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
)
2370 p
->sched_class
->task_woken(rq
, p
);
2372 task_rq_unlock(rq
, p
, &flags
);
2375 #ifdef CONFIG_PREEMPT_NOTIFIERS
2377 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2379 void preempt_notifier_inc(void)
2381 static_key_slow_inc(&preempt_notifier_key
);
2383 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2385 void preempt_notifier_dec(void)
2387 static_key_slow_dec(&preempt_notifier_key
);
2389 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2392 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2393 * @notifier: notifier struct to register
2395 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2397 if (!static_key_false(&preempt_notifier_key
))
2398 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2400 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2402 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2405 * preempt_notifier_unregister - no longer interested in preemption notifications
2406 * @notifier: notifier struct to unregister
2408 * This is *not* safe to call from within a preemption notifier.
2410 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2412 hlist_del(¬ifier
->link
);
2414 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2416 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2418 struct preempt_notifier
*notifier
;
2420 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2421 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2424 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2426 if (static_key_false(&preempt_notifier_key
))
2427 __fire_sched_in_preempt_notifiers(curr
);
2431 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2432 struct task_struct
*next
)
2434 struct preempt_notifier
*notifier
;
2436 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2437 notifier
->ops
->sched_out(notifier
, next
);
2440 static __always_inline
void
2441 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2442 struct task_struct
*next
)
2444 if (static_key_false(&preempt_notifier_key
))
2445 __fire_sched_out_preempt_notifiers(curr
, next
);
2448 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2450 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2455 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2456 struct task_struct
*next
)
2460 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2463 * prepare_task_switch - prepare to switch tasks
2464 * @rq: the runqueue preparing to switch
2465 * @prev: the current task that is being switched out
2466 * @next: the task we are going to switch to.
2468 * This is called with the rq lock held and interrupts off. It must
2469 * be paired with a subsequent finish_task_switch after the context
2472 * prepare_task_switch sets up locking and calls architecture specific
2476 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2477 struct task_struct
*next
)
2479 trace_sched_switch(prev
, next
);
2480 sched_info_switch(rq
, prev
, next
);
2481 perf_event_task_sched_out(prev
, next
);
2482 fire_sched_out_preempt_notifiers(prev
, next
);
2483 prepare_lock_switch(rq
, next
);
2484 prepare_arch_switch(next
);
2488 * finish_task_switch - clean up after a task-switch
2489 * @prev: the thread we just switched away from.
2491 * finish_task_switch must be called after the context switch, paired
2492 * with a prepare_task_switch call before the context switch.
2493 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2494 * and do any other architecture-specific cleanup actions.
2496 * Note that we may have delayed dropping an mm in context_switch(). If
2497 * so, we finish that here outside of the runqueue lock. (Doing it
2498 * with the lock held can cause deadlocks; see schedule() for
2501 * The context switch have flipped the stack from under us and restored the
2502 * local variables which were saved when this task called schedule() in the
2503 * past. prev == current is still correct but we need to recalculate this_rq
2504 * because prev may have moved to another CPU.
2506 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2507 __releases(rq
->lock
)
2509 struct rq
*rq
= this_rq();
2510 struct mm_struct
*mm
= rq
->prev_mm
;
2516 * A task struct has one reference for the use as "current".
2517 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2518 * schedule one last time. The schedule call will never return, and
2519 * the scheduled task must drop that reference.
2521 * We must observe prev->state before clearing prev->on_cpu (in
2522 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2523 * running on another CPU and we could rave with its RUNNING -> DEAD
2524 * transition, resulting in a double drop.
2526 prev_state
= prev
->state
;
2527 vtime_task_switch(prev
);
2528 perf_event_task_sched_in(prev
, current
);
2529 finish_lock_switch(rq
, prev
);
2530 finish_arch_post_lock_switch();
2532 fire_sched_in_preempt_notifiers(current
);
2535 if (unlikely(prev_state
== TASK_DEAD
)) {
2536 if (prev
->sched_class
->task_dead
)
2537 prev
->sched_class
->task_dead(prev
);
2540 * Remove function-return probe instances associated with this
2541 * task and put them back on the free list.
2543 kprobe_flush_task(prev
);
2544 put_task_struct(prev
);
2547 tick_nohz_task_switch();
2553 /* rq->lock is NOT held, but preemption is disabled */
2554 static void __balance_callback(struct rq
*rq
)
2556 struct callback_head
*head
, *next
;
2557 void (*func
)(struct rq
*rq
);
2558 unsigned long flags
;
2560 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2561 head
= rq
->balance_callback
;
2562 rq
->balance_callback
= NULL
;
2564 func
= (void (*)(struct rq
*))head
->func
;
2571 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2574 static inline void balance_callback(struct rq
*rq
)
2576 if (unlikely(rq
->balance_callback
))
2577 __balance_callback(rq
);
2582 static inline void balance_callback(struct rq
*rq
)
2589 * schedule_tail - first thing a freshly forked thread must call.
2590 * @prev: the thread we just switched away from.
2592 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2593 __releases(rq
->lock
)
2597 /* finish_task_switch() drops rq->lock and enables preemtion */
2599 rq
= finish_task_switch(prev
);
2600 balance_callback(rq
);
2603 if (current
->set_child_tid
)
2604 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2608 * context_switch - switch to the new MM and the new thread's register state.
2610 static inline struct rq
*
2611 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2612 struct task_struct
*next
)
2614 struct mm_struct
*mm
, *oldmm
;
2616 prepare_task_switch(rq
, prev
, next
);
2619 oldmm
= prev
->active_mm
;
2621 * For paravirt, this is coupled with an exit in switch_to to
2622 * combine the page table reload and the switch backend into
2625 arch_start_context_switch(prev
);
2628 next
->active_mm
= oldmm
;
2629 atomic_inc(&oldmm
->mm_count
);
2630 enter_lazy_tlb(oldmm
, next
);
2632 switch_mm(oldmm
, mm
, next
);
2635 prev
->active_mm
= NULL
;
2636 rq
->prev_mm
= oldmm
;
2639 * Since the runqueue lock will be released by the next
2640 * task (which is an invalid locking op but in the case
2641 * of the scheduler it's an obvious special-case), so we
2642 * do an early lockdep release here:
2644 lockdep_unpin_lock(&rq
->lock
);
2645 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2647 /* Here we just switch the register state and the stack. */
2648 switch_to(prev
, next
, prev
);
2651 return finish_task_switch(prev
);
2655 * nr_running and nr_context_switches:
2657 * externally visible scheduler statistics: current number of runnable
2658 * threads, total number of context switches performed since bootup.
2660 unsigned long nr_running(void)
2662 unsigned long i
, sum
= 0;
2664 for_each_online_cpu(i
)
2665 sum
+= cpu_rq(i
)->nr_running
;
2671 * Check if only the current task is running on the cpu.
2673 * Caution: this function does not check that the caller has disabled
2674 * preemption, thus the result might have a time-of-check-to-time-of-use
2675 * race. The caller is responsible to use it correctly, for example:
2677 * - from a non-preemptable section (of course)
2679 * - from a thread that is bound to a single CPU
2681 * - in a loop with very short iterations (e.g. a polling loop)
2683 bool single_task_running(void)
2685 return raw_rq()->nr_running
== 1;
2687 EXPORT_SYMBOL(single_task_running
);
2689 unsigned long long nr_context_switches(void)
2692 unsigned long long sum
= 0;
2694 for_each_possible_cpu(i
)
2695 sum
+= cpu_rq(i
)->nr_switches
;
2700 unsigned long nr_iowait(void)
2702 unsigned long i
, sum
= 0;
2704 for_each_possible_cpu(i
)
2705 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2710 unsigned long nr_iowait_cpu(int cpu
)
2712 struct rq
*this = cpu_rq(cpu
);
2713 return atomic_read(&this->nr_iowait
);
2716 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2718 struct rq
*rq
= this_rq();
2719 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2720 *load
= rq
->load
.weight
;
2726 * sched_exec - execve() is a valuable balancing opportunity, because at
2727 * this point the task has the smallest effective memory and cache footprint.
2729 void sched_exec(void)
2731 struct task_struct
*p
= current
;
2732 unsigned long flags
;
2735 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2736 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2737 if (dest_cpu
== smp_processor_id())
2740 if (likely(cpu_active(dest_cpu
))) {
2741 struct migration_arg arg
= { p
, dest_cpu
};
2743 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2744 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2748 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2753 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2754 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2756 EXPORT_PER_CPU_SYMBOL(kstat
);
2757 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2760 * Return accounted runtime for the task.
2761 * In case the task is currently running, return the runtime plus current's
2762 * pending runtime that have not been accounted yet.
2764 unsigned long long task_sched_runtime(struct task_struct
*p
)
2766 unsigned long flags
;
2770 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2772 * 64-bit doesn't need locks to atomically read a 64bit value.
2773 * So we have a optimization chance when the task's delta_exec is 0.
2774 * Reading ->on_cpu is racy, but this is ok.
2776 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2777 * If we race with it entering cpu, unaccounted time is 0. This is
2778 * indistinguishable from the read occurring a few cycles earlier.
2779 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2780 * been accounted, so we're correct here as well.
2782 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2783 return p
->se
.sum_exec_runtime
;
2786 rq
= task_rq_lock(p
, &flags
);
2788 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2789 * project cycles that may never be accounted to this
2790 * thread, breaking clock_gettime().
2792 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2793 update_rq_clock(rq
);
2794 p
->sched_class
->update_curr(rq
);
2796 ns
= p
->se
.sum_exec_runtime
;
2797 task_rq_unlock(rq
, p
, &flags
);
2803 * This function gets called by the timer code, with HZ frequency.
2804 * We call it with interrupts disabled.
2806 void scheduler_tick(void)
2808 int cpu
= smp_processor_id();
2809 struct rq
*rq
= cpu_rq(cpu
);
2810 struct task_struct
*curr
= rq
->curr
;
2814 raw_spin_lock(&rq
->lock
);
2815 update_rq_clock(rq
);
2816 curr
->sched_class
->task_tick(rq
, curr
, 0);
2817 update_cpu_load_active(rq
);
2818 calc_global_load_tick(rq
);
2819 raw_spin_unlock(&rq
->lock
);
2821 perf_event_task_tick();
2824 rq
->idle_balance
= idle_cpu(cpu
);
2825 trigger_load_balance(rq
);
2827 rq_last_tick_reset(rq
);
2830 #ifdef CONFIG_NO_HZ_FULL
2832 * scheduler_tick_max_deferment
2834 * Keep at least one tick per second when a single
2835 * active task is running because the scheduler doesn't
2836 * yet completely support full dynticks environment.
2838 * This makes sure that uptime, CFS vruntime, load
2839 * balancing, etc... continue to move forward, even
2840 * with a very low granularity.
2842 * Return: Maximum deferment in nanoseconds.
2844 u64
scheduler_tick_max_deferment(void)
2846 struct rq
*rq
= this_rq();
2847 unsigned long next
, now
= READ_ONCE(jiffies
);
2849 next
= rq
->last_sched_tick
+ HZ
;
2851 if (time_before_eq(next
, now
))
2854 return jiffies_to_nsecs(next
- now
);
2858 notrace
unsigned long get_parent_ip(unsigned long addr
)
2860 if (in_lock_functions(addr
)) {
2861 addr
= CALLER_ADDR2
;
2862 if (in_lock_functions(addr
))
2863 addr
= CALLER_ADDR3
;
2868 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2869 defined(CONFIG_PREEMPT_TRACER))
2871 void preempt_count_add(int val
)
2873 #ifdef CONFIG_DEBUG_PREEMPT
2877 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2880 __preempt_count_add(val
);
2881 #ifdef CONFIG_DEBUG_PREEMPT
2883 * Spinlock count overflowing soon?
2885 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2888 if (preempt_count() == val
) {
2889 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2890 #ifdef CONFIG_DEBUG_PREEMPT
2891 current
->preempt_disable_ip
= ip
;
2893 trace_preempt_off(CALLER_ADDR0
, ip
);
2896 EXPORT_SYMBOL(preempt_count_add
);
2897 NOKPROBE_SYMBOL(preempt_count_add
);
2899 void preempt_count_sub(int val
)
2901 #ifdef CONFIG_DEBUG_PREEMPT
2905 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2908 * Is the spinlock portion underflowing?
2910 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2911 !(preempt_count() & PREEMPT_MASK
)))
2915 if (preempt_count() == val
)
2916 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2917 __preempt_count_sub(val
);
2919 EXPORT_SYMBOL(preempt_count_sub
);
2920 NOKPROBE_SYMBOL(preempt_count_sub
);
2925 * Print scheduling while atomic bug:
2927 static noinline
void __schedule_bug(struct task_struct
*prev
)
2929 if (oops_in_progress
)
2932 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2933 prev
->comm
, prev
->pid
, preempt_count());
2935 debug_show_held_locks(prev
);
2937 if (irqs_disabled())
2938 print_irqtrace_events(prev
);
2939 #ifdef CONFIG_DEBUG_PREEMPT
2940 if (in_atomic_preempt_off()) {
2941 pr_err("Preemption disabled at:");
2942 print_ip_sym(current
->preempt_disable_ip
);
2947 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2951 * Various schedule()-time debugging checks and statistics:
2953 static inline void schedule_debug(struct task_struct
*prev
)
2955 #ifdef CONFIG_SCHED_STACK_END_CHECK
2956 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2959 * Test if we are atomic. Since do_exit() needs to call into
2960 * schedule() atomically, we ignore that path. Otherwise whine
2961 * if we are scheduling when we should not.
2963 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2964 __schedule_bug(prev
);
2967 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2969 schedstat_inc(this_rq(), sched_count
);
2973 * Pick up the highest-prio task:
2975 static inline struct task_struct
*
2976 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2978 const struct sched_class
*class = &fair_sched_class
;
2979 struct task_struct
*p
;
2982 * Optimization: we know that if all tasks are in
2983 * the fair class we can call that function directly:
2985 if (likely(prev
->sched_class
== class &&
2986 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2987 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2988 if (unlikely(p
== RETRY_TASK
))
2991 /* assumes fair_sched_class->next == idle_sched_class */
2993 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2999 for_each_class(class) {
3000 p
= class->pick_next_task(rq
, prev
);
3002 if (unlikely(p
== RETRY_TASK
))
3008 BUG(); /* the idle class will always have a runnable task */
3012 * __schedule() is the main scheduler function.
3014 * The main means of driving the scheduler and thus entering this function are:
3016 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3018 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3019 * paths. For example, see arch/x86/entry_64.S.
3021 * To drive preemption between tasks, the scheduler sets the flag in timer
3022 * interrupt handler scheduler_tick().
3024 * 3. Wakeups don't really cause entry into schedule(). They add a
3025 * task to the run-queue and that's it.
3027 * Now, if the new task added to the run-queue preempts the current
3028 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3029 * called on the nearest possible occasion:
3031 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3033 * - in syscall or exception context, at the next outmost
3034 * preempt_enable(). (this might be as soon as the wake_up()'s
3037 * - in IRQ context, return from interrupt-handler to
3038 * preemptible context
3040 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3043 * - cond_resched() call
3044 * - explicit schedule() call
3045 * - return from syscall or exception to user-space
3046 * - return from interrupt-handler to user-space
3048 * WARNING: must be called with preemption disabled!
3050 static void __sched
__schedule(void)
3052 struct task_struct
*prev
, *next
;
3053 unsigned long *switch_count
;
3057 cpu
= smp_processor_id();
3059 rcu_note_context_switch();
3062 schedule_debug(prev
);
3064 if (sched_feat(HRTICK
))
3068 * Make sure that signal_pending_state()->signal_pending() below
3069 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3070 * done by the caller to avoid the race with signal_wake_up().
3072 smp_mb__before_spinlock();
3073 raw_spin_lock_irq(&rq
->lock
);
3074 lockdep_pin_lock(&rq
->lock
);
3076 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3078 switch_count
= &prev
->nivcsw
;
3079 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3080 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3081 prev
->state
= TASK_RUNNING
;
3083 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3087 * If a worker went to sleep, notify and ask workqueue
3088 * whether it wants to wake up a task to maintain
3091 if (prev
->flags
& PF_WQ_WORKER
) {
3092 struct task_struct
*to_wakeup
;
3094 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3096 try_to_wake_up_local(to_wakeup
);
3099 switch_count
= &prev
->nvcsw
;
3102 if (task_on_rq_queued(prev
))
3103 update_rq_clock(rq
);
3105 next
= pick_next_task(rq
, prev
);
3106 clear_tsk_need_resched(prev
);
3107 clear_preempt_need_resched();
3108 rq
->clock_skip_update
= 0;
3110 if (likely(prev
!= next
)) {
3115 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3118 lockdep_unpin_lock(&rq
->lock
);
3119 raw_spin_unlock_irq(&rq
->lock
);
3122 balance_callback(rq
);
3125 static inline void sched_submit_work(struct task_struct
*tsk
)
3127 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3130 * If we are going to sleep and we have plugged IO queued,
3131 * make sure to submit it to avoid deadlocks.
3133 if (blk_needs_flush_plug(tsk
))
3134 blk_schedule_flush_plug(tsk
);
3137 asmlinkage __visible
void __sched
schedule(void)
3139 struct task_struct
*tsk
= current
;
3141 sched_submit_work(tsk
);
3145 sched_preempt_enable_no_resched();
3146 } while (need_resched());
3148 EXPORT_SYMBOL(schedule
);
3150 #ifdef CONFIG_CONTEXT_TRACKING
3151 asmlinkage __visible
void __sched
schedule_user(void)
3154 * If we come here after a random call to set_need_resched(),
3155 * or we have been woken up remotely but the IPI has not yet arrived,
3156 * we haven't yet exited the RCU idle mode. Do it here manually until
3157 * we find a better solution.
3159 * NB: There are buggy callers of this function. Ideally we
3160 * should warn if prev_state != CONTEXT_USER, but that will trigger
3161 * too frequently to make sense yet.
3163 enum ctx_state prev_state
= exception_enter();
3165 exception_exit(prev_state
);
3170 * schedule_preempt_disabled - called with preemption disabled
3172 * Returns with preemption disabled. Note: preempt_count must be 1
3174 void __sched
schedule_preempt_disabled(void)
3176 sched_preempt_enable_no_resched();
3181 static void __sched notrace
preempt_schedule_common(void)
3184 preempt_active_enter();
3186 preempt_active_exit();
3189 * Check again in case we missed a preemption opportunity
3190 * between schedule and now.
3192 } while (need_resched());
3195 #ifdef CONFIG_PREEMPT
3197 * this is the entry point to schedule() from in-kernel preemption
3198 * off of preempt_enable. Kernel preemptions off return from interrupt
3199 * occur there and call schedule directly.
3201 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3204 * If there is a non-zero preempt_count or interrupts are disabled,
3205 * we do not want to preempt the current task. Just return..
3207 if (likely(!preemptible()))
3210 preempt_schedule_common();
3212 NOKPROBE_SYMBOL(preempt_schedule
);
3213 EXPORT_SYMBOL(preempt_schedule
);
3216 * preempt_schedule_notrace - preempt_schedule called by tracing
3218 * The tracing infrastructure uses preempt_enable_notrace to prevent
3219 * recursion and tracing preempt enabling caused by the tracing
3220 * infrastructure itself. But as tracing can happen in areas coming
3221 * from userspace or just about to enter userspace, a preempt enable
3222 * can occur before user_exit() is called. This will cause the scheduler
3223 * to be called when the system is still in usermode.
3225 * To prevent this, the preempt_enable_notrace will use this function
3226 * instead of preempt_schedule() to exit user context if needed before
3227 * calling the scheduler.
3229 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3231 enum ctx_state prev_ctx
;
3233 if (likely(!preemptible()))
3238 * Use raw __prempt_count() ops that don't call function.
3239 * We can't call functions before disabling preemption which
3240 * disarm preemption tracing recursions.
3242 __preempt_count_add(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3245 * Needs preempt disabled in case user_exit() is traced
3246 * and the tracer calls preempt_enable_notrace() causing
3247 * an infinite recursion.
3249 prev_ctx
= exception_enter();
3251 exception_exit(prev_ctx
);
3254 __preempt_count_sub(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3255 } while (need_resched());
3257 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3259 #endif /* CONFIG_PREEMPT */
3262 * this is the entry point to schedule() from kernel preemption
3263 * off of irq context.
3264 * Note, that this is called and return with irqs disabled. This will
3265 * protect us against recursive calling from irq.
3267 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3269 enum ctx_state prev_state
;
3271 /* Catch callers which need to be fixed */
3272 BUG_ON(preempt_count() || !irqs_disabled());
3274 prev_state
= exception_enter();
3277 preempt_active_enter();
3280 local_irq_disable();
3281 preempt_active_exit();
3282 } while (need_resched());
3284 exception_exit(prev_state
);
3287 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3290 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3292 EXPORT_SYMBOL(default_wake_function
);
3294 #ifdef CONFIG_RT_MUTEXES
3297 * rt_mutex_setprio - set the current priority of a task
3299 * @prio: prio value (kernel-internal form)
3301 * This function changes the 'effective' priority of a task. It does
3302 * not touch ->normal_prio like __setscheduler().
3304 * Used by the rt_mutex code to implement priority inheritance
3305 * logic. Call site only calls if the priority of the task changed.
3307 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3309 int oldprio
, queued
, running
, enqueue_flag
= 0;
3311 const struct sched_class
*prev_class
;
3313 BUG_ON(prio
> MAX_PRIO
);
3315 rq
= __task_rq_lock(p
);
3318 * Idle task boosting is a nono in general. There is one
3319 * exception, when PREEMPT_RT and NOHZ is active:
3321 * The idle task calls get_next_timer_interrupt() and holds
3322 * the timer wheel base->lock on the CPU and another CPU wants
3323 * to access the timer (probably to cancel it). We can safely
3324 * ignore the boosting request, as the idle CPU runs this code
3325 * with interrupts disabled and will complete the lock
3326 * protected section without being interrupted. So there is no
3327 * real need to boost.
3329 if (unlikely(p
== rq
->idle
)) {
3330 WARN_ON(p
!= rq
->curr
);
3331 WARN_ON(p
->pi_blocked_on
);
3335 trace_sched_pi_setprio(p
, prio
);
3337 prev_class
= p
->sched_class
;
3338 queued
= task_on_rq_queued(p
);
3339 running
= task_current(rq
, p
);
3341 dequeue_task(rq
, p
, 0);
3343 put_prev_task(rq
, p
);
3346 * Boosting condition are:
3347 * 1. -rt task is running and holds mutex A
3348 * --> -dl task blocks on mutex A
3350 * 2. -dl task is running and holds mutex A
3351 * --> -dl task blocks on mutex A and could preempt the
3354 if (dl_prio(prio
)) {
3355 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3356 if (!dl_prio(p
->normal_prio
) ||
3357 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3358 p
->dl
.dl_boosted
= 1;
3359 enqueue_flag
= ENQUEUE_REPLENISH
;
3361 p
->dl
.dl_boosted
= 0;
3362 p
->sched_class
= &dl_sched_class
;
3363 } else if (rt_prio(prio
)) {
3364 if (dl_prio(oldprio
))
3365 p
->dl
.dl_boosted
= 0;
3367 enqueue_flag
= ENQUEUE_HEAD
;
3368 p
->sched_class
= &rt_sched_class
;
3370 if (dl_prio(oldprio
))
3371 p
->dl
.dl_boosted
= 0;
3372 if (rt_prio(oldprio
))
3374 p
->sched_class
= &fair_sched_class
;
3380 p
->sched_class
->set_curr_task(rq
);
3382 enqueue_task(rq
, p
, enqueue_flag
);
3384 check_class_changed(rq
, p
, prev_class
, oldprio
);
3386 preempt_disable(); /* avoid rq from going away on us */
3387 __task_rq_unlock(rq
);
3389 balance_callback(rq
);
3394 void set_user_nice(struct task_struct
*p
, long nice
)
3396 int old_prio
, delta
, queued
;
3397 unsigned long flags
;
3400 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3403 * We have to be careful, if called from sys_setpriority(),
3404 * the task might be in the middle of scheduling on another CPU.
3406 rq
= task_rq_lock(p
, &flags
);
3408 * The RT priorities are set via sched_setscheduler(), but we still
3409 * allow the 'normal' nice value to be set - but as expected
3410 * it wont have any effect on scheduling until the task is
3411 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3413 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3414 p
->static_prio
= NICE_TO_PRIO(nice
);
3417 queued
= task_on_rq_queued(p
);
3419 dequeue_task(rq
, p
, 0);
3421 p
->static_prio
= NICE_TO_PRIO(nice
);
3424 p
->prio
= effective_prio(p
);
3425 delta
= p
->prio
- old_prio
;
3428 enqueue_task(rq
, p
, 0);
3430 * If the task increased its priority or is running and
3431 * lowered its priority, then reschedule its CPU:
3433 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3437 task_rq_unlock(rq
, p
, &flags
);
3439 EXPORT_SYMBOL(set_user_nice
);
3442 * can_nice - check if a task can reduce its nice value
3446 int can_nice(const struct task_struct
*p
, const int nice
)
3448 /* convert nice value [19,-20] to rlimit style value [1,40] */
3449 int nice_rlim
= nice_to_rlimit(nice
);
3451 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3452 capable(CAP_SYS_NICE
));
3455 #ifdef __ARCH_WANT_SYS_NICE
3458 * sys_nice - change the priority of the current process.
3459 * @increment: priority increment
3461 * sys_setpriority is a more generic, but much slower function that
3462 * does similar things.
3464 SYSCALL_DEFINE1(nice
, int, increment
)
3469 * Setpriority might change our priority at the same moment.
3470 * We don't have to worry. Conceptually one call occurs first
3471 * and we have a single winner.
3473 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3474 nice
= task_nice(current
) + increment
;
3476 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3477 if (increment
< 0 && !can_nice(current
, nice
))
3480 retval
= security_task_setnice(current
, nice
);
3484 set_user_nice(current
, nice
);
3491 * task_prio - return the priority value of a given task.
3492 * @p: the task in question.
3494 * Return: The priority value as seen by users in /proc.
3495 * RT tasks are offset by -200. Normal tasks are centered
3496 * around 0, value goes from -16 to +15.
3498 int task_prio(const struct task_struct
*p
)
3500 return p
->prio
- MAX_RT_PRIO
;
3504 * idle_cpu - is a given cpu idle currently?
3505 * @cpu: the processor in question.
3507 * Return: 1 if the CPU is currently idle. 0 otherwise.
3509 int idle_cpu(int cpu
)
3511 struct rq
*rq
= cpu_rq(cpu
);
3513 if (rq
->curr
!= rq
->idle
)
3520 if (!llist_empty(&rq
->wake_list
))
3528 * idle_task - return the idle task for a given cpu.
3529 * @cpu: the processor in question.
3531 * Return: The idle task for the cpu @cpu.
3533 struct task_struct
*idle_task(int cpu
)
3535 return cpu_rq(cpu
)->idle
;
3539 * find_process_by_pid - find a process with a matching PID value.
3540 * @pid: the pid in question.
3542 * The task of @pid, if found. %NULL otherwise.
3544 static struct task_struct
*find_process_by_pid(pid_t pid
)
3546 return pid
? find_task_by_vpid(pid
) : current
;
3550 * This function initializes the sched_dl_entity of a newly becoming
3551 * SCHED_DEADLINE task.
3553 * Only the static values are considered here, the actual runtime and the
3554 * absolute deadline will be properly calculated when the task is enqueued
3555 * for the first time with its new policy.
3558 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3560 struct sched_dl_entity
*dl_se
= &p
->dl
;
3562 dl_se
->dl_runtime
= attr
->sched_runtime
;
3563 dl_se
->dl_deadline
= attr
->sched_deadline
;
3564 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3565 dl_se
->flags
= attr
->sched_flags
;
3566 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3569 * Changing the parameters of a task is 'tricky' and we're not doing
3570 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3572 * What we SHOULD do is delay the bandwidth release until the 0-lag
3573 * point. This would include retaining the task_struct until that time
3574 * and change dl_overflow() to not immediately decrement the current
3577 * Instead we retain the current runtime/deadline and let the new
3578 * parameters take effect after the current reservation period lapses.
3579 * This is safe (albeit pessimistic) because the 0-lag point is always
3580 * before the current scheduling deadline.
3582 * We can still have temporary overloads because we do not delay the
3583 * change in bandwidth until that time; so admission control is
3584 * not on the safe side. It does however guarantee tasks will never
3585 * consume more than promised.
3590 * sched_setparam() passes in -1 for its policy, to let the functions
3591 * it calls know not to change it.
3593 #define SETPARAM_POLICY -1
3595 static void __setscheduler_params(struct task_struct
*p
,
3596 const struct sched_attr
*attr
)
3598 int policy
= attr
->sched_policy
;
3600 if (policy
== SETPARAM_POLICY
)
3605 if (dl_policy(policy
))
3606 __setparam_dl(p
, attr
);
3607 else if (fair_policy(policy
))
3608 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3611 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3612 * !rt_policy. Always setting this ensures that things like
3613 * getparam()/getattr() don't report silly values for !rt tasks.
3615 p
->rt_priority
= attr
->sched_priority
;
3616 p
->normal_prio
= normal_prio(p
);
3620 /* Actually do priority change: must hold pi & rq lock. */
3621 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3622 const struct sched_attr
*attr
, bool keep_boost
)
3624 __setscheduler_params(p
, attr
);
3627 * Keep a potential priority boosting if called from
3628 * sched_setscheduler().
3631 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3633 p
->prio
= normal_prio(p
);
3635 if (dl_prio(p
->prio
))
3636 p
->sched_class
= &dl_sched_class
;
3637 else if (rt_prio(p
->prio
))
3638 p
->sched_class
= &rt_sched_class
;
3640 p
->sched_class
= &fair_sched_class
;
3644 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3646 struct sched_dl_entity
*dl_se
= &p
->dl
;
3648 attr
->sched_priority
= p
->rt_priority
;
3649 attr
->sched_runtime
= dl_se
->dl_runtime
;
3650 attr
->sched_deadline
= dl_se
->dl_deadline
;
3651 attr
->sched_period
= dl_se
->dl_period
;
3652 attr
->sched_flags
= dl_se
->flags
;
3656 * This function validates the new parameters of a -deadline task.
3657 * We ask for the deadline not being zero, and greater or equal
3658 * than the runtime, as well as the period of being zero or
3659 * greater than deadline. Furthermore, we have to be sure that
3660 * user parameters are above the internal resolution of 1us (we
3661 * check sched_runtime only since it is always the smaller one) and
3662 * below 2^63 ns (we have to check both sched_deadline and
3663 * sched_period, as the latter can be zero).
3666 __checkparam_dl(const struct sched_attr
*attr
)
3669 if (attr
->sched_deadline
== 0)
3673 * Since we truncate DL_SCALE bits, make sure we're at least
3676 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3680 * Since we use the MSB for wrap-around and sign issues, make
3681 * sure it's not set (mind that period can be equal to zero).
3683 if (attr
->sched_deadline
& (1ULL << 63) ||
3684 attr
->sched_period
& (1ULL << 63))
3687 /* runtime <= deadline <= period (if period != 0) */
3688 if ((attr
->sched_period
!= 0 &&
3689 attr
->sched_period
< attr
->sched_deadline
) ||
3690 attr
->sched_deadline
< attr
->sched_runtime
)
3697 * check the target process has a UID that matches the current process's
3699 static bool check_same_owner(struct task_struct
*p
)
3701 const struct cred
*cred
= current_cred(), *pcred
;
3705 pcred
= __task_cred(p
);
3706 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3707 uid_eq(cred
->euid
, pcred
->uid
));
3712 static bool dl_param_changed(struct task_struct
*p
,
3713 const struct sched_attr
*attr
)
3715 struct sched_dl_entity
*dl_se
= &p
->dl
;
3717 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3718 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3719 dl_se
->dl_period
!= attr
->sched_period
||
3720 dl_se
->flags
!= attr
->sched_flags
)
3726 static int __sched_setscheduler(struct task_struct
*p
,
3727 const struct sched_attr
*attr
,
3730 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3731 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3732 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3733 int new_effective_prio
, policy
= attr
->sched_policy
;
3734 unsigned long flags
;
3735 const struct sched_class
*prev_class
;
3739 /* may grab non-irq protected spin_locks */
3740 BUG_ON(in_interrupt());
3742 /* double check policy once rq lock held */
3744 reset_on_fork
= p
->sched_reset_on_fork
;
3745 policy
= oldpolicy
= p
->policy
;
3747 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3749 if (policy
!= SCHED_DEADLINE
&&
3750 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3751 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3752 policy
!= SCHED_IDLE
)
3756 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3760 * Valid priorities for SCHED_FIFO and SCHED_RR are
3761 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3762 * SCHED_BATCH and SCHED_IDLE is 0.
3764 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3765 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3767 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3768 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3772 * Allow unprivileged RT tasks to decrease priority:
3774 if (user
&& !capable(CAP_SYS_NICE
)) {
3775 if (fair_policy(policy
)) {
3776 if (attr
->sched_nice
< task_nice(p
) &&
3777 !can_nice(p
, attr
->sched_nice
))
3781 if (rt_policy(policy
)) {
3782 unsigned long rlim_rtprio
=
3783 task_rlimit(p
, RLIMIT_RTPRIO
);
3785 /* can't set/change the rt policy */
3786 if (policy
!= p
->policy
&& !rlim_rtprio
)
3789 /* can't increase priority */
3790 if (attr
->sched_priority
> p
->rt_priority
&&
3791 attr
->sched_priority
> rlim_rtprio
)
3796 * Can't set/change SCHED_DEADLINE policy at all for now
3797 * (safest behavior); in the future we would like to allow
3798 * unprivileged DL tasks to increase their relative deadline
3799 * or reduce their runtime (both ways reducing utilization)
3801 if (dl_policy(policy
))
3805 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3806 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3808 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3809 if (!can_nice(p
, task_nice(p
)))
3813 /* can't change other user's priorities */
3814 if (!check_same_owner(p
))
3817 /* Normal users shall not reset the sched_reset_on_fork flag */
3818 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3823 retval
= security_task_setscheduler(p
);
3829 * make sure no PI-waiters arrive (or leave) while we are
3830 * changing the priority of the task:
3832 * To be able to change p->policy safely, the appropriate
3833 * runqueue lock must be held.
3835 rq
= task_rq_lock(p
, &flags
);
3838 * Changing the policy of the stop threads its a very bad idea
3840 if (p
== rq
->stop
) {
3841 task_rq_unlock(rq
, p
, &flags
);
3846 * If not changing anything there's no need to proceed further,
3847 * but store a possible modification of reset_on_fork.
3849 if (unlikely(policy
== p
->policy
)) {
3850 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3852 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3854 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3857 p
->sched_reset_on_fork
= reset_on_fork
;
3858 task_rq_unlock(rq
, p
, &flags
);
3864 #ifdef CONFIG_RT_GROUP_SCHED
3866 * Do not allow realtime tasks into groups that have no runtime
3869 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3870 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3871 !task_group_is_autogroup(task_group(p
))) {
3872 task_rq_unlock(rq
, p
, &flags
);
3877 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3878 cpumask_t
*span
= rq
->rd
->span
;
3881 * Don't allow tasks with an affinity mask smaller than
3882 * the entire root_domain to become SCHED_DEADLINE. We
3883 * will also fail if there's no bandwidth available.
3885 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3886 rq
->rd
->dl_bw
.bw
== 0) {
3887 task_rq_unlock(rq
, p
, &flags
);
3894 /* recheck policy now with rq lock held */
3895 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3896 policy
= oldpolicy
= -1;
3897 task_rq_unlock(rq
, p
, &flags
);
3902 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3903 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3906 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3907 task_rq_unlock(rq
, p
, &flags
);
3911 p
->sched_reset_on_fork
= reset_on_fork
;
3916 * Take priority boosted tasks into account. If the new
3917 * effective priority is unchanged, we just store the new
3918 * normal parameters and do not touch the scheduler class and
3919 * the runqueue. This will be done when the task deboost
3922 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
3923 if (new_effective_prio
== oldprio
) {
3924 __setscheduler_params(p
, attr
);
3925 task_rq_unlock(rq
, p
, &flags
);
3930 queued
= task_on_rq_queued(p
);
3931 running
= task_current(rq
, p
);
3933 dequeue_task(rq
, p
, 0);
3935 put_prev_task(rq
, p
);
3937 prev_class
= p
->sched_class
;
3938 __setscheduler(rq
, p
, attr
, pi
);
3941 p
->sched_class
->set_curr_task(rq
);
3944 * We enqueue to tail when the priority of a task is
3945 * increased (user space view).
3947 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3950 check_class_changed(rq
, p
, prev_class
, oldprio
);
3951 preempt_disable(); /* avoid rq from going away on us */
3952 task_rq_unlock(rq
, p
, &flags
);
3955 rt_mutex_adjust_pi(p
);
3958 * Run balance callbacks after we've adjusted the PI chain.
3960 balance_callback(rq
);
3966 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3967 const struct sched_param
*param
, bool check
)
3969 struct sched_attr attr
= {
3970 .sched_policy
= policy
,
3971 .sched_priority
= param
->sched_priority
,
3972 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3975 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3976 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3977 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3978 policy
&= ~SCHED_RESET_ON_FORK
;
3979 attr
.sched_policy
= policy
;
3982 return __sched_setscheduler(p
, &attr
, check
, true);
3985 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3986 * @p: the task in question.
3987 * @policy: new policy.
3988 * @param: structure containing the new RT priority.
3990 * Return: 0 on success. An error code otherwise.
3992 * NOTE that the task may be already dead.
3994 int sched_setscheduler(struct task_struct
*p
, int policy
,
3995 const struct sched_param
*param
)
3997 return _sched_setscheduler(p
, policy
, param
, true);
3999 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4001 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4003 return __sched_setscheduler(p
, attr
, true, true);
4005 EXPORT_SYMBOL_GPL(sched_setattr
);
4008 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4009 * @p: the task in question.
4010 * @policy: new policy.
4011 * @param: structure containing the new RT priority.
4013 * Just like sched_setscheduler, only don't bother checking if the
4014 * current context has permission. For example, this is needed in
4015 * stop_machine(): we create temporary high priority worker threads,
4016 * but our caller might not have that capability.
4018 * Return: 0 on success. An error code otherwise.
4020 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4021 const struct sched_param
*param
)
4023 return _sched_setscheduler(p
, policy
, param
, false);
4027 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4029 struct sched_param lparam
;
4030 struct task_struct
*p
;
4033 if (!param
|| pid
< 0)
4035 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4040 p
= find_process_by_pid(pid
);
4042 retval
= sched_setscheduler(p
, policy
, &lparam
);
4049 * Mimics kernel/events/core.c perf_copy_attr().
4051 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4052 struct sched_attr
*attr
)
4057 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4061 * zero the full structure, so that a short copy will be nice.
4063 memset(attr
, 0, sizeof(*attr
));
4065 ret
= get_user(size
, &uattr
->size
);
4069 if (size
> PAGE_SIZE
) /* silly large */
4072 if (!size
) /* abi compat */
4073 size
= SCHED_ATTR_SIZE_VER0
;
4075 if (size
< SCHED_ATTR_SIZE_VER0
)
4079 * If we're handed a bigger struct than we know of,
4080 * ensure all the unknown bits are 0 - i.e. new
4081 * user-space does not rely on any kernel feature
4082 * extensions we dont know about yet.
4084 if (size
> sizeof(*attr
)) {
4085 unsigned char __user
*addr
;
4086 unsigned char __user
*end
;
4089 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4090 end
= (void __user
*)uattr
+ size
;
4092 for (; addr
< end
; addr
++) {
4093 ret
= get_user(val
, addr
);
4099 size
= sizeof(*attr
);
4102 ret
= copy_from_user(attr
, uattr
, size
);
4107 * XXX: do we want to be lenient like existing syscalls; or do we want
4108 * to be strict and return an error on out-of-bounds values?
4110 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4115 put_user(sizeof(*attr
), &uattr
->size
);
4120 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4121 * @pid: the pid in question.
4122 * @policy: new policy.
4123 * @param: structure containing the new RT priority.
4125 * Return: 0 on success. An error code otherwise.
4127 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4128 struct sched_param __user
*, param
)
4130 /* negative values for policy are not valid */
4134 return do_sched_setscheduler(pid
, policy
, param
);
4138 * sys_sched_setparam - set/change the RT priority of a thread
4139 * @pid: the pid in question.
4140 * @param: structure containing the new RT priority.
4142 * Return: 0 on success. An error code otherwise.
4144 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4146 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4150 * sys_sched_setattr - same as above, but with extended sched_attr
4151 * @pid: the pid in question.
4152 * @uattr: structure containing the extended parameters.
4153 * @flags: for future extension.
4155 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4156 unsigned int, flags
)
4158 struct sched_attr attr
;
4159 struct task_struct
*p
;
4162 if (!uattr
|| pid
< 0 || flags
)
4165 retval
= sched_copy_attr(uattr
, &attr
);
4169 if ((int)attr
.sched_policy
< 0)
4174 p
= find_process_by_pid(pid
);
4176 retval
= sched_setattr(p
, &attr
);
4183 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4184 * @pid: the pid in question.
4186 * Return: On success, the policy of the thread. Otherwise, a negative error
4189 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4191 struct task_struct
*p
;
4199 p
= find_process_by_pid(pid
);
4201 retval
= security_task_getscheduler(p
);
4204 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4211 * sys_sched_getparam - get the RT priority of a thread
4212 * @pid: the pid in question.
4213 * @param: structure containing the RT priority.
4215 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4218 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4220 struct sched_param lp
= { .sched_priority
= 0 };
4221 struct task_struct
*p
;
4224 if (!param
|| pid
< 0)
4228 p
= find_process_by_pid(pid
);
4233 retval
= security_task_getscheduler(p
);
4237 if (task_has_rt_policy(p
))
4238 lp
.sched_priority
= p
->rt_priority
;
4242 * This one might sleep, we cannot do it with a spinlock held ...
4244 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4253 static int sched_read_attr(struct sched_attr __user
*uattr
,
4254 struct sched_attr
*attr
,
4259 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4263 * If we're handed a smaller struct than we know of,
4264 * ensure all the unknown bits are 0 - i.e. old
4265 * user-space does not get uncomplete information.
4267 if (usize
< sizeof(*attr
)) {
4268 unsigned char *addr
;
4271 addr
= (void *)attr
+ usize
;
4272 end
= (void *)attr
+ sizeof(*attr
);
4274 for (; addr
< end
; addr
++) {
4282 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4290 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4291 * @pid: the pid in question.
4292 * @uattr: structure containing the extended parameters.
4293 * @size: sizeof(attr) for fwd/bwd comp.
4294 * @flags: for future extension.
4296 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4297 unsigned int, size
, unsigned int, flags
)
4299 struct sched_attr attr
= {
4300 .size
= sizeof(struct sched_attr
),
4302 struct task_struct
*p
;
4305 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4306 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4310 p
= find_process_by_pid(pid
);
4315 retval
= security_task_getscheduler(p
);
4319 attr
.sched_policy
= p
->policy
;
4320 if (p
->sched_reset_on_fork
)
4321 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4322 if (task_has_dl_policy(p
))
4323 __getparam_dl(p
, &attr
);
4324 else if (task_has_rt_policy(p
))
4325 attr
.sched_priority
= p
->rt_priority
;
4327 attr
.sched_nice
= task_nice(p
);
4331 retval
= sched_read_attr(uattr
, &attr
, size
);
4339 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4341 cpumask_var_t cpus_allowed
, new_mask
;
4342 struct task_struct
*p
;
4347 p
= find_process_by_pid(pid
);
4353 /* Prevent p going away */
4357 if (p
->flags
& PF_NO_SETAFFINITY
) {
4361 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4365 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4367 goto out_free_cpus_allowed
;
4370 if (!check_same_owner(p
)) {
4372 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4374 goto out_free_new_mask
;
4379 retval
= security_task_setscheduler(p
);
4381 goto out_free_new_mask
;
4384 cpuset_cpus_allowed(p
, cpus_allowed
);
4385 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4388 * Since bandwidth control happens on root_domain basis,
4389 * if admission test is enabled, we only admit -deadline
4390 * tasks allowed to run on all the CPUs in the task's
4394 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4396 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4399 goto out_free_new_mask
;
4405 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4408 cpuset_cpus_allowed(p
, cpus_allowed
);
4409 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4411 * We must have raced with a concurrent cpuset
4412 * update. Just reset the cpus_allowed to the
4413 * cpuset's cpus_allowed
4415 cpumask_copy(new_mask
, cpus_allowed
);
4420 free_cpumask_var(new_mask
);
4421 out_free_cpus_allowed
:
4422 free_cpumask_var(cpus_allowed
);
4428 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4429 struct cpumask
*new_mask
)
4431 if (len
< cpumask_size())
4432 cpumask_clear(new_mask
);
4433 else if (len
> cpumask_size())
4434 len
= cpumask_size();
4436 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4440 * sys_sched_setaffinity - set the cpu affinity of a process
4441 * @pid: pid of the process
4442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4443 * @user_mask_ptr: user-space pointer to the new cpu mask
4445 * Return: 0 on success. An error code otherwise.
4447 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4448 unsigned long __user
*, user_mask_ptr
)
4450 cpumask_var_t new_mask
;
4453 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4456 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4458 retval
= sched_setaffinity(pid
, new_mask
);
4459 free_cpumask_var(new_mask
);
4463 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4465 struct task_struct
*p
;
4466 unsigned long flags
;
4472 p
= find_process_by_pid(pid
);
4476 retval
= security_task_getscheduler(p
);
4480 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4481 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4482 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4491 * sys_sched_getaffinity - get the cpu affinity of a process
4492 * @pid: pid of the process
4493 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4494 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4496 * Return: 0 on success. An error code otherwise.
4498 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4499 unsigned long __user
*, user_mask_ptr
)
4504 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4506 if (len
& (sizeof(unsigned long)-1))
4509 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4512 ret
= sched_getaffinity(pid
, mask
);
4514 size_t retlen
= min_t(size_t, len
, cpumask_size());
4516 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4521 free_cpumask_var(mask
);
4527 * sys_sched_yield - yield the current processor to other threads.
4529 * This function yields the current CPU to other tasks. If there are no
4530 * other threads running on this CPU then this function will return.
4534 SYSCALL_DEFINE0(sched_yield
)
4536 struct rq
*rq
= this_rq_lock();
4538 schedstat_inc(rq
, yld_count
);
4539 current
->sched_class
->yield_task(rq
);
4542 * Since we are going to call schedule() anyway, there's
4543 * no need to preempt or enable interrupts:
4545 __release(rq
->lock
);
4546 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4547 do_raw_spin_unlock(&rq
->lock
);
4548 sched_preempt_enable_no_resched();
4555 int __sched
_cond_resched(void)
4557 if (should_resched(0)) {
4558 preempt_schedule_common();
4563 EXPORT_SYMBOL(_cond_resched
);
4566 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4567 * call schedule, and on return reacquire the lock.
4569 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4570 * operations here to prevent schedule() from being called twice (once via
4571 * spin_unlock(), once by hand).
4573 int __cond_resched_lock(spinlock_t
*lock
)
4575 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4578 lockdep_assert_held(lock
);
4580 if (spin_needbreak(lock
) || resched
) {
4583 preempt_schedule_common();
4591 EXPORT_SYMBOL(__cond_resched_lock
);
4593 int __sched
__cond_resched_softirq(void)
4595 BUG_ON(!in_softirq());
4597 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4599 preempt_schedule_common();
4605 EXPORT_SYMBOL(__cond_resched_softirq
);
4608 * yield - yield the current processor to other threads.
4610 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4612 * The scheduler is at all times free to pick the calling task as the most
4613 * eligible task to run, if removing the yield() call from your code breaks
4614 * it, its already broken.
4616 * Typical broken usage is:
4621 * where one assumes that yield() will let 'the other' process run that will
4622 * make event true. If the current task is a SCHED_FIFO task that will never
4623 * happen. Never use yield() as a progress guarantee!!
4625 * If you want to use yield() to wait for something, use wait_event().
4626 * If you want to use yield() to be 'nice' for others, use cond_resched().
4627 * If you still want to use yield(), do not!
4629 void __sched
yield(void)
4631 set_current_state(TASK_RUNNING
);
4634 EXPORT_SYMBOL(yield
);
4637 * yield_to - yield the current processor to another thread in
4638 * your thread group, or accelerate that thread toward the
4639 * processor it's on.
4641 * @preempt: whether task preemption is allowed or not
4643 * It's the caller's job to ensure that the target task struct
4644 * can't go away on us before we can do any checks.
4647 * true (>0) if we indeed boosted the target task.
4648 * false (0) if we failed to boost the target.
4649 * -ESRCH if there's no task to yield to.
4651 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4653 struct task_struct
*curr
= current
;
4654 struct rq
*rq
, *p_rq
;
4655 unsigned long flags
;
4658 local_irq_save(flags
);
4664 * If we're the only runnable task on the rq and target rq also
4665 * has only one task, there's absolutely no point in yielding.
4667 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4672 double_rq_lock(rq
, p_rq
);
4673 if (task_rq(p
) != p_rq
) {
4674 double_rq_unlock(rq
, p_rq
);
4678 if (!curr
->sched_class
->yield_to_task
)
4681 if (curr
->sched_class
!= p
->sched_class
)
4684 if (task_running(p_rq
, p
) || p
->state
)
4687 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4689 schedstat_inc(rq
, yld_count
);
4691 * Make p's CPU reschedule; pick_next_entity takes care of
4694 if (preempt
&& rq
!= p_rq
)
4699 double_rq_unlock(rq
, p_rq
);
4701 local_irq_restore(flags
);
4708 EXPORT_SYMBOL_GPL(yield_to
);
4711 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4712 * that process accounting knows that this is a task in IO wait state.
4714 long __sched
io_schedule_timeout(long timeout
)
4716 int old_iowait
= current
->in_iowait
;
4720 current
->in_iowait
= 1;
4721 blk_schedule_flush_plug(current
);
4723 delayacct_blkio_start();
4725 atomic_inc(&rq
->nr_iowait
);
4726 ret
= schedule_timeout(timeout
);
4727 current
->in_iowait
= old_iowait
;
4728 atomic_dec(&rq
->nr_iowait
);
4729 delayacct_blkio_end();
4733 EXPORT_SYMBOL(io_schedule_timeout
);
4736 * sys_sched_get_priority_max - return maximum RT priority.
4737 * @policy: scheduling class.
4739 * Return: On success, this syscall returns the maximum
4740 * rt_priority that can be used by a given scheduling class.
4741 * On failure, a negative error code is returned.
4743 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4750 ret
= MAX_USER_RT_PRIO
-1;
4752 case SCHED_DEADLINE
:
4763 * sys_sched_get_priority_min - return minimum RT priority.
4764 * @policy: scheduling class.
4766 * Return: On success, this syscall returns the minimum
4767 * rt_priority that can be used by a given scheduling class.
4768 * On failure, a negative error code is returned.
4770 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4779 case SCHED_DEADLINE
:
4789 * sys_sched_rr_get_interval - return the default timeslice of a process.
4790 * @pid: pid of the process.
4791 * @interval: userspace pointer to the timeslice value.
4793 * this syscall writes the default timeslice value of a given process
4794 * into the user-space timespec buffer. A value of '0' means infinity.
4796 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4799 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4800 struct timespec __user
*, interval
)
4802 struct task_struct
*p
;
4803 unsigned int time_slice
;
4804 unsigned long flags
;
4814 p
= find_process_by_pid(pid
);
4818 retval
= security_task_getscheduler(p
);
4822 rq
= task_rq_lock(p
, &flags
);
4824 if (p
->sched_class
->get_rr_interval
)
4825 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4826 task_rq_unlock(rq
, p
, &flags
);
4829 jiffies_to_timespec(time_slice
, &t
);
4830 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4838 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4840 void sched_show_task(struct task_struct
*p
)
4842 unsigned long free
= 0;
4844 unsigned long state
= p
->state
;
4847 state
= __ffs(state
) + 1;
4848 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4849 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4850 #if BITS_PER_LONG == 32
4851 if (state
== TASK_RUNNING
)
4852 printk(KERN_CONT
" running ");
4854 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4856 if (state
== TASK_RUNNING
)
4857 printk(KERN_CONT
" running task ");
4859 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4861 #ifdef CONFIG_DEBUG_STACK_USAGE
4862 free
= stack_not_used(p
);
4867 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4869 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4870 task_pid_nr(p
), ppid
,
4871 (unsigned long)task_thread_info(p
)->flags
);
4873 print_worker_info(KERN_INFO
, p
);
4874 show_stack(p
, NULL
);
4877 void show_state_filter(unsigned long state_filter
)
4879 struct task_struct
*g
, *p
;
4881 #if BITS_PER_LONG == 32
4883 " task PC stack pid father\n");
4886 " task PC stack pid father\n");
4889 for_each_process_thread(g
, p
) {
4891 * reset the NMI-timeout, listing all files on a slow
4892 * console might take a lot of time:
4894 touch_nmi_watchdog();
4895 if (!state_filter
|| (p
->state
& state_filter
))
4899 touch_all_softlockup_watchdogs();
4901 #ifdef CONFIG_SCHED_DEBUG
4902 sysrq_sched_debug_show();
4906 * Only show locks if all tasks are dumped:
4909 debug_show_all_locks();
4912 void init_idle_bootup_task(struct task_struct
*idle
)
4914 idle
->sched_class
= &idle_sched_class
;
4918 * init_idle - set up an idle thread for a given CPU
4919 * @idle: task in question
4920 * @cpu: cpu the idle task belongs to
4922 * NOTE: this function does not set the idle thread's NEED_RESCHED
4923 * flag, to make booting more robust.
4925 void init_idle(struct task_struct
*idle
, int cpu
)
4927 struct rq
*rq
= cpu_rq(cpu
);
4928 unsigned long flags
;
4930 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
4931 raw_spin_lock(&rq
->lock
);
4933 __sched_fork(0, idle
);
4934 idle
->state
= TASK_RUNNING
;
4935 idle
->se
.exec_start
= sched_clock();
4939 * Its possible that init_idle() gets called multiple times on a task,
4940 * in that case do_set_cpus_allowed() will not do the right thing.
4942 * And since this is boot we can forgo the serialization.
4944 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
4947 * We're having a chicken and egg problem, even though we are
4948 * holding rq->lock, the cpu isn't yet set to this cpu so the
4949 * lockdep check in task_group() will fail.
4951 * Similar case to sched_fork(). / Alternatively we could
4952 * use task_rq_lock() here and obtain the other rq->lock.
4957 __set_task_cpu(idle
, cpu
);
4960 rq
->curr
= rq
->idle
= idle
;
4961 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4965 raw_spin_unlock(&rq
->lock
);
4966 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
4968 /* Set the preempt count _outside_ the spinlocks! */
4969 init_idle_preempt_count(idle
, cpu
);
4972 * The idle tasks have their own, simple scheduling class:
4974 idle
->sched_class
= &idle_sched_class
;
4975 ftrace_graph_init_idle_task(idle
, cpu
);
4976 vtime_init_idle(idle
, cpu
);
4978 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4982 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4983 const struct cpumask
*trial
)
4985 int ret
= 1, trial_cpus
;
4986 struct dl_bw
*cur_dl_b
;
4987 unsigned long flags
;
4989 if (!cpumask_weight(cur
))
4992 rcu_read_lock_sched();
4993 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4994 trial_cpus
= cpumask_weight(trial
);
4996 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4997 if (cur_dl_b
->bw
!= -1 &&
4998 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5000 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5001 rcu_read_unlock_sched();
5006 int task_can_attach(struct task_struct
*p
,
5007 const struct cpumask
*cs_cpus_allowed
)
5012 * Kthreads which disallow setaffinity shouldn't be moved
5013 * to a new cpuset; we don't want to change their cpu
5014 * affinity and isolating such threads by their set of
5015 * allowed nodes is unnecessary. Thus, cpusets are not
5016 * applicable for such threads. This prevents checking for
5017 * success of set_cpus_allowed_ptr() on all attached tasks
5018 * before cpus_allowed may be changed.
5020 if (p
->flags
& PF_NO_SETAFFINITY
) {
5026 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5028 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5033 unsigned long flags
;
5035 rcu_read_lock_sched();
5036 dl_b
= dl_bw_of(dest_cpu
);
5037 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5038 cpus
= dl_bw_cpus(dest_cpu
);
5039 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5044 * We reserve space for this task in the destination
5045 * root_domain, as we can't fail after this point.
5046 * We will free resources in the source root_domain
5047 * later on (see set_cpus_allowed_dl()).
5049 __dl_add(dl_b
, p
->dl
.dl_bw
);
5051 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5052 rcu_read_unlock_sched();
5062 #ifdef CONFIG_NUMA_BALANCING
5063 /* Migrate current task p to target_cpu */
5064 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5066 struct migration_arg arg
= { p
, target_cpu
};
5067 int curr_cpu
= task_cpu(p
);
5069 if (curr_cpu
== target_cpu
)
5072 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5075 /* TODO: This is not properly updating schedstats */
5077 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5078 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5082 * Requeue a task on a given node and accurately track the number of NUMA
5083 * tasks on the runqueues
5085 void sched_setnuma(struct task_struct
*p
, int nid
)
5088 unsigned long flags
;
5089 bool queued
, running
;
5091 rq
= task_rq_lock(p
, &flags
);
5092 queued
= task_on_rq_queued(p
);
5093 running
= task_current(rq
, p
);
5096 dequeue_task(rq
, p
, 0);
5098 put_prev_task(rq
, p
);
5100 p
->numa_preferred_nid
= nid
;
5103 p
->sched_class
->set_curr_task(rq
);
5105 enqueue_task(rq
, p
, 0);
5106 task_rq_unlock(rq
, p
, &flags
);
5108 #endif /* CONFIG_NUMA_BALANCING */
5110 #ifdef CONFIG_HOTPLUG_CPU
5112 * Ensures that the idle task is using init_mm right before its cpu goes
5115 void idle_task_exit(void)
5117 struct mm_struct
*mm
= current
->active_mm
;
5119 BUG_ON(cpu_online(smp_processor_id()));
5121 if (mm
!= &init_mm
) {
5122 switch_mm(mm
, &init_mm
, current
);
5123 finish_arch_post_lock_switch();
5129 * Since this CPU is going 'away' for a while, fold any nr_active delta
5130 * we might have. Assumes we're called after migrate_tasks() so that the
5131 * nr_active count is stable.
5133 * Also see the comment "Global load-average calculations".
5135 static void calc_load_migrate(struct rq
*rq
)
5137 long delta
= calc_load_fold_active(rq
);
5139 atomic_long_add(delta
, &calc_load_tasks
);
5142 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5146 static const struct sched_class fake_sched_class
= {
5147 .put_prev_task
= put_prev_task_fake
,
5150 static struct task_struct fake_task
= {
5152 * Avoid pull_{rt,dl}_task()
5154 .prio
= MAX_PRIO
+ 1,
5155 .sched_class
= &fake_sched_class
,
5159 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5160 * try_to_wake_up()->select_task_rq().
5162 * Called with rq->lock held even though we'er in stop_machine() and
5163 * there's no concurrency possible, we hold the required locks anyway
5164 * because of lock validation efforts.
5166 static void migrate_tasks(struct rq
*dead_rq
)
5168 struct rq
*rq
= dead_rq
;
5169 struct task_struct
*next
, *stop
= rq
->stop
;
5173 * Fudge the rq selection such that the below task selection loop
5174 * doesn't get stuck on the currently eligible stop task.
5176 * We're currently inside stop_machine() and the rq is either stuck
5177 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5178 * either way we should never end up calling schedule() until we're
5184 * put_prev_task() and pick_next_task() sched
5185 * class method both need to have an up-to-date
5186 * value of rq->clock[_task]
5188 update_rq_clock(rq
);
5192 * There's this thread running, bail when that's the only
5195 if (rq
->nr_running
== 1)
5199 * pick_next_task assumes pinned rq->lock.
5201 lockdep_pin_lock(&rq
->lock
);
5202 next
= pick_next_task(rq
, &fake_task
);
5204 next
->sched_class
->put_prev_task(rq
, next
);
5207 * Rules for changing task_struct::cpus_allowed are holding
5208 * both pi_lock and rq->lock, such that holding either
5209 * stabilizes the mask.
5211 * Drop rq->lock is not quite as disastrous as it usually is
5212 * because !cpu_active at this point, which means load-balance
5213 * will not interfere. Also, stop-machine.
5215 lockdep_unpin_lock(&rq
->lock
);
5216 raw_spin_unlock(&rq
->lock
);
5217 raw_spin_lock(&next
->pi_lock
);
5218 raw_spin_lock(&rq
->lock
);
5221 * Since we're inside stop-machine, _nothing_ should have
5222 * changed the task, WARN if weird stuff happened, because in
5223 * that case the above rq->lock drop is a fail too.
5225 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5226 raw_spin_unlock(&next
->pi_lock
);
5230 /* Find suitable destination for @next, with force if needed. */
5231 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5233 rq
= __migrate_task(rq
, next
, dest_cpu
);
5234 if (rq
!= dead_rq
) {
5235 raw_spin_unlock(&rq
->lock
);
5237 raw_spin_lock(&rq
->lock
);
5239 raw_spin_unlock(&next
->pi_lock
);
5244 #endif /* CONFIG_HOTPLUG_CPU */
5246 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5248 static struct ctl_table sd_ctl_dir
[] = {
5250 .procname
= "sched_domain",
5256 static struct ctl_table sd_ctl_root
[] = {
5258 .procname
= "kernel",
5260 .child
= sd_ctl_dir
,
5265 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5267 struct ctl_table
*entry
=
5268 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5273 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5275 struct ctl_table
*entry
;
5278 * In the intermediate directories, both the child directory and
5279 * procname are dynamically allocated and could fail but the mode
5280 * will always be set. In the lowest directory the names are
5281 * static strings and all have proc handlers.
5283 for (entry
= *tablep
; entry
->mode
; entry
++) {
5285 sd_free_ctl_entry(&entry
->child
);
5286 if (entry
->proc_handler
== NULL
)
5287 kfree(entry
->procname
);
5294 static int min_load_idx
= 0;
5295 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5298 set_table_entry(struct ctl_table
*entry
,
5299 const char *procname
, void *data
, int maxlen
,
5300 umode_t mode
, proc_handler
*proc_handler
,
5303 entry
->procname
= procname
;
5305 entry
->maxlen
= maxlen
;
5307 entry
->proc_handler
= proc_handler
;
5310 entry
->extra1
= &min_load_idx
;
5311 entry
->extra2
= &max_load_idx
;
5315 static struct ctl_table
*
5316 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5318 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5323 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5324 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5325 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5326 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5327 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5328 sizeof(int), 0644, proc_dointvec_minmax
, true);
5329 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5330 sizeof(int), 0644, proc_dointvec_minmax
, true);
5331 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5332 sizeof(int), 0644, proc_dointvec_minmax
, true);
5333 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5334 sizeof(int), 0644, proc_dointvec_minmax
, true);
5335 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5336 sizeof(int), 0644, proc_dointvec_minmax
, true);
5337 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5338 sizeof(int), 0644, proc_dointvec_minmax
, false);
5339 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5340 sizeof(int), 0644, proc_dointvec_minmax
, false);
5341 set_table_entry(&table
[9], "cache_nice_tries",
5342 &sd
->cache_nice_tries
,
5343 sizeof(int), 0644, proc_dointvec_minmax
, false);
5344 set_table_entry(&table
[10], "flags", &sd
->flags
,
5345 sizeof(int), 0644, proc_dointvec_minmax
, false);
5346 set_table_entry(&table
[11], "max_newidle_lb_cost",
5347 &sd
->max_newidle_lb_cost
,
5348 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5349 set_table_entry(&table
[12], "name", sd
->name
,
5350 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5351 /* &table[13] is terminator */
5356 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5358 struct ctl_table
*entry
, *table
;
5359 struct sched_domain
*sd
;
5360 int domain_num
= 0, i
;
5363 for_each_domain(cpu
, sd
)
5365 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5370 for_each_domain(cpu
, sd
) {
5371 snprintf(buf
, 32, "domain%d", i
);
5372 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5374 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5381 static struct ctl_table_header
*sd_sysctl_header
;
5382 static void register_sched_domain_sysctl(void)
5384 int i
, cpu_num
= num_possible_cpus();
5385 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5388 WARN_ON(sd_ctl_dir
[0].child
);
5389 sd_ctl_dir
[0].child
= entry
;
5394 for_each_possible_cpu(i
) {
5395 snprintf(buf
, 32, "cpu%d", i
);
5396 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5398 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5402 WARN_ON(sd_sysctl_header
);
5403 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5406 /* may be called multiple times per register */
5407 static void unregister_sched_domain_sysctl(void)
5409 unregister_sysctl_table(sd_sysctl_header
);
5410 sd_sysctl_header
= NULL
;
5411 if (sd_ctl_dir
[0].child
)
5412 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5415 static void register_sched_domain_sysctl(void)
5418 static void unregister_sched_domain_sysctl(void)
5421 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5423 static void set_rq_online(struct rq
*rq
)
5426 const struct sched_class
*class;
5428 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5431 for_each_class(class) {
5432 if (class->rq_online
)
5433 class->rq_online(rq
);
5438 static void set_rq_offline(struct rq
*rq
)
5441 const struct sched_class
*class;
5443 for_each_class(class) {
5444 if (class->rq_offline
)
5445 class->rq_offline(rq
);
5448 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5454 * migration_call - callback that gets triggered when a CPU is added.
5455 * Here we can start up the necessary migration thread for the new CPU.
5458 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5460 int cpu
= (long)hcpu
;
5461 unsigned long flags
;
5462 struct rq
*rq
= cpu_rq(cpu
);
5464 switch (action
& ~CPU_TASKS_FROZEN
) {
5466 case CPU_UP_PREPARE
:
5467 rq
->calc_load_update
= calc_load_update
;
5471 /* Update our root-domain */
5472 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5474 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5478 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5481 #ifdef CONFIG_HOTPLUG_CPU
5483 sched_ttwu_pending();
5484 /* Update our root-domain */
5485 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5487 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5491 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5492 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5496 calc_load_migrate(rq
);
5501 update_max_interval();
5507 * Register at high priority so that task migration (migrate_all_tasks)
5508 * happens before everything else. This has to be lower priority than
5509 * the notifier in the perf_event subsystem, though.
5511 static struct notifier_block migration_notifier
= {
5512 .notifier_call
= migration_call
,
5513 .priority
= CPU_PRI_MIGRATION
,
5516 static void set_cpu_rq_start_time(void)
5518 int cpu
= smp_processor_id();
5519 struct rq
*rq
= cpu_rq(cpu
);
5520 rq
->age_stamp
= sched_clock_cpu(cpu
);
5523 static int sched_cpu_active(struct notifier_block
*nfb
,
5524 unsigned long action
, void *hcpu
)
5526 switch (action
& ~CPU_TASKS_FROZEN
) {
5528 set_cpu_rq_start_time();
5532 * At this point a starting CPU has marked itself as online via
5533 * set_cpu_online(). But it might not yet have marked itself
5534 * as active, which is essential from here on.
5536 * Thus, fall-through and help the starting CPU along.
5538 case CPU_DOWN_FAILED
:
5539 set_cpu_active((long)hcpu
, true);
5546 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5547 unsigned long action
, void *hcpu
)
5549 switch (action
& ~CPU_TASKS_FROZEN
) {
5550 case CPU_DOWN_PREPARE
:
5551 set_cpu_active((long)hcpu
, false);
5558 static int __init
migration_init(void)
5560 void *cpu
= (void *)(long)smp_processor_id();
5563 /* Initialize migration for the boot CPU */
5564 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5565 BUG_ON(err
== NOTIFY_BAD
);
5566 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5567 register_cpu_notifier(&migration_notifier
);
5569 /* Register cpu active notifiers */
5570 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5571 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5575 early_initcall(migration_init
);
5577 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5579 #ifdef CONFIG_SCHED_DEBUG
5581 static __read_mostly
int sched_debug_enabled
;
5583 static int __init
sched_debug_setup(char *str
)
5585 sched_debug_enabled
= 1;
5589 early_param("sched_debug", sched_debug_setup
);
5591 static inline bool sched_debug(void)
5593 return sched_debug_enabled
;
5596 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5597 struct cpumask
*groupmask
)
5599 struct sched_group
*group
= sd
->groups
;
5601 cpumask_clear(groupmask
);
5603 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5605 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5606 printk("does not load-balance\n");
5608 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5613 printk(KERN_CONT
"span %*pbl level %s\n",
5614 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5616 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5617 printk(KERN_ERR
"ERROR: domain->span does not contain "
5620 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5621 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5625 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5629 printk(KERN_ERR
"ERROR: group is NULL\n");
5633 if (!cpumask_weight(sched_group_cpus(group
))) {
5634 printk(KERN_CONT
"\n");
5635 printk(KERN_ERR
"ERROR: empty group\n");
5639 if (!(sd
->flags
& SD_OVERLAP
) &&
5640 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5641 printk(KERN_CONT
"\n");
5642 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5646 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5648 printk(KERN_CONT
" %*pbl",
5649 cpumask_pr_args(sched_group_cpus(group
)));
5650 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5651 printk(KERN_CONT
" (cpu_capacity = %d)",
5652 group
->sgc
->capacity
);
5655 group
= group
->next
;
5656 } while (group
!= sd
->groups
);
5657 printk(KERN_CONT
"\n");
5659 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5660 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5663 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5664 printk(KERN_ERR
"ERROR: parent span is not a superset "
5665 "of domain->span\n");
5669 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5673 if (!sched_debug_enabled
)
5677 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5681 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5684 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5692 #else /* !CONFIG_SCHED_DEBUG */
5693 # define sched_domain_debug(sd, cpu) do { } while (0)
5694 static inline bool sched_debug(void)
5698 #endif /* CONFIG_SCHED_DEBUG */
5700 static int sd_degenerate(struct sched_domain
*sd
)
5702 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5705 /* Following flags need at least 2 groups */
5706 if (sd
->flags
& (SD_LOAD_BALANCE
|
5707 SD_BALANCE_NEWIDLE
|
5710 SD_SHARE_CPUCAPACITY
|
5711 SD_SHARE_PKG_RESOURCES
|
5712 SD_SHARE_POWERDOMAIN
)) {
5713 if (sd
->groups
!= sd
->groups
->next
)
5717 /* Following flags don't use groups */
5718 if (sd
->flags
& (SD_WAKE_AFFINE
))
5725 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5727 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5729 if (sd_degenerate(parent
))
5732 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5735 /* Flags needing groups don't count if only 1 group in parent */
5736 if (parent
->groups
== parent
->groups
->next
) {
5737 pflags
&= ~(SD_LOAD_BALANCE
|
5738 SD_BALANCE_NEWIDLE
|
5741 SD_SHARE_CPUCAPACITY
|
5742 SD_SHARE_PKG_RESOURCES
|
5744 SD_SHARE_POWERDOMAIN
);
5745 if (nr_node_ids
== 1)
5746 pflags
&= ~SD_SERIALIZE
;
5748 if (~cflags
& pflags
)
5754 static void free_rootdomain(struct rcu_head
*rcu
)
5756 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5758 cpupri_cleanup(&rd
->cpupri
);
5759 cpudl_cleanup(&rd
->cpudl
);
5760 free_cpumask_var(rd
->dlo_mask
);
5761 free_cpumask_var(rd
->rto_mask
);
5762 free_cpumask_var(rd
->online
);
5763 free_cpumask_var(rd
->span
);
5767 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5769 struct root_domain
*old_rd
= NULL
;
5770 unsigned long flags
;
5772 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5777 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5780 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5783 * If we dont want to free the old_rd yet then
5784 * set old_rd to NULL to skip the freeing later
5787 if (!atomic_dec_and_test(&old_rd
->refcount
))
5791 atomic_inc(&rd
->refcount
);
5794 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5795 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5798 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5801 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5804 static int init_rootdomain(struct root_domain
*rd
)
5806 memset(rd
, 0, sizeof(*rd
));
5808 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5810 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5812 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5814 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5817 init_dl_bw(&rd
->dl_bw
);
5818 if (cpudl_init(&rd
->cpudl
) != 0)
5821 if (cpupri_init(&rd
->cpupri
) != 0)
5826 free_cpumask_var(rd
->rto_mask
);
5828 free_cpumask_var(rd
->dlo_mask
);
5830 free_cpumask_var(rd
->online
);
5832 free_cpumask_var(rd
->span
);
5838 * By default the system creates a single root-domain with all cpus as
5839 * members (mimicking the global state we have today).
5841 struct root_domain def_root_domain
;
5843 static void init_defrootdomain(void)
5845 init_rootdomain(&def_root_domain
);
5847 atomic_set(&def_root_domain
.refcount
, 1);
5850 static struct root_domain
*alloc_rootdomain(void)
5852 struct root_domain
*rd
;
5854 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5858 if (init_rootdomain(rd
) != 0) {
5866 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5868 struct sched_group
*tmp
, *first
;
5877 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5882 } while (sg
!= first
);
5885 static void free_sched_domain(struct rcu_head
*rcu
)
5887 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5890 * If its an overlapping domain it has private groups, iterate and
5893 if (sd
->flags
& SD_OVERLAP
) {
5894 free_sched_groups(sd
->groups
, 1);
5895 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5896 kfree(sd
->groups
->sgc
);
5902 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5904 call_rcu(&sd
->rcu
, free_sched_domain
);
5907 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5909 for (; sd
; sd
= sd
->parent
)
5910 destroy_sched_domain(sd
, cpu
);
5914 * Keep a special pointer to the highest sched_domain that has
5915 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5916 * allows us to avoid some pointer chasing select_idle_sibling().
5918 * Also keep a unique ID per domain (we use the first cpu number in
5919 * the cpumask of the domain), this allows us to quickly tell if
5920 * two cpus are in the same cache domain, see cpus_share_cache().
5922 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5923 DEFINE_PER_CPU(int, sd_llc_size
);
5924 DEFINE_PER_CPU(int, sd_llc_id
);
5925 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5926 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5927 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5929 static void update_top_cache_domain(int cpu
)
5931 struct sched_domain
*sd
;
5932 struct sched_domain
*busy_sd
= NULL
;
5936 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5938 id
= cpumask_first(sched_domain_span(sd
));
5939 size
= cpumask_weight(sched_domain_span(sd
));
5940 busy_sd
= sd
->parent
; /* sd_busy */
5942 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5944 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5945 per_cpu(sd_llc_size
, cpu
) = size
;
5946 per_cpu(sd_llc_id
, cpu
) = id
;
5948 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5949 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5951 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5952 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5956 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5957 * hold the hotplug lock.
5960 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5962 struct rq
*rq
= cpu_rq(cpu
);
5963 struct sched_domain
*tmp
;
5965 /* Remove the sched domains which do not contribute to scheduling. */
5966 for (tmp
= sd
; tmp
; ) {
5967 struct sched_domain
*parent
= tmp
->parent
;
5971 if (sd_parent_degenerate(tmp
, parent
)) {
5972 tmp
->parent
= parent
->parent
;
5974 parent
->parent
->child
= tmp
;
5976 * Transfer SD_PREFER_SIBLING down in case of a
5977 * degenerate parent; the spans match for this
5978 * so the property transfers.
5980 if (parent
->flags
& SD_PREFER_SIBLING
)
5981 tmp
->flags
|= SD_PREFER_SIBLING
;
5982 destroy_sched_domain(parent
, cpu
);
5987 if (sd
&& sd_degenerate(sd
)) {
5990 destroy_sched_domain(tmp
, cpu
);
5995 sched_domain_debug(sd
, cpu
);
5997 rq_attach_root(rq
, rd
);
5999 rcu_assign_pointer(rq
->sd
, sd
);
6000 destroy_sched_domains(tmp
, cpu
);
6002 update_top_cache_domain(cpu
);
6005 /* Setup the mask of cpus configured for isolated domains */
6006 static int __init
isolated_cpu_setup(char *str
)
6008 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6009 cpulist_parse(str
, cpu_isolated_map
);
6013 __setup("isolcpus=", isolated_cpu_setup
);
6016 struct sched_domain
** __percpu sd
;
6017 struct root_domain
*rd
;
6028 * Build an iteration mask that can exclude certain CPUs from the upwards
6031 * Asymmetric node setups can result in situations where the domain tree is of
6032 * unequal depth, make sure to skip domains that already cover the entire
6035 * In that case build_sched_domains() will have terminated the iteration early
6036 * and our sibling sd spans will be empty. Domains should always include the
6037 * cpu they're built on, so check that.
6040 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6042 const struct cpumask
*span
= sched_domain_span(sd
);
6043 struct sd_data
*sdd
= sd
->private;
6044 struct sched_domain
*sibling
;
6047 for_each_cpu(i
, span
) {
6048 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6049 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6052 cpumask_set_cpu(i
, sched_group_mask(sg
));
6057 * Return the canonical balance cpu for this group, this is the first cpu
6058 * of this group that's also in the iteration mask.
6060 int group_balance_cpu(struct sched_group
*sg
)
6062 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6066 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6068 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6069 const struct cpumask
*span
= sched_domain_span(sd
);
6070 struct cpumask
*covered
= sched_domains_tmpmask
;
6071 struct sd_data
*sdd
= sd
->private;
6072 struct sched_domain
*sibling
;
6075 cpumask_clear(covered
);
6077 for_each_cpu(i
, span
) {
6078 struct cpumask
*sg_span
;
6080 if (cpumask_test_cpu(i
, covered
))
6083 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6085 /* See the comment near build_group_mask(). */
6086 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6089 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6090 GFP_KERNEL
, cpu_to_node(cpu
));
6095 sg_span
= sched_group_cpus(sg
);
6097 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6099 cpumask_set_cpu(i
, sg_span
);
6101 cpumask_or(covered
, covered
, sg_span
);
6103 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6104 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6105 build_group_mask(sd
, sg
);
6108 * Initialize sgc->capacity such that even if we mess up the
6109 * domains and no possible iteration will get us here, we won't
6112 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6115 * Make sure the first group of this domain contains the
6116 * canonical balance cpu. Otherwise the sched_domain iteration
6117 * breaks. See update_sg_lb_stats().
6119 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6120 group_balance_cpu(sg
) == cpu
)
6130 sd
->groups
= groups
;
6135 free_sched_groups(first
, 0);
6140 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6142 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6143 struct sched_domain
*child
= sd
->child
;
6146 cpu
= cpumask_first(sched_domain_span(child
));
6149 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6150 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6151 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6158 * build_sched_groups will build a circular linked list of the groups
6159 * covered by the given span, and will set each group's ->cpumask correctly,
6160 * and ->cpu_capacity to 0.
6162 * Assumes the sched_domain tree is fully constructed
6165 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6167 struct sched_group
*first
= NULL
, *last
= NULL
;
6168 struct sd_data
*sdd
= sd
->private;
6169 const struct cpumask
*span
= sched_domain_span(sd
);
6170 struct cpumask
*covered
;
6173 get_group(cpu
, sdd
, &sd
->groups
);
6174 atomic_inc(&sd
->groups
->ref
);
6176 if (cpu
!= cpumask_first(span
))
6179 lockdep_assert_held(&sched_domains_mutex
);
6180 covered
= sched_domains_tmpmask
;
6182 cpumask_clear(covered
);
6184 for_each_cpu(i
, span
) {
6185 struct sched_group
*sg
;
6188 if (cpumask_test_cpu(i
, covered
))
6191 group
= get_group(i
, sdd
, &sg
);
6192 cpumask_setall(sched_group_mask(sg
));
6194 for_each_cpu(j
, span
) {
6195 if (get_group(j
, sdd
, NULL
) != group
)
6198 cpumask_set_cpu(j
, covered
);
6199 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6214 * Initialize sched groups cpu_capacity.
6216 * cpu_capacity indicates the capacity of sched group, which is used while
6217 * distributing the load between different sched groups in a sched domain.
6218 * Typically cpu_capacity for all the groups in a sched domain will be same
6219 * unless there are asymmetries in the topology. If there are asymmetries,
6220 * group having more cpu_capacity will pickup more load compared to the
6221 * group having less cpu_capacity.
6223 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6225 struct sched_group
*sg
= sd
->groups
;
6230 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6232 } while (sg
!= sd
->groups
);
6234 if (cpu
!= group_balance_cpu(sg
))
6237 update_group_capacity(sd
, cpu
);
6238 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6242 * Initializers for schedule domains
6243 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6246 static int default_relax_domain_level
= -1;
6247 int sched_domain_level_max
;
6249 static int __init
setup_relax_domain_level(char *str
)
6251 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6252 pr_warn("Unable to set relax_domain_level\n");
6256 __setup("relax_domain_level=", setup_relax_domain_level
);
6258 static void set_domain_attribute(struct sched_domain
*sd
,
6259 struct sched_domain_attr
*attr
)
6263 if (!attr
|| attr
->relax_domain_level
< 0) {
6264 if (default_relax_domain_level
< 0)
6267 request
= default_relax_domain_level
;
6269 request
= attr
->relax_domain_level
;
6270 if (request
< sd
->level
) {
6271 /* turn off idle balance on this domain */
6272 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6274 /* turn on idle balance on this domain */
6275 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6279 static void __sdt_free(const struct cpumask
*cpu_map
);
6280 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6282 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6283 const struct cpumask
*cpu_map
)
6287 if (!atomic_read(&d
->rd
->refcount
))
6288 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6290 free_percpu(d
->sd
); /* fall through */
6292 __sdt_free(cpu_map
); /* fall through */
6298 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6299 const struct cpumask
*cpu_map
)
6301 memset(d
, 0, sizeof(*d
));
6303 if (__sdt_alloc(cpu_map
))
6304 return sa_sd_storage
;
6305 d
->sd
= alloc_percpu(struct sched_domain
*);
6307 return sa_sd_storage
;
6308 d
->rd
= alloc_rootdomain();
6311 return sa_rootdomain
;
6315 * NULL the sd_data elements we've used to build the sched_domain and
6316 * sched_group structure so that the subsequent __free_domain_allocs()
6317 * will not free the data we're using.
6319 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6321 struct sd_data
*sdd
= sd
->private;
6323 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6324 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6326 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6327 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6329 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6330 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6334 static int sched_domains_numa_levels
;
6335 enum numa_topology_type sched_numa_topology_type
;
6336 static int *sched_domains_numa_distance
;
6337 int sched_max_numa_distance
;
6338 static struct cpumask
***sched_domains_numa_masks
;
6339 static int sched_domains_curr_level
;
6343 * SD_flags allowed in topology descriptions.
6345 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6346 * SD_SHARE_PKG_RESOURCES - describes shared caches
6347 * SD_NUMA - describes NUMA topologies
6348 * SD_SHARE_POWERDOMAIN - describes shared power domain
6351 * SD_ASYM_PACKING - describes SMT quirks
6353 #define TOPOLOGY_SD_FLAGS \
6354 (SD_SHARE_CPUCAPACITY | \
6355 SD_SHARE_PKG_RESOURCES | \
6358 SD_SHARE_POWERDOMAIN)
6360 static struct sched_domain
*
6361 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6363 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6364 int sd_weight
, sd_flags
= 0;
6368 * Ugly hack to pass state to sd_numa_mask()...
6370 sched_domains_curr_level
= tl
->numa_level
;
6373 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6376 sd_flags
= (*tl
->sd_flags
)();
6377 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6378 "wrong sd_flags in topology description\n"))
6379 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6381 *sd
= (struct sched_domain
){
6382 .min_interval
= sd_weight
,
6383 .max_interval
= 2*sd_weight
,
6385 .imbalance_pct
= 125,
6387 .cache_nice_tries
= 0,
6394 .flags
= 1*SD_LOAD_BALANCE
6395 | 1*SD_BALANCE_NEWIDLE
6400 | 0*SD_SHARE_CPUCAPACITY
6401 | 0*SD_SHARE_PKG_RESOURCES
6403 | 0*SD_PREFER_SIBLING
6408 .last_balance
= jiffies
,
6409 .balance_interval
= sd_weight
,
6411 .max_newidle_lb_cost
= 0,
6412 .next_decay_max_lb_cost
= jiffies
,
6413 #ifdef CONFIG_SCHED_DEBUG
6419 * Convert topological properties into behaviour.
6422 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6423 sd
->flags
|= SD_PREFER_SIBLING
;
6424 sd
->imbalance_pct
= 110;
6425 sd
->smt_gain
= 1178; /* ~15% */
6427 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6428 sd
->imbalance_pct
= 117;
6429 sd
->cache_nice_tries
= 1;
6433 } else if (sd
->flags
& SD_NUMA
) {
6434 sd
->cache_nice_tries
= 2;
6438 sd
->flags
|= SD_SERIALIZE
;
6439 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6440 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6447 sd
->flags
|= SD_PREFER_SIBLING
;
6448 sd
->cache_nice_tries
= 1;
6453 sd
->private = &tl
->data
;
6459 * Topology list, bottom-up.
6461 static struct sched_domain_topology_level default_topology
[] = {
6462 #ifdef CONFIG_SCHED_SMT
6463 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6465 #ifdef CONFIG_SCHED_MC
6466 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6468 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6472 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6474 #define for_each_sd_topology(tl) \
6475 for (tl = sched_domain_topology; tl->mask; tl++)
6477 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6479 sched_domain_topology
= tl
;
6484 static const struct cpumask
*sd_numa_mask(int cpu
)
6486 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6489 static void sched_numa_warn(const char *str
)
6491 static int done
= false;
6499 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6501 for (i
= 0; i
< nr_node_ids
; i
++) {
6502 printk(KERN_WARNING
" ");
6503 for (j
= 0; j
< nr_node_ids
; j
++)
6504 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6505 printk(KERN_CONT
"\n");
6507 printk(KERN_WARNING
"\n");
6510 bool find_numa_distance(int distance
)
6514 if (distance
== node_distance(0, 0))
6517 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6518 if (sched_domains_numa_distance
[i
] == distance
)
6526 * A system can have three types of NUMA topology:
6527 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6528 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6529 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6531 * The difference between a glueless mesh topology and a backplane
6532 * topology lies in whether communication between not directly
6533 * connected nodes goes through intermediary nodes (where programs
6534 * could run), or through backplane controllers. This affects
6535 * placement of programs.
6537 * The type of topology can be discerned with the following tests:
6538 * - If the maximum distance between any nodes is 1 hop, the system
6539 * is directly connected.
6540 * - If for two nodes A and B, located N > 1 hops away from each other,
6541 * there is an intermediary node C, which is < N hops away from both
6542 * nodes A and B, the system is a glueless mesh.
6544 static void init_numa_topology_type(void)
6548 n
= sched_max_numa_distance
;
6550 if (sched_domains_numa_levels
<= 1) {
6551 sched_numa_topology_type
= NUMA_DIRECT
;
6555 for_each_online_node(a
) {
6556 for_each_online_node(b
) {
6557 /* Find two nodes furthest removed from each other. */
6558 if (node_distance(a
, b
) < n
)
6561 /* Is there an intermediary node between a and b? */
6562 for_each_online_node(c
) {
6563 if (node_distance(a
, c
) < n
&&
6564 node_distance(b
, c
) < n
) {
6565 sched_numa_topology_type
=
6571 sched_numa_topology_type
= NUMA_BACKPLANE
;
6577 static void sched_init_numa(void)
6579 int next_distance
, curr_distance
= node_distance(0, 0);
6580 struct sched_domain_topology_level
*tl
;
6584 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6585 if (!sched_domains_numa_distance
)
6589 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6590 * unique distances in the node_distance() table.
6592 * Assumes node_distance(0,j) includes all distances in
6593 * node_distance(i,j) in order to avoid cubic time.
6595 next_distance
= curr_distance
;
6596 for (i
= 0; i
< nr_node_ids
; i
++) {
6597 for (j
= 0; j
< nr_node_ids
; j
++) {
6598 for (k
= 0; k
< nr_node_ids
; k
++) {
6599 int distance
= node_distance(i
, k
);
6601 if (distance
> curr_distance
&&
6602 (distance
< next_distance
||
6603 next_distance
== curr_distance
))
6604 next_distance
= distance
;
6607 * While not a strong assumption it would be nice to know
6608 * about cases where if node A is connected to B, B is not
6609 * equally connected to A.
6611 if (sched_debug() && node_distance(k
, i
) != distance
)
6612 sched_numa_warn("Node-distance not symmetric");
6614 if (sched_debug() && i
&& !find_numa_distance(distance
))
6615 sched_numa_warn("Node-0 not representative");
6617 if (next_distance
!= curr_distance
) {
6618 sched_domains_numa_distance
[level
++] = next_distance
;
6619 sched_domains_numa_levels
= level
;
6620 curr_distance
= next_distance
;
6625 * In case of sched_debug() we verify the above assumption.
6635 * 'level' contains the number of unique distances, excluding the
6636 * identity distance node_distance(i,i).
6638 * The sched_domains_numa_distance[] array includes the actual distance
6643 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6644 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6645 * the array will contain less then 'level' members. This could be
6646 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6647 * in other functions.
6649 * We reset it to 'level' at the end of this function.
6651 sched_domains_numa_levels
= 0;
6653 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6654 if (!sched_domains_numa_masks
)
6658 * Now for each level, construct a mask per node which contains all
6659 * cpus of nodes that are that many hops away from us.
6661 for (i
= 0; i
< level
; i
++) {
6662 sched_domains_numa_masks
[i
] =
6663 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6664 if (!sched_domains_numa_masks
[i
])
6667 for (j
= 0; j
< nr_node_ids
; j
++) {
6668 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6672 sched_domains_numa_masks
[i
][j
] = mask
;
6674 for (k
= 0; k
< nr_node_ids
; k
++) {
6675 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6678 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6683 /* Compute default topology size */
6684 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6686 tl
= kzalloc((i
+ level
+ 1) *
6687 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6692 * Copy the default topology bits..
6694 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6695 tl
[i
] = sched_domain_topology
[i
];
6698 * .. and append 'j' levels of NUMA goodness.
6700 for (j
= 0; j
< level
; i
++, j
++) {
6701 tl
[i
] = (struct sched_domain_topology_level
){
6702 .mask
= sd_numa_mask
,
6703 .sd_flags
= cpu_numa_flags
,
6704 .flags
= SDTL_OVERLAP
,
6710 sched_domain_topology
= tl
;
6712 sched_domains_numa_levels
= level
;
6713 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6715 init_numa_topology_type();
6718 static void sched_domains_numa_masks_set(int cpu
)
6721 int node
= cpu_to_node(cpu
);
6723 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6724 for (j
= 0; j
< nr_node_ids
; j
++) {
6725 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6726 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6731 static void sched_domains_numa_masks_clear(int cpu
)
6734 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6735 for (j
= 0; j
< nr_node_ids
; j
++)
6736 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6741 * Update sched_domains_numa_masks[level][node] array when new cpus
6744 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6745 unsigned long action
,
6748 int cpu
= (long)hcpu
;
6750 switch (action
& ~CPU_TASKS_FROZEN
) {
6752 sched_domains_numa_masks_set(cpu
);
6756 sched_domains_numa_masks_clear(cpu
);
6766 static inline void sched_init_numa(void)
6770 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6771 unsigned long action
,
6776 #endif /* CONFIG_NUMA */
6778 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6780 struct sched_domain_topology_level
*tl
;
6783 for_each_sd_topology(tl
) {
6784 struct sd_data
*sdd
= &tl
->data
;
6786 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6790 sdd
->sg
= alloc_percpu(struct sched_group
*);
6794 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6798 for_each_cpu(j
, cpu_map
) {
6799 struct sched_domain
*sd
;
6800 struct sched_group
*sg
;
6801 struct sched_group_capacity
*sgc
;
6803 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6804 GFP_KERNEL
, cpu_to_node(j
));
6808 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6810 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6811 GFP_KERNEL
, cpu_to_node(j
));
6817 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6819 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6820 GFP_KERNEL
, cpu_to_node(j
));
6824 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6831 static void __sdt_free(const struct cpumask
*cpu_map
)
6833 struct sched_domain_topology_level
*tl
;
6836 for_each_sd_topology(tl
) {
6837 struct sd_data
*sdd
= &tl
->data
;
6839 for_each_cpu(j
, cpu_map
) {
6840 struct sched_domain
*sd
;
6843 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6844 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6845 free_sched_groups(sd
->groups
, 0);
6846 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6850 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6852 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6854 free_percpu(sdd
->sd
);
6856 free_percpu(sdd
->sg
);
6858 free_percpu(sdd
->sgc
);
6863 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6864 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6865 struct sched_domain
*child
, int cpu
)
6867 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6871 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6873 sd
->level
= child
->level
+ 1;
6874 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6878 if (!cpumask_subset(sched_domain_span(child
),
6879 sched_domain_span(sd
))) {
6880 pr_err("BUG: arch topology borken\n");
6881 #ifdef CONFIG_SCHED_DEBUG
6882 pr_err(" the %s domain not a subset of the %s domain\n",
6883 child
->name
, sd
->name
);
6885 /* Fixup, ensure @sd has at least @child cpus. */
6886 cpumask_or(sched_domain_span(sd
),
6887 sched_domain_span(sd
),
6888 sched_domain_span(child
));
6892 set_domain_attribute(sd
, attr
);
6898 * Build sched domains for a given set of cpus and attach the sched domains
6899 * to the individual cpus
6901 static int build_sched_domains(const struct cpumask
*cpu_map
,
6902 struct sched_domain_attr
*attr
)
6904 enum s_alloc alloc_state
;
6905 struct sched_domain
*sd
;
6907 int i
, ret
= -ENOMEM
;
6909 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6910 if (alloc_state
!= sa_rootdomain
)
6913 /* Set up domains for cpus specified by the cpu_map. */
6914 for_each_cpu(i
, cpu_map
) {
6915 struct sched_domain_topology_level
*tl
;
6918 for_each_sd_topology(tl
) {
6919 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6920 if (tl
== sched_domain_topology
)
6921 *per_cpu_ptr(d
.sd
, i
) = sd
;
6922 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6923 sd
->flags
|= SD_OVERLAP
;
6924 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6929 /* Build the groups for the domains */
6930 for_each_cpu(i
, cpu_map
) {
6931 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6932 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6933 if (sd
->flags
& SD_OVERLAP
) {
6934 if (build_overlap_sched_groups(sd
, i
))
6937 if (build_sched_groups(sd
, i
))
6943 /* Calculate CPU capacity for physical packages and nodes */
6944 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6945 if (!cpumask_test_cpu(i
, cpu_map
))
6948 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6949 claim_allocations(i
, sd
);
6950 init_sched_groups_capacity(i
, sd
);
6954 /* Attach the domains */
6956 for_each_cpu(i
, cpu_map
) {
6957 sd
= *per_cpu_ptr(d
.sd
, i
);
6958 cpu_attach_domain(sd
, d
.rd
, i
);
6964 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6968 static cpumask_var_t
*doms_cur
; /* current sched domains */
6969 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6970 static struct sched_domain_attr
*dattr_cur
;
6971 /* attribues of custom domains in 'doms_cur' */
6974 * Special case: If a kmalloc of a doms_cur partition (array of
6975 * cpumask) fails, then fallback to a single sched domain,
6976 * as determined by the single cpumask fallback_doms.
6978 static cpumask_var_t fallback_doms
;
6981 * arch_update_cpu_topology lets virtualized architectures update the
6982 * cpu core maps. It is supposed to return 1 if the topology changed
6983 * or 0 if it stayed the same.
6985 int __weak
arch_update_cpu_topology(void)
6990 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6993 cpumask_var_t
*doms
;
6995 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6998 for (i
= 0; i
< ndoms
; i
++) {
6999 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7000 free_sched_domains(doms
, i
);
7007 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7010 for (i
= 0; i
< ndoms
; i
++)
7011 free_cpumask_var(doms
[i
]);
7016 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7017 * For now this just excludes isolated cpus, but could be used to
7018 * exclude other special cases in the future.
7020 static int init_sched_domains(const struct cpumask
*cpu_map
)
7024 arch_update_cpu_topology();
7026 doms_cur
= alloc_sched_domains(ndoms_cur
);
7028 doms_cur
= &fallback_doms
;
7029 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7030 err
= build_sched_domains(doms_cur
[0], NULL
);
7031 register_sched_domain_sysctl();
7037 * Detach sched domains from a group of cpus specified in cpu_map
7038 * These cpus will now be attached to the NULL domain
7040 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7045 for_each_cpu(i
, cpu_map
)
7046 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7050 /* handle null as "default" */
7051 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7052 struct sched_domain_attr
*new, int idx_new
)
7054 struct sched_domain_attr tmp
;
7061 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7062 new ? (new + idx_new
) : &tmp
,
7063 sizeof(struct sched_domain_attr
));
7067 * Partition sched domains as specified by the 'ndoms_new'
7068 * cpumasks in the array doms_new[] of cpumasks. This compares
7069 * doms_new[] to the current sched domain partitioning, doms_cur[].
7070 * It destroys each deleted domain and builds each new domain.
7072 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7073 * The masks don't intersect (don't overlap.) We should setup one
7074 * sched domain for each mask. CPUs not in any of the cpumasks will
7075 * not be load balanced. If the same cpumask appears both in the
7076 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7079 * The passed in 'doms_new' should be allocated using
7080 * alloc_sched_domains. This routine takes ownership of it and will
7081 * free_sched_domains it when done with it. If the caller failed the
7082 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7083 * and partition_sched_domains() will fallback to the single partition
7084 * 'fallback_doms', it also forces the domains to be rebuilt.
7086 * If doms_new == NULL it will be replaced with cpu_online_mask.
7087 * ndoms_new == 0 is a special case for destroying existing domains,
7088 * and it will not create the default domain.
7090 * Call with hotplug lock held
7092 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7093 struct sched_domain_attr
*dattr_new
)
7098 mutex_lock(&sched_domains_mutex
);
7100 /* always unregister in case we don't destroy any domains */
7101 unregister_sched_domain_sysctl();
7103 /* Let architecture update cpu core mappings. */
7104 new_topology
= arch_update_cpu_topology();
7106 n
= doms_new
? ndoms_new
: 0;
7108 /* Destroy deleted domains */
7109 for (i
= 0; i
< ndoms_cur
; i
++) {
7110 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7111 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7112 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7115 /* no match - a current sched domain not in new doms_new[] */
7116 detach_destroy_domains(doms_cur
[i
]);
7122 if (doms_new
== NULL
) {
7124 doms_new
= &fallback_doms
;
7125 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7126 WARN_ON_ONCE(dattr_new
);
7129 /* Build new domains */
7130 for (i
= 0; i
< ndoms_new
; i
++) {
7131 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7132 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7133 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7136 /* no match - add a new doms_new */
7137 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7142 /* Remember the new sched domains */
7143 if (doms_cur
!= &fallback_doms
)
7144 free_sched_domains(doms_cur
, ndoms_cur
);
7145 kfree(dattr_cur
); /* kfree(NULL) is safe */
7146 doms_cur
= doms_new
;
7147 dattr_cur
= dattr_new
;
7148 ndoms_cur
= ndoms_new
;
7150 register_sched_domain_sysctl();
7152 mutex_unlock(&sched_domains_mutex
);
7155 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7158 * Update cpusets according to cpu_active mask. If cpusets are
7159 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7160 * around partition_sched_domains().
7162 * If we come here as part of a suspend/resume, don't touch cpusets because we
7163 * want to restore it back to its original state upon resume anyway.
7165 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7169 case CPU_ONLINE_FROZEN
:
7170 case CPU_DOWN_FAILED_FROZEN
:
7173 * num_cpus_frozen tracks how many CPUs are involved in suspend
7174 * resume sequence. As long as this is not the last online
7175 * operation in the resume sequence, just build a single sched
7176 * domain, ignoring cpusets.
7179 if (likely(num_cpus_frozen
)) {
7180 partition_sched_domains(1, NULL
, NULL
);
7185 * This is the last CPU online operation. So fall through and
7186 * restore the original sched domains by considering the
7187 * cpuset configurations.
7191 cpuset_update_active_cpus(true);
7199 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7202 unsigned long flags
;
7203 long cpu
= (long)hcpu
;
7209 case CPU_DOWN_PREPARE
:
7210 rcu_read_lock_sched();
7211 dl_b
= dl_bw_of(cpu
);
7213 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7214 cpus
= dl_bw_cpus(cpu
);
7215 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7216 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7218 rcu_read_unlock_sched();
7221 return notifier_from_errno(-EBUSY
);
7222 cpuset_update_active_cpus(false);
7224 case CPU_DOWN_PREPARE_FROZEN
:
7226 partition_sched_domains(1, NULL
, NULL
);
7234 void __init
sched_init_smp(void)
7236 cpumask_var_t non_isolated_cpus
;
7238 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7239 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7244 * There's no userspace yet to cause hotplug operations; hence all the
7245 * cpu masks are stable and all blatant races in the below code cannot
7248 mutex_lock(&sched_domains_mutex
);
7249 init_sched_domains(cpu_active_mask
);
7250 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7251 if (cpumask_empty(non_isolated_cpus
))
7252 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7253 mutex_unlock(&sched_domains_mutex
);
7255 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7256 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7257 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7261 /* Move init over to a non-isolated CPU */
7262 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7264 sched_init_granularity();
7265 free_cpumask_var(non_isolated_cpus
);
7267 init_sched_rt_class();
7268 init_sched_dl_class();
7271 void __init
sched_init_smp(void)
7273 sched_init_granularity();
7275 #endif /* CONFIG_SMP */
7277 int in_sched_functions(unsigned long addr
)
7279 return in_lock_functions(addr
) ||
7280 (addr
>= (unsigned long)__sched_text_start
7281 && addr
< (unsigned long)__sched_text_end
);
7284 #ifdef CONFIG_CGROUP_SCHED
7286 * Default task group.
7287 * Every task in system belongs to this group at bootup.
7289 struct task_group root_task_group
;
7290 LIST_HEAD(task_groups
);
7293 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7295 void __init
sched_init(void)
7298 unsigned long alloc_size
= 0, ptr
;
7300 #ifdef CONFIG_FAIR_GROUP_SCHED
7301 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7303 #ifdef CONFIG_RT_GROUP_SCHED
7304 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7307 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7309 #ifdef CONFIG_FAIR_GROUP_SCHED
7310 root_task_group
.se
= (struct sched_entity
**)ptr
;
7311 ptr
+= nr_cpu_ids
* sizeof(void **);
7313 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7314 ptr
+= nr_cpu_ids
* sizeof(void **);
7316 #endif /* CONFIG_FAIR_GROUP_SCHED */
7317 #ifdef CONFIG_RT_GROUP_SCHED
7318 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7319 ptr
+= nr_cpu_ids
* sizeof(void **);
7321 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7322 ptr
+= nr_cpu_ids
* sizeof(void **);
7324 #endif /* CONFIG_RT_GROUP_SCHED */
7326 #ifdef CONFIG_CPUMASK_OFFSTACK
7327 for_each_possible_cpu(i
) {
7328 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7329 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7331 #endif /* CONFIG_CPUMASK_OFFSTACK */
7333 init_rt_bandwidth(&def_rt_bandwidth
,
7334 global_rt_period(), global_rt_runtime());
7335 init_dl_bandwidth(&def_dl_bandwidth
,
7336 global_rt_period(), global_rt_runtime());
7339 init_defrootdomain();
7342 #ifdef CONFIG_RT_GROUP_SCHED
7343 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7344 global_rt_period(), global_rt_runtime());
7345 #endif /* CONFIG_RT_GROUP_SCHED */
7347 #ifdef CONFIG_CGROUP_SCHED
7348 list_add(&root_task_group
.list
, &task_groups
);
7349 INIT_LIST_HEAD(&root_task_group
.children
);
7350 INIT_LIST_HEAD(&root_task_group
.siblings
);
7351 autogroup_init(&init_task
);
7353 #endif /* CONFIG_CGROUP_SCHED */
7355 for_each_possible_cpu(i
) {
7359 raw_spin_lock_init(&rq
->lock
);
7361 rq
->calc_load_active
= 0;
7362 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7363 init_cfs_rq(&rq
->cfs
);
7364 init_rt_rq(&rq
->rt
);
7365 init_dl_rq(&rq
->dl
);
7366 #ifdef CONFIG_FAIR_GROUP_SCHED
7367 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7368 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7370 * How much cpu bandwidth does root_task_group get?
7372 * In case of task-groups formed thr' the cgroup filesystem, it
7373 * gets 100% of the cpu resources in the system. This overall
7374 * system cpu resource is divided among the tasks of
7375 * root_task_group and its child task-groups in a fair manner,
7376 * based on each entity's (task or task-group's) weight
7377 * (se->load.weight).
7379 * In other words, if root_task_group has 10 tasks of weight
7380 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7381 * then A0's share of the cpu resource is:
7383 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7385 * We achieve this by letting root_task_group's tasks sit
7386 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7388 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7389 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7390 #endif /* CONFIG_FAIR_GROUP_SCHED */
7392 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7393 #ifdef CONFIG_RT_GROUP_SCHED
7394 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7397 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7398 rq
->cpu_load
[j
] = 0;
7400 rq
->last_load_update_tick
= jiffies
;
7405 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7406 rq
->balance_callback
= NULL
;
7407 rq
->active_balance
= 0;
7408 rq
->next_balance
= jiffies
;
7413 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7414 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7416 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7418 rq_attach_root(rq
, &def_root_domain
);
7419 #ifdef CONFIG_NO_HZ_COMMON
7422 #ifdef CONFIG_NO_HZ_FULL
7423 rq
->last_sched_tick
= 0;
7427 atomic_set(&rq
->nr_iowait
, 0);
7430 set_load_weight(&init_task
);
7432 #ifdef CONFIG_PREEMPT_NOTIFIERS
7433 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7437 * The boot idle thread does lazy MMU switching as well:
7439 atomic_inc(&init_mm
.mm_count
);
7440 enter_lazy_tlb(&init_mm
, current
);
7443 * During early bootup we pretend to be a normal task:
7445 current
->sched_class
= &fair_sched_class
;
7448 * Make us the idle thread. Technically, schedule() should not be
7449 * called from this thread, however somewhere below it might be,
7450 * but because we are the idle thread, we just pick up running again
7451 * when this runqueue becomes "idle".
7453 init_idle(current
, smp_processor_id());
7455 calc_load_update
= jiffies
+ LOAD_FREQ
;
7458 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7459 /* May be allocated at isolcpus cmdline parse time */
7460 if (cpu_isolated_map
== NULL
)
7461 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7462 idle_thread_set_boot_cpu();
7463 set_cpu_rq_start_time();
7465 init_sched_fair_class();
7467 scheduler_running
= 1;
7470 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7471 static inline int preempt_count_equals(int preempt_offset
)
7473 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7475 return (nested
== preempt_offset
);
7478 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7481 * Blocking primitives will set (and therefore destroy) current->state,
7482 * since we will exit with TASK_RUNNING make sure we enter with it,
7483 * otherwise we will destroy state.
7485 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7486 "do not call blocking ops when !TASK_RUNNING; "
7487 "state=%lx set at [<%p>] %pS\n",
7489 (void *)current
->task_state_change
,
7490 (void *)current
->task_state_change
);
7492 ___might_sleep(file
, line
, preempt_offset
);
7494 EXPORT_SYMBOL(__might_sleep
);
7496 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7498 static unsigned long prev_jiffy
; /* ratelimiting */
7500 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7501 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7502 !is_idle_task(current
)) ||
7503 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7505 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7507 prev_jiffy
= jiffies
;
7510 "BUG: sleeping function called from invalid context at %s:%d\n",
7513 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7514 in_atomic(), irqs_disabled(),
7515 current
->pid
, current
->comm
);
7517 if (task_stack_end_corrupted(current
))
7518 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7520 debug_show_held_locks(current
);
7521 if (irqs_disabled())
7522 print_irqtrace_events(current
);
7523 #ifdef CONFIG_DEBUG_PREEMPT
7524 if (!preempt_count_equals(preempt_offset
)) {
7525 pr_err("Preemption disabled at:");
7526 print_ip_sym(current
->preempt_disable_ip
);
7532 EXPORT_SYMBOL(___might_sleep
);
7535 #ifdef CONFIG_MAGIC_SYSRQ
7536 void normalize_rt_tasks(void)
7538 struct task_struct
*g
, *p
;
7539 struct sched_attr attr
= {
7540 .sched_policy
= SCHED_NORMAL
,
7543 read_lock(&tasklist_lock
);
7544 for_each_process_thread(g
, p
) {
7546 * Only normalize user tasks:
7548 if (p
->flags
& PF_KTHREAD
)
7551 p
->se
.exec_start
= 0;
7552 #ifdef CONFIG_SCHEDSTATS
7553 p
->se
.statistics
.wait_start
= 0;
7554 p
->se
.statistics
.sleep_start
= 0;
7555 p
->se
.statistics
.block_start
= 0;
7558 if (!dl_task(p
) && !rt_task(p
)) {
7560 * Renice negative nice level userspace
7563 if (task_nice(p
) < 0)
7564 set_user_nice(p
, 0);
7568 __sched_setscheduler(p
, &attr
, false, false);
7570 read_unlock(&tasklist_lock
);
7573 #endif /* CONFIG_MAGIC_SYSRQ */
7575 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7577 * These functions are only useful for the IA64 MCA handling, or kdb.
7579 * They can only be called when the whole system has been
7580 * stopped - every CPU needs to be quiescent, and no scheduling
7581 * activity can take place. Using them for anything else would
7582 * be a serious bug, and as a result, they aren't even visible
7583 * under any other configuration.
7587 * curr_task - return the current task for a given cpu.
7588 * @cpu: the processor in question.
7590 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7592 * Return: The current task for @cpu.
7594 struct task_struct
*curr_task(int cpu
)
7596 return cpu_curr(cpu
);
7599 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7603 * set_curr_task - set the current task for a given cpu.
7604 * @cpu: the processor in question.
7605 * @p: the task pointer to set.
7607 * Description: This function must only be used when non-maskable interrupts
7608 * are serviced on a separate stack. It allows the architecture to switch the
7609 * notion of the current task on a cpu in a non-blocking manner. This function
7610 * must be called with all CPU's synchronized, and interrupts disabled, the
7611 * and caller must save the original value of the current task (see
7612 * curr_task() above) and restore that value before reenabling interrupts and
7613 * re-starting the system.
7615 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7617 void set_curr_task(int cpu
, struct task_struct
*p
)
7624 #ifdef CONFIG_CGROUP_SCHED
7625 /* task_group_lock serializes the addition/removal of task groups */
7626 static DEFINE_SPINLOCK(task_group_lock
);
7628 static void free_sched_group(struct task_group
*tg
)
7630 free_fair_sched_group(tg
);
7631 free_rt_sched_group(tg
);
7636 /* allocate runqueue etc for a new task group */
7637 struct task_group
*sched_create_group(struct task_group
*parent
)
7639 struct task_group
*tg
;
7641 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7643 return ERR_PTR(-ENOMEM
);
7645 if (!alloc_fair_sched_group(tg
, parent
))
7648 if (!alloc_rt_sched_group(tg
, parent
))
7654 free_sched_group(tg
);
7655 return ERR_PTR(-ENOMEM
);
7658 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7660 unsigned long flags
;
7662 spin_lock_irqsave(&task_group_lock
, flags
);
7663 list_add_rcu(&tg
->list
, &task_groups
);
7665 WARN_ON(!parent
); /* root should already exist */
7667 tg
->parent
= parent
;
7668 INIT_LIST_HEAD(&tg
->children
);
7669 list_add_rcu(&tg
->siblings
, &parent
->children
);
7670 spin_unlock_irqrestore(&task_group_lock
, flags
);
7673 /* rcu callback to free various structures associated with a task group */
7674 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7676 /* now it should be safe to free those cfs_rqs */
7677 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7680 /* Destroy runqueue etc associated with a task group */
7681 void sched_destroy_group(struct task_group
*tg
)
7683 /* wait for possible concurrent references to cfs_rqs complete */
7684 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7687 void sched_offline_group(struct task_group
*tg
)
7689 unsigned long flags
;
7692 /* end participation in shares distribution */
7693 for_each_possible_cpu(i
)
7694 unregister_fair_sched_group(tg
, i
);
7696 spin_lock_irqsave(&task_group_lock
, flags
);
7697 list_del_rcu(&tg
->list
);
7698 list_del_rcu(&tg
->siblings
);
7699 spin_unlock_irqrestore(&task_group_lock
, flags
);
7702 /* change task's runqueue when it moves between groups.
7703 * The caller of this function should have put the task in its new group
7704 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7705 * reflect its new group.
7707 void sched_move_task(struct task_struct
*tsk
)
7709 struct task_group
*tg
;
7710 int queued
, running
;
7711 unsigned long flags
;
7714 rq
= task_rq_lock(tsk
, &flags
);
7716 running
= task_current(rq
, tsk
);
7717 queued
= task_on_rq_queued(tsk
);
7720 dequeue_task(rq
, tsk
, 0);
7721 if (unlikely(running
))
7722 put_prev_task(rq
, tsk
);
7725 * All callers are synchronized by task_rq_lock(); we do not use RCU
7726 * which is pointless here. Thus, we pass "true" to task_css_check()
7727 * to prevent lockdep warnings.
7729 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7730 struct task_group
, css
);
7731 tg
= autogroup_task_group(tsk
, tg
);
7732 tsk
->sched_task_group
= tg
;
7734 #ifdef CONFIG_FAIR_GROUP_SCHED
7735 if (tsk
->sched_class
->task_move_group
)
7736 tsk
->sched_class
->task_move_group(tsk
, queued
);
7739 set_task_rq(tsk
, task_cpu(tsk
));
7741 if (unlikely(running
))
7742 tsk
->sched_class
->set_curr_task(rq
);
7744 enqueue_task(rq
, tsk
, 0);
7746 task_rq_unlock(rq
, tsk
, &flags
);
7748 #endif /* CONFIG_CGROUP_SCHED */
7750 #ifdef CONFIG_RT_GROUP_SCHED
7752 * Ensure that the real time constraints are schedulable.
7754 static DEFINE_MUTEX(rt_constraints_mutex
);
7756 /* Must be called with tasklist_lock held */
7757 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7759 struct task_struct
*g
, *p
;
7762 * Autogroups do not have RT tasks; see autogroup_create().
7764 if (task_group_is_autogroup(tg
))
7767 for_each_process_thread(g
, p
) {
7768 if (rt_task(p
) && task_group(p
) == tg
)
7775 struct rt_schedulable_data
{
7776 struct task_group
*tg
;
7781 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7783 struct rt_schedulable_data
*d
= data
;
7784 struct task_group
*child
;
7785 unsigned long total
, sum
= 0;
7786 u64 period
, runtime
;
7788 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7789 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7792 period
= d
->rt_period
;
7793 runtime
= d
->rt_runtime
;
7797 * Cannot have more runtime than the period.
7799 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7803 * Ensure we don't starve existing RT tasks.
7805 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7808 total
= to_ratio(period
, runtime
);
7811 * Nobody can have more than the global setting allows.
7813 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7817 * The sum of our children's runtime should not exceed our own.
7819 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7820 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7821 runtime
= child
->rt_bandwidth
.rt_runtime
;
7823 if (child
== d
->tg
) {
7824 period
= d
->rt_period
;
7825 runtime
= d
->rt_runtime
;
7828 sum
+= to_ratio(period
, runtime
);
7837 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7841 struct rt_schedulable_data data
= {
7843 .rt_period
= period
,
7844 .rt_runtime
= runtime
,
7848 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7854 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7855 u64 rt_period
, u64 rt_runtime
)
7860 * Disallowing the root group RT runtime is BAD, it would disallow the
7861 * kernel creating (and or operating) RT threads.
7863 if (tg
== &root_task_group
&& rt_runtime
== 0)
7866 /* No period doesn't make any sense. */
7870 mutex_lock(&rt_constraints_mutex
);
7871 read_lock(&tasklist_lock
);
7872 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7876 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7877 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7878 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7880 for_each_possible_cpu(i
) {
7881 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7883 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7884 rt_rq
->rt_runtime
= rt_runtime
;
7885 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7887 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7889 read_unlock(&tasklist_lock
);
7890 mutex_unlock(&rt_constraints_mutex
);
7895 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7897 u64 rt_runtime
, rt_period
;
7899 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7900 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7901 if (rt_runtime_us
< 0)
7902 rt_runtime
= RUNTIME_INF
;
7904 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7907 static long sched_group_rt_runtime(struct task_group
*tg
)
7911 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7914 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7915 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7916 return rt_runtime_us
;
7919 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7921 u64 rt_runtime
, rt_period
;
7923 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7924 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7926 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7929 static long sched_group_rt_period(struct task_group
*tg
)
7933 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7934 do_div(rt_period_us
, NSEC_PER_USEC
);
7935 return rt_period_us
;
7937 #endif /* CONFIG_RT_GROUP_SCHED */
7939 #ifdef CONFIG_RT_GROUP_SCHED
7940 static int sched_rt_global_constraints(void)
7944 mutex_lock(&rt_constraints_mutex
);
7945 read_lock(&tasklist_lock
);
7946 ret
= __rt_schedulable(NULL
, 0, 0);
7947 read_unlock(&tasklist_lock
);
7948 mutex_unlock(&rt_constraints_mutex
);
7953 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7955 /* Don't accept realtime tasks when there is no way for them to run */
7956 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7962 #else /* !CONFIG_RT_GROUP_SCHED */
7963 static int sched_rt_global_constraints(void)
7965 unsigned long flags
;
7968 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7969 for_each_possible_cpu(i
) {
7970 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7972 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7973 rt_rq
->rt_runtime
= global_rt_runtime();
7974 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7976 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7980 #endif /* CONFIG_RT_GROUP_SCHED */
7982 static int sched_dl_global_validate(void)
7984 u64 runtime
= global_rt_runtime();
7985 u64 period
= global_rt_period();
7986 u64 new_bw
= to_ratio(period
, runtime
);
7989 unsigned long flags
;
7992 * Here we want to check the bandwidth not being set to some
7993 * value smaller than the currently allocated bandwidth in
7994 * any of the root_domains.
7996 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7997 * cycling on root_domains... Discussion on different/better
7998 * solutions is welcome!
8000 for_each_possible_cpu(cpu
) {
8001 rcu_read_lock_sched();
8002 dl_b
= dl_bw_of(cpu
);
8004 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8005 if (new_bw
< dl_b
->total_bw
)
8007 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8009 rcu_read_unlock_sched();
8018 static void sched_dl_do_global(void)
8023 unsigned long flags
;
8025 def_dl_bandwidth
.dl_period
= global_rt_period();
8026 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8028 if (global_rt_runtime() != RUNTIME_INF
)
8029 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8032 * FIXME: As above...
8034 for_each_possible_cpu(cpu
) {
8035 rcu_read_lock_sched();
8036 dl_b
= dl_bw_of(cpu
);
8038 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8040 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8042 rcu_read_unlock_sched();
8046 static int sched_rt_global_validate(void)
8048 if (sysctl_sched_rt_period
<= 0)
8051 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8052 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8058 static void sched_rt_do_global(void)
8060 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8061 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8064 int sched_rt_handler(struct ctl_table
*table
, int write
,
8065 void __user
*buffer
, size_t *lenp
,
8068 int old_period
, old_runtime
;
8069 static DEFINE_MUTEX(mutex
);
8073 old_period
= sysctl_sched_rt_period
;
8074 old_runtime
= sysctl_sched_rt_runtime
;
8076 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8078 if (!ret
&& write
) {
8079 ret
= sched_rt_global_validate();
8083 ret
= sched_dl_global_validate();
8087 ret
= sched_rt_global_constraints();
8091 sched_rt_do_global();
8092 sched_dl_do_global();
8096 sysctl_sched_rt_period
= old_period
;
8097 sysctl_sched_rt_runtime
= old_runtime
;
8099 mutex_unlock(&mutex
);
8104 int sched_rr_handler(struct ctl_table
*table
, int write
,
8105 void __user
*buffer
, size_t *lenp
,
8109 static DEFINE_MUTEX(mutex
);
8112 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8113 /* make sure that internally we keep jiffies */
8114 /* also, writing zero resets timeslice to default */
8115 if (!ret
&& write
) {
8116 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8117 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8119 mutex_unlock(&mutex
);
8123 #ifdef CONFIG_CGROUP_SCHED
8125 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8127 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8130 static struct cgroup_subsys_state
*
8131 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8133 struct task_group
*parent
= css_tg(parent_css
);
8134 struct task_group
*tg
;
8137 /* This is early initialization for the top cgroup */
8138 return &root_task_group
.css
;
8141 tg
= sched_create_group(parent
);
8143 return ERR_PTR(-ENOMEM
);
8148 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8150 struct task_group
*tg
= css_tg(css
);
8151 struct task_group
*parent
= css_tg(css
->parent
);
8154 sched_online_group(tg
, parent
);
8158 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8160 struct task_group
*tg
= css_tg(css
);
8162 sched_destroy_group(tg
);
8165 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8167 struct task_group
*tg
= css_tg(css
);
8169 sched_offline_group(tg
);
8172 static void cpu_cgroup_fork(struct task_struct
*task
, void *private)
8174 sched_move_task(task
);
8177 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
8178 struct cgroup_taskset
*tset
)
8180 struct task_struct
*task
;
8182 cgroup_taskset_for_each(task
, tset
) {
8183 #ifdef CONFIG_RT_GROUP_SCHED
8184 if (!sched_rt_can_attach(css_tg(css
), task
))
8187 /* We don't support RT-tasks being in separate groups */
8188 if (task
->sched_class
!= &fair_sched_class
)
8195 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8196 struct cgroup_taskset
*tset
)
8198 struct task_struct
*task
;
8200 cgroup_taskset_for_each(task
, tset
)
8201 sched_move_task(task
);
8204 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8205 struct cgroup_subsys_state
*old_css
,
8206 struct task_struct
*task
)
8209 * cgroup_exit() is called in the copy_process() failure path.
8210 * Ignore this case since the task hasn't ran yet, this avoids
8211 * trying to poke a half freed task state from generic code.
8213 if (!(task
->flags
& PF_EXITING
))
8216 sched_move_task(task
);
8219 #ifdef CONFIG_FAIR_GROUP_SCHED
8220 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8221 struct cftype
*cftype
, u64 shareval
)
8223 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8226 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8229 struct task_group
*tg
= css_tg(css
);
8231 return (u64
) scale_load_down(tg
->shares
);
8234 #ifdef CONFIG_CFS_BANDWIDTH
8235 static DEFINE_MUTEX(cfs_constraints_mutex
);
8237 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8238 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8240 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8242 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8244 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8245 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8247 if (tg
== &root_task_group
)
8251 * Ensure we have at some amount of bandwidth every period. This is
8252 * to prevent reaching a state of large arrears when throttled via
8253 * entity_tick() resulting in prolonged exit starvation.
8255 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8259 * Likewise, bound things on the otherside by preventing insane quota
8260 * periods. This also allows us to normalize in computing quota
8263 if (period
> max_cfs_quota_period
)
8267 * Prevent race between setting of cfs_rq->runtime_enabled and
8268 * unthrottle_offline_cfs_rqs().
8271 mutex_lock(&cfs_constraints_mutex
);
8272 ret
= __cfs_schedulable(tg
, period
, quota
);
8276 runtime_enabled
= quota
!= RUNTIME_INF
;
8277 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8279 * If we need to toggle cfs_bandwidth_used, off->on must occur
8280 * before making related changes, and on->off must occur afterwards
8282 if (runtime_enabled
&& !runtime_was_enabled
)
8283 cfs_bandwidth_usage_inc();
8284 raw_spin_lock_irq(&cfs_b
->lock
);
8285 cfs_b
->period
= ns_to_ktime(period
);
8286 cfs_b
->quota
= quota
;
8288 __refill_cfs_bandwidth_runtime(cfs_b
);
8289 /* restart the period timer (if active) to handle new period expiry */
8290 if (runtime_enabled
)
8291 start_cfs_bandwidth(cfs_b
);
8292 raw_spin_unlock_irq(&cfs_b
->lock
);
8294 for_each_online_cpu(i
) {
8295 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8296 struct rq
*rq
= cfs_rq
->rq
;
8298 raw_spin_lock_irq(&rq
->lock
);
8299 cfs_rq
->runtime_enabled
= runtime_enabled
;
8300 cfs_rq
->runtime_remaining
= 0;
8302 if (cfs_rq
->throttled
)
8303 unthrottle_cfs_rq(cfs_rq
);
8304 raw_spin_unlock_irq(&rq
->lock
);
8306 if (runtime_was_enabled
&& !runtime_enabled
)
8307 cfs_bandwidth_usage_dec();
8309 mutex_unlock(&cfs_constraints_mutex
);
8315 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8319 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8320 if (cfs_quota_us
< 0)
8321 quota
= RUNTIME_INF
;
8323 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8325 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8328 long tg_get_cfs_quota(struct task_group
*tg
)
8332 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8335 quota_us
= tg
->cfs_bandwidth
.quota
;
8336 do_div(quota_us
, NSEC_PER_USEC
);
8341 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8345 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8346 quota
= tg
->cfs_bandwidth
.quota
;
8348 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8351 long tg_get_cfs_period(struct task_group
*tg
)
8355 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8356 do_div(cfs_period_us
, NSEC_PER_USEC
);
8358 return cfs_period_us
;
8361 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8364 return tg_get_cfs_quota(css_tg(css
));
8367 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8368 struct cftype
*cftype
, s64 cfs_quota_us
)
8370 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8373 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8376 return tg_get_cfs_period(css_tg(css
));
8379 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8380 struct cftype
*cftype
, u64 cfs_period_us
)
8382 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8385 struct cfs_schedulable_data
{
8386 struct task_group
*tg
;
8391 * normalize group quota/period to be quota/max_period
8392 * note: units are usecs
8394 static u64
normalize_cfs_quota(struct task_group
*tg
,
8395 struct cfs_schedulable_data
*d
)
8403 period
= tg_get_cfs_period(tg
);
8404 quota
= tg_get_cfs_quota(tg
);
8407 /* note: these should typically be equivalent */
8408 if (quota
== RUNTIME_INF
|| quota
== -1)
8411 return to_ratio(period
, quota
);
8414 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8416 struct cfs_schedulable_data
*d
= data
;
8417 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8418 s64 quota
= 0, parent_quota
= -1;
8421 quota
= RUNTIME_INF
;
8423 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8425 quota
= normalize_cfs_quota(tg
, d
);
8426 parent_quota
= parent_b
->hierarchical_quota
;
8429 * ensure max(child_quota) <= parent_quota, inherit when no
8432 if (quota
== RUNTIME_INF
)
8433 quota
= parent_quota
;
8434 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8437 cfs_b
->hierarchical_quota
= quota
;
8442 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8445 struct cfs_schedulable_data data
= {
8451 if (quota
!= RUNTIME_INF
) {
8452 do_div(data
.period
, NSEC_PER_USEC
);
8453 do_div(data
.quota
, NSEC_PER_USEC
);
8457 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8463 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8465 struct task_group
*tg
= css_tg(seq_css(sf
));
8466 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8468 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8469 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8470 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8474 #endif /* CONFIG_CFS_BANDWIDTH */
8475 #endif /* CONFIG_FAIR_GROUP_SCHED */
8477 #ifdef CONFIG_RT_GROUP_SCHED
8478 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8479 struct cftype
*cft
, s64 val
)
8481 return sched_group_set_rt_runtime(css_tg(css
), val
);
8484 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8487 return sched_group_rt_runtime(css_tg(css
));
8490 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8491 struct cftype
*cftype
, u64 rt_period_us
)
8493 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8496 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8499 return sched_group_rt_period(css_tg(css
));
8501 #endif /* CONFIG_RT_GROUP_SCHED */
8503 static struct cftype cpu_files
[] = {
8504 #ifdef CONFIG_FAIR_GROUP_SCHED
8507 .read_u64
= cpu_shares_read_u64
,
8508 .write_u64
= cpu_shares_write_u64
,
8511 #ifdef CONFIG_CFS_BANDWIDTH
8513 .name
= "cfs_quota_us",
8514 .read_s64
= cpu_cfs_quota_read_s64
,
8515 .write_s64
= cpu_cfs_quota_write_s64
,
8518 .name
= "cfs_period_us",
8519 .read_u64
= cpu_cfs_period_read_u64
,
8520 .write_u64
= cpu_cfs_period_write_u64
,
8524 .seq_show
= cpu_stats_show
,
8527 #ifdef CONFIG_RT_GROUP_SCHED
8529 .name
= "rt_runtime_us",
8530 .read_s64
= cpu_rt_runtime_read
,
8531 .write_s64
= cpu_rt_runtime_write
,
8534 .name
= "rt_period_us",
8535 .read_u64
= cpu_rt_period_read_uint
,
8536 .write_u64
= cpu_rt_period_write_uint
,
8542 struct cgroup_subsys cpu_cgrp_subsys
= {
8543 .css_alloc
= cpu_cgroup_css_alloc
,
8544 .css_free
= cpu_cgroup_css_free
,
8545 .css_online
= cpu_cgroup_css_online
,
8546 .css_offline
= cpu_cgroup_css_offline
,
8547 .fork
= cpu_cgroup_fork
,
8548 .can_attach
= cpu_cgroup_can_attach
,
8549 .attach
= cpu_cgroup_attach
,
8550 .exit
= cpu_cgroup_exit
,
8551 .legacy_cftypes
= cpu_files
,
8555 #endif /* CONFIG_CGROUP_SCHED */
8557 void dump_cpu_task(int cpu
)
8559 pr_info("Task dump for CPU %d:\n", cpu
);
8560 sched_show_task(cpu_curr(cpu
));