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
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.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/context_tracking.h>
75 #include <linux/compiler.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 DEFINE_MUTEX(sched_domains_mutex
);
93 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
95 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
97 void update_rq_clock(struct rq
*rq
)
101 lockdep_assert_held(&rq
->lock
);
103 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
106 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
110 update_rq_clock_task(rq
, delta
);
114 * Debugging: various feature bits
117 #define SCHED_FEAT(name, enabled) \
118 (1UL << __SCHED_FEAT_##name) * enabled |
120 const_debug
unsigned int sysctl_sched_features
=
121 #include "features.h"
127 * Number of tasks to iterate in a single balance run.
128 * Limited because this is done with IRQs disabled.
130 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
133 * period over which we average the RT time consumption, measured
138 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
141 * period over which we measure -rt task cpu usage in us.
144 unsigned int sysctl_sched_rt_period
= 1000000;
146 __read_mostly
int scheduler_running
;
149 * part of the period that we allow rt tasks to run in us.
152 int sysctl_sched_rt_runtime
= 950000;
154 /* cpus with isolated domains */
155 cpumask_var_t cpu_isolated_map
;
158 * this_rq_lock - lock this runqueue and disable interrupts.
160 static struct rq
*this_rq_lock(void)
167 raw_spin_lock(&rq
->lock
);
172 #ifdef CONFIG_SCHED_HRTICK
174 * Use HR-timers to deliver accurate preemption points.
177 static void hrtick_clear(struct rq
*rq
)
179 if (hrtimer_active(&rq
->hrtick_timer
))
180 hrtimer_cancel(&rq
->hrtick_timer
);
184 * High-resolution timer tick.
185 * Runs from hardirq context with interrupts disabled.
187 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
189 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
191 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
193 raw_spin_lock(&rq
->lock
);
195 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
196 raw_spin_unlock(&rq
->lock
);
198 return HRTIMER_NORESTART
;
203 static void __hrtick_restart(struct rq
*rq
)
205 struct hrtimer
*timer
= &rq
->hrtick_timer
;
207 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
211 * called from hardirq (IPI) context
213 static void __hrtick_start(void *arg
)
217 raw_spin_lock(&rq
->lock
);
218 __hrtick_restart(rq
);
219 rq
->hrtick_csd_pending
= 0;
220 raw_spin_unlock(&rq
->lock
);
224 * Called to set the hrtick timer state.
226 * called with rq->lock held and irqs disabled
228 void hrtick_start(struct rq
*rq
, u64 delay
)
230 struct hrtimer
*timer
= &rq
->hrtick_timer
;
235 * Don't schedule slices shorter than 10000ns, that just
236 * doesn't make sense and can cause timer DoS.
238 delta
= max_t(s64
, delay
, 10000LL);
239 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
241 hrtimer_set_expires(timer
, time
);
243 if (rq
== this_rq()) {
244 __hrtick_restart(rq
);
245 } else if (!rq
->hrtick_csd_pending
) {
246 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
247 rq
->hrtick_csd_pending
= 1;
252 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
254 int cpu
= (int)(long)hcpu
;
257 case CPU_UP_CANCELED
:
258 case CPU_UP_CANCELED_FROZEN
:
259 case CPU_DOWN_PREPARE
:
260 case CPU_DOWN_PREPARE_FROZEN
:
262 case CPU_DEAD_FROZEN
:
263 hrtick_clear(cpu_rq(cpu
));
270 static __init
void init_hrtick(void)
272 hotcpu_notifier(hotplug_hrtick
, 0);
276 * Called to set the hrtick timer state.
278 * called with rq->lock held and irqs disabled
280 void hrtick_start(struct rq
*rq
, u64 delay
)
283 * Don't schedule slices shorter than 10000ns, that just
284 * doesn't make sense. Rely on vruntime for fairness.
286 delay
= max_t(u64
, delay
, 10000LL);
287 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
288 HRTIMER_MODE_REL_PINNED
);
291 static inline void init_hrtick(void)
294 #endif /* CONFIG_SMP */
296 static void init_rq_hrtick(struct rq
*rq
)
299 rq
->hrtick_csd_pending
= 0;
301 rq
->hrtick_csd
.flags
= 0;
302 rq
->hrtick_csd
.func
= __hrtick_start
;
303 rq
->hrtick_csd
.info
= rq
;
306 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
307 rq
->hrtick_timer
.function
= hrtick
;
309 #else /* CONFIG_SCHED_HRTICK */
310 static inline void hrtick_clear(struct rq
*rq
)
314 static inline void init_rq_hrtick(struct rq
*rq
)
318 static inline void init_hrtick(void)
321 #endif /* CONFIG_SCHED_HRTICK */
323 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
325 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
326 * this avoids any races wrt polling state changes and thereby avoids
329 static bool set_nr_and_not_polling(struct task_struct
*p
)
331 struct thread_info
*ti
= task_thread_info(p
);
332 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
336 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
338 * If this returns true, then the idle task promises to call
339 * sched_ttwu_pending() and reschedule soon.
341 static bool set_nr_if_polling(struct task_struct
*p
)
343 struct thread_info
*ti
= task_thread_info(p
);
344 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
347 if (!(val
& _TIF_POLLING_NRFLAG
))
349 if (val
& _TIF_NEED_RESCHED
)
351 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
360 static bool set_nr_and_not_polling(struct task_struct
*p
)
362 set_tsk_need_resched(p
);
367 static bool set_nr_if_polling(struct task_struct
*p
)
374 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
376 struct wake_q_node
*node
= &task
->wake_q
;
379 * Atomically grab the task, if ->wake_q is !nil already it means
380 * its already queued (either by us or someone else) and will get the
381 * wakeup due to that.
383 * This cmpxchg() implies a full barrier, which pairs with the write
384 * barrier implied by the wakeup in wake_up_list().
386 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
389 get_task_struct(task
);
392 * The head is context local, there can be no concurrency.
395 head
->lastp
= &node
->next
;
398 void wake_up_q(struct wake_q_head
*head
)
400 struct wake_q_node
*node
= head
->first
;
402 while (node
!= WAKE_Q_TAIL
) {
403 struct task_struct
*task
;
405 task
= container_of(node
, struct task_struct
, wake_q
);
407 /* task can safely be re-inserted now */
409 task
->wake_q
.next
= NULL
;
412 * wake_up_process() implies a wmb() to pair with the queueing
413 * in wake_q_add() so as not to miss wakeups.
415 wake_up_process(task
);
416 put_task_struct(task
);
421 * resched_curr - mark rq's current task 'to be rescheduled now'.
423 * On UP this means the setting of the need_resched flag, on SMP it
424 * might also involve a cross-CPU call to trigger the scheduler on
427 void resched_curr(struct rq
*rq
)
429 struct task_struct
*curr
= rq
->curr
;
432 lockdep_assert_held(&rq
->lock
);
434 if (test_tsk_need_resched(curr
))
439 if (cpu
== smp_processor_id()) {
440 set_tsk_need_resched(curr
);
441 set_preempt_need_resched();
445 if (set_nr_and_not_polling(curr
))
446 smp_send_reschedule(cpu
);
448 trace_sched_wake_idle_without_ipi(cpu
);
451 void resched_cpu(int cpu
)
453 struct rq
*rq
= cpu_rq(cpu
);
456 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
459 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
463 #ifdef CONFIG_NO_HZ_COMMON
465 * In the semi idle case, use the nearest busy cpu for migrating timers
466 * from an idle cpu. This is good for power-savings.
468 * We don't do similar optimization for completely idle system, as
469 * selecting an idle cpu will add more delays to the timers than intended
470 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
472 int get_nohz_timer_target(void)
474 int i
, cpu
= smp_processor_id();
475 struct sched_domain
*sd
;
477 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
481 for_each_domain(cpu
, sd
) {
482 for_each_cpu(i
, sched_domain_span(sd
)) {
483 if (!idle_cpu(i
) && is_housekeeping_cpu(cpu
)) {
490 if (!is_housekeeping_cpu(cpu
))
491 cpu
= housekeeping_any_cpu();
497 * When add_timer_on() enqueues a timer into the timer wheel of an
498 * idle CPU then this timer might expire before the next timer event
499 * which is scheduled to wake up that CPU. In case of a completely
500 * idle system the next event might even be infinite time into the
501 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
502 * leaves the inner idle loop so the newly added timer is taken into
503 * account when the CPU goes back to idle and evaluates the timer
504 * wheel for the next timer event.
506 static void wake_up_idle_cpu(int cpu
)
508 struct rq
*rq
= cpu_rq(cpu
);
510 if (cpu
== smp_processor_id())
513 if (set_nr_and_not_polling(rq
->idle
))
514 smp_send_reschedule(cpu
);
516 trace_sched_wake_idle_without_ipi(cpu
);
519 static bool wake_up_full_nohz_cpu(int cpu
)
522 * We just need the target to call irq_exit() and re-evaluate
523 * the next tick. The nohz full kick at least implies that.
524 * If needed we can still optimize that later with an
527 if (tick_nohz_full_cpu(cpu
)) {
528 if (cpu
!= smp_processor_id() ||
529 tick_nohz_tick_stopped())
530 tick_nohz_full_kick_cpu(cpu
);
537 void wake_up_nohz_cpu(int cpu
)
539 if (!wake_up_full_nohz_cpu(cpu
))
540 wake_up_idle_cpu(cpu
);
543 static inline bool got_nohz_idle_kick(void)
545 int cpu
= smp_processor_id();
547 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
550 if (idle_cpu(cpu
) && !need_resched())
554 * We can't run Idle Load Balance on this CPU for this time so we
555 * cancel it and clear NOHZ_BALANCE_KICK
557 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
561 #else /* CONFIG_NO_HZ_COMMON */
563 static inline bool got_nohz_idle_kick(void)
568 #endif /* CONFIG_NO_HZ_COMMON */
570 #ifdef CONFIG_NO_HZ_FULL
571 bool sched_can_stop_tick(struct rq
*rq
)
575 /* Deadline tasks, even if single, need the tick */
576 if (rq
->dl
.dl_nr_running
)
580 * FIFO realtime policy runs the highest priority task (after DEADLINE).
581 * Other runnable tasks are of a lower priority. The scheduler tick
584 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
589 * Round-robin realtime tasks time slice with other tasks at the same
592 if (rq
->rt
.rr_nr_running
) {
593 if (rq
->rt
.rr_nr_running
== 1)
599 /* Normal multitasking need periodic preemption checks */
600 if (rq
->cfs
.nr_running
> 1)
605 #endif /* CONFIG_NO_HZ_FULL */
607 void sched_avg_update(struct rq
*rq
)
609 s64 period
= sched_avg_period();
611 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
613 * Inline assembly required to prevent the compiler
614 * optimising this loop into a divmod call.
615 * See __iter_div_u64_rem() for another example of this.
617 asm("" : "+rm" (rq
->age_stamp
));
618 rq
->age_stamp
+= period
;
623 #endif /* CONFIG_SMP */
625 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
626 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
628 * Iterate task_group tree rooted at *from, calling @down when first entering a
629 * node and @up when leaving it for the final time.
631 * Caller must hold rcu_lock or sufficient equivalent.
633 int walk_tg_tree_from(struct task_group
*from
,
634 tg_visitor down
, tg_visitor up
, void *data
)
636 struct task_group
*parent
, *child
;
642 ret
= (*down
)(parent
, data
);
645 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
652 ret
= (*up
)(parent
, data
);
653 if (ret
|| parent
== from
)
657 parent
= parent
->parent
;
664 int tg_nop(struct task_group
*tg
, void *data
)
670 static void set_load_weight(struct task_struct
*p
)
672 int prio
= p
->static_prio
- MAX_RT_PRIO
;
673 struct load_weight
*load
= &p
->se
.load
;
676 * SCHED_IDLE tasks get minimal weight:
678 if (idle_policy(p
->policy
)) {
679 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
680 load
->inv_weight
= WMULT_IDLEPRIO
;
684 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
685 load
->inv_weight
= sched_prio_to_wmult
[prio
];
688 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
691 if (!(flags
& ENQUEUE_RESTORE
))
692 sched_info_queued(rq
, p
);
693 p
->sched_class
->enqueue_task(rq
, p
, flags
);
696 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
699 if (!(flags
& DEQUEUE_SAVE
))
700 sched_info_dequeued(rq
, p
);
701 p
->sched_class
->dequeue_task(rq
, p
, flags
);
704 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
706 if (task_contributes_to_load(p
))
707 rq
->nr_uninterruptible
--;
709 enqueue_task(rq
, p
, flags
);
712 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
714 if (task_contributes_to_load(p
))
715 rq
->nr_uninterruptible
++;
717 dequeue_task(rq
, p
, flags
);
720 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
723 * In theory, the compile should just see 0 here, and optimize out the call
724 * to sched_rt_avg_update. But I don't trust it...
726 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
727 s64 steal
= 0, irq_delta
= 0;
729 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
730 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
733 * Since irq_time is only updated on {soft,}irq_exit, we might run into
734 * this case when a previous update_rq_clock() happened inside a
737 * When this happens, we stop ->clock_task and only update the
738 * prev_irq_time stamp to account for the part that fit, so that a next
739 * update will consume the rest. This ensures ->clock_task is
742 * It does however cause some slight miss-attribution of {soft,}irq
743 * time, a more accurate solution would be to update the irq_time using
744 * the current rq->clock timestamp, except that would require using
747 if (irq_delta
> delta
)
750 rq
->prev_irq_time
+= irq_delta
;
753 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
754 if (static_key_false((¶virt_steal_rq_enabled
))) {
755 steal
= paravirt_steal_clock(cpu_of(rq
));
756 steal
-= rq
->prev_steal_time_rq
;
758 if (unlikely(steal
> delta
))
761 rq
->prev_steal_time_rq
+= steal
;
766 rq
->clock_task
+= delta
;
768 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
769 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
770 sched_rt_avg_update(rq
, irq_delta
+ steal
);
774 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
776 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
777 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
781 * Make it appear like a SCHED_FIFO task, its something
782 * userspace knows about and won't get confused about.
784 * Also, it will make PI more or less work without too
785 * much confusion -- but then, stop work should not
786 * rely on PI working anyway.
788 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
790 stop
->sched_class
= &stop_sched_class
;
793 cpu_rq(cpu
)->stop
= stop
;
797 * Reset it back to a normal scheduling class so that
798 * it can die in pieces.
800 old_stop
->sched_class
= &rt_sched_class
;
805 * __normal_prio - return the priority that is based on the static prio
807 static inline int __normal_prio(struct task_struct
*p
)
809 return p
->static_prio
;
813 * Calculate the expected normal priority: i.e. priority
814 * without taking RT-inheritance into account. Might be
815 * boosted by interactivity modifiers. Changes upon fork,
816 * setprio syscalls, and whenever the interactivity
817 * estimator recalculates.
819 static inline int normal_prio(struct task_struct
*p
)
823 if (task_has_dl_policy(p
))
824 prio
= MAX_DL_PRIO
-1;
825 else if (task_has_rt_policy(p
))
826 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
828 prio
= __normal_prio(p
);
833 * Calculate the current priority, i.e. the priority
834 * taken into account by the scheduler. This value might
835 * be boosted by RT tasks, or might be boosted by
836 * interactivity modifiers. Will be RT if the task got
837 * RT-boosted. If not then it returns p->normal_prio.
839 static int effective_prio(struct task_struct
*p
)
841 p
->normal_prio
= normal_prio(p
);
843 * If we are RT tasks or we were boosted to RT priority,
844 * keep the priority unchanged. Otherwise, update priority
845 * to the normal priority:
847 if (!rt_prio(p
->prio
))
848 return p
->normal_prio
;
853 * task_curr - is this task currently executing on a CPU?
854 * @p: the task in question.
856 * Return: 1 if the task is currently executing. 0 otherwise.
858 inline int task_curr(const struct task_struct
*p
)
860 return cpu_curr(task_cpu(p
)) == p
;
864 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
865 * use the balance_callback list if you want balancing.
867 * this means any call to check_class_changed() must be followed by a call to
868 * balance_callback().
870 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
871 const struct sched_class
*prev_class
,
874 if (prev_class
!= p
->sched_class
) {
875 if (prev_class
->switched_from
)
876 prev_class
->switched_from(rq
, p
);
878 p
->sched_class
->switched_to(rq
, p
);
879 } else if (oldprio
!= p
->prio
|| dl_task(p
))
880 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
883 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
885 const struct sched_class
*class;
887 if (p
->sched_class
== rq
->curr
->sched_class
) {
888 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
890 for_each_class(class) {
891 if (class == rq
->curr
->sched_class
)
893 if (class == p
->sched_class
) {
901 * A queue event has occurred, and we're going to schedule. In
902 * this case, we can save a useless back to back clock update.
904 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
905 rq_clock_skip_update(rq
, true);
910 * This is how migration works:
912 * 1) we invoke migration_cpu_stop() on the target CPU using
914 * 2) stopper starts to run (implicitly forcing the migrated thread
916 * 3) it checks whether the migrated task is still in the wrong runqueue.
917 * 4) if it's in the wrong runqueue then the migration thread removes
918 * it and puts it into the right queue.
919 * 5) stopper completes and stop_one_cpu() returns and the migration
924 * move_queued_task - move a queued task to new rq.
926 * Returns (locked) new rq. Old rq's lock is released.
928 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
930 lockdep_assert_held(&rq
->lock
);
932 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
933 dequeue_task(rq
, p
, 0);
934 set_task_cpu(p
, new_cpu
);
935 raw_spin_unlock(&rq
->lock
);
937 rq
= cpu_rq(new_cpu
);
939 raw_spin_lock(&rq
->lock
);
940 BUG_ON(task_cpu(p
) != new_cpu
);
941 enqueue_task(rq
, p
, 0);
942 p
->on_rq
= TASK_ON_RQ_QUEUED
;
943 check_preempt_curr(rq
, p
, 0);
948 struct migration_arg
{
949 struct task_struct
*task
;
954 * Move (not current) task off this cpu, onto dest cpu. We're doing
955 * this because either it can't run here any more (set_cpus_allowed()
956 * away from this CPU, or CPU going down), or because we're
957 * attempting to rebalance this task on exec (sched_exec).
959 * So we race with normal scheduler movements, but that's OK, as long
960 * as the task is no longer on this CPU.
962 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
964 if (unlikely(!cpu_active(dest_cpu
)))
967 /* Affinity changed (again). */
968 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
971 rq
= move_queued_task(rq
, p
, dest_cpu
);
977 * migration_cpu_stop - this will be executed by a highprio stopper thread
978 * and performs thread migration by bumping thread off CPU then
979 * 'pushing' onto another runqueue.
981 static int migration_cpu_stop(void *data
)
983 struct migration_arg
*arg
= data
;
984 struct task_struct
*p
= arg
->task
;
985 struct rq
*rq
= this_rq();
988 * The original target cpu might have gone down and we might
989 * be on another cpu but it doesn't matter.
993 * We need to explicitly wake pending tasks before running
994 * __migrate_task() such that we will not miss enforcing cpus_allowed
995 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
997 sched_ttwu_pending();
999 raw_spin_lock(&p
->pi_lock
);
1000 raw_spin_lock(&rq
->lock
);
1002 * If task_rq(p) != rq, it cannot be migrated here, because we're
1003 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1004 * we're holding p->pi_lock.
1006 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1007 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1008 raw_spin_unlock(&rq
->lock
);
1009 raw_spin_unlock(&p
->pi_lock
);
1016 * sched_class::set_cpus_allowed must do the below, but is not required to
1017 * actually call this function.
1019 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1021 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1022 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1025 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1027 struct rq
*rq
= task_rq(p
);
1028 bool queued
, running
;
1030 lockdep_assert_held(&p
->pi_lock
);
1032 queued
= task_on_rq_queued(p
);
1033 running
= task_current(rq
, p
);
1037 * Because __kthread_bind() calls this on blocked tasks without
1040 lockdep_assert_held(&rq
->lock
);
1041 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1044 put_prev_task(rq
, p
);
1046 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1049 p
->sched_class
->set_curr_task(rq
);
1051 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1055 * Change a given task's CPU affinity. Migrate the thread to a
1056 * proper CPU and schedule it away if the CPU it's executing on
1057 * is removed from the allowed bitmask.
1059 * NOTE: the caller must have a valid reference to the task, the
1060 * task must not exit() & deallocate itself prematurely. The
1061 * call is not atomic; no spinlocks may be held.
1063 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1064 const struct cpumask
*new_mask
, bool check
)
1066 unsigned long flags
;
1068 unsigned int dest_cpu
;
1071 rq
= task_rq_lock(p
, &flags
);
1074 * Must re-check here, to close a race against __kthread_bind(),
1075 * sched_setaffinity() is not guaranteed to observe the flag.
1077 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1082 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1085 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1090 do_set_cpus_allowed(p
, new_mask
);
1092 /* Can the task run on the task's current CPU? If so, we're done */
1093 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1096 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1097 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1098 struct migration_arg arg
= { p
, dest_cpu
};
1099 /* Need help from migration thread: drop lock and wait. */
1100 task_rq_unlock(rq
, p
, &flags
);
1101 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1102 tlb_migrate_finish(p
->mm
);
1104 } else if (task_on_rq_queued(p
)) {
1106 * OK, since we're going to drop the lock immediately
1107 * afterwards anyway.
1109 lockdep_unpin_lock(&rq
->lock
);
1110 rq
= move_queued_task(rq
, p
, dest_cpu
);
1111 lockdep_pin_lock(&rq
->lock
);
1114 task_rq_unlock(rq
, p
, &flags
);
1119 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1121 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1123 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1125 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1127 #ifdef CONFIG_SCHED_DEBUG
1129 * We should never call set_task_cpu() on a blocked task,
1130 * ttwu() will sort out the placement.
1132 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1136 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1137 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1138 * time relying on p->on_rq.
1140 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1141 p
->sched_class
== &fair_sched_class
&&
1142 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1144 #ifdef CONFIG_LOCKDEP
1146 * The caller should hold either p->pi_lock or rq->lock, when changing
1147 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1149 * sched_move_task() holds both and thus holding either pins the cgroup,
1152 * Furthermore, all task_rq users should acquire both locks, see
1155 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1156 lockdep_is_held(&task_rq(p
)->lock
)));
1160 trace_sched_migrate_task(p
, new_cpu
);
1162 if (task_cpu(p
) != new_cpu
) {
1163 if (p
->sched_class
->migrate_task_rq
)
1164 p
->sched_class
->migrate_task_rq(p
);
1165 p
->se
.nr_migrations
++;
1166 perf_event_task_migrate(p
);
1169 __set_task_cpu(p
, new_cpu
);
1172 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1174 if (task_on_rq_queued(p
)) {
1175 struct rq
*src_rq
, *dst_rq
;
1177 src_rq
= task_rq(p
);
1178 dst_rq
= cpu_rq(cpu
);
1180 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1181 deactivate_task(src_rq
, p
, 0);
1182 set_task_cpu(p
, cpu
);
1183 activate_task(dst_rq
, p
, 0);
1184 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1185 check_preempt_curr(dst_rq
, p
, 0);
1188 * Task isn't running anymore; make it appear like we migrated
1189 * it before it went to sleep. This means on wakeup we make the
1190 * previous cpu our targer instead of where it really is.
1196 struct migration_swap_arg
{
1197 struct task_struct
*src_task
, *dst_task
;
1198 int src_cpu
, dst_cpu
;
1201 static int migrate_swap_stop(void *data
)
1203 struct migration_swap_arg
*arg
= data
;
1204 struct rq
*src_rq
, *dst_rq
;
1207 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1210 src_rq
= cpu_rq(arg
->src_cpu
);
1211 dst_rq
= cpu_rq(arg
->dst_cpu
);
1213 double_raw_lock(&arg
->src_task
->pi_lock
,
1214 &arg
->dst_task
->pi_lock
);
1215 double_rq_lock(src_rq
, dst_rq
);
1217 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1220 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1223 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1226 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1229 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1230 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1235 double_rq_unlock(src_rq
, dst_rq
);
1236 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1237 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1243 * Cross migrate two tasks
1245 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1247 struct migration_swap_arg arg
;
1250 arg
= (struct migration_swap_arg
){
1252 .src_cpu
= task_cpu(cur
),
1254 .dst_cpu
= task_cpu(p
),
1257 if (arg
.src_cpu
== arg
.dst_cpu
)
1261 * These three tests are all lockless; this is OK since all of them
1262 * will be re-checked with proper locks held further down the line.
1264 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1267 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1270 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1273 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1274 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1281 * wait_task_inactive - wait for a thread to unschedule.
1283 * If @match_state is nonzero, it's the @p->state value just checked and
1284 * not expected to change. If it changes, i.e. @p might have woken up,
1285 * then return zero. When we succeed in waiting for @p to be off its CPU,
1286 * we return a positive number (its total switch count). If a second call
1287 * a short while later returns the same number, the caller can be sure that
1288 * @p has remained unscheduled the whole time.
1290 * The caller must ensure that the task *will* unschedule sometime soon,
1291 * else this function might spin for a *long* time. This function can't
1292 * be called with interrupts off, or it may introduce deadlock with
1293 * smp_call_function() if an IPI is sent by the same process we are
1294 * waiting to become inactive.
1296 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1298 unsigned long flags
;
1299 int running
, queued
;
1305 * We do the initial early heuristics without holding
1306 * any task-queue locks at all. We'll only try to get
1307 * the runqueue lock when things look like they will
1313 * If the task is actively running on another CPU
1314 * still, just relax and busy-wait without holding
1317 * NOTE! Since we don't hold any locks, it's not
1318 * even sure that "rq" stays as the right runqueue!
1319 * But we don't care, since "task_running()" will
1320 * return false if the runqueue has changed and p
1321 * is actually now running somewhere else!
1323 while (task_running(rq
, p
)) {
1324 if (match_state
&& unlikely(p
->state
!= match_state
))
1330 * Ok, time to look more closely! We need the rq
1331 * lock now, to be *sure*. If we're wrong, we'll
1332 * just go back and repeat.
1334 rq
= task_rq_lock(p
, &flags
);
1335 trace_sched_wait_task(p
);
1336 running
= task_running(rq
, p
);
1337 queued
= task_on_rq_queued(p
);
1339 if (!match_state
|| p
->state
== match_state
)
1340 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1341 task_rq_unlock(rq
, p
, &flags
);
1344 * If it changed from the expected state, bail out now.
1346 if (unlikely(!ncsw
))
1350 * Was it really running after all now that we
1351 * checked with the proper locks actually held?
1353 * Oops. Go back and try again..
1355 if (unlikely(running
)) {
1361 * It's not enough that it's not actively running,
1362 * it must be off the runqueue _entirely_, and not
1365 * So if it was still runnable (but just not actively
1366 * running right now), it's preempted, and we should
1367 * yield - it could be a while.
1369 if (unlikely(queued
)) {
1370 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1372 set_current_state(TASK_UNINTERRUPTIBLE
);
1373 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1378 * Ahh, all good. It wasn't running, and it wasn't
1379 * runnable, which means that it will never become
1380 * running in the future either. We're all done!
1389 * kick_process - kick a running thread to enter/exit the kernel
1390 * @p: the to-be-kicked thread
1392 * Cause a process which is running on another CPU to enter
1393 * kernel-mode, without any delay. (to get signals handled.)
1395 * NOTE: this function doesn't have to take the runqueue lock,
1396 * because all it wants to ensure is that the remote task enters
1397 * the kernel. If the IPI races and the task has been migrated
1398 * to another CPU then no harm is done and the purpose has been
1401 void kick_process(struct task_struct
*p
)
1407 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1408 smp_send_reschedule(cpu
);
1411 EXPORT_SYMBOL_GPL(kick_process
);
1414 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1416 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1418 int nid
= cpu_to_node(cpu
);
1419 const struct cpumask
*nodemask
= NULL
;
1420 enum { cpuset
, possible
, fail
} state
= cpuset
;
1424 * If the node that the cpu is on has been offlined, cpu_to_node()
1425 * will return -1. There is no cpu on the node, and we should
1426 * select the cpu on the other node.
1429 nodemask
= cpumask_of_node(nid
);
1431 /* Look for allowed, online CPU in same node. */
1432 for_each_cpu(dest_cpu
, nodemask
) {
1433 if (!cpu_online(dest_cpu
))
1435 if (!cpu_active(dest_cpu
))
1437 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1443 /* Any allowed, online CPU? */
1444 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1445 if (!cpu_online(dest_cpu
))
1447 if (!cpu_active(dest_cpu
))
1452 /* No more Mr. Nice Guy. */
1455 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1456 cpuset_cpus_allowed_fallback(p
);
1462 do_set_cpus_allowed(p
, cpu_possible_mask
);
1473 if (state
!= cpuset
) {
1475 * Don't tell them about moving exiting tasks or
1476 * kernel threads (both mm NULL), since they never
1479 if (p
->mm
&& printk_ratelimit()) {
1480 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1481 task_pid_nr(p
), p
->comm
, cpu
);
1489 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1492 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1494 lockdep_assert_held(&p
->pi_lock
);
1496 if (p
->nr_cpus_allowed
> 1)
1497 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1500 * In order not to call set_task_cpu() on a blocking task we need
1501 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1504 * Since this is common to all placement strategies, this lives here.
1506 * [ this allows ->select_task() to simply return task_cpu(p) and
1507 * not worry about this generic constraint ]
1509 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1511 cpu
= select_fallback_rq(task_cpu(p
), p
);
1516 static void update_avg(u64
*avg
, u64 sample
)
1518 s64 diff
= sample
- *avg
;
1524 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1525 const struct cpumask
*new_mask
, bool check
)
1527 return set_cpus_allowed_ptr(p
, new_mask
);
1530 #endif /* CONFIG_SMP */
1533 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1535 #ifdef CONFIG_SCHEDSTATS
1536 struct rq
*rq
= this_rq();
1539 int this_cpu
= smp_processor_id();
1541 if (cpu
== this_cpu
) {
1542 schedstat_inc(rq
, ttwu_local
);
1543 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1545 struct sched_domain
*sd
;
1547 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1549 for_each_domain(this_cpu
, sd
) {
1550 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1551 schedstat_inc(sd
, ttwu_wake_remote
);
1558 if (wake_flags
& WF_MIGRATED
)
1559 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1561 #endif /* CONFIG_SMP */
1563 schedstat_inc(rq
, ttwu_count
);
1564 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1566 if (wake_flags
& WF_SYNC
)
1567 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1569 #endif /* CONFIG_SCHEDSTATS */
1572 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1574 activate_task(rq
, p
, en_flags
);
1575 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1577 /* if a worker is waking up, notify workqueue */
1578 if (p
->flags
& PF_WQ_WORKER
)
1579 wq_worker_waking_up(p
, cpu_of(rq
));
1583 * Mark the task runnable and perform wakeup-preemption.
1586 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1588 check_preempt_curr(rq
, p
, wake_flags
);
1589 p
->state
= TASK_RUNNING
;
1590 trace_sched_wakeup(p
);
1593 if (p
->sched_class
->task_woken
) {
1595 * Our task @p is fully woken up and running; so its safe to
1596 * drop the rq->lock, hereafter rq is only used for statistics.
1598 lockdep_unpin_lock(&rq
->lock
);
1599 p
->sched_class
->task_woken(rq
, p
);
1600 lockdep_pin_lock(&rq
->lock
);
1603 if (rq
->idle_stamp
) {
1604 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1605 u64 max
= 2*rq
->max_idle_balance_cost
;
1607 update_avg(&rq
->avg_idle
, delta
);
1609 if (rq
->avg_idle
> max
)
1618 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1620 lockdep_assert_held(&rq
->lock
);
1623 if (p
->sched_contributes_to_load
)
1624 rq
->nr_uninterruptible
--;
1627 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1628 ttwu_do_wakeup(rq
, p
, wake_flags
);
1632 * Called in case the task @p isn't fully descheduled from its runqueue,
1633 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1634 * since all we need to do is flip p->state to TASK_RUNNING, since
1635 * the task is still ->on_rq.
1637 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1642 rq
= __task_rq_lock(p
);
1643 if (task_on_rq_queued(p
)) {
1644 /* check_preempt_curr() may use rq clock */
1645 update_rq_clock(rq
);
1646 ttwu_do_wakeup(rq
, p
, wake_flags
);
1649 __task_rq_unlock(rq
);
1655 void sched_ttwu_pending(void)
1657 struct rq
*rq
= this_rq();
1658 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1659 struct task_struct
*p
;
1660 unsigned long flags
;
1665 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1666 lockdep_pin_lock(&rq
->lock
);
1669 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1670 llist
= llist_next(llist
);
1671 ttwu_do_activate(rq
, p
, 0);
1674 lockdep_unpin_lock(&rq
->lock
);
1675 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1678 void scheduler_ipi(void)
1681 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1682 * TIF_NEED_RESCHED remotely (for the first time) will also send
1685 preempt_fold_need_resched();
1687 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1691 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1692 * traditionally all their work was done from the interrupt return
1693 * path. Now that we actually do some work, we need to make sure
1696 * Some archs already do call them, luckily irq_enter/exit nest
1699 * Arguably we should visit all archs and update all handlers,
1700 * however a fair share of IPIs are still resched only so this would
1701 * somewhat pessimize the simple resched case.
1704 sched_ttwu_pending();
1707 * Check if someone kicked us for doing the nohz idle load balance.
1709 if (unlikely(got_nohz_idle_kick())) {
1710 this_rq()->idle_balance
= 1;
1711 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1716 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1718 struct rq
*rq
= cpu_rq(cpu
);
1720 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1721 if (!set_nr_if_polling(rq
->idle
))
1722 smp_send_reschedule(cpu
);
1724 trace_sched_wake_idle_without_ipi(cpu
);
1728 void wake_up_if_idle(int cpu
)
1730 struct rq
*rq
= cpu_rq(cpu
);
1731 unsigned long flags
;
1735 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1738 if (set_nr_if_polling(rq
->idle
)) {
1739 trace_sched_wake_idle_without_ipi(cpu
);
1741 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1742 if (is_idle_task(rq
->curr
))
1743 smp_send_reschedule(cpu
);
1744 /* Else cpu is not in idle, do nothing here */
1745 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1752 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1754 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1756 #endif /* CONFIG_SMP */
1758 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1760 struct rq
*rq
= cpu_rq(cpu
);
1762 #if defined(CONFIG_SMP)
1763 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1764 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1765 ttwu_queue_remote(p
, cpu
);
1770 raw_spin_lock(&rq
->lock
);
1771 lockdep_pin_lock(&rq
->lock
);
1772 ttwu_do_activate(rq
, p
, 0);
1773 lockdep_unpin_lock(&rq
->lock
);
1774 raw_spin_unlock(&rq
->lock
);
1778 * Notes on Program-Order guarantees on SMP systems.
1782 * The basic program-order guarantee on SMP systems is that when a task [t]
1783 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1784 * execution on its new cpu [c1].
1786 * For migration (of runnable tasks) this is provided by the following means:
1788 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1789 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1790 * rq(c1)->lock (if not at the same time, then in that order).
1791 * C) LOCK of the rq(c1)->lock scheduling in task
1793 * Transitivity guarantees that B happens after A and C after B.
1794 * Note: we only require RCpc transitivity.
1795 * Note: the cpu doing B need not be c0 or c1
1804 * UNLOCK rq(0)->lock
1806 * LOCK rq(0)->lock // orders against CPU0
1808 * UNLOCK rq(0)->lock
1812 * UNLOCK rq(1)->lock
1814 * LOCK rq(1)->lock // orders against CPU2
1817 * UNLOCK rq(1)->lock
1820 * BLOCKING -- aka. SLEEP + WAKEUP
1822 * For blocking we (obviously) need to provide the same guarantee as for
1823 * migration. However the means are completely different as there is no lock
1824 * chain to provide order. Instead we do:
1826 * 1) smp_store_release(X->on_cpu, 0)
1827 * 2) smp_cond_acquire(!X->on_cpu)
1831 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1833 * LOCK rq(0)->lock LOCK X->pi_lock
1836 * smp_store_release(X->on_cpu, 0);
1838 * smp_cond_acquire(!X->on_cpu);
1844 * X->state = RUNNING
1845 * UNLOCK rq(2)->lock
1847 * LOCK rq(2)->lock // orders against CPU1
1850 * UNLOCK rq(2)->lock
1853 * UNLOCK rq(0)->lock
1856 * However; for wakeups there is a second guarantee we must provide, namely we
1857 * must observe the state that lead to our wakeup. That is, not only must our
1858 * task observe its own prior state, it must also observe the stores prior to
1861 * This means that any means of doing remote wakeups must order the CPU doing
1862 * the wakeup against the CPU the task is going to end up running on. This,
1863 * however, is already required for the regular Program-Order guarantee above,
1864 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1869 * try_to_wake_up - wake up a thread
1870 * @p: the thread to be awakened
1871 * @state: the mask of task states that can be woken
1872 * @wake_flags: wake modifier flags (WF_*)
1874 * Put it on the run-queue if it's not already there. The "current"
1875 * thread is always on the run-queue (except when the actual
1876 * re-schedule is in progress), and as such you're allowed to do
1877 * the simpler "current->state = TASK_RUNNING" to mark yourself
1878 * runnable without the overhead of this.
1880 * Return: %true if @p was woken up, %false if it was already running.
1881 * or @state didn't match @p's state.
1884 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1886 unsigned long flags
;
1887 int cpu
, success
= 0;
1890 * If we are going to wake up a thread waiting for CONDITION we
1891 * need to ensure that CONDITION=1 done by the caller can not be
1892 * reordered with p->state check below. This pairs with mb() in
1893 * set_current_state() the waiting thread does.
1895 smp_mb__before_spinlock();
1896 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1897 if (!(p
->state
& state
))
1900 trace_sched_waking(p
);
1902 success
= 1; /* we're going to change ->state */
1905 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1910 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1911 * possible to, falsely, observe p->on_cpu == 0.
1913 * One must be running (->on_cpu == 1) in order to remove oneself
1914 * from the runqueue.
1916 * [S] ->on_cpu = 1; [L] ->on_rq
1920 * [S] ->on_rq = 0; [L] ->on_cpu
1922 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1923 * from the consecutive calls to schedule(); the first switching to our
1924 * task, the second putting it to sleep.
1929 * If the owning (remote) cpu is still in the middle of schedule() with
1930 * this task as prev, wait until its done referencing the task.
1932 * Pairs with the smp_store_release() in finish_lock_switch().
1934 * This ensures that tasks getting woken will be fully ordered against
1935 * their previous state and preserve Program Order.
1937 smp_cond_acquire(!p
->on_cpu
);
1939 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1940 p
->state
= TASK_WAKING
;
1942 if (p
->sched_class
->task_waking
)
1943 p
->sched_class
->task_waking(p
);
1945 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1946 if (task_cpu(p
) != cpu
) {
1947 wake_flags
|= WF_MIGRATED
;
1948 set_task_cpu(p
, cpu
);
1950 #endif /* CONFIG_SMP */
1954 if (schedstat_enabled())
1955 ttwu_stat(p
, cpu
, wake_flags
);
1957 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1963 * try_to_wake_up_local - try to wake up a local task with rq lock held
1964 * @p: the thread to be awakened
1966 * Put @p on the run-queue if it's not already there. The caller must
1967 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1970 static void try_to_wake_up_local(struct task_struct
*p
)
1972 struct rq
*rq
= task_rq(p
);
1974 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1975 WARN_ON_ONCE(p
== current
))
1978 lockdep_assert_held(&rq
->lock
);
1980 if (!raw_spin_trylock(&p
->pi_lock
)) {
1982 * This is OK, because current is on_cpu, which avoids it being
1983 * picked for load-balance and preemption/IRQs are still
1984 * disabled avoiding further scheduler activity on it and we've
1985 * not yet picked a replacement task.
1987 lockdep_unpin_lock(&rq
->lock
);
1988 raw_spin_unlock(&rq
->lock
);
1989 raw_spin_lock(&p
->pi_lock
);
1990 raw_spin_lock(&rq
->lock
);
1991 lockdep_pin_lock(&rq
->lock
);
1994 if (!(p
->state
& TASK_NORMAL
))
1997 trace_sched_waking(p
);
1999 if (!task_on_rq_queued(p
))
2000 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2002 ttwu_do_wakeup(rq
, p
, 0);
2003 if (schedstat_enabled())
2004 ttwu_stat(p
, smp_processor_id(), 0);
2006 raw_spin_unlock(&p
->pi_lock
);
2010 * wake_up_process - Wake up a specific process
2011 * @p: The process to be woken up.
2013 * Attempt to wake up the nominated process and move it to the set of runnable
2016 * Return: 1 if the process was woken up, 0 if it was already running.
2018 * It may be assumed that this function implies a write memory barrier before
2019 * changing the task state if and only if any tasks are woken up.
2021 int wake_up_process(struct task_struct
*p
)
2023 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2025 EXPORT_SYMBOL(wake_up_process
);
2027 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2029 return try_to_wake_up(p
, state
, 0);
2033 * This function clears the sched_dl_entity static params.
2035 void __dl_clear_params(struct task_struct
*p
)
2037 struct sched_dl_entity
*dl_se
= &p
->dl
;
2039 dl_se
->dl_runtime
= 0;
2040 dl_se
->dl_deadline
= 0;
2041 dl_se
->dl_period
= 0;
2045 dl_se
->dl_throttled
= 0;
2046 dl_se
->dl_yielded
= 0;
2050 * Perform scheduler related setup for a newly forked process p.
2051 * p is forked by current.
2053 * __sched_fork() is basic setup used by init_idle() too:
2055 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2060 p
->se
.exec_start
= 0;
2061 p
->se
.sum_exec_runtime
= 0;
2062 p
->se
.prev_sum_exec_runtime
= 0;
2063 p
->se
.nr_migrations
= 0;
2065 INIT_LIST_HEAD(&p
->se
.group_node
);
2067 #ifdef CONFIG_FAIR_GROUP_SCHED
2068 p
->se
.cfs_rq
= NULL
;
2071 #ifdef CONFIG_SCHEDSTATS
2072 /* Even if schedstat is disabled, there should not be garbage */
2073 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2076 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2077 init_dl_task_timer(&p
->dl
);
2078 __dl_clear_params(p
);
2080 INIT_LIST_HEAD(&p
->rt
.run_list
);
2082 p
->rt
.time_slice
= sched_rr_timeslice
;
2086 #ifdef CONFIG_PREEMPT_NOTIFIERS
2087 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2090 #ifdef CONFIG_NUMA_BALANCING
2091 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2092 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2093 p
->mm
->numa_scan_seq
= 0;
2096 if (clone_flags
& CLONE_VM
)
2097 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2099 p
->numa_preferred_nid
= -1;
2101 p
->node_stamp
= 0ULL;
2102 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2103 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2104 p
->numa_work
.next
= &p
->numa_work
;
2105 p
->numa_faults
= NULL
;
2106 p
->last_task_numa_placement
= 0;
2107 p
->last_sum_exec_runtime
= 0;
2109 p
->numa_group
= NULL
;
2110 #endif /* CONFIG_NUMA_BALANCING */
2113 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2115 #ifdef CONFIG_NUMA_BALANCING
2117 void set_numabalancing_state(bool enabled
)
2120 static_branch_enable(&sched_numa_balancing
);
2122 static_branch_disable(&sched_numa_balancing
);
2125 #ifdef CONFIG_PROC_SYSCTL
2126 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2127 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2131 int state
= static_branch_likely(&sched_numa_balancing
);
2133 if (write
&& !capable(CAP_SYS_ADMIN
))
2138 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2142 set_numabalancing_state(state
);
2148 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2150 #ifdef CONFIG_SCHEDSTATS
2151 static void set_schedstats(bool enabled
)
2154 static_branch_enable(&sched_schedstats
);
2156 static_branch_disable(&sched_schedstats
);
2159 void force_schedstat_enabled(void)
2161 if (!schedstat_enabled()) {
2162 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2163 static_branch_enable(&sched_schedstats
);
2167 static int __init
setup_schedstats(char *str
)
2173 if (!strcmp(str
, "enable")) {
2174 set_schedstats(true);
2176 } else if (!strcmp(str
, "disable")) {
2177 set_schedstats(false);
2182 pr_warn("Unable to parse schedstats=\n");
2186 __setup("schedstats=", setup_schedstats
);
2188 #ifdef CONFIG_PROC_SYSCTL
2189 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2190 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2194 int state
= static_branch_likely(&sched_schedstats
);
2196 if (write
&& !capable(CAP_SYS_ADMIN
))
2201 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2205 set_schedstats(state
);
2212 * fork()/clone()-time setup:
2214 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2216 unsigned long flags
;
2217 int cpu
= get_cpu();
2219 __sched_fork(clone_flags
, p
);
2221 * We mark the process as running here. This guarantees that
2222 * nobody will actually run it, and a signal or other external
2223 * event cannot wake it up and insert it on the runqueue either.
2225 p
->state
= TASK_RUNNING
;
2228 * Make sure we do not leak PI boosting priority to the child.
2230 p
->prio
= current
->normal_prio
;
2233 * Revert to default priority/policy on fork if requested.
2235 if (unlikely(p
->sched_reset_on_fork
)) {
2236 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2237 p
->policy
= SCHED_NORMAL
;
2238 p
->static_prio
= NICE_TO_PRIO(0);
2240 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2241 p
->static_prio
= NICE_TO_PRIO(0);
2243 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2247 * We don't need the reset flag anymore after the fork. It has
2248 * fulfilled its duty:
2250 p
->sched_reset_on_fork
= 0;
2253 if (dl_prio(p
->prio
)) {
2256 } else if (rt_prio(p
->prio
)) {
2257 p
->sched_class
= &rt_sched_class
;
2259 p
->sched_class
= &fair_sched_class
;
2262 if (p
->sched_class
->task_fork
)
2263 p
->sched_class
->task_fork(p
);
2266 * The child is not yet in the pid-hash so no cgroup attach races,
2267 * and the cgroup is pinned to this child due to cgroup_fork()
2268 * is ran before sched_fork().
2270 * Silence PROVE_RCU.
2272 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2273 set_task_cpu(p
, cpu
);
2274 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2276 #ifdef CONFIG_SCHED_INFO
2277 if (likely(sched_info_on()))
2278 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2280 #if defined(CONFIG_SMP)
2283 init_task_preempt_count(p
);
2285 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2286 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2293 unsigned long to_ratio(u64 period
, u64 runtime
)
2295 if (runtime
== RUNTIME_INF
)
2299 * Doing this here saves a lot of checks in all
2300 * the calling paths, and returning zero seems
2301 * safe for them anyway.
2306 return div64_u64(runtime
<< 20, period
);
2310 inline struct dl_bw
*dl_bw_of(int i
)
2312 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2313 "sched RCU must be held");
2314 return &cpu_rq(i
)->rd
->dl_bw
;
2317 static inline int dl_bw_cpus(int i
)
2319 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2322 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2323 "sched RCU must be held");
2324 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2330 inline struct dl_bw
*dl_bw_of(int i
)
2332 return &cpu_rq(i
)->dl
.dl_bw
;
2335 static inline int dl_bw_cpus(int i
)
2342 * We must be sure that accepting a new task (or allowing changing the
2343 * parameters of an existing one) is consistent with the bandwidth
2344 * constraints. If yes, this function also accordingly updates the currently
2345 * allocated bandwidth to reflect the new situation.
2347 * This function is called while holding p's rq->lock.
2349 * XXX we should delay bw change until the task's 0-lag point, see
2352 static int dl_overflow(struct task_struct
*p
, int policy
,
2353 const struct sched_attr
*attr
)
2356 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2357 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2358 u64 runtime
= attr
->sched_runtime
;
2359 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2362 if (new_bw
== p
->dl
.dl_bw
)
2366 * Either if a task, enters, leave, or stays -deadline but changes
2367 * its parameters, we may need to update accordingly the total
2368 * allocated bandwidth of the container.
2370 raw_spin_lock(&dl_b
->lock
);
2371 cpus
= dl_bw_cpus(task_cpu(p
));
2372 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2373 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2374 __dl_add(dl_b
, new_bw
);
2376 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2377 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2378 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2379 __dl_add(dl_b
, new_bw
);
2381 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2382 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2385 raw_spin_unlock(&dl_b
->lock
);
2390 extern void init_dl_bw(struct dl_bw
*dl_b
);
2393 * wake_up_new_task - wake up a newly created task for the first time.
2395 * This function will do some initial scheduler statistics housekeeping
2396 * that must be done for every newly created context, then puts the task
2397 * on the runqueue and wakes it.
2399 void wake_up_new_task(struct task_struct
*p
)
2401 unsigned long flags
;
2404 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2405 /* Initialize new task's runnable average */
2406 init_entity_runnable_average(&p
->se
);
2409 * Fork balancing, do it here and not earlier because:
2410 * - cpus_allowed can change in the fork path
2411 * - any previously selected cpu might disappear through hotplug
2413 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2416 rq
= __task_rq_lock(p
);
2417 activate_task(rq
, p
, 0);
2418 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2419 trace_sched_wakeup_new(p
);
2420 check_preempt_curr(rq
, p
, WF_FORK
);
2422 if (p
->sched_class
->task_woken
) {
2424 * Nothing relies on rq->lock after this, so its fine to
2427 lockdep_unpin_lock(&rq
->lock
);
2428 p
->sched_class
->task_woken(rq
, p
);
2429 lockdep_pin_lock(&rq
->lock
);
2432 task_rq_unlock(rq
, p
, &flags
);
2435 #ifdef CONFIG_PREEMPT_NOTIFIERS
2437 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2439 void preempt_notifier_inc(void)
2441 static_key_slow_inc(&preempt_notifier_key
);
2443 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2445 void preempt_notifier_dec(void)
2447 static_key_slow_dec(&preempt_notifier_key
);
2449 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2452 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2453 * @notifier: notifier struct to register
2455 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2457 if (!static_key_false(&preempt_notifier_key
))
2458 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2460 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2462 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2465 * preempt_notifier_unregister - no longer interested in preemption notifications
2466 * @notifier: notifier struct to unregister
2468 * This is *not* safe to call from within a preemption notifier.
2470 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2472 hlist_del(¬ifier
->link
);
2474 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2476 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2478 struct preempt_notifier
*notifier
;
2480 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2481 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2484 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2486 if (static_key_false(&preempt_notifier_key
))
2487 __fire_sched_in_preempt_notifiers(curr
);
2491 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2492 struct task_struct
*next
)
2494 struct preempt_notifier
*notifier
;
2496 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2497 notifier
->ops
->sched_out(notifier
, next
);
2500 static __always_inline
void
2501 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2502 struct task_struct
*next
)
2504 if (static_key_false(&preempt_notifier_key
))
2505 __fire_sched_out_preempt_notifiers(curr
, next
);
2508 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2510 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2515 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2516 struct task_struct
*next
)
2520 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2523 * prepare_task_switch - prepare to switch tasks
2524 * @rq: the runqueue preparing to switch
2525 * @prev: the current task that is being switched out
2526 * @next: the task we are going to switch to.
2528 * This is called with the rq lock held and interrupts off. It must
2529 * be paired with a subsequent finish_task_switch after the context
2532 * prepare_task_switch sets up locking and calls architecture specific
2536 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2537 struct task_struct
*next
)
2539 sched_info_switch(rq
, prev
, next
);
2540 perf_event_task_sched_out(prev
, next
);
2541 fire_sched_out_preempt_notifiers(prev
, next
);
2542 prepare_lock_switch(rq
, next
);
2543 prepare_arch_switch(next
);
2547 * finish_task_switch - clean up after a task-switch
2548 * @prev: the thread we just switched away from.
2550 * finish_task_switch must be called after the context switch, paired
2551 * with a prepare_task_switch call before the context switch.
2552 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2553 * and do any other architecture-specific cleanup actions.
2555 * Note that we may have delayed dropping an mm in context_switch(). If
2556 * so, we finish that here outside of the runqueue lock. (Doing it
2557 * with the lock held can cause deadlocks; see schedule() for
2560 * The context switch have flipped the stack from under us and restored the
2561 * local variables which were saved when this task called schedule() in the
2562 * past. prev == current is still correct but we need to recalculate this_rq
2563 * because prev may have moved to another CPU.
2565 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2566 __releases(rq
->lock
)
2568 struct rq
*rq
= this_rq();
2569 struct mm_struct
*mm
= rq
->prev_mm
;
2573 * The previous task will have left us with a preempt_count of 2
2574 * because it left us after:
2577 * preempt_disable(); // 1
2579 * raw_spin_lock_irq(&rq->lock) // 2
2581 * Also, see FORK_PREEMPT_COUNT.
2583 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2584 "corrupted preempt_count: %s/%d/0x%x\n",
2585 current
->comm
, current
->pid
, preempt_count()))
2586 preempt_count_set(FORK_PREEMPT_COUNT
);
2591 * A task struct has one reference for the use as "current".
2592 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2593 * schedule one last time. The schedule call will never return, and
2594 * the scheduled task must drop that reference.
2596 * We must observe prev->state before clearing prev->on_cpu (in
2597 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2598 * running on another CPU and we could rave with its RUNNING -> DEAD
2599 * transition, resulting in a double drop.
2601 prev_state
= prev
->state
;
2602 vtime_task_switch(prev
);
2603 perf_event_task_sched_in(prev
, current
);
2604 finish_lock_switch(rq
, prev
);
2605 finish_arch_post_lock_switch();
2607 fire_sched_in_preempt_notifiers(current
);
2610 if (unlikely(prev_state
== TASK_DEAD
)) {
2611 if (prev
->sched_class
->task_dead
)
2612 prev
->sched_class
->task_dead(prev
);
2615 * Remove function-return probe instances associated with this
2616 * task and put them back on the free list.
2618 kprobe_flush_task(prev
);
2619 put_task_struct(prev
);
2622 tick_nohz_task_switch();
2628 /* rq->lock is NOT held, but preemption is disabled */
2629 static void __balance_callback(struct rq
*rq
)
2631 struct callback_head
*head
, *next
;
2632 void (*func
)(struct rq
*rq
);
2633 unsigned long flags
;
2635 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2636 head
= rq
->balance_callback
;
2637 rq
->balance_callback
= NULL
;
2639 func
= (void (*)(struct rq
*))head
->func
;
2646 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2649 static inline void balance_callback(struct rq
*rq
)
2651 if (unlikely(rq
->balance_callback
))
2652 __balance_callback(rq
);
2657 static inline void balance_callback(struct rq
*rq
)
2664 * schedule_tail - first thing a freshly forked thread must call.
2665 * @prev: the thread we just switched away from.
2667 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2668 __releases(rq
->lock
)
2673 * New tasks start with FORK_PREEMPT_COUNT, see there and
2674 * finish_task_switch() for details.
2676 * finish_task_switch() will drop rq->lock() and lower preempt_count
2677 * and the preempt_enable() will end up enabling preemption (on
2678 * PREEMPT_COUNT kernels).
2681 rq
= finish_task_switch(prev
);
2682 balance_callback(rq
);
2685 if (current
->set_child_tid
)
2686 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2690 * context_switch - switch to the new MM and the new thread's register state.
2692 static inline struct rq
*
2693 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2694 struct task_struct
*next
)
2696 struct mm_struct
*mm
, *oldmm
;
2698 prepare_task_switch(rq
, prev
, next
);
2701 oldmm
= prev
->active_mm
;
2703 * For paravirt, this is coupled with an exit in switch_to to
2704 * combine the page table reload and the switch backend into
2707 arch_start_context_switch(prev
);
2710 next
->active_mm
= oldmm
;
2711 atomic_inc(&oldmm
->mm_count
);
2712 enter_lazy_tlb(oldmm
, next
);
2714 switch_mm(oldmm
, mm
, next
);
2717 prev
->active_mm
= NULL
;
2718 rq
->prev_mm
= oldmm
;
2721 * Since the runqueue lock will be released by the next
2722 * task (which is an invalid locking op but in the case
2723 * of the scheduler it's an obvious special-case), so we
2724 * do an early lockdep release here:
2726 lockdep_unpin_lock(&rq
->lock
);
2727 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2729 /* Here we just switch the register state and the stack. */
2730 switch_to(prev
, next
, prev
);
2733 return finish_task_switch(prev
);
2737 * nr_running and nr_context_switches:
2739 * externally visible scheduler statistics: current number of runnable
2740 * threads, total number of context switches performed since bootup.
2742 unsigned long nr_running(void)
2744 unsigned long i
, sum
= 0;
2746 for_each_online_cpu(i
)
2747 sum
+= cpu_rq(i
)->nr_running
;
2753 * Check if only the current task is running on the cpu.
2755 * Caution: this function does not check that the caller has disabled
2756 * preemption, thus the result might have a time-of-check-to-time-of-use
2757 * race. The caller is responsible to use it correctly, for example:
2759 * - from a non-preemptable section (of course)
2761 * - from a thread that is bound to a single CPU
2763 * - in a loop with very short iterations (e.g. a polling loop)
2765 bool single_task_running(void)
2767 return raw_rq()->nr_running
== 1;
2769 EXPORT_SYMBOL(single_task_running
);
2771 unsigned long long nr_context_switches(void)
2774 unsigned long long sum
= 0;
2776 for_each_possible_cpu(i
)
2777 sum
+= cpu_rq(i
)->nr_switches
;
2782 unsigned long nr_iowait(void)
2784 unsigned long i
, sum
= 0;
2786 for_each_possible_cpu(i
)
2787 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2792 unsigned long nr_iowait_cpu(int cpu
)
2794 struct rq
*this = cpu_rq(cpu
);
2795 return atomic_read(&this->nr_iowait
);
2798 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2800 struct rq
*rq
= this_rq();
2801 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2802 *load
= rq
->load
.weight
;
2808 * sched_exec - execve() is a valuable balancing opportunity, because at
2809 * this point the task has the smallest effective memory and cache footprint.
2811 void sched_exec(void)
2813 struct task_struct
*p
= current
;
2814 unsigned long flags
;
2817 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2818 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2819 if (dest_cpu
== smp_processor_id())
2822 if (likely(cpu_active(dest_cpu
))) {
2823 struct migration_arg arg
= { p
, dest_cpu
};
2825 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2826 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2830 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2835 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2836 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2838 EXPORT_PER_CPU_SYMBOL(kstat
);
2839 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2842 * Return accounted runtime for the task.
2843 * In case the task is currently running, return the runtime plus current's
2844 * pending runtime that have not been accounted yet.
2846 unsigned long long task_sched_runtime(struct task_struct
*p
)
2848 unsigned long flags
;
2852 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2854 * 64-bit doesn't need locks to atomically read a 64bit value.
2855 * So we have a optimization chance when the task's delta_exec is 0.
2856 * Reading ->on_cpu is racy, but this is ok.
2858 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2859 * If we race with it entering cpu, unaccounted time is 0. This is
2860 * indistinguishable from the read occurring a few cycles earlier.
2861 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2862 * been accounted, so we're correct here as well.
2864 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2865 return p
->se
.sum_exec_runtime
;
2868 rq
= task_rq_lock(p
, &flags
);
2870 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2871 * project cycles that may never be accounted to this
2872 * thread, breaking clock_gettime().
2874 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2875 update_rq_clock(rq
);
2876 p
->sched_class
->update_curr(rq
);
2878 ns
= p
->se
.sum_exec_runtime
;
2879 task_rq_unlock(rq
, p
, &flags
);
2885 * This function gets called by the timer code, with HZ frequency.
2886 * We call it with interrupts disabled.
2888 void scheduler_tick(void)
2890 int cpu
= smp_processor_id();
2891 struct rq
*rq
= cpu_rq(cpu
);
2892 struct task_struct
*curr
= rq
->curr
;
2896 raw_spin_lock(&rq
->lock
);
2897 update_rq_clock(rq
);
2898 curr
->sched_class
->task_tick(rq
, curr
, 0);
2899 update_cpu_load_active(rq
);
2900 calc_global_load_tick(rq
);
2901 raw_spin_unlock(&rq
->lock
);
2903 perf_event_task_tick();
2906 rq
->idle_balance
= idle_cpu(cpu
);
2907 trigger_load_balance(rq
);
2909 rq_last_tick_reset(rq
);
2912 #ifdef CONFIG_NO_HZ_FULL
2914 * scheduler_tick_max_deferment
2916 * Keep at least one tick per second when a single
2917 * active task is running because the scheduler doesn't
2918 * yet completely support full dynticks environment.
2920 * This makes sure that uptime, CFS vruntime, load
2921 * balancing, etc... continue to move forward, even
2922 * with a very low granularity.
2924 * Return: Maximum deferment in nanoseconds.
2926 u64
scheduler_tick_max_deferment(void)
2928 struct rq
*rq
= this_rq();
2929 unsigned long next
, now
= READ_ONCE(jiffies
);
2931 next
= rq
->last_sched_tick
+ HZ
;
2933 if (time_before_eq(next
, now
))
2936 return jiffies_to_nsecs(next
- now
);
2940 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2941 defined(CONFIG_PREEMPT_TRACER))
2943 void preempt_count_add(int val
)
2945 #ifdef CONFIG_DEBUG_PREEMPT
2949 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2952 __preempt_count_add(val
);
2953 #ifdef CONFIG_DEBUG_PREEMPT
2955 * Spinlock count overflowing soon?
2957 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2960 if (preempt_count() == val
) {
2961 unsigned long ip
= get_lock_parent_ip();
2962 #ifdef CONFIG_DEBUG_PREEMPT
2963 current
->preempt_disable_ip
= ip
;
2965 trace_preempt_off(CALLER_ADDR0
, ip
);
2968 EXPORT_SYMBOL(preempt_count_add
);
2969 NOKPROBE_SYMBOL(preempt_count_add
);
2971 void preempt_count_sub(int val
)
2973 #ifdef CONFIG_DEBUG_PREEMPT
2977 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2980 * Is the spinlock portion underflowing?
2982 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2983 !(preempt_count() & PREEMPT_MASK
)))
2987 if (preempt_count() == val
)
2988 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
2989 __preempt_count_sub(val
);
2991 EXPORT_SYMBOL(preempt_count_sub
);
2992 NOKPROBE_SYMBOL(preempt_count_sub
);
2997 * Print scheduling while atomic bug:
2999 static noinline
void __schedule_bug(struct task_struct
*prev
)
3001 if (oops_in_progress
)
3004 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3005 prev
->comm
, prev
->pid
, preempt_count());
3007 debug_show_held_locks(prev
);
3009 if (irqs_disabled())
3010 print_irqtrace_events(prev
);
3011 #ifdef CONFIG_DEBUG_PREEMPT
3012 if (in_atomic_preempt_off()) {
3013 pr_err("Preemption disabled at:");
3014 print_ip_sym(current
->preempt_disable_ip
);
3019 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3023 * Various schedule()-time debugging checks and statistics:
3025 static inline void schedule_debug(struct task_struct
*prev
)
3027 #ifdef CONFIG_SCHED_STACK_END_CHECK
3028 BUG_ON(task_stack_end_corrupted(prev
));
3031 if (unlikely(in_atomic_preempt_off())) {
3032 __schedule_bug(prev
);
3033 preempt_count_set(PREEMPT_DISABLED
);
3037 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3039 schedstat_inc(this_rq(), sched_count
);
3043 * Pick up the highest-prio task:
3045 static inline struct task_struct
*
3046 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3048 const struct sched_class
*class = &fair_sched_class
;
3049 struct task_struct
*p
;
3052 * Optimization: we know that if all tasks are in
3053 * the fair class we can call that function directly:
3055 if (likely(prev
->sched_class
== class &&
3056 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3057 p
= fair_sched_class
.pick_next_task(rq
, prev
);
3058 if (unlikely(p
== RETRY_TASK
))
3061 /* assumes fair_sched_class->next == idle_sched_class */
3063 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3069 for_each_class(class) {
3070 p
= class->pick_next_task(rq
, prev
);
3072 if (unlikely(p
== RETRY_TASK
))
3078 BUG(); /* the idle class will always have a runnable task */
3082 * __schedule() is the main scheduler function.
3084 * The main means of driving the scheduler and thus entering this function are:
3086 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3088 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3089 * paths. For example, see arch/x86/entry_64.S.
3091 * To drive preemption between tasks, the scheduler sets the flag in timer
3092 * interrupt handler scheduler_tick().
3094 * 3. Wakeups don't really cause entry into schedule(). They add a
3095 * task to the run-queue and that's it.
3097 * Now, if the new task added to the run-queue preempts the current
3098 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3099 * called on the nearest possible occasion:
3101 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3103 * - in syscall or exception context, at the next outmost
3104 * preempt_enable(). (this might be as soon as the wake_up()'s
3107 * - in IRQ context, return from interrupt-handler to
3108 * preemptible context
3110 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3113 * - cond_resched() call
3114 * - explicit schedule() call
3115 * - return from syscall or exception to user-space
3116 * - return from interrupt-handler to user-space
3118 * WARNING: must be called with preemption disabled!
3120 static void __sched notrace
__schedule(bool preempt
)
3122 struct task_struct
*prev
, *next
;
3123 unsigned long *switch_count
;
3127 cpu
= smp_processor_id();
3132 * do_exit() calls schedule() with preemption disabled as an exception;
3133 * however we must fix that up, otherwise the next task will see an
3134 * inconsistent (higher) preempt count.
3136 * It also avoids the below schedule_debug() test from complaining
3139 if (unlikely(prev
->state
== TASK_DEAD
))
3140 preempt_enable_no_resched_notrace();
3142 schedule_debug(prev
);
3144 if (sched_feat(HRTICK
))
3147 local_irq_disable();
3148 rcu_note_context_switch();
3151 * Make sure that signal_pending_state()->signal_pending() below
3152 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3153 * done by the caller to avoid the race with signal_wake_up().
3155 smp_mb__before_spinlock();
3156 raw_spin_lock(&rq
->lock
);
3157 lockdep_pin_lock(&rq
->lock
);
3159 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3161 switch_count
= &prev
->nivcsw
;
3162 if (!preempt
&& prev
->state
) {
3163 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3164 prev
->state
= TASK_RUNNING
;
3166 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3170 * If a worker went to sleep, notify and ask workqueue
3171 * whether it wants to wake up a task to maintain
3174 if (prev
->flags
& PF_WQ_WORKER
) {
3175 struct task_struct
*to_wakeup
;
3177 to_wakeup
= wq_worker_sleeping(prev
);
3179 try_to_wake_up_local(to_wakeup
);
3182 switch_count
= &prev
->nvcsw
;
3185 if (task_on_rq_queued(prev
))
3186 update_rq_clock(rq
);
3188 next
= pick_next_task(rq
, prev
);
3189 clear_tsk_need_resched(prev
);
3190 clear_preempt_need_resched();
3191 rq
->clock_skip_update
= 0;
3193 if (likely(prev
!= next
)) {
3198 trace_sched_switch(preempt
, prev
, next
);
3199 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3201 lockdep_unpin_lock(&rq
->lock
);
3202 raw_spin_unlock_irq(&rq
->lock
);
3205 balance_callback(rq
);
3208 static inline void sched_submit_work(struct task_struct
*tsk
)
3210 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3213 * If we are going to sleep and we have plugged IO queued,
3214 * make sure to submit it to avoid deadlocks.
3216 if (blk_needs_flush_plug(tsk
))
3217 blk_schedule_flush_plug(tsk
);
3220 asmlinkage __visible
void __sched
schedule(void)
3222 struct task_struct
*tsk
= current
;
3224 sched_submit_work(tsk
);
3228 sched_preempt_enable_no_resched();
3229 } while (need_resched());
3231 EXPORT_SYMBOL(schedule
);
3233 #ifdef CONFIG_CONTEXT_TRACKING
3234 asmlinkage __visible
void __sched
schedule_user(void)
3237 * If we come here after a random call to set_need_resched(),
3238 * or we have been woken up remotely but the IPI has not yet arrived,
3239 * we haven't yet exited the RCU idle mode. Do it here manually until
3240 * we find a better solution.
3242 * NB: There are buggy callers of this function. Ideally we
3243 * should warn if prev_state != CONTEXT_USER, but that will trigger
3244 * too frequently to make sense yet.
3246 enum ctx_state prev_state
= exception_enter();
3248 exception_exit(prev_state
);
3253 * schedule_preempt_disabled - called with preemption disabled
3255 * Returns with preemption disabled. Note: preempt_count must be 1
3257 void __sched
schedule_preempt_disabled(void)
3259 sched_preempt_enable_no_resched();
3264 static void __sched notrace
preempt_schedule_common(void)
3267 preempt_disable_notrace();
3269 preempt_enable_no_resched_notrace();
3272 * Check again in case we missed a preemption opportunity
3273 * between schedule and now.
3275 } while (need_resched());
3278 #ifdef CONFIG_PREEMPT
3280 * this is the entry point to schedule() from in-kernel preemption
3281 * off of preempt_enable. Kernel preemptions off return from interrupt
3282 * occur there and call schedule directly.
3284 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3287 * If there is a non-zero preempt_count or interrupts are disabled,
3288 * we do not want to preempt the current task. Just return..
3290 if (likely(!preemptible()))
3293 preempt_schedule_common();
3295 NOKPROBE_SYMBOL(preempt_schedule
);
3296 EXPORT_SYMBOL(preempt_schedule
);
3299 * preempt_schedule_notrace - preempt_schedule called by tracing
3301 * The tracing infrastructure uses preempt_enable_notrace to prevent
3302 * recursion and tracing preempt enabling caused by the tracing
3303 * infrastructure itself. But as tracing can happen in areas coming
3304 * from userspace or just about to enter userspace, a preempt enable
3305 * can occur before user_exit() is called. This will cause the scheduler
3306 * to be called when the system is still in usermode.
3308 * To prevent this, the preempt_enable_notrace will use this function
3309 * instead of preempt_schedule() to exit user context if needed before
3310 * calling the scheduler.
3312 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3314 enum ctx_state prev_ctx
;
3316 if (likely(!preemptible()))
3320 preempt_disable_notrace();
3322 * Needs preempt disabled in case user_exit() is traced
3323 * and the tracer calls preempt_enable_notrace() causing
3324 * an infinite recursion.
3326 prev_ctx
= exception_enter();
3328 exception_exit(prev_ctx
);
3330 preempt_enable_no_resched_notrace();
3331 } while (need_resched());
3333 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3335 #endif /* CONFIG_PREEMPT */
3338 * this is the entry point to schedule() from kernel preemption
3339 * off of irq context.
3340 * Note, that this is called and return with irqs disabled. This will
3341 * protect us against recursive calling from irq.
3343 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3345 enum ctx_state prev_state
;
3347 /* Catch callers which need to be fixed */
3348 BUG_ON(preempt_count() || !irqs_disabled());
3350 prev_state
= exception_enter();
3356 local_irq_disable();
3357 sched_preempt_enable_no_resched();
3358 } while (need_resched());
3360 exception_exit(prev_state
);
3363 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3366 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3368 EXPORT_SYMBOL(default_wake_function
);
3370 #ifdef CONFIG_RT_MUTEXES
3373 * rt_mutex_setprio - set the current priority of a task
3375 * @prio: prio value (kernel-internal form)
3377 * This function changes the 'effective' priority of a task. It does
3378 * not touch ->normal_prio like __setscheduler().
3380 * Used by the rt_mutex code to implement priority inheritance
3381 * logic. Call site only calls if the priority of the task changed.
3383 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3385 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3387 const struct sched_class
*prev_class
;
3389 BUG_ON(prio
> MAX_PRIO
);
3391 rq
= __task_rq_lock(p
);
3394 * Idle task boosting is a nono in general. There is one
3395 * exception, when PREEMPT_RT and NOHZ is active:
3397 * The idle task calls get_next_timer_interrupt() and holds
3398 * the timer wheel base->lock on the CPU and another CPU wants
3399 * to access the timer (probably to cancel it). We can safely
3400 * ignore the boosting request, as the idle CPU runs this code
3401 * with interrupts disabled and will complete the lock
3402 * protected section without being interrupted. So there is no
3403 * real need to boost.
3405 if (unlikely(p
== rq
->idle
)) {
3406 WARN_ON(p
!= rq
->curr
);
3407 WARN_ON(p
->pi_blocked_on
);
3411 trace_sched_pi_setprio(p
, prio
);
3414 if (oldprio
== prio
)
3415 queue_flag
&= ~DEQUEUE_MOVE
;
3417 prev_class
= p
->sched_class
;
3418 queued
= task_on_rq_queued(p
);
3419 running
= task_current(rq
, p
);
3421 dequeue_task(rq
, p
, queue_flag
);
3423 put_prev_task(rq
, p
);
3426 * Boosting condition are:
3427 * 1. -rt task is running and holds mutex A
3428 * --> -dl task blocks on mutex A
3430 * 2. -dl task is running and holds mutex A
3431 * --> -dl task blocks on mutex A and could preempt the
3434 if (dl_prio(prio
)) {
3435 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3436 if (!dl_prio(p
->normal_prio
) ||
3437 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3438 p
->dl
.dl_boosted
= 1;
3439 queue_flag
|= ENQUEUE_REPLENISH
;
3441 p
->dl
.dl_boosted
= 0;
3442 p
->sched_class
= &dl_sched_class
;
3443 } else if (rt_prio(prio
)) {
3444 if (dl_prio(oldprio
))
3445 p
->dl
.dl_boosted
= 0;
3447 queue_flag
|= ENQUEUE_HEAD
;
3448 p
->sched_class
= &rt_sched_class
;
3450 if (dl_prio(oldprio
))
3451 p
->dl
.dl_boosted
= 0;
3452 if (rt_prio(oldprio
))
3454 p
->sched_class
= &fair_sched_class
;
3460 p
->sched_class
->set_curr_task(rq
);
3462 enqueue_task(rq
, p
, queue_flag
);
3464 check_class_changed(rq
, p
, prev_class
, oldprio
);
3466 preempt_disable(); /* avoid rq from going away on us */
3467 __task_rq_unlock(rq
);
3469 balance_callback(rq
);
3474 void set_user_nice(struct task_struct
*p
, long nice
)
3476 int old_prio
, delta
, queued
;
3477 unsigned long flags
;
3480 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3483 * We have to be careful, if called from sys_setpriority(),
3484 * the task might be in the middle of scheduling on another CPU.
3486 rq
= task_rq_lock(p
, &flags
);
3488 * The RT priorities are set via sched_setscheduler(), but we still
3489 * allow the 'normal' nice value to be set - but as expected
3490 * it wont have any effect on scheduling until the task is
3491 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3493 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3494 p
->static_prio
= NICE_TO_PRIO(nice
);
3497 queued
= task_on_rq_queued(p
);
3499 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3501 p
->static_prio
= NICE_TO_PRIO(nice
);
3504 p
->prio
= effective_prio(p
);
3505 delta
= p
->prio
- old_prio
;
3508 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3510 * If the task increased its priority or is running and
3511 * lowered its priority, then reschedule its CPU:
3513 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3517 task_rq_unlock(rq
, p
, &flags
);
3519 EXPORT_SYMBOL(set_user_nice
);
3522 * can_nice - check if a task can reduce its nice value
3526 int can_nice(const struct task_struct
*p
, const int nice
)
3528 /* convert nice value [19,-20] to rlimit style value [1,40] */
3529 int nice_rlim
= nice_to_rlimit(nice
);
3531 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3532 capable(CAP_SYS_NICE
));
3535 #ifdef __ARCH_WANT_SYS_NICE
3538 * sys_nice - change the priority of the current process.
3539 * @increment: priority increment
3541 * sys_setpriority is a more generic, but much slower function that
3542 * does similar things.
3544 SYSCALL_DEFINE1(nice
, int, increment
)
3549 * Setpriority might change our priority at the same moment.
3550 * We don't have to worry. Conceptually one call occurs first
3551 * and we have a single winner.
3553 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3554 nice
= task_nice(current
) + increment
;
3556 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3557 if (increment
< 0 && !can_nice(current
, nice
))
3560 retval
= security_task_setnice(current
, nice
);
3564 set_user_nice(current
, nice
);
3571 * task_prio - return the priority value of a given task.
3572 * @p: the task in question.
3574 * Return: The priority value as seen by users in /proc.
3575 * RT tasks are offset by -200. Normal tasks are centered
3576 * around 0, value goes from -16 to +15.
3578 int task_prio(const struct task_struct
*p
)
3580 return p
->prio
- MAX_RT_PRIO
;
3584 * idle_cpu - is a given cpu idle currently?
3585 * @cpu: the processor in question.
3587 * Return: 1 if the CPU is currently idle. 0 otherwise.
3589 int idle_cpu(int cpu
)
3591 struct rq
*rq
= cpu_rq(cpu
);
3593 if (rq
->curr
!= rq
->idle
)
3600 if (!llist_empty(&rq
->wake_list
))
3608 * idle_task - return the idle task for a given cpu.
3609 * @cpu: the processor in question.
3611 * Return: The idle task for the cpu @cpu.
3613 struct task_struct
*idle_task(int cpu
)
3615 return cpu_rq(cpu
)->idle
;
3619 * find_process_by_pid - find a process with a matching PID value.
3620 * @pid: the pid in question.
3622 * The task of @pid, if found. %NULL otherwise.
3624 static struct task_struct
*find_process_by_pid(pid_t pid
)
3626 return pid
? find_task_by_vpid(pid
) : current
;
3630 * This function initializes the sched_dl_entity of a newly becoming
3631 * SCHED_DEADLINE task.
3633 * Only the static values are considered here, the actual runtime and the
3634 * absolute deadline will be properly calculated when the task is enqueued
3635 * for the first time with its new policy.
3638 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3640 struct sched_dl_entity
*dl_se
= &p
->dl
;
3642 dl_se
->dl_runtime
= attr
->sched_runtime
;
3643 dl_se
->dl_deadline
= attr
->sched_deadline
;
3644 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3645 dl_se
->flags
= attr
->sched_flags
;
3646 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3649 * Changing the parameters of a task is 'tricky' and we're not doing
3650 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3652 * What we SHOULD do is delay the bandwidth release until the 0-lag
3653 * point. This would include retaining the task_struct until that time
3654 * and change dl_overflow() to not immediately decrement the current
3657 * Instead we retain the current runtime/deadline and let the new
3658 * parameters take effect after the current reservation period lapses.
3659 * This is safe (albeit pessimistic) because the 0-lag point is always
3660 * before the current scheduling deadline.
3662 * We can still have temporary overloads because we do not delay the
3663 * change in bandwidth until that time; so admission control is
3664 * not on the safe side. It does however guarantee tasks will never
3665 * consume more than promised.
3670 * sched_setparam() passes in -1 for its policy, to let the functions
3671 * it calls know not to change it.
3673 #define SETPARAM_POLICY -1
3675 static void __setscheduler_params(struct task_struct
*p
,
3676 const struct sched_attr
*attr
)
3678 int policy
= attr
->sched_policy
;
3680 if (policy
== SETPARAM_POLICY
)
3685 if (dl_policy(policy
))
3686 __setparam_dl(p
, attr
);
3687 else if (fair_policy(policy
))
3688 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3691 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3692 * !rt_policy. Always setting this ensures that things like
3693 * getparam()/getattr() don't report silly values for !rt tasks.
3695 p
->rt_priority
= attr
->sched_priority
;
3696 p
->normal_prio
= normal_prio(p
);
3700 /* Actually do priority change: must hold pi & rq lock. */
3701 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3702 const struct sched_attr
*attr
, bool keep_boost
)
3704 __setscheduler_params(p
, attr
);
3707 * Keep a potential priority boosting if called from
3708 * sched_setscheduler().
3711 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3713 p
->prio
= normal_prio(p
);
3715 if (dl_prio(p
->prio
))
3716 p
->sched_class
= &dl_sched_class
;
3717 else if (rt_prio(p
->prio
))
3718 p
->sched_class
= &rt_sched_class
;
3720 p
->sched_class
= &fair_sched_class
;
3724 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3726 struct sched_dl_entity
*dl_se
= &p
->dl
;
3728 attr
->sched_priority
= p
->rt_priority
;
3729 attr
->sched_runtime
= dl_se
->dl_runtime
;
3730 attr
->sched_deadline
= dl_se
->dl_deadline
;
3731 attr
->sched_period
= dl_se
->dl_period
;
3732 attr
->sched_flags
= dl_se
->flags
;
3736 * This function validates the new parameters of a -deadline task.
3737 * We ask for the deadline not being zero, and greater or equal
3738 * than the runtime, as well as the period of being zero or
3739 * greater than deadline. Furthermore, we have to be sure that
3740 * user parameters are above the internal resolution of 1us (we
3741 * check sched_runtime only since it is always the smaller one) and
3742 * below 2^63 ns (we have to check both sched_deadline and
3743 * sched_period, as the latter can be zero).
3746 __checkparam_dl(const struct sched_attr
*attr
)
3749 if (attr
->sched_deadline
== 0)
3753 * Since we truncate DL_SCALE bits, make sure we're at least
3756 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3760 * Since we use the MSB for wrap-around and sign issues, make
3761 * sure it's not set (mind that period can be equal to zero).
3763 if (attr
->sched_deadline
& (1ULL << 63) ||
3764 attr
->sched_period
& (1ULL << 63))
3767 /* runtime <= deadline <= period (if period != 0) */
3768 if ((attr
->sched_period
!= 0 &&
3769 attr
->sched_period
< attr
->sched_deadline
) ||
3770 attr
->sched_deadline
< attr
->sched_runtime
)
3777 * check the target process has a UID that matches the current process's
3779 static bool check_same_owner(struct task_struct
*p
)
3781 const struct cred
*cred
= current_cred(), *pcred
;
3785 pcred
= __task_cred(p
);
3786 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3787 uid_eq(cred
->euid
, pcred
->uid
));
3792 static bool dl_param_changed(struct task_struct
*p
,
3793 const struct sched_attr
*attr
)
3795 struct sched_dl_entity
*dl_se
= &p
->dl
;
3797 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3798 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3799 dl_se
->dl_period
!= attr
->sched_period
||
3800 dl_se
->flags
!= attr
->sched_flags
)
3806 static int __sched_setscheduler(struct task_struct
*p
,
3807 const struct sched_attr
*attr
,
3810 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3811 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3812 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3813 int new_effective_prio
, policy
= attr
->sched_policy
;
3814 unsigned long flags
;
3815 const struct sched_class
*prev_class
;
3818 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3820 /* may grab non-irq protected spin_locks */
3821 BUG_ON(in_interrupt());
3823 /* double check policy once rq lock held */
3825 reset_on_fork
= p
->sched_reset_on_fork
;
3826 policy
= oldpolicy
= p
->policy
;
3828 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3830 if (!valid_policy(policy
))
3834 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3838 * Valid priorities for SCHED_FIFO and SCHED_RR are
3839 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3840 * SCHED_BATCH and SCHED_IDLE is 0.
3842 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3843 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3845 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3846 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3850 * Allow unprivileged RT tasks to decrease priority:
3852 if (user
&& !capable(CAP_SYS_NICE
)) {
3853 if (fair_policy(policy
)) {
3854 if (attr
->sched_nice
< task_nice(p
) &&
3855 !can_nice(p
, attr
->sched_nice
))
3859 if (rt_policy(policy
)) {
3860 unsigned long rlim_rtprio
=
3861 task_rlimit(p
, RLIMIT_RTPRIO
);
3863 /* can't set/change the rt policy */
3864 if (policy
!= p
->policy
&& !rlim_rtprio
)
3867 /* can't increase priority */
3868 if (attr
->sched_priority
> p
->rt_priority
&&
3869 attr
->sched_priority
> rlim_rtprio
)
3874 * Can't set/change SCHED_DEADLINE policy at all for now
3875 * (safest behavior); in the future we would like to allow
3876 * unprivileged DL tasks to increase their relative deadline
3877 * or reduce their runtime (both ways reducing utilization)
3879 if (dl_policy(policy
))
3883 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3884 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3886 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
3887 if (!can_nice(p
, task_nice(p
)))
3891 /* can't change other user's priorities */
3892 if (!check_same_owner(p
))
3895 /* Normal users shall not reset the sched_reset_on_fork flag */
3896 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3901 retval
= security_task_setscheduler(p
);
3907 * make sure no PI-waiters arrive (or leave) while we are
3908 * changing the priority of the task:
3910 * To be able to change p->policy safely, the appropriate
3911 * runqueue lock must be held.
3913 rq
= task_rq_lock(p
, &flags
);
3916 * Changing the policy of the stop threads its a very bad idea
3918 if (p
== rq
->stop
) {
3919 task_rq_unlock(rq
, p
, &flags
);
3924 * If not changing anything there's no need to proceed further,
3925 * but store a possible modification of reset_on_fork.
3927 if (unlikely(policy
== p
->policy
)) {
3928 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3930 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3932 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3935 p
->sched_reset_on_fork
= reset_on_fork
;
3936 task_rq_unlock(rq
, p
, &flags
);
3942 #ifdef CONFIG_RT_GROUP_SCHED
3944 * Do not allow realtime tasks into groups that have no runtime
3947 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3948 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3949 !task_group_is_autogroup(task_group(p
))) {
3950 task_rq_unlock(rq
, p
, &flags
);
3955 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3956 cpumask_t
*span
= rq
->rd
->span
;
3959 * Don't allow tasks with an affinity mask smaller than
3960 * the entire root_domain to become SCHED_DEADLINE. We
3961 * will also fail if there's no bandwidth available.
3963 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3964 rq
->rd
->dl_bw
.bw
== 0) {
3965 task_rq_unlock(rq
, p
, &flags
);
3972 /* recheck policy now with rq lock held */
3973 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3974 policy
= oldpolicy
= -1;
3975 task_rq_unlock(rq
, p
, &flags
);
3980 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3981 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3984 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3985 task_rq_unlock(rq
, p
, &flags
);
3989 p
->sched_reset_on_fork
= reset_on_fork
;
3994 * Take priority boosted tasks into account. If the new
3995 * effective priority is unchanged, we just store the new
3996 * normal parameters and do not touch the scheduler class and
3997 * the runqueue. This will be done when the task deboost
4000 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4001 if (new_effective_prio
== oldprio
)
4002 queue_flags
&= ~DEQUEUE_MOVE
;
4005 queued
= task_on_rq_queued(p
);
4006 running
= task_current(rq
, p
);
4008 dequeue_task(rq
, p
, queue_flags
);
4010 put_prev_task(rq
, p
);
4012 prev_class
= p
->sched_class
;
4013 __setscheduler(rq
, p
, attr
, pi
);
4016 p
->sched_class
->set_curr_task(rq
);
4019 * We enqueue to tail when the priority of a task is
4020 * increased (user space view).
4022 if (oldprio
< p
->prio
)
4023 queue_flags
|= ENQUEUE_HEAD
;
4025 enqueue_task(rq
, p
, queue_flags
);
4028 check_class_changed(rq
, p
, prev_class
, oldprio
);
4029 preempt_disable(); /* avoid rq from going away on us */
4030 task_rq_unlock(rq
, p
, &flags
);
4033 rt_mutex_adjust_pi(p
);
4036 * Run balance callbacks after we've adjusted the PI chain.
4038 balance_callback(rq
);
4044 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4045 const struct sched_param
*param
, bool check
)
4047 struct sched_attr attr
= {
4048 .sched_policy
= policy
,
4049 .sched_priority
= param
->sched_priority
,
4050 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4053 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4054 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4055 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4056 policy
&= ~SCHED_RESET_ON_FORK
;
4057 attr
.sched_policy
= policy
;
4060 return __sched_setscheduler(p
, &attr
, check
, true);
4063 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4064 * @p: the task in question.
4065 * @policy: new policy.
4066 * @param: structure containing the new RT priority.
4068 * Return: 0 on success. An error code otherwise.
4070 * NOTE that the task may be already dead.
4072 int sched_setscheduler(struct task_struct
*p
, int policy
,
4073 const struct sched_param
*param
)
4075 return _sched_setscheduler(p
, policy
, param
, true);
4077 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4079 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4081 return __sched_setscheduler(p
, attr
, true, true);
4083 EXPORT_SYMBOL_GPL(sched_setattr
);
4086 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4087 * @p: the task in question.
4088 * @policy: new policy.
4089 * @param: structure containing the new RT priority.
4091 * Just like sched_setscheduler, only don't bother checking if the
4092 * current context has permission. For example, this is needed in
4093 * stop_machine(): we create temporary high priority worker threads,
4094 * but our caller might not have that capability.
4096 * Return: 0 on success. An error code otherwise.
4098 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4099 const struct sched_param
*param
)
4101 return _sched_setscheduler(p
, policy
, param
, false);
4103 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4106 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4108 struct sched_param lparam
;
4109 struct task_struct
*p
;
4112 if (!param
|| pid
< 0)
4114 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4119 p
= find_process_by_pid(pid
);
4121 retval
= sched_setscheduler(p
, policy
, &lparam
);
4128 * Mimics kernel/events/core.c perf_copy_attr().
4130 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4131 struct sched_attr
*attr
)
4136 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4140 * zero the full structure, so that a short copy will be nice.
4142 memset(attr
, 0, sizeof(*attr
));
4144 ret
= get_user(size
, &uattr
->size
);
4148 if (size
> PAGE_SIZE
) /* silly large */
4151 if (!size
) /* abi compat */
4152 size
= SCHED_ATTR_SIZE_VER0
;
4154 if (size
< SCHED_ATTR_SIZE_VER0
)
4158 * If we're handed a bigger struct than we know of,
4159 * ensure all the unknown bits are 0 - i.e. new
4160 * user-space does not rely on any kernel feature
4161 * extensions we dont know about yet.
4163 if (size
> sizeof(*attr
)) {
4164 unsigned char __user
*addr
;
4165 unsigned char __user
*end
;
4168 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4169 end
= (void __user
*)uattr
+ size
;
4171 for (; addr
< end
; addr
++) {
4172 ret
= get_user(val
, addr
);
4178 size
= sizeof(*attr
);
4181 ret
= copy_from_user(attr
, uattr
, size
);
4186 * XXX: do we want to be lenient like existing syscalls; or do we want
4187 * to be strict and return an error on out-of-bounds values?
4189 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4194 put_user(sizeof(*attr
), &uattr
->size
);
4199 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4200 * @pid: the pid in question.
4201 * @policy: new policy.
4202 * @param: structure containing the new RT priority.
4204 * Return: 0 on success. An error code otherwise.
4206 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4207 struct sched_param __user
*, param
)
4209 /* negative values for policy are not valid */
4213 return do_sched_setscheduler(pid
, policy
, param
);
4217 * sys_sched_setparam - set/change the RT priority of a thread
4218 * @pid: the pid in question.
4219 * @param: structure containing the new RT priority.
4221 * Return: 0 on success. An error code otherwise.
4223 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4225 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4229 * sys_sched_setattr - same as above, but with extended sched_attr
4230 * @pid: the pid in question.
4231 * @uattr: structure containing the extended parameters.
4232 * @flags: for future extension.
4234 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4235 unsigned int, flags
)
4237 struct sched_attr attr
;
4238 struct task_struct
*p
;
4241 if (!uattr
|| pid
< 0 || flags
)
4244 retval
= sched_copy_attr(uattr
, &attr
);
4248 if ((int)attr
.sched_policy
< 0)
4253 p
= find_process_by_pid(pid
);
4255 retval
= sched_setattr(p
, &attr
);
4262 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4263 * @pid: the pid in question.
4265 * Return: On success, the policy of the thread. Otherwise, a negative error
4268 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4270 struct task_struct
*p
;
4278 p
= find_process_by_pid(pid
);
4280 retval
= security_task_getscheduler(p
);
4283 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4290 * sys_sched_getparam - get the RT priority of a thread
4291 * @pid: the pid in question.
4292 * @param: structure containing the RT priority.
4294 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4297 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4299 struct sched_param lp
= { .sched_priority
= 0 };
4300 struct task_struct
*p
;
4303 if (!param
|| pid
< 0)
4307 p
= find_process_by_pid(pid
);
4312 retval
= security_task_getscheduler(p
);
4316 if (task_has_rt_policy(p
))
4317 lp
.sched_priority
= p
->rt_priority
;
4321 * This one might sleep, we cannot do it with a spinlock held ...
4323 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4332 static int sched_read_attr(struct sched_attr __user
*uattr
,
4333 struct sched_attr
*attr
,
4338 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4342 * If we're handed a smaller struct than we know of,
4343 * ensure all the unknown bits are 0 - i.e. old
4344 * user-space does not get uncomplete information.
4346 if (usize
< sizeof(*attr
)) {
4347 unsigned char *addr
;
4350 addr
= (void *)attr
+ usize
;
4351 end
= (void *)attr
+ sizeof(*attr
);
4353 for (; addr
< end
; addr
++) {
4361 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4369 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4370 * @pid: the pid in question.
4371 * @uattr: structure containing the extended parameters.
4372 * @size: sizeof(attr) for fwd/bwd comp.
4373 * @flags: for future extension.
4375 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4376 unsigned int, size
, unsigned int, flags
)
4378 struct sched_attr attr
= {
4379 .size
= sizeof(struct sched_attr
),
4381 struct task_struct
*p
;
4384 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4385 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4389 p
= find_process_by_pid(pid
);
4394 retval
= security_task_getscheduler(p
);
4398 attr
.sched_policy
= p
->policy
;
4399 if (p
->sched_reset_on_fork
)
4400 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4401 if (task_has_dl_policy(p
))
4402 __getparam_dl(p
, &attr
);
4403 else if (task_has_rt_policy(p
))
4404 attr
.sched_priority
= p
->rt_priority
;
4406 attr
.sched_nice
= task_nice(p
);
4410 retval
= sched_read_attr(uattr
, &attr
, size
);
4418 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4420 cpumask_var_t cpus_allowed
, new_mask
;
4421 struct task_struct
*p
;
4426 p
= find_process_by_pid(pid
);
4432 /* Prevent p going away */
4436 if (p
->flags
& PF_NO_SETAFFINITY
) {
4440 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4444 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4446 goto out_free_cpus_allowed
;
4449 if (!check_same_owner(p
)) {
4451 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4453 goto out_free_new_mask
;
4458 retval
= security_task_setscheduler(p
);
4460 goto out_free_new_mask
;
4463 cpuset_cpus_allowed(p
, cpus_allowed
);
4464 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4467 * Since bandwidth control happens on root_domain basis,
4468 * if admission test is enabled, we only admit -deadline
4469 * tasks allowed to run on all the CPUs in the task's
4473 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4475 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4478 goto out_free_new_mask
;
4484 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4487 cpuset_cpus_allowed(p
, cpus_allowed
);
4488 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4490 * We must have raced with a concurrent cpuset
4491 * update. Just reset the cpus_allowed to the
4492 * cpuset's cpus_allowed
4494 cpumask_copy(new_mask
, cpus_allowed
);
4499 free_cpumask_var(new_mask
);
4500 out_free_cpus_allowed
:
4501 free_cpumask_var(cpus_allowed
);
4507 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4508 struct cpumask
*new_mask
)
4510 if (len
< cpumask_size())
4511 cpumask_clear(new_mask
);
4512 else if (len
> cpumask_size())
4513 len
= cpumask_size();
4515 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4519 * sys_sched_setaffinity - set the cpu affinity of a process
4520 * @pid: pid of the process
4521 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4522 * @user_mask_ptr: user-space pointer to the new cpu mask
4524 * Return: 0 on success. An error code otherwise.
4526 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4527 unsigned long __user
*, user_mask_ptr
)
4529 cpumask_var_t new_mask
;
4532 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4535 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4537 retval
= sched_setaffinity(pid
, new_mask
);
4538 free_cpumask_var(new_mask
);
4542 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4544 struct task_struct
*p
;
4545 unsigned long flags
;
4551 p
= find_process_by_pid(pid
);
4555 retval
= security_task_getscheduler(p
);
4559 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4560 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4561 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4570 * sys_sched_getaffinity - get the cpu affinity of a process
4571 * @pid: pid of the process
4572 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4573 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4575 * Return: 0 on success. An error code otherwise.
4577 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4578 unsigned long __user
*, user_mask_ptr
)
4583 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4585 if (len
& (sizeof(unsigned long)-1))
4588 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4591 ret
= sched_getaffinity(pid
, mask
);
4593 size_t retlen
= min_t(size_t, len
, cpumask_size());
4595 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4600 free_cpumask_var(mask
);
4606 * sys_sched_yield - yield the current processor to other threads.
4608 * This function yields the current CPU to other tasks. If there are no
4609 * other threads running on this CPU then this function will return.
4613 SYSCALL_DEFINE0(sched_yield
)
4615 struct rq
*rq
= this_rq_lock();
4617 schedstat_inc(rq
, yld_count
);
4618 current
->sched_class
->yield_task(rq
);
4621 * Since we are going to call schedule() anyway, there's
4622 * no need to preempt or enable interrupts:
4624 __release(rq
->lock
);
4625 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4626 do_raw_spin_unlock(&rq
->lock
);
4627 sched_preempt_enable_no_resched();
4634 int __sched
_cond_resched(void)
4636 if (should_resched(0)) {
4637 preempt_schedule_common();
4642 EXPORT_SYMBOL(_cond_resched
);
4645 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4646 * call schedule, and on return reacquire the lock.
4648 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4649 * operations here to prevent schedule() from being called twice (once via
4650 * spin_unlock(), once by hand).
4652 int __cond_resched_lock(spinlock_t
*lock
)
4654 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4657 lockdep_assert_held(lock
);
4659 if (spin_needbreak(lock
) || resched
) {
4662 preempt_schedule_common();
4670 EXPORT_SYMBOL(__cond_resched_lock
);
4672 int __sched
__cond_resched_softirq(void)
4674 BUG_ON(!in_softirq());
4676 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4678 preempt_schedule_common();
4684 EXPORT_SYMBOL(__cond_resched_softirq
);
4687 * yield - yield the current processor to other threads.
4689 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4691 * The scheduler is at all times free to pick the calling task as the most
4692 * eligible task to run, if removing the yield() call from your code breaks
4693 * it, its already broken.
4695 * Typical broken usage is:
4700 * where one assumes that yield() will let 'the other' process run that will
4701 * make event true. If the current task is a SCHED_FIFO task that will never
4702 * happen. Never use yield() as a progress guarantee!!
4704 * If you want to use yield() to wait for something, use wait_event().
4705 * If you want to use yield() to be 'nice' for others, use cond_resched().
4706 * If you still want to use yield(), do not!
4708 void __sched
yield(void)
4710 set_current_state(TASK_RUNNING
);
4713 EXPORT_SYMBOL(yield
);
4716 * yield_to - yield the current processor to another thread in
4717 * your thread group, or accelerate that thread toward the
4718 * processor it's on.
4720 * @preempt: whether task preemption is allowed or not
4722 * It's the caller's job to ensure that the target task struct
4723 * can't go away on us before we can do any checks.
4726 * true (>0) if we indeed boosted the target task.
4727 * false (0) if we failed to boost the target.
4728 * -ESRCH if there's no task to yield to.
4730 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4732 struct task_struct
*curr
= current
;
4733 struct rq
*rq
, *p_rq
;
4734 unsigned long flags
;
4737 local_irq_save(flags
);
4743 * If we're the only runnable task on the rq and target rq also
4744 * has only one task, there's absolutely no point in yielding.
4746 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4751 double_rq_lock(rq
, p_rq
);
4752 if (task_rq(p
) != p_rq
) {
4753 double_rq_unlock(rq
, p_rq
);
4757 if (!curr
->sched_class
->yield_to_task
)
4760 if (curr
->sched_class
!= p
->sched_class
)
4763 if (task_running(p_rq
, p
) || p
->state
)
4766 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4768 schedstat_inc(rq
, yld_count
);
4770 * Make p's CPU reschedule; pick_next_entity takes care of
4773 if (preempt
&& rq
!= p_rq
)
4778 double_rq_unlock(rq
, p_rq
);
4780 local_irq_restore(flags
);
4787 EXPORT_SYMBOL_GPL(yield_to
);
4790 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4791 * that process accounting knows that this is a task in IO wait state.
4793 long __sched
io_schedule_timeout(long timeout
)
4795 int old_iowait
= current
->in_iowait
;
4799 current
->in_iowait
= 1;
4800 blk_schedule_flush_plug(current
);
4802 delayacct_blkio_start();
4804 atomic_inc(&rq
->nr_iowait
);
4805 ret
= schedule_timeout(timeout
);
4806 current
->in_iowait
= old_iowait
;
4807 atomic_dec(&rq
->nr_iowait
);
4808 delayacct_blkio_end();
4812 EXPORT_SYMBOL(io_schedule_timeout
);
4815 * sys_sched_get_priority_max - return maximum RT priority.
4816 * @policy: scheduling class.
4818 * Return: On success, this syscall returns the maximum
4819 * rt_priority that can be used by a given scheduling class.
4820 * On failure, a negative error code is returned.
4822 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4829 ret
= MAX_USER_RT_PRIO
-1;
4831 case SCHED_DEADLINE
:
4842 * sys_sched_get_priority_min - return minimum RT priority.
4843 * @policy: scheduling class.
4845 * Return: On success, this syscall returns the minimum
4846 * rt_priority that can be used by a given scheduling class.
4847 * On failure, a negative error code is returned.
4849 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4858 case SCHED_DEADLINE
:
4868 * sys_sched_rr_get_interval - return the default timeslice of a process.
4869 * @pid: pid of the process.
4870 * @interval: userspace pointer to the timeslice value.
4872 * this syscall writes the default timeslice value of a given process
4873 * into the user-space timespec buffer. A value of '0' means infinity.
4875 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4878 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4879 struct timespec __user
*, interval
)
4881 struct task_struct
*p
;
4882 unsigned int time_slice
;
4883 unsigned long flags
;
4893 p
= find_process_by_pid(pid
);
4897 retval
= security_task_getscheduler(p
);
4901 rq
= task_rq_lock(p
, &flags
);
4903 if (p
->sched_class
->get_rr_interval
)
4904 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4905 task_rq_unlock(rq
, p
, &flags
);
4908 jiffies_to_timespec(time_slice
, &t
);
4909 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4917 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4919 void sched_show_task(struct task_struct
*p
)
4921 unsigned long free
= 0;
4923 unsigned long state
= p
->state
;
4926 state
= __ffs(state
) + 1;
4927 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4928 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4929 #if BITS_PER_LONG == 32
4930 if (state
== TASK_RUNNING
)
4931 printk(KERN_CONT
" running ");
4933 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4935 if (state
== TASK_RUNNING
)
4936 printk(KERN_CONT
" running task ");
4938 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4940 #ifdef CONFIG_DEBUG_STACK_USAGE
4941 free
= stack_not_used(p
);
4946 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4948 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4949 task_pid_nr(p
), ppid
,
4950 (unsigned long)task_thread_info(p
)->flags
);
4952 print_worker_info(KERN_INFO
, p
);
4953 show_stack(p
, NULL
);
4956 void show_state_filter(unsigned long state_filter
)
4958 struct task_struct
*g
, *p
;
4960 #if BITS_PER_LONG == 32
4962 " task PC stack pid father\n");
4965 " task PC stack pid father\n");
4968 for_each_process_thread(g
, p
) {
4970 * reset the NMI-timeout, listing all files on a slow
4971 * console might take a lot of time:
4973 touch_nmi_watchdog();
4974 if (!state_filter
|| (p
->state
& state_filter
))
4978 touch_all_softlockup_watchdogs();
4980 #ifdef CONFIG_SCHED_DEBUG
4981 sysrq_sched_debug_show();
4985 * Only show locks if all tasks are dumped:
4988 debug_show_all_locks();
4991 void init_idle_bootup_task(struct task_struct
*idle
)
4993 idle
->sched_class
= &idle_sched_class
;
4997 * init_idle - set up an idle thread for a given CPU
4998 * @idle: task in question
4999 * @cpu: cpu the idle task belongs to
5001 * NOTE: this function does not set the idle thread's NEED_RESCHED
5002 * flag, to make booting more robust.
5004 void init_idle(struct task_struct
*idle
, int cpu
)
5006 struct rq
*rq
= cpu_rq(cpu
);
5007 unsigned long flags
;
5009 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5010 raw_spin_lock(&rq
->lock
);
5012 __sched_fork(0, idle
);
5013 idle
->state
= TASK_RUNNING
;
5014 idle
->se
.exec_start
= sched_clock();
5016 kasan_unpoison_task_stack(idle
);
5020 * Its possible that init_idle() gets called multiple times on a task,
5021 * in that case do_set_cpus_allowed() will not do the right thing.
5023 * And since this is boot we can forgo the serialization.
5025 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5028 * We're having a chicken and egg problem, even though we are
5029 * holding rq->lock, the cpu isn't yet set to this cpu so the
5030 * lockdep check in task_group() will fail.
5032 * Similar case to sched_fork(). / Alternatively we could
5033 * use task_rq_lock() here and obtain the other rq->lock.
5038 __set_task_cpu(idle
, cpu
);
5041 rq
->curr
= rq
->idle
= idle
;
5042 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5046 raw_spin_unlock(&rq
->lock
);
5047 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5049 /* Set the preempt count _outside_ the spinlocks! */
5050 init_idle_preempt_count(idle
, cpu
);
5053 * The idle tasks have their own, simple scheduling class:
5055 idle
->sched_class
= &idle_sched_class
;
5056 ftrace_graph_init_idle_task(idle
, cpu
);
5057 vtime_init_idle(idle
, cpu
);
5059 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5063 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5064 const struct cpumask
*trial
)
5066 int ret
= 1, trial_cpus
;
5067 struct dl_bw
*cur_dl_b
;
5068 unsigned long flags
;
5070 if (!cpumask_weight(cur
))
5073 rcu_read_lock_sched();
5074 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5075 trial_cpus
= cpumask_weight(trial
);
5077 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5078 if (cur_dl_b
->bw
!= -1 &&
5079 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5081 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5082 rcu_read_unlock_sched();
5087 int task_can_attach(struct task_struct
*p
,
5088 const struct cpumask
*cs_cpus_allowed
)
5093 * Kthreads which disallow setaffinity shouldn't be moved
5094 * to a new cpuset; we don't want to change their cpu
5095 * affinity and isolating such threads by their set of
5096 * allowed nodes is unnecessary. Thus, cpusets are not
5097 * applicable for such threads. This prevents checking for
5098 * success of set_cpus_allowed_ptr() on all attached tasks
5099 * before cpus_allowed may be changed.
5101 if (p
->flags
& PF_NO_SETAFFINITY
) {
5107 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5109 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5114 unsigned long flags
;
5116 rcu_read_lock_sched();
5117 dl_b
= dl_bw_of(dest_cpu
);
5118 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5119 cpus
= dl_bw_cpus(dest_cpu
);
5120 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5125 * We reserve space for this task in the destination
5126 * root_domain, as we can't fail after this point.
5127 * We will free resources in the source root_domain
5128 * later on (see set_cpus_allowed_dl()).
5130 __dl_add(dl_b
, p
->dl
.dl_bw
);
5132 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5133 rcu_read_unlock_sched();
5143 #ifdef CONFIG_NUMA_BALANCING
5144 /* Migrate current task p to target_cpu */
5145 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5147 struct migration_arg arg
= { p
, target_cpu
};
5148 int curr_cpu
= task_cpu(p
);
5150 if (curr_cpu
== target_cpu
)
5153 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5156 /* TODO: This is not properly updating schedstats */
5158 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5159 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5163 * Requeue a task on a given node and accurately track the number of NUMA
5164 * tasks on the runqueues
5166 void sched_setnuma(struct task_struct
*p
, int nid
)
5169 unsigned long flags
;
5170 bool queued
, running
;
5172 rq
= task_rq_lock(p
, &flags
);
5173 queued
= task_on_rq_queued(p
);
5174 running
= task_current(rq
, p
);
5177 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5179 put_prev_task(rq
, p
);
5181 p
->numa_preferred_nid
= nid
;
5184 p
->sched_class
->set_curr_task(rq
);
5186 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5187 task_rq_unlock(rq
, p
, &flags
);
5189 #endif /* CONFIG_NUMA_BALANCING */
5191 #ifdef CONFIG_HOTPLUG_CPU
5193 * Ensures that the idle task is using init_mm right before its cpu goes
5196 void idle_task_exit(void)
5198 struct mm_struct
*mm
= current
->active_mm
;
5200 BUG_ON(cpu_online(smp_processor_id()));
5202 if (mm
!= &init_mm
) {
5203 switch_mm(mm
, &init_mm
, current
);
5204 finish_arch_post_lock_switch();
5210 * Since this CPU is going 'away' for a while, fold any nr_active delta
5211 * we might have. Assumes we're called after migrate_tasks() so that the
5212 * nr_active count is stable.
5214 * Also see the comment "Global load-average calculations".
5216 static void calc_load_migrate(struct rq
*rq
)
5218 long delta
= calc_load_fold_active(rq
);
5220 atomic_long_add(delta
, &calc_load_tasks
);
5223 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5227 static const struct sched_class fake_sched_class
= {
5228 .put_prev_task
= put_prev_task_fake
,
5231 static struct task_struct fake_task
= {
5233 * Avoid pull_{rt,dl}_task()
5235 .prio
= MAX_PRIO
+ 1,
5236 .sched_class
= &fake_sched_class
,
5240 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5241 * try_to_wake_up()->select_task_rq().
5243 * Called with rq->lock held even though we'er in stop_machine() and
5244 * there's no concurrency possible, we hold the required locks anyway
5245 * because of lock validation efforts.
5247 static void migrate_tasks(struct rq
*dead_rq
)
5249 struct rq
*rq
= dead_rq
;
5250 struct task_struct
*next
, *stop
= rq
->stop
;
5254 * Fudge the rq selection such that the below task selection loop
5255 * doesn't get stuck on the currently eligible stop task.
5257 * We're currently inside stop_machine() and the rq is either stuck
5258 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5259 * either way we should never end up calling schedule() until we're
5265 * put_prev_task() and pick_next_task() sched
5266 * class method both need to have an up-to-date
5267 * value of rq->clock[_task]
5269 update_rq_clock(rq
);
5273 * There's this thread running, bail when that's the only
5276 if (rq
->nr_running
== 1)
5280 * pick_next_task assumes pinned rq->lock.
5282 lockdep_pin_lock(&rq
->lock
);
5283 next
= pick_next_task(rq
, &fake_task
);
5285 next
->sched_class
->put_prev_task(rq
, next
);
5288 * Rules for changing task_struct::cpus_allowed are holding
5289 * both pi_lock and rq->lock, such that holding either
5290 * stabilizes the mask.
5292 * Drop rq->lock is not quite as disastrous as it usually is
5293 * because !cpu_active at this point, which means load-balance
5294 * will not interfere. Also, stop-machine.
5296 lockdep_unpin_lock(&rq
->lock
);
5297 raw_spin_unlock(&rq
->lock
);
5298 raw_spin_lock(&next
->pi_lock
);
5299 raw_spin_lock(&rq
->lock
);
5302 * Since we're inside stop-machine, _nothing_ should have
5303 * changed the task, WARN if weird stuff happened, because in
5304 * that case the above rq->lock drop is a fail too.
5306 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5307 raw_spin_unlock(&next
->pi_lock
);
5311 /* Find suitable destination for @next, with force if needed. */
5312 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5314 rq
= __migrate_task(rq
, next
, dest_cpu
);
5315 if (rq
!= dead_rq
) {
5316 raw_spin_unlock(&rq
->lock
);
5318 raw_spin_lock(&rq
->lock
);
5320 raw_spin_unlock(&next
->pi_lock
);
5325 #endif /* CONFIG_HOTPLUG_CPU */
5327 static void set_rq_online(struct rq
*rq
)
5330 const struct sched_class
*class;
5332 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5335 for_each_class(class) {
5336 if (class->rq_online
)
5337 class->rq_online(rq
);
5342 static void set_rq_offline(struct rq
*rq
)
5345 const struct sched_class
*class;
5347 for_each_class(class) {
5348 if (class->rq_offline
)
5349 class->rq_offline(rq
);
5352 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5358 * migration_call - callback that gets triggered when a CPU is added.
5359 * Here we can start up the necessary migration thread for the new CPU.
5362 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5364 int cpu
= (long)hcpu
;
5365 unsigned long flags
;
5366 struct rq
*rq
= cpu_rq(cpu
);
5368 switch (action
& ~CPU_TASKS_FROZEN
) {
5370 case CPU_UP_PREPARE
:
5371 rq
->calc_load_update
= calc_load_update
;
5375 /* Update our root-domain */
5376 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5378 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5382 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5385 #ifdef CONFIG_HOTPLUG_CPU
5387 sched_ttwu_pending();
5388 /* Update our root-domain */
5389 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5391 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5395 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5396 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5400 calc_load_migrate(rq
);
5405 update_max_interval();
5411 * Register at high priority so that task migration (migrate_all_tasks)
5412 * happens before everything else. This has to be lower priority than
5413 * the notifier in the perf_event subsystem, though.
5415 static struct notifier_block migration_notifier
= {
5416 .notifier_call
= migration_call
,
5417 .priority
= CPU_PRI_MIGRATION
,
5420 static void set_cpu_rq_start_time(void)
5422 int cpu
= smp_processor_id();
5423 struct rq
*rq
= cpu_rq(cpu
);
5424 rq
->age_stamp
= sched_clock_cpu(cpu
);
5427 static int sched_cpu_active(struct notifier_block
*nfb
,
5428 unsigned long action
, void *hcpu
)
5430 int cpu
= (long)hcpu
;
5432 switch (action
& ~CPU_TASKS_FROZEN
) {
5434 set_cpu_rq_start_time();
5437 case CPU_DOWN_FAILED
:
5438 set_cpu_active(cpu
, true);
5446 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5447 unsigned long action
, void *hcpu
)
5449 switch (action
& ~CPU_TASKS_FROZEN
) {
5450 case CPU_DOWN_PREPARE
:
5451 set_cpu_active((long)hcpu
, false);
5458 static int __init
migration_init(void)
5460 void *cpu
= (void *)(long)smp_processor_id();
5463 /* Initialize migration for the boot CPU */
5464 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5465 BUG_ON(err
== NOTIFY_BAD
);
5466 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5467 register_cpu_notifier(&migration_notifier
);
5469 /* Register cpu active notifiers */
5470 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5471 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5475 early_initcall(migration_init
);
5477 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5479 #ifdef CONFIG_SCHED_DEBUG
5481 static __read_mostly
int sched_debug_enabled
;
5483 static int __init
sched_debug_setup(char *str
)
5485 sched_debug_enabled
= 1;
5489 early_param("sched_debug", sched_debug_setup
);
5491 static inline bool sched_debug(void)
5493 return sched_debug_enabled
;
5496 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5497 struct cpumask
*groupmask
)
5499 struct sched_group
*group
= sd
->groups
;
5501 cpumask_clear(groupmask
);
5503 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5505 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5506 printk("does not load-balance\n");
5508 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5513 printk(KERN_CONT
"span %*pbl level %s\n",
5514 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5516 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5517 printk(KERN_ERR
"ERROR: domain->span does not contain "
5520 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5521 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5525 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5529 printk(KERN_ERR
"ERROR: group is NULL\n");
5533 if (!cpumask_weight(sched_group_cpus(group
))) {
5534 printk(KERN_CONT
"\n");
5535 printk(KERN_ERR
"ERROR: empty group\n");
5539 if (!(sd
->flags
& SD_OVERLAP
) &&
5540 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5541 printk(KERN_CONT
"\n");
5542 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5546 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5548 printk(KERN_CONT
" %*pbl",
5549 cpumask_pr_args(sched_group_cpus(group
)));
5550 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5551 printk(KERN_CONT
" (cpu_capacity = %d)",
5552 group
->sgc
->capacity
);
5555 group
= group
->next
;
5556 } while (group
!= sd
->groups
);
5557 printk(KERN_CONT
"\n");
5559 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5560 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5563 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5564 printk(KERN_ERR
"ERROR: parent span is not a superset "
5565 "of domain->span\n");
5569 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5573 if (!sched_debug_enabled
)
5577 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5581 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5584 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5592 #else /* !CONFIG_SCHED_DEBUG */
5593 # define sched_domain_debug(sd, cpu) do { } while (0)
5594 static inline bool sched_debug(void)
5598 #endif /* CONFIG_SCHED_DEBUG */
5600 static int sd_degenerate(struct sched_domain
*sd
)
5602 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5605 /* Following flags need at least 2 groups */
5606 if (sd
->flags
& (SD_LOAD_BALANCE
|
5607 SD_BALANCE_NEWIDLE
|
5610 SD_SHARE_CPUCAPACITY
|
5611 SD_SHARE_PKG_RESOURCES
|
5612 SD_SHARE_POWERDOMAIN
)) {
5613 if (sd
->groups
!= sd
->groups
->next
)
5617 /* Following flags don't use groups */
5618 if (sd
->flags
& (SD_WAKE_AFFINE
))
5625 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5627 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5629 if (sd_degenerate(parent
))
5632 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5635 /* Flags needing groups don't count if only 1 group in parent */
5636 if (parent
->groups
== parent
->groups
->next
) {
5637 pflags
&= ~(SD_LOAD_BALANCE
|
5638 SD_BALANCE_NEWIDLE
|
5641 SD_SHARE_CPUCAPACITY
|
5642 SD_SHARE_PKG_RESOURCES
|
5644 SD_SHARE_POWERDOMAIN
);
5645 if (nr_node_ids
== 1)
5646 pflags
&= ~SD_SERIALIZE
;
5648 if (~cflags
& pflags
)
5654 static void free_rootdomain(struct rcu_head
*rcu
)
5656 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5658 cpupri_cleanup(&rd
->cpupri
);
5659 cpudl_cleanup(&rd
->cpudl
);
5660 free_cpumask_var(rd
->dlo_mask
);
5661 free_cpumask_var(rd
->rto_mask
);
5662 free_cpumask_var(rd
->online
);
5663 free_cpumask_var(rd
->span
);
5667 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5669 struct root_domain
*old_rd
= NULL
;
5670 unsigned long flags
;
5672 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5677 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5680 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5683 * If we dont want to free the old_rd yet then
5684 * set old_rd to NULL to skip the freeing later
5687 if (!atomic_dec_and_test(&old_rd
->refcount
))
5691 atomic_inc(&rd
->refcount
);
5694 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5695 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5698 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5701 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5704 static int init_rootdomain(struct root_domain
*rd
)
5706 memset(rd
, 0, sizeof(*rd
));
5708 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5710 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5712 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5714 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5717 init_dl_bw(&rd
->dl_bw
);
5718 if (cpudl_init(&rd
->cpudl
) != 0)
5721 if (cpupri_init(&rd
->cpupri
) != 0)
5726 free_cpumask_var(rd
->rto_mask
);
5728 free_cpumask_var(rd
->dlo_mask
);
5730 free_cpumask_var(rd
->online
);
5732 free_cpumask_var(rd
->span
);
5738 * By default the system creates a single root-domain with all cpus as
5739 * members (mimicking the global state we have today).
5741 struct root_domain def_root_domain
;
5743 static void init_defrootdomain(void)
5745 init_rootdomain(&def_root_domain
);
5747 atomic_set(&def_root_domain
.refcount
, 1);
5750 static struct root_domain
*alloc_rootdomain(void)
5752 struct root_domain
*rd
;
5754 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5758 if (init_rootdomain(rd
) != 0) {
5766 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5768 struct sched_group
*tmp
, *first
;
5777 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5782 } while (sg
!= first
);
5785 static void free_sched_domain(struct rcu_head
*rcu
)
5787 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5790 * If its an overlapping domain it has private groups, iterate and
5793 if (sd
->flags
& SD_OVERLAP
) {
5794 free_sched_groups(sd
->groups
, 1);
5795 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5796 kfree(sd
->groups
->sgc
);
5802 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5804 call_rcu(&sd
->rcu
, free_sched_domain
);
5807 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5809 for (; sd
; sd
= sd
->parent
)
5810 destroy_sched_domain(sd
, cpu
);
5814 * Keep a special pointer to the highest sched_domain that has
5815 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5816 * allows us to avoid some pointer chasing select_idle_sibling().
5818 * Also keep a unique ID per domain (we use the first cpu number in
5819 * the cpumask of the domain), this allows us to quickly tell if
5820 * two cpus are in the same cache domain, see cpus_share_cache().
5822 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5823 DEFINE_PER_CPU(int, sd_llc_size
);
5824 DEFINE_PER_CPU(int, sd_llc_id
);
5825 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5826 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5827 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5829 static void update_top_cache_domain(int cpu
)
5831 struct sched_domain
*sd
;
5832 struct sched_domain
*busy_sd
= NULL
;
5836 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5838 id
= cpumask_first(sched_domain_span(sd
));
5839 size
= cpumask_weight(sched_domain_span(sd
));
5840 busy_sd
= sd
->parent
; /* sd_busy */
5842 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5844 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5845 per_cpu(sd_llc_size
, cpu
) = size
;
5846 per_cpu(sd_llc_id
, cpu
) = id
;
5848 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5849 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5851 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5852 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5856 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5857 * hold the hotplug lock.
5860 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5862 struct rq
*rq
= cpu_rq(cpu
);
5863 struct sched_domain
*tmp
;
5865 /* Remove the sched domains which do not contribute to scheduling. */
5866 for (tmp
= sd
; tmp
; ) {
5867 struct sched_domain
*parent
= tmp
->parent
;
5871 if (sd_parent_degenerate(tmp
, parent
)) {
5872 tmp
->parent
= parent
->parent
;
5874 parent
->parent
->child
= tmp
;
5876 * Transfer SD_PREFER_SIBLING down in case of a
5877 * degenerate parent; the spans match for this
5878 * so the property transfers.
5880 if (parent
->flags
& SD_PREFER_SIBLING
)
5881 tmp
->flags
|= SD_PREFER_SIBLING
;
5882 destroy_sched_domain(parent
, cpu
);
5887 if (sd
&& sd_degenerate(sd
)) {
5890 destroy_sched_domain(tmp
, cpu
);
5895 sched_domain_debug(sd
, cpu
);
5897 rq_attach_root(rq
, rd
);
5899 rcu_assign_pointer(rq
->sd
, sd
);
5900 destroy_sched_domains(tmp
, cpu
);
5902 update_top_cache_domain(cpu
);
5905 /* Setup the mask of cpus configured for isolated domains */
5906 static int __init
isolated_cpu_setup(char *str
)
5910 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5911 ret
= cpulist_parse(str
, cpu_isolated_map
);
5913 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5918 __setup("isolcpus=", isolated_cpu_setup
);
5921 struct sched_domain
** __percpu sd
;
5922 struct root_domain
*rd
;
5933 * Build an iteration mask that can exclude certain CPUs from the upwards
5936 * Asymmetric node setups can result in situations where the domain tree is of
5937 * unequal depth, make sure to skip domains that already cover the entire
5940 * In that case build_sched_domains() will have terminated the iteration early
5941 * and our sibling sd spans will be empty. Domains should always include the
5942 * cpu they're built on, so check that.
5945 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5947 const struct cpumask
*span
= sched_domain_span(sd
);
5948 struct sd_data
*sdd
= sd
->private;
5949 struct sched_domain
*sibling
;
5952 for_each_cpu(i
, span
) {
5953 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5954 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5957 cpumask_set_cpu(i
, sched_group_mask(sg
));
5962 * Return the canonical balance cpu for this group, this is the first cpu
5963 * of this group that's also in the iteration mask.
5965 int group_balance_cpu(struct sched_group
*sg
)
5967 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5971 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5973 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5974 const struct cpumask
*span
= sched_domain_span(sd
);
5975 struct cpumask
*covered
= sched_domains_tmpmask
;
5976 struct sd_data
*sdd
= sd
->private;
5977 struct sched_domain
*sibling
;
5980 cpumask_clear(covered
);
5982 for_each_cpu(i
, span
) {
5983 struct cpumask
*sg_span
;
5985 if (cpumask_test_cpu(i
, covered
))
5988 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5990 /* See the comment near build_group_mask(). */
5991 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5994 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5995 GFP_KERNEL
, cpu_to_node(cpu
));
6000 sg_span
= sched_group_cpus(sg
);
6002 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6004 cpumask_set_cpu(i
, sg_span
);
6006 cpumask_or(covered
, covered
, sg_span
);
6008 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6009 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6010 build_group_mask(sd
, sg
);
6013 * Initialize sgc->capacity such that even if we mess up the
6014 * domains and no possible iteration will get us here, we won't
6017 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6020 * Make sure the first group of this domain contains the
6021 * canonical balance cpu. Otherwise the sched_domain iteration
6022 * breaks. See update_sg_lb_stats().
6024 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6025 group_balance_cpu(sg
) == cpu
)
6035 sd
->groups
= groups
;
6040 free_sched_groups(first
, 0);
6045 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6047 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6048 struct sched_domain
*child
= sd
->child
;
6051 cpu
= cpumask_first(sched_domain_span(child
));
6054 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6055 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6056 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6063 * build_sched_groups will build a circular linked list of the groups
6064 * covered by the given span, and will set each group's ->cpumask correctly,
6065 * and ->cpu_capacity to 0.
6067 * Assumes the sched_domain tree is fully constructed
6070 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6072 struct sched_group
*first
= NULL
, *last
= NULL
;
6073 struct sd_data
*sdd
= sd
->private;
6074 const struct cpumask
*span
= sched_domain_span(sd
);
6075 struct cpumask
*covered
;
6078 get_group(cpu
, sdd
, &sd
->groups
);
6079 atomic_inc(&sd
->groups
->ref
);
6081 if (cpu
!= cpumask_first(span
))
6084 lockdep_assert_held(&sched_domains_mutex
);
6085 covered
= sched_domains_tmpmask
;
6087 cpumask_clear(covered
);
6089 for_each_cpu(i
, span
) {
6090 struct sched_group
*sg
;
6093 if (cpumask_test_cpu(i
, covered
))
6096 group
= get_group(i
, sdd
, &sg
);
6097 cpumask_setall(sched_group_mask(sg
));
6099 for_each_cpu(j
, span
) {
6100 if (get_group(j
, sdd
, NULL
) != group
)
6103 cpumask_set_cpu(j
, covered
);
6104 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6119 * Initialize sched groups cpu_capacity.
6121 * cpu_capacity indicates the capacity of sched group, which is used while
6122 * distributing the load between different sched groups in a sched domain.
6123 * Typically cpu_capacity for all the groups in a sched domain will be same
6124 * unless there are asymmetries in the topology. If there are asymmetries,
6125 * group having more cpu_capacity will pickup more load compared to the
6126 * group having less cpu_capacity.
6128 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6130 struct sched_group
*sg
= sd
->groups
;
6135 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6137 } while (sg
!= sd
->groups
);
6139 if (cpu
!= group_balance_cpu(sg
))
6142 update_group_capacity(sd
, cpu
);
6143 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6147 * Initializers for schedule domains
6148 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6151 static int default_relax_domain_level
= -1;
6152 int sched_domain_level_max
;
6154 static int __init
setup_relax_domain_level(char *str
)
6156 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6157 pr_warn("Unable to set relax_domain_level\n");
6161 __setup("relax_domain_level=", setup_relax_domain_level
);
6163 static void set_domain_attribute(struct sched_domain
*sd
,
6164 struct sched_domain_attr
*attr
)
6168 if (!attr
|| attr
->relax_domain_level
< 0) {
6169 if (default_relax_domain_level
< 0)
6172 request
= default_relax_domain_level
;
6174 request
= attr
->relax_domain_level
;
6175 if (request
< sd
->level
) {
6176 /* turn off idle balance on this domain */
6177 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6179 /* turn on idle balance on this domain */
6180 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6184 static void __sdt_free(const struct cpumask
*cpu_map
);
6185 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6187 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6188 const struct cpumask
*cpu_map
)
6192 if (!atomic_read(&d
->rd
->refcount
))
6193 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6195 free_percpu(d
->sd
); /* fall through */
6197 __sdt_free(cpu_map
); /* fall through */
6203 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6204 const struct cpumask
*cpu_map
)
6206 memset(d
, 0, sizeof(*d
));
6208 if (__sdt_alloc(cpu_map
))
6209 return sa_sd_storage
;
6210 d
->sd
= alloc_percpu(struct sched_domain
*);
6212 return sa_sd_storage
;
6213 d
->rd
= alloc_rootdomain();
6216 return sa_rootdomain
;
6220 * NULL the sd_data elements we've used to build the sched_domain and
6221 * sched_group structure so that the subsequent __free_domain_allocs()
6222 * will not free the data we're using.
6224 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6226 struct sd_data
*sdd
= sd
->private;
6228 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6229 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6231 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6232 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6234 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6235 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6239 static int sched_domains_numa_levels
;
6240 enum numa_topology_type sched_numa_topology_type
;
6241 static int *sched_domains_numa_distance
;
6242 int sched_max_numa_distance
;
6243 static struct cpumask
***sched_domains_numa_masks
;
6244 static int sched_domains_curr_level
;
6248 * SD_flags allowed in topology descriptions.
6250 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6251 * SD_SHARE_PKG_RESOURCES - describes shared caches
6252 * SD_NUMA - describes NUMA topologies
6253 * SD_SHARE_POWERDOMAIN - describes shared power domain
6256 * SD_ASYM_PACKING - describes SMT quirks
6258 #define TOPOLOGY_SD_FLAGS \
6259 (SD_SHARE_CPUCAPACITY | \
6260 SD_SHARE_PKG_RESOURCES | \
6263 SD_SHARE_POWERDOMAIN)
6265 static struct sched_domain
*
6266 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6268 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6269 int sd_weight
, sd_flags
= 0;
6273 * Ugly hack to pass state to sd_numa_mask()...
6275 sched_domains_curr_level
= tl
->numa_level
;
6278 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6281 sd_flags
= (*tl
->sd_flags
)();
6282 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6283 "wrong sd_flags in topology description\n"))
6284 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6286 *sd
= (struct sched_domain
){
6287 .min_interval
= sd_weight
,
6288 .max_interval
= 2*sd_weight
,
6290 .imbalance_pct
= 125,
6292 .cache_nice_tries
= 0,
6299 .flags
= 1*SD_LOAD_BALANCE
6300 | 1*SD_BALANCE_NEWIDLE
6305 | 0*SD_SHARE_CPUCAPACITY
6306 | 0*SD_SHARE_PKG_RESOURCES
6308 | 0*SD_PREFER_SIBLING
6313 .last_balance
= jiffies
,
6314 .balance_interval
= sd_weight
,
6316 .max_newidle_lb_cost
= 0,
6317 .next_decay_max_lb_cost
= jiffies
,
6318 #ifdef CONFIG_SCHED_DEBUG
6324 * Convert topological properties into behaviour.
6327 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6328 sd
->flags
|= SD_PREFER_SIBLING
;
6329 sd
->imbalance_pct
= 110;
6330 sd
->smt_gain
= 1178; /* ~15% */
6332 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6333 sd
->imbalance_pct
= 117;
6334 sd
->cache_nice_tries
= 1;
6338 } else if (sd
->flags
& SD_NUMA
) {
6339 sd
->cache_nice_tries
= 2;
6343 sd
->flags
|= SD_SERIALIZE
;
6344 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6345 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6352 sd
->flags
|= SD_PREFER_SIBLING
;
6353 sd
->cache_nice_tries
= 1;
6358 sd
->private = &tl
->data
;
6364 * Topology list, bottom-up.
6366 static struct sched_domain_topology_level default_topology
[] = {
6367 #ifdef CONFIG_SCHED_SMT
6368 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6370 #ifdef CONFIG_SCHED_MC
6371 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6373 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6377 static struct sched_domain_topology_level
*sched_domain_topology
=
6380 #define for_each_sd_topology(tl) \
6381 for (tl = sched_domain_topology; tl->mask; tl++)
6383 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6385 sched_domain_topology
= tl
;
6390 static const struct cpumask
*sd_numa_mask(int cpu
)
6392 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6395 static void sched_numa_warn(const char *str
)
6397 static int done
= false;
6405 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6407 for (i
= 0; i
< nr_node_ids
; i
++) {
6408 printk(KERN_WARNING
" ");
6409 for (j
= 0; j
< nr_node_ids
; j
++)
6410 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6411 printk(KERN_CONT
"\n");
6413 printk(KERN_WARNING
"\n");
6416 bool find_numa_distance(int distance
)
6420 if (distance
== node_distance(0, 0))
6423 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6424 if (sched_domains_numa_distance
[i
] == distance
)
6432 * A system can have three types of NUMA topology:
6433 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6434 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6435 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6437 * The difference between a glueless mesh topology and a backplane
6438 * topology lies in whether communication between not directly
6439 * connected nodes goes through intermediary nodes (where programs
6440 * could run), or through backplane controllers. This affects
6441 * placement of programs.
6443 * The type of topology can be discerned with the following tests:
6444 * - If the maximum distance between any nodes is 1 hop, the system
6445 * is directly connected.
6446 * - If for two nodes A and B, located N > 1 hops away from each other,
6447 * there is an intermediary node C, which is < N hops away from both
6448 * nodes A and B, the system is a glueless mesh.
6450 static void init_numa_topology_type(void)
6454 n
= sched_max_numa_distance
;
6456 if (sched_domains_numa_levels
<= 1) {
6457 sched_numa_topology_type
= NUMA_DIRECT
;
6461 for_each_online_node(a
) {
6462 for_each_online_node(b
) {
6463 /* Find two nodes furthest removed from each other. */
6464 if (node_distance(a
, b
) < n
)
6467 /* Is there an intermediary node between a and b? */
6468 for_each_online_node(c
) {
6469 if (node_distance(a
, c
) < n
&&
6470 node_distance(b
, c
) < n
) {
6471 sched_numa_topology_type
=
6477 sched_numa_topology_type
= NUMA_BACKPLANE
;
6483 static void sched_init_numa(void)
6485 int next_distance
, curr_distance
= node_distance(0, 0);
6486 struct sched_domain_topology_level
*tl
;
6490 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6491 if (!sched_domains_numa_distance
)
6495 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6496 * unique distances in the node_distance() table.
6498 * Assumes node_distance(0,j) includes all distances in
6499 * node_distance(i,j) in order to avoid cubic time.
6501 next_distance
= curr_distance
;
6502 for (i
= 0; i
< nr_node_ids
; i
++) {
6503 for (j
= 0; j
< nr_node_ids
; j
++) {
6504 for (k
= 0; k
< nr_node_ids
; k
++) {
6505 int distance
= node_distance(i
, k
);
6507 if (distance
> curr_distance
&&
6508 (distance
< next_distance
||
6509 next_distance
== curr_distance
))
6510 next_distance
= distance
;
6513 * While not a strong assumption it would be nice to know
6514 * about cases where if node A is connected to B, B is not
6515 * equally connected to A.
6517 if (sched_debug() && node_distance(k
, i
) != distance
)
6518 sched_numa_warn("Node-distance not symmetric");
6520 if (sched_debug() && i
&& !find_numa_distance(distance
))
6521 sched_numa_warn("Node-0 not representative");
6523 if (next_distance
!= curr_distance
) {
6524 sched_domains_numa_distance
[level
++] = next_distance
;
6525 sched_domains_numa_levels
= level
;
6526 curr_distance
= next_distance
;
6531 * In case of sched_debug() we verify the above assumption.
6541 * 'level' contains the number of unique distances, excluding the
6542 * identity distance node_distance(i,i).
6544 * The sched_domains_numa_distance[] array includes the actual distance
6549 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6550 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6551 * the array will contain less then 'level' members. This could be
6552 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6553 * in other functions.
6555 * We reset it to 'level' at the end of this function.
6557 sched_domains_numa_levels
= 0;
6559 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6560 if (!sched_domains_numa_masks
)
6564 * Now for each level, construct a mask per node which contains all
6565 * cpus of nodes that are that many hops away from us.
6567 for (i
= 0; i
< level
; i
++) {
6568 sched_domains_numa_masks
[i
] =
6569 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6570 if (!sched_domains_numa_masks
[i
])
6573 for (j
= 0; j
< nr_node_ids
; j
++) {
6574 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6578 sched_domains_numa_masks
[i
][j
] = mask
;
6581 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6584 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6589 /* Compute default topology size */
6590 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6592 tl
= kzalloc((i
+ level
+ 1) *
6593 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6598 * Copy the default topology bits..
6600 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6601 tl
[i
] = sched_domain_topology
[i
];
6604 * .. and append 'j' levels of NUMA goodness.
6606 for (j
= 0; j
< level
; i
++, j
++) {
6607 tl
[i
] = (struct sched_domain_topology_level
){
6608 .mask
= sd_numa_mask
,
6609 .sd_flags
= cpu_numa_flags
,
6610 .flags
= SDTL_OVERLAP
,
6616 sched_domain_topology
= tl
;
6618 sched_domains_numa_levels
= level
;
6619 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6621 init_numa_topology_type();
6624 static void sched_domains_numa_masks_set(int cpu
)
6627 int node
= cpu_to_node(cpu
);
6629 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6630 for (j
= 0; j
< nr_node_ids
; j
++) {
6631 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6632 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6637 static void sched_domains_numa_masks_clear(int cpu
)
6640 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6641 for (j
= 0; j
< nr_node_ids
; j
++)
6642 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6647 * Update sched_domains_numa_masks[level][node] array when new cpus
6650 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6651 unsigned long action
,
6654 int cpu
= (long)hcpu
;
6656 switch (action
& ~CPU_TASKS_FROZEN
) {
6658 sched_domains_numa_masks_set(cpu
);
6662 sched_domains_numa_masks_clear(cpu
);
6672 static inline void sched_init_numa(void)
6676 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6677 unsigned long action
,
6682 #endif /* CONFIG_NUMA */
6684 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6686 struct sched_domain_topology_level
*tl
;
6689 for_each_sd_topology(tl
) {
6690 struct sd_data
*sdd
= &tl
->data
;
6692 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6696 sdd
->sg
= alloc_percpu(struct sched_group
*);
6700 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6704 for_each_cpu(j
, cpu_map
) {
6705 struct sched_domain
*sd
;
6706 struct sched_group
*sg
;
6707 struct sched_group_capacity
*sgc
;
6709 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6710 GFP_KERNEL
, cpu_to_node(j
));
6714 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6716 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6717 GFP_KERNEL
, cpu_to_node(j
));
6723 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6725 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6726 GFP_KERNEL
, cpu_to_node(j
));
6730 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6737 static void __sdt_free(const struct cpumask
*cpu_map
)
6739 struct sched_domain_topology_level
*tl
;
6742 for_each_sd_topology(tl
) {
6743 struct sd_data
*sdd
= &tl
->data
;
6745 for_each_cpu(j
, cpu_map
) {
6746 struct sched_domain
*sd
;
6749 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6750 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6751 free_sched_groups(sd
->groups
, 0);
6752 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6756 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6758 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6760 free_percpu(sdd
->sd
);
6762 free_percpu(sdd
->sg
);
6764 free_percpu(sdd
->sgc
);
6769 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6770 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6771 struct sched_domain
*child
, int cpu
)
6773 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6777 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6779 sd
->level
= child
->level
+ 1;
6780 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6784 if (!cpumask_subset(sched_domain_span(child
),
6785 sched_domain_span(sd
))) {
6786 pr_err("BUG: arch topology borken\n");
6787 #ifdef CONFIG_SCHED_DEBUG
6788 pr_err(" the %s domain not a subset of the %s domain\n",
6789 child
->name
, sd
->name
);
6791 /* Fixup, ensure @sd has at least @child cpus. */
6792 cpumask_or(sched_domain_span(sd
),
6793 sched_domain_span(sd
),
6794 sched_domain_span(child
));
6798 set_domain_attribute(sd
, attr
);
6804 * Build sched domains for a given set of cpus and attach the sched domains
6805 * to the individual cpus
6807 static int build_sched_domains(const struct cpumask
*cpu_map
,
6808 struct sched_domain_attr
*attr
)
6810 enum s_alloc alloc_state
;
6811 struct sched_domain
*sd
;
6813 int i
, ret
= -ENOMEM
;
6815 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6816 if (alloc_state
!= sa_rootdomain
)
6819 /* Set up domains for cpus specified by the cpu_map. */
6820 for_each_cpu(i
, cpu_map
) {
6821 struct sched_domain_topology_level
*tl
;
6824 for_each_sd_topology(tl
) {
6825 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6826 if (tl
== sched_domain_topology
)
6827 *per_cpu_ptr(d
.sd
, i
) = sd
;
6828 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6829 sd
->flags
|= SD_OVERLAP
;
6830 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6835 /* Build the groups for the domains */
6836 for_each_cpu(i
, cpu_map
) {
6837 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6838 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6839 if (sd
->flags
& SD_OVERLAP
) {
6840 if (build_overlap_sched_groups(sd
, i
))
6843 if (build_sched_groups(sd
, i
))
6849 /* Calculate CPU capacity for physical packages and nodes */
6850 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6851 if (!cpumask_test_cpu(i
, cpu_map
))
6854 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6855 claim_allocations(i
, sd
);
6856 init_sched_groups_capacity(i
, sd
);
6860 /* Attach the domains */
6862 for_each_cpu(i
, cpu_map
) {
6863 sd
= *per_cpu_ptr(d
.sd
, i
);
6864 cpu_attach_domain(sd
, d
.rd
, i
);
6870 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6874 static cpumask_var_t
*doms_cur
; /* current sched domains */
6875 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6876 static struct sched_domain_attr
*dattr_cur
;
6877 /* attribues of custom domains in 'doms_cur' */
6880 * Special case: If a kmalloc of a doms_cur partition (array of
6881 * cpumask) fails, then fallback to a single sched domain,
6882 * as determined by the single cpumask fallback_doms.
6884 static cpumask_var_t fallback_doms
;
6887 * arch_update_cpu_topology lets virtualized architectures update the
6888 * cpu core maps. It is supposed to return 1 if the topology changed
6889 * or 0 if it stayed the same.
6891 int __weak
arch_update_cpu_topology(void)
6896 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6899 cpumask_var_t
*doms
;
6901 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6904 for (i
= 0; i
< ndoms
; i
++) {
6905 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6906 free_sched_domains(doms
, i
);
6913 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6916 for (i
= 0; i
< ndoms
; i
++)
6917 free_cpumask_var(doms
[i
]);
6922 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6923 * For now this just excludes isolated cpus, but could be used to
6924 * exclude other special cases in the future.
6926 static int init_sched_domains(const struct cpumask
*cpu_map
)
6930 arch_update_cpu_topology();
6932 doms_cur
= alloc_sched_domains(ndoms_cur
);
6934 doms_cur
= &fallback_doms
;
6935 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6936 err
= build_sched_domains(doms_cur
[0], NULL
);
6937 register_sched_domain_sysctl();
6943 * Detach sched domains from a group of cpus specified in cpu_map
6944 * These cpus will now be attached to the NULL domain
6946 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6951 for_each_cpu(i
, cpu_map
)
6952 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6956 /* handle null as "default" */
6957 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6958 struct sched_domain_attr
*new, int idx_new
)
6960 struct sched_domain_attr tmp
;
6967 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6968 new ? (new + idx_new
) : &tmp
,
6969 sizeof(struct sched_domain_attr
));
6973 * Partition sched domains as specified by the 'ndoms_new'
6974 * cpumasks in the array doms_new[] of cpumasks. This compares
6975 * doms_new[] to the current sched domain partitioning, doms_cur[].
6976 * It destroys each deleted domain and builds each new domain.
6978 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6979 * The masks don't intersect (don't overlap.) We should setup one
6980 * sched domain for each mask. CPUs not in any of the cpumasks will
6981 * not be load balanced. If the same cpumask appears both in the
6982 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6985 * The passed in 'doms_new' should be allocated using
6986 * alloc_sched_domains. This routine takes ownership of it and will
6987 * free_sched_domains it when done with it. If the caller failed the
6988 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6989 * and partition_sched_domains() will fallback to the single partition
6990 * 'fallback_doms', it also forces the domains to be rebuilt.
6992 * If doms_new == NULL it will be replaced with cpu_online_mask.
6993 * ndoms_new == 0 is a special case for destroying existing domains,
6994 * and it will not create the default domain.
6996 * Call with hotplug lock held
6998 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6999 struct sched_domain_attr
*dattr_new
)
7004 mutex_lock(&sched_domains_mutex
);
7006 /* always unregister in case we don't destroy any domains */
7007 unregister_sched_domain_sysctl();
7009 /* Let architecture update cpu core mappings. */
7010 new_topology
= arch_update_cpu_topology();
7012 n
= doms_new
? ndoms_new
: 0;
7014 /* Destroy deleted domains */
7015 for (i
= 0; i
< ndoms_cur
; i
++) {
7016 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7017 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7018 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7021 /* no match - a current sched domain not in new doms_new[] */
7022 detach_destroy_domains(doms_cur
[i
]);
7028 if (doms_new
== NULL
) {
7030 doms_new
= &fallback_doms
;
7031 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7032 WARN_ON_ONCE(dattr_new
);
7035 /* Build new domains */
7036 for (i
= 0; i
< ndoms_new
; i
++) {
7037 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7038 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7039 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7042 /* no match - add a new doms_new */
7043 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7048 /* Remember the new sched domains */
7049 if (doms_cur
!= &fallback_doms
)
7050 free_sched_domains(doms_cur
, ndoms_cur
);
7051 kfree(dattr_cur
); /* kfree(NULL) is safe */
7052 doms_cur
= doms_new
;
7053 dattr_cur
= dattr_new
;
7054 ndoms_cur
= ndoms_new
;
7056 register_sched_domain_sysctl();
7058 mutex_unlock(&sched_domains_mutex
);
7061 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7064 * Update cpusets according to cpu_active mask. If cpusets are
7065 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7066 * around partition_sched_domains().
7068 * If we come here as part of a suspend/resume, don't touch cpusets because we
7069 * want to restore it back to its original state upon resume anyway.
7071 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7075 case CPU_ONLINE_FROZEN
:
7076 case CPU_DOWN_FAILED_FROZEN
:
7079 * num_cpus_frozen tracks how many CPUs are involved in suspend
7080 * resume sequence. As long as this is not the last online
7081 * operation in the resume sequence, just build a single sched
7082 * domain, ignoring cpusets.
7085 if (likely(num_cpus_frozen
)) {
7086 partition_sched_domains(1, NULL
, NULL
);
7091 * This is the last CPU online operation. So fall through and
7092 * restore the original sched domains by considering the
7093 * cpuset configurations.
7097 cpuset_update_active_cpus(true);
7105 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7108 unsigned long flags
;
7109 long cpu
= (long)hcpu
;
7115 case CPU_DOWN_PREPARE
:
7116 rcu_read_lock_sched();
7117 dl_b
= dl_bw_of(cpu
);
7119 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7120 cpus
= dl_bw_cpus(cpu
);
7121 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7122 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7124 rcu_read_unlock_sched();
7127 return notifier_from_errno(-EBUSY
);
7128 cpuset_update_active_cpus(false);
7130 case CPU_DOWN_PREPARE_FROZEN
:
7132 partition_sched_domains(1, NULL
, NULL
);
7140 void __init
sched_init_smp(void)
7142 cpumask_var_t non_isolated_cpus
;
7144 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7145 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7150 * There's no userspace yet to cause hotplug operations; hence all the
7151 * cpu masks are stable and all blatant races in the below code cannot
7154 mutex_lock(&sched_domains_mutex
);
7155 init_sched_domains(cpu_active_mask
);
7156 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7157 if (cpumask_empty(non_isolated_cpus
))
7158 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7159 mutex_unlock(&sched_domains_mutex
);
7161 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7162 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7163 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7167 /* Move init over to a non-isolated CPU */
7168 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7170 sched_init_granularity();
7171 free_cpumask_var(non_isolated_cpus
);
7173 init_sched_rt_class();
7174 init_sched_dl_class();
7177 void __init
sched_init_smp(void)
7179 sched_init_granularity();
7181 #endif /* CONFIG_SMP */
7183 int in_sched_functions(unsigned long addr
)
7185 return in_lock_functions(addr
) ||
7186 (addr
>= (unsigned long)__sched_text_start
7187 && addr
< (unsigned long)__sched_text_end
);
7190 #ifdef CONFIG_CGROUP_SCHED
7192 * Default task group.
7193 * Every task in system belongs to this group at bootup.
7195 struct task_group root_task_group
;
7196 LIST_HEAD(task_groups
);
7198 /* Cacheline aligned slab cache for task_group */
7199 static struct kmem_cache
*task_group_cache __read_mostly
;
7202 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7204 void __init
sched_init(void)
7207 unsigned long alloc_size
= 0, ptr
;
7209 #ifdef CONFIG_FAIR_GROUP_SCHED
7210 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7212 #ifdef CONFIG_RT_GROUP_SCHED
7213 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7216 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7218 #ifdef CONFIG_FAIR_GROUP_SCHED
7219 root_task_group
.se
= (struct sched_entity
**)ptr
;
7220 ptr
+= nr_cpu_ids
* sizeof(void **);
7222 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7223 ptr
+= nr_cpu_ids
* sizeof(void **);
7225 #endif /* CONFIG_FAIR_GROUP_SCHED */
7226 #ifdef CONFIG_RT_GROUP_SCHED
7227 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7228 ptr
+= nr_cpu_ids
* sizeof(void **);
7230 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7231 ptr
+= nr_cpu_ids
* sizeof(void **);
7233 #endif /* CONFIG_RT_GROUP_SCHED */
7235 #ifdef CONFIG_CPUMASK_OFFSTACK
7236 for_each_possible_cpu(i
) {
7237 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7238 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7240 #endif /* CONFIG_CPUMASK_OFFSTACK */
7242 init_rt_bandwidth(&def_rt_bandwidth
,
7243 global_rt_period(), global_rt_runtime());
7244 init_dl_bandwidth(&def_dl_bandwidth
,
7245 global_rt_period(), global_rt_runtime());
7248 init_defrootdomain();
7251 #ifdef CONFIG_RT_GROUP_SCHED
7252 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7253 global_rt_period(), global_rt_runtime());
7254 #endif /* CONFIG_RT_GROUP_SCHED */
7256 #ifdef CONFIG_CGROUP_SCHED
7257 task_group_cache
= KMEM_CACHE(task_group
, 0);
7259 list_add(&root_task_group
.list
, &task_groups
);
7260 INIT_LIST_HEAD(&root_task_group
.children
);
7261 INIT_LIST_HEAD(&root_task_group
.siblings
);
7262 autogroup_init(&init_task
);
7263 #endif /* CONFIG_CGROUP_SCHED */
7265 for_each_possible_cpu(i
) {
7269 raw_spin_lock_init(&rq
->lock
);
7271 rq
->calc_load_active
= 0;
7272 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7273 init_cfs_rq(&rq
->cfs
);
7274 init_rt_rq(&rq
->rt
);
7275 init_dl_rq(&rq
->dl
);
7276 #ifdef CONFIG_FAIR_GROUP_SCHED
7277 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7278 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7280 * How much cpu bandwidth does root_task_group get?
7282 * In case of task-groups formed thr' the cgroup filesystem, it
7283 * gets 100% of the cpu resources in the system. This overall
7284 * system cpu resource is divided among the tasks of
7285 * root_task_group and its child task-groups in a fair manner,
7286 * based on each entity's (task or task-group's) weight
7287 * (se->load.weight).
7289 * In other words, if root_task_group has 10 tasks of weight
7290 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7291 * then A0's share of the cpu resource is:
7293 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7295 * We achieve this by letting root_task_group's tasks sit
7296 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7298 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7299 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7300 #endif /* CONFIG_FAIR_GROUP_SCHED */
7302 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7303 #ifdef CONFIG_RT_GROUP_SCHED
7304 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7307 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7308 rq
->cpu_load
[j
] = 0;
7310 rq
->last_load_update_tick
= jiffies
;
7315 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7316 rq
->balance_callback
= NULL
;
7317 rq
->active_balance
= 0;
7318 rq
->next_balance
= jiffies
;
7323 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7324 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7326 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7328 rq_attach_root(rq
, &def_root_domain
);
7329 #ifdef CONFIG_NO_HZ_COMMON
7332 #ifdef CONFIG_NO_HZ_FULL
7333 rq
->last_sched_tick
= 0;
7337 atomic_set(&rq
->nr_iowait
, 0);
7340 set_load_weight(&init_task
);
7342 #ifdef CONFIG_PREEMPT_NOTIFIERS
7343 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7347 * The boot idle thread does lazy MMU switching as well:
7349 atomic_inc(&init_mm
.mm_count
);
7350 enter_lazy_tlb(&init_mm
, current
);
7353 * During early bootup we pretend to be a normal task:
7355 current
->sched_class
= &fair_sched_class
;
7358 * Make us the idle thread. Technically, schedule() should not be
7359 * called from this thread, however somewhere below it might be,
7360 * but because we are the idle thread, we just pick up running again
7361 * when this runqueue becomes "idle".
7363 init_idle(current
, smp_processor_id());
7365 calc_load_update
= jiffies
+ LOAD_FREQ
;
7368 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7369 /* May be allocated at isolcpus cmdline parse time */
7370 if (cpu_isolated_map
== NULL
)
7371 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7372 idle_thread_set_boot_cpu();
7373 set_cpu_rq_start_time();
7375 init_sched_fair_class();
7377 scheduler_running
= 1;
7380 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7381 static inline int preempt_count_equals(int preempt_offset
)
7383 int nested
= preempt_count() + rcu_preempt_depth();
7385 return (nested
== preempt_offset
);
7388 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7391 * Blocking primitives will set (and therefore destroy) current->state,
7392 * since we will exit with TASK_RUNNING make sure we enter with it,
7393 * otherwise we will destroy state.
7395 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7396 "do not call blocking ops when !TASK_RUNNING; "
7397 "state=%lx set at [<%p>] %pS\n",
7399 (void *)current
->task_state_change
,
7400 (void *)current
->task_state_change
);
7402 ___might_sleep(file
, line
, preempt_offset
);
7404 EXPORT_SYMBOL(__might_sleep
);
7406 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7408 static unsigned long prev_jiffy
; /* ratelimiting */
7410 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7411 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7412 !is_idle_task(current
)) ||
7413 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7415 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7417 prev_jiffy
= jiffies
;
7420 "BUG: sleeping function called from invalid context at %s:%d\n",
7423 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7424 in_atomic(), irqs_disabled(),
7425 current
->pid
, current
->comm
);
7427 if (task_stack_end_corrupted(current
))
7428 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7430 debug_show_held_locks(current
);
7431 if (irqs_disabled())
7432 print_irqtrace_events(current
);
7433 #ifdef CONFIG_DEBUG_PREEMPT
7434 if (!preempt_count_equals(preempt_offset
)) {
7435 pr_err("Preemption disabled at:");
7436 print_ip_sym(current
->preempt_disable_ip
);
7442 EXPORT_SYMBOL(___might_sleep
);
7445 #ifdef CONFIG_MAGIC_SYSRQ
7446 void normalize_rt_tasks(void)
7448 struct task_struct
*g
, *p
;
7449 struct sched_attr attr
= {
7450 .sched_policy
= SCHED_NORMAL
,
7453 read_lock(&tasklist_lock
);
7454 for_each_process_thread(g
, p
) {
7456 * Only normalize user tasks:
7458 if (p
->flags
& PF_KTHREAD
)
7461 p
->se
.exec_start
= 0;
7462 #ifdef CONFIG_SCHEDSTATS
7463 p
->se
.statistics
.wait_start
= 0;
7464 p
->se
.statistics
.sleep_start
= 0;
7465 p
->se
.statistics
.block_start
= 0;
7468 if (!dl_task(p
) && !rt_task(p
)) {
7470 * Renice negative nice level userspace
7473 if (task_nice(p
) < 0)
7474 set_user_nice(p
, 0);
7478 __sched_setscheduler(p
, &attr
, false, false);
7480 read_unlock(&tasklist_lock
);
7483 #endif /* CONFIG_MAGIC_SYSRQ */
7485 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7487 * These functions are only useful for the IA64 MCA handling, or kdb.
7489 * They can only be called when the whole system has been
7490 * stopped - every CPU needs to be quiescent, and no scheduling
7491 * activity can take place. Using them for anything else would
7492 * be a serious bug, and as a result, they aren't even visible
7493 * under any other configuration.
7497 * curr_task - return the current task for a given cpu.
7498 * @cpu: the processor in question.
7500 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7502 * Return: The current task for @cpu.
7504 struct task_struct
*curr_task(int cpu
)
7506 return cpu_curr(cpu
);
7509 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7513 * set_curr_task - set the current task for a given cpu.
7514 * @cpu: the processor in question.
7515 * @p: the task pointer to set.
7517 * Description: This function must only be used when non-maskable interrupts
7518 * are serviced on a separate stack. It allows the architecture to switch the
7519 * notion of the current task on a cpu in a non-blocking manner. This function
7520 * must be called with all CPU's synchronized, and interrupts disabled, the
7521 * and caller must save the original value of the current task (see
7522 * curr_task() above) and restore that value before reenabling interrupts and
7523 * re-starting the system.
7525 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7527 void set_curr_task(int cpu
, struct task_struct
*p
)
7534 #ifdef CONFIG_CGROUP_SCHED
7535 /* task_group_lock serializes the addition/removal of task groups */
7536 static DEFINE_SPINLOCK(task_group_lock
);
7538 static void free_sched_group(struct task_group
*tg
)
7540 free_fair_sched_group(tg
);
7541 free_rt_sched_group(tg
);
7543 kmem_cache_free(task_group_cache
, tg
);
7546 /* allocate runqueue etc for a new task group */
7547 struct task_group
*sched_create_group(struct task_group
*parent
)
7549 struct task_group
*tg
;
7551 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7553 return ERR_PTR(-ENOMEM
);
7555 if (!alloc_fair_sched_group(tg
, parent
))
7558 if (!alloc_rt_sched_group(tg
, parent
))
7564 free_sched_group(tg
);
7565 return ERR_PTR(-ENOMEM
);
7568 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7570 unsigned long flags
;
7572 spin_lock_irqsave(&task_group_lock
, flags
);
7573 list_add_rcu(&tg
->list
, &task_groups
);
7575 WARN_ON(!parent
); /* root should already exist */
7577 tg
->parent
= parent
;
7578 INIT_LIST_HEAD(&tg
->children
);
7579 list_add_rcu(&tg
->siblings
, &parent
->children
);
7580 spin_unlock_irqrestore(&task_group_lock
, flags
);
7583 /* rcu callback to free various structures associated with a task group */
7584 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7586 /* now it should be safe to free those cfs_rqs */
7587 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7590 /* Destroy runqueue etc associated with a task group */
7591 void sched_destroy_group(struct task_group
*tg
)
7593 /* wait for possible concurrent references to cfs_rqs complete */
7594 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7597 void sched_offline_group(struct task_group
*tg
)
7599 unsigned long flags
;
7601 /* end participation in shares distribution */
7602 unregister_fair_sched_group(tg
);
7604 spin_lock_irqsave(&task_group_lock
, flags
);
7605 list_del_rcu(&tg
->list
);
7606 list_del_rcu(&tg
->siblings
);
7607 spin_unlock_irqrestore(&task_group_lock
, flags
);
7610 /* change task's runqueue when it moves between groups.
7611 * The caller of this function should have put the task in its new group
7612 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7613 * reflect its new group.
7615 void sched_move_task(struct task_struct
*tsk
)
7617 struct task_group
*tg
;
7618 int queued
, running
;
7619 unsigned long flags
;
7622 rq
= task_rq_lock(tsk
, &flags
);
7624 running
= task_current(rq
, tsk
);
7625 queued
= task_on_rq_queued(tsk
);
7628 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7629 if (unlikely(running
))
7630 put_prev_task(rq
, tsk
);
7633 * All callers are synchronized by task_rq_lock(); we do not use RCU
7634 * which is pointless here. Thus, we pass "true" to task_css_check()
7635 * to prevent lockdep warnings.
7637 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7638 struct task_group
, css
);
7639 tg
= autogroup_task_group(tsk
, tg
);
7640 tsk
->sched_task_group
= tg
;
7642 #ifdef CONFIG_FAIR_GROUP_SCHED
7643 if (tsk
->sched_class
->task_move_group
)
7644 tsk
->sched_class
->task_move_group(tsk
);
7647 set_task_rq(tsk
, task_cpu(tsk
));
7649 if (unlikely(running
))
7650 tsk
->sched_class
->set_curr_task(rq
);
7652 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7654 task_rq_unlock(rq
, tsk
, &flags
);
7656 #endif /* CONFIG_CGROUP_SCHED */
7658 #ifdef CONFIG_RT_GROUP_SCHED
7660 * Ensure that the real time constraints are schedulable.
7662 static DEFINE_MUTEX(rt_constraints_mutex
);
7664 /* Must be called with tasklist_lock held */
7665 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7667 struct task_struct
*g
, *p
;
7670 * Autogroups do not have RT tasks; see autogroup_create().
7672 if (task_group_is_autogroup(tg
))
7675 for_each_process_thread(g
, p
) {
7676 if (rt_task(p
) && task_group(p
) == tg
)
7683 struct rt_schedulable_data
{
7684 struct task_group
*tg
;
7689 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7691 struct rt_schedulable_data
*d
= data
;
7692 struct task_group
*child
;
7693 unsigned long total
, sum
= 0;
7694 u64 period
, runtime
;
7696 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7697 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7700 period
= d
->rt_period
;
7701 runtime
= d
->rt_runtime
;
7705 * Cannot have more runtime than the period.
7707 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7711 * Ensure we don't starve existing RT tasks.
7713 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7716 total
= to_ratio(period
, runtime
);
7719 * Nobody can have more than the global setting allows.
7721 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7725 * The sum of our children's runtime should not exceed our own.
7727 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7728 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7729 runtime
= child
->rt_bandwidth
.rt_runtime
;
7731 if (child
== d
->tg
) {
7732 period
= d
->rt_period
;
7733 runtime
= d
->rt_runtime
;
7736 sum
+= to_ratio(period
, runtime
);
7745 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7749 struct rt_schedulable_data data
= {
7751 .rt_period
= period
,
7752 .rt_runtime
= runtime
,
7756 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7762 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7763 u64 rt_period
, u64 rt_runtime
)
7768 * Disallowing the root group RT runtime is BAD, it would disallow the
7769 * kernel creating (and or operating) RT threads.
7771 if (tg
== &root_task_group
&& rt_runtime
== 0)
7774 /* No period doesn't make any sense. */
7778 mutex_lock(&rt_constraints_mutex
);
7779 read_lock(&tasklist_lock
);
7780 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7784 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7785 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7786 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7788 for_each_possible_cpu(i
) {
7789 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7791 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7792 rt_rq
->rt_runtime
= rt_runtime
;
7793 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7795 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7797 read_unlock(&tasklist_lock
);
7798 mutex_unlock(&rt_constraints_mutex
);
7803 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7805 u64 rt_runtime
, rt_period
;
7807 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7808 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7809 if (rt_runtime_us
< 0)
7810 rt_runtime
= RUNTIME_INF
;
7812 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7815 static long sched_group_rt_runtime(struct task_group
*tg
)
7819 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7822 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7823 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7824 return rt_runtime_us
;
7827 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7829 u64 rt_runtime
, rt_period
;
7831 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7832 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7834 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7837 static long sched_group_rt_period(struct task_group
*tg
)
7841 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7842 do_div(rt_period_us
, NSEC_PER_USEC
);
7843 return rt_period_us
;
7845 #endif /* CONFIG_RT_GROUP_SCHED */
7847 #ifdef CONFIG_RT_GROUP_SCHED
7848 static int sched_rt_global_constraints(void)
7852 mutex_lock(&rt_constraints_mutex
);
7853 read_lock(&tasklist_lock
);
7854 ret
= __rt_schedulable(NULL
, 0, 0);
7855 read_unlock(&tasklist_lock
);
7856 mutex_unlock(&rt_constraints_mutex
);
7861 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7863 /* Don't accept realtime tasks when there is no way for them to run */
7864 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7870 #else /* !CONFIG_RT_GROUP_SCHED */
7871 static int sched_rt_global_constraints(void)
7873 unsigned long flags
;
7876 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7877 for_each_possible_cpu(i
) {
7878 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7880 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7881 rt_rq
->rt_runtime
= global_rt_runtime();
7882 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7884 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7888 #endif /* CONFIG_RT_GROUP_SCHED */
7890 static int sched_dl_global_validate(void)
7892 u64 runtime
= global_rt_runtime();
7893 u64 period
= global_rt_period();
7894 u64 new_bw
= to_ratio(period
, runtime
);
7897 unsigned long flags
;
7900 * Here we want to check the bandwidth not being set to some
7901 * value smaller than the currently allocated bandwidth in
7902 * any of the root_domains.
7904 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7905 * cycling on root_domains... Discussion on different/better
7906 * solutions is welcome!
7908 for_each_possible_cpu(cpu
) {
7909 rcu_read_lock_sched();
7910 dl_b
= dl_bw_of(cpu
);
7912 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7913 if (new_bw
< dl_b
->total_bw
)
7915 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7917 rcu_read_unlock_sched();
7926 static void sched_dl_do_global(void)
7931 unsigned long flags
;
7933 def_dl_bandwidth
.dl_period
= global_rt_period();
7934 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7936 if (global_rt_runtime() != RUNTIME_INF
)
7937 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7940 * FIXME: As above...
7942 for_each_possible_cpu(cpu
) {
7943 rcu_read_lock_sched();
7944 dl_b
= dl_bw_of(cpu
);
7946 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7948 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7950 rcu_read_unlock_sched();
7954 static int sched_rt_global_validate(void)
7956 if (sysctl_sched_rt_period
<= 0)
7959 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7960 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7966 static void sched_rt_do_global(void)
7968 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7969 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7972 int sched_rt_handler(struct ctl_table
*table
, int write
,
7973 void __user
*buffer
, size_t *lenp
,
7976 int old_period
, old_runtime
;
7977 static DEFINE_MUTEX(mutex
);
7981 old_period
= sysctl_sched_rt_period
;
7982 old_runtime
= sysctl_sched_rt_runtime
;
7984 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7986 if (!ret
&& write
) {
7987 ret
= sched_rt_global_validate();
7991 ret
= sched_dl_global_validate();
7995 ret
= sched_rt_global_constraints();
7999 sched_rt_do_global();
8000 sched_dl_do_global();
8004 sysctl_sched_rt_period
= old_period
;
8005 sysctl_sched_rt_runtime
= old_runtime
;
8007 mutex_unlock(&mutex
);
8012 int sched_rr_handler(struct ctl_table
*table
, int write
,
8013 void __user
*buffer
, size_t *lenp
,
8017 static DEFINE_MUTEX(mutex
);
8020 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8021 /* make sure that internally we keep jiffies */
8022 /* also, writing zero resets timeslice to default */
8023 if (!ret
&& write
) {
8024 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8025 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8027 mutex_unlock(&mutex
);
8031 #ifdef CONFIG_CGROUP_SCHED
8033 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8035 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8038 static struct cgroup_subsys_state
*
8039 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8041 struct task_group
*parent
= css_tg(parent_css
);
8042 struct task_group
*tg
;
8045 /* This is early initialization for the top cgroup */
8046 return &root_task_group
.css
;
8049 tg
= sched_create_group(parent
);
8051 return ERR_PTR(-ENOMEM
);
8056 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8058 struct task_group
*tg
= css_tg(css
);
8059 struct task_group
*parent
= css_tg(css
->parent
);
8062 sched_online_group(tg
, parent
);
8066 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8068 struct task_group
*tg
= css_tg(css
);
8070 sched_destroy_group(tg
);
8073 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8075 struct task_group
*tg
= css_tg(css
);
8077 sched_offline_group(tg
);
8080 static void cpu_cgroup_fork(struct task_struct
*task
)
8082 sched_move_task(task
);
8085 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8087 struct task_struct
*task
;
8088 struct cgroup_subsys_state
*css
;
8090 cgroup_taskset_for_each(task
, css
, tset
) {
8091 #ifdef CONFIG_RT_GROUP_SCHED
8092 if (!sched_rt_can_attach(css_tg(css
), task
))
8095 /* We don't support RT-tasks being in separate groups */
8096 if (task
->sched_class
!= &fair_sched_class
)
8103 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8105 struct task_struct
*task
;
8106 struct cgroup_subsys_state
*css
;
8108 cgroup_taskset_for_each(task
, css
, tset
)
8109 sched_move_task(task
);
8112 #ifdef CONFIG_FAIR_GROUP_SCHED
8113 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8114 struct cftype
*cftype
, u64 shareval
)
8116 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8119 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8122 struct task_group
*tg
= css_tg(css
);
8124 return (u64
) scale_load_down(tg
->shares
);
8127 #ifdef CONFIG_CFS_BANDWIDTH
8128 static DEFINE_MUTEX(cfs_constraints_mutex
);
8130 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8131 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8133 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8135 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8137 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8138 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8140 if (tg
== &root_task_group
)
8144 * Ensure we have at some amount of bandwidth every period. This is
8145 * to prevent reaching a state of large arrears when throttled via
8146 * entity_tick() resulting in prolonged exit starvation.
8148 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8152 * Likewise, bound things on the otherside by preventing insane quota
8153 * periods. This also allows us to normalize in computing quota
8156 if (period
> max_cfs_quota_period
)
8160 * Prevent race between setting of cfs_rq->runtime_enabled and
8161 * unthrottle_offline_cfs_rqs().
8164 mutex_lock(&cfs_constraints_mutex
);
8165 ret
= __cfs_schedulable(tg
, period
, quota
);
8169 runtime_enabled
= quota
!= RUNTIME_INF
;
8170 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8172 * If we need to toggle cfs_bandwidth_used, off->on must occur
8173 * before making related changes, and on->off must occur afterwards
8175 if (runtime_enabled
&& !runtime_was_enabled
)
8176 cfs_bandwidth_usage_inc();
8177 raw_spin_lock_irq(&cfs_b
->lock
);
8178 cfs_b
->period
= ns_to_ktime(period
);
8179 cfs_b
->quota
= quota
;
8181 __refill_cfs_bandwidth_runtime(cfs_b
);
8182 /* restart the period timer (if active) to handle new period expiry */
8183 if (runtime_enabled
)
8184 start_cfs_bandwidth(cfs_b
);
8185 raw_spin_unlock_irq(&cfs_b
->lock
);
8187 for_each_online_cpu(i
) {
8188 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8189 struct rq
*rq
= cfs_rq
->rq
;
8191 raw_spin_lock_irq(&rq
->lock
);
8192 cfs_rq
->runtime_enabled
= runtime_enabled
;
8193 cfs_rq
->runtime_remaining
= 0;
8195 if (cfs_rq
->throttled
)
8196 unthrottle_cfs_rq(cfs_rq
);
8197 raw_spin_unlock_irq(&rq
->lock
);
8199 if (runtime_was_enabled
&& !runtime_enabled
)
8200 cfs_bandwidth_usage_dec();
8202 mutex_unlock(&cfs_constraints_mutex
);
8208 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8212 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8213 if (cfs_quota_us
< 0)
8214 quota
= RUNTIME_INF
;
8216 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8218 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8221 long tg_get_cfs_quota(struct task_group
*tg
)
8225 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8228 quota_us
= tg
->cfs_bandwidth
.quota
;
8229 do_div(quota_us
, NSEC_PER_USEC
);
8234 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8238 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8239 quota
= tg
->cfs_bandwidth
.quota
;
8241 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8244 long tg_get_cfs_period(struct task_group
*tg
)
8248 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8249 do_div(cfs_period_us
, NSEC_PER_USEC
);
8251 return cfs_period_us
;
8254 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8257 return tg_get_cfs_quota(css_tg(css
));
8260 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8261 struct cftype
*cftype
, s64 cfs_quota_us
)
8263 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8266 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8269 return tg_get_cfs_period(css_tg(css
));
8272 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8273 struct cftype
*cftype
, u64 cfs_period_us
)
8275 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8278 struct cfs_schedulable_data
{
8279 struct task_group
*tg
;
8284 * normalize group quota/period to be quota/max_period
8285 * note: units are usecs
8287 static u64
normalize_cfs_quota(struct task_group
*tg
,
8288 struct cfs_schedulable_data
*d
)
8296 period
= tg_get_cfs_period(tg
);
8297 quota
= tg_get_cfs_quota(tg
);
8300 /* note: these should typically be equivalent */
8301 if (quota
== RUNTIME_INF
|| quota
== -1)
8304 return to_ratio(period
, quota
);
8307 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8309 struct cfs_schedulable_data
*d
= data
;
8310 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8311 s64 quota
= 0, parent_quota
= -1;
8314 quota
= RUNTIME_INF
;
8316 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8318 quota
= normalize_cfs_quota(tg
, d
);
8319 parent_quota
= parent_b
->hierarchical_quota
;
8322 * ensure max(child_quota) <= parent_quota, inherit when no
8325 if (quota
== RUNTIME_INF
)
8326 quota
= parent_quota
;
8327 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8330 cfs_b
->hierarchical_quota
= quota
;
8335 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8338 struct cfs_schedulable_data data
= {
8344 if (quota
!= RUNTIME_INF
) {
8345 do_div(data
.period
, NSEC_PER_USEC
);
8346 do_div(data
.quota
, NSEC_PER_USEC
);
8350 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8356 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8358 struct task_group
*tg
= css_tg(seq_css(sf
));
8359 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8361 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8362 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8363 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8367 #endif /* CONFIG_CFS_BANDWIDTH */
8368 #endif /* CONFIG_FAIR_GROUP_SCHED */
8370 #ifdef CONFIG_RT_GROUP_SCHED
8371 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8372 struct cftype
*cft
, s64 val
)
8374 return sched_group_set_rt_runtime(css_tg(css
), val
);
8377 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8380 return sched_group_rt_runtime(css_tg(css
));
8383 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8384 struct cftype
*cftype
, u64 rt_period_us
)
8386 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8389 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8392 return sched_group_rt_period(css_tg(css
));
8394 #endif /* CONFIG_RT_GROUP_SCHED */
8396 static struct cftype cpu_files
[] = {
8397 #ifdef CONFIG_FAIR_GROUP_SCHED
8400 .read_u64
= cpu_shares_read_u64
,
8401 .write_u64
= cpu_shares_write_u64
,
8404 #ifdef CONFIG_CFS_BANDWIDTH
8406 .name
= "cfs_quota_us",
8407 .read_s64
= cpu_cfs_quota_read_s64
,
8408 .write_s64
= cpu_cfs_quota_write_s64
,
8411 .name
= "cfs_period_us",
8412 .read_u64
= cpu_cfs_period_read_u64
,
8413 .write_u64
= cpu_cfs_period_write_u64
,
8417 .seq_show
= cpu_stats_show
,
8420 #ifdef CONFIG_RT_GROUP_SCHED
8422 .name
= "rt_runtime_us",
8423 .read_s64
= cpu_rt_runtime_read
,
8424 .write_s64
= cpu_rt_runtime_write
,
8427 .name
= "rt_period_us",
8428 .read_u64
= cpu_rt_period_read_uint
,
8429 .write_u64
= cpu_rt_period_write_uint
,
8435 struct cgroup_subsys cpu_cgrp_subsys
= {
8436 .css_alloc
= cpu_cgroup_css_alloc
,
8437 .css_free
= cpu_cgroup_css_free
,
8438 .css_online
= cpu_cgroup_css_online
,
8439 .css_offline
= cpu_cgroup_css_offline
,
8440 .fork
= cpu_cgroup_fork
,
8441 .can_attach
= cpu_cgroup_can_attach
,
8442 .attach
= cpu_cgroup_attach
,
8443 .legacy_cftypes
= cpu_files
,
8447 #endif /* CONFIG_CGROUP_SCHED */
8449 void dump_cpu_task(int cpu
)
8451 pr_info("Task dump for CPU %d:\n", cpu
);
8452 sched_show_task(cpu_curr(cpu
));
8456 * Nice levels are multiplicative, with a gentle 10% change for every
8457 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8458 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8459 * that remained on nice 0.
8461 * The "10% effect" is relative and cumulative: from _any_ nice level,
8462 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8463 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8464 * If a task goes up by ~10% and another task goes down by ~10% then
8465 * the relative distance between them is ~25%.)
8467 const int sched_prio_to_weight
[40] = {
8468 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8469 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8470 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8471 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8472 /* 0 */ 1024, 820, 655, 526, 423,
8473 /* 5 */ 335, 272, 215, 172, 137,
8474 /* 10 */ 110, 87, 70, 56, 45,
8475 /* 15 */ 36, 29, 23, 18, 15,
8479 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8481 * In cases where the weight does not change often, we can use the
8482 * precalculated inverse to speed up arithmetics by turning divisions
8483 * into multiplications:
8485 const u32 sched_prio_to_wmult
[40] = {
8486 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8487 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8488 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8489 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8490 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8491 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8492 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8493 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,