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
= ACCESS_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
)
524 * resched_curr - mark rq's current task 'to be rescheduled now'.
526 * On UP this means the setting of the need_resched flag, on SMP it
527 * might also involve a cross-CPU call to trigger the scheduler on
530 void resched_curr(struct rq
*rq
)
532 struct task_struct
*curr
= rq
->curr
;
535 lockdep_assert_held(&rq
->lock
);
537 if (test_tsk_need_resched(curr
))
542 if (cpu
== smp_processor_id()) {
543 set_tsk_need_resched(curr
);
544 set_preempt_need_resched();
548 if (set_nr_and_not_polling(curr
))
549 smp_send_reschedule(cpu
);
551 trace_sched_wake_idle_without_ipi(cpu
);
554 void resched_cpu(int cpu
)
556 struct rq
*rq
= cpu_rq(cpu
);
559 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
562 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
566 #ifdef CONFIG_NO_HZ_COMMON
568 * In the semi idle case, use the nearest busy cpu for migrating timers
569 * from an idle cpu. This is good for power-savings.
571 * We don't do similar optimization for completely idle system, as
572 * selecting an idle cpu will add more delays to the timers than intended
573 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
575 int get_nohz_timer_target(int pinned
)
577 int cpu
= smp_processor_id();
579 struct sched_domain
*sd
;
581 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
585 for_each_domain(cpu
, sd
) {
586 for_each_cpu(i
, sched_domain_span(sd
)) {
598 * When add_timer_on() enqueues a timer into the timer wheel of an
599 * idle CPU then this timer might expire before the next timer event
600 * which is scheduled to wake up that CPU. In case of a completely
601 * idle system the next event might even be infinite time into the
602 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
603 * leaves the inner idle loop so the newly added timer is taken into
604 * account when the CPU goes back to idle and evaluates the timer
605 * wheel for the next timer event.
607 static void wake_up_idle_cpu(int cpu
)
609 struct rq
*rq
= cpu_rq(cpu
);
611 if (cpu
== smp_processor_id())
614 if (set_nr_and_not_polling(rq
->idle
))
615 smp_send_reschedule(cpu
);
617 trace_sched_wake_idle_without_ipi(cpu
);
620 static bool wake_up_full_nohz_cpu(int cpu
)
623 * We just need the target to call irq_exit() and re-evaluate
624 * the next tick. The nohz full kick at least implies that.
625 * If needed we can still optimize that later with an
628 if (tick_nohz_full_cpu(cpu
)) {
629 if (cpu
!= smp_processor_id() ||
630 tick_nohz_tick_stopped())
631 tick_nohz_full_kick_cpu(cpu
);
638 void wake_up_nohz_cpu(int cpu
)
640 if (!wake_up_full_nohz_cpu(cpu
))
641 wake_up_idle_cpu(cpu
);
644 static inline bool got_nohz_idle_kick(void)
646 int cpu
= smp_processor_id();
648 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
651 if (idle_cpu(cpu
) && !need_resched())
655 * We can't run Idle Load Balance on this CPU for this time so we
656 * cancel it and clear NOHZ_BALANCE_KICK
658 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
662 #else /* CONFIG_NO_HZ_COMMON */
664 static inline bool got_nohz_idle_kick(void)
669 #endif /* CONFIG_NO_HZ_COMMON */
671 #ifdef CONFIG_NO_HZ_FULL
672 bool sched_can_stop_tick(void)
675 * FIFO realtime policy runs the highest priority task. Other runnable
676 * tasks are of a lower priority. The scheduler tick does nothing.
678 if (current
->policy
== SCHED_FIFO
)
682 * Round-robin realtime tasks time slice with other tasks at the same
683 * realtime priority. Is this task the only one at this priority?
685 if (current
->policy
== SCHED_RR
) {
686 struct sched_rt_entity
*rt_se
= ¤t
->rt
;
688 return rt_se
->run_list
.prev
== rt_se
->run_list
.next
;
692 * More than one running task need preemption.
693 * nr_running update is assumed to be visible
694 * after IPI is sent from wakers.
696 if (this_rq()->nr_running
> 1)
701 #endif /* CONFIG_NO_HZ_FULL */
703 void sched_avg_update(struct rq
*rq
)
705 s64 period
= sched_avg_period();
707 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
709 * Inline assembly required to prevent the compiler
710 * optimising this loop into a divmod call.
711 * See __iter_div_u64_rem() for another example of this.
713 asm("" : "+rm" (rq
->age_stamp
));
714 rq
->age_stamp
+= period
;
719 #endif /* CONFIG_SMP */
721 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
722 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
724 * Iterate task_group tree rooted at *from, calling @down when first entering a
725 * node and @up when leaving it for the final time.
727 * Caller must hold rcu_lock or sufficient equivalent.
729 int walk_tg_tree_from(struct task_group
*from
,
730 tg_visitor down
, tg_visitor up
, void *data
)
732 struct task_group
*parent
, *child
;
738 ret
= (*down
)(parent
, data
);
741 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
748 ret
= (*up
)(parent
, data
);
749 if (ret
|| parent
== from
)
753 parent
= parent
->parent
;
760 int tg_nop(struct task_group
*tg
, void *data
)
766 static void set_load_weight(struct task_struct
*p
)
768 int prio
= p
->static_prio
- MAX_RT_PRIO
;
769 struct load_weight
*load
= &p
->se
.load
;
772 * SCHED_IDLE tasks get minimal weight:
774 if (p
->policy
== SCHED_IDLE
) {
775 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
776 load
->inv_weight
= WMULT_IDLEPRIO
;
780 load
->weight
= scale_load(prio_to_weight
[prio
]);
781 load
->inv_weight
= prio_to_wmult
[prio
];
784 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
787 sched_info_queued(rq
, p
);
788 p
->sched_class
->enqueue_task(rq
, p
, flags
);
791 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
794 sched_info_dequeued(rq
, p
);
795 p
->sched_class
->dequeue_task(rq
, p
, flags
);
798 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
800 if (task_contributes_to_load(p
))
801 rq
->nr_uninterruptible
--;
803 enqueue_task(rq
, p
, flags
);
806 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
808 if (task_contributes_to_load(p
))
809 rq
->nr_uninterruptible
++;
811 dequeue_task(rq
, p
, flags
);
814 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
817 * In theory, the compile should just see 0 here, and optimize out the call
818 * to sched_rt_avg_update. But I don't trust it...
820 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
821 s64 steal
= 0, irq_delta
= 0;
823 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
824 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
827 * Since irq_time is only updated on {soft,}irq_exit, we might run into
828 * this case when a previous update_rq_clock() happened inside a
831 * When this happens, we stop ->clock_task and only update the
832 * prev_irq_time stamp to account for the part that fit, so that a next
833 * update will consume the rest. This ensures ->clock_task is
836 * It does however cause some slight miss-attribution of {soft,}irq
837 * time, a more accurate solution would be to update the irq_time using
838 * the current rq->clock timestamp, except that would require using
841 if (irq_delta
> delta
)
844 rq
->prev_irq_time
+= irq_delta
;
847 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
848 if (static_key_false((¶virt_steal_rq_enabled
))) {
849 steal
= paravirt_steal_clock(cpu_of(rq
));
850 steal
-= rq
->prev_steal_time_rq
;
852 if (unlikely(steal
> delta
))
855 rq
->prev_steal_time_rq
+= steal
;
860 rq
->clock_task
+= delta
;
862 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
863 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
864 sched_rt_avg_update(rq
, irq_delta
+ steal
);
868 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
870 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
871 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
875 * Make it appear like a SCHED_FIFO task, its something
876 * userspace knows about and won't get confused about.
878 * Also, it will make PI more or less work without too
879 * much confusion -- but then, stop work should not
880 * rely on PI working anyway.
882 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
884 stop
->sched_class
= &stop_sched_class
;
887 cpu_rq(cpu
)->stop
= stop
;
891 * Reset it back to a normal scheduling class so that
892 * it can die in pieces.
894 old_stop
->sched_class
= &rt_sched_class
;
899 * __normal_prio - return the priority that is based on the static prio
901 static inline int __normal_prio(struct task_struct
*p
)
903 return p
->static_prio
;
907 * Calculate the expected normal priority: i.e. priority
908 * without taking RT-inheritance into account. Might be
909 * boosted by interactivity modifiers. Changes upon fork,
910 * setprio syscalls, and whenever the interactivity
911 * estimator recalculates.
913 static inline int normal_prio(struct task_struct
*p
)
917 if (task_has_dl_policy(p
))
918 prio
= MAX_DL_PRIO
-1;
919 else if (task_has_rt_policy(p
))
920 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
922 prio
= __normal_prio(p
);
927 * Calculate the current priority, i.e. the priority
928 * taken into account by the scheduler. This value might
929 * be boosted by RT tasks, or might be boosted by
930 * interactivity modifiers. Will be RT if the task got
931 * RT-boosted. If not then it returns p->normal_prio.
933 static int effective_prio(struct task_struct
*p
)
935 p
->normal_prio
= normal_prio(p
);
937 * If we are RT tasks or we were boosted to RT priority,
938 * keep the priority unchanged. Otherwise, update priority
939 * to the normal priority:
941 if (!rt_prio(p
->prio
))
942 return p
->normal_prio
;
947 * task_curr - is this task currently executing on a CPU?
948 * @p: the task in question.
950 * Return: 1 if the task is currently executing. 0 otherwise.
952 inline int task_curr(const struct task_struct
*p
)
954 return cpu_curr(task_cpu(p
)) == p
;
958 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
960 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
961 const struct sched_class
*prev_class
,
964 if (prev_class
!= p
->sched_class
) {
965 if (prev_class
->switched_from
)
966 prev_class
->switched_from(rq
, p
);
967 /* Possble rq->lock 'hole'. */
968 p
->sched_class
->switched_to(rq
, p
);
969 } else if (oldprio
!= p
->prio
|| dl_task(p
))
970 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
973 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
975 const struct sched_class
*class;
977 if (p
->sched_class
== rq
->curr
->sched_class
) {
978 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
980 for_each_class(class) {
981 if (class == rq
->curr
->sched_class
)
983 if (class == p
->sched_class
) {
991 * A queue event has occurred, and we're going to schedule. In
992 * this case, we can save a useless back to back clock update.
994 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
995 rq_clock_skip_update(rq
, true);
999 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1001 #ifdef CONFIG_SCHED_DEBUG
1003 * We should never call set_task_cpu() on a blocked task,
1004 * ttwu() will sort out the placement.
1006 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1009 #ifdef CONFIG_LOCKDEP
1011 * The caller should hold either p->pi_lock or rq->lock, when changing
1012 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1014 * sched_move_task() holds both and thus holding either pins the cgroup,
1017 * Furthermore, all task_rq users should acquire both locks, see
1020 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1021 lockdep_is_held(&task_rq(p
)->lock
)));
1025 trace_sched_migrate_task(p
, new_cpu
);
1027 if (task_cpu(p
) != new_cpu
) {
1028 if (p
->sched_class
->migrate_task_rq
)
1029 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1030 p
->se
.nr_migrations
++;
1031 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 0);
1034 __set_task_cpu(p
, new_cpu
);
1037 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1039 if (task_on_rq_queued(p
)) {
1040 struct rq
*src_rq
, *dst_rq
;
1042 src_rq
= task_rq(p
);
1043 dst_rq
= cpu_rq(cpu
);
1045 deactivate_task(src_rq
, p
, 0);
1046 set_task_cpu(p
, cpu
);
1047 activate_task(dst_rq
, p
, 0);
1048 check_preempt_curr(dst_rq
, p
, 0);
1051 * Task isn't running anymore; make it appear like we migrated
1052 * it before it went to sleep. This means on wakeup we make the
1053 * previous cpu our targer instead of where it really is.
1059 struct migration_swap_arg
{
1060 struct task_struct
*src_task
, *dst_task
;
1061 int src_cpu
, dst_cpu
;
1064 static int migrate_swap_stop(void *data
)
1066 struct migration_swap_arg
*arg
= data
;
1067 struct rq
*src_rq
, *dst_rq
;
1070 src_rq
= cpu_rq(arg
->src_cpu
);
1071 dst_rq
= cpu_rq(arg
->dst_cpu
);
1073 double_raw_lock(&arg
->src_task
->pi_lock
,
1074 &arg
->dst_task
->pi_lock
);
1075 double_rq_lock(src_rq
, dst_rq
);
1076 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1079 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1082 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1085 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1088 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1089 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1094 double_rq_unlock(src_rq
, dst_rq
);
1095 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1096 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1102 * Cross migrate two tasks
1104 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1106 struct migration_swap_arg arg
;
1109 arg
= (struct migration_swap_arg
){
1111 .src_cpu
= task_cpu(cur
),
1113 .dst_cpu
= task_cpu(p
),
1116 if (arg
.src_cpu
== arg
.dst_cpu
)
1120 * These three tests are all lockless; this is OK since all of them
1121 * will be re-checked with proper locks held further down the line.
1123 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1126 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1129 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1132 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1133 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1139 struct migration_arg
{
1140 struct task_struct
*task
;
1144 static int migration_cpu_stop(void *data
);
1147 * wait_task_inactive - wait for a thread to unschedule.
1149 * If @match_state is nonzero, it's the @p->state value just checked and
1150 * not expected to change. If it changes, i.e. @p might have woken up,
1151 * then return zero. When we succeed in waiting for @p to be off its CPU,
1152 * we return a positive number (its total switch count). If a second call
1153 * a short while later returns the same number, the caller can be sure that
1154 * @p has remained unscheduled the whole time.
1156 * The caller must ensure that the task *will* unschedule sometime soon,
1157 * else this function might spin for a *long* time. This function can't
1158 * be called with interrupts off, or it may introduce deadlock with
1159 * smp_call_function() if an IPI is sent by the same process we are
1160 * waiting to become inactive.
1162 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1164 unsigned long flags
;
1165 int running
, queued
;
1171 * We do the initial early heuristics without holding
1172 * any task-queue locks at all. We'll only try to get
1173 * the runqueue lock when things look like they will
1179 * If the task is actively running on another CPU
1180 * still, just relax and busy-wait without holding
1183 * NOTE! Since we don't hold any locks, it's not
1184 * even sure that "rq" stays as the right runqueue!
1185 * But we don't care, since "task_running()" will
1186 * return false if the runqueue has changed and p
1187 * is actually now running somewhere else!
1189 while (task_running(rq
, p
)) {
1190 if (match_state
&& unlikely(p
->state
!= match_state
))
1196 * Ok, time to look more closely! We need the rq
1197 * lock now, to be *sure*. If we're wrong, we'll
1198 * just go back and repeat.
1200 rq
= task_rq_lock(p
, &flags
);
1201 trace_sched_wait_task(p
);
1202 running
= task_running(rq
, p
);
1203 queued
= task_on_rq_queued(p
);
1205 if (!match_state
|| p
->state
== match_state
)
1206 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1207 task_rq_unlock(rq
, p
, &flags
);
1210 * If it changed from the expected state, bail out now.
1212 if (unlikely(!ncsw
))
1216 * Was it really running after all now that we
1217 * checked with the proper locks actually held?
1219 * Oops. Go back and try again..
1221 if (unlikely(running
)) {
1227 * It's not enough that it's not actively running,
1228 * it must be off the runqueue _entirely_, and not
1231 * So if it was still runnable (but just not actively
1232 * running right now), it's preempted, and we should
1233 * yield - it could be a while.
1235 if (unlikely(queued
)) {
1236 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1238 set_current_state(TASK_UNINTERRUPTIBLE
);
1239 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1244 * Ahh, all good. It wasn't running, and it wasn't
1245 * runnable, which means that it will never become
1246 * running in the future either. We're all done!
1255 * kick_process - kick a running thread to enter/exit the kernel
1256 * @p: the to-be-kicked thread
1258 * Cause a process which is running on another CPU to enter
1259 * kernel-mode, without any delay. (to get signals handled.)
1261 * NOTE: this function doesn't have to take the runqueue lock,
1262 * because all it wants to ensure is that the remote task enters
1263 * the kernel. If the IPI races and the task has been migrated
1264 * to another CPU then no harm is done and the purpose has been
1267 void kick_process(struct task_struct
*p
)
1273 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1274 smp_send_reschedule(cpu
);
1277 EXPORT_SYMBOL_GPL(kick_process
);
1278 #endif /* CONFIG_SMP */
1282 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1284 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1286 int nid
= cpu_to_node(cpu
);
1287 const struct cpumask
*nodemask
= NULL
;
1288 enum { cpuset
, possible
, fail
} state
= cpuset
;
1292 * If the node that the cpu is on has been offlined, cpu_to_node()
1293 * will return -1. There is no cpu on the node, and we should
1294 * select the cpu on the other node.
1297 nodemask
= cpumask_of_node(nid
);
1299 /* Look for allowed, online CPU in same node. */
1300 for_each_cpu(dest_cpu
, nodemask
) {
1301 if (!cpu_online(dest_cpu
))
1303 if (!cpu_active(dest_cpu
))
1305 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1311 /* Any allowed, online CPU? */
1312 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1313 if (!cpu_online(dest_cpu
))
1315 if (!cpu_active(dest_cpu
))
1322 /* No more Mr. Nice Guy. */
1323 cpuset_cpus_allowed_fallback(p
);
1328 do_set_cpus_allowed(p
, cpu_possible_mask
);
1339 if (state
!= cpuset
) {
1341 * Don't tell them about moving exiting tasks or
1342 * kernel threads (both mm NULL), since they never
1345 if (p
->mm
&& printk_ratelimit()) {
1346 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1347 task_pid_nr(p
), p
->comm
, cpu
);
1355 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1358 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1360 if (p
->nr_cpus_allowed
> 1)
1361 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1364 * In order not to call set_task_cpu() on a blocking task we need
1365 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1368 * Since this is common to all placement strategies, this lives here.
1370 * [ this allows ->select_task() to simply return task_cpu(p) and
1371 * not worry about this generic constraint ]
1373 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1375 cpu
= select_fallback_rq(task_cpu(p
), p
);
1380 static void update_avg(u64
*avg
, u64 sample
)
1382 s64 diff
= sample
- *avg
;
1388 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1390 #ifdef CONFIG_SCHEDSTATS
1391 struct rq
*rq
= this_rq();
1394 int this_cpu
= smp_processor_id();
1396 if (cpu
== this_cpu
) {
1397 schedstat_inc(rq
, ttwu_local
);
1398 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1400 struct sched_domain
*sd
;
1402 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1404 for_each_domain(this_cpu
, sd
) {
1405 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1406 schedstat_inc(sd
, ttwu_wake_remote
);
1413 if (wake_flags
& WF_MIGRATED
)
1414 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1416 #endif /* CONFIG_SMP */
1418 schedstat_inc(rq
, ttwu_count
);
1419 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1421 if (wake_flags
& WF_SYNC
)
1422 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1424 #endif /* CONFIG_SCHEDSTATS */
1427 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1429 activate_task(rq
, p
, en_flags
);
1430 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1432 /* if a worker is waking up, notify workqueue */
1433 if (p
->flags
& PF_WQ_WORKER
)
1434 wq_worker_waking_up(p
, cpu_of(rq
));
1438 * Mark the task runnable and perform wakeup-preemption.
1441 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1443 check_preempt_curr(rq
, p
, wake_flags
);
1444 trace_sched_wakeup(p
, true);
1446 p
->state
= TASK_RUNNING
;
1448 if (p
->sched_class
->task_woken
)
1449 p
->sched_class
->task_woken(rq
, p
);
1451 if (rq
->idle_stamp
) {
1452 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1453 u64 max
= 2*rq
->max_idle_balance_cost
;
1455 update_avg(&rq
->avg_idle
, delta
);
1457 if (rq
->avg_idle
> max
)
1466 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1469 if (p
->sched_contributes_to_load
)
1470 rq
->nr_uninterruptible
--;
1473 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1474 ttwu_do_wakeup(rq
, p
, wake_flags
);
1478 * Called in case the task @p isn't fully descheduled from its runqueue,
1479 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1480 * since all we need to do is flip p->state to TASK_RUNNING, since
1481 * the task is still ->on_rq.
1483 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1488 rq
= __task_rq_lock(p
);
1489 if (task_on_rq_queued(p
)) {
1490 /* check_preempt_curr() may use rq clock */
1491 update_rq_clock(rq
);
1492 ttwu_do_wakeup(rq
, p
, wake_flags
);
1495 __task_rq_unlock(rq
);
1501 void sched_ttwu_pending(void)
1503 struct rq
*rq
= this_rq();
1504 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1505 struct task_struct
*p
;
1506 unsigned long flags
;
1511 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1514 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1515 llist
= llist_next(llist
);
1516 ttwu_do_activate(rq
, p
, 0);
1519 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1522 void scheduler_ipi(void)
1525 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1526 * TIF_NEED_RESCHED remotely (for the first time) will also send
1529 preempt_fold_need_resched();
1531 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1535 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1536 * traditionally all their work was done from the interrupt return
1537 * path. Now that we actually do some work, we need to make sure
1540 * Some archs already do call them, luckily irq_enter/exit nest
1543 * Arguably we should visit all archs and update all handlers,
1544 * however a fair share of IPIs are still resched only so this would
1545 * somewhat pessimize the simple resched case.
1548 sched_ttwu_pending();
1551 * Check if someone kicked us for doing the nohz idle load balance.
1553 if (unlikely(got_nohz_idle_kick())) {
1554 this_rq()->idle_balance
= 1;
1555 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1560 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1562 struct rq
*rq
= cpu_rq(cpu
);
1564 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1565 if (!set_nr_if_polling(rq
->idle
))
1566 smp_send_reschedule(cpu
);
1568 trace_sched_wake_idle_without_ipi(cpu
);
1572 void wake_up_if_idle(int cpu
)
1574 struct rq
*rq
= cpu_rq(cpu
);
1575 unsigned long flags
;
1579 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1582 if (set_nr_if_polling(rq
->idle
)) {
1583 trace_sched_wake_idle_without_ipi(cpu
);
1585 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1586 if (is_idle_task(rq
->curr
))
1587 smp_send_reschedule(cpu
);
1588 /* Else cpu is not in idle, do nothing here */
1589 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1596 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1598 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1600 #endif /* CONFIG_SMP */
1602 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1604 struct rq
*rq
= cpu_rq(cpu
);
1606 #if defined(CONFIG_SMP)
1607 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1608 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1609 ttwu_queue_remote(p
, cpu
);
1614 raw_spin_lock(&rq
->lock
);
1615 ttwu_do_activate(rq
, p
, 0);
1616 raw_spin_unlock(&rq
->lock
);
1620 * try_to_wake_up - wake up a thread
1621 * @p: the thread to be awakened
1622 * @state: the mask of task states that can be woken
1623 * @wake_flags: wake modifier flags (WF_*)
1625 * Put it on the run-queue if it's not already there. The "current"
1626 * thread is always on the run-queue (except when the actual
1627 * re-schedule is in progress), and as such you're allowed to do
1628 * the simpler "current->state = TASK_RUNNING" to mark yourself
1629 * runnable without the overhead of this.
1631 * Return: %true if @p was woken up, %false if it was already running.
1632 * or @state didn't match @p's state.
1635 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1637 unsigned long flags
;
1638 int cpu
, success
= 0;
1641 * If we are going to wake up a thread waiting for CONDITION we
1642 * need to ensure that CONDITION=1 done by the caller can not be
1643 * reordered with p->state check below. This pairs with mb() in
1644 * set_current_state() the waiting thread does.
1646 smp_mb__before_spinlock();
1647 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1648 if (!(p
->state
& state
))
1651 success
= 1; /* we're going to change ->state */
1654 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1659 * If the owning (remote) cpu is still in the middle of schedule() with
1660 * this task as prev, wait until its done referencing the task.
1665 * Pairs with the smp_wmb() in finish_lock_switch().
1669 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1670 p
->state
= TASK_WAKING
;
1672 if (p
->sched_class
->task_waking
)
1673 p
->sched_class
->task_waking(p
);
1675 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1676 if (task_cpu(p
) != cpu
) {
1677 wake_flags
|= WF_MIGRATED
;
1678 set_task_cpu(p
, cpu
);
1680 #endif /* CONFIG_SMP */
1684 ttwu_stat(p
, cpu
, wake_flags
);
1686 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1692 * try_to_wake_up_local - try to wake up a local task with rq lock held
1693 * @p: the thread to be awakened
1695 * Put @p on the run-queue if it's not already there. The caller must
1696 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1699 static void try_to_wake_up_local(struct task_struct
*p
)
1701 struct rq
*rq
= task_rq(p
);
1703 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1704 WARN_ON_ONCE(p
== current
))
1707 lockdep_assert_held(&rq
->lock
);
1709 if (!raw_spin_trylock(&p
->pi_lock
)) {
1710 raw_spin_unlock(&rq
->lock
);
1711 raw_spin_lock(&p
->pi_lock
);
1712 raw_spin_lock(&rq
->lock
);
1715 if (!(p
->state
& TASK_NORMAL
))
1718 if (!task_on_rq_queued(p
))
1719 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1721 ttwu_do_wakeup(rq
, p
, 0);
1722 ttwu_stat(p
, smp_processor_id(), 0);
1724 raw_spin_unlock(&p
->pi_lock
);
1728 * wake_up_process - Wake up a specific process
1729 * @p: The process to be woken up.
1731 * Attempt to wake up the nominated process and move it to the set of runnable
1734 * Return: 1 if the process was woken up, 0 if it was already running.
1736 * It may be assumed that this function implies a write memory barrier before
1737 * changing the task state if and only if any tasks are woken up.
1739 int wake_up_process(struct task_struct
*p
)
1741 WARN_ON(task_is_stopped_or_traced(p
));
1742 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1744 EXPORT_SYMBOL(wake_up_process
);
1746 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1748 return try_to_wake_up(p
, state
, 0);
1752 * This function clears the sched_dl_entity static params.
1754 void __dl_clear_params(struct task_struct
*p
)
1756 struct sched_dl_entity
*dl_se
= &p
->dl
;
1758 dl_se
->dl_runtime
= 0;
1759 dl_se
->dl_deadline
= 0;
1760 dl_se
->dl_period
= 0;
1764 dl_se
->dl_throttled
= 0;
1766 dl_se
->dl_yielded
= 0;
1770 * Perform scheduler related setup for a newly forked process p.
1771 * p is forked by current.
1773 * __sched_fork() is basic setup used by init_idle() too:
1775 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1780 p
->se
.exec_start
= 0;
1781 p
->se
.sum_exec_runtime
= 0;
1782 p
->se
.prev_sum_exec_runtime
= 0;
1783 p
->se
.nr_migrations
= 0;
1786 p
->se
.avg
.decay_count
= 0;
1788 INIT_LIST_HEAD(&p
->se
.group_node
);
1790 #ifdef CONFIG_SCHEDSTATS
1791 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1794 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1795 init_dl_task_timer(&p
->dl
);
1796 __dl_clear_params(p
);
1798 INIT_LIST_HEAD(&p
->rt
.run_list
);
1800 #ifdef CONFIG_PREEMPT_NOTIFIERS
1801 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1804 #ifdef CONFIG_NUMA_BALANCING
1805 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1806 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1807 p
->mm
->numa_scan_seq
= 0;
1810 if (clone_flags
& CLONE_VM
)
1811 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1813 p
->numa_preferred_nid
= -1;
1815 p
->node_stamp
= 0ULL;
1816 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1817 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1818 p
->numa_work
.next
= &p
->numa_work
;
1819 p
->numa_faults
= NULL
;
1820 p
->last_task_numa_placement
= 0;
1821 p
->last_sum_exec_runtime
= 0;
1823 p
->numa_group
= NULL
;
1824 #endif /* CONFIG_NUMA_BALANCING */
1827 #ifdef CONFIG_NUMA_BALANCING
1828 #ifdef CONFIG_SCHED_DEBUG
1829 void set_numabalancing_state(bool enabled
)
1832 sched_feat_set("NUMA");
1834 sched_feat_set("NO_NUMA");
1837 __read_mostly
bool numabalancing_enabled
;
1839 void set_numabalancing_state(bool enabled
)
1841 numabalancing_enabled
= enabled
;
1843 #endif /* CONFIG_SCHED_DEBUG */
1845 #ifdef CONFIG_PROC_SYSCTL
1846 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1847 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1851 int state
= numabalancing_enabled
;
1853 if (write
&& !capable(CAP_SYS_ADMIN
))
1858 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1862 set_numabalancing_state(state
);
1869 * fork()/clone()-time setup:
1871 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1873 unsigned long flags
;
1874 int cpu
= get_cpu();
1876 __sched_fork(clone_flags
, p
);
1878 * We mark the process as running here. This guarantees that
1879 * nobody will actually run it, and a signal or other external
1880 * event cannot wake it up and insert it on the runqueue either.
1882 p
->state
= TASK_RUNNING
;
1885 * Make sure we do not leak PI boosting priority to the child.
1887 p
->prio
= current
->normal_prio
;
1890 * Revert to default priority/policy on fork if requested.
1892 if (unlikely(p
->sched_reset_on_fork
)) {
1893 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1894 p
->policy
= SCHED_NORMAL
;
1895 p
->static_prio
= NICE_TO_PRIO(0);
1897 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1898 p
->static_prio
= NICE_TO_PRIO(0);
1900 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1904 * We don't need the reset flag anymore after the fork. It has
1905 * fulfilled its duty:
1907 p
->sched_reset_on_fork
= 0;
1910 if (dl_prio(p
->prio
)) {
1913 } else if (rt_prio(p
->prio
)) {
1914 p
->sched_class
= &rt_sched_class
;
1916 p
->sched_class
= &fair_sched_class
;
1919 if (p
->sched_class
->task_fork
)
1920 p
->sched_class
->task_fork(p
);
1923 * The child is not yet in the pid-hash so no cgroup attach races,
1924 * and the cgroup is pinned to this child due to cgroup_fork()
1925 * is ran before sched_fork().
1927 * Silence PROVE_RCU.
1929 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1930 set_task_cpu(p
, cpu
);
1931 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1933 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1934 if (likely(sched_info_on()))
1935 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1937 #if defined(CONFIG_SMP)
1940 init_task_preempt_count(p
);
1942 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1943 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
1950 unsigned long to_ratio(u64 period
, u64 runtime
)
1952 if (runtime
== RUNTIME_INF
)
1956 * Doing this here saves a lot of checks in all
1957 * the calling paths, and returning zero seems
1958 * safe for them anyway.
1963 return div64_u64(runtime
<< 20, period
);
1967 inline struct dl_bw
*dl_bw_of(int i
)
1969 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1970 "sched RCU must be held");
1971 return &cpu_rq(i
)->rd
->dl_bw
;
1974 static inline int dl_bw_cpus(int i
)
1976 struct root_domain
*rd
= cpu_rq(i
)->rd
;
1979 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1980 "sched RCU must be held");
1981 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
1987 inline struct dl_bw
*dl_bw_of(int i
)
1989 return &cpu_rq(i
)->dl
.dl_bw
;
1992 static inline int dl_bw_cpus(int i
)
1999 * We must be sure that accepting a new task (or allowing changing the
2000 * parameters of an existing one) is consistent with the bandwidth
2001 * constraints. If yes, this function also accordingly updates the currently
2002 * allocated bandwidth to reflect the new situation.
2004 * This function is called while holding p's rq->lock.
2006 * XXX we should delay bw change until the task's 0-lag point, see
2009 static int dl_overflow(struct task_struct
*p
, int policy
,
2010 const struct sched_attr
*attr
)
2013 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2014 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2015 u64 runtime
= attr
->sched_runtime
;
2016 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2019 if (new_bw
== p
->dl
.dl_bw
)
2023 * Either if a task, enters, leave, or stays -deadline but changes
2024 * its parameters, we may need to update accordingly the total
2025 * allocated bandwidth of the container.
2027 raw_spin_lock(&dl_b
->lock
);
2028 cpus
= dl_bw_cpus(task_cpu(p
));
2029 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2030 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2031 __dl_add(dl_b
, new_bw
);
2033 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2034 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2035 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2036 __dl_add(dl_b
, new_bw
);
2038 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2039 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2042 raw_spin_unlock(&dl_b
->lock
);
2047 extern void init_dl_bw(struct dl_bw
*dl_b
);
2050 * wake_up_new_task - wake up a newly created task for the first time.
2052 * This function will do some initial scheduler statistics housekeeping
2053 * that must be done for every newly created context, then puts the task
2054 * on the runqueue and wakes it.
2056 void wake_up_new_task(struct task_struct
*p
)
2058 unsigned long flags
;
2061 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2064 * Fork balancing, do it here and not earlier because:
2065 * - cpus_allowed can change in the fork path
2066 * - any previously selected cpu might disappear through hotplug
2068 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2071 /* Initialize new task's runnable average */
2072 init_task_runnable_average(p
);
2073 rq
= __task_rq_lock(p
);
2074 activate_task(rq
, p
, 0);
2075 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2076 trace_sched_wakeup_new(p
, true);
2077 check_preempt_curr(rq
, p
, WF_FORK
);
2079 if (p
->sched_class
->task_woken
)
2080 p
->sched_class
->task_woken(rq
, p
);
2082 task_rq_unlock(rq
, p
, &flags
);
2085 #ifdef CONFIG_PREEMPT_NOTIFIERS
2088 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2089 * @notifier: notifier struct to register
2091 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2093 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2095 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2098 * preempt_notifier_unregister - no longer interested in preemption notifications
2099 * @notifier: notifier struct to unregister
2101 * This is safe to call from within a preemption notifier.
2103 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2105 hlist_del(¬ifier
->link
);
2107 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2109 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2111 struct preempt_notifier
*notifier
;
2113 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2114 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2118 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2119 struct task_struct
*next
)
2121 struct preempt_notifier
*notifier
;
2123 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2124 notifier
->ops
->sched_out(notifier
, next
);
2127 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2129 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2134 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2135 struct task_struct
*next
)
2139 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2142 * prepare_task_switch - prepare to switch tasks
2143 * @rq: the runqueue preparing to switch
2144 * @prev: the current task that is being switched out
2145 * @next: the task we are going to switch to.
2147 * This is called with the rq lock held and interrupts off. It must
2148 * be paired with a subsequent finish_task_switch after the context
2151 * prepare_task_switch sets up locking and calls architecture specific
2155 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2156 struct task_struct
*next
)
2158 trace_sched_switch(prev
, next
);
2159 sched_info_switch(rq
, prev
, next
);
2160 perf_event_task_sched_out(prev
, next
);
2161 fire_sched_out_preempt_notifiers(prev
, next
);
2162 prepare_lock_switch(rq
, next
);
2163 prepare_arch_switch(next
);
2167 * finish_task_switch - clean up after a task-switch
2168 * @prev: the thread we just switched away from.
2170 * finish_task_switch must be called after the context switch, paired
2171 * with a prepare_task_switch call before the context switch.
2172 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2173 * and do any other architecture-specific cleanup actions.
2175 * Note that we may have delayed dropping an mm in context_switch(). If
2176 * so, we finish that here outside of the runqueue lock. (Doing it
2177 * with the lock held can cause deadlocks; see schedule() for
2180 * The context switch have flipped the stack from under us and restored the
2181 * local variables which were saved when this task called schedule() in the
2182 * past. prev == current is still correct but we need to recalculate this_rq
2183 * because prev may have moved to another CPU.
2185 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2186 __releases(rq
->lock
)
2188 struct rq
*rq
= this_rq();
2189 struct mm_struct
*mm
= rq
->prev_mm
;
2195 * A task struct has one reference for the use as "current".
2196 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2197 * schedule one last time. The schedule call will never return, and
2198 * the scheduled task must drop that reference.
2199 * The test for TASK_DEAD must occur while the runqueue locks are
2200 * still held, otherwise prev could be scheduled on another cpu, die
2201 * there before we look at prev->state, and then the reference would
2203 * Manfred Spraul <manfred@colorfullife.com>
2205 prev_state
= prev
->state
;
2206 vtime_task_switch(prev
);
2207 finish_arch_switch(prev
);
2208 perf_event_task_sched_in(prev
, current
);
2209 finish_lock_switch(rq
, prev
);
2210 finish_arch_post_lock_switch();
2212 fire_sched_in_preempt_notifiers(current
);
2215 if (unlikely(prev_state
== TASK_DEAD
)) {
2216 if (prev
->sched_class
->task_dead
)
2217 prev
->sched_class
->task_dead(prev
);
2220 * Remove function-return probe instances associated with this
2221 * task and put them back on the free list.
2223 kprobe_flush_task(prev
);
2224 put_task_struct(prev
);
2227 tick_nohz_task_switch(current
);
2233 /* rq->lock is NOT held, but preemption is disabled */
2234 static inline void post_schedule(struct rq
*rq
)
2236 if (rq
->post_schedule
) {
2237 unsigned long flags
;
2239 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2240 if (rq
->curr
->sched_class
->post_schedule
)
2241 rq
->curr
->sched_class
->post_schedule(rq
);
2242 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2244 rq
->post_schedule
= 0;
2250 static inline void post_schedule(struct rq
*rq
)
2257 * schedule_tail - first thing a freshly forked thread must call.
2258 * @prev: the thread we just switched away from.
2260 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2261 __releases(rq
->lock
)
2265 /* finish_task_switch() drops rq->lock and enables preemtion */
2267 rq
= finish_task_switch(prev
);
2271 if (current
->set_child_tid
)
2272 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2276 * context_switch - switch to the new MM and the new thread's register state.
2278 static inline struct rq
*
2279 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2280 struct task_struct
*next
)
2282 struct mm_struct
*mm
, *oldmm
;
2284 prepare_task_switch(rq
, prev
, next
);
2287 oldmm
= prev
->active_mm
;
2289 * For paravirt, this is coupled with an exit in switch_to to
2290 * combine the page table reload and the switch backend into
2293 arch_start_context_switch(prev
);
2296 next
->active_mm
= oldmm
;
2297 atomic_inc(&oldmm
->mm_count
);
2298 enter_lazy_tlb(oldmm
, next
);
2300 switch_mm(oldmm
, mm
, next
);
2303 prev
->active_mm
= NULL
;
2304 rq
->prev_mm
= oldmm
;
2307 * Since the runqueue lock will be released by the next
2308 * task (which is an invalid locking op but in the case
2309 * of the scheduler it's an obvious special-case), so we
2310 * do an early lockdep release here:
2312 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2314 context_tracking_task_switch(prev
, next
);
2315 /* Here we just switch the register state and the stack. */
2316 switch_to(prev
, next
, prev
);
2319 return finish_task_switch(prev
);
2323 * nr_running and nr_context_switches:
2325 * externally visible scheduler statistics: current number of runnable
2326 * threads, total number of context switches performed since bootup.
2328 unsigned long nr_running(void)
2330 unsigned long i
, sum
= 0;
2332 for_each_online_cpu(i
)
2333 sum
+= cpu_rq(i
)->nr_running
;
2339 * Check if only the current task is running on the cpu.
2341 bool single_task_running(void)
2343 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2348 EXPORT_SYMBOL(single_task_running
);
2350 unsigned long long nr_context_switches(void)
2353 unsigned long long sum
= 0;
2355 for_each_possible_cpu(i
)
2356 sum
+= cpu_rq(i
)->nr_switches
;
2361 unsigned long nr_iowait(void)
2363 unsigned long i
, sum
= 0;
2365 for_each_possible_cpu(i
)
2366 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2371 unsigned long nr_iowait_cpu(int cpu
)
2373 struct rq
*this = cpu_rq(cpu
);
2374 return atomic_read(&this->nr_iowait
);
2377 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2379 struct rq
*this = this_rq();
2380 *nr_waiters
= atomic_read(&this->nr_iowait
);
2381 *load
= this->cpu_load
[0];
2387 * sched_exec - execve() is a valuable balancing opportunity, because at
2388 * this point the task has the smallest effective memory and cache footprint.
2390 void sched_exec(void)
2392 struct task_struct
*p
= current
;
2393 unsigned long flags
;
2396 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2397 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2398 if (dest_cpu
== smp_processor_id())
2401 if (likely(cpu_active(dest_cpu
))) {
2402 struct migration_arg arg
= { p
, dest_cpu
};
2404 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2405 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2409 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2414 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2415 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2417 EXPORT_PER_CPU_SYMBOL(kstat
);
2418 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2421 * Return accounted runtime for the task.
2422 * In case the task is currently running, return the runtime plus current's
2423 * pending runtime that have not been accounted yet.
2425 unsigned long long task_sched_runtime(struct task_struct
*p
)
2427 unsigned long flags
;
2431 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2433 * 64-bit doesn't need locks to atomically read a 64bit value.
2434 * So we have a optimization chance when the task's delta_exec is 0.
2435 * Reading ->on_cpu is racy, but this is ok.
2437 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2438 * If we race with it entering cpu, unaccounted time is 0. This is
2439 * indistinguishable from the read occurring a few cycles earlier.
2440 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2441 * been accounted, so we're correct here as well.
2443 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2444 return p
->se
.sum_exec_runtime
;
2447 rq
= task_rq_lock(p
, &flags
);
2449 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2450 * project cycles that may never be accounted to this
2451 * thread, breaking clock_gettime().
2453 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2454 update_rq_clock(rq
);
2455 p
->sched_class
->update_curr(rq
);
2457 ns
= p
->se
.sum_exec_runtime
;
2458 task_rq_unlock(rq
, p
, &flags
);
2464 * This function gets called by the timer code, with HZ frequency.
2465 * We call it with interrupts disabled.
2467 void scheduler_tick(void)
2469 int cpu
= smp_processor_id();
2470 struct rq
*rq
= cpu_rq(cpu
);
2471 struct task_struct
*curr
= rq
->curr
;
2475 raw_spin_lock(&rq
->lock
);
2476 update_rq_clock(rq
);
2477 curr
->sched_class
->task_tick(rq
, curr
, 0);
2478 update_cpu_load_active(rq
);
2479 raw_spin_unlock(&rq
->lock
);
2481 perf_event_task_tick();
2484 rq
->idle_balance
= idle_cpu(cpu
);
2485 trigger_load_balance(rq
);
2487 rq_last_tick_reset(rq
);
2490 #ifdef CONFIG_NO_HZ_FULL
2492 * scheduler_tick_max_deferment
2494 * Keep at least one tick per second when a single
2495 * active task is running because the scheduler doesn't
2496 * yet completely support full dynticks environment.
2498 * This makes sure that uptime, CFS vruntime, load
2499 * balancing, etc... continue to move forward, even
2500 * with a very low granularity.
2502 * Return: Maximum deferment in nanoseconds.
2504 u64
scheduler_tick_max_deferment(void)
2506 struct rq
*rq
= this_rq();
2507 unsigned long next
, now
= ACCESS_ONCE(jiffies
);
2509 next
= rq
->last_sched_tick
+ HZ
;
2511 if (time_before_eq(next
, now
))
2514 return jiffies_to_nsecs(next
- now
);
2518 notrace
unsigned long get_parent_ip(unsigned long addr
)
2520 if (in_lock_functions(addr
)) {
2521 addr
= CALLER_ADDR2
;
2522 if (in_lock_functions(addr
))
2523 addr
= CALLER_ADDR3
;
2528 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2529 defined(CONFIG_PREEMPT_TRACER))
2531 void preempt_count_add(int val
)
2533 #ifdef CONFIG_DEBUG_PREEMPT
2537 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2540 __preempt_count_add(val
);
2541 #ifdef CONFIG_DEBUG_PREEMPT
2543 * Spinlock count overflowing soon?
2545 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2548 if (preempt_count() == val
) {
2549 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2550 #ifdef CONFIG_DEBUG_PREEMPT
2551 current
->preempt_disable_ip
= ip
;
2553 trace_preempt_off(CALLER_ADDR0
, ip
);
2556 EXPORT_SYMBOL(preempt_count_add
);
2557 NOKPROBE_SYMBOL(preempt_count_add
);
2559 void preempt_count_sub(int val
)
2561 #ifdef CONFIG_DEBUG_PREEMPT
2565 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2568 * Is the spinlock portion underflowing?
2570 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2571 !(preempt_count() & PREEMPT_MASK
)))
2575 if (preempt_count() == val
)
2576 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2577 __preempt_count_sub(val
);
2579 EXPORT_SYMBOL(preempt_count_sub
);
2580 NOKPROBE_SYMBOL(preempt_count_sub
);
2585 * Print scheduling while atomic bug:
2587 static noinline
void __schedule_bug(struct task_struct
*prev
)
2589 if (oops_in_progress
)
2592 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2593 prev
->comm
, prev
->pid
, preempt_count());
2595 debug_show_held_locks(prev
);
2597 if (irqs_disabled())
2598 print_irqtrace_events(prev
);
2599 #ifdef CONFIG_DEBUG_PREEMPT
2600 if (in_atomic_preempt_off()) {
2601 pr_err("Preemption disabled at:");
2602 print_ip_sym(current
->preempt_disable_ip
);
2607 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2611 * Various schedule()-time debugging checks and statistics:
2613 static inline void schedule_debug(struct task_struct
*prev
)
2615 #ifdef CONFIG_SCHED_STACK_END_CHECK
2616 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2619 * Test if we are atomic. Since do_exit() needs to call into
2620 * schedule() atomically, we ignore that path. Otherwise whine
2621 * if we are scheduling when we should not.
2623 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2624 __schedule_bug(prev
);
2627 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2629 schedstat_inc(this_rq(), sched_count
);
2633 * Pick up the highest-prio task:
2635 static inline struct task_struct
*
2636 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2638 const struct sched_class
*class = &fair_sched_class
;
2639 struct task_struct
*p
;
2642 * Optimization: we know that if all tasks are in
2643 * the fair class we can call that function directly:
2645 if (likely(prev
->sched_class
== class &&
2646 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2647 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2648 if (unlikely(p
== RETRY_TASK
))
2651 /* assumes fair_sched_class->next == idle_sched_class */
2653 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2659 for_each_class(class) {
2660 p
= class->pick_next_task(rq
, prev
);
2662 if (unlikely(p
== RETRY_TASK
))
2668 BUG(); /* the idle class will always have a runnable task */
2672 * __schedule() is the main scheduler function.
2674 * The main means of driving the scheduler and thus entering this function are:
2676 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2678 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2679 * paths. For example, see arch/x86/entry_64.S.
2681 * To drive preemption between tasks, the scheduler sets the flag in timer
2682 * interrupt handler scheduler_tick().
2684 * 3. Wakeups don't really cause entry into schedule(). They add a
2685 * task to the run-queue and that's it.
2687 * Now, if the new task added to the run-queue preempts the current
2688 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2689 * called on the nearest possible occasion:
2691 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2693 * - in syscall or exception context, at the next outmost
2694 * preempt_enable(). (this might be as soon as the wake_up()'s
2697 * - in IRQ context, return from interrupt-handler to
2698 * preemptible context
2700 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2703 * - cond_resched() call
2704 * - explicit schedule() call
2705 * - return from syscall or exception to user-space
2706 * - return from interrupt-handler to user-space
2708 * WARNING: all callers must re-check need_resched() afterward and reschedule
2709 * accordingly in case an event triggered the need for rescheduling (such as
2710 * an interrupt waking up a task) while preemption was disabled in __schedule().
2712 static void __sched
__schedule(void)
2714 struct task_struct
*prev
, *next
;
2715 unsigned long *switch_count
;
2720 cpu
= smp_processor_id();
2722 rcu_note_context_switch();
2725 schedule_debug(prev
);
2727 if (sched_feat(HRTICK
))
2731 * Make sure that signal_pending_state()->signal_pending() below
2732 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2733 * done by the caller to avoid the race with signal_wake_up().
2735 smp_mb__before_spinlock();
2736 raw_spin_lock_irq(&rq
->lock
);
2738 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
2740 switch_count
= &prev
->nivcsw
;
2741 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2742 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2743 prev
->state
= TASK_RUNNING
;
2745 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2749 * If a worker went to sleep, notify and ask workqueue
2750 * whether it wants to wake up a task to maintain
2753 if (prev
->flags
& PF_WQ_WORKER
) {
2754 struct task_struct
*to_wakeup
;
2756 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2758 try_to_wake_up_local(to_wakeup
);
2761 switch_count
= &prev
->nvcsw
;
2764 if (task_on_rq_queued(prev
))
2765 update_rq_clock(rq
);
2767 next
= pick_next_task(rq
, prev
);
2768 clear_tsk_need_resched(prev
);
2769 clear_preempt_need_resched();
2770 rq
->clock_skip_update
= 0;
2772 if (likely(prev
!= next
)) {
2777 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
2780 raw_spin_unlock_irq(&rq
->lock
);
2784 sched_preempt_enable_no_resched();
2787 static inline void sched_submit_work(struct task_struct
*tsk
)
2789 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2792 * If we are going to sleep and we have plugged IO queued,
2793 * make sure to submit it to avoid deadlocks.
2795 if (blk_needs_flush_plug(tsk
))
2796 blk_schedule_flush_plug(tsk
);
2799 asmlinkage __visible
void __sched
schedule(void)
2801 struct task_struct
*tsk
= current
;
2803 sched_submit_work(tsk
);
2806 } while (need_resched());
2808 EXPORT_SYMBOL(schedule
);
2810 #ifdef CONFIG_CONTEXT_TRACKING
2811 asmlinkage __visible
void __sched
schedule_user(void)
2814 * If we come here after a random call to set_need_resched(),
2815 * or we have been woken up remotely but the IPI has not yet arrived,
2816 * we haven't yet exited the RCU idle mode. Do it here manually until
2817 * we find a better solution.
2819 * NB: There are buggy callers of this function. Ideally we
2820 * should warn if prev_state != CONTEXT_USER, but that will trigger
2821 * too frequently to make sense yet.
2823 enum ctx_state prev_state
= exception_enter();
2825 exception_exit(prev_state
);
2830 * schedule_preempt_disabled - called with preemption disabled
2832 * Returns with preemption disabled. Note: preempt_count must be 1
2834 void __sched
schedule_preempt_disabled(void)
2836 sched_preempt_enable_no_resched();
2841 static void __sched notrace
preempt_schedule_common(void)
2844 __preempt_count_add(PREEMPT_ACTIVE
);
2846 __preempt_count_sub(PREEMPT_ACTIVE
);
2849 * Check again in case we missed a preemption opportunity
2850 * between schedule and now.
2853 } while (need_resched());
2856 #ifdef CONFIG_PREEMPT
2858 * this is the entry point to schedule() from in-kernel preemption
2859 * off of preempt_enable. Kernel preemptions off return from interrupt
2860 * occur there and call schedule directly.
2862 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2865 * If there is a non-zero preempt_count or interrupts are disabled,
2866 * we do not want to preempt the current task. Just return..
2868 if (likely(!preemptible()))
2871 preempt_schedule_common();
2873 NOKPROBE_SYMBOL(preempt_schedule
);
2874 EXPORT_SYMBOL(preempt_schedule
);
2876 #ifdef CONFIG_CONTEXT_TRACKING
2878 * preempt_schedule_context - preempt_schedule called by tracing
2880 * The tracing infrastructure uses preempt_enable_notrace to prevent
2881 * recursion and tracing preempt enabling caused by the tracing
2882 * infrastructure itself. But as tracing can happen in areas coming
2883 * from userspace or just about to enter userspace, a preempt enable
2884 * can occur before user_exit() is called. This will cause the scheduler
2885 * to be called when the system is still in usermode.
2887 * To prevent this, the preempt_enable_notrace will use this function
2888 * instead of preempt_schedule() to exit user context if needed before
2889 * calling the scheduler.
2891 asmlinkage __visible
void __sched notrace
preempt_schedule_context(void)
2893 enum ctx_state prev_ctx
;
2895 if (likely(!preemptible()))
2899 __preempt_count_add(PREEMPT_ACTIVE
);
2901 * Needs preempt disabled in case user_exit() is traced
2902 * and the tracer calls preempt_enable_notrace() causing
2903 * an infinite recursion.
2905 prev_ctx
= exception_enter();
2907 exception_exit(prev_ctx
);
2909 __preempt_count_sub(PREEMPT_ACTIVE
);
2911 } while (need_resched());
2913 EXPORT_SYMBOL_GPL(preempt_schedule_context
);
2914 #endif /* CONFIG_CONTEXT_TRACKING */
2916 #endif /* CONFIG_PREEMPT */
2919 * this is the entry point to schedule() from kernel preemption
2920 * off of irq context.
2921 * Note, that this is called and return with irqs disabled. This will
2922 * protect us against recursive calling from irq.
2924 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2926 enum ctx_state prev_state
;
2928 /* Catch callers which need to be fixed */
2929 BUG_ON(preempt_count() || !irqs_disabled());
2931 prev_state
= exception_enter();
2934 __preempt_count_add(PREEMPT_ACTIVE
);
2937 local_irq_disable();
2938 __preempt_count_sub(PREEMPT_ACTIVE
);
2941 * Check again in case we missed a preemption opportunity
2942 * between schedule and now.
2945 } while (need_resched());
2947 exception_exit(prev_state
);
2950 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
2953 return try_to_wake_up(curr
->private, mode
, wake_flags
);
2955 EXPORT_SYMBOL(default_wake_function
);
2957 #ifdef CONFIG_RT_MUTEXES
2960 * rt_mutex_setprio - set the current priority of a task
2962 * @prio: prio value (kernel-internal form)
2964 * This function changes the 'effective' priority of a task. It does
2965 * not touch ->normal_prio like __setscheduler().
2967 * Used by the rt_mutex code to implement priority inheritance
2968 * logic. Call site only calls if the priority of the task changed.
2970 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
2972 int oldprio
, queued
, running
, enqueue_flag
= 0;
2974 const struct sched_class
*prev_class
;
2976 BUG_ON(prio
> MAX_PRIO
);
2978 rq
= __task_rq_lock(p
);
2981 * Idle task boosting is a nono in general. There is one
2982 * exception, when PREEMPT_RT and NOHZ is active:
2984 * The idle task calls get_next_timer_interrupt() and holds
2985 * the timer wheel base->lock on the CPU and another CPU wants
2986 * to access the timer (probably to cancel it). We can safely
2987 * ignore the boosting request, as the idle CPU runs this code
2988 * with interrupts disabled and will complete the lock
2989 * protected section without being interrupted. So there is no
2990 * real need to boost.
2992 if (unlikely(p
== rq
->idle
)) {
2993 WARN_ON(p
!= rq
->curr
);
2994 WARN_ON(p
->pi_blocked_on
);
2998 trace_sched_pi_setprio(p
, prio
);
3000 prev_class
= p
->sched_class
;
3001 queued
= task_on_rq_queued(p
);
3002 running
= task_current(rq
, p
);
3004 dequeue_task(rq
, p
, 0);
3006 put_prev_task(rq
, p
);
3009 * Boosting condition are:
3010 * 1. -rt task is running and holds mutex A
3011 * --> -dl task blocks on mutex A
3013 * 2. -dl task is running and holds mutex A
3014 * --> -dl task blocks on mutex A and could preempt the
3017 if (dl_prio(prio
)) {
3018 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3019 if (!dl_prio(p
->normal_prio
) ||
3020 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3021 p
->dl
.dl_boosted
= 1;
3022 p
->dl
.dl_throttled
= 0;
3023 enqueue_flag
= ENQUEUE_REPLENISH
;
3025 p
->dl
.dl_boosted
= 0;
3026 p
->sched_class
= &dl_sched_class
;
3027 } else if (rt_prio(prio
)) {
3028 if (dl_prio(oldprio
))
3029 p
->dl
.dl_boosted
= 0;
3031 enqueue_flag
= ENQUEUE_HEAD
;
3032 p
->sched_class
= &rt_sched_class
;
3034 if (dl_prio(oldprio
))
3035 p
->dl
.dl_boosted
= 0;
3036 if (rt_prio(oldprio
))
3038 p
->sched_class
= &fair_sched_class
;
3044 p
->sched_class
->set_curr_task(rq
);
3046 enqueue_task(rq
, p
, enqueue_flag
);
3048 check_class_changed(rq
, p
, prev_class
, oldprio
);
3050 __task_rq_unlock(rq
);
3054 void set_user_nice(struct task_struct
*p
, long nice
)
3056 int old_prio
, delta
, queued
;
3057 unsigned long flags
;
3060 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3063 * We have to be careful, if called from sys_setpriority(),
3064 * the task might be in the middle of scheduling on another CPU.
3066 rq
= task_rq_lock(p
, &flags
);
3068 * The RT priorities are set via sched_setscheduler(), but we still
3069 * allow the 'normal' nice value to be set - but as expected
3070 * it wont have any effect on scheduling until the task is
3071 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3073 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3074 p
->static_prio
= NICE_TO_PRIO(nice
);
3077 queued
= task_on_rq_queued(p
);
3079 dequeue_task(rq
, p
, 0);
3081 p
->static_prio
= NICE_TO_PRIO(nice
);
3084 p
->prio
= effective_prio(p
);
3085 delta
= p
->prio
- old_prio
;
3088 enqueue_task(rq
, p
, 0);
3090 * If the task increased its priority or is running and
3091 * lowered its priority, then reschedule its CPU:
3093 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3097 task_rq_unlock(rq
, p
, &flags
);
3099 EXPORT_SYMBOL(set_user_nice
);
3102 * can_nice - check if a task can reduce its nice value
3106 int can_nice(const struct task_struct
*p
, const int nice
)
3108 /* convert nice value [19,-20] to rlimit style value [1,40] */
3109 int nice_rlim
= nice_to_rlimit(nice
);
3111 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3112 capable(CAP_SYS_NICE
));
3115 #ifdef __ARCH_WANT_SYS_NICE
3118 * sys_nice - change the priority of the current process.
3119 * @increment: priority increment
3121 * sys_setpriority is a more generic, but much slower function that
3122 * does similar things.
3124 SYSCALL_DEFINE1(nice
, int, increment
)
3129 * Setpriority might change our priority at the same moment.
3130 * We don't have to worry. Conceptually one call occurs first
3131 * and we have a single winner.
3133 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3134 nice
= task_nice(current
) + increment
;
3136 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3137 if (increment
< 0 && !can_nice(current
, nice
))
3140 retval
= security_task_setnice(current
, nice
);
3144 set_user_nice(current
, nice
);
3151 * task_prio - return the priority value of a given task.
3152 * @p: the task in question.
3154 * Return: The priority value as seen by users in /proc.
3155 * RT tasks are offset by -200. Normal tasks are centered
3156 * around 0, value goes from -16 to +15.
3158 int task_prio(const struct task_struct
*p
)
3160 return p
->prio
- MAX_RT_PRIO
;
3164 * idle_cpu - is a given cpu idle currently?
3165 * @cpu: the processor in question.
3167 * Return: 1 if the CPU is currently idle. 0 otherwise.
3169 int idle_cpu(int cpu
)
3171 struct rq
*rq
= cpu_rq(cpu
);
3173 if (rq
->curr
!= rq
->idle
)
3180 if (!llist_empty(&rq
->wake_list
))
3188 * idle_task - return the idle task for a given cpu.
3189 * @cpu: the processor in question.
3191 * Return: The idle task for the cpu @cpu.
3193 struct task_struct
*idle_task(int cpu
)
3195 return cpu_rq(cpu
)->idle
;
3199 * find_process_by_pid - find a process with a matching PID value.
3200 * @pid: the pid in question.
3202 * The task of @pid, if found. %NULL otherwise.
3204 static struct task_struct
*find_process_by_pid(pid_t pid
)
3206 return pid
? find_task_by_vpid(pid
) : current
;
3210 * This function initializes the sched_dl_entity of a newly becoming
3211 * SCHED_DEADLINE task.
3213 * Only the static values are considered here, the actual runtime and the
3214 * absolute deadline will be properly calculated when the task is enqueued
3215 * for the first time with its new policy.
3218 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3220 struct sched_dl_entity
*dl_se
= &p
->dl
;
3222 dl_se
->dl_runtime
= attr
->sched_runtime
;
3223 dl_se
->dl_deadline
= attr
->sched_deadline
;
3224 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3225 dl_se
->flags
= attr
->sched_flags
;
3226 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3229 * Changing the parameters of a task is 'tricky' and we're not doing
3230 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3232 * What we SHOULD do is delay the bandwidth release until the 0-lag
3233 * point. This would include retaining the task_struct until that time
3234 * and change dl_overflow() to not immediately decrement the current
3237 * Instead we retain the current runtime/deadline and let the new
3238 * parameters take effect after the current reservation period lapses.
3239 * This is safe (albeit pessimistic) because the 0-lag point is always
3240 * before the current scheduling deadline.
3242 * We can still have temporary overloads because we do not delay the
3243 * change in bandwidth until that time; so admission control is
3244 * not on the safe side. It does however guarantee tasks will never
3245 * consume more than promised.
3250 * sched_setparam() passes in -1 for its policy, to let the functions
3251 * it calls know not to change it.
3253 #define SETPARAM_POLICY -1
3255 static void __setscheduler_params(struct task_struct
*p
,
3256 const struct sched_attr
*attr
)
3258 int policy
= attr
->sched_policy
;
3260 if (policy
== SETPARAM_POLICY
)
3265 if (dl_policy(policy
))
3266 __setparam_dl(p
, attr
);
3267 else if (fair_policy(policy
))
3268 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3271 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3272 * !rt_policy. Always setting this ensures that things like
3273 * getparam()/getattr() don't report silly values for !rt tasks.
3275 p
->rt_priority
= attr
->sched_priority
;
3276 p
->normal_prio
= normal_prio(p
);
3280 /* Actually do priority change: must hold pi & rq lock. */
3281 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3282 const struct sched_attr
*attr
, bool keep_boost
)
3284 __setscheduler_params(p
, attr
);
3287 * Keep a potential priority boosting if called from
3288 * sched_setscheduler().
3291 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3293 p
->prio
= normal_prio(p
);
3295 if (dl_prio(p
->prio
))
3296 p
->sched_class
= &dl_sched_class
;
3297 else if (rt_prio(p
->prio
))
3298 p
->sched_class
= &rt_sched_class
;
3300 p
->sched_class
= &fair_sched_class
;
3304 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3306 struct sched_dl_entity
*dl_se
= &p
->dl
;
3308 attr
->sched_priority
= p
->rt_priority
;
3309 attr
->sched_runtime
= dl_se
->dl_runtime
;
3310 attr
->sched_deadline
= dl_se
->dl_deadline
;
3311 attr
->sched_period
= dl_se
->dl_period
;
3312 attr
->sched_flags
= dl_se
->flags
;
3316 * This function validates the new parameters of a -deadline task.
3317 * We ask for the deadline not being zero, and greater or equal
3318 * than the runtime, as well as the period of being zero or
3319 * greater than deadline. Furthermore, we have to be sure that
3320 * user parameters are above the internal resolution of 1us (we
3321 * check sched_runtime only since it is always the smaller one) and
3322 * below 2^63 ns (we have to check both sched_deadline and
3323 * sched_period, as the latter can be zero).
3326 __checkparam_dl(const struct sched_attr
*attr
)
3329 if (attr
->sched_deadline
== 0)
3333 * Since we truncate DL_SCALE bits, make sure we're at least
3336 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3340 * Since we use the MSB for wrap-around and sign issues, make
3341 * sure it's not set (mind that period can be equal to zero).
3343 if (attr
->sched_deadline
& (1ULL << 63) ||
3344 attr
->sched_period
& (1ULL << 63))
3347 /* runtime <= deadline <= period (if period != 0) */
3348 if ((attr
->sched_period
!= 0 &&
3349 attr
->sched_period
< attr
->sched_deadline
) ||
3350 attr
->sched_deadline
< attr
->sched_runtime
)
3357 * check the target process has a UID that matches the current process's
3359 static bool check_same_owner(struct task_struct
*p
)
3361 const struct cred
*cred
= current_cred(), *pcred
;
3365 pcred
= __task_cred(p
);
3366 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3367 uid_eq(cred
->euid
, pcred
->uid
));
3372 static bool dl_param_changed(struct task_struct
*p
,
3373 const struct sched_attr
*attr
)
3375 struct sched_dl_entity
*dl_se
= &p
->dl
;
3377 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3378 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3379 dl_se
->dl_period
!= attr
->sched_period
||
3380 dl_se
->flags
!= attr
->sched_flags
)
3386 static int __sched_setscheduler(struct task_struct
*p
,
3387 const struct sched_attr
*attr
,
3390 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3391 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3392 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3393 int new_effective_prio
, policy
= attr
->sched_policy
;
3394 unsigned long flags
;
3395 const struct sched_class
*prev_class
;
3399 /* may grab non-irq protected spin_locks */
3400 BUG_ON(in_interrupt());
3402 /* double check policy once rq lock held */
3404 reset_on_fork
= p
->sched_reset_on_fork
;
3405 policy
= oldpolicy
= p
->policy
;
3407 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3409 if (policy
!= SCHED_DEADLINE
&&
3410 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3411 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3412 policy
!= SCHED_IDLE
)
3416 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3420 * Valid priorities for SCHED_FIFO and SCHED_RR are
3421 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3422 * SCHED_BATCH and SCHED_IDLE is 0.
3424 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3425 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3427 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3428 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3432 * Allow unprivileged RT tasks to decrease priority:
3434 if (user
&& !capable(CAP_SYS_NICE
)) {
3435 if (fair_policy(policy
)) {
3436 if (attr
->sched_nice
< task_nice(p
) &&
3437 !can_nice(p
, attr
->sched_nice
))
3441 if (rt_policy(policy
)) {
3442 unsigned long rlim_rtprio
=
3443 task_rlimit(p
, RLIMIT_RTPRIO
);
3445 /* can't set/change the rt policy */
3446 if (policy
!= p
->policy
&& !rlim_rtprio
)
3449 /* can't increase priority */
3450 if (attr
->sched_priority
> p
->rt_priority
&&
3451 attr
->sched_priority
> rlim_rtprio
)
3456 * Can't set/change SCHED_DEADLINE policy at all for now
3457 * (safest behavior); in the future we would like to allow
3458 * unprivileged DL tasks to increase their relative deadline
3459 * or reduce their runtime (both ways reducing utilization)
3461 if (dl_policy(policy
))
3465 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3466 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3468 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3469 if (!can_nice(p
, task_nice(p
)))
3473 /* can't change other user's priorities */
3474 if (!check_same_owner(p
))
3477 /* Normal users shall not reset the sched_reset_on_fork flag */
3478 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3483 retval
= security_task_setscheduler(p
);
3489 * make sure no PI-waiters arrive (or leave) while we are
3490 * changing the priority of the task:
3492 * To be able to change p->policy safely, the appropriate
3493 * runqueue lock must be held.
3495 rq
= task_rq_lock(p
, &flags
);
3498 * Changing the policy of the stop threads its a very bad idea
3500 if (p
== rq
->stop
) {
3501 task_rq_unlock(rq
, p
, &flags
);
3506 * If not changing anything there's no need to proceed further,
3507 * but store a possible modification of reset_on_fork.
3509 if (unlikely(policy
== p
->policy
)) {
3510 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3512 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3514 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3517 p
->sched_reset_on_fork
= reset_on_fork
;
3518 task_rq_unlock(rq
, p
, &flags
);
3524 #ifdef CONFIG_RT_GROUP_SCHED
3526 * Do not allow realtime tasks into groups that have no runtime
3529 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3530 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3531 !task_group_is_autogroup(task_group(p
))) {
3532 task_rq_unlock(rq
, p
, &flags
);
3537 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3538 cpumask_t
*span
= rq
->rd
->span
;
3541 * Don't allow tasks with an affinity mask smaller than
3542 * the entire root_domain to become SCHED_DEADLINE. We
3543 * will also fail if there's no bandwidth available.
3545 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3546 rq
->rd
->dl_bw
.bw
== 0) {
3547 task_rq_unlock(rq
, p
, &flags
);
3554 /* recheck policy now with rq lock held */
3555 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3556 policy
= oldpolicy
= -1;
3557 task_rq_unlock(rq
, p
, &flags
);
3562 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3563 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3566 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3567 task_rq_unlock(rq
, p
, &flags
);
3571 p
->sched_reset_on_fork
= reset_on_fork
;
3575 * Take priority boosted tasks into account. If the new
3576 * effective priority is unchanged, we just store the new
3577 * normal parameters and do not touch the scheduler class and
3578 * the runqueue. This will be done when the task deboost
3581 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
3582 if (new_effective_prio
== oldprio
) {
3583 __setscheduler_params(p
, attr
);
3584 task_rq_unlock(rq
, p
, &flags
);
3588 queued
= task_on_rq_queued(p
);
3589 running
= task_current(rq
, p
);
3591 dequeue_task(rq
, p
, 0);
3593 put_prev_task(rq
, p
);
3595 prev_class
= p
->sched_class
;
3596 __setscheduler(rq
, p
, attr
, true);
3599 p
->sched_class
->set_curr_task(rq
);
3602 * We enqueue to tail when the priority of a task is
3603 * increased (user space view).
3605 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3608 check_class_changed(rq
, p
, prev_class
, oldprio
);
3609 task_rq_unlock(rq
, p
, &flags
);
3611 rt_mutex_adjust_pi(p
);
3616 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3617 const struct sched_param
*param
, bool check
)
3619 struct sched_attr attr
= {
3620 .sched_policy
= policy
,
3621 .sched_priority
= param
->sched_priority
,
3622 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3625 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3626 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3627 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3628 policy
&= ~SCHED_RESET_ON_FORK
;
3629 attr
.sched_policy
= policy
;
3632 return __sched_setscheduler(p
, &attr
, check
);
3635 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3636 * @p: the task in question.
3637 * @policy: new policy.
3638 * @param: structure containing the new RT priority.
3640 * Return: 0 on success. An error code otherwise.
3642 * NOTE that the task may be already dead.
3644 int sched_setscheduler(struct task_struct
*p
, int policy
,
3645 const struct sched_param
*param
)
3647 return _sched_setscheduler(p
, policy
, param
, true);
3649 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3651 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3653 return __sched_setscheduler(p
, attr
, true);
3655 EXPORT_SYMBOL_GPL(sched_setattr
);
3658 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3659 * @p: the task in question.
3660 * @policy: new policy.
3661 * @param: structure containing the new RT priority.
3663 * Just like sched_setscheduler, only don't bother checking if the
3664 * current context has permission. For example, this is needed in
3665 * stop_machine(): we create temporary high priority worker threads,
3666 * but our caller might not have that capability.
3668 * Return: 0 on success. An error code otherwise.
3670 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3671 const struct sched_param
*param
)
3673 return _sched_setscheduler(p
, policy
, param
, false);
3677 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3679 struct sched_param lparam
;
3680 struct task_struct
*p
;
3683 if (!param
|| pid
< 0)
3685 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3690 p
= find_process_by_pid(pid
);
3692 retval
= sched_setscheduler(p
, policy
, &lparam
);
3699 * Mimics kernel/events/core.c perf_copy_attr().
3701 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3702 struct sched_attr
*attr
)
3707 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3711 * zero the full structure, so that a short copy will be nice.
3713 memset(attr
, 0, sizeof(*attr
));
3715 ret
= get_user(size
, &uattr
->size
);
3719 if (size
> PAGE_SIZE
) /* silly large */
3722 if (!size
) /* abi compat */
3723 size
= SCHED_ATTR_SIZE_VER0
;
3725 if (size
< SCHED_ATTR_SIZE_VER0
)
3729 * If we're handed a bigger struct than we know of,
3730 * ensure all the unknown bits are 0 - i.e. new
3731 * user-space does not rely on any kernel feature
3732 * extensions we dont know about yet.
3734 if (size
> sizeof(*attr
)) {
3735 unsigned char __user
*addr
;
3736 unsigned char __user
*end
;
3739 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3740 end
= (void __user
*)uattr
+ size
;
3742 for (; addr
< end
; addr
++) {
3743 ret
= get_user(val
, addr
);
3749 size
= sizeof(*attr
);
3752 ret
= copy_from_user(attr
, uattr
, size
);
3757 * XXX: do we want to be lenient like existing syscalls; or do we want
3758 * to be strict and return an error on out-of-bounds values?
3760 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3765 put_user(sizeof(*attr
), &uattr
->size
);
3770 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3771 * @pid: the pid in question.
3772 * @policy: new policy.
3773 * @param: structure containing the new RT priority.
3775 * Return: 0 on success. An error code otherwise.
3777 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3778 struct sched_param __user
*, param
)
3780 /* negative values for policy are not valid */
3784 return do_sched_setscheduler(pid
, policy
, param
);
3788 * sys_sched_setparam - set/change the RT priority of a thread
3789 * @pid: the pid in question.
3790 * @param: structure containing the new RT priority.
3792 * Return: 0 on success. An error code otherwise.
3794 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3796 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
3800 * sys_sched_setattr - same as above, but with extended sched_attr
3801 * @pid: the pid in question.
3802 * @uattr: structure containing the extended parameters.
3803 * @flags: for future extension.
3805 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3806 unsigned int, flags
)
3808 struct sched_attr attr
;
3809 struct task_struct
*p
;
3812 if (!uattr
|| pid
< 0 || flags
)
3815 retval
= sched_copy_attr(uattr
, &attr
);
3819 if ((int)attr
.sched_policy
< 0)
3824 p
= find_process_by_pid(pid
);
3826 retval
= sched_setattr(p
, &attr
);
3833 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3834 * @pid: the pid in question.
3836 * Return: On success, the policy of the thread. Otherwise, a negative error
3839 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3841 struct task_struct
*p
;
3849 p
= find_process_by_pid(pid
);
3851 retval
= security_task_getscheduler(p
);
3854 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3861 * sys_sched_getparam - get the RT priority of a thread
3862 * @pid: the pid in question.
3863 * @param: structure containing the RT priority.
3865 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3868 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3870 struct sched_param lp
= { .sched_priority
= 0 };
3871 struct task_struct
*p
;
3874 if (!param
|| pid
< 0)
3878 p
= find_process_by_pid(pid
);
3883 retval
= security_task_getscheduler(p
);
3887 if (task_has_rt_policy(p
))
3888 lp
.sched_priority
= p
->rt_priority
;
3892 * This one might sleep, we cannot do it with a spinlock held ...
3894 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3903 static int sched_read_attr(struct sched_attr __user
*uattr
,
3904 struct sched_attr
*attr
,
3909 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3913 * If we're handed a smaller struct than we know of,
3914 * ensure all the unknown bits are 0 - i.e. old
3915 * user-space does not get uncomplete information.
3917 if (usize
< sizeof(*attr
)) {
3918 unsigned char *addr
;
3921 addr
= (void *)attr
+ usize
;
3922 end
= (void *)attr
+ sizeof(*attr
);
3924 for (; addr
< end
; addr
++) {
3932 ret
= copy_to_user(uattr
, attr
, attr
->size
);
3940 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3941 * @pid: the pid in question.
3942 * @uattr: structure containing the extended parameters.
3943 * @size: sizeof(attr) for fwd/bwd comp.
3944 * @flags: for future extension.
3946 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3947 unsigned int, size
, unsigned int, flags
)
3949 struct sched_attr attr
= {
3950 .size
= sizeof(struct sched_attr
),
3952 struct task_struct
*p
;
3955 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
3956 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
3960 p
= find_process_by_pid(pid
);
3965 retval
= security_task_getscheduler(p
);
3969 attr
.sched_policy
= p
->policy
;
3970 if (p
->sched_reset_on_fork
)
3971 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3972 if (task_has_dl_policy(p
))
3973 __getparam_dl(p
, &attr
);
3974 else if (task_has_rt_policy(p
))
3975 attr
.sched_priority
= p
->rt_priority
;
3977 attr
.sched_nice
= task_nice(p
);
3981 retval
= sched_read_attr(uattr
, &attr
, size
);
3989 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
3991 cpumask_var_t cpus_allowed
, new_mask
;
3992 struct task_struct
*p
;
3997 p
= find_process_by_pid(pid
);
4003 /* Prevent p going away */
4007 if (p
->flags
& PF_NO_SETAFFINITY
) {
4011 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4015 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4017 goto out_free_cpus_allowed
;
4020 if (!check_same_owner(p
)) {
4022 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4024 goto out_free_new_mask
;
4029 retval
= security_task_setscheduler(p
);
4031 goto out_free_new_mask
;
4034 cpuset_cpus_allowed(p
, cpus_allowed
);
4035 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4038 * Since bandwidth control happens on root_domain basis,
4039 * if admission test is enabled, we only admit -deadline
4040 * tasks allowed to run on all the CPUs in the task's
4044 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4046 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4049 goto out_free_new_mask
;
4055 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4058 cpuset_cpus_allowed(p
, cpus_allowed
);
4059 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4061 * We must have raced with a concurrent cpuset
4062 * update. Just reset the cpus_allowed to the
4063 * cpuset's cpus_allowed
4065 cpumask_copy(new_mask
, cpus_allowed
);
4070 free_cpumask_var(new_mask
);
4071 out_free_cpus_allowed
:
4072 free_cpumask_var(cpus_allowed
);
4078 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4079 struct cpumask
*new_mask
)
4081 if (len
< cpumask_size())
4082 cpumask_clear(new_mask
);
4083 else if (len
> cpumask_size())
4084 len
= cpumask_size();
4086 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4090 * sys_sched_setaffinity - set the cpu affinity of a process
4091 * @pid: pid of the process
4092 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4093 * @user_mask_ptr: user-space pointer to the new cpu mask
4095 * Return: 0 on success. An error code otherwise.
4097 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4098 unsigned long __user
*, user_mask_ptr
)
4100 cpumask_var_t new_mask
;
4103 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4106 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4108 retval
= sched_setaffinity(pid
, new_mask
);
4109 free_cpumask_var(new_mask
);
4113 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4115 struct task_struct
*p
;
4116 unsigned long flags
;
4122 p
= find_process_by_pid(pid
);
4126 retval
= security_task_getscheduler(p
);
4130 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4131 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4132 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4141 * sys_sched_getaffinity - get the cpu affinity of a process
4142 * @pid: pid of the process
4143 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4144 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4146 * Return: 0 on success. An error code otherwise.
4148 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4149 unsigned long __user
*, user_mask_ptr
)
4154 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4156 if (len
& (sizeof(unsigned long)-1))
4159 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4162 ret
= sched_getaffinity(pid
, mask
);
4164 size_t retlen
= min_t(size_t, len
, cpumask_size());
4166 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4171 free_cpumask_var(mask
);
4177 * sys_sched_yield - yield the current processor to other threads.
4179 * This function yields the current CPU to other tasks. If there are no
4180 * other threads running on this CPU then this function will return.
4184 SYSCALL_DEFINE0(sched_yield
)
4186 struct rq
*rq
= this_rq_lock();
4188 schedstat_inc(rq
, yld_count
);
4189 current
->sched_class
->yield_task(rq
);
4192 * Since we are going to call schedule() anyway, there's
4193 * no need to preempt or enable interrupts:
4195 __release(rq
->lock
);
4196 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4197 do_raw_spin_unlock(&rq
->lock
);
4198 sched_preempt_enable_no_resched();
4205 int __sched
_cond_resched(void)
4207 if (should_resched()) {
4208 preempt_schedule_common();
4213 EXPORT_SYMBOL(_cond_resched
);
4216 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4217 * call schedule, and on return reacquire the lock.
4219 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4220 * operations here to prevent schedule() from being called twice (once via
4221 * spin_unlock(), once by hand).
4223 int __cond_resched_lock(spinlock_t
*lock
)
4225 int resched
= should_resched();
4228 lockdep_assert_held(lock
);
4230 if (spin_needbreak(lock
) || resched
) {
4233 preempt_schedule_common();
4241 EXPORT_SYMBOL(__cond_resched_lock
);
4243 int __sched
__cond_resched_softirq(void)
4245 BUG_ON(!in_softirq());
4247 if (should_resched()) {
4249 preempt_schedule_common();
4255 EXPORT_SYMBOL(__cond_resched_softirq
);
4258 * yield - yield the current processor to other threads.
4260 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4262 * The scheduler is at all times free to pick the calling task as the most
4263 * eligible task to run, if removing the yield() call from your code breaks
4264 * it, its already broken.
4266 * Typical broken usage is:
4271 * where one assumes that yield() will let 'the other' process run that will
4272 * make event true. If the current task is a SCHED_FIFO task that will never
4273 * happen. Never use yield() as a progress guarantee!!
4275 * If you want to use yield() to wait for something, use wait_event().
4276 * If you want to use yield() to be 'nice' for others, use cond_resched().
4277 * If you still want to use yield(), do not!
4279 void __sched
yield(void)
4281 set_current_state(TASK_RUNNING
);
4284 EXPORT_SYMBOL(yield
);
4287 * yield_to - yield the current processor to another thread in
4288 * your thread group, or accelerate that thread toward the
4289 * processor it's on.
4291 * @preempt: whether task preemption is allowed or not
4293 * It's the caller's job to ensure that the target task struct
4294 * can't go away on us before we can do any checks.
4297 * true (>0) if we indeed boosted the target task.
4298 * false (0) if we failed to boost the target.
4299 * -ESRCH if there's no task to yield to.
4301 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4303 struct task_struct
*curr
= current
;
4304 struct rq
*rq
, *p_rq
;
4305 unsigned long flags
;
4308 local_irq_save(flags
);
4314 * If we're the only runnable task on the rq and target rq also
4315 * has only one task, there's absolutely no point in yielding.
4317 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4322 double_rq_lock(rq
, p_rq
);
4323 if (task_rq(p
) != p_rq
) {
4324 double_rq_unlock(rq
, p_rq
);
4328 if (!curr
->sched_class
->yield_to_task
)
4331 if (curr
->sched_class
!= p
->sched_class
)
4334 if (task_running(p_rq
, p
) || p
->state
)
4337 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4339 schedstat_inc(rq
, yld_count
);
4341 * Make p's CPU reschedule; pick_next_entity takes care of
4344 if (preempt
&& rq
!= p_rq
)
4349 double_rq_unlock(rq
, p_rq
);
4351 local_irq_restore(flags
);
4358 EXPORT_SYMBOL_GPL(yield_to
);
4361 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4362 * that process accounting knows that this is a task in IO wait state.
4364 long __sched
io_schedule_timeout(long timeout
)
4366 int old_iowait
= current
->in_iowait
;
4370 current
->in_iowait
= 1;
4372 blk_schedule_flush_plug(current
);
4374 blk_flush_plug(current
);
4376 delayacct_blkio_start();
4378 atomic_inc(&rq
->nr_iowait
);
4379 ret
= schedule_timeout(timeout
);
4380 current
->in_iowait
= old_iowait
;
4381 atomic_dec(&rq
->nr_iowait
);
4382 delayacct_blkio_end();
4386 EXPORT_SYMBOL(io_schedule_timeout
);
4389 * sys_sched_get_priority_max - return maximum RT priority.
4390 * @policy: scheduling class.
4392 * Return: On success, this syscall returns the maximum
4393 * rt_priority that can be used by a given scheduling class.
4394 * On failure, a negative error code is returned.
4396 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4403 ret
= MAX_USER_RT_PRIO
-1;
4405 case SCHED_DEADLINE
:
4416 * sys_sched_get_priority_min - return minimum RT priority.
4417 * @policy: scheduling class.
4419 * Return: On success, this syscall returns the minimum
4420 * rt_priority that can be used by a given scheduling class.
4421 * On failure, a negative error code is returned.
4423 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4432 case SCHED_DEADLINE
:
4442 * sys_sched_rr_get_interval - return the default timeslice of a process.
4443 * @pid: pid of the process.
4444 * @interval: userspace pointer to the timeslice value.
4446 * this syscall writes the default timeslice value of a given process
4447 * into the user-space timespec buffer. A value of '0' means infinity.
4449 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4452 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4453 struct timespec __user
*, interval
)
4455 struct task_struct
*p
;
4456 unsigned int time_slice
;
4457 unsigned long flags
;
4467 p
= find_process_by_pid(pid
);
4471 retval
= security_task_getscheduler(p
);
4475 rq
= task_rq_lock(p
, &flags
);
4477 if (p
->sched_class
->get_rr_interval
)
4478 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4479 task_rq_unlock(rq
, p
, &flags
);
4482 jiffies_to_timespec(time_slice
, &t
);
4483 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4491 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4493 void sched_show_task(struct task_struct
*p
)
4495 unsigned long free
= 0;
4497 unsigned long state
= p
->state
;
4500 state
= __ffs(state
) + 1;
4501 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4502 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4503 #if BITS_PER_LONG == 32
4504 if (state
== TASK_RUNNING
)
4505 printk(KERN_CONT
" running ");
4507 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4509 if (state
== TASK_RUNNING
)
4510 printk(KERN_CONT
" running task ");
4512 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4514 #ifdef CONFIG_DEBUG_STACK_USAGE
4515 free
= stack_not_used(p
);
4520 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4522 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4523 task_pid_nr(p
), ppid
,
4524 (unsigned long)task_thread_info(p
)->flags
);
4526 print_worker_info(KERN_INFO
, p
);
4527 show_stack(p
, NULL
);
4530 void show_state_filter(unsigned long state_filter
)
4532 struct task_struct
*g
, *p
;
4534 #if BITS_PER_LONG == 32
4536 " task PC stack pid father\n");
4539 " task PC stack pid father\n");
4542 for_each_process_thread(g
, p
) {
4544 * reset the NMI-timeout, listing all files on a slow
4545 * console might take a lot of time:
4547 touch_nmi_watchdog();
4548 if (!state_filter
|| (p
->state
& state_filter
))
4552 touch_all_softlockup_watchdogs();
4554 #ifdef CONFIG_SCHED_DEBUG
4555 sysrq_sched_debug_show();
4559 * Only show locks if all tasks are dumped:
4562 debug_show_all_locks();
4565 void init_idle_bootup_task(struct task_struct
*idle
)
4567 idle
->sched_class
= &idle_sched_class
;
4571 * init_idle - set up an idle thread for a given CPU
4572 * @idle: task in question
4573 * @cpu: cpu the idle task belongs to
4575 * NOTE: this function does not set the idle thread's NEED_RESCHED
4576 * flag, to make booting more robust.
4578 void init_idle(struct task_struct
*idle
, int cpu
)
4580 struct rq
*rq
= cpu_rq(cpu
);
4581 unsigned long flags
;
4583 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4585 __sched_fork(0, idle
);
4586 idle
->state
= TASK_RUNNING
;
4587 idle
->se
.exec_start
= sched_clock();
4589 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4591 * We're having a chicken and egg problem, even though we are
4592 * holding rq->lock, the cpu isn't yet set to this cpu so the
4593 * lockdep check in task_group() will fail.
4595 * Similar case to sched_fork(). / Alternatively we could
4596 * use task_rq_lock() here and obtain the other rq->lock.
4601 __set_task_cpu(idle
, cpu
);
4604 rq
->curr
= rq
->idle
= idle
;
4605 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4606 #if defined(CONFIG_SMP)
4609 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4611 /* Set the preempt count _outside_ the spinlocks! */
4612 init_idle_preempt_count(idle
, cpu
);
4615 * The idle tasks have their own, simple scheduling class:
4617 idle
->sched_class
= &idle_sched_class
;
4618 ftrace_graph_init_idle_task(idle
, cpu
);
4619 vtime_init_idle(idle
, cpu
);
4620 #if defined(CONFIG_SMP)
4621 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4625 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4626 const struct cpumask
*trial
)
4628 int ret
= 1, trial_cpus
;
4629 struct dl_bw
*cur_dl_b
;
4630 unsigned long flags
;
4632 if (!cpumask_weight(cur
))
4635 rcu_read_lock_sched();
4636 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4637 trial_cpus
= cpumask_weight(trial
);
4639 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4640 if (cur_dl_b
->bw
!= -1 &&
4641 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
4643 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
4644 rcu_read_unlock_sched();
4649 int task_can_attach(struct task_struct
*p
,
4650 const struct cpumask
*cs_cpus_allowed
)
4655 * Kthreads which disallow setaffinity shouldn't be moved
4656 * to a new cpuset; we don't want to change their cpu
4657 * affinity and isolating such threads by their set of
4658 * allowed nodes is unnecessary. Thus, cpusets are not
4659 * applicable for such threads. This prevents checking for
4660 * success of set_cpus_allowed_ptr() on all attached tasks
4661 * before cpus_allowed may be changed.
4663 if (p
->flags
& PF_NO_SETAFFINITY
) {
4669 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
4671 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
4676 unsigned long flags
;
4678 rcu_read_lock_sched();
4679 dl_b
= dl_bw_of(dest_cpu
);
4680 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
4681 cpus
= dl_bw_cpus(dest_cpu
);
4682 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
4687 * We reserve space for this task in the destination
4688 * root_domain, as we can't fail after this point.
4689 * We will free resources in the source root_domain
4690 * later on (see set_cpus_allowed_dl()).
4692 __dl_add(dl_b
, p
->dl
.dl_bw
);
4694 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
4695 rcu_read_unlock_sched();
4705 * move_queued_task - move a queued task to new rq.
4707 * Returns (locked) new rq. Old rq's lock is released.
4709 static struct rq
*move_queued_task(struct task_struct
*p
, int new_cpu
)
4711 struct rq
*rq
= task_rq(p
);
4713 lockdep_assert_held(&rq
->lock
);
4715 dequeue_task(rq
, p
, 0);
4716 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
4717 set_task_cpu(p
, new_cpu
);
4718 raw_spin_unlock(&rq
->lock
);
4720 rq
= cpu_rq(new_cpu
);
4722 raw_spin_lock(&rq
->lock
);
4723 BUG_ON(task_cpu(p
) != new_cpu
);
4724 p
->on_rq
= TASK_ON_RQ_QUEUED
;
4725 enqueue_task(rq
, p
, 0);
4726 check_preempt_curr(rq
, p
, 0);
4731 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4733 if (p
->sched_class
->set_cpus_allowed
)
4734 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4736 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4737 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4741 * This is how migration works:
4743 * 1) we invoke migration_cpu_stop() on the target CPU using
4745 * 2) stopper starts to run (implicitly forcing the migrated thread
4747 * 3) it checks whether the migrated task is still in the wrong runqueue.
4748 * 4) if it's in the wrong runqueue then the migration thread removes
4749 * it and puts it into the right queue.
4750 * 5) stopper completes and stop_one_cpu() returns and the migration
4755 * Change a given task's CPU affinity. Migrate the thread to a
4756 * proper CPU and schedule it away if the CPU it's executing on
4757 * is removed from the allowed bitmask.
4759 * NOTE: the caller must have a valid reference to the task, the
4760 * task must not exit() & deallocate itself prematurely. The
4761 * call is not atomic; no spinlocks may be held.
4763 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4765 unsigned long flags
;
4767 unsigned int dest_cpu
;
4770 rq
= task_rq_lock(p
, &flags
);
4772 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4775 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4780 do_set_cpus_allowed(p
, new_mask
);
4782 /* Can the task run on the task's current CPU? If so, we're done */
4783 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4786 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4787 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
4788 struct migration_arg arg
= { p
, dest_cpu
};
4789 /* Need help from migration thread: drop lock and wait. */
4790 task_rq_unlock(rq
, p
, &flags
);
4791 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4792 tlb_migrate_finish(p
->mm
);
4794 } else if (task_on_rq_queued(p
))
4795 rq
= move_queued_task(p
, dest_cpu
);
4797 task_rq_unlock(rq
, p
, &flags
);
4801 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4804 * Move (not current) task off this cpu, onto dest cpu. We're doing
4805 * this because either it can't run here any more (set_cpus_allowed()
4806 * away from this CPU, or CPU going down), or because we're
4807 * attempting to rebalance this task on exec (sched_exec).
4809 * So we race with normal scheduler movements, but that's OK, as long
4810 * as the task is no longer on this CPU.
4812 * Returns non-zero if task was successfully migrated.
4814 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4819 if (unlikely(!cpu_active(dest_cpu
)))
4822 rq
= cpu_rq(src_cpu
);
4824 raw_spin_lock(&p
->pi_lock
);
4825 raw_spin_lock(&rq
->lock
);
4826 /* Already moved. */
4827 if (task_cpu(p
) != src_cpu
)
4830 /* Affinity changed (again). */
4831 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4835 * If we're not on a rq, the next wake-up will ensure we're
4838 if (task_on_rq_queued(p
))
4839 rq
= move_queued_task(p
, dest_cpu
);
4843 raw_spin_unlock(&rq
->lock
);
4844 raw_spin_unlock(&p
->pi_lock
);
4848 #ifdef CONFIG_NUMA_BALANCING
4849 /* Migrate current task p to target_cpu */
4850 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4852 struct migration_arg arg
= { p
, target_cpu
};
4853 int curr_cpu
= task_cpu(p
);
4855 if (curr_cpu
== target_cpu
)
4858 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4861 /* TODO: This is not properly updating schedstats */
4863 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4864 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4868 * Requeue a task on a given node and accurately track the number of NUMA
4869 * tasks on the runqueues
4871 void sched_setnuma(struct task_struct
*p
, int nid
)
4874 unsigned long flags
;
4875 bool queued
, running
;
4877 rq
= task_rq_lock(p
, &flags
);
4878 queued
= task_on_rq_queued(p
);
4879 running
= task_current(rq
, p
);
4882 dequeue_task(rq
, p
, 0);
4884 put_prev_task(rq
, p
);
4886 p
->numa_preferred_nid
= nid
;
4889 p
->sched_class
->set_curr_task(rq
);
4891 enqueue_task(rq
, p
, 0);
4892 task_rq_unlock(rq
, p
, &flags
);
4897 * migration_cpu_stop - this will be executed by a highprio stopper thread
4898 * and performs thread migration by bumping thread off CPU then
4899 * 'pushing' onto another runqueue.
4901 static int migration_cpu_stop(void *data
)
4903 struct migration_arg
*arg
= data
;
4906 * The original target cpu might have gone down and we might
4907 * be on another cpu but it doesn't matter.
4909 local_irq_disable();
4911 * We need to explicitly wake pending tasks before running
4912 * __migrate_task() such that we will not miss enforcing cpus_allowed
4913 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4915 sched_ttwu_pending();
4916 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4921 #ifdef CONFIG_HOTPLUG_CPU
4924 * Ensures that the idle task is using init_mm right before its cpu goes
4927 void idle_task_exit(void)
4929 struct mm_struct
*mm
= current
->active_mm
;
4931 BUG_ON(cpu_online(smp_processor_id()));
4933 if (mm
!= &init_mm
) {
4934 switch_mm(mm
, &init_mm
, current
);
4935 finish_arch_post_lock_switch();
4941 * Since this CPU is going 'away' for a while, fold any nr_active delta
4942 * we might have. Assumes we're called after migrate_tasks() so that the
4943 * nr_active count is stable.
4945 * Also see the comment "Global load-average calculations".
4947 static void calc_load_migrate(struct rq
*rq
)
4949 long delta
= calc_load_fold_active(rq
);
4951 atomic_long_add(delta
, &calc_load_tasks
);
4954 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
4958 static const struct sched_class fake_sched_class
= {
4959 .put_prev_task
= put_prev_task_fake
,
4962 static struct task_struct fake_task
= {
4964 * Avoid pull_{rt,dl}_task()
4966 .prio
= MAX_PRIO
+ 1,
4967 .sched_class
= &fake_sched_class
,
4971 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4972 * try_to_wake_up()->select_task_rq().
4974 * Called with rq->lock held even though we'er in stop_machine() and
4975 * there's no concurrency possible, we hold the required locks anyway
4976 * because of lock validation efforts.
4978 static void migrate_tasks(unsigned int dead_cpu
)
4980 struct rq
*rq
= cpu_rq(dead_cpu
);
4981 struct task_struct
*next
, *stop
= rq
->stop
;
4985 * Fudge the rq selection such that the below task selection loop
4986 * doesn't get stuck on the currently eligible stop task.
4988 * We're currently inside stop_machine() and the rq is either stuck
4989 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4990 * either way we should never end up calling schedule() until we're
4996 * put_prev_task() and pick_next_task() sched
4997 * class method both need to have an up-to-date
4998 * value of rq->clock[_task]
5000 update_rq_clock(rq
);
5004 * There's this thread running, bail when that's the only
5007 if (rq
->nr_running
== 1)
5010 next
= pick_next_task(rq
, &fake_task
);
5012 next
->sched_class
->put_prev_task(rq
, next
);
5014 /* Find suitable destination for @next, with force if needed. */
5015 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5016 raw_spin_unlock(&rq
->lock
);
5018 __migrate_task(next
, dead_cpu
, dest_cpu
);
5020 raw_spin_lock(&rq
->lock
);
5026 #endif /* CONFIG_HOTPLUG_CPU */
5028 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5030 static struct ctl_table sd_ctl_dir
[] = {
5032 .procname
= "sched_domain",
5038 static struct ctl_table sd_ctl_root
[] = {
5040 .procname
= "kernel",
5042 .child
= sd_ctl_dir
,
5047 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5049 struct ctl_table
*entry
=
5050 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5055 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5057 struct ctl_table
*entry
;
5060 * In the intermediate directories, both the child directory and
5061 * procname are dynamically allocated and could fail but the mode
5062 * will always be set. In the lowest directory the names are
5063 * static strings and all have proc handlers.
5065 for (entry
= *tablep
; entry
->mode
; entry
++) {
5067 sd_free_ctl_entry(&entry
->child
);
5068 if (entry
->proc_handler
== NULL
)
5069 kfree(entry
->procname
);
5076 static int min_load_idx
= 0;
5077 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5080 set_table_entry(struct ctl_table
*entry
,
5081 const char *procname
, void *data
, int maxlen
,
5082 umode_t mode
, proc_handler
*proc_handler
,
5085 entry
->procname
= procname
;
5087 entry
->maxlen
= maxlen
;
5089 entry
->proc_handler
= proc_handler
;
5092 entry
->extra1
= &min_load_idx
;
5093 entry
->extra2
= &max_load_idx
;
5097 static struct ctl_table
*
5098 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5100 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5105 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5106 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5107 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5108 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5109 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5110 sizeof(int), 0644, proc_dointvec_minmax
, true);
5111 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5112 sizeof(int), 0644, proc_dointvec_minmax
, true);
5113 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5114 sizeof(int), 0644, proc_dointvec_minmax
, true);
5115 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5116 sizeof(int), 0644, proc_dointvec_minmax
, true);
5117 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5118 sizeof(int), 0644, proc_dointvec_minmax
, true);
5119 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5120 sizeof(int), 0644, proc_dointvec_minmax
, false);
5121 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5122 sizeof(int), 0644, proc_dointvec_minmax
, false);
5123 set_table_entry(&table
[9], "cache_nice_tries",
5124 &sd
->cache_nice_tries
,
5125 sizeof(int), 0644, proc_dointvec_minmax
, false);
5126 set_table_entry(&table
[10], "flags", &sd
->flags
,
5127 sizeof(int), 0644, proc_dointvec_minmax
, false);
5128 set_table_entry(&table
[11], "max_newidle_lb_cost",
5129 &sd
->max_newidle_lb_cost
,
5130 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5131 set_table_entry(&table
[12], "name", sd
->name
,
5132 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5133 /* &table[13] is terminator */
5138 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5140 struct ctl_table
*entry
, *table
;
5141 struct sched_domain
*sd
;
5142 int domain_num
= 0, i
;
5145 for_each_domain(cpu
, sd
)
5147 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5152 for_each_domain(cpu
, sd
) {
5153 snprintf(buf
, 32, "domain%d", i
);
5154 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5156 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5163 static struct ctl_table_header
*sd_sysctl_header
;
5164 static void register_sched_domain_sysctl(void)
5166 int i
, cpu_num
= num_possible_cpus();
5167 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5170 WARN_ON(sd_ctl_dir
[0].child
);
5171 sd_ctl_dir
[0].child
= entry
;
5176 for_each_possible_cpu(i
) {
5177 snprintf(buf
, 32, "cpu%d", i
);
5178 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5180 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5184 WARN_ON(sd_sysctl_header
);
5185 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5188 /* may be called multiple times per register */
5189 static void unregister_sched_domain_sysctl(void)
5191 if (sd_sysctl_header
)
5192 unregister_sysctl_table(sd_sysctl_header
);
5193 sd_sysctl_header
= NULL
;
5194 if (sd_ctl_dir
[0].child
)
5195 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5198 static void register_sched_domain_sysctl(void)
5201 static void unregister_sched_domain_sysctl(void)
5206 static void set_rq_online(struct rq
*rq
)
5209 const struct sched_class
*class;
5211 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5214 for_each_class(class) {
5215 if (class->rq_online
)
5216 class->rq_online(rq
);
5221 static void set_rq_offline(struct rq
*rq
)
5224 const struct sched_class
*class;
5226 for_each_class(class) {
5227 if (class->rq_offline
)
5228 class->rq_offline(rq
);
5231 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5237 * migration_call - callback that gets triggered when a CPU is added.
5238 * Here we can start up the necessary migration thread for the new CPU.
5241 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5243 int cpu
= (long)hcpu
;
5244 unsigned long flags
;
5245 struct rq
*rq
= cpu_rq(cpu
);
5247 switch (action
& ~CPU_TASKS_FROZEN
) {
5249 case CPU_UP_PREPARE
:
5250 rq
->calc_load_update
= calc_load_update
;
5254 /* Update our root-domain */
5255 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5257 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5261 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5264 #ifdef CONFIG_HOTPLUG_CPU
5266 sched_ttwu_pending();
5267 /* Update our root-domain */
5268 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5270 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5274 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5275 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5279 calc_load_migrate(rq
);
5284 update_max_interval();
5290 * Register at high priority so that task migration (migrate_all_tasks)
5291 * happens before everything else. This has to be lower priority than
5292 * the notifier in the perf_event subsystem, though.
5294 static struct notifier_block migration_notifier
= {
5295 .notifier_call
= migration_call
,
5296 .priority
= CPU_PRI_MIGRATION
,
5299 static void __cpuinit
set_cpu_rq_start_time(void)
5301 int cpu
= smp_processor_id();
5302 struct rq
*rq
= cpu_rq(cpu
);
5303 rq
->age_stamp
= sched_clock_cpu(cpu
);
5306 static int sched_cpu_active(struct notifier_block
*nfb
,
5307 unsigned long action
, void *hcpu
)
5309 switch (action
& ~CPU_TASKS_FROZEN
) {
5311 set_cpu_rq_start_time();
5313 case CPU_DOWN_FAILED
:
5314 set_cpu_active((long)hcpu
, true);
5321 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5322 unsigned long action
, void *hcpu
)
5324 switch (action
& ~CPU_TASKS_FROZEN
) {
5325 case CPU_DOWN_PREPARE
:
5326 set_cpu_active((long)hcpu
, false);
5333 static int __init
migration_init(void)
5335 void *cpu
= (void *)(long)smp_processor_id();
5338 /* Initialize migration for the boot CPU */
5339 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5340 BUG_ON(err
== NOTIFY_BAD
);
5341 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5342 register_cpu_notifier(&migration_notifier
);
5344 /* Register cpu active notifiers */
5345 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5346 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5350 early_initcall(migration_init
);
5355 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5357 #ifdef CONFIG_SCHED_DEBUG
5359 static __read_mostly
int sched_debug_enabled
;
5361 static int __init
sched_debug_setup(char *str
)
5363 sched_debug_enabled
= 1;
5367 early_param("sched_debug", sched_debug_setup
);
5369 static inline bool sched_debug(void)
5371 return sched_debug_enabled
;
5374 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5375 struct cpumask
*groupmask
)
5377 struct sched_group
*group
= sd
->groups
;
5379 cpumask_clear(groupmask
);
5381 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5383 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5384 printk("does not load-balance\n");
5386 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5391 printk(KERN_CONT
"span %*pbl level %s\n",
5392 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5394 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5395 printk(KERN_ERR
"ERROR: domain->span does not contain "
5398 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5399 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5403 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5407 printk(KERN_ERR
"ERROR: group is NULL\n");
5411 if (!cpumask_weight(sched_group_cpus(group
))) {
5412 printk(KERN_CONT
"\n");
5413 printk(KERN_ERR
"ERROR: empty group\n");
5417 if (!(sd
->flags
& SD_OVERLAP
) &&
5418 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5419 printk(KERN_CONT
"\n");
5420 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5424 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5426 printk(KERN_CONT
" %*pbl",
5427 cpumask_pr_args(sched_group_cpus(group
)));
5428 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5429 printk(KERN_CONT
" (cpu_capacity = %d)",
5430 group
->sgc
->capacity
);
5433 group
= group
->next
;
5434 } while (group
!= sd
->groups
);
5435 printk(KERN_CONT
"\n");
5437 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5438 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5441 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5442 printk(KERN_ERR
"ERROR: parent span is not a superset "
5443 "of domain->span\n");
5447 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5451 if (!sched_debug_enabled
)
5455 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5459 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5462 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5470 #else /* !CONFIG_SCHED_DEBUG */
5471 # define sched_domain_debug(sd, cpu) do { } while (0)
5472 static inline bool sched_debug(void)
5476 #endif /* CONFIG_SCHED_DEBUG */
5478 static int sd_degenerate(struct sched_domain
*sd
)
5480 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5483 /* Following flags need at least 2 groups */
5484 if (sd
->flags
& (SD_LOAD_BALANCE
|
5485 SD_BALANCE_NEWIDLE
|
5488 SD_SHARE_CPUCAPACITY
|
5489 SD_SHARE_PKG_RESOURCES
|
5490 SD_SHARE_POWERDOMAIN
)) {
5491 if (sd
->groups
!= sd
->groups
->next
)
5495 /* Following flags don't use groups */
5496 if (sd
->flags
& (SD_WAKE_AFFINE
))
5503 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5505 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5507 if (sd_degenerate(parent
))
5510 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5513 /* Flags needing groups don't count if only 1 group in parent */
5514 if (parent
->groups
== parent
->groups
->next
) {
5515 pflags
&= ~(SD_LOAD_BALANCE
|
5516 SD_BALANCE_NEWIDLE
|
5519 SD_SHARE_CPUCAPACITY
|
5520 SD_SHARE_PKG_RESOURCES
|
5522 SD_SHARE_POWERDOMAIN
);
5523 if (nr_node_ids
== 1)
5524 pflags
&= ~SD_SERIALIZE
;
5526 if (~cflags
& pflags
)
5532 static void free_rootdomain(struct rcu_head
*rcu
)
5534 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5536 cpupri_cleanup(&rd
->cpupri
);
5537 cpudl_cleanup(&rd
->cpudl
);
5538 free_cpumask_var(rd
->dlo_mask
);
5539 free_cpumask_var(rd
->rto_mask
);
5540 free_cpumask_var(rd
->online
);
5541 free_cpumask_var(rd
->span
);
5545 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5547 struct root_domain
*old_rd
= NULL
;
5548 unsigned long flags
;
5550 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5555 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5558 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5561 * If we dont want to free the old_rd yet then
5562 * set old_rd to NULL to skip the freeing later
5565 if (!atomic_dec_and_test(&old_rd
->refcount
))
5569 atomic_inc(&rd
->refcount
);
5572 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5573 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5576 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5579 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5582 static int init_rootdomain(struct root_domain
*rd
)
5584 memset(rd
, 0, sizeof(*rd
));
5586 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5588 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5590 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5592 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5595 init_dl_bw(&rd
->dl_bw
);
5596 if (cpudl_init(&rd
->cpudl
) != 0)
5599 if (cpupri_init(&rd
->cpupri
) != 0)
5604 free_cpumask_var(rd
->rto_mask
);
5606 free_cpumask_var(rd
->dlo_mask
);
5608 free_cpumask_var(rd
->online
);
5610 free_cpumask_var(rd
->span
);
5616 * By default the system creates a single root-domain with all cpus as
5617 * members (mimicking the global state we have today).
5619 struct root_domain def_root_domain
;
5621 static void init_defrootdomain(void)
5623 init_rootdomain(&def_root_domain
);
5625 atomic_set(&def_root_domain
.refcount
, 1);
5628 static struct root_domain
*alloc_rootdomain(void)
5630 struct root_domain
*rd
;
5632 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5636 if (init_rootdomain(rd
) != 0) {
5644 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5646 struct sched_group
*tmp
, *first
;
5655 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5660 } while (sg
!= first
);
5663 static void free_sched_domain(struct rcu_head
*rcu
)
5665 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5668 * If its an overlapping domain it has private groups, iterate and
5671 if (sd
->flags
& SD_OVERLAP
) {
5672 free_sched_groups(sd
->groups
, 1);
5673 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5674 kfree(sd
->groups
->sgc
);
5680 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5682 call_rcu(&sd
->rcu
, free_sched_domain
);
5685 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5687 for (; sd
; sd
= sd
->parent
)
5688 destroy_sched_domain(sd
, cpu
);
5692 * Keep a special pointer to the highest sched_domain that has
5693 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5694 * allows us to avoid some pointer chasing select_idle_sibling().
5696 * Also keep a unique ID per domain (we use the first cpu number in
5697 * the cpumask of the domain), this allows us to quickly tell if
5698 * two cpus are in the same cache domain, see cpus_share_cache().
5700 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5701 DEFINE_PER_CPU(int, sd_llc_size
);
5702 DEFINE_PER_CPU(int, sd_llc_id
);
5703 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5704 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5705 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5707 static void update_top_cache_domain(int cpu
)
5709 struct sched_domain
*sd
;
5710 struct sched_domain
*busy_sd
= NULL
;
5714 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5716 id
= cpumask_first(sched_domain_span(sd
));
5717 size
= cpumask_weight(sched_domain_span(sd
));
5718 busy_sd
= sd
->parent
; /* sd_busy */
5720 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5722 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5723 per_cpu(sd_llc_size
, cpu
) = size
;
5724 per_cpu(sd_llc_id
, cpu
) = id
;
5726 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5727 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5729 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5730 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5734 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5735 * hold the hotplug lock.
5738 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5740 struct rq
*rq
= cpu_rq(cpu
);
5741 struct sched_domain
*tmp
;
5743 /* Remove the sched domains which do not contribute to scheduling. */
5744 for (tmp
= sd
; tmp
; ) {
5745 struct sched_domain
*parent
= tmp
->parent
;
5749 if (sd_parent_degenerate(tmp
, parent
)) {
5750 tmp
->parent
= parent
->parent
;
5752 parent
->parent
->child
= tmp
;
5754 * Transfer SD_PREFER_SIBLING down in case of a
5755 * degenerate parent; the spans match for this
5756 * so the property transfers.
5758 if (parent
->flags
& SD_PREFER_SIBLING
)
5759 tmp
->flags
|= SD_PREFER_SIBLING
;
5760 destroy_sched_domain(parent
, cpu
);
5765 if (sd
&& sd_degenerate(sd
)) {
5768 destroy_sched_domain(tmp
, cpu
);
5773 sched_domain_debug(sd
, cpu
);
5775 rq_attach_root(rq
, rd
);
5777 rcu_assign_pointer(rq
->sd
, sd
);
5778 destroy_sched_domains(tmp
, cpu
);
5780 update_top_cache_domain(cpu
);
5783 /* Setup the mask of cpus configured for isolated domains */
5784 static int __init
isolated_cpu_setup(char *str
)
5786 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5787 cpulist_parse(str
, cpu_isolated_map
);
5791 __setup("isolcpus=", isolated_cpu_setup
);
5794 struct sched_domain
** __percpu sd
;
5795 struct root_domain
*rd
;
5806 * Build an iteration mask that can exclude certain CPUs from the upwards
5809 * Asymmetric node setups can result in situations where the domain tree is of
5810 * unequal depth, make sure to skip domains that already cover the entire
5813 * In that case build_sched_domains() will have terminated the iteration early
5814 * and our sibling sd spans will be empty. Domains should always include the
5815 * cpu they're built on, so check that.
5818 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5820 const struct cpumask
*span
= sched_domain_span(sd
);
5821 struct sd_data
*sdd
= sd
->private;
5822 struct sched_domain
*sibling
;
5825 for_each_cpu(i
, span
) {
5826 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5827 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5830 cpumask_set_cpu(i
, sched_group_mask(sg
));
5835 * Return the canonical balance cpu for this group, this is the first cpu
5836 * of this group that's also in the iteration mask.
5838 int group_balance_cpu(struct sched_group
*sg
)
5840 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5844 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5846 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5847 const struct cpumask
*span
= sched_domain_span(sd
);
5848 struct cpumask
*covered
= sched_domains_tmpmask
;
5849 struct sd_data
*sdd
= sd
->private;
5850 struct sched_domain
*sibling
;
5853 cpumask_clear(covered
);
5855 for_each_cpu(i
, span
) {
5856 struct cpumask
*sg_span
;
5858 if (cpumask_test_cpu(i
, covered
))
5861 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5863 /* See the comment near build_group_mask(). */
5864 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5867 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5868 GFP_KERNEL
, cpu_to_node(cpu
));
5873 sg_span
= sched_group_cpus(sg
);
5875 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
5877 cpumask_set_cpu(i
, sg_span
);
5879 cpumask_or(covered
, covered
, sg_span
);
5881 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
5882 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
5883 build_group_mask(sd
, sg
);
5886 * Initialize sgc->capacity such that even if we mess up the
5887 * domains and no possible iteration will get us here, we won't
5890 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
5893 * Make sure the first group of this domain contains the
5894 * canonical balance cpu. Otherwise the sched_domain iteration
5895 * breaks. See update_sg_lb_stats().
5897 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5898 group_balance_cpu(sg
) == cpu
)
5908 sd
->groups
= groups
;
5913 free_sched_groups(first
, 0);
5918 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5920 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5921 struct sched_domain
*child
= sd
->child
;
5924 cpu
= cpumask_first(sched_domain_span(child
));
5927 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
5928 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
5929 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
5936 * build_sched_groups will build a circular linked list of the groups
5937 * covered by the given span, and will set each group's ->cpumask correctly,
5938 * and ->cpu_capacity to 0.
5940 * Assumes the sched_domain tree is fully constructed
5943 build_sched_groups(struct sched_domain
*sd
, int cpu
)
5945 struct sched_group
*first
= NULL
, *last
= NULL
;
5946 struct sd_data
*sdd
= sd
->private;
5947 const struct cpumask
*span
= sched_domain_span(sd
);
5948 struct cpumask
*covered
;
5951 get_group(cpu
, sdd
, &sd
->groups
);
5952 atomic_inc(&sd
->groups
->ref
);
5954 if (cpu
!= cpumask_first(span
))
5957 lockdep_assert_held(&sched_domains_mutex
);
5958 covered
= sched_domains_tmpmask
;
5960 cpumask_clear(covered
);
5962 for_each_cpu(i
, span
) {
5963 struct sched_group
*sg
;
5966 if (cpumask_test_cpu(i
, covered
))
5969 group
= get_group(i
, sdd
, &sg
);
5970 cpumask_setall(sched_group_mask(sg
));
5972 for_each_cpu(j
, span
) {
5973 if (get_group(j
, sdd
, NULL
) != group
)
5976 cpumask_set_cpu(j
, covered
);
5977 cpumask_set_cpu(j
, sched_group_cpus(sg
));
5992 * Initialize sched groups cpu_capacity.
5994 * cpu_capacity indicates the capacity of sched group, which is used while
5995 * distributing the load between different sched groups in a sched domain.
5996 * Typically cpu_capacity for all the groups in a sched domain will be same
5997 * unless there are asymmetries in the topology. If there are asymmetries,
5998 * group having more cpu_capacity will pickup more load compared to the
5999 * group having less cpu_capacity.
6001 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6003 struct sched_group
*sg
= sd
->groups
;
6008 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6010 } while (sg
!= sd
->groups
);
6012 if (cpu
!= group_balance_cpu(sg
))
6015 update_group_capacity(sd
, cpu
);
6016 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6020 * Initializers for schedule domains
6021 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6024 static int default_relax_domain_level
= -1;
6025 int sched_domain_level_max
;
6027 static int __init
setup_relax_domain_level(char *str
)
6029 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6030 pr_warn("Unable to set relax_domain_level\n");
6034 __setup("relax_domain_level=", setup_relax_domain_level
);
6036 static void set_domain_attribute(struct sched_domain
*sd
,
6037 struct sched_domain_attr
*attr
)
6041 if (!attr
|| attr
->relax_domain_level
< 0) {
6042 if (default_relax_domain_level
< 0)
6045 request
= default_relax_domain_level
;
6047 request
= attr
->relax_domain_level
;
6048 if (request
< sd
->level
) {
6049 /* turn off idle balance on this domain */
6050 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6052 /* turn on idle balance on this domain */
6053 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6057 static void __sdt_free(const struct cpumask
*cpu_map
);
6058 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6060 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6061 const struct cpumask
*cpu_map
)
6065 if (!atomic_read(&d
->rd
->refcount
))
6066 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6068 free_percpu(d
->sd
); /* fall through */
6070 __sdt_free(cpu_map
); /* fall through */
6076 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6077 const struct cpumask
*cpu_map
)
6079 memset(d
, 0, sizeof(*d
));
6081 if (__sdt_alloc(cpu_map
))
6082 return sa_sd_storage
;
6083 d
->sd
= alloc_percpu(struct sched_domain
*);
6085 return sa_sd_storage
;
6086 d
->rd
= alloc_rootdomain();
6089 return sa_rootdomain
;
6093 * NULL the sd_data elements we've used to build the sched_domain and
6094 * sched_group structure so that the subsequent __free_domain_allocs()
6095 * will not free the data we're using.
6097 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6099 struct sd_data
*sdd
= sd
->private;
6101 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6102 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6104 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6105 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6107 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6108 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6112 static int sched_domains_numa_levels
;
6113 enum numa_topology_type sched_numa_topology_type
;
6114 static int *sched_domains_numa_distance
;
6115 int sched_max_numa_distance
;
6116 static struct cpumask
***sched_domains_numa_masks
;
6117 static int sched_domains_curr_level
;
6121 * SD_flags allowed in topology descriptions.
6123 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6124 * SD_SHARE_PKG_RESOURCES - describes shared caches
6125 * SD_NUMA - describes NUMA topologies
6126 * SD_SHARE_POWERDOMAIN - describes shared power domain
6129 * SD_ASYM_PACKING - describes SMT quirks
6131 #define TOPOLOGY_SD_FLAGS \
6132 (SD_SHARE_CPUCAPACITY | \
6133 SD_SHARE_PKG_RESOURCES | \
6136 SD_SHARE_POWERDOMAIN)
6138 static struct sched_domain
*
6139 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6141 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6142 int sd_weight
, sd_flags
= 0;
6146 * Ugly hack to pass state to sd_numa_mask()...
6148 sched_domains_curr_level
= tl
->numa_level
;
6151 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6154 sd_flags
= (*tl
->sd_flags
)();
6155 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6156 "wrong sd_flags in topology description\n"))
6157 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6159 *sd
= (struct sched_domain
){
6160 .min_interval
= sd_weight
,
6161 .max_interval
= 2*sd_weight
,
6163 .imbalance_pct
= 125,
6165 .cache_nice_tries
= 0,
6172 .flags
= 1*SD_LOAD_BALANCE
6173 | 1*SD_BALANCE_NEWIDLE
6178 | 0*SD_SHARE_CPUCAPACITY
6179 | 0*SD_SHARE_PKG_RESOURCES
6181 | 0*SD_PREFER_SIBLING
6186 .last_balance
= jiffies
,
6187 .balance_interval
= sd_weight
,
6189 .max_newidle_lb_cost
= 0,
6190 .next_decay_max_lb_cost
= jiffies
,
6191 #ifdef CONFIG_SCHED_DEBUG
6197 * Convert topological properties into behaviour.
6200 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6201 sd
->flags
|= SD_PREFER_SIBLING
;
6202 sd
->imbalance_pct
= 110;
6203 sd
->smt_gain
= 1178; /* ~15% */
6205 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6206 sd
->imbalance_pct
= 117;
6207 sd
->cache_nice_tries
= 1;
6211 } else if (sd
->flags
& SD_NUMA
) {
6212 sd
->cache_nice_tries
= 2;
6216 sd
->flags
|= SD_SERIALIZE
;
6217 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6218 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6225 sd
->flags
|= SD_PREFER_SIBLING
;
6226 sd
->cache_nice_tries
= 1;
6231 sd
->private = &tl
->data
;
6237 * Topology list, bottom-up.
6239 static struct sched_domain_topology_level default_topology
[] = {
6240 #ifdef CONFIG_SCHED_SMT
6241 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6243 #ifdef CONFIG_SCHED_MC
6244 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6246 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6250 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6252 #define for_each_sd_topology(tl) \
6253 for (tl = sched_domain_topology; tl->mask; tl++)
6255 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6257 sched_domain_topology
= tl
;
6262 static const struct cpumask
*sd_numa_mask(int cpu
)
6264 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6267 static void sched_numa_warn(const char *str
)
6269 static int done
= false;
6277 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6279 for (i
= 0; i
< nr_node_ids
; i
++) {
6280 printk(KERN_WARNING
" ");
6281 for (j
= 0; j
< nr_node_ids
; j
++)
6282 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6283 printk(KERN_CONT
"\n");
6285 printk(KERN_WARNING
"\n");
6288 bool find_numa_distance(int distance
)
6292 if (distance
== node_distance(0, 0))
6295 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6296 if (sched_domains_numa_distance
[i
] == distance
)
6304 * A system can have three types of NUMA topology:
6305 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6306 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6307 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6309 * The difference between a glueless mesh topology and a backplane
6310 * topology lies in whether communication between not directly
6311 * connected nodes goes through intermediary nodes (where programs
6312 * could run), or through backplane controllers. This affects
6313 * placement of programs.
6315 * The type of topology can be discerned with the following tests:
6316 * - If the maximum distance between any nodes is 1 hop, the system
6317 * is directly connected.
6318 * - If for two nodes A and B, located N > 1 hops away from each other,
6319 * there is an intermediary node C, which is < N hops away from both
6320 * nodes A and B, the system is a glueless mesh.
6322 static void init_numa_topology_type(void)
6326 n
= sched_max_numa_distance
;
6329 sched_numa_topology_type
= NUMA_DIRECT
;
6331 for_each_online_node(a
) {
6332 for_each_online_node(b
) {
6333 /* Find two nodes furthest removed from each other. */
6334 if (node_distance(a
, b
) < n
)
6337 /* Is there an intermediary node between a and b? */
6338 for_each_online_node(c
) {
6339 if (node_distance(a
, c
) < n
&&
6340 node_distance(b
, c
) < n
) {
6341 sched_numa_topology_type
=
6347 sched_numa_topology_type
= NUMA_BACKPLANE
;
6353 static void sched_init_numa(void)
6355 int next_distance
, curr_distance
= node_distance(0, 0);
6356 struct sched_domain_topology_level
*tl
;
6360 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6361 if (!sched_domains_numa_distance
)
6365 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6366 * unique distances in the node_distance() table.
6368 * Assumes node_distance(0,j) includes all distances in
6369 * node_distance(i,j) in order to avoid cubic time.
6371 next_distance
= curr_distance
;
6372 for (i
= 0; i
< nr_node_ids
; i
++) {
6373 for (j
= 0; j
< nr_node_ids
; j
++) {
6374 for (k
= 0; k
< nr_node_ids
; k
++) {
6375 int distance
= node_distance(i
, k
);
6377 if (distance
> curr_distance
&&
6378 (distance
< next_distance
||
6379 next_distance
== curr_distance
))
6380 next_distance
= distance
;
6383 * While not a strong assumption it would be nice to know
6384 * about cases where if node A is connected to B, B is not
6385 * equally connected to A.
6387 if (sched_debug() && node_distance(k
, i
) != distance
)
6388 sched_numa_warn("Node-distance not symmetric");
6390 if (sched_debug() && i
&& !find_numa_distance(distance
))
6391 sched_numa_warn("Node-0 not representative");
6393 if (next_distance
!= curr_distance
) {
6394 sched_domains_numa_distance
[level
++] = next_distance
;
6395 sched_domains_numa_levels
= level
;
6396 curr_distance
= next_distance
;
6401 * In case of sched_debug() we verify the above assumption.
6411 * 'level' contains the number of unique distances, excluding the
6412 * identity distance node_distance(i,i).
6414 * The sched_domains_numa_distance[] array includes the actual distance
6419 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6420 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6421 * the array will contain less then 'level' members. This could be
6422 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6423 * in other functions.
6425 * We reset it to 'level' at the end of this function.
6427 sched_domains_numa_levels
= 0;
6429 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6430 if (!sched_domains_numa_masks
)
6434 * Now for each level, construct a mask per node which contains all
6435 * cpus of nodes that are that many hops away from us.
6437 for (i
= 0; i
< level
; i
++) {
6438 sched_domains_numa_masks
[i
] =
6439 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6440 if (!sched_domains_numa_masks
[i
])
6443 for (j
= 0; j
< nr_node_ids
; j
++) {
6444 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6448 sched_domains_numa_masks
[i
][j
] = mask
;
6450 for (k
= 0; k
< nr_node_ids
; k
++) {
6451 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6454 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6459 /* Compute default topology size */
6460 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6462 tl
= kzalloc((i
+ level
+ 1) *
6463 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6468 * Copy the default topology bits..
6470 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6471 tl
[i
] = sched_domain_topology
[i
];
6474 * .. and append 'j' levels of NUMA goodness.
6476 for (j
= 0; j
< level
; i
++, j
++) {
6477 tl
[i
] = (struct sched_domain_topology_level
){
6478 .mask
= sd_numa_mask
,
6479 .sd_flags
= cpu_numa_flags
,
6480 .flags
= SDTL_OVERLAP
,
6486 sched_domain_topology
= tl
;
6488 sched_domains_numa_levels
= level
;
6489 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6491 init_numa_topology_type();
6494 static void sched_domains_numa_masks_set(int cpu
)
6497 int node
= cpu_to_node(cpu
);
6499 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6500 for (j
= 0; j
< nr_node_ids
; j
++) {
6501 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6502 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6507 static void sched_domains_numa_masks_clear(int cpu
)
6510 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6511 for (j
= 0; j
< nr_node_ids
; j
++)
6512 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6517 * Update sched_domains_numa_masks[level][node] array when new cpus
6520 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6521 unsigned long action
,
6524 int cpu
= (long)hcpu
;
6526 switch (action
& ~CPU_TASKS_FROZEN
) {
6528 sched_domains_numa_masks_set(cpu
);
6532 sched_domains_numa_masks_clear(cpu
);
6542 static inline void sched_init_numa(void)
6546 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6547 unsigned long action
,
6552 #endif /* CONFIG_NUMA */
6554 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6556 struct sched_domain_topology_level
*tl
;
6559 for_each_sd_topology(tl
) {
6560 struct sd_data
*sdd
= &tl
->data
;
6562 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6566 sdd
->sg
= alloc_percpu(struct sched_group
*);
6570 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6574 for_each_cpu(j
, cpu_map
) {
6575 struct sched_domain
*sd
;
6576 struct sched_group
*sg
;
6577 struct sched_group_capacity
*sgc
;
6579 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6580 GFP_KERNEL
, cpu_to_node(j
));
6584 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6586 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6587 GFP_KERNEL
, cpu_to_node(j
));
6593 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6595 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6596 GFP_KERNEL
, cpu_to_node(j
));
6600 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6607 static void __sdt_free(const struct cpumask
*cpu_map
)
6609 struct sched_domain_topology_level
*tl
;
6612 for_each_sd_topology(tl
) {
6613 struct sd_data
*sdd
= &tl
->data
;
6615 for_each_cpu(j
, cpu_map
) {
6616 struct sched_domain
*sd
;
6619 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6620 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6621 free_sched_groups(sd
->groups
, 0);
6622 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6626 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6628 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6630 free_percpu(sdd
->sd
);
6632 free_percpu(sdd
->sg
);
6634 free_percpu(sdd
->sgc
);
6639 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6640 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6641 struct sched_domain
*child
, int cpu
)
6643 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6647 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6649 sd
->level
= child
->level
+ 1;
6650 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6654 if (!cpumask_subset(sched_domain_span(child
),
6655 sched_domain_span(sd
))) {
6656 pr_err("BUG: arch topology borken\n");
6657 #ifdef CONFIG_SCHED_DEBUG
6658 pr_err(" the %s domain not a subset of the %s domain\n",
6659 child
->name
, sd
->name
);
6661 /* Fixup, ensure @sd has at least @child cpus. */
6662 cpumask_or(sched_domain_span(sd
),
6663 sched_domain_span(sd
),
6664 sched_domain_span(child
));
6668 set_domain_attribute(sd
, attr
);
6674 * Build sched domains for a given set of cpus and attach the sched domains
6675 * to the individual cpus
6677 static int build_sched_domains(const struct cpumask
*cpu_map
,
6678 struct sched_domain_attr
*attr
)
6680 enum s_alloc alloc_state
;
6681 struct sched_domain
*sd
;
6683 int i
, ret
= -ENOMEM
;
6685 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6686 if (alloc_state
!= sa_rootdomain
)
6689 /* Set up domains for cpus specified by the cpu_map. */
6690 for_each_cpu(i
, cpu_map
) {
6691 struct sched_domain_topology_level
*tl
;
6694 for_each_sd_topology(tl
) {
6695 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6696 if (tl
== sched_domain_topology
)
6697 *per_cpu_ptr(d
.sd
, i
) = sd
;
6698 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6699 sd
->flags
|= SD_OVERLAP
;
6700 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6705 /* Build the groups for the domains */
6706 for_each_cpu(i
, cpu_map
) {
6707 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6708 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6709 if (sd
->flags
& SD_OVERLAP
) {
6710 if (build_overlap_sched_groups(sd
, i
))
6713 if (build_sched_groups(sd
, i
))
6719 /* Calculate CPU capacity for physical packages and nodes */
6720 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6721 if (!cpumask_test_cpu(i
, cpu_map
))
6724 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6725 claim_allocations(i
, sd
);
6726 init_sched_groups_capacity(i
, sd
);
6730 /* Attach the domains */
6732 for_each_cpu(i
, cpu_map
) {
6733 sd
= *per_cpu_ptr(d
.sd
, i
);
6734 cpu_attach_domain(sd
, d
.rd
, i
);
6740 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6744 static cpumask_var_t
*doms_cur
; /* current sched domains */
6745 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6746 static struct sched_domain_attr
*dattr_cur
;
6747 /* attribues of custom domains in 'doms_cur' */
6750 * Special case: If a kmalloc of a doms_cur partition (array of
6751 * cpumask) fails, then fallback to a single sched domain,
6752 * as determined by the single cpumask fallback_doms.
6754 static cpumask_var_t fallback_doms
;
6757 * arch_update_cpu_topology lets virtualized architectures update the
6758 * cpu core maps. It is supposed to return 1 if the topology changed
6759 * or 0 if it stayed the same.
6761 int __weak
arch_update_cpu_topology(void)
6766 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6769 cpumask_var_t
*doms
;
6771 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6774 for (i
= 0; i
< ndoms
; i
++) {
6775 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6776 free_sched_domains(doms
, i
);
6783 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6786 for (i
= 0; i
< ndoms
; i
++)
6787 free_cpumask_var(doms
[i
]);
6792 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6793 * For now this just excludes isolated cpus, but could be used to
6794 * exclude other special cases in the future.
6796 static int init_sched_domains(const struct cpumask
*cpu_map
)
6800 arch_update_cpu_topology();
6802 doms_cur
= alloc_sched_domains(ndoms_cur
);
6804 doms_cur
= &fallback_doms
;
6805 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6806 err
= build_sched_domains(doms_cur
[0], NULL
);
6807 register_sched_domain_sysctl();
6813 * Detach sched domains from a group of cpus specified in cpu_map
6814 * These cpus will now be attached to the NULL domain
6816 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6821 for_each_cpu(i
, cpu_map
)
6822 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6826 /* handle null as "default" */
6827 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6828 struct sched_domain_attr
*new, int idx_new
)
6830 struct sched_domain_attr tmp
;
6837 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6838 new ? (new + idx_new
) : &tmp
,
6839 sizeof(struct sched_domain_attr
));
6843 * Partition sched domains as specified by the 'ndoms_new'
6844 * cpumasks in the array doms_new[] of cpumasks. This compares
6845 * doms_new[] to the current sched domain partitioning, doms_cur[].
6846 * It destroys each deleted domain and builds each new domain.
6848 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6849 * The masks don't intersect (don't overlap.) We should setup one
6850 * sched domain for each mask. CPUs not in any of the cpumasks will
6851 * not be load balanced. If the same cpumask appears both in the
6852 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6855 * The passed in 'doms_new' should be allocated using
6856 * alloc_sched_domains. This routine takes ownership of it and will
6857 * free_sched_domains it when done with it. If the caller failed the
6858 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6859 * and partition_sched_domains() will fallback to the single partition
6860 * 'fallback_doms', it also forces the domains to be rebuilt.
6862 * If doms_new == NULL it will be replaced with cpu_online_mask.
6863 * ndoms_new == 0 is a special case for destroying existing domains,
6864 * and it will not create the default domain.
6866 * Call with hotplug lock held
6868 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6869 struct sched_domain_attr
*dattr_new
)
6874 mutex_lock(&sched_domains_mutex
);
6876 /* always unregister in case we don't destroy any domains */
6877 unregister_sched_domain_sysctl();
6879 /* Let architecture update cpu core mappings. */
6880 new_topology
= arch_update_cpu_topology();
6882 n
= doms_new
? ndoms_new
: 0;
6884 /* Destroy deleted domains */
6885 for (i
= 0; i
< ndoms_cur
; i
++) {
6886 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6887 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6888 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6891 /* no match - a current sched domain not in new doms_new[] */
6892 detach_destroy_domains(doms_cur
[i
]);
6898 if (doms_new
== NULL
) {
6900 doms_new
= &fallback_doms
;
6901 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6902 WARN_ON_ONCE(dattr_new
);
6905 /* Build new domains */
6906 for (i
= 0; i
< ndoms_new
; i
++) {
6907 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6908 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6909 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6912 /* no match - add a new doms_new */
6913 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6918 /* Remember the new sched domains */
6919 if (doms_cur
!= &fallback_doms
)
6920 free_sched_domains(doms_cur
, ndoms_cur
);
6921 kfree(dattr_cur
); /* kfree(NULL) is safe */
6922 doms_cur
= doms_new
;
6923 dattr_cur
= dattr_new
;
6924 ndoms_cur
= ndoms_new
;
6926 register_sched_domain_sysctl();
6928 mutex_unlock(&sched_domains_mutex
);
6931 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
6934 * Update cpusets according to cpu_active mask. If cpusets are
6935 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6936 * around partition_sched_domains().
6938 * If we come here as part of a suspend/resume, don't touch cpusets because we
6939 * want to restore it back to its original state upon resume anyway.
6941 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
6945 case CPU_ONLINE_FROZEN
:
6946 case CPU_DOWN_FAILED_FROZEN
:
6949 * num_cpus_frozen tracks how many CPUs are involved in suspend
6950 * resume sequence. As long as this is not the last online
6951 * operation in the resume sequence, just build a single sched
6952 * domain, ignoring cpusets.
6955 if (likely(num_cpus_frozen
)) {
6956 partition_sched_domains(1, NULL
, NULL
);
6961 * This is the last CPU online operation. So fall through and
6962 * restore the original sched domains by considering the
6963 * cpuset configurations.
6967 cpuset_update_active_cpus(true);
6975 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
6978 unsigned long flags
;
6979 long cpu
= (long)hcpu
;
6985 case CPU_DOWN_PREPARE
:
6986 rcu_read_lock_sched();
6987 dl_b
= dl_bw_of(cpu
);
6989 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
6990 cpus
= dl_bw_cpus(cpu
);
6991 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
6992 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
6994 rcu_read_unlock_sched();
6997 return notifier_from_errno(-EBUSY
);
6998 cpuset_update_active_cpus(false);
7000 case CPU_DOWN_PREPARE_FROZEN
:
7002 partition_sched_domains(1, NULL
, NULL
);
7010 void __init
sched_init_smp(void)
7012 cpumask_var_t non_isolated_cpus
;
7014 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7015 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7020 * There's no userspace yet to cause hotplug operations; hence all the
7021 * cpu masks are stable and all blatant races in the below code cannot
7024 mutex_lock(&sched_domains_mutex
);
7025 init_sched_domains(cpu_active_mask
);
7026 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7027 if (cpumask_empty(non_isolated_cpus
))
7028 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7029 mutex_unlock(&sched_domains_mutex
);
7031 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7032 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7033 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7037 /* Move init over to a non-isolated CPU */
7038 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7040 sched_init_granularity();
7041 free_cpumask_var(non_isolated_cpus
);
7043 init_sched_rt_class();
7044 init_sched_dl_class();
7047 void __init
sched_init_smp(void)
7049 sched_init_granularity();
7051 #endif /* CONFIG_SMP */
7053 const_debug
unsigned int sysctl_timer_migration
= 1;
7055 int in_sched_functions(unsigned long addr
)
7057 return in_lock_functions(addr
) ||
7058 (addr
>= (unsigned long)__sched_text_start
7059 && addr
< (unsigned long)__sched_text_end
);
7062 #ifdef CONFIG_CGROUP_SCHED
7064 * Default task group.
7065 * Every task in system belongs to this group at bootup.
7067 struct task_group root_task_group
;
7068 LIST_HEAD(task_groups
);
7071 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7073 void __init
sched_init(void)
7076 unsigned long alloc_size
= 0, ptr
;
7078 #ifdef CONFIG_FAIR_GROUP_SCHED
7079 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7081 #ifdef CONFIG_RT_GROUP_SCHED
7082 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7085 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7087 #ifdef CONFIG_FAIR_GROUP_SCHED
7088 root_task_group
.se
= (struct sched_entity
**)ptr
;
7089 ptr
+= nr_cpu_ids
* sizeof(void **);
7091 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7092 ptr
+= nr_cpu_ids
* sizeof(void **);
7094 #endif /* CONFIG_FAIR_GROUP_SCHED */
7095 #ifdef CONFIG_RT_GROUP_SCHED
7096 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7097 ptr
+= nr_cpu_ids
* sizeof(void **);
7099 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7100 ptr
+= nr_cpu_ids
* sizeof(void **);
7102 #endif /* CONFIG_RT_GROUP_SCHED */
7104 #ifdef CONFIG_CPUMASK_OFFSTACK
7105 for_each_possible_cpu(i
) {
7106 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7107 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7109 #endif /* CONFIG_CPUMASK_OFFSTACK */
7111 init_rt_bandwidth(&def_rt_bandwidth
,
7112 global_rt_period(), global_rt_runtime());
7113 init_dl_bandwidth(&def_dl_bandwidth
,
7114 global_rt_period(), global_rt_runtime());
7117 init_defrootdomain();
7120 #ifdef CONFIG_RT_GROUP_SCHED
7121 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7122 global_rt_period(), global_rt_runtime());
7123 #endif /* CONFIG_RT_GROUP_SCHED */
7125 #ifdef CONFIG_CGROUP_SCHED
7126 list_add(&root_task_group
.list
, &task_groups
);
7127 INIT_LIST_HEAD(&root_task_group
.children
);
7128 INIT_LIST_HEAD(&root_task_group
.siblings
);
7129 autogroup_init(&init_task
);
7131 #endif /* CONFIG_CGROUP_SCHED */
7133 for_each_possible_cpu(i
) {
7137 raw_spin_lock_init(&rq
->lock
);
7139 rq
->calc_load_active
= 0;
7140 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7141 init_cfs_rq(&rq
->cfs
);
7142 init_rt_rq(&rq
->rt
);
7143 init_dl_rq(&rq
->dl
);
7144 #ifdef CONFIG_FAIR_GROUP_SCHED
7145 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7146 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7148 * How much cpu bandwidth does root_task_group get?
7150 * In case of task-groups formed thr' the cgroup filesystem, it
7151 * gets 100% of the cpu resources in the system. This overall
7152 * system cpu resource is divided among the tasks of
7153 * root_task_group and its child task-groups in a fair manner,
7154 * based on each entity's (task or task-group's) weight
7155 * (se->load.weight).
7157 * In other words, if root_task_group has 10 tasks of weight
7158 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7159 * then A0's share of the cpu resource is:
7161 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7163 * We achieve this by letting root_task_group's tasks sit
7164 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7166 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7167 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7168 #endif /* CONFIG_FAIR_GROUP_SCHED */
7170 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7171 #ifdef CONFIG_RT_GROUP_SCHED
7172 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7175 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7176 rq
->cpu_load
[j
] = 0;
7178 rq
->last_load_update_tick
= jiffies
;
7183 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7184 rq
->post_schedule
= 0;
7185 rq
->active_balance
= 0;
7186 rq
->next_balance
= jiffies
;
7191 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7192 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7194 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7196 rq_attach_root(rq
, &def_root_domain
);
7197 #ifdef CONFIG_NO_HZ_COMMON
7200 #ifdef CONFIG_NO_HZ_FULL
7201 rq
->last_sched_tick
= 0;
7205 atomic_set(&rq
->nr_iowait
, 0);
7208 set_load_weight(&init_task
);
7210 #ifdef CONFIG_PREEMPT_NOTIFIERS
7211 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7215 * The boot idle thread does lazy MMU switching as well:
7217 atomic_inc(&init_mm
.mm_count
);
7218 enter_lazy_tlb(&init_mm
, current
);
7221 * During early bootup we pretend to be a normal task:
7223 current
->sched_class
= &fair_sched_class
;
7226 * Make us the idle thread. Technically, schedule() should not be
7227 * called from this thread, however somewhere below it might be,
7228 * but because we are the idle thread, we just pick up running again
7229 * when this runqueue becomes "idle".
7231 init_idle(current
, smp_processor_id());
7233 calc_load_update
= jiffies
+ LOAD_FREQ
;
7236 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7237 /* May be allocated at isolcpus cmdline parse time */
7238 if (cpu_isolated_map
== NULL
)
7239 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7240 idle_thread_set_boot_cpu();
7241 set_cpu_rq_start_time();
7243 init_sched_fair_class();
7245 scheduler_running
= 1;
7248 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7249 static inline int preempt_count_equals(int preempt_offset
)
7251 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7253 return (nested
== preempt_offset
);
7256 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7259 * Blocking primitives will set (and therefore destroy) current->state,
7260 * since we will exit with TASK_RUNNING make sure we enter with it,
7261 * otherwise we will destroy state.
7263 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7264 "do not call blocking ops when !TASK_RUNNING; "
7265 "state=%lx set at [<%p>] %pS\n",
7267 (void *)current
->task_state_change
,
7268 (void *)current
->task_state_change
);
7270 ___might_sleep(file
, line
, preempt_offset
);
7272 EXPORT_SYMBOL(__might_sleep
);
7274 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7276 static unsigned long prev_jiffy
; /* ratelimiting */
7278 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7279 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7280 !is_idle_task(current
)) ||
7281 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7283 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7285 prev_jiffy
= jiffies
;
7288 "BUG: sleeping function called from invalid context at %s:%d\n",
7291 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7292 in_atomic(), irqs_disabled(),
7293 current
->pid
, current
->comm
);
7295 if (task_stack_end_corrupted(current
))
7296 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7298 debug_show_held_locks(current
);
7299 if (irqs_disabled())
7300 print_irqtrace_events(current
);
7301 #ifdef CONFIG_DEBUG_PREEMPT
7302 if (!preempt_count_equals(preempt_offset
)) {
7303 pr_err("Preemption disabled at:");
7304 print_ip_sym(current
->preempt_disable_ip
);
7310 EXPORT_SYMBOL(___might_sleep
);
7313 #ifdef CONFIG_MAGIC_SYSRQ
7314 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7316 const struct sched_class
*prev_class
= p
->sched_class
;
7317 struct sched_attr attr
= {
7318 .sched_policy
= SCHED_NORMAL
,
7320 int old_prio
= p
->prio
;
7323 queued
= task_on_rq_queued(p
);
7325 dequeue_task(rq
, p
, 0);
7326 __setscheduler(rq
, p
, &attr
, false);
7328 enqueue_task(rq
, p
, 0);
7332 check_class_changed(rq
, p
, prev_class
, old_prio
);
7335 void normalize_rt_tasks(void)
7337 struct task_struct
*g
, *p
;
7338 unsigned long flags
;
7341 read_lock(&tasklist_lock
);
7342 for_each_process_thread(g
, p
) {
7344 * Only normalize user tasks:
7346 if (p
->flags
& PF_KTHREAD
)
7349 p
->se
.exec_start
= 0;
7350 #ifdef CONFIG_SCHEDSTATS
7351 p
->se
.statistics
.wait_start
= 0;
7352 p
->se
.statistics
.sleep_start
= 0;
7353 p
->se
.statistics
.block_start
= 0;
7356 if (!dl_task(p
) && !rt_task(p
)) {
7358 * Renice negative nice level userspace
7361 if (task_nice(p
) < 0)
7362 set_user_nice(p
, 0);
7366 rq
= task_rq_lock(p
, &flags
);
7367 normalize_task(rq
, p
);
7368 task_rq_unlock(rq
, p
, &flags
);
7370 read_unlock(&tasklist_lock
);
7373 #endif /* CONFIG_MAGIC_SYSRQ */
7375 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7377 * These functions are only useful for the IA64 MCA handling, or kdb.
7379 * They can only be called when the whole system has been
7380 * stopped - every CPU needs to be quiescent, and no scheduling
7381 * activity can take place. Using them for anything else would
7382 * be a serious bug, and as a result, they aren't even visible
7383 * under any other configuration.
7387 * curr_task - return the current task for a given cpu.
7388 * @cpu: the processor in question.
7390 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7392 * Return: The current task for @cpu.
7394 struct task_struct
*curr_task(int cpu
)
7396 return cpu_curr(cpu
);
7399 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7403 * set_curr_task - set the current task for a given cpu.
7404 * @cpu: the processor in question.
7405 * @p: the task pointer to set.
7407 * Description: This function must only be used when non-maskable interrupts
7408 * are serviced on a separate stack. It allows the architecture to switch the
7409 * notion of the current task on a cpu in a non-blocking manner. This function
7410 * must be called with all CPU's synchronized, and interrupts disabled, the
7411 * and caller must save the original value of the current task (see
7412 * curr_task() above) and restore that value before reenabling interrupts and
7413 * re-starting the system.
7415 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7417 void set_curr_task(int cpu
, struct task_struct
*p
)
7424 #ifdef CONFIG_CGROUP_SCHED
7425 /* task_group_lock serializes the addition/removal of task groups */
7426 static DEFINE_SPINLOCK(task_group_lock
);
7428 static void free_sched_group(struct task_group
*tg
)
7430 free_fair_sched_group(tg
);
7431 free_rt_sched_group(tg
);
7436 /* allocate runqueue etc for a new task group */
7437 struct task_group
*sched_create_group(struct task_group
*parent
)
7439 struct task_group
*tg
;
7441 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7443 return ERR_PTR(-ENOMEM
);
7445 if (!alloc_fair_sched_group(tg
, parent
))
7448 if (!alloc_rt_sched_group(tg
, parent
))
7454 free_sched_group(tg
);
7455 return ERR_PTR(-ENOMEM
);
7458 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7460 unsigned long flags
;
7462 spin_lock_irqsave(&task_group_lock
, flags
);
7463 list_add_rcu(&tg
->list
, &task_groups
);
7465 WARN_ON(!parent
); /* root should already exist */
7467 tg
->parent
= parent
;
7468 INIT_LIST_HEAD(&tg
->children
);
7469 list_add_rcu(&tg
->siblings
, &parent
->children
);
7470 spin_unlock_irqrestore(&task_group_lock
, flags
);
7473 /* rcu callback to free various structures associated with a task group */
7474 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7476 /* now it should be safe to free those cfs_rqs */
7477 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7480 /* Destroy runqueue etc associated with a task group */
7481 void sched_destroy_group(struct task_group
*tg
)
7483 /* wait for possible concurrent references to cfs_rqs complete */
7484 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7487 void sched_offline_group(struct task_group
*tg
)
7489 unsigned long flags
;
7492 /* end participation in shares distribution */
7493 for_each_possible_cpu(i
)
7494 unregister_fair_sched_group(tg
, i
);
7496 spin_lock_irqsave(&task_group_lock
, flags
);
7497 list_del_rcu(&tg
->list
);
7498 list_del_rcu(&tg
->siblings
);
7499 spin_unlock_irqrestore(&task_group_lock
, flags
);
7502 /* change task's runqueue when it moves between groups.
7503 * The caller of this function should have put the task in its new group
7504 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7505 * reflect its new group.
7507 void sched_move_task(struct task_struct
*tsk
)
7509 struct task_group
*tg
;
7510 int queued
, running
;
7511 unsigned long flags
;
7514 rq
= task_rq_lock(tsk
, &flags
);
7516 running
= task_current(rq
, tsk
);
7517 queued
= task_on_rq_queued(tsk
);
7520 dequeue_task(rq
, tsk
, 0);
7521 if (unlikely(running
))
7522 put_prev_task(rq
, tsk
);
7525 * All callers are synchronized by task_rq_lock(); we do not use RCU
7526 * which is pointless here. Thus, we pass "true" to task_css_check()
7527 * to prevent lockdep warnings.
7529 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7530 struct task_group
, css
);
7531 tg
= autogroup_task_group(tsk
, tg
);
7532 tsk
->sched_task_group
= tg
;
7534 #ifdef CONFIG_FAIR_GROUP_SCHED
7535 if (tsk
->sched_class
->task_move_group
)
7536 tsk
->sched_class
->task_move_group(tsk
, queued
);
7539 set_task_rq(tsk
, task_cpu(tsk
));
7541 if (unlikely(running
))
7542 tsk
->sched_class
->set_curr_task(rq
);
7544 enqueue_task(rq
, tsk
, 0);
7546 task_rq_unlock(rq
, tsk
, &flags
);
7548 #endif /* CONFIG_CGROUP_SCHED */
7550 #ifdef CONFIG_RT_GROUP_SCHED
7552 * Ensure that the real time constraints are schedulable.
7554 static DEFINE_MUTEX(rt_constraints_mutex
);
7556 /* Must be called with tasklist_lock held */
7557 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7559 struct task_struct
*g
, *p
;
7562 * Autogroups do not have RT tasks; see autogroup_create().
7564 if (task_group_is_autogroup(tg
))
7567 for_each_process_thread(g
, p
) {
7568 if (rt_task(p
) && task_group(p
) == tg
)
7575 struct rt_schedulable_data
{
7576 struct task_group
*tg
;
7581 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7583 struct rt_schedulable_data
*d
= data
;
7584 struct task_group
*child
;
7585 unsigned long total
, sum
= 0;
7586 u64 period
, runtime
;
7588 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7589 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7592 period
= d
->rt_period
;
7593 runtime
= d
->rt_runtime
;
7597 * Cannot have more runtime than the period.
7599 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7603 * Ensure we don't starve existing RT tasks.
7605 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7608 total
= to_ratio(period
, runtime
);
7611 * Nobody can have more than the global setting allows.
7613 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7617 * The sum of our children's runtime should not exceed our own.
7619 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7620 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7621 runtime
= child
->rt_bandwidth
.rt_runtime
;
7623 if (child
== d
->tg
) {
7624 period
= d
->rt_period
;
7625 runtime
= d
->rt_runtime
;
7628 sum
+= to_ratio(period
, runtime
);
7637 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7641 struct rt_schedulable_data data
= {
7643 .rt_period
= period
,
7644 .rt_runtime
= runtime
,
7648 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7654 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7655 u64 rt_period
, u64 rt_runtime
)
7660 * Disallowing the root group RT runtime is BAD, it would disallow the
7661 * kernel creating (and or operating) RT threads.
7663 if (tg
== &root_task_group
&& rt_runtime
== 0)
7666 /* No period doesn't make any sense. */
7670 mutex_lock(&rt_constraints_mutex
);
7671 read_lock(&tasklist_lock
);
7672 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7676 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7677 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7678 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7680 for_each_possible_cpu(i
) {
7681 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7683 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7684 rt_rq
->rt_runtime
= rt_runtime
;
7685 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7687 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7689 read_unlock(&tasklist_lock
);
7690 mutex_unlock(&rt_constraints_mutex
);
7695 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7697 u64 rt_runtime
, rt_period
;
7699 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7700 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7701 if (rt_runtime_us
< 0)
7702 rt_runtime
= RUNTIME_INF
;
7704 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7707 static long sched_group_rt_runtime(struct task_group
*tg
)
7711 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7714 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7715 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7716 return rt_runtime_us
;
7719 static int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
7721 u64 rt_runtime
, rt_period
;
7723 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
7724 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7726 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7729 static long sched_group_rt_period(struct task_group
*tg
)
7733 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7734 do_div(rt_period_us
, NSEC_PER_USEC
);
7735 return rt_period_us
;
7737 #endif /* CONFIG_RT_GROUP_SCHED */
7739 #ifdef CONFIG_RT_GROUP_SCHED
7740 static int sched_rt_global_constraints(void)
7744 mutex_lock(&rt_constraints_mutex
);
7745 read_lock(&tasklist_lock
);
7746 ret
= __rt_schedulable(NULL
, 0, 0);
7747 read_unlock(&tasklist_lock
);
7748 mutex_unlock(&rt_constraints_mutex
);
7753 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7755 /* Don't accept realtime tasks when there is no way for them to run */
7756 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7762 #else /* !CONFIG_RT_GROUP_SCHED */
7763 static int sched_rt_global_constraints(void)
7765 unsigned long flags
;
7768 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7769 for_each_possible_cpu(i
) {
7770 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7772 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7773 rt_rq
->rt_runtime
= global_rt_runtime();
7774 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7776 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7780 #endif /* CONFIG_RT_GROUP_SCHED */
7782 static int sched_dl_global_validate(void)
7784 u64 runtime
= global_rt_runtime();
7785 u64 period
= global_rt_period();
7786 u64 new_bw
= to_ratio(period
, runtime
);
7789 unsigned long flags
;
7792 * Here we want to check the bandwidth not being set to some
7793 * value smaller than the currently allocated bandwidth in
7794 * any of the root_domains.
7796 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7797 * cycling on root_domains... Discussion on different/better
7798 * solutions is welcome!
7800 for_each_possible_cpu(cpu
) {
7801 rcu_read_lock_sched();
7802 dl_b
= dl_bw_of(cpu
);
7804 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7805 if (new_bw
< dl_b
->total_bw
)
7807 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7809 rcu_read_unlock_sched();
7818 static void sched_dl_do_global(void)
7823 unsigned long flags
;
7825 def_dl_bandwidth
.dl_period
= global_rt_period();
7826 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7828 if (global_rt_runtime() != RUNTIME_INF
)
7829 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7832 * FIXME: As above...
7834 for_each_possible_cpu(cpu
) {
7835 rcu_read_lock_sched();
7836 dl_b
= dl_bw_of(cpu
);
7838 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7840 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7842 rcu_read_unlock_sched();
7846 static int sched_rt_global_validate(void)
7848 if (sysctl_sched_rt_period
<= 0)
7851 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7852 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7858 static void sched_rt_do_global(void)
7860 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7861 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7864 int sched_rt_handler(struct ctl_table
*table
, int write
,
7865 void __user
*buffer
, size_t *lenp
,
7868 int old_period
, old_runtime
;
7869 static DEFINE_MUTEX(mutex
);
7873 old_period
= sysctl_sched_rt_period
;
7874 old_runtime
= sysctl_sched_rt_runtime
;
7876 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7878 if (!ret
&& write
) {
7879 ret
= sched_rt_global_validate();
7883 ret
= sched_dl_global_validate();
7887 ret
= sched_rt_global_constraints();
7891 sched_rt_do_global();
7892 sched_dl_do_global();
7896 sysctl_sched_rt_period
= old_period
;
7897 sysctl_sched_rt_runtime
= old_runtime
;
7899 mutex_unlock(&mutex
);
7904 int sched_rr_handler(struct ctl_table
*table
, int write
,
7905 void __user
*buffer
, size_t *lenp
,
7909 static DEFINE_MUTEX(mutex
);
7912 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7913 /* make sure that internally we keep jiffies */
7914 /* also, writing zero resets timeslice to default */
7915 if (!ret
&& write
) {
7916 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7917 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7919 mutex_unlock(&mutex
);
7923 #ifdef CONFIG_CGROUP_SCHED
7925 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7927 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7930 static struct cgroup_subsys_state
*
7931 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7933 struct task_group
*parent
= css_tg(parent_css
);
7934 struct task_group
*tg
;
7937 /* This is early initialization for the top cgroup */
7938 return &root_task_group
.css
;
7941 tg
= sched_create_group(parent
);
7943 return ERR_PTR(-ENOMEM
);
7948 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7950 struct task_group
*tg
= css_tg(css
);
7951 struct task_group
*parent
= css_tg(css
->parent
);
7954 sched_online_group(tg
, parent
);
7958 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7960 struct task_group
*tg
= css_tg(css
);
7962 sched_destroy_group(tg
);
7965 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
7967 struct task_group
*tg
= css_tg(css
);
7969 sched_offline_group(tg
);
7972 static void cpu_cgroup_fork(struct task_struct
*task
)
7974 sched_move_task(task
);
7977 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7978 struct cgroup_taskset
*tset
)
7980 struct task_struct
*task
;
7982 cgroup_taskset_for_each(task
, tset
) {
7983 #ifdef CONFIG_RT_GROUP_SCHED
7984 if (!sched_rt_can_attach(css_tg(css
), task
))
7987 /* We don't support RT-tasks being in separate groups */
7988 if (task
->sched_class
!= &fair_sched_class
)
7995 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
7996 struct cgroup_taskset
*tset
)
7998 struct task_struct
*task
;
8000 cgroup_taskset_for_each(task
, tset
)
8001 sched_move_task(task
);
8004 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8005 struct cgroup_subsys_state
*old_css
,
8006 struct task_struct
*task
)
8009 * cgroup_exit() is called in the copy_process() failure path.
8010 * Ignore this case since the task hasn't ran yet, this avoids
8011 * trying to poke a half freed task state from generic code.
8013 if (!(task
->flags
& PF_EXITING
))
8016 sched_move_task(task
);
8019 #ifdef CONFIG_FAIR_GROUP_SCHED
8020 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8021 struct cftype
*cftype
, u64 shareval
)
8023 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8026 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8029 struct task_group
*tg
= css_tg(css
);
8031 return (u64
) scale_load_down(tg
->shares
);
8034 #ifdef CONFIG_CFS_BANDWIDTH
8035 static DEFINE_MUTEX(cfs_constraints_mutex
);
8037 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8038 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8040 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8042 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8044 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8045 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8047 if (tg
== &root_task_group
)
8051 * Ensure we have at some amount of bandwidth every period. This is
8052 * to prevent reaching a state of large arrears when throttled via
8053 * entity_tick() resulting in prolonged exit starvation.
8055 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8059 * Likewise, bound things on the otherside by preventing insane quota
8060 * periods. This also allows us to normalize in computing quota
8063 if (period
> max_cfs_quota_period
)
8067 * Prevent race between setting of cfs_rq->runtime_enabled and
8068 * unthrottle_offline_cfs_rqs().
8071 mutex_lock(&cfs_constraints_mutex
);
8072 ret
= __cfs_schedulable(tg
, period
, quota
);
8076 runtime_enabled
= quota
!= RUNTIME_INF
;
8077 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8079 * If we need to toggle cfs_bandwidth_used, off->on must occur
8080 * before making related changes, and on->off must occur afterwards
8082 if (runtime_enabled
&& !runtime_was_enabled
)
8083 cfs_bandwidth_usage_inc();
8084 raw_spin_lock_irq(&cfs_b
->lock
);
8085 cfs_b
->period
= ns_to_ktime(period
);
8086 cfs_b
->quota
= quota
;
8088 __refill_cfs_bandwidth_runtime(cfs_b
);
8089 /* restart the period timer (if active) to handle new period expiry */
8090 if (runtime_enabled
)
8091 start_cfs_bandwidth(cfs_b
);
8092 raw_spin_unlock_irq(&cfs_b
->lock
);
8094 for_each_online_cpu(i
) {
8095 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8096 struct rq
*rq
= cfs_rq
->rq
;
8098 raw_spin_lock_irq(&rq
->lock
);
8099 cfs_rq
->runtime_enabled
= runtime_enabled
;
8100 cfs_rq
->runtime_remaining
= 0;
8102 if (cfs_rq
->throttled
)
8103 unthrottle_cfs_rq(cfs_rq
);
8104 raw_spin_unlock_irq(&rq
->lock
);
8106 if (runtime_was_enabled
&& !runtime_enabled
)
8107 cfs_bandwidth_usage_dec();
8109 mutex_unlock(&cfs_constraints_mutex
);
8115 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8119 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8120 if (cfs_quota_us
< 0)
8121 quota
= RUNTIME_INF
;
8123 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8125 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8128 long tg_get_cfs_quota(struct task_group
*tg
)
8132 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8135 quota_us
= tg
->cfs_bandwidth
.quota
;
8136 do_div(quota_us
, NSEC_PER_USEC
);
8141 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8145 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8146 quota
= tg
->cfs_bandwidth
.quota
;
8148 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8151 long tg_get_cfs_period(struct task_group
*tg
)
8155 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8156 do_div(cfs_period_us
, NSEC_PER_USEC
);
8158 return cfs_period_us
;
8161 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8164 return tg_get_cfs_quota(css_tg(css
));
8167 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8168 struct cftype
*cftype
, s64 cfs_quota_us
)
8170 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8173 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8176 return tg_get_cfs_period(css_tg(css
));
8179 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8180 struct cftype
*cftype
, u64 cfs_period_us
)
8182 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8185 struct cfs_schedulable_data
{
8186 struct task_group
*tg
;
8191 * normalize group quota/period to be quota/max_period
8192 * note: units are usecs
8194 static u64
normalize_cfs_quota(struct task_group
*tg
,
8195 struct cfs_schedulable_data
*d
)
8203 period
= tg_get_cfs_period(tg
);
8204 quota
= tg_get_cfs_quota(tg
);
8207 /* note: these should typically be equivalent */
8208 if (quota
== RUNTIME_INF
|| quota
== -1)
8211 return to_ratio(period
, quota
);
8214 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8216 struct cfs_schedulable_data
*d
= data
;
8217 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8218 s64 quota
= 0, parent_quota
= -1;
8221 quota
= RUNTIME_INF
;
8223 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8225 quota
= normalize_cfs_quota(tg
, d
);
8226 parent_quota
= parent_b
->hierarchical_quota
;
8229 * ensure max(child_quota) <= parent_quota, inherit when no
8232 if (quota
== RUNTIME_INF
)
8233 quota
= parent_quota
;
8234 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8237 cfs_b
->hierarchical_quota
= quota
;
8242 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8245 struct cfs_schedulable_data data
= {
8251 if (quota
!= RUNTIME_INF
) {
8252 do_div(data
.period
, NSEC_PER_USEC
);
8253 do_div(data
.quota
, NSEC_PER_USEC
);
8257 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8263 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8265 struct task_group
*tg
= css_tg(seq_css(sf
));
8266 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8268 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8269 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8270 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8274 #endif /* CONFIG_CFS_BANDWIDTH */
8275 #endif /* CONFIG_FAIR_GROUP_SCHED */
8277 #ifdef CONFIG_RT_GROUP_SCHED
8278 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8279 struct cftype
*cft
, s64 val
)
8281 return sched_group_set_rt_runtime(css_tg(css
), val
);
8284 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8287 return sched_group_rt_runtime(css_tg(css
));
8290 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8291 struct cftype
*cftype
, u64 rt_period_us
)
8293 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8296 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8299 return sched_group_rt_period(css_tg(css
));
8301 #endif /* CONFIG_RT_GROUP_SCHED */
8303 static struct cftype cpu_files
[] = {
8304 #ifdef CONFIG_FAIR_GROUP_SCHED
8307 .read_u64
= cpu_shares_read_u64
,
8308 .write_u64
= cpu_shares_write_u64
,
8311 #ifdef CONFIG_CFS_BANDWIDTH
8313 .name
= "cfs_quota_us",
8314 .read_s64
= cpu_cfs_quota_read_s64
,
8315 .write_s64
= cpu_cfs_quota_write_s64
,
8318 .name
= "cfs_period_us",
8319 .read_u64
= cpu_cfs_period_read_u64
,
8320 .write_u64
= cpu_cfs_period_write_u64
,
8324 .seq_show
= cpu_stats_show
,
8327 #ifdef CONFIG_RT_GROUP_SCHED
8329 .name
= "rt_runtime_us",
8330 .read_s64
= cpu_rt_runtime_read
,
8331 .write_s64
= cpu_rt_runtime_write
,
8334 .name
= "rt_period_us",
8335 .read_u64
= cpu_rt_period_read_uint
,
8336 .write_u64
= cpu_rt_period_write_uint
,
8342 struct cgroup_subsys cpu_cgrp_subsys
= {
8343 .css_alloc
= cpu_cgroup_css_alloc
,
8344 .css_free
= cpu_cgroup_css_free
,
8345 .css_online
= cpu_cgroup_css_online
,
8346 .css_offline
= cpu_cgroup_css_offline
,
8347 .fork
= cpu_cgroup_fork
,
8348 .can_attach
= cpu_cgroup_can_attach
,
8349 .attach
= cpu_cgroup_attach
,
8350 .exit
= cpu_cgroup_exit
,
8351 .legacy_cftypes
= cpu_files
,
8355 #endif /* CONFIG_CGROUP_SCHED */
8357 void dump_cpu_task(int cpu
)
8359 pr_info("Task dump for CPU %d:\n", cpu
);
8360 sched_show_task(cpu_curr(cpu
));