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 static_key_disable(&sched_feat_keys
[i
]);
170 static void sched_feat_enable(int i
)
172 static_key_enable(&sched_feat_keys
[i
]);
175 static void sched_feat_disable(int i
) { };
176 static void sched_feat_enable(int i
) { };
177 #endif /* HAVE_JUMP_LABEL */
179 static int sched_feat_set(char *cmp
)
184 if (strncmp(cmp
, "NO_", 3) == 0) {
189 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
190 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
192 sysctl_sched_features
&= ~(1UL << i
);
193 sched_feat_disable(i
);
195 sysctl_sched_features
|= (1UL << i
);
196 sched_feat_enable(i
);
206 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
207 size_t cnt
, loff_t
*ppos
)
217 if (copy_from_user(&buf
, ubuf
, cnt
))
223 /* Ensure the static_key remains in a consistent state */
224 inode
= file_inode(filp
);
225 mutex_lock(&inode
->i_mutex
);
226 i
= sched_feat_set(cmp
);
227 mutex_unlock(&inode
->i_mutex
);
228 if (i
== __SCHED_FEAT_NR
)
236 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
238 return single_open(filp
, sched_feat_show
, NULL
);
241 static const struct file_operations sched_feat_fops
= {
242 .open
= sched_feat_open
,
243 .write
= sched_feat_write
,
246 .release
= single_release
,
249 static __init
int sched_init_debug(void)
251 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
256 late_initcall(sched_init_debug
);
257 #endif /* CONFIG_SCHED_DEBUG */
260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
263 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
266 * period over which we average the RT time consumption, measured
271 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
274 * period over which we measure -rt task cpu usage in us.
277 unsigned int sysctl_sched_rt_period
= 1000000;
279 __read_mostly
int scheduler_running
;
282 * part of the period that we allow rt tasks to run in us.
285 int sysctl_sched_rt_runtime
= 950000;
287 /* cpus with isolated domains */
288 cpumask_var_t cpu_isolated_map
;
291 * this_rq_lock - lock this runqueue and disable interrupts.
293 static struct rq
*this_rq_lock(void)
300 raw_spin_lock(&rq
->lock
);
305 #ifdef CONFIG_SCHED_HRTICK
307 * Use HR-timers to deliver accurate preemption points.
310 static void hrtick_clear(struct rq
*rq
)
312 if (hrtimer_active(&rq
->hrtick_timer
))
313 hrtimer_cancel(&rq
->hrtick_timer
);
317 * High-resolution timer tick.
318 * Runs from hardirq context with interrupts disabled.
320 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
322 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
324 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
326 raw_spin_lock(&rq
->lock
);
328 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
329 raw_spin_unlock(&rq
->lock
);
331 return HRTIMER_NORESTART
;
336 static void __hrtick_restart(struct rq
*rq
)
338 struct hrtimer
*timer
= &rq
->hrtick_timer
;
340 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
344 * called from hardirq (IPI) context
346 static void __hrtick_start(void *arg
)
350 raw_spin_lock(&rq
->lock
);
351 __hrtick_restart(rq
);
352 rq
->hrtick_csd_pending
= 0;
353 raw_spin_unlock(&rq
->lock
);
357 * Called to set the hrtick timer state.
359 * called with rq->lock held and irqs disabled
361 void hrtick_start(struct rq
*rq
, u64 delay
)
363 struct hrtimer
*timer
= &rq
->hrtick_timer
;
368 * Don't schedule slices shorter than 10000ns, that just
369 * doesn't make sense and can cause timer DoS.
371 delta
= max_t(s64
, delay
, 10000LL);
372 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
374 hrtimer_set_expires(timer
, time
);
376 if (rq
== this_rq()) {
377 __hrtick_restart(rq
);
378 } else if (!rq
->hrtick_csd_pending
) {
379 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
380 rq
->hrtick_csd_pending
= 1;
385 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
387 int cpu
= (int)(long)hcpu
;
390 case CPU_UP_CANCELED
:
391 case CPU_UP_CANCELED_FROZEN
:
392 case CPU_DOWN_PREPARE
:
393 case CPU_DOWN_PREPARE_FROZEN
:
395 case CPU_DEAD_FROZEN
:
396 hrtick_clear(cpu_rq(cpu
));
403 static __init
void init_hrtick(void)
405 hotcpu_notifier(hotplug_hrtick
, 0);
409 * Called to set the hrtick timer state.
411 * called with rq->lock held and irqs disabled
413 void hrtick_start(struct rq
*rq
, u64 delay
)
416 * Don't schedule slices shorter than 10000ns, that just
417 * doesn't make sense. Rely on vruntime for fairness.
419 delay
= max_t(u64
, delay
, 10000LL);
420 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
421 HRTIMER_MODE_REL_PINNED
);
424 static inline void init_hrtick(void)
427 #endif /* CONFIG_SMP */
429 static void init_rq_hrtick(struct rq
*rq
)
432 rq
->hrtick_csd_pending
= 0;
434 rq
->hrtick_csd
.flags
= 0;
435 rq
->hrtick_csd
.func
= __hrtick_start
;
436 rq
->hrtick_csd
.info
= rq
;
439 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
440 rq
->hrtick_timer
.function
= hrtick
;
442 #else /* CONFIG_SCHED_HRTICK */
443 static inline void hrtick_clear(struct rq
*rq
)
447 static inline void init_rq_hrtick(struct rq
*rq
)
451 static inline void init_hrtick(void)
454 #endif /* CONFIG_SCHED_HRTICK */
457 * cmpxchg based fetch_or, macro so it works for different integer types
459 #define fetch_or(ptr, val) \
460 ({ typeof(*(ptr)) __old, __val = *(ptr); \
462 __old = cmpxchg((ptr), __val, __val | (val)); \
463 if (__old == __val) \
470 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
472 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
473 * this avoids any races wrt polling state changes and thereby avoids
476 static bool set_nr_and_not_polling(struct task_struct
*p
)
478 struct thread_info
*ti
= task_thread_info(p
);
479 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
483 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
485 * If this returns true, then the idle task promises to call
486 * sched_ttwu_pending() and reschedule soon.
488 static bool set_nr_if_polling(struct task_struct
*p
)
490 struct thread_info
*ti
= task_thread_info(p
);
491 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
494 if (!(val
& _TIF_POLLING_NRFLAG
))
496 if (val
& _TIF_NEED_RESCHED
)
498 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
507 static bool set_nr_and_not_polling(struct task_struct
*p
)
509 set_tsk_need_resched(p
);
514 static bool set_nr_if_polling(struct task_struct
*p
)
521 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
523 struct wake_q_node
*node
= &task
->wake_q
;
526 * Atomically grab the task, if ->wake_q is !nil already it means
527 * its already queued (either by us or someone else) and will get the
528 * wakeup due to that.
530 * This cmpxchg() implies a full barrier, which pairs with the write
531 * barrier implied by the wakeup in wake_up_list().
533 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
536 get_task_struct(task
);
539 * The head is context local, there can be no concurrency.
542 head
->lastp
= &node
->next
;
545 void wake_up_q(struct wake_q_head
*head
)
547 struct wake_q_node
*node
= head
->first
;
549 while (node
!= WAKE_Q_TAIL
) {
550 struct task_struct
*task
;
552 task
= container_of(node
, struct task_struct
, wake_q
);
554 /* task can safely be re-inserted now */
556 task
->wake_q
.next
= NULL
;
559 * wake_up_process() implies a wmb() to pair with the queueing
560 * in wake_q_add() so as not to miss wakeups.
562 wake_up_process(task
);
563 put_task_struct(task
);
568 * resched_curr - mark rq's current task 'to be rescheduled now'.
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
574 void resched_curr(struct rq
*rq
)
576 struct task_struct
*curr
= rq
->curr
;
579 lockdep_assert_held(&rq
->lock
);
581 if (test_tsk_need_resched(curr
))
586 if (cpu
== smp_processor_id()) {
587 set_tsk_need_resched(curr
);
588 set_preempt_need_resched();
592 if (set_nr_and_not_polling(curr
))
593 smp_send_reschedule(cpu
);
595 trace_sched_wake_idle_without_ipi(cpu
);
598 void resched_cpu(int cpu
)
600 struct rq
*rq
= cpu_rq(cpu
);
603 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
606 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
610 #ifdef CONFIG_NO_HZ_COMMON
612 * In the semi idle case, use the nearest busy cpu for migrating timers
613 * from an idle cpu. This is good for power-savings.
615 * We don't do similar optimization for completely idle system, as
616 * selecting an idle cpu will add more delays to the timers than intended
617 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
619 int get_nohz_timer_target(void)
621 int i
, cpu
= smp_processor_id();
622 struct sched_domain
*sd
;
624 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
628 for_each_domain(cpu
, sd
) {
629 for_each_cpu(i
, sched_domain_span(sd
)) {
630 if (!idle_cpu(i
) && is_housekeeping_cpu(cpu
)) {
637 if (!is_housekeeping_cpu(cpu
))
638 cpu
= housekeeping_any_cpu();
644 * When add_timer_on() enqueues a timer into the timer wheel of an
645 * idle CPU then this timer might expire before the next timer event
646 * which is scheduled to wake up that CPU. In case of a completely
647 * idle system the next event might even be infinite time into the
648 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
649 * leaves the inner idle loop so the newly added timer is taken into
650 * account when the CPU goes back to idle and evaluates the timer
651 * wheel for the next timer event.
653 static void wake_up_idle_cpu(int cpu
)
655 struct rq
*rq
= cpu_rq(cpu
);
657 if (cpu
== smp_processor_id())
660 if (set_nr_and_not_polling(rq
->idle
))
661 smp_send_reschedule(cpu
);
663 trace_sched_wake_idle_without_ipi(cpu
);
666 static bool wake_up_full_nohz_cpu(int cpu
)
669 * We just need the target to call irq_exit() and re-evaluate
670 * the next tick. The nohz full kick at least implies that.
671 * If needed we can still optimize that later with an
674 if (tick_nohz_full_cpu(cpu
)) {
675 if (cpu
!= smp_processor_id() ||
676 tick_nohz_tick_stopped())
677 tick_nohz_full_kick_cpu(cpu
);
684 void wake_up_nohz_cpu(int cpu
)
686 if (!wake_up_full_nohz_cpu(cpu
))
687 wake_up_idle_cpu(cpu
);
690 static inline bool got_nohz_idle_kick(void)
692 int cpu
= smp_processor_id();
694 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
697 if (idle_cpu(cpu
) && !need_resched())
701 * We can't run Idle Load Balance on this CPU for this time so we
702 * cancel it and clear NOHZ_BALANCE_KICK
704 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
708 #else /* CONFIG_NO_HZ_COMMON */
710 static inline bool got_nohz_idle_kick(void)
715 #endif /* CONFIG_NO_HZ_COMMON */
717 #ifdef CONFIG_NO_HZ_FULL
718 bool sched_can_stop_tick(void)
721 * FIFO realtime policy runs the highest priority task. Other runnable
722 * tasks are of a lower priority. The scheduler tick does nothing.
724 if (current
->policy
== SCHED_FIFO
)
728 * Round-robin realtime tasks time slice with other tasks at the same
729 * realtime priority. Is this task the only one at this priority?
731 if (current
->policy
== SCHED_RR
) {
732 struct sched_rt_entity
*rt_se
= ¤t
->rt
;
734 return rt_se
->run_list
.prev
== rt_se
->run_list
.next
;
738 * More than one running task need preemption.
739 * nr_running update is assumed to be visible
740 * after IPI is sent from wakers.
742 if (this_rq()->nr_running
> 1)
747 #endif /* CONFIG_NO_HZ_FULL */
749 void sched_avg_update(struct rq
*rq
)
751 s64 period
= sched_avg_period();
753 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
755 * Inline assembly required to prevent the compiler
756 * optimising this loop into a divmod call.
757 * See __iter_div_u64_rem() for another example of this.
759 asm("" : "+rm" (rq
->age_stamp
));
760 rq
->age_stamp
+= period
;
765 #endif /* CONFIG_SMP */
767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
768 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
770 * Iterate task_group tree rooted at *from, calling @down when first entering a
771 * node and @up when leaving it for the final time.
773 * Caller must hold rcu_lock or sufficient equivalent.
775 int walk_tg_tree_from(struct task_group
*from
,
776 tg_visitor down
, tg_visitor up
, void *data
)
778 struct task_group
*parent
, *child
;
784 ret
= (*down
)(parent
, data
);
787 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
794 ret
= (*up
)(parent
, data
);
795 if (ret
|| parent
== from
)
799 parent
= parent
->parent
;
806 int tg_nop(struct task_group
*tg
, void *data
)
812 static void set_load_weight(struct task_struct
*p
)
814 int prio
= p
->static_prio
- MAX_RT_PRIO
;
815 struct load_weight
*load
= &p
->se
.load
;
818 * SCHED_IDLE tasks get minimal weight:
820 if (p
->policy
== SCHED_IDLE
) {
821 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
822 load
->inv_weight
= WMULT_IDLEPRIO
;
826 load
->weight
= scale_load(prio_to_weight
[prio
]);
827 load
->inv_weight
= prio_to_wmult
[prio
];
830 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
833 sched_info_queued(rq
, p
);
834 p
->sched_class
->enqueue_task(rq
, p
, flags
);
837 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
840 sched_info_dequeued(rq
, p
);
841 p
->sched_class
->dequeue_task(rq
, p
, flags
);
844 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
846 if (task_contributes_to_load(p
))
847 rq
->nr_uninterruptible
--;
849 enqueue_task(rq
, p
, flags
);
852 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
854 if (task_contributes_to_load(p
))
855 rq
->nr_uninterruptible
++;
857 dequeue_task(rq
, p
, flags
);
860 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
863 * In theory, the compile should just see 0 here, and optimize out the call
864 * to sched_rt_avg_update. But I don't trust it...
866 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
867 s64 steal
= 0, irq_delta
= 0;
869 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
870 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
873 * Since irq_time is only updated on {soft,}irq_exit, we might run into
874 * this case when a previous update_rq_clock() happened inside a
877 * When this happens, we stop ->clock_task and only update the
878 * prev_irq_time stamp to account for the part that fit, so that a next
879 * update will consume the rest. This ensures ->clock_task is
882 * It does however cause some slight miss-attribution of {soft,}irq
883 * time, a more accurate solution would be to update the irq_time using
884 * the current rq->clock timestamp, except that would require using
887 if (irq_delta
> delta
)
890 rq
->prev_irq_time
+= irq_delta
;
893 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
894 if (static_key_false((¶virt_steal_rq_enabled
))) {
895 steal
= paravirt_steal_clock(cpu_of(rq
));
896 steal
-= rq
->prev_steal_time_rq
;
898 if (unlikely(steal
> delta
))
901 rq
->prev_steal_time_rq
+= steal
;
906 rq
->clock_task
+= delta
;
908 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
909 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
910 sched_rt_avg_update(rq
, irq_delta
+ steal
);
914 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
916 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
917 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
921 * Make it appear like a SCHED_FIFO task, its something
922 * userspace knows about and won't get confused about.
924 * Also, it will make PI more or less work without too
925 * much confusion -- but then, stop work should not
926 * rely on PI working anyway.
928 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
930 stop
->sched_class
= &stop_sched_class
;
933 cpu_rq(cpu
)->stop
= stop
;
937 * Reset it back to a normal scheduling class so that
938 * it can die in pieces.
940 old_stop
->sched_class
= &rt_sched_class
;
945 * __normal_prio - return the priority that is based on the static prio
947 static inline int __normal_prio(struct task_struct
*p
)
949 return p
->static_prio
;
953 * Calculate the expected normal priority: i.e. priority
954 * without taking RT-inheritance into account. Might be
955 * boosted by interactivity modifiers. Changes upon fork,
956 * setprio syscalls, and whenever the interactivity
957 * estimator recalculates.
959 static inline int normal_prio(struct task_struct
*p
)
963 if (task_has_dl_policy(p
))
964 prio
= MAX_DL_PRIO
-1;
965 else if (task_has_rt_policy(p
))
966 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
968 prio
= __normal_prio(p
);
973 * Calculate the current priority, i.e. the priority
974 * taken into account by the scheduler. This value might
975 * be boosted by RT tasks, or might be boosted by
976 * interactivity modifiers. Will be RT if the task got
977 * RT-boosted. If not then it returns p->normal_prio.
979 static int effective_prio(struct task_struct
*p
)
981 p
->normal_prio
= normal_prio(p
);
983 * If we are RT tasks or we were boosted to RT priority,
984 * keep the priority unchanged. Otherwise, update priority
985 * to the normal priority:
987 if (!rt_prio(p
->prio
))
988 return p
->normal_prio
;
993 * task_curr - is this task currently executing on a CPU?
994 * @p: the task in question.
996 * Return: 1 if the task is currently executing. 0 otherwise.
998 inline int task_curr(const struct task_struct
*p
)
1000 return cpu_curr(task_cpu(p
)) == p
;
1004 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1005 * use the balance_callback list if you want balancing.
1007 * this means any call to check_class_changed() must be followed by a call to
1008 * balance_callback().
1010 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1011 const struct sched_class
*prev_class
,
1014 if (prev_class
!= p
->sched_class
) {
1015 if (prev_class
->switched_from
)
1016 prev_class
->switched_from(rq
, p
);
1018 p
->sched_class
->switched_to(rq
, p
);
1019 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1020 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1023 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1025 const struct sched_class
*class;
1027 if (p
->sched_class
== rq
->curr
->sched_class
) {
1028 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1030 for_each_class(class) {
1031 if (class == rq
->curr
->sched_class
)
1033 if (class == p
->sched_class
) {
1041 * A queue event has occurred, and we're going to schedule. In
1042 * this case, we can save a useless back to back clock update.
1044 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1045 rq_clock_skip_update(rq
, true);
1050 * This is how migration works:
1052 * 1) we invoke migration_cpu_stop() on the target CPU using
1054 * 2) stopper starts to run (implicitly forcing the migrated thread
1056 * 3) it checks whether the migrated task is still in the wrong runqueue.
1057 * 4) if it's in the wrong runqueue then the migration thread removes
1058 * it and puts it into the right queue.
1059 * 5) stopper completes and stop_one_cpu() returns and the migration
1064 * move_queued_task - move a queued task to new rq.
1066 * Returns (locked) new rq. Old rq's lock is released.
1068 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
1070 lockdep_assert_held(&rq
->lock
);
1072 dequeue_task(rq
, p
, 0);
1073 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1074 set_task_cpu(p
, new_cpu
);
1075 raw_spin_unlock(&rq
->lock
);
1077 rq
= cpu_rq(new_cpu
);
1079 raw_spin_lock(&rq
->lock
);
1080 BUG_ON(task_cpu(p
) != new_cpu
);
1081 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1082 enqueue_task(rq
, p
, 0);
1083 check_preempt_curr(rq
, p
, 0);
1088 struct migration_arg
{
1089 struct task_struct
*task
;
1094 * Move (not current) task off this cpu, onto dest cpu. We're doing
1095 * this because either it can't run here any more (set_cpus_allowed()
1096 * away from this CPU, or CPU going down), or because we're
1097 * attempting to rebalance this task on exec (sched_exec).
1099 * So we race with normal scheduler movements, but that's OK, as long
1100 * as the task is no longer on this CPU.
1102 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1104 if (unlikely(!cpu_active(dest_cpu
)))
1107 /* Affinity changed (again). */
1108 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1111 rq
= move_queued_task(rq
, p
, dest_cpu
);
1117 * migration_cpu_stop - this will be executed by a highprio stopper thread
1118 * and performs thread migration by bumping thread off CPU then
1119 * 'pushing' onto another runqueue.
1121 static int migration_cpu_stop(void *data
)
1123 struct migration_arg
*arg
= data
;
1124 struct task_struct
*p
= arg
->task
;
1125 struct rq
*rq
= this_rq();
1128 * The original target cpu might have gone down and we might
1129 * be on another cpu but it doesn't matter.
1131 local_irq_disable();
1133 * We need to explicitly wake pending tasks before running
1134 * __migrate_task() such that we will not miss enforcing cpus_allowed
1135 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1137 sched_ttwu_pending();
1139 raw_spin_lock(&p
->pi_lock
);
1140 raw_spin_lock(&rq
->lock
);
1142 * If task_rq(p) != rq, it cannot be migrated here, because we're
1143 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1144 * we're holding p->pi_lock.
1146 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1147 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1148 raw_spin_unlock(&rq
->lock
);
1149 raw_spin_unlock(&p
->pi_lock
);
1156 * sched_class::set_cpus_allowed must do the below, but is not required to
1157 * actually call this function.
1159 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1161 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1162 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1165 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1167 struct rq
*rq
= task_rq(p
);
1168 bool queued
, running
;
1170 lockdep_assert_held(&p
->pi_lock
);
1172 queued
= task_on_rq_queued(p
);
1173 running
= task_current(rq
, p
);
1177 * Because __kthread_bind() calls this on blocked tasks without
1180 lockdep_assert_held(&rq
->lock
);
1181 dequeue_task(rq
, p
, 0);
1184 put_prev_task(rq
, p
);
1186 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1189 p
->sched_class
->set_curr_task(rq
);
1191 enqueue_task(rq
, p
, 0);
1195 * Change a given task's CPU affinity. Migrate the thread to a
1196 * proper CPU and schedule it away if the CPU it's executing on
1197 * is removed from the allowed bitmask.
1199 * NOTE: the caller must have a valid reference to the task, the
1200 * task must not exit() & deallocate itself prematurely. The
1201 * call is not atomic; no spinlocks may be held.
1203 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1204 const struct cpumask
*new_mask
, bool check
)
1206 unsigned long flags
;
1208 unsigned int dest_cpu
;
1211 rq
= task_rq_lock(p
, &flags
);
1214 * Must re-check here, to close a race against __kthread_bind(),
1215 * sched_setaffinity() is not guaranteed to observe the flag.
1217 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1222 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1225 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1230 do_set_cpus_allowed(p
, new_mask
);
1232 /* Can the task run on the task's current CPU? If so, we're done */
1233 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1236 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1237 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1238 struct migration_arg arg
= { p
, dest_cpu
};
1239 /* Need help from migration thread: drop lock and wait. */
1240 task_rq_unlock(rq
, p
, &flags
);
1241 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1242 tlb_migrate_finish(p
->mm
);
1244 } else if (task_on_rq_queued(p
)) {
1246 * OK, since we're going to drop the lock immediately
1247 * afterwards anyway.
1249 lockdep_unpin_lock(&rq
->lock
);
1250 rq
= move_queued_task(rq
, p
, dest_cpu
);
1251 lockdep_pin_lock(&rq
->lock
);
1254 task_rq_unlock(rq
, p
, &flags
);
1259 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1261 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1263 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1265 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1267 #ifdef CONFIG_SCHED_DEBUG
1269 * We should never call set_task_cpu() on a blocked task,
1270 * ttwu() will sort out the placement.
1272 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1275 #ifdef CONFIG_LOCKDEP
1277 * The caller should hold either p->pi_lock or rq->lock, when changing
1278 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1280 * sched_move_task() holds both and thus holding either pins the cgroup,
1283 * Furthermore, all task_rq users should acquire both locks, see
1286 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1287 lockdep_is_held(&task_rq(p
)->lock
)));
1291 trace_sched_migrate_task(p
, new_cpu
);
1293 if (task_cpu(p
) != new_cpu
) {
1294 if (p
->sched_class
->migrate_task_rq
)
1295 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1296 p
->se
.nr_migrations
++;
1297 perf_event_task_migrate(p
);
1300 __set_task_cpu(p
, new_cpu
);
1303 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1305 if (task_on_rq_queued(p
)) {
1306 struct rq
*src_rq
, *dst_rq
;
1308 src_rq
= task_rq(p
);
1309 dst_rq
= cpu_rq(cpu
);
1311 deactivate_task(src_rq
, p
, 0);
1312 set_task_cpu(p
, cpu
);
1313 activate_task(dst_rq
, p
, 0);
1314 check_preempt_curr(dst_rq
, p
, 0);
1317 * Task isn't running anymore; make it appear like we migrated
1318 * it before it went to sleep. This means on wakeup we make the
1319 * previous cpu our targer instead of where it really is.
1325 struct migration_swap_arg
{
1326 struct task_struct
*src_task
, *dst_task
;
1327 int src_cpu
, dst_cpu
;
1330 static int migrate_swap_stop(void *data
)
1332 struct migration_swap_arg
*arg
= data
;
1333 struct rq
*src_rq
, *dst_rq
;
1336 src_rq
= cpu_rq(arg
->src_cpu
);
1337 dst_rq
= cpu_rq(arg
->dst_cpu
);
1339 double_raw_lock(&arg
->src_task
->pi_lock
,
1340 &arg
->dst_task
->pi_lock
);
1341 double_rq_lock(src_rq
, dst_rq
);
1342 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1345 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1348 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1351 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1354 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1355 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1360 double_rq_unlock(src_rq
, dst_rq
);
1361 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1362 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1368 * Cross migrate two tasks
1370 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1372 struct migration_swap_arg arg
;
1375 arg
= (struct migration_swap_arg
){
1377 .src_cpu
= task_cpu(cur
),
1379 .dst_cpu
= task_cpu(p
),
1382 if (arg
.src_cpu
== arg
.dst_cpu
)
1386 * These three tests are all lockless; this is OK since all of them
1387 * will be re-checked with proper locks held further down the line.
1389 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1392 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1395 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1398 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1399 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1406 * wait_task_inactive - wait for a thread to unschedule.
1408 * If @match_state is nonzero, it's the @p->state value just checked and
1409 * not expected to change. If it changes, i.e. @p might have woken up,
1410 * then return zero. When we succeed in waiting for @p to be off its CPU,
1411 * we return a positive number (its total switch count). If a second call
1412 * a short while later returns the same number, the caller can be sure that
1413 * @p has remained unscheduled the whole time.
1415 * The caller must ensure that the task *will* unschedule sometime soon,
1416 * else this function might spin for a *long* time. This function can't
1417 * be called with interrupts off, or it may introduce deadlock with
1418 * smp_call_function() if an IPI is sent by the same process we are
1419 * waiting to become inactive.
1421 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1423 unsigned long flags
;
1424 int running
, queued
;
1430 * We do the initial early heuristics without holding
1431 * any task-queue locks at all. We'll only try to get
1432 * the runqueue lock when things look like they will
1438 * If the task is actively running on another CPU
1439 * still, just relax and busy-wait without holding
1442 * NOTE! Since we don't hold any locks, it's not
1443 * even sure that "rq" stays as the right runqueue!
1444 * But we don't care, since "task_running()" will
1445 * return false if the runqueue has changed and p
1446 * is actually now running somewhere else!
1448 while (task_running(rq
, p
)) {
1449 if (match_state
&& unlikely(p
->state
!= match_state
))
1455 * Ok, time to look more closely! We need the rq
1456 * lock now, to be *sure*. If we're wrong, we'll
1457 * just go back and repeat.
1459 rq
= task_rq_lock(p
, &flags
);
1460 trace_sched_wait_task(p
);
1461 running
= task_running(rq
, p
);
1462 queued
= task_on_rq_queued(p
);
1464 if (!match_state
|| p
->state
== match_state
)
1465 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1466 task_rq_unlock(rq
, p
, &flags
);
1469 * If it changed from the expected state, bail out now.
1471 if (unlikely(!ncsw
))
1475 * Was it really running after all now that we
1476 * checked with the proper locks actually held?
1478 * Oops. Go back and try again..
1480 if (unlikely(running
)) {
1486 * It's not enough that it's not actively running,
1487 * it must be off the runqueue _entirely_, and not
1490 * So if it was still runnable (but just not actively
1491 * running right now), it's preempted, and we should
1492 * yield - it could be a while.
1494 if (unlikely(queued
)) {
1495 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1497 set_current_state(TASK_UNINTERRUPTIBLE
);
1498 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1503 * Ahh, all good. It wasn't running, and it wasn't
1504 * runnable, which means that it will never become
1505 * running in the future either. We're all done!
1514 * kick_process - kick a running thread to enter/exit the kernel
1515 * @p: the to-be-kicked thread
1517 * Cause a process which is running on another CPU to enter
1518 * kernel-mode, without any delay. (to get signals handled.)
1520 * NOTE: this function doesn't have to take the runqueue lock,
1521 * because all it wants to ensure is that the remote task enters
1522 * the kernel. If the IPI races and the task has been migrated
1523 * to another CPU then no harm is done and the purpose has been
1526 void kick_process(struct task_struct
*p
)
1532 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1533 smp_send_reschedule(cpu
);
1536 EXPORT_SYMBOL_GPL(kick_process
);
1539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1541 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1543 int nid
= cpu_to_node(cpu
);
1544 const struct cpumask
*nodemask
= NULL
;
1545 enum { cpuset
, possible
, fail
} state
= cpuset
;
1549 * If the node that the cpu is on has been offlined, cpu_to_node()
1550 * will return -1. There is no cpu on the node, and we should
1551 * select the cpu on the other node.
1554 nodemask
= cpumask_of_node(nid
);
1556 /* Look for allowed, online CPU in same node. */
1557 for_each_cpu(dest_cpu
, nodemask
) {
1558 if (!cpu_online(dest_cpu
))
1560 if (!cpu_active(dest_cpu
))
1562 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1568 /* Any allowed, online CPU? */
1569 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1570 if (!cpu_online(dest_cpu
))
1572 if (!cpu_active(dest_cpu
))
1579 /* No more Mr. Nice Guy. */
1580 cpuset_cpus_allowed_fallback(p
);
1585 do_set_cpus_allowed(p
, cpu_possible_mask
);
1596 if (state
!= cpuset
) {
1598 * Don't tell them about moving exiting tasks or
1599 * kernel threads (both mm NULL), since they never
1602 if (p
->mm
&& printk_ratelimit()) {
1603 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1604 task_pid_nr(p
), p
->comm
, cpu
);
1612 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1615 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1617 lockdep_assert_held(&p
->pi_lock
);
1619 if (p
->nr_cpus_allowed
> 1)
1620 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1623 * In order not to call set_task_cpu() on a blocking task we need
1624 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1627 * Since this is common to all placement strategies, this lives here.
1629 * [ this allows ->select_task() to simply return task_cpu(p) and
1630 * not worry about this generic constraint ]
1632 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1634 cpu
= select_fallback_rq(task_cpu(p
), p
);
1639 static void update_avg(u64
*avg
, u64 sample
)
1641 s64 diff
= sample
- *avg
;
1647 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1648 const struct cpumask
*new_mask
, bool check
)
1650 return set_cpus_allowed_ptr(p
, new_mask
);
1653 #endif /* CONFIG_SMP */
1656 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1658 #ifdef CONFIG_SCHEDSTATS
1659 struct rq
*rq
= this_rq();
1662 int this_cpu
= smp_processor_id();
1664 if (cpu
== this_cpu
) {
1665 schedstat_inc(rq
, ttwu_local
);
1666 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1668 struct sched_domain
*sd
;
1670 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1672 for_each_domain(this_cpu
, sd
) {
1673 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1674 schedstat_inc(sd
, ttwu_wake_remote
);
1681 if (wake_flags
& WF_MIGRATED
)
1682 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1684 #endif /* CONFIG_SMP */
1686 schedstat_inc(rq
, ttwu_count
);
1687 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1689 if (wake_flags
& WF_SYNC
)
1690 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1692 #endif /* CONFIG_SCHEDSTATS */
1695 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1697 activate_task(rq
, p
, en_flags
);
1698 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1700 /* if a worker is waking up, notify workqueue */
1701 if (p
->flags
& PF_WQ_WORKER
)
1702 wq_worker_waking_up(p
, cpu_of(rq
));
1706 * Mark the task runnable and perform wakeup-preemption.
1709 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1711 check_preempt_curr(rq
, p
, wake_flags
);
1712 p
->state
= TASK_RUNNING
;
1713 trace_sched_wakeup(p
);
1716 if (p
->sched_class
->task_woken
) {
1718 * Our task @p is fully woken up and running; so its safe to
1719 * drop the rq->lock, hereafter rq is only used for statistics.
1721 lockdep_unpin_lock(&rq
->lock
);
1722 p
->sched_class
->task_woken(rq
, p
);
1723 lockdep_pin_lock(&rq
->lock
);
1726 if (rq
->idle_stamp
) {
1727 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1728 u64 max
= 2*rq
->max_idle_balance_cost
;
1730 update_avg(&rq
->avg_idle
, delta
);
1732 if (rq
->avg_idle
> max
)
1741 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1743 lockdep_assert_held(&rq
->lock
);
1746 if (p
->sched_contributes_to_load
)
1747 rq
->nr_uninterruptible
--;
1750 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1751 ttwu_do_wakeup(rq
, p
, wake_flags
);
1755 * Called in case the task @p isn't fully descheduled from its runqueue,
1756 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1757 * since all we need to do is flip p->state to TASK_RUNNING, since
1758 * the task is still ->on_rq.
1760 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1765 rq
= __task_rq_lock(p
);
1766 if (task_on_rq_queued(p
)) {
1767 /* check_preempt_curr() may use rq clock */
1768 update_rq_clock(rq
);
1769 ttwu_do_wakeup(rq
, p
, wake_flags
);
1772 __task_rq_unlock(rq
);
1778 void sched_ttwu_pending(void)
1780 struct rq
*rq
= this_rq();
1781 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1782 struct task_struct
*p
;
1783 unsigned long flags
;
1788 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1789 lockdep_pin_lock(&rq
->lock
);
1792 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1793 llist
= llist_next(llist
);
1794 ttwu_do_activate(rq
, p
, 0);
1797 lockdep_unpin_lock(&rq
->lock
);
1798 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1801 void scheduler_ipi(void)
1804 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1805 * TIF_NEED_RESCHED remotely (for the first time) will also send
1808 preempt_fold_need_resched();
1810 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1814 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1815 * traditionally all their work was done from the interrupt return
1816 * path. Now that we actually do some work, we need to make sure
1819 * Some archs already do call them, luckily irq_enter/exit nest
1822 * Arguably we should visit all archs and update all handlers,
1823 * however a fair share of IPIs are still resched only so this would
1824 * somewhat pessimize the simple resched case.
1827 sched_ttwu_pending();
1830 * Check if someone kicked us for doing the nohz idle load balance.
1832 if (unlikely(got_nohz_idle_kick())) {
1833 this_rq()->idle_balance
= 1;
1834 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1839 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1841 struct rq
*rq
= cpu_rq(cpu
);
1843 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1844 if (!set_nr_if_polling(rq
->idle
))
1845 smp_send_reschedule(cpu
);
1847 trace_sched_wake_idle_without_ipi(cpu
);
1851 void wake_up_if_idle(int cpu
)
1853 struct rq
*rq
= cpu_rq(cpu
);
1854 unsigned long flags
;
1858 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1861 if (set_nr_if_polling(rq
->idle
)) {
1862 trace_sched_wake_idle_without_ipi(cpu
);
1864 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1865 if (is_idle_task(rq
->curr
))
1866 smp_send_reschedule(cpu
);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1875 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1877 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1879 #endif /* CONFIG_SMP */
1881 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1883 struct rq
*rq
= cpu_rq(cpu
);
1885 #if defined(CONFIG_SMP)
1886 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1887 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1888 ttwu_queue_remote(p
, cpu
);
1893 raw_spin_lock(&rq
->lock
);
1894 lockdep_pin_lock(&rq
->lock
);
1895 ttwu_do_activate(rq
, p
, 0);
1896 lockdep_unpin_lock(&rq
->lock
);
1897 raw_spin_unlock(&rq
->lock
);
1901 * try_to_wake_up - wake up a thread
1902 * @p: the thread to be awakened
1903 * @state: the mask of task states that can be woken
1904 * @wake_flags: wake modifier flags (WF_*)
1906 * Put it on the run-queue if it's not already there. The "current"
1907 * thread is always on the run-queue (except when the actual
1908 * re-schedule is in progress), and as such you're allowed to do
1909 * the simpler "current->state = TASK_RUNNING" to mark yourself
1910 * runnable without the overhead of this.
1912 * Return: %true if @p was woken up, %false if it was already running.
1913 * or @state didn't match @p's state.
1916 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1918 unsigned long flags
;
1919 int cpu
, success
= 0;
1922 * If we are going to wake up a thread waiting for CONDITION we
1923 * need to ensure that CONDITION=1 done by the caller can not be
1924 * reordered with p->state check below. This pairs with mb() in
1925 * set_current_state() the waiting thread does.
1927 smp_mb__before_spinlock();
1928 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1929 if (!(p
->state
& state
))
1932 trace_sched_waking(p
);
1934 success
= 1; /* we're going to change ->state */
1937 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1942 * If the owning (remote) cpu is still in the middle of schedule() with
1943 * this task as prev, wait until its done referencing the task.
1948 * Pairs with the smp_wmb() in finish_lock_switch().
1952 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1953 p
->state
= TASK_WAKING
;
1955 if (p
->sched_class
->task_waking
)
1956 p
->sched_class
->task_waking(p
);
1958 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1959 if (task_cpu(p
) != cpu
) {
1960 wake_flags
|= WF_MIGRATED
;
1961 set_task_cpu(p
, cpu
);
1963 #endif /* CONFIG_SMP */
1967 ttwu_stat(p
, cpu
, wake_flags
);
1969 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1975 * try_to_wake_up_local - try to wake up a local task with rq lock held
1976 * @p: the thread to be awakened
1978 * Put @p on the run-queue if it's not already there. The caller must
1979 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1982 static void try_to_wake_up_local(struct task_struct
*p
)
1984 struct rq
*rq
= task_rq(p
);
1986 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1987 WARN_ON_ONCE(p
== current
))
1990 lockdep_assert_held(&rq
->lock
);
1992 if (!raw_spin_trylock(&p
->pi_lock
)) {
1994 * This is OK, because current is on_cpu, which avoids it being
1995 * picked for load-balance and preemption/IRQs are still
1996 * disabled avoiding further scheduler activity on it and we've
1997 * not yet picked a replacement task.
1999 lockdep_unpin_lock(&rq
->lock
);
2000 raw_spin_unlock(&rq
->lock
);
2001 raw_spin_lock(&p
->pi_lock
);
2002 raw_spin_lock(&rq
->lock
);
2003 lockdep_pin_lock(&rq
->lock
);
2006 if (!(p
->state
& TASK_NORMAL
))
2009 trace_sched_waking(p
);
2011 if (!task_on_rq_queued(p
))
2012 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2014 ttwu_do_wakeup(rq
, p
, 0);
2015 ttwu_stat(p
, smp_processor_id(), 0);
2017 raw_spin_unlock(&p
->pi_lock
);
2021 * wake_up_process - Wake up a specific process
2022 * @p: The process to be woken up.
2024 * Attempt to wake up the nominated process and move it to the set of runnable
2027 * Return: 1 if the process was woken up, 0 if it was already running.
2029 * It may be assumed that this function implies a write memory barrier before
2030 * changing the task state if and only if any tasks are woken up.
2032 int wake_up_process(struct task_struct
*p
)
2034 WARN_ON(task_is_stopped_or_traced(p
));
2035 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2037 EXPORT_SYMBOL(wake_up_process
);
2039 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2041 return try_to_wake_up(p
, state
, 0);
2045 * This function clears the sched_dl_entity static params.
2047 void __dl_clear_params(struct task_struct
*p
)
2049 struct sched_dl_entity
*dl_se
= &p
->dl
;
2051 dl_se
->dl_runtime
= 0;
2052 dl_se
->dl_deadline
= 0;
2053 dl_se
->dl_period
= 0;
2057 dl_se
->dl_throttled
= 0;
2059 dl_se
->dl_yielded
= 0;
2063 * Perform scheduler related setup for a newly forked process p.
2064 * p is forked by current.
2066 * __sched_fork() is basic setup used by init_idle() too:
2068 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2073 p
->se
.exec_start
= 0;
2074 p
->se
.sum_exec_runtime
= 0;
2075 p
->se
.prev_sum_exec_runtime
= 0;
2076 p
->se
.nr_migrations
= 0;
2078 INIT_LIST_HEAD(&p
->se
.group_node
);
2080 #ifdef CONFIG_SCHEDSTATS
2081 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2084 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2085 init_dl_task_timer(&p
->dl
);
2086 __dl_clear_params(p
);
2088 INIT_LIST_HEAD(&p
->rt
.run_list
);
2090 #ifdef CONFIG_PREEMPT_NOTIFIERS
2091 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2094 #ifdef CONFIG_NUMA_BALANCING
2095 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2096 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2097 p
->mm
->numa_scan_seq
= 0;
2100 if (clone_flags
& CLONE_VM
)
2101 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2103 p
->numa_preferred_nid
= -1;
2105 p
->node_stamp
= 0ULL;
2106 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2107 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2108 p
->numa_work
.next
= &p
->numa_work
;
2109 p
->numa_faults
= NULL
;
2110 p
->last_task_numa_placement
= 0;
2111 p
->last_sum_exec_runtime
= 0;
2113 p
->numa_group
= NULL
;
2114 #endif /* CONFIG_NUMA_BALANCING */
2117 #ifdef CONFIG_NUMA_BALANCING
2118 #ifdef CONFIG_SCHED_DEBUG
2119 void set_numabalancing_state(bool enabled
)
2122 sched_feat_set("NUMA");
2124 sched_feat_set("NO_NUMA");
2127 __read_mostly
bool numabalancing_enabled
;
2129 void set_numabalancing_state(bool enabled
)
2131 numabalancing_enabled
= enabled
;
2133 #endif /* CONFIG_SCHED_DEBUG */
2135 #ifdef CONFIG_PROC_SYSCTL
2136 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2137 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2141 int state
= numabalancing_enabled
;
2143 if (write
&& !capable(CAP_SYS_ADMIN
))
2148 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2152 set_numabalancing_state(state
);
2159 * fork()/clone()-time setup:
2161 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2163 unsigned long flags
;
2164 int cpu
= get_cpu();
2166 __sched_fork(clone_flags
, p
);
2168 * We mark the process as running here. This guarantees that
2169 * nobody will actually run it, and a signal or other external
2170 * event cannot wake it up and insert it on the runqueue either.
2172 p
->state
= TASK_RUNNING
;
2175 * Make sure we do not leak PI boosting priority to the child.
2177 p
->prio
= current
->normal_prio
;
2180 * Revert to default priority/policy on fork if requested.
2182 if (unlikely(p
->sched_reset_on_fork
)) {
2183 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2184 p
->policy
= SCHED_NORMAL
;
2185 p
->static_prio
= NICE_TO_PRIO(0);
2187 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2188 p
->static_prio
= NICE_TO_PRIO(0);
2190 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2194 * We don't need the reset flag anymore after the fork. It has
2195 * fulfilled its duty:
2197 p
->sched_reset_on_fork
= 0;
2200 if (dl_prio(p
->prio
)) {
2203 } else if (rt_prio(p
->prio
)) {
2204 p
->sched_class
= &rt_sched_class
;
2206 p
->sched_class
= &fair_sched_class
;
2209 if (p
->sched_class
->task_fork
)
2210 p
->sched_class
->task_fork(p
);
2213 * The child is not yet in the pid-hash so no cgroup attach races,
2214 * and the cgroup is pinned to this child due to cgroup_fork()
2215 * is ran before sched_fork().
2217 * Silence PROVE_RCU.
2219 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2220 set_task_cpu(p
, cpu
);
2221 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2223 #ifdef CONFIG_SCHED_INFO
2224 if (likely(sched_info_on()))
2225 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2227 #if defined(CONFIG_SMP)
2230 init_task_preempt_count(p
);
2232 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2233 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2240 unsigned long to_ratio(u64 period
, u64 runtime
)
2242 if (runtime
== RUNTIME_INF
)
2246 * Doing this here saves a lot of checks in all
2247 * the calling paths, and returning zero seems
2248 * safe for them anyway.
2253 return div64_u64(runtime
<< 20, period
);
2257 inline struct dl_bw
*dl_bw_of(int i
)
2259 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2260 "sched RCU must be held");
2261 return &cpu_rq(i
)->rd
->dl_bw
;
2264 static inline int dl_bw_cpus(int i
)
2266 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2269 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2270 "sched RCU must be held");
2271 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2277 inline struct dl_bw
*dl_bw_of(int i
)
2279 return &cpu_rq(i
)->dl
.dl_bw
;
2282 static inline int dl_bw_cpus(int i
)
2289 * We must be sure that accepting a new task (or allowing changing the
2290 * parameters of an existing one) is consistent with the bandwidth
2291 * constraints. If yes, this function also accordingly updates the currently
2292 * allocated bandwidth to reflect the new situation.
2294 * This function is called while holding p's rq->lock.
2296 * XXX we should delay bw change until the task's 0-lag point, see
2299 static int dl_overflow(struct task_struct
*p
, int policy
,
2300 const struct sched_attr
*attr
)
2303 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2304 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2305 u64 runtime
= attr
->sched_runtime
;
2306 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2309 if (new_bw
== p
->dl
.dl_bw
)
2313 * Either if a task, enters, leave, or stays -deadline but changes
2314 * its parameters, we may need to update accordingly the total
2315 * allocated bandwidth of the container.
2317 raw_spin_lock(&dl_b
->lock
);
2318 cpus
= dl_bw_cpus(task_cpu(p
));
2319 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2320 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2321 __dl_add(dl_b
, new_bw
);
2323 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2324 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2325 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2326 __dl_add(dl_b
, new_bw
);
2328 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2329 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2332 raw_spin_unlock(&dl_b
->lock
);
2337 extern void init_dl_bw(struct dl_bw
*dl_b
);
2340 * wake_up_new_task - wake up a newly created task for the first time.
2342 * This function will do some initial scheduler statistics housekeeping
2343 * that must be done for every newly created context, then puts the task
2344 * on the runqueue and wakes it.
2346 void wake_up_new_task(struct task_struct
*p
)
2348 unsigned long flags
;
2351 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2354 * Fork balancing, do it here and not earlier because:
2355 * - cpus_allowed can change in the fork path
2356 * - any previously selected cpu might disappear through hotplug
2358 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2361 /* Initialize new task's runnable average */
2362 init_entity_runnable_average(&p
->se
);
2363 rq
= __task_rq_lock(p
);
2364 activate_task(rq
, p
, 0);
2365 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2366 trace_sched_wakeup_new(p
);
2367 check_preempt_curr(rq
, p
, WF_FORK
);
2369 if (p
->sched_class
->task_woken
) {
2371 * Nothing relies on rq->lock after this, so its fine to
2374 lockdep_unpin_lock(&rq
->lock
);
2375 p
->sched_class
->task_woken(rq
, p
);
2376 lockdep_pin_lock(&rq
->lock
);
2379 task_rq_unlock(rq
, p
, &flags
);
2382 #ifdef CONFIG_PREEMPT_NOTIFIERS
2384 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2386 void preempt_notifier_inc(void)
2388 static_key_slow_inc(&preempt_notifier_key
);
2390 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2392 void preempt_notifier_dec(void)
2394 static_key_slow_dec(&preempt_notifier_key
);
2396 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2399 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2400 * @notifier: notifier struct to register
2402 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2404 if (!static_key_false(&preempt_notifier_key
))
2405 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2407 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2409 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2412 * preempt_notifier_unregister - no longer interested in preemption notifications
2413 * @notifier: notifier struct to unregister
2415 * This is *not* safe to call from within a preemption notifier.
2417 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2419 hlist_del(¬ifier
->link
);
2421 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2423 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2425 struct preempt_notifier
*notifier
;
2427 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2428 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2431 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2433 if (static_key_false(&preempt_notifier_key
))
2434 __fire_sched_in_preempt_notifiers(curr
);
2438 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2439 struct task_struct
*next
)
2441 struct preempt_notifier
*notifier
;
2443 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2444 notifier
->ops
->sched_out(notifier
, next
);
2447 static __always_inline
void
2448 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2449 struct task_struct
*next
)
2451 if (static_key_false(&preempt_notifier_key
))
2452 __fire_sched_out_preempt_notifiers(curr
, next
);
2455 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2457 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2462 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2463 struct task_struct
*next
)
2467 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2470 * prepare_task_switch - prepare to switch tasks
2471 * @rq: the runqueue preparing to switch
2472 * @prev: the current task that is being switched out
2473 * @next: the task we are going to switch to.
2475 * This is called with the rq lock held and interrupts off. It must
2476 * be paired with a subsequent finish_task_switch after the context
2479 * prepare_task_switch sets up locking and calls architecture specific
2483 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2484 struct task_struct
*next
)
2486 trace_sched_switch(prev
, next
);
2487 sched_info_switch(rq
, prev
, next
);
2488 perf_event_task_sched_out(prev
, next
);
2489 fire_sched_out_preempt_notifiers(prev
, next
);
2490 prepare_lock_switch(rq
, next
);
2491 prepare_arch_switch(next
);
2495 * finish_task_switch - clean up after a task-switch
2496 * @prev: the thread we just switched away from.
2498 * finish_task_switch must be called after the context switch, paired
2499 * with a prepare_task_switch call before the context switch.
2500 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2501 * and do any other architecture-specific cleanup actions.
2503 * Note that we may have delayed dropping an mm in context_switch(). If
2504 * so, we finish that here outside of the runqueue lock. (Doing it
2505 * with the lock held can cause deadlocks; see schedule() for
2508 * The context switch have flipped the stack from under us and restored the
2509 * local variables which were saved when this task called schedule() in the
2510 * past. prev == current is still correct but we need to recalculate this_rq
2511 * because prev may have moved to another CPU.
2513 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2514 __releases(rq
->lock
)
2516 struct rq
*rq
= this_rq();
2517 struct mm_struct
*mm
= rq
->prev_mm
;
2523 * A task struct has one reference for the use as "current".
2524 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2525 * schedule one last time. The schedule call will never return, and
2526 * the scheduled task must drop that reference.
2528 * We must observe prev->state before clearing prev->on_cpu (in
2529 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2530 * running on another CPU and we could rave with its RUNNING -> DEAD
2531 * transition, resulting in a double drop.
2533 prev_state
= prev
->state
;
2534 vtime_task_switch(prev
);
2535 perf_event_task_sched_in(prev
, current
);
2536 finish_lock_switch(rq
, prev
);
2537 finish_arch_post_lock_switch();
2539 fire_sched_in_preempt_notifiers(current
);
2542 if (unlikely(prev_state
== TASK_DEAD
)) {
2543 if (prev
->sched_class
->task_dead
)
2544 prev
->sched_class
->task_dead(prev
);
2547 * Remove function-return probe instances associated with this
2548 * task and put them back on the free list.
2550 kprobe_flush_task(prev
);
2551 put_task_struct(prev
);
2554 tick_nohz_task_switch();
2560 /* rq->lock is NOT held, but preemption is disabled */
2561 static void __balance_callback(struct rq
*rq
)
2563 struct callback_head
*head
, *next
;
2564 void (*func
)(struct rq
*rq
);
2565 unsigned long flags
;
2567 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2568 head
= rq
->balance_callback
;
2569 rq
->balance_callback
= NULL
;
2571 func
= (void (*)(struct rq
*))head
->func
;
2578 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2581 static inline void balance_callback(struct rq
*rq
)
2583 if (unlikely(rq
->balance_callback
))
2584 __balance_callback(rq
);
2589 static inline void balance_callback(struct rq
*rq
)
2596 * schedule_tail - first thing a freshly forked thread must call.
2597 * @prev: the thread we just switched away from.
2599 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2600 __releases(rq
->lock
)
2604 /* finish_task_switch() drops rq->lock and enables preemtion */
2606 rq
= finish_task_switch(prev
);
2607 balance_callback(rq
);
2610 if (current
->set_child_tid
)
2611 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2615 * context_switch - switch to the new MM and the new thread's register state.
2617 static inline struct rq
*
2618 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2619 struct task_struct
*next
)
2621 struct mm_struct
*mm
, *oldmm
;
2623 prepare_task_switch(rq
, prev
, next
);
2626 oldmm
= prev
->active_mm
;
2628 * For paravirt, this is coupled with an exit in switch_to to
2629 * combine the page table reload and the switch backend into
2632 arch_start_context_switch(prev
);
2635 next
->active_mm
= oldmm
;
2636 atomic_inc(&oldmm
->mm_count
);
2637 enter_lazy_tlb(oldmm
, next
);
2639 switch_mm(oldmm
, mm
, next
);
2642 prev
->active_mm
= NULL
;
2643 rq
->prev_mm
= oldmm
;
2646 * Since the runqueue lock will be released by the next
2647 * task (which is an invalid locking op but in the case
2648 * of the scheduler it's an obvious special-case), so we
2649 * do an early lockdep release here:
2651 lockdep_unpin_lock(&rq
->lock
);
2652 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2654 /* Here we just switch the register state and the stack. */
2655 switch_to(prev
, next
, prev
);
2658 return finish_task_switch(prev
);
2662 * nr_running and nr_context_switches:
2664 * externally visible scheduler statistics: current number of runnable
2665 * threads, total number of context switches performed since bootup.
2667 unsigned long nr_running(void)
2669 unsigned long i
, sum
= 0;
2671 for_each_online_cpu(i
)
2672 sum
+= cpu_rq(i
)->nr_running
;
2678 * Check if only the current task is running on the cpu.
2680 * Caution: this function does not check that the caller has disabled
2681 * preemption, thus the result might have a time-of-check-to-time-of-use
2682 * race. The caller is responsible to use it correctly, for example:
2684 * - from a non-preemptable section (of course)
2686 * - from a thread that is bound to a single CPU
2688 * - in a loop with very short iterations (e.g. a polling loop)
2690 bool single_task_running(void)
2692 return raw_rq()->nr_running
== 1;
2694 EXPORT_SYMBOL(single_task_running
);
2696 unsigned long long nr_context_switches(void)
2699 unsigned long long sum
= 0;
2701 for_each_possible_cpu(i
)
2702 sum
+= cpu_rq(i
)->nr_switches
;
2707 unsigned long nr_iowait(void)
2709 unsigned long i
, sum
= 0;
2711 for_each_possible_cpu(i
)
2712 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2717 unsigned long nr_iowait_cpu(int cpu
)
2719 struct rq
*this = cpu_rq(cpu
);
2720 return atomic_read(&this->nr_iowait
);
2723 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2725 struct rq
*rq
= this_rq();
2726 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2727 *load
= rq
->load
.weight
;
2733 * sched_exec - execve() is a valuable balancing opportunity, because at
2734 * this point the task has the smallest effective memory and cache footprint.
2736 void sched_exec(void)
2738 struct task_struct
*p
= current
;
2739 unsigned long flags
;
2742 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2743 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2744 if (dest_cpu
== smp_processor_id())
2747 if (likely(cpu_active(dest_cpu
))) {
2748 struct migration_arg arg
= { p
, dest_cpu
};
2750 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2751 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2755 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2760 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2761 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2763 EXPORT_PER_CPU_SYMBOL(kstat
);
2764 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2767 * Return accounted runtime for the task.
2768 * In case the task is currently running, return the runtime plus current's
2769 * pending runtime that have not been accounted yet.
2771 unsigned long long task_sched_runtime(struct task_struct
*p
)
2773 unsigned long flags
;
2777 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2779 * 64-bit doesn't need locks to atomically read a 64bit value.
2780 * So we have a optimization chance when the task's delta_exec is 0.
2781 * Reading ->on_cpu is racy, but this is ok.
2783 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2784 * If we race with it entering cpu, unaccounted time is 0. This is
2785 * indistinguishable from the read occurring a few cycles earlier.
2786 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2787 * been accounted, so we're correct here as well.
2789 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2790 return p
->se
.sum_exec_runtime
;
2793 rq
= task_rq_lock(p
, &flags
);
2795 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2796 * project cycles that may never be accounted to this
2797 * thread, breaking clock_gettime().
2799 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2800 update_rq_clock(rq
);
2801 p
->sched_class
->update_curr(rq
);
2803 ns
= p
->se
.sum_exec_runtime
;
2804 task_rq_unlock(rq
, p
, &flags
);
2810 * This function gets called by the timer code, with HZ frequency.
2811 * We call it with interrupts disabled.
2813 void scheduler_tick(void)
2815 int cpu
= smp_processor_id();
2816 struct rq
*rq
= cpu_rq(cpu
);
2817 struct task_struct
*curr
= rq
->curr
;
2821 raw_spin_lock(&rq
->lock
);
2822 update_rq_clock(rq
);
2823 curr
->sched_class
->task_tick(rq
, curr
, 0);
2824 update_cpu_load_active(rq
);
2825 calc_global_load_tick(rq
);
2826 raw_spin_unlock(&rq
->lock
);
2828 perf_event_task_tick();
2831 rq
->idle_balance
= idle_cpu(cpu
);
2832 trigger_load_balance(rq
);
2834 rq_last_tick_reset(rq
);
2837 #ifdef CONFIG_NO_HZ_FULL
2839 * scheduler_tick_max_deferment
2841 * Keep at least one tick per second when a single
2842 * active task is running because the scheduler doesn't
2843 * yet completely support full dynticks environment.
2845 * This makes sure that uptime, CFS vruntime, load
2846 * balancing, etc... continue to move forward, even
2847 * with a very low granularity.
2849 * Return: Maximum deferment in nanoseconds.
2851 u64
scheduler_tick_max_deferment(void)
2853 struct rq
*rq
= this_rq();
2854 unsigned long next
, now
= READ_ONCE(jiffies
);
2856 next
= rq
->last_sched_tick
+ HZ
;
2858 if (time_before_eq(next
, now
))
2861 return jiffies_to_nsecs(next
- now
);
2865 notrace
unsigned long get_parent_ip(unsigned long addr
)
2867 if (in_lock_functions(addr
)) {
2868 addr
= CALLER_ADDR2
;
2869 if (in_lock_functions(addr
))
2870 addr
= CALLER_ADDR3
;
2875 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2876 defined(CONFIG_PREEMPT_TRACER))
2878 void preempt_count_add(int val
)
2880 #ifdef CONFIG_DEBUG_PREEMPT
2884 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2887 __preempt_count_add(val
);
2888 #ifdef CONFIG_DEBUG_PREEMPT
2890 * Spinlock count overflowing soon?
2892 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2895 if (preempt_count() == val
) {
2896 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2897 #ifdef CONFIG_DEBUG_PREEMPT
2898 current
->preempt_disable_ip
= ip
;
2900 trace_preempt_off(CALLER_ADDR0
, ip
);
2903 EXPORT_SYMBOL(preempt_count_add
);
2904 NOKPROBE_SYMBOL(preempt_count_add
);
2906 void preempt_count_sub(int val
)
2908 #ifdef CONFIG_DEBUG_PREEMPT
2912 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2915 * Is the spinlock portion underflowing?
2917 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2918 !(preempt_count() & PREEMPT_MASK
)))
2922 if (preempt_count() == val
)
2923 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2924 __preempt_count_sub(val
);
2926 EXPORT_SYMBOL(preempt_count_sub
);
2927 NOKPROBE_SYMBOL(preempt_count_sub
);
2932 * Print scheduling while atomic bug:
2934 static noinline
void __schedule_bug(struct task_struct
*prev
)
2936 if (oops_in_progress
)
2939 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2940 prev
->comm
, prev
->pid
, preempt_count());
2942 debug_show_held_locks(prev
);
2944 if (irqs_disabled())
2945 print_irqtrace_events(prev
);
2946 #ifdef CONFIG_DEBUG_PREEMPT
2947 if (in_atomic_preempt_off()) {
2948 pr_err("Preemption disabled at:");
2949 print_ip_sym(current
->preempt_disable_ip
);
2954 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2958 * Various schedule()-time debugging checks and statistics:
2960 static inline void schedule_debug(struct task_struct
*prev
)
2962 #ifdef CONFIG_SCHED_STACK_END_CHECK
2963 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2966 * Test if we are atomic. Since do_exit() needs to call into
2967 * schedule() atomically, we ignore that path. Otherwise whine
2968 * if we are scheduling when we should not.
2970 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2971 __schedule_bug(prev
);
2974 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2976 schedstat_inc(this_rq(), sched_count
);
2980 * Pick up the highest-prio task:
2982 static inline struct task_struct
*
2983 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2985 const struct sched_class
*class = &fair_sched_class
;
2986 struct task_struct
*p
;
2989 * Optimization: we know that if all tasks are in
2990 * the fair class we can call that function directly:
2992 if (likely(prev
->sched_class
== class &&
2993 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2994 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2995 if (unlikely(p
== RETRY_TASK
))
2998 /* assumes fair_sched_class->next == idle_sched_class */
3000 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3006 for_each_class(class) {
3007 p
= class->pick_next_task(rq
, prev
);
3009 if (unlikely(p
== RETRY_TASK
))
3015 BUG(); /* the idle class will always have a runnable task */
3019 * __schedule() is the main scheduler function.
3021 * The main means of driving the scheduler and thus entering this function are:
3023 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3025 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3026 * paths. For example, see arch/x86/entry_64.S.
3028 * To drive preemption between tasks, the scheduler sets the flag in timer
3029 * interrupt handler scheduler_tick().
3031 * 3. Wakeups don't really cause entry into schedule(). They add a
3032 * task to the run-queue and that's it.
3034 * Now, if the new task added to the run-queue preempts the current
3035 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3036 * called on the nearest possible occasion:
3038 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3040 * - in syscall or exception context, at the next outmost
3041 * preempt_enable(). (this might be as soon as the wake_up()'s
3044 * - in IRQ context, return from interrupt-handler to
3045 * preemptible context
3047 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3050 * - cond_resched() call
3051 * - explicit schedule() call
3052 * - return from syscall or exception to user-space
3053 * - return from interrupt-handler to user-space
3055 * WARNING: must be called with preemption disabled!
3057 static void __sched
__schedule(void)
3059 struct task_struct
*prev
, *next
;
3060 unsigned long *switch_count
;
3064 cpu
= smp_processor_id();
3066 rcu_note_context_switch();
3069 schedule_debug(prev
);
3071 if (sched_feat(HRTICK
))
3075 * Make sure that signal_pending_state()->signal_pending() below
3076 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3077 * done by the caller to avoid the race with signal_wake_up().
3079 smp_mb__before_spinlock();
3080 raw_spin_lock_irq(&rq
->lock
);
3081 lockdep_pin_lock(&rq
->lock
);
3083 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3085 switch_count
= &prev
->nivcsw
;
3086 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3087 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3088 prev
->state
= TASK_RUNNING
;
3090 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3094 * If a worker went to sleep, notify and ask workqueue
3095 * whether it wants to wake up a task to maintain
3098 if (prev
->flags
& PF_WQ_WORKER
) {
3099 struct task_struct
*to_wakeup
;
3101 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3103 try_to_wake_up_local(to_wakeup
);
3106 switch_count
= &prev
->nvcsw
;
3109 if (task_on_rq_queued(prev
))
3110 update_rq_clock(rq
);
3112 next
= pick_next_task(rq
, prev
);
3113 clear_tsk_need_resched(prev
);
3114 clear_preempt_need_resched();
3115 rq
->clock_skip_update
= 0;
3117 if (likely(prev
!= next
)) {
3122 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3125 lockdep_unpin_lock(&rq
->lock
);
3126 raw_spin_unlock_irq(&rq
->lock
);
3129 balance_callback(rq
);
3132 static inline void sched_submit_work(struct task_struct
*tsk
)
3134 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3137 * If we are going to sleep and we have plugged IO queued,
3138 * make sure to submit it to avoid deadlocks.
3140 if (blk_needs_flush_plug(tsk
))
3141 blk_schedule_flush_plug(tsk
);
3144 asmlinkage __visible
void __sched
schedule(void)
3146 struct task_struct
*tsk
= current
;
3148 sched_submit_work(tsk
);
3152 sched_preempt_enable_no_resched();
3153 } while (need_resched());
3155 EXPORT_SYMBOL(schedule
);
3157 #ifdef CONFIG_CONTEXT_TRACKING
3158 asmlinkage __visible
void __sched
schedule_user(void)
3161 * If we come here after a random call to set_need_resched(),
3162 * or we have been woken up remotely but the IPI has not yet arrived,
3163 * we haven't yet exited the RCU idle mode. Do it here manually until
3164 * we find a better solution.
3166 * NB: There are buggy callers of this function. Ideally we
3167 * should warn if prev_state != CONTEXT_USER, but that will trigger
3168 * too frequently to make sense yet.
3170 enum ctx_state prev_state
= exception_enter();
3172 exception_exit(prev_state
);
3177 * schedule_preempt_disabled - called with preemption disabled
3179 * Returns with preemption disabled. Note: preempt_count must be 1
3181 void __sched
schedule_preempt_disabled(void)
3183 sched_preempt_enable_no_resched();
3188 static void __sched notrace
preempt_schedule_common(void)
3191 preempt_active_enter();
3193 preempt_active_exit();
3196 * Check again in case we missed a preemption opportunity
3197 * between schedule and now.
3199 } while (need_resched());
3202 #ifdef CONFIG_PREEMPT
3204 * this is the entry point to schedule() from in-kernel preemption
3205 * off of preempt_enable. Kernel preemptions off return from interrupt
3206 * occur there and call schedule directly.
3208 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3211 * If there is a non-zero preempt_count or interrupts are disabled,
3212 * we do not want to preempt the current task. Just return..
3214 if (likely(!preemptible()))
3217 preempt_schedule_common();
3219 NOKPROBE_SYMBOL(preempt_schedule
);
3220 EXPORT_SYMBOL(preempt_schedule
);
3223 * preempt_schedule_notrace - preempt_schedule called by tracing
3225 * The tracing infrastructure uses preempt_enable_notrace to prevent
3226 * recursion and tracing preempt enabling caused by the tracing
3227 * infrastructure itself. But as tracing can happen in areas coming
3228 * from userspace or just about to enter userspace, a preempt enable
3229 * can occur before user_exit() is called. This will cause the scheduler
3230 * to be called when the system is still in usermode.
3232 * To prevent this, the preempt_enable_notrace will use this function
3233 * instead of preempt_schedule() to exit user context if needed before
3234 * calling the scheduler.
3236 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3238 enum ctx_state prev_ctx
;
3240 if (likely(!preemptible()))
3245 * Use raw __prempt_count() ops that don't call function.
3246 * We can't call functions before disabling preemption which
3247 * disarm preemption tracing recursions.
3249 __preempt_count_add(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3252 * Needs preempt disabled in case user_exit() is traced
3253 * and the tracer calls preempt_enable_notrace() causing
3254 * an infinite recursion.
3256 prev_ctx
= exception_enter();
3258 exception_exit(prev_ctx
);
3261 __preempt_count_sub(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3262 } while (need_resched());
3264 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3266 #endif /* CONFIG_PREEMPT */
3269 * this is the entry point to schedule() from kernel preemption
3270 * off of irq context.
3271 * Note, that this is called and return with irqs disabled. This will
3272 * protect us against recursive calling from irq.
3274 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3276 enum ctx_state prev_state
;
3278 /* Catch callers which need to be fixed */
3279 BUG_ON(preempt_count() || !irqs_disabled());
3281 prev_state
= exception_enter();
3284 preempt_active_enter();
3287 local_irq_disable();
3288 preempt_active_exit();
3289 } while (need_resched());
3291 exception_exit(prev_state
);
3294 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3297 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3299 EXPORT_SYMBOL(default_wake_function
);
3301 #ifdef CONFIG_RT_MUTEXES
3304 * rt_mutex_setprio - set the current priority of a task
3306 * @prio: prio value (kernel-internal form)
3308 * This function changes the 'effective' priority of a task. It does
3309 * not touch ->normal_prio like __setscheduler().
3311 * Used by the rt_mutex code to implement priority inheritance
3312 * logic. Call site only calls if the priority of the task changed.
3314 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3316 int oldprio
, queued
, running
, enqueue_flag
= 0;
3318 const struct sched_class
*prev_class
;
3320 BUG_ON(prio
> MAX_PRIO
);
3322 rq
= __task_rq_lock(p
);
3325 * Idle task boosting is a nono in general. There is one
3326 * exception, when PREEMPT_RT and NOHZ is active:
3328 * The idle task calls get_next_timer_interrupt() and holds
3329 * the timer wheel base->lock on the CPU and another CPU wants
3330 * to access the timer (probably to cancel it). We can safely
3331 * ignore the boosting request, as the idle CPU runs this code
3332 * with interrupts disabled and will complete the lock
3333 * protected section without being interrupted. So there is no
3334 * real need to boost.
3336 if (unlikely(p
== rq
->idle
)) {
3337 WARN_ON(p
!= rq
->curr
);
3338 WARN_ON(p
->pi_blocked_on
);
3342 trace_sched_pi_setprio(p
, prio
);
3344 prev_class
= p
->sched_class
;
3345 queued
= task_on_rq_queued(p
);
3346 running
= task_current(rq
, p
);
3348 dequeue_task(rq
, p
, 0);
3350 put_prev_task(rq
, p
);
3353 * Boosting condition are:
3354 * 1. -rt task is running and holds mutex A
3355 * --> -dl task blocks on mutex A
3357 * 2. -dl task is running and holds mutex A
3358 * --> -dl task blocks on mutex A and could preempt the
3361 if (dl_prio(prio
)) {
3362 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3363 if (!dl_prio(p
->normal_prio
) ||
3364 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3365 p
->dl
.dl_boosted
= 1;
3366 enqueue_flag
= ENQUEUE_REPLENISH
;
3368 p
->dl
.dl_boosted
= 0;
3369 p
->sched_class
= &dl_sched_class
;
3370 } else if (rt_prio(prio
)) {
3371 if (dl_prio(oldprio
))
3372 p
->dl
.dl_boosted
= 0;
3374 enqueue_flag
= ENQUEUE_HEAD
;
3375 p
->sched_class
= &rt_sched_class
;
3377 if (dl_prio(oldprio
))
3378 p
->dl
.dl_boosted
= 0;
3379 if (rt_prio(oldprio
))
3381 p
->sched_class
= &fair_sched_class
;
3387 p
->sched_class
->set_curr_task(rq
);
3389 enqueue_task(rq
, p
, enqueue_flag
);
3391 check_class_changed(rq
, p
, prev_class
, oldprio
);
3393 preempt_disable(); /* avoid rq from going away on us */
3394 __task_rq_unlock(rq
);
3396 balance_callback(rq
);
3401 void set_user_nice(struct task_struct
*p
, long nice
)
3403 int old_prio
, delta
, queued
;
3404 unsigned long flags
;
3407 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3410 * We have to be careful, if called from sys_setpriority(),
3411 * the task might be in the middle of scheduling on another CPU.
3413 rq
= task_rq_lock(p
, &flags
);
3415 * The RT priorities are set via sched_setscheduler(), but we still
3416 * allow the 'normal' nice value to be set - but as expected
3417 * it wont have any effect on scheduling until the task is
3418 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3420 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3421 p
->static_prio
= NICE_TO_PRIO(nice
);
3424 queued
= task_on_rq_queued(p
);
3426 dequeue_task(rq
, p
, 0);
3428 p
->static_prio
= NICE_TO_PRIO(nice
);
3431 p
->prio
= effective_prio(p
);
3432 delta
= p
->prio
- old_prio
;
3435 enqueue_task(rq
, p
, 0);
3437 * If the task increased its priority or is running and
3438 * lowered its priority, then reschedule its CPU:
3440 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3444 task_rq_unlock(rq
, p
, &flags
);
3446 EXPORT_SYMBOL(set_user_nice
);
3449 * can_nice - check if a task can reduce its nice value
3453 int can_nice(const struct task_struct
*p
, const int nice
)
3455 /* convert nice value [19,-20] to rlimit style value [1,40] */
3456 int nice_rlim
= nice_to_rlimit(nice
);
3458 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3459 capable(CAP_SYS_NICE
));
3462 #ifdef __ARCH_WANT_SYS_NICE
3465 * sys_nice - change the priority of the current process.
3466 * @increment: priority increment
3468 * sys_setpriority is a more generic, but much slower function that
3469 * does similar things.
3471 SYSCALL_DEFINE1(nice
, int, increment
)
3476 * Setpriority might change our priority at the same moment.
3477 * We don't have to worry. Conceptually one call occurs first
3478 * and we have a single winner.
3480 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3481 nice
= task_nice(current
) + increment
;
3483 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3484 if (increment
< 0 && !can_nice(current
, nice
))
3487 retval
= security_task_setnice(current
, nice
);
3491 set_user_nice(current
, nice
);
3498 * task_prio - return the priority value of a given task.
3499 * @p: the task in question.
3501 * Return: The priority value as seen by users in /proc.
3502 * RT tasks are offset by -200. Normal tasks are centered
3503 * around 0, value goes from -16 to +15.
3505 int task_prio(const struct task_struct
*p
)
3507 return p
->prio
- MAX_RT_PRIO
;
3511 * idle_cpu - is a given cpu idle currently?
3512 * @cpu: the processor in question.
3514 * Return: 1 if the CPU is currently idle. 0 otherwise.
3516 int idle_cpu(int cpu
)
3518 struct rq
*rq
= cpu_rq(cpu
);
3520 if (rq
->curr
!= rq
->idle
)
3527 if (!llist_empty(&rq
->wake_list
))
3535 * idle_task - return the idle task for a given cpu.
3536 * @cpu: the processor in question.
3538 * Return: The idle task for the cpu @cpu.
3540 struct task_struct
*idle_task(int cpu
)
3542 return cpu_rq(cpu
)->idle
;
3546 * find_process_by_pid - find a process with a matching PID value.
3547 * @pid: the pid in question.
3549 * The task of @pid, if found. %NULL otherwise.
3551 static struct task_struct
*find_process_by_pid(pid_t pid
)
3553 return pid
? find_task_by_vpid(pid
) : current
;
3557 * This function initializes the sched_dl_entity of a newly becoming
3558 * SCHED_DEADLINE task.
3560 * Only the static values are considered here, the actual runtime and the
3561 * absolute deadline will be properly calculated when the task is enqueued
3562 * for the first time with its new policy.
3565 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3567 struct sched_dl_entity
*dl_se
= &p
->dl
;
3569 dl_se
->dl_runtime
= attr
->sched_runtime
;
3570 dl_se
->dl_deadline
= attr
->sched_deadline
;
3571 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3572 dl_se
->flags
= attr
->sched_flags
;
3573 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3576 * Changing the parameters of a task is 'tricky' and we're not doing
3577 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3579 * What we SHOULD do is delay the bandwidth release until the 0-lag
3580 * point. This would include retaining the task_struct until that time
3581 * and change dl_overflow() to not immediately decrement the current
3584 * Instead we retain the current runtime/deadline and let the new
3585 * parameters take effect after the current reservation period lapses.
3586 * This is safe (albeit pessimistic) because the 0-lag point is always
3587 * before the current scheduling deadline.
3589 * We can still have temporary overloads because we do not delay the
3590 * change in bandwidth until that time; so admission control is
3591 * not on the safe side. It does however guarantee tasks will never
3592 * consume more than promised.
3597 * sched_setparam() passes in -1 for its policy, to let the functions
3598 * it calls know not to change it.
3600 #define SETPARAM_POLICY -1
3602 static void __setscheduler_params(struct task_struct
*p
,
3603 const struct sched_attr
*attr
)
3605 int policy
= attr
->sched_policy
;
3607 if (policy
== SETPARAM_POLICY
)
3612 if (dl_policy(policy
))
3613 __setparam_dl(p
, attr
);
3614 else if (fair_policy(policy
))
3615 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3618 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3619 * !rt_policy. Always setting this ensures that things like
3620 * getparam()/getattr() don't report silly values for !rt tasks.
3622 p
->rt_priority
= attr
->sched_priority
;
3623 p
->normal_prio
= normal_prio(p
);
3627 /* Actually do priority change: must hold pi & rq lock. */
3628 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3629 const struct sched_attr
*attr
, bool keep_boost
)
3631 __setscheduler_params(p
, attr
);
3634 * Keep a potential priority boosting if called from
3635 * sched_setscheduler().
3638 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3640 p
->prio
= normal_prio(p
);
3642 if (dl_prio(p
->prio
))
3643 p
->sched_class
= &dl_sched_class
;
3644 else if (rt_prio(p
->prio
))
3645 p
->sched_class
= &rt_sched_class
;
3647 p
->sched_class
= &fair_sched_class
;
3651 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3653 struct sched_dl_entity
*dl_se
= &p
->dl
;
3655 attr
->sched_priority
= p
->rt_priority
;
3656 attr
->sched_runtime
= dl_se
->dl_runtime
;
3657 attr
->sched_deadline
= dl_se
->dl_deadline
;
3658 attr
->sched_period
= dl_se
->dl_period
;
3659 attr
->sched_flags
= dl_se
->flags
;
3663 * This function validates the new parameters of a -deadline task.
3664 * We ask for the deadline not being zero, and greater or equal
3665 * than the runtime, as well as the period of being zero or
3666 * greater than deadline. Furthermore, we have to be sure that
3667 * user parameters are above the internal resolution of 1us (we
3668 * check sched_runtime only since it is always the smaller one) and
3669 * below 2^63 ns (we have to check both sched_deadline and
3670 * sched_period, as the latter can be zero).
3673 __checkparam_dl(const struct sched_attr
*attr
)
3676 if (attr
->sched_deadline
== 0)
3680 * Since we truncate DL_SCALE bits, make sure we're at least
3683 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3687 * Since we use the MSB for wrap-around and sign issues, make
3688 * sure it's not set (mind that period can be equal to zero).
3690 if (attr
->sched_deadline
& (1ULL << 63) ||
3691 attr
->sched_period
& (1ULL << 63))
3694 /* runtime <= deadline <= period (if period != 0) */
3695 if ((attr
->sched_period
!= 0 &&
3696 attr
->sched_period
< attr
->sched_deadline
) ||
3697 attr
->sched_deadline
< attr
->sched_runtime
)
3704 * check the target process has a UID that matches the current process's
3706 static bool check_same_owner(struct task_struct
*p
)
3708 const struct cred
*cred
= current_cred(), *pcred
;
3712 pcred
= __task_cred(p
);
3713 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3714 uid_eq(cred
->euid
, pcred
->uid
));
3719 static bool dl_param_changed(struct task_struct
*p
,
3720 const struct sched_attr
*attr
)
3722 struct sched_dl_entity
*dl_se
= &p
->dl
;
3724 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3725 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3726 dl_se
->dl_period
!= attr
->sched_period
||
3727 dl_se
->flags
!= attr
->sched_flags
)
3733 static int __sched_setscheduler(struct task_struct
*p
,
3734 const struct sched_attr
*attr
,
3737 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3738 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3739 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3740 int new_effective_prio
, policy
= attr
->sched_policy
;
3741 unsigned long flags
;
3742 const struct sched_class
*prev_class
;
3746 /* may grab non-irq protected spin_locks */
3747 BUG_ON(in_interrupt());
3749 /* double check policy once rq lock held */
3751 reset_on_fork
= p
->sched_reset_on_fork
;
3752 policy
= oldpolicy
= p
->policy
;
3754 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3756 if (policy
!= SCHED_DEADLINE
&&
3757 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3758 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3759 policy
!= SCHED_IDLE
)
3763 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3767 * Valid priorities for SCHED_FIFO and SCHED_RR are
3768 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3769 * SCHED_BATCH and SCHED_IDLE is 0.
3771 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3772 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3774 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3775 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3779 * Allow unprivileged RT tasks to decrease priority:
3781 if (user
&& !capable(CAP_SYS_NICE
)) {
3782 if (fair_policy(policy
)) {
3783 if (attr
->sched_nice
< task_nice(p
) &&
3784 !can_nice(p
, attr
->sched_nice
))
3788 if (rt_policy(policy
)) {
3789 unsigned long rlim_rtprio
=
3790 task_rlimit(p
, RLIMIT_RTPRIO
);
3792 /* can't set/change the rt policy */
3793 if (policy
!= p
->policy
&& !rlim_rtprio
)
3796 /* can't increase priority */
3797 if (attr
->sched_priority
> p
->rt_priority
&&
3798 attr
->sched_priority
> rlim_rtprio
)
3803 * Can't set/change SCHED_DEADLINE policy at all for now
3804 * (safest behavior); in the future we would like to allow
3805 * unprivileged DL tasks to increase their relative deadline
3806 * or reduce their runtime (both ways reducing utilization)
3808 if (dl_policy(policy
))
3812 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3813 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3815 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3816 if (!can_nice(p
, task_nice(p
)))
3820 /* can't change other user's priorities */
3821 if (!check_same_owner(p
))
3824 /* Normal users shall not reset the sched_reset_on_fork flag */
3825 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3830 retval
= security_task_setscheduler(p
);
3836 * make sure no PI-waiters arrive (or leave) while we are
3837 * changing the priority of the task:
3839 * To be able to change p->policy safely, the appropriate
3840 * runqueue lock must be held.
3842 rq
= task_rq_lock(p
, &flags
);
3845 * Changing the policy of the stop threads its a very bad idea
3847 if (p
== rq
->stop
) {
3848 task_rq_unlock(rq
, p
, &flags
);
3853 * If not changing anything there's no need to proceed further,
3854 * but store a possible modification of reset_on_fork.
3856 if (unlikely(policy
== p
->policy
)) {
3857 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3859 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3861 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3864 p
->sched_reset_on_fork
= reset_on_fork
;
3865 task_rq_unlock(rq
, p
, &flags
);
3871 #ifdef CONFIG_RT_GROUP_SCHED
3873 * Do not allow realtime tasks into groups that have no runtime
3876 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3877 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3878 !task_group_is_autogroup(task_group(p
))) {
3879 task_rq_unlock(rq
, p
, &flags
);
3884 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3885 cpumask_t
*span
= rq
->rd
->span
;
3888 * Don't allow tasks with an affinity mask smaller than
3889 * the entire root_domain to become SCHED_DEADLINE. We
3890 * will also fail if there's no bandwidth available.
3892 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3893 rq
->rd
->dl_bw
.bw
== 0) {
3894 task_rq_unlock(rq
, p
, &flags
);
3901 /* recheck policy now with rq lock held */
3902 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3903 policy
= oldpolicy
= -1;
3904 task_rq_unlock(rq
, p
, &flags
);
3909 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3910 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3913 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3914 task_rq_unlock(rq
, p
, &flags
);
3918 p
->sched_reset_on_fork
= reset_on_fork
;
3923 * Take priority boosted tasks into account. If the new
3924 * effective priority is unchanged, we just store the new
3925 * normal parameters and do not touch the scheduler class and
3926 * the runqueue. This will be done when the task deboost
3929 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
3930 if (new_effective_prio
== oldprio
) {
3931 __setscheduler_params(p
, attr
);
3932 task_rq_unlock(rq
, p
, &flags
);
3937 queued
= task_on_rq_queued(p
);
3938 running
= task_current(rq
, p
);
3940 dequeue_task(rq
, p
, 0);
3942 put_prev_task(rq
, p
);
3944 prev_class
= p
->sched_class
;
3945 __setscheduler(rq
, p
, attr
, pi
);
3948 p
->sched_class
->set_curr_task(rq
);
3951 * We enqueue to tail when the priority of a task is
3952 * increased (user space view).
3954 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3957 check_class_changed(rq
, p
, prev_class
, oldprio
);
3958 preempt_disable(); /* avoid rq from going away on us */
3959 task_rq_unlock(rq
, p
, &flags
);
3962 rt_mutex_adjust_pi(p
);
3965 * Run balance callbacks after we've adjusted the PI chain.
3967 balance_callback(rq
);
3973 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3974 const struct sched_param
*param
, bool check
)
3976 struct sched_attr attr
= {
3977 .sched_policy
= policy
,
3978 .sched_priority
= param
->sched_priority
,
3979 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3982 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3983 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3984 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3985 policy
&= ~SCHED_RESET_ON_FORK
;
3986 attr
.sched_policy
= policy
;
3989 return __sched_setscheduler(p
, &attr
, check
, true);
3992 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3993 * @p: the task in question.
3994 * @policy: new policy.
3995 * @param: structure containing the new RT priority.
3997 * Return: 0 on success. An error code otherwise.
3999 * NOTE that the task may be already dead.
4001 int sched_setscheduler(struct task_struct
*p
, int policy
,
4002 const struct sched_param
*param
)
4004 return _sched_setscheduler(p
, policy
, param
, true);
4006 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4008 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4010 return __sched_setscheduler(p
, attr
, true, true);
4012 EXPORT_SYMBOL_GPL(sched_setattr
);
4015 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4016 * @p: the task in question.
4017 * @policy: new policy.
4018 * @param: structure containing the new RT priority.
4020 * Just like sched_setscheduler, only don't bother checking if the
4021 * current context has permission. For example, this is needed in
4022 * stop_machine(): we create temporary high priority worker threads,
4023 * but our caller might not have that capability.
4025 * Return: 0 on success. An error code otherwise.
4027 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4028 const struct sched_param
*param
)
4030 return _sched_setscheduler(p
, policy
, param
, false);
4034 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4036 struct sched_param lparam
;
4037 struct task_struct
*p
;
4040 if (!param
|| pid
< 0)
4042 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4047 p
= find_process_by_pid(pid
);
4049 retval
= sched_setscheduler(p
, policy
, &lparam
);
4056 * Mimics kernel/events/core.c perf_copy_attr().
4058 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4059 struct sched_attr
*attr
)
4064 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4068 * zero the full structure, so that a short copy will be nice.
4070 memset(attr
, 0, sizeof(*attr
));
4072 ret
= get_user(size
, &uattr
->size
);
4076 if (size
> PAGE_SIZE
) /* silly large */
4079 if (!size
) /* abi compat */
4080 size
= SCHED_ATTR_SIZE_VER0
;
4082 if (size
< SCHED_ATTR_SIZE_VER0
)
4086 * If we're handed a bigger struct than we know of,
4087 * ensure all the unknown bits are 0 - i.e. new
4088 * user-space does not rely on any kernel feature
4089 * extensions we dont know about yet.
4091 if (size
> sizeof(*attr
)) {
4092 unsigned char __user
*addr
;
4093 unsigned char __user
*end
;
4096 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4097 end
= (void __user
*)uattr
+ size
;
4099 for (; addr
< end
; addr
++) {
4100 ret
= get_user(val
, addr
);
4106 size
= sizeof(*attr
);
4109 ret
= copy_from_user(attr
, uattr
, size
);
4114 * XXX: do we want to be lenient like existing syscalls; or do we want
4115 * to be strict and return an error on out-of-bounds values?
4117 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4122 put_user(sizeof(*attr
), &uattr
->size
);
4127 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4128 * @pid: the pid in question.
4129 * @policy: new policy.
4130 * @param: structure containing the new RT priority.
4132 * Return: 0 on success. An error code otherwise.
4134 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4135 struct sched_param __user
*, param
)
4137 /* negative values for policy are not valid */
4141 return do_sched_setscheduler(pid
, policy
, param
);
4145 * sys_sched_setparam - set/change the RT priority of a thread
4146 * @pid: the pid in question.
4147 * @param: structure containing the new RT priority.
4149 * Return: 0 on success. An error code otherwise.
4151 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4153 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4157 * sys_sched_setattr - same as above, but with extended sched_attr
4158 * @pid: the pid in question.
4159 * @uattr: structure containing the extended parameters.
4160 * @flags: for future extension.
4162 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4163 unsigned int, flags
)
4165 struct sched_attr attr
;
4166 struct task_struct
*p
;
4169 if (!uattr
|| pid
< 0 || flags
)
4172 retval
= sched_copy_attr(uattr
, &attr
);
4176 if ((int)attr
.sched_policy
< 0)
4181 p
= find_process_by_pid(pid
);
4183 retval
= sched_setattr(p
, &attr
);
4190 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4191 * @pid: the pid in question.
4193 * Return: On success, the policy of the thread. Otherwise, a negative error
4196 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4198 struct task_struct
*p
;
4206 p
= find_process_by_pid(pid
);
4208 retval
= security_task_getscheduler(p
);
4211 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4218 * sys_sched_getparam - get the RT priority of a thread
4219 * @pid: the pid in question.
4220 * @param: structure containing the RT priority.
4222 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4225 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4227 struct sched_param lp
= { .sched_priority
= 0 };
4228 struct task_struct
*p
;
4231 if (!param
|| pid
< 0)
4235 p
= find_process_by_pid(pid
);
4240 retval
= security_task_getscheduler(p
);
4244 if (task_has_rt_policy(p
))
4245 lp
.sched_priority
= p
->rt_priority
;
4249 * This one might sleep, we cannot do it with a spinlock held ...
4251 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4260 static int sched_read_attr(struct sched_attr __user
*uattr
,
4261 struct sched_attr
*attr
,
4266 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4270 * If we're handed a smaller struct than we know of,
4271 * ensure all the unknown bits are 0 - i.e. old
4272 * user-space does not get uncomplete information.
4274 if (usize
< sizeof(*attr
)) {
4275 unsigned char *addr
;
4278 addr
= (void *)attr
+ usize
;
4279 end
= (void *)attr
+ sizeof(*attr
);
4281 for (; addr
< end
; addr
++) {
4289 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4297 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4298 * @pid: the pid in question.
4299 * @uattr: structure containing the extended parameters.
4300 * @size: sizeof(attr) for fwd/bwd comp.
4301 * @flags: for future extension.
4303 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4304 unsigned int, size
, unsigned int, flags
)
4306 struct sched_attr attr
= {
4307 .size
= sizeof(struct sched_attr
),
4309 struct task_struct
*p
;
4312 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4313 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4317 p
= find_process_by_pid(pid
);
4322 retval
= security_task_getscheduler(p
);
4326 attr
.sched_policy
= p
->policy
;
4327 if (p
->sched_reset_on_fork
)
4328 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4329 if (task_has_dl_policy(p
))
4330 __getparam_dl(p
, &attr
);
4331 else if (task_has_rt_policy(p
))
4332 attr
.sched_priority
= p
->rt_priority
;
4334 attr
.sched_nice
= task_nice(p
);
4338 retval
= sched_read_attr(uattr
, &attr
, size
);
4346 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4348 cpumask_var_t cpus_allowed
, new_mask
;
4349 struct task_struct
*p
;
4354 p
= find_process_by_pid(pid
);
4360 /* Prevent p going away */
4364 if (p
->flags
& PF_NO_SETAFFINITY
) {
4368 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4372 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4374 goto out_free_cpus_allowed
;
4377 if (!check_same_owner(p
)) {
4379 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4381 goto out_free_new_mask
;
4386 retval
= security_task_setscheduler(p
);
4388 goto out_free_new_mask
;
4391 cpuset_cpus_allowed(p
, cpus_allowed
);
4392 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4395 * Since bandwidth control happens on root_domain basis,
4396 * if admission test is enabled, we only admit -deadline
4397 * tasks allowed to run on all the CPUs in the task's
4401 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4403 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4406 goto out_free_new_mask
;
4412 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4415 cpuset_cpus_allowed(p
, cpus_allowed
);
4416 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4418 * We must have raced with a concurrent cpuset
4419 * update. Just reset the cpus_allowed to the
4420 * cpuset's cpus_allowed
4422 cpumask_copy(new_mask
, cpus_allowed
);
4427 free_cpumask_var(new_mask
);
4428 out_free_cpus_allowed
:
4429 free_cpumask_var(cpus_allowed
);
4435 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4436 struct cpumask
*new_mask
)
4438 if (len
< cpumask_size())
4439 cpumask_clear(new_mask
);
4440 else if (len
> cpumask_size())
4441 len
= cpumask_size();
4443 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4447 * sys_sched_setaffinity - set the cpu affinity of a process
4448 * @pid: pid of the process
4449 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4450 * @user_mask_ptr: user-space pointer to the new cpu mask
4452 * Return: 0 on success. An error code otherwise.
4454 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4455 unsigned long __user
*, user_mask_ptr
)
4457 cpumask_var_t new_mask
;
4460 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4463 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4465 retval
= sched_setaffinity(pid
, new_mask
);
4466 free_cpumask_var(new_mask
);
4470 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4472 struct task_struct
*p
;
4473 unsigned long flags
;
4479 p
= find_process_by_pid(pid
);
4483 retval
= security_task_getscheduler(p
);
4487 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4488 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4489 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4498 * sys_sched_getaffinity - get the cpu affinity of a process
4499 * @pid: pid of the process
4500 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4501 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4503 * Return: 0 on success. An error code otherwise.
4505 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4506 unsigned long __user
*, user_mask_ptr
)
4511 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4513 if (len
& (sizeof(unsigned long)-1))
4516 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4519 ret
= sched_getaffinity(pid
, mask
);
4521 size_t retlen
= min_t(size_t, len
, cpumask_size());
4523 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4528 free_cpumask_var(mask
);
4534 * sys_sched_yield - yield the current processor to other threads.
4536 * This function yields the current CPU to other tasks. If there are no
4537 * other threads running on this CPU then this function will return.
4541 SYSCALL_DEFINE0(sched_yield
)
4543 struct rq
*rq
= this_rq_lock();
4545 schedstat_inc(rq
, yld_count
);
4546 current
->sched_class
->yield_task(rq
);
4549 * Since we are going to call schedule() anyway, there's
4550 * no need to preempt or enable interrupts:
4552 __release(rq
->lock
);
4553 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4554 do_raw_spin_unlock(&rq
->lock
);
4555 sched_preempt_enable_no_resched();
4562 int __sched
_cond_resched(void)
4564 if (should_resched(0)) {
4565 preempt_schedule_common();
4570 EXPORT_SYMBOL(_cond_resched
);
4573 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4574 * call schedule, and on return reacquire the lock.
4576 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4577 * operations here to prevent schedule() from being called twice (once via
4578 * spin_unlock(), once by hand).
4580 int __cond_resched_lock(spinlock_t
*lock
)
4582 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4585 lockdep_assert_held(lock
);
4587 if (spin_needbreak(lock
) || resched
) {
4590 preempt_schedule_common();
4598 EXPORT_SYMBOL(__cond_resched_lock
);
4600 int __sched
__cond_resched_softirq(void)
4602 BUG_ON(!in_softirq());
4604 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4606 preempt_schedule_common();
4612 EXPORT_SYMBOL(__cond_resched_softirq
);
4615 * yield - yield the current processor to other threads.
4617 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4619 * The scheduler is at all times free to pick the calling task as the most
4620 * eligible task to run, if removing the yield() call from your code breaks
4621 * it, its already broken.
4623 * Typical broken usage is:
4628 * where one assumes that yield() will let 'the other' process run that will
4629 * make event true. If the current task is a SCHED_FIFO task that will never
4630 * happen. Never use yield() as a progress guarantee!!
4632 * If you want to use yield() to wait for something, use wait_event().
4633 * If you want to use yield() to be 'nice' for others, use cond_resched().
4634 * If you still want to use yield(), do not!
4636 void __sched
yield(void)
4638 set_current_state(TASK_RUNNING
);
4641 EXPORT_SYMBOL(yield
);
4644 * yield_to - yield the current processor to another thread in
4645 * your thread group, or accelerate that thread toward the
4646 * processor it's on.
4648 * @preempt: whether task preemption is allowed or not
4650 * It's the caller's job to ensure that the target task struct
4651 * can't go away on us before we can do any checks.
4654 * true (>0) if we indeed boosted the target task.
4655 * false (0) if we failed to boost the target.
4656 * -ESRCH if there's no task to yield to.
4658 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4660 struct task_struct
*curr
= current
;
4661 struct rq
*rq
, *p_rq
;
4662 unsigned long flags
;
4665 local_irq_save(flags
);
4671 * If we're the only runnable task on the rq and target rq also
4672 * has only one task, there's absolutely no point in yielding.
4674 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4679 double_rq_lock(rq
, p_rq
);
4680 if (task_rq(p
) != p_rq
) {
4681 double_rq_unlock(rq
, p_rq
);
4685 if (!curr
->sched_class
->yield_to_task
)
4688 if (curr
->sched_class
!= p
->sched_class
)
4691 if (task_running(p_rq
, p
) || p
->state
)
4694 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4696 schedstat_inc(rq
, yld_count
);
4698 * Make p's CPU reschedule; pick_next_entity takes care of
4701 if (preempt
&& rq
!= p_rq
)
4706 double_rq_unlock(rq
, p_rq
);
4708 local_irq_restore(flags
);
4715 EXPORT_SYMBOL_GPL(yield_to
);
4718 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4719 * that process accounting knows that this is a task in IO wait state.
4721 long __sched
io_schedule_timeout(long timeout
)
4723 int old_iowait
= current
->in_iowait
;
4727 current
->in_iowait
= 1;
4728 blk_schedule_flush_plug(current
);
4730 delayacct_blkio_start();
4732 atomic_inc(&rq
->nr_iowait
);
4733 ret
= schedule_timeout(timeout
);
4734 current
->in_iowait
= old_iowait
;
4735 atomic_dec(&rq
->nr_iowait
);
4736 delayacct_blkio_end();
4740 EXPORT_SYMBOL(io_schedule_timeout
);
4743 * sys_sched_get_priority_max - return maximum RT priority.
4744 * @policy: scheduling class.
4746 * Return: On success, this syscall returns the maximum
4747 * rt_priority that can be used by a given scheduling class.
4748 * On failure, a negative error code is returned.
4750 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4757 ret
= MAX_USER_RT_PRIO
-1;
4759 case SCHED_DEADLINE
:
4770 * sys_sched_get_priority_min - return minimum RT priority.
4771 * @policy: scheduling class.
4773 * Return: On success, this syscall returns the minimum
4774 * rt_priority that can be used by a given scheduling class.
4775 * On failure, a negative error code is returned.
4777 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4786 case SCHED_DEADLINE
:
4796 * sys_sched_rr_get_interval - return the default timeslice of a process.
4797 * @pid: pid of the process.
4798 * @interval: userspace pointer to the timeslice value.
4800 * this syscall writes the default timeslice value of a given process
4801 * into the user-space timespec buffer. A value of '0' means infinity.
4803 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4806 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4807 struct timespec __user
*, interval
)
4809 struct task_struct
*p
;
4810 unsigned int time_slice
;
4811 unsigned long flags
;
4821 p
= find_process_by_pid(pid
);
4825 retval
= security_task_getscheduler(p
);
4829 rq
= task_rq_lock(p
, &flags
);
4831 if (p
->sched_class
->get_rr_interval
)
4832 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4833 task_rq_unlock(rq
, p
, &flags
);
4836 jiffies_to_timespec(time_slice
, &t
);
4837 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4845 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4847 void sched_show_task(struct task_struct
*p
)
4849 unsigned long free
= 0;
4851 unsigned long state
= p
->state
;
4854 state
= __ffs(state
) + 1;
4855 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4856 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4857 #if BITS_PER_LONG == 32
4858 if (state
== TASK_RUNNING
)
4859 printk(KERN_CONT
" running ");
4861 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4863 if (state
== TASK_RUNNING
)
4864 printk(KERN_CONT
" running task ");
4866 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4868 #ifdef CONFIG_DEBUG_STACK_USAGE
4869 free
= stack_not_used(p
);
4874 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4876 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4877 task_pid_nr(p
), ppid
,
4878 (unsigned long)task_thread_info(p
)->flags
);
4880 print_worker_info(KERN_INFO
, p
);
4881 show_stack(p
, NULL
);
4884 void show_state_filter(unsigned long state_filter
)
4886 struct task_struct
*g
, *p
;
4888 #if BITS_PER_LONG == 32
4890 " task PC stack pid father\n");
4893 " task PC stack pid father\n");
4896 for_each_process_thread(g
, p
) {
4898 * reset the NMI-timeout, listing all files on a slow
4899 * console might take a lot of time:
4901 touch_nmi_watchdog();
4902 if (!state_filter
|| (p
->state
& state_filter
))
4906 touch_all_softlockup_watchdogs();
4908 #ifdef CONFIG_SCHED_DEBUG
4909 sysrq_sched_debug_show();
4913 * Only show locks if all tasks are dumped:
4916 debug_show_all_locks();
4919 void init_idle_bootup_task(struct task_struct
*idle
)
4921 idle
->sched_class
= &idle_sched_class
;
4925 * init_idle - set up an idle thread for a given CPU
4926 * @idle: task in question
4927 * @cpu: cpu the idle task belongs to
4929 * NOTE: this function does not set the idle thread's NEED_RESCHED
4930 * flag, to make booting more robust.
4932 void init_idle(struct task_struct
*idle
, int cpu
)
4934 struct rq
*rq
= cpu_rq(cpu
);
4935 unsigned long flags
;
4937 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
4938 raw_spin_lock(&rq
->lock
);
4940 __sched_fork(0, idle
);
4941 idle
->state
= TASK_RUNNING
;
4942 idle
->se
.exec_start
= sched_clock();
4946 * Its possible that init_idle() gets called multiple times on a task,
4947 * in that case do_set_cpus_allowed() will not do the right thing.
4949 * And since this is boot we can forgo the serialization.
4951 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
4954 * We're having a chicken and egg problem, even though we are
4955 * holding rq->lock, the cpu isn't yet set to this cpu so the
4956 * lockdep check in task_group() will fail.
4958 * Similar case to sched_fork(). / Alternatively we could
4959 * use task_rq_lock() here and obtain the other rq->lock.
4964 __set_task_cpu(idle
, cpu
);
4967 rq
->curr
= rq
->idle
= idle
;
4968 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4972 raw_spin_unlock(&rq
->lock
);
4973 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
4975 /* Set the preempt count _outside_ the spinlocks! */
4976 init_idle_preempt_count(idle
, cpu
);
4979 * The idle tasks have their own, simple scheduling class:
4981 idle
->sched_class
= &idle_sched_class
;
4982 ftrace_graph_init_idle_task(idle
, cpu
);
4983 vtime_init_idle(idle
, cpu
);
4985 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4989 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4990 const struct cpumask
*trial
)
4992 int ret
= 1, trial_cpus
;
4993 struct dl_bw
*cur_dl_b
;
4994 unsigned long flags
;
4996 if (!cpumask_weight(cur
))
4999 rcu_read_lock_sched();
5000 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5001 trial_cpus
= cpumask_weight(trial
);
5003 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5004 if (cur_dl_b
->bw
!= -1 &&
5005 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5007 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5008 rcu_read_unlock_sched();
5013 int task_can_attach(struct task_struct
*p
,
5014 const struct cpumask
*cs_cpus_allowed
)
5019 * Kthreads which disallow setaffinity shouldn't be moved
5020 * to a new cpuset; we don't want to change their cpu
5021 * affinity and isolating such threads by their set of
5022 * allowed nodes is unnecessary. Thus, cpusets are not
5023 * applicable for such threads. This prevents checking for
5024 * success of set_cpus_allowed_ptr() on all attached tasks
5025 * before cpus_allowed may be changed.
5027 if (p
->flags
& PF_NO_SETAFFINITY
) {
5033 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5035 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5040 unsigned long flags
;
5042 rcu_read_lock_sched();
5043 dl_b
= dl_bw_of(dest_cpu
);
5044 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5045 cpus
= dl_bw_cpus(dest_cpu
);
5046 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5051 * We reserve space for this task in the destination
5052 * root_domain, as we can't fail after this point.
5053 * We will free resources in the source root_domain
5054 * later on (see set_cpus_allowed_dl()).
5056 __dl_add(dl_b
, p
->dl
.dl_bw
);
5058 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5059 rcu_read_unlock_sched();
5069 #ifdef CONFIG_NUMA_BALANCING
5070 /* Migrate current task p to target_cpu */
5071 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5073 struct migration_arg arg
= { p
, target_cpu
};
5074 int curr_cpu
= task_cpu(p
);
5076 if (curr_cpu
== target_cpu
)
5079 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5082 /* TODO: This is not properly updating schedstats */
5084 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5085 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5089 * Requeue a task on a given node and accurately track the number of NUMA
5090 * tasks on the runqueues
5092 void sched_setnuma(struct task_struct
*p
, int nid
)
5095 unsigned long flags
;
5096 bool queued
, running
;
5098 rq
= task_rq_lock(p
, &flags
);
5099 queued
= task_on_rq_queued(p
);
5100 running
= task_current(rq
, p
);
5103 dequeue_task(rq
, p
, 0);
5105 put_prev_task(rq
, p
);
5107 p
->numa_preferred_nid
= nid
;
5110 p
->sched_class
->set_curr_task(rq
);
5112 enqueue_task(rq
, p
, 0);
5113 task_rq_unlock(rq
, p
, &flags
);
5115 #endif /* CONFIG_NUMA_BALANCING */
5117 #ifdef CONFIG_HOTPLUG_CPU
5119 * Ensures that the idle task is using init_mm right before its cpu goes
5122 void idle_task_exit(void)
5124 struct mm_struct
*mm
= current
->active_mm
;
5126 BUG_ON(cpu_online(smp_processor_id()));
5128 if (mm
!= &init_mm
) {
5129 switch_mm(mm
, &init_mm
, current
);
5130 finish_arch_post_lock_switch();
5136 * Since this CPU is going 'away' for a while, fold any nr_active delta
5137 * we might have. Assumes we're called after migrate_tasks() so that the
5138 * nr_active count is stable.
5140 * Also see the comment "Global load-average calculations".
5142 static void calc_load_migrate(struct rq
*rq
)
5144 long delta
= calc_load_fold_active(rq
);
5146 atomic_long_add(delta
, &calc_load_tasks
);
5149 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5153 static const struct sched_class fake_sched_class
= {
5154 .put_prev_task
= put_prev_task_fake
,
5157 static struct task_struct fake_task
= {
5159 * Avoid pull_{rt,dl}_task()
5161 .prio
= MAX_PRIO
+ 1,
5162 .sched_class
= &fake_sched_class
,
5166 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5167 * try_to_wake_up()->select_task_rq().
5169 * Called with rq->lock held even though we'er in stop_machine() and
5170 * there's no concurrency possible, we hold the required locks anyway
5171 * because of lock validation efforts.
5173 static void migrate_tasks(struct rq
*dead_rq
)
5175 struct rq
*rq
= dead_rq
;
5176 struct task_struct
*next
, *stop
= rq
->stop
;
5180 * Fudge the rq selection such that the below task selection loop
5181 * doesn't get stuck on the currently eligible stop task.
5183 * We're currently inside stop_machine() and the rq is either stuck
5184 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5185 * either way we should never end up calling schedule() until we're
5191 * put_prev_task() and pick_next_task() sched
5192 * class method both need to have an up-to-date
5193 * value of rq->clock[_task]
5195 update_rq_clock(rq
);
5199 * There's this thread running, bail when that's the only
5202 if (rq
->nr_running
== 1)
5206 * pick_next_task assumes pinned rq->lock.
5208 lockdep_pin_lock(&rq
->lock
);
5209 next
= pick_next_task(rq
, &fake_task
);
5211 next
->sched_class
->put_prev_task(rq
, next
);
5214 * Rules for changing task_struct::cpus_allowed are holding
5215 * both pi_lock and rq->lock, such that holding either
5216 * stabilizes the mask.
5218 * Drop rq->lock is not quite as disastrous as it usually is
5219 * because !cpu_active at this point, which means load-balance
5220 * will not interfere. Also, stop-machine.
5222 lockdep_unpin_lock(&rq
->lock
);
5223 raw_spin_unlock(&rq
->lock
);
5224 raw_spin_lock(&next
->pi_lock
);
5225 raw_spin_lock(&rq
->lock
);
5228 * Since we're inside stop-machine, _nothing_ should have
5229 * changed the task, WARN if weird stuff happened, because in
5230 * that case the above rq->lock drop is a fail too.
5232 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5233 raw_spin_unlock(&next
->pi_lock
);
5237 /* Find suitable destination for @next, with force if needed. */
5238 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5240 rq
= __migrate_task(rq
, next
, dest_cpu
);
5241 if (rq
!= dead_rq
) {
5242 raw_spin_unlock(&rq
->lock
);
5244 raw_spin_lock(&rq
->lock
);
5246 raw_spin_unlock(&next
->pi_lock
);
5251 #endif /* CONFIG_HOTPLUG_CPU */
5253 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5255 static struct ctl_table sd_ctl_dir
[] = {
5257 .procname
= "sched_domain",
5263 static struct ctl_table sd_ctl_root
[] = {
5265 .procname
= "kernel",
5267 .child
= sd_ctl_dir
,
5272 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5274 struct ctl_table
*entry
=
5275 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5280 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5282 struct ctl_table
*entry
;
5285 * In the intermediate directories, both the child directory and
5286 * procname are dynamically allocated and could fail but the mode
5287 * will always be set. In the lowest directory the names are
5288 * static strings and all have proc handlers.
5290 for (entry
= *tablep
; entry
->mode
; entry
++) {
5292 sd_free_ctl_entry(&entry
->child
);
5293 if (entry
->proc_handler
== NULL
)
5294 kfree(entry
->procname
);
5301 static int min_load_idx
= 0;
5302 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5305 set_table_entry(struct ctl_table
*entry
,
5306 const char *procname
, void *data
, int maxlen
,
5307 umode_t mode
, proc_handler
*proc_handler
,
5310 entry
->procname
= procname
;
5312 entry
->maxlen
= maxlen
;
5314 entry
->proc_handler
= proc_handler
;
5317 entry
->extra1
= &min_load_idx
;
5318 entry
->extra2
= &max_load_idx
;
5322 static struct ctl_table
*
5323 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5325 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5330 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5331 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5332 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5333 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5334 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5335 sizeof(int), 0644, proc_dointvec_minmax
, true);
5336 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5337 sizeof(int), 0644, proc_dointvec_minmax
, true);
5338 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5339 sizeof(int), 0644, proc_dointvec_minmax
, true);
5340 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5341 sizeof(int), 0644, proc_dointvec_minmax
, true);
5342 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5343 sizeof(int), 0644, proc_dointvec_minmax
, true);
5344 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5345 sizeof(int), 0644, proc_dointvec_minmax
, false);
5346 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5347 sizeof(int), 0644, proc_dointvec_minmax
, false);
5348 set_table_entry(&table
[9], "cache_nice_tries",
5349 &sd
->cache_nice_tries
,
5350 sizeof(int), 0644, proc_dointvec_minmax
, false);
5351 set_table_entry(&table
[10], "flags", &sd
->flags
,
5352 sizeof(int), 0644, proc_dointvec_minmax
, false);
5353 set_table_entry(&table
[11], "max_newidle_lb_cost",
5354 &sd
->max_newidle_lb_cost
,
5355 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5356 set_table_entry(&table
[12], "name", sd
->name
,
5357 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5358 /* &table[13] is terminator */
5363 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5365 struct ctl_table
*entry
, *table
;
5366 struct sched_domain
*sd
;
5367 int domain_num
= 0, i
;
5370 for_each_domain(cpu
, sd
)
5372 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5377 for_each_domain(cpu
, sd
) {
5378 snprintf(buf
, 32, "domain%d", i
);
5379 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5381 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5388 static struct ctl_table_header
*sd_sysctl_header
;
5389 static void register_sched_domain_sysctl(void)
5391 int i
, cpu_num
= num_possible_cpus();
5392 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5395 WARN_ON(sd_ctl_dir
[0].child
);
5396 sd_ctl_dir
[0].child
= entry
;
5401 for_each_possible_cpu(i
) {
5402 snprintf(buf
, 32, "cpu%d", i
);
5403 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5405 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5409 WARN_ON(sd_sysctl_header
);
5410 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5413 /* may be called multiple times per register */
5414 static void unregister_sched_domain_sysctl(void)
5416 unregister_sysctl_table(sd_sysctl_header
);
5417 sd_sysctl_header
= NULL
;
5418 if (sd_ctl_dir
[0].child
)
5419 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5422 static void register_sched_domain_sysctl(void)
5425 static void unregister_sched_domain_sysctl(void)
5428 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5430 static void set_rq_online(struct rq
*rq
)
5433 const struct sched_class
*class;
5435 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5438 for_each_class(class) {
5439 if (class->rq_online
)
5440 class->rq_online(rq
);
5445 static void set_rq_offline(struct rq
*rq
)
5448 const struct sched_class
*class;
5450 for_each_class(class) {
5451 if (class->rq_offline
)
5452 class->rq_offline(rq
);
5455 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5461 * migration_call - callback that gets triggered when a CPU is added.
5462 * Here we can start up the necessary migration thread for the new CPU.
5465 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5467 int cpu
= (long)hcpu
;
5468 unsigned long flags
;
5469 struct rq
*rq
= cpu_rq(cpu
);
5471 switch (action
& ~CPU_TASKS_FROZEN
) {
5473 case CPU_UP_PREPARE
:
5474 rq
->calc_load_update
= calc_load_update
;
5478 /* Update our root-domain */
5479 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5481 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5485 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5488 #ifdef CONFIG_HOTPLUG_CPU
5490 sched_ttwu_pending();
5491 /* Update our root-domain */
5492 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5494 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5498 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5499 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5503 calc_load_migrate(rq
);
5508 update_max_interval();
5514 * Register at high priority so that task migration (migrate_all_tasks)
5515 * happens before everything else. This has to be lower priority than
5516 * the notifier in the perf_event subsystem, though.
5518 static struct notifier_block migration_notifier
= {
5519 .notifier_call
= migration_call
,
5520 .priority
= CPU_PRI_MIGRATION
,
5523 static void set_cpu_rq_start_time(void)
5525 int cpu
= smp_processor_id();
5526 struct rq
*rq
= cpu_rq(cpu
);
5527 rq
->age_stamp
= sched_clock_cpu(cpu
);
5530 static int sched_cpu_active(struct notifier_block
*nfb
,
5531 unsigned long action
, void *hcpu
)
5533 switch (action
& ~CPU_TASKS_FROZEN
) {
5535 set_cpu_rq_start_time();
5539 * At this point a starting CPU has marked itself as online via
5540 * set_cpu_online(). But it might not yet have marked itself
5541 * as active, which is essential from here on.
5543 * Thus, fall-through and help the starting CPU along.
5545 case CPU_DOWN_FAILED
:
5546 set_cpu_active((long)hcpu
, true);
5553 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5554 unsigned long action
, void *hcpu
)
5556 switch (action
& ~CPU_TASKS_FROZEN
) {
5557 case CPU_DOWN_PREPARE
:
5558 set_cpu_active((long)hcpu
, false);
5565 static int __init
migration_init(void)
5567 void *cpu
= (void *)(long)smp_processor_id();
5570 /* Initialize migration for the boot CPU */
5571 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5572 BUG_ON(err
== NOTIFY_BAD
);
5573 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5574 register_cpu_notifier(&migration_notifier
);
5576 /* Register cpu active notifiers */
5577 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5578 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5582 early_initcall(migration_init
);
5584 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5586 #ifdef CONFIG_SCHED_DEBUG
5588 static __read_mostly
int sched_debug_enabled
;
5590 static int __init
sched_debug_setup(char *str
)
5592 sched_debug_enabled
= 1;
5596 early_param("sched_debug", sched_debug_setup
);
5598 static inline bool sched_debug(void)
5600 return sched_debug_enabled
;
5603 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5604 struct cpumask
*groupmask
)
5606 struct sched_group
*group
= sd
->groups
;
5608 cpumask_clear(groupmask
);
5610 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5612 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5613 printk("does not load-balance\n");
5615 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5620 printk(KERN_CONT
"span %*pbl level %s\n",
5621 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5623 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5624 printk(KERN_ERR
"ERROR: domain->span does not contain "
5627 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5628 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5632 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5636 printk(KERN_ERR
"ERROR: group is NULL\n");
5640 if (!cpumask_weight(sched_group_cpus(group
))) {
5641 printk(KERN_CONT
"\n");
5642 printk(KERN_ERR
"ERROR: empty group\n");
5646 if (!(sd
->flags
& SD_OVERLAP
) &&
5647 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5648 printk(KERN_CONT
"\n");
5649 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5653 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5655 printk(KERN_CONT
" %*pbl",
5656 cpumask_pr_args(sched_group_cpus(group
)));
5657 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5658 printk(KERN_CONT
" (cpu_capacity = %d)",
5659 group
->sgc
->capacity
);
5662 group
= group
->next
;
5663 } while (group
!= sd
->groups
);
5664 printk(KERN_CONT
"\n");
5666 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5667 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5670 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5671 printk(KERN_ERR
"ERROR: parent span is not a superset "
5672 "of domain->span\n");
5676 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5680 if (!sched_debug_enabled
)
5684 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5688 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5691 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5699 #else /* !CONFIG_SCHED_DEBUG */
5700 # define sched_domain_debug(sd, cpu) do { } while (0)
5701 static inline bool sched_debug(void)
5705 #endif /* CONFIG_SCHED_DEBUG */
5707 static int sd_degenerate(struct sched_domain
*sd
)
5709 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5712 /* Following flags need at least 2 groups */
5713 if (sd
->flags
& (SD_LOAD_BALANCE
|
5714 SD_BALANCE_NEWIDLE
|
5717 SD_SHARE_CPUCAPACITY
|
5718 SD_SHARE_PKG_RESOURCES
|
5719 SD_SHARE_POWERDOMAIN
)) {
5720 if (sd
->groups
!= sd
->groups
->next
)
5724 /* Following flags don't use groups */
5725 if (sd
->flags
& (SD_WAKE_AFFINE
))
5732 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5734 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5736 if (sd_degenerate(parent
))
5739 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5742 /* Flags needing groups don't count if only 1 group in parent */
5743 if (parent
->groups
== parent
->groups
->next
) {
5744 pflags
&= ~(SD_LOAD_BALANCE
|
5745 SD_BALANCE_NEWIDLE
|
5748 SD_SHARE_CPUCAPACITY
|
5749 SD_SHARE_PKG_RESOURCES
|
5751 SD_SHARE_POWERDOMAIN
);
5752 if (nr_node_ids
== 1)
5753 pflags
&= ~SD_SERIALIZE
;
5755 if (~cflags
& pflags
)
5761 static void free_rootdomain(struct rcu_head
*rcu
)
5763 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5765 cpupri_cleanup(&rd
->cpupri
);
5766 cpudl_cleanup(&rd
->cpudl
);
5767 free_cpumask_var(rd
->dlo_mask
);
5768 free_cpumask_var(rd
->rto_mask
);
5769 free_cpumask_var(rd
->online
);
5770 free_cpumask_var(rd
->span
);
5774 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5776 struct root_domain
*old_rd
= NULL
;
5777 unsigned long flags
;
5779 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5784 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5787 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5790 * If we dont want to free the old_rd yet then
5791 * set old_rd to NULL to skip the freeing later
5794 if (!atomic_dec_and_test(&old_rd
->refcount
))
5798 atomic_inc(&rd
->refcount
);
5801 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5802 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5805 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5808 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5811 static int init_rootdomain(struct root_domain
*rd
)
5813 memset(rd
, 0, sizeof(*rd
));
5815 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5817 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5819 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5821 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5824 init_dl_bw(&rd
->dl_bw
);
5825 if (cpudl_init(&rd
->cpudl
) != 0)
5828 if (cpupri_init(&rd
->cpupri
) != 0)
5833 free_cpumask_var(rd
->rto_mask
);
5835 free_cpumask_var(rd
->dlo_mask
);
5837 free_cpumask_var(rd
->online
);
5839 free_cpumask_var(rd
->span
);
5845 * By default the system creates a single root-domain with all cpus as
5846 * members (mimicking the global state we have today).
5848 struct root_domain def_root_domain
;
5850 static void init_defrootdomain(void)
5852 init_rootdomain(&def_root_domain
);
5854 atomic_set(&def_root_domain
.refcount
, 1);
5857 static struct root_domain
*alloc_rootdomain(void)
5859 struct root_domain
*rd
;
5861 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5865 if (init_rootdomain(rd
) != 0) {
5873 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5875 struct sched_group
*tmp
, *first
;
5884 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5889 } while (sg
!= first
);
5892 static void free_sched_domain(struct rcu_head
*rcu
)
5894 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5897 * If its an overlapping domain it has private groups, iterate and
5900 if (sd
->flags
& SD_OVERLAP
) {
5901 free_sched_groups(sd
->groups
, 1);
5902 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5903 kfree(sd
->groups
->sgc
);
5909 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5911 call_rcu(&sd
->rcu
, free_sched_domain
);
5914 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5916 for (; sd
; sd
= sd
->parent
)
5917 destroy_sched_domain(sd
, cpu
);
5921 * Keep a special pointer to the highest sched_domain that has
5922 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5923 * allows us to avoid some pointer chasing select_idle_sibling().
5925 * Also keep a unique ID per domain (we use the first cpu number in
5926 * the cpumask of the domain), this allows us to quickly tell if
5927 * two cpus are in the same cache domain, see cpus_share_cache().
5929 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5930 DEFINE_PER_CPU(int, sd_llc_size
);
5931 DEFINE_PER_CPU(int, sd_llc_id
);
5932 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5933 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5934 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5936 static void update_top_cache_domain(int cpu
)
5938 struct sched_domain
*sd
;
5939 struct sched_domain
*busy_sd
= NULL
;
5943 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5945 id
= cpumask_first(sched_domain_span(sd
));
5946 size
= cpumask_weight(sched_domain_span(sd
));
5947 busy_sd
= sd
->parent
; /* sd_busy */
5949 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5951 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5952 per_cpu(sd_llc_size
, cpu
) = size
;
5953 per_cpu(sd_llc_id
, cpu
) = id
;
5955 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5956 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5958 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5959 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5963 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5964 * hold the hotplug lock.
5967 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5969 struct rq
*rq
= cpu_rq(cpu
);
5970 struct sched_domain
*tmp
;
5972 /* Remove the sched domains which do not contribute to scheduling. */
5973 for (tmp
= sd
; tmp
; ) {
5974 struct sched_domain
*parent
= tmp
->parent
;
5978 if (sd_parent_degenerate(tmp
, parent
)) {
5979 tmp
->parent
= parent
->parent
;
5981 parent
->parent
->child
= tmp
;
5983 * Transfer SD_PREFER_SIBLING down in case of a
5984 * degenerate parent; the spans match for this
5985 * so the property transfers.
5987 if (parent
->flags
& SD_PREFER_SIBLING
)
5988 tmp
->flags
|= SD_PREFER_SIBLING
;
5989 destroy_sched_domain(parent
, cpu
);
5994 if (sd
&& sd_degenerate(sd
)) {
5997 destroy_sched_domain(tmp
, cpu
);
6002 sched_domain_debug(sd
, cpu
);
6004 rq_attach_root(rq
, rd
);
6006 rcu_assign_pointer(rq
->sd
, sd
);
6007 destroy_sched_domains(tmp
, cpu
);
6009 update_top_cache_domain(cpu
);
6012 /* Setup the mask of cpus configured for isolated domains */
6013 static int __init
isolated_cpu_setup(char *str
)
6015 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6016 cpulist_parse(str
, cpu_isolated_map
);
6020 __setup("isolcpus=", isolated_cpu_setup
);
6023 struct sched_domain
** __percpu sd
;
6024 struct root_domain
*rd
;
6035 * Build an iteration mask that can exclude certain CPUs from the upwards
6038 * Asymmetric node setups can result in situations where the domain tree is of
6039 * unequal depth, make sure to skip domains that already cover the entire
6042 * In that case build_sched_domains() will have terminated the iteration early
6043 * and our sibling sd spans will be empty. Domains should always include the
6044 * cpu they're built on, so check that.
6047 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6049 const struct cpumask
*span
= sched_domain_span(sd
);
6050 struct sd_data
*sdd
= sd
->private;
6051 struct sched_domain
*sibling
;
6054 for_each_cpu(i
, span
) {
6055 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6056 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6059 cpumask_set_cpu(i
, sched_group_mask(sg
));
6064 * Return the canonical balance cpu for this group, this is the first cpu
6065 * of this group that's also in the iteration mask.
6067 int group_balance_cpu(struct sched_group
*sg
)
6069 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6073 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6075 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6076 const struct cpumask
*span
= sched_domain_span(sd
);
6077 struct cpumask
*covered
= sched_domains_tmpmask
;
6078 struct sd_data
*sdd
= sd
->private;
6079 struct sched_domain
*sibling
;
6082 cpumask_clear(covered
);
6084 for_each_cpu(i
, span
) {
6085 struct cpumask
*sg_span
;
6087 if (cpumask_test_cpu(i
, covered
))
6090 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6092 /* See the comment near build_group_mask(). */
6093 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6096 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6097 GFP_KERNEL
, cpu_to_node(cpu
));
6102 sg_span
= sched_group_cpus(sg
);
6104 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6106 cpumask_set_cpu(i
, sg_span
);
6108 cpumask_or(covered
, covered
, sg_span
);
6110 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6111 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6112 build_group_mask(sd
, sg
);
6115 * Initialize sgc->capacity such that even if we mess up the
6116 * domains and no possible iteration will get us here, we won't
6119 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6122 * Make sure the first group of this domain contains the
6123 * canonical balance cpu. Otherwise the sched_domain iteration
6124 * breaks. See update_sg_lb_stats().
6126 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6127 group_balance_cpu(sg
) == cpu
)
6137 sd
->groups
= groups
;
6142 free_sched_groups(first
, 0);
6147 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6149 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6150 struct sched_domain
*child
= sd
->child
;
6153 cpu
= cpumask_first(sched_domain_span(child
));
6156 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6157 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6158 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6165 * build_sched_groups will build a circular linked list of the groups
6166 * covered by the given span, and will set each group's ->cpumask correctly,
6167 * and ->cpu_capacity to 0.
6169 * Assumes the sched_domain tree is fully constructed
6172 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6174 struct sched_group
*first
= NULL
, *last
= NULL
;
6175 struct sd_data
*sdd
= sd
->private;
6176 const struct cpumask
*span
= sched_domain_span(sd
);
6177 struct cpumask
*covered
;
6180 get_group(cpu
, sdd
, &sd
->groups
);
6181 atomic_inc(&sd
->groups
->ref
);
6183 if (cpu
!= cpumask_first(span
))
6186 lockdep_assert_held(&sched_domains_mutex
);
6187 covered
= sched_domains_tmpmask
;
6189 cpumask_clear(covered
);
6191 for_each_cpu(i
, span
) {
6192 struct sched_group
*sg
;
6195 if (cpumask_test_cpu(i
, covered
))
6198 group
= get_group(i
, sdd
, &sg
);
6199 cpumask_setall(sched_group_mask(sg
));
6201 for_each_cpu(j
, span
) {
6202 if (get_group(j
, sdd
, NULL
) != group
)
6205 cpumask_set_cpu(j
, covered
);
6206 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6221 * Initialize sched groups cpu_capacity.
6223 * cpu_capacity indicates the capacity of sched group, which is used while
6224 * distributing the load between different sched groups in a sched domain.
6225 * Typically cpu_capacity for all the groups in a sched domain will be same
6226 * unless there are asymmetries in the topology. If there are asymmetries,
6227 * group having more cpu_capacity will pickup more load compared to the
6228 * group having less cpu_capacity.
6230 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6232 struct sched_group
*sg
= sd
->groups
;
6237 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6239 } while (sg
!= sd
->groups
);
6241 if (cpu
!= group_balance_cpu(sg
))
6244 update_group_capacity(sd
, cpu
);
6245 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6249 * Initializers for schedule domains
6250 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6253 static int default_relax_domain_level
= -1;
6254 int sched_domain_level_max
;
6256 static int __init
setup_relax_domain_level(char *str
)
6258 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6259 pr_warn("Unable to set relax_domain_level\n");
6263 __setup("relax_domain_level=", setup_relax_domain_level
);
6265 static void set_domain_attribute(struct sched_domain
*sd
,
6266 struct sched_domain_attr
*attr
)
6270 if (!attr
|| attr
->relax_domain_level
< 0) {
6271 if (default_relax_domain_level
< 0)
6274 request
= default_relax_domain_level
;
6276 request
= attr
->relax_domain_level
;
6277 if (request
< sd
->level
) {
6278 /* turn off idle balance on this domain */
6279 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6281 /* turn on idle balance on this domain */
6282 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6286 static void __sdt_free(const struct cpumask
*cpu_map
);
6287 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6289 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6290 const struct cpumask
*cpu_map
)
6294 if (!atomic_read(&d
->rd
->refcount
))
6295 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6297 free_percpu(d
->sd
); /* fall through */
6299 __sdt_free(cpu_map
); /* fall through */
6305 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6306 const struct cpumask
*cpu_map
)
6308 memset(d
, 0, sizeof(*d
));
6310 if (__sdt_alloc(cpu_map
))
6311 return sa_sd_storage
;
6312 d
->sd
= alloc_percpu(struct sched_domain
*);
6314 return sa_sd_storage
;
6315 d
->rd
= alloc_rootdomain();
6318 return sa_rootdomain
;
6322 * NULL the sd_data elements we've used to build the sched_domain and
6323 * sched_group structure so that the subsequent __free_domain_allocs()
6324 * will not free the data we're using.
6326 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6328 struct sd_data
*sdd
= sd
->private;
6330 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6331 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6333 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6334 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6336 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6337 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6341 static int sched_domains_numa_levels
;
6342 enum numa_topology_type sched_numa_topology_type
;
6343 static int *sched_domains_numa_distance
;
6344 int sched_max_numa_distance
;
6345 static struct cpumask
***sched_domains_numa_masks
;
6346 static int sched_domains_curr_level
;
6350 * SD_flags allowed in topology descriptions.
6352 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6353 * SD_SHARE_PKG_RESOURCES - describes shared caches
6354 * SD_NUMA - describes NUMA topologies
6355 * SD_SHARE_POWERDOMAIN - describes shared power domain
6358 * SD_ASYM_PACKING - describes SMT quirks
6360 #define TOPOLOGY_SD_FLAGS \
6361 (SD_SHARE_CPUCAPACITY | \
6362 SD_SHARE_PKG_RESOURCES | \
6365 SD_SHARE_POWERDOMAIN)
6367 static struct sched_domain
*
6368 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6370 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6371 int sd_weight
, sd_flags
= 0;
6375 * Ugly hack to pass state to sd_numa_mask()...
6377 sched_domains_curr_level
= tl
->numa_level
;
6380 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6383 sd_flags
= (*tl
->sd_flags
)();
6384 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6385 "wrong sd_flags in topology description\n"))
6386 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6388 *sd
= (struct sched_domain
){
6389 .min_interval
= sd_weight
,
6390 .max_interval
= 2*sd_weight
,
6392 .imbalance_pct
= 125,
6394 .cache_nice_tries
= 0,
6401 .flags
= 1*SD_LOAD_BALANCE
6402 | 1*SD_BALANCE_NEWIDLE
6407 | 0*SD_SHARE_CPUCAPACITY
6408 | 0*SD_SHARE_PKG_RESOURCES
6410 | 0*SD_PREFER_SIBLING
6415 .last_balance
= jiffies
,
6416 .balance_interval
= sd_weight
,
6418 .max_newidle_lb_cost
= 0,
6419 .next_decay_max_lb_cost
= jiffies
,
6420 #ifdef CONFIG_SCHED_DEBUG
6426 * Convert topological properties into behaviour.
6429 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6430 sd
->flags
|= SD_PREFER_SIBLING
;
6431 sd
->imbalance_pct
= 110;
6432 sd
->smt_gain
= 1178; /* ~15% */
6434 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6435 sd
->imbalance_pct
= 117;
6436 sd
->cache_nice_tries
= 1;
6440 } else if (sd
->flags
& SD_NUMA
) {
6441 sd
->cache_nice_tries
= 2;
6445 sd
->flags
|= SD_SERIALIZE
;
6446 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6447 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6454 sd
->flags
|= SD_PREFER_SIBLING
;
6455 sd
->cache_nice_tries
= 1;
6460 sd
->private = &tl
->data
;
6466 * Topology list, bottom-up.
6468 static struct sched_domain_topology_level default_topology
[] = {
6469 #ifdef CONFIG_SCHED_SMT
6470 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6472 #ifdef CONFIG_SCHED_MC
6473 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6475 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6479 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6481 #define for_each_sd_topology(tl) \
6482 for (tl = sched_domain_topology; tl->mask; tl++)
6484 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6486 sched_domain_topology
= tl
;
6491 static const struct cpumask
*sd_numa_mask(int cpu
)
6493 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6496 static void sched_numa_warn(const char *str
)
6498 static int done
= false;
6506 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6508 for (i
= 0; i
< nr_node_ids
; i
++) {
6509 printk(KERN_WARNING
" ");
6510 for (j
= 0; j
< nr_node_ids
; j
++)
6511 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6512 printk(KERN_CONT
"\n");
6514 printk(KERN_WARNING
"\n");
6517 bool find_numa_distance(int distance
)
6521 if (distance
== node_distance(0, 0))
6524 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6525 if (sched_domains_numa_distance
[i
] == distance
)
6533 * A system can have three types of NUMA topology:
6534 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6535 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6536 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6538 * The difference between a glueless mesh topology and a backplane
6539 * topology lies in whether communication between not directly
6540 * connected nodes goes through intermediary nodes (where programs
6541 * could run), or through backplane controllers. This affects
6542 * placement of programs.
6544 * The type of topology can be discerned with the following tests:
6545 * - If the maximum distance between any nodes is 1 hop, the system
6546 * is directly connected.
6547 * - If for two nodes A and B, located N > 1 hops away from each other,
6548 * there is an intermediary node C, which is < N hops away from both
6549 * nodes A and B, the system is a glueless mesh.
6551 static void init_numa_topology_type(void)
6555 n
= sched_max_numa_distance
;
6557 if (sched_domains_numa_levels
<= 1) {
6558 sched_numa_topology_type
= NUMA_DIRECT
;
6562 for_each_online_node(a
) {
6563 for_each_online_node(b
) {
6564 /* Find two nodes furthest removed from each other. */
6565 if (node_distance(a
, b
) < n
)
6568 /* Is there an intermediary node between a and b? */
6569 for_each_online_node(c
) {
6570 if (node_distance(a
, c
) < n
&&
6571 node_distance(b
, c
) < n
) {
6572 sched_numa_topology_type
=
6578 sched_numa_topology_type
= NUMA_BACKPLANE
;
6584 static void sched_init_numa(void)
6586 int next_distance
, curr_distance
= node_distance(0, 0);
6587 struct sched_domain_topology_level
*tl
;
6591 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6592 if (!sched_domains_numa_distance
)
6596 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6597 * unique distances in the node_distance() table.
6599 * Assumes node_distance(0,j) includes all distances in
6600 * node_distance(i,j) in order to avoid cubic time.
6602 next_distance
= curr_distance
;
6603 for (i
= 0; i
< nr_node_ids
; i
++) {
6604 for (j
= 0; j
< nr_node_ids
; j
++) {
6605 for (k
= 0; k
< nr_node_ids
; k
++) {
6606 int distance
= node_distance(i
, k
);
6608 if (distance
> curr_distance
&&
6609 (distance
< next_distance
||
6610 next_distance
== curr_distance
))
6611 next_distance
= distance
;
6614 * While not a strong assumption it would be nice to know
6615 * about cases where if node A is connected to B, B is not
6616 * equally connected to A.
6618 if (sched_debug() && node_distance(k
, i
) != distance
)
6619 sched_numa_warn("Node-distance not symmetric");
6621 if (sched_debug() && i
&& !find_numa_distance(distance
))
6622 sched_numa_warn("Node-0 not representative");
6624 if (next_distance
!= curr_distance
) {
6625 sched_domains_numa_distance
[level
++] = next_distance
;
6626 sched_domains_numa_levels
= level
;
6627 curr_distance
= next_distance
;
6632 * In case of sched_debug() we verify the above assumption.
6642 * 'level' contains the number of unique distances, excluding the
6643 * identity distance node_distance(i,i).
6645 * The sched_domains_numa_distance[] array includes the actual distance
6650 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6651 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6652 * the array will contain less then 'level' members. This could be
6653 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6654 * in other functions.
6656 * We reset it to 'level' at the end of this function.
6658 sched_domains_numa_levels
= 0;
6660 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6661 if (!sched_domains_numa_masks
)
6665 * Now for each level, construct a mask per node which contains all
6666 * cpus of nodes that are that many hops away from us.
6668 for (i
= 0; i
< level
; i
++) {
6669 sched_domains_numa_masks
[i
] =
6670 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6671 if (!sched_domains_numa_masks
[i
])
6674 for (j
= 0; j
< nr_node_ids
; j
++) {
6675 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6679 sched_domains_numa_masks
[i
][j
] = mask
;
6681 for (k
= 0; k
< nr_node_ids
; k
++) {
6682 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6685 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6690 /* Compute default topology size */
6691 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6693 tl
= kzalloc((i
+ level
+ 1) *
6694 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6699 * Copy the default topology bits..
6701 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6702 tl
[i
] = sched_domain_topology
[i
];
6705 * .. and append 'j' levels of NUMA goodness.
6707 for (j
= 0; j
< level
; i
++, j
++) {
6708 tl
[i
] = (struct sched_domain_topology_level
){
6709 .mask
= sd_numa_mask
,
6710 .sd_flags
= cpu_numa_flags
,
6711 .flags
= SDTL_OVERLAP
,
6717 sched_domain_topology
= tl
;
6719 sched_domains_numa_levels
= level
;
6720 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6722 init_numa_topology_type();
6725 static void sched_domains_numa_masks_set(int cpu
)
6728 int node
= cpu_to_node(cpu
);
6730 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6731 for (j
= 0; j
< nr_node_ids
; j
++) {
6732 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6733 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6738 static void sched_domains_numa_masks_clear(int cpu
)
6741 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6742 for (j
= 0; j
< nr_node_ids
; j
++)
6743 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6748 * Update sched_domains_numa_masks[level][node] array when new cpus
6751 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6752 unsigned long action
,
6755 int cpu
= (long)hcpu
;
6757 switch (action
& ~CPU_TASKS_FROZEN
) {
6759 sched_domains_numa_masks_set(cpu
);
6763 sched_domains_numa_masks_clear(cpu
);
6773 static inline void sched_init_numa(void)
6777 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6778 unsigned long action
,
6783 #endif /* CONFIG_NUMA */
6785 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6787 struct sched_domain_topology_level
*tl
;
6790 for_each_sd_topology(tl
) {
6791 struct sd_data
*sdd
= &tl
->data
;
6793 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6797 sdd
->sg
= alloc_percpu(struct sched_group
*);
6801 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6805 for_each_cpu(j
, cpu_map
) {
6806 struct sched_domain
*sd
;
6807 struct sched_group
*sg
;
6808 struct sched_group_capacity
*sgc
;
6810 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6811 GFP_KERNEL
, cpu_to_node(j
));
6815 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6817 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6818 GFP_KERNEL
, cpu_to_node(j
));
6824 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6826 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6827 GFP_KERNEL
, cpu_to_node(j
));
6831 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6838 static void __sdt_free(const struct cpumask
*cpu_map
)
6840 struct sched_domain_topology_level
*tl
;
6843 for_each_sd_topology(tl
) {
6844 struct sd_data
*sdd
= &tl
->data
;
6846 for_each_cpu(j
, cpu_map
) {
6847 struct sched_domain
*sd
;
6850 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6851 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6852 free_sched_groups(sd
->groups
, 0);
6853 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6857 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6859 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6861 free_percpu(sdd
->sd
);
6863 free_percpu(sdd
->sg
);
6865 free_percpu(sdd
->sgc
);
6870 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6871 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6872 struct sched_domain
*child
, int cpu
)
6874 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6878 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6880 sd
->level
= child
->level
+ 1;
6881 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6885 if (!cpumask_subset(sched_domain_span(child
),
6886 sched_domain_span(sd
))) {
6887 pr_err("BUG: arch topology borken\n");
6888 #ifdef CONFIG_SCHED_DEBUG
6889 pr_err(" the %s domain not a subset of the %s domain\n",
6890 child
->name
, sd
->name
);
6892 /* Fixup, ensure @sd has at least @child cpus. */
6893 cpumask_or(sched_domain_span(sd
),
6894 sched_domain_span(sd
),
6895 sched_domain_span(child
));
6899 set_domain_attribute(sd
, attr
);
6905 * Build sched domains for a given set of cpus and attach the sched domains
6906 * to the individual cpus
6908 static int build_sched_domains(const struct cpumask
*cpu_map
,
6909 struct sched_domain_attr
*attr
)
6911 enum s_alloc alloc_state
;
6912 struct sched_domain
*sd
;
6914 int i
, ret
= -ENOMEM
;
6916 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6917 if (alloc_state
!= sa_rootdomain
)
6920 /* Set up domains for cpus specified by the cpu_map. */
6921 for_each_cpu(i
, cpu_map
) {
6922 struct sched_domain_topology_level
*tl
;
6925 for_each_sd_topology(tl
) {
6926 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6927 if (tl
== sched_domain_topology
)
6928 *per_cpu_ptr(d
.sd
, i
) = sd
;
6929 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6930 sd
->flags
|= SD_OVERLAP
;
6931 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6936 /* Build the groups for the domains */
6937 for_each_cpu(i
, cpu_map
) {
6938 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6939 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6940 if (sd
->flags
& SD_OVERLAP
) {
6941 if (build_overlap_sched_groups(sd
, i
))
6944 if (build_sched_groups(sd
, i
))
6950 /* Calculate CPU capacity for physical packages and nodes */
6951 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6952 if (!cpumask_test_cpu(i
, cpu_map
))
6955 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6956 claim_allocations(i
, sd
);
6957 init_sched_groups_capacity(i
, sd
);
6961 /* Attach the domains */
6963 for_each_cpu(i
, cpu_map
) {
6964 sd
= *per_cpu_ptr(d
.sd
, i
);
6965 cpu_attach_domain(sd
, d
.rd
, i
);
6971 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6975 static cpumask_var_t
*doms_cur
; /* current sched domains */
6976 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6977 static struct sched_domain_attr
*dattr_cur
;
6978 /* attribues of custom domains in 'doms_cur' */
6981 * Special case: If a kmalloc of a doms_cur partition (array of
6982 * cpumask) fails, then fallback to a single sched domain,
6983 * as determined by the single cpumask fallback_doms.
6985 static cpumask_var_t fallback_doms
;
6988 * arch_update_cpu_topology lets virtualized architectures update the
6989 * cpu core maps. It is supposed to return 1 if the topology changed
6990 * or 0 if it stayed the same.
6992 int __weak
arch_update_cpu_topology(void)
6997 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7000 cpumask_var_t
*doms
;
7002 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7005 for (i
= 0; i
< ndoms
; i
++) {
7006 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7007 free_sched_domains(doms
, i
);
7014 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7017 for (i
= 0; i
< ndoms
; i
++)
7018 free_cpumask_var(doms
[i
]);
7023 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7024 * For now this just excludes isolated cpus, but could be used to
7025 * exclude other special cases in the future.
7027 static int init_sched_domains(const struct cpumask
*cpu_map
)
7031 arch_update_cpu_topology();
7033 doms_cur
= alloc_sched_domains(ndoms_cur
);
7035 doms_cur
= &fallback_doms
;
7036 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7037 err
= build_sched_domains(doms_cur
[0], NULL
);
7038 register_sched_domain_sysctl();
7044 * Detach sched domains from a group of cpus specified in cpu_map
7045 * These cpus will now be attached to the NULL domain
7047 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7052 for_each_cpu(i
, cpu_map
)
7053 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7057 /* handle null as "default" */
7058 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7059 struct sched_domain_attr
*new, int idx_new
)
7061 struct sched_domain_attr tmp
;
7068 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7069 new ? (new + idx_new
) : &tmp
,
7070 sizeof(struct sched_domain_attr
));
7074 * Partition sched domains as specified by the 'ndoms_new'
7075 * cpumasks in the array doms_new[] of cpumasks. This compares
7076 * doms_new[] to the current sched domain partitioning, doms_cur[].
7077 * It destroys each deleted domain and builds each new domain.
7079 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7080 * The masks don't intersect (don't overlap.) We should setup one
7081 * sched domain for each mask. CPUs not in any of the cpumasks will
7082 * not be load balanced. If the same cpumask appears both in the
7083 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7086 * The passed in 'doms_new' should be allocated using
7087 * alloc_sched_domains. This routine takes ownership of it and will
7088 * free_sched_domains it when done with it. If the caller failed the
7089 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7090 * and partition_sched_domains() will fallback to the single partition
7091 * 'fallback_doms', it also forces the domains to be rebuilt.
7093 * If doms_new == NULL it will be replaced with cpu_online_mask.
7094 * ndoms_new == 0 is a special case for destroying existing domains,
7095 * and it will not create the default domain.
7097 * Call with hotplug lock held
7099 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7100 struct sched_domain_attr
*dattr_new
)
7105 mutex_lock(&sched_domains_mutex
);
7107 /* always unregister in case we don't destroy any domains */
7108 unregister_sched_domain_sysctl();
7110 /* Let architecture update cpu core mappings. */
7111 new_topology
= arch_update_cpu_topology();
7113 n
= doms_new
? ndoms_new
: 0;
7115 /* Destroy deleted domains */
7116 for (i
= 0; i
< ndoms_cur
; i
++) {
7117 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7118 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7119 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7122 /* no match - a current sched domain not in new doms_new[] */
7123 detach_destroy_domains(doms_cur
[i
]);
7129 if (doms_new
== NULL
) {
7131 doms_new
= &fallback_doms
;
7132 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7133 WARN_ON_ONCE(dattr_new
);
7136 /* Build new domains */
7137 for (i
= 0; i
< ndoms_new
; i
++) {
7138 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7139 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7140 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7143 /* no match - add a new doms_new */
7144 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7149 /* Remember the new sched domains */
7150 if (doms_cur
!= &fallback_doms
)
7151 free_sched_domains(doms_cur
, ndoms_cur
);
7152 kfree(dattr_cur
); /* kfree(NULL) is safe */
7153 doms_cur
= doms_new
;
7154 dattr_cur
= dattr_new
;
7155 ndoms_cur
= ndoms_new
;
7157 register_sched_domain_sysctl();
7159 mutex_unlock(&sched_domains_mutex
);
7162 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7165 * Update cpusets according to cpu_active mask. If cpusets are
7166 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7167 * around partition_sched_domains().
7169 * If we come here as part of a suspend/resume, don't touch cpusets because we
7170 * want to restore it back to its original state upon resume anyway.
7172 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7176 case CPU_ONLINE_FROZEN
:
7177 case CPU_DOWN_FAILED_FROZEN
:
7180 * num_cpus_frozen tracks how many CPUs are involved in suspend
7181 * resume sequence. As long as this is not the last online
7182 * operation in the resume sequence, just build a single sched
7183 * domain, ignoring cpusets.
7186 if (likely(num_cpus_frozen
)) {
7187 partition_sched_domains(1, NULL
, NULL
);
7192 * This is the last CPU online operation. So fall through and
7193 * restore the original sched domains by considering the
7194 * cpuset configurations.
7198 cpuset_update_active_cpus(true);
7206 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7209 unsigned long flags
;
7210 long cpu
= (long)hcpu
;
7216 case CPU_DOWN_PREPARE
:
7217 rcu_read_lock_sched();
7218 dl_b
= dl_bw_of(cpu
);
7220 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7221 cpus
= dl_bw_cpus(cpu
);
7222 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7223 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7225 rcu_read_unlock_sched();
7228 return notifier_from_errno(-EBUSY
);
7229 cpuset_update_active_cpus(false);
7231 case CPU_DOWN_PREPARE_FROZEN
:
7233 partition_sched_domains(1, NULL
, NULL
);
7241 void __init
sched_init_smp(void)
7243 cpumask_var_t non_isolated_cpus
;
7245 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7246 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7251 * There's no userspace yet to cause hotplug operations; hence all the
7252 * cpu masks are stable and all blatant races in the below code cannot
7255 mutex_lock(&sched_domains_mutex
);
7256 init_sched_domains(cpu_active_mask
);
7257 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7258 if (cpumask_empty(non_isolated_cpus
))
7259 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7260 mutex_unlock(&sched_domains_mutex
);
7262 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7263 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7264 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7268 /* Move init over to a non-isolated CPU */
7269 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7271 sched_init_granularity();
7272 free_cpumask_var(non_isolated_cpus
);
7274 init_sched_rt_class();
7275 init_sched_dl_class();
7278 void __init
sched_init_smp(void)
7280 sched_init_granularity();
7282 #endif /* CONFIG_SMP */
7284 int in_sched_functions(unsigned long addr
)
7286 return in_lock_functions(addr
) ||
7287 (addr
>= (unsigned long)__sched_text_start
7288 && addr
< (unsigned long)__sched_text_end
);
7291 #ifdef CONFIG_CGROUP_SCHED
7293 * Default task group.
7294 * Every task in system belongs to this group at bootup.
7296 struct task_group root_task_group
;
7297 LIST_HEAD(task_groups
);
7300 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7302 void __init
sched_init(void)
7305 unsigned long alloc_size
= 0, ptr
;
7307 #ifdef CONFIG_FAIR_GROUP_SCHED
7308 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7310 #ifdef CONFIG_RT_GROUP_SCHED
7311 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7314 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7316 #ifdef CONFIG_FAIR_GROUP_SCHED
7317 root_task_group
.se
= (struct sched_entity
**)ptr
;
7318 ptr
+= nr_cpu_ids
* sizeof(void **);
7320 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7321 ptr
+= nr_cpu_ids
* sizeof(void **);
7323 #endif /* CONFIG_FAIR_GROUP_SCHED */
7324 #ifdef CONFIG_RT_GROUP_SCHED
7325 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7326 ptr
+= nr_cpu_ids
* sizeof(void **);
7328 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7329 ptr
+= nr_cpu_ids
* sizeof(void **);
7331 #endif /* CONFIG_RT_GROUP_SCHED */
7333 #ifdef CONFIG_CPUMASK_OFFSTACK
7334 for_each_possible_cpu(i
) {
7335 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7336 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7338 #endif /* CONFIG_CPUMASK_OFFSTACK */
7340 init_rt_bandwidth(&def_rt_bandwidth
,
7341 global_rt_period(), global_rt_runtime());
7342 init_dl_bandwidth(&def_dl_bandwidth
,
7343 global_rt_period(), global_rt_runtime());
7346 init_defrootdomain();
7349 #ifdef CONFIG_RT_GROUP_SCHED
7350 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7351 global_rt_period(), global_rt_runtime());
7352 #endif /* CONFIG_RT_GROUP_SCHED */
7354 #ifdef CONFIG_CGROUP_SCHED
7355 list_add(&root_task_group
.list
, &task_groups
);
7356 INIT_LIST_HEAD(&root_task_group
.children
);
7357 INIT_LIST_HEAD(&root_task_group
.siblings
);
7358 autogroup_init(&init_task
);
7360 #endif /* CONFIG_CGROUP_SCHED */
7362 for_each_possible_cpu(i
) {
7366 raw_spin_lock_init(&rq
->lock
);
7368 rq
->calc_load_active
= 0;
7369 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7370 init_cfs_rq(&rq
->cfs
);
7371 init_rt_rq(&rq
->rt
);
7372 init_dl_rq(&rq
->dl
);
7373 #ifdef CONFIG_FAIR_GROUP_SCHED
7374 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7375 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7377 * How much cpu bandwidth does root_task_group get?
7379 * In case of task-groups formed thr' the cgroup filesystem, it
7380 * gets 100% of the cpu resources in the system. This overall
7381 * system cpu resource is divided among the tasks of
7382 * root_task_group and its child task-groups in a fair manner,
7383 * based on each entity's (task or task-group's) weight
7384 * (se->load.weight).
7386 * In other words, if root_task_group has 10 tasks of weight
7387 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7388 * then A0's share of the cpu resource is:
7390 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7392 * We achieve this by letting root_task_group's tasks sit
7393 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7395 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7396 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7397 #endif /* CONFIG_FAIR_GROUP_SCHED */
7399 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7400 #ifdef CONFIG_RT_GROUP_SCHED
7401 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7404 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7405 rq
->cpu_load
[j
] = 0;
7407 rq
->last_load_update_tick
= jiffies
;
7412 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7413 rq
->balance_callback
= NULL
;
7414 rq
->active_balance
= 0;
7415 rq
->next_balance
= jiffies
;
7420 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7421 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7423 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7425 rq_attach_root(rq
, &def_root_domain
);
7426 #ifdef CONFIG_NO_HZ_COMMON
7429 #ifdef CONFIG_NO_HZ_FULL
7430 rq
->last_sched_tick
= 0;
7434 atomic_set(&rq
->nr_iowait
, 0);
7437 set_load_weight(&init_task
);
7439 #ifdef CONFIG_PREEMPT_NOTIFIERS
7440 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7444 * The boot idle thread does lazy MMU switching as well:
7446 atomic_inc(&init_mm
.mm_count
);
7447 enter_lazy_tlb(&init_mm
, current
);
7450 * During early bootup we pretend to be a normal task:
7452 current
->sched_class
= &fair_sched_class
;
7455 * Make us the idle thread. Technically, schedule() should not be
7456 * called from this thread, however somewhere below it might be,
7457 * but because we are the idle thread, we just pick up running again
7458 * when this runqueue becomes "idle".
7460 init_idle(current
, smp_processor_id());
7462 calc_load_update
= jiffies
+ LOAD_FREQ
;
7465 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7466 /* May be allocated at isolcpus cmdline parse time */
7467 if (cpu_isolated_map
== NULL
)
7468 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7469 idle_thread_set_boot_cpu();
7470 set_cpu_rq_start_time();
7472 init_sched_fair_class();
7474 scheduler_running
= 1;
7477 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7478 static inline int preempt_count_equals(int preempt_offset
)
7480 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7482 return (nested
== preempt_offset
);
7485 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7488 * Blocking primitives will set (and therefore destroy) current->state,
7489 * since we will exit with TASK_RUNNING make sure we enter with it,
7490 * otherwise we will destroy state.
7492 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7493 "do not call blocking ops when !TASK_RUNNING; "
7494 "state=%lx set at [<%p>] %pS\n",
7496 (void *)current
->task_state_change
,
7497 (void *)current
->task_state_change
);
7499 ___might_sleep(file
, line
, preempt_offset
);
7501 EXPORT_SYMBOL(__might_sleep
);
7503 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7505 static unsigned long prev_jiffy
; /* ratelimiting */
7507 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7508 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7509 !is_idle_task(current
)) ||
7510 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7512 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7514 prev_jiffy
= jiffies
;
7517 "BUG: sleeping function called from invalid context at %s:%d\n",
7520 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7521 in_atomic(), irqs_disabled(),
7522 current
->pid
, current
->comm
);
7524 if (task_stack_end_corrupted(current
))
7525 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7527 debug_show_held_locks(current
);
7528 if (irqs_disabled())
7529 print_irqtrace_events(current
);
7530 #ifdef CONFIG_DEBUG_PREEMPT
7531 if (!preempt_count_equals(preempt_offset
)) {
7532 pr_err("Preemption disabled at:");
7533 print_ip_sym(current
->preempt_disable_ip
);
7539 EXPORT_SYMBOL(___might_sleep
);
7542 #ifdef CONFIG_MAGIC_SYSRQ
7543 void normalize_rt_tasks(void)
7545 struct task_struct
*g
, *p
;
7546 struct sched_attr attr
= {
7547 .sched_policy
= SCHED_NORMAL
,
7550 read_lock(&tasklist_lock
);
7551 for_each_process_thread(g
, p
) {
7553 * Only normalize user tasks:
7555 if (p
->flags
& PF_KTHREAD
)
7558 p
->se
.exec_start
= 0;
7559 #ifdef CONFIG_SCHEDSTATS
7560 p
->se
.statistics
.wait_start
= 0;
7561 p
->se
.statistics
.sleep_start
= 0;
7562 p
->se
.statistics
.block_start
= 0;
7565 if (!dl_task(p
) && !rt_task(p
)) {
7567 * Renice negative nice level userspace
7570 if (task_nice(p
) < 0)
7571 set_user_nice(p
, 0);
7575 __sched_setscheduler(p
, &attr
, false, false);
7577 read_unlock(&tasklist_lock
);
7580 #endif /* CONFIG_MAGIC_SYSRQ */
7582 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7584 * These functions are only useful for the IA64 MCA handling, or kdb.
7586 * They can only be called when the whole system has been
7587 * stopped - every CPU needs to be quiescent, and no scheduling
7588 * activity can take place. Using them for anything else would
7589 * be a serious bug, and as a result, they aren't even visible
7590 * under any other configuration.
7594 * curr_task - return the current task for a given cpu.
7595 * @cpu: the processor in question.
7597 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7599 * Return: The current task for @cpu.
7601 struct task_struct
*curr_task(int cpu
)
7603 return cpu_curr(cpu
);
7606 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7610 * set_curr_task - set the current task for a given cpu.
7611 * @cpu: the processor in question.
7612 * @p: the task pointer to set.
7614 * Description: This function must only be used when non-maskable interrupts
7615 * are serviced on a separate stack. It allows the architecture to switch the
7616 * notion of the current task on a cpu in a non-blocking manner. This function
7617 * must be called with all CPU's synchronized, and interrupts disabled, the
7618 * and caller must save the original value of the current task (see
7619 * curr_task() above) and restore that value before reenabling interrupts and
7620 * re-starting the system.
7622 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7624 void set_curr_task(int cpu
, struct task_struct
*p
)
7631 #ifdef CONFIG_CGROUP_SCHED
7632 /* task_group_lock serializes the addition/removal of task groups */
7633 static DEFINE_SPINLOCK(task_group_lock
);
7635 static void free_sched_group(struct task_group
*tg
)
7637 free_fair_sched_group(tg
);
7638 free_rt_sched_group(tg
);
7643 /* allocate runqueue etc for a new task group */
7644 struct task_group
*sched_create_group(struct task_group
*parent
)
7646 struct task_group
*tg
;
7648 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7650 return ERR_PTR(-ENOMEM
);
7652 if (!alloc_fair_sched_group(tg
, parent
))
7655 if (!alloc_rt_sched_group(tg
, parent
))
7661 free_sched_group(tg
);
7662 return ERR_PTR(-ENOMEM
);
7665 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7667 unsigned long flags
;
7669 spin_lock_irqsave(&task_group_lock
, flags
);
7670 list_add_rcu(&tg
->list
, &task_groups
);
7672 WARN_ON(!parent
); /* root should already exist */
7674 tg
->parent
= parent
;
7675 INIT_LIST_HEAD(&tg
->children
);
7676 list_add_rcu(&tg
->siblings
, &parent
->children
);
7677 spin_unlock_irqrestore(&task_group_lock
, flags
);
7680 /* rcu callback to free various structures associated with a task group */
7681 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7683 /* now it should be safe to free those cfs_rqs */
7684 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7687 /* Destroy runqueue etc associated with a task group */
7688 void sched_destroy_group(struct task_group
*tg
)
7690 /* wait for possible concurrent references to cfs_rqs complete */
7691 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7694 void sched_offline_group(struct task_group
*tg
)
7696 unsigned long flags
;
7699 /* end participation in shares distribution */
7700 for_each_possible_cpu(i
)
7701 unregister_fair_sched_group(tg
, i
);
7703 spin_lock_irqsave(&task_group_lock
, flags
);
7704 list_del_rcu(&tg
->list
);
7705 list_del_rcu(&tg
->siblings
);
7706 spin_unlock_irqrestore(&task_group_lock
, flags
);
7709 /* change task's runqueue when it moves between groups.
7710 * The caller of this function should have put the task in its new group
7711 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7712 * reflect its new group.
7714 void sched_move_task(struct task_struct
*tsk
)
7716 struct task_group
*tg
;
7717 int queued
, running
;
7718 unsigned long flags
;
7721 rq
= task_rq_lock(tsk
, &flags
);
7723 running
= task_current(rq
, tsk
);
7724 queued
= task_on_rq_queued(tsk
);
7727 dequeue_task(rq
, tsk
, 0);
7728 if (unlikely(running
))
7729 put_prev_task(rq
, tsk
);
7732 * All callers are synchronized by task_rq_lock(); we do not use RCU
7733 * which is pointless here. Thus, we pass "true" to task_css_check()
7734 * to prevent lockdep warnings.
7736 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7737 struct task_group
, css
);
7738 tg
= autogroup_task_group(tsk
, tg
);
7739 tsk
->sched_task_group
= tg
;
7741 #ifdef CONFIG_FAIR_GROUP_SCHED
7742 if (tsk
->sched_class
->task_move_group
)
7743 tsk
->sched_class
->task_move_group(tsk
, queued
);
7746 set_task_rq(tsk
, task_cpu(tsk
));
7748 if (unlikely(running
))
7749 tsk
->sched_class
->set_curr_task(rq
);
7751 enqueue_task(rq
, tsk
, 0);
7753 task_rq_unlock(rq
, tsk
, &flags
);
7755 #endif /* CONFIG_CGROUP_SCHED */
7757 #ifdef CONFIG_RT_GROUP_SCHED
7759 * Ensure that the real time constraints are schedulable.
7761 static DEFINE_MUTEX(rt_constraints_mutex
);
7763 /* Must be called with tasklist_lock held */
7764 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7766 struct task_struct
*g
, *p
;
7769 * Autogroups do not have RT tasks; see autogroup_create().
7771 if (task_group_is_autogroup(tg
))
7774 for_each_process_thread(g
, p
) {
7775 if (rt_task(p
) && task_group(p
) == tg
)
7782 struct rt_schedulable_data
{
7783 struct task_group
*tg
;
7788 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7790 struct rt_schedulable_data
*d
= data
;
7791 struct task_group
*child
;
7792 unsigned long total
, sum
= 0;
7793 u64 period
, runtime
;
7795 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7796 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7799 period
= d
->rt_period
;
7800 runtime
= d
->rt_runtime
;
7804 * Cannot have more runtime than the period.
7806 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7810 * Ensure we don't starve existing RT tasks.
7812 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7815 total
= to_ratio(period
, runtime
);
7818 * Nobody can have more than the global setting allows.
7820 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7824 * The sum of our children's runtime should not exceed our own.
7826 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7827 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7828 runtime
= child
->rt_bandwidth
.rt_runtime
;
7830 if (child
== d
->tg
) {
7831 period
= d
->rt_period
;
7832 runtime
= d
->rt_runtime
;
7835 sum
+= to_ratio(period
, runtime
);
7844 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7848 struct rt_schedulable_data data
= {
7850 .rt_period
= period
,
7851 .rt_runtime
= runtime
,
7855 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7861 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7862 u64 rt_period
, u64 rt_runtime
)
7867 * Disallowing the root group RT runtime is BAD, it would disallow the
7868 * kernel creating (and or operating) RT threads.
7870 if (tg
== &root_task_group
&& rt_runtime
== 0)
7873 /* No period doesn't make any sense. */
7877 mutex_lock(&rt_constraints_mutex
);
7878 read_lock(&tasklist_lock
);
7879 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7883 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7884 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7885 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7887 for_each_possible_cpu(i
) {
7888 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7890 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7891 rt_rq
->rt_runtime
= rt_runtime
;
7892 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7894 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7896 read_unlock(&tasklist_lock
);
7897 mutex_unlock(&rt_constraints_mutex
);
7902 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7904 u64 rt_runtime
, rt_period
;
7906 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7907 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7908 if (rt_runtime_us
< 0)
7909 rt_runtime
= RUNTIME_INF
;
7911 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7914 static long sched_group_rt_runtime(struct task_group
*tg
)
7918 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7921 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7922 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7923 return rt_runtime_us
;
7926 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7928 u64 rt_runtime
, rt_period
;
7930 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7931 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7933 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7936 static long sched_group_rt_period(struct task_group
*tg
)
7940 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7941 do_div(rt_period_us
, NSEC_PER_USEC
);
7942 return rt_period_us
;
7944 #endif /* CONFIG_RT_GROUP_SCHED */
7946 #ifdef CONFIG_RT_GROUP_SCHED
7947 static int sched_rt_global_constraints(void)
7951 mutex_lock(&rt_constraints_mutex
);
7952 read_lock(&tasklist_lock
);
7953 ret
= __rt_schedulable(NULL
, 0, 0);
7954 read_unlock(&tasklist_lock
);
7955 mutex_unlock(&rt_constraints_mutex
);
7960 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7962 /* Don't accept realtime tasks when there is no way for them to run */
7963 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7969 #else /* !CONFIG_RT_GROUP_SCHED */
7970 static int sched_rt_global_constraints(void)
7972 unsigned long flags
;
7975 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7976 for_each_possible_cpu(i
) {
7977 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7979 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7980 rt_rq
->rt_runtime
= global_rt_runtime();
7981 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7983 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7987 #endif /* CONFIG_RT_GROUP_SCHED */
7989 static int sched_dl_global_validate(void)
7991 u64 runtime
= global_rt_runtime();
7992 u64 period
= global_rt_period();
7993 u64 new_bw
= to_ratio(period
, runtime
);
7996 unsigned long flags
;
7999 * Here we want to check the bandwidth not being set to some
8000 * value smaller than the currently allocated bandwidth in
8001 * any of the root_domains.
8003 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8004 * cycling on root_domains... Discussion on different/better
8005 * solutions is welcome!
8007 for_each_possible_cpu(cpu
) {
8008 rcu_read_lock_sched();
8009 dl_b
= dl_bw_of(cpu
);
8011 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8012 if (new_bw
< dl_b
->total_bw
)
8014 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8016 rcu_read_unlock_sched();
8025 static void sched_dl_do_global(void)
8030 unsigned long flags
;
8032 def_dl_bandwidth
.dl_period
= global_rt_period();
8033 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8035 if (global_rt_runtime() != RUNTIME_INF
)
8036 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8039 * FIXME: As above...
8041 for_each_possible_cpu(cpu
) {
8042 rcu_read_lock_sched();
8043 dl_b
= dl_bw_of(cpu
);
8045 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8047 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8049 rcu_read_unlock_sched();
8053 static int sched_rt_global_validate(void)
8055 if (sysctl_sched_rt_period
<= 0)
8058 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8059 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8065 static void sched_rt_do_global(void)
8067 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8068 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8071 int sched_rt_handler(struct ctl_table
*table
, int write
,
8072 void __user
*buffer
, size_t *lenp
,
8075 int old_period
, old_runtime
;
8076 static DEFINE_MUTEX(mutex
);
8080 old_period
= sysctl_sched_rt_period
;
8081 old_runtime
= sysctl_sched_rt_runtime
;
8083 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8085 if (!ret
&& write
) {
8086 ret
= sched_rt_global_validate();
8090 ret
= sched_dl_global_validate();
8094 ret
= sched_rt_global_constraints();
8098 sched_rt_do_global();
8099 sched_dl_do_global();
8103 sysctl_sched_rt_period
= old_period
;
8104 sysctl_sched_rt_runtime
= old_runtime
;
8106 mutex_unlock(&mutex
);
8111 int sched_rr_handler(struct ctl_table
*table
, int write
,
8112 void __user
*buffer
, size_t *lenp
,
8116 static DEFINE_MUTEX(mutex
);
8119 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8120 /* make sure that internally we keep jiffies */
8121 /* also, writing zero resets timeslice to default */
8122 if (!ret
&& write
) {
8123 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8124 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8126 mutex_unlock(&mutex
);
8130 #ifdef CONFIG_CGROUP_SCHED
8132 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8134 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8137 static struct cgroup_subsys_state
*
8138 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8140 struct task_group
*parent
= css_tg(parent_css
);
8141 struct task_group
*tg
;
8144 /* This is early initialization for the top cgroup */
8145 return &root_task_group
.css
;
8148 tg
= sched_create_group(parent
);
8150 return ERR_PTR(-ENOMEM
);
8155 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8157 struct task_group
*tg
= css_tg(css
);
8158 struct task_group
*parent
= css_tg(css
->parent
);
8161 sched_online_group(tg
, parent
);
8165 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8167 struct task_group
*tg
= css_tg(css
);
8169 sched_destroy_group(tg
);
8172 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8174 struct task_group
*tg
= css_tg(css
);
8176 sched_offline_group(tg
);
8179 static void cpu_cgroup_fork(struct task_struct
*task
, void *private)
8181 sched_move_task(task
);
8184 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
8185 struct cgroup_taskset
*tset
)
8187 struct task_struct
*task
;
8189 cgroup_taskset_for_each(task
, tset
) {
8190 #ifdef CONFIG_RT_GROUP_SCHED
8191 if (!sched_rt_can_attach(css_tg(css
), task
))
8194 /* We don't support RT-tasks being in separate groups */
8195 if (task
->sched_class
!= &fair_sched_class
)
8202 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8203 struct cgroup_taskset
*tset
)
8205 struct task_struct
*task
;
8207 cgroup_taskset_for_each(task
, tset
)
8208 sched_move_task(task
);
8211 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8212 struct cgroup_subsys_state
*old_css
,
8213 struct task_struct
*task
)
8216 * cgroup_exit() is called in the copy_process() failure path.
8217 * Ignore this case since the task hasn't ran yet, this avoids
8218 * trying to poke a half freed task state from generic code.
8220 if (!(task
->flags
& PF_EXITING
))
8223 sched_move_task(task
);
8226 #ifdef CONFIG_FAIR_GROUP_SCHED
8227 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8228 struct cftype
*cftype
, u64 shareval
)
8230 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8233 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8236 struct task_group
*tg
= css_tg(css
);
8238 return (u64
) scale_load_down(tg
->shares
);
8241 #ifdef CONFIG_CFS_BANDWIDTH
8242 static DEFINE_MUTEX(cfs_constraints_mutex
);
8244 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8245 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8247 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8249 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8251 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8252 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8254 if (tg
== &root_task_group
)
8258 * Ensure we have at some amount of bandwidth every period. This is
8259 * to prevent reaching a state of large arrears when throttled via
8260 * entity_tick() resulting in prolonged exit starvation.
8262 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8266 * Likewise, bound things on the otherside by preventing insane quota
8267 * periods. This also allows us to normalize in computing quota
8270 if (period
> max_cfs_quota_period
)
8274 * Prevent race between setting of cfs_rq->runtime_enabled and
8275 * unthrottle_offline_cfs_rqs().
8278 mutex_lock(&cfs_constraints_mutex
);
8279 ret
= __cfs_schedulable(tg
, period
, quota
);
8283 runtime_enabled
= quota
!= RUNTIME_INF
;
8284 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8286 * If we need to toggle cfs_bandwidth_used, off->on must occur
8287 * before making related changes, and on->off must occur afterwards
8289 if (runtime_enabled
&& !runtime_was_enabled
)
8290 cfs_bandwidth_usage_inc();
8291 raw_spin_lock_irq(&cfs_b
->lock
);
8292 cfs_b
->period
= ns_to_ktime(period
);
8293 cfs_b
->quota
= quota
;
8295 __refill_cfs_bandwidth_runtime(cfs_b
);
8296 /* restart the period timer (if active) to handle new period expiry */
8297 if (runtime_enabled
)
8298 start_cfs_bandwidth(cfs_b
);
8299 raw_spin_unlock_irq(&cfs_b
->lock
);
8301 for_each_online_cpu(i
) {
8302 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8303 struct rq
*rq
= cfs_rq
->rq
;
8305 raw_spin_lock_irq(&rq
->lock
);
8306 cfs_rq
->runtime_enabled
= runtime_enabled
;
8307 cfs_rq
->runtime_remaining
= 0;
8309 if (cfs_rq
->throttled
)
8310 unthrottle_cfs_rq(cfs_rq
);
8311 raw_spin_unlock_irq(&rq
->lock
);
8313 if (runtime_was_enabled
&& !runtime_enabled
)
8314 cfs_bandwidth_usage_dec();
8316 mutex_unlock(&cfs_constraints_mutex
);
8322 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8326 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8327 if (cfs_quota_us
< 0)
8328 quota
= RUNTIME_INF
;
8330 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8332 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8335 long tg_get_cfs_quota(struct task_group
*tg
)
8339 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8342 quota_us
= tg
->cfs_bandwidth
.quota
;
8343 do_div(quota_us
, NSEC_PER_USEC
);
8348 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8352 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8353 quota
= tg
->cfs_bandwidth
.quota
;
8355 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8358 long tg_get_cfs_period(struct task_group
*tg
)
8362 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8363 do_div(cfs_period_us
, NSEC_PER_USEC
);
8365 return cfs_period_us
;
8368 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8371 return tg_get_cfs_quota(css_tg(css
));
8374 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8375 struct cftype
*cftype
, s64 cfs_quota_us
)
8377 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8380 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8383 return tg_get_cfs_period(css_tg(css
));
8386 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8387 struct cftype
*cftype
, u64 cfs_period_us
)
8389 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8392 struct cfs_schedulable_data
{
8393 struct task_group
*tg
;
8398 * normalize group quota/period to be quota/max_period
8399 * note: units are usecs
8401 static u64
normalize_cfs_quota(struct task_group
*tg
,
8402 struct cfs_schedulable_data
*d
)
8410 period
= tg_get_cfs_period(tg
);
8411 quota
= tg_get_cfs_quota(tg
);
8414 /* note: these should typically be equivalent */
8415 if (quota
== RUNTIME_INF
|| quota
== -1)
8418 return to_ratio(period
, quota
);
8421 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8423 struct cfs_schedulable_data
*d
= data
;
8424 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8425 s64 quota
= 0, parent_quota
= -1;
8428 quota
= RUNTIME_INF
;
8430 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8432 quota
= normalize_cfs_quota(tg
, d
);
8433 parent_quota
= parent_b
->hierarchical_quota
;
8436 * ensure max(child_quota) <= parent_quota, inherit when no
8439 if (quota
== RUNTIME_INF
)
8440 quota
= parent_quota
;
8441 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8444 cfs_b
->hierarchical_quota
= quota
;
8449 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8452 struct cfs_schedulable_data data
= {
8458 if (quota
!= RUNTIME_INF
) {
8459 do_div(data
.period
, NSEC_PER_USEC
);
8460 do_div(data
.quota
, NSEC_PER_USEC
);
8464 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8470 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8472 struct task_group
*tg
= css_tg(seq_css(sf
));
8473 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8475 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8476 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8477 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8481 #endif /* CONFIG_CFS_BANDWIDTH */
8482 #endif /* CONFIG_FAIR_GROUP_SCHED */
8484 #ifdef CONFIG_RT_GROUP_SCHED
8485 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8486 struct cftype
*cft
, s64 val
)
8488 return sched_group_set_rt_runtime(css_tg(css
), val
);
8491 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8494 return sched_group_rt_runtime(css_tg(css
));
8497 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8498 struct cftype
*cftype
, u64 rt_period_us
)
8500 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8503 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8506 return sched_group_rt_period(css_tg(css
));
8508 #endif /* CONFIG_RT_GROUP_SCHED */
8510 static struct cftype cpu_files
[] = {
8511 #ifdef CONFIG_FAIR_GROUP_SCHED
8514 .read_u64
= cpu_shares_read_u64
,
8515 .write_u64
= cpu_shares_write_u64
,
8518 #ifdef CONFIG_CFS_BANDWIDTH
8520 .name
= "cfs_quota_us",
8521 .read_s64
= cpu_cfs_quota_read_s64
,
8522 .write_s64
= cpu_cfs_quota_write_s64
,
8525 .name
= "cfs_period_us",
8526 .read_u64
= cpu_cfs_period_read_u64
,
8527 .write_u64
= cpu_cfs_period_write_u64
,
8531 .seq_show
= cpu_stats_show
,
8534 #ifdef CONFIG_RT_GROUP_SCHED
8536 .name
= "rt_runtime_us",
8537 .read_s64
= cpu_rt_runtime_read
,
8538 .write_s64
= cpu_rt_runtime_write
,
8541 .name
= "rt_period_us",
8542 .read_u64
= cpu_rt_period_read_uint
,
8543 .write_u64
= cpu_rt_period_write_uint
,
8549 struct cgroup_subsys cpu_cgrp_subsys
= {
8550 .css_alloc
= cpu_cgroup_css_alloc
,
8551 .css_free
= cpu_cgroup_css_free
,
8552 .css_online
= cpu_cgroup_css_online
,
8553 .css_offline
= cpu_cgroup_css_offline
,
8554 .fork
= cpu_cgroup_fork
,
8555 .can_attach
= cpu_cgroup_can_attach
,
8556 .attach
= cpu_cgroup_attach
,
8557 .exit
= cpu_cgroup_exit
,
8558 .legacy_cftypes
= cpu_files
,
8562 #endif /* CONFIG_CGROUP_SCHED */
8564 void dump_cpu_task(int cpu
)
8566 pr_info("Task dump for CPU %d:\n", cpu
);
8567 sched_show_task(cpu_curr(cpu
));