4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex
);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
96 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
98 void update_rq_clock(struct rq
*rq
)
102 lockdep_assert_held(&rq
->lock
);
104 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
107 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
111 update_rq_clock_task(rq
, delta
);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug
unsigned int sysctl_sched_features
=
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names
[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file
*m
, void *v
)
141 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
142 if (!(sysctl_sched_features
& (1UL << i
)))
144 seq_printf(m
, "%s ", sched_feat_names
[i
]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
160 #include "features.h"
165 static void sched_feat_disable(int i
)
167 if (static_key_enabled(&sched_feat_keys
[i
]))
168 static_key_slow_dec(&sched_feat_keys
[i
]);
171 static void sched_feat_enable(int i
)
173 if (!static_key_enabled(&sched_feat_keys
[i
]))
174 static_key_slow_inc(&sched_feat_keys
[i
]);
177 static void sched_feat_disable(int i
) { };
178 static void sched_feat_enable(int i
) { };
179 #endif /* HAVE_JUMP_LABEL */
181 static int sched_feat_set(char *cmp
)
186 if (strncmp(cmp
, "NO_", 3) == 0) {
191 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
192 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
194 sysctl_sched_features
&= ~(1UL << i
);
195 sched_feat_disable(i
);
197 sysctl_sched_features
|= (1UL << i
);
198 sched_feat_enable(i
);
208 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
209 size_t cnt
, loff_t
*ppos
)
219 if (copy_from_user(&buf
, ubuf
, cnt
))
225 /* Ensure the static_key remains in a consistent state */
226 inode
= file_inode(filp
);
227 mutex_lock(&inode
->i_mutex
);
228 i
= sched_feat_set(cmp
);
229 mutex_unlock(&inode
->i_mutex
);
230 if (i
== __SCHED_FEAT_NR
)
238 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
240 return single_open(filp
, sched_feat_show
, NULL
);
243 static const struct file_operations sched_feat_fops
= {
244 .open
= sched_feat_open
,
245 .write
= sched_feat_write
,
248 .release
= single_release
,
251 static __init
int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
258 late_initcall(sched_init_debug
);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
268 * period over which we average the RT time consumption, measured
273 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period
= 1000000;
281 __read_mostly
int scheduler_running
;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime
= 950000;
289 /* cpus with isolated domains */
290 cpumask_var_t cpu_isolated_map
;
293 * this_rq_lock - lock this runqueue and disable interrupts.
295 static struct rq
*this_rq_lock(void)
302 raw_spin_lock(&rq
->lock
);
307 #ifdef CONFIG_SCHED_HRTICK
309 * Use HR-timers to deliver accurate preemption points.
312 static void hrtick_clear(struct rq
*rq
)
314 if (hrtimer_active(&rq
->hrtick_timer
))
315 hrtimer_cancel(&rq
->hrtick_timer
);
319 * High-resolution timer tick.
320 * Runs from hardirq context with interrupts disabled.
322 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
324 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
326 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
328 raw_spin_lock(&rq
->lock
);
330 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
331 raw_spin_unlock(&rq
->lock
);
333 return HRTIMER_NORESTART
;
338 static void __hrtick_restart(struct rq
*rq
)
340 struct hrtimer
*timer
= &rq
->hrtick_timer
;
342 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
346 * called from hardirq (IPI) context
348 static void __hrtick_start(void *arg
)
352 raw_spin_lock(&rq
->lock
);
353 __hrtick_restart(rq
);
354 rq
->hrtick_csd_pending
= 0;
355 raw_spin_unlock(&rq
->lock
);
359 * Called to set the hrtick timer state.
361 * called with rq->lock held and irqs disabled
363 void hrtick_start(struct rq
*rq
, u64 delay
)
365 struct hrtimer
*timer
= &rq
->hrtick_timer
;
370 * Don't schedule slices shorter than 10000ns, that just
371 * doesn't make sense and can cause timer DoS.
373 delta
= max_t(s64
, delay
, 10000LL);
374 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
376 hrtimer_set_expires(timer
, time
);
378 if (rq
== this_rq()) {
379 __hrtick_restart(rq
);
380 } else if (!rq
->hrtick_csd_pending
) {
381 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
382 rq
->hrtick_csd_pending
= 1;
387 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
389 int cpu
= (int)(long)hcpu
;
392 case CPU_UP_CANCELED
:
393 case CPU_UP_CANCELED_FROZEN
:
394 case CPU_DOWN_PREPARE
:
395 case CPU_DOWN_PREPARE_FROZEN
:
397 case CPU_DEAD_FROZEN
:
398 hrtick_clear(cpu_rq(cpu
));
405 static __init
void init_hrtick(void)
407 hotcpu_notifier(hotplug_hrtick
, 0);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq
*rq
, u64 delay
)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay
= max_t(u64
, delay
, 10000LL);
422 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
423 HRTIMER_MODE_REL_PINNED
);
426 static inline void init_hrtick(void)
429 #endif /* CONFIG_SMP */
431 static void init_rq_hrtick(struct rq
*rq
)
434 rq
->hrtick_csd_pending
= 0;
436 rq
->hrtick_csd
.flags
= 0;
437 rq
->hrtick_csd
.func
= __hrtick_start
;
438 rq
->hrtick_csd
.info
= rq
;
441 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
442 rq
->hrtick_timer
.function
= hrtick
;
444 #else /* CONFIG_SCHED_HRTICK */
445 static inline void hrtick_clear(struct rq
*rq
)
449 static inline void init_rq_hrtick(struct rq
*rq
)
453 static inline void init_hrtick(void)
456 #endif /* CONFIG_SCHED_HRTICK */
459 * cmpxchg based fetch_or, macro so it works for different integer types
461 #define fetch_or(ptr, val) \
462 ({ typeof(*(ptr)) __old, __val = *(ptr); \
464 __old = cmpxchg((ptr), __val, __val | (val)); \
465 if (__old == __val) \
472 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
474 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
475 * this avoids any races wrt polling state changes and thereby avoids
478 static bool set_nr_and_not_polling(struct task_struct
*p
)
480 struct thread_info
*ti
= task_thread_info(p
);
481 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
485 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
487 * If this returns true, then the idle task promises to call
488 * sched_ttwu_pending() and reschedule soon.
490 static bool set_nr_if_polling(struct task_struct
*p
)
492 struct thread_info
*ti
= task_thread_info(p
);
493 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
496 if (!(val
& _TIF_POLLING_NRFLAG
))
498 if (val
& _TIF_NEED_RESCHED
)
500 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
509 static bool set_nr_and_not_polling(struct task_struct
*p
)
511 set_tsk_need_resched(p
);
516 static bool set_nr_if_polling(struct task_struct
*p
)
523 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
525 struct wake_q_node
*node
= &task
->wake_q
;
528 * Atomically grab the task, if ->wake_q is !nil already it means
529 * its already queued (either by us or someone else) and will get the
530 * wakeup due to that.
532 * This cmpxchg() implies a full barrier, which pairs with the write
533 * barrier implied by the wakeup in wake_up_list().
535 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
538 get_task_struct(task
);
541 * The head is context local, there can be no concurrency.
544 head
->lastp
= &node
->next
;
547 void wake_up_q(struct wake_q_head
*head
)
549 struct wake_q_node
*node
= head
->first
;
551 while (node
!= WAKE_Q_TAIL
) {
552 struct task_struct
*task
;
554 task
= container_of(node
, struct task_struct
, wake_q
);
556 /* task can safely be re-inserted now */
558 task
->wake_q
.next
= NULL
;
561 * wake_up_process() implies a wmb() to pair with the queueing
562 * in wake_q_add() so as not to miss wakeups.
564 wake_up_process(task
);
565 put_task_struct(task
);
570 * resched_curr - mark rq's current task 'to be rescheduled now'.
572 * On UP this means the setting of the need_resched flag, on SMP it
573 * might also involve a cross-CPU call to trigger the scheduler on
576 void resched_curr(struct rq
*rq
)
578 struct task_struct
*curr
= rq
->curr
;
581 lockdep_assert_held(&rq
->lock
);
583 if (test_tsk_need_resched(curr
))
588 if (cpu
== smp_processor_id()) {
589 set_tsk_need_resched(curr
);
590 set_preempt_need_resched();
594 if (set_nr_and_not_polling(curr
))
595 smp_send_reschedule(cpu
);
597 trace_sched_wake_idle_without_ipi(cpu
);
600 void resched_cpu(int cpu
)
602 struct rq
*rq
= cpu_rq(cpu
);
605 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
608 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
612 #ifdef CONFIG_NO_HZ_COMMON
614 * In the semi idle case, use the nearest busy cpu for migrating timers
615 * from an idle cpu. This is good for power-savings.
617 * We don't do similar optimization for completely idle system, as
618 * selecting an idle cpu will add more delays to the timers than intended
619 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
621 int get_nohz_timer_target(void)
623 int i
, cpu
= smp_processor_id();
624 struct sched_domain
*sd
;
630 for_each_domain(cpu
, sd
) {
631 for_each_cpu(i
, sched_domain_span(sd
)) {
643 * When add_timer_on() enqueues a timer into the timer wheel of an
644 * idle CPU then this timer might expire before the next timer event
645 * which is scheduled to wake up that CPU. In case of a completely
646 * idle system the next event might even be infinite time into the
647 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
648 * leaves the inner idle loop so the newly added timer is taken into
649 * account when the CPU goes back to idle and evaluates the timer
650 * wheel for the next timer event.
652 static void wake_up_idle_cpu(int cpu
)
654 struct rq
*rq
= cpu_rq(cpu
);
656 if (cpu
== smp_processor_id())
659 if (set_nr_and_not_polling(rq
->idle
))
660 smp_send_reschedule(cpu
);
662 trace_sched_wake_idle_without_ipi(cpu
);
665 static bool wake_up_full_nohz_cpu(int cpu
)
668 * We just need the target to call irq_exit() and re-evaluate
669 * the next tick. The nohz full kick at least implies that.
670 * If needed we can still optimize that later with an
673 if (tick_nohz_full_cpu(cpu
)) {
674 if (cpu
!= smp_processor_id() ||
675 tick_nohz_tick_stopped())
676 tick_nohz_full_kick_cpu(cpu
);
683 void wake_up_nohz_cpu(int cpu
)
685 if (!wake_up_full_nohz_cpu(cpu
))
686 wake_up_idle_cpu(cpu
);
689 static inline bool got_nohz_idle_kick(void)
691 int cpu
= smp_processor_id();
693 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
696 if (idle_cpu(cpu
) && !need_resched())
700 * We can't run Idle Load Balance on this CPU for this time so we
701 * cancel it and clear NOHZ_BALANCE_KICK
703 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
707 #else /* CONFIG_NO_HZ_COMMON */
709 static inline bool got_nohz_idle_kick(void)
714 #endif /* CONFIG_NO_HZ_COMMON */
716 #ifdef CONFIG_NO_HZ_FULL
717 bool sched_can_stop_tick(void)
720 * FIFO realtime policy runs the highest priority task. Other runnable
721 * tasks are of a lower priority. The scheduler tick does nothing.
723 if (current
->policy
== SCHED_FIFO
)
727 * Round-robin realtime tasks time slice with other tasks at the same
728 * realtime priority. Is this task the only one at this priority?
730 if (current
->policy
== SCHED_RR
) {
731 struct sched_rt_entity
*rt_se
= ¤t
->rt
;
733 return rt_se
->run_list
.prev
== rt_se
->run_list
.next
;
737 * More than one running task need preemption.
738 * nr_running update is assumed to be visible
739 * after IPI is sent from wakers.
741 if (this_rq()->nr_running
> 1)
746 #endif /* CONFIG_NO_HZ_FULL */
748 void sched_avg_update(struct rq
*rq
)
750 s64 period
= sched_avg_period();
752 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
754 * Inline assembly required to prevent the compiler
755 * optimising this loop into a divmod call.
756 * See __iter_div_u64_rem() for another example of this.
758 asm("" : "+rm" (rq
->age_stamp
));
759 rq
->age_stamp
+= period
;
764 #endif /* CONFIG_SMP */
766 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
767 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
769 * Iterate task_group tree rooted at *from, calling @down when first entering a
770 * node and @up when leaving it for the final time.
772 * Caller must hold rcu_lock or sufficient equivalent.
774 int walk_tg_tree_from(struct task_group
*from
,
775 tg_visitor down
, tg_visitor up
, void *data
)
777 struct task_group
*parent
, *child
;
783 ret
= (*down
)(parent
, data
);
786 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
793 ret
= (*up
)(parent
, data
);
794 if (ret
|| parent
== from
)
798 parent
= parent
->parent
;
805 int tg_nop(struct task_group
*tg
, void *data
)
811 static void set_load_weight(struct task_struct
*p
)
813 int prio
= p
->static_prio
- MAX_RT_PRIO
;
814 struct load_weight
*load
= &p
->se
.load
;
817 * SCHED_IDLE tasks get minimal weight:
819 if (p
->policy
== SCHED_IDLE
) {
820 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
821 load
->inv_weight
= WMULT_IDLEPRIO
;
825 load
->weight
= scale_load(prio_to_weight
[prio
]);
826 load
->inv_weight
= prio_to_wmult
[prio
];
829 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
832 sched_info_queued(rq
, p
);
833 p
->sched_class
->enqueue_task(rq
, p
, flags
);
836 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
839 sched_info_dequeued(rq
, p
);
840 p
->sched_class
->dequeue_task(rq
, p
, flags
);
843 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
845 if (task_contributes_to_load(p
))
846 rq
->nr_uninterruptible
--;
848 enqueue_task(rq
, p
, flags
);
851 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
853 if (task_contributes_to_load(p
))
854 rq
->nr_uninterruptible
++;
856 dequeue_task(rq
, p
, flags
);
859 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
862 * In theory, the compile should just see 0 here, and optimize out the call
863 * to sched_rt_avg_update. But I don't trust it...
865 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
866 s64 steal
= 0, irq_delta
= 0;
868 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
869 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
872 * Since irq_time is only updated on {soft,}irq_exit, we might run into
873 * this case when a previous update_rq_clock() happened inside a
876 * When this happens, we stop ->clock_task and only update the
877 * prev_irq_time stamp to account for the part that fit, so that a next
878 * update will consume the rest. This ensures ->clock_task is
881 * It does however cause some slight miss-attribution of {soft,}irq
882 * time, a more accurate solution would be to update the irq_time using
883 * the current rq->clock timestamp, except that would require using
886 if (irq_delta
> delta
)
889 rq
->prev_irq_time
+= irq_delta
;
892 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
893 if (static_key_false((¶virt_steal_rq_enabled
))) {
894 steal
= paravirt_steal_clock(cpu_of(rq
));
895 steal
-= rq
->prev_steal_time_rq
;
897 if (unlikely(steal
> delta
))
900 rq
->prev_steal_time_rq
+= steal
;
905 rq
->clock_task
+= delta
;
907 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
908 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
909 sched_rt_avg_update(rq
, irq_delta
+ steal
);
913 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
915 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
916 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
920 * Make it appear like a SCHED_FIFO task, its something
921 * userspace knows about and won't get confused about.
923 * Also, it will make PI more or less work without too
924 * much confusion -- but then, stop work should not
925 * rely on PI working anyway.
927 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
929 stop
->sched_class
= &stop_sched_class
;
932 cpu_rq(cpu
)->stop
= stop
;
936 * Reset it back to a normal scheduling class so that
937 * it can die in pieces.
939 old_stop
->sched_class
= &rt_sched_class
;
944 * __normal_prio - return the priority that is based on the static prio
946 static inline int __normal_prio(struct task_struct
*p
)
948 return p
->static_prio
;
952 * Calculate the expected normal priority: i.e. priority
953 * without taking RT-inheritance into account. Might be
954 * boosted by interactivity modifiers. Changes upon fork,
955 * setprio syscalls, and whenever the interactivity
956 * estimator recalculates.
958 static inline int normal_prio(struct task_struct
*p
)
962 if (task_has_dl_policy(p
))
963 prio
= MAX_DL_PRIO
-1;
964 else if (task_has_rt_policy(p
))
965 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
967 prio
= __normal_prio(p
);
972 * Calculate the current priority, i.e. the priority
973 * taken into account by the scheduler. This value might
974 * be boosted by RT tasks, or might be boosted by
975 * interactivity modifiers. Will be RT if the task got
976 * RT-boosted. If not then it returns p->normal_prio.
978 static int effective_prio(struct task_struct
*p
)
980 p
->normal_prio
= normal_prio(p
);
982 * If we are RT tasks or we were boosted to RT priority,
983 * keep the priority unchanged. Otherwise, update priority
984 * to the normal priority:
986 if (!rt_prio(p
->prio
))
987 return p
->normal_prio
;
992 * task_curr - is this task currently executing on a CPU?
993 * @p: the task in question.
995 * Return: 1 if the task is currently executing. 0 otherwise.
997 inline int task_curr(const struct task_struct
*p
)
999 return cpu_curr(task_cpu(p
)) == p
;
1003 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1004 * use the balance_callback list if you want balancing.
1006 * this means any call to check_class_changed() must be followed by a call to
1007 * balance_callback().
1009 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1010 const struct sched_class
*prev_class
,
1013 if (prev_class
!= p
->sched_class
) {
1014 if (prev_class
->switched_from
)
1015 prev_class
->switched_from(rq
, p
);
1017 p
->sched_class
->switched_to(rq
, p
);
1018 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1019 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1022 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1024 const struct sched_class
*class;
1026 if (p
->sched_class
== rq
->curr
->sched_class
) {
1027 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1029 for_each_class(class) {
1030 if (class == rq
->curr
->sched_class
)
1032 if (class == p
->sched_class
) {
1040 * A queue event has occurred, and we're going to schedule. In
1041 * this case, we can save a useless back to back clock update.
1043 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1044 rq_clock_skip_update(rq
, true);
1049 * This is how migration works:
1051 * 1) we invoke migration_cpu_stop() on the target CPU using
1053 * 2) stopper starts to run (implicitly forcing the migrated thread
1055 * 3) it checks whether the migrated task is still in the wrong runqueue.
1056 * 4) if it's in the wrong runqueue then the migration thread removes
1057 * it and puts it into the right queue.
1058 * 5) stopper completes and stop_one_cpu() returns and the migration
1063 * move_queued_task - move a queued task to new rq.
1065 * Returns (locked) new rq. Old rq's lock is released.
1067 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
1069 lockdep_assert_held(&rq
->lock
);
1071 dequeue_task(rq
, p
, 0);
1072 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1073 set_task_cpu(p
, new_cpu
);
1074 raw_spin_unlock(&rq
->lock
);
1076 rq
= cpu_rq(new_cpu
);
1078 raw_spin_lock(&rq
->lock
);
1079 BUG_ON(task_cpu(p
) != new_cpu
);
1080 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1081 enqueue_task(rq
, p
, 0);
1082 check_preempt_curr(rq
, p
, 0);
1087 struct migration_arg
{
1088 struct task_struct
*task
;
1093 * Move (not current) task off this cpu, onto dest cpu. We're doing
1094 * this because either it can't run here any more (set_cpus_allowed()
1095 * away from this CPU, or CPU going down), or because we're
1096 * attempting to rebalance this task on exec (sched_exec).
1098 * So we race with normal scheduler movements, but that's OK, as long
1099 * as the task is no longer on this CPU.
1101 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1103 if (unlikely(!cpu_active(dest_cpu
)))
1106 /* Affinity changed (again). */
1107 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1110 rq
= move_queued_task(rq
, p
, dest_cpu
);
1116 * migration_cpu_stop - this will be executed by a highprio stopper thread
1117 * and performs thread migration by bumping thread off CPU then
1118 * 'pushing' onto another runqueue.
1120 static int migration_cpu_stop(void *data
)
1122 struct migration_arg
*arg
= data
;
1123 struct task_struct
*p
= arg
->task
;
1124 struct rq
*rq
= this_rq();
1127 * The original target cpu might have gone down and we might
1128 * be on another cpu but it doesn't matter.
1130 local_irq_disable();
1132 * We need to explicitly wake pending tasks before running
1133 * __migrate_task() such that we will not miss enforcing cpus_allowed
1134 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1136 sched_ttwu_pending();
1138 raw_spin_lock(&p
->pi_lock
);
1139 raw_spin_lock(&rq
->lock
);
1141 * If task_rq(p) != rq, it cannot be migrated here, because we're
1142 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1143 * we're holding p->pi_lock.
1145 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1146 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1147 raw_spin_unlock(&rq
->lock
);
1148 raw_spin_unlock(&p
->pi_lock
);
1154 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1156 lockdep_assert_held(&p
->pi_lock
);
1158 if (p
->sched_class
->set_cpus_allowed
)
1159 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1161 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1162 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1166 * Change a given task's CPU affinity. Migrate the thread to a
1167 * proper CPU and schedule it away if the CPU it's executing on
1168 * is removed from the allowed bitmask.
1170 * NOTE: the caller must have a valid reference to the task, the
1171 * task must not exit() & deallocate itself prematurely. The
1172 * call is not atomic; no spinlocks may be held.
1174 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1175 const struct cpumask
*new_mask
, bool check
)
1177 unsigned long flags
;
1179 unsigned int dest_cpu
;
1182 rq
= task_rq_lock(p
, &flags
);
1185 * Must re-check here, to close a race against __kthread_bind(),
1186 * sched_setaffinity() is not guaranteed to observe the flag.
1188 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1193 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1196 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1201 do_set_cpus_allowed(p
, new_mask
);
1203 /* Can the task run on the task's current CPU? If so, we're done */
1204 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1207 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1208 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1209 struct migration_arg arg
= { p
, dest_cpu
};
1210 /* Need help from migration thread: drop lock and wait. */
1211 task_rq_unlock(rq
, p
, &flags
);
1212 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1213 tlb_migrate_finish(p
->mm
);
1215 } else if (task_on_rq_queued(p
)) {
1217 * OK, since we're going to drop the lock immediately
1218 * afterwards anyway.
1220 lockdep_unpin_lock(&rq
->lock
);
1221 rq
= move_queued_task(rq
, p
, dest_cpu
);
1222 lockdep_pin_lock(&rq
->lock
);
1225 task_rq_unlock(rq
, p
, &flags
);
1230 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1232 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1234 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1236 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1238 #ifdef CONFIG_SCHED_DEBUG
1240 * We should never call set_task_cpu() on a blocked task,
1241 * ttwu() will sort out the placement.
1243 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1246 #ifdef CONFIG_LOCKDEP
1248 * The caller should hold either p->pi_lock or rq->lock, when changing
1249 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1251 * sched_move_task() holds both and thus holding either pins the cgroup,
1254 * Furthermore, all task_rq users should acquire both locks, see
1257 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1258 lockdep_is_held(&task_rq(p
)->lock
)));
1262 trace_sched_migrate_task(p
, new_cpu
);
1264 if (task_cpu(p
) != new_cpu
) {
1265 if (p
->sched_class
->migrate_task_rq
)
1266 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1267 p
->se
.nr_migrations
++;
1268 perf_event_task_migrate(p
);
1271 __set_task_cpu(p
, new_cpu
);
1274 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1276 if (task_on_rq_queued(p
)) {
1277 struct rq
*src_rq
, *dst_rq
;
1279 src_rq
= task_rq(p
);
1280 dst_rq
= cpu_rq(cpu
);
1282 deactivate_task(src_rq
, p
, 0);
1283 set_task_cpu(p
, cpu
);
1284 activate_task(dst_rq
, p
, 0);
1285 check_preempt_curr(dst_rq
, p
, 0);
1288 * Task isn't running anymore; make it appear like we migrated
1289 * it before it went to sleep. This means on wakeup we make the
1290 * previous cpu our targer instead of where it really is.
1296 struct migration_swap_arg
{
1297 struct task_struct
*src_task
, *dst_task
;
1298 int src_cpu
, dst_cpu
;
1301 static int migrate_swap_stop(void *data
)
1303 struct migration_swap_arg
*arg
= data
;
1304 struct rq
*src_rq
, *dst_rq
;
1307 src_rq
= cpu_rq(arg
->src_cpu
);
1308 dst_rq
= cpu_rq(arg
->dst_cpu
);
1310 double_raw_lock(&arg
->src_task
->pi_lock
,
1311 &arg
->dst_task
->pi_lock
);
1312 double_rq_lock(src_rq
, dst_rq
);
1313 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1316 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1319 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1322 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1325 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1326 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1331 double_rq_unlock(src_rq
, dst_rq
);
1332 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1333 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1339 * Cross migrate two tasks
1341 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1343 struct migration_swap_arg arg
;
1346 arg
= (struct migration_swap_arg
){
1348 .src_cpu
= task_cpu(cur
),
1350 .dst_cpu
= task_cpu(p
),
1353 if (arg
.src_cpu
== arg
.dst_cpu
)
1357 * These three tests are all lockless; this is OK since all of them
1358 * will be re-checked with proper locks held further down the line.
1360 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1363 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1366 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1369 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1370 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1377 * wait_task_inactive - wait for a thread to unschedule.
1379 * If @match_state is nonzero, it's the @p->state value just checked and
1380 * not expected to change. If it changes, i.e. @p might have woken up,
1381 * then return zero. When we succeed in waiting for @p to be off its CPU,
1382 * we return a positive number (its total switch count). If a second call
1383 * a short while later returns the same number, the caller can be sure that
1384 * @p has remained unscheduled the whole time.
1386 * The caller must ensure that the task *will* unschedule sometime soon,
1387 * else this function might spin for a *long* time. This function can't
1388 * be called with interrupts off, or it may introduce deadlock with
1389 * smp_call_function() if an IPI is sent by the same process we are
1390 * waiting to become inactive.
1392 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1394 unsigned long flags
;
1395 int running
, queued
;
1401 * We do the initial early heuristics without holding
1402 * any task-queue locks at all. We'll only try to get
1403 * the runqueue lock when things look like they will
1409 * If the task is actively running on another CPU
1410 * still, just relax and busy-wait without holding
1413 * NOTE! Since we don't hold any locks, it's not
1414 * even sure that "rq" stays as the right runqueue!
1415 * But we don't care, since "task_running()" will
1416 * return false if the runqueue has changed and p
1417 * is actually now running somewhere else!
1419 while (task_running(rq
, p
)) {
1420 if (match_state
&& unlikely(p
->state
!= match_state
))
1426 * Ok, time to look more closely! We need the rq
1427 * lock now, to be *sure*. If we're wrong, we'll
1428 * just go back and repeat.
1430 rq
= task_rq_lock(p
, &flags
);
1431 trace_sched_wait_task(p
);
1432 running
= task_running(rq
, p
);
1433 queued
= task_on_rq_queued(p
);
1435 if (!match_state
|| p
->state
== match_state
)
1436 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1437 task_rq_unlock(rq
, p
, &flags
);
1440 * If it changed from the expected state, bail out now.
1442 if (unlikely(!ncsw
))
1446 * Was it really running after all now that we
1447 * checked with the proper locks actually held?
1449 * Oops. Go back and try again..
1451 if (unlikely(running
)) {
1457 * It's not enough that it's not actively running,
1458 * it must be off the runqueue _entirely_, and not
1461 * So if it was still runnable (but just not actively
1462 * running right now), it's preempted, and we should
1463 * yield - it could be a while.
1465 if (unlikely(queued
)) {
1466 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1468 set_current_state(TASK_UNINTERRUPTIBLE
);
1469 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1474 * Ahh, all good. It wasn't running, and it wasn't
1475 * runnable, which means that it will never become
1476 * running in the future either. We're all done!
1485 * kick_process - kick a running thread to enter/exit the kernel
1486 * @p: the to-be-kicked thread
1488 * Cause a process which is running on another CPU to enter
1489 * kernel-mode, without any delay. (to get signals handled.)
1491 * NOTE: this function doesn't have to take the runqueue lock,
1492 * because all it wants to ensure is that the remote task enters
1493 * the kernel. If the IPI races and the task has been migrated
1494 * to another CPU then no harm is done and the purpose has been
1497 void kick_process(struct task_struct
*p
)
1503 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1504 smp_send_reschedule(cpu
);
1507 EXPORT_SYMBOL_GPL(kick_process
);
1510 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1512 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1514 int nid
= cpu_to_node(cpu
);
1515 const struct cpumask
*nodemask
= NULL
;
1516 enum { cpuset
, possible
, fail
} state
= cpuset
;
1520 * If the node that the cpu is on has been offlined, cpu_to_node()
1521 * will return -1. There is no cpu on the node, and we should
1522 * select the cpu on the other node.
1525 nodemask
= cpumask_of_node(nid
);
1527 /* Look for allowed, online CPU in same node. */
1528 for_each_cpu(dest_cpu
, nodemask
) {
1529 if (!cpu_online(dest_cpu
))
1531 if (!cpu_active(dest_cpu
))
1533 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1539 /* Any allowed, online CPU? */
1540 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1541 if (!cpu_online(dest_cpu
))
1543 if (!cpu_active(dest_cpu
))
1550 /* No more Mr. Nice Guy. */
1551 cpuset_cpus_allowed_fallback(p
);
1556 do_set_cpus_allowed(p
, cpu_possible_mask
);
1567 if (state
!= cpuset
) {
1569 * Don't tell them about moving exiting tasks or
1570 * kernel threads (both mm NULL), since they never
1573 if (p
->mm
&& printk_ratelimit()) {
1574 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1575 task_pid_nr(p
), p
->comm
, cpu
);
1583 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1586 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1588 lockdep_assert_held(&p
->pi_lock
);
1590 if (p
->nr_cpus_allowed
> 1)
1591 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1594 * In order not to call set_task_cpu() on a blocking task we need
1595 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1598 * Since this is common to all placement strategies, this lives here.
1600 * [ this allows ->select_task() to simply return task_cpu(p) and
1601 * not worry about this generic constraint ]
1603 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1605 cpu
= select_fallback_rq(task_cpu(p
), p
);
1610 static void update_avg(u64
*avg
, u64 sample
)
1612 s64 diff
= sample
- *avg
;
1618 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1619 const struct cpumask
*new_mask
, bool check
)
1621 return set_cpus_allowed_ptr(p
, new_mask
);
1624 #endif /* CONFIG_SMP */
1627 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1629 #ifdef CONFIG_SCHEDSTATS
1630 struct rq
*rq
= this_rq();
1633 int this_cpu
= smp_processor_id();
1635 if (cpu
== this_cpu
) {
1636 schedstat_inc(rq
, ttwu_local
);
1637 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1639 struct sched_domain
*sd
;
1641 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1643 for_each_domain(this_cpu
, sd
) {
1644 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1645 schedstat_inc(sd
, ttwu_wake_remote
);
1652 if (wake_flags
& WF_MIGRATED
)
1653 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1655 #endif /* CONFIG_SMP */
1657 schedstat_inc(rq
, ttwu_count
);
1658 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1660 if (wake_flags
& WF_SYNC
)
1661 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1663 #endif /* CONFIG_SCHEDSTATS */
1666 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1668 activate_task(rq
, p
, en_flags
);
1669 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1671 /* if a worker is waking up, notify workqueue */
1672 if (p
->flags
& PF_WQ_WORKER
)
1673 wq_worker_waking_up(p
, cpu_of(rq
));
1677 * Mark the task runnable and perform wakeup-preemption.
1680 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1682 check_preempt_curr(rq
, p
, wake_flags
);
1683 p
->state
= TASK_RUNNING
;
1684 trace_sched_wakeup(p
);
1687 if (p
->sched_class
->task_woken
) {
1689 * Our task @p is fully woken up and running; so its safe to
1690 * drop the rq->lock, hereafter rq is only used for statistics.
1692 lockdep_unpin_lock(&rq
->lock
);
1693 p
->sched_class
->task_woken(rq
, p
);
1694 lockdep_pin_lock(&rq
->lock
);
1697 if (rq
->idle_stamp
) {
1698 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1699 u64 max
= 2*rq
->max_idle_balance_cost
;
1701 update_avg(&rq
->avg_idle
, delta
);
1703 if (rq
->avg_idle
> max
)
1712 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1714 lockdep_assert_held(&rq
->lock
);
1717 if (p
->sched_contributes_to_load
)
1718 rq
->nr_uninterruptible
--;
1721 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1722 ttwu_do_wakeup(rq
, p
, wake_flags
);
1726 * Called in case the task @p isn't fully descheduled from its runqueue,
1727 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1728 * since all we need to do is flip p->state to TASK_RUNNING, since
1729 * the task is still ->on_rq.
1731 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1736 rq
= __task_rq_lock(p
);
1737 if (task_on_rq_queued(p
)) {
1738 /* check_preempt_curr() may use rq clock */
1739 update_rq_clock(rq
);
1740 ttwu_do_wakeup(rq
, p
, wake_flags
);
1743 __task_rq_unlock(rq
);
1749 void sched_ttwu_pending(void)
1751 struct rq
*rq
= this_rq();
1752 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1753 struct task_struct
*p
;
1754 unsigned long flags
;
1759 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1760 lockdep_pin_lock(&rq
->lock
);
1763 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1764 llist
= llist_next(llist
);
1765 ttwu_do_activate(rq
, p
, 0);
1768 lockdep_unpin_lock(&rq
->lock
);
1769 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1772 void scheduler_ipi(void)
1775 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1776 * TIF_NEED_RESCHED remotely (for the first time) will also send
1779 preempt_fold_need_resched();
1781 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1785 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1786 * traditionally all their work was done from the interrupt return
1787 * path. Now that we actually do some work, we need to make sure
1790 * Some archs already do call them, luckily irq_enter/exit nest
1793 * Arguably we should visit all archs and update all handlers,
1794 * however a fair share of IPIs are still resched only so this would
1795 * somewhat pessimize the simple resched case.
1798 sched_ttwu_pending();
1801 * Check if someone kicked us for doing the nohz idle load balance.
1803 if (unlikely(got_nohz_idle_kick())) {
1804 this_rq()->idle_balance
= 1;
1805 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1810 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1812 struct rq
*rq
= cpu_rq(cpu
);
1814 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1815 if (!set_nr_if_polling(rq
->idle
))
1816 smp_send_reschedule(cpu
);
1818 trace_sched_wake_idle_without_ipi(cpu
);
1822 void wake_up_if_idle(int cpu
)
1824 struct rq
*rq
= cpu_rq(cpu
);
1825 unsigned long flags
;
1829 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1832 if (set_nr_if_polling(rq
->idle
)) {
1833 trace_sched_wake_idle_without_ipi(cpu
);
1835 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1836 if (is_idle_task(rq
->curr
))
1837 smp_send_reschedule(cpu
);
1838 /* Else cpu is not in idle, do nothing here */
1839 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1846 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1848 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1850 #endif /* CONFIG_SMP */
1852 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1854 struct rq
*rq
= cpu_rq(cpu
);
1856 #if defined(CONFIG_SMP)
1857 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1858 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1859 ttwu_queue_remote(p
, cpu
);
1864 raw_spin_lock(&rq
->lock
);
1865 lockdep_pin_lock(&rq
->lock
);
1866 ttwu_do_activate(rq
, p
, 0);
1867 lockdep_unpin_lock(&rq
->lock
);
1868 raw_spin_unlock(&rq
->lock
);
1872 * try_to_wake_up - wake up a thread
1873 * @p: the thread to be awakened
1874 * @state: the mask of task states that can be woken
1875 * @wake_flags: wake modifier flags (WF_*)
1877 * Put it on the run-queue if it's not already there. The "current"
1878 * thread is always on the run-queue (except when the actual
1879 * re-schedule is in progress), and as such you're allowed to do
1880 * the simpler "current->state = TASK_RUNNING" to mark yourself
1881 * runnable without the overhead of this.
1883 * Return: %true if @p was woken up, %false if it was already running.
1884 * or @state didn't match @p's state.
1887 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1889 unsigned long flags
;
1890 int cpu
, success
= 0;
1893 * If we are going to wake up a thread waiting for CONDITION we
1894 * need to ensure that CONDITION=1 done by the caller can not be
1895 * reordered with p->state check below. This pairs with mb() in
1896 * set_current_state() the waiting thread does.
1898 smp_mb__before_spinlock();
1899 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1900 if (!(p
->state
& state
))
1903 trace_sched_waking(p
);
1905 success
= 1; /* we're going to change ->state */
1908 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1913 * If the owning (remote) cpu is still in the middle of schedule() with
1914 * this task as prev, wait until its done referencing the task.
1919 * Pairs with the smp_wmb() in finish_lock_switch().
1923 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1924 p
->state
= TASK_WAKING
;
1926 if (p
->sched_class
->task_waking
)
1927 p
->sched_class
->task_waking(p
);
1929 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1930 if (task_cpu(p
) != cpu
) {
1931 wake_flags
|= WF_MIGRATED
;
1932 set_task_cpu(p
, cpu
);
1934 #endif /* CONFIG_SMP */
1938 ttwu_stat(p
, cpu
, wake_flags
);
1940 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1946 * try_to_wake_up_local - try to wake up a local task with rq lock held
1947 * @p: the thread to be awakened
1949 * Put @p on the run-queue if it's not already there. The caller must
1950 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1953 static void try_to_wake_up_local(struct task_struct
*p
)
1955 struct rq
*rq
= task_rq(p
);
1957 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1958 WARN_ON_ONCE(p
== current
))
1961 lockdep_assert_held(&rq
->lock
);
1963 if (!raw_spin_trylock(&p
->pi_lock
)) {
1965 * This is OK, because current is on_cpu, which avoids it being
1966 * picked for load-balance and preemption/IRQs are still
1967 * disabled avoiding further scheduler activity on it and we've
1968 * not yet picked a replacement task.
1970 lockdep_unpin_lock(&rq
->lock
);
1971 raw_spin_unlock(&rq
->lock
);
1972 raw_spin_lock(&p
->pi_lock
);
1973 raw_spin_lock(&rq
->lock
);
1974 lockdep_pin_lock(&rq
->lock
);
1977 if (!(p
->state
& TASK_NORMAL
))
1980 trace_sched_waking(p
);
1982 if (!task_on_rq_queued(p
))
1983 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1985 ttwu_do_wakeup(rq
, p
, 0);
1986 ttwu_stat(p
, smp_processor_id(), 0);
1988 raw_spin_unlock(&p
->pi_lock
);
1992 * wake_up_process - Wake up a specific process
1993 * @p: The process to be woken up.
1995 * Attempt to wake up the nominated process and move it to the set of runnable
1998 * Return: 1 if the process was woken up, 0 if it was already running.
2000 * It may be assumed that this function implies a write memory barrier before
2001 * changing the task state if and only if any tasks are woken up.
2003 int wake_up_process(struct task_struct
*p
)
2005 WARN_ON(task_is_stopped_or_traced(p
));
2006 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2008 EXPORT_SYMBOL(wake_up_process
);
2010 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2012 return try_to_wake_up(p
, state
, 0);
2016 * This function clears the sched_dl_entity static params.
2018 void __dl_clear_params(struct task_struct
*p
)
2020 struct sched_dl_entity
*dl_se
= &p
->dl
;
2022 dl_se
->dl_runtime
= 0;
2023 dl_se
->dl_deadline
= 0;
2024 dl_se
->dl_period
= 0;
2028 dl_se
->dl_throttled
= 0;
2030 dl_se
->dl_yielded
= 0;
2034 * Perform scheduler related setup for a newly forked process p.
2035 * p is forked by current.
2037 * __sched_fork() is basic setup used by init_idle() too:
2039 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2044 p
->se
.exec_start
= 0;
2045 p
->se
.sum_exec_runtime
= 0;
2046 p
->se
.prev_sum_exec_runtime
= 0;
2047 p
->se
.nr_migrations
= 0;
2049 INIT_LIST_HEAD(&p
->se
.group_node
);
2051 #ifdef CONFIG_SCHEDSTATS
2052 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2055 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2056 init_dl_task_timer(&p
->dl
);
2057 __dl_clear_params(p
);
2059 INIT_LIST_HEAD(&p
->rt
.run_list
);
2061 #ifdef CONFIG_PREEMPT_NOTIFIERS
2062 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2065 #ifdef CONFIG_NUMA_BALANCING
2066 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2067 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2068 p
->mm
->numa_scan_seq
= 0;
2071 if (clone_flags
& CLONE_VM
)
2072 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2074 p
->numa_preferred_nid
= -1;
2076 p
->node_stamp
= 0ULL;
2077 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2078 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2079 p
->numa_work
.next
= &p
->numa_work
;
2080 p
->numa_faults
= NULL
;
2081 p
->last_task_numa_placement
= 0;
2082 p
->last_sum_exec_runtime
= 0;
2084 p
->numa_group
= NULL
;
2085 #endif /* CONFIG_NUMA_BALANCING */
2088 #ifdef CONFIG_NUMA_BALANCING
2089 #ifdef CONFIG_SCHED_DEBUG
2090 void set_numabalancing_state(bool enabled
)
2093 sched_feat_set("NUMA");
2095 sched_feat_set("NO_NUMA");
2098 __read_mostly
bool numabalancing_enabled
;
2100 void set_numabalancing_state(bool enabled
)
2102 numabalancing_enabled
= enabled
;
2104 #endif /* CONFIG_SCHED_DEBUG */
2106 #ifdef CONFIG_PROC_SYSCTL
2107 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2108 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2112 int state
= numabalancing_enabled
;
2114 if (write
&& !capable(CAP_SYS_ADMIN
))
2119 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2123 set_numabalancing_state(state
);
2130 * fork()/clone()-time setup:
2132 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2134 unsigned long flags
;
2135 int cpu
= get_cpu();
2137 __sched_fork(clone_flags
, p
);
2139 * We mark the process as running here. This guarantees that
2140 * nobody will actually run it, and a signal or other external
2141 * event cannot wake it up and insert it on the runqueue either.
2143 p
->state
= TASK_RUNNING
;
2146 * Make sure we do not leak PI boosting priority to the child.
2148 p
->prio
= current
->normal_prio
;
2151 * Revert to default priority/policy on fork if requested.
2153 if (unlikely(p
->sched_reset_on_fork
)) {
2154 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2155 p
->policy
= SCHED_NORMAL
;
2156 p
->static_prio
= NICE_TO_PRIO(0);
2158 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2159 p
->static_prio
= NICE_TO_PRIO(0);
2161 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2165 * We don't need the reset flag anymore after the fork. It has
2166 * fulfilled its duty:
2168 p
->sched_reset_on_fork
= 0;
2171 if (dl_prio(p
->prio
)) {
2174 } else if (rt_prio(p
->prio
)) {
2175 p
->sched_class
= &rt_sched_class
;
2177 p
->sched_class
= &fair_sched_class
;
2180 if (p
->sched_class
->task_fork
)
2181 p
->sched_class
->task_fork(p
);
2184 * The child is not yet in the pid-hash so no cgroup attach races,
2185 * and the cgroup is pinned to this child due to cgroup_fork()
2186 * is ran before sched_fork().
2188 * Silence PROVE_RCU.
2190 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2191 set_task_cpu(p
, cpu
);
2192 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2194 #ifdef CONFIG_SCHED_INFO
2195 if (likely(sched_info_on()))
2196 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2198 #if defined(CONFIG_SMP)
2201 init_task_preempt_count(p
);
2203 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2204 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2211 unsigned long to_ratio(u64 period
, u64 runtime
)
2213 if (runtime
== RUNTIME_INF
)
2217 * Doing this here saves a lot of checks in all
2218 * the calling paths, and returning zero seems
2219 * safe for them anyway.
2224 return div64_u64(runtime
<< 20, period
);
2228 inline struct dl_bw
*dl_bw_of(int i
)
2230 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2231 "sched RCU must be held");
2232 return &cpu_rq(i
)->rd
->dl_bw
;
2235 static inline int dl_bw_cpus(int i
)
2237 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2240 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2241 "sched RCU must be held");
2242 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2248 inline struct dl_bw
*dl_bw_of(int i
)
2250 return &cpu_rq(i
)->dl
.dl_bw
;
2253 static inline int dl_bw_cpus(int i
)
2260 * We must be sure that accepting a new task (or allowing changing the
2261 * parameters of an existing one) is consistent with the bandwidth
2262 * constraints. If yes, this function also accordingly updates the currently
2263 * allocated bandwidth to reflect the new situation.
2265 * This function is called while holding p's rq->lock.
2267 * XXX we should delay bw change until the task's 0-lag point, see
2270 static int dl_overflow(struct task_struct
*p
, int policy
,
2271 const struct sched_attr
*attr
)
2274 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2275 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2276 u64 runtime
= attr
->sched_runtime
;
2277 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2280 if (new_bw
== p
->dl
.dl_bw
)
2284 * Either if a task, enters, leave, or stays -deadline but changes
2285 * its parameters, we may need to update accordingly the total
2286 * allocated bandwidth of the container.
2288 raw_spin_lock(&dl_b
->lock
);
2289 cpus
= dl_bw_cpus(task_cpu(p
));
2290 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2291 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2292 __dl_add(dl_b
, new_bw
);
2294 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2295 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2296 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2297 __dl_add(dl_b
, new_bw
);
2299 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2300 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2303 raw_spin_unlock(&dl_b
->lock
);
2308 extern void init_dl_bw(struct dl_bw
*dl_b
);
2311 * wake_up_new_task - wake up a newly created task for the first time.
2313 * This function will do some initial scheduler statistics housekeeping
2314 * that must be done for every newly created context, then puts the task
2315 * on the runqueue and wakes it.
2317 void wake_up_new_task(struct task_struct
*p
)
2319 unsigned long flags
;
2322 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2325 * Fork balancing, do it here and not earlier because:
2326 * - cpus_allowed can change in the fork path
2327 * - any previously selected cpu might disappear through hotplug
2329 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2332 /* Initialize new task's runnable average */
2333 init_entity_runnable_average(&p
->se
);
2334 rq
= __task_rq_lock(p
);
2335 activate_task(rq
, p
, 0);
2336 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2337 trace_sched_wakeup_new(p
);
2338 check_preempt_curr(rq
, p
, WF_FORK
);
2340 if (p
->sched_class
->task_woken
)
2341 p
->sched_class
->task_woken(rq
, p
);
2343 task_rq_unlock(rq
, p
, &flags
);
2346 #ifdef CONFIG_PREEMPT_NOTIFIERS
2348 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2350 void preempt_notifier_inc(void)
2352 static_key_slow_inc(&preempt_notifier_key
);
2354 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2356 void preempt_notifier_dec(void)
2358 static_key_slow_dec(&preempt_notifier_key
);
2360 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2363 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2364 * @notifier: notifier struct to register
2366 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2368 if (!static_key_false(&preempt_notifier_key
))
2369 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2371 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2373 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2376 * preempt_notifier_unregister - no longer interested in preemption notifications
2377 * @notifier: notifier struct to unregister
2379 * This is *not* safe to call from within a preemption notifier.
2381 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2383 hlist_del(¬ifier
->link
);
2385 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2387 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2389 struct preempt_notifier
*notifier
;
2391 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2392 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2395 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2397 if (static_key_false(&preempt_notifier_key
))
2398 __fire_sched_in_preempt_notifiers(curr
);
2402 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2403 struct task_struct
*next
)
2405 struct preempt_notifier
*notifier
;
2407 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2408 notifier
->ops
->sched_out(notifier
, next
);
2411 static __always_inline
void
2412 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2413 struct task_struct
*next
)
2415 if (static_key_false(&preempt_notifier_key
))
2416 __fire_sched_out_preempt_notifiers(curr
, next
);
2419 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2421 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2426 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2427 struct task_struct
*next
)
2431 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2434 * prepare_task_switch - prepare to switch tasks
2435 * @rq: the runqueue preparing to switch
2436 * @prev: the current task that is being switched out
2437 * @next: the task we are going to switch to.
2439 * This is called with the rq lock held and interrupts off. It must
2440 * be paired with a subsequent finish_task_switch after the context
2443 * prepare_task_switch sets up locking and calls architecture specific
2447 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2448 struct task_struct
*next
)
2450 trace_sched_switch(prev
, next
);
2451 sched_info_switch(rq
, prev
, next
);
2452 perf_event_task_sched_out(prev
, next
);
2453 fire_sched_out_preempt_notifiers(prev
, next
);
2454 prepare_lock_switch(rq
, next
);
2455 prepare_arch_switch(next
);
2459 * finish_task_switch - clean up after a task-switch
2460 * @prev: the thread we just switched away from.
2462 * finish_task_switch must be called after the context switch, paired
2463 * with a prepare_task_switch call before the context switch.
2464 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2465 * and do any other architecture-specific cleanup actions.
2467 * Note that we may have delayed dropping an mm in context_switch(). If
2468 * so, we finish that here outside of the runqueue lock. (Doing it
2469 * with the lock held can cause deadlocks; see schedule() for
2472 * The context switch have flipped the stack from under us and restored the
2473 * local variables which were saved when this task called schedule() in the
2474 * past. prev == current is still correct but we need to recalculate this_rq
2475 * because prev may have moved to another CPU.
2477 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2478 __releases(rq
->lock
)
2480 struct rq
*rq
= this_rq();
2481 struct mm_struct
*mm
= rq
->prev_mm
;
2487 * A task struct has one reference for the use as "current".
2488 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2489 * schedule one last time. The schedule call will never return, and
2490 * the scheduled task must drop that reference.
2491 * The test for TASK_DEAD must occur while the runqueue locks are
2492 * still held, otherwise prev could be scheduled on another cpu, die
2493 * there before we look at prev->state, and then the reference would
2495 * Manfred Spraul <manfred@colorfullife.com>
2497 prev_state
= prev
->state
;
2498 vtime_task_switch(prev
);
2499 perf_event_task_sched_in(prev
, current
);
2500 finish_lock_switch(rq
, prev
);
2501 finish_arch_post_lock_switch();
2503 fire_sched_in_preempt_notifiers(current
);
2506 if (unlikely(prev_state
== TASK_DEAD
)) {
2507 if (prev
->sched_class
->task_dead
)
2508 prev
->sched_class
->task_dead(prev
);
2511 * Remove function-return probe instances associated with this
2512 * task and put them back on the free list.
2514 kprobe_flush_task(prev
);
2515 put_task_struct(prev
);
2518 tick_nohz_task_switch(current
);
2524 /* rq->lock is NOT held, but preemption is disabled */
2525 static void __balance_callback(struct rq
*rq
)
2527 struct callback_head
*head
, *next
;
2528 void (*func
)(struct rq
*rq
);
2529 unsigned long flags
;
2531 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2532 head
= rq
->balance_callback
;
2533 rq
->balance_callback
= NULL
;
2535 func
= (void (*)(struct rq
*))head
->func
;
2542 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2545 static inline void balance_callback(struct rq
*rq
)
2547 if (unlikely(rq
->balance_callback
))
2548 __balance_callback(rq
);
2553 static inline void balance_callback(struct rq
*rq
)
2560 * schedule_tail - first thing a freshly forked thread must call.
2561 * @prev: the thread we just switched away from.
2563 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2564 __releases(rq
->lock
)
2568 /* finish_task_switch() drops rq->lock and enables preemtion */
2570 rq
= finish_task_switch(prev
);
2571 balance_callback(rq
);
2574 if (current
->set_child_tid
)
2575 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2579 * context_switch - switch to the new MM and the new thread's register state.
2581 static inline struct rq
*
2582 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2583 struct task_struct
*next
)
2585 struct mm_struct
*mm
, *oldmm
;
2587 prepare_task_switch(rq
, prev
, next
);
2590 oldmm
= prev
->active_mm
;
2592 * For paravirt, this is coupled with an exit in switch_to to
2593 * combine the page table reload and the switch backend into
2596 arch_start_context_switch(prev
);
2599 next
->active_mm
= oldmm
;
2600 atomic_inc(&oldmm
->mm_count
);
2601 enter_lazy_tlb(oldmm
, next
);
2603 switch_mm(oldmm
, mm
, next
);
2606 prev
->active_mm
= NULL
;
2607 rq
->prev_mm
= oldmm
;
2610 * Since the runqueue lock will be released by the next
2611 * task (which is an invalid locking op but in the case
2612 * of the scheduler it's an obvious special-case), so we
2613 * do an early lockdep release here:
2615 lockdep_unpin_lock(&rq
->lock
);
2616 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2618 /* Here we just switch the register state and the stack. */
2619 switch_to(prev
, next
, prev
);
2622 return finish_task_switch(prev
);
2626 * nr_running and nr_context_switches:
2628 * externally visible scheduler statistics: current number of runnable
2629 * threads, total number of context switches performed since bootup.
2631 unsigned long nr_running(void)
2633 unsigned long i
, sum
= 0;
2635 for_each_online_cpu(i
)
2636 sum
+= cpu_rq(i
)->nr_running
;
2642 * Check if only the current task is running on the cpu.
2644 bool single_task_running(void)
2646 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2651 EXPORT_SYMBOL(single_task_running
);
2653 unsigned long long nr_context_switches(void)
2656 unsigned long long sum
= 0;
2658 for_each_possible_cpu(i
)
2659 sum
+= cpu_rq(i
)->nr_switches
;
2664 unsigned long nr_iowait(void)
2666 unsigned long i
, sum
= 0;
2668 for_each_possible_cpu(i
)
2669 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2674 unsigned long nr_iowait_cpu(int cpu
)
2676 struct rq
*this = cpu_rq(cpu
);
2677 return atomic_read(&this->nr_iowait
);
2680 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2682 struct rq
*rq
= this_rq();
2683 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2684 *load
= rq
->load
.weight
;
2690 * sched_exec - execve() is a valuable balancing opportunity, because at
2691 * this point the task has the smallest effective memory and cache footprint.
2693 void sched_exec(void)
2695 struct task_struct
*p
= current
;
2696 unsigned long flags
;
2699 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2700 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2701 if (dest_cpu
== smp_processor_id())
2704 if (likely(cpu_active(dest_cpu
))) {
2705 struct migration_arg arg
= { p
, dest_cpu
};
2707 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2708 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2712 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2717 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2718 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2720 EXPORT_PER_CPU_SYMBOL(kstat
);
2721 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2724 * Return accounted runtime for the task.
2725 * In case the task is currently running, return the runtime plus current's
2726 * pending runtime that have not been accounted yet.
2728 unsigned long long task_sched_runtime(struct task_struct
*p
)
2730 unsigned long flags
;
2734 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2736 * 64-bit doesn't need locks to atomically read a 64bit value.
2737 * So we have a optimization chance when the task's delta_exec is 0.
2738 * Reading ->on_cpu is racy, but this is ok.
2740 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2741 * If we race with it entering cpu, unaccounted time is 0. This is
2742 * indistinguishable from the read occurring a few cycles earlier.
2743 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2744 * been accounted, so we're correct here as well.
2746 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2747 return p
->se
.sum_exec_runtime
;
2750 rq
= task_rq_lock(p
, &flags
);
2752 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2753 * project cycles that may never be accounted to this
2754 * thread, breaking clock_gettime().
2756 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2757 update_rq_clock(rq
);
2758 p
->sched_class
->update_curr(rq
);
2760 ns
= p
->se
.sum_exec_runtime
;
2761 task_rq_unlock(rq
, p
, &flags
);
2767 * This function gets called by the timer code, with HZ frequency.
2768 * We call it with interrupts disabled.
2770 void scheduler_tick(void)
2772 int cpu
= smp_processor_id();
2773 struct rq
*rq
= cpu_rq(cpu
);
2774 struct task_struct
*curr
= rq
->curr
;
2778 raw_spin_lock(&rq
->lock
);
2779 update_rq_clock(rq
);
2780 curr
->sched_class
->task_tick(rq
, curr
, 0);
2781 update_cpu_load_active(rq
);
2782 calc_global_load_tick(rq
);
2783 raw_spin_unlock(&rq
->lock
);
2785 perf_event_task_tick();
2788 rq
->idle_balance
= idle_cpu(cpu
);
2789 trigger_load_balance(rq
);
2791 rq_last_tick_reset(rq
);
2794 #ifdef CONFIG_NO_HZ_FULL
2796 * scheduler_tick_max_deferment
2798 * Keep at least one tick per second when a single
2799 * active task is running because the scheduler doesn't
2800 * yet completely support full dynticks environment.
2802 * This makes sure that uptime, CFS vruntime, load
2803 * balancing, etc... continue to move forward, even
2804 * with a very low granularity.
2806 * Return: Maximum deferment in nanoseconds.
2808 u64
scheduler_tick_max_deferment(void)
2810 struct rq
*rq
= this_rq();
2811 unsigned long next
, now
= READ_ONCE(jiffies
);
2813 next
= rq
->last_sched_tick
+ HZ
;
2815 if (time_before_eq(next
, now
))
2818 return jiffies_to_nsecs(next
- now
);
2822 notrace
unsigned long get_parent_ip(unsigned long addr
)
2824 if (in_lock_functions(addr
)) {
2825 addr
= CALLER_ADDR2
;
2826 if (in_lock_functions(addr
))
2827 addr
= CALLER_ADDR3
;
2832 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2833 defined(CONFIG_PREEMPT_TRACER))
2835 void preempt_count_add(int val
)
2837 #ifdef CONFIG_DEBUG_PREEMPT
2841 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2844 __preempt_count_add(val
);
2845 #ifdef CONFIG_DEBUG_PREEMPT
2847 * Spinlock count overflowing soon?
2849 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2852 if (preempt_count() == val
) {
2853 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2854 #ifdef CONFIG_DEBUG_PREEMPT
2855 current
->preempt_disable_ip
= ip
;
2857 trace_preempt_off(CALLER_ADDR0
, ip
);
2860 EXPORT_SYMBOL(preempt_count_add
);
2861 NOKPROBE_SYMBOL(preempt_count_add
);
2863 void preempt_count_sub(int val
)
2865 #ifdef CONFIG_DEBUG_PREEMPT
2869 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2872 * Is the spinlock portion underflowing?
2874 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2875 !(preempt_count() & PREEMPT_MASK
)))
2879 if (preempt_count() == val
)
2880 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2881 __preempt_count_sub(val
);
2883 EXPORT_SYMBOL(preempt_count_sub
);
2884 NOKPROBE_SYMBOL(preempt_count_sub
);
2889 * Print scheduling while atomic bug:
2891 static noinline
void __schedule_bug(struct task_struct
*prev
)
2893 if (oops_in_progress
)
2896 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2897 prev
->comm
, prev
->pid
, preempt_count());
2899 debug_show_held_locks(prev
);
2901 if (irqs_disabled())
2902 print_irqtrace_events(prev
);
2903 #ifdef CONFIG_DEBUG_PREEMPT
2904 if (in_atomic_preempt_off()) {
2905 pr_err("Preemption disabled at:");
2906 print_ip_sym(current
->preempt_disable_ip
);
2911 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2915 * Various schedule()-time debugging checks and statistics:
2917 static inline void schedule_debug(struct task_struct
*prev
)
2919 #ifdef CONFIG_SCHED_STACK_END_CHECK
2920 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2923 * Test if we are atomic. Since do_exit() needs to call into
2924 * schedule() atomically, we ignore that path. Otherwise whine
2925 * if we are scheduling when we should not.
2927 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2928 __schedule_bug(prev
);
2931 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2933 schedstat_inc(this_rq(), sched_count
);
2937 * Pick up the highest-prio task:
2939 static inline struct task_struct
*
2940 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2942 const struct sched_class
*class = &fair_sched_class
;
2943 struct task_struct
*p
;
2946 * Optimization: we know that if all tasks are in
2947 * the fair class we can call that function directly:
2949 if (likely(prev
->sched_class
== class &&
2950 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2951 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2952 if (unlikely(p
== RETRY_TASK
))
2955 /* assumes fair_sched_class->next == idle_sched_class */
2957 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2963 for_each_class(class) {
2964 p
= class->pick_next_task(rq
, prev
);
2966 if (unlikely(p
== RETRY_TASK
))
2972 BUG(); /* the idle class will always have a runnable task */
2976 * __schedule() is the main scheduler function.
2978 * The main means of driving the scheduler and thus entering this function are:
2980 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2982 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2983 * paths. For example, see arch/x86/entry_64.S.
2985 * To drive preemption between tasks, the scheduler sets the flag in timer
2986 * interrupt handler scheduler_tick().
2988 * 3. Wakeups don't really cause entry into schedule(). They add a
2989 * task to the run-queue and that's it.
2991 * Now, if the new task added to the run-queue preempts the current
2992 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2993 * called on the nearest possible occasion:
2995 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2997 * - in syscall or exception context, at the next outmost
2998 * preempt_enable(). (this might be as soon as the wake_up()'s
3001 * - in IRQ context, return from interrupt-handler to
3002 * preemptible context
3004 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3007 * - cond_resched() call
3008 * - explicit schedule() call
3009 * - return from syscall or exception to user-space
3010 * - return from interrupt-handler to user-space
3012 * WARNING: must be called with preemption disabled!
3014 static void __sched
__schedule(void)
3016 struct task_struct
*prev
, *next
;
3017 unsigned long *switch_count
;
3021 cpu
= smp_processor_id();
3023 rcu_note_context_switch();
3026 schedule_debug(prev
);
3028 if (sched_feat(HRTICK
))
3032 * Make sure that signal_pending_state()->signal_pending() below
3033 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3034 * done by the caller to avoid the race with signal_wake_up().
3036 smp_mb__before_spinlock();
3037 raw_spin_lock_irq(&rq
->lock
);
3038 lockdep_pin_lock(&rq
->lock
);
3040 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3042 switch_count
= &prev
->nivcsw
;
3043 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3044 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3045 prev
->state
= TASK_RUNNING
;
3047 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3051 * If a worker went to sleep, notify and ask workqueue
3052 * whether it wants to wake up a task to maintain
3055 if (prev
->flags
& PF_WQ_WORKER
) {
3056 struct task_struct
*to_wakeup
;
3058 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3060 try_to_wake_up_local(to_wakeup
);
3063 switch_count
= &prev
->nvcsw
;
3066 if (task_on_rq_queued(prev
))
3067 update_rq_clock(rq
);
3069 next
= pick_next_task(rq
, prev
);
3070 clear_tsk_need_resched(prev
);
3071 clear_preempt_need_resched();
3072 rq
->clock_skip_update
= 0;
3074 if (likely(prev
!= next
)) {
3079 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3082 lockdep_unpin_lock(&rq
->lock
);
3083 raw_spin_unlock_irq(&rq
->lock
);
3086 balance_callback(rq
);
3089 static inline void sched_submit_work(struct task_struct
*tsk
)
3091 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3094 * If we are going to sleep and we have plugged IO queued,
3095 * make sure to submit it to avoid deadlocks.
3097 if (blk_needs_flush_plug(tsk
))
3098 blk_schedule_flush_plug(tsk
);
3101 asmlinkage __visible
void __sched
schedule(void)
3103 struct task_struct
*tsk
= current
;
3105 sched_submit_work(tsk
);
3109 sched_preempt_enable_no_resched();
3110 } while (need_resched());
3112 EXPORT_SYMBOL(schedule
);
3114 #ifdef CONFIG_CONTEXT_TRACKING
3115 asmlinkage __visible
void __sched
schedule_user(void)
3118 * If we come here after a random call to set_need_resched(),
3119 * or we have been woken up remotely but the IPI has not yet arrived,
3120 * we haven't yet exited the RCU idle mode. Do it here manually until
3121 * we find a better solution.
3123 * NB: There are buggy callers of this function. Ideally we
3124 * should warn if prev_state != CONTEXT_USER, but that will trigger
3125 * too frequently to make sense yet.
3127 enum ctx_state prev_state
= exception_enter();
3129 exception_exit(prev_state
);
3134 * schedule_preempt_disabled - called with preemption disabled
3136 * Returns with preemption disabled. Note: preempt_count must be 1
3138 void __sched
schedule_preempt_disabled(void)
3140 sched_preempt_enable_no_resched();
3145 static void __sched notrace
preempt_schedule_common(void)
3148 preempt_active_enter();
3150 preempt_active_exit();
3153 * Check again in case we missed a preemption opportunity
3154 * between schedule and now.
3156 } while (need_resched());
3159 #ifdef CONFIG_PREEMPT
3161 * this is the entry point to schedule() from in-kernel preemption
3162 * off of preempt_enable. Kernel preemptions off return from interrupt
3163 * occur there and call schedule directly.
3165 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3168 * If there is a non-zero preempt_count or interrupts are disabled,
3169 * we do not want to preempt the current task. Just return..
3171 if (likely(!preemptible()))
3174 preempt_schedule_common();
3176 NOKPROBE_SYMBOL(preempt_schedule
);
3177 EXPORT_SYMBOL(preempt_schedule
);
3180 * preempt_schedule_notrace - preempt_schedule called by tracing
3182 * The tracing infrastructure uses preempt_enable_notrace to prevent
3183 * recursion and tracing preempt enabling caused by the tracing
3184 * infrastructure itself. But as tracing can happen in areas coming
3185 * from userspace or just about to enter userspace, a preempt enable
3186 * can occur before user_exit() is called. This will cause the scheduler
3187 * to be called when the system is still in usermode.
3189 * To prevent this, the preempt_enable_notrace will use this function
3190 * instead of preempt_schedule() to exit user context if needed before
3191 * calling the scheduler.
3193 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3195 enum ctx_state prev_ctx
;
3197 if (likely(!preemptible()))
3202 * Use raw __prempt_count() ops that don't call function.
3203 * We can't call functions before disabling preemption which
3204 * disarm preemption tracing recursions.
3206 __preempt_count_add(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3209 * Needs preempt disabled in case user_exit() is traced
3210 * and the tracer calls preempt_enable_notrace() causing
3211 * an infinite recursion.
3213 prev_ctx
= exception_enter();
3215 exception_exit(prev_ctx
);
3218 __preempt_count_sub(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
3219 } while (need_resched());
3221 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3223 #endif /* CONFIG_PREEMPT */
3226 * this is the entry point to schedule() from kernel preemption
3227 * off of irq context.
3228 * Note, that this is called and return with irqs disabled. This will
3229 * protect us against recursive calling from irq.
3231 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3233 enum ctx_state prev_state
;
3235 /* Catch callers which need to be fixed */
3236 BUG_ON(preempt_count() || !irqs_disabled());
3238 prev_state
= exception_enter();
3241 preempt_active_enter();
3244 local_irq_disable();
3245 preempt_active_exit();
3246 } while (need_resched());
3248 exception_exit(prev_state
);
3251 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3254 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3256 EXPORT_SYMBOL(default_wake_function
);
3258 #ifdef CONFIG_RT_MUTEXES
3261 * rt_mutex_setprio - set the current priority of a task
3263 * @prio: prio value (kernel-internal form)
3265 * This function changes the 'effective' priority of a task. It does
3266 * not touch ->normal_prio like __setscheduler().
3268 * Used by the rt_mutex code to implement priority inheritance
3269 * logic. Call site only calls if the priority of the task changed.
3271 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3273 int oldprio
, queued
, running
, enqueue_flag
= 0;
3275 const struct sched_class
*prev_class
;
3277 BUG_ON(prio
> MAX_PRIO
);
3279 rq
= __task_rq_lock(p
);
3282 * Idle task boosting is a nono in general. There is one
3283 * exception, when PREEMPT_RT and NOHZ is active:
3285 * The idle task calls get_next_timer_interrupt() and holds
3286 * the timer wheel base->lock on the CPU and another CPU wants
3287 * to access the timer (probably to cancel it). We can safely
3288 * ignore the boosting request, as the idle CPU runs this code
3289 * with interrupts disabled and will complete the lock
3290 * protected section without being interrupted. So there is no
3291 * real need to boost.
3293 if (unlikely(p
== rq
->idle
)) {
3294 WARN_ON(p
!= rq
->curr
);
3295 WARN_ON(p
->pi_blocked_on
);
3299 trace_sched_pi_setprio(p
, prio
);
3301 prev_class
= p
->sched_class
;
3302 queued
= task_on_rq_queued(p
);
3303 running
= task_current(rq
, p
);
3305 dequeue_task(rq
, p
, 0);
3307 put_prev_task(rq
, p
);
3310 * Boosting condition are:
3311 * 1. -rt task is running and holds mutex A
3312 * --> -dl task blocks on mutex A
3314 * 2. -dl task is running and holds mutex A
3315 * --> -dl task blocks on mutex A and could preempt the
3318 if (dl_prio(prio
)) {
3319 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3320 if (!dl_prio(p
->normal_prio
) ||
3321 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3322 p
->dl
.dl_boosted
= 1;
3323 enqueue_flag
= ENQUEUE_REPLENISH
;
3325 p
->dl
.dl_boosted
= 0;
3326 p
->sched_class
= &dl_sched_class
;
3327 } else if (rt_prio(prio
)) {
3328 if (dl_prio(oldprio
))
3329 p
->dl
.dl_boosted
= 0;
3331 enqueue_flag
= ENQUEUE_HEAD
;
3332 p
->sched_class
= &rt_sched_class
;
3334 if (dl_prio(oldprio
))
3335 p
->dl
.dl_boosted
= 0;
3336 if (rt_prio(oldprio
))
3338 p
->sched_class
= &fair_sched_class
;
3344 p
->sched_class
->set_curr_task(rq
);
3346 enqueue_task(rq
, p
, enqueue_flag
);
3348 check_class_changed(rq
, p
, prev_class
, oldprio
);
3350 preempt_disable(); /* avoid rq from going away on us */
3351 __task_rq_unlock(rq
);
3353 balance_callback(rq
);
3358 void set_user_nice(struct task_struct
*p
, long nice
)
3360 int old_prio
, delta
, queued
;
3361 unsigned long flags
;
3364 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3367 * We have to be careful, if called from sys_setpriority(),
3368 * the task might be in the middle of scheduling on another CPU.
3370 rq
= task_rq_lock(p
, &flags
);
3372 * The RT priorities are set via sched_setscheduler(), but we still
3373 * allow the 'normal' nice value to be set - but as expected
3374 * it wont have any effect on scheduling until the task is
3375 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3377 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3378 p
->static_prio
= NICE_TO_PRIO(nice
);
3381 queued
= task_on_rq_queued(p
);
3383 dequeue_task(rq
, p
, 0);
3385 p
->static_prio
= NICE_TO_PRIO(nice
);
3388 p
->prio
= effective_prio(p
);
3389 delta
= p
->prio
- old_prio
;
3392 enqueue_task(rq
, p
, 0);
3394 * If the task increased its priority or is running and
3395 * lowered its priority, then reschedule its CPU:
3397 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3401 task_rq_unlock(rq
, p
, &flags
);
3403 EXPORT_SYMBOL(set_user_nice
);
3406 * can_nice - check if a task can reduce its nice value
3410 int can_nice(const struct task_struct
*p
, const int nice
)
3412 /* convert nice value [19,-20] to rlimit style value [1,40] */
3413 int nice_rlim
= nice_to_rlimit(nice
);
3415 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3416 capable(CAP_SYS_NICE
));
3419 #ifdef __ARCH_WANT_SYS_NICE
3422 * sys_nice - change the priority of the current process.
3423 * @increment: priority increment
3425 * sys_setpriority is a more generic, but much slower function that
3426 * does similar things.
3428 SYSCALL_DEFINE1(nice
, int, increment
)
3433 * Setpriority might change our priority at the same moment.
3434 * We don't have to worry. Conceptually one call occurs first
3435 * and we have a single winner.
3437 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3438 nice
= task_nice(current
) + increment
;
3440 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3441 if (increment
< 0 && !can_nice(current
, nice
))
3444 retval
= security_task_setnice(current
, nice
);
3448 set_user_nice(current
, nice
);
3455 * task_prio - return the priority value of a given task.
3456 * @p: the task in question.
3458 * Return: The priority value as seen by users in /proc.
3459 * RT tasks are offset by -200. Normal tasks are centered
3460 * around 0, value goes from -16 to +15.
3462 int task_prio(const struct task_struct
*p
)
3464 return p
->prio
- MAX_RT_PRIO
;
3468 * idle_cpu - is a given cpu idle currently?
3469 * @cpu: the processor in question.
3471 * Return: 1 if the CPU is currently idle. 0 otherwise.
3473 int idle_cpu(int cpu
)
3475 struct rq
*rq
= cpu_rq(cpu
);
3477 if (rq
->curr
!= rq
->idle
)
3484 if (!llist_empty(&rq
->wake_list
))
3492 * idle_task - return the idle task for a given cpu.
3493 * @cpu: the processor in question.
3495 * Return: The idle task for the cpu @cpu.
3497 struct task_struct
*idle_task(int cpu
)
3499 return cpu_rq(cpu
)->idle
;
3503 * find_process_by_pid - find a process with a matching PID value.
3504 * @pid: the pid in question.
3506 * The task of @pid, if found. %NULL otherwise.
3508 static struct task_struct
*find_process_by_pid(pid_t pid
)
3510 return pid
? find_task_by_vpid(pid
) : current
;
3514 * This function initializes the sched_dl_entity of a newly becoming
3515 * SCHED_DEADLINE task.
3517 * Only the static values are considered here, the actual runtime and the
3518 * absolute deadline will be properly calculated when the task is enqueued
3519 * for the first time with its new policy.
3522 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3524 struct sched_dl_entity
*dl_se
= &p
->dl
;
3526 dl_se
->dl_runtime
= attr
->sched_runtime
;
3527 dl_se
->dl_deadline
= attr
->sched_deadline
;
3528 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3529 dl_se
->flags
= attr
->sched_flags
;
3530 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3533 * Changing the parameters of a task is 'tricky' and we're not doing
3534 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3536 * What we SHOULD do is delay the bandwidth release until the 0-lag
3537 * point. This would include retaining the task_struct until that time
3538 * and change dl_overflow() to not immediately decrement the current
3541 * Instead we retain the current runtime/deadline and let the new
3542 * parameters take effect after the current reservation period lapses.
3543 * This is safe (albeit pessimistic) because the 0-lag point is always
3544 * before the current scheduling deadline.
3546 * We can still have temporary overloads because we do not delay the
3547 * change in bandwidth until that time; so admission control is
3548 * not on the safe side. It does however guarantee tasks will never
3549 * consume more than promised.
3554 * sched_setparam() passes in -1 for its policy, to let the functions
3555 * it calls know not to change it.
3557 #define SETPARAM_POLICY -1
3559 static void __setscheduler_params(struct task_struct
*p
,
3560 const struct sched_attr
*attr
)
3562 int policy
= attr
->sched_policy
;
3564 if (policy
== SETPARAM_POLICY
)
3569 if (dl_policy(policy
))
3570 __setparam_dl(p
, attr
);
3571 else if (fair_policy(policy
))
3572 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3575 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3576 * !rt_policy. Always setting this ensures that things like
3577 * getparam()/getattr() don't report silly values for !rt tasks.
3579 p
->rt_priority
= attr
->sched_priority
;
3580 p
->normal_prio
= normal_prio(p
);
3584 /* Actually do priority change: must hold pi & rq lock. */
3585 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3586 const struct sched_attr
*attr
, bool keep_boost
)
3588 __setscheduler_params(p
, attr
);
3591 * Keep a potential priority boosting if called from
3592 * sched_setscheduler().
3595 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3597 p
->prio
= normal_prio(p
);
3599 if (dl_prio(p
->prio
))
3600 p
->sched_class
= &dl_sched_class
;
3601 else if (rt_prio(p
->prio
))
3602 p
->sched_class
= &rt_sched_class
;
3604 p
->sched_class
= &fair_sched_class
;
3608 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3610 struct sched_dl_entity
*dl_se
= &p
->dl
;
3612 attr
->sched_priority
= p
->rt_priority
;
3613 attr
->sched_runtime
= dl_se
->dl_runtime
;
3614 attr
->sched_deadline
= dl_se
->dl_deadline
;
3615 attr
->sched_period
= dl_se
->dl_period
;
3616 attr
->sched_flags
= dl_se
->flags
;
3620 * This function validates the new parameters of a -deadline task.
3621 * We ask for the deadline not being zero, and greater or equal
3622 * than the runtime, as well as the period of being zero or
3623 * greater than deadline. Furthermore, we have to be sure that
3624 * user parameters are above the internal resolution of 1us (we
3625 * check sched_runtime only since it is always the smaller one) and
3626 * below 2^63 ns (we have to check both sched_deadline and
3627 * sched_period, as the latter can be zero).
3630 __checkparam_dl(const struct sched_attr
*attr
)
3633 if (attr
->sched_deadline
== 0)
3637 * Since we truncate DL_SCALE bits, make sure we're at least
3640 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3644 * Since we use the MSB for wrap-around and sign issues, make
3645 * sure it's not set (mind that period can be equal to zero).
3647 if (attr
->sched_deadline
& (1ULL << 63) ||
3648 attr
->sched_period
& (1ULL << 63))
3651 /* runtime <= deadline <= period (if period != 0) */
3652 if ((attr
->sched_period
!= 0 &&
3653 attr
->sched_period
< attr
->sched_deadline
) ||
3654 attr
->sched_deadline
< attr
->sched_runtime
)
3661 * check the target process has a UID that matches the current process's
3663 static bool check_same_owner(struct task_struct
*p
)
3665 const struct cred
*cred
= current_cred(), *pcred
;
3669 pcred
= __task_cred(p
);
3670 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3671 uid_eq(cred
->euid
, pcred
->uid
));
3676 static bool dl_param_changed(struct task_struct
*p
,
3677 const struct sched_attr
*attr
)
3679 struct sched_dl_entity
*dl_se
= &p
->dl
;
3681 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3682 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3683 dl_se
->dl_period
!= attr
->sched_period
||
3684 dl_se
->flags
!= attr
->sched_flags
)
3690 static int __sched_setscheduler(struct task_struct
*p
,
3691 const struct sched_attr
*attr
,
3694 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3695 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3696 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3697 int new_effective_prio
, policy
= attr
->sched_policy
;
3698 unsigned long flags
;
3699 const struct sched_class
*prev_class
;
3703 /* may grab non-irq protected spin_locks */
3704 BUG_ON(in_interrupt());
3706 /* double check policy once rq lock held */
3708 reset_on_fork
= p
->sched_reset_on_fork
;
3709 policy
= oldpolicy
= p
->policy
;
3711 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3713 if (policy
!= SCHED_DEADLINE
&&
3714 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3715 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3716 policy
!= SCHED_IDLE
)
3720 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3724 * Valid priorities for SCHED_FIFO and SCHED_RR are
3725 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3726 * SCHED_BATCH and SCHED_IDLE is 0.
3728 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3729 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3731 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3732 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3736 * Allow unprivileged RT tasks to decrease priority:
3738 if (user
&& !capable(CAP_SYS_NICE
)) {
3739 if (fair_policy(policy
)) {
3740 if (attr
->sched_nice
< task_nice(p
) &&
3741 !can_nice(p
, attr
->sched_nice
))
3745 if (rt_policy(policy
)) {
3746 unsigned long rlim_rtprio
=
3747 task_rlimit(p
, RLIMIT_RTPRIO
);
3749 /* can't set/change the rt policy */
3750 if (policy
!= p
->policy
&& !rlim_rtprio
)
3753 /* can't increase priority */
3754 if (attr
->sched_priority
> p
->rt_priority
&&
3755 attr
->sched_priority
> rlim_rtprio
)
3760 * Can't set/change SCHED_DEADLINE policy at all for now
3761 * (safest behavior); in the future we would like to allow
3762 * unprivileged DL tasks to increase their relative deadline
3763 * or reduce their runtime (both ways reducing utilization)
3765 if (dl_policy(policy
))
3769 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3770 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3772 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3773 if (!can_nice(p
, task_nice(p
)))
3777 /* can't change other user's priorities */
3778 if (!check_same_owner(p
))
3781 /* Normal users shall not reset the sched_reset_on_fork flag */
3782 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3787 retval
= security_task_setscheduler(p
);
3793 * make sure no PI-waiters arrive (or leave) while we are
3794 * changing the priority of the task:
3796 * To be able to change p->policy safely, the appropriate
3797 * runqueue lock must be held.
3799 rq
= task_rq_lock(p
, &flags
);
3802 * Changing the policy of the stop threads its a very bad idea
3804 if (p
== rq
->stop
) {
3805 task_rq_unlock(rq
, p
, &flags
);
3810 * If not changing anything there's no need to proceed further,
3811 * but store a possible modification of reset_on_fork.
3813 if (unlikely(policy
== p
->policy
)) {
3814 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3816 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3818 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3821 p
->sched_reset_on_fork
= reset_on_fork
;
3822 task_rq_unlock(rq
, p
, &flags
);
3828 #ifdef CONFIG_RT_GROUP_SCHED
3830 * Do not allow realtime tasks into groups that have no runtime
3833 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3834 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3835 !task_group_is_autogroup(task_group(p
))) {
3836 task_rq_unlock(rq
, p
, &flags
);
3841 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3842 cpumask_t
*span
= rq
->rd
->span
;
3845 * Don't allow tasks with an affinity mask smaller than
3846 * the entire root_domain to become SCHED_DEADLINE. We
3847 * will also fail if there's no bandwidth available.
3849 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3850 rq
->rd
->dl_bw
.bw
== 0) {
3851 task_rq_unlock(rq
, p
, &flags
);
3858 /* recheck policy now with rq lock held */
3859 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3860 policy
= oldpolicy
= -1;
3861 task_rq_unlock(rq
, p
, &flags
);
3866 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3867 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3870 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3871 task_rq_unlock(rq
, p
, &flags
);
3875 p
->sched_reset_on_fork
= reset_on_fork
;
3880 * Take priority boosted tasks into account. If the new
3881 * effective priority is unchanged, we just store the new
3882 * normal parameters and do not touch the scheduler class and
3883 * the runqueue. This will be done when the task deboost
3886 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
3887 if (new_effective_prio
== oldprio
) {
3888 __setscheduler_params(p
, attr
);
3889 task_rq_unlock(rq
, p
, &flags
);
3894 queued
= task_on_rq_queued(p
);
3895 running
= task_current(rq
, p
);
3897 dequeue_task(rq
, p
, 0);
3899 put_prev_task(rq
, p
);
3901 prev_class
= p
->sched_class
;
3902 __setscheduler(rq
, p
, attr
, pi
);
3905 p
->sched_class
->set_curr_task(rq
);
3908 * We enqueue to tail when the priority of a task is
3909 * increased (user space view).
3911 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3914 check_class_changed(rq
, p
, prev_class
, oldprio
);
3915 preempt_disable(); /* avoid rq from going away on us */
3916 task_rq_unlock(rq
, p
, &flags
);
3919 rt_mutex_adjust_pi(p
);
3922 * Run balance callbacks after we've adjusted the PI chain.
3924 balance_callback(rq
);
3930 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3931 const struct sched_param
*param
, bool check
)
3933 struct sched_attr attr
= {
3934 .sched_policy
= policy
,
3935 .sched_priority
= param
->sched_priority
,
3936 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3939 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3940 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3941 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3942 policy
&= ~SCHED_RESET_ON_FORK
;
3943 attr
.sched_policy
= policy
;
3946 return __sched_setscheduler(p
, &attr
, check
, true);
3949 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3950 * @p: the task in question.
3951 * @policy: new policy.
3952 * @param: structure containing the new RT priority.
3954 * Return: 0 on success. An error code otherwise.
3956 * NOTE that the task may be already dead.
3958 int sched_setscheduler(struct task_struct
*p
, int policy
,
3959 const struct sched_param
*param
)
3961 return _sched_setscheduler(p
, policy
, param
, true);
3963 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3965 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3967 return __sched_setscheduler(p
, attr
, true, true);
3969 EXPORT_SYMBOL_GPL(sched_setattr
);
3972 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3973 * @p: the task in question.
3974 * @policy: new policy.
3975 * @param: structure containing the new RT priority.
3977 * Just like sched_setscheduler, only don't bother checking if the
3978 * current context has permission. For example, this is needed in
3979 * stop_machine(): we create temporary high priority worker threads,
3980 * but our caller might not have that capability.
3982 * Return: 0 on success. An error code otherwise.
3984 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3985 const struct sched_param
*param
)
3987 return _sched_setscheduler(p
, policy
, param
, false);
3991 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3993 struct sched_param lparam
;
3994 struct task_struct
*p
;
3997 if (!param
|| pid
< 0)
3999 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4004 p
= find_process_by_pid(pid
);
4006 retval
= sched_setscheduler(p
, policy
, &lparam
);
4013 * Mimics kernel/events/core.c perf_copy_attr().
4015 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4016 struct sched_attr
*attr
)
4021 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4025 * zero the full structure, so that a short copy will be nice.
4027 memset(attr
, 0, sizeof(*attr
));
4029 ret
= get_user(size
, &uattr
->size
);
4033 if (size
> PAGE_SIZE
) /* silly large */
4036 if (!size
) /* abi compat */
4037 size
= SCHED_ATTR_SIZE_VER0
;
4039 if (size
< SCHED_ATTR_SIZE_VER0
)
4043 * If we're handed a bigger struct than we know of,
4044 * ensure all the unknown bits are 0 - i.e. new
4045 * user-space does not rely on any kernel feature
4046 * extensions we dont know about yet.
4048 if (size
> sizeof(*attr
)) {
4049 unsigned char __user
*addr
;
4050 unsigned char __user
*end
;
4053 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4054 end
= (void __user
*)uattr
+ size
;
4056 for (; addr
< end
; addr
++) {
4057 ret
= get_user(val
, addr
);
4063 size
= sizeof(*attr
);
4066 ret
= copy_from_user(attr
, uattr
, size
);
4071 * XXX: do we want to be lenient like existing syscalls; or do we want
4072 * to be strict and return an error on out-of-bounds values?
4074 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4079 put_user(sizeof(*attr
), &uattr
->size
);
4084 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4085 * @pid: the pid in question.
4086 * @policy: new policy.
4087 * @param: structure containing the new RT priority.
4089 * Return: 0 on success. An error code otherwise.
4091 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4092 struct sched_param __user
*, param
)
4094 /* negative values for policy are not valid */
4098 return do_sched_setscheduler(pid
, policy
, param
);
4102 * sys_sched_setparam - set/change the RT priority of a thread
4103 * @pid: the pid in question.
4104 * @param: structure containing the new RT priority.
4106 * Return: 0 on success. An error code otherwise.
4108 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4110 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4114 * sys_sched_setattr - same as above, but with extended sched_attr
4115 * @pid: the pid in question.
4116 * @uattr: structure containing the extended parameters.
4117 * @flags: for future extension.
4119 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4120 unsigned int, flags
)
4122 struct sched_attr attr
;
4123 struct task_struct
*p
;
4126 if (!uattr
|| pid
< 0 || flags
)
4129 retval
= sched_copy_attr(uattr
, &attr
);
4133 if ((int)attr
.sched_policy
< 0)
4138 p
= find_process_by_pid(pid
);
4140 retval
= sched_setattr(p
, &attr
);
4147 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4148 * @pid: the pid in question.
4150 * Return: On success, the policy of the thread. Otherwise, a negative error
4153 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4155 struct task_struct
*p
;
4163 p
= find_process_by_pid(pid
);
4165 retval
= security_task_getscheduler(p
);
4168 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4175 * sys_sched_getparam - get the RT priority of a thread
4176 * @pid: the pid in question.
4177 * @param: structure containing the RT priority.
4179 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4182 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4184 struct sched_param lp
= { .sched_priority
= 0 };
4185 struct task_struct
*p
;
4188 if (!param
|| pid
< 0)
4192 p
= find_process_by_pid(pid
);
4197 retval
= security_task_getscheduler(p
);
4201 if (task_has_rt_policy(p
))
4202 lp
.sched_priority
= p
->rt_priority
;
4206 * This one might sleep, we cannot do it with a spinlock held ...
4208 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4217 static int sched_read_attr(struct sched_attr __user
*uattr
,
4218 struct sched_attr
*attr
,
4223 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4227 * If we're handed a smaller struct than we know of,
4228 * ensure all the unknown bits are 0 - i.e. old
4229 * user-space does not get uncomplete information.
4231 if (usize
< sizeof(*attr
)) {
4232 unsigned char *addr
;
4235 addr
= (void *)attr
+ usize
;
4236 end
= (void *)attr
+ sizeof(*attr
);
4238 for (; addr
< end
; addr
++) {
4246 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4254 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4255 * @pid: the pid in question.
4256 * @uattr: structure containing the extended parameters.
4257 * @size: sizeof(attr) for fwd/bwd comp.
4258 * @flags: for future extension.
4260 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4261 unsigned int, size
, unsigned int, flags
)
4263 struct sched_attr attr
= {
4264 .size
= sizeof(struct sched_attr
),
4266 struct task_struct
*p
;
4269 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4270 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4274 p
= find_process_by_pid(pid
);
4279 retval
= security_task_getscheduler(p
);
4283 attr
.sched_policy
= p
->policy
;
4284 if (p
->sched_reset_on_fork
)
4285 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4286 if (task_has_dl_policy(p
))
4287 __getparam_dl(p
, &attr
);
4288 else if (task_has_rt_policy(p
))
4289 attr
.sched_priority
= p
->rt_priority
;
4291 attr
.sched_nice
= task_nice(p
);
4295 retval
= sched_read_attr(uattr
, &attr
, size
);
4303 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4305 cpumask_var_t cpus_allowed
, new_mask
;
4306 struct task_struct
*p
;
4311 p
= find_process_by_pid(pid
);
4317 /* Prevent p going away */
4321 if (p
->flags
& PF_NO_SETAFFINITY
) {
4325 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4329 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4331 goto out_free_cpus_allowed
;
4334 if (!check_same_owner(p
)) {
4336 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4338 goto out_free_new_mask
;
4343 retval
= security_task_setscheduler(p
);
4345 goto out_free_new_mask
;
4348 cpuset_cpus_allowed(p
, cpus_allowed
);
4349 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4352 * Since bandwidth control happens on root_domain basis,
4353 * if admission test is enabled, we only admit -deadline
4354 * tasks allowed to run on all the CPUs in the task's
4358 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4360 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4363 goto out_free_new_mask
;
4369 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4372 cpuset_cpus_allowed(p
, cpus_allowed
);
4373 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4375 * We must have raced with a concurrent cpuset
4376 * update. Just reset the cpus_allowed to the
4377 * cpuset's cpus_allowed
4379 cpumask_copy(new_mask
, cpus_allowed
);
4384 free_cpumask_var(new_mask
);
4385 out_free_cpus_allowed
:
4386 free_cpumask_var(cpus_allowed
);
4392 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4393 struct cpumask
*new_mask
)
4395 if (len
< cpumask_size())
4396 cpumask_clear(new_mask
);
4397 else if (len
> cpumask_size())
4398 len
= cpumask_size();
4400 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4404 * sys_sched_setaffinity - set the cpu affinity of a process
4405 * @pid: pid of the process
4406 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4407 * @user_mask_ptr: user-space pointer to the new cpu mask
4409 * Return: 0 on success. An error code otherwise.
4411 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4412 unsigned long __user
*, user_mask_ptr
)
4414 cpumask_var_t new_mask
;
4417 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4420 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4422 retval
= sched_setaffinity(pid
, new_mask
);
4423 free_cpumask_var(new_mask
);
4427 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4429 struct task_struct
*p
;
4430 unsigned long flags
;
4436 p
= find_process_by_pid(pid
);
4440 retval
= security_task_getscheduler(p
);
4444 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4445 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4446 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4455 * sys_sched_getaffinity - get the cpu affinity of a process
4456 * @pid: pid of the process
4457 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4458 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4460 * Return: 0 on success. An error code otherwise.
4462 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4463 unsigned long __user
*, user_mask_ptr
)
4468 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4470 if (len
& (sizeof(unsigned long)-1))
4473 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4476 ret
= sched_getaffinity(pid
, mask
);
4478 size_t retlen
= min_t(size_t, len
, cpumask_size());
4480 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4485 free_cpumask_var(mask
);
4491 * sys_sched_yield - yield the current processor to other threads.
4493 * This function yields the current CPU to other tasks. If there are no
4494 * other threads running on this CPU then this function will return.
4498 SYSCALL_DEFINE0(sched_yield
)
4500 struct rq
*rq
= this_rq_lock();
4502 schedstat_inc(rq
, yld_count
);
4503 current
->sched_class
->yield_task(rq
);
4506 * Since we are going to call schedule() anyway, there's
4507 * no need to preempt or enable interrupts:
4509 __release(rq
->lock
);
4510 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4511 do_raw_spin_unlock(&rq
->lock
);
4512 sched_preempt_enable_no_resched();
4519 int __sched
_cond_resched(void)
4521 if (should_resched(0)) {
4522 preempt_schedule_common();
4527 EXPORT_SYMBOL(_cond_resched
);
4530 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4531 * call schedule, and on return reacquire the lock.
4533 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4534 * operations here to prevent schedule() from being called twice (once via
4535 * spin_unlock(), once by hand).
4537 int __cond_resched_lock(spinlock_t
*lock
)
4539 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4542 lockdep_assert_held(lock
);
4544 if (spin_needbreak(lock
) || resched
) {
4547 preempt_schedule_common();
4555 EXPORT_SYMBOL(__cond_resched_lock
);
4557 int __sched
__cond_resched_softirq(void)
4559 BUG_ON(!in_softirq());
4561 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4563 preempt_schedule_common();
4569 EXPORT_SYMBOL(__cond_resched_softirq
);
4572 * yield - yield the current processor to other threads.
4574 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4576 * The scheduler is at all times free to pick the calling task as the most
4577 * eligible task to run, if removing the yield() call from your code breaks
4578 * it, its already broken.
4580 * Typical broken usage is:
4585 * where one assumes that yield() will let 'the other' process run that will
4586 * make event true. If the current task is a SCHED_FIFO task that will never
4587 * happen. Never use yield() as a progress guarantee!!
4589 * If you want to use yield() to wait for something, use wait_event().
4590 * If you want to use yield() to be 'nice' for others, use cond_resched().
4591 * If you still want to use yield(), do not!
4593 void __sched
yield(void)
4595 set_current_state(TASK_RUNNING
);
4598 EXPORT_SYMBOL(yield
);
4601 * yield_to - yield the current processor to another thread in
4602 * your thread group, or accelerate that thread toward the
4603 * processor it's on.
4605 * @preempt: whether task preemption is allowed or not
4607 * It's the caller's job to ensure that the target task struct
4608 * can't go away on us before we can do any checks.
4611 * true (>0) if we indeed boosted the target task.
4612 * false (0) if we failed to boost the target.
4613 * -ESRCH if there's no task to yield to.
4615 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4617 struct task_struct
*curr
= current
;
4618 struct rq
*rq
, *p_rq
;
4619 unsigned long flags
;
4622 local_irq_save(flags
);
4628 * If we're the only runnable task on the rq and target rq also
4629 * has only one task, there's absolutely no point in yielding.
4631 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4636 double_rq_lock(rq
, p_rq
);
4637 if (task_rq(p
) != p_rq
) {
4638 double_rq_unlock(rq
, p_rq
);
4642 if (!curr
->sched_class
->yield_to_task
)
4645 if (curr
->sched_class
!= p
->sched_class
)
4648 if (task_running(p_rq
, p
) || p
->state
)
4651 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4653 schedstat_inc(rq
, yld_count
);
4655 * Make p's CPU reschedule; pick_next_entity takes care of
4658 if (preempt
&& rq
!= p_rq
)
4663 double_rq_unlock(rq
, p_rq
);
4665 local_irq_restore(flags
);
4672 EXPORT_SYMBOL_GPL(yield_to
);
4675 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4676 * that process accounting knows that this is a task in IO wait state.
4678 long __sched
io_schedule_timeout(long timeout
)
4680 int old_iowait
= current
->in_iowait
;
4684 current
->in_iowait
= 1;
4685 blk_schedule_flush_plug(current
);
4687 delayacct_blkio_start();
4689 atomic_inc(&rq
->nr_iowait
);
4690 ret
= schedule_timeout(timeout
);
4691 current
->in_iowait
= old_iowait
;
4692 atomic_dec(&rq
->nr_iowait
);
4693 delayacct_blkio_end();
4697 EXPORT_SYMBOL(io_schedule_timeout
);
4700 * sys_sched_get_priority_max - return maximum RT priority.
4701 * @policy: scheduling class.
4703 * Return: On success, this syscall returns the maximum
4704 * rt_priority that can be used by a given scheduling class.
4705 * On failure, a negative error code is returned.
4707 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4714 ret
= MAX_USER_RT_PRIO
-1;
4716 case SCHED_DEADLINE
:
4727 * sys_sched_get_priority_min - return minimum RT priority.
4728 * @policy: scheduling class.
4730 * Return: On success, this syscall returns the minimum
4731 * rt_priority that can be used by a given scheduling class.
4732 * On failure, a negative error code is returned.
4734 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4743 case SCHED_DEADLINE
:
4753 * sys_sched_rr_get_interval - return the default timeslice of a process.
4754 * @pid: pid of the process.
4755 * @interval: userspace pointer to the timeslice value.
4757 * this syscall writes the default timeslice value of a given process
4758 * into the user-space timespec buffer. A value of '0' means infinity.
4760 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4763 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4764 struct timespec __user
*, interval
)
4766 struct task_struct
*p
;
4767 unsigned int time_slice
;
4768 unsigned long flags
;
4778 p
= find_process_by_pid(pid
);
4782 retval
= security_task_getscheduler(p
);
4786 rq
= task_rq_lock(p
, &flags
);
4788 if (p
->sched_class
->get_rr_interval
)
4789 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4790 task_rq_unlock(rq
, p
, &flags
);
4793 jiffies_to_timespec(time_slice
, &t
);
4794 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4802 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4804 void sched_show_task(struct task_struct
*p
)
4806 unsigned long free
= 0;
4808 unsigned long state
= p
->state
;
4811 state
= __ffs(state
) + 1;
4812 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4813 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4814 #if BITS_PER_LONG == 32
4815 if (state
== TASK_RUNNING
)
4816 printk(KERN_CONT
" running ");
4818 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4820 if (state
== TASK_RUNNING
)
4821 printk(KERN_CONT
" running task ");
4823 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4825 #ifdef CONFIG_DEBUG_STACK_USAGE
4826 free
= stack_not_used(p
);
4831 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4833 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4834 task_pid_nr(p
), ppid
,
4835 (unsigned long)task_thread_info(p
)->flags
);
4837 print_worker_info(KERN_INFO
, p
);
4838 show_stack(p
, NULL
);
4841 void show_state_filter(unsigned long state_filter
)
4843 struct task_struct
*g
, *p
;
4845 #if BITS_PER_LONG == 32
4847 " task PC stack pid father\n");
4850 " task PC stack pid father\n");
4853 for_each_process_thread(g
, p
) {
4855 * reset the NMI-timeout, listing all files on a slow
4856 * console might take a lot of time:
4858 touch_nmi_watchdog();
4859 if (!state_filter
|| (p
->state
& state_filter
))
4863 touch_all_softlockup_watchdogs();
4865 #ifdef CONFIG_SCHED_DEBUG
4866 sysrq_sched_debug_show();
4870 * Only show locks if all tasks are dumped:
4873 debug_show_all_locks();
4876 void init_idle_bootup_task(struct task_struct
*idle
)
4878 idle
->sched_class
= &idle_sched_class
;
4882 * init_idle - set up an idle thread for a given CPU
4883 * @idle: task in question
4884 * @cpu: cpu the idle task belongs to
4886 * NOTE: this function does not set the idle thread's NEED_RESCHED
4887 * flag, to make booting more robust.
4889 void init_idle(struct task_struct
*idle
, int cpu
)
4891 struct rq
*rq
= cpu_rq(cpu
);
4892 unsigned long flags
;
4894 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
4895 raw_spin_lock(&rq
->lock
);
4897 __sched_fork(0, idle
);
4898 idle
->state
= TASK_RUNNING
;
4899 idle
->se
.exec_start
= sched_clock();
4901 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4903 * We're having a chicken and egg problem, even though we are
4904 * holding rq->lock, the cpu isn't yet set to this cpu so the
4905 * lockdep check in task_group() will fail.
4907 * Similar case to sched_fork(). / Alternatively we could
4908 * use task_rq_lock() here and obtain the other rq->lock.
4913 __set_task_cpu(idle
, cpu
);
4916 rq
->curr
= rq
->idle
= idle
;
4917 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4918 #if defined(CONFIG_SMP)
4921 raw_spin_unlock(&rq
->lock
);
4922 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
4924 /* Set the preempt count _outside_ the spinlocks! */
4925 init_idle_preempt_count(idle
, cpu
);
4928 * The idle tasks have their own, simple scheduling class:
4930 idle
->sched_class
= &idle_sched_class
;
4931 ftrace_graph_init_idle_task(idle
, cpu
);
4932 vtime_init_idle(idle
, cpu
);
4933 #if defined(CONFIG_SMP)
4934 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4938 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4939 const struct cpumask
*trial
)
4941 int ret
= 1, trial_cpus
;
4942 struct dl_bw
*cur_dl_b
;
4943 unsigned long flags
;
4945 if (!cpumask_weight(cur
))
4948 rcu_read_lock_sched();
4949 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4950 trial_cpus
= cpumask_weight(trial
);
4952 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4953 if (cur_dl_b
->bw
!= -1 &&
4954 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
4956 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
4957 rcu_read_unlock_sched();
4962 int task_can_attach(struct task_struct
*p
,
4963 const struct cpumask
*cs_cpus_allowed
)
4968 * Kthreads which disallow setaffinity shouldn't be moved
4969 * to a new cpuset; we don't want to change their cpu
4970 * affinity and isolating such threads by their set of
4971 * allowed nodes is unnecessary. Thus, cpusets are not
4972 * applicable for such threads. This prevents checking for
4973 * success of set_cpus_allowed_ptr() on all attached tasks
4974 * before cpus_allowed may be changed.
4976 if (p
->flags
& PF_NO_SETAFFINITY
) {
4982 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
4984 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
4989 unsigned long flags
;
4991 rcu_read_lock_sched();
4992 dl_b
= dl_bw_of(dest_cpu
);
4993 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
4994 cpus
= dl_bw_cpus(dest_cpu
);
4995 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5000 * We reserve space for this task in the destination
5001 * root_domain, as we can't fail after this point.
5002 * We will free resources in the source root_domain
5003 * later on (see set_cpus_allowed_dl()).
5005 __dl_add(dl_b
, p
->dl
.dl_bw
);
5007 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5008 rcu_read_unlock_sched();
5018 #ifdef CONFIG_NUMA_BALANCING
5019 /* Migrate current task p to target_cpu */
5020 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5022 struct migration_arg arg
= { p
, target_cpu
};
5023 int curr_cpu
= task_cpu(p
);
5025 if (curr_cpu
== target_cpu
)
5028 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5031 /* TODO: This is not properly updating schedstats */
5033 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5034 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5038 * Requeue a task on a given node and accurately track the number of NUMA
5039 * tasks on the runqueues
5041 void sched_setnuma(struct task_struct
*p
, int nid
)
5044 unsigned long flags
;
5045 bool queued
, running
;
5047 rq
= task_rq_lock(p
, &flags
);
5048 queued
= task_on_rq_queued(p
);
5049 running
= task_current(rq
, p
);
5052 dequeue_task(rq
, p
, 0);
5054 put_prev_task(rq
, p
);
5056 p
->numa_preferred_nid
= nid
;
5059 p
->sched_class
->set_curr_task(rq
);
5061 enqueue_task(rq
, p
, 0);
5062 task_rq_unlock(rq
, p
, &flags
);
5064 #endif /* CONFIG_NUMA_BALANCING */
5066 #ifdef CONFIG_HOTPLUG_CPU
5068 * Ensures that the idle task is using init_mm right before its cpu goes
5071 void idle_task_exit(void)
5073 struct mm_struct
*mm
= current
->active_mm
;
5075 BUG_ON(cpu_online(smp_processor_id()));
5077 if (mm
!= &init_mm
) {
5078 switch_mm(mm
, &init_mm
, current
);
5079 finish_arch_post_lock_switch();
5085 * Since this CPU is going 'away' for a while, fold any nr_active delta
5086 * we might have. Assumes we're called after migrate_tasks() so that the
5087 * nr_active count is stable.
5089 * Also see the comment "Global load-average calculations".
5091 static void calc_load_migrate(struct rq
*rq
)
5093 long delta
= calc_load_fold_active(rq
);
5095 atomic_long_add(delta
, &calc_load_tasks
);
5098 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5102 static const struct sched_class fake_sched_class
= {
5103 .put_prev_task
= put_prev_task_fake
,
5106 static struct task_struct fake_task
= {
5108 * Avoid pull_{rt,dl}_task()
5110 .prio
= MAX_PRIO
+ 1,
5111 .sched_class
= &fake_sched_class
,
5115 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5116 * try_to_wake_up()->select_task_rq().
5118 * Called with rq->lock held even though we'er in stop_machine() and
5119 * there's no concurrency possible, we hold the required locks anyway
5120 * because of lock validation efforts.
5122 static void migrate_tasks(struct rq
*dead_rq
)
5124 struct rq
*rq
= dead_rq
;
5125 struct task_struct
*next
, *stop
= rq
->stop
;
5129 * Fudge the rq selection such that the below task selection loop
5130 * doesn't get stuck on the currently eligible stop task.
5132 * We're currently inside stop_machine() and the rq is either stuck
5133 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5134 * either way we should never end up calling schedule() until we're
5140 * put_prev_task() and pick_next_task() sched
5141 * class method both need to have an up-to-date
5142 * value of rq->clock[_task]
5144 update_rq_clock(rq
);
5148 * There's this thread running, bail when that's the only
5151 if (rq
->nr_running
== 1)
5155 * Ensure rq->lock covers the entire task selection
5156 * until the migration.
5158 lockdep_pin_lock(&rq
->lock
);
5159 next
= pick_next_task(rq
, &fake_task
);
5161 next
->sched_class
->put_prev_task(rq
, next
);
5163 /* Find suitable destination for @next, with force if needed. */
5164 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5166 lockdep_unpin_lock(&rq
->lock
);
5167 rq
= __migrate_task(rq
, next
, dest_cpu
);
5168 if (rq
!= dead_rq
) {
5169 raw_spin_unlock(&rq
->lock
);
5171 raw_spin_lock(&rq
->lock
);
5177 #endif /* CONFIG_HOTPLUG_CPU */
5179 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5181 static struct ctl_table sd_ctl_dir
[] = {
5183 .procname
= "sched_domain",
5189 static struct ctl_table sd_ctl_root
[] = {
5191 .procname
= "kernel",
5193 .child
= sd_ctl_dir
,
5198 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5200 struct ctl_table
*entry
=
5201 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5206 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5208 struct ctl_table
*entry
;
5211 * In the intermediate directories, both the child directory and
5212 * procname are dynamically allocated and could fail but the mode
5213 * will always be set. In the lowest directory the names are
5214 * static strings and all have proc handlers.
5216 for (entry
= *tablep
; entry
->mode
; entry
++) {
5218 sd_free_ctl_entry(&entry
->child
);
5219 if (entry
->proc_handler
== NULL
)
5220 kfree(entry
->procname
);
5227 static int min_load_idx
= 0;
5228 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5231 set_table_entry(struct ctl_table
*entry
,
5232 const char *procname
, void *data
, int maxlen
,
5233 umode_t mode
, proc_handler
*proc_handler
,
5236 entry
->procname
= procname
;
5238 entry
->maxlen
= maxlen
;
5240 entry
->proc_handler
= proc_handler
;
5243 entry
->extra1
= &min_load_idx
;
5244 entry
->extra2
= &max_load_idx
;
5248 static struct ctl_table
*
5249 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5251 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5256 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5257 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5258 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5259 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5260 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5261 sizeof(int), 0644, proc_dointvec_minmax
, true);
5262 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5263 sizeof(int), 0644, proc_dointvec_minmax
, true);
5264 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5265 sizeof(int), 0644, proc_dointvec_minmax
, true);
5266 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5267 sizeof(int), 0644, proc_dointvec_minmax
, true);
5268 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5269 sizeof(int), 0644, proc_dointvec_minmax
, true);
5270 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5271 sizeof(int), 0644, proc_dointvec_minmax
, false);
5272 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5273 sizeof(int), 0644, proc_dointvec_minmax
, false);
5274 set_table_entry(&table
[9], "cache_nice_tries",
5275 &sd
->cache_nice_tries
,
5276 sizeof(int), 0644, proc_dointvec_minmax
, false);
5277 set_table_entry(&table
[10], "flags", &sd
->flags
,
5278 sizeof(int), 0644, proc_dointvec_minmax
, false);
5279 set_table_entry(&table
[11], "max_newidle_lb_cost",
5280 &sd
->max_newidle_lb_cost
,
5281 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5282 set_table_entry(&table
[12], "name", sd
->name
,
5283 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5284 /* &table[13] is terminator */
5289 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5291 struct ctl_table
*entry
, *table
;
5292 struct sched_domain
*sd
;
5293 int domain_num
= 0, i
;
5296 for_each_domain(cpu
, sd
)
5298 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5303 for_each_domain(cpu
, sd
) {
5304 snprintf(buf
, 32, "domain%d", i
);
5305 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5307 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5314 static struct ctl_table_header
*sd_sysctl_header
;
5315 static void register_sched_domain_sysctl(void)
5317 int i
, cpu_num
= num_possible_cpus();
5318 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5321 WARN_ON(sd_ctl_dir
[0].child
);
5322 sd_ctl_dir
[0].child
= entry
;
5327 for_each_possible_cpu(i
) {
5328 snprintf(buf
, 32, "cpu%d", i
);
5329 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5331 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5335 WARN_ON(sd_sysctl_header
);
5336 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5339 /* may be called multiple times per register */
5340 static void unregister_sched_domain_sysctl(void)
5342 unregister_sysctl_table(sd_sysctl_header
);
5343 sd_sysctl_header
= NULL
;
5344 if (sd_ctl_dir
[0].child
)
5345 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5348 static void register_sched_domain_sysctl(void)
5351 static void unregister_sched_domain_sysctl(void)
5354 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5356 static void set_rq_online(struct rq
*rq
)
5359 const struct sched_class
*class;
5361 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5364 for_each_class(class) {
5365 if (class->rq_online
)
5366 class->rq_online(rq
);
5371 static void set_rq_offline(struct rq
*rq
)
5374 const struct sched_class
*class;
5376 for_each_class(class) {
5377 if (class->rq_offline
)
5378 class->rq_offline(rq
);
5381 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5387 * migration_call - callback that gets triggered when a CPU is added.
5388 * Here we can start up the necessary migration thread for the new CPU.
5391 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5393 int cpu
= (long)hcpu
;
5394 unsigned long flags
;
5395 struct rq
*rq
= cpu_rq(cpu
);
5397 switch (action
& ~CPU_TASKS_FROZEN
) {
5399 case CPU_UP_PREPARE
:
5400 rq
->calc_load_update
= calc_load_update
;
5404 /* Update our root-domain */
5405 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5407 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5411 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5414 #ifdef CONFIG_HOTPLUG_CPU
5416 sched_ttwu_pending();
5417 /* Update our root-domain */
5418 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5420 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5424 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5425 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5429 calc_load_migrate(rq
);
5434 update_max_interval();
5440 * Register at high priority so that task migration (migrate_all_tasks)
5441 * happens before everything else. This has to be lower priority than
5442 * the notifier in the perf_event subsystem, though.
5444 static struct notifier_block migration_notifier
= {
5445 .notifier_call
= migration_call
,
5446 .priority
= CPU_PRI_MIGRATION
,
5449 static void set_cpu_rq_start_time(void)
5451 int cpu
= smp_processor_id();
5452 struct rq
*rq
= cpu_rq(cpu
);
5453 rq
->age_stamp
= sched_clock_cpu(cpu
);
5456 static int sched_cpu_active(struct notifier_block
*nfb
,
5457 unsigned long action
, void *hcpu
)
5459 switch (action
& ~CPU_TASKS_FROZEN
) {
5461 set_cpu_rq_start_time();
5463 case CPU_DOWN_FAILED
:
5464 set_cpu_active((long)hcpu
, true);
5471 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5472 unsigned long action
, void *hcpu
)
5474 switch (action
& ~CPU_TASKS_FROZEN
) {
5475 case CPU_DOWN_PREPARE
:
5476 set_cpu_active((long)hcpu
, false);
5483 static int __init
migration_init(void)
5485 void *cpu
= (void *)(long)smp_processor_id();
5488 /* Initialize migration for the boot CPU */
5489 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5490 BUG_ON(err
== NOTIFY_BAD
);
5491 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5492 register_cpu_notifier(&migration_notifier
);
5494 /* Register cpu active notifiers */
5495 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5496 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5500 early_initcall(migration_init
);
5502 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5504 #ifdef CONFIG_SCHED_DEBUG
5506 static __read_mostly
int sched_debug_enabled
;
5508 static int __init
sched_debug_setup(char *str
)
5510 sched_debug_enabled
= 1;
5514 early_param("sched_debug", sched_debug_setup
);
5516 static inline bool sched_debug(void)
5518 return sched_debug_enabled
;
5521 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5522 struct cpumask
*groupmask
)
5524 struct sched_group
*group
= sd
->groups
;
5526 cpumask_clear(groupmask
);
5528 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5530 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5531 printk("does not load-balance\n");
5533 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5538 printk(KERN_CONT
"span %*pbl level %s\n",
5539 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5541 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5542 printk(KERN_ERR
"ERROR: domain->span does not contain "
5545 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5546 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5550 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5554 printk(KERN_ERR
"ERROR: group is NULL\n");
5558 if (!cpumask_weight(sched_group_cpus(group
))) {
5559 printk(KERN_CONT
"\n");
5560 printk(KERN_ERR
"ERROR: empty group\n");
5564 if (!(sd
->flags
& SD_OVERLAP
) &&
5565 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5566 printk(KERN_CONT
"\n");
5567 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5571 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5573 printk(KERN_CONT
" %*pbl",
5574 cpumask_pr_args(sched_group_cpus(group
)));
5575 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5576 printk(KERN_CONT
" (cpu_capacity = %d)",
5577 group
->sgc
->capacity
);
5580 group
= group
->next
;
5581 } while (group
!= sd
->groups
);
5582 printk(KERN_CONT
"\n");
5584 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5585 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5588 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5589 printk(KERN_ERR
"ERROR: parent span is not a superset "
5590 "of domain->span\n");
5594 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5598 if (!sched_debug_enabled
)
5602 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5606 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5609 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5617 #else /* !CONFIG_SCHED_DEBUG */
5618 # define sched_domain_debug(sd, cpu) do { } while (0)
5619 static inline bool sched_debug(void)
5623 #endif /* CONFIG_SCHED_DEBUG */
5625 static int sd_degenerate(struct sched_domain
*sd
)
5627 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5630 /* Following flags need at least 2 groups */
5631 if (sd
->flags
& (SD_LOAD_BALANCE
|
5632 SD_BALANCE_NEWIDLE
|
5635 SD_SHARE_CPUCAPACITY
|
5636 SD_SHARE_PKG_RESOURCES
|
5637 SD_SHARE_POWERDOMAIN
)) {
5638 if (sd
->groups
!= sd
->groups
->next
)
5642 /* Following flags don't use groups */
5643 if (sd
->flags
& (SD_WAKE_AFFINE
))
5650 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5652 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5654 if (sd_degenerate(parent
))
5657 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5660 /* Flags needing groups don't count if only 1 group in parent */
5661 if (parent
->groups
== parent
->groups
->next
) {
5662 pflags
&= ~(SD_LOAD_BALANCE
|
5663 SD_BALANCE_NEWIDLE
|
5666 SD_SHARE_CPUCAPACITY
|
5667 SD_SHARE_PKG_RESOURCES
|
5669 SD_SHARE_POWERDOMAIN
);
5670 if (nr_node_ids
== 1)
5671 pflags
&= ~SD_SERIALIZE
;
5673 if (~cflags
& pflags
)
5679 static void free_rootdomain(struct rcu_head
*rcu
)
5681 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5683 cpupri_cleanup(&rd
->cpupri
);
5684 cpudl_cleanup(&rd
->cpudl
);
5685 free_cpumask_var(rd
->dlo_mask
);
5686 free_cpumask_var(rd
->rto_mask
);
5687 free_cpumask_var(rd
->online
);
5688 free_cpumask_var(rd
->span
);
5692 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5694 struct root_domain
*old_rd
= NULL
;
5695 unsigned long flags
;
5697 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5702 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5705 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5708 * If we dont want to free the old_rd yet then
5709 * set old_rd to NULL to skip the freeing later
5712 if (!atomic_dec_and_test(&old_rd
->refcount
))
5716 atomic_inc(&rd
->refcount
);
5719 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5720 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5723 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5726 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5729 static int init_rootdomain(struct root_domain
*rd
)
5731 memset(rd
, 0, sizeof(*rd
));
5733 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5735 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5737 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5739 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5742 init_dl_bw(&rd
->dl_bw
);
5743 if (cpudl_init(&rd
->cpudl
) != 0)
5746 if (cpupri_init(&rd
->cpupri
) != 0)
5751 free_cpumask_var(rd
->rto_mask
);
5753 free_cpumask_var(rd
->dlo_mask
);
5755 free_cpumask_var(rd
->online
);
5757 free_cpumask_var(rd
->span
);
5763 * By default the system creates a single root-domain with all cpus as
5764 * members (mimicking the global state we have today).
5766 struct root_domain def_root_domain
;
5768 static void init_defrootdomain(void)
5770 init_rootdomain(&def_root_domain
);
5772 atomic_set(&def_root_domain
.refcount
, 1);
5775 static struct root_domain
*alloc_rootdomain(void)
5777 struct root_domain
*rd
;
5779 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5783 if (init_rootdomain(rd
) != 0) {
5791 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5793 struct sched_group
*tmp
, *first
;
5802 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5807 } while (sg
!= first
);
5810 static void free_sched_domain(struct rcu_head
*rcu
)
5812 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5815 * If its an overlapping domain it has private groups, iterate and
5818 if (sd
->flags
& SD_OVERLAP
) {
5819 free_sched_groups(sd
->groups
, 1);
5820 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5821 kfree(sd
->groups
->sgc
);
5827 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5829 call_rcu(&sd
->rcu
, free_sched_domain
);
5832 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5834 for (; sd
; sd
= sd
->parent
)
5835 destroy_sched_domain(sd
, cpu
);
5839 * Keep a special pointer to the highest sched_domain that has
5840 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5841 * allows us to avoid some pointer chasing select_idle_sibling().
5843 * Also keep a unique ID per domain (we use the first cpu number in
5844 * the cpumask of the domain), this allows us to quickly tell if
5845 * two cpus are in the same cache domain, see cpus_share_cache().
5847 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5848 DEFINE_PER_CPU(int, sd_llc_size
);
5849 DEFINE_PER_CPU(int, sd_llc_id
);
5850 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5851 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5852 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5854 static void update_top_cache_domain(int cpu
)
5856 struct sched_domain
*sd
;
5857 struct sched_domain
*busy_sd
= NULL
;
5861 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5863 id
= cpumask_first(sched_domain_span(sd
));
5864 size
= cpumask_weight(sched_domain_span(sd
));
5865 busy_sd
= sd
->parent
; /* sd_busy */
5867 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5869 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5870 per_cpu(sd_llc_size
, cpu
) = size
;
5871 per_cpu(sd_llc_id
, cpu
) = id
;
5873 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5874 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5876 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5877 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5881 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5882 * hold the hotplug lock.
5885 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5887 struct rq
*rq
= cpu_rq(cpu
);
5888 struct sched_domain
*tmp
;
5890 /* Remove the sched domains which do not contribute to scheduling. */
5891 for (tmp
= sd
; tmp
; ) {
5892 struct sched_domain
*parent
= tmp
->parent
;
5896 if (sd_parent_degenerate(tmp
, parent
)) {
5897 tmp
->parent
= parent
->parent
;
5899 parent
->parent
->child
= tmp
;
5901 * Transfer SD_PREFER_SIBLING down in case of a
5902 * degenerate parent; the spans match for this
5903 * so the property transfers.
5905 if (parent
->flags
& SD_PREFER_SIBLING
)
5906 tmp
->flags
|= SD_PREFER_SIBLING
;
5907 destroy_sched_domain(parent
, cpu
);
5912 if (sd
&& sd_degenerate(sd
)) {
5915 destroy_sched_domain(tmp
, cpu
);
5920 sched_domain_debug(sd
, cpu
);
5922 rq_attach_root(rq
, rd
);
5924 rcu_assign_pointer(rq
->sd
, sd
);
5925 destroy_sched_domains(tmp
, cpu
);
5927 update_top_cache_domain(cpu
);
5930 /* Setup the mask of cpus configured for isolated domains */
5931 static int __init
isolated_cpu_setup(char *str
)
5933 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5934 cpulist_parse(str
, cpu_isolated_map
);
5938 __setup("isolcpus=", isolated_cpu_setup
);
5941 struct sched_domain
** __percpu sd
;
5942 struct root_domain
*rd
;
5953 * Build an iteration mask that can exclude certain CPUs from the upwards
5956 * Asymmetric node setups can result in situations where the domain tree is of
5957 * unequal depth, make sure to skip domains that already cover the entire
5960 * In that case build_sched_domains() will have terminated the iteration early
5961 * and our sibling sd spans will be empty. Domains should always include the
5962 * cpu they're built on, so check that.
5965 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5967 const struct cpumask
*span
= sched_domain_span(sd
);
5968 struct sd_data
*sdd
= sd
->private;
5969 struct sched_domain
*sibling
;
5972 for_each_cpu(i
, span
) {
5973 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5974 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5977 cpumask_set_cpu(i
, sched_group_mask(sg
));
5982 * Return the canonical balance cpu for this group, this is the first cpu
5983 * of this group that's also in the iteration mask.
5985 int group_balance_cpu(struct sched_group
*sg
)
5987 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5991 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5993 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5994 const struct cpumask
*span
= sched_domain_span(sd
);
5995 struct cpumask
*covered
= sched_domains_tmpmask
;
5996 struct sd_data
*sdd
= sd
->private;
5997 struct sched_domain
*sibling
;
6000 cpumask_clear(covered
);
6002 for_each_cpu(i
, span
) {
6003 struct cpumask
*sg_span
;
6005 if (cpumask_test_cpu(i
, covered
))
6008 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6010 /* See the comment near build_group_mask(). */
6011 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6014 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6015 GFP_KERNEL
, cpu_to_node(cpu
));
6020 sg_span
= sched_group_cpus(sg
);
6022 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6024 cpumask_set_cpu(i
, sg_span
);
6026 cpumask_or(covered
, covered
, sg_span
);
6028 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6029 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6030 build_group_mask(sd
, sg
);
6033 * Initialize sgc->capacity such that even if we mess up the
6034 * domains and no possible iteration will get us here, we won't
6037 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6040 * Make sure the first group of this domain contains the
6041 * canonical balance cpu. Otherwise the sched_domain iteration
6042 * breaks. See update_sg_lb_stats().
6044 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6045 group_balance_cpu(sg
) == cpu
)
6055 sd
->groups
= groups
;
6060 free_sched_groups(first
, 0);
6065 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6067 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6068 struct sched_domain
*child
= sd
->child
;
6071 cpu
= cpumask_first(sched_domain_span(child
));
6074 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6075 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6076 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6083 * build_sched_groups will build a circular linked list of the groups
6084 * covered by the given span, and will set each group's ->cpumask correctly,
6085 * and ->cpu_capacity to 0.
6087 * Assumes the sched_domain tree is fully constructed
6090 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6092 struct sched_group
*first
= NULL
, *last
= NULL
;
6093 struct sd_data
*sdd
= sd
->private;
6094 const struct cpumask
*span
= sched_domain_span(sd
);
6095 struct cpumask
*covered
;
6098 get_group(cpu
, sdd
, &sd
->groups
);
6099 atomic_inc(&sd
->groups
->ref
);
6101 if (cpu
!= cpumask_first(span
))
6104 lockdep_assert_held(&sched_domains_mutex
);
6105 covered
= sched_domains_tmpmask
;
6107 cpumask_clear(covered
);
6109 for_each_cpu(i
, span
) {
6110 struct sched_group
*sg
;
6113 if (cpumask_test_cpu(i
, covered
))
6116 group
= get_group(i
, sdd
, &sg
);
6117 cpumask_setall(sched_group_mask(sg
));
6119 for_each_cpu(j
, span
) {
6120 if (get_group(j
, sdd
, NULL
) != group
)
6123 cpumask_set_cpu(j
, covered
);
6124 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6139 * Initialize sched groups cpu_capacity.
6141 * cpu_capacity indicates the capacity of sched group, which is used while
6142 * distributing the load between different sched groups in a sched domain.
6143 * Typically cpu_capacity for all the groups in a sched domain will be same
6144 * unless there are asymmetries in the topology. If there are asymmetries,
6145 * group having more cpu_capacity will pickup more load compared to the
6146 * group having less cpu_capacity.
6148 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6150 struct sched_group
*sg
= sd
->groups
;
6155 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6157 } while (sg
!= sd
->groups
);
6159 if (cpu
!= group_balance_cpu(sg
))
6162 update_group_capacity(sd
, cpu
);
6163 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6167 * Initializers for schedule domains
6168 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6171 static int default_relax_domain_level
= -1;
6172 int sched_domain_level_max
;
6174 static int __init
setup_relax_domain_level(char *str
)
6176 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6177 pr_warn("Unable to set relax_domain_level\n");
6181 __setup("relax_domain_level=", setup_relax_domain_level
);
6183 static void set_domain_attribute(struct sched_domain
*sd
,
6184 struct sched_domain_attr
*attr
)
6188 if (!attr
|| attr
->relax_domain_level
< 0) {
6189 if (default_relax_domain_level
< 0)
6192 request
= default_relax_domain_level
;
6194 request
= attr
->relax_domain_level
;
6195 if (request
< sd
->level
) {
6196 /* turn off idle balance on this domain */
6197 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6199 /* turn on idle balance on this domain */
6200 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6204 static void __sdt_free(const struct cpumask
*cpu_map
);
6205 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6207 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6208 const struct cpumask
*cpu_map
)
6212 if (!atomic_read(&d
->rd
->refcount
))
6213 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6215 free_percpu(d
->sd
); /* fall through */
6217 __sdt_free(cpu_map
); /* fall through */
6223 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6224 const struct cpumask
*cpu_map
)
6226 memset(d
, 0, sizeof(*d
));
6228 if (__sdt_alloc(cpu_map
))
6229 return sa_sd_storage
;
6230 d
->sd
= alloc_percpu(struct sched_domain
*);
6232 return sa_sd_storage
;
6233 d
->rd
= alloc_rootdomain();
6236 return sa_rootdomain
;
6240 * NULL the sd_data elements we've used to build the sched_domain and
6241 * sched_group structure so that the subsequent __free_domain_allocs()
6242 * will not free the data we're using.
6244 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6246 struct sd_data
*sdd
= sd
->private;
6248 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6249 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6251 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6252 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6254 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6255 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6259 static int sched_domains_numa_levels
;
6260 enum numa_topology_type sched_numa_topology_type
;
6261 static int *sched_domains_numa_distance
;
6262 int sched_max_numa_distance
;
6263 static struct cpumask
***sched_domains_numa_masks
;
6264 static int sched_domains_curr_level
;
6268 * SD_flags allowed in topology descriptions.
6270 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6271 * SD_SHARE_PKG_RESOURCES - describes shared caches
6272 * SD_NUMA - describes NUMA topologies
6273 * SD_SHARE_POWERDOMAIN - describes shared power domain
6276 * SD_ASYM_PACKING - describes SMT quirks
6278 #define TOPOLOGY_SD_FLAGS \
6279 (SD_SHARE_CPUCAPACITY | \
6280 SD_SHARE_PKG_RESOURCES | \
6283 SD_SHARE_POWERDOMAIN)
6285 static struct sched_domain
*
6286 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6288 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6289 int sd_weight
, sd_flags
= 0;
6293 * Ugly hack to pass state to sd_numa_mask()...
6295 sched_domains_curr_level
= tl
->numa_level
;
6298 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6301 sd_flags
= (*tl
->sd_flags
)();
6302 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6303 "wrong sd_flags in topology description\n"))
6304 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6306 *sd
= (struct sched_domain
){
6307 .min_interval
= sd_weight
,
6308 .max_interval
= 2*sd_weight
,
6310 .imbalance_pct
= 125,
6312 .cache_nice_tries
= 0,
6319 .flags
= 1*SD_LOAD_BALANCE
6320 | 1*SD_BALANCE_NEWIDLE
6325 | 0*SD_SHARE_CPUCAPACITY
6326 | 0*SD_SHARE_PKG_RESOURCES
6328 | 0*SD_PREFER_SIBLING
6333 .last_balance
= jiffies
,
6334 .balance_interval
= sd_weight
,
6336 .max_newidle_lb_cost
= 0,
6337 .next_decay_max_lb_cost
= jiffies
,
6338 #ifdef CONFIG_SCHED_DEBUG
6344 * Convert topological properties into behaviour.
6347 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6348 sd
->flags
|= SD_PREFER_SIBLING
;
6349 sd
->imbalance_pct
= 110;
6350 sd
->smt_gain
= 1178; /* ~15% */
6352 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6353 sd
->imbalance_pct
= 117;
6354 sd
->cache_nice_tries
= 1;
6358 } else if (sd
->flags
& SD_NUMA
) {
6359 sd
->cache_nice_tries
= 2;
6363 sd
->flags
|= SD_SERIALIZE
;
6364 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6365 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6372 sd
->flags
|= SD_PREFER_SIBLING
;
6373 sd
->cache_nice_tries
= 1;
6378 sd
->private = &tl
->data
;
6384 * Topology list, bottom-up.
6386 static struct sched_domain_topology_level default_topology
[] = {
6387 #ifdef CONFIG_SCHED_SMT
6388 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6390 #ifdef CONFIG_SCHED_MC
6391 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6393 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6397 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6399 #define for_each_sd_topology(tl) \
6400 for (tl = sched_domain_topology; tl->mask; tl++)
6402 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6404 sched_domain_topology
= tl
;
6409 static const struct cpumask
*sd_numa_mask(int cpu
)
6411 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6414 static void sched_numa_warn(const char *str
)
6416 static int done
= false;
6424 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6426 for (i
= 0; i
< nr_node_ids
; i
++) {
6427 printk(KERN_WARNING
" ");
6428 for (j
= 0; j
< nr_node_ids
; j
++)
6429 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6430 printk(KERN_CONT
"\n");
6432 printk(KERN_WARNING
"\n");
6435 bool find_numa_distance(int distance
)
6439 if (distance
== node_distance(0, 0))
6442 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6443 if (sched_domains_numa_distance
[i
] == distance
)
6451 * A system can have three types of NUMA topology:
6452 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6453 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6454 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6456 * The difference between a glueless mesh topology and a backplane
6457 * topology lies in whether communication between not directly
6458 * connected nodes goes through intermediary nodes (where programs
6459 * could run), or through backplane controllers. This affects
6460 * placement of programs.
6462 * The type of topology can be discerned with the following tests:
6463 * - If the maximum distance between any nodes is 1 hop, the system
6464 * is directly connected.
6465 * - If for two nodes A and B, located N > 1 hops away from each other,
6466 * there is an intermediary node C, which is < N hops away from both
6467 * nodes A and B, the system is a glueless mesh.
6469 static void init_numa_topology_type(void)
6473 n
= sched_max_numa_distance
;
6475 if (sched_domains_numa_levels
<= 1) {
6476 sched_numa_topology_type
= NUMA_DIRECT
;
6480 for_each_online_node(a
) {
6481 for_each_online_node(b
) {
6482 /* Find two nodes furthest removed from each other. */
6483 if (node_distance(a
, b
) < n
)
6486 /* Is there an intermediary node between a and b? */
6487 for_each_online_node(c
) {
6488 if (node_distance(a
, c
) < n
&&
6489 node_distance(b
, c
) < n
) {
6490 sched_numa_topology_type
=
6496 sched_numa_topology_type
= NUMA_BACKPLANE
;
6502 static void sched_init_numa(void)
6504 int next_distance
, curr_distance
= node_distance(0, 0);
6505 struct sched_domain_topology_level
*tl
;
6509 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6510 if (!sched_domains_numa_distance
)
6514 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6515 * unique distances in the node_distance() table.
6517 * Assumes node_distance(0,j) includes all distances in
6518 * node_distance(i,j) in order to avoid cubic time.
6520 next_distance
= curr_distance
;
6521 for (i
= 0; i
< nr_node_ids
; i
++) {
6522 for (j
= 0; j
< nr_node_ids
; j
++) {
6523 for (k
= 0; k
< nr_node_ids
; k
++) {
6524 int distance
= node_distance(i
, k
);
6526 if (distance
> curr_distance
&&
6527 (distance
< next_distance
||
6528 next_distance
== curr_distance
))
6529 next_distance
= distance
;
6532 * While not a strong assumption it would be nice to know
6533 * about cases where if node A is connected to B, B is not
6534 * equally connected to A.
6536 if (sched_debug() && node_distance(k
, i
) != distance
)
6537 sched_numa_warn("Node-distance not symmetric");
6539 if (sched_debug() && i
&& !find_numa_distance(distance
))
6540 sched_numa_warn("Node-0 not representative");
6542 if (next_distance
!= curr_distance
) {
6543 sched_domains_numa_distance
[level
++] = next_distance
;
6544 sched_domains_numa_levels
= level
;
6545 curr_distance
= next_distance
;
6550 * In case of sched_debug() we verify the above assumption.
6560 * 'level' contains the number of unique distances, excluding the
6561 * identity distance node_distance(i,i).
6563 * The sched_domains_numa_distance[] array includes the actual distance
6568 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6569 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6570 * the array will contain less then 'level' members. This could be
6571 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6572 * in other functions.
6574 * We reset it to 'level' at the end of this function.
6576 sched_domains_numa_levels
= 0;
6578 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6579 if (!sched_domains_numa_masks
)
6583 * Now for each level, construct a mask per node which contains all
6584 * cpus of nodes that are that many hops away from us.
6586 for (i
= 0; i
< level
; i
++) {
6587 sched_domains_numa_masks
[i
] =
6588 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6589 if (!sched_domains_numa_masks
[i
])
6592 for (j
= 0; j
< nr_node_ids
; j
++) {
6593 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6597 sched_domains_numa_masks
[i
][j
] = mask
;
6599 for (k
= 0; k
< nr_node_ids
; k
++) {
6600 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6603 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6608 /* Compute default topology size */
6609 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6611 tl
= kzalloc((i
+ level
+ 1) *
6612 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6617 * Copy the default topology bits..
6619 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6620 tl
[i
] = sched_domain_topology
[i
];
6623 * .. and append 'j' levels of NUMA goodness.
6625 for (j
= 0; j
< level
; i
++, j
++) {
6626 tl
[i
] = (struct sched_domain_topology_level
){
6627 .mask
= sd_numa_mask
,
6628 .sd_flags
= cpu_numa_flags
,
6629 .flags
= SDTL_OVERLAP
,
6635 sched_domain_topology
= tl
;
6637 sched_domains_numa_levels
= level
;
6638 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6640 init_numa_topology_type();
6643 static void sched_domains_numa_masks_set(int cpu
)
6646 int node
= cpu_to_node(cpu
);
6648 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6649 for (j
= 0; j
< nr_node_ids
; j
++) {
6650 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6651 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6656 static void sched_domains_numa_masks_clear(int cpu
)
6659 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6660 for (j
= 0; j
< nr_node_ids
; j
++)
6661 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6666 * Update sched_domains_numa_masks[level][node] array when new cpus
6669 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6670 unsigned long action
,
6673 int cpu
= (long)hcpu
;
6675 switch (action
& ~CPU_TASKS_FROZEN
) {
6677 sched_domains_numa_masks_set(cpu
);
6681 sched_domains_numa_masks_clear(cpu
);
6691 static inline void sched_init_numa(void)
6695 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6696 unsigned long action
,
6701 #endif /* CONFIG_NUMA */
6703 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6705 struct sched_domain_topology_level
*tl
;
6708 for_each_sd_topology(tl
) {
6709 struct sd_data
*sdd
= &tl
->data
;
6711 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6715 sdd
->sg
= alloc_percpu(struct sched_group
*);
6719 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6723 for_each_cpu(j
, cpu_map
) {
6724 struct sched_domain
*sd
;
6725 struct sched_group
*sg
;
6726 struct sched_group_capacity
*sgc
;
6728 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6729 GFP_KERNEL
, cpu_to_node(j
));
6733 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6735 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6736 GFP_KERNEL
, cpu_to_node(j
));
6742 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6744 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6745 GFP_KERNEL
, cpu_to_node(j
));
6749 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6756 static void __sdt_free(const struct cpumask
*cpu_map
)
6758 struct sched_domain_topology_level
*tl
;
6761 for_each_sd_topology(tl
) {
6762 struct sd_data
*sdd
= &tl
->data
;
6764 for_each_cpu(j
, cpu_map
) {
6765 struct sched_domain
*sd
;
6768 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6769 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6770 free_sched_groups(sd
->groups
, 0);
6771 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6775 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6777 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6779 free_percpu(sdd
->sd
);
6781 free_percpu(sdd
->sg
);
6783 free_percpu(sdd
->sgc
);
6788 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6789 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6790 struct sched_domain
*child
, int cpu
)
6792 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6796 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6798 sd
->level
= child
->level
+ 1;
6799 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6803 if (!cpumask_subset(sched_domain_span(child
),
6804 sched_domain_span(sd
))) {
6805 pr_err("BUG: arch topology borken\n");
6806 #ifdef CONFIG_SCHED_DEBUG
6807 pr_err(" the %s domain not a subset of the %s domain\n",
6808 child
->name
, sd
->name
);
6810 /* Fixup, ensure @sd has at least @child cpus. */
6811 cpumask_or(sched_domain_span(sd
),
6812 sched_domain_span(sd
),
6813 sched_domain_span(child
));
6817 set_domain_attribute(sd
, attr
);
6823 * Build sched domains for a given set of cpus and attach the sched domains
6824 * to the individual cpus
6826 static int build_sched_domains(const struct cpumask
*cpu_map
,
6827 struct sched_domain_attr
*attr
)
6829 enum s_alloc alloc_state
;
6830 struct sched_domain
*sd
;
6832 int i
, ret
= -ENOMEM
;
6834 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6835 if (alloc_state
!= sa_rootdomain
)
6838 /* Set up domains for cpus specified by the cpu_map. */
6839 for_each_cpu(i
, cpu_map
) {
6840 struct sched_domain_topology_level
*tl
;
6843 for_each_sd_topology(tl
) {
6844 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6845 if (tl
== sched_domain_topology
)
6846 *per_cpu_ptr(d
.sd
, i
) = sd
;
6847 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6848 sd
->flags
|= SD_OVERLAP
;
6849 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6854 /* Build the groups for the domains */
6855 for_each_cpu(i
, cpu_map
) {
6856 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6857 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6858 if (sd
->flags
& SD_OVERLAP
) {
6859 if (build_overlap_sched_groups(sd
, i
))
6862 if (build_sched_groups(sd
, i
))
6868 /* Calculate CPU capacity for physical packages and nodes */
6869 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6870 if (!cpumask_test_cpu(i
, cpu_map
))
6873 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6874 claim_allocations(i
, sd
);
6875 init_sched_groups_capacity(i
, sd
);
6879 /* Attach the domains */
6881 for_each_cpu(i
, cpu_map
) {
6882 sd
= *per_cpu_ptr(d
.sd
, i
);
6883 cpu_attach_domain(sd
, d
.rd
, i
);
6889 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6893 static cpumask_var_t
*doms_cur
; /* current sched domains */
6894 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6895 static struct sched_domain_attr
*dattr_cur
;
6896 /* attribues of custom domains in 'doms_cur' */
6899 * Special case: If a kmalloc of a doms_cur partition (array of
6900 * cpumask) fails, then fallback to a single sched domain,
6901 * as determined by the single cpumask fallback_doms.
6903 static cpumask_var_t fallback_doms
;
6906 * arch_update_cpu_topology lets virtualized architectures update the
6907 * cpu core maps. It is supposed to return 1 if the topology changed
6908 * or 0 if it stayed the same.
6910 int __weak
arch_update_cpu_topology(void)
6915 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6918 cpumask_var_t
*doms
;
6920 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6923 for (i
= 0; i
< ndoms
; i
++) {
6924 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6925 free_sched_domains(doms
, i
);
6932 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6935 for (i
= 0; i
< ndoms
; i
++)
6936 free_cpumask_var(doms
[i
]);
6941 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6942 * For now this just excludes isolated cpus, but could be used to
6943 * exclude other special cases in the future.
6945 static int init_sched_domains(const struct cpumask
*cpu_map
)
6949 arch_update_cpu_topology();
6951 doms_cur
= alloc_sched_domains(ndoms_cur
);
6953 doms_cur
= &fallback_doms
;
6954 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6955 err
= build_sched_domains(doms_cur
[0], NULL
);
6956 register_sched_domain_sysctl();
6962 * Detach sched domains from a group of cpus specified in cpu_map
6963 * These cpus will now be attached to the NULL domain
6965 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6970 for_each_cpu(i
, cpu_map
)
6971 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6975 /* handle null as "default" */
6976 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6977 struct sched_domain_attr
*new, int idx_new
)
6979 struct sched_domain_attr tmp
;
6986 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6987 new ? (new + idx_new
) : &tmp
,
6988 sizeof(struct sched_domain_attr
));
6992 * Partition sched domains as specified by the 'ndoms_new'
6993 * cpumasks in the array doms_new[] of cpumasks. This compares
6994 * doms_new[] to the current sched domain partitioning, doms_cur[].
6995 * It destroys each deleted domain and builds each new domain.
6997 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6998 * The masks don't intersect (don't overlap.) We should setup one
6999 * sched domain for each mask. CPUs not in any of the cpumasks will
7000 * not be load balanced. If the same cpumask appears both in the
7001 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7004 * The passed in 'doms_new' should be allocated using
7005 * alloc_sched_domains. This routine takes ownership of it and will
7006 * free_sched_domains it when done with it. If the caller failed the
7007 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7008 * and partition_sched_domains() will fallback to the single partition
7009 * 'fallback_doms', it also forces the domains to be rebuilt.
7011 * If doms_new == NULL it will be replaced with cpu_online_mask.
7012 * ndoms_new == 0 is a special case for destroying existing domains,
7013 * and it will not create the default domain.
7015 * Call with hotplug lock held
7017 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7018 struct sched_domain_attr
*dattr_new
)
7023 mutex_lock(&sched_domains_mutex
);
7025 /* always unregister in case we don't destroy any domains */
7026 unregister_sched_domain_sysctl();
7028 /* Let architecture update cpu core mappings. */
7029 new_topology
= arch_update_cpu_topology();
7031 n
= doms_new
? ndoms_new
: 0;
7033 /* Destroy deleted domains */
7034 for (i
= 0; i
< ndoms_cur
; i
++) {
7035 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7036 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7037 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7040 /* no match - a current sched domain not in new doms_new[] */
7041 detach_destroy_domains(doms_cur
[i
]);
7047 if (doms_new
== NULL
) {
7049 doms_new
= &fallback_doms
;
7050 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7051 WARN_ON_ONCE(dattr_new
);
7054 /* Build new domains */
7055 for (i
= 0; i
< ndoms_new
; i
++) {
7056 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7057 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7058 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7061 /* no match - add a new doms_new */
7062 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7067 /* Remember the new sched domains */
7068 if (doms_cur
!= &fallback_doms
)
7069 free_sched_domains(doms_cur
, ndoms_cur
);
7070 kfree(dattr_cur
); /* kfree(NULL) is safe */
7071 doms_cur
= doms_new
;
7072 dattr_cur
= dattr_new
;
7073 ndoms_cur
= ndoms_new
;
7075 register_sched_domain_sysctl();
7077 mutex_unlock(&sched_domains_mutex
);
7080 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7083 * Update cpusets according to cpu_active mask. If cpusets are
7084 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7085 * around partition_sched_domains().
7087 * If we come here as part of a suspend/resume, don't touch cpusets because we
7088 * want to restore it back to its original state upon resume anyway.
7090 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7094 case CPU_ONLINE_FROZEN
:
7095 case CPU_DOWN_FAILED_FROZEN
:
7098 * num_cpus_frozen tracks how many CPUs are involved in suspend
7099 * resume sequence. As long as this is not the last online
7100 * operation in the resume sequence, just build a single sched
7101 * domain, ignoring cpusets.
7104 if (likely(num_cpus_frozen
)) {
7105 partition_sched_domains(1, NULL
, NULL
);
7110 * This is the last CPU online operation. So fall through and
7111 * restore the original sched domains by considering the
7112 * cpuset configurations.
7116 cpuset_update_active_cpus(true);
7124 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7127 unsigned long flags
;
7128 long cpu
= (long)hcpu
;
7134 case CPU_DOWN_PREPARE
:
7135 rcu_read_lock_sched();
7136 dl_b
= dl_bw_of(cpu
);
7138 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7139 cpus
= dl_bw_cpus(cpu
);
7140 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7141 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7143 rcu_read_unlock_sched();
7146 return notifier_from_errno(-EBUSY
);
7147 cpuset_update_active_cpus(false);
7149 case CPU_DOWN_PREPARE_FROZEN
:
7151 partition_sched_domains(1, NULL
, NULL
);
7159 void __init
sched_init_smp(void)
7161 cpumask_var_t non_isolated_cpus
;
7163 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7164 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7166 /* nohz_full won't take effect without isolating the cpus. */
7167 tick_nohz_full_add_cpus_to(cpu_isolated_map
);
7172 * There's no userspace yet to cause hotplug operations; hence all the
7173 * cpu masks are stable and all blatant races in the below code cannot
7176 mutex_lock(&sched_domains_mutex
);
7177 init_sched_domains(cpu_active_mask
);
7178 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7179 if (cpumask_empty(non_isolated_cpus
))
7180 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7181 mutex_unlock(&sched_domains_mutex
);
7183 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7184 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7185 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7189 /* Move init over to a non-isolated CPU */
7190 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7192 sched_init_granularity();
7193 free_cpumask_var(non_isolated_cpus
);
7195 init_sched_rt_class();
7196 init_sched_dl_class();
7199 void __init
sched_init_smp(void)
7201 sched_init_granularity();
7203 #endif /* CONFIG_SMP */
7205 int in_sched_functions(unsigned long addr
)
7207 return in_lock_functions(addr
) ||
7208 (addr
>= (unsigned long)__sched_text_start
7209 && addr
< (unsigned long)__sched_text_end
);
7212 #ifdef CONFIG_CGROUP_SCHED
7214 * Default task group.
7215 * Every task in system belongs to this group at bootup.
7217 struct task_group root_task_group
;
7218 LIST_HEAD(task_groups
);
7221 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7223 void __init
sched_init(void)
7226 unsigned long alloc_size
= 0, ptr
;
7228 #ifdef CONFIG_FAIR_GROUP_SCHED
7229 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7231 #ifdef CONFIG_RT_GROUP_SCHED
7232 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7235 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7237 #ifdef CONFIG_FAIR_GROUP_SCHED
7238 root_task_group
.se
= (struct sched_entity
**)ptr
;
7239 ptr
+= nr_cpu_ids
* sizeof(void **);
7241 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7242 ptr
+= nr_cpu_ids
* sizeof(void **);
7244 #endif /* CONFIG_FAIR_GROUP_SCHED */
7245 #ifdef CONFIG_RT_GROUP_SCHED
7246 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7247 ptr
+= nr_cpu_ids
* sizeof(void **);
7249 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7250 ptr
+= nr_cpu_ids
* sizeof(void **);
7252 #endif /* CONFIG_RT_GROUP_SCHED */
7254 #ifdef CONFIG_CPUMASK_OFFSTACK
7255 for_each_possible_cpu(i
) {
7256 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7257 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7259 #endif /* CONFIG_CPUMASK_OFFSTACK */
7261 init_rt_bandwidth(&def_rt_bandwidth
,
7262 global_rt_period(), global_rt_runtime());
7263 init_dl_bandwidth(&def_dl_bandwidth
,
7264 global_rt_period(), global_rt_runtime());
7267 init_defrootdomain();
7270 #ifdef CONFIG_RT_GROUP_SCHED
7271 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7272 global_rt_period(), global_rt_runtime());
7273 #endif /* CONFIG_RT_GROUP_SCHED */
7275 #ifdef CONFIG_CGROUP_SCHED
7276 list_add(&root_task_group
.list
, &task_groups
);
7277 INIT_LIST_HEAD(&root_task_group
.children
);
7278 INIT_LIST_HEAD(&root_task_group
.siblings
);
7279 autogroup_init(&init_task
);
7281 #endif /* CONFIG_CGROUP_SCHED */
7283 for_each_possible_cpu(i
) {
7287 raw_spin_lock_init(&rq
->lock
);
7289 rq
->calc_load_active
= 0;
7290 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7291 init_cfs_rq(&rq
->cfs
);
7292 init_rt_rq(&rq
->rt
);
7293 init_dl_rq(&rq
->dl
);
7294 #ifdef CONFIG_FAIR_GROUP_SCHED
7295 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7296 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7298 * How much cpu bandwidth does root_task_group get?
7300 * In case of task-groups formed thr' the cgroup filesystem, it
7301 * gets 100% of the cpu resources in the system. This overall
7302 * system cpu resource is divided among the tasks of
7303 * root_task_group and its child task-groups in a fair manner,
7304 * based on each entity's (task or task-group's) weight
7305 * (se->load.weight).
7307 * In other words, if root_task_group has 10 tasks of weight
7308 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7309 * then A0's share of the cpu resource is:
7311 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7313 * We achieve this by letting root_task_group's tasks sit
7314 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7316 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7317 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7318 #endif /* CONFIG_FAIR_GROUP_SCHED */
7320 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7321 #ifdef CONFIG_RT_GROUP_SCHED
7322 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7325 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7326 rq
->cpu_load
[j
] = 0;
7328 rq
->last_load_update_tick
= jiffies
;
7333 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7334 rq
->balance_callback
= NULL
;
7335 rq
->active_balance
= 0;
7336 rq
->next_balance
= jiffies
;
7341 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7342 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7344 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7346 rq_attach_root(rq
, &def_root_domain
);
7347 #ifdef CONFIG_NO_HZ_COMMON
7350 #ifdef CONFIG_NO_HZ_FULL
7351 rq
->last_sched_tick
= 0;
7355 atomic_set(&rq
->nr_iowait
, 0);
7358 set_load_weight(&init_task
);
7360 #ifdef CONFIG_PREEMPT_NOTIFIERS
7361 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7365 * The boot idle thread does lazy MMU switching as well:
7367 atomic_inc(&init_mm
.mm_count
);
7368 enter_lazy_tlb(&init_mm
, current
);
7371 * During early bootup we pretend to be a normal task:
7373 current
->sched_class
= &fair_sched_class
;
7376 * Make us the idle thread. Technically, schedule() should not be
7377 * called from this thread, however somewhere below it might be,
7378 * but because we are the idle thread, we just pick up running again
7379 * when this runqueue becomes "idle".
7381 init_idle(current
, smp_processor_id());
7383 calc_load_update
= jiffies
+ LOAD_FREQ
;
7386 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7387 /* May be allocated at isolcpus cmdline parse time */
7388 if (cpu_isolated_map
== NULL
)
7389 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7390 idle_thread_set_boot_cpu();
7391 set_cpu_rq_start_time();
7393 init_sched_fair_class();
7395 scheduler_running
= 1;
7398 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7399 static inline int preempt_count_equals(int preempt_offset
)
7401 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7403 return (nested
== preempt_offset
);
7406 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7409 * Blocking primitives will set (and therefore destroy) current->state,
7410 * since we will exit with TASK_RUNNING make sure we enter with it,
7411 * otherwise we will destroy state.
7413 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7414 "do not call blocking ops when !TASK_RUNNING; "
7415 "state=%lx set at [<%p>] %pS\n",
7417 (void *)current
->task_state_change
,
7418 (void *)current
->task_state_change
);
7420 ___might_sleep(file
, line
, preempt_offset
);
7422 EXPORT_SYMBOL(__might_sleep
);
7424 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7426 static unsigned long prev_jiffy
; /* ratelimiting */
7428 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7429 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7430 !is_idle_task(current
)) ||
7431 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7433 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7435 prev_jiffy
= jiffies
;
7438 "BUG: sleeping function called from invalid context at %s:%d\n",
7441 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7442 in_atomic(), irqs_disabled(),
7443 current
->pid
, current
->comm
);
7445 if (task_stack_end_corrupted(current
))
7446 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7448 debug_show_held_locks(current
);
7449 if (irqs_disabled())
7450 print_irqtrace_events(current
);
7451 #ifdef CONFIG_DEBUG_PREEMPT
7452 if (!preempt_count_equals(preempt_offset
)) {
7453 pr_err("Preemption disabled at:");
7454 print_ip_sym(current
->preempt_disable_ip
);
7460 EXPORT_SYMBOL(___might_sleep
);
7463 #ifdef CONFIG_MAGIC_SYSRQ
7464 void normalize_rt_tasks(void)
7466 struct task_struct
*g
, *p
;
7467 struct sched_attr attr
= {
7468 .sched_policy
= SCHED_NORMAL
,
7471 read_lock(&tasklist_lock
);
7472 for_each_process_thread(g
, p
) {
7474 * Only normalize user tasks:
7476 if (p
->flags
& PF_KTHREAD
)
7479 p
->se
.exec_start
= 0;
7480 #ifdef CONFIG_SCHEDSTATS
7481 p
->se
.statistics
.wait_start
= 0;
7482 p
->se
.statistics
.sleep_start
= 0;
7483 p
->se
.statistics
.block_start
= 0;
7486 if (!dl_task(p
) && !rt_task(p
)) {
7488 * Renice negative nice level userspace
7491 if (task_nice(p
) < 0)
7492 set_user_nice(p
, 0);
7496 __sched_setscheduler(p
, &attr
, false, false);
7498 read_unlock(&tasklist_lock
);
7501 #endif /* CONFIG_MAGIC_SYSRQ */
7503 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7505 * These functions are only useful for the IA64 MCA handling, or kdb.
7507 * They can only be called when the whole system has been
7508 * stopped - every CPU needs to be quiescent, and no scheduling
7509 * activity can take place. Using them for anything else would
7510 * be a serious bug, and as a result, they aren't even visible
7511 * under any other configuration.
7515 * curr_task - return the current task for a given cpu.
7516 * @cpu: the processor in question.
7518 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7520 * Return: The current task for @cpu.
7522 struct task_struct
*curr_task(int cpu
)
7524 return cpu_curr(cpu
);
7527 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7531 * set_curr_task - set the current task for a given cpu.
7532 * @cpu: the processor in question.
7533 * @p: the task pointer to set.
7535 * Description: This function must only be used when non-maskable interrupts
7536 * are serviced on a separate stack. It allows the architecture to switch the
7537 * notion of the current task on a cpu in a non-blocking manner. This function
7538 * must be called with all CPU's synchronized, and interrupts disabled, the
7539 * and caller must save the original value of the current task (see
7540 * curr_task() above) and restore that value before reenabling interrupts and
7541 * re-starting the system.
7543 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7545 void set_curr_task(int cpu
, struct task_struct
*p
)
7552 #ifdef CONFIG_CGROUP_SCHED
7553 /* task_group_lock serializes the addition/removal of task groups */
7554 static DEFINE_SPINLOCK(task_group_lock
);
7556 static void free_sched_group(struct task_group
*tg
)
7558 free_fair_sched_group(tg
);
7559 free_rt_sched_group(tg
);
7564 /* allocate runqueue etc for a new task group */
7565 struct task_group
*sched_create_group(struct task_group
*parent
)
7567 struct task_group
*tg
;
7569 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7571 return ERR_PTR(-ENOMEM
);
7573 if (!alloc_fair_sched_group(tg
, parent
))
7576 if (!alloc_rt_sched_group(tg
, parent
))
7582 free_sched_group(tg
);
7583 return ERR_PTR(-ENOMEM
);
7586 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7588 unsigned long flags
;
7590 spin_lock_irqsave(&task_group_lock
, flags
);
7591 list_add_rcu(&tg
->list
, &task_groups
);
7593 WARN_ON(!parent
); /* root should already exist */
7595 tg
->parent
= parent
;
7596 INIT_LIST_HEAD(&tg
->children
);
7597 list_add_rcu(&tg
->siblings
, &parent
->children
);
7598 spin_unlock_irqrestore(&task_group_lock
, flags
);
7601 /* rcu callback to free various structures associated with a task group */
7602 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7604 /* now it should be safe to free those cfs_rqs */
7605 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7608 /* Destroy runqueue etc associated with a task group */
7609 void sched_destroy_group(struct task_group
*tg
)
7611 /* wait for possible concurrent references to cfs_rqs complete */
7612 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7615 void sched_offline_group(struct task_group
*tg
)
7617 unsigned long flags
;
7620 /* end participation in shares distribution */
7621 for_each_possible_cpu(i
)
7622 unregister_fair_sched_group(tg
, i
);
7624 spin_lock_irqsave(&task_group_lock
, flags
);
7625 list_del_rcu(&tg
->list
);
7626 list_del_rcu(&tg
->siblings
);
7627 spin_unlock_irqrestore(&task_group_lock
, flags
);
7630 /* change task's runqueue when it moves between groups.
7631 * The caller of this function should have put the task in its new group
7632 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7633 * reflect its new group.
7635 void sched_move_task(struct task_struct
*tsk
)
7637 struct task_group
*tg
;
7638 int queued
, running
;
7639 unsigned long flags
;
7642 rq
= task_rq_lock(tsk
, &flags
);
7644 running
= task_current(rq
, tsk
);
7645 queued
= task_on_rq_queued(tsk
);
7648 dequeue_task(rq
, tsk
, 0);
7649 if (unlikely(running
))
7650 put_prev_task(rq
, tsk
);
7653 * All callers are synchronized by task_rq_lock(); we do not use RCU
7654 * which is pointless here. Thus, we pass "true" to task_css_check()
7655 * to prevent lockdep warnings.
7657 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7658 struct task_group
, css
);
7659 tg
= autogroup_task_group(tsk
, tg
);
7660 tsk
->sched_task_group
= tg
;
7662 #ifdef CONFIG_FAIR_GROUP_SCHED
7663 if (tsk
->sched_class
->task_move_group
)
7664 tsk
->sched_class
->task_move_group(tsk
, queued
);
7667 set_task_rq(tsk
, task_cpu(tsk
));
7669 if (unlikely(running
))
7670 tsk
->sched_class
->set_curr_task(rq
);
7672 enqueue_task(rq
, tsk
, 0);
7674 task_rq_unlock(rq
, tsk
, &flags
);
7676 #endif /* CONFIG_CGROUP_SCHED */
7678 #ifdef CONFIG_RT_GROUP_SCHED
7680 * Ensure that the real time constraints are schedulable.
7682 static DEFINE_MUTEX(rt_constraints_mutex
);
7684 /* Must be called with tasklist_lock held */
7685 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7687 struct task_struct
*g
, *p
;
7690 * Autogroups do not have RT tasks; see autogroup_create().
7692 if (task_group_is_autogroup(tg
))
7695 for_each_process_thread(g
, p
) {
7696 if (rt_task(p
) && task_group(p
) == tg
)
7703 struct rt_schedulable_data
{
7704 struct task_group
*tg
;
7709 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7711 struct rt_schedulable_data
*d
= data
;
7712 struct task_group
*child
;
7713 unsigned long total
, sum
= 0;
7714 u64 period
, runtime
;
7716 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7717 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7720 period
= d
->rt_period
;
7721 runtime
= d
->rt_runtime
;
7725 * Cannot have more runtime than the period.
7727 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7731 * Ensure we don't starve existing RT tasks.
7733 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7736 total
= to_ratio(period
, runtime
);
7739 * Nobody can have more than the global setting allows.
7741 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7745 * The sum of our children's runtime should not exceed our own.
7747 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7748 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7749 runtime
= child
->rt_bandwidth
.rt_runtime
;
7751 if (child
== d
->tg
) {
7752 period
= d
->rt_period
;
7753 runtime
= d
->rt_runtime
;
7756 sum
+= to_ratio(period
, runtime
);
7765 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7769 struct rt_schedulable_data data
= {
7771 .rt_period
= period
,
7772 .rt_runtime
= runtime
,
7776 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7782 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7783 u64 rt_period
, u64 rt_runtime
)
7788 * Disallowing the root group RT runtime is BAD, it would disallow the
7789 * kernel creating (and or operating) RT threads.
7791 if (tg
== &root_task_group
&& rt_runtime
== 0)
7794 /* No period doesn't make any sense. */
7798 mutex_lock(&rt_constraints_mutex
);
7799 read_lock(&tasklist_lock
);
7800 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7804 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7805 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7806 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7808 for_each_possible_cpu(i
) {
7809 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7811 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7812 rt_rq
->rt_runtime
= rt_runtime
;
7813 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7815 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7817 read_unlock(&tasklist_lock
);
7818 mutex_unlock(&rt_constraints_mutex
);
7823 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7825 u64 rt_runtime
, rt_period
;
7827 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7828 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7829 if (rt_runtime_us
< 0)
7830 rt_runtime
= RUNTIME_INF
;
7832 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7835 static long sched_group_rt_runtime(struct task_group
*tg
)
7839 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7842 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7843 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7844 return rt_runtime_us
;
7847 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7849 u64 rt_runtime
, rt_period
;
7851 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7852 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7854 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7857 static long sched_group_rt_period(struct task_group
*tg
)
7861 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7862 do_div(rt_period_us
, NSEC_PER_USEC
);
7863 return rt_period_us
;
7865 #endif /* CONFIG_RT_GROUP_SCHED */
7867 #ifdef CONFIG_RT_GROUP_SCHED
7868 static int sched_rt_global_constraints(void)
7872 mutex_lock(&rt_constraints_mutex
);
7873 read_lock(&tasklist_lock
);
7874 ret
= __rt_schedulable(NULL
, 0, 0);
7875 read_unlock(&tasklist_lock
);
7876 mutex_unlock(&rt_constraints_mutex
);
7881 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7883 /* Don't accept realtime tasks when there is no way for them to run */
7884 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7890 #else /* !CONFIG_RT_GROUP_SCHED */
7891 static int sched_rt_global_constraints(void)
7893 unsigned long flags
;
7896 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7897 for_each_possible_cpu(i
) {
7898 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7900 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7901 rt_rq
->rt_runtime
= global_rt_runtime();
7902 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7904 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7908 #endif /* CONFIG_RT_GROUP_SCHED */
7910 static int sched_dl_global_validate(void)
7912 u64 runtime
= global_rt_runtime();
7913 u64 period
= global_rt_period();
7914 u64 new_bw
= to_ratio(period
, runtime
);
7917 unsigned long flags
;
7920 * Here we want to check the bandwidth not being set to some
7921 * value smaller than the currently allocated bandwidth in
7922 * any of the root_domains.
7924 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7925 * cycling on root_domains... Discussion on different/better
7926 * solutions is welcome!
7928 for_each_possible_cpu(cpu
) {
7929 rcu_read_lock_sched();
7930 dl_b
= dl_bw_of(cpu
);
7932 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7933 if (new_bw
< dl_b
->total_bw
)
7935 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7937 rcu_read_unlock_sched();
7946 static void sched_dl_do_global(void)
7951 unsigned long flags
;
7953 def_dl_bandwidth
.dl_period
= global_rt_period();
7954 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7956 if (global_rt_runtime() != RUNTIME_INF
)
7957 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7960 * FIXME: As above...
7962 for_each_possible_cpu(cpu
) {
7963 rcu_read_lock_sched();
7964 dl_b
= dl_bw_of(cpu
);
7966 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7968 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7970 rcu_read_unlock_sched();
7974 static int sched_rt_global_validate(void)
7976 if (sysctl_sched_rt_period
<= 0)
7979 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7980 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7986 static void sched_rt_do_global(void)
7988 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7989 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7992 int sched_rt_handler(struct ctl_table
*table
, int write
,
7993 void __user
*buffer
, size_t *lenp
,
7996 int old_period
, old_runtime
;
7997 static DEFINE_MUTEX(mutex
);
8001 old_period
= sysctl_sched_rt_period
;
8002 old_runtime
= sysctl_sched_rt_runtime
;
8004 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8006 if (!ret
&& write
) {
8007 ret
= sched_rt_global_validate();
8011 ret
= sched_dl_global_validate();
8015 ret
= sched_rt_global_constraints();
8019 sched_rt_do_global();
8020 sched_dl_do_global();
8024 sysctl_sched_rt_period
= old_period
;
8025 sysctl_sched_rt_runtime
= old_runtime
;
8027 mutex_unlock(&mutex
);
8032 int sched_rr_handler(struct ctl_table
*table
, int write
,
8033 void __user
*buffer
, size_t *lenp
,
8037 static DEFINE_MUTEX(mutex
);
8040 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8041 /* make sure that internally we keep jiffies */
8042 /* also, writing zero resets timeslice to default */
8043 if (!ret
&& write
) {
8044 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8045 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8047 mutex_unlock(&mutex
);
8051 #ifdef CONFIG_CGROUP_SCHED
8053 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8055 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8058 static struct cgroup_subsys_state
*
8059 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8061 struct task_group
*parent
= css_tg(parent_css
);
8062 struct task_group
*tg
;
8065 /* This is early initialization for the top cgroup */
8066 return &root_task_group
.css
;
8069 tg
= sched_create_group(parent
);
8071 return ERR_PTR(-ENOMEM
);
8076 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8078 struct task_group
*tg
= css_tg(css
);
8079 struct task_group
*parent
= css_tg(css
->parent
);
8082 sched_online_group(tg
, parent
);
8086 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8088 struct task_group
*tg
= css_tg(css
);
8090 sched_destroy_group(tg
);
8093 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8095 struct task_group
*tg
= css_tg(css
);
8097 sched_offline_group(tg
);
8100 static void cpu_cgroup_fork(struct task_struct
*task
)
8102 sched_move_task(task
);
8105 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
8106 struct cgroup_taskset
*tset
)
8108 struct task_struct
*task
;
8110 cgroup_taskset_for_each(task
, tset
) {
8111 #ifdef CONFIG_RT_GROUP_SCHED
8112 if (!sched_rt_can_attach(css_tg(css
), task
))
8115 /* We don't support RT-tasks being in separate groups */
8116 if (task
->sched_class
!= &fair_sched_class
)
8123 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8124 struct cgroup_taskset
*tset
)
8126 struct task_struct
*task
;
8128 cgroup_taskset_for_each(task
, tset
)
8129 sched_move_task(task
);
8132 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8133 struct cgroup_subsys_state
*old_css
,
8134 struct task_struct
*task
)
8137 * cgroup_exit() is called in the copy_process() failure path.
8138 * Ignore this case since the task hasn't ran yet, this avoids
8139 * trying to poke a half freed task state from generic code.
8141 if (!(task
->flags
& PF_EXITING
))
8144 sched_move_task(task
);
8147 #ifdef CONFIG_FAIR_GROUP_SCHED
8148 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8149 struct cftype
*cftype
, u64 shareval
)
8151 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8154 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8157 struct task_group
*tg
= css_tg(css
);
8159 return (u64
) scale_load_down(tg
->shares
);
8162 #ifdef CONFIG_CFS_BANDWIDTH
8163 static DEFINE_MUTEX(cfs_constraints_mutex
);
8165 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8166 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8168 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8170 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8172 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8173 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8175 if (tg
== &root_task_group
)
8179 * Ensure we have at some amount of bandwidth every period. This is
8180 * to prevent reaching a state of large arrears when throttled via
8181 * entity_tick() resulting in prolonged exit starvation.
8183 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8187 * Likewise, bound things on the otherside by preventing insane quota
8188 * periods. This also allows us to normalize in computing quota
8191 if (period
> max_cfs_quota_period
)
8195 * Prevent race between setting of cfs_rq->runtime_enabled and
8196 * unthrottle_offline_cfs_rqs().
8199 mutex_lock(&cfs_constraints_mutex
);
8200 ret
= __cfs_schedulable(tg
, period
, quota
);
8204 runtime_enabled
= quota
!= RUNTIME_INF
;
8205 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8207 * If we need to toggle cfs_bandwidth_used, off->on must occur
8208 * before making related changes, and on->off must occur afterwards
8210 if (runtime_enabled
&& !runtime_was_enabled
)
8211 cfs_bandwidth_usage_inc();
8212 raw_spin_lock_irq(&cfs_b
->lock
);
8213 cfs_b
->period
= ns_to_ktime(period
);
8214 cfs_b
->quota
= quota
;
8216 __refill_cfs_bandwidth_runtime(cfs_b
);
8217 /* restart the period timer (if active) to handle new period expiry */
8218 if (runtime_enabled
)
8219 start_cfs_bandwidth(cfs_b
);
8220 raw_spin_unlock_irq(&cfs_b
->lock
);
8222 for_each_online_cpu(i
) {
8223 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8224 struct rq
*rq
= cfs_rq
->rq
;
8226 raw_spin_lock_irq(&rq
->lock
);
8227 cfs_rq
->runtime_enabled
= runtime_enabled
;
8228 cfs_rq
->runtime_remaining
= 0;
8230 if (cfs_rq
->throttled
)
8231 unthrottle_cfs_rq(cfs_rq
);
8232 raw_spin_unlock_irq(&rq
->lock
);
8234 if (runtime_was_enabled
&& !runtime_enabled
)
8235 cfs_bandwidth_usage_dec();
8237 mutex_unlock(&cfs_constraints_mutex
);
8243 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8247 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8248 if (cfs_quota_us
< 0)
8249 quota
= RUNTIME_INF
;
8251 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8253 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8256 long tg_get_cfs_quota(struct task_group
*tg
)
8260 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8263 quota_us
= tg
->cfs_bandwidth
.quota
;
8264 do_div(quota_us
, NSEC_PER_USEC
);
8269 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8273 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8274 quota
= tg
->cfs_bandwidth
.quota
;
8276 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8279 long tg_get_cfs_period(struct task_group
*tg
)
8283 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8284 do_div(cfs_period_us
, NSEC_PER_USEC
);
8286 return cfs_period_us
;
8289 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8292 return tg_get_cfs_quota(css_tg(css
));
8295 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8296 struct cftype
*cftype
, s64 cfs_quota_us
)
8298 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8301 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8304 return tg_get_cfs_period(css_tg(css
));
8307 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8308 struct cftype
*cftype
, u64 cfs_period_us
)
8310 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8313 struct cfs_schedulable_data
{
8314 struct task_group
*tg
;
8319 * normalize group quota/period to be quota/max_period
8320 * note: units are usecs
8322 static u64
normalize_cfs_quota(struct task_group
*tg
,
8323 struct cfs_schedulable_data
*d
)
8331 period
= tg_get_cfs_period(tg
);
8332 quota
= tg_get_cfs_quota(tg
);
8335 /* note: these should typically be equivalent */
8336 if (quota
== RUNTIME_INF
|| quota
== -1)
8339 return to_ratio(period
, quota
);
8342 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8344 struct cfs_schedulable_data
*d
= data
;
8345 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8346 s64 quota
= 0, parent_quota
= -1;
8349 quota
= RUNTIME_INF
;
8351 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8353 quota
= normalize_cfs_quota(tg
, d
);
8354 parent_quota
= parent_b
->hierarchical_quota
;
8357 * ensure max(child_quota) <= parent_quota, inherit when no
8360 if (quota
== RUNTIME_INF
)
8361 quota
= parent_quota
;
8362 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8365 cfs_b
->hierarchical_quota
= quota
;
8370 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8373 struct cfs_schedulable_data data
= {
8379 if (quota
!= RUNTIME_INF
) {
8380 do_div(data
.period
, NSEC_PER_USEC
);
8381 do_div(data
.quota
, NSEC_PER_USEC
);
8385 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8391 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8393 struct task_group
*tg
= css_tg(seq_css(sf
));
8394 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8396 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8397 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8398 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8402 #endif /* CONFIG_CFS_BANDWIDTH */
8403 #endif /* CONFIG_FAIR_GROUP_SCHED */
8405 #ifdef CONFIG_RT_GROUP_SCHED
8406 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8407 struct cftype
*cft
, s64 val
)
8409 return sched_group_set_rt_runtime(css_tg(css
), val
);
8412 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8415 return sched_group_rt_runtime(css_tg(css
));
8418 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8419 struct cftype
*cftype
, u64 rt_period_us
)
8421 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8424 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8427 return sched_group_rt_period(css_tg(css
));
8429 #endif /* CONFIG_RT_GROUP_SCHED */
8431 static struct cftype cpu_files
[] = {
8432 #ifdef CONFIG_FAIR_GROUP_SCHED
8435 .read_u64
= cpu_shares_read_u64
,
8436 .write_u64
= cpu_shares_write_u64
,
8439 #ifdef CONFIG_CFS_BANDWIDTH
8441 .name
= "cfs_quota_us",
8442 .read_s64
= cpu_cfs_quota_read_s64
,
8443 .write_s64
= cpu_cfs_quota_write_s64
,
8446 .name
= "cfs_period_us",
8447 .read_u64
= cpu_cfs_period_read_u64
,
8448 .write_u64
= cpu_cfs_period_write_u64
,
8452 .seq_show
= cpu_stats_show
,
8455 #ifdef CONFIG_RT_GROUP_SCHED
8457 .name
= "rt_runtime_us",
8458 .read_s64
= cpu_rt_runtime_read
,
8459 .write_s64
= cpu_rt_runtime_write
,
8462 .name
= "rt_period_us",
8463 .read_u64
= cpu_rt_period_read_uint
,
8464 .write_u64
= cpu_rt_period_write_uint
,
8470 struct cgroup_subsys cpu_cgrp_subsys
= {
8471 .css_alloc
= cpu_cgroup_css_alloc
,
8472 .css_free
= cpu_cgroup_css_free
,
8473 .css_online
= cpu_cgroup_css_online
,
8474 .css_offline
= cpu_cgroup_css_offline
,
8475 .fork
= cpu_cgroup_fork
,
8476 .can_attach
= cpu_cgroup_can_attach
,
8477 .attach
= cpu_cgroup_attach
,
8478 .exit
= cpu_cgroup_exit
,
8479 .legacy_cftypes
= cpu_files
,
8483 #endif /* CONFIG_CGROUP_SCHED */
8485 void dump_cpu_task(int cpu
)
8487 pr_info("Task dump for CPU %d:\n", cpu
);
8488 sched_show_task(cpu_curr(cpu
));