4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex
);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
96 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
98 void update_rq_clock(struct rq
*rq
)
102 lockdep_assert_held(&rq
->lock
);
104 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
107 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
111 update_rq_clock_task(rq
, delta
);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug
unsigned int sysctl_sched_features
=
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names
[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file
*m
, void *v
)
141 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
142 if (!(sysctl_sched_features
& (1UL << i
)))
144 seq_printf(m
, "%s ", sched_feat_names
[i
]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
160 #include "features.h"
165 static void sched_feat_disable(int i
)
167 if (static_key_enabled(&sched_feat_keys
[i
]))
168 static_key_slow_dec(&sched_feat_keys
[i
]);
171 static void sched_feat_enable(int i
)
173 if (!static_key_enabled(&sched_feat_keys
[i
]))
174 static_key_slow_inc(&sched_feat_keys
[i
]);
177 static void sched_feat_disable(int i
) { };
178 static void sched_feat_enable(int i
) { };
179 #endif /* HAVE_JUMP_LABEL */
181 static int sched_feat_set(char *cmp
)
186 if (strncmp(cmp
, "NO_", 3) == 0) {
191 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
192 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
194 sysctl_sched_features
&= ~(1UL << i
);
195 sched_feat_disable(i
);
197 sysctl_sched_features
|= (1UL << i
);
198 sched_feat_enable(i
);
208 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
209 size_t cnt
, loff_t
*ppos
)
219 if (copy_from_user(&buf
, ubuf
, cnt
))
225 /* Ensure the static_key remains in a consistent state */
226 inode
= file_inode(filp
);
227 mutex_lock(&inode
->i_mutex
);
228 i
= sched_feat_set(cmp
);
229 mutex_unlock(&inode
->i_mutex
);
230 if (i
== __SCHED_FEAT_NR
)
238 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
240 return single_open(filp
, sched_feat_show
, NULL
);
243 static const struct file_operations sched_feat_fops
= {
244 .open
= sched_feat_open
,
245 .write
= sched_feat_write
,
248 .release
= single_release
,
251 static __init
int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
258 late_initcall(sched_init_debug
);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
268 * period over which we average the RT time consumption, measured
273 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period
= 1000000;
281 __read_mostly
int scheduler_running
;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime
= 950000;
289 /* cpus with isolated domains */
290 cpumask_var_t cpu_isolated_map
;
293 * this_rq_lock - lock this runqueue and disable interrupts.
295 static struct rq
*this_rq_lock(void)
302 raw_spin_lock(&rq
->lock
);
307 #ifdef CONFIG_SCHED_HRTICK
309 * Use HR-timers to deliver accurate preemption points.
312 static void hrtick_clear(struct rq
*rq
)
314 if (hrtimer_active(&rq
->hrtick_timer
))
315 hrtimer_cancel(&rq
->hrtick_timer
);
319 * High-resolution timer tick.
320 * Runs from hardirq context with interrupts disabled.
322 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
324 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
326 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
328 raw_spin_lock(&rq
->lock
);
330 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
331 raw_spin_unlock(&rq
->lock
);
333 return HRTIMER_NORESTART
;
338 static void __hrtick_restart(struct rq
*rq
)
340 struct hrtimer
*timer
= &rq
->hrtick_timer
;
342 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
346 * called from hardirq (IPI) context
348 static void __hrtick_start(void *arg
)
352 raw_spin_lock(&rq
->lock
);
353 __hrtick_restart(rq
);
354 rq
->hrtick_csd_pending
= 0;
355 raw_spin_unlock(&rq
->lock
);
359 * Called to set the hrtick timer state.
361 * called with rq->lock held and irqs disabled
363 void hrtick_start(struct rq
*rq
, u64 delay
)
365 struct hrtimer
*timer
= &rq
->hrtick_timer
;
370 * Don't schedule slices shorter than 10000ns, that just
371 * doesn't make sense and can cause timer DoS.
373 delta
= max_t(s64
, delay
, 10000LL);
374 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
376 hrtimer_set_expires(timer
, time
);
378 if (rq
== this_rq()) {
379 __hrtick_restart(rq
);
380 } else if (!rq
->hrtick_csd_pending
) {
381 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
382 rq
->hrtick_csd_pending
= 1;
387 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
389 int cpu
= (int)(long)hcpu
;
392 case CPU_UP_CANCELED
:
393 case CPU_UP_CANCELED_FROZEN
:
394 case CPU_DOWN_PREPARE
:
395 case CPU_DOWN_PREPARE_FROZEN
:
397 case CPU_DEAD_FROZEN
:
398 hrtick_clear(cpu_rq(cpu
));
405 static __init
void init_hrtick(void)
407 hotcpu_notifier(hotplug_hrtick
, 0);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq
*rq
, u64 delay
)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay
= max_t(u64
, delay
, 10000LL);
422 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
423 HRTIMER_MODE_REL_PINNED
);
426 static inline void init_hrtick(void)
429 #endif /* CONFIG_SMP */
431 static void init_rq_hrtick(struct rq
*rq
)
434 rq
->hrtick_csd_pending
= 0;
436 rq
->hrtick_csd
.flags
= 0;
437 rq
->hrtick_csd
.func
= __hrtick_start
;
438 rq
->hrtick_csd
.info
= rq
;
441 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
442 rq
->hrtick_timer
.function
= hrtick
;
444 #else /* CONFIG_SCHED_HRTICK */
445 static inline void hrtick_clear(struct rq
*rq
)
449 static inline void init_rq_hrtick(struct rq
*rq
)
453 static inline void init_hrtick(void)
456 #endif /* CONFIG_SCHED_HRTICK */
459 * cmpxchg based fetch_or, macro so it works for different integer types
461 #define fetch_or(ptr, val) \
462 ({ typeof(*(ptr)) __old, __val = *(ptr); \
464 __old = cmpxchg((ptr), __val, __val | (val)); \
465 if (__old == __val) \
472 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
474 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
475 * this avoids any races wrt polling state changes and thereby avoids
478 static bool set_nr_and_not_polling(struct task_struct
*p
)
480 struct thread_info
*ti
= task_thread_info(p
);
481 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
485 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
487 * If this returns true, then the idle task promises to call
488 * sched_ttwu_pending() and reschedule soon.
490 static bool set_nr_if_polling(struct task_struct
*p
)
492 struct thread_info
*ti
= task_thread_info(p
);
493 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
496 if (!(val
& _TIF_POLLING_NRFLAG
))
498 if (val
& _TIF_NEED_RESCHED
)
500 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
509 static bool set_nr_and_not_polling(struct task_struct
*p
)
511 set_tsk_need_resched(p
);
516 static bool set_nr_if_polling(struct task_struct
*p
)
523 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
525 struct wake_q_node
*node
= &task
->wake_q
;
528 * Atomically grab the task, if ->wake_q is !nil already it means
529 * its already queued (either by us or someone else) and will get the
530 * wakeup due to that.
532 * This cmpxchg() implies a full barrier, which pairs with the write
533 * barrier implied by the wakeup in wake_up_list().
535 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
538 get_task_struct(task
);
541 * The head is context local, there can be no concurrency.
544 head
->lastp
= &node
->next
;
547 void wake_up_q(struct wake_q_head
*head
)
549 struct wake_q_node
*node
= head
->first
;
551 while (node
!= WAKE_Q_TAIL
) {
552 struct task_struct
*task
;
554 task
= container_of(node
, struct task_struct
, wake_q
);
556 /* task can safely be re-inserted now */
558 task
->wake_q
.next
= NULL
;
561 * wake_up_process() implies a wmb() to pair with the queueing
562 * in wake_q_add() so as not to miss wakeups.
564 wake_up_process(task
);
565 put_task_struct(task
);
570 * resched_curr - mark rq's current task 'to be rescheduled now'.
572 * On UP this means the setting of the need_resched flag, on SMP it
573 * might also involve a cross-CPU call to trigger the scheduler on
576 void resched_curr(struct rq
*rq
)
578 struct task_struct
*curr
= rq
->curr
;
581 lockdep_assert_held(&rq
->lock
);
583 if (test_tsk_need_resched(curr
))
588 if (cpu
== smp_processor_id()) {
589 set_tsk_need_resched(curr
);
590 set_preempt_need_resched();
594 if (set_nr_and_not_polling(curr
))
595 smp_send_reschedule(cpu
);
597 trace_sched_wake_idle_without_ipi(cpu
);
600 void resched_cpu(int cpu
)
602 struct rq
*rq
= cpu_rq(cpu
);
605 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
608 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
612 #ifdef CONFIG_NO_HZ_COMMON
614 * In the semi idle case, use the nearest busy cpu for migrating timers
615 * from an idle cpu. This is good for power-savings.
617 * We don't do similar optimization for completely idle system, as
618 * selecting an idle cpu will add more delays to the timers than intended
619 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
621 int get_nohz_timer_target(void)
623 int i
, cpu
= smp_processor_id();
624 struct sched_domain
*sd
;
630 for_each_domain(cpu
, sd
) {
631 for_each_cpu(i
, sched_domain_span(sd
)) {
643 * When add_timer_on() enqueues a timer into the timer wheel of an
644 * idle CPU then this timer might expire before the next timer event
645 * which is scheduled to wake up that CPU. In case of a completely
646 * idle system the next event might even be infinite time into the
647 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
648 * leaves the inner idle loop so the newly added timer is taken into
649 * account when the CPU goes back to idle and evaluates the timer
650 * wheel for the next timer event.
652 static void wake_up_idle_cpu(int cpu
)
654 struct rq
*rq
= cpu_rq(cpu
);
656 if (cpu
== smp_processor_id())
659 if (set_nr_and_not_polling(rq
->idle
))
660 smp_send_reschedule(cpu
);
662 trace_sched_wake_idle_without_ipi(cpu
);
665 static bool wake_up_full_nohz_cpu(int cpu
)
668 * We just need the target to call irq_exit() and re-evaluate
669 * the next tick. The nohz full kick at least implies that.
670 * If needed we can still optimize that later with an
673 if (tick_nohz_full_cpu(cpu
)) {
674 if (cpu
!= smp_processor_id() ||
675 tick_nohz_tick_stopped())
676 tick_nohz_full_kick_cpu(cpu
);
683 void wake_up_nohz_cpu(int cpu
)
685 if (!wake_up_full_nohz_cpu(cpu
))
686 wake_up_idle_cpu(cpu
);
689 static inline bool got_nohz_idle_kick(void)
691 int cpu
= smp_processor_id();
693 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
696 if (idle_cpu(cpu
) && !need_resched())
700 * We can't run Idle Load Balance on this CPU for this time so we
701 * cancel it and clear NOHZ_BALANCE_KICK
703 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
707 #else /* CONFIG_NO_HZ_COMMON */
709 static inline bool got_nohz_idle_kick(void)
714 #endif /* CONFIG_NO_HZ_COMMON */
716 #ifdef CONFIG_NO_HZ_FULL
717 bool sched_can_stop_tick(void)
720 * FIFO realtime policy runs the highest priority task. Other runnable
721 * tasks are of a lower priority. The scheduler tick does nothing.
723 if (current
->policy
== SCHED_FIFO
)
727 * Round-robin realtime tasks time slice with other tasks at the same
728 * realtime priority. Is this task the only one at this priority?
730 if (current
->policy
== SCHED_RR
) {
731 struct sched_rt_entity
*rt_se
= ¤t
->rt
;
733 return rt_se
->run_list
.prev
== rt_se
->run_list
.next
;
737 * More than one running task need preemption.
738 * nr_running update is assumed to be visible
739 * after IPI is sent from wakers.
741 if (this_rq()->nr_running
> 1)
746 #endif /* CONFIG_NO_HZ_FULL */
748 void sched_avg_update(struct rq
*rq
)
750 s64 period
= sched_avg_period();
752 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
754 * Inline assembly required to prevent the compiler
755 * optimising this loop into a divmod call.
756 * See __iter_div_u64_rem() for another example of this.
758 asm("" : "+rm" (rq
->age_stamp
));
759 rq
->age_stamp
+= period
;
764 #endif /* CONFIG_SMP */
766 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
767 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
769 * Iterate task_group tree rooted at *from, calling @down when first entering a
770 * node and @up when leaving it for the final time.
772 * Caller must hold rcu_lock or sufficient equivalent.
774 int walk_tg_tree_from(struct task_group
*from
,
775 tg_visitor down
, tg_visitor up
, void *data
)
777 struct task_group
*parent
, *child
;
783 ret
= (*down
)(parent
, data
);
786 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
793 ret
= (*up
)(parent
, data
);
794 if (ret
|| parent
== from
)
798 parent
= parent
->parent
;
805 int tg_nop(struct task_group
*tg
, void *data
)
811 static void set_load_weight(struct task_struct
*p
)
813 int prio
= p
->static_prio
- MAX_RT_PRIO
;
814 struct load_weight
*load
= &p
->se
.load
;
817 * SCHED_IDLE tasks get minimal weight:
819 if (p
->policy
== SCHED_IDLE
) {
820 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
821 load
->inv_weight
= WMULT_IDLEPRIO
;
825 load
->weight
= scale_load(prio_to_weight
[prio
]);
826 load
->inv_weight
= prio_to_wmult
[prio
];
829 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
832 sched_info_queued(rq
, p
);
833 p
->sched_class
->enqueue_task(rq
, p
, flags
);
836 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
839 sched_info_dequeued(rq
, p
);
840 p
->sched_class
->dequeue_task(rq
, p
, flags
);
843 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
845 if (task_contributes_to_load(p
))
846 rq
->nr_uninterruptible
--;
848 enqueue_task(rq
, p
, flags
);
851 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
853 if (task_contributes_to_load(p
))
854 rq
->nr_uninterruptible
++;
856 dequeue_task(rq
, p
, flags
);
859 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
862 * In theory, the compile should just see 0 here, and optimize out the call
863 * to sched_rt_avg_update. But I don't trust it...
865 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
866 s64 steal
= 0, irq_delta
= 0;
868 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
869 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
872 * Since irq_time is only updated on {soft,}irq_exit, we might run into
873 * this case when a previous update_rq_clock() happened inside a
876 * When this happens, we stop ->clock_task and only update the
877 * prev_irq_time stamp to account for the part that fit, so that a next
878 * update will consume the rest. This ensures ->clock_task is
881 * It does however cause some slight miss-attribution of {soft,}irq
882 * time, a more accurate solution would be to update the irq_time using
883 * the current rq->clock timestamp, except that would require using
886 if (irq_delta
> delta
)
889 rq
->prev_irq_time
+= irq_delta
;
892 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
893 if (static_key_false((¶virt_steal_rq_enabled
))) {
894 steal
= paravirt_steal_clock(cpu_of(rq
));
895 steal
-= rq
->prev_steal_time_rq
;
897 if (unlikely(steal
> delta
))
900 rq
->prev_steal_time_rq
+= steal
;
905 rq
->clock_task
+= delta
;
907 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
908 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
909 sched_rt_avg_update(rq
, irq_delta
+ steal
);
913 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
915 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
916 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
920 * Make it appear like a SCHED_FIFO task, its something
921 * userspace knows about and won't get confused about.
923 * Also, it will make PI more or less work without too
924 * much confusion -- but then, stop work should not
925 * rely on PI working anyway.
927 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
929 stop
->sched_class
= &stop_sched_class
;
932 cpu_rq(cpu
)->stop
= stop
;
936 * Reset it back to a normal scheduling class so that
937 * it can die in pieces.
939 old_stop
->sched_class
= &rt_sched_class
;
944 * __normal_prio - return the priority that is based on the static prio
946 static inline int __normal_prio(struct task_struct
*p
)
948 return p
->static_prio
;
952 * Calculate the expected normal priority: i.e. priority
953 * without taking RT-inheritance into account. Might be
954 * boosted by interactivity modifiers. Changes upon fork,
955 * setprio syscalls, and whenever the interactivity
956 * estimator recalculates.
958 static inline int normal_prio(struct task_struct
*p
)
962 if (task_has_dl_policy(p
))
963 prio
= MAX_DL_PRIO
-1;
964 else if (task_has_rt_policy(p
))
965 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
967 prio
= __normal_prio(p
);
972 * Calculate the current priority, i.e. the priority
973 * taken into account by the scheduler. This value might
974 * be boosted by RT tasks, or might be boosted by
975 * interactivity modifiers. Will be RT if the task got
976 * RT-boosted. If not then it returns p->normal_prio.
978 static int effective_prio(struct task_struct
*p
)
980 p
->normal_prio
= normal_prio(p
);
982 * If we are RT tasks or we were boosted to RT priority,
983 * keep the priority unchanged. Otherwise, update priority
984 * to the normal priority:
986 if (!rt_prio(p
->prio
))
987 return p
->normal_prio
;
992 * task_curr - is this task currently executing on a CPU?
993 * @p: the task in question.
995 * Return: 1 if the task is currently executing. 0 otherwise.
997 inline int task_curr(const struct task_struct
*p
)
999 return cpu_curr(task_cpu(p
)) == p
;
1003 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1004 * use the balance_callback list if you want balancing.
1006 * this means any call to check_class_changed() must be followed by a call to
1007 * balance_callback().
1009 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1010 const struct sched_class
*prev_class
,
1013 if (prev_class
!= p
->sched_class
) {
1014 if (prev_class
->switched_from
)
1015 prev_class
->switched_from(rq
, p
);
1017 p
->sched_class
->switched_to(rq
, p
);
1018 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1019 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1022 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1024 const struct sched_class
*class;
1026 if (p
->sched_class
== rq
->curr
->sched_class
) {
1027 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1029 for_each_class(class) {
1030 if (class == rq
->curr
->sched_class
)
1032 if (class == p
->sched_class
) {
1040 * A queue event has occurred, and we're going to schedule. In
1041 * this case, we can save a useless back to back clock update.
1043 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1044 rq_clock_skip_update(rq
, true);
1049 * This is how migration works:
1051 * 1) we invoke migration_cpu_stop() on the target CPU using
1053 * 2) stopper starts to run (implicitly forcing the migrated thread
1055 * 3) it checks whether the migrated task is still in the wrong runqueue.
1056 * 4) if it's in the wrong runqueue then the migration thread removes
1057 * it and puts it into the right queue.
1058 * 5) stopper completes and stop_one_cpu() returns and the migration
1063 * move_queued_task - move a queued task to new rq.
1065 * Returns (locked) new rq. Old rq's lock is released.
1067 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
1069 lockdep_assert_held(&rq
->lock
);
1071 dequeue_task(rq
, p
, 0);
1072 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1073 set_task_cpu(p
, new_cpu
);
1074 raw_spin_unlock(&rq
->lock
);
1076 rq
= cpu_rq(new_cpu
);
1078 raw_spin_lock(&rq
->lock
);
1079 BUG_ON(task_cpu(p
) != new_cpu
);
1080 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1081 enqueue_task(rq
, p
, 0);
1082 check_preempt_curr(rq
, p
, 0);
1087 struct migration_arg
{
1088 struct task_struct
*task
;
1093 * Move (not current) task off this cpu, onto dest cpu. We're doing
1094 * this because either it can't run here any more (set_cpus_allowed()
1095 * away from this CPU, or CPU going down), or because we're
1096 * attempting to rebalance this task on exec (sched_exec).
1098 * So we race with normal scheduler movements, but that's OK, as long
1099 * as the task is no longer on this CPU.
1101 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1103 if (unlikely(!cpu_active(dest_cpu
)))
1106 /* Affinity changed (again). */
1107 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1110 rq
= move_queued_task(rq
, p
, dest_cpu
);
1116 * migration_cpu_stop - this will be executed by a highprio stopper thread
1117 * and performs thread migration by bumping thread off CPU then
1118 * 'pushing' onto another runqueue.
1120 static int migration_cpu_stop(void *data
)
1122 struct migration_arg
*arg
= data
;
1123 struct task_struct
*p
= arg
->task
;
1124 struct rq
*rq
= this_rq();
1127 * The original target cpu might have gone down and we might
1128 * be on another cpu but it doesn't matter.
1130 local_irq_disable();
1132 * We need to explicitly wake pending tasks before running
1133 * __migrate_task() such that we will not miss enforcing cpus_allowed
1134 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1136 sched_ttwu_pending();
1138 raw_spin_lock(&p
->pi_lock
);
1139 raw_spin_lock(&rq
->lock
);
1141 * If task_rq(p) != rq, it cannot be migrated here, because we're
1142 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1143 * we're holding p->pi_lock.
1145 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1146 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1147 raw_spin_unlock(&rq
->lock
);
1148 raw_spin_unlock(&p
->pi_lock
);
1155 * sched_class::set_cpus_allowed must do the below, but is not required to
1156 * actually call this function.
1158 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1160 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1161 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1164 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1166 lockdep_assert_held(&p
->pi_lock
);
1167 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1171 * Change a given task's CPU affinity. Migrate the thread to a
1172 * proper CPU and schedule it away if the CPU it's executing on
1173 * is removed from the allowed bitmask.
1175 * NOTE: the caller must have a valid reference to the task, the
1176 * task must not exit() & deallocate itself prematurely. The
1177 * call is not atomic; no spinlocks may be held.
1179 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1180 const struct cpumask
*new_mask
, bool check
)
1182 unsigned long flags
;
1184 unsigned int dest_cpu
;
1187 rq
= task_rq_lock(p
, &flags
);
1190 * Must re-check here, to close a race against __kthread_bind(),
1191 * sched_setaffinity() is not guaranteed to observe the flag.
1193 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1198 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1201 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1206 do_set_cpus_allowed(p
, new_mask
);
1208 /* Can the task run on the task's current CPU? If so, we're done */
1209 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1212 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1213 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1214 struct migration_arg arg
= { p
, dest_cpu
};
1215 /* Need help from migration thread: drop lock and wait. */
1216 task_rq_unlock(rq
, p
, &flags
);
1217 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1218 tlb_migrate_finish(p
->mm
);
1220 } else if (task_on_rq_queued(p
)) {
1222 * OK, since we're going to drop the lock immediately
1223 * afterwards anyway.
1225 lockdep_unpin_lock(&rq
->lock
);
1226 rq
= move_queued_task(rq
, p
, dest_cpu
);
1227 lockdep_pin_lock(&rq
->lock
);
1230 task_rq_unlock(rq
, p
, &flags
);
1235 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1237 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1239 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1241 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1243 #ifdef CONFIG_SCHED_DEBUG
1245 * We should never call set_task_cpu() on a blocked task,
1246 * ttwu() will sort out the placement.
1248 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1251 #ifdef CONFIG_LOCKDEP
1253 * The caller should hold either p->pi_lock or rq->lock, when changing
1254 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1256 * sched_move_task() holds both and thus holding either pins the cgroup,
1259 * Furthermore, all task_rq users should acquire both locks, see
1262 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1263 lockdep_is_held(&task_rq(p
)->lock
)));
1267 trace_sched_migrate_task(p
, new_cpu
);
1269 if (task_cpu(p
) != new_cpu
) {
1270 if (p
->sched_class
->migrate_task_rq
)
1271 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1272 p
->se
.nr_migrations
++;
1273 perf_event_task_migrate(p
);
1276 __set_task_cpu(p
, new_cpu
);
1279 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1281 if (task_on_rq_queued(p
)) {
1282 struct rq
*src_rq
, *dst_rq
;
1284 src_rq
= task_rq(p
);
1285 dst_rq
= cpu_rq(cpu
);
1287 deactivate_task(src_rq
, p
, 0);
1288 set_task_cpu(p
, cpu
);
1289 activate_task(dst_rq
, p
, 0);
1290 check_preempt_curr(dst_rq
, p
, 0);
1293 * Task isn't running anymore; make it appear like we migrated
1294 * it before it went to sleep. This means on wakeup we make the
1295 * previous cpu our targer instead of where it really is.
1301 struct migration_swap_arg
{
1302 struct task_struct
*src_task
, *dst_task
;
1303 int src_cpu
, dst_cpu
;
1306 static int migrate_swap_stop(void *data
)
1308 struct migration_swap_arg
*arg
= data
;
1309 struct rq
*src_rq
, *dst_rq
;
1312 src_rq
= cpu_rq(arg
->src_cpu
);
1313 dst_rq
= cpu_rq(arg
->dst_cpu
);
1315 double_raw_lock(&arg
->src_task
->pi_lock
,
1316 &arg
->dst_task
->pi_lock
);
1317 double_rq_lock(src_rq
, dst_rq
);
1318 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1321 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1324 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1327 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1330 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1331 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1336 double_rq_unlock(src_rq
, dst_rq
);
1337 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1338 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1344 * Cross migrate two tasks
1346 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1348 struct migration_swap_arg arg
;
1351 arg
= (struct migration_swap_arg
){
1353 .src_cpu
= task_cpu(cur
),
1355 .dst_cpu
= task_cpu(p
),
1358 if (arg
.src_cpu
== arg
.dst_cpu
)
1362 * These three tests are all lockless; this is OK since all of them
1363 * will be re-checked with proper locks held further down the line.
1365 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1368 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1371 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1374 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1375 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1382 * wait_task_inactive - wait for a thread to unschedule.
1384 * If @match_state is nonzero, it's the @p->state value just checked and
1385 * not expected to change. If it changes, i.e. @p might have woken up,
1386 * then return zero. When we succeed in waiting for @p to be off its CPU,
1387 * we return a positive number (its total switch count). If a second call
1388 * a short while later returns the same number, the caller can be sure that
1389 * @p has remained unscheduled the whole time.
1391 * The caller must ensure that the task *will* unschedule sometime soon,
1392 * else this function might spin for a *long* time. This function can't
1393 * be called with interrupts off, or it may introduce deadlock with
1394 * smp_call_function() if an IPI is sent by the same process we are
1395 * waiting to become inactive.
1397 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1399 unsigned long flags
;
1400 int running
, queued
;
1406 * We do the initial early heuristics without holding
1407 * any task-queue locks at all. We'll only try to get
1408 * the runqueue lock when things look like they will
1414 * If the task is actively running on another CPU
1415 * still, just relax and busy-wait without holding
1418 * NOTE! Since we don't hold any locks, it's not
1419 * even sure that "rq" stays as the right runqueue!
1420 * But we don't care, since "task_running()" will
1421 * return false if the runqueue has changed and p
1422 * is actually now running somewhere else!
1424 while (task_running(rq
, p
)) {
1425 if (match_state
&& unlikely(p
->state
!= match_state
))
1431 * Ok, time to look more closely! We need the rq
1432 * lock now, to be *sure*. If we're wrong, we'll
1433 * just go back and repeat.
1435 rq
= task_rq_lock(p
, &flags
);
1436 trace_sched_wait_task(p
);
1437 running
= task_running(rq
, p
);
1438 queued
= task_on_rq_queued(p
);
1440 if (!match_state
|| p
->state
== match_state
)
1441 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1442 task_rq_unlock(rq
, p
, &flags
);
1445 * If it changed from the expected state, bail out now.
1447 if (unlikely(!ncsw
))
1451 * Was it really running after all now that we
1452 * checked with the proper locks actually held?
1454 * Oops. Go back and try again..
1456 if (unlikely(running
)) {
1462 * It's not enough that it's not actively running,
1463 * it must be off the runqueue _entirely_, and not
1466 * So if it was still runnable (but just not actively
1467 * running right now), it's preempted, and we should
1468 * yield - it could be a while.
1470 if (unlikely(queued
)) {
1471 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1473 set_current_state(TASK_UNINTERRUPTIBLE
);
1474 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1479 * Ahh, all good. It wasn't running, and it wasn't
1480 * runnable, which means that it will never become
1481 * running in the future either. We're all done!
1490 * kick_process - kick a running thread to enter/exit the kernel
1491 * @p: the to-be-kicked thread
1493 * Cause a process which is running on another CPU to enter
1494 * kernel-mode, without any delay. (to get signals handled.)
1496 * NOTE: this function doesn't have to take the runqueue lock,
1497 * because all it wants to ensure is that the remote task enters
1498 * the kernel. If the IPI races and the task has been migrated
1499 * to another CPU then no harm is done and the purpose has been
1502 void kick_process(struct task_struct
*p
)
1508 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1509 smp_send_reschedule(cpu
);
1512 EXPORT_SYMBOL_GPL(kick_process
);
1515 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1517 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1519 int nid
= cpu_to_node(cpu
);
1520 const struct cpumask
*nodemask
= NULL
;
1521 enum { cpuset
, possible
, fail
} state
= cpuset
;
1525 * If the node that the cpu is on has been offlined, cpu_to_node()
1526 * will return -1. There is no cpu on the node, and we should
1527 * select the cpu on the other node.
1530 nodemask
= cpumask_of_node(nid
);
1532 /* Look for allowed, online CPU in same node. */
1533 for_each_cpu(dest_cpu
, nodemask
) {
1534 if (!cpu_online(dest_cpu
))
1536 if (!cpu_active(dest_cpu
))
1538 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1544 /* Any allowed, online CPU? */
1545 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1546 if (!cpu_online(dest_cpu
))
1548 if (!cpu_active(dest_cpu
))
1555 /* No more Mr. Nice Guy. */
1556 cpuset_cpus_allowed_fallback(p
);
1561 do_set_cpus_allowed(p
, cpu_possible_mask
);
1572 if (state
!= cpuset
) {
1574 * Don't tell them about moving exiting tasks or
1575 * kernel threads (both mm NULL), since they never
1578 if (p
->mm
&& printk_ratelimit()) {
1579 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1580 task_pid_nr(p
), p
->comm
, cpu
);
1588 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1591 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1593 lockdep_assert_held(&p
->pi_lock
);
1595 if (p
->nr_cpus_allowed
> 1)
1596 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1599 * In order not to call set_task_cpu() on a blocking task we need
1600 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1603 * Since this is common to all placement strategies, this lives here.
1605 * [ this allows ->select_task() to simply return task_cpu(p) and
1606 * not worry about this generic constraint ]
1608 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1610 cpu
= select_fallback_rq(task_cpu(p
), p
);
1615 static void update_avg(u64
*avg
, u64 sample
)
1617 s64 diff
= sample
- *avg
;
1623 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1624 const struct cpumask
*new_mask
, bool check
)
1626 return set_cpus_allowed_ptr(p
, new_mask
);
1629 #endif /* CONFIG_SMP */
1632 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1634 #ifdef CONFIG_SCHEDSTATS
1635 struct rq
*rq
= this_rq();
1638 int this_cpu
= smp_processor_id();
1640 if (cpu
== this_cpu
) {
1641 schedstat_inc(rq
, ttwu_local
);
1642 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1644 struct sched_domain
*sd
;
1646 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1648 for_each_domain(this_cpu
, sd
) {
1649 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1650 schedstat_inc(sd
, ttwu_wake_remote
);
1657 if (wake_flags
& WF_MIGRATED
)
1658 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1660 #endif /* CONFIG_SMP */
1662 schedstat_inc(rq
, ttwu_count
);
1663 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1665 if (wake_flags
& WF_SYNC
)
1666 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1668 #endif /* CONFIG_SCHEDSTATS */
1671 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1673 activate_task(rq
, p
, en_flags
);
1674 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1676 /* if a worker is waking up, notify workqueue */
1677 if (p
->flags
& PF_WQ_WORKER
)
1678 wq_worker_waking_up(p
, cpu_of(rq
));
1682 * Mark the task runnable and perform wakeup-preemption.
1685 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1687 check_preempt_curr(rq
, p
, wake_flags
);
1688 p
->state
= TASK_RUNNING
;
1689 trace_sched_wakeup(p
);
1692 if (p
->sched_class
->task_woken
) {
1694 * Our task @p is fully woken up and running; so its safe to
1695 * drop the rq->lock, hereafter rq is only used for statistics.
1697 lockdep_unpin_lock(&rq
->lock
);
1698 p
->sched_class
->task_woken(rq
, p
);
1699 lockdep_pin_lock(&rq
->lock
);
1702 if (rq
->idle_stamp
) {
1703 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1704 u64 max
= 2*rq
->max_idle_balance_cost
;
1706 update_avg(&rq
->avg_idle
, delta
);
1708 if (rq
->avg_idle
> max
)
1717 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1719 lockdep_assert_held(&rq
->lock
);
1722 if (p
->sched_contributes_to_load
)
1723 rq
->nr_uninterruptible
--;
1726 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1727 ttwu_do_wakeup(rq
, p
, wake_flags
);
1731 * Called in case the task @p isn't fully descheduled from its runqueue,
1732 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1733 * since all we need to do is flip p->state to TASK_RUNNING, since
1734 * the task is still ->on_rq.
1736 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1741 rq
= __task_rq_lock(p
);
1742 if (task_on_rq_queued(p
)) {
1743 /* check_preempt_curr() may use rq clock */
1744 update_rq_clock(rq
);
1745 ttwu_do_wakeup(rq
, p
, wake_flags
);
1748 __task_rq_unlock(rq
);
1754 void sched_ttwu_pending(void)
1756 struct rq
*rq
= this_rq();
1757 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1758 struct task_struct
*p
;
1759 unsigned long flags
;
1764 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1765 lockdep_pin_lock(&rq
->lock
);
1768 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1769 llist
= llist_next(llist
);
1770 ttwu_do_activate(rq
, p
, 0);
1773 lockdep_unpin_lock(&rq
->lock
);
1774 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1777 void scheduler_ipi(void)
1780 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1781 * TIF_NEED_RESCHED remotely (for the first time) will also send
1784 preempt_fold_need_resched();
1786 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1790 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1791 * traditionally all their work was done from the interrupt return
1792 * path. Now that we actually do some work, we need to make sure
1795 * Some archs already do call them, luckily irq_enter/exit nest
1798 * Arguably we should visit all archs and update all handlers,
1799 * however a fair share of IPIs are still resched only so this would
1800 * somewhat pessimize the simple resched case.
1803 sched_ttwu_pending();
1806 * Check if someone kicked us for doing the nohz idle load balance.
1808 if (unlikely(got_nohz_idle_kick())) {
1809 this_rq()->idle_balance
= 1;
1810 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1815 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1817 struct rq
*rq
= cpu_rq(cpu
);
1819 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1820 if (!set_nr_if_polling(rq
->idle
))
1821 smp_send_reschedule(cpu
);
1823 trace_sched_wake_idle_without_ipi(cpu
);
1827 void wake_up_if_idle(int cpu
)
1829 struct rq
*rq
= cpu_rq(cpu
);
1830 unsigned long flags
;
1834 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1837 if (set_nr_if_polling(rq
->idle
)) {
1838 trace_sched_wake_idle_without_ipi(cpu
);
1840 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1841 if (is_idle_task(rq
->curr
))
1842 smp_send_reschedule(cpu
);
1843 /* Else cpu is not in idle, do nothing here */
1844 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1851 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1853 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1855 #endif /* CONFIG_SMP */
1857 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1859 struct rq
*rq
= cpu_rq(cpu
);
1861 #if defined(CONFIG_SMP)
1862 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1863 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1864 ttwu_queue_remote(p
, cpu
);
1869 raw_spin_lock(&rq
->lock
);
1870 lockdep_pin_lock(&rq
->lock
);
1871 ttwu_do_activate(rq
, p
, 0);
1872 lockdep_unpin_lock(&rq
->lock
);
1873 raw_spin_unlock(&rq
->lock
);
1877 * try_to_wake_up - wake up a thread
1878 * @p: the thread to be awakened
1879 * @state: the mask of task states that can be woken
1880 * @wake_flags: wake modifier flags (WF_*)
1882 * Put it on the run-queue if it's not already there. The "current"
1883 * thread is always on the run-queue (except when the actual
1884 * re-schedule is in progress), and as such you're allowed to do
1885 * the simpler "current->state = TASK_RUNNING" to mark yourself
1886 * runnable without the overhead of this.
1888 * Return: %true if @p was woken up, %false if it was already running.
1889 * or @state didn't match @p's state.
1892 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1894 unsigned long flags
;
1895 int cpu
, success
= 0;
1898 * If we are going to wake up a thread waiting for CONDITION we
1899 * need to ensure that CONDITION=1 done by the caller can not be
1900 * reordered with p->state check below. This pairs with mb() in
1901 * set_current_state() the waiting thread does.
1903 smp_mb__before_spinlock();
1904 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1905 if (!(p
->state
& state
))
1908 trace_sched_waking(p
);
1910 success
= 1; /* we're going to change ->state */
1913 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1918 * If the owning (remote) cpu is still in the middle of schedule() with
1919 * this task as prev, wait until its done referencing the task.
1924 * Pairs with the smp_wmb() in finish_lock_switch().
1928 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1929 p
->state
= TASK_WAKING
;
1931 if (p
->sched_class
->task_waking
)
1932 p
->sched_class
->task_waking(p
);
1934 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1935 if (task_cpu(p
) != cpu
) {
1936 wake_flags
|= WF_MIGRATED
;
1937 set_task_cpu(p
, cpu
);
1939 #endif /* CONFIG_SMP */
1943 ttwu_stat(p
, cpu
, wake_flags
);
1945 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1951 * try_to_wake_up_local - try to wake up a local task with rq lock held
1952 * @p: the thread to be awakened
1954 * Put @p on the run-queue if it's not already there. The caller must
1955 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1958 static void try_to_wake_up_local(struct task_struct
*p
)
1960 struct rq
*rq
= task_rq(p
);
1962 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1963 WARN_ON_ONCE(p
== current
))
1966 lockdep_assert_held(&rq
->lock
);
1968 if (!raw_spin_trylock(&p
->pi_lock
)) {
1970 * This is OK, because current is on_cpu, which avoids it being
1971 * picked for load-balance and preemption/IRQs are still
1972 * disabled avoiding further scheduler activity on it and we've
1973 * not yet picked a replacement task.
1975 lockdep_unpin_lock(&rq
->lock
);
1976 raw_spin_unlock(&rq
->lock
);
1977 raw_spin_lock(&p
->pi_lock
);
1978 raw_spin_lock(&rq
->lock
);
1979 lockdep_pin_lock(&rq
->lock
);
1982 if (!(p
->state
& TASK_NORMAL
))
1985 trace_sched_waking(p
);
1987 if (!task_on_rq_queued(p
))
1988 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1990 ttwu_do_wakeup(rq
, p
, 0);
1991 ttwu_stat(p
, smp_processor_id(), 0);
1993 raw_spin_unlock(&p
->pi_lock
);
1997 * wake_up_process - Wake up a specific process
1998 * @p: The process to be woken up.
2000 * Attempt to wake up the nominated process and move it to the set of runnable
2003 * Return: 1 if the process was woken up, 0 if it was already running.
2005 * It may be assumed that this function implies a write memory barrier before
2006 * changing the task state if and only if any tasks are woken up.
2008 int wake_up_process(struct task_struct
*p
)
2010 WARN_ON(task_is_stopped_or_traced(p
));
2011 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2013 EXPORT_SYMBOL(wake_up_process
);
2015 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2017 return try_to_wake_up(p
, state
, 0);
2021 * This function clears the sched_dl_entity static params.
2023 void __dl_clear_params(struct task_struct
*p
)
2025 struct sched_dl_entity
*dl_se
= &p
->dl
;
2027 dl_se
->dl_runtime
= 0;
2028 dl_se
->dl_deadline
= 0;
2029 dl_se
->dl_period
= 0;
2033 dl_se
->dl_throttled
= 0;
2035 dl_se
->dl_yielded
= 0;
2039 * Perform scheduler related setup for a newly forked process p.
2040 * p is forked by current.
2042 * __sched_fork() is basic setup used by init_idle() too:
2044 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2049 p
->se
.exec_start
= 0;
2050 p
->se
.sum_exec_runtime
= 0;
2051 p
->se
.prev_sum_exec_runtime
= 0;
2052 p
->se
.nr_migrations
= 0;
2054 INIT_LIST_HEAD(&p
->se
.group_node
);
2056 #ifdef CONFIG_SCHEDSTATS
2057 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2060 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2061 init_dl_task_timer(&p
->dl
);
2062 __dl_clear_params(p
);
2064 INIT_LIST_HEAD(&p
->rt
.run_list
);
2066 #ifdef CONFIG_PREEMPT_NOTIFIERS
2067 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2070 #ifdef CONFIG_NUMA_BALANCING
2071 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2072 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2073 p
->mm
->numa_scan_seq
= 0;
2076 if (clone_flags
& CLONE_VM
)
2077 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2079 p
->numa_preferred_nid
= -1;
2081 p
->node_stamp
= 0ULL;
2082 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2083 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2084 p
->numa_work
.next
= &p
->numa_work
;
2085 p
->numa_faults
= NULL
;
2086 p
->last_task_numa_placement
= 0;
2087 p
->last_sum_exec_runtime
= 0;
2089 p
->numa_group
= NULL
;
2090 #endif /* CONFIG_NUMA_BALANCING */
2093 #ifdef CONFIG_NUMA_BALANCING
2094 #ifdef CONFIG_SCHED_DEBUG
2095 void set_numabalancing_state(bool enabled
)
2098 sched_feat_set("NUMA");
2100 sched_feat_set("NO_NUMA");
2103 __read_mostly
bool numabalancing_enabled
;
2105 void set_numabalancing_state(bool enabled
)
2107 numabalancing_enabled
= enabled
;
2109 #endif /* CONFIG_SCHED_DEBUG */
2111 #ifdef CONFIG_PROC_SYSCTL
2112 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2113 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2117 int state
= numabalancing_enabled
;
2119 if (write
&& !capable(CAP_SYS_ADMIN
))
2124 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2128 set_numabalancing_state(state
);
2135 * fork()/clone()-time setup:
2137 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2139 unsigned long flags
;
2140 int cpu
= get_cpu();
2142 __sched_fork(clone_flags
, p
);
2144 * We mark the process as running here. This guarantees that
2145 * nobody will actually run it, and a signal or other external
2146 * event cannot wake it up and insert it on the runqueue either.
2148 p
->state
= TASK_RUNNING
;
2151 * Make sure we do not leak PI boosting priority to the child.
2153 p
->prio
= current
->normal_prio
;
2156 * Revert to default priority/policy on fork if requested.
2158 if (unlikely(p
->sched_reset_on_fork
)) {
2159 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2160 p
->policy
= SCHED_NORMAL
;
2161 p
->static_prio
= NICE_TO_PRIO(0);
2163 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2164 p
->static_prio
= NICE_TO_PRIO(0);
2166 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2170 * We don't need the reset flag anymore after the fork. It has
2171 * fulfilled its duty:
2173 p
->sched_reset_on_fork
= 0;
2176 if (dl_prio(p
->prio
)) {
2179 } else if (rt_prio(p
->prio
)) {
2180 p
->sched_class
= &rt_sched_class
;
2182 p
->sched_class
= &fair_sched_class
;
2185 if (p
->sched_class
->task_fork
)
2186 p
->sched_class
->task_fork(p
);
2189 * The child is not yet in the pid-hash so no cgroup attach races,
2190 * and the cgroup is pinned to this child due to cgroup_fork()
2191 * is ran before sched_fork().
2193 * Silence PROVE_RCU.
2195 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2196 set_task_cpu(p
, cpu
);
2197 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2199 #ifdef CONFIG_SCHED_INFO
2200 if (likely(sched_info_on()))
2201 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2203 #if defined(CONFIG_SMP)
2206 init_task_preempt_count(p
);
2208 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2209 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2216 unsigned long to_ratio(u64 period
, u64 runtime
)
2218 if (runtime
== RUNTIME_INF
)
2222 * Doing this here saves a lot of checks in all
2223 * the calling paths, and returning zero seems
2224 * safe for them anyway.
2229 return div64_u64(runtime
<< 20, period
);
2233 inline struct dl_bw
*dl_bw_of(int i
)
2235 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2236 "sched RCU must be held");
2237 return &cpu_rq(i
)->rd
->dl_bw
;
2240 static inline int dl_bw_cpus(int i
)
2242 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2245 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2246 "sched RCU must be held");
2247 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2253 inline struct dl_bw
*dl_bw_of(int i
)
2255 return &cpu_rq(i
)->dl
.dl_bw
;
2258 static inline int dl_bw_cpus(int i
)
2265 * We must be sure that accepting a new task (or allowing changing the
2266 * parameters of an existing one) is consistent with the bandwidth
2267 * constraints. If yes, this function also accordingly updates the currently
2268 * allocated bandwidth to reflect the new situation.
2270 * This function is called while holding p's rq->lock.
2272 * XXX we should delay bw change until the task's 0-lag point, see
2275 static int dl_overflow(struct task_struct
*p
, int policy
,
2276 const struct sched_attr
*attr
)
2279 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2280 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2281 u64 runtime
= attr
->sched_runtime
;
2282 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2285 if (new_bw
== p
->dl
.dl_bw
)
2289 * Either if a task, enters, leave, or stays -deadline but changes
2290 * its parameters, we may need to update accordingly the total
2291 * allocated bandwidth of the container.
2293 raw_spin_lock(&dl_b
->lock
);
2294 cpus
= dl_bw_cpus(task_cpu(p
));
2295 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2296 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2297 __dl_add(dl_b
, new_bw
);
2299 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2300 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2301 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2302 __dl_add(dl_b
, new_bw
);
2304 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2305 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2308 raw_spin_unlock(&dl_b
->lock
);
2313 extern void init_dl_bw(struct dl_bw
*dl_b
);
2316 * wake_up_new_task - wake up a newly created task for the first time.
2318 * This function will do some initial scheduler statistics housekeeping
2319 * that must be done for every newly created context, then puts the task
2320 * on the runqueue and wakes it.
2322 void wake_up_new_task(struct task_struct
*p
)
2324 unsigned long flags
;
2327 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2330 * Fork balancing, do it here and not earlier because:
2331 * - cpus_allowed can change in the fork path
2332 * - any previously selected cpu might disappear through hotplug
2334 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2337 /* Initialize new task's runnable average */
2338 init_entity_runnable_average(&p
->se
);
2339 rq
= __task_rq_lock(p
);
2340 activate_task(rq
, p
, 0);
2341 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2342 trace_sched_wakeup_new(p
);
2343 check_preempt_curr(rq
, p
, WF_FORK
);
2345 if (p
->sched_class
->task_woken
)
2346 p
->sched_class
->task_woken(rq
, p
);
2348 task_rq_unlock(rq
, p
, &flags
);
2351 #ifdef CONFIG_PREEMPT_NOTIFIERS
2353 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2355 void preempt_notifier_inc(void)
2357 static_key_slow_inc(&preempt_notifier_key
);
2359 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2361 void preempt_notifier_dec(void)
2363 static_key_slow_dec(&preempt_notifier_key
);
2365 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2368 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2369 * @notifier: notifier struct to register
2371 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2373 if (!static_key_false(&preempt_notifier_key
))
2374 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2376 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2378 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2381 * preempt_notifier_unregister - no longer interested in preemption notifications
2382 * @notifier: notifier struct to unregister
2384 * This is *not* safe to call from within a preemption notifier.
2386 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2388 hlist_del(¬ifier
->link
);
2390 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2392 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2394 struct preempt_notifier
*notifier
;
2396 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2397 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2400 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2402 if (static_key_false(&preempt_notifier_key
))
2403 __fire_sched_in_preempt_notifiers(curr
);
2407 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2408 struct task_struct
*next
)
2410 struct preempt_notifier
*notifier
;
2412 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2413 notifier
->ops
->sched_out(notifier
, next
);
2416 static __always_inline
void
2417 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2418 struct task_struct
*next
)
2420 if (static_key_false(&preempt_notifier_key
))
2421 __fire_sched_out_preempt_notifiers(curr
, next
);
2424 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2426 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2431 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2432 struct task_struct
*next
)
2436 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2439 * prepare_task_switch - prepare to switch tasks
2440 * @rq: the runqueue preparing to switch
2441 * @prev: the current task that is being switched out
2442 * @next: the task we are going to switch to.
2444 * This is called with the rq lock held and interrupts off. It must
2445 * be paired with a subsequent finish_task_switch after the context
2448 * prepare_task_switch sets up locking and calls architecture specific
2452 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2453 struct task_struct
*next
)
2455 trace_sched_switch(prev
, next
);
2456 sched_info_switch(rq
, prev
, next
);
2457 perf_event_task_sched_out(prev
, next
);
2458 fire_sched_out_preempt_notifiers(prev
, next
);
2459 prepare_lock_switch(rq
, next
);
2460 prepare_arch_switch(next
);
2464 * finish_task_switch - clean up after a task-switch
2465 * @prev: the thread we just switched away from.
2467 * finish_task_switch must be called after the context switch, paired
2468 * with a prepare_task_switch call before the context switch.
2469 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2470 * and do any other architecture-specific cleanup actions.
2472 * Note that we may have delayed dropping an mm in context_switch(). If
2473 * so, we finish that here outside of the runqueue lock. (Doing it
2474 * with the lock held can cause deadlocks; see schedule() for
2477 * The context switch have flipped the stack from under us and restored the
2478 * local variables which were saved when this task called schedule() in the
2479 * past. prev == current is still correct but we need to recalculate this_rq
2480 * because prev may have moved to another CPU.
2482 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2483 __releases(rq
->lock
)
2485 struct rq
*rq
= this_rq();
2486 struct mm_struct
*mm
= rq
->prev_mm
;
2492 * A task struct has one reference for the use as "current".
2493 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2494 * schedule one last time. The schedule call will never return, and
2495 * the scheduled task must drop that reference.
2496 * The test for TASK_DEAD must occur while the runqueue locks are
2497 * still held, otherwise prev could be scheduled on another cpu, die
2498 * there before we look at prev->state, and then the reference would
2500 * Manfred Spraul <manfred@colorfullife.com>
2502 prev_state
= prev
->state
;
2503 vtime_task_switch(prev
);
2504 perf_event_task_sched_in(prev
, current
);
2505 finish_lock_switch(rq
, prev
);
2506 finish_arch_post_lock_switch();
2508 fire_sched_in_preempt_notifiers(current
);
2511 if (unlikely(prev_state
== TASK_DEAD
)) {
2512 if (prev
->sched_class
->task_dead
)
2513 prev
->sched_class
->task_dead(prev
);
2516 * Remove function-return probe instances associated with this
2517 * task and put them back on the free list.
2519 kprobe_flush_task(prev
);
2520 put_task_struct(prev
);
2523 tick_nohz_task_switch(current
);
2529 /* rq->lock is NOT held, but preemption is disabled */
2530 static void __balance_callback(struct rq
*rq
)
2532 struct callback_head
*head
, *next
;
2533 void (*func
)(struct rq
*rq
);
2534 unsigned long flags
;
2536 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2537 head
= rq
->balance_callback
;
2538 rq
->balance_callback
= NULL
;
2540 func
= (void (*)(struct rq
*))head
->func
;
2547 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2550 static inline void balance_callback(struct rq
*rq
)
2552 if (unlikely(rq
->balance_callback
))
2553 __balance_callback(rq
);
2558 static inline void balance_callback(struct rq
*rq
)
2565 * schedule_tail - first thing a freshly forked thread must call.
2566 * @prev: the thread we just switched away from.
2568 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2569 __releases(rq
->lock
)
2573 /* finish_task_switch() drops rq->lock and enables preemtion */
2575 rq
= finish_task_switch(prev
);
2576 balance_callback(rq
);
2579 if (current
->set_child_tid
)
2580 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2584 * context_switch - switch to the new MM and the new thread's register state.
2586 static inline struct rq
*
2587 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2588 struct task_struct
*next
)
2590 struct mm_struct
*mm
, *oldmm
;
2592 prepare_task_switch(rq
, prev
, next
);
2595 oldmm
= prev
->active_mm
;
2597 * For paravirt, this is coupled with an exit in switch_to to
2598 * combine the page table reload and the switch backend into
2601 arch_start_context_switch(prev
);
2604 next
->active_mm
= oldmm
;
2605 atomic_inc(&oldmm
->mm_count
);
2606 enter_lazy_tlb(oldmm
, next
);
2608 switch_mm(oldmm
, mm
, next
);
2611 prev
->active_mm
= NULL
;
2612 rq
->prev_mm
= oldmm
;
2615 * Since the runqueue lock will be released by the next
2616 * task (which is an invalid locking op but in the case
2617 * of the scheduler it's an obvious special-case), so we
2618 * do an early lockdep release here:
2620 lockdep_unpin_lock(&rq
->lock
);
2621 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2623 /* Here we just switch the register state and the stack. */
2624 switch_to(prev
, next
, prev
);
2627 return finish_task_switch(prev
);
2631 * nr_running and nr_context_switches:
2633 * externally visible scheduler statistics: current number of runnable
2634 * threads, total number of context switches performed since bootup.
2636 unsigned long nr_running(void)
2638 unsigned long i
, sum
= 0;
2640 for_each_online_cpu(i
)
2641 sum
+= cpu_rq(i
)->nr_running
;
2647 * Check if only the current task is running on the cpu.
2649 bool single_task_running(void)
2651 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2656 EXPORT_SYMBOL(single_task_running
);
2658 unsigned long long nr_context_switches(void)
2661 unsigned long long sum
= 0;
2663 for_each_possible_cpu(i
)
2664 sum
+= cpu_rq(i
)->nr_switches
;
2669 unsigned long nr_iowait(void)
2671 unsigned long i
, sum
= 0;
2673 for_each_possible_cpu(i
)
2674 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2679 unsigned long nr_iowait_cpu(int cpu
)
2681 struct rq
*this = cpu_rq(cpu
);
2682 return atomic_read(&this->nr_iowait
);
2685 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2687 struct rq
*rq
= this_rq();
2688 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2689 *load
= rq
->load
.weight
;
2695 * sched_exec - execve() is a valuable balancing opportunity, because at
2696 * this point the task has the smallest effective memory and cache footprint.
2698 void sched_exec(void)
2700 struct task_struct
*p
= current
;
2701 unsigned long flags
;
2704 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2705 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2706 if (dest_cpu
== smp_processor_id())
2709 if (likely(cpu_active(dest_cpu
))) {
2710 struct migration_arg arg
= { p
, dest_cpu
};
2712 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2713 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2717 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2722 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2723 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2725 EXPORT_PER_CPU_SYMBOL(kstat
);
2726 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2729 * Return accounted runtime for the task.
2730 * In case the task is currently running, return the runtime plus current's
2731 * pending runtime that have not been accounted yet.
2733 unsigned long long task_sched_runtime(struct task_struct
*p
)
2735 unsigned long flags
;
2739 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2741 * 64-bit doesn't need locks to atomically read a 64bit value.
2742 * So we have a optimization chance when the task's delta_exec is 0.
2743 * Reading ->on_cpu is racy, but this is ok.
2745 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2746 * If we race with it entering cpu, unaccounted time is 0. This is
2747 * indistinguishable from the read occurring a few cycles earlier.
2748 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2749 * been accounted, so we're correct here as well.
2751 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2752 return p
->se
.sum_exec_runtime
;
2755 rq
= task_rq_lock(p
, &flags
);
2757 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2758 * project cycles that may never be accounted to this
2759 * thread, breaking clock_gettime().
2761 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2762 update_rq_clock(rq
);
2763 p
->sched_class
->update_curr(rq
);
2765 ns
= p
->se
.sum_exec_runtime
;
2766 task_rq_unlock(rq
, p
, &flags
);
2772 * This function gets called by the timer code, with HZ frequency.
2773 * We call it with interrupts disabled.
2775 void scheduler_tick(void)
2777 int cpu
= smp_processor_id();
2778 struct rq
*rq
= cpu_rq(cpu
);
2779 struct task_struct
*curr
= rq
->curr
;
2783 raw_spin_lock(&rq
->lock
);
2784 update_rq_clock(rq
);
2785 curr
->sched_class
->task_tick(rq
, curr
, 0);
2786 update_cpu_load_active(rq
);
2787 calc_global_load_tick(rq
);
2788 raw_spin_unlock(&rq
->lock
);
2790 perf_event_task_tick();
2793 rq
->idle_balance
= idle_cpu(cpu
);
2794 trigger_load_balance(rq
);
2796 rq_last_tick_reset(rq
);
2799 #ifdef CONFIG_NO_HZ_FULL
2801 * scheduler_tick_max_deferment
2803 * Keep at least one tick per second when a single
2804 * active task is running because the scheduler doesn't
2805 * yet completely support full dynticks environment.
2807 * This makes sure that uptime, CFS vruntime, load
2808 * balancing, etc... continue to move forward, even
2809 * with a very low granularity.
2811 * Return: Maximum deferment in nanoseconds.
2813 u64
scheduler_tick_max_deferment(void)
2815 struct rq
*rq
= this_rq();
2816 unsigned long next
, now
= READ_ONCE(jiffies
);
2818 next
= rq
->last_sched_tick
+ HZ
;
2820 if (time_before_eq(next
, now
))
2823 return jiffies_to_nsecs(next
- now
);
2827 notrace
unsigned long get_parent_ip(unsigned long addr
)
2829 if (in_lock_functions(addr
)) {
2830 addr
= CALLER_ADDR2
;
2831 if (in_lock_functions(addr
))
2832 addr
= CALLER_ADDR3
;
2837 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2838 defined(CONFIG_PREEMPT_TRACER))
2840 void preempt_count_add(int val
)
2842 #ifdef CONFIG_DEBUG_PREEMPT
2846 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2849 __preempt_count_add(val
);
2850 #ifdef CONFIG_DEBUG_PREEMPT
2852 * Spinlock count overflowing soon?
2854 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2857 if (preempt_count() == val
) {
2858 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2859 #ifdef CONFIG_DEBUG_PREEMPT
2860 current
->preempt_disable_ip
= ip
;
2862 trace_preempt_off(CALLER_ADDR0
, ip
);
2865 EXPORT_SYMBOL(preempt_count_add
);
2866 NOKPROBE_SYMBOL(preempt_count_add
);
2868 void preempt_count_sub(int val
)
2870 #ifdef CONFIG_DEBUG_PREEMPT
2874 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2877 * Is the spinlock portion underflowing?
2879 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2880 !(preempt_count() & PREEMPT_MASK
)))
2884 if (preempt_count() == val
)
2885 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2886 __preempt_count_sub(val
);
2888 EXPORT_SYMBOL(preempt_count_sub
);
2889 NOKPROBE_SYMBOL(preempt_count_sub
);
2894 * Print scheduling while atomic bug:
2896 static noinline
void __schedule_bug(struct task_struct
*prev
)
2898 if (oops_in_progress
)
2901 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2902 prev
->comm
, prev
->pid
, preempt_count());
2904 debug_show_held_locks(prev
);
2906 if (irqs_disabled())
2907 print_irqtrace_events(prev
);
2908 #ifdef CONFIG_DEBUG_PREEMPT
2909 if (in_atomic_preempt_off()) {
2910 pr_err("Preemption disabled at:");
2911 print_ip_sym(current
->preempt_disable_ip
);
2916 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2920 * Various schedule()-time debugging checks and statistics:
2922 static inline void schedule_debug(struct task_struct
*prev
)
2924 #ifdef CONFIG_SCHED_STACK_END_CHECK
2925 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2928 * Test if we are atomic. Since do_exit() needs to call into
2929 * schedule() atomically, we ignore that path. Otherwise whine
2930 * if we are scheduling when we should not.
2932 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2933 __schedule_bug(prev
);
2936 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2938 schedstat_inc(this_rq(), sched_count
);
2942 * Pick up the highest-prio task:
2944 static inline struct task_struct
*
2945 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2947 const struct sched_class
*class = &fair_sched_class
;
2948 struct task_struct
*p
;
2951 * Optimization: we know that if all tasks are in
2952 * the fair class we can call that function directly:
2954 if (likely(prev
->sched_class
== class &&
2955 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2956 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2957 if (unlikely(p
== RETRY_TASK
))
2960 /* assumes fair_sched_class->next == idle_sched_class */
2962 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2968 for_each_class(class) {
2969 p
= class->pick_next_task(rq
, prev
);
2971 if (unlikely(p
== RETRY_TASK
))
2977 BUG(); /* the idle class will always have a runnable task */
2981 * __schedule() is the main scheduler function.
2983 * The main means of driving the scheduler and thus entering this function are:
2985 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2987 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2988 * paths. For example, see arch/x86/entry_64.S.
2990 * To drive preemption between tasks, the scheduler sets the flag in timer
2991 * interrupt handler scheduler_tick().
2993 * 3. Wakeups don't really cause entry into schedule(). They add a
2994 * task to the run-queue and that's it.
2996 * Now, if the new task added to the run-queue preempts the current
2997 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2998 * called on the nearest possible occasion:
3000 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3002 * - in syscall or exception context, at the next outmost
3003 * preempt_enable(). (this might be as soon as the wake_up()'s
3006 * - in IRQ context, return from interrupt-handler to
3007 * preemptible context
3009 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3012 * - cond_resched() call
3013 * - explicit schedule() call
3014 * - return from syscall or exception to user-space
3015 * - return from interrupt-handler to user-space
3017 * WARNING: must be called with preemption disabled!
3019 static void __sched
__schedule(void)
3021 struct task_struct
*prev
, *next
;
3022 unsigned long *switch_count
;
3026 cpu
= smp_processor_id();
3028 rcu_note_context_switch();
3031 schedule_debug(prev
);
3033 if (sched_feat(HRTICK
))
3037 * Make sure that signal_pending_state()->signal_pending() below
3038 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3039 * done by the caller to avoid the race with signal_wake_up().
3041 smp_mb__before_spinlock();
3042 raw_spin_lock_irq(&rq
->lock
);
3043 lockdep_pin_lock(&rq
->lock
);
3045 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3047 switch_count
= &prev
->nivcsw
;
3048 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3049 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3050 prev
->state
= TASK_RUNNING
;
3052 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3056 * If a worker went to sleep, notify and ask workqueue
3057 * whether it wants to wake up a task to maintain
3060 if (prev
->flags
& PF_WQ_WORKER
) {
3061 struct task_struct
*to_wakeup
;
3063 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3065 try_to_wake_up_local(to_wakeup
);
3068 switch_count
= &prev
->nvcsw
;
3071 if (task_on_rq_queued(prev
))
3072 update_rq_clock(rq
);
3074 next
= pick_next_task(rq
, prev
);
3075 clear_tsk_need_resched(prev
);
3076 clear_preempt_need_resched();
3077 rq
->clock_skip_update
= 0;
3079 if (likely(prev
!= next
)) {
3084 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3087 lockdep_unpin_lock(&rq
->lock
);
3088 raw_spin_unlock_irq(&rq
->lock
);
3091 balance_callback(rq
);
3094 static inline void sched_submit_work(struct task_struct
*tsk
)
3096 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3099 * If we are going to sleep and we have plugged IO queued,
3100 * make sure to submit it to avoid deadlocks.
3102 if (blk_needs_flush_plug(tsk
))
3103 blk_schedule_flush_plug(tsk
);
3106 asmlinkage __visible
void __sched
schedule(void)
3108 struct task_struct
*tsk
= current
;
3110 sched_submit_work(tsk
);
3114 sched_preempt_enable_no_resched();
3115 } while (need_resched());
3117 EXPORT_SYMBOL(schedule
);
3119 #ifdef CONFIG_CONTEXT_TRACKING
3120 asmlinkage __visible
void __sched
schedule_user(void)
3123 * If we come here after a random call to set_need_resched(),
3124 * or we have been woken up remotely but the IPI has not yet arrived,
3125 * we haven't yet exited the RCU idle mode. Do it here manually until
3126 * we find a better solution.
3128 * NB: There are buggy callers of this function. Ideally we
3129 * should warn if prev_state != CONTEXT_USER, but that will trigger
3130 * too frequently to make sense yet.
3132 enum ctx_state prev_state
= exception_enter();
3134 exception_exit(prev_state
);
3139 * schedule_preempt_disabled - called with preemption disabled
3141 * Returns with preemption disabled. Note: preempt_count must be 1
3143 void __sched
schedule_preempt_disabled(void)
3145 sched_preempt_enable_no_resched();
3150 static void __sched notrace
preempt_schedule_common(void)
3153 preempt_active_enter();
3155 preempt_active_exit();
3158 * Check again in case we missed a preemption opportunity
3159 * between schedule and now.
3161 } while (need_resched());
3164 #ifdef CONFIG_PREEMPT
3166 * this is the entry point to schedule() from in-kernel preemption
3167 * off of preempt_enable. Kernel preemptions off return from interrupt
3168 * occur there and call schedule directly.
3170 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3173 * If there is a non-zero preempt_count or interrupts are disabled,
3174 * we do not want to preempt the current task. Just return..
3176 if (likely(!preemptible()))
3179 preempt_schedule_common();
3181 NOKPROBE_SYMBOL(preempt_schedule
);
3182 EXPORT_SYMBOL(preempt_schedule
);
3185 * preempt_schedule_notrace - preempt_schedule called by tracing
3187 * The tracing infrastructure uses preempt_enable_notrace to prevent
3188 * recursion and tracing preempt enabling caused by the tracing
3189 * infrastructure itself. But as tracing can happen in areas coming
3190 * from userspace or just about to enter userspace, a preempt enable
3191 * can occur before user_exit() is called. This will cause the scheduler
3192 * to be called when the system is still in usermode.
3194 * To prevent this, the preempt_enable_notrace will use this function
3195 * instead of preempt_schedule() to exit user context if needed before
3196 * calling the scheduler.
3198 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3200 enum ctx_state prev_ctx
;
3202 if (likely(!preemptible()))
3207 * Use raw __prempt_count() ops that don't call function.
3208 * We can't call functions before disabling preemption which
3209 * disarm preemption tracing recursions.
3211 __preempt_count_add(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3214 * Needs preempt disabled in case user_exit() is traced
3215 * and the tracer calls preempt_enable_notrace() causing
3216 * an infinite recursion.
3218 prev_ctx
= exception_enter();
3220 exception_exit(prev_ctx
);
3223 __preempt_count_sub(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3224 } while (need_resched());
3226 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3228 #endif /* CONFIG_PREEMPT */
3231 * this is the entry point to schedule() from kernel preemption
3232 * off of irq context.
3233 * Note, that this is called and return with irqs disabled. This will
3234 * protect us against recursive calling from irq.
3236 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3238 enum ctx_state prev_state
;
3240 /* Catch callers which need to be fixed */
3241 BUG_ON(preempt_count() || !irqs_disabled());
3243 prev_state
= exception_enter();
3246 preempt_active_enter();
3249 local_irq_disable();
3250 preempt_active_exit();
3251 } while (need_resched());
3253 exception_exit(prev_state
);
3256 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3259 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3261 EXPORT_SYMBOL(default_wake_function
);
3263 #ifdef CONFIG_RT_MUTEXES
3266 * rt_mutex_setprio - set the current priority of a task
3268 * @prio: prio value (kernel-internal form)
3270 * This function changes the 'effective' priority of a task. It does
3271 * not touch ->normal_prio like __setscheduler().
3273 * Used by the rt_mutex code to implement priority inheritance
3274 * logic. Call site only calls if the priority of the task changed.
3276 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3278 int oldprio
, queued
, running
, enqueue_flag
= 0;
3280 const struct sched_class
*prev_class
;
3282 BUG_ON(prio
> MAX_PRIO
);
3284 rq
= __task_rq_lock(p
);
3287 * Idle task boosting is a nono in general. There is one
3288 * exception, when PREEMPT_RT and NOHZ is active:
3290 * The idle task calls get_next_timer_interrupt() and holds
3291 * the timer wheel base->lock on the CPU and another CPU wants
3292 * to access the timer (probably to cancel it). We can safely
3293 * ignore the boosting request, as the idle CPU runs this code
3294 * with interrupts disabled and will complete the lock
3295 * protected section without being interrupted. So there is no
3296 * real need to boost.
3298 if (unlikely(p
== rq
->idle
)) {
3299 WARN_ON(p
!= rq
->curr
);
3300 WARN_ON(p
->pi_blocked_on
);
3304 trace_sched_pi_setprio(p
, prio
);
3306 prev_class
= p
->sched_class
;
3307 queued
= task_on_rq_queued(p
);
3308 running
= task_current(rq
, p
);
3310 dequeue_task(rq
, p
, 0);
3312 put_prev_task(rq
, p
);
3315 * Boosting condition are:
3316 * 1. -rt task is running and holds mutex A
3317 * --> -dl task blocks on mutex A
3319 * 2. -dl task is running and holds mutex A
3320 * --> -dl task blocks on mutex A and could preempt the
3323 if (dl_prio(prio
)) {
3324 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3325 if (!dl_prio(p
->normal_prio
) ||
3326 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3327 p
->dl
.dl_boosted
= 1;
3328 enqueue_flag
= ENQUEUE_REPLENISH
;
3330 p
->dl
.dl_boosted
= 0;
3331 p
->sched_class
= &dl_sched_class
;
3332 } else if (rt_prio(prio
)) {
3333 if (dl_prio(oldprio
))
3334 p
->dl
.dl_boosted
= 0;
3336 enqueue_flag
= ENQUEUE_HEAD
;
3337 p
->sched_class
= &rt_sched_class
;
3339 if (dl_prio(oldprio
))
3340 p
->dl
.dl_boosted
= 0;
3341 if (rt_prio(oldprio
))
3343 p
->sched_class
= &fair_sched_class
;
3349 p
->sched_class
->set_curr_task(rq
);
3351 enqueue_task(rq
, p
, enqueue_flag
);
3353 check_class_changed(rq
, p
, prev_class
, oldprio
);
3355 preempt_disable(); /* avoid rq from going away on us */
3356 __task_rq_unlock(rq
);
3358 balance_callback(rq
);
3363 void set_user_nice(struct task_struct
*p
, long nice
)
3365 int old_prio
, delta
, queued
;
3366 unsigned long flags
;
3369 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3372 * We have to be careful, if called from sys_setpriority(),
3373 * the task might be in the middle of scheduling on another CPU.
3375 rq
= task_rq_lock(p
, &flags
);
3377 * The RT priorities are set via sched_setscheduler(), but we still
3378 * allow the 'normal' nice value to be set - but as expected
3379 * it wont have any effect on scheduling until the task is
3380 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3382 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3383 p
->static_prio
= NICE_TO_PRIO(nice
);
3386 queued
= task_on_rq_queued(p
);
3388 dequeue_task(rq
, p
, 0);
3390 p
->static_prio
= NICE_TO_PRIO(nice
);
3393 p
->prio
= effective_prio(p
);
3394 delta
= p
->prio
- old_prio
;
3397 enqueue_task(rq
, p
, 0);
3399 * If the task increased its priority or is running and
3400 * lowered its priority, then reschedule its CPU:
3402 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3406 task_rq_unlock(rq
, p
, &flags
);
3408 EXPORT_SYMBOL(set_user_nice
);
3411 * can_nice - check if a task can reduce its nice value
3415 int can_nice(const struct task_struct
*p
, const int nice
)
3417 /* convert nice value [19,-20] to rlimit style value [1,40] */
3418 int nice_rlim
= nice_to_rlimit(nice
);
3420 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3421 capable(CAP_SYS_NICE
));
3424 #ifdef __ARCH_WANT_SYS_NICE
3427 * sys_nice - change the priority of the current process.
3428 * @increment: priority increment
3430 * sys_setpriority is a more generic, but much slower function that
3431 * does similar things.
3433 SYSCALL_DEFINE1(nice
, int, increment
)
3438 * Setpriority might change our priority at the same moment.
3439 * We don't have to worry. Conceptually one call occurs first
3440 * and we have a single winner.
3442 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3443 nice
= task_nice(current
) + increment
;
3445 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3446 if (increment
< 0 && !can_nice(current
, nice
))
3449 retval
= security_task_setnice(current
, nice
);
3453 set_user_nice(current
, nice
);
3460 * task_prio - return the priority value of a given task.
3461 * @p: the task in question.
3463 * Return: The priority value as seen by users in /proc.
3464 * RT tasks are offset by -200. Normal tasks are centered
3465 * around 0, value goes from -16 to +15.
3467 int task_prio(const struct task_struct
*p
)
3469 return p
->prio
- MAX_RT_PRIO
;
3473 * idle_cpu - is a given cpu idle currently?
3474 * @cpu: the processor in question.
3476 * Return: 1 if the CPU is currently idle. 0 otherwise.
3478 int idle_cpu(int cpu
)
3480 struct rq
*rq
= cpu_rq(cpu
);
3482 if (rq
->curr
!= rq
->idle
)
3489 if (!llist_empty(&rq
->wake_list
))
3497 * idle_task - return the idle task for a given cpu.
3498 * @cpu: the processor in question.
3500 * Return: The idle task for the cpu @cpu.
3502 struct task_struct
*idle_task(int cpu
)
3504 return cpu_rq(cpu
)->idle
;
3508 * find_process_by_pid - find a process with a matching PID value.
3509 * @pid: the pid in question.
3511 * The task of @pid, if found. %NULL otherwise.
3513 static struct task_struct
*find_process_by_pid(pid_t pid
)
3515 return pid
? find_task_by_vpid(pid
) : current
;
3519 * This function initializes the sched_dl_entity of a newly becoming
3520 * SCHED_DEADLINE task.
3522 * Only the static values are considered here, the actual runtime and the
3523 * absolute deadline will be properly calculated when the task is enqueued
3524 * for the first time with its new policy.
3527 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3529 struct sched_dl_entity
*dl_se
= &p
->dl
;
3531 dl_se
->dl_runtime
= attr
->sched_runtime
;
3532 dl_se
->dl_deadline
= attr
->sched_deadline
;
3533 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3534 dl_se
->flags
= attr
->sched_flags
;
3535 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3538 * Changing the parameters of a task is 'tricky' and we're not doing
3539 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3541 * What we SHOULD do is delay the bandwidth release until the 0-lag
3542 * point. This would include retaining the task_struct until that time
3543 * and change dl_overflow() to not immediately decrement the current
3546 * Instead we retain the current runtime/deadline and let the new
3547 * parameters take effect after the current reservation period lapses.
3548 * This is safe (albeit pessimistic) because the 0-lag point is always
3549 * before the current scheduling deadline.
3551 * We can still have temporary overloads because we do not delay the
3552 * change in bandwidth until that time; so admission control is
3553 * not on the safe side. It does however guarantee tasks will never
3554 * consume more than promised.
3559 * sched_setparam() passes in -1 for its policy, to let the functions
3560 * it calls know not to change it.
3562 #define SETPARAM_POLICY -1
3564 static void __setscheduler_params(struct task_struct
*p
,
3565 const struct sched_attr
*attr
)
3567 int policy
= attr
->sched_policy
;
3569 if (policy
== SETPARAM_POLICY
)
3574 if (dl_policy(policy
))
3575 __setparam_dl(p
, attr
);
3576 else if (fair_policy(policy
))
3577 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3580 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3581 * !rt_policy. Always setting this ensures that things like
3582 * getparam()/getattr() don't report silly values for !rt tasks.
3584 p
->rt_priority
= attr
->sched_priority
;
3585 p
->normal_prio
= normal_prio(p
);
3589 /* Actually do priority change: must hold pi & rq lock. */
3590 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3591 const struct sched_attr
*attr
, bool keep_boost
)
3593 __setscheduler_params(p
, attr
);
3596 * Keep a potential priority boosting if called from
3597 * sched_setscheduler().
3600 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3602 p
->prio
= normal_prio(p
);
3604 if (dl_prio(p
->prio
))
3605 p
->sched_class
= &dl_sched_class
;
3606 else if (rt_prio(p
->prio
))
3607 p
->sched_class
= &rt_sched_class
;
3609 p
->sched_class
= &fair_sched_class
;
3613 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3615 struct sched_dl_entity
*dl_se
= &p
->dl
;
3617 attr
->sched_priority
= p
->rt_priority
;
3618 attr
->sched_runtime
= dl_se
->dl_runtime
;
3619 attr
->sched_deadline
= dl_se
->dl_deadline
;
3620 attr
->sched_period
= dl_se
->dl_period
;
3621 attr
->sched_flags
= dl_se
->flags
;
3625 * This function validates the new parameters of a -deadline task.
3626 * We ask for the deadline not being zero, and greater or equal
3627 * than the runtime, as well as the period of being zero or
3628 * greater than deadline. Furthermore, we have to be sure that
3629 * user parameters are above the internal resolution of 1us (we
3630 * check sched_runtime only since it is always the smaller one) and
3631 * below 2^63 ns (we have to check both sched_deadline and
3632 * sched_period, as the latter can be zero).
3635 __checkparam_dl(const struct sched_attr
*attr
)
3638 if (attr
->sched_deadline
== 0)
3642 * Since we truncate DL_SCALE bits, make sure we're at least
3645 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3649 * Since we use the MSB for wrap-around and sign issues, make
3650 * sure it's not set (mind that period can be equal to zero).
3652 if (attr
->sched_deadline
& (1ULL << 63) ||
3653 attr
->sched_period
& (1ULL << 63))
3656 /* runtime <= deadline <= period (if period != 0) */
3657 if ((attr
->sched_period
!= 0 &&
3658 attr
->sched_period
< attr
->sched_deadline
) ||
3659 attr
->sched_deadline
< attr
->sched_runtime
)
3666 * check the target process has a UID that matches the current process's
3668 static bool check_same_owner(struct task_struct
*p
)
3670 const struct cred
*cred
= current_cred(), *pcred
;
3674 pcred
= __task_cred(p
);
3675 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3676 uid_eq(cred
->euid
, pcred
->uid
));
3681 static bool dl_param_changed(struct task_struct
*p
,
3682 const struct sched_attr
*attr
)
3684 struct sched_dl_entity
*dl_se
= &p
->dl
;
3686 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3687 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3688 dl_se
->dl_period
!= attr
->sched_period
||
3689 dl_se
->flags
!= attr
->sched_flags
)
3695 static int __sched_setscheduler(struct task_struct
*p
,
3696 const struct sched_attr
*attr
,
3699 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3700 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3701 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3702 int new_effective_prio
, policy
= attr
->sched_policy
;
3703 unsigned long flags
;
3704 const struct sched_class
*prev_class
;
3708 /* may grab non-irq protected spin_locks */
3709 BUG_ON(in_interrupt());
3711 /* double check policy once rq lock held */
3713 reset_on_fork
= p
->sched_reset_on_fork
;
3714 policy
= oldpolicy
= p
->policy
;
3716 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3718 if (policy
!= SCHED_DEADLINE
&&
3719 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3720 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3721 policy
!= SCHED_IDLE
)
3725 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3729 * Valid priorities for SCHED_FIFO and SCHED_RR are
3730 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3731 * SCHED_BATCH and SCHED_IDLE is 0.
3733 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3734 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3736 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3737 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3741 * Allow unprivileged RT tasks to decrease priority:
3743 if (user
&& !capable(CAP_SYS_NICE
)) {
3744 if (fair_policy(policy
)) {
3745 if (attr
->sched_nice
< task_nice(p
) &&
3746 !can_nice(p
, attr
->sched_nice
))
3750 if (rt_policy(policy
)) {
3751 unsigned long rlim_rtprio
=
3752 task_rlimit(p
, RLIMIT_RTPRIO
);
3754 /* can't set/change the rt policy */
3755 if (policy
!= p
->policy
&& !rlim_rtprio
)
3758 /* can't increase priority */
3759 if (attr
->sched_priority
> p
->rt_priority
&&
3760 attr
->sched_priority
> rlim_rtprio
)
3765 * Can't set/change SCHED_DEADLINE policy at all for now
3766 * (safest behavior); in the future we would like to allow
3767 * unprivileged DL tasks to increase their relative deadline
3768 * or reduce their runtime (both ways reducing utilization)
3770 if (dl_policy(policy
))
3774 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3775 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3777 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3778 if (!can_nice(p
, task_nice(p
)))
3782 /* can't change other user's priorities */
3783 if (!check_same_owner(p
))
3786 /* Normal users shall not reset the sched_reset_on_fork flag */
3787 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3792 retval
= security_task_setscheduler(p
);
3798 * make sure no PI-waiters arrive (or leave) while we are
3799 * changing the priority of the task:
3801 * To be able to change p->policy safely, the appropriate
3802 * runqueue lock must be held.
3804 rq
= task_rq_lock(p
, &flags
);
3807 * Changing the policy of the stop threads its a very bad idea
3809 if (p
== rq
->stop
) {
3810 task_rq_unlock(rq
, p
, &flags
);
3815 * If not changing anything there's no need to proceed further,
3816 * but store a possible modification of reset_on_fork.
3818 if (unlikely(policy
== p
->policy
)) {
3819 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3821 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3823 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3826 p
->sched_reset_on_fork
= reset_on_fork
;
3827 task_rq_unlock(rq
, p
, &flags
);
3833 #ifdef CONFIG_RT_GROUP_SCHED
3835 * Do not allow realtime tasks into groups that have no runtime
3838 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3839 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3840 !task_group_is_autogroup(task_group(p
))) {
3841 task_rq_unlock(rq
, p
, &flags
);
3846 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3847 cpumask_t
*span
= rq
->rd
->span
;
3850 * Don't allow tasks with an affinity mask smaller than
3851 * the entire root_domain to become SCHED_DEADLINE. We
3852 * will also fail if there's no bandwidth available.
3854 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3855 rq
->rd
->dl_bw
.bw
== 0) {
3856 task_rq_unlock(rq
, p
, &flags
);
3863 /* recheck policy now with rq lock held */
3864 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3865 policy
= oldpolicy
= -1;
3866 task_rq_unlock(rq
, p
, &flags
);
3871 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3872 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3875 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3876 task_rq_unlock(rq
, p
, &flags
);
3880 p
->sched_reset_on_fork
= reset_on_fork
;
3885 * Take priority boosted tasks into account. If the new
3886 * effective priority is unchanged, we just store the new
3887 * normal parameters and do not touch the scheduler class and
3888 * the runqueue. This will be done when the task deboost
3891 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
3892 if (new_effective_prio
== oldprio
) {
3893 __setscheduler_params(p
, attr
);
3894 task_rq_unlock(rq
, p
, &flags
);
3899 queued
= task_on_rq_queued(p
);
3900 running
= task_current(rq
, p
);
3902 dequeue_task(rq
, p
, 0);
3904 put_prev_task(rq
, p
);
3906 prev_class
= p
->sched_class
;
3907 __setscheduler(rq
, p
, attr
, pi
);
3910 p
->sched_class
->set_curr_task(rq
);
3913 * We enqueue to tail when the priority of a task is
3914 * increased (user space view).
3916 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3919 check_class_changed(rq
, p
, prev_class
, oldprio
);
3920 preempt_disable(); /* avoid rq from going away on us */
3921 task_rq_unlock(rq
, p
, &flags
);
3924 rt_mutex_adjust_pi(p
);
3927 * Run balance callbacks after we've adjusted the PI chain.
3929 balance_callback(rq
);
3935 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3936 const struct sched_param
*param
, bool check
)
3938 struct sched_attr attr
= {
3939 .sched_policy
= policy
,
3940 .sched_priority
= param
->sched_priority
,
3941 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3944 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3945 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3946 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3947 policy
&= ~SCHED_RESET_ON_FORK
;
3948 attr
.sched_policy
= policy
;
3951 return __sched_setscheduler(p
, &attr
, check
, true);
3954 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3955 * @p: the task in question.
3956 * @policy: new policy.
3957 * @param: structure containing the new RT priority.
3959 * Return: 0 on success. An error code otherwise.
3961 * NOTE that the task may be already dead.
3963 int sched_setscheduler(struct task_struct
*p
, int policy
,
3964 const struct sched_param
*param
)
3966 return _sched_setscheduler(p
, policy
, param
, true);
3968 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3970 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3972 return __sched_setscheduler(p
, attr
, true, true);
3974 EXPORT_SYMBOL_GPL(sched_setattr
);
3977 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3978 * @p: the task in question.
3979 * @policy: new policy.
3980 * @param: structure containing the new RT priority.
3982 * Just like sched_setscheduler, only don't bother checking if the
3983 * current context has permission. For example, this is needed in
3984 * stop_machine(): we create temporary high priority worker threads,
3985 * but our caller might not have that capability.
3987 * Return: 0 on success. An error code otherwise.
3989 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3990 const struct sched_param
*param
)
3992 return _sched_setscheduler(p
, policy
, param
, false);
3996 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3998 struct sched_param lparam
;
3999 struct task_struct
*p
;
4002 if (!param
|| pid
< 0)
4004 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4009 p
= find_process_by_pid(pid
);
4011 retval
= sched_setscheduler(p
, policy
, &lparam
);
4018 * Mimics kernel/events/core.c perf_copy_attr().
4020 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4021 struct sched_attr
*attr
)
4026 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4030 * zero the full structure, so that a short copy will be nice.
4032 memset(attr
, 0, sizeof(*attr
));
4034 ret
= get_user(size
, &uattr
->size
);
4038 if (size
> PAGE_SIZE
) /* silly large */
4041 if (!size
) /* abi compat */
4042 size
= SCHED_ATTR_SIZE_VER0
;
4044 if (size
< SCHED_ATTR_SIZE_VER0
)
4048 * If we're handed a bigger struct than we know of,
4049 * ensure all the unknown bits are 0 - i.e. new
4050 * user-space does not rely on any kernel feature
4051 * extensions we dont know about yet.
4053 if (size
> sizeof(*attr
)) {
4054 unsigned char __user
*addr
;
4055 unsigned char __user
*end
;
4058 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4059 end
= (void __user
*)uattr
+ size
;
4061 for (; addr
< end
; addr
++) {
4062 ret
= get_user(val
, addr
);
4068 size
= sizeof(*attr
);
4071 ret
= copy_from_user(attr
, uattr
, size
);
4076 * XXX: do we want to be lenient like existing syscalls; or do we want
4077 * to be strict and return an error on out-of-bounds values?
4079 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4084 put_user(sizeof(*attr
), &uattr
->size
);
4089 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4090 * @pid: the pid in question.
4091 * @policy: new policy.
4092 * @param: structure containing the new RT priority.
4094 * Return: 0 on success. An error code otherwise.
4096 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4097 struct sched_param __user
*, param
)
4099 /* negative values for policy are not valid */
4103 return do_sched_setscheduler(pid
, policy
, param
);
4107 * sys_sched_setparam - set/change the RT priority of a thread
4108 * @pid: the pid in question.
4109 * @param: structure containing the new RT priority.
4111 * Return: 0 on success. An error code otherwise.
4113 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4115 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4119 * sys_sched_setattr - same as above, but with extended sched_attr
4120 * @pid: the pid in question.
4121 * @uattr: structure containing the extended parameters.
4122 * @flags: for future extension.
4124 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4125 unsigned int, flags
)
4127 struct sched_attr attr
;
4128 struct task_struct
*p
;
4131 if (!uattr
|| pid
< 0 || flags
)
4134 retval
= sched_copy_attr(uattr
, &attr
);
4138 if ((int)attr
.sched_policy
< 0)
4143 p
= find_process_by_pid(pid
);
4145 retval
= sched_setattr(p
, &attr
);
4152 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4153 * @pid: the pid in question.
4155 * Return: On success, the policy of the thread. Otherwise, a negative error
4158 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4160 struct task_struct
*p
;
4168 p
= find_process_by_pid(pid
);
4170 retval
= security_task_getscheduler(p
);
4173 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4180 * sys_sched_getparam - get the RT priority of a thread
4181 * @pid: the pid in question.
4182 * @param: structure containing the RT priority.
4184 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4187 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4189 struct sched_param lp
= { .sched_priority
= 0 };
4190 struct task_struct
*p
;
4193 if (!param
|| pid
< 0)
4197 p
= find_process_by_pid(pid
);
4202 retval
= security_task_getscheduler(p
);
4206 if (task_has_rt_policy(p
))
4207 lp
.sched_priority
= p
->rt_priority
;
4211 * This one might sleep, we cannot do it with a spinlock held ...
4213 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4222 static int sched_read_attr(struct sched_attr __user
*uattr
,
4223 struct sched_attr
*attr
,
4228 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4232 * If we're handed a smaller struct than we know of,
4233 * ensure all the unknown bits are 0 - i.e. old
4234 * user-space does not get uncomplete information.
4236 if (usize
< sizeof(*attr
)) {
4237 unsigned char *addr
;
4240 addr
= (void *)attr
+ usize
;
4241 end
= (void *)attr
+ sizeof(*attr
);
4243 for (; addr
< end
; addr
++) {
4251 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4259 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4260 * @pid: the pid in question.
4261 * @uattr: structure containing the extended parameters.
4262 * @size: sizeof(attr) for fwd/bwd comp.
4263 * @flags: for future extension.
4265 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4266 unsigned int, size
, unsigned int, flags
)
4268 struct sched_attr attr
= {
4269 .size
= sizeof(struct sched_attr
),
4271 struct task_struct
*p
;
4274 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4275 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4279 p
= find_process_by_pid(pid
);
4284 retval
= security_task_getscheduler(p
);
4288 attr
.sched_policy
= p
->policy
;
4289 if (p
->sched_reset_on_fork
)
4290 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4291 if (task_has_dl_policy(p
))
4292 __getparam_dl(p
, &attr
);
4293 else if (task_has_rt_policy(p
))
4294 attr
.sched_priority
= p
->rt_priority
;
4296 attr
.sched_nice
= task_nice(p
);
4300 retval
= sched_read_attr(uattr
, &attr
, size
);
4308 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4310 cpumask_var_t cpus_allowed
, new_mask
;
4311 struct task_struct
*p
;
4316 p
= find_process_by_pid(pid
);
4322 /* Prevent p going away */
4326 if (p
->flags
& PF_NO_SETAFFINITY
) {
4330 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4334 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4336 goto out_free_cpus_allowed
;
4339 if (!check_same_owner(p
)) {
4341 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4343 goto out_free_new_mask
;
4348 retval
= security_task_setscheduler(p
);
4350 goto out_free_new_mask
;
4353 cpuset_cpus_allowed(p
, cpus_allowed
);
4354 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4357 * Since bandwidth control happens on root_domain basis,
4358 * if admission test is enabled, we only admit -deadline
4359 * tasks allowed to run on all the CPUs in the task's
4363 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4365 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4368 goto out_free_new_mask
;
4374 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4377 cpuset_cpus_allowed(p
, cpus_allowed
);
4378 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4380 * We must have raced with a concurrent cpuset
4381 * update. Just reset the cpus_allowed to the
4382 * cpuset's cpus_allowed
4384 cpumask_copy(new_mask
, cpus_allowed
);
4389 free_cpumask_var(new_mask
);
4390 out_free_cpus_allowed
:
4391 free_cpumask_var(cpus_allowed
);
4397 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4398 struct cpumask
*new_mask
)
4400 if (len
< cpumask_size())
4401 cpumask_clear(new_mask
);
4402 else if (len
> cpumask_size())
4403 len
= cpumask_size();
4405 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4409 * sys_sched_setaffinity - set the cpu affinity of a process
4410 * @pid: pid of the process
4411 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4412 * @user_mask_ptr: user-space pointer to the new cpu mask
4414 * Return: 0 on success. An error code otherwise.
4416 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4417 unsigned long __user
*, user_mask_ptr
)
4419 cpumask_var_t new_mask
;
4422 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4425 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4427 retval
= sched_setaffinity(pid
, new_mask
);
4428 free_cpumask_var(new_mask
);
4432 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4434 struct task_struct
*p
;
4435 unsigned long flags
;
4441 p
= find_process_by_pid(pid
);
4445 retval
= security_task_getscheduler(p
);
4449 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4450 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4451 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4460 * sys_sched_getaffinity - get the cpu affinity of a process
4461 * @pid: pid of the process
4462 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4463 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4465 * Return: 0 on success. An error code otherwise.
4467 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4468 unsigned long __user
*, user_mask_ptr
)
4473 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4475 if (len
& (sizeof(unsigned long)-1))
4478 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4481 ret
= sched_getaffinity(pid
, mask
);
4483 size_t retlen
= min_t(size_t, len
, cpumask_size());
4485 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4490 free_cpumask_var(mask
);
4496 * sys_sched_yield - yield the current processor to other threads.
4498 * This function yields the current CPU to other tasks. If there are no
4499 * other threads running on this CPU then this function will return.
4503 SYSCALL_DEFINE0(sched_yield
)
4505 struct rq
*rq
= this_rq_lock();
4507 schedstat_inc(rq
, yld_count
);
4508 current
->sched_class
->yield_task(rq
);
4511 * Since we are going to call schedule() anyway, there's
4512 * no need to preempt or enable interrupts:
4514 __release(rq
->lock
);
4515 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4516 do_raw_spin_unlock(&rq
->lock
);
4517 sched_preempt_enable_no_resched();
4524 int __sched
_cond_resched(void)
4526 if (should_resched(0)) {
4527 preempt_schedule_common();
4532 EXPORT_SYMBOL(_cond_resched
);
4535 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4536 * call schedule, and on return reacquire the lock.
4538 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4539 * operations here to prevent schedule() from being called twice (once via
4540 * spin_unlock(), once by hand).
4542 int __cond_resched_lock(spinlock_t
*lock
)
4544 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4547 lockdep_assert_held(lock
);
4549 if (spin_needbreak(lock
) || resched
) {
4552 preempt_schedule_common();
4560 EXPORT_SYMBOL(__cond_resched_lock
);
4562 int __sched
__cond_resched_softirq(void)
4564 BUG_ON(!in_softirq());
4566 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4568 preempt_schedule_common();
4574 EXPORT_SYMBOL(__cond_resched_softirq
);
4577 * yield - yield the current processor to other threads.
4579 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4581 * The scheduler is at all times free to pick the calling task as the most
4582 * eligible task to run, if removing the yield() call from your code breaks
4583 * it, its already broken.
4585 * Typical broken usage is:
4590 * where one assumes that yield() will let 'the other' process run that will
4591 * make event true. If the current task is a SCHED_FIFO task that will never
4592 * happen. Never use yield() as a progress guarantee!!
4594 * If you want to use yield() to wait for something, use wait_event().
4595 * If you want to use yield() to be 'nice' for others, use cond_resched().
4596 * If you still want to use yield(), do not!
4598 void __sched
yield(void)
4600 set_current_state(TASK_RUNNING
);
4603 EXPORT_SYMBOL(yield
);
4606 * yield_to - yield the current processor to another thread in
4607 * your thread group, or accelerate that thread toward the
4608 * processor it's on.
4610 * @preempt: whether task preemption is allowed or not
4612 * It's the caller's job to ensure that the target task struct
4613 * can't go away on us before we can do any checks.
4616 * true (>0) if we indeed boosted the target task.
4617 * false (0) if we failed to boost the target.
4618 * -ESRCH if there's no task to yield to.
4620 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4622 struct task_struct
*curr
= current
;
4623 struct rq
*rq
, *p_rq
;
4624 unsigned long flags
;
4627 local_irq_save(flags
);
4633 * If we're the only runnable task on the rq and target rq also
4634 * has only one task, there's absolutely no point in yielding.
4636 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4641 double_rq_lock(rq
, p_rq
);
4642 if (task_rq(p
) != p_rq
) {
4643 double_rq_unlock(rq
, p_rq
);
4647 if (!curr
->sched_class
->yield_to_task
)
4650 if (curr
->sched_class
!= p
->sched_class
)
4653 if (task_running(p_rq
, p
) || p
->state
)
4656 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4658 schedstat_inc(rq
, yld_count
);
4660 * Make p's CPU reschedule; pick_next_entity takes care of
4663 if (preempt
&& rq
!= p_rq
)
4668 double_rq_unlock(rq
, p_rq
);
4670 local_irq_restore(flags
);
4677 EXPORT_SYMBOL_GPL(yield_to
);
4680 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4681 * that process accounting knows that this is a task in IO wait state.
4683 long __sched
io_schedule_timeout(long timeout
)
4685 int old_iowait
= current
->in_iowait
;
4689 current
->in_iowait
= 1;
4690 blk_schedule_flush_plug(current
);
4692 delayacct_blkio_start();
4694 atomic_inc(&rq
->nr_iowait
);
4695 ret
= schedule_timeout(timeout
);
4696 current
->in_iowait
= old_iowait
;
4697 atomic_dec(&rq
->nr_iowait
);
4698 delayacct_blkio_end();
4702 EXPORT_SYMBOL(io_schedule_timeout
);
4705 * sys_sched_get_priority_max - return maximum RT priority.
4706 * @policy: scheduling class.
4708 * Return: On success, this syscall returns the maximum
4709 * rt_priority that can be used by a given scheduling class.
4710 * On failure, a negative error code is returned.
4712 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4719 ret
= MAX_USER_RT_PRIO
-1;
4721 case SCHED_DEADLINE
:
4732 * sys_sched_get_priority_min - return minimum RT priority.
4733 * @policy: scheduling class.
4735 * Return: On success, this syscall returns the minimum
4736 * rt_priority that can be used by a given scheduling class.
4737 * On failure, a negative error code is returned.
4739 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4748 case SCHED_DEADLINE
:
4758 * sys_sched_rr_get_interval - return the default timeslice of a process.
4759 * @pid: pid of the process.
4760 * @interval: userspace pointer to the timeslice value.
4762 * this syscall writes the default timeslice value of a given process
4763 * into the user-space timespec buffer. A value of '0' means infinity.
4765 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4768 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4769 struct timespec __user
*, interval
)
4771 struct task_struct
*p
;
4772 unsigned int time_slice
;
4773 unsigned long flags
;
4783 p
= find_process_by_pid(pid
);
4787 retval
= security_task_getscheduler(p
);
4791 rq
= task_rq_lock(p
, &flags
);
4793 if (p
->sched_class
->get_rr_interval
)
4794 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4795 task_rq_unlock(rq
, p
, &flags
);
4798 jiffies_to_timespec(time_slice
, &t
);
4799 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4807 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4809 void sched_show_task(struct task_struct
*p
)
4811 unsigned long free
= 0;
4813 unsigned long state
= p
->state
;
4816 state
= __ffs(state
) + 1;
4817 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4818 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4819 #if BITS_PER_LONG == 32
4820 if (state
== TASK_RUNNING
)
4821 printk(KERN_CONT
" running ");
4823 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4825 if (state
== TASK_RUNNING
)
4826 printk(KERN_CONT
" running task ");
4828 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4830 #ifdef CONFIG_DEBUG_STACK_USAGE
4831 free
= stack_not_used(p
);
4836 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4838 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4839 task_pid_nr(p
), ppid
,
4840 (unsigned long)task_thread_info(p
)->flags
);
4842 print_worker_info(KERN_INFO
, p
);
4843 show_stack(p
, NULL
);
4846 void show_state_filter(unsigned long state_filter
)
4848 struct task_struct
*g
, *p
;
4850 #if BITS_PER_LONG == 32
4852 " task PC stack pid father\n");
4855 " task PC stack pid father\n");
4858 for_each_process_thread(g
, p
) {
4860 * reset the NMI-timeout, listing all files on a slow
4861 * console might take a lot of time:
4863 touch_nmi_watchdog();
4864 if (!state_filter
|| (p
->state
& state_filter
))
4868 touch_all_softlockup_watchdogs();
4870 #ifdef CONFIG_SCHED_DEBUG
4871 sysrq_sched_debug_show();
4875 * Only show locks if all tasks are dumped:
4878 debug_show_all_locks();
4881 void init_idle_bootup_task(struct task_struct
*idle
)
4883 idle
->sched_class
= &idle_sched_class
;
4887 * init_idle - set up an idle thread for a given CPU
4888 * @idle: task in question
4889 * @cpu: cpu the idle task belongs to
4891 * NOTE: this function does not set the idle thread's NEED_RESCHED
4892 * flag, to make booting more robust.
4894 void init_idle(struct task_struct
*idle
, int cpu
)
4896 struct rq
*rq
= cpu_rq(cpu
);
4897 unsigned long flags
;
4899 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
4900 raw_spin_lock(&rq
->lock
);
4902 __sched_fork(0, idle
);
4903 idle
->state
= TASK_RUNNING
;
4904 idle
->se
.exec_start
= sched_clock();
4906 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4908 * We're having a chicken and egg problem, even though we are
4909 * holding rq->lock, the cpu isn't yet set to this cpu so the
4910 * lockdep check in task_group() will fail.
4912 * Similar case to sched_fork(). / Alternatively we could
4913 * use task_rq_lock() here and obtain the other rq->lock.
4918 __set_task_cpu(idle
, cpu
);
4921 rq
->curr
= rq
->idle
= idle
;
4922 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4923 #if defined(CONFIG_SMP)
4926 raw_spin_unlock(&rq
->lock
);
4927 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
4929 /* Set the preempt count _outside_ the spinlocks! */
4930 init_idle_preempt_count(idle
, cpu
);
4933 * The idle tasks have their own, simple scheduling class:
4935 idle
->sched_class
= &idle_sched_class
;
4936 ftrace_graph_init_idle_task(idle
, cpu
);
4937 vtime_init_idle(idle
, cpu
);
4938 #if defined(CONFIG_SMP)
4939 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4943 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4944 const struct cpumask
*trial
)
4946 int ret
= 1, trial_cpus
;
4947 struct dl_bw
*cur_dl_b
;
4948 unsigned long flags
;
4950 if (!cpumask_weight(cur
))
4953 rcu_read_lock_sched();
4954 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4955 trial_cpus
= cpumask_weight(trial
);
4957 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4958 if (cur_dl_b
->bw
!= -1 &&
4959 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
4961 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
4962 rcu_read_unlock_sched();
4967 int task_can_attach(struct task_struct
*p
,
4968 const struct cpumask
*cs_cpus_allowed
)
4973 * Kthreads which disallow setaffinity shouldn't be moved
4974 * to a new cpuset; we don't want to change their cpu
4975 * affinity and isolating such threads by their set of
4976 * allowed nodes is unnecessary. Thus, cpusets are not
4977 * applicable for such threads. This prevents checking for
4978 * success of set_cpus_allowed_ptr() on all attached tasks
4979 * before cpus_allowed may be changed.
4981 if (p
->flags
& PF_NO_SETAFFINITY
) {
4987 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
4989 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
4994 unsigned long flags
;
4996 rcu_read_lock_sched();
4997 dl_b
= dl_bw_of(dest_cpu
);
4998 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
4999 cpus
= dl_bw_cpus(dest_cpu
);
5000 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5005 * We reserve space for this task in the destination
5006 * root_domain, as we can't fail after this point.
5007 * We will free resources in the source root_domain
5008 * later on (see set_cpus_allowed_dl()).
5010 __dl_add(dl_b
, p
->dl
.dl_bw
);
5012 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5013 rcu_read_unlock_sched();
5023 #ifdef CONFIG_NUMA_BALANCING
5024 /* Migrate current task p to target_cpu */
5025 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5027 struct migration_arg arg
= { p
, target_cpu
};
5028 int curr_cpu
= task_cpu(p
);
5030 if (curr_cpu
== target_cpu
)
5033 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5036 /* TODO: This is not properly updating schedstats */
5038 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5039 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5043 * Requeue a task on a given node and accurately track the number of NUMA
5044 * tasks on the runqueues
5046 void sched_setnuma(struct task_struct
*p
, int nid
)
5049 unsigned long flags
;
5050 bool queued
, running
;
5052 rq
= task_rq_lock(p
, &flags
);
5053 queued
= task_on_rq_queued(p
);
5054 running
= task_current(rq
, p
);
5057 dequeue_task(rq
, p
, 0);
5059 put_prev_task(rq
, p
);
5061 p
->numa_preferred_nid
= nid
;
5064 p
->sched_class
->set_curr_task(rq
);
5066 enqueue_task(rq
, p
, 0);
5067 task_rq_unlock(rq
, p
, &flags
);
5069 #endif /* CONFIG_NUMA_BALANCING */
5071 #ifdef CONFIG_HOTPLUG_CPU
5073 * Ensures that the idle task is using init_mm right before its cpu goes
5076 void idle_task_exit(void)
5078 struct mm_struct
*mm
= current
->active_mm
;
5080 BUG_ON(cpu_online(smp_processor_id()));
5082 if (mm
!= &init_mm
) {
5083 switch_mm(mm
, &init_mm
, current
);
5084 finish_arch_post_lock_switch();
5090 * Since this CPU is going 'away' for a while, fold any nr_active delta
5091 * we might have. Assumes we're called after migrate_tasks() so that the
5092 * nr_active count is stable.
5094 * Also see the comment "Global load-average calculations".
5096 static void calc_load_migrate(struct rq
*rq
)
5098 long delta
= calc_load_fold_active(rq
);
5100 atomic_long_add(delta
, &calc_load_tasks
);
5103 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5107 static const struct sched_class fake_sched_class
= {
5108 .put_prev_task
= put_prev_task_fake
,
5111 static struct task_struct fake_task
= {
5113 * Avoid pull_{rt,dl}_task()
5115 .prio
= MAX_PRIO
+ 1,
5116 .sched_class
= &fake_sched_class
,
5120 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5121 * try_to_wake_up()->select_task_rq().
5123 * Called with rq->lock held even though we'er in stop_machine() and
5124 * there's no concurrency possible, we hold the required locks anyway
5125 * because of lock validation efforts.
5127 static void migrate_tasks(struct rq
*dead_rq
)
5129 struct rq
*rq
= dead_rq
;
5130 struct task_struct
*next
, *stop
= rq
->stop
;
5134 * Fudge the rq selection such that the below task selection loop
5135 * doesn't get stuck on the currently eligible stop task.
5137 * We're currently inside stop_machine() and the rq is either stuck
5138 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5139 * either way we should never end up calling schedule() until we're
5145 * put_prev_task() and pick_next_task() sched
5146 * class method both need to have an up-to-date
5147 * value of rq->clock[_task]
5149 update_rq_clock(rq
);
5153 * There's this thread running, bail when that's the only
5156 if (rq
->nr_running
== 1)
5160 * Ensure rq->lock covers the entire task selection
5161 * until the migration.
5163 lockdep_pin_lock(&rq
->lock
);
5164 next
= pick_next_task(rq
, &fake_task
);
5166 next
->sched_class
->put_prev_task(rq
, next
);
5168 /* Find suitable destination for @next, with force if needed. */
5169 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5171 lockdep_unpin_lock(&rq
->lock
);
5172 rq
= __migrate_task(rq
, next
, dest_cpu
);
5173 if (rq
!= dead_rq
) {
5174 raw_spin_unlock(&rq
->lock
);
5176 raw_spin_lock(&rq
->lock
);
5182 #endif /* CONFIG_HOTPLUG_CPU */
5184 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5186 static struct ctl_table sd_ctl_dir
[] = {
5188 .procname
= "sched_domain",
5194 static struct ctl_table sd_ctl_root
[] = {
5196 .procname
= "kernel",
5198 .child
= sd_ctl_dir
,
5203 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5205 struct ctl_table
*entry
=
5206 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5211 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5213 struct ctl_table
*entry
;
5216 * In the intermediate directories, both the child directory and
5217 * procname are dynamically allocated and could fail but the mode
5218 * will always be set. In the lowest directory the names are
5219 * static strings and all have proc handlers.
5221 for (entry
= *tablep
; entry
->mode
; entry
++) {
5223 sd_free_ctl_entry(&entry
->child
);
5224 if (entry
->proc_handler
== NULL
)
5225 kfree(entry
->procname
);
5232 static int min_load_idx
= 0;
5233 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5236 set_table_entry(struct ctl_table
*entry
,
5237 const char *procname
, void *data
, int maxlen
,
5238 umode_t mode
, proc_handler
*proc_handler
,
5241 entry
->procname
= procname
;
5243 entry
->maxlen
= maxlen
;
5245 entry
->proc_handler
= proc_handler
;
5248 entry
->extra1
= &min_load_idx
;
5249 entry
->extra2
= &max_load_idx
;
5253 static struct ctl_table
*
5254 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5256 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5261 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5262 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5263 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5264 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5265 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5266 sizeof(int), 0644, proc_dointvec_minmax
, true);
5267 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5268 sizeof(int), 0644, proc_dointvec_minmax
, true);
5269 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5270 sizeof(int), 0644, proc_dointvec_minmax
, true);
5271 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5272 sizeof(int), 0644, proc_dointvec_minmax
, true);
5273 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5274 sizeof(int), 0644, proc_dointvec_minmax
, true);
5275 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5276 sizeof(int), 0644, proc_dointvec_minmax
, false);
5277 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5278 sizeof(int), 0644, proc_dointvec_minmax
, false);
5279 set_table_entry(&table
[9], "cache_nice_tries",
5280 &sd
->cache_nice_tries
,
5281 sizeof(int), 0644, proc_dointvec_minmax
, false);
5282 set_table_entry(&table
[10], "flags", &sd
->flags
,
5283 sizeof(int), 0644, proc_dointvec_minmax
, false);
5284 set_table_entry(&table
[11], "max_newidle_lb_cost",
5285 &sd
->max_newidle_lb_cost
,
5286 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5287 set_table_entry(&table
[12], "name", sd
->name
,
5288 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5289 /* &table[13] is terminator */
5294 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5296 struct ctl_table
*entry
, *table
;
5297 struct sched_domain
*sd
;
5298 int domain_num
= 0, i
;
5301 for_each_domain(cpu
, sd
)
5303 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5308 for_each_domain(cpu
, sd
) {
5309 snprintf(buf
, 32, "domain%d", i
);
5310 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5312 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5319 static struct ctl_table_header
*sd_sysctl_header
;
5320 static void register_sched_domain_sysctl(void)
5322 int i
, cpu_num
= num_possible_cpus();
5323 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5326 WARN_ON(sd_ctl_dir
[0].child
);
5327 sd_ctl_dir
[0].child
= entry
;
5332 for_each_possible_cpu(i
) {
5333 snprintf(buf
, 32, "cpu%d", i
);
5334 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5336 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5340 WARN_ON(sd_sysctl_header
);
5341 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5344 /* may be called multiple times per register */
5345 static void unregister_sched_domain_sysctl(void)
5347 unregister_sysctl_table(sd_sysctl_header
);
5348 sd_sysctl_header
= NULL
;
5349 if (sd_ctl_dir
[0].child
)
5350 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5353 static void register_sched_domain_sysctl(void)
5356 static void unregister_sched_domain_sysctl(void)
5359 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5361 static void set_rq_online(struct rq
*rq
)
5364 const struct sched_class
*class;
5366 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5369 for_each_class(class) {
5370 if (class->rq_online
)
5371 class->rq_online(rq
);
5376 static void set_rq_offline(struct rq
*rq
)
5379 const struct sched_class
*class;
5381 for_each_class(class) {
5382 if (class->rq_offline
)
5383 class->rq_offline(rq
);
5386 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5392 * migration_call - callback that gets triggered when a CPU is added.
5393 * Here we can start up the necessary migration thread for the new CPU.
5396 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5398 int cpu
= (long)hcpu
;
5399 unsigned long flags
;
5400 struct rq
*rq
= cpu_rq(cpu
);
5402 switch (action
& ~CPU_TASKS_FROZEN
) {
5404 case CPU_UP_PREPARE
:
5405 rq
->calc_load_update
= calc_load_update
;
5409 /* Update our root-domain */
5410 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5412 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5416 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5419 #ifdef CONFIG_HOTPLUG_CPU
5421 sched_ttwu_pending();
5422 /* Update our root-domain */
5423 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5425 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5429 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5430 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5434 calc_load_migrate(rq
);
5439 update_max_interval();
5445 * Register at high priority so that task migration (migrate_all_tasks)
5446 * happens before everything else. This has to be lower priority than
5447 * the notifier in the perf_event subsystem, though.
5449 static struct notifier_block migration_notifier
= {
5450 .notifier_call
= migration_call
,
5451 .priority
= CPU_PRI_MIGRATION
,
5454 static void set_cpu_rq_start_time(void)
5456 int cpu
= smp_processor_id();
5457 struct rq
*rq
= cpu_rq(cpu
);
5458 rq
->age_stamp
= sched_clock_cpu(cpu
);
5461 static int sched_cpu_active(struct notifier_block
*nfb
,
5462 unsigned long action
, void *hcpu
)
5464 switch (action
& ~CPU_TASKS_FROZEN
) {
5466 set_cpu_rq_start_time();
5468 case CPU_DOWN_FAILED
:
5469 set_cpu_active((long)hcpu
, true);
5476 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5477 unsigned long action
, void *hcpu
)
5479 switch (action
& ~CPU_TASKS_FROZEN
) {
5480 case CPU_DOWN_PREPARE
:
5481 set_cpu_active((long)hcpu
, false);
5488 static int __init
migration_init(void)
5490 void *cpu
= (void *)(long)smp_processor_id();
5493 /* Initialize migration for the boot CPU */
5494 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5495 BUG_ON(err
== NOTIFY_BAD
);
5496 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5497 register_cpu_notifier(&migration_notifier
);
5499 /* Register cpu active notifiers */
5500 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5501 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5505 early_initcall(migration_init
);
5507 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5509 #ifdef CONFIG_SCHED_DEBUG
5511 static __read_mostly
int sched_debug_enabled
;
5513 static int __init
sched_debug_setup(char *str
)
5515 sched_debug_enabled
= 1;
5519 early_param("sched_debug", sched_debug_setup
);
5521 static inline bool sched_debug(void)
5523 return sched_debug_enabled
;
5526 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5527 struct cpumask
*groupmask
)
5529 struct sched_group
*group
= sd
->groups
;
5531 cpumask_clear(groupmask
);
5533 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5535 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5536 printk("does not load-balance\n");
5538 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5543 printk(KERN_CONT
"span %*pbl level %s\n",
5544 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5546 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5547 printk(KERN_ERR
"ERROR: domain->span does not contain "
5550 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5551 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5555 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5559 printk(KERN_ERR
"ERROR: group is NULL\n");
5563 if (!cpumask_weight(sched_group_cpus(group
))) {
5564 printk(KERN_CONT
"\n");
5565 printk(KERN_ERR
"ERROR: empty group\n");
5569 if (!(sd
->flags
& SD_OVERLAP
) &&
5570 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5571 printk(KERN_CONT
"\n");
5572 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5576 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5578 printk(KERN_CONT
" %*pbl",
5579 cpumask_pr_args(sched_group_cpus(group
)));
5580 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5581 printk(KERN_CONT
" (cpu_capacity = %d)",
5582 group
->sgc
->capacity
);
5585 group
= group
->next
;
5586 } while (group
!= sd
->groups
);
5587 printk(KERN_CONT
"\n");
5589 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5590 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5593 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5594 printk(KERN_ERR
"ERROR: parent span is not a superset "
5595 "of domain->span\n");
5599 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5603 if (!sched_debug_enabled
)
5607 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5611 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5614 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5622 #else /* !CONFIG_SCHED_DEBUG */
5623 # define sched_domain_debug(sd, cpu) do { } while (0)
5624 static inline bool sched_debug(void)
5628 #endif /* CONFIG_SCHED_DEBUG */
5630 static int sd_degenerate(struct sched_domain
*sd
)
5632 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5635 /* Following flags need at least 2 groups */
5636 if (sd
->flags
& (SD_LOAD_BALANCE
|
5637 SD_BALANCE_NEWIDLE
|
5640 SD_SHARE_CPUCAPACITY
|
5641 SD_SHARE_PKG_RESOURCES
|
5642 SD_SHARE_POWERDOMAIN
)) {
5643 if (sd
->groups
!= sd
->groups
->next
)
5647 /* Following flags don't use groups */
5648 if (sd
->flags
& (SD_WAKE_AFFINE
))
5655 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5657 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5659 if (sd_degenerate(parent
))
5662 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5665 /* Flags needing groups don't count if only 1 group in parent */
5666 if (parent
->groups
== parent
->groups
->next
) {
5667 pflags
&= ~(SD_LOAD_BALANCE
|
5668 SD_BALANCE_NEWIDLE
|
5671 SD_SHARE_CPUCAPACITY
|
5672 SD_SHARE_PKG_RESOURCES
|
5674 SD_SHARE_POWERDOMAIN
);
5675 if (nr_node_ids
== 1)
5676 pflags
&= ~SD_SERIALIZE
;
5678 if (~cflags
& pflags
)
5684 static void free_rootdomain(struct rcu_head
*rcu
)
5686 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5688 cpupri_cleanup(&rd
->cpupri
);
5689 cpudl_cleanup(&rd
->cpudl
);
5690 free_cpumask_var(rd
->dlo_mask
);
5691 free_cpumask_var(rd
->rto_mask
);
5692 free_cpumask_var(rd
->online
);
5693 free_cpumask_var(rd
->span
);
5697 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5699 struct root_domain
*old_rd
= NULL
;
5700 unsigned long flags
;
5702 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5707 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5710 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5713 * If we dont want to free the old_rd yet then
5714 * set old_rd to NULL to skip the freeing later
5717 if (!atomic_dec_and_test(&old_rd
->refcount
))
5721 atomic_inc(&rd
->refcount
);
5724 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5725 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5728 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5731 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5734 static int init_rootdomain(struct root_domain
*rd
)
5736 memset(rd
, 0, sizeof(*rd
));
5738 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5740 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5742 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5744 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5747 init_dl_bw(&rd
->dl_bw
);
5748 if (cpudl_init(&rd
->cpudl
) != 0)
5751 if (cpupri_init(&rd
->cpupri
) != 0)
5756 free_cpumask_var(rd
->rto_mask
);
5758 free_cpumask_var(rd
->dlo_mask
);
5760 free_cpumask_var(rd
->online
);
5762 free_cpumask_var(rd
->span
);
5768 * By default the system creates a single root-domain with all cpus as
5769 * members (mimicking the global state we have today).
5771 struct root_domain def_root_domain
;
5773 static void init_defrootdomain(void)
5775 init_rootdomain(&def_root_domain
);
5777 atomic_set(&def_root_domain
.refcount
, 1);
5780 static struct root_domain
*alloc_rootdomain(void)
5782 struct root_domain
*rd
;
5784 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5788 if (init_rootdomain(rd
) != 0) {
5796 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5798 struct sched_group
*tmp
, *first
;
5807 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5812 } while (sg
!= first
);
5815 static void free_sched_domain(struct rcu_head
*rcu
)
5817 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5820 * If its an overlapping domain it has private groups, iterate and
5823 if (sd
->flags
& SD_OVERLAP
) {
5824 free_sched_groups(sd
->groups
, 1);
5825 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5826 kfree(sd
->groups
->sgc
);
5832 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5834 call_rcu(&sd
->rcu
, free_sched_domain
);
5837 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5839 for (; sd
; sd
= sd
->parent
)
5840 destroy_sched_domain(sd
, cpu
);
5844 * Keep a special pointer to the highest sched_domain that has
5845 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5846 * allows us to avoid some pointer chasing select_idle_sibling().
5848 * Also keep a unique ID per domain (we use the first cpu number in
5849 * the cpumask of the domain), this allows us to quickly tell if
5850 * two cpus are in the same cache domain, see cpus_share_cache().
5852 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5853 DEFINE_PER_CPU(int, sd_llc_size
);
5854 DEFINE_PER_CPU(int, sd_llc_id
);
5855 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5856 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5857 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5859 static void update_top_cache_domain(int cpu
)
5861 struct sched_domain
*sd
;
5862 struct sched_domain
*busy_sd
= NULL
;
5866 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5868 id
= cpumask_first(sched_domain_span(sd
));
5869 size
= cpumask_weight(sched_domain_span(sd
));
5870 busy_sd
= sd
->parent
; /* sd_busy */
5872 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5874 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5875 per_cpu(sd_llc_size
, cpu
) = size
;
5876 per_cpu(sd_llc_id
, cpu
) = id
;
5878 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5879 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5881 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5882 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5886 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5887 * hold the hotplug lock.
5890 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5892 struct rq
*rq
= cpu_rq(cpu
);
5893 struct sched_domain
*tmp
;
5895 /* Remove the sched domains which do not contribute to scheduling. */
5896 for (tmp
= sd
; tmp
; ) {
5897 struct sched_domain
*parent
= tmp
->parent
;
5901 if (sd_parent_degenerate(tmp
, parent
)) {
5902 tmp
->parent
= parent
->parent
;
5904 parent
->parent
->child
= tmp
;
5906 * Transfer SD_PREFER_SIBLING down in case of a
5907 * degenerate parent; the spans match for this
5908 * so the property transfers.
5910 if (parent
->flags
& SD_PREFER_SIBLING
)
5911 tmp
->flags
|= SD_PREFER_SIBLING
;
5912 destroy_sched_domain(parent
, cpu
);
5917 if (sd
&& sd_degenerate(sd
)) {
5920 destroy_sched_domain(tmp
, cpu
);
5925 sched_domain_debug(sd
, cpu
);
5927 rq_attach_root(rq
, rd
);
5929 rcu_assign_pointer(rq
->sd
, sd
);
5930 destroy_sched_domains(tmp
, cpu
);
5932 update_top_cache_domain(cpu
);
5935 /* Setup the mask of cpus configured for isolated domains */
5936 static int __init
isolated_cpu_setup(char *str
)
5938 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5939 cpulist_parse(str
, cpu_isolated_map
);
5943 __setup("isolcpus=", isolated_cpu_setup
);
5946 struct sched_domain
** __percpu sd
;
5947 struct root_domain
*rd
;
5958 * Build an iteration mask that can exclude certain CPUs from the upwards
5961 * Asymmetric node setups can result in situations where the domain tree is of
5962 * unequal depth, make sure to skip domains that already cover the entire
5965 * In that case build_sched_domains() will have terminated the iteration early
5966 * and our sibling sd spans will be empty. Domains should always include the
5967 * cpu they're built on, so check that.
5970 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5972 const struct cpumask
*span
= sched_domain_span(sd
);
5973 struct sd_data
*sdd
= sd
->private;
5974 struct sched_domain
*sibling
;
5977 for_each_cpu(i
, span
) {
5978 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5979 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5982 cpumask_set_cpu(i
, sched_group_mask(sg
));
5987 * Return the canonical balance cpu for this group, this is the first cpu
5988 * of this group that's also in the iteration mask.
5990 int group_balance_cpu(struct sched_group
*sg
)
5992 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5996 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5998 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5999 const struct cpumask
*span
= sched_domain_span(sd
);
6000 struct cpumask
*covered
= sched_domains_tmpmask
;
6001 struct sd_data
*sdd
= sd
->private;
6002 struct sched_domain
*sibling
;
6005 cpumask_clear(covered
);
6007 for_each_cpu(i
, span
) {
6008 struct cpumask
*sg_span
;
6010 if (cpumask_test_cpu(i
, covered
))
6013 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6015 /* See the comment near build_group_mask(). */
6016 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6019 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6020 GFP_KERNEL
, cpu_to_node(cpu
));
6025 sg_span
= sched_group_cpus(sg
);
6027 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6029 cpumask_set_cpu(i
, sg_span
);
6031 cpumask_or(covered
, covered
, sg_span
);
6033 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6034 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6035 build_group_mask(sd
, sg
);
6038 * Initialize sgc->capacity such that even if we mess up the
6039 * domains and no possible iteration will get us here, we won't
6042 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6045 * Make sure the first group of this domain contains the
6046 * canonical balance cpu. Otherwise the sched_domain iteration
6047 * breaks. See update_sg_lb_stats().
6049 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6050 group_balance_cpu(sg
) == cpu
)
6060 sd
->groups
= groups
;
6065 free_sched_groups(first
, 0);
6070 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6072 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6073 struct sched_domain
*child
= sd
->child
;
6076 cpu
= cpumask_first(sched_domain_span(child
));
6079 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6080 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6081 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6088 * build_sched_groups will build a circular linked list of the groups
6089 * covered by the given span, and will set each group's ->cpumask correctly,
6090 * and ->cpu_capacity to 0.
6092 * Assumes the sched_domain tree is fully constructed
6095 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6097 struct sched_group
*first
= NULL
, *last
= NULL
;
6098 struct sd_data
*sdd
= sd
->private;
6099 const struct cpumask
*span
= sched_domain_span(sd
);
6100 struct cpumask
*covered
;
6103 get_group(cpu
, sdd
, &sd
->groups
);
6104 atomic_inc(&sd
->groups
->ref
);
6106 if (cpu
!= cpumask_first(span
))
6109 lockdep_assert_held(&sched_domains_mutex
);
6110 covered
= sched_domains_tmpmask
;
6112 cpumask_clear(covered
);
6114 for_each_cpu(i
, span
) {
6115 struct sched_group
*sg
;
6118 if (cpumask_test_cpu(i
, covered
))
6121 group
= get_group(i
, sdd
, &sg
);
6122 cpumask_setall(sched_group_mask(sg
));
6124 for_each_cpu(j
, span
) {
6125 if (get_group(j
, sdd
, NULL
) != group
)
6128 cpumask_set_cpu(j
, covered
);
6129 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6144 * Initialize sched groups cpu_capacity.
6146 * cpu_capacity indicates the capacity of sched group, which is used while
6147 * distributing the load between different sched groups in a sched domain.
6148 * Typically cpu_capacity for all the groups in a sched domain will be same
6149 * unless there are asymmetries in the topology. If there are asymmetries,
6150 * group having more cpu_capacity will pickup more load compared to the
6151 * group having less cpu_capacity.
6153 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6155 struct sched_group
*sg
= sd
->groups
;
6160 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6162 } while (sg
!= sd
->groups
);
6164 if (cpu
!= group_balance_cpu(sg
))
6167 update_group_capacity(sd
, cpu
);
6168 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6172 * Initializers for schedule domains
6173 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6176 static int default_relax_domain_level
= -1;
6177 int sched_domain_level_max
;
6179 static int __init
setup_relax_domain_level(char *str
)
6181 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6182 pr_warn("Unable to set relax_domain_level\n");
6186 __setup("relax_domain_level=", setup_relax_domain_level
);
6188 static void set_domain_attribute(struct sched_domain
*sd
,
6189 struct sched_domain_attr
*attr
)
6193 if (!attr
|| attr
->relax_domain_level
< 0) {
6194 if (default_relax_domain_level
< 0)
6197 request
= default_relax_domain_level
;
6199 request
= attr
->relax_domain_level
;
6200 if (request
< sd
->level
) {
6201 /* turn off idle balance on this domain */
6202 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6204 /* turn on idle balance on this domain */
6205 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6209 static void __sdt_free(const struct cpumask
*cpu_map
);
6210 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6212 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6213 const struct cpumask
*cpu_map
)
6217 if (!atomic_read(&d
->rd
->refcount
))
6218 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6220 free_percpu(d
->sd
); /* fall through */
6222 __sdt_free(cpu_map
); /* fall through */
6228 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6229 const struct cpumask
*cpu_map
)
6231 memset(d
, 0, sizeof(*d
));
6233 if (__sdt_alloc(cpu_map
))
6234 return sa_sd_storage
;
6235 d
->sd
= alloc_percpu(struct sched_domain
*);
6237 return sa_sd_storage
;
6238 d
->rd
= alloc_rootdomain();
6241 return sa_rootdomain
;
6245 * NULL the sd_data elements we've used to build the sched_domain and
6246 * sched_group structure so that the subsequent __free_domain_allocs()
6247 * will not free the data we're using.
6249 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6251 struct sd_data
*sdd
= sd
->private;
6253 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6254 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6256 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6257 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6259 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6260 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6264 static int sched_domains_numa_levels
;
6265 enum numa_topology_type sched_numa_topology_type
;
6266 static int *sched_domains_numa_distance
;
6267 int sched_max_numa_distance
;
6268 static struct cpumask
***sched_domains_numa_masks
;
6269 static int sched_domains_curr_level
;
6273 * SD_flags allowed in topology descriptions.
6275 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6276 * SD_SHARE_PKG_RESOURCES - describes shared caches
6277 * SD_NUMA - describes NUMA topologies
6278 * SD_SHARE_POWERDOMAIN - describes shared power domain
6281 * SD_ASYM_PACKING - describes SMT quirks
6283 #define TOPOLOGY_SD_FLAGS \
6284 (SD_SHARE_CPUCAPACITY | \
6285 SD_SHARE_PKG_RESOURCES | \
6288 SD_SHARE_POWERDOMAIN)
6290 static struct sched_domain
*
6291 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6293 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6294 int sd_weight
, sd_flags
= 0;
6298 * Ugly hack to pass state to sd_numa_mask()...
6300 sched_domains_curr_level
= tl
->numa_level
;
6303 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6306 sd_flags
= (*tl
->sd_flags
)();
6307 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6308 "wrong sd_flags in topology description\n"))
6309 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6311 *sd
= (struct sched_domain
){
6312 .min_interval
= sd_weight
,
6313 .max_interval
= 2*sd_weight
,
6315 .imbalance_pct
= 125,
6317 .cache_nice_tries
= 0,
6324 .flags
= 1*SD_LOAD_BALANCE
6325 | 1*SD_BALANCE_NEWIDLE
6330 | 0*SD_SHARE_CPUCAPACITY
6331 | 0*SD_SHARE_PKG_RESOURCES
6333 | 0*SD_PREFER_SIBLING
6338 .last_balance
= jiffies
,
6339 .balance_interval
= sd_weight
,
6341 .max_newidle_lb_cost
= 0,
6342 .next_decay_max_lb_cost
= jiffies
,
6343 #ifdef CONFIG_SCHED_DEBUG
6349 * Convert topological properties into behaviour.
6352 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6353 sd
->flags
|= SD_PREFER_SIBLING
;
6354 sd
->imbalance_pct
= 110;
6355 sd
->smt_gain
= 1178; /* ~15% */
6357 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6358 sd
->imbalance_pct
= 117;
6359 sd
->cache_nice_tries
= 1;
6363 } else if (sd
->flags
& SD_NUMA
) {
6364 sd
->cache_nice_tries
= 2;
6368 sd
->flags
|= SD_SERIALIZE
;
6369 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6370 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6377 sd
->flags
|= SD_PREFER_SIBLING
;
6378 sd
->cache_nice_tries
= 1;
6383 sd
->private = &tl
->data
;
6389 * Topology list, bottom-up.
6391 static struct sched_domain_topology_level default_topology
[] = {
6392 #ifdef CONFIG_SCHED_SMT
6393 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6395 #ifdef CONFIG_SCHED_MC
6396 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6398 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6402 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6404 #define for_each_sd_topology(tl) \
6405 for (tl = sched_domain_topology; tl->mask; tl++)
6407 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6409 sched_domain_topology
= tl
;
6414 static const struct cpumask
*sd_numa_mask(int cpu
)
6416 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6419 static void sched_numa_warn(const char *str
)
6421 static int done
= false;
6429 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6431 for (i
= 0; i
< nr_node_ids
; i
++) {
6432 printk(KERN_WARNING
" ");
6433 for (j
= 0; j
< nr_node_ids
; j
++)
6434 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6435 printk(KERN_CONT
"\n");
6437 printk(KERN_WARNING
"\n");
6440 bool find_numa_distance(int distance
)
6444 if (distance
== node_distance(0, 0))
6447 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6448 if (sched_domains_numa_distance
[i
] == distance
)
6456 * A system can have three types of NUMA topology:
6457 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6458 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6459 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6461 * The difference between a glueless mesh topology and a backplane
6462 * topology lies in whether communication between not directly
6463 * connected nodes goes through intermediary nodes (where programs
6464 * could run), or through backplane controllers. This affects
6465 * placement of programs.
6467 * The type of topology can be discerned with the following tests:
6468 * - If the maximum distance between any nodes is 1 hop, the system
6469 * is directly connected.
6470 * - If for two nodes A and B, located N > 1 hops away from each other,
6471 * there is an intermediary node C, which is < N hops away from both
6472 * nodes A and B, the system is a glueless mesh.
6474 static void init_numa_topology_type(void)
6478 n
= sched_max_numa_distance
;
6480 if (sched_domains_numa_levels
<= 1) {
6481 sched_numa_topology_type
= NUMA_DIRECT
;
6485 for_each_online_node(a
) {
6486 for_each_online_node(b
) {
6487 /* Find two nodes furthest removed from each other. */
6488 if (node_distance(a
, b
) < n
)
6491 /* Is there an intermediary node between a and b? */
6492 for_each_online_node(c
) {
6493 if (node_distance(a
, c
) < n
&&
6494 node_distance(b
, c
) < n
) {
6495 sched_numa_topology_type
=
6501 sched_numa_topology_type
= NUMA_BACKPLANE
;
6507 static void sched_init_numa(void)
6509 int next_distance
, curr_distance
= node_distance(0, 0);
6510 struct sched_domain_topology_level
*tl
;
6514 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6515 if (!sched_domains_numa_distance
)
6519 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6520 * unique distances in the node_distance() table.
6522 * Assumes node_distance(0,j) includes all distances in
6523 * node_distance(i,j) in order to avoid cubic time.
6525 next_distance
= curr_distance
;
6526 for (i
= 0; i
< nr_node_ids
; i
++) {
6527 for (j
= 0; j
< nr_node_ids
; j
++) {
6528 for (k
= 0; k
< nr_node_ids
; k
++) {
6529 int distance
= node_distance(i
, k
);
6531 if (distance
> curr_distance
&&
6532 (distance
< next_distance
||
6533 next_distance
== curr_distance
))
6534 next_distance
= distance
;
6537 * While not a strong assumption it would be nice to know
6538 * about cases where if node A is connected to B, B is not
6539 * equally connected to A.
6541 if (sched_debug() && node_distance(k
, i
) != distance
)
6542 sched_numa_warn("Node-distance not symmetric");
6544 if (sched_debug() && i
&& !find_numa_distance(distance
))
6545 sched_numa_warn("Node-0 not representative");
6547 if (next_distance
!= curr_distance
) {
6548 sched_domains_numa_distance
[level
++] = next_distance
;
6549 sched_domains_numa_levels
= level
;
6550 curr_distance
= next_distance
;
6555 * In case of sched_debug() we verify the above assumption.
6565 * 'level' contains the number of unique distances, excluding the
6566 * identity distance node_distance(i,i).
6568 * The sched_domains_numa_distance[] array includes the actual distance
6573 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6574 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6575 * the array will contain less then 'level' members. This could be
6576 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6577 * in other functions.
6579 * We reset it to 'level' at the end of this function.
6581 sched_domains_numa_levels
= 0;
6583 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6584 if (!sched_domains_numa_masks
)
6588 * Now for each level, construct a mask per node which contains all
6589 * cpus of nodes that are that many hops away from us.
6591 for (i
= 0; i
< level
; i
++) {
6592 sched_domains_numa_masks
[i
] =
6593 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6594 if (!sched_domains_numa_masks
[i
])
6597 for (j
= 0; j
< nr_node_ids
; j
++) {
6598 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6602 sched_domains_numa_masks
[i
][j
] = mask
;
6604 for (k
= 0; k
< nr_node_ids
; k
++) {
6605 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6608 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6613 /* Compute default topology size */
6614 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6616 tl
= kzalloc((i
+ level
+ 1) *
6617 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6622 * Copy the default topology bits..
6624 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6625 tl
[i
] = sched_domain_topology
[i
];
6628 * .. and append 'j' levels of NUMA goodness.
6630 for (j
= 0; j
< level
; i
++, j
++) {
6631 tl
[i
] = (struct sched_domain_topology_level
){
6632 .mask
= sd_numa_mask
,
6633 .sd_flags
= cpu_numa_flags
,
6634 .flags
= SDTL_OVERLAP
,
6640 sched_domain_topology
= tl
;
6642 sched_domains_numa_levels
= level
;
6643 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6645 init_numa_topology_type();
6648 static void sched_domains_numa_masks_set(int cpu
)
6651 int node
= cpu_to_node(cpu
);
6653 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6654 for (j
= 0; j
< nr_node_ids
; j
++) {
6655 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6656 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6661 static void sched_domains_numa_masks_clear(int cpu
)
6664 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6665 for (j
= 0; j
< nr_node_ids
; j
++)
6666 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6671 * Update sched_domains_numa_masks[level][node] array when new cpus
6674 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6675 unsigned long action
,
6678 int cpu
= (long)hcpu
;
6680 switch (action
& ~CPU_TASKS_FROZEN
) {
6682 sched_domains_numa_masks_set(cpu
);
6686 sched_domains_numa_masks_clear(cpu
);
6696 static inline void sched_init_numa(void)
6700 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6701 unsigned long action
,
6706 #endif /* CONFIG_NUMA */
6708 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6710 struct sched_domain_topology_level
*tl
;
6713 for_each_sd_topology(tl
) {
6714 struct sd_data
*sdd
= &tl
->data
;
6716 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6720 sdd
->sg
= alloc_percpu(struct sched_group
*);
6724 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6728 for_each_cpu(j
, cpu_map
) {
6729 struct sched_domain
*sd
;
6730 struct sched_group
*sg
;
6731 struct sched_group_capacity
*sgc
;
6733 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6734 GFP_KERNEL
, cpu_to_node(j
));
6738 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6740 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6741 GFP_KERNEL
, cpu_to_node(j
));
6747 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6749 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6750 GFP_KERNEL
, cpu_to_node(j
));
6754 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6761 static void __sdt_free(const struct cpumask
*cpu_map
)
6763 struct sched_domain_topology_level
*tl
;
6766 for_each_sd_topology(tl
) {
6767 struct sd_data
*sdd
= &tl
->data
;
6769 for_each_cpu(j
, cpu_map
) {
6770 struct sched_domain
*sd
;
6773 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6774 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6775 free_sched_groups(sd
->groups
, 0);
6776 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6780 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6782 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6784 free_percpu(sdd
->sd
);
6786 free_percpu(sdd
->sg
);
6788 free_percpu(sdd
->sgc
);
6793 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6794 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6795 struct sched_domain
*child
, int cpu
)
6797 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6801 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6803 sd
->level
= child
->level
+ 1;
6804 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6808 if (!cpumask_subset(sched_domain_span(child
),
6809 sched_domain_span(sd
))) {
6810 pr_err("BUG: arch topology borken\n");
6811 #ifdef CONFIG_SCHED_DEBUG
6812 pr_err(" the %s domain not a subset of the %s domain\n",
6813 child
->name
, sd
->name
);
6815 /* Fixup, ensure @sd has at least @child cpus. */
6816 cpumask_or(sched_domain_span(sd
),
6817 sched_domain_span(sd
),
6818 sched_domain_span(child
));
6822 set_domain_attribute(sd
, attr
);
6828 * Build sched domains for a given set of cpus and attach the sched domains
6829 * to the individual cpus
6831 static int build_sched_domains(const struct cpumask
*cpu_map
,
6832 struct sched_domain_attr
*attr
)
6834 enum s_alloc alloc_state
;
6835 struct sched_domain
*sd
;
6837 int i
, ret
= -ENOMEM
;
6839 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6840 if (alloc_state
!= sa_rootdomain
)
6843 /* Set up domains for cpus specified by the cpu_map. */
6844 for_each_cpu(i
, cpu_map
) {
6845 struct sched_domain_topology_level
*tl
;
6848 for_each_sd_topology(tl
) {
6849 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6850 if (tl
== sched_domain_topology
)
6851 *per_cpu_ptr(d
.sd
, i
) = sd
;
6852 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6853 sd
->flags
|= SD_OVERLAP
;
6854 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6859 /* Build the groups for the domains */
6860 for_each_cpu(i
, cpu_map
) {
6861 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6862 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6863 if (sd
->flags
& SD_OVERLAP
) {
6864 if (build_overlap_sched_groups(sd
, i
))
6867 if (build_sched_groups(sd
, i
))
6873 /* Calculate CPU capacity for physical packages and nodes */
6874 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6875 if (!cpumask_test_cpu(i
, cpu_map
))
6878 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6879 claim_allocations(i
, sd
);
6880 init_sched_groups_capacity(i
, sd
);
6884 /* Attach the domains */
6886 for_each_cpu(i
, cpu_map
) {
6887 sd
= *per_cpu_ptr(d
.sd
, i
);
6888 cpu_attach_domain(sd
, d
.rd
, i
);
6894 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6898 static cpumask_var_t
*doms_cur
; /* current sched domains */
6899 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6900 static struct sched_domain_attr
*dattr_cur
;
6901 /* attribues of custom domains in 'doms_cur' */
6904 * Special case: If a kmalloc of a doms_cur partition (array of
6905 * cpumask) fails, then fallback to a single sched domain,
6906 * as determined by the single cpumask fallback_doms.
6908 static cpumask_var_t fallback_doms
;
6911 * arch_update_cpu_topology lets virtualized architectures update the
6912 * cpu core maps. It is supposed to return 1 if the topology changed
6913 * or 0 if it stayed the same.
6915 int __weak
arch_update_cpu_topology(void)
6920 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6923 cpumask_var_t
*doms
;
6925 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6928 for (i
= 0; i
< ndoms
; i
++) {
6929 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6930 free_sched_domains(doms
, i
);
6937 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6940 for (i
= 0; i
< ndoms
; i
++)
6941 free_cpumask_var(doms
[i
]);
6946 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6947 * For now this just excludes isolated cpus, but could be used to
6948 * exclude other special cases in the future.
6950 static int init_sched_domains(const struct cpumask
*cpu_map
)
6954 arch_update_cpu_topology();
6956 doms_cur
= alloc_sched_domains(ndoms_cur
);
6958 doms_cur
= &fallback_doms
;
6959 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6960 err
= build_sched_domains(doms_cur
[0], NULL
);
6961 register_sched_domain_sysctl();
6967 * Detach sched domains from a group of cpus specified in cpu_map
6968 * These cpus will now be attached to the NULL domain
6970 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6975 for_each_cpu(i
, cpu_map
)
6976 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6980 /* handle null as "default" */
6981 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6982 struct sched_domain_attr
*new, int idx_new
)
6984 struct sched_domain_attr tmp
;
6991 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6992 new ? (new + idx_new
) : &tmp
,
6993 sizeof(struct sched_domain_attr
));
6997 * Partition sched domains as specified by the 'ndoms_new'
6998 * cpumasks in the array doms_new[] of cpumasks. This compares
6999 * doms_new[] to the current sched domain partitioning, doms_cur[].
7000 * It destroys each deleted domain and builds each new domain.
7002 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7003 * The masks don't intersect (don't overlap.) We should setup one
7004 * sched domain for each mask. CPUs not in any of the cpumasks will
7005 * not be load balanced. If the same cpumask appears both in the
7006 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7009 * The passed in 'doms_new' should be allocated using
7010 * alloc_sched_domains. This routine takes ownership of it and will
7011 * free_sched_domains it when done with it. If the caller failed the
7012 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7013 * and partition_sched_domains() will fallback to the single partition
7014 * 'fallback_doms', it also forces the domains to be rebuilt.
7016 * If doms_new == NULL it will be replaced with cpu_online_mask.
7017 * ndoms_new == 0 is a special case for destroying existing domains,
7018 * and it will not create the default domain.
7020 * Call with hotplug lock held
7022 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7023 struct sched_domain_attr
*dattr_new
)
7028 mutex_lock(&sched_domains_mutex
);
7030 /* always unregister in case we don't destroy any domains */
7031 unregister_sched_domain_sysctl();
7033 /* Let architecture update cpu core mappings. */
7034 new_topology
= arch_update_cpu_topology();
7036 n
= doms_new
? ndoms_new
: 0;
7038 /* Destroy deleted domains */
7039 for (i
= 0; i
< ndoms_cur
; i
++) {
7040 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7041 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7042 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7045 /* no match - a current sched domain not in new doms_new[] */
7046 detach_destroy_domains(doms_cur
[i
]);
7052 if (doms_new
== NULL
) {
7054 doms_new
= &fallback_doms
;
7055 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7056 WARN_ON_ONCE(dattr_new
);
7059 /* Build new domains */
7060 for (i
= 0; i
< ndoms_new
; i
++) {
7061 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7062 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7063 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7066 /* no match - add a new doms_new */
7067 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7072 /* Remember the new sched domains */
7073 if (doms_cur
!= &fallback_doms
)
7074 free_sched_domains(doms_cur
, ndoms_cur
);
7075 kfree(dattr_cur
); /* kfree(NULL) is safe */
7076 doms_cur
= doms_new
;
7077 dattr_cur
= dattr_new
;
7078 ndoms_cur
= ndoms_new
;
7080 register_sched_domain_sysctl();
7082 mutex_unlock(&sched_domains_mutex
);
7085 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7088 * Update cpusets according to cpu_active mask. If cpusets are
7089 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7090 * around partition_sched_domains().
7092 * If we come here as part of a suspend/resume, don't touch cpusets because we
7093 * want to restore it back to its original state upon resume anyway.
7095 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7099 case CPU_ONLINE_FROZEN
:
7100 case CPU_DOWN_FAILED_FROZEN
:
7103 * num_cpus_frozen tracks how many CPUs are involved in suspend
7104 * resume sequence. As long as this is not the last online
7105 * operation in the resume sequence, just build a single sched
7106 * domain, ignoring cpusets.
7109 if (likely(num_cpus_frozen
)) {
7110 partition_sched_domains(1, NULL
, NULL
);
7115 * This is the last CPU online operation. So fall through and
7116 * restore the original sched domains by considering the
7117 * cpuset configurations.
7121 cpuset_update_active_cpus(true);
7129 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7132 unsigned long flags
;
7133 long cpu
= (long)hcpu
;
7139 case CPU_DOWN_PREPARE
:
7140 rcu_read_lock_sched();
7141 dl_b
= dl_bw_of(cpu
);
7143 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7144 cpus
= dl_bw_cpus(cpu
);
7145 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7146 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7148 rcu_read_unlock_sched();
7151 return notifier_from_errno(-EBUSY
);
7152 cpuset_update_active_cpus(false);
7154 case CPU_DOWN_PREPARE_FROZEN
:
7156 partition_sched_domains(1, NULL
, NULL
);
7164 void __init
sched_init_smp(void)
7166 cpumask_var_t non_isolated_cpus
;
7168 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7169 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7171 /* nohz_full won't take effect without isolating the cpus. */
7172 tick_nohz_full_add_cpus_to(cpu_isolated_map
);
7177 * There's no userspace yet to cause hotplug operations; hence all the
7178 * cpu masks are stable and all blatant races in the below code cannot
7181 mutex_lock(&sched_domains_mutex
);
7182 init_sched_domains(cpu_active_mask
);
7183 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7184 if (cpumask_empty(non_isolated_cpus
))
7185 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7186 mutex_unlock(&sched_domains_mutex
);
7188 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7189 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7190 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7194 /* Move init over to a non-isolated CPU */
7195 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7197 sched_init_granularity();
7198 free_cpumask_var(non_isolated_cpus
);
7200 init_sched_rt_class();
7201 init_sched_dl_class();
7204 void __init
sched_init_smp(void)
7206 sched_init_granularity();
7208 #endif /* CONFIG_SMP */
7210 int in_sched_functions(unsigned long addr
)
7212 return in_lock_functions(addr
) ||
7213 (addr
>= (unsigned long)__sched_text_start
7214 && addr
< (unsigned long)__sched_text_end
);
7217 #ifdef CONFIG_CGROUP_SCHED
7219 * Default task group.
7220 * Every task in system belongs to this group at bootup.
7222 struct task_group root_task_group
;
7223 LIST_HEAD(task_groups
);
7226 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7228 void __init
sched_init(void)
7231 unsigned long alloc_size
= 0, ptr
;
7233 #ifdef CONFIG_FAIR_GROUP_SCHED
7234 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7236 #ifdef CONFIG_RT_GROUP_SCHED
7237 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7240 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7242 #ifdef CONFIG_FAIR_GROUP_SCHED
7243 root_task_group
.se
= (struct sched_entity
**)ptr
;
7244 ptr
+= nr_cpu_ids
* sizeof(void **);
7246 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7247 ptr
+= nr_cpu_ids
* sizeof(void **);
7249 #endif /* CONFIG_FAIR_GROUP_SCHED */
7250 #ifdef CONFIG_RT_GROUP_SCHED
7251 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7252 ptr
+= nr_cpu_ids
* sizeof(void **);
7254 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7255 ptr
+= nr_cpu_ids
* sizeof(void **);
7257 #endif /* CONFIG_RT_GROUP_SCHED */
7259 #ifdef CONFIG_CPUMASK_OFFSTACK
7260 for_each_possible_cpu(i
) {
7261 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7262 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7264 #endif /* CONFIG_CPUMASK_OFFSTACK */
7266 init_rt_bandwidth(&def_rt_bandwidth
,
7267 global_rt_period(), global_rt_runtime());
7268 init_dl_bandwidth(&def_dl_bandwidth
,
7269 global_rt_period(), global_rt_runtime());
7272 init_defrootdomain();
7275 #ifdef CONFIG_RT_GROUP_SCHED
7276 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7277 global_rt_period(), global_rt_runtime());
7278 #endif /* CONFIG_RT_GROUP_SCHED */
7280 #ifdef CONFIG_CGROUP_SCHED
7281 list_add(&root_task_group
.list
, &task_groups
);
7282 INIT_LIST_HEAD(&root_task_group
.children
);
7283 INIT_LIST_HEAD(&root_task_group
.siblings
);
7284 autogroup_init(&init_task
);
7286 #endif /* CONFIG_CGROUP_SCHED */
7288 for_each_possible_cpu(i
) {
7292 raw_spin_lock_init(&rq
->lock
);
7294 rq
->calc_load_active
= 0;
7295 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7296 init_cfs_rq(&rq
->cfs
);
7297 init_rt_rq(&rq
->rt
);
7298 init_dl_rq(&rq
->dl
);
7299 #ifdef CONFIG_FAIR_GROUP_SCHED
7300 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7301 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7303 * How much cpu bandwidth does root_task_group get?
7305 * In case of task-groups formed thr' the cgroup filesystem, it
7306 * gets 100% of the cpu resources in the system. This overall
7307 * system cpu resource is divided among the tasks of
7308 * root_task_group and its child task-groups in a fair manner,
7309 * based on each entity's (task or task-group's) weight
7310 * (se->load.weight).
7312 * In other words, if root_task_group has 10 tasks of weight
7313 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7314 * then A0's share of the cpu resource is:
7316 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7318 * We achieve this by letting root_task_group's tasks sit
7319 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7321 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7322 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7323 #endif /* CONFIG_FAIR_GROUP_SCHED */
7325 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7326 #ifdef CONFIG_RT_GROUP_SCHED
7327 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7330 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7331 rq
->cpu_load
[j
] = 0;
7333 rq
->last_load_update_tick
= jiffies
;
7338 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7339 rq
->balance_callback
= NULL
;
7340 rq
->active_balance
= 0;
7341 rq
->next_balance
= jiffies
;
7346 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7347 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7349 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7351 rq_attach_root(rq
, &def_root_domain
);
7352 #ifdef CONFIG_NO_HZ_COMMON
7355 #ifdef CONFIG_NO_HZ_FULL
7356 rq
->last_sched_tick
= 0;
7360 atomic_set(&rq
->nr_iowait
, 0);
7363 set_load_weight(&init_task
);
7365 #ifdef CONFIG_PREEMPT_NOTIFIERS
7366 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7370 * The boot idle thread does lazy MMU switching as well:
7372 atomic_inc(&init_mm
.mm_count
);
7373 enter_lazy_tlb(&init_mm
, current
);
7376 * During early bootup we pretend to be a normal task:
7378 current
->sched_class
= &fair_sched_class
;
7381 * Make us the idle thread. Technically, schedule() should not be
7382 * called from this thread, however somewhere below it might be,
7383 * but because we are the idle thread, we just pick up running again
7384 * when this runqueue becomes "idle".
7386 init_idle(current
, smp_processor_id());
7388 calc_load_update
= jiffies
+ LOAD_FREQ
;
7391 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7392 /* May be allocated at isolcpus cmdline parse time */
7393 if (cpu_isolated_map
== NULL
)
7394 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7395 idle_thread_set_boot_cpu();
7396 set_cpu_rq_start_time();
7398 init_sched_fair_class();
7400 scheduler_running
= 1;
7403 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7404 static inline int preempt_count_equals(int preempt_offset
)
7406 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7408 return (nested
== preempt_offset
);
7411 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7414 * Blocking primitives will set (and therefore destroy) current->state,
7415 * since we will exit with TASK_RUNNING make sure we enter with it,
7416 * otherwise we will destroy state.
7418 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7419 "do not call blocking ops when !TASK_RUNNING; "
7420 "state=%lx set at [<%p>] %pS\n",
7422 (void *)current
->task_state_change
,
7423 (void *)current
->task_state_change
);
7425 ___might_sleep(file
, line
, preempt_offset
);
7427 EXPORT_SYMBOL(__might_sleep
);
7429 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7431 static unsigned long prev_jiffy
; /* ratelimiting */
7433 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7434 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7435 !is_idle_task(current
)) ||
7436 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7438 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7440 prev_jiffy
= jiffies
;
7443 "BUG: sleeping function called from invalid context at %s:%d\n",
7446 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7447 in_atomic(), irqs_disabled(),
7448 current
->pid
, current
->comm
);
7450 if (task_stack_end_corrupted(current
))
7451 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7453 debug_show_held_locks(current
);
7454 if (irqs_disabled())
7455 print_irqtrace_events(current
);
7456 #ifdef CONFIG_DEBUG_PREEMPT
7457 if (!preempt_count_equals(preempt_offset
)) {
7458 pr_err("Preemption disabled at:");
7459 print_ip_sym(current
->preempt_disable_ip
);
7465 EXPORT_SYMBOL(___might_sleep
);
7468 #ifdef CONFIG_MAGIC_SYSRQ
7469 void normalize_rt_tasks(void)
7471 struct task_struct
*g
, *p
;
7472 struct sched_attr attr
= {
7473 .sched_policy
= SCHED_NORMAL
,
7476 read_lock(&tasklist_lock
);
7477 for_each_process_thread(g
, p
) {
7479 * Only normalize user tasks:
7481 if (p
->flags
& PF_KTHREAD
)
7484 p
->se
.exec_start
= 0;
7485 #ifdef CONFIG_SCHEDSTATS
7486 p
->se
.statistics
.wait_start
= 0;
7487 p
->se
.statistics
.sleep_start
= 0;
7488 p
->se
.statistics
.block_start
= 0;
7491 if (!dl_task(p
) && !rt_task(p
)) {
7493 * Renice negative nice level userspace
7496 if (task_nice(p
) < 0)
7497 set_user_nice(p
, 0);
7501 __sched_setscheduler(p
, &attr
, false, false);
7503 read_unlock(&tasklist_lock
);
7506 #endif /* CONFIG_MAGIC_SYSRQ */
7508 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7510 * These functions are only useful for the IA64 MCA handling, or kdb.
7512 * They can only be called when the whole system has been
7513 * stopped - every CPU needs to be quiescent, and no scheduling
7514 * activity can take place. Using them for anything else would
7515 * be a serious bug, and as a result, they aren't even visible
7516 * under any other configuration.
7520 * curr_task - return the current task for a given cpu.
7521 * @cpu: the processor in question.
7523 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7525 * Return: The current task for @cpu.
7527 struct task_struct
*curr_task(int cpu
)
7529 return cpu_curr(cpu
);
7532 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7536 * set_curr_task - set the current task for a given cpu.
7537 * @cpu: the processor in question.
7538 * @p: the task pointer to set.
7540 * Description: This function must only be used when non-maskable interrupts
7541 * are serviced on a separate stack. It allows the architecture to switch the
7542 * notion of the current task on a cpu in a non-blocking manner. This function
7543 * must be called with all CPU's synchronized, and interrupts disabled, the
7544 * and caller must save the original value of the current task (see
7545 * curr_task() above) and restore that value before reenabling interrupts and
7546 * re-starting the system.
7548 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7550 void set_curr_task(int cpu
, struct task_struct
*p
)
7557 #ifdef CONFIG_CGROUP_SCHED
7558 /* task_group_lock serializes the addition/removal of task groups */
7559 static DEFINE_SPINLOCK(task_group_lock
);
7561 static void free_sched_group(struct task_group
*tg
)
7563 free_fair_sched_group(tg
);
7564 free_rt_sched_group(tg
);
7569 /* allocate runqueue etc for a new task group */
7570 struct task_group
*sched_create_group(struct task_group
*parent
)
7572 struct task_group
*tg
;
7574 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7576 return ERR_PTR(-ENOMEM
);
7578 if (!alloc_fair_sched_group(tg
, parent
))
7581 if (!alloc_rt_sched_group(tg
, parent
))
7587 free_sched_group(tg
);
7588 return ERR_PTR(-ENOMEM
);
7591 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7593 unsigned long flags
;
7595 spin_lock_irqsave(&task_group_lock
, flags
);
7596 list_add_rcu(&tg
->list
, &task_groups
);
7598 WARN_ON(!parent
); /* root should already exist */
7600 tg
->parent
= parent
;
7601 INIT_LIST_HEAD(&tg
->children
);
7602 list_add_rcu(&tg
->siblings
, &parent
->children
);
7603 spin_unlock_irqrestore(&task_group_lock
, flags
);
7606 /* rcu callback to free various structures associated with a task group */
7607 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7609 /* now it should be safe to free those cfs_rqs */
7610 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7613 /* Destroy runqueue etc associated with a task group */
7614 void sched_destroy_group(struct task_group
*tg
)
7616 /* wait for possible concurrent references to cfs_rqs complete */
7617 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7620 void sched_offline_group(struct task_group
*tg
)
7622 unsigned long flags
;
7625 /* end participation in shares distribution */
7626 for_each_possible_cpu(i
)
7627 unregister_fair_sched_group(tg
, i
);
7629 spin_lock_irqsave(&task_group_lock
, flags
);
7630 list_del_rcu(&tg
->list
);
7631 list_del_rcu(&tg
->siblings
);
7632 spin_unlock_irqrestore(&task_group_lock
, flags
);
7635 /* change task's runqueue when it moves between groups.
7636 * The caller of this function should have put the task in its new group
7637 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7638 * reflect its new group.
7640 void sched_move_task(struct task_struct
*tsk
)
7642 struct task_group
*tg
;
7643 int queued
, running
;
7644 unsigned long flags
;
7647 rq
= task_rq_lock(tsk
, &flags
);
7649 running
= task_current(rq
, tsk
);
7650 queued
= task_on_rq_queued(tsk
);
7653 dequeue_task(rq
, tsk
, 0);
7654 if (unlikely(running
))
7655 put_prev_task(rq
, tsk
);
7658 * All callers are synchronized by task_rq_lock(); we do not use RCU
7659 * which is pointless here. Thus, we pass "true" to task_css_check()
7660 * to prevent lockdep warnings.
7662 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7663 struct task_group
, css
);
7664 tg
= autogroup_task_group(tsk
, tg
);
7665 tsk
->sched_task_group
= tg
;
7667 #ifdef CONFIG_FAIR_GROUP_SCHED
7668 if (tsk
->sched_class
->task_move_group
)
7669 tsk
->sched_class
->task_move_group(tsk
, queued
);
7672 set_task_rq(tsk
, task_cpu(tsk
));
7674 if (unlikely(running
))
7675 tsk
->sched_class
->set_curr_task(rq
);
7677 enqueue_task(rq
, tsk
, 0);
7679 task_rq_unlock(rq
, tsk
, &flags
);
7681 #endif /* CONFIG_CGROUP_SCHED */
7683 #ifdef CONFIG_RT_GROUP_SCHED
7685 * Ensure that the real time constraints are schedulable.
7687 static DEFINE_MUTEX(rt_constraints_mutex
);
7689 /* Must be called with tasklist_lock held */
7690 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7692 struct task_struct
*g
, *p
;
7695 * Autogroups do not have RT tasks; see autogroup_create().
7697 if (task_group_is_autogroup(tg
))
7700 for_each_process_thread(g
, p
) {
7701 if (rt_task(p
) && task_group(p
) == tg
)
7708 struct rt_schedulable_data
{
7709 struct task_group
*tg
;
7714 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7716 struct rt_schedulable_data
*d
= data
;
7717 struct task_group
*child
;
7718 unsigned long total
, sum
= 0;
7719 u64 period
, runtime
;
7721 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7722 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7725 period
= d
->rt_period
;
7726 runtime
= d
->rt_runtime
;
7730 * Cannot have more runtime than the period.
7732 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7736 * Ensure we don't starve existing RT tasks.
7738 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7741 total
= to_ratio(period
, runtime
);
7744 * Nobody can have more than the global setting allows.
7746 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7750 * The sum of our children's runtime should not exceed our own.
7752 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7753 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7754 runtime
= child
->rt_bandwidth
.rt_runtime
;
7756 if (child
== d
->tg
) {
7757 period
= d
->rt_period
;
7758 runtime
= d
->rt_runtime
;
7761 sum
+= to_ratio(period
, runtime
);
7770 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7774 struct rt_schedulable_data data
= {
7776 .rt_period
= period
,
7777 .rt_runtime
= runtime
,
7781 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7787 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7788 u64 rt_period
, u64 rt_runtime
)
7793 * Disallowing the root group RT runtime is BAD, it would disallow the
7794 * kernel creating (and or operating) RT threads.
7796 if (tg
== &root_task_group
&& rt_runtime
== 0)
7799 /* No period doesn't make any sense. */
7803 mutex_lock(&rt_constraints_mutex
);
7804 read_lock(&tasklist_lock
);
7805 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7809 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7810 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7811 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7813 for_each_possible_cpu(i
) {
7814 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7816 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7817 rt_rq
->rt_runtime
= rt_runtime
;
7818 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7820 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7822 read_unlock(&tasklist_lock
);
7823 mutex_unlock(&rt_constraints_mutex
);
7828 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7830 u64 rt_runtime
, rt_period
;
7832 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7833 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7834 if (rt_runtime_us
< 0)
7835 rt_runtime
= RUNTIME_INF
;
7837 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7840 static long sched_group_rt_runtime(struct task_group
*tg
)
7844 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7847 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7848 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7849 return rt_runtime_us
;
7852 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7854 u64 rt_runtime
, rt_period
;
7856 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7857 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7859 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7862 static long sched_group_rt_period(struct task_group
*tg
)
7866 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7867 do_div(rt_period_us
, NSEC_PER_USEC
);
7868 return rt_period_us
;
7870 #endif /* CONFIG_RT_GROUP_SCHED */
7872 #ifdef CONFIG_RT_GROUP_SCHED
7873 static int sched_rt_global_constraints(void)
7877 mutex_lock(&rt_constraints_mutex
);
7878 read_lock(&tasklist_lock
);
7879 ret
= __rt_schedulable(NULL
, 0, 0);
7880 read_unlock(&tasklist_lock
);
7881 mutex_unlock(&rt_constraints_mutex
);
7886 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7888 /* Don't accept realtime tasks when there is no way for them to run */
7889 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7895 #else /* !CONFIG_RT_GROUP_SCHED */
7896 static int sched_rt_global_constraints(void)
7898 unsigned long flags
;
7901 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7902 for_each_possible_cpu(i
) {
7903 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7905 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7906 rt_rq
->rt_runtime
= global_rt_runtime();
7907 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7909 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7913 #endif /* CONFIG_RT_GROUP_SCHED */
7915 static int sched_dl_global_validate(void)
7917 u64 runtime
= global_rt_runtime();
7918 u64 period
= global_rt_period();
7919 u64 new_bw
= to_ratio(period
, runtime
);
7922 unsigned long flags
;
7925 * Here we want to check the bandwidth not being set to some
7926 * value smaller than the currently allocated bandwidth in
7927 * any of the root_domains.
7929 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7930 * cycling on root_domains... Discussion on different/better
7931 * solutions is welcome!
7933 for_each_possible_cpu(cpu
) {
7934 rcu_read_lock_sched();
7935 dl_b
= dl_bw_of(cpu
);
7937 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7938 if (new_bw
< dl_b
->total_bw
)
7940 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7942 rcu_read_unlock_sched();
7951 static void sched_dl_do_global(void)
7956 unsigned long flags
;
7958 def_dl_bandwidth
.dl_period
= global_rt_period();
7959 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7961 if (global_rt_runtime() != RUNTIME_INF
)
7962 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7965 * FIXME: As above...
7967 for_each_possible_cpu(cpu
) {
7968 rcu_read_lock_sched();
7969 dl_b
= dl_bw_of(cpu
);
7971 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7973 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7975 rcu_read_unlock_sched();
7979 static int sched_rt_global_validate(void)
7981 if (sysctl_sched_rt_period
<= 0)
7984 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7985 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7991 static void sched_rt_do_global(void)
7993 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7994 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7997 int sched_rt_handler(struct ctl_table
*table
, int write
,
7998 void __user
*buffer
, size_t *lenp
,
8001 int old_period
, old_runtime
;
8002 static DEFINE_MUTEX(mutex
);
8006 old_period
= sysctl_sched_rt_period
;
8007 old_runtime
= sysctl_sched_rt_runtime
;
8009 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8011 if (!ret
&& write
) {
8012 ret
= sched_rt_global_validate();
8016 ret
= sched_dl_global_validate();
8020 ret
= sched_rt_global_constraints();
8024 sched_rt_do_global();
8025 sched_dl_do_global();
8029 sysctl_sched_rt_period
= old_period
;
8030 sysctl_sched_rt_runtime
= old_runtime
;
8032 mutex_unlock(&mutex
);
8037 int sched_rr_handler(struct ctl_table
*table
, int write
,
8038 void __user
*buffer
, size_t *lenp
,
8042 static DEFINE_MUTEX(mutex
);
8045 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8046 /* make sure that internally we keep jiffies */
8047 /* also, writing zero resets timeslice to default */
8048 if (!ret
&& write
) {
8049 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8050 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8052 mutex_unlock(&mutex
);
8056 #ifdef CONFIG_CGROUP_SCHED
8058 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8060 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8063 static struct cgroup_subsys_state
*
8064 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8066 struct task_group
*parent
= css_tg(parent_css
);
8067 struct task_group
*tg
;
8070 /* This is early initialization for the top cgroup */
8071 return &root_task_group
.css
;
8074 tg
= sched_create_group(parent
);
8076 return ERR_PTR(-ENOMEM
);
8081 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8083 struct task_group
*tg
= css_tg(css
);
8084 struct task_group
*parent
= css_tg(css
->parent
);
8087 sched_online_group(tg
, parent
);
8091 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8093 struct task_group
*tg
= css_tg(css
);
8095 sched_destroy_group(tg
);
8098 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8100 struct task_group
*tg
= css_tg(css
);
8102 sched_offline_group(tg
);
8105 static void cpu_cgroup_fork(struct task_struct
*task
)
8107 sched_move_task(task
);
8110 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
8111 struct cgroup_taskset
*tset
)
8113 struct task_struct
*task
;
8115 cgroup_taskset_for_each(task
, tset
) {
8116 #ifdef CONFIG_RT_GROUP_SCHED
8117 if (!sched_rt_can_attach(css_tg(css
), task
))
8120 /* We don't support RT-tasks being in separate groups */
8121 if (task
->sched_class
!= &fair_sched_class
)
8128 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8129 struct cgroup_taskset
*tset
)
8131 struct task_struct
*task
;
8133 cgroup_taskset_for_each(task
, tset
)
8134 sched_move_task(task
);
8137 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8138 struct cgroup_subsys_state
*old_css
,
8139 struct task_struct
*task
)
8142 * cgroup_exit() is called in the copy_process() failure path.
8143 * Ignore this case since the task hasn't ran yet, this avoids
8144 * trying to poke a half freed task state from generic code.
8146 if (!(task
->flags
& PF_EXITING
))
8149 sched_move_task(task
);
8152 #ifdef CONFIG_FAIR_GROUP_SCHED
8153 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8154 struct cftype
*cftype
, u64 shareval
)
8156 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8159 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8162 struct task_group
*tg
= css_tg(css
);
8164 return (u64
) scale_load_down(tg
->shares
);
8167 #ifdef CONFIG_CFS_BANDWIDTH
8168 static DEFINE_MUTEX(cfs_constraints_mutex
);
8170 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8171 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8173 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8175 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8177 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8178 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8180 if (tg
== &root_task_group
)
8184 * Ensure we have at some amount of bandwidth every period. This is
8185 * to prevent reaching a state of large arrears when throttled via
8186 * entity_tick() resulting in prolonged exit starvation.
8188 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8192 * Likewise, bound things on the otherside by preventing insane quota
8193 * periods. This also allows us to normalize in computing quota
8196 if (period
> max_cfs_quota_period
)
8200 * Prevent race between setting of cfs_rq->runtime_enabled and
8201 * unthrottle_offline_cfs_rqs().
8204 mutex_lock(&cfs_constraints_mutex
);
8205 ret
= __cfs_schedulable(tg
, period
, quota
);
8209 runtime_enabled
= quota
!= RUNTIME_INF
;
8210 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8212 * If we need to toggle cfs_bandwidth_used, off->on must occur
8213 * before making related changes, and on->off must occur afterwards
8215 if (runtime_enabled
&& !runtime_was_enabled
)
8216 cfs_bandwidth_usage_inc();
8217 raw_spin_lock_irq(&cfs_b
->lock
);
8218 cfs_b
->period
= ns_to_ktime(period
);
8219 cfs_b
->quota
= quota
;
8221 __refill_cfs_bandwidth_runtime(cfs_b
);
8222 /* restart the period timer (if active) to handle new period expiry */
8223 if (runtime_enabled
)
8224 start_cfs_bandwidth(cfs_b
);
8225 raw_spin_unlock_irq(&cfs_b
->lock
);
8227 for_each_online_cpu(i
) {
8228 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8229 struct rq
*rq
= cfs_rq
->rq
;
8231 raw_spin_lock_irq(&rq
->lock
);
8232 cfs_rq
->runtime_enabled
= runtime_enabled
;
8233 cfs_rq
->runtime_remaining
= 0;
8235 if (cfs_rq
->throttled
)
8236 unthrottle_cfs_rq(cfs_rq
);
8237 raw_spin_unlock_irq(&rq
->lock
);
8239 if (runtime_was_enabled
&& !runtime_enabled
)
8240 cfs_bandwidth_usage_dec();
8242 mutex_unlock(&cfs_constraints_mutex
);
8248 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8252 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8253 if (cfs_quota_us
< 0)
8254 quota
= RUNTIME_INF
;
8256 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8258 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8261 long tg_get_cfs_quota(struct task_group
*tg
)
8265 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8268 quota_us
= tg
->cfs_bandwidth
.quota
;
8269 do_div(quota_us
, NSEC_PER_USEC
);
8274 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8278 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8279 quota
= tg
->cfs_bandwidth
.quota
;
8281 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8284 long tg_get_cfs_period(struct task_group
*tg
)
8288 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8289 do_div(cfs_period_us
, NSEC_PER_USEC
);
8291 return cfs_period_us
;
8294 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8297 return tg_get_cfs_quota(css_tg(css
));
8300 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8301 struct cftype
*cftype
, s64 cfs_quota_us
)
8303 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8306 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8309 return tg_get_cfs_period(css_tg(css
));
8312 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8313 struct cftype
*cftype
, u64 cfs_period_us
)
8315 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8318 struct cfs_schedulable_data
{
8319 struct task_group
*tg
;
8324 * normalize group quota/period to be quota/max_period
8325 * note: units are usecs
8327 static u64
normalize_cfs_quota(struct task_group
*tg
,
8328 struct cfs_schedulable_data
*d
)
8336 period
= tg_get_cfs_period(tg
);
8337 quota
= tg_get_cfs_quota(tg
);
8340 /* note: these should typically be equivalent */
8341 if (quota
== RUNTIME_INF
|| quota
== -1)
8344 return to_ratio(period
, quota
);
8347 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8349 struct cfs_schedulable_data
*d
= data
;
8350 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8351 s64 quota
= 0, parent_quota
= -1;
8354 quota
= RUNTIME_INF
;
8356 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8358 quota
= normalize_cfs_quota(tg
, d
);
8359 parent_quota
= parent_b
->hierarchical_quota
;
8362 * ensure max(child_quota) <= parent_quota, inherit when no
8365 if (quota
== RUNTIME_INF
)
8366 quota
= parent_quota
;
8367 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8370 cfs_b
->hierarchical_quota
= quota
;
8375 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8378 struct cfs_schedulable_data data
= {
8384 if (quota
!= RUNTIME_INF
) {
8385 do_div(data
.period
, NSEC_PER_USEC
);
8386 do_div(data
.quota
, NSEC_PER_USEC
);
8390 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8396 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8398 struct task_group
*tg
= css_tg(seq_css(sf
));
8399 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8401 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8402 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8403 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8407 #endif /* CONFIG_CFS_BANDWIDTH */
8408 #endif /* CONFIG_FAIR_GROUP_SCHED */
8410 #ifdef CONFIG_RT_GROUP_SCHED
8411 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8412 struct cftype
*cft
, s64 val
)
8414 return sched_group_set_rt_runtime(css_tg(css
), val
);
8417 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8420 return sched_group_rt_runtime(css_tg(css
));
8423 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8424 struct cftype
*cftype
, u64 rt_period_us
)
8426 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8429 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8432 return sched_group_rt_period(css_tg(css
));
8434 #endif /* CONFIG_RT_GROUP_SCHED */
8436 static struct cftype cpu_files
[] = {
8437 #ifdef CONFIG_FAIR_GROUP_SCHED
8440 .read_u64
= cpu_shares_read_u64
,
8441 .write_u64
= cpu_shares_write_u64
,
8444 #ifdef CONFIG_CFS_BANDWIDTH
8446 .name
= "cfs_quota_us",
8447 .read_s64
= cpu_cfs_quota_read_s64
,
8448 .write_s64
= cpu_cfs_quota_write_s64
,
8451 .name
= "cfs_period_us",
8452 .read_u64
= cpu_cfs_period_read_u64
,
8453 .write_u64
= cpu_cfs_period_write_u64
,
8457 .seq_show
= cpu_stats_show
,
8460 #ifdef CONFIG_RT_GROUP_SCHED
8462 .name
= "rt_runtime_us",
8463 .read_s64
= cpu_rt_runtime_read
,
8464 .write_s64
= cpu_rt_runtime_write
,
8467 .name
= "rt_period_us",
8468 .read_u64
= cpu_rt_period_read_uint
,
8469 .write_u64
= cpu_rt_period_write_uint
,
8475 struct cgroup_subsys cpu_cgrp_subsys
= {
8476 .css_alloc
= cpu_cgroup_css_alloc
,
8477 .css_free
= cpu_cgroup_css_free
,
8478 .css_online
= cpu_cgroup_css_online
,
8479 .css_offline
= cpu_cgroup_css_offline
,
8480 .fork
= cpu_cgroup_fork
,
8481 .can_attach
= cpu_cgroup_can_attach
,
8482 .attach
= cpu_cgroup_attach
,
8483 .exit
= cpu_cgroup_exit
,
8484 .legacy_cftypes
= cpu_files
,
8488 #endif /* CONFIG_CGROUP_SCHED */
8490 void dump_cpu_task(int cpu
)
8492 pr_info("Task dump for CPU %d:\n", cpu
);
8493 sched_show_task(cpu_curr(cpu
));