();
6814 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
6816 #ifdef CONFIG_FAIR_GROUP_SCHED
6817 root_task_group
.se
= (struct sched_entity
**)ptr
;
6818 ptr
+= nr_cpu_ids
* sizeof(void **);
6820 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6821 ptr
+= nr_cpu_ids
* sizeof(void **);
6823 #endif /* CONFIG_FAIR_GROUP_SCHED */
6824 #ifdef CONFIG_RT_GROUP_SCHED
6825 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6826 ptr
+= nr_cpu_ids
* sizeof(void **);
6828 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6829 ptr
+= nr_cpu_ids
* sizeof(void **);
6831 #endif /* CONFIG_RT_GROUP_SCHED */
6832 #ifdef CONFIG_CPUMASK_OFFSTACK
6833 for_each_possible_cpu(i
) {
6834 per_cpu(load_balance_mask
, i
) = (void *)ptr
;
6835 ptr
+= cpumask_size();
6837 #endif /* CONFIG_CPUMASK_OFFSTACK */
6840 init_rt_bandwidth(&def_rt_bandwidth
,
6841 global_rt_period(), global_rt_runtime());
6842 init_dl_bandwidth(&def_dl_bandwidth
,
6843 global_rt_period(), global_rt_runtime());
6846 init_defrootdomain();
6849 #ifdef CONFIG_RT_GROUP_SCHED
6850 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6851 global_rt_period(), global_rt_runtime());
6852 #endif /* CONFIG_RT_GROUP_SCHED */
6854 #ifdef CONFIG_CGROUP_SCHED
6855 list_add(&root_task_group
.list
, &task_groups
);
6856 INIT_LIST_HEAD(&root_task_group
.children
);
6857 INIT_LIST_HEAD(&root_task_group
.siblings
);
6858 autogroup_init(&init_task
);
6860 #endif /* CONFIG_CGROUP_SCHED */
6862 for_each_possible_cpu(i
) {
6866 raw_spin_lock_init(&rq
->lock
);
6868 rq
->calc_load_active
= 0;
6869 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6870 init_cfs_rq(&rq
->cfs
);
6871 init_rt_rq(&rq
->rt
, rq
);
6872 init_dl_rq(&rq
->dl
, rq
);
6873 #ifdef CONFIG_FAIR_GROUP_SCHED
6874 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6875 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6877 * How much cpu bandwidth does root_task_group get?
6879 * In case of task-groups formed thr' the cgroup filesystem, it
6880 * gets 100% of the cpu resources in the system. This overall
6881 * system cpu resource is divided among the tasks of
6882 * root_task_group and its child task-groups in a fair manner,
6883 * based on each entity's (task or task-group's) weight
6884 * (se->load.weight).
6886 * In other words, if root_task_group has 10 tasks of weight
6887 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6888 * then A0's share of the cpu resource is:
6890 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6892 * We achieve this by letting root_task_group's tasks sit
6893 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6895 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6896 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6897 #endif /* CONFIG_FAIR_GROUP_SCHED */
6899 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6900 #ifdef CONFIG_RT_GROUP_SCHED
6901 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6904 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6905 rq
->cpu_load
[j
] = 0;
6907 rq
->last_load_update_tick
= jiffies
;
6912 rq
->cpu_power
= SCHED_POWER_SCALE
;
6913 rq
->post_schedule
= 0;
6914 rq
->active_balance
= 0;
6915 rq
->next_balance
= jiffies
;
6920 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6921 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6923 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6925 rq_attach_root(rq
, &def_root_domain
);
6926 #ifdef CONFIG_NO_HZ_COMMON
6929 #ifdef CONFIG_NO_HZ_FULL
6930 rq
->last_sched_tick
= 0;
6934 atomic_set(&rq
->nr_iowait
, 0);
6937 set_load_weight(&init_task
);
6939 #ifdef CONFIG_PREEMPT_NOTIFIERS
6940 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
6944 * The boot idle thread does lazy MMU switching as well:
6946 atomic_inc(&init_mm
.mm_count
);
6947 enter_lazy_tlb(&init_mm
, current
);
6950 * Make us the idle thread. Technically, schedule() should not be
6951 * called from this thread, however somewhere below it might be,
6952 * but because we are the idle thread, we just pick up running again
6953 * when this runqueue becomes "idle".
6955 init_idle(current
, smp_processor_id());
6957 calc_load_update
= jiffies
+ LOAD_FREQ
;
6960 * During early bootup we pretend to be a normal task:
6962 current
->sched_class
= &fair_sched_class
;
6965 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6966 /* May be allocated at isolcpus cmdline parse time */
6967 if (cpu_isolated_map
== NULL
)
6968 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6969 idle_thread_set_boot_cpu();
6971 init_sched_fair_class();
6973 scheduler_running
= 1;
6976 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6977 static inline int preempt_count_equals(int preempt_offset
)
6979 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
6981 return (nested
== preempt_offset
);
6984 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6986 static unsigned long prev_jiffy
; /* ratelimiting */
6988 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6989 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6990 !is_idle_task(current
)) ||
6991 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6993 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6995 prev_jiffy
= jiffies
;
6998 "BUG: sleeping function called from invalid context at %s:%d\n",
7001 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7002 in_atomic(), irqs_disabled(),
7003 current
->pid
, current
->comm
);
7005 debug_show_held_locks(current
);
7006 if (irqs_disabled())
7007 print_irqtrace_events(current
);
7008 #ifdef CONFIG_DEBUG_PREEMPT
7009 if (!preempt_count_equals(preempt_offset
)) {
7010 pr_err("Preemption disabled at:");
7011 print_ip_sym(current
->preempt_disable_ip
);
7017 EXPORT_SYMBOL(__might_sleep
);
7020 #ifdef CONFIG_MAGIC_SYSRQ
7021 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7023 const struct sched_class
*prev_class
= p
->sched_class
;
7024 struct sched_attr attr
= {
7025 .sched_policy
= SCHED_NORMAL
,
7027 int old_prio
= p
->prio
;
7032 dequeue_task(rq
, p
, 0);
7033 __setscheduler(rq
, p
, &attr
);
7035 enqueue_task(rq
, p
, 0);
7036 resched_task(rq
->curr
);
7039 check_class_changed(rq
, p
, prev_class
, old_prio
);
7042 void normalize_rt_tasks(void)
7044 struct task_struct
*g
, *p
;
7045 unsigned long flags
;
7048 read_lock_irqsave(&tasklist_lock
, flags
);
7049 do_each_thread(g
, p
) {
7051 * Only normalize user tasks:
7056 p
->se
.exec_start
= 0;
7057 #ifdef CONFIG_SCHEDSTATS
7058 p
->se
.statistics
.wait_start
= 0;
7059 p
->se
.statistics
.sleep_start
= 0;
7060 p
->se
.statistics
.block_start
= 0;
7063 if (!dl_task(p
) && !rt_task(p
)) {
7065 * Renice negative nice level userspace
7068 if (task_nice(p
) < 0 && p
->mm
)
7069 set_user_nice(p
, 0);
7073 raw_spin_lock(&p
->pi_lock
);
7074 rq
= __task_rq_lock(p
);
7076 normalize_task(rq
, p
);
7078 __task_rq_unlock(rq
);
7079 raw_spin_unlock(&p
->pi_lock
);
7080 } while_each_thread(g
, p
);
7082 read_unlock_irqrestore(&tasklist_lock
, flags
);
7085 #endif /* CONFIG_MAGIC_SYSRQ */
7087 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7089 * These functions are only useful for the IA64 MCA handling, or kdb.
7091 * They can only be called when the whole system has been
7092 * stopped - every CPU needs to be quiescent, and no scheduling
7093 * activity can take place. Using them for anything else would
7094 * be a serious bug, and as a result, they aren't even visible
7095 * under any other configuration.
7099 * curr_task - return the current task for a given cpu.
7100 * @cpu: the processor in question.
7102 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7104 * Return: The current task for @cpu.
7106 struct task_struct
*curr_task(int cpu
)
7108 return cpu_curr(cpu
);
7111 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7115 * set_curr_task - set the current task for a given cpu.
7116 * @cpu: the processor in question.
7117 * @p: the task pointer to set.
7119 * Description: This function must only be used when non-maskable interrupts
7120 * are serviced on a separate stack. It allows the architecture to switch the
7121 * notion of the current task on a cpu in a non-blocking manner. This function
7122 * must be called with all CPU's synchronized, and interrupts disabled, the
7123 * and caller must save the original value of the current task (see
7124 * curr_task() above) and restore that value before reenabling interrupts and
7125 * re-starting the system.
7127 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7129 void set_curr_task(int cpu
, struct task_struct
*p
)
7136 #ifdef CONFIG_CGROUP_SCHED
7137 /* task_group_lock serializes the addition/removal of task groups */
7138 static DEFINE_SPINLOCK(task_group_lock
);
7140 static void free_sched_group(struct task_group
*tg
)
7142 free_fair_sched_group(tg
);
7143 free_rt_sched_group(tg
);
7148 /* allocate runqueue etc for a new task group */
7149 struct task_group
*sched_create_group(struct task_group
*parent
)
7151 struct task_group
*tg
;
7153 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7155 return ERR_PTR(-ENOMEM
);
7157 if (!alloc_fair_sched_group(tg
, parent
))
7160 if (!alloc_rt_sched_group(tg
, parent
))
7166 free_sched_group(tg
);
7167 return ERR_PTR(-ENOMEM
);
7170 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7172 unsigned long flags
;
7174 spin_lock_irqsave(&task_group_lock
, flags
);
7175 list_add_rcu(&tg
->list
, &task_groups
);
7177 WARN_ON(!parent
); /* root should already exist */
7179 tg
->parent
= parent
;
7180 INIT_LIST_HEAD(&tg
->children
);
7181 list_add_rcu(&tg
->siblings
, &parent
->children
);
7182 spin_unlock_irqrestore(&task_group_lock
, flags
);
7185 /* rcu callback to free various structures associated with a task group */
7186 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7188 /* now it should be safe to free those cfs_rqs */
7189 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7192 /* Destroy runqueue etc associated with a task group */
7193 void sched_destroy_group(struct task_group
*tg
)
7195 /* wait for possible concurrent references to cfs_rqs complete */
7196 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7199 void sched_offline_group(struct task_group
*tg
)
7201 unsigned long flags
;
7204 /* end participation in shares distribution */
7205 for_each_possible_cpu(i
)
7206 unregister_fair_sched_group(tg
, i
);
7208 spin_lock_irqsave(&task_group_lock
, flags
);
7209 list_del_rcu(&tg
->list
);
7210 list_del_rcu(&tg
->siblings
);
7211 spin_unlock_irqrestore(&task_group_lock
, flags
);
7214 /* change task's runqueue when it moves between groups.
7215 * The caller of this function should have put the task in its new group
7216 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7217 * reflect its new group.
7219 void sched_move_task(struct task_struct
*tsk
)
7221 struct task_group
*tg
;
7223 unsigned long flags
;
7226 rq
= task_rq_lock(tsk
, &flags
);
7228 running
= task_current(rq
, tsk
);
7232 dequeue_task(rq
, tsk
, 0);
7233 if (unlikely(running
))
7234 tsk
->sched_class
->put_prev_task(rq
, tsk
);
7236 tg
= container_of(task_css_check(tsk
, cpu_cgroup_subsys_id
,
7237 lockdep_is_held(&tsk
->sighand
->siglock
)),
7238 struct task_group
, css
);
7239 tg
= autogroup_task_group(tsk
, tg
);
7240 tsk
->sched_task_group
= tg
;
7242 #ifdef CONFIG_FAIR_GROUP_SCHED
7243 if (tsk
->sched_class
->task_move_group
)
7244 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
7247 set_task_rq(tsk
, task_cpu(tsk
));
7249 if (unlikely(running
))
7250 tsk
->sched_class
->set_curr_task(rq
);
7252 enqueue_task(rq
, tsk
, 0);
7254 task_rq_unlock(rq
, tsk
, &flags
);
7256 #endif /* CONFIG_CGROUP_SCHED */
7258 #ifdef CONFIG_RT_GROUP_SCHED
7260 * Ensure that the real time constraints are schedulable.
7262 static DEFINE_MUTEX(rt_constraints_mutex
);
7264 /* Must be called with tasklist_lock held */
7265 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7267 struct task_struct
*g
, *p
;
7269 do_each_thread(g
, p
) {
7270 if (rt_task(p
) && task_rq(p
)->rt
.tg
== tg
)
7272 } while_each_thread(g
, p
);
7277 struct rt_schedulable_data
{
7278 struct task_group
*tg
;
7283 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7285 struct rt_schedulable_data
*d
= data
;
7286 struct task_group
*child
;
7287 unsigned long total
, sum
= 0;
7288 u64 period
, runtime
;
7290 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7291 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7294 period
= d
->rt_period
;
7295 runtime
= d
->rt_runtime
;
7299 * Cannot have more runtime than the period.
7301 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7305 * Ensure we don't starve existing RT tasks.
7307 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7310 total
= to_ratio(period
, runtime
);
7313 * Nobody can have more than the global setting allows.
7315 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7319 * The sum of our children's runtime should not exceed our own.
7321 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7322 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7323 runtime
= child
->rt_bandwidth
.rt_runtime
;
7325 if (child
== d
->tg
) {
7326 period
= d
->rt_period
;
7327 runtime
= d
->rt_runtime
;
7330 sum
+= to_ratio(period
, runtime
);
7339 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7343 struct rt_schedulable_data data
= {
7345 .rt_period
= period
,
7346 .rt_runtime
= runtime
,
7350 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7356 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7357 u64 rt_period
, u64 rt_runtime
)
7361 mutex_lock(&rt_constraints_mutex
);
7362 read_lock(&tasklist_lock
);
7363 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7367 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7368 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7369 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7371 for_each_possible_cpu(i
) {
7372 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7374 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7375 rt_rq
->rt_runtime
= rt_runtime
;
7376 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7378 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7380 read_unlock(&tasklist_lock
);
7381 mutex_unlock(&rt_constraints_mutex
);
7386 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7388 u64 rt_runtime
, rt_period
;
7390 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7391 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7392 if (rt_runtime_us
< 0)
7393 rt_runtime
= RUNTIME_INF
;
7395 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7398 static long sched_group_rt_runtime(struct task_group
*tg
)
7402 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7405 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7406 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7407 return rt_runtime_us
;
7410 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7412 u64 rt_runtime
, rt_period
;
7414 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7415 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7420 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7423 static long sched_group_rt_period(struct task_group
*tg
)
7427 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7428 do_div(rt_period_us
, NSEC_PER_USEC
);
7429 return rt_period_us
;
7431 #endif /* CONFIG_RT_GROUP_SCHED */
7433 #ifdef CONFIG_RT_GROUP_SCHED
7434 static int sched_rt_global_constraints(void)
7438 mutex_lock(&rt_constraints_mutex
);
7439 read_lock(&tasklist_lock
);
7440 ret
= __rt_schedulable(NULL
, 0, 0);
7441 read_unlock(&tasklist_lock
);
7442 mutex_unlock(&rt_constraints_mutex
);
7447 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7449 /* Don't accept realtime tasks when there is no way for them to run */
7450 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7456 #else /* !CONFIG_RT_GROUP_SCHED */
7457 static int sched_rt_global_constraints(void)
7459 unsigned long flags
;
7462 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7463 for_each_possible_cpu(i
) {
7464 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7466 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7467 rt_rq
->rt_runtime
= global_rt_runtime();
7468 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7470 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7474 #endif /* CONFIG_RT_GROUP_SCHED */
7476 static int sched_dl_global_constraints(void)
7478 u64 runtime
= global_rt_runtime();
7479 u64 period
= global_rt_period();
7480 u64 new_bw
= to_ratio(period
, runtime
);
7482 unsigned long flags
;
7485 * Here we want to check the bandwidth not being set to some
7486 * value smaller than the currently allocated bandwidth in
7487 * any of the root_domains.
7489 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7490 * cycling on root_domains... Discussion on different/better
7491 * solutions is welcome!
7493 for_each_possible_cpu(cpu
) {
7494 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7496 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7497 if (new_bw
< dl_b
->total_bw
)
7499 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7508 static void sched_dl_do_global(void)
7512 unsigned long flags
;
7514 def_dl_bandwidth
.dl_period
= global_rt_period();
7515 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7517 if (global_rt_runtime() != RUNTIME_INF
)
7518 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7521 * FIXME: As above...
7523 for_each_possible_cpu(cpu
) {
7524 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
7526 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7528 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7532 static int sched_rt_global_validate(void)
7534 if (sysctl_sched_rt_period
<= 0)
7537 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7538 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7544 static void sched_rt_do_global(void)
7546 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7547 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7550 int sched_rt_handler(struct ctl_table
*table
, int write
,
7551 void __user
*buffer
, size_t *lenp
,
7554 int old_period
, old_runtime
;
7555 static DEFINE_MUTEX(mutex
);
7559 old_period
= sysctl_sched_rt_period
;
7560 old_runtime
= sysctl_sched_rt_runtime
;
7562 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7564 if (!ret
&& write
) {
7565 ret
= sched_rt_global_validate();
7569 ret
= sched_rt_global_constraints();
7573 ret
= sched_dl_global_constraints();
7577 sched_rt_do_global();
7578 sched_dl_do_global();
7582 sysctl_sched_rt_period
= old_period
;
7583 sysctl_sched_rt_runtime
= old_runtime
;
7585 mutex_unlock(&mutex
);
7590 int sched_rr_handler(struct ctl_table
*table
, int write
,
7591 void __user
*buffer
, size_t *lenp
,
7595 static DEFINE_MUTEX(mutex
);
7598 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7599 /* make sure that internally we keep jiffies */
7600 /* also, writing zero resets timeslice to default */
7601 if (!ret
&& write
) {
7602 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7603 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7605 mutex_unlock(&mutex
);
7609 #ifdef CONFIG_CGROUP_SCHED
7611 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7613 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7616 static struct cgroup_subsys_state
*
7617 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7619 struct task_group
*parent
= css_tg(parent_css
);
7620 struct task_group
*tg
;
7623 /* This is early initialization for the top cgroup */
7624 return &root_task_group
.css
;
7627 tg
= sched_create_group(parent
);
7629 return ERR_PTR(-ENOMEM
);
7634 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7636 struct task_group
*tg
= css_tg(css
);
7637 struct task_group
*parent
= css_tg(css_parent(css
));
7640 sched_online_group(tg
, parent
);
7644 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7646 struct task_group
*tg
= css_tg(css
);
7648 sched_destroy_group(tg
);
7651 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7653 struct task_group
*tg
= css_tg(css
);
7655 sched_offline_group(tg
);
7658 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7659 struct cgroup_taskset
*tset
)
7661 struct task_struct
*task
;
7663 cgroup_taskset_for_each(task
, css
, tset
) {
7664 #ifdef CONFIG_RT_GROUP_SCHED
7665 if (!sched_rt_can_attach(css_tg(css
), task
))
7668 /* We don't support RT-tasks being in separate groups */
7669 if (task
->sched_class
!= &fair_sched_class
)
7676 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7677 struct cgroup_taskset
*tset
)
7679 struct task_struct
*task
;
7681 cgroup_taskset_for_each(task
, css
, tset
)
7682 sched_move_task(task
);
7685 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
7686 struct cgroup_subsys_state
*old_css
,
7687 struct task_struct
*task
)
7690 * cgroup_exit() is called in the copy_process() failure path.
7691 * Ignore this case since the task hasn't ran yet, this avoids
7692 * trying to poke a half freed task state from generic code.
7694 if (!(task
->flags
& PF_EXITING
))
7697 sched_move_task(task
);
7700 #ifdef CONFIG_FAIR_GROUP_SCHED
7701 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7702 struct cftype
*cftype
, u64 shareval
)
7704 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7707 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7710 struct task_group
*tg
= css_tg(css
);
7712 return (u64
) scale_load_down(tg
->shares
);
7715 #ifdef CONFIG_CFS_BANDWIDTH
7716 static DEFINE_MUTEX(cfs_constraints_mutex
);
7718 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7719 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7721 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7723 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7725 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7726 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7728 if (tg
== &root_task_group
)
7732 * Ensure we have at some amount of bandwidth every period. This is
7733 * to prevent reaching a state of large arrears when throttled via
7734 * entity_tick() resulting in prolonged exit starvation.
7736 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7740 * Likewise, bound things on the otherside by preventing insane quota
7741 * periods. This also allows us to normalize in computing quota
7744 if (period
> max_cfs_quota_period
)
7747 mutex_lock(&cfs_constraints_mutex
);
7748 ret
= __cfs_schedulable(tg
, period
, quota
);
7752 runtime_enabled
= quota
!= RUNTIME_INF
;
7753 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7755 * If we need to toggle cfs_bandwidth_used, off->on must occur
7756 * before making related changes, and on->off must occur afterwards
7758 if (runtime_enabled
&& !runtime_was_enabled
)
7759 cfs_bandwidth_usage_inc();
7760 raw_spin_lock_irq(&cfs_b
->lock
);
7761 cfs_b
->period
= ns_to_ktime(period
);
7762 cfs_b
->quota
= quota
;
7764 __refill_cfs_bandwidth_runtime(cfs_b
);
7765 /* restart the period timer (if active) to handle new period expiry */
7766 if (runtime_enabled
&& cfs_b
->timer_active
) {
7767 /* force a reprogram */
7768 cfs_b
->timer_active
= 0;
7769 __start_cfs_bandwidth(cfs_b
);
7771 raw_spin_unlock_irq(&cfs_b
->lock
);
7773 for_each_possible_cpu(i
) {
7774 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7775 struct rq
*rq
= cfs_rq
->rq
;
7777 raw_spin_lock_irq(&rq
->lock
);
7778 cfs_rq
->runtime_enabled
= runtime_enabled
;
7779 cfs_rq
->runtime_remaining
= 0;
7781 if (cfs_rq
->throttled
)
7782 unthrottle_cfs_rq(cfs_rq
);
7783 raw_spin_unlock_irq(&rq
->lock
);
7785 if (runtime_was_enabled
&& !runtime_enabled
)
7786 cfs_bandwidth_usage_dec();
7788 mutex_unlock(&cfs_constraints_mutex
);
7793 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7797 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7798 if (cfs_quota_us
< 0)
7799 quota
= RUNTIME_INF
;
7801 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7803 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7806 long tg_get_cfs_quota(struct task_group
*tg
)
7810 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7813 quota_us
= tg
->cfs_bandwidth
.quota
;
7814 do_div(quota_us
, NSEC_PER_USEC
);
7819 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7823 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7824 quota
= tg
->cfs_bandwidth
.quota
;
7826 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7829 long tg_get_cfs_period(struct task_group
*tg
)
7833 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7834 do_div(cfs_period_us
, NSEC_PER_USEC
);
7836 return cfs_period_us
;
7839 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7842 return tg_get_cfs_quota(css_tg(css
));
7845 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7846 struct cftype
*cftype
, s64 cfs_quota_us
)
7848 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7851 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7854 return tg_get_cfs_period(css_tg(css
));
7857 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7858 struct cftype
*cftype
, u64 cfs_period_us
)
7860 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7863 struct cfs_schedulable_data
{
7864 struct task_group
*tg
;
7869 * normalize group quota/period to be quota/max_period
7870 * note: units are usecs
7872 static u64
normalize_cfs_quota(struct task_group
*tg
,
7873 struct cfs_schedulable_data
*d
)
7881 period
= tg_get_cfs_period(tg
);
7882 quota
= tg_get_cfs_quota(tg
);
7885 /* note: these should typically be equivalent */
7886 if (quota
== RUNTIME_INF
|| quota
== -1)
7889 return to_ratio(period
, quota
);
7892 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7894 struct cfs_schedulable_data
*d
= data
;
7895 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7896 s64 quota
= 0, parent_quota
= -1;
7899 quota
= RUNTIME_INF
;
7901 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7903 quota
= normalize_cfs_quota(tg
, d
);
7904 parent_quota
= parent_b
->hierarchal_quota
;
7907 * ensure max(child_quota) <= parent_quota, inherit when no
7910 if (quota
== RUNTIME_INF
)
7911 quota
= parent_quota
;
7912 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7915 cfs_b
->hierarchal_quota
= quota
;
7920 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7923 struct cfs_schedulable_data data
= {
7929 if (quota
!= RUNTIME_INF
) {
7930 do_div(data
.period
, NSEC_PER_USEC
);
7931 do_div(data
.quota
, NSEC_PER_USEC
);
7935 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7941 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7943 struct task_group
*tg
= css_tg(seq_css(sf
));
7944 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7946 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7947 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7948 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7952 #endif /* CONFIG_CFS_BANDWIDTH */
7953 #endif /* CONFIG_FAIR_GROUP_SCHED */
7955 #ifdef CONFIG_RT_GROUP_SCHED
7956 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7957 struct cftype
*cft
, s64 val
)
7959 return sched_group_set_rt_runtime(css_tg(css
), val
);
7962 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7965 return sched_group_rt_runtime(css_tg(css
));
7968 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7969 struct cftype
*cftype
, u64 rt_period_us
)
7971 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7974 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7977 return sched_group_rt_period(css_tg(css
));
7979 #endif /* CONFIG_RT_GROUP_SCHED */
7981 static struct cftype cpu_files
[] = {
7982 #ifdef CONFIG_FAIR_GROUP_SCHED
7985 .read_u64
= cpu_shares_read_u64
,
7986 .write_u64
= cpu_shares_write_u64
,
7989 #ifdef CONFIG_CFS_BANDWIDTH
7991 .name
= "cfs_quota_us",
7992 .read_s64
= cpu_cfs_quota_read_s64
,
7993 .write_s64
= cpu_cfs_quota_write_s64
,
7996 .name
= "cfs_period_us",
7997 .read_u64
= cpu_cfs_period_read_u64
,
7998 .write_u64
= cpu_cfs_period_write_u64
,
8002 .seq_show
= cpu_stats_show
,
8005 #ifdef CONFIG_RT_GROUP_SCHED
8007 .name
= "rt_runtime_us",
8008 .read_s64
= cpu_rt_runtime_read
,
8009 .write_s64
= cpu_rt_runtime_write
,
8012 .name
= "rt_period_us",
8013 .read_u64
= cpu_rt_period_read_uint
,
8014 .write_u64
= cpu_rt_period_write_uint
,
8020 struct cgroup_subsys cpu_cgroup_subsys
= {
8022 .css_alloc
= cpu_cgroup_css_alloc
,
8023 .css_free
= cpu_cgroup_css_free
,
8024 .css_online
= cpu_cgroup_css_online
,
8025 .css_offline
= cpu_cgroup_css_offline
,
8026 .can_attach
= cpu_cgroup_can_attach
,
8027 .attach
= cpu_cgroup_attach
,
8028 .exit
= cpu_cgroup_exit
,
8029 .subsys_id
= cpu_cgroup_subsys_id
,
8030 .base_cftypes
= cpu_files
,
8034 #endif /* CONFIG_CGROUP_SCHED */
8036 void dump_cpu_task(int cpu
)
8038 pr_info("Task dump for CPU %d:\n", cpu
);
8039 sched_show_task(cpu_curr(cpu
));