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 <linux/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>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
78 #include <linux/mutex.h>
80 #include <asm/switch_to.h>
82 #include <asm/irq_regs.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex
);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
97 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
99 void update_rq_clock(struct rq
*rq
)
103 lockdep_assert_held(&rq
->lock
);
105 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
108 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
112 update_rq_clock_task(rq
, delta
);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug
unsigned int sysctl_sched_features
=
123 #include "features.h"
129 * Number of tasks to iterate in a single balance run.
130 * Limited because this is done with IRQs disabled.
132 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
135 * period over which we average the RT time consumption, measured
140 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
143 * period over which we measure -rt task cpu usage in us.
146 unsigned int sysctl_sched_rt_period
= 1000000;
148 __read_mostly
int scheduler_running
;
151 * part of the period that we allow rt tasks to run in us.
154 int sysctl_sched_rt_runtime
= 950000;
156 /* cpus with isolated domains */
157 cpumask_var_t cpu_isolated_map
;
160 * this_rq_lock - lock this runqueue and disable interrupts.
162 static struct rq
*this_rq_lock(void)
169 raw_spin_lock(&rq
->lock
);
175 * __task_rq_lock - lock the rq @p resides on.
177 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
182 lockdep_assert_held(&p
->pi_lock
);
186 raw_spin_lock(&rq
->lock
);
187 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
191 raw_spin_unlock(&rq
->lock
);
193 while (unlikely(task_on_rq_migrating(p
)))
199 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
201 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
202 __acquires(p
->pi_lock
)
208 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
210 raw_spin_lock(&rq
->lock
);
212 * move_queued_task() task_rq_lock()
215 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
216 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
217 * [S] ->cpu = new_cpu [L] task_rq()
221 * If we observe the old cpu in task_rq_lock, the acquire of
222 * the old rq->lock will fully serialize against the stores.
224 * If we observe the new cpu in task_rq_lock, the acquire will
225 * pair with the WMB to ensure we must then also see migrating.
227 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
231 raw_spin_unlock(&rq
->lock
);
232 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
234 while (unlikely(task_on_rq_migrating(p
)))
239 #ifdef CONFIG_SCHED_HRTICK
241 * Use HR-timers to deliver accurate preemption points.
244 static void hrtick_clear(struct rq
*rq
)
246 if (hrtimer_active(&rq
->hrtick_timer
))
247 hrtimer_cancel(&rq
->hrtick_timer
);
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
254 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
256 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
258 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
260 raw_spin_lock(&rq
->lock
);
262 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
263 raw_spin_unlock(&rq
->lock
);
265 return HRTIMER_NORESTART
;
270 static void __hrtick_restart(struct rq
*rq
)
272 struct hrtimer
*timer
= &rq
->hrtick_timer
;
274 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
278 * called from hardirq (IPI) context
280 static void __hrtick_start(void *arg
)
284 raw_spin_lock(&rq
->lock
);
285 __hrtick_restart(rq
);
286 rq
->hrtick_csd_pending
= 0;
287 raw_spin_unlock(&rq
->lock
);
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq
*rq
, u64 delay
)
297 struct hrtimer
*timer
= &rq
->hrtick_timer
;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta
= max_t(s64
, delay
, 10000LL);
306 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
308 hrtimer_set_expires(timer
, time
);
310 if (rq
== this_rq()) {
311 __hrtick_restart(rq
);
312 } else if (!rq
->hrtick_csd_pending
) {
313 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
314 rq
->hrtick_csd_pending
= 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq
*rq
, u64 delay
)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay
= max_t(u64
, delay
, 10000LL);
331 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
332 HRTIMER_MODE_REL_PINNED
);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq
*rq
)
339 rq
->hrtick_csd_pending
= 0;
341 rq
->hrtick_csd
.flags
= 0;
342 rq
->hrtick_csd
.func
= __hrtick_start
;
343 rq
->hrtick_csd
.info
= rq
;
346 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
347 rq
->hrtick_timer
.function
= hrtick
;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq
*rq
)
354 static inline void init_rq_hrtick(struct rq
*rq
)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct
*p
)
385 struct thread_info
*ti
= task_thread_info(p
);
386 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct
*p
)
397 struct thread_info
*ti
= task_thread_info(p
);
398 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
401 if (!(val
& _TIF_POLLING_NRFLAG
))
403 if (val
& _TIF_NEED_RESCHED
)
405 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
414 static bool set_nr_and_not_polling(struct task_struct
*p
)
416 set_tsk_need_resched(p
);
421 static bool set_nr_if_polling(struct task_struct
*p
)
428 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
430 struct wake_q_node
*node
= &task
->wake_q
;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
443 get_task_struct(task
);
446 * The head is context local, there can be no concurrency.
449 head
->lastp
= &node
->next
;
452 void wake_up_q(struct wake_q_head
*head
)
454 struct wake_q_node
*node
= head
->first
;
456 while (node
!= WAKE_Q_TAIL
) {
457 struct task_struct
*task
;
459 task
= container_of(node
, struct task_struct
, wake_q
);
461 /* task can safely be re-inserted now */
463 task
->wake_q
.next
= NULL
;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task
);
470 put_task_struct(task
);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq
*rq
)
483 struct task_struct
*curr
= rq
->curr
;
486 lockdep_assert_held(&rq
->lock
);
488 if (test_tsk_need_resched(curr
))
493 if (cpu
== smp_processor_id()) {
494 set_tsk_need_resched(curr
);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr
))
500 smp_send_reschedule(cpu
);
502 trace_sched_wake_idle_without_ipi(cpu
);
505 void resched_cpu(int cpu
)
507 struct rq
*rq
= cpu_rq(cpu
);
510 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
513 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i
, cpu
= smp_processor_id();
529 struct sched_domain
*sd
;
531 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
535 for_each_domain(cpu
, sd
) {
536 for_each_cpu(i
, sched_domain_span(sd
)) {
540 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
547 if (!is_housekeeping_cpu(cpu
))
548 cpu
= housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu
)
565 struct rq
*rq
= cpu_rq(cpu
);
567 if (cpu
== smp_processor_id())
570 if (set_nr_and_not_polling(rq
->idle
))
571 smp_send_reschedule(cpu
);
573 trace_sched_wake_idle_without_ipi(cpu
);
576 static bool wake_up_full_nohz_cpu(int cpu
)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (cpu_is_offline(cpu
))
585 return true; /* Don't try to wake offline CPUs. */
586 if (tick_nohz_full_cpu(cpu
)) {
587 if (cpu
!= smp_processor_id() ||
588 tick_nohz_tick_stopped())
589 tick_nohz_full_kick_cpu(cpu
);
597 * Wake up the specified CPU. If the CPU is going offline, it is the
598 * caller's responsibility to deal with the lost wakeup, for example,
599 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
601 void wake_up_nohz_cpu(int cpu
)
603 if (!wake_up_full_nohz_cpu(cpu
))
604 wake_up_idle_cpu(cpu
);
607 static inline bool got_nohz_idle_kick(void)
609 int cpu
= smp_processor_id();
611 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
614 if (idle_cpu(cpu
) && !need_resched())
618 * We can't run Idle Load Balance on this CPU for this time so we
619 * cancel it and clear NOHZ_BALANCE_KICK
621 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
625 #else /* CONFIG_NO_HZ_COMMON */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ_COMMON */
634 #ifdef CONFIG_NO_HZ_FULL
635 bool sched_can_stop_tick(struct rq
*rq
)
639 /* Deadline tasks, even if single, need the tick */
640 if (rq
->dl
.dl_nr_running
)
644 * If there are more than one RR tasks, we need the tick to effect the
645 * actual RR behaviour.
647 if (rq
->rt
.rr_nr_running
) {
648 if (rq
->rt
.rr_nr_running
== 1)
655 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 * forced preemption between FIFO tasks.
658 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
663 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 * if there's more than one we need the tick for involuntary
667 if (rq
->nr_running
> 1)
672 #endif /* CONFIG_NO_HZ_FULL */
674 void sched_avg_update(struct rq
*rq
)
676 s64 period
= sched_avg_period();
678 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
680 * Inline assembly required to prevent the compiler
681 * optimising this loop into a divmod call.
682 * See __iter_div_u64_rem() for another example of this.
684 asm("" : "+rm" (rq
->age_stamp
));
685 rq
->age_stamp
+= period
;
690 #endif /* CONFIG_SMP */
692 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
695 * Iterate task_group tree rooted at *from, calling @down when first entering a
696 * node and @up when leaving it for the final time.
698 * Caller must hold rcu_lock or sufficient equivalent.
700 int walk_tg_tree_from(struct task_group
*from
,
701 tg_visitor down
, tg_visitor up
, void *data
)
703 struct task_group
*parent
, *child
;
709 ret
= (*down
)(parent
, data
);
712 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
719 ret
= (*up
)(parent
, data
);
720 if (ret
|| parent
== from
)
724 parent
= parent
->parent
;
731 int tg_nop(struct task_group
*tg
, void *data
)
737 static void set_load_weight(struct task_struct
*p
)
739 int prio
= p
->static_prio
- MAX_RT_PRIO
;
740 struct load_weight
*load
= &p
->se
.load
;
743 * SCHED_IDLE tasks get minimal weight:
745 if (idle_policy(p
->policy
)) {
746 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
747 load
->inv_weight
= WMULT_IDLEPRIO
;
751 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
752 load
->inv_weight
= sched_prio_to_wmult
[prio
];
755 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
758 if (!(flags
& ENQUEUE_RESTORE
))
759 sched_info_queued(rq
, p
);
760 p
->sched_class
->enqueue_task(rq
, p
, flags
);
763 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 if (!(flags
& DEQUEUE_SAVE
))
767 sched_info_dequeued(rq
, p
);
768 p
->sched_class
->dequeue_task(rq
, p
, flags
);
771 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
773 if (task_contributes_to_load(p
))
774 rq
->nr_uninterruptible
--;
776 enqueue_task(rq
, p
, flags
);
779 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
781 if (task_contributes_to_load(p
))
782 rq
->nr_uninterruptible
++;
784 dequeue_task(rq
, p
, flags
);
787 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
790 * In theory, the compile should just see 0 here, and optimize out the call
791 * to sched_rt_avg_update. But I don't trust it...
793 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
794 s64 steal
= 0, irq_delta
= 0;
796 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
797 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
800 * Since irq_time is only updated on {soft,}irq_exit, we might run into
801 * this case when a previous update_rq_clock() happened inside a
804 * When this happens, we stop ->clock_task and only update the
805 * prev_irq_time stamp to account for the part that fit, so that a next
806 * update will consume the rest. This ensures ->clock_task is
809 * It does however cause some slight miss-attribution of {soft,}irq
810 * time, a more accurate solution would be to update the irq_time using
811 * the current rq->clock timestamp, except that would require using
814 if (irq_delta
> delta
)
817 rq
->prev_irq_time
+= irq_delta
;
820 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
821 if (static_key_false((¶virt_steal_rq_enabled
))) {
822 steal
= paravirt_steal_clock(cpu_of(rq
));
823 steal
-= rq
->prev_steal_time_rq
;
825 if (unlikely(steal
> delta
))
828 rq
->prev_steal_time_rq
+= steal
;
833 rq
->clock_task
+= delta
;
835 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
836 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
837 sched_rt_avg_update(rq
, irq_delta
+ steal
);
841 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
843 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
844 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
848 * Make it appear like a SCHED_FIFO task, its something
849 * userspace knows about and won't get confused about.
851 * Also, it will make PI more or less work without too
852 * much confusion -- but then, stop work should not
853 * rely on PI working anyway.
855 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
857 stop
->sched_class
= &stop_sched_class
;
860 cpu_rq(cpu
)->stop
= stop
;
864 * Reset it back to a normal scheduling class so that
865 * it can die in pieces.
867 old_stop
->sched_class
= &rt_sched_class
;
872 * __normal_prio - return the priority that is based on the static prio
874 static inline int __normal_prio(struct task_struct
*p
)
876 return p
->static_prio
;
880 * Calculate the expected normal priority: i.e. priority
881 * without taking RT-inheritance into account. Might be
882 * boosted by interactivity modifiers. Changes upon fork,
883 * setprio syscalls, and whenever the interactivity
884 * estimator recalculates.
886 static inline int normal_prio(struct task_struct
*p
)
890 if (task_has_dl_policy(p
))
891 prio
= MAX_DL_PRIO
-1;
892 else if (task_has_rt_policy(p
))
893 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
895 prio
= __normal_prio(p
);
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
906 static int effective_prio(struct task_struct
*p
)
908 p
->normal_prio
= normal_prio(p
);
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
914 if (!rt_prio(p
->prio
))
915 return p
->normal_prio
;
920 * task_curr - is this task currently executing on a CPU?
921 * @p: the task in question.
923 * Return: 1 if the task is currently executing. 0 otherwise.
925 inline int task_curr(const struct task_struct
*p
)
927 return cpu_curr(task_cpu(p
)) == p
;
931 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
932 * use the balance_callback list if you want balancing.
934 * this means any call to check_class_changed() must be followed by a call to
935 * balance_callback().
937 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
938 const struct sched_class
*prev_class
,
941 if (prev_class
!= p
->sched_class
) {
942 if (prev_class
->switched_from
)
943 prev_class
->switched_from(rq
, p
);
945 p
->sched_class
->switched_to(rq
, p
);
946 } else if (oldprio
!= p
->prio
|| dl_task(p
))
947 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
950 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
952 const struct sched_class
*class;
954 if (p
->sched_class
== rq
->curr
->sched_class
) {
955 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
957 for_each_class(class) {
958 if (class == rq
->curr
->sched_class
)
960 if (class == p
->sched_class
) {
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
972 rq_clock_skip_update(rq
, true);
977 * This is how migration works:
979 * 1) we invoke migration_cpu_stop() on the target CPU using
981 * 2) stopper starts to run (implicitly forcing the migrated thread
983 * 3) it checks whether the migrated task is still in the wrong runqueue.
984 * 4) if it's in the wrong runqueue then the migration thread removes
985 * it and puts it into the right queue.
986 * 5) stopper completes and stop_one_cpu() returns and the migration
991 * move_queued_task - move a queued task to new rq.
993 * Returns (locked) new rq. Old rq's lock is released.
995 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
997 lockdep_assert_held(&rq
->lock
);
999 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1000 dequeue_task(rq
, p
, 0);
1001 set_task_cpu(p
, new_cpu
);
1002 raw_spin_unlock(&rq
->lock
);
1004 rq
= cpu_rq(new_cpu
);
1006 raw_spin_lock(&rq
->lock
);
1007 BUG_ON(task_cpu(p
) != new_cpu
);
1008 enqueue_task(rq
, p
, 0);
1009 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1010 check_preempt_curr(rq
, p
, 0);
1015 struct migration_arg
{
1016 struct task_struct
*task
;
1021 * Move (not current) task off this cpu, onto dest cpu. We're doing
1022 * this because either it can't run here any more (set_cpus_allowed()
1023 * away from this CPU, or CPU going down), or because we're
1024 * attempting to rebalance this task on exec (sched_exec).
1026 * So we race with normal scheduler movements, but that's OK, as long
1027 * as the task is no longer on this CPU.
1029 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1031 if (unlikely(!cpu_active(dest_cpu
)))
1034 /* Affinity changed (again). */
1035 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1038 rq
= move_queued_task(rq
, p
, dest_cpu
);
1044 * migration_cpu_stop - this will be executed by a highprio stopper thread
1045 * and performs thread migration by bumping thread off CPU then
1046 * 'pushing' onto another runqueue.
1048 static int migration_cpu_stop(void *data
)
1050 struct migration_arg
*arg
= data
;
1051 struct task_struct
*p
= arg
->task
;
1052 struct rq
*rq
= this_rq();
1055 * The original target cpu might have gone down and we might
1056 * be on another cpu but it doesn't matter.
1058 local_irq_disable();
1060 * We need to explicitly wake pending tasks before running
1061 * __migrate_task() such that we will not miss enforcing cpus_allowed
1062 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1064 sched_ttwu_pending();
1066 raw_spin_lock(&p
->pi_lock
);
1067 raw_spin_lock(&rq
->lock
);
1069 * If task_rq(p) != rq, it cannot be migrated here, because we're
1070 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1071 * we're holding p->pi_lock.
1073 if (task_rq(p
) == rq
) {
1074 if (task_on_rq_queued(p
))
1075 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1077 p
->wake_cpu
= arg
->dest_cpu
;
1079 raw_spin_unlock(&rq
->lock
);
1080 raw_spin_unlock(&p
->pi_lock
);
1087 * sched_class::set_cpus_allowed must do the below, but is not required to
1088 * actually call this function.
1090 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1092 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1093 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1096 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1098 struct rq
*rq
= task_rq(p
);
1099 bool queued
, running
;
1101 lockdep_assert_held(&p
->pi_lock
);
1103 queued
= task_on_rq_queued(p
);
1104 running
= task_current(rq
, p
);
1108 * Because __kthread_bind() calls this on blocked tasks without
1111 lockdep_assert_held(&rq
->lock
);
1112 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1115 put_prev_task(rq
, p
);
1117 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1120 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1122 set_curr_task(rq
, p
);
1126 * Change a given task's CPU affinity. Migrate the thread to a
1127 * proper CPU and schedule it away if the CPU it's executing on
1128 * is removed from the allowed bitmask.
1130 * NOTE: the caller must have a valid reference to the task, the
1131 * task must not exit() & deallocate itself prematurely. The
1132 * call is not atomic; no spinlocks may be held.
1134 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1135 const struct cpumask
*new_mask
, bool check
)
1137 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1138 unsigned int dest_cpu
;
1143 rq
= task_rq_lock(p
, &rf
);
1145 if (p
->flags
& PF_KTHREAD
) {
1147 * Kernel threads are allowed on online && !active CPUs
1149 cpu_valid_mask
= cpu_online_mask
;
1153 * Must re-check here, to close a race against __kthread_bind(),
1154 * sched_setaffinity() is not guaranteed to observe the flag.
1156 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1161 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1164 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1169 do_set_cpus_allowed(p
, new_mask
);
1171 if (p
->flags
& PF_KTHREAD
) {
1173 * For kernel threads that do indeed end up on online &&
1174 * !active we want to ensure they are strict per-cpu threads.
1176 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1177 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1178 p
->nr_cpus_allowed
!= 1);
1181 /* Can the task run on the task's current CPU? If so, we're done */
1182 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1185 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1186 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1187 struct migration_arg arg
= { p
, dest_cpu
};
1188 /* Need help from migration thread: drop lock and wait. */
1189 task_rq_unlock(rq
, p
, &rf
);
1190 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1191 tlb_migrate_finish(p
->mm
);
1193 } else if (task_on_rq_queued(p
)) {
1195 * OK, since we're going to drop the lock immediately
1196 * afterwards anyway.
1198 rq_unpin_lock(rq
, &rf
);
1199 rq
= move_queued_task(rq
, p
, dest_cpu
);
1200 rq_repin_lock(rq
, &rf
);
1203 task_rq_unlock(rq
, p
, &rf
);
1208 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1210 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1212 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1214 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1216 #ifdef CONFIG_SCHED_DEBUG
1218 * We should never call set_task_cpu() on a blocked task,
1219 * ttwu() will sort out the placement.
1221 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1225 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1226 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1227 * time relying on p->on_rq.
1229 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1230 p
->sched_class
== &fair_sched_class
&&
1231 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1233 #ifdef CONFIG_LOCKDEP
1235 * The caller should hold either p->pi_lock or rq->lock, when changing
1236 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1238 * sched_move_task() holds both and thus holding either pins the cgroup,
1241 * Furthermore, all task_rq users should acquire both locks, see
1244 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1245 lockdep_is_held(&task_rq(p
)->lock
)));
1249 trace_sched_migrate_task(p
, new_cpu
);
1251 if (task_cpu(p
) != new_cpu
) {
1252 if (p
->sched_class
->migrate_task_rq
)
1253 p
->sched_class
->migrate_task_rq(p
);
1254 p
->se
.nr_migrations
++;
1255 perf_event_task_migrate(p
);
1258 __set_task_cpu(p
, new_cpu
);
1261 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1263 if (task_on_rq_queued(p
)) {
1264 struct rq
*src_rq
, *dst_rq
;
1266 src_rq
= task_rq(p
);
1267 dst_rq
= cpu_rq(cpu
);
1269 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1270 deactivate_task(src_rq
, p
, 0);
1271 set_task_cpu(p
, cpu
);
1272 activate_task(dst_rq
, p
, 0);
1273 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1274 check_preempt_curr(dst_rq
, p
, 0);
1277 * Task isn't running anymore; make it appear like we migrated
1278 * it before it went to sleep. This means on wakeup we make the
1279 * previous cpu our target instead of where it really is.
1285 struct migration_swap_arg
{
1286 struct task_struct
*src_task
, *dst_task
;
1287 int src_cpu
, dst_cpu
;
1290 static int migrate_swap_stop(void *data
)
1292 struct migration_swap_arg
*arg
= data
;
1293 struct rq
*src_rq
, *dst_rq
;
1296 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1299 src_rq
= cpu_rq(arg
->src_cpu
);
1300 dst_rq
= cpu_rq(arg
->dst_cpu
);
1302 double_raw_lock(&arg
->src_task
->pi_lock
,
1303 &arg
->dst_task
->pi_lock
);
1304 double_rq_lock(src_rq
, dst_rq
);
1306 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1309 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1312 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1315 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1318 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1319 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1324 double_rq_unlock(src_rq
, dst_rq
);
1325 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1326 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1332 * Cross migrate two tasks
1334 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1336 struct migration_swap_arg arg
;
1339 arg
= (struct migration_swap_arg
){
1341 .src_cpu
= task_cpu(cur
),
1343 .dst_cpu
= task_cpu(p
),
1346 if (arg
.src_cpu
== arg
.dst_cpu
)
1350 * These three tests are all lockless; this is OK since all of them
1351 * will be re-checked with proper locks held further down the line.
1353 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1356 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1359 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1362 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1363 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1370 * wait_task_inactive - wait for a thread to unschedule.
1372 * If @match_state is nonzero, it's the @p->state value just checked and
1373 * not expected to change. If it changes, i.e. @p might have woken up,
1374 * then return zero. When we succeed in waiting for @p to be off its CPU,
1375 * we return a positive number (its total switch count). If a second call
1376 * a short while later returns the same number, the caller can be sure that
1377 * @p has remained unscheduled the whole time.
1379 * The caller must ensure that the task *will* unschedule sometime soon,
1380 * else this function might spin for a *long* time. This function can't
1381 * be called with interrupts off, or it may introduce deadlock with
1382 * smp_call_function() if an IPI is sent by the same process we are
1383 * waiting to become inactive.
1385 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1387 int running
, queued
;
1394 * We do the initial early heuristics without holding
1395 * any task-queue locks at all. We'll only try to get
1396 * the runqueue lock when things look like they will
1402 * If the task is actively running on another CPU
1403 * still, just relax and busy-wait without holding
1406 * NOTE! Since we don't hold any locks, it's not
1407 * even sure that "rq" stays as the right runqueue!
1408 * But we don't care, since "task_running()" will
1409 * return false if the runqueue has changed and p
1410 * is actually now running somewhere else!
1412 while (task_running(rq
, p
)) {
1413 if (match_state
&& unlikely(p
->state
!= match_state
))
1419 * Ok, time to look more closely! We need the rq
1420 * lock now, to be *sure*. If we're wrong, we'll
1421 * just go back and repeat.
1423 rq
= task_rq_lock(p
, &rf
);
1424 trace_sched_wait_task(p
);
1425 running
= task_running(rq
, p
);
1426 queued
= task_on_rq_queued(p
);
1428 if (!match_state
|| p
->state
== match_state
)
1429 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1430 task_rq_unlock(rq
, p
, &rf
);
1433 * If it changed from the expected state, bail out now.
1435 if (unlikely(!ncsw
))
1439 * Was it really running after all now that we
1440 * checked with the proper locks actually held?
1442 * Oops. Go back and try again..
1444 if (unlikely(running
)) {
1450 * It's not enough that it's not actively running,
1451 * it must be off the runqueue _entirely_, and not
1454 * So if it was still runnable (but just not actively
1455 * running right now), it's preempted, and we should
1456 * yield - it could be a while.
1458 if (unlikely(queued
)) {
1459 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1461 set_current_state(TASK_UNINTERRUPTIBLE
);
1462 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1467 * Ahh, all good. It wasn't running, and it wasn't
1468 * runnable, which means that it will never become
1469 * running in the future either. We're all done!
1478 * kick_process - kick a running thread to enter/exit the kernel
1479 * @p: the to-be-kicked thread
1481 * Cause a process which is running on another CPU to enter
1482 * kernel-mode, without any delay. (to get signals handled.)
1484 * NOTE: this function doesn't have to take the runqueue lock,
1485 * because all it wants to ensure is that the remote task enters
1486 * the kernel. If the IPI races and the task has been migrated
1487 * to another CPU then no harm is done and the purpose has been
1490 void kick_process(struct task_struct
*p
)
1496 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1497 smp_send_reschedule(cpu
);
1500 EXPORT_SYMBOL_GPL(kick_process
);
1503 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1505 * A few notes on cpu_active vs cpu_online:
1507 * - cpu_active must be a subset of cpu_online
1509 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1510 * see __set_cpus_allowed_ptr(). At this point the newly online
1511 * cpu isn't yet part of the sched domains, and balancing will not
1514 * - on cpu-down we clear cpu_active() to mask the sched domains and
1515 * avoid the load balancer to place new tasks on the to be removed
1516 * cpu. Existing tasks will remain running there and will be taken
1519 * This means that fallback selection must not select !active CPUs.
1520 * And can assume that any active CPU must be online. Conversely
1521 * select_task_rq() below may allow selection of !active CPUs in order
1522 * to satisfy the above rules.
1524 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1526 int nid
= cpu_to_node(cpu
);
1527 const struct cpumask
*nodemask
= NULL
;
1528 enum { cpuset
, possible
, fail
} state
= cpuset
;
1532 * If the node that the cpu is on has been offlined, cpu_to_node()
1533 * will return -1. There is no cpu on the node, and we should
1534 * select the cpu on the other node.
1537 nodemask
= cpumask_of_node(nid
);
1539 /* Look for allowed, online CPU in same node. */
1540 for_each_cpu(dest_cpu
, nodemask
) {
1541 if (!cpu_active(dest_cpu
))
1543 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1549 /* Any allowed, online CPU? */
1550 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1551 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1553 if (!cpu_online(dest_cpu
))
1558 /* No more Mr. Nice Guy. */
1561 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1562 cpuset_cpus_allowed_fallback(p
);
1568 do_set_cpus_allowed(p
, cpu_possible_mask
);
1579 if (state
!= cpuset
) {
1581 * Don't tell them about moving exiting tasks or
1582 * kernel threads (both mm NULL), since they never
1585 if (p
->mm
&& printk_ratelimit()) {
1586 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1587 task_pid_nr(p
), p
->comm
, cpu
);
1595 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1598 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1600 lockdep_assert_held(&p
->pi_lock
);
1602 if (tsk_nr_cpus_allowed(p
) > 1)
1603 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1605 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1608 * In order not to call set_task_cpu() on a blocking task we need
1609 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1612 * Since this is common to all placement strategies, this lives here.
1614 * [ this allows ->select_task() to simply return task_cpu(p) and
1615 * not worry about this generic constraint ]
1617 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1619 cpu
= select_fallback_rq(task_cpu(p
), p
);
1624 static void update_avg(u64
*avg
, u64 sample
)
1626 s64 diff
= sample
- *avg
;
1632 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1633 const struct cpumask
*new_mask
, bool check
)
1635 return set_cpus_allowed_ptr(p
, new_mask
);
1638 #endif /* CONFIG_SMP */
1641 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1645 if (!schedstat_enabled())
1651 if (cpu
== rq
->cpu
) {
1652 schedstat_inc(rq
->ttwu_local
);
1653 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1655 struct sched_domain
*sd
;
1657 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1659 for_each_domain(rq
->cpu
, sd
) {
1660 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1661 schedstat_inc(sd
->ttwu_wake_remote
);
1668 if (wake_flags
& WF_MIGRATED
)
1669 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1670 #endif /* CONFIG_SMP */
1672 schedstat_inc(rq
->ttwu_count
);
1673 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1675 if (wake_flags
& WF_SYNC
)
1676 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1679 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1681 activate_task(rq
, p
, en_flags
);
1682 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1684 /* if a worker is waking up, notify workqueue */
1685 if (p
->flags
& PF_WQ_WORKER
)
1686 wq_worker_waking_up(p
, cpu_of(rq
));
1690 * Mark the task runnable and perform wakeup-preemption.
1692 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1693 struct rq_flags
*rf
)
1695 check_preempt_curr(rq
, p
, wake_flags
);
1696 p
->state
= TASK_RUNNING
;
1697 trace_sched_wakeup(p
);
1700 if (p
->sched_class
->task_woken
) {
1702 * Our task @p is fully woken up and running; so its safe to
1703 * drop the rq->lock, hereafter rq is only used for statistics.
1705 rq_unpin_lock(rq
, rf
);
1706 p
->sched_class
->task_woken(rq
, p
);
1707 rq_repin_lock(rq
, rf
);
1710 if (rq
->idle_stamp
) {
1711 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1712 u64 max
= 2*rq
->max_idle_balance_cost
;
1714 update_avg(&rq
->avg_idle
, delta
);
1716 if (rq
->avg_idle
> max
)
1725 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1726 struct rq_flags
*rf
)
1728 int en_flags
= ENQUEUE_WAKEUP
;
1730 lockdep_assert_held(&rq
->lock
);
1733 if (p
->sched_contributes_to_load
)
1734 rq
->nr_uninterruptible
--;
1736 if (wake_flags
& WF_MIGRATED
)
1737 en_flags
|= ENQUEUE_MIGRATED
;
1740 ttwu_activate(rq
, p
, en_flags
);
1741 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1745 * Called in case the task @p isn't fully descheduled from its runqueue,
1746 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1747 * since all we need to do is flip p->state to TASK_RUNNING, since
1748 * the task is still ->on_rq.
1750 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1756 rq
= __task_rq_lock(p
, &rf
);
1757 if (task_on_rq_queued(p
)) {
1758 /* check_preempt_curr() may use rq clock */
1759 update_rq_clock(rq
);
1760 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1763 __task_rq_unlock(rq
, &rf
);
1769 void sched_ttwu_pending(void)
1771 struct rq
*rq
= this_rq();
1772 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1773 struct task_struct
*p
;
1774 unsigned long flags
;
1780 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1781 rq_pin_lock(rq
, &rf
);
1786 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1787 llist
= llist_next(llist
);
1789 if (p
->sched_remote_wakeup
)
1790 wake_flags
= WF_MIGRATED
;
1792 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1795 rq_unpin_lock(rq
, &rf
);
1796 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1799 void scheduler_ipi(void)
1802 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1803 * TIF_NEED_RESCHED remotely (for the first time) will also send
1806 preempt_fold_need_resched();
1808 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1812 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1813 * traditionally all their work was done from the interrupt return
1814 * path. Now that we actually do some work, we need to make sure
1817 * Some archs already do call them, luckily irq_enter/exit nest
1820 * Arguably we should visit all archs and update all handlers,
1821 * however a fair share of IPIs are still resched only so this would
1822 * somewhat pessimize the simple resched case.
1825 sched_ttwu_pending();
1828 * Check if someone kicked us for doing the nohz idle load balance.
1830 if (unlikely(got_nohz_idle_kick())) {
1831 this_rq()->idle_balance
= 1;
1832 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1837 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1839 struct rq
*rq
= cpu_rq(cpu
);
1841 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1843 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1844 if (!set_nr_if_polling(rq
->idle
))
1845 smp_send_reschedule(cpu
);
1847 trace_sched_wake_idle_without_ipi(cpu
);
1851 void wake_up_if_idle(int cpu
)
1853 struct rq
*rq
= cpu_rq(cpu
);
1854 unsigned long flags
;
1858 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1861 if (set_nr_if_polling(rq
->idle
)) {
1862 trace_sched_wake_idle_without_ipi(cpu
);
1864 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1865 if (is_idle_task(rq
->curr
))
1866 smp_send_reschedule(cpu
);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1875 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1877 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1879 #endif /* CONFIG_SMP */
1881 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1883 struct rq
*rq
= cpu_rq(cpu
);
1886 #if defined(CONFIG_SMP)
1887 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1888 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1889 ttwu_queue_remote(p
, cpu
, wake_flags
);
1894 raw_spin_lock(&rq
->lock
);
1895 rq_pin_lock(rq
, &rf
);
1896 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1897 rq_unpin_lock(rq
, &rf
);
1898 raw_spin_unlock(&rq
->lock
);
1902 * Notes on Program-Order guarantees on SMP systems.
1906 * The basic program-order guarantee on SMP systems is that when a task [t]
1907 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1908 * execution on its new cpu [c1].
1910 * For migration (of runnable tasks) this is provided by the following means:
1912 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1913 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1914 * rq(c1)->lock (if not at the same time, then in that order).
1915 * C) LOCK of the rq(c1)->lock scheduling in task
1917 * Transitivity guarantees that B happens after A and C after B.
1918 * Note: we only require RCpc transitivity.
1919 * Note: the cpu doing B need not be c0 or c1
1928 * UNLOCK rq(0)->lock
1930 * LOCK rq(0)->lock // orders against CPU0
1932 * UNLOCK rq(0)->lock
1936 * UNLOCK rq(1)->lock
1938 * LOCK rq(1)->lock // orders against CPU2
1941 * UNLOCK rq(1)->lock
1944 * BLOCKING -- aka. SLEEP + WAKEUP
1946 * For blocking we (obviously) need to provide the same guarantee as for
1947 * migration. However the means are completely different as there is no lock
1948 * chain to provide order. Instead we do:
1950 * 1) smp_store_release(X->on_cpu, 0)
1951 * 2) smp_cond_load_acquire(!X->on_cpu)
1955 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1957 * LOCK rq(0)->lock LOCK X->pi_lock
1960 * smp_store_release(X->on_cpu, 0);
1962 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1968 * X->state = RUNNING
1969 * UNLOCK rq(2)->lock
1971 * LOCK rq(2)->lock // orders against CPU1
1974 * UNLOCK rq(2)->lock
1977 * UNLOCK rq(0)->lock
1980 * However; for wakeups there is a second guarantee we must provide, namely we
1981 * must observe the state that lead to our wakeup. That is, not only must our
1982 * task observe its own prior state, it must also observe the stores prior to
1985 * This means that any means of doing remote wakeups must order the CPU doing
1986 * the wakeup against the CPU the task is going to end up running on. This,
1987 * however, is already required for the regular Program-Order guarantee above,
1988 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1993 * try_to_wake_up - wake up a thread
1994 * @p: the thread to be awakened
1995 * @state: the mask of task states that can be woken
1996 * @wake_flags: wake modifier flags (WF_*)
1998 * If (@state & @p->state) @p->state = TASK_RUNNING.
2000 * If the task was not queued/runnable, also place it back on a runqueue.
2002 * Atomic against schedule() which would dequeue a task, also see
2003 * set_current_state().
2005 * Return: %true if @p->state changes (an actual wakeup was done),
2009 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2011 unsigned long flags
;
2012 int cpu
, success
= 0;
2015 * If we are going to wake up a thread waiting for CONDITION we
2016 * need to ensure that CONDITION=1 done by the caller can not be
2017 * reordered with p->state check below. This pairs with mb() in
2018 * set_current_state() the waiting thread does.
2020 smp_mb__before_spinlock();
2021 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2022 if (!(p
->state
& state
))
2025 trace_sched_waking(p
);
2027 success
= 1; /* we're going to change ->state */
2031 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2032 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2033 * in smp_cond_load_acquire() below.
2035 * sched_ttwu_pending() try_to_wake_up()
2036 * [S] p->on_rq = 1; [L] P->state
2037 * UNLOCK rq->lock -----.
2041 * LOCK rq->lock -----'
2045 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2047 * Pairs with the UNLOCK+LOCK on rq->lock from the
2048 * last wakeup of our task and the schedule that got our task
2052 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2057 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2058 * possible to, falsely, observe p->on_cpu == 0.
2060 * One must be running (->on_cpu == 1) in order to remove oneself
2061 * from the runqueue.
2063 * [S] ->on_cpu = 1; [L] ->on_rq
2067 * [S] ->on_rq = 0; [L] ->on_cpu
2069 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2070 * from the consecutive calls to schedule(); the first switching to our
2071 * task, the second putting it to sleep.
2076 * If the owning (remote) cpu is still in the middle of schedule() with
2077 * this task as prev, wait until its done referencing the task.
2079 * Pairs with the smp_store_release() in finish_lock_switch().
2081 * This ensures that tasks getting woken will be fully ordered against
2082 * their previous state and preserve Program Order.
2084 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2086 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2087 p
->state
= TASK_WAKING
;
2089 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2090 if (task_cpu(p
) != cpu
) {
2091 wake_flags
|= WF_MIGRATED
;
2092 set_task_cpu(p
, cpu
);
2094 #endif /* CONFIG_SMP */
2096 ttwu_queue(p
, cpu
, wake_flags
);
2098 ttwu_stat(p
, cpu
, wake_flags
);
2100 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2106 * try_to_wake_up_local - try to wake up a local task with rq lock held
2107 * @p: the thread to be awakened
2108 * @cookie: context's cookie for pinning
2110 * Put @p on the run-queue if it's not already there. The caller must
2111 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2114 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2116 struct rq
*rq
= task_rq(p
);
2118 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2119 WARN_ON_ONCE(p
== current
))
2122 lockdep_assert_held(&rq
->lock
);
2124 if (!raw_spin_trylock(&p
->pi_lock
)) {
2126 * This is OK, because current is on_cpu, which avoids it being
2127 * picked for load-balance and preemption/IRQs are still
2128 * disabled avoiding further scheduler activity on it and we've
2129 * not yet picked a replacement task.
2131 rq_unpin_lock(rq
, rf
);
2132 raw_spin_unlock(&rq
->lock
);
2133 raw_spin_lock(&p
->pi_lock
);
2134 raw_spin_lock(&rq
->lock
);
2135 rq_repin_lock(rq
, rf
);
2138 if (!(p
->state
& TASK_NORMAL
))
2141 trace_sched_waking(p
);
2143 if (!task_on_rq_queued(p
))
2144 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2146 ttwu_do_wakeup(rq
, p
, 0, rf
);
2147 ttwu_stat(p
, smp_processor_id(), 0);
2149 raw_spin_unlock(&p
->pi_lock
);
2153 * wake_up_process - Wake up a specific process
2154 * @p: The process to be woken up.
2156 * Attempt to wake up the nominated process and move it to the set of runnable
2159 * Return: 1 if the process was woken up, 0 if it was already running.
2161 * It may be assumed that this function implies a write memory barrier before
2162 * changing the task state if and only if any tasks are woken up.
2164 int wake_up_process(struct task_struct
*p
)
2166 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2168 EXPORT_SYMBOL(wake_up_process
);
2170 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2172 return try_to_wake_up(p
, state
, 0);
2176 * This function clears the sched_dl_entity static params.
2178 void __dl_clear_params(struct task_struct
*p
)
2180 struct sched_dl_entity
*dl_se
= &p
->dl
;
2182 dl_se
->dl_runtime
= 0;
2183 dl_se
->dl_deadline
= 0;
2184 dl_se
->dl_period
= 0;
2188 dl_se
->dl_throttled
= 0;
2189 dl_se
->dl_yielded
= 0;
2193 * Perform scheduler related setup for a newly forked process p.
2194 * p is forked by current.
2196 * __sched_fork() is basic setup used by init_idle() too:
2198 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2203 p
->se
.exec_start
= 0;
2204 p
->se
.sum_exec_runtime
= 0;
2205 p
->se
.prev_sum_exec_runtime
= 0;
2206 p
->se
.nr_migrations
= 0;
2208 INIT_LIST_HEAD(&p
->se
.group_node
);
2210 #ifdef CONFIG_FAIR_GROUP_SCHED
2211 p
->se
.cfs_rq
= NULL
;
2214 #ifdef CONFIG_SCHEDSTATS
2215 /* Even if schedstat is disabled, there should not be garbage */
2216 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2219 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2220 init_dl_task_timer(&p
->dl
);
2221 __dl_clear_params(p
);
2223 INIT_LIST_HEAD(&p
->rt
.run_list
);
2225 p
->rt
.time_slice
= sched_rr_timeslice
;
2229 #ifdef CONFIG_PREEMPT_NOTIFIERS
2230 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2233 #ifdef CONFIG_NUMA_BALANCING
2234 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2235 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2236 p
->mm
->numa_scan_seq
= 0;
2239 if (clone_flags
& CLONE_VM
)
2240 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2242 p
->numa_preferred_nid
= -1;
2244 p
->node_stamp
= 0ULL;
2245 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2246 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2247 p
->numa_work
.next
= &p
->numa_work
;
2248 p
->numa_faults
= NULL
;
2249 p
->last_task_numa_placement
= 0;
2250 p
->last_sum_exec_runtime
= 0;
2252 p
->numa_group
= NULL
;
2253 #endif /* CONFIG_NUMA_BALANCING */
2256 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2258 #ifdef CONFIG_NUMA_BALANCING
2260 void set_numabalancing_state(bool enabled
)
2263 static_branch_enable(&sched_numa_balancing
);
2265 static_branch_disable(&sched_numa_balancing
);
2268 #ifdef CONFIG_PROC_SYSCTL
2269 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2270 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2274 int state
= static_branch_likely(&sched_numa_balancing
);
2276 if (write
&& !capable(CAP_SYS_ADMIN
))
2281 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2285 set_numabalancing_state(state
);
2291 #ifdef CONFIG_SCHEDSTATS
2293 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2294 static bool __initdata __sched_schedstats
= false;
2296 static void set_schedstats(bool enabled
)
2299 static_branch_enable(&sched_schedstats
);
2301 static_branch_disable(&sched_schedstats
);
2304 void force_schedstat_enabled(void)
2306 if (!schedstat_enabled()) {
2307 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2308 static_branch_enable(&sched_schedstats
);
2312 static int __init
setup_schedstats(char *str
)
2319 * This code is called before jump labels have been set up, so we can't
2320 * change the static branch directly just yet. Instead set a temporary
2321 * variable so init_schedstats() can do it later.
2323 if (!strcmp(str
, "enable")) {
2324 __sched_schedstats
= true;
2326 } else if (!strcmp(str
, "disable")) {
2327 __sched_schedstats
= false;
2332 pr_warn("Unable to parse schedstats=\n");
2336 __setup("schedstats=", setup_schedstats
);
2338 static void __init
init_schedstats(void)
2340 set_schedstats(__sched_schedstats
);
2343 #ifdef CONFIG_PROC_SYSCTL
2344 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2345 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2349 int state
= static_branch_likely(&sched_schedstats
);
2351 if (write
&& !capable(CAP_SYS_ADMIN
))
2356 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2360 set_schedstats(state
);
2363 #endif /* CONFIG_PROC_SYSCTL */
2364 #else /* !CONFIG_SCHEDSTATS */
2365 static inline void init_schedstats(void) {}
2366 #endif /* CONFIG_SCHEDSTATS */
2369 * fork()/clone()-time setup:
2371 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2373 unsigned long flags
;
2374 int cpu
= get_cpu();
2376 __sched_fork(clone_flags
, p
);
2378 * We mark the process as NEW here. This guarantees that
2379 * nobody will actually run it, and a signal or other external
2380 * event cannot wake it up and insert it on the runqueue either.
2382 p
->state
= TASK_NEW
;
2385 * Make sure we do not leak PI boosting priority to the child.
2387 p
->prio
= current
->normal_prio
;
2390 * Revert to default priority/policy on fork if requested.
2392 if (unlikely(p
->sched_reset_on_fork
)) {
2393 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2394 p
->policy
= SCHED_NORMAL
;
2395 p
->static_prio
= NICE_TO_PRIO(0);
2397 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2398 p
->static_prio
= NICE_TO_PRIO(0);
2400 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2404 * We don't need the reset flag anymore after the fork. It has
2405 * fulfilled its duty:
2407 p
->sched_reset_on_fork
= 0;
2410 if (dl_prio(p
->prio
)) {
2413 } else if (rt_prio(p
->prio
)) {
2414 p
->sched_class
= &rt_sched_class
;
2416 p
->sched_class
= &fair_sched_class
;
2419 init_entity_runnable_average(&p
->se
);
2422 * The child is not yet in the pid-hash so no cgroup attach races,
2423 * and the cgroup is pinned to this child due to cgroup_fork()
2424 * is ran before sched_fork().
2426 * Silence PROVE_RCU.
2428 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2430 * We're setting the cpu for the first time, we don't migrate,
2431 * so use __set_task_cpu().
2433 __set_task_cpu(p
, cpu
);
2434 if (p
->sched_class
->task_fork
)
2435 p
->sched_class
->task_fork(p
);
2436 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2438 #ifdef CONFIG_SCHED_INFO
2439 if (likely(sched_info_on()))
2440 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2442 #if defined(CONFIG_SMP)
2445 init_task_preempt_count(p
);
2447 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2448 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2455 unsigned long to_ratio(u64 period
, u64 runtime
)
2457 if (runtime
== RUNTIME_INF
)
2461 * Doing this here saves a lot of checks in all
2462 * the calling paths, and returning zero seems
2463 * safe for them anyway.
2468 return div64_u64(runtime
<< 20, period
);
2472 inline struct dl_bw
*dl_bw_of(int i
)
2474 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2475 "sched RCU must be held");
2476 return &cpu_rq(i
)->rd
->dl_bw
;
2479 static inline int dl_bw_cpus(int i
)
2481 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2484 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2485 "sched RCU must be held");
2486 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2492 inline struct dl_bw
*dl_bw_of(int i
)
2494 return &cpu_rq(i
)->dl
.dl_bw
;
2497 static inline int dl_bw_cpus(int i
)
2504 * We must be sure that accepting a new task (or allowing changing the
2505 * parameters of an existing one) is consistent with the bandwidth
2506 * constraints. If yes, this function also accordingly updates the currently
2507 * allocated bandwidth to reflect the new situation.
2509 * This function is called while holding p's rq->lock.
2511 * XXX we should delay bw change until the task's 0-lag point, see
2514 static int dl_overflow(struct task_struct
*p
, int policy
,
2515 const struct sched_attr
*attr
)
2518 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2519 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2520 u64 runtime
= attr
->sched_runtime
;
2521 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2524 /* !deadline task may carry old deadline bandwidth */
2525 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2529 * Either if a task, enters, leave, or stays -deadline but changes
2530 * its parameters, we may need to update accordingly the total
2531 * allocated bandwidth of the container.
2533 raw_spin_lock(&dl_b
->lock
);
2534 cpus
= dl_bw_cpus(task_cpu(p
));
2535 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2536 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2537 __dl_add(dl_b
, new_bw
);
2539 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2540 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2541 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2542 __dl_add(dl_b
, new_bw
);
2544 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2545 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2548 raw_spin_unlock(&dl_b
->lock
);
2553 extern void init_dl_bw(struct dl_bw
*dl_b
);
2556 * wake_up_new_task - wake up a newly created task for the first time.
2558 * This function will do some initial scheduler statistics housekeeping
2559 * that must be done for every newly created context, then puts the task
2560 * on the runqueue and wakes it.
2562 void wake_up_new_task(struct task_struct
*p
)
2567 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2568 p
->state
= TASK_RUNNING
;
2571 * Fork balancing, do it here and not earlier because:
2572 * - cpus_allowed can change in the fork path
2573 * - any previously selected cpu might disappear through hotplug
2575 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2576 * as we're not fully set-up yet.
2578 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2580 rq
= __task_rq_lock(p
, &rf
);
2581 update_rq_clock(rq
);
2582 post_init_entity_util_avg(&p
->se
);
2584 activate_task(rq
, p
, 0);
2585 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2586 trace_sched_wakeup_new(p
);
2587 check_preempt_curr(rq
, p
, WF_FORK
);
2589 if (p
->sched_class
->task_woken
) {
2591 * Nothing relies on rq->lock after this, so its fine to
2594 rq_unpin_lock(rq
, &rf
);
2595 p
->sched_class
->task_woken(rq
, p
);
2596 rq_repin_lock(rq
, &rf
);
2599 task_rq_unlock(rq
, p
, &rf
);
2602 #ifdef CONFIG_PREEMPT_NOTIFIERS
2604 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2606 void preempt_notifier_inc(void)
2608 static_key_slow_inc(&preempt_notifier_key
);
2610 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2612 void preempt_notifier_dec(void)
2614 static_key_slow_dec(&preempt_notifier_key
);
2616 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2619 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2620 * @notifier: notifier struct to register
2622 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2624 if (!static_key_false(&preempt_notifier_key
))
2625 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2627 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2629 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2632 * preempt_notifier_unregister - no longer interested in preemption notifications
2633 * @notifier: notifier struct to unregister
2635 * This is *not* safe to call from within a preemption notifier.
2637 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2639 hlist_del(¬ifier
->link
);
2641 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2643 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2645 struct preempt_notifier
*notifier
;
2647 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2648 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2651 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2653 if (static_key_false(&preempt_notifier_key
))
2654 __fire_sched_in_preempt_notifiers(curr
);
2658 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2659 struct task_struct
*next
)
2661 struct preempt_notifier
*notifier
;
2663 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2664 notifier
->ops
->sched_out(notifier
, next
);
2667 static __always_inline
void
2668 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2669 struct task_struct
*next
)
2671 if (static_key_false(&preempt_notifier_key
))
2672 __fire_sched_out_preempt_notifiers(curr
, next
);
2675 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2677 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2682 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2683 struct task_struct
*next
)
2687 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2690 * prepare_task_switch - prepare to switch tasks
2691 * @rq: the runqueue preparing to switch
2692 * @prev: the current task that is being switched out
2693 * @next: the task we are going to switch to.
2695 * This is called with the rq lock held and interrupts off. It must
2696 * be paired with a subsequent finish_task_switch after the context
2699 * prepare_task_switch sets up locking and calls architecture specific
2703 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2704 struct task_struct
*next
)
2706 sched_info_switch(rq
, prev
, next
);
2707 perf_event_task_sched_out(prev
, next
);
2708 fire_sched_out_preempt_notifiers(prev
, next
);
2709 prepare_lock_switch(rq
, next
);
2710 prepare_arch_switch(next
);
2714 * finish_task_switch - clean up after a task-switch
2715 * @prev: the thread we just switched away from.
2717 * finish_task_switch must be called after the context switch, paired
2718 * with a prepare_task_switch call before the context switch.
2719 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2720 * and do any other architecture-specific cleanup actions.
2722 * Note that we may have delayed dropping an mm in context_switch(). If
2723 * so, we finish that here outside of the runqueue lock. (Doing it
2724 * with the lock held can cause deadlocks; see schedule() for
2727 * The context switch have flipped the stack from under us and restored the
2728 * local variables which were saved when this task called schedule() in the
2729 * past. prev == current is still correct but we need to recalculate this_rq
2730 * because prev may have moved to another CPU.
2732 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2733 __releases(rq
->lock
)
2735 struct rq
*rq
= this_rq();
2736 struct mm_struct
*mm
= rq
->prev_mm
;
2740 * The previous task will have left us with a preempt_count of 2
2741 * because it left us after:
2744 * preempt_disable(); // 1
2746 * raw_spin_lock_irq(&rq->lock) // 2
2748 * Also, see FORK_PREEMPT_COUNT.
2750 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2751 "corrupted preempt_count: %s/%d/0x%x\n",
2752 current
->comm
, current
->pid
, preempt_count()))
2753 preempt_count_set(FORK_PREEMPT_COUNT
);
2758 * A task struct has one reference for the use as "current".
2759 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2760 * schedule one last time. The schedule call will never return, and
2761 * the scheduled task must drop that reference.
2763 * We must observe prev->state before clearing prev->on_cpu (in
2764 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2765 * running on another CPU and we could rave with its RUNNING -> DEAD
2766 * transition, resulting in a double drop.
2768 prev_state
= prev
->state
;
2769 vtime_task_switch(prev
);
2770 perf_event_task_sched_in(prev
, current
);
2771 finish_lock_switch(rq
, prev
);
2772 finish_arch_post_lock_switch();
2774 fire_sched_in_preempt_notifiers(current
);
2777 if (unlikely(prev_state
== TASK_DEAD
)) {
2778 if (prev
->sched_class
->task_dead
)
2779 prev
->sched_class
->task_dead(prev
);
2782 * Remove function-return probe instances associated with this
2783 * task and put them back on the free list.
2785 kprobe_flush_task(prev
);
2787 /* Task is done with its stack. */
2788 put_task_stack(prev
);
2790 put_task_struct(prev
);
2793 tick_nohz_task_switch();
2799 /* rq->lock is NOT held, but preemption is disabled */
2800 static void __balance_callback(struct rq
*rq
)
2802 struct callback_head
*head
, *next
;
2803 void (*func
)(struct rq
*rq
);
2804 unsigned long flags
;
2806 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2807 head
= rq
->balance_callback
;
2808 rq
->balance_callback
= NULL
;
2810 func
= (void (*)(struct rq
*))head
->func
;
2817 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2820 static inline void balance_callback(struct rq
*rq
)
2822 if (unlikely(rq
->balance_callback
))
2823 __balance_callback(rq
);
2828 static inline void balance_callback(struct rq
*rq
)
2835 * schedule_tail - first thing a freshly forked thread must call.
2836 * @prev: the thread we just switched away from.
2838 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2839 __releases(rq
->lock
)
2844 * New tasks start with FORK_PREEMPT_COUNT, see there and
2845 * finish_task_switch() for details.
2847 * finish_task_switch() will drop rq->lock() and lower preempt_count
2848 * and the preempt_enable() will end up enabling preemption (on
2849 * PREEMPT_COUNT kernels).
2852 rq
= finish_task_switch(prev
);
2853 balance_callback(rq
);
2856 if (current
->set_child_tid
)
2857 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2861 * context_switch - switch to the new MM and the new thread's register state.
2863 static __always_inline
struct rq
*
2864 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2865 struct task_struct
*next
, struct rq_flags
*rf
)
2867 struct mm_struct
*mm
, *oldmm
;
2869 prepare_task_switch(rq
, prev
, next
);
2872 oldmm
= prev
->active_mm
;
2874 * For paravirt, this is coupled with an exit in switch_to to
2875 * combine the page table reload and the switch backend into
2878 arch_start_context_switch(prev
);
2881 next
->active_mm
= oldmm
;
2882 atomic_inc(&oldmm
->mm_count
);
2883 enter_lazy_tlb(oldmm
, next
);
2885 switch_mm_irqs_off(oldmm
, mm
, next
);
2888 prev
->active_mm
= NULL
;
2889 rq
->prev_mm
= oldmm
;
2892 rq
->clock_skip_update
= 0;
2895 * Since the runqueue lock will be released by the next
2896 * task (which is an invalid locking op but in the case
2897 * of the scheduler it's an obvious special-case), so we
2898 * do an early lockdep release here:
2900 rq_unpin_lock(rq
, rf
);
2901 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2903 /* Here we just switch the register state and the stack. */
2904 switch_to(prev
, next
, prev
);
2907 return finish_task_switch(prev
);
2911 * nr_running and nr_context_switches:
2913 * externally visible scheduler statistics: current number of runnable
2914 * threads, total number of context switches performed since bootup.
2916 unsigned long nr_running(void)
2918 unsigned long i
, sum
= 0;
2920 for_each_online_cpu(i
)
2921 sum
+= cpu_rq(i
)->nr_running
;
2927 * Check if only the current task is running on the cpu.
2929 * Caution: this function does not check that the caller has disabled
2930 * preemption, thus the result might have a time-of-check-to-time-of-use
2931 * race. The caller is responsible to use it correctly, for example:
2933 * - from a non-preemptable section (of course)
2935 * - from a thread that is bound to a single CPU
2937 * - in a loop with very short iterations (e.g. a polling loop)
2939 bool single_task_running(void)
2941 return raw_rq()->nr_running
== 1;
2943 EXPORT_SYMBOL(single_task_running
);
2945 unsigned long long nr_context_switches(void)
2948 unsigned long long sum
= 0;
2950 for_each_possible_cpu(i
)
2951 sum
+= cpu_rq(i
)->nr_switches
;
2956 unsigned long nr_iowait(void)
2958 unsigned long i
, sum
= 0;
2960 for_each_possible_cpu(i
)
2961 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2966 unsigned long nr_iowait_cpu(int cpu
)
2968 struct rq
*this = cpu_rq(cpu
);
2969 return atomic_read(&this->nr_iowait
);
2972 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2974 struct rq
*rq
= this_rq();
2975 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2976 *load
= rq
->load
.weight
;
2982 * sched_exec - execve() is a valuable balancing opportunity, because at
2983 * this point the task has the smallest effective memory and cache footprint.
2985 void sched_exec(void)
2987 struct task_struct
*p
= current
;
2988 unsigned long flags
;
2991 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2992 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2993 if (dest_cpu
== smp_processor_id())
2996 if (likely(cpu_active(dest_cpu
))) {
2997 struct migration_arg arg
= { p
, dest_cpu
};
2999 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3000 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3004 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3009 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3010 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3012 EXPORT_PER_CPU_SYMBOL(kstat
);
3013 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3016 * The function fair_sched_class.update_curr accesses the struct curr
3017 * and its field curr->exec_start; when called from task_sched_runtime(),
3018 * we observe a high rate of cache misses in practice.
3019 * Prefetching this data results in improved performance.
3021 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3023 #ifdef CONFIG_FAIR_GROUP_SCHED
3024 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3026 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3029 prefetch(&curr
->exec_start
);
3033 * Return accounted runtime for the task.
3034 * In case the task is currently running, return the runtime plus current's
3035 * pending runtime that have not been accounted yet.
3037 unsigned long long task_sched_runtime(struct task_struct
*p
)
3043 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3045 * 64-bit doesn't need locks to atomically read a 64bit value.
3046 * So we have a optimization chance when the task's delta_exec is 0.
3047 * Reading ->on_cpu is racy, but this is ok.
3049 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3050 * If we race with it entering cpu, unaccounted time is 0. This is
3051 * indistinguishable from the read occurring a few cycles earlier.
3052 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3053 * been accounted, so we're correct here as well.
3055 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3056 return p
->se
.sum_exec_runtime
;
3059 rq
= task_rq_lock(p
, &rf
);
3061 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3062 * project cycles that may never be accounted to this
3063 * thread, breaking clock_gettime().
3065 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3066 prefetch_curr_exec_start(p
);
3067 update_rq_clock(rq
);
3068 p
->sched_class
->update_curr(rq
);
3070 ns
= p
->se
.sum_exec_runtime
;
3071 task_rq_unlock(rq
, p
, &rf
);
3077 * This function gets called by the timer code, with HZ frequency.
3078 * We call it with interrupts disabled.
3080 void scheduler_tick(void)
3082 int cpu
= smp_processor_id();
3083 struct rq
*rq
= cpu_rq(cpu
);
3084 struct task_struct
*curr
= rq
->curr
;
3088 raw_spin_lock(&rq
->lock
);
3089 update_rq_clock(rq
);
3090 curr
->sched_class
->task_tick(rq
, curr
, 0);
3091 cpu_load_update_active(rq
);
3092 calc_global_load_tick(rq
);
3093 raw_spin_unlock(&rq
->lock
);
3095 perf_event_task_tick();
3098 rq
->idle_balance
= idle_cpu(cpu
);
3099 trigger_load_balance(rq
);
3101 rq_last_tick_reset(rq
);
3104 #ifdef CONFIG_NO_HZ_FULL
3106 * scheduler_tick_max_deferment
3108 * Keep at least one tick per second when a single
3109 * active task is running because the scheduler doesn't
3110 * yet completely support full dynticks environment.
3112 * This makes sure that uptime, CFS vruntime, load
3113 * balancing, etc... continue to move forward, even
3114 * with a very low granularity.
3116 * Return: Maximum deferment in nanoseconds.
3118 u64
scheduler_tick_max_deferment(void)
3120 struct rq
*rq
= this_rq();
3121 unsigned long next
, now
= READ_ONCE(jiffies
);
3123 next
= rq
->last_sched_tick
+ HZ
;
3125 if (time_before_eq(next
, now
))
3128 return jiffies_to_nsecs(next
- now
);
3132 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3133 defined(CONFIG_PREEMPT_TRACER))
3135 * If the value passed in is equal to the current preempt count
3136 * then we just disabled preemption. Start timing the latency.
3138 static inline void preempt_latency_start(int val
)
3140 if (preempt_count() == val
) {
3141 unsigned long ip
= get_lock_parent_ip();
3142 #ifdef CONFIG_DEBUG_PREEMPT
3143 current
->preempt_disable_ip
= ip
;
3145 trace_preempt_off(CALLER_ADDR0
, ip
);
3149 void preempt_count_add(int val
)
3151 #ifdef CONFIG_DEBUG_PREEMPT
3155 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3158 __preempt_count_add(val
);
3159 #ifdef CONFIG_DEBUG_PREEMPT
3161 * Spinlock count overflowing soon?
3163 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3166 preempt_latency_start(val
);
3168 EXPORT_SYMBOL(preempt_count_add
);
3169 NOKPROBE_SYMBOL(preempt_count_add
);
3172 * If the value passed in equals to the current preempt count
3173 * then we just enabled preemption. Stop timing the latency.
3175 static inline void preempt_latency_stop(int val
)
3177 if (preempt_count() == val
)
3178 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3181 void preempt_count_sub(int val
)
3183 #ifdef CONFIG_DEBUG_PREEMPT
3187 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3190 * Is the spinlock portion underflowing?
3192 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3193 !(preempt_count() & PREEMPT_MASK
)))
3197 preempt_latency_stop(val
);
3198 __preempt_count_sub(val
);
3200 EXPORT_SYMBOL(preempt_count_sub
);
3201 NOKPROBE_SYMBOL(preempt_count_sub
);
3204 static inline void preempt_latency_start(int val
) { }
3205 static inline void preempt_latency_stop(int val
) { }
3209 * Print scheduling while atomic bug:
3211 static noinline
void __schedule_bug(struct task_struct
*prev
)
3213 /* Save this before calling printk(), since that will clobber it */
3214 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3216 if (oops_in_progress
)
3219 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3220 prev
->comm
, prev
->pid
, preempt_count());
3222 debug_show_held_locks(prev
);
3224 if (irqs_disabled())
3225 print_irqtrace_events(prev
);
3226 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3227 && in_atomic_preempt_off()) {
3228 pr_err("Preemption disabled at:");
3229 print_ip_sym(preempt_disable_ip
);
3233 panic("scheduling while atomic\n");
3236 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3240 * Various schedule()-time debugging checks and statistics:
3242 static inline void schedule_debug(struct task_struct
*prev
)
3244 #ifdef CONFIG_SCHED_STACK_END_CHECK
3245 if (task_stack_end_corrupted(prev
))
3246 panic("corrupted stack end detected inside scheduler\n");
3249 if (unlikely(in_atomic_preempt_off())) {
3250 __schedule_bug(prev
);
3251 preempt_count_set(PREEMPT_DISABLED
);
3255 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3257 schedstat_inc(this_rq()->sched_count
);
3261 * Pick up the highest-prio task:
3263 static inline struct task_struct
*
3264 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3266 const struct sched_class
*class = &fair_sched_class
;
3267 struct task_struct
*p
;
3270 * Optimization: we know that if all tasks are in
3271 * the fair class we can call that function directly:
3273 if (likely(prev
->sched_class
== class &&
3274 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3275 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3276 if (unlikely(p
== RETRY_TASK
))
3279 /* assumes fair_sched_class->next == idle_sched_class */
3281 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3287 for_each_class(class) {
3288 p
= class->pick_next_task(rq
, prev
, rf
);
3290 if (unlikely(p
== RETRY_TASK
))
3296 BUG(); /* the idle class will always have a runnable task */
3300 * __schedule() is the main scheduler function.
3302 * The main means of driving the scheduler and thus entering this function are:
3304 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3306 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3307 * paths. For example, see arch/x86/entry_64.S.
3309 * To drive preemption between tasks, the scheduler sets the flag in timer
3310 * interrupt handler scheduler_tick().
3312 * 3. Wakeups don't really cause entry into schedule(). They add a
3313 * task to the run-queue and that's it.
3315 * Now, if the new task added to the run-queue preempts the current
3316 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3317 * called on the nearest possible occasion:
3319 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3321 * - in syscall or exception context, at the next outmost
3322 * preempt_enable(). (this might be as soon as the wake_up()'s
3325 * - in IRQ context, return from interrupt-handler to
3326 * preemptible context
3328 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3331 * - cond_resched() call
3332 * - explicit schedule() call
3333 * - return from syscall or exception to user-space
3334 * - return from interrupt-handler to user-space
3336 * WARNING: must be called with preemption disabled!
3338 static void __sched notrace
__schedule(bool preempt
)
3340 struct task_struct
*prev
, *next
;
3341 unsigned long *switch_count
;
3346 cpu
= smp_processor_id();
3350 schedule_debug(prev
);
3352 if (sched_feat(HRTICK
))
3355 local_irq_disable();
3356 rcu_note_context_switch();
3359 * Make sure that signal_pending_state()->signal_pending() below
3360 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3361 * done by the caller to avoid the race with signal_wake_up().
3363 smp_mb__before_spinlock();
3364 raw_spin_lock(&rq
->lock
);
3365 rq_pin_lock(rq
, &rf
);
3367 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3369 switch_count
= &prev
->nivcsw
;
3370 if (!preempt
&& prev
->state
) {
3371 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3372 prev
->state
= TASK_RUNNING
;
3374 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3378 * If a worker went to sleep, notify and ask workqueue
3379 * whether it wants to wake up a task to maintain
3382 if (prev
->flags
& PF_WQ_WORKER
) {
3383 struct task_struct
*to_wakeup
;
3385 to_wakeup
= wq_worker_sleeping(prev
);
3387 try_to_wake_up_local(to_wakeup
, &rf
);
3390 switch_count
= &prev
->nvcsw
;
3393 if (task_on_rq_queued(prev
))
3394 update_rq_clock(rq
);
3396 next
= pick_next_task(rq
, prev
, &rf
);
3397 clear_tsk_need_resched(prev
);
3398 clear_preempt_need_resched();
3400 if (likely(prev
!= next
)) {
3405 trace_sched_switch(preempt
, prev
, next
);
3406 rq
= context_switch(rq
, prev
, next
, &rf
); /* unlocks the rq */
3408 rq
->clock_skip_update
= 0;
3409 rq_unpin_lock(rq
, &rf
);
3410 raw_spin_unlock_irq(&rq
->lock
);
3413 balance_callback(rq
);
3416 void __noreturn
do_task_dead(void)
3419 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3420 * when the following two conditions become true.
3421 * - There is race condition of mmap_sem (It is acquired by
3423 * - SMI occurs before setting TASK_RUNINNG.
3424 * (or hypervisor of virtual machine switches to other guest)
3425 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3427 * To avoid it, we have to wait for releasing tsk->pi_lock which
3428 * is held by try_to_wake_up()
3431 raw_spin_unlock_wait(¤t
->pi_lock
);
3433 /* causes final put_task_struct in finish_task_switch(). */
3434 __set_current_state(TASK_DEAD
);
3435 current
->flags
|= PF_NOFREEZE
; /* tell freezer to ignore us */
3438 /* Avoid "noreturn function does return". */
3440 cpu_relax(); /* For when BUG is null */
3443 static inline void sched_submit_work(struct task_struct
*tsk
)
3445 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3448 * If we are going to sleep and we have plugged IO queued,
3449 * make sure to submit it to avoid deadlocks.
3451 if (blk_needs_flush_plug(tsk
))
3452 blk_schedule_flush_plug(tsk
);
3455 asmlinkage __visible
void __sched
schedule(void)
3457 struct task_struct
*tsk
= current
;
3459 sched_submit_work(tsk
);
3463 sched_preempt_enable_no_resched();
3464 } while (need_resched());
3466 EXPORT_SYMBOL(schedule
);
3468 #ifdef CONFIG_CONTEXT_TRACKING
3469 asmlinkage __visible
void __sched
schedule_user(void)
3472 * If we come here after a random call to set_need_resched(),
3473 * or we have been woken up remotely but the IPI has not yet arrived,
3474 * we haven't yet exited the RCU idle mode. Do it here manually until
3475 * we find a better solution.
3477 * NB: There are buggy callers of this function. Ideally we
3478 * should warn if prev_state != CONTEXT_USER, but that will trigger
3479 * too frequently to make sense yet.
3481 enum ctx_state prev_state
= exception_enter();
3483 exception_exit(prev_state
);
3488 * schedule_preempt_disabled - called with preemption disabled
3490 * Returns with preemption disabled. Note: preempt_count must be 1
3492 void __sched
schedule_preempt_disabled(void)
3494 sched_preempt_enable_no_resched();
3499 static void __sched notrace
preempt_schedule_common(void)
3503 * Because the function tracer can trace preempt_count_sub()
3504 * and it also uses preempt_enable/disable_notrace(), if
3505 * NEED_RESCHED is set, the preempt_enable_notrace() called
3506 * by the function tracer will call this function again and
3507 * cause infinite recursion.
3509 * Preemption must be disabled here before the function
3510 * tracer can trace. Break up preempt_disable() into two
3511 * calls. One to disable preemption without fear of being
3512 * traced. The other to still record the preemption latency,
3513 * which can also be traced by the function tracer.
3515 preempt_disable_notrace();
3516 preempt_latency_start(1);
3518 preempt_latency_stop(1);
3519 preempt_enable_no_resched_notrace();
3522 * Check again in case we missed a preemption opportunity
3523 * between schedule and now.
3525 } while (need_resched());
3528 #ifdef CONFIG_PREEMPT
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3534 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3537 * If there is a non-zero preempt_count or interrupts are disabled,
3538 * we do not want to preempt the current task. Just return..
3540 if (likely(!preemptible()))
3543 preempt_schedule_common();
3545 NOKPROBE_SYMBOL(preempt_schedule
);
3546 EXPORT_SYMBOL(preempt_schedule
);
3549 * preempt_schedule_notrace - preempt_schedule called by tracing
3551 * The tracing infrastructure uses preempt_enable_notrace to prevent
3552 * recursion and tracing preempt enabling caused by the tracing
3553 * infrastructure itself. But as tracing can happen in areas coming
3554 * from userspace or just about to enter userspace, a preempt enable
3555 * can occur before user_exit() is called. This will cause the scheduler
3556 * to be called when the system is still in usermode.
3558 * To prevent this, the preempt_enable_notrace will use this function
3559 * instead of preempt_schedule() to exit user context if needed before
3560 * calling the scheduler.
3562 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3564 enum ctx_state prev_ctx
;
3566 if (likely(!preemptible()))
3571 * Because the function tracer can trace preempt_count_sub()
3572 * and it also uses preempt_enable/disable_notrace(), if
3573 * NEED_RESCHED is set, the preempt_enable_notrace() called
3574 * by the function tracer will call this function again and
3575 * cause infinite recursion.
3577 * Preemption must be disabled here before the function
3578 * tracer can trace. Break up preempt_disable() into two
3579 * calls. One to disable preemption without fear of being
3580 * traced. The other to still record the preemption latency,
3581 * which can also be traced by the function tracer.
3583 preempt_disable_notrace();
3584 preempt_latency_start(1);
3586 * Needs preempt disabled in case user_exit() is traced
3587 * and the tracer calls preempt_enable_notrace() causing
3588 * an infinite recursion.
3590 prev_ctx
= exception_enter();
3592 exception_exit(prev_ctx
);
3594 preempt_latency_stop(1);
3595 preempt_enable_no_resched_notrace();
3596 } while (need_resched());
3598 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3600 #endif /* CONFIG_PREEMPT */
3603 * this is the entry point to schedule() from kernel preemption
3604 * off of irq context.
3605 * Note, that this is called and return with irqs disabled. This will
3606 * protect us against recursive calling from irq.
3608 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3610 enum ctx_state prev_state
;
3612 /* Catch callers which need to be fixed */
3613 BUG_ON(preempt_count() || !irqs_disabled());
3615 prev_state
= exception_enter();
3621 local_irq_disable();
3622 sched_preempt_enable_no_resched();
3623 } while (need_resched());
3625 exception_exit(prev_state
);
3628 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3631 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3633 EXPORT_SYMBOL(default_wake_function
);
3635 #ifdef CONFIG_RT_MUTEXES
3638 * rt_mutex_setprio - set the current priority of a task
3640 * @prio: prio value (kernel-internal form)
3642 * This function changes the 'effective' priority of a task. It does
3643 * not touch ->normal_prio like __setscheduler().
3645 * Used by the rt_mutex code to implement priority inheritance
3646 * logic. Call site only calls if the priority of the task changed.
3648 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3650 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3651 const struct sched_class
*prev_class
;
3655 BUG_ON(prio
> MAX_PRIO
);
3657 rq
= __task_rq_lock(p
, &rf
);
3660 * Idle task boosting is a nono in general. There is one
3661 * exception, when PREEMPT_RT and NOHZ is active:
3663 * The idle task calls get_next_timer_interrupt() and holds
3664 * the timer wheel base->lock on the CPU and another CPU wants
3665 * to access the timer (probably to cancel it). We can safely
3666 * ignore the boosting request, as the idle CPU runs this code
3667 * with interrupts disabled and will complete the lock
3668 * protected section without being interrupted. So there is no
3669 * real need to boost.
3671 if (unlikely(p
== rq
->idle
)) {
3672 WARN_ON(p
!= rq
->curr
);
3673 WARN_ON(p
->pi_blocked_on
);
3677 trace_sched_pi_setprio(p
, prio
);
3680 if (oldprio
== prio
)
3681 queue_flag
&= ~DEQUEUE_MOVE
;
3683 prev_class
= p
->sched_class
;
3684 queued
= task_on_rq_queued(p
);
3685 running
= task_current(rq
, p
);
3687 dequeue_task(rq
, p
, queue_flag
);
3689 put_prev_task(rq
, p
);
3692 * Boosting condition are:
3693 * 1. -rt task is running and holds mutex A
3694 * --> -dl task blocks on mutex A
3696 * 2. -dl task is running and holds mutex A
3697 * --> -dl task blocks on mutex A and could preempt the
3700 if (dl_prio(prio
)) {
3701 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3702 if (!dl_prio(p
->normal_prio
) ||
3703 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3704 p
->dl
.dl_boosted
= 1;
3705 queue_flag
|= ENQUEUE_REPLENISH
;
3707 p
->dl
.dl_boosted
= 0;
3708 p
->sched_class
= &dl_sched_class
;
3709 } else if (rt_prio(prio
)) {
3710 if (dl_prio(oldprio
))
3711 p
->dl
.dl_boosted
= 0;
3713 queue_flag
|= ENQUEUE_HEAD
;
3714 p
->sched_class
= &rt_sched_class
;
3716 if (dl_prio(oldprio
))
3717 p
->dl
.dl_boosted
= 0;
3718 if (rt_prio(oldprio
))
3720 p
->sched_class
= &fair_sched_class
;
3726 enqueue_task(rq
, p
, queue_flag
);
3728 set_curr_task(rq
, p
);
3730 check_class_changed(rq
, p
, prev_class
, oldprio
);
3732 preempt_disable(); /* avoid rq from going away on us */
3733 __task_rq_unlock(rq
, &rf
);
3735 balance_callback(rq
);
3740 void set_user_nice(struct task_struct
*p
, long nice
)
3742 bool queued
, running
;
3743 int old_prio
, delta
;
3747 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3750 * We have to be careful, if called from sys_setpriority(),
3751 * the task might be in the middle of scheduling on another CPU.
3753 rq
= task_rq_lock(p
, &rf
);
3755 * The RT priorities are set via sched_setscheduler(), but we still
3756 * allow the 'normal' nice value to be set - but as expected
3757 * it wont have any effect on scheduling until the task is
3758 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3760 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3761 p
->static_prio
= NICE_TO_PRIO(nice
);
3764 queued
= task_on_rq_queued(p
);
3765 running
= task_current(rq
, p
);
3767 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3769 put_prev_task(rq
, p
);
3771 p
->static_prio
= NICE_TO_PRIO(nice
);
3774 p
->prio
= effective_prio(p
);
3775 delta
= p
->prio
- old_prio
;
3778 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3780 * If the task increased its priority or is running and
3781 * lowered its priority, then reschedule its CPU:
3783 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3787 set_curr_task(rq
, p
);
3789 task_rq_unlock(rq
, p
, &rf
);
3791 EXPORT_SYMBOL(set_user_nice
);
3794 * can_nice - check if a task can reduce its nice value
3798 int can_nice(const struct task_struct
*p
, const int nice
)
3800 /* convert nice value [19,-20] to rlimit style value [1,40] */
3801 int nice_rlim
= nice_to_rlimit(nice
);
3803 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3804 capable(CAP_SYS_NICE
));
3807 #ifdef __ARCH_WANT_SYS_NICE
3810 * sys_nice - change the priority of the current process.
3811 * @increment: priority increment
3813 * sys_setpriority is a more generic, but much slower function that
3814 * does similar things.
3816 SYSCALL_DEFINE1(nice
, int, increment
)
3821 * Setpriority might change our priority at the same moment.
3822 * We don't have to worry. Conceptually one call occurs first
3823 * and we have a single winner.
3825 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3826 nice
= task_nice(current
) + increment
;
3828 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3829 if (increment
< 0 && !can_nice(current
, nice
))
3832 retval
= security_task_setnice(current
, nice
);
3836 set_user_nice(current
, nice
);
3843 * task_prio - return the priority value of a given task.
3844 * @p: the task in question.
3846 * Return: The priority value as seen by users in /proc.
3847 * RT tasks are offset by -200. Normal tasks are centered
3848 * around 0, value goes from -16 to +15.
3850 int task_prio(const struct task_struct
*p
)
3852 return p
->prio
- MAX_RT_PRIO
;
3856 * idle_cpu - is a given cpu idle currently?
3857 * @cpu: the processor in question.
3859 * Return: 1 if the CPU is currently idle. 0 otherwise.
3861 int idle_cpu(int cpu
)
3863 struct rq
*rq
= cpu_rq(cpu
);
3865 if (rq
->curr
!= rq
->idle
)
3872 if (!llist_empty(&rq
->wake_list
))
3880 * idle_task - return the idle task for a given cpu.
3881 * @cpu: the processor in question.
3883 * Return: The idle task for the cpu @cpu.
3885 struct task_struct
*idle_task(int cpu
)
3887 return cpu_rq(cpu
)->idle
;
3891 * find_process_by_pid - find a process with a matching PID value.
3892 * @pid: the pid in question.
3894 * The task of @pid, if found. %NULL otherwise.
3896 static struct task_struct
*find_process_by_pid(pid_t pid
)
3898 return pid
? find_task_by_vpid(pid
) : current
;
3902 * This function initializes the sched_dl_entity of a newly becoming
3903 * SCHED_DEADLINE task.
3905 * Only the static values are considered here, the actual runtime and the
3906 * absolute deadline will be properly calculated when the task is enqueued
3907 * for the first time with its new policy.
3910 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3912 struct sched_dl_entity
*dl_se
= &p
->dl
;
3914 dl_se
->dl_runtime
= attr
->sched_runtime
;
3915 dl_se
->dl_deadline
= attr
->sched_deadline
;
3916 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3917 dl_se
->flags
= attr
->sched_flags
;
3918 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3921 * Changing the parameters of a task is 'tricky' and we're not doing
3922 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3924 * What we SHOULD do is delay the bandwidth release until the 0-lag
3925 * point. This would include retaining the task_struct until that time
3926 * and change dl_overflow() to not immediately decrement the current
3929 * Instead we retain the current runtime/deadline and let the new
3930 * parameters take effect after the current reservation period lapses.
3931 * This is safe (albeit pessimistic) because the 0-lag point is always
3932 * before the current scheduling deadline.
3934 * We can still have temporary overloads because we do not delay the
3935 * change in bandwidth until that time; so admission control is
3936 * not on the safe side. It does however guarantee tasks will never
3937 * consume more than promised.
3942 * sched_setparam() passes in -1 for its policy, to let the functions
3943 * it calls know not to change it.
3945 #define SETPARAM_POLICY -1
3947 static void __setscheduler_params(struct task_struct
*p
,
3948 const struct sched_attr
*attr
)
3950 int policy
= attr
->sched_policy
;
3952 if (policy
== SETPARAM_POLICY
)
3957 if (dl_policy(policy
))
3958 __setparam_dl(p
, attr
);
3959 else if (fair_policy(policy
))
3960 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3963 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3964 * !rt_policy. Always setting this ensures that things like
3965 * getparam()/getattr() don't report silly values for !rt tasks.
3967 p
->rt_priority
= attr
->sched_priority
;
3968 p
->normal_prio
= normal_prio(p
);
3972 /* Actually do priority change: must hold pi & rq lock. */
3973 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3974 const struct sched_attr
*attr
, bool keep_boost
)
3976 __setscheduler_params(p
, attr
);
3979 * Keep a potential priority boosting if called from
3980 * sched_setscheduler().
3983 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3985 p
->prio
= normal_prio(p
);
3987 if (dl_prio(p
->prio
))
3988 p
->sched_class
= &dl_sched_class
;
3989 else if (rt_prio(p
->prio
))
3990 p
->sched_class
= &rt_sched_class
;
3992 p
->sched_class
= &fair_sched_class
;
3996 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3998 struct sched_dl_entity
*dl_se
= &p
->dl
;
4000 attr
->sched_priority
= p
->rt_priority
;
4001 attr
->sched_runtime
= dl_se
->dl_runtime
;
4002 attr
->sched_deadline
= dl_se
->dl_deadline
;
4003 attr
->sched_period
= dl_se
->dl_period
;
4004 attr
->sched_flags
= dl_se
->flags
;
4008 * This function validates the new parameters of a -deadline task.
4009 * We ask for the deadline not being zero, and greater or equal
4010 * than the runtime, as well as the period of being zero or
4011 * greater than deadline. Furthermore, we have to be sure that
4012 * user parameters are above the internal resolution of 1us (we
4013 * check sched_runtime only since it is always the smaller one) and
4014 * below 2^63 ns (we have to check both sched_deadline and
4015 * sched_period, as the latter can be zero).
4018 __checkparam_dl(const struct sched_attr
*attr
)
4021 if (attr
->sched_deadline
== 0)
4025 * Since we truncate DL_SCALE bits, make sure we're at least
4028 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
4032 * Since we use the MSB for wrap-around and sign issues, make
4033 * sure it's not set (mind that period can be equal to zero).
4035 if (attr
->sched_deadline
& (1ULL << 63) ||
4036 attr
->sched_period
& (1ULL << 63))
4039 /* runtime <= deadline <= period (if period != 0) */
4040 if ((attr
->sched_period
!= 0 &&
4041 attr
->sched_period
< attr
->sched_deadline
) ||
4042 attr
->sched_deadline
< attr
->sched_runtime
)
4049 * check the target process has a UID that matches the current process's
4051 static bool check_same_owner(struct task_struct
*p
)
4053 const struct cred
*cred
= current_cred(), *pcred
;
4057 pcred
= __task_cred(p
);
4058 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4059 uid_eq(cred
->euid
, pcred
->uid
));
4064 static bool dl_param_changed(struct task_struct
*p
,
4065 const struct sched_attr
*attr
)
4067 struct sched_dl_entity
*dl_se
= &p
->dl
;
4069 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
4070 dl_se
->dl_deadline
!= attr
->sched_deadline
||
4071 dl_se
->dl_period
!= attr
->sched_period
||
4072 dl_se
->flags
!= attr
->sched_flags
)
4078 static int __sched_setscheduler(struct task_struct
*p
,
4079 const struct sched_attr
*attr
,
4082 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4083 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4084 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4085 int new_effective_prio
, policy
= attr
->sched_policy
;
4086 const struct sched_class
*prev_class
;
4089 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
4092 /* may grab non-irq protected spin_locks */
4093 BUG_ON(in_interrupt());
4095 /* double check policy once rq lock held */
4097 reset_on_fork
= p
->sched_reset_on_fork
;
4098 policy
= oldpolicy
= p
->policy
;
4100 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4102 if (!valid_policy(policy
))
4106 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4110 * Valid priorities for SCHED_FIFO and SCHED_RR are
4111 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4112 * SCHED_BATCH and SCHED_IDLE is 0.
4114 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4115 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4117 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4118 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4122 * Allow unprivileged RT tasks to decrease priority:
4124 if (user
&& !capable(CAP_SYS_NICE
)) {
4125 if (fair_policy(policy
)) {
4126 if (attr
->sched_nice
< task_nice(p
) &&
4127 !can_nice(p
, attr
->sched_nice
))
4131 if (rt_policy(policy
)) {
4132 unsigned long rlim_rtprio
=
4133 task_rlimit(p
, RLIMIT_RTPRIO
);
4135 /* can't set/change the rt policy */
4136 if (policy
!= p
->policy
&& !rlim_rtprio
)
4139 /* can't increase priority */
4140 if (attr
->sched_priority
> p
->rt_priority
&&
4141 attr
->sched_priority
> rlim_rtprio
)
4146 * Can't set/change SCHED_DEADLINE policy at all for now
4147 * (safest behavior); in the future we would like to allow
4148 * unprivileged DL tasks to increase their relative deadline
4149 * or reduce their runtime (both ways reducing utilization)
4151 if (dl_policy(policy
))
4155 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4156 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4158 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4159 if (!can_nice(p
, task_nice(p
)))
4163 /* can't change other user's priorities */
4164 if (!check_same_owner(p
))
4167 /* Normal users shall not reset the sched_reset_on_fork flag */
4168 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4173 retval
= security_task_setscheduler(p
);
4179 * make sure no PI-waiters arrive (or leave) while we are
4180 * changing the priority of the task:
4182 * To be able to change p->policy safely, the appropriate
4183 * runqueue lock must be held.
4185 rq
= task_rq_lock(p
, &rf
);
4188 * Changing the policy of the stop threads its a very bad idea
4190 if (p
== rq
->stop
) {
4191 task_rq_unlock(rq
, p
, &rf
);
4196 * If not changing anything there's no need to proceed further,
4197 * but store a possible modification of reset_on_fork.
4199 if (unlikely(policy
== p
->policy
)) {
4200 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4202 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4204 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4207 p
->sched_reset_on_fork
= reset_on_fork
;
4208 task_rq_unlock(rq
, p
, &rf
);
4214 #ifdef CONFIG_RT_GROUP_SCHED
4216 * Do not allow realtime tasks into groups that have no runtime
4219 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4220 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4221 !task_group_is_autogroup(task_group(p
))) {
4222 task_rq_unlock(rq
, p
, &rf
);
4227 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4228 cpumask_t
*span
= rq
->rd
->span
;
4231 * Don't allow tasks with an affinity mask smaller than
4232 * the entire root_domain to become SCHED_DEADLINE. We
4233 * will also fail if there's no bandwidth available.
4235 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4236 rq
->rd
->dl_bw
.bw
== 0) {
4237 task_rq_unlock(rq
, p
, &rf
);
4244 /* recheck policy now with rq lock held */
4245 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4246 policy
= oldpolicy
= -1;
4247 task_rq_unlock(rq
, p
, &rf
);
4252 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4253 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4256 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4257 task_rq_unlock(rq
, p
, &rf
);
4261 p
->sched_reset_on_fork
= reset_on_fork
;
4266 * Take priority boosted tasks into account. If the new
4267 * effective priority is unchanged, we just store the new
4268 * normal parameters and do not touch the scheduler class and
4269 * the runqueue. This will be done when the task deboost
4272 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4273 if (new_effective_prio
== oldprio
)
4274 queue_flags
&= ~DEQUEUE_MOVE
;
4277 queued
= task_on_rq_queued(p
);
4278 running
= task_current(rq
, p
);
4280 dequeue_task(rq
, p
, queue_flags
);
4282 put_prev_task(rq
, p
);
4284 prev_class
= p
->sched_class
;
4285 __setscheduler(rq
, p
, attr
, pi
);
4289 * We enqueue to tail when the priority of a task is
4290 * increased (user space view).
4292 if (oldprio
< p
->prio
)
4293 queue_flags
|= ENQUEUE_HEAD
;
4295 enqueue_task(rq
, p
, queue_flags
);
4298 set_curr_task(rq
, p
);
4300 check_class_changed(rq
, p
, prev_class
, oldprio
);
4301 preempt_disable(); /* avoid rq from going away on us */
4302 task_rq_unlock(rq
, p
, &rf
);
4305 rt_mutex_adjust_pi(p
);
4308 * Run balance callbacks after we've adjusted the PI chain.
4310 balance_callback(rq
);
4316 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4317 const struct sched_param
*param
, bool check
)
4319 struct sched_attr attr
= {
4320 .sched_policy
= policy
,
4321 .sched_priority
= param
->sched_priority
,
4322 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4325 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4326 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4327 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4328 policy
&= ~SCHED_RESET_ON_FORK
;
4329 attr
.sched_policy
= policy
;
4332 return __sched_setscheduler(p
, &attr
, check
, true);
4335 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4336 * @p: the task in question.
4337 * @policy: new policy.
4338 * @param: structure containing the new RT priority.
4340 * Return: 0 on success. An error code otherwise.
4342 * NOTE that the task may be already dead.
4344 int sched_setscheduler(struct task_struct
*p
, int policy
,
4345 const struct sched_param
*param
)
4347 return _sched_setscheduler(p
, policy
, param
, true);
4349 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4351 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4353 return __sched_setscheduler(p
, attr
, true, true);
4355 EXPORT_SYMBOL_GPL(sched_setattr
);
4358 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4359 * @p: the task in question.
4360 * @policy: new policy.
4361 * @param: structure containing the new RT priority.
4363 * Just like sched_setscheduler, only don't bother checking if the
4364 * current context has permission. For example, this is needed in
4365 * stop_machine(): we create temporary high priority worker threads,
4366 * but our caller might not have that capability.
4368 * Return: 0 on success. An error code otherwise.
4370 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4371 const struct sched_param
*param
)
4373 return _sched_setscheduler(p
, policy
, param
, false);
4375 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4378 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4380 struct sched_param lparam
;
4381 struct task_struct
*p
;
4384 if (!param
|| pid
< 0)
4386 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4391 p
= find_process_by_pid(pid
);
4393 retval
= sched_setscheduler(p
, policy
, &lparam
);
4400 * Mimics kernel/events/core.c perf_copy_attr().
4402 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4403 struct sched_attr
*attr
)
4408 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4412 * zero the full structure, so that a short copy will be nice.
4414 memset(attr
, 0, sizeof(*attr
));
4416 ret
= get_user(size
, &uattr
->size
);
4420 if (size
> PAGE_SIZE
) /* silly large */
4423 if (!size
) /* abi compat */
4424 size
= SCHED_ATTR_SIZE_VER0
;
4426 if (size
< SCHED_ATTR_SIZE_VER0
)
4430 * If we're handed a bigger struct than we know of,
4431 * ensure all the unknown bits are 0 - i.e. new
4432 * user-space does not rely on any kernel feature
4433 * extensions we dont know about yet.
4435 if (size
> sizeof(*attr
)) {
4436 unsigned char __user
*addr
;
4437 unsigned char __user
*end
;
4440 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4441 end
= (void __user
*)uattr
+ size
;
4443 for (; addr
< end
; addr
++) {
4444 ret
= get_user(val
, addr
);
4450 size
= sizeof(*attr
);
4453 ret
= copy_from_user(attr
, uattr
, size
);
4458 * XXX: do we want to be lenient like existing syscalls; or do we want
4459 * to be strict and return an error on out-of-bounds values?
4461 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4466 put_user(sizeof(*attr
), &uattr
->size
);
4471 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4472 * @pid: the pid in question.
4473 * @policy: new policy.
4474 * @param: structure containing the new RT priority.
4476 * Return: 0 on success. An error code otherwise.
4478 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4479 struct sched_param __user
*, param
)
4481 /* negative values for policy are not valid */
4485 return do_sched_setscheduler(pid
, policy
, param
);
4489 * sys_sched_setparam - set/change the RT priority of a thread
4490 * @pid: the pid in question.
4491 * @param: structure containing the new RT priority.
4493 * Return: 0 on success. An error code otherwise.
4495 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4497 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4501 * sys_sched_setattr - same as above, but with extended sched_attr
4502 * @pid: the pid in question.
4503 * @uattr: structure containing the extended parameters.
4504 * @flags: for future extension.
4506 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4507 unsigned int, flags
)
4509 struct sched_attr attr
;
4510 struct task_struct
*p
;
4513 if (!uattr
|| pid
< 0 || flags
)
4516 retval
= sched_copy_attr(uattr
, &attr
);
4520 if ((int)attr
.sched_policy
< 0)
4525 p
= find_process_by_pid(pid
);
4527 retval
= sched_setattr(p
, &attr
);
4534 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4535 * @pid: the pid in question.
4537 * Return: On success, the policy of the thread. Otherwise, a negative error
4540 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4542 struct task_struct
*p
;
4550 p
= find_process_by_pid(pid
);
4552 retval
= security_task_getscheduler(p
);
4555 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4562 * sys_sched_getparam - get the RT priority of a thread
4563 * @pid: the pid in question.
4564 * @param: structure containing the RT priority.
4566 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4569 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4571 struct sched_param lp
= { .sched_priority
= 0 };
4572 struct task_struct
*p
;
4575 if (!param
|| pid
< 0)
4579 p
= find_process_by_pid(pid
);
4584 retval
= security_task_getscheduler(p
);
4588 if (task_has_rt_policy(p
))
4589 lp
.sched_priority
= p
->rt_priority
;
4593 * This one might sleep, we cannot do it with a spinlock held ...
4595 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4604 static int sched_read_attr(struct sched_attr __user
*uattr
,
4605 struct sched_attr
*attr
,
4610 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4614 * If we're handed a smaller struct than we know of,
4615 * ensure all the unknown bits are 0 - i.e. old
4616 * user-space does not get uncomplete information.
4618 if (usize
< sizeof(*attr
)) {
4619 unsigned char *addr
;
4622 addr
= (void *)attr
+ usize
;
4623 end
= (void *)attr
+ sizeof(*attr
);
4625 for (; addr
< end
; addr
++) {
4633 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4641 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4642 * @pid: the pid in question.
4643 * @uattr: structure containing the extended parameters.
4644 * @size: sizeof(attr) for fwd/bwd comp.
4645 * @flags: for future extension.
4647 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4648 unsigned int, size
, unsigned int, flags
)
4650 struct sched_attr attr
= {
4651 .size
= sizeof(struct sched_attr
),
4653 struct task_struct
*p
;
4656 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4657 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4661 p
= find_process_by_pid(pid
);
4666 retval
= security_task_getscheduler(p
);
4670 attr
.sched_policy
= p
->policy
;
4671 if (p
->sched_reset_on_fork
)
4672 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4673 if (task_has_dl_policy(p
))
4674 __getparam_dl(p
, &attr
);
4675 else if (task_has_rt_policy(p
))
4676 attr
.sched_priority
= p
->rt_priority
;
4678 attr
.sched_nice
= task_nice(p
);
4682 retval
= sched_read_attr(uattr
, &attr
, size
);
4690 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4692 cpumask_var_t cpus_allowed
, new_mask
;
4693 struct task_struct
*p
;
4698 p
= find_process_by_pid(pid
);
4704 /* Prevent p going away */
4708 if (p
->flags
& PF_NO_SETAFFINITY
) {
4712 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4716 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4718 goto out_free_cpus_allowed
;
4721 if (!check_same_owner(p
)) {
4723 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4725 goto out_free_new_mask
;
4730 retval
= security_task_setscheduler(p
);
4732 goto out_free_new_mask
;
4735 cpuset_cpus_allowed(p
, cpus_allowed
);
4736 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4739 * Since bandwidth control happens on root_domain basis,
4740 * if admission test is enabled, we only admit -deadline
4741 * tasks allowed to run on all the CPUs in the task's
4745 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4747 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4750 goto out_free_new_mask
;
4756 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4759 cpuset_cpus_allowed(p
, cpus_allowed
);
4760 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4762 * We must have raced with a concurrent cpuset
4763 * update. Just reset the cpus_allowed to the
4764 * cpuset's cpus_allowed
4766 cpumask_copy(new_mask
, cpus_allowed
);
4771 free_cpumask_var(new_mask
);
4772 out_free_cpus_allowed
:
4773 free_cpumask_var(cpus_allowed
);
4779 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4780 struct cpumask
*new_mask
)
4782 if (len
< cpumask_size())
4783 cpumask_clear(new_mask
);
4784 else if (len
> cpumask_size())
4785 len
= cpumask_size();
4787 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4791 * sys_sched_setaffinity - set the cpu affinity of a process
4792 * @pid: pid of the process
4793 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4794 * @user_mask_ptr: user-space pointer to the new cpu mask
4796 * Return: 0 on success. An error code otherwise.
4798 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4799 unsigned long __user
*, user_mask_ptr
)
4801 cpumask_var_t new_mask
;
4804 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4807 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4809 retval
= sched_setaffinity(pid
, new_mask
);
4810 free_cpumask_var(new_mask
);
4814 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4816 struct task_struct
*p
;
4817 unsigned long flags
;
4823 p
= find_process_by_pid(pid
);
4827 retval
= security_task_getscheduler(p
);
4831 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4832 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4833 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4842 * sys_sched_getaffinity - get the cpu affinity of a process
4843 * @pid: pid of the process
4844 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4845 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4847 * Return: size of CPU mask copied to user_mask_ptr on success. An
4848 * error code otherwise.
4850 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4851 unsigned long __user
*, user_mask_ptr
)
4856 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4858 if (len
& (sizeof(unsigned long)-1))
4861 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4864 ret
= sched_getaffinity(pid
, mask
);
4866 size_t retlen
= min_t(size_t, len
, cpumask_size());
4868 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4873 free_cpumask_var(mask
);
4879 * sys_sched_yield - yield the current processor to other threads.
4881 * This function yields the current CPU to other tasks. If there are no
4882 * other threads running on this CPU then this function will return.
4886 SYSCALL_DEFINE0(sched_yield
)
4888 struct rq
*rq
= this_rq_lock();
4890 schedstat_inc(rq
->yld_count
);
4891 current
->sched_class
->yield_task(rq
);
4894 * Since we are going to call schedule() anyway, there's
4895 * no need to preempt or enable interrupts:
4897 __release(rq
->lock
);
4898 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4899 do_raw_spin_unlock(&rq
->lock
);
4900 sched_preempt_enable_no_resched();
4907 #ifndef CONFIG_PREEMPT
4908 int __sched
_cond_resched(void)
4910 if (should_resched(0)) {
4911 preempt_schedule_common();
4916 EXPORT_SYMBOL(_cond_resched
);
4920 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4921 * call schedule, and on return reacquire the lock.
4923 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4924 * operations here to prevent schedule() from being called twice (once via
4925 * spin_unlock(), once by hand).
4927 int __cond_resched_lock(spinlock_t
*lock
)
4929 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4932 lockdep_assert_held(lock
);
4934 if (spin_needbreak(lock
) || resched
) {
4937 preempt_schedule_common();
4945 EXPORT_SYMBOL(__cond_resched_lock
);
4947 int __sched
__cond_resched_softirq(void)
4949 BUG_ON(!in_softirq());
4951 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4953 preempt_schedule_common();
4959 EXPORT_SYMBOL(__cond_resched_softirq
);
4962 * yield - yield the current processor to other threads.
4964 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4966 * The scheduler is at all times free to pick the calling task as the most
4967 * eligible task to run, if removing the yield() call from your code breaks
4968 * it, its already broken.
4970 * Typical broken usage is:
4975 * where one assumes that yield() will let 'the other' process run that will
4976 * make event true. If the current task is a SCHED_FIFO task that will never
4977 * happen. Never use yield() as a progress guarantee!!
4979 * If you want to use yield() to wait for something, use wait_event().
4980 * If you want to use yield() to be 'nice' for others, use cond_resched().
4981 * If you still want to use yield(), do not!
4983 void __sched
yield(void)
4985 set_current_state(TASK_RUNNING
);
4988 EXPORT_SYMBOL(yield
);
4991 * yield_to - yield the current processor to another thread in
4992 * your thread group, or accelerate that thread toward the
4993 * processor it's on.
4995 * @preempt: whether task preemption is allowed or not
4997 * It's the caller's job to ensure that the target task struct
4998 * can't go away on us before we can do any checks.
5001 * true (>0) if we indeed boosted the target task.
5002 * false (0) if we failed to boost the target.
5003 * -ESRCH if there's no task to yield to.
5005 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5007 struct task_struct
*curr
= current
;
5008 struct rq
*rq
, *p_rq
;
5009 unsigned long flags
;
5012 local_irq_save(flags
);
5018 * If we're the only runnable task on the rq and target rq also
5019 * has only one task, there's absolutely no point in yielding.
5021 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5026 double_rq_lock(rq
, p_rq
);
5027 if (task_rq(p
) != p_rq
) {
5028 double_rq_unlock(rq
, p_rq
);
5032 if (!curr
->sched_class
->yield_to_task
)
5035 if (curr
->sched_class
!= p
->sched_class
)
5038 if (task_running(p_rq
, p
) || p
->state
)
5041 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5043 schedstat_inc(rq
->yld_count
);
5045 * Make p's CPU reschedule; pick_next_entity takes care of
5048 if (preempt
&& rq
!= p_rq
)
5053 double_rq_unlock(rq
, p_rq
);
5055 local_irq_restore(flags
);
5062 EXPORT_SYMBOL_GPL(yield_to
);
5065 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5066 * that process accounting knows that this is a task in IO wait state.
5068 long __sched
io_schedule_timeout(long timeout
)
5070 int old_iowait
= current
->in_iowait
;
5074 current
->in_iowait
= 1;
5075 blk_schedule_flush_plug(current
);
5077 delayacct_blkio_start();
5079 atomic_inc(&rq
->nr_iowait
);
5080 ret
= schedule_timeout(timeout
);
5081 current
->in_iowait
= old_iowait
;
5082 atomic_dec(&rq
->nr_iowait
);
5083 delayacct_blkio_end();
5087 EXPORT_SYMBOL(io_schedule_timeout
);
5090 * sys_sched_get_priority_max - return maximum RT priority.
5091 * @policy: scheduling class.
5093 * Return: On success, this syscall returns the maximum
5094 * rt_priority that can be used by a given scheduling class.
5095 * On failure, a negative error code is returned.
5097 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5104 ret
= MAX_USER_RT_PRIO
-1;
5106 case SCHED_DEADLINE
:
5117 * sys_sched_get_priority_min - return minimum RT priority.
5118 * @policy: scheduling class.
5120 * Return: On success, this syscall returns the minimum
5121 * rt_priority that can be used by a given scheduling class.
5122 * On failure, a negative error code is returned.
5124 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5133 case SCHED_DEADLINE
:
5143 * sys_sched_rr_get_interval - return the default timeslice of a process.
5144 * @pid: pid of the process.
5145 * @interval: userspace pointer to the timeslice value.
5147 * this syscall writes the default timeslice value of a given process
5148 * into the user-space timespec buffer. A value of '0' means infinity.
5150 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5153 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5154 struct timespec __user
*, interval
)
5156 struct task_struct
*p
;
5157 unsigned int time_slice
;
5168 p
= find_process_by_pid(pid
);
5172 retval
= security_task_getscheduler(p
);
5176 rq
= task_rq_lock(p
, &rf
);
5178 if (p
->sched_class
->get_rr_interval
)
5179 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5180 task_rq_unlock(rq
, p
, &rf
);
5183 jiffies_to_timespec(time_slice
, &t
);
5184 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5192 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5194 void sched_show_task(struct task_struct
*p
)
5196 unsigned long free
= 0;
5198 unsigned long state
= p
->state
;
5200 if (!try_get_task_stack(p
))
5203 state
= __ffs(state
) + 1;
5204 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5205 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5206 if (state
== TASK_RUNNING
)
5207 printk(KERN_CONT
" running task ");
5208 #ifdef CONFIG_DEBUG_STACK_USAGE
5209 free
= stack_not_used(p
);
5214 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5216 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5217 task_pid_nr(p
), ppid
,
5218 (unsigned long)task_thread_info(p
)->flags
);
5220 print_worker_info(KERN_INFO
, p
);
5221 show_stack(p
, NULL
);
5225 void show_state_filter(unsigned long state_filter
)
5227 struct task_struct
*g
, *p
;
5229 #if BITS_PER_LONG == 32
5231 " task PC stack pid father\n");
5234 " task PC stack pid father\n");
5237 for_each_process_thread(g
, p
) {
5239 * reset the NMI-timeout, listing all files on a slow
5240 * console might take a lot of time:
5241 * Also, reset softlockup watchdogs on all CPUs, because
5242 * another CPU might be blocked waiting for us to process
5245 touch_nmi_watchdog();
5246 touch_all_softlockup_watchdogs();
5247 if (!state_filter
|| (p
->state
& state_filter
))
5251 #ifdef CONFIG_SCHED_DEBUG
5253 sysrq_sched_debug_show();
5257 * Only show locks if all tasks are dumped:
5260 debug_show_all_locks();
5263 void init_idle_bootup_task(struct task_struct
*idle
)
5265 idle
->sched_class
= &idle_sched_class
;
5269 * init_idle - set up an idle thread for a given CPU
5270 * @idle: task in question
5271 * @cpu: cpu the idle task belongs to
5273 * NOTE: this function does not set the idle thread's NEED_RESCHED
5274 * flag, to make booting more robust.
5276 void init_idle(struct task_struct
*idle
, int cpu
)
5278 struct rq
*rq
= cpu_rq(cpu
);
5279 unsigned long flags
;
5281 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5282 raw_spin_lock(&rq
->lock
);
5284 __sched_fork(0, idle
);
5285 idle
->state
= TASK_RUNNING
;
5286 idle
->se
.exec_start
= sched_clock();
5287 idle
->flags
|= PF_IDLE
;
5289 kasan_unpoison_task_stack(idle
);
5293 * Its possible that init_idle() gets called multiple times on a task,
5294 * in that case do_set_cpus_allowed() will not do the right thing.
5296 * And since this is boot we can forgo the serialization.
5298 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5301 * We're having a chicken and egg problem, even though we are
5302 * holding rq->lock, the cpu isn't yet set to this cpu so the
5303 * lockdep check in task_group() will fail.
5305 * Similar case to sched_fork(). / Alternatively we could
5306 * use task_rq_lock() here and obtain the other rq->lock.
5311 __set_task_cpu(idle
, cpu
);
5314 rq
->curr
= rq
->idle
= idle
;
5315 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5319 raw_spin_unlock(&rq
->lock
);
5320 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5322 /* Set the preempt count _outside_ the spinlocks! */
5323 init_idle_preempt_count(idle
, cpu
);
5326 * The idle tasks have their own, simple scheduling class:
5328 idle
->sched_class
= &idle_sched_class
;
5329 ftrace_graph_init_idle_task(idle
, cpu
);
5330 vtime_init_idle(idle
, cpu
);
5332 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5336 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5337 const struct cpumask
*trial
)
5339 int ret
= 1, trial_cpus
;
5340 struct dl_bw
*cur_dl_b
;
5341 unsigned long flags
;
5343 if (!cpumask_weight(cur
))
5346 rcu_read_lock_sched();
5347 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5348 trial_cpus
= cpumask_weight(trial
);
5350 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5351 if (cur_dl_b
->bw
!= -1 &&
5352 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5354 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5355 rcu_read_unlock_sched();
5360 int task_can_attach(struct task_struct
*p
,
5361 const struct cpumask
*cs_cpus_allowed
)
5366 * Kthreads which disallow setaffinity shouldn't be moved
5367 * to a new cpuset; we don't want to change their cpu
5368 * affinity and isolating such threads by their set of
5369 * allowed nodes is unnecessary. Thus, cpusets are not
5370 * applicable for such threads. This prevents checking for
5371 * success of set_cpus_allowed_ptr() on all attached tasks
5372 * before cpus_allowed may be changed.
5374 if (p
->flags
& PF_NO_SETAFFINITY
) {
5380 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5382 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5387 unsigned long flags
;
5389 rcu_read_lock_sched();
5390 dl_b
= dl_bw_of(dest_cpu
);
5391 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5392 cpus
= dl_bw_cpus(dest_cpu
);
5393 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5398 * We reserve space for this task in the destination
5399 * root_domain, as we can't fail after this point.
5400 * We will free resources in the source root_domain
5401 * later on (see set_cpus_allowed_dl()).
5403 __dl_add(dl_b
, p
->dl
.dl_bw
);
5405 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5406 rcu_read_unlock_sched();
5416 static bool sched_smp_initialized __read_mostly
;
5418 #ifdef CONFIG_NUMA_BALANCING
5419 /* Migrate current task p to target_cpu */
5420 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5422 struct migration_arg arg
= { p
, target_cpu
};
5423 int curr_cpu
= task_cpu(p
);
5425 if (curr_cpu
== target_cpu
)
5428 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5431 /* TODO: This is not properly updating schedstats */
5433 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5434 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5438 * Requeue a task on a given node and accurately track the number of NUMA
5439 * tasks on the runqueues
5441 void sched_setnuma(struct task_struct
*p
, int nid
)
5443 bool queued
, running
;
5447 rq
= task_rq_lock(p
, &rf
);
5448 queued
= task_on_rq_queued(p
);
5449 running
= task_current(rq
, p
);
5452 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5454 put_prev_task(rq
, p
);
5456 p
->numa_preferred_nid
= nid
;
5459 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5461 set_curr_task(rq
, p
);
5462 task_rq_unlock(rq
, p
, &rf
);
5464 #endif /* CONFIG_NUMA_BALANCING */
5466 #ifdef CONFIG_HOTPLUG_CPU
5468 * Ensures that the idle task is using init_mm right before its cpu goes
5471 void idle_task_exit(void)
5473 struct mm_struct
*mm
= current
->active_mm
;
5475 BUG_ON(cpu_online(smp_processor_id()));
5477 if (mm
!= &init_mm
) {
5478 switch_mm_irqs_off(mm
, &init_mm
, current
);
5479 finish_arch_post_lock_switch();
5485 * Since this CPU is going 'away' for a while, fold any nr_active delta
5486 * we might have. Assumes we're called after migrate_tasks() so that the
5487 * nr_active count is stable. We need to take the teardown thread which
5488 * is calling this into account, so we hand in adjust = 1 to the load
5491 * Also see the comment "Global load-average calculations".
5493 static void calc_load_migrate(struct rq
*rq
)
5495 long delta
= calc_load_fold_active(rq
, 1);
5497 atomic_long_add(delta
, &calc_load_tasks
);
5500 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5504 static const struct sched_class fake_sched_class
= {
5505 .put_prev_task
= put_prev_task_fake
,
5508 static struct task_struct fake_task
= {
5510 * Avoid pull_{rt,dl}_task()
5512 .prio
= MAX_PRIO
+ 1,
5513 .sched_class
= &fake_sched_class
,
5517 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5518 * try_to_wake_up()->select_task_rq().
5520 * Called with rq->lock held even though we'er in stop_machine() and
5521 * there's no concurrency possible, we hold the required locks anyway
5522 * because of lock validation efforts.
5524 static void migrate_tasks(struct rq
*dead_rq
)
5526 struct rq
*rq
= dead_rq
;
5527 struct task_struct
*next
, *stop
= rq
->stop
;
5532 * Fudge the rq selection such that the below task selection loop
5533 * doesn't get stuck on the currently eligible stop task.
5535 * We're currently inside stop_machine() and the rq is either stuck
5536 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5537 * either way we should never end up calling schedule() until we're
5543 * put_prev_task() and pick_next_task() sched
5544 * class method both need to have an up-to-date
5545 * value of rq->clock[_task]
5547 update_rq_clock(rq
);
5551 * There's this thread running, bail when that's the only
5554 if (rq
->nr_running
== 1)
5558 * pick_next_task assumes pinned rq->lock.
5560 rq_pin_lock(rq
, &rf
);
5561 next
= pick_next_task(rq
, &fake_task
, &rf
);
5563 next
->sched_class
->put_prev_task(rq
, next
);
5566 * Rules for changing task_struct::cpus_allowed are holding
5567 * both pi_lock and rq->lock, such that holding either
5568 * stabilizes the mask.
5570 * Drop rq->lock is not quite as disastrous as it usually is
5571 * because !cpu_active at this point, which means load-balance
5572 * will not interfere. Also, stop-machine.
5574 rq_unpin_lock(rq
, &rf
);
5575 raw_spin_unlock(&rq
->lock
);
5576 raw_spin_lock(&next
->pi_lock
);
5577 raw_spin_lock(&rq
->lock
);
5580 * Since we're inside stop-machine, _nothing_ should have
5581 * changed the task, WARN if weird stuff happened, because in
5582 * that case the above rq->lock drop is a fail too.
5584 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5585 raw_spin_unlock(&next
->pi_lock
);
5589 /* Find suitable destination for @next, with force if needed. */
5590 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5592 rq
= __migrate_task(rq
, next
, dest_cpu
);
5593 if (rq
!= dead_rq
) {
5594 raw_spin_unlock(&rq
->lock
);
5596 raw_spin_lock(&rq
->lock
);
5598 raw_spin_unlock(&next
->pi_lock
);
5603 #endif /* CONFIG_HOTPLUG_CPU */
5605 static void set_rq_online(struct rq
*rq
)
5608 const struct sched_class
*class;
5610 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5613 for_each_class(class) {
5614 if (class->rq_online
)
5615 class->rq_online(rq
);
5620 static void set_rq_offline(struct rq
*rq
)
5623 const struct sched_class
*class;
5625 for_each_class(class) {
5626 if (class->rq_offline
)
5627 class->rq_offline(rq
);
5630 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5635 static void set_cpu_rq_start_time(unsigned int cpu
)
5637 struct rq
*rq
= cpu_rq(cpu
);
5639 rq
->age_stamp
= sched_clock_cpu(cpu
);
5642 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5644 #ifdef CONFIG_SCHED_DEBUG
5646 static __read_mostly
int sched_debug_enabled
;
5648 static int __init
sched_debug_setup(char *str
)
5650 sched_debug_enabled
= 1;
5654 early_param("sched_debug", sched_debug_setup
);
5656 static inline bool sched_debug(void)
5658 return sched_debug_enabled
;
5661 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5662 struct cpumask
*groupmask
)
5664 struct sched_group
*group
= sd
->groups
;
5666 cpumask_clear(groupmask
);
5668 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5670 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5671 printk("does not load-balance\n");
5673 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5678 printk(KERN_CONT
"span %*pbl level %s\n",
5679 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5681 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5682 printk(KERN_ERR
"ERROR: domain->span does not contain "
5685 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5686 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5690 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5694 printk(KERN_ERR
"ERROR: group is NULL\n");
5698 if (!cpumask_weight(sched_group_cpus(group
))) {
5699 printk(KERN_CONT
"\n");
5700 printk(KERN_ERR
"ERROR: empty group\n");
5704 if (!(sd
->flags
& SD_OVERLAP
) &&
5705 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5706 printk(KERN_CONT
"\n");
5707 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5711 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5713 printk(KERN_CONT
" %*pbl",
5714 cpumask_pr_args(sched_group_cpus(group
)));
5715 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5716 printk(KERN_CONT
" (cpu_capacity = %lu)",
5717 group
->sgc
->capacity
);
5720 group
= group
->next
;
5721 } while (group
!= sd
->groups
);
5722 printk(KERN_CONT
"\n");
5724 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5725 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5728 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5729 printk(KERN_ERR
"ERROR: parent span is not a superset "
5730 "of domain->span\n");
5734 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5738 if (!sched_debug_enabled
)
5742 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5746 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5749 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5757 #else /* !CONFIG_SCHED_DEBUG */
5759 # define sched_debug_enabled 0
5760 # define sched_domain_debug(sd, cpu) do { } while (0)
5761 static inline bool sched_debug(void)
5765 #endif /* CONFIG_SCHED_DEBUG */
5767 static int sd_degenerate(struct sched_domain
*sd
)
5769 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5772 /* Following flags need at least 2 groups */
5773 if (sd
->flags
& (SD_LOAD_BALANCE
|
5774 SD_BALANCE_NEWIDLE
|
5777 SD_SHARE_CPUCAPACITY
|
5778 SD_ASYM_CPUCAPACITY
|
5779 SD_SHARE_PKG_RESOURCES
|
5780 SD_SHARE_POWERDOMAIN
)) {
5781 if (sd
->groups
!= sd
->groups
->next
)
5785 /* Following flags don't use groups */
5786 if (sd
->flags
& (SD_WAKE_AFFINE
))
5793 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5795 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5797 if (sd_degenerate(parent
))
5800 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5803 /* Flags needing groups don't count if only 1 group in parent */
5804 if (parent
->groups
== parent
->groups
->next
) {
5805 pflags
&= ~(SD_LOAD_BALANCE
|
5806 SD_BALANCE_NEWIDLE
|
5809 SD_ASYM_CPUCAPACITY
|
5810 SD_SHARE_CPUCAPACITY
|
5811 SD_SHARE_PKG_RESOURCES
|
5813 SD_SHARE_POWERDOMAIN
);
5814 if (nr_node_ids
== 1)
5815 pflags
&= ~SD_SERIALIZE
;
5817 if (~cflags
& pflags
)
5823 static void free_rootdomain(struct rcu_head
*rcu
)
5825 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5827 cpupri_cleanup(&rd
->cpupri
);
5828 cpudl_cleanup(&rd
->cpudl
);
5829 free_cpumask_var(rd
->dlo_mask
);
5830 free_cpumask_var(rd
->rto_mask
);
5831 free_cpumask_var(rd
->online
);
5832 free_cpumask_var(rd
->span
);
5836 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5838 struct root_domain
*old_rd
= NULL
;
5839 unsigned long flags
;
5841 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5846 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5849 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5852 * If we dont want to free the old_rd yet then
5853 * set old_rd to NULL to skip the freeing later
5856 if (!atomic_dec_and_test(&old_rd
->refcount
))
5860 atomic_inc(&rd
->refcount
);
5863 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5864 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5867 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5870 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5873 static int init_rootdomain(struct root_domain
*rd
)
5875 memset(rd
, 0, sizeof(*rd
));
5877 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5879 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5881 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5883 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5886 init_dl_bw(&rd
->dl_bw
);
5887 if (cpudl_init(&rd
->cpudl
) != 0)
5890 if (cpupri_init(&rd
->cpupri
) != 0)
5895 free_cpumask_var(rd
->rto_mask
);
5897 free_cpumask_var(rd
->dlo_mask
);
5899 free_cpumask_var(rd
->online
);
5901 free_cpumask_var(rd
->span
);
5907 * By default the system creates a single root-domain with all cpus as
5908 * members (mimicking the global state we have today).
5910 struct root_domain def_root_domain
;
5912 static void init_defrootdomain(void)
5914 init_rootdomain(&def_root_domain
);
5916 atomic_set(&def_root_domain
.refcount
, 1);
5919 static struct root_domain
*alloc_rootdomain(void)
5921 struct root_domain
*rd
;
5923 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5927 if (init_rootdomain(rd
) != 0) {
5935 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5937 struct sched_group
*tmp
, *first
;
5946 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5951 } while (sg
!= first
);
5954 static void destroy_sched_domain(struct sched_domain
*sd
)
5957 * If its an overlapping domain it has private groups, iterate and
5960 if (sd
->flags
& SD_OVERLAP
) {
5961 free_sched_groups(sd
->groups
, 1);
5962 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5963 kfree(sd
->groups
->sgc
);
5966 if (sd
->shared
&& atomic_dec_and_test(&sd
->shared
->ref
))
5971 static void destroy_sched_domains_rcu(struct rcu_head
*rcu
)
5973 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5976 struct sched_domain
*parent
= sd
->parent
;
5977 destroy_sched_domain(sd
);
5982 static void destroy_sched_domains(struct sched_domain
*sd
)
5985 call_rcu(&sd
->rcu
, destroy_sched_domains_rcu
);
5989 * Keep a special pointer to the highest sched_domain that has
5990 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5991 * allows us to avoid some pointer chasing select_idle_sibling().
5993 * Also keep a unique ID per domain (we use the first cpu number in
5994 * the cpumask of the domain), this allows us to quickly tell if
5995 * two cpus are in the same cache domain, see cpus_share_cache().
5997 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5998 DEFINE_PER_CPU(int, sd_llc_size
);
5999 DEFINE_PER_CPU(int, sd_llc_id
);
6000 DEFINE_PER_CPU(struct sched_domain_shared
*, sd_llc_shared
);
6001 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
6002 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
6004 static void update_top_cache_domain(int cpu
)
6006 struct sched_domain_shared
*sds
= NULL
;
6007 struct sched_domain
*sd
;
6011 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
6013 id
= cpumask_first(sched_domain_span(sd
));
6014 size
= cpumask_weight(sched_domain_span(sd
));
6018 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
6019 per_cpu(sd_llc_size
, cpu
) = size
;
6020 per_cpu(sd_llc_id
, cpu
) = id
;
6021 rcu_assign_pointer(per_cpu(sd_llc_shared
, cpu
), sds
);
6023 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
6024 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
6026 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
6027 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
6031 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6032 * hold the hotplug lock.
6035 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6037 struct rq
*rq
= cpu_rq(cpu
);
6038 struct sched_domain
*tmp
;
6040 /* Remove the sched domains which do not contribute to scheduling. */
6041 for (tmp
= sd
; tmp
; ) {
6042 struct sched_domain
*parent
= tmp
->parent
;
6046 if (sd_parent_degenerate(tmp
, parent
)) {
6047 tmp
->parent
= parent
->parent
;
6049 parent
->parent
->child
= tmp
;
6051 * Transfer SD_PREFER_SIBLING down in case of a
6052 * degenerate parent; the spans match for this
6053 * so the property transfers.
6055 if (parent
->flags
& SD_PREFER_SIBLING
)
6056 tmp
->flags
|= SD_PREFER_SIBLING
;
6057 destroy_sched_domain(parent
);
6062 if (sd
&& sd_degenerate(sd
)) {
6065 destroy_sched_domain(tmp
);
6070 sched_domain_debug(sd
, cpu
);
6072 rq_attach_root(rq
, rd
);
6074 rcu_assign_pointer(rq
->sd
, sd
);
6075 destroy_sched_domains(tmp
);
6077 update_top_cache_domain(cpu
);
6080 /* Setup the mask of cpus configured for isolated domains */
6081 static int __init
isolated_cpu_setup(char *str
)
6085 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6086 ret
= cpulist_parse(str
, cpu_isolated_map
);
6088 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
6093 __setup("isolcpus=", isolated_cpu_setup
);
6096 struct sched_domain
** __percpu sd
;
6097 struct root_domain
*rd
;
6108 * Build an iteration mask that can exclude certain CPUs from the upwards
6111 * Asymmetric node setups can result in situations where the domain tree is of
6112 * unequal depth, make sure to skip domains that already cover the entire
6115 * In that case build_sched_domains() will have terminated the iteration early
6116 * and our sibling sd spans will be empty. Domains should always include the
6117 * cpu they're built on, so check that.
6120 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6122 const struct cpumask
*span
= sched_domain_span(sd
);
6123 struct sd_data
*sdd
= sd
->private;
6124 struct sched_domain
*sibling
;
6127 for_each_cpu(i
, span
) {
6128 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6129 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6132 cpumask_set_cpu(i
, sched_group_mask(sg
));
6137 * Return the canonical balance cpu for this group, this is the first cpu
6138 * of this group that's also in the iteration mask.
6140 int group_balance_cpu(struct sched_group
*sg
)
6142 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6146 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6148 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6149 const struct cpumask
*span
= sched_domain_span(sd
);
6150 struct cpumask
*covered
= sched_domains_tmpmask
;
6151 struct sd_data
*sdd
= sd
->private;
6152 struct sched_domain
*sibling
;
6155 cpumask_clear(covered
);
6157 for_each_cpu(i
, span
) {
6158 struct cpumask
*sg_span
;
6160 if (cpumask_test_cpu(i
, covered
))
6163 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6165 /* See the comment near build_group_mask(). */
6166 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6169 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6170 GFP_KERNEL
, cpu_to_node(cpu
));
6175 sg_span
= sched_group_cpus(sg
);
6177 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6179 cpumask_set_cpu(i
, sg_span
);
6181 cpumask_or(covered
, covered
, sg_span
);
6183 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6184 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6185 build_group_mask(sd
, sg
);
6188 * Initialize sgc->capacity such that even if we mess up the
6189 * domains and no possible iteration will get us here, we won't
6192 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6193 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
6196 * Make sure the first group of this domain contains the
6197 * canonical balance cpu. Otherwise the sched_domain iteration
6198 * breaks. See update_sg_lb_stats().
6200 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6201 group_balance_cpu(sg
) == cpu
)
6211 sd
->groups
= groups
;
6216 free_sched_groups(first
, 0);
6221 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6223 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6224 struct sched_domain
*child
= sd
->child
;
6227 cpu
= cpumask_first(sched_domain_span(child
));
6230 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6231 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6232 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6239 * build_sched_groups will build a circular linked list of the groups
6240 * covered by the given span, and will set each group's ->cpumask correctly,
6241 * and ->cpu_capacity to 0.
6243 * Assumes the sched_domain tree is fully constructed
6246 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6248 struct sched_group
*first
= NULL
, *last
= NULL
;
6249 struct sd_data
*sdd
= sd
->private;
6250 const struct cpumask
*span
= sched_domain_span(sd
);
6251 struct cpumask
*covered
;
6254 get_group(cpu
, sdd
, &sd
->groups
);
6255 atomic_inc(&sd
->groups
->ref
);
6257 if (cpu
!= cpumask_first(span
))
6260 lockdep_assert_held(&sched_domains_mutex
);
6261 covered
= sched_domains_tmpmask
;
6263 cpumask_clear(covered
);
6265 for_each_cpu(i
, span
) {
6266 struct sched_group
*sg
;
6269 if (cpumask_test_cpu(i
, covered
))
6272 group
= get_group(i
, sdd
, &sg
);
6273 cpumask_setall(sched_group_mask(sg
));
6275 for_each_cpu(j
, span
) {
6276 if (get_group(j
, sdd
, NULL
) != group
)
6279 cpumask_set_cpu(j
, covered
);
6280 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6295 * Initialize sched groups cpu_capacity.
6297 * cpu_capacity indicates the capacity of sched group, which is used while
6298 * distributing the load between different sched groups in a sched domain.
6299 * Typically cpu_capacity for all the groups in a sched domain will be same
6300 * unless there are asymmetries in the topology. If there are asymmetries,
6301 * group having more cpu_capacity will pickup more load compared to the
6302 * group having less cpu_capacity.
6304 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6306 struct sched_group
*sg
= sd
->groups
;
6311 int cpu
, max_cpu
= -1;
6313 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6315 if (!(sd
->flags
& SD_ASYM_PACKING
))
6318 for_each_cpu(cpu
, sched_group_cpus(sg
)) {
6321 else if (sched_asym_prefer(cpu
, max_cpu
))
6324 sg
->asym_prefer_cpu
= max_cpu
;
6328 } while (sg
!= sd
->groups
);
6330 if (cpu
!= group_balance_cpu(sg
))
6333 update_group_capacity(sd
, cpu
);
6337 * Initializers for schedule domains
6338 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6341 static int default_relax_domain_level
= -1;
6342 int sched_domain_level_max
;
6344 static int __init
setup_relax_domain_level(char *str
)
6346 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6347 pr_warn("Unable to set relax_domain_level\n");
6351 __setup("relax_domain_level=", setup_relax_domain_level
);
6353 static void set_domain_attribute(struct sched_domain
*sd
,
6354 struct sched_domain_attr
*attr
)
6358 if (!attr
|| attr
->relax_domain_level
< 0) {
6359 if (default_relax_domain_level
< 0)
6362 request
= default_relax_domain_level
;
6364 request
= attr
->relax_domain_level
;
6365 if (request
< sd
->level
) {
6366 /* turn off idle balance on this domain */
6367 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6369 /* turn on idle balance on this domain */
6370 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6374 static void __sdt_free(const struct cpumask
*cpu_map
);
6375 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6377 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6378 const struct cpumask
*cpu_map
)
6382 if (!atomic_read(&d
->rd
->refcount
))
6383 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6385 free_percpu(d
->sd
); /* fall through */
6387 __sdt_free(cpu_map
); /* fall through */
6393 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6394 const struct cpumask
*cpu_map
)
6396 memset(d
, 0, sizeof(*d
));
6398 if (__sdt_alloc(cpu_map
))
6399 return sa_sd_storage
;
6400 d
->sd
= alloc_percpu(struct sched_domain
*);
6402 return sa_sd_storage
;
6403 d
->rd
= alloc_rootdomain();
6406 return sa_rootdomain
;
6410 * NULL the sd_data elements we've used to build the sched_domain and
6411 * sched_group structure so that the subsequent __free_domain_allocs()
6412 * will not free the data we're using.
6414 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6416 struct sd_data
*sdd
= sd
->private;
6418 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6419 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6421 if (atomic_read(&(*per_cpu_ptr(sdd
->sds
, cpu
))->ref
))
6422 *per_cpu_ptr(sdd
->sds
, cpu
) = NULL
;
6424 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6425 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6427 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6428 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6432 static int sched_domains_numa_levels
;
6433 enum numa_topology_type sched_numa_topology_type
;
6434 static int *sched_domains_numa_distance
;
6435 int sched_max_numa_distance
;
6436 static struct cpumask
***sched_domains_numa_masks
;
6437 static int sched_domains_curr_level
;
6441 * SD_flags allowed in topology descriptions.
6443 * These flags are purely descriptive of the topology and do not prescribe
6444 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6447 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6448 * SD_SHARE_PKG_RESOURCES - describes shared caches
6449 * SD_NUMA - describes NUMA topologies
6450 * SD_SHARE_POWERDOMAIN - describes shared power domain
6451 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6453 * Odd one out, which beside describing the topology has a quirk also
6454 * prescribes the desired behaviour that goes along with it:
6456 * SD_ASYM_PACKING - describes SMT quirks
6458 #define TOPOLOGY_SD_FLAGS \
6459 (SD_SHARE_CPUCAPACITY | \
6460 SD_SHARE_PKG_RESOURCES | \
6463 SD_ASYM_CPUCAPACITY | \
6464 SD_SHARE_POWERDOMAIN)
6466 static struct sched_domain
*
6467 sd_init(struct sched_domain_topology_level
*tl
,
6468 const struct cpumask
*cpu_map
,
6469 struct sched_domain
*child
, int cpu
)
6471 struct sd_data
*sdd
= &tl
->data
;
6472 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6473 int sd_id
, sd_weight
, sd_flags
= 0;
6477 * Ugly hack to pass state to sd_numa_mask()...
6479 sched_domains_curr_level
= tl
->numa_level
;
6482 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6485 sd_flags
= (*tl
->sd_flags
)();
6486 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6487 "wrong sd_flags in topology description\n"))
6488 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6490 *sd
= (struct sched_domain
){
6491 .min_interval
= sd_weight
,
6492 .max_interval
= 2*sd_weight
,
6494 .imbalance_pct
= 125,
6496 .cache_nice_tries
= 0,
6503 .flags
= 1*SD_LOAD_BALANCE
6504 | 1*SD_BALANCE_NEWIDLE
6509 | 0*SD_SHARE_CPUCAPACITY
6510 | 0*SD_SHARE_PKG_RESOURCES
6512 | 0*SD_PREFER_SIBLING
6517 .last_balance
= jiffies
,
6518 .balance_interval
= sd_weight
,
6520 .max_newidle_lb_cost
= 0,
6521 .next_decay_max_lb_cost
= jiffies
,
6523 #ifdef CONFIG_SCHED_DEBUG
6528 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6529 sd_id
= cpumask_first(sched_domain_span(sd
));
6532 * Convert topological properties into behaviour.
6535 if (sd
->flags
& SD_ASYM_CPUCAPACITY
) {
6536 struct sched_domain
*t
= sd
;
6538 for_each_lower_domain(t
)
6539 t
->flags
|= SD_BALANCE_WAKE
;
6542 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6543 sd
->flags
|= SD_PREFER_SIBLING
;
6544 sd
->imbalance_pct
= 110;
6545 sd
->smt_gain
= 1178; /* ~15% */
6547 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6548 sd
->imbalance_pct
= 117;
6549 sd
->cache_nice_tries
= 1;
6553 } else if (sd
->flags
& SD_NUMA
) {
6554 sd
->cache_nice_tries
= 2;
6558 sd
->flags
|= SD_SERIALIZE
;
6559 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6560 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6567 sd
->flags
|= SD_PREFER_SIBLING
;
6568 sd
->cache_nice_tries
= 1;
6574 * For all levels sharing cache; connect a sched_domain_shared
6577 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6578 sd
->shared
= *per_cpu_ptr(sdd
->sds
, sd_id
);
6579 atomic_inc(&sd
->shared
->ref
);
6580 atomic_set(&sd
->shared
->nr_busy_cpus
, sd_weight
);
6589 * Topology list, bottom-up.
6591 static struct sched_domain_topology_level default_topology
[] = {
6592 #ifdef CONFIG_SCHED_SMT
6593 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6595 #ifdef CONFIG_SCHED_MC
6596 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6598 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6602 static struct sched_domain_topology_level
*sched_domain_topology
=
6605 #define for_each_sd_topology(tl) \
6606 for (tl = sched_domain_topology; tl->mask; tl++)
6608 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6610 if (WARN_ON_ONCE(sched_smp_initialized
))
6613 sched_domain_topology
= tl
;
6618 static const struct cpumask
*sd_numa_mask(int cpu
)
6620 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6623 static void sched_numa_warn(const char *str
)
6625 static int done
= false;
6633 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6635 for (i
= 0; i
< nr_node_ids
; i
++) {
6636 printk(KERN_WARNING
" ");
6637 for (j
= 0; j
< nr_node_ids
; j
++)
6638 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6639 printk(KERN_CONT
"\n");
6641 printk(KERN_WARNING
"\n");
6644 bool find_numa_distance(int distance
)
6648 if (distance
== node_distance(0, 0))
6651 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6652 if (sched_domains_numa_distance
[i
] == distance
)
6660 * A system can have three types of NUMA topology:
6661 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6662 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6663 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6665 * The difference between a glueless mesh topology and a backplane
6666 * topology lies in whether communication between not directly
6667 * connected nodes goes through intermediary nodes (where programs
6668 * could run), or through backplane controllers. This affects
6669 * placement of programs.
6671 * The type of topology can be discerned with the following tests:
6672 * - If the maximum distance between any nodes is 1 hop, the system
6673 * is directly connected.
6674 * - If for two nodes A and B, located N > 1 hops away from each other,
6675 * there is an intermediary node C, which is < N hops away from both
6676 * nodes A and B, the system is a glueless mesh.
6678 static void init_numa_topology_type(void)
6682 n
= sched_max_numa_distance
;
6684 if (sched_domains_numa_levels
<= 1) {
6685 sched_numa_topology_type
= NUMA_DIRECT
;
6689 for_each_online_node(a
) {
6690 for_each_online_node(b
) {
6691 /* Find two nodes furthest removed from each other. */
6692 if (node_distance(a
, b
) < n
)
6695 /* Is there an intermediary node between a and b? */
6696 for_each_online_node(c
) {
6697 if (node_distance(a
, c
) < n
&&
6698 node_distance(b
, c
) < n
) {
6699 sched_numa_topology_type
=
6705 sched_numa_topology_type
= NUMA_BACKPLANE
;
6711 static void sched_init_numa(void)
6713 int next_distance
, curr_distance
= node_distance(0, 0);
6714 struct sched_domain_topology_level
*tl
;
6718 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6719 if (!sched_domains_numa_distance
)
6723 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6724 * unique distances in the node_distance() table.
6726 * Assumes node_distance(0,j) includes all distances in
6727 * node_distance(i,j) in order to avoid cubic time.
6729 next_distance
= curr_distance
;
6730 for (i
= 0; i
< nr_node_ids
; i
++) {
6731 for (j
= 0; j
< nr_node_ids
; j
++) {
6732 for (k
= 0; k
< nr_node_ids
; k
++) {
6733 int distance
= node_distance(i
, k
);
6735 if (distance
> curr_distance
&&
6736 (distance
< next_distance
||
6737 next_distance
== curr_distance
))
6738 next_distance
= distance
;
6741 * While not a strong assumption it would be nice to know
6742 * about cases where if node A is connected to B, B is not
6743 * equally connected to A.
6745 if (sched_debug() && node_distance(k
, i
) != distance
)
6746 sched_numa_warn("Node-distance not symmetric");
6748 if (sched_debug() && i
&& !find_numa_distance(distance
))
6749 sched_numa_warn("Node-0 not representative");
6751 if (next_distance
!= curr_distance
) {
6752 sched_domains_numa_distance
[level
++] = next_distance
;
6753 sched_domains_numa_levels
= level
;
6754 curr_distance
= next_distance
;
6759 * In case of sched_debug() we verify the above assumption.
6769 * 'level' contains the number of unique distances, excluding the
6770 * identity distance node_distance(i,i).
6772 * The sched_domains_numa_distance[] array includes the actual distance
6777 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6778 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6779 * the array will contain less then 'level' members. This could be
6780 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6781 * in other functions.
6783 * We reset it to 'level' at the end of this function.
6785 sched_domains_numa_levels
= 0;
6787 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6788 if (!sched_domains_numa_masks
)
6792 * Now for each level, construct a mask per node which contains all
6793 * cpus of nodes that are that many hops away from us.
6795 for (i
= 0; i
< level
; i
++) {
6796 sched_domains_numa_masks
[i
] =
6797 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6798 if (!sched_domains_numa_masks
[i
])
6801 for (j
= 0; j
< nr_node_ids
; j
++) {
6802 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6806 sched_domains_numa_masks
[i
][j
] = mask
;
6809 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6812 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6817 /* Compute default topology size */
6818 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6820 tl
= kzalloc((i
+ level
+ 1) *
6821 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6826 * Copy the default topology bits..
6828 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6829 tl
[i
] = sched_domain_topology
[i
];
6832 * .. and append 'j' levels of NUMA goodness.
6834 for (j
= 0; j
< level
; i
++, j
++) {
6835 tl
[i
] = (struct sched_domain_topology_level
){
6836 .mask
= sd_numa_mask
,
6837 .sd_flags
= cpu_numa_flags
,
6838 .flags
= SDTL_OVERLAP
,
6844 sched_domain_topology
= tl
;
6846 sched_domains_numa_levels
= level
;
6847 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6849 init_numa_topology_type();
6852 static void sched_domains_numa_masks_set(unsigned int cpu
)
6854 int node
= cpu_to_node(cpu
);
6857 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6858 for (j
= 0; j
< nr_node_ids
; j
++) {
6859 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6860 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6865 static void sched_domains_numa_masks_clear(unsigned int cpu
)
6869 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6870 for (j
= 0; j
< nr_node_ids
; j
++)
6871 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6876 static inline void sched_init_numa(void) { }
6877 static void sched_domains_numa_masks_set(unsigned int cpu
) { }
6878 static void sched_domains_numa_masks_clear(unsigned int cpu
) { }
6879 #endif /* CONFIG_NUMA */
6881 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6883 struct sched_domain_topology_level
*tl
;
6886 for_each_sd_topology(tl
) {
6887 struct sd_data
*sdd
= &tl
->data
;
6889 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6893 sdd
->sds
= alloc_percpu(struct sched_domain_shared
*);
6897 sdd
->sg
= alloc_percpu(struct sched_group
*);
6901 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6905 for_each_cpu(j
, cpu_map
) {
6906 struct sched_domain
*sd
;
6907 struct sched_domain_shared
*sds
;
6908 struct sched_group
*sg
;
6909 struct sched_group_capacity
*sgc
;
6911 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6912 GFP_KERNEL
, cpu_to_node(j
));
6916 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6918 sds
= kzalloc_node(sizeof(struct sched_domain_shared
),
6919 GFP_KERNEL
, cpu_to_node(j
));
6923 *per_cpu_ptr(sdd
->sds
, j
) = sds
;
6925 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6926 GFP_KERNEL
, cpu_to_node(j
));
6932 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6934 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6935 GFP_KERNEL
, cpu_to_node(j
));
6939 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6946 static void __sdt_free(const struct cpumask
*cpu_map
)
6948 struct sched_domain_topology_level
*tl
;
6951 for_each_sd_topology(tl
) {
6952 struct sd_data
*sdd
= &tl
->data
;
6954 for_each_cpu(j
, cpu_map
) {
6955 struct sched_domain
*sd
;
6958 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6959 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6960 free_sched_groups(sd
->groups
, 0);
6961 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6965 kfree(*per_cpu_ptr(sdd
->sds
, j
));
6967 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6969 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6971 free_percpu(sdd
->sd
);
6973 free_percpu(sdd
->sds
);
6975 free_percpu(sdd
->sg
);
6977 free_percpu(sdd
->sgc
);
6982 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6983 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6984 struct sched_domain
*child
, int cpu
)
6986 struct sched_domain
*sd
= sd_init(tl
, cpu_map
, child
, cpu
);
6989 sd
->level
= child
->level
+ 1;
6990 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6993 if (!cpumask_subset(sched_domain_span(child
),
6994 sched_domain_span(sd
))) {
6995 pr_err("BUG: arch topology borken\n");
6996 #ifdef CONFIG_SCHED_DEBUG
6997 pr_err(" the %s domain not a subset of the %s domain\n",
6998 child
->name
, sd
->name
);
7000 /* Fixup, ensure @sd has at least @child cpus. */
7001 cpumask_or(sched_domain_span(sd
),
7002 sched_domain_span(sd
),
7003 sched_domain_span(child
));
7007 set_domain_attribute(sd
, attr
);
7013 * Build sched domains for a given set of cpus and attach the sched domains
7014 * to the individual cpus
7016 static int build_sched_domains(const struct cpumask
*cpu_map
,
7017 struct sched_domain_attr
*attr
)
7019 enum s_alloc alloc_state
;
7020 struct sched_domain
*sd
;
7022 struct rq
*rq
= NULL
;
7023 int i
, ret
= -ENOMEM
;
7025 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7026 if (alloc_state
!= sa_rootdomain
)
7029 /* Set up domains for cpus specified by the cpu_map. */
7030 for_each_cpu(i
, cpu_map
) {
7031 struct sched_domain_topology_level
*tl
;
7034 for_each_sd_topology(tl
) {
7035 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
7036 if (tl
== sched_domain_topology
)
7037 *per_cpu_ptr(d
.sd
, i
) = sd
;
7038 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
7039 sd
->flags
|= SD_OVERLAP
;
7040 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
7045 /* Build the groups for the domains */
7046 for_each_cpu(i
, cpu_map
) {
7047 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7048 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
7049 if (sd
->flags
& SD_OVERLAP
) {
7050 if (build_overlap_sched_groups(sd
, i
))
7053 if (build_sched_groups(sd
, i
))
7059 /* Calculate CPU capacity for physical packages and nodes */
7060 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
7061 if (!cpumask_test_cpu(i
, cpu_map
))
7064 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7065 claim_allocations(i
, sd
);
7066 init_sched_groups_capacity(i
, sd
);
7070 /* Attach the domains */
7072 for_each_cpu(i
, cpu_map
) {
7074 sd
= *per_cpu_ptr(d
.sd
, i
);
7076 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7077 if (rq
->cpu_capacity_orig
> READ_ONCE(d
.rd
->max_cpu_capacity
))
7078 WRITE_ONCE(d
.rd
->max_cpu_capacity
, rq
->cpu_capacity_orig
);
7080 cpu_attach_domain(sd
, d
.rd
, i
);
7084 if (rq
&& sched_debug_enabled
) {
7085 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7086 cpumask_pr_args(cpu_map
), rq
->rd
->max_cpu_capacity
);
7091 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7095 static cpumask_var_t
*doms_cur
; /* current sched domains */
7096 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7097 static struct sched_domain_attr
*dattr_cur
;
7098 /* attribues of custom domains in 'doms_cur' */
7101 * Special case: If a kmalloc of a doms_cur partition (array of
7102 * cpumask) fails, then fallback to a single sched domain,
7103 * as determined by the single cpumask fallback_doms.
7105 static cpumask_var_t fallback_doms
;
7108 * arch_update_cpu_topology lets virtualized architectures update the
7109 * cpu core maps. It is supposed to return 1 if the topology changed
7110 * or 0 if it stayed the same.
7112 int __weak
arch_update_cpu_topology(void)
7117 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7120 cpumask_var_t
*doms
;
7122 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7125 for (i
= 0; i
< ndoms
; i
++) {
7126 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7127 free_sched_domains(doms
, i
);
7134 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7137 for (i
= 0; i
< ndoms
; i
++)
7138 free_cpumask_var(doms
[i
]);
7143 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7144 * For now this just excludes isolated cpus, but could be used to
7145 * exclude other special cases in the future.
7147 static int init_sched_domains(const struct cpumask
*cpu_map
)
7151 arch_update_cpu_topology();
7153 doms_cur
= alloc_sched_domains(ndoms_cur
);
7155 doms_cur
= &fallback_doms
;
7156 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7157 err
= build_sched_domains(doms_cur
[0], NULL
);
7158 register_sched_domain_sysctl();
7164 * Detach sched domains from a group of cpus specified in cpu_map
7165 * These cpus will now be attached to the NULL domain
7167 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7172 for_each_cpu(i
, cpu_map
)
7173 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7177 /* handle null as "default" */
7178 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7179 struct sched_domain_attr
*new, int idx_new
)
7181 struct sched_domain_attr tmp
;
7188 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7189 new ? (new + idx_new
) : &tmp
,
7190 sizeof(struct sched_domain_attr
));
7194 * Partition sched domains as specified by the 'ndoms_new'
7195 * cpumasks in the array doms_new[] of cpumasks. This compares
7196 * doms_new[] to the current sched domain partitioning, doms_cur[].
7197 * It destroys each deleted domain and builds each new domain.
7199 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7200 * The masks don't intersect (don't overlap.) We should setup one
7201 * sched domain for each mask. CPUs not in any of the cpumasks will
7202 * not be load balanced. If the same cpumask appears both in the
7203 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7206 * The passed in 'doms_new' should be allocated using
7207 * alloc_sched_domains. This routine takes ownership of it and will
7208 * free_sched_domains it when done with it. If the caller failed the
7209 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7210 * and partition_sched_domains() will fallback to the single partition
7211 * 'fallback_doms', it also forces the domains to be rebuilt.
7213 * If doms_new == NULL it will be replaced with cpu_online_mask.
7214 * ndoms_new == 0 is a special case for destroying existing domains,
7215 * and it will not create the default domain.
7217 * Call with hotplug lock held
7219 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7220 struct sched_domain_attr
*dattr_new
)
7225 mutex_lock(&sched_domains_mutex
);
7227 /* always unregister in case we don't destroy any domains */
7228 unregister_sched_domain_sysctl();
7230 /* Let architecture update cpu core mappings. */
7231 new_topology
= arch_update_cpu_topology();
7233 n
= doms_new
? ndoms_new
: 0;
7235 /* Destroy deleted domains */
7236 for (i
= 0; i
< ndoms_cur
; i
++) {
7237 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7238 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7239 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7242 /* no match - a current sched domain not in new doms_new[] */
7243 detach_destroy_domains(doms_cur
[i
]);
7249 if (doms_new
== NULL
) {
7251 doms_new
= &fallback_doms
;
7252 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7253 WARN_ON_ONCE(dattr_new
);
7256 /* Build new domains */
7257 for (i
= 0; i
< ndoms_new
; i
++) {
7258 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7259 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7260 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7263 /* no match - add a new doms_new */
7264 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7269 /* Remember the new sched domains */
7270 if (doms_cur
!= &fallback_doms
)
7271 free_sched_domains(doms_cur
, ndoms_cur
);
7272 kfree(dattr_cur
); /* kfree(NULL) is safe */
7273 doms_cur
= doms_new
;
7274 dattr_cur
= dattr_new
;
7275 ndoms_cur
= ndoms_new
;
7277 register_sched_domain_sysctl();
7279 mutex_unlock(&sched_domains_mutex
);
7282 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7285 * Update cpusets according to cpu_active mask. If cpusets are
7286 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7287 * around partition_sched_domains().
7289 * If we come here as part of a suspend/resume, don't touch cpusets because we
7290 * want to restore it back to its original state upon resume anyway.
7292 static void cpuset_cpu_active(void)
7294 if (cpuhp_tasks_frozen
) {
7296 * num_cpus_frozen tracks how many CPUs are involved in suspend
7297 * resume sequence. As long as this is not the last online
7298 * operation in the resume sequence, just build a single sched
7299 * domain, ignoring cpusets.
7302 if (likely(num_cpus_frozen
)) {
7303 partition_sched_domains(1, NULL
, NULL
);
7307 * This is the last CPU online operation. So fall through and
7308 * restore the original sched domains by considering the
7309 * cpuset configurations.
7312 cpuset_update_active_cpus(true);
7315 static int cpuset_cpu_inactive(unsigned int cpu
)
7317 unsigned long flags
;
7322 if (!cpuhp_tasks_frozen
) {
7323 rcu_read_lock_sched();
7324 dl_b
= dl_bw_of(cpu
);
7326 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7327 cpus
= dl_bw_cpus(cpu
);
7328 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7329 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7331 rcu_read_unlock_sched();
7335 cpuset_update_active_cpus(false);
7338 partition_sched_domains(1, NULL
, NULL
);
7343 int sched_cpu_activate(unsigned int cpu
)
7345 struct rq
*rq
= cpu_rq(cpu
);
7346 unsigned long flags
;
7348 set_cpu_active(cpu
, true);
7350 if (sched_smp_initialized
) {
7351 sched_domains_numa_masks_set(cpu
);
7352 cpuset_cpu_active();
7356 * Put the rq online, if not already. This happens:
7358 * 1) In the early boot process, because we build the real domains
7359 * after all cpus have been brought up.
7361 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7364 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7366 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7369 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7371 update_max_interval();
7376 int sched_cpu_deactivate(unsigned int cpu
)
7380 set_cpu_active(cpu
, false);
7382 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7383 * users of this state to go away such that all new such users will
7386 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7387 * not imply sync_sched(), so wait for both.
7389 * Do sync before park smpboot threads to take care the rcu boost case.
7391 if (IS_ENABLED(CONFIG_PREEMPT
))
7392 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
7396 if (!sched_smp_initialized
)
7399 ret
= cpuset_cpu_inactive(cpu
);
7401 set_cpu_active(cpu
, true);
7404 sched_domains_numa_masks_clear(cpu
);
7408 static void sched_rq_cpu_starting(unsigned int cpu
)
7410 struct rq
*rq
= cpu_rq(cpu
);
7412 rq
->calc_load_update
= calc_load_update
;
7413 update_max_interval();
7416 int sched_cpu_starting(unsigned int cpu
)
7418 set_cpu_rq_start_time(cpu
);
7419 sched_rq_cpu_starting(cpu
);
7423 #ifdef CONFIG_HOTPLUG_CPU
7424 int sched_cpu_dying(unsigned int cpu
)
7426 struct rq
*rq
= cpu_rq(cpu
);
7427 unsigned long flags
;
7429 /* Handle pending wakeups and then migrate everything off */
7430 sched_ttwu_pending();
7431 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7433 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7437 BUG_ON(rq
->nr_running
!= 1);
7438 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7439 calc_load_migrate(rq
);
7440 update_max_interval();
7441 nohz_balance_exit_idle(cpu
);
7447 #ifdef CONFIG_SCHED_SMT
7448 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
7450 static void sched_init_smt(void)
7453 * We've enumerated all CPUs and will assume that if any CPU
7454 * has SMT siblings, CPU0 will too.
7456 if (cpumask_weight(cpu_smt_mask(0)) > 1)
7457 static_branch_enable(&sched_smt_present
);
7460 static inline void sched_init_smt(void) { }
7463 void __init
sched_init_smp(void)
7465 cpumask_var_t non_isolated_cpus
;
7467 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7468 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7473 * There's no userspace yet to cause hotplug operations; hence all the
7474 * cpu masks are stable and all blatant races in the below code cannot
7477 mutex_lock(&sched_domains_mutex
);
7478 init_sched_domains(cpu_active_mask
);
7479 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7480 if (cpumask_empty(non_isolated_cpus
))
7481 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7482 mutex_unlock(&sched_domains_mutex
);
7484 /* Move init over to a non-isolated CPU */
7485 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7487 sched_init_granularity();
7488 free_cpumask_var(non_isolated_cpus
);
7490 init_sched_rt_class();
7491 init_sched_dl_class();
7495 sched_smp_initialized
= true;
7498 static int __init
migration_init(void)
7500 sched_rq_cpu_starting(smp_processor_id());
7503 early_initcall(migration_init
);
7506 void __init
sched_init_smp(void)
7508 sched_init_granularity();
7510 #endif /* CONFIG_SMP */
7512 int in_sched_functions(unsigned long addr
)
7514 return in_lock_functions(addr
) ||
7515 (addr
>= (unsigned long)__sched_text_start
7516 && addr
< (unsigned long)__sched_text_end
);
7519 #ifdef CONFIG_CGROUP_SCHED
7521 * Default task group.
7522 * Every task in system belongs to this group at bootup.
7524 struct task_group root_task_group
;
7525 LIST_HEAD(task_groups
);
7527 /* Cacheline aligned slab cache for task_group */
7528 static struct kmem_cache
*task_group_cache __read_mostly
;
7531 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7532 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
7534 #define WAIT_TABLE_BITS 8
7535 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7536 static wait_queue_head_t bit_wait_table
[WAIT_TABLE_SIZE
] __cacheline_aligned
;
7538 wait_queue_head_t
*bit_waitqueue(void *word
, int bit
)
7540 const int shift
= BITS_PER_LONG
== 32 ? 5 : 6;
7541 unsigned long val
= (unsigned long)word
<< shift
| bit
;
7543 return bit_wait_table
+ hash_long(val
, WAIT_TABLE_BITS
);
7545 EXPORT_SYMBOL(bit_waitqueue
);
7547 void __init
sched_init(void)
7550 unsigned long alloc_size
= 0, ptr
;
7552 for (i
= 0; i
< WAIT_TABLE_SIZE
; i
++)
7553 init_waitqueue_head(bit_wait_table
+ i
);
7555 #ifdef CONFIG_FAIR_GROUP_SCHED
7556 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7558 #ifdef CONFIG_RT_GROUP_SCHED
7559 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7562 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7564 #ifdef CONFIG_FAIR_GROUP_SCHED
7565 root_task_group
.se
= (struct sched_entity
**)ptr
;
7566 ptr
+= nr_cpu_ids
* sizeof(void **);
7568 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7569 ptr
+= nr_cpu_ids
* sizeof(void **);
7571 #endif /* CONFIG_FAIR_GROUP_SCHED */
7572 #ifdef CONFIG_RT_GROUP_SCHED
7573 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7574 ptr
+= nr_cpu_ids
* sizeof(void **);
7576 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7577 ptr
+= nr_cpu_ids
* sizeof(void **);
7579 #endif /* CONFIG_RT_GROUP_SCHED */
7581 #ifdef CONFIG_CPUMASK_OFFSTACK
7582 for_each_possible_cpu(i
) {
7583 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7584 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7585 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7586 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7588 #endif /* CONFIG_CPUMASK_OFFSTACK */
7590 init_rt_bandwidth(&def_rt_bandwidth
,
7591 global_rt_period(), global_rt_runtime());
7592 init_dl_bandwidth(&def_dl_bandwidth
,
7593 global_rt_period(), global_rt_runtime());
7596 init_defrootdomain();
7599 #ifdef CONFIG_RT_GROUP_SCHED
7600 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7601 global_rt_period(), global_rt_runtime());
7602 #endif /* CONFIG_RT_GROUP_SCHED */
7604 #ifdef CONFIG_CGROUP_SCHED
7605 task_group_cache
= KMEM_CACHE(task_group
, 0);
7607 list_add(&root_task_group
.list
, &task_groups
);
7608 INIT_LIST_HEAD(&root_task_group
.children
);
7609 INIT_LIST_HEAD(&root_task_group
.siblings
);
7610 autogroup_init(&init_task
);
7611 #endif /* CONFIG_CGROUP_SCHED */
7613 for_each_possible_cpu(i
) {
7617 raw_spin_lock_init(&rq
->lock
);
7619 rq
->calc_load_active
= 0;
7620 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7621 init_cfs_rq(&rq
->cfs
);
7622 init_rt_rq(&rq
->rt
);
7623 init_dl_rq(&rq
->dl
);
7624 #ifdef CONFIG_FAIR_GROUP_SCHED
7625 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7626 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7627 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
7629 * How much cpu bandwidth does root_task_group get?
7631 * In case of task-groups formed thr' the cgroup filesystem, it
7632 * gets 100% of the cpu resources in the system. This overall
7633 * system cpu resource is divided among the tasks of
7634 * root_task_group and its child task-groups in a fair manner,
7635 * based on each entity's (task or task-group's) weight
7636 * (se->load.weight).
7638 * In other words, if root_task_group has 10 tasks of weight
7639 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7640 * then A0's share of the cpu resource is:
7642 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7644 * We achieve this by letting root_task_group's tasks sit
7645 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7647 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7648 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7649 #endif /* CONFIG_FAIR_GROUP_SCHED */
7651 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7652 #ifdef CONFIG_RT_GROUP_SCHED
7653 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7656 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7657 rq
->cpu_load
[j
] = 0;
7662 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7663 rq
->balance_callback
= NULL
;
7664 rq
->active_balance
= 0;
7665 rq
->next_balance
= jiffies
;
7670 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7671 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7673 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7675 rq_attach_root(rq
, &def_root_domain
);
7676 #ifdef CONFIG_NO_HZ_COMMON
7677 rq
->last_load_update_tick
= jiffies
;
7680 #ifdef CONFIG_NO_HZ_FULL
7681 rq
->last_sched_tick
= 0;
7683 #endif /* CONFIG_SMP */
7685 atomic_set(&rq
->nr_iowait
, 0);
7688 set_load_weight(&init_task
);
7691 * The boot idle thread does lazy MMU switching as well:
7693 atomic_inc(&init_mm
.mm_count
);
7694 enter_lazy_tlb(&init_mm
, current
);
7697 * Make us the idle thread. Technically, schedule() should not be
7698 * called from this thread, however somewhere below it might be,
7699 * but because we are the idle thread, we just pick up running again
7700 * when this runqueue becomes "idle".
7702 init_idle(current
, smp_processor_id());
7704 calc_load_update
= jiffies
+ LOAD_FREQ
;
7707 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7708 /* May be allocated at isolcpus cmdline parse time */
7709 if (cpu_isolated_map
== NULL
)
7710 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7711 idle_thread_set_boot_cpu();
7712 set_cpu_rq_start_time(smp_processor_id());
7714 init_sched_fair_class();
7718 scheduler_running
= 1;
7721 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7722 static inline int preempt_count_equals(int preempt_offset
)
7724 int nested
= preempt_count() + rcu_preempt_depth();
7726 return (nested
== preempt_offset
);
7729 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7732 * Blocking primitives will set (and therefore destroy) current->state,
7733 * since we will exit with TASK_RUNNING make sure we enter with it,
7734 * otherwise we will destroy state.
7736 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7737 "do not call blocking ops when !TASK_RUNNING; "
7738 "state=%lx set at [<%p>] %pS\n",
7740 (void *)current
->task_state_change
,
7741 (void *)current
->task_state_change
);
7743 ___might_sleep(file
, line
, preempt_offset
);
7745 EXPORT_SYMBOL(__might_sleep
);
7747 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7749 static unsigned long prev_jiffy
; /* ratelimiting */
7750 unsigned long preempt_disable_ip
;
7752 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7753 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7754 !is_idle_task(current
)) ||
7755 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7757 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7759 prev_jiffy
= jiffies
;
7761 /* Save this before calling printk(), since that will clobber it */
7762 preempt_disable_ip
= get_preempt_disable_ip(current
);
7765 "BUG: sleeping function called from invalid context at %s:%d\n",
7768 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7769 in_atomic(), irqs_disabled(),
7770 current
->pid
, current
->comm
);
7772 if (task_stack_end_corrupted(current
))
7773 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7775 debug_show_held_locks(current
);
7776 if (irqs_disabled())
7777 print_irqtrace_events(current
);
7778 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
7779 && !preempt_count_equals(preempt_offset
)) {
7780 pr_err("Preemption disabled at:");
7781 print_ip_sym(preempt_disable_ip
);
7785 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7787 EXPORT_SYMBOL(___might_sleep
);
7790 #ifdef CONFIG_MAGIC_SYSRQ
7791 void normalize_rt_tasks(void)
7793 struct task_struct
*g
, *p
;
7794 struct sched_attr attr
= {
7795 .sched_policy
= SCHED_NORMAL
,
7798 read_lock(&tasklist_lock
);
7799 for_each_process_thread(g
, p
) {
7801 * Only normalize user tasks:
7803 if (p
->flags
& PF_KTHREAD
)
7806 p
->se
.exec_start
= 0;
7807 schedstat_set(p
->se
.statistics
.wait_start
, 0);
7808 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
7809 schedstat_set(p
->se
.statistics
.block_start
, 0);
7811 if (!dl_task(p
) && !rt_task(p
)) {
7813 * Renice negative nice level userspace
7816 if (task_nice(p
) < 0)
7817 set_user_nice(p
, 0);
7821 __sched_setscheduler(p
, &attr
, false, false);
7823 read_unlock(&tasklist_lock
);
7826 #endif /* CONFIG_MAGIC_SYSRQ */
7828 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7830 * These functions are only useful for the IA64 MCA handling, or kdb.
7832 * They can only be called when the whole system has been
7833 * stopped - every CPU needs to be quiescent, and no scheduling
7834 * activity can take place. Using them for anything else would
7835 * be a serious bug, and as a result, they aren't even visible
7836 * under any other configuration.
7840 * curr_task - return the current task for a given cpu.
7841 * @cpu: the processor in question.
7843 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7845 * Return: The current task for @cpu.
7847 struct task_struct
*curr_task(int cpu
)
7849 return cpu_curr(cpu
);
7852 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7856 * set_curr_task - set the current task for a given cpu.
7857 * @cpu: the processor in question.
7858 * @p: the task pointer to set.
7860 * Description: This function must only be used when non-maskable interrupts
7861 * are serviced on a separate stack. It allows the architecture to switch the
7862 * notion of the current task on a cpu in a non-blocking manner. This function
7863 * must be called with all CPU's synchronized, and interrupts disabled, the
7864 * and caller must save the original value of the current task (see
7865 * curr_task() above) and restore that value before reenabling interrupts and
7866 * re-starting the system.
7868 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7870 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
7877 #ifdef CONFIG_CGROUP_SCHED
7878 /* task_group_lock serializes the addition/removal of task groups */
7879 static DEFINE_SPINLOCK(task_group_lock
);
7881 static void sched_free_group(struct task_group
*tg
)
7883 free_fair_sched_group(tg
);
7884 free_rt_sched_group(tg
);
7886 kmem_cache_free(task_group_cache
, tg
);
7889 /* allocate runqueue etc for a new task group */
7890 struct task_group
*sched_create_group(struct task_group
*parent
)
7892 struct task_group
*tg
;
7894 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7896 return ERR_PTR(-ENOMEM
);
7898 if (!alloc_fair_sched_group(tg
, parent
))
7901 if (!alloc_rt_sched_group(tg
, parent
))
7907 sched_free_group(tg
);
7908 return ERR_PTR(-ENOMEM
);
7911 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7913 unsigned long flags
;
7915 spin_lock_irqsave(&task_group_lock
, flags
);
7916 list_add_rcu(&tg
->list
, &task_groups
);
7918 WARN_ON(!parent
); /* root should already exist */
7920 tg
->parent
= parent
;
7921 INIT_LIST_HEAD(&tg
->children
);
7922 list_add_rcu(&tg
->siblings
, &parent
->children
);
7923 spin_unlock_irqrestore(&task_group_lock
, flags
);
7925 online_fair_sched_group(tg
);
7928 /* rcu callback to free various structures associated with a task group */
7929 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7931 /* now it should be safe to free those cfs_rqs */
7932 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7935 void sched_destroy_group(struct task_group
*tg
)
7937 /* wait for possible concurrent references to cfs_rqs complete */
7938 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7941 void sched_offline_group(struct task_group
*tg
)
7943 unsigned long flags
;
7945 /* end participation in shares distribution */
7946 unregister_fair_sched_group(tg
);
7948 spin_lock_irqsave(&task_group_lock
, flags
);
7949 list_del_rcu(&tg
->list
);
7950 list_del_rcu(&tg
->siblings
);
7951 spin_unlock_irqrestore(&task_group_lock
, flags
);
7954 static void sched_change_group(struct task_struct
*tsk
, int type
)
7956 struct task_group
*tg
;
7959 * All callers are synchronized by task_rq_lock(); we do not use RCU
7960 * which is pointless here. Thus, we pass "true" to task_css_check()
7961 * to prevent lockdep warnings.
7963 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7964 struct task_group
, css
);
7965 tg
= autogroup_task_group(tsk
, tg
);
7966 tsk
->sched_task_group
= tg
;
7968 #ifdef CONFIG_FAIR_GROUP_SCHED
7969 if (tsk
->sched_class
->task_change_group
)
7970 tsk
->sched_class
->task_change_group(tsk
, type
);
7973 set_task_rq(tsk
, task_cpu(tsk
));
7977 * Change task's runqueue when it moves between groups.
7979 * The caller of this function should have put the task in its new group by
7980 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7983 void sched_move_task(struct task_struct
*tsk
)
7985 int queued
, running
;
7989 rq
= task_rq_lock(tsk
, &rf
);
7991 running
= task_current(rq
, tsk
);
7992 queued
= task_on_rq_queued(tsk
);
7995 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7996 if (unlikely(running
))
7997 put_prev_task(rq
, tsk
);
7999 sched_change_group(tsk
, TASK_MOVE_GROUP
);
8002 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
8003 if (unlikely(running
))
8004 set_curr_task(rq
, tsk
);
8006 task_rq_unlock(rq
, tsk
, &rf
);
8008 #endif /* CONFIG_CGROUP_SCHED */
8010 #ifdef CONFIG_RT_GROUP_SCHED
8012 * Ensure that the real time constraints are schedulable.
8014 static DEFINE_MUTEX(rt_constraints_mutex
);
8016 /* Must be called with tasklist_lock held */
8017 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8019 struct task_struct
*g
, *p
;
8022 * Autogroups do not have RT tasks; see autogroup_create().
8024 if (task_group_is_autogroup(tg
))
8027 for_each_process_thread(g
, p
) {
8028 if (rt_task(p
) && task_group(p
) == tg
)
8035 struct rt_schedulable_data
{
8036 struct task_group
*tg
;
8041 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
8043 struct rt_schedulable_data
*d
= data
;
8044 struct task_group
*child
;
8045 unsigned long total
, sum
= 0;
8046 u64 period
, runtime
;
8048 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8049 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8052 period
= d
->rt_period
;
8053 runtime
= d
->rt_runtime
;
8057 * Cannot have more runtime than the period.
8059 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8063 * Ensure we don't starve existing RT tasks.
8065 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8068 total
= to_ratio(period
, runtime
);
8071 * Nobody can have more than the global setting allows.
8073 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8077 * The sum of our children's runtime should not exceed our own.
8079 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8080 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8081 runtime
= child
->rt_bandwidth
.rt_runtime
;
8083 if (child
== d
->tg
) {
8084 period
= d
->rt_period
;
8085 runtime
= d
->rt_runtime
;
8088 sum
+= to_ratio(period
, runtime
);
8097 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8101 struct rt_schedulable_data data
= {
8103 .rt_period
= period
,
8104 .rt_runtime
= runtime
,
8108 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
8114 static int tg_set_rt_bandwidth(struct task_group
*tg
,
8115 u64 rt_period
, u64 rt_runtime
)
8120 * Disallowing the root group RT runtime is BAD, it would disallow the
8121 * kernel creating (and or operating) RT threads.
8123 if (tg
== &root_task_group
&& rt_runtime
== 0)
8126 /* No period doesn't make any sense. */
8130 mutex_lock(&rt_constraints_mutex
);
8131 read_lock(&tasklist_lock
);
8132 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8136 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8137 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8138 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8140 for_each_possible_cpu(i
) {
8141 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8143 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8144 rt_rq
->rt_runtime
= rt_runtime
;
8145 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8147 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8149 read_unlock(&tasklist_lock
);
8150 mutex_unlock(&rt_constraints_mutex
);
8155 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8157 u64 rt_runtime
, rt_period
;
8159 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8160 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8161 if (rt_runtime_us
< 0)
8162 rt_runtime
= RUNTIME_INF
;
8164 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8167 static long sched_group_rt_runtime(struct task_group
*tg
)
8171 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8174 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8175 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8176 return rt_runtime_us
;
8179 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
8181 u64 rt_runtime
, rt_period
;
8183 rt_period
= rt_period_us
* NSEC_PER_USEC
;
8184 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8186 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8189 static long sched_group_rt_period(struct task_group
*tg
)
8193 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8194 do_div(rt_period_us
, NSEC_PER_USEC
);
8195 return rt_period_us
;
8197 #endif /* CONFIG_RT_GROUP_SCHED */
8199 #ifdef CONFIG_RT_GROUP_SCHED
8200 static int sched_rt_global_constraints(void)
8204 mutex_lock(&rt_constraints_mutex
);
8205 read_lock(&tasklist_lock
);
8206 ret
= __rt_schedulable(NULL
, 0, 0);
8207 read_unlock(&tasklist_lock
);
8208 mutex_unlock(&rt_constraints_mutex
);
8213 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8215 /* Don't accept realtime tasks when there is no way for them to run */
8216 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8222 #else /* !CONFIG_RT_GROUP_SCHED */
8223 static int sched_rt_global_constraints(void)
8225 unsigned long flags
;
8228 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8229 for_each_possible_cpu(i
) {
8230 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8232 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8233 rt_rq
->rt_runtime
= global_rt_runtime();
8234 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8236 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8240 #endif /* CONFIG_RT_GROUP_SCHED */
8242 static int sched_dl_global_validate(void)
8244 u64 runtime
= global_rt_runtime();
8245 u64 period
= global_rt_period();
8246 u64 new_bw
= to_ratio(period
, runtime
);
8249 unsigned long flags
;
8252 * Here we want to check the bandwidth not being set to some
8253 * value smaller than the currently allocated bandwidth in
8254 * any of the root_domains.
8256 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8257 * cycling on root_domains... Discussion on different/better
8258 * solutions is welcome!
8260 for_each_possible_cpu(cpu
) {
8261 rcu_read_lock_sched();
8262 dl_b
= dl_bw_of(cpu
);
8264 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8265 if (new_bw
< dl_b
->total_bw
)
8267 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8269 rcu_read_unlock_sched();
8278 static void sched_dl_do_global(void)
8283 unsigned long flags
;
8285 def_dl_bandwidth
.dl_period
= global_rt_period();
8286 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8288 if (global_rt_runtime() != RUNTIME_INF
)
8289 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8292 * FIXME: As above...
8294 for_each_possible_cpu(cpu
) {
8295 rcu_read_lock_sched();
8296 dl_b
= dl_bw_of(cpu
);
8298 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8300 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8302 rcu_read_unlock_sched();
8306 static int sched_rt_global_validate(void)
8308 if (sysctl_sched_rt_period
<= 0)
8311 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8312 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8318 static void sched_rt_do_global(void)
8320 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8321 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8324 int sched_rt_handler(struct ctl_table
*table
, int write
,
8325 void __user
*buffer
, size_t *lenp
,
8328 int old_period
, old_runtime
;
8329 static DEFINE_MUTEX(mutex
);
8333 old_period
= sysctl_sched_rt_period
;
8334 old_runtime
= sysctl_sched_rt_runtime
;
8336 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8338 if (!ret
&& write
) {
8339 ret
= sched_rt_global_validate();
8343 ret
= sched_dl_global_validate();
8347 ret
= sched_rt_global_constraints();
8351 sched_rt_do_global();
8352 sched_dl_do_global();
8356 sysctl_sched_rt_period
= old_period
;
8357 sysctl_sched_rt_runtime
= old_runtime
;
8359 mutex_unlock(&mutex
);
8364 int sched_rr_handler(struct ctl_table
*table
, int write
,
8365 void __user
*buffer
, size_t *lenp
,
8369 static DEFINE_MUTEX(mutex
);
8372 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8373 /* make sure that internally we keep jiffies */
8374 /* also, writing zero resets timeslice to default */
8375 if (!ret
&& write
) {
8376 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8377 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8379 mutex_unlock(&mutex
);
8383 #ifdef CONFIG_CGROUP_SCHED
8385 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8387 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8390 static struct cgroup_subsys_state
*
8391 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8393 struct task_group
*parent
= css_tg(parent_css
);
8394 struct task_group
*tg
;
8397 /* This is early initialization for the top cgroup */
8398 return &root_task_group
.css
;
8401 tg
= sched_create_group(parent
);
8403 return ERR_PTR(-ENOMEM
);
8405 sched_online_group(tg
, parent
);
8410 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8412 struct task_group
*tg
= css_tg(css
);
8414 sched_offline_group(tg
);
8417 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8419 struct task_group
*tg
= css_tg(css
);
8422 * Relies on the RCU grace period between css_released() and this.
8424 sched_free_group(tg
);
8428 * This is called before wake_up_new_task(), therefore we really only
8429 * have to set its group bits, all the other stuff does not apply.
8431 static void cpu_cgroup_fork(struct task_struct
*task
)
8436 rq
= task_rq_lock(task
, &rf
);
8438 sched_change_group(task
, TASK_SET_GROUP
);
8440 task_rq_unlock(rq
, task
, &rf
);
8443 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8445 struct task_struct
*task
;
8446 struct cgroup_subsys_state
*css
;
8449 cgroup_taskset_for_each(task
, css
, tset
) {
8450 #ifdef CONFIG_RT_GROUP_SCHED
8451 if (!sched_rt_can_attach(css_tg(css
), task
))
8454 /* We don't support RT-tasks being in separate groups */
8455 if (task
->sched_class
!= &fair_sched_class
)
8459 * Serialize against wake_up_new_task() such that if its
8460 * running, we're sure to observe its full state.
8462 raw_spin_lock_irq(&task
->pi_lock
);
8464 * Avoid calling sched_move_task() before wake_up_new_task()
8465 * has happened. This would lead to problems with PELT, due to
8466 * move wanting to detach+attach while we're not attached yet.
8468 if (task
->state
== TASK_NEW
)
8470 raw_spin_unlock_irq(&task
->pi_lock
);
8478 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8480 struct task_struct
*task
;
8481 struct cgroup_subsys_state
*css
;
8483 cgroup_taskset_for_each(task
, css
, tset
)
8484 sched_move_task(task
);
8487 #ifdef CONFIG_FAIR_GROUP_SCHED
8488 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8489 struct cftype
*cftype
, u64 shareval
)
8491 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8494 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8497 struct task_group
*tg
= css_tg(css
);
8499 return (u64
) scale_load_down(tg
->shares
);
8502 #ifdef CONFIG_CFS_BANDWIDTH
8503 static DEFINE_MUTEX(cfs_constraints_mutex
);
8505 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8506 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8508 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8510 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8512 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8513 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8515 if (tg
== &root_task_group
)
8519 * Ensure we have at some amount of bandwidth every period. This is
8520 * to prevent reaching a state of large arrears when throttled via
8521 * entity_tick() resulting in prolonged exit starvation.
8523 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8527 * Likewise, bound things on the otherside by preventing insane quota
8528 * periods. This also allows us to normalize in computing quota
8531 if (period
> max_cfs_quota_period
)
8535 * Prevent race between setting of cfs_rq->runtime_enabled and
8536 * unthrottle_offline_cfs_rqs().
8539 mutex_lock(&cfs_constraints_mutex
);
8540 ret
= __cfs_schedulable(tg
, period
, quota
);
8544 runtime_enabled
= quota
!= RUNTIME_INF
;
8545 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8547 * If we need to toggle cfs_bandwidth_used, off->on must occur
8548 * before making related changes, and on->off must occur afterwards
8550 if (runtime_enabled
&& !runtime_was_enabled
)
8551 cfs_bandwidth_usage_inc();
8552 raw_spin_lock_irq(&cfs_b
->lock
);
8553 cfs_b
->period
= ns_to_ktime(period
);
8554 cfs_b
->quota
= quota
;
8556 __refill_cfs_bandwidth_runtime(cfs_b
);
8557 /* restart the period timer (if active) to handle new period expiry */
8558 if (runtime_enabled
)
8559 start_cfs_bandwidth(cfs_b
);
8560 raw_spin_unlock_irq(&cfs_b
->lock
);
8562 for_each_online_cpu(i
) {
8563 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8564 struct rq
*rq
= cfs_rq
->rq
;
8566 raw_spin_lock_irq(&rq
->lock
);
8567 cfs_rq
->runtime_enabled
= runtime_enabled
;
8568 cfs_rq
->runtime_remaining
= 0;
8570 if (cfs_rq
->throttled
)
8571 unthrottle_cfs_rq(cfs_rq
);
8572 raw_spin_unlock_irq(&rq
->lock
);
8574 if (runtime_was_enabled
&& !runtime_enabled
)
8575 cfs_bandwidth_usage_dec();
8577 mutex_unlock(&cfs_constraints_mutex
);
8583 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8587 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8588 if (cfs_quota_us
< 0)
8589 quota
= RUNTIME_INF
;
8591 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8593 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8596 long tg_get_cfs_quota(struct task_group
*tg
)
8600 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8603 quota_us
= tg
->cfs_bandwidth
.quota
;
8604 do_div(quota_us
, NSEC_PER_USEC
);
8609 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8613 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8614 quota
= tg
->cfs_bandwidth
.quota
;
8616 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8619 long tg_get_cfs_period(struct task_group
*tg
)
8623 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8624 do_div(cfs_period_us
, NSEC_PER_USEC
);
8626 return cfs_period_us
;
8629 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8632 return tg_get_cfs_quota(css_tg(css
));
8635 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8636 struct cftype
*cftype
, s64 cfs_quota_us
)
8638 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8641 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8644 return tg_get_cfs_period(css_tg(css
));
8647 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8648 struct cftype
*cftype
, u64 cfs_period_us
)
8650 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8653 struct cfs_schedulable_data
{
8654 struct task_group
*tg
;
8659 * normalize group quota/period to be quota/max_period
8660 * note: units are usecs
8662 static u64
normalize_cfs_quota(struct task_group
*tg
,
8663 struct cfs_schedulable_data
*d
)
8671 period
= tg_get_cfs_period(tg
);
8672 quota
= tg_get_cfs_quota(tg
);
8675 /* note: these should typically be equivalent */
8676 if (quota
== RUNTIME_INF
|| quota
== -1)
8679 return to_ratio(period
, quota
);
8682 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8684 struct cfs_schedulable_data
*d
= data
;
8685 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8686 s64 quota
= 0, parent_quota
= -1;
8689 quota
= RUNTIME_INF
;
8691 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8693 quota
= normalize_cfs_quota(tg
, d
);
8694 parent_quota
= parent_b
->hierarchical_quota
;
8697 * ensure max(child_quota) <= parent_quota, inherit when no
8700 if (quota
== RUNTIME_INF
)
8701 quota
= parent_quota
;
8702 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8705 cfs_b
->hierarchical_quota
= quota
;
8710 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8713 struct cfs_schedulable_data data
= {
8719 if (quota
!= RUNTIME_INF
) {
8720 do_div(data
.period
, NSEC_PER_USEC
);
8721 do_div(data
.quota
, NSEC_PER_USEC
);
8725 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8731 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8733 struct task_group
*tg
= css_tg(seq_css(sf
));
8734 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8736 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8737 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8738 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8742 #endif /* CONFIG_CFS_BANDWIDTH */
8743 #endif /* CONFIG_FAIR_GROUP_SCHED */
8745 #ifdef CONFIG_RT_GROUP_SCHED
8746 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8747 struct cftype
*cft
, s64 val
)
8749 return sched_group_set_rt_runtime(css_tg(css
), val
);
8752 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8755 return sched_group_rt_runtime(css_tg(css
));
8758 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8759 struct cftype
*cftype
, u64 rt_period_us
)
8761 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8764 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8767 return sched_group_rt_period(css_tg(css
));
8769 #endif /* CONFIG_RT_GROUP_SCHED */
8771 static struct cftype cpu_files
[] = {
8772 #ifdef CONFIG_FAIR_GROUP_SCHED
8775 .read_u64
= cpu_shares_read_u64
,
8776 .write_u64
= cpu_shares_write_u64
,
8779 #ifdef CONFIG_CFS_BANDWIDTH
8781 .name
= "cfs_quota_us",
8782 .read_s64
= cpu_cfs_quota_read_s64
,
8783 .write_s64
= cpu_cfs_quota_write_s64
,
8786 .name
= "cfs_period_us",
8787 .read_u64
= cpu_cfs_period_read_u64
,
8788 .write_u64
= cpu_cfs_period_write_u64
,
8792 .seq_show
= cpu_stats_show
,
8795 #ifdef CONFIG_RT_GROUP_SCHED
8797 .name
= "rt_runtime_us",
8798 .read_s64
= cpu_rt_runtime_read
,
8799 .write_s64
= cpu_rt_runtime_write
,
8802 .name
= "rt_period_us",
8803 .read_u64
= cpu_rt_period_read_uint
,
8804 .write_u64
= cpu_rt_period_write_uint
,
8810 struct cgroup_subsys cpu_cgrp_subsys
= {
8811 .css_alloc
= cpu_cgroup_css_alloc
,
8812 .css_released
= cpu_cgroup_css_released
,
8813 .css_free
= cpu_cgroup_css_free
,
8814 .fork
= cpu_cgroup_fork
,
8815 .can_attach
= cpu_cgroup_can_attach
,
8816 .attach
= cpu_cgroup_attach
,
8817 .legacy_cftypes
= cpu_files
,
8821 #endif /* CONFIG_CGROUP_SCHED */
8823 void dump_cpu_task(int cpu
)
8825 pr_info("Task dump for CPU %d:\n", cpu
);
8826 sched_show_task(cpu_curr(cpu
));
8830 * Nice levels are multiplicative, with a gentle 10% change for every
8831 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8832 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8833 * that remained on nice 0.
8835 * The "10% effect" is relative and cumulative: from _any_ nice level,
8836 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8837 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8838 * If a task goes up by ~10% and another task goes down by ~10% then
8839 * the relative distance between them is ~25%.)
8841 const int sched_prio_to_weight
[40] = {
8842 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8843 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8844 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8845 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8846 /* 0 */ 1024, 820, 655, 526, 423,
8847 /* 5 */ 335, 272, 215, 172, 137,
8848 /* 10 */ 110, 87, 70, 56, 45,
8849 /* 15 */ 36, 29, 23, 18, 15,
8853 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8855 * In cases where the weight does not change often, we can use the
8856 * precalculated inverse to speed up arithmetics by turning divisions
8857 * into multiplications:
8859 const u32 sched_prio_to_wmult
[40] = {
8860 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8861 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8862 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8863 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8864 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8865 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8866 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8867 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,