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>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex
);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
96 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
98 void update_rq_clock(struct rq
*rq
)
102 lockdep_assert_held(&rq
->lock
);
104 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
107 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
111 update_rq_clock_task(rq
, delta
);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug
unsigned int sysctl_sched_features
=
122 #include "features.h"
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
131 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
134 * period over which we average the RT time consumption, measured
139 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
142 * period over which we measure -rt task cpu usage in us.
145 unsigned int sysctl_sched_rt_period
= 1000000;
147 __read_mostly
int scheduler_running
;
150 * part of the period that we allow rt tasks to run in us.
153 int sysctl_sched_rt_runtime
= 950000;
155 /* cpus with isolated domains */
156 cpumask_var_t cpu_isolated_map
;
159 * this_rq_lock - lock this runqueue and disable interrupts.
161 static struct rq
*this_rq_lock(void)
168 raw_spin_lock(&rq
->lock
);
174 * __task_rq_lock - lock the rq @p resides on.
176 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
181 lockdep_assert_held(&p
->pi_lock
);
185 raw_spin_lock(&rq
->lock
);
186 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
187 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
190 raw_spin_unlock(&rq
->lock
);
192 while (unlikely(task_on_rq_migrating(p
)))
198 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
200 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
201 __acquires(p
->pi_lock
)
207 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
209 raw_spin_lock(&rq
->lock
);
211 * move_queued_task() task_rq_lock()
214 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
215 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
216 * [S] ->cpu = new_cpu [L] task_rq()
220 * If we observe the old cpu in task_rq_lock, the acquire of
221 * the old rq->lock will fully serialize against the stores.
223 * If we observe the new cpu in task_rq_lock, the acquire will
224 * pair with the WMB to ensure we must then also see migrating.
226 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
227 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
230 raw_spin_unlock(&rq
->lock
);
231 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
233 while (unlikely(task_on_rq_migrating(p
)))
238 #ifdef CONFIG_SCHED_HRTICK
240 * Use HR-timers to deliver accurate preemption points.
243 static void hrtick_clear(struct rq
*rq
)
245 if (hrtimer_active(&rq
->hrtick_timer
))
246 hrtimer_cancel(&rq
->hrtick_timer
);
250 * High-resolution timer tick.
251 * Runs from hardirq context with interrupts disabled.
253 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
255 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
257 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
259 raw_spin_lock(&rq
->lock
);
261 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
262 raw_spin_unlock(&rq
->lock
);
264 return HRTIMER_NORESTART
;
269 static void __hrtick_restart(struct rq
*rq
)
271 struct hrtimer
*timer
= &rq
->hrtick_timer
;
273 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
277 * called from hardirq (IPI) context
279 static void __hrtick_start(void *arg
)
283 raw_spin_lock(&rq
->lock
);
284 __hrtick_restart(rq
);
285 rq
->hrtick_csd_pending
= 0;
286 raw_spin_unlock(&rq
->lock
);
290 * Called to set the hrtick timer state.
292 * called with rq->lock held and irqs disabled
294 void hrtick_start(struct rq
*rq
, u64 delay
)
296 struct hrtimer
*timer
= &rq
->hrtick_timer
;
301 * Don't schedule slices shorter than 10000ns, that just
302 * doesn't make sense and can cause timer DoS.
304 delta
= max_t(s64
, delay
, 10000LL);
305 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
307 hrtimer_set_expires(timer
, time
);
309 if (rq
== this_rq()) {
310 __hrtick_restart(rq
);
311 } else if (!rq
->hrtick_csd_pending
) {
312 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
313 rq
->hrtick_csd_pending
= 1;
319 * Called to set the hrtick timer state.
321 * called with rq->lock held and irqs disabled
323 void hrtick_start(struct rq
*rq
, u64 delay
)
326 * Don't schedule slices shorter than 10000ns, that just
327 * doesn't make sense. Rely on vruntime for fairness.
329 delay
= max_t(u64
, delay
, 10000LL);
330 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
331 HRTIMER_MODE_REL_PINNED
);
333 #endif /* CONFIG_SMP */
335 static void init_rq_hrtick(struct rq
*rq
)
338 rq
->hrtick_csd_pending
= 0;
340 rq
->hrtick_csd
.flags
= 0;
341 rq
->hrtick_csd
.func
= __hrtick_start
;
342 rq
->hrtick_csd
.info
= rq
;
345 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
346 rq
->hrtick_timer
.function
= hrtick
;
348 #else /* CONFIG_SCHED_HRTICK */
349 static inline void hrtick_clear(struct rq
*rq
)
353 static inline void init_rq_hrtick(struct rq
*rq
)
356 #endif /* CONFIG_SCHED_HRTICK */
359 * cmpxchg based fetch_or, macro so it works for different integer types
361 #define fetch_or(ptr, mask) \
363 typeof(ptr) _ptr = (ptr); \
364 typeof(mask) _mask = (mask); \
365 typeof(*_ptr) _old, _val = *_ptr; \
368 _old = cmpxchg(_ptr, _val, _val | _mask); \
376 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
378 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
379 * this avoids any races wrt polling state changes and thereby avoids
382 static bool set_nr_and_not_polling(struct task_struct
*p
)
384 struct thread_info
*ti
= task_thread_info(p
);
385 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
389 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
391 * If this returns true, then the idle task promises to call
392 * sched_ttwu_pending() and reschedule soon.
394 static bool set_nr_if_polling(struct task_struct
*p
)
396 struct thread_info
*ti
= task_thread_info(p
);
397 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
400 if (!(val
& _TIF_POLLING_NRFLAG
))
402 if (val
& _TIF_NEED_RESCHED
)
404 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
413 static bool set_nr_and_not_polling(struct task_struct
*p
)
415 set_tsk_need_resched(p
);
420 static bool set_nr_if_polling(struct task_struct
*p
)
427 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
429 struct wake_q_node
*node
= &task
->wake_q
;
432 * Atomically grab the task, if ->wake_q is !nil already it means
433 * its already queued (either by us or someone else) and will get the
434 * wakeup due to that.
436 * This cmpxchg() implies a full barrier, which pairs with the write
437 * barrier implied by the wakeup in wake_up_q().
439 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
442 get_task_struct(task
);
445 * The head is context local, there can be no concurrency.
448 head
->lastp
= &node
->next
;
451 void wake_up_q(struct wake_q_head
*head
)
453 struct wake_q_node
*node
= head
->first
;
455 while (node
!= WAKE_Q_TAIL
) {
456 struct task_struct
*task
;
458 task
= container_of(node
, struct task_struct
, wake_q
);
460 /* task can safely be re-inserted now */
462 task
->wake_q
.next
= NULL
;
465 * wake_up_process() implies a wmb() to pair with the queueing
466 * in wake_q_add() so as not to miss wakeups.
468 wake_up_process(task
);
469 put_task_struct(task
);
474 * resched_curr - mark rq's current task 'to be rescheduled now'.
476 * On UP this means the setting of the need_resched flag, on SMP it
477 * might also involve a cross-CPU call to trigger the scheduler on
480 void resched_curr(struct rq
*rq
)
482 struct task_struct
*curr
= rq
->curr
;
485 lockdep_assert_held(&rq
->lock
);
487 if (test_tsk_need_resched(curr
))
492 if (cpu
== smp_processor_id()) {
493 set_tsk_need_resched(curr
);
494 set_preempt_need_resched();
498 if (set_nr_and_not_polling(curr
))
499 smp_send_reschedule(cpu
);
501 trace_sched_wake_idle_without_ipi(cpu
);
504 void resched_cpu(int cpu
)
506 struct rq
*rq
= cpu_rq(cpu
);
509 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
512 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
516 #ifdef CONFIG_NO_HZ_COMMON
518 * In the semi idle case, use the nearest busy cpu for migrating timers
519 * from an idle cpu. This is good for power-savings.
521 * We don't do similar optimization for completely idle system, as
522 * selecting an idle cpu will add more delays to the timers than intended
523 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
525 int get_nohz_timer_target(void)
527 int i
, cpu
= smp_processor_id();
528 struct sched_domain
*sd
;
530 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
534 for_each_domain(cpu
, sd
) {
535 for_each_cpu(i
, sched_domain_span(sd
)) {
539 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
546 if (!is_housekeeping_cpu(cpu
))
547 cpu
= housekeeping_any_cpu();
553 * When add_timer_on() enqueues a timer into the timer wheel of an
554 * idle CPU then this timer might expire before the next timer event
555 * which is scheduled to wake up that CPU. In case of a completely
556 * idle system the next event might even be infinite time into the
557 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
558 * leaves the inner idle loop so the newly added timer is taken into
559 * account when the CPU goes back to idle and evaluates the timer
560 * wheel for the next timer event.
562 static void wake_up_idle_cpu(int cpu
)
564 struct rq
*rq
= cpu_rq(cpu
);
566 if (cpu
== smp_processor_id())
569 if (set_nr_and_not_polling(rq
->idle
))
570 smp_send_reschedule(cpu
);
572 trace_sched_wake_idle_without_ipi(cpu
);
575 static bool wake_up_full_nohz_cpu(int cpu
)
578 * We just need the target to call irq_exit() and re-evaluate
579 * the next tick. The nohz full kick at least implies that.
580 * If needed we can still optimize that later with an
583 if (tick_nohz_full_cpu(cpu
)) {
584 if (cpu
!= smp_processor_id() ||
585 tick_nohz_tick_stopped())
586 tick_nohz_full_kick_cpu(cpu
);
593 void wake_up_nohz_cpu(int cpu
)
595 if (!wake_up_full_nohz_cpu(cpu
))
596 wake_up_idle_cpu(cpu
);
599 static inline bool got_nohz_idle_kick(void)
601 int cpu
= smp_processor_id();
603 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
606 if (idle_cpu(cpu
) && !need_resched())
610 * We can't run Idle Load Balance on this CPU for this time so we
611 * cancel it and clear NOHZ_BALANCE_KICK
613 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
617 #else /* CONFIG_NO_HZ_COMMON */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ_COMMON */
626 #ifdef CONFIG_NO_HZ_FULL
627 bool sched_can_stop_tick(struct rq
*rq
)
631 /* Deadline tasks, even if single, need the tick */
632 if (rq
->dl
.dl_nr_running
)
636 * If there are more than one RR tasks, we need the tick to effect the
637 * actual RR behaviour.
639 if (rq
->rt
.rr_nr_running
) {
640 if (rq
->rt
.rr_nr_running
== 1)
647 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
648 * forced preemption between FIFO tasks.
650 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
655 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
656 * if there's more than one we need the tick for involuntary
659 if (rq
->nr_running
> 1)
664 #endif /* CONFIG_NO_HZ_FULL */
666 void sched_avg_update(struct rq
*rq
)
668 s64 period
= sched_avg_period();
670 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
672 * Inline assembly required to prevent the compiler
673 * optimising this loop into a divmod call.
674 * See __iter_div_u64_rem() for another example of this.
676 asm("" : "+rm" (rq
->age_stamp
));
677 rq
->age_stamp
+= period
;
682 #endif /* CONFIG_SMP */
684 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
685 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
687 * Iterate task_group tree rooted at *from, calling @down when first entering a
688 * node and @up when leaving it for the final time.
690 * Caller must hold rcu_lock or sufficient equivalent.
692 int walk_tg_tree_from(struct task_group
*from
,
693 tg_visitor down
, tg_visitor up
, void *data
)
695 struct task_group
*parent
, *child
;
701 ret
= (*down
)(parent
, data
);
704 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
711 ret
= (*up
)(parent
, data
);
712 if (ret
|| parent
== from
)
716 parent
= parent
->parent
;
723 int tg_nop(struct task_group
*tg
, void *data
)
729 static void set_load_weight(struct task_struct
*p
)
731 int prio
= p
->static_prio
- MAX_RT_PRIO
;
732 struct load_weight
*load
= &p
->se
.load
;
735 * SCHED_IDLE tasks get minimal weight:
737 if (idle_policy(p
->policy
)) {
738 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
739 load
->inv_weight
= WMULT_IDLEPRIO
;
743 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
744 load
->inv_weight
= sched_prio_to_wmult
[prio
];
747 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
750 if (!(flags
& ENQUEUE_RESTORE
))
751 sched_info_queued(rq
, p
);
752 p
->sched_class
->enqueue_task(rq
, p
, flags
);
755 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
758 if (!(flags
& DEQUEUE_SAVE
))
759 sched_info_dequeued(rq
, p
);
760 p
->sched_class
->dequeue_task(rq
, p
, flags
);
763 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
765 if (task_contributes_to_load(p
))
766 rq
->nr_uninterruptible
--;
768 enqueue_task(rq
, p
, flags
);
771 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
773 if (task_contributes_to_load(p
))
774 rq
->nr_uninterruptible
++;
776 dequeue_task(rq
, p
, flags
);
779 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
782 * In theory, the compile should just see 0 here, and optimize out the call
783 * to sched_rt_avg_update. But I don't trust it...
785 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
786 s64 steal
= 0, irq_delta
= 0;
788 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
789 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
792 * Since irq_time is only updated on {soft,}irq_exit, we might run into
793 * this case when a previous update_rq_clock() happened inside a
796 * When this happens, we stop ->clock_task and only update the
797 * prev_irq_time stamp to account for the part that fit, so that a next
798 * update will consume the rest. This ensures ->clock_task is
801 * It does however cause some slight miss-attribution of {soft,}irq
802 * time, a more accurate solution would be to update the irq_time using
803 * the current rq->clock timestamp, except that would require using
806 if (irq_delta
> delta
)
809 rq
->prev_irq_time
+= irq_delta
;
812 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
813 if (static_key_false((¶virt_steal_rq_enabled
))) {
814 steal
= paravirt_steal_clock(cpu_of(rq
));
815 steal
-= rq
->prev_steal_time_rq
;
817 if (unlikely(steal
> delta
))
820 rq
->prev_steal_time_rq
+= steal
;
825 rq
->clock_task
+= delta
;
827 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
828 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
829 sched_rt_avg_update(rq
, irq_delta
+ steal
);
833 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
835 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
836 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
840 * Make it appear like a SCHED_FIFO task, its something
841 * userspace knows about and won't get confused about.
843 * Also, it will make PI more or less work without too
844 * much confusion -- but then, stop work should not
845 * rely on PI working anyway.
847 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
849 stop
->sched_class
= &stop_sched_class
;
852 cpu_rq(cpu
)->stop
= stop
;
856 * Reset it back to a normal scheduling class so that
857 * it can die in pieces.
859 old_stop
->sched_class
= &rt_sched_class
;
864 * __normal_prio - return the priority that is based on the static prio
866 static inline int __normal_prio(struct task_struct
*p
)
868 return p
->static_prio
;
872 * Calculate the expected normal priority: i.e. priority
873 * without taking RT-inheritance into account. Might be
874 * boosted by interactivity modifiers. Changes upon fork,
875 * setprio syscalls, and whenever the interactivity
876 * estimator recalculates.
878 static inline int normal_prio(struct task_struct
*p
)
882 if (task_has_dl_policy(p
))
883 prio
= MAX_DL_PRIO
-1;
884 else if (task_has_rt_policy(p
))
885 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
887 prio
= __normal_prio(p
);
892 * Calculate the current priority, i.e. the priority
893 * taken into account by the scheduler. This value might
894 * be boosted by RT tasks, or might be boosted by
895 * interactivity modifiers. Will be RT if the task got
896 * RT-boosted. If not then it returns p->normal_prio.
898 static int effective_prio(struct task_struct
*p
)
900 p
->normal_prio
= normal_prio(p
);
902 * If we are RT tasks or we were boosted to RT priority,
903 * keep the priority unchanged. Otherwise, update priority
904 * to the normal priority:
906 if (!rt_prio(p
->prio
))
907 return p
->normal_prio
;
912 * task_curr - is this task currently executing on a CPU?
913 * @p: the task in question.
915 * Return: 1 if the task is currently executing. 0 otherwise.
917 inline int task_curr(const struct task_struct
*p
)
919 return cpu_curr(task_cpu(p
)) == p
;
923 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
924 * use the balance_callback list if you want balancing.
926 * this means any call to check_class_changed() must be followed by a call to
927 * balance_callback().
929 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
930 const struct sched_class
*prev_class
,
933 if (prev_class
!= p
->sched_class
) {
934 if (prev_class
->switched_from
)
935 prev_class
->switched_from(rq
, p
);
937 p
->sched_class
->switched_to(rq
, p
);
938 } else if (oldprio
!= p
->prio
|| dl_task(p
))
939 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
942 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
944 const struct sched_class
*class;
946 if (p
->sched_class
== rq
->curr
->sched_class
) {
947 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
949 for_each_class(class) {
950 if (class == rq
->curr
->sched_class
)
952 if (class == p
->sched_class
) {
960 * A queue event has occurred, and we're going to schedule. In
961 * this case, we can save a useless back to back clock update.
963 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
964 rq_clock_skip_update(rq
, true);
969 * This is how migration works:
971 * 1) we invoke migration_cpu_stop() on the target CPU using
973 * 2) stopper starts to run (implicitly forcing the migrated thread
975 * 3) it checks whether the migrated task is still in the wrong runqueue.
976 * 4) if it's in the wrong runqueue then the migration thread removes
977 * it and puts it into the right queue.
978 * 5) stopper completes and stop_one_cpu() returns and the migration
983 * move_queued_task - move a queued task to new rq.
985 * Returns (locked) new rq. Old rq's lock is released.
987 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
989 lockdep_assert_held(&rq
->lock
);
991 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
992 dequeue_task(rq
, p
, 0);
993 set_task_cpu(p
, new_cpu
);
994 raw_spin_unlock(&rq
->lock
);
996 rq
= cpu_rq(new_cpu
);
998 raw_spin_lock(&rq
->lock
);
999 BUG_ON(task_cpu(p
) != new_cpu
);
1000 enqueue_task(rq
, p
, 0);
1001 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1002 check_preempt_curr(rq
, p
, 0);
1007 struct migration_arg
{
1008 struct task_struct
*task
;
1013 * Move (not current) task off this cpu, onto dest cpu. We're doing
1014 * this because either it can't run here any more (set_cpus_allowed()
1015 * away from this CPU, or CPU going down), or because we're
1016 * attempting to rebalance this task on exec (sched_exec).
1018 * So we race with normal scheduler movements, but that's OK, as long
1019 * as the task is no longer on this CPU.
1021 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1023 if (unlikely(!cpu_active(dest_cpu
)))
1026 /* Affinity changed (again). */
1027 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1030 rq
= move_queued_task(rq
, p
, dest_cpu
);
1036 * migration_cpu_stop - this will be executed by a highprio stopper thread
1037 * and performs thread migration by bumping thread off CPU then
1038 * 'pushing' onto another runqueue.
1040 static int migration_cpu_stop(void *data
)
1042 struct migration_arg
*arg
= data
;
1043 struct task_struct
*p
= arg
->task
;
1044 struct rq
*rq
= this_rq();
1047 * The original target cpu might have gone down and we might
1048 * be on another cpu but it doesn't matter.
1050 local_irq_disable();
1052 * We need to explicitly wake pending tasks before running
1053 * __migrate_task() such that we will not miss enforcing cpus_allowed
1054 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1056 sched_ttwu_pending();
1058 raw_spin_lock(&p
->pi_lock
);
1059 raw_spin_lock(&rq
->lock
);
1061 * If task_rq(p) != rq, it cannot be migrated here, because we're
1062 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1063 * we're holding p->pi_lock.
1065 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1066 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1067 raw_spin_unlock(&rq
->lock
);
1068 raw_spin_unlock(&p
->pi_lock
);
1075 * sched_class::set_cpus_allowed must do the below, but is not required to
1076 * actually call this function.
1078 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1080 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1081 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1084 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1086 struct rq
*rq
= task_rq(p
);
1087 bool queued
, running
;
1089 lockdep_assert_held(&p
->pi_lock
);
1091 queued
= task_on_rq_queued(p
);
1092 running
= task_current(rq
, p
);
1096 * Because __kthread_bind() calls this on blocked tasks without
1099 lockdep_assert_held(&rq
->lock
);
1100 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1103 put_prev_task(rq
, p
);
1105 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1108 p
->sched_class
->set_curr_task(rq
);
1110 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1114 * Change a given task's CPU affinity. Migrate the thread to a
1115 * proper CPU and schedule it away if the CPU it's executing on
1116 * is removed from the allowed bitmask.
1118 * NOTE: the caller must have a valid reference to the task, the
1119 * task must not exit() & deallocate itself prematurely. The
1120 * call is not atomic; no spinlocks may be held.
1122 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1123 const struct cpumask
*new_mask
, bool check
)
1125 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1126 unsigned int dest_cpu
;
1131 rq
= task_rq_lock(p
, &rf
);
1133 if (p
->flags
& PF_KTHREAD
) {
1135 * Kernel threads are allowed on online && !active CPUs
1137 cpu_valid_mask
= cpu_online_mask
;
1141 * Must re-check here, to close a race against __kthread_bind(),
1142 * sched_setaffinity() is not guaranteed to observe the flag.
1144 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1149 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1152 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1157 do_set_cpus_allowed(p
, new_mask
);
1159 if (p
->flags
& PF_KTHREAD
) {
1161 * For kernel threads that do indeed end up on online &&
1162 * !active we want to ensure they are strict per-cpu threads.
1164 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1165 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1166 p
->nr_cpus_allowed
!= 1);
1169 /* Can the task run on the task's current CPU? If so, we're done */
1170 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1173 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1174 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1175 struct migration_arg arg
= { p
, dest_cpu
};
1176 /* Need help from migration thread: drop lock and wait. */
1177 task_rq_unlock(rq
, p
, &rf
);
1178 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1179 tlb_migrate_finish(p
->mm
);
1181 } else if (task_on_rq_queued(p
)) {
1183 * OK, since we're going to drop the lock immediately
1184 * afterwards anyway.
1186 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
1187 rq
= move_queued_task(rq
, p
, dest_cpu
);
1188 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
1191 task_rq_unlock(rq
, p
, &rf
);
1196 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1198 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1200 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1202 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1204 #ifdef CONFIG_SCHED_DEBUG
1206 * We should never call set_task_cpu() on a blocked task,
1207 * ttwu() will sort out the placement.
1209 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1213 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1214 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1215 * time relying on p->on_rq.
1217 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1218 p
->sched_class
== &fair_sched_class
&&
1219 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1221 #ifdef CONFIG_LOCKDEP
1223 * The caller should hold either p->pi_lock or rq->lock, when changing
1224 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1226 * sched_move_task() holds both and thus holding either pins the cgroup,
1229 * Furthermore, all task_rq users should acquire both locks, see
1232 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1233 lockdep_is_held(&task_rq(p
)->lock
)));
1237 trace_sched_migrate_task(p
, new_cpu
);
1239 if (task_cpu(p
) != new_cpu
) {
1240 if (p
->sched_class
->migrate_task_rq
)
1241 p
->sched_class
->migrate_task_rq(p
);
1242 p
->se
.nr_migrations
++;
1243 perf_event_task_migrate(p
);
1246 __set_task_cpu(p
, new_cpu
);
1249 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1251 if (task_on_rq_queued(p
)) {
1252 struct rq
*src_rq
, *dst_rq
;
1254 src_rq
= task_rq(p
);
1255 dst_rq
= cpu_rq(cpu
);
1257 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1258 deactivate_task(src_rq
, p
, 0);
1259 set_task_cpu(p
, cpu
);
1260 activate_task(dst_rq
, p
, 0);
1261 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1262 check_preempt_curr(dst_rq
, p
, 0);
1265 * Task isn't running anymore; make it appear like we migrated
1266 * it before it went to sleep. This means on wakeup we make the
1267 * previous cpu our targer instead of where it really is.
1273 struct migration_swap_arg
{
1274 struct task_struct
*src_task
, *dst_task
;
1275 int src_cpu
, dst_cpu
;
1278 static int migrate_swap_stop(void *data
)
1280 struct migration_swap_arg
*arg
= data
;
1281 struct rq
*src_rq
, *dst_rq
;
1284 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1287 src_rq
= cpu_rq(arg
->src_cpu
);
1288 dst_rq
= cpu_rq(arg
->dst_cpu
);
1290 double_raw_lock(&arg
->src_task
->pi_lock
,
1291 &arg
->dst_task
->pi_lock
);
1292 double_rq_lock(src_rq
, dst_rq
);
1294 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1297 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1300 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1303 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1306 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1307 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1312 double_rq_unlock(src_rq
, dst_rq
);
1313 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1314 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1320 * Cross migrate two tasks
1322 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1324 struct migration_swap_arg arg
;
1327 arg
= (struct migration_swap_arg
){
1329 .src_cpu
= task_cpu(cur
),
1331 .dst_cpu
= task_cpu(p
),
1334 if (arg
.src_cpu
== arg
.dst_cpu
)
1338 * These three tests are all lockless; this is OK since all of them
1339 * will be re-checked with proper locks held further down the line.
1341 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1344 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1347 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1350 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1351 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1358 * wait_task_inactive - wait for a thread to unschedule.
1360 * If @match_state is nonzero, it's the @p->state value just checked and
1361 * not expected to change. If it changes, i.e. @p might have woken up,
1362 * then return zero. When we succeed in waiting for @p to be off its CPU,
1363 * we return a positive number (its total switch count). If a second call
1364 * a short while later returns the same number, the caller can be sure that
1365 * @p has remained unscheduled the whole time.
1367 * The caller must ensure that the task *will* unschedule sometime soon,
1368 * else this function might spin for a *long* time. This function can't
1369 * be called with interrupts off, or it may introduce deadlock with
1370 * smp_call_function() if an IPI is sent by the same process we are
1371 * waiting to become inactive.
1373 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1375 int running
, queued
;
1382 * We do the initial early heuristics without holding
1383 * any task-queue locks at all. We'll only try to get
1384 * the runqueue lock when things look like they will
1390 * If the task is actively running on another CPU
1391 * still, just relax and busy-wait without holding
1394 * NOTE! Since we don't hold any locks, it's not
1395 * even sure that "rq" stays as the right runqueue!
1396 * But we don't care, since "task_running()" will
1397 * return false if the runqueue has changed and p
1398 * is actually now running somewhere else!
1400 while (task_running(rq
, p
)) {
1401 if (match_state
&& unlikely(p
->state
!= match_state
))
1407 * Ok, time to look more closely! We need the rq
1408 * lock now, to be *sure*. If we're wrong, we'll
1409 * just go back and repeat.
1411 rq
= task_rq_lock(p
, &rf
);
1412 trace_sched_wait_task(p
);
1413 running
= task_running(rq
, p
);
1414 queued
= task_on_rq_queued(p
);
1416 if (!match_state
|| p
->state
== match_state
)
1417 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1418 task_rq_unlock(rq
, p
, &rf
);
1421 * If it changed from the expected state, bail out now.
1423 if (unlikely(!ncsw
))
1427 * Was it really running after all now that we
1428 * checked with the proper locks actually held?
1430 * Oops. Go back and try again..
1432 if (unlikely(running
)) {
1438 * It's not enough that it's not actively running,
1439 * it must be off the runqueue _entirely_, and not
1442 * So if it was still runnable (but just not actively
1443 * running right now), it's preempted, and we should
1444 * yield - it could be a while.
1446 if (unlikely(queued
)) {
1447 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1449 set_current_state(TASK_UNINTERRUPTIBLE
);
1450 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1455 * Ahh, all good. It wasn't running, and it wasn't
1456 * runnable, which means that it will never become
1457 * running in the future either. We're all done!
1466 * kick_process - kick a running thread to enter/exit the kernel
1467 * @p: the to-be-kicked thread
1469 * Cause a process which is running on another CPU to enter
1470 * kernel-mode, without any delay. (to get signals handled.)
1472 * NOTE: this function doesn't have to take the runqueue lock,
1473 * because all it wants to ensure is that the remote task enters
1474 * the kernel. If the IPI races and the task has been migrated
1475 * to another CPU then no harm is done and the purpose has been
1478 void kick_process(struct task_struct
*p
)
1484 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1485 smp_send_reschedule(cpu
);
1488 EXPORT_SYMBOL_GPL(kick_process
);
1491 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1493 * A few notes on cpu_active vs cpu_online:
1495 * - cpu_active must be a subset of cpu_online
1497 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1498 * see __set_cpus_allowed_ptr(). At this point the newly online
1499 * cpu isn't yet part of the sched domains, and balancing will not
1502 * - on cpu-down we clear cpu_active() to mask the sched domains and
1503 * avoid the load balancer to place new tasks on the to be removed
1504 * cpu. Existing tasks will remain running there and will be taken
1507 * This means that fallback selection must not select !active CPUs.
1508 * And can assume that any active CPU must be online. Conversely
1509 * select_task_rq() below may allow selection of !active CPUs in order
1510 * to satisfy the above rules.
1512 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1514 int nid
= cpu_to_node(cpu
);
1515 const struct cpumask
*nodemask
= NULL
;
1516 enum { cpuset
, possible
, fail
} state
= cpuset
;
1520 * If the node that the cpu is on has been offlined, cpu_to_node()
1521 * will return -1. There is no cpu on the node, and we should
1522 * select the cpu on the other node.
1525 nodemask
= cpumask_of_node(nid
);
1527 /* Look for allowed, online CPU in same node. */
1528 for_each_cpu(dest_cpu
, nodemask
) {
1529 if (!cpu_active(dest_cpu
))
1531 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1537 /* Any allowed, online CPU? */
1538 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1539 if (!cpu_active(dest_cpu
))
1544 /* No more Mr. Nice Guy. */
1547 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1548 cpuset_cpus_allowed_fallback(p
);
1554 do_set_cpus_allowed(p
, cpu_possible_mask
);
1565 if (state
!= cpuset
) {
1567 * Don't tell them about moving exiting tasks or
1568 * kernel threads (both mm NULL), since they never
1571 if (p
->mm
&& printk_ratelimit()) {
1572 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1573 task_pid_nr(p
), p
->comm
, cpu
);
1581 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1584 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1586 lockdep_assert_held(&p
->pi_lock
);
1588 if (tsk_nr_cpus_allowed(p
) > 1)
1589 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1591 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1594 * In order not to call set_task_cpu() on a blocking task we need
1595 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1598 * Since this is common to all placement strategies, this lives here.
1600 * [ this allows ->select_task() to simply return task_cpu(p) and
1601 * not worry about this generic constraint ]
1603 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1605 cpu
= select_fallback_rq(task_cpu(p
), p
);
1610 static void update_avg(u64
*avg
, u64 sample
)
1612 s64 diff
= sample
- *avg
;
1618 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1619 const struct cpumask
*new_mask
, bool check
)
1621 return set_cpus_allowed_ptr(p
, new_mask
);
1624 #endif /* CONFIG_SMP */
1627 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1629 #ifdef CONFIG_SCHEDSTATS
1630 struct rq
*rq
= this_rq();
1633 int this_cpu
= smp_processor_id();
1635 if (cpu
== this_cpu
) {
1636 schedstat_inc(rq
, ttwu_local
);
1637 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1639 struct sched_domain
*sd
;
1641 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1643 for_each_domain(this_cpu
, sd
) {
1644 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1645 schedstat_inc(sd
, ttwu_wake_remote
);
1652 if (wake_flags
& WF_MIGRATED
)
1653 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1655 #endif /* CONFIG_SMP */
1657 schedstat_inc(rq
, ttwu_count
);
1658 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1660 if (wake_flags
& WF_SYNC
)
1661 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1663 #endif /* CONFIG_SCHEDSTATS */
1666 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1668 activate_task(rq
, p
, en_flags
);
1669 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1671 /* if a worker is waking up, notify workqueue */
1672 if (p
->flags
& PF_WQ_WORKER
)
1673 wq_worker_waking_up(p
, cpu_of(rq
));
1677 * Mark the task runnable and perform wakeup-preemption.
1679 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1680 struct pin_cookie cookie
)
1682 check_preempt_curr(rq
, p
, wake_flags
);
1683 p
->state
= TASK_RUNNING
;
1684 trace_sched_wakeup(p
);
1687 if (p
->sched_class
->task_woken
) {
1689 * Our task @p is fully woken up and running; so its safe to
1690 * drop the rq->lock, hereafter rq is only used for statistics.
1692 lockdep_unpin_lock(&rq
->lock
, cookie
);
1693 p
->sched_class
->task_woken(rq
, p
);
1694 lockdep_repin_lock(&rq
->lock
, cookie
);
1697 if (rq
->idle_stamp
) {
1698 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1699 u64 max
= 2*rq
->max_idle_balance_cost
;
1701 update_avg(&rq
->avg_idle
, delta
);
1703 if (rq
->avg_idle
> max
)
1712 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1713 struct pin_cookie cookie
)
1715 int en_flags
= ENQUEUE_WAKEUP
;
1717 lockdep_assert_held(&rq
->lock
);
1720 if (p
->sched_contributes_to_load
)
1721 rq
->nr_uninterruptible
--;
1723 if (wake_flags
& WF_MIGRATED
)
1724 en_flags
|= ENQUEUE_MIGRATED
;
1727 ttwu_activate(rq
, p
, en_flags
);
1728 ttwu_do_wakeup(rq
, p
, wake_flags
, cookie
);
1732 * Called in case the task @p isn't fully descheduled from its runqueue,
1733 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1734 * since all we need to do is flip p->state to TASK_RUNNING, since
1735 * the task is still ->on_rq.
1737 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1743 rq
= __task_rq_lock(p
, &rf
);
1744 if (task_on_rq_queued(p
)) {
1745 /* check_preempt_curr() may use rq clock */
1746 update_rq_clock(rq
);
1747 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
.cookie
);
1750 __task_rq_unlock(rq
, &rf
);
1756 void sched_ttwu_pending(void)
1758 struct rq
*rq
= this_rq();
1759 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1760 struct pin_cookie cookie
;
1761 struct task_struct
*p
;
1762 unsigned long flags
;
1767 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1768 cookie
= lockdep_pin_lock(&rq
->lock
);
1773 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1774 llist
= llist_next(llist
);
1776 if (p
->sched_remote_wakeup
)
1777 wake_flags
= WF_MIGRATED
;
1779 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1782 lockdep_unpin_lock(&rq
->lock
, cookie
);
1783 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1786 void scheduler_ipi(void)
1789 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1790 * TIF_NEED_RESCHED remotely (for the first time) will also send
1793 preempt_fold_need_resched();
1795 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1799 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1800 * traditionally all their work was done from the interrupt return
1801 * path. Now that we actually do some work, we need to make sure
1804 * Some archs already do call them, luckily irq_enter/exit nest
1807 * Arguably we should visit all archs and update all handlers,
1808 * however a fair share of IPIs are still resched only so this would
1809 * somewhat pessimize the simple resched case.
1812 sched_ttwu_pending();
1815 * Check if someone kicked us for doing the nohz idle load balance.
1817 if (unlikely(got_nohz_idle_kick())) {
1818 this_rq()->idle_balance
= 1;
1819 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1824 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1826 struct rq
*rq
= cpu_rq(cpu
);
1828 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1830 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1831 if (!set_nr_if_polling(rq
->idle
))
1832 smp_send_reschedule(cpu
);
1834 trace_sched_wake_idle_without_ipi(cpu
);
1838 void wake_up_if_idle(int cpu
)
1840 struct rq
*rq
= cpu_rq(cpu
);
1841 unsigned long flags
;
1845 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1848 if (set_nr_if_polling(rq
->idle
)) {
1849 trace_sched_wake_idle_without_ipi(cpu
);
1851 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1852 if (is_idle_task(rq
->curr
))
1853 smp_send_reschedule(cpu
);
1854 /* Else cpu is not in idle, do nothing here */
1855 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1862 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1864 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1866 #endif /* CONFIG_SMP */
1868 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1870 struct rq
*rq
= cpu_rq(cpu
);
1871 struct pin_cookie cookie
;
1873 #if defined(CONFIG_SMP)
1874 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1875 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1876 ttwu_queue_remote(p
, cpu
, wake_flags
);
1881 raw_spin_lock(&rq
->lock
);
1882 cookie
= lockdep_pin_lock(&rq
->lock
);
1883 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1884 lockdep_unpin_lock(&rq
->lock
, cookie
);
1885 raw_spin_unlock(&rq
->lock
);
1889 * Notes on Program-Order guarantees on SMP systems.
1893 * The basic program-order guarantee on SMP systems is that when a task [t]
1894 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1895 * execution on its new cpu [c1].
1897 * For migration (of runnable tasks) this is provided by the following means:
1899 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1900 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1901 * rq(c1)->lock (if not at the same time, then in that order).
1902 * C) LOCK of the rq(c1)->lock scheduling in task
1904 * Transitivity guarantees that B happens after A and C after B.
1905 * Note: we only require RCpc transitivity.
1906 * Note: the cpu doing B need not be c0 or c1
1915 * UNLOCK rq(0)->lock
1917 * LOCK rq(0)->lock // orders against CPU0
1919 * UNLOCK rq(0)->lock
1923 * UNLOCK rq(1)->lock
1925 * LOCK rq(1)->lock // orders against CPU2
1928 * UNLOCK rq(1)->lock
1931 * BLOCKING -- aka. SLEEP + WAKEUP
1933 * For blocking we (obviously) need to provide the same guarantee as for
1934 * migration. However the means are completely different as there is no lock
1935 * chain to provide order. Instead we do:
1937 * 1) smp_store_release(X->on_cpu, 0)
1938 * 2) smp_cond_acquire(!X->on_cpu)
1942 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1944 * LOCK rq(0)->lock LOCK X->pi_lock
1947 * smp_store_release(X->on_cpu, 0);
1949 * smp_cond_acquire(!X->on_cpu);
1955 * X->state = RUNNING
1956 * UNLOCK rq(2)->lock
1958 * LOCK rq(2)->lock // orders against CPU1
1961 * UNLOCK rq(2)->lock
1964 * UNLOCK rq(0)->lock
1967 * However; for wakeups there is a second guarantee we must provide, namely we
1968 * must observe the state that lead to our wakeup. That is, not only must our
1969 * task observe its own prior state, it must also observe the stores prior to
1972 * This means that any means of doing remote wakeups must order the CPU doing
1973 * the wakeup against the CPU the task is going to end up running on. This,
1974 * however, is already required for the regular Program-Order guarantee above,
1975 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1980 * try_to_wake_up - wake up a thread
1981 * @p: the thread to be awakened
1982 * @state: the mask of task states that can be woken
1983 * @wake_flags: wake modifier flags (WF_*)
1985 * Put it on the run-queue if it's not already there. The "current"
1986 * thread is always on the run-queue (except when the actual
1987 * re-schedule is in progress), and as such you're allowed to do
1988 * the simpler "current->state = TASK_RUNNING" to mark yourself
1989 * runnable without the overhead of this.
1991 * Return: %true if @p was woken up, %false if it was already running.
1992 * or @state didn't match @p's state.
1995 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1997 unsigned long flags
;
1998 int cpu
, success
= 0;
2001 * If we are going to wake up a thread waiting for CONDITION we
2002 * need to ensure that CONDITION=1 done by the caller can not be
2003 * reordered with p->state check below. This pairs with mb() in
2004 * set_current_state() the waiting thread does.
2006 smp_mb__before_spinlock();
2007 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2008 if (!(p
->state
& state
))
2011 trace_sched_waking(p
);
2013 success
= 1; /* we're going to change ->state */
2016 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2021 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2022 * possible to, falsely, observe p->on_cpu == 0.
2024 * One must be running (->on_cpu == 1) in order to remove oneself
2025 * from the runqueue.
2027 * [S] ->on_cpu = 1; [L] ->on_rq
2031 * [S] ->on_rq = 0; [L] ->on_cpu
2033 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2034 * from the consecutive calls to schedule(); the first switching to our
2035 * task, the second putting it to sleep.
2040 * If the owning (remote) cpu is still in the middle of schedule() with
2041 * this task as prev, wait until its done referencing the task.
2043 * Pairs with the smp_store_release() in finish_lock_switch().
2045 * This ensures that tasks getting woken will be fully ordered against
2046 * their previous state and preserve Program Order.
2048 smp_cond_acquire(!p
->on_cpu
);
2050 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2051 p
->state
= TASK_WAKING
;
2053 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2054 if (task_cpu(p
) != cpu
) {
2055 wake_flags
|= WF_MIGRATED
;
2056 set_task_cpu(p
, cpu
);
2058 #endif /* CONFIG_SMP */
2060 ttwu_queue(p
, cpu
, wake_flags
);
2062 if (schedstat_enabled())
2063 ttwu_stat(p
, cpu
, wake_flags
);
2065 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2071 * try_to_wake_up_local - try to wake up a local task with rq lock held
2072 * @p: the thread to be awakened
2074 * Put @p on the run-queue if it's not already there. The caller must
2075 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2078 static void try_to_wake_up_local(struct task_struct
*p
, struct pin_cookie cookie
)
2080 struct rq
*rq
= task_rq(p
);
2082 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2083 WARN_ON_ONCE(p
== current
))
2086 lockdep_assert_held(&rq
->lock
);
2088 if (!raw_spin_trylock(&p
->pi_lock
)) {
2090 * This is OK, because current is on_cpu, which avoids it being
2091 * picked for load-balance and preemption/IRQs are still
2092 * disabled avoiding further scheduler activity on it and we've
2093 * not yet picked a replacement task.
2095 lockdep_unpin_lock(&rq
->lock
, cookie
);
2096 raw_spin_unlock(&rq
->lock
);
2097 raw_spin_lock(&p
->pi_lock
);
2098 raw_spin_lock(&rq
->lock
);
2099 lockdep_repin_lock(&rq
->lock
, cookie
);
2102 if (!(p
->state
& TASK_NORMAL
))
2105 trace_sched_waking(p
);
2107 if (!task_on_rq_queued(p
))
2108 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2110 ttwu_do_wakeup(rq
, p
, 0, cookie
);
2111 if (schedstat_enabled())
2112 ttwu_stat(p
, smp_processor_id(), 0);
2114 raw_spin_unlock(&p
->pi_lock
);
2118 * wake_up_process - Wake up a specific process
2119 * @p: The process to be woken up.
2121 * Attempt to wake up the nominated process and move it to the set of runnable
2124 * Return: 1 if the process was woken up, 0 if it was already running.
2126 * It may be assumed that this function implies a write memory barrier before
2127 * changing the task state if and only if any tasks are woken up.
2129 int wake_up_process(struct task_struct
*p
)
2131 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2133 EXPORT_SYMBOL(wake_up_process
);
2135 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2137 return try_to_wake_up(p
, state
, 0);
2141 * This function clears the sched_dl_entity static params.
2143 void __dl_clear_params(struct task_struct
*p
)
2145 struct sched_dl_entity
*dl_se
= &p
->dl
;
2147 dl_se
->dl_runtime
= 0;
2148 dl_se
->dl_deadline
= 0;
2149 dl_se
->dl_period
= 0;
2153 dl_se
->dl_throttled
= 0;
2154 dl_se
->dl_yielded
= 0;
2158 * Perform scheduler related setup for a newly forked process p.
2159 * p is forked by current.
2161 * __sched_fork() is basic setup used by init_idle() too:
2163 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2168 p
->se
.exec_start
= 0;
2169 p
->se
.sum_exec_runtime
= 0;
2170 p
->se
.prev_sum_exec_runtime
= 0;
2171 p
->se
.nr_migrations
= 0;
2173 INIT_LIST_HEAD(&p
->se
.group_node
);
2175 #ifdef CONFIG_FAIR_GROUP_SCHED
2176 p
->se
.cfs_rq
= NULL
;
2179 #ifdef CONFIG_SCHEDSTATS
2180 /* Even if schedstat is disabled, there should not be garbage */
2181 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2184 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2185 init_dl_task_timer(&p
->dl
);
2186 __dl_clear_params(p
);
2188 INIT_LIST_HEAD(&p
->rt
.run_list
);
2190 p
->rt
.time_slice
= sched_rr_timeslice
;
2194 #ifdef CONFIG_PREEMPT_NOTIFIERS
2195 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2198 #ifdef CONFIG_NUMA_BALANCING
2199 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2200 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2201 p
->mm
->numa_scan_seq
= 0;
2204 if (clone_flags
& CLONE_VM
)
2205 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2207 p
->numa_preferred_nid
= -1;
2209 p
->node_stamp
= 0ULL;
2210 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2211 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2212 p
->numa_work
.next
= &p
->numa_work
;
2213 p
->numa_faults
= NULL
;
2214 p
->last_task_numa_placement
= 0;
2215 p
->last_sum_exec_runtime
= 0;
2217 p
->numa_group
= NULL
;
2218 #endif /* CONFIG_NUMA_BALANCING */
2221 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2223 #ifdef CONFIG_NUMA_BALANCING
2225 void set_numabalancing_state(bool enabled
)
2228 static_branch_enable(&sched_numa_balancing
);
2230 static_branch_disable(&sched_numa_balancing
);
2233 #ifdef CONFIG_PROC_SYSCTL
2234 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2235 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2239 int state
= static_branch_likely(&sched_numa_balancing
);
2241 if (write
&& !capable(CAP_SYS_ADMIN
))
2246 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2250 set_numabalancing_state(state
);
2256 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2258 #ifdef CONFIG_SCHEDSTATS
2259 static void set_schedstats(bool enabled
)
2262 static_branch_enable(&sched_schedstats
);
2264 static_branch_disable(&sched_schedstats
);
2267 void force_schedstat_enabled(void)
2269 if (!schedstat_enabled()) {
2270 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2271 static_branch_enable(&sched_schedstats
);
2275 static int __init
setup_schedstats(char *str
)
2281 if (!strcmp(str
, "enable")) {
2282 set_schedstats(true);
2284 } else if (!strcmp(str
, "disable")) {
2285 set_schedstats(false);
2290 pr_warn("Unable to parse schedstats=\n");
2294 __setup("schedstats=", setup_schedstats
);
2296 #ifdef CONFIG_PROC_SYSCTL
2297 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2298 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2302 int state
= static_branch_likely(&sched_schedstats
);
2304 if (write
&& !capable(CAP_SYS_ADMIN
))
2309 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2313 set_schedstats(state
);
2320 * fork()/clone()-time setup:
2322 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2324 unsigned long flags
;
2325 int cpu
= get_cpu();
2327 __sched_fork(clone_flags
, p
);
2329 * We mark the process as running here. This guarantees that
2330 * nobody will actually run it, and a signal or other external
2331 * event cannot wake it up and insert it on the runqueue either.
2333 p
->state
= TASK_RUNNING
;
2336 * Make sure we do not leak PI boosting priority to the child.
2338 p
->prio
= current
->normal_prio
;
2341 * Revert to default priority/policy on fork if requested.
2343 if (unlikely(p
->sched_reset_on_fork
)) {
2344 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2345 p
->policy
= SCHED_NORMAL
;
2346 p
->static_prio
= NICE_TO_PRIO(0);
2348 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2349 p
->static_prio
= NICE_TO_PRIO(0);
2351 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2355 * We don't need the reset flag anymore after the fork. It has
2356 * fulfilled its duty:
2358 p
->sched_reset_on_fork
= 0;
2361 if (dl_prio(p
->prio
)) {
2364 } else if (rt_prio(p
->prio
)) {
2365 p
->sched_class
= &rt_sched_class
;
2367 p
->sched_class
= &fair_sched_class
;
2370 if (p
->sched_class
->task_fork
)
2371 p
->sched_class
->task_fork(p
);
2374 * The child is not yet in the pid-hash so no cgroup attach races,
2375 * and the cgroup is pinned to this child due to cgroup_fork()
2376 * is ran before sched_fork().
2378 * Silence PROVE_RCU.
2380 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2381 set_task_cpu(p
, cpu
);
2382 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2384 #ifdef CONFIG_SCHED_INFO
2385 if (likely(sched_info_on()))
2386 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2388 #if defined(CONFIG_SMP)
2391 init_task_preempt_count(p
);
2393 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2394 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2401 unsigned long to_ratio(u64 period
, u64 runtime
)
2403 if (runtime
== RUNTIME_INF
)
2407 * Doing this here saves a lot of checks in all
2408 * the calling paths, and returning zero seems
2409 * safe for them anyway.
2414 return div64_u64(runtime
<< 20, period
);
2418 inline struct dl_bw
*dl_bw_of(int i
)
2420 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2421 "sched RCU must be held");
2422 return &cpu_rq(i
)->rd
->dl_bw
;
2425 static inline int dl_bw_cpus(int i
)
2427 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2430 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2431 "sched RCU must be held");
2432 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2438 inline struct dl_bw
*dl_bw_of(int i
)
2440 return &cpu_rq(i
)->dl
.dl_bw
;
2443 static inline int dl_bw_cpus(int i
)
2450 * We must be sure that accepting a new task (or allowing changing the
2451 * parameters of an existing one) is consistent with the bandwidth
2452 * constraints. If yes, this function also accordingly updates the currently
2453 * allocated bandwidth to reflect the new situation.
2455 * This function is called while holding p's rq->lock.
2457 * XXX we should delay bw change until the task's 0-lag point, see
2460 static int dl_overflow(struct task_struct
*p
, int policy
,
2461 const struct sched_attr
*attr
)
2464 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2465 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2466 u64 runtime
= attr
->sched_runtime
;
2467 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2470 /* !deadline task may carry old deadline bandwidth */
2471 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2475 * Either if a task, enters, leave, or stays -deadline but changes
2476 * its parameters, we may need to update accordingly the total
2477 * allocated bandwidth of the container.
2479 raw_spin_lock(&dl_b
->lock
);
2480 cpus
= dl_bw_cpus(task_cpu(p
));
2481 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2482 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2483 __dl_add(dl_b
, new_bw
);
2485 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2486 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2487 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2488 __dl_add(dl_b
, new_bw
);
2490 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2491 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2494 raw_spin_unlock(&dl_b
->lock
);
2499 extern void init_dl_bw(struct dl_bw
*dl_b
);
2502 * wake_up_new_task - wake up a newly created task for the first time.
2504 * This function will do some initial scheduler statistics housekeeping
2505 * that must be done for every newly created context, then puts the task
2506 * on the runqueue and wakes it.
2508 void wake_up_new_task(struct task_struct
*p
)
2513 /* Initialize new task's runnable average */
2514 init_entity_runnable_average(&p
->se
);
2515 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2518 * Fork balancing, do it here and not earlier because:
2519 * - cpus_allowed can change in the fork path
2520 * - any previously selected cpu might disappear through hotplug
2522 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2524 /* Post initialize new task's util average when its cfs_rq is set */
2525 post_init_entity_util_avg(&p
->se
);
2527 rq
= __task_rq_lock(p
, &rf
);
2528 activate_task(rq
, p
, 0);
2529 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2530 trace_sched_wakeup_new(p
);
2531 check_preempt_curr(rq
, p
, WF_FORK
);
2533 if (p
->sched_class
->task_woken
) {
2535 * Nothing relies on rq->lock after this, so its fine to
2538 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
2539 p
->sched_class
->task_woken(rq
, p
);
2540 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
2543 task_rq_unlock(rq
, p
, &rf
);
2546 #ifdef CONFIG_PREEMPT_NOTIFIERS
2548 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2550 void preempt_notifier_inc(void)
2552 static_key_slow_inc(&preempt_notifier_key
);
2554 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2556 void preempt_notifier_dec(void)
2558 static_key_slow_dec(&preempt_notifier_key
);
2560 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2563 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2564 * @notifier: notifier struct to register
2566 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2568 if (!static_key_false(&preempt_notifier_key
))
2569 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2571 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2573 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2576 * preempt_notifier_unregister - no longer interested in preemption notifications
2577 * @notifier: notifier struct to unregister
2579 * This is *not* safe to call from within a preemption notifier.
2581 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2583 hlist_del(¬ifier
->link
);
2585 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2587 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2589 struct preempt_notifier
*notifier
;
2591 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2592 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2595 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2597 if (static_key_false(&preempt_notifier_key
))
2598 __fire_sched_in_preempt_notifiers(curr
);
2602 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2603 struct task_struct
*next
)
2605 struct preempt_notifier
*notifier
;
2607 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2608 notifier
->ops
->sched_out(notifier
, next
);
2611 static __always_inline
void
2612 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2613 struct task_struct
*next
)
2615 if (static_key_false(&preempt_notifier_key
))
2616 __fire_sched_out_preempt_notifiers(curr
, next
);
2619 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2621 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2626 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2627 struct task_struct
*next
)
2631 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2634 * prepare_task_switch - prepare to switch tasks
2635 * @rq: the runqueue preparing to switch
2636 * @prev: the current task that is being switched out
2637 * @next: the task we are going to switch to.
2639 * This is called with the rq lock held and interrupts off. It must
2640 * be paired with a subsequent finish_task_switch after the context
2643 * prepare_task_switch sets up locking and calls architecture specific
2647 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2648 struct task_struct
*next
)
2650 sched_info_switch(rq
, prev
, next
);
2651 perf_event_task_sched_out(prev
, next
);
2652 fire_sched_out_preempt_notifiers(prev
, next
);
2653 prepare_lock_switch(rq
, next
);
2654 prepare_arch_switch(next
);
2658 * finish_task_switch - clean up after a task-switch
2659 * @prev: the thread we just switched away from.
2661 * finish_task_switch must be called after the context switch, paired
2662 * with a prepare_task_switch call before the context switch.
2663 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2664 * and do any other architecture-specific cleanup actions.
2666 * Note that we may have delayed dropping an mm in context_switch(). If
2667 * so, we finish that here outside of the runqueue lock. (Doing it
2668 * with the lock held can cause deadlocks; see schedule() for
2671 * The context switch have flipped the stack from under us and restored the
2672 * local variables which were saved when this task called schedule() in the
2673 * past. prev == current is still correct but we need to recalculate this_rq
2674 * because prev may have moved to another CPU.
2676 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2677 __releases(rq
->lock
)
2679 struct rq
*rq
= this_rq();
2680 struct mm_struct
*mm
= rq
->prev_mm
;
2684 * The previous task will have left us with a preempt_count of 2
2685 * because it left us after:
2688 * preempt_disable(); // 1
2690 * raw_spin_lock_irq(&rq->lock) // 2
2692 * Also, see FORK_PREEMPT_COUNT.
2694 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2695 "corrupted preempt_count: %s/%d/0x%x\n",
2696 current
->comm
, current
->pid
, preempt_count()))
2697 preempt_count_set(FORK_PREEMPT_COUNT
);
2702 * A task struct has one reference for the use as "current".
2703 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2704 * schedule one last time. The schedule call will never return, and
2705 * the scheduled task must drop that reference.
2707 * We must observe prev->state before clearing prev->on_cpu (in
2708 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2709 * running on another CPU and we could rave with its RUNNING -> DEAD
2710 * transition, resulting in a double drop.
2712 prev_state
= prev
->state
;
2713 vtime_task_switch(prev
);
2714 perf_event_task_sched_in(prev
, current
);
2715 finish_lock_switch(rq
, prev
);
2716 finish_arch_post_lock_switch();
2718 fire_sched_in_preempt_notifiers(current
);
2721 if (unlikely(prev_state
== TASK_DEAD
)) {
2722 if (prev
->sched_class
->task_dead
)
2723 prev
->sched_class
->task_dead(prev
);
2726 * Remove function-return probe instances associated with this
2727 * task and put them back on the free list.
2729 kprobe_flush_task(prev
);
2730 put_task_struct(prev
);
2733 tick_nohz_task_switch();
2739 /* rq->lock is NOT held, but preemption is disabled */
2740 static void __balance_callback(struct rq
*rq
)
2742 struct callback_head
*head
, *next
;
2743 void (*func
)(struct rq
*rq
);
2744 unsigned long flags
;
2746 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2747 head
= rq
->balance_callback
;
2748 rq
->balance_callback
= NULL
;
2750 func
= (void (*)(struct rq
*))head
->func
;
2757 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2760 static inline void balance_callback(struct rq
*rq
)
2762 if (unlikely(rq
->balance_callback
))
2763 __balance_callback(rq
);
2768 static inline void balance_callback(struct rq
*rq
)
2775 * schedule_tail - first thing a freshly forked thread must call.
2776 * @prev: the thread we just switched away from.
2778 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2779 __releases(rq
->lock
)
2784 * New tasks start with FORK_PREEMPT_COUNT, see there and
2785 * finish_task_switch() for details.
2787 * finish_task_switch() will drop rq->lock() and lower preempt_count
2788 * and the preempt_enable() will end up enabling preemption (on
2789 * PREEMPT_COUNT kernels).
2792 rq
= finish_task_switch(prev
);
2793 balance_callback(rq
);
2796 if (current
->set_child_tid
)
2797 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2801 * context_switch - switch to the new MM and the new thread's register state.
2803 static __always_inline
struct rq
*
2804 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2805 struct task_struct
*next
, struct pin_cookie cookie
)
2807 struct mm_struct
*mm
, *oldmm
;
2809 prepare_task_switch(rq
, prev
, next
);
2812 oldmm
= prev
->active_mm
;
2814 * For paravirt, this is coupled with an exit in switch_to to
2815 * combine the page table reload and the switch backend into
2818 arch_start_context_switch(prev
);
2821 next
->active_mm
= oldmm
;
2822 atomic_inc(&oldmm
->mm_count
);
2823 enter_lazy_tlb(oldmm
, next
);
2825 switch_mm_irqs_off(oldmm
, mm
, next
);
2828 prev
->active_mm
= NULL
;
2829 rq
->prev_mm
= oldmm
;
2832 * Since the runqueue lock will be released by the next
2833 * task (which is an invalid locking op but in the case
2834 * of the scheduler it's an obvious special-case), so we
2835 * do an early lockdep release here:
2837 lockdep_unpin_lock(&rq
->lock
, cookie
);
2838 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2840 /* Here we just switch the register state and the stack. */
2841 switch_to(prev
, next
, prev
);
2844 return finish_task_switch(prev
);
2848 * nr_running and nr_context_switches:
2850 * externally visible scheduler statistics: current number of runnable
2851 * threads, total number of context switches performed since bootup.
2853 unsigned long nr_running(void)
2855 unsigned long i
, sum
= 0;
2857 for_each_online_cpu(i
)
2858 sum
+= cpu_rq(i
)->nr_running
;
2864 * Check if only the current task is running on the cpu.
2866 * Caution: this function does not check that the caller has disabled
2867 * preemption, thus the result might have a time-of-check-to-time-of-use
2868 * race. The caller is responsible to use it correctly, for example:
2870 * - from a non-preemptable section (of course)
2872 * - from a thread that is bound to a single CPU
2874 * - in a loop with very short iterations (e.g. a polling loop)
2876 bool single_task_running(void)
2878 return raw_rq()->nr_running
== 1;
2880 EXPORT_SYMBOL(single_task_running
);
2882 unsigned long long nr_context_switches(void)
2885 unsigned long long sum
= 0;
2887 for_each_possible_cpu(i
)
2888 sum
+= cpu_rq(i
)->nr_switches
;
2893 unsigned long nr_iowait(void)
2895 unsigned long i
, sum
= 0;
2897 for_each_possible_cpu(i
)
2898 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2903 unsigned long nr_iowait_cpu(int cpu
)
2905 struct rq
*this = cpu_rq(cpu
);
2906 return atomic_read(&this->nr_iowait
);
2909 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2911 struct rq
*rq
= this_rq();
2912 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2913 *load
= rq
->load
.weight
;
2919 * sched_exec - execve() is a valuable balancing opportunity, because at
2920 * this point the task has the smallest effective memory and cache footprint.
2922 void sched_exec(void)
2924 struct task_struct
*p
= current
;
2925 unsigned long flags
;
2928 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2929 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2930 if (dest_cpu
== smp_processor_id())
2933 if (likely(cpu_active(dest_cpu
))) {
2934 struct migration_arg arg
= { p
, dest_cpu
};
2936 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2937 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2941 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2946 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2947 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2949 EXPORT_PER_CPU_SYMBOL(kstat
);
2950 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2953 * Return accounted runtime for the task.
2954 * In case the task is currently running, return the runtime plus current's
2955 * pending runtime that have not been accounted yet.
2957 unsigned long long task_sched_runtime(struct task_struct
*p
)
2963 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2965 * 64-bit doesn't need locks to atomically read a 64bit value.
2966 * So we have a optimization chance when the task's delta_exec is 0.
2967 * Reading ->on_cpu is racy, but this is ok.
2969 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2970 * If we race with it entering cpu, unaccounted time is 0. This is
2971 * indistinguishable from the read occurring a few cycles earlier.
2972 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2973 * been accounted, so we're correct here as well.
2975 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2976 return p
->se
.sum_exec_runtime
;
2979 rq
= task_rq_lock(p
, &rf
);
2981 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2982 * project cycles that may never be accounted to this
2983 * thread, breaking clock_gettime().
2985 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2986 update_rq_clock(rq
);
2987 p
->sched_class
->update_curr(rq
);
2989 ns
= p
->se
.sum_exec_runtime
;
2990 task_rq_unlock(rq
, p
, &rf
);
2996 * This function gets called by the timer code, with HZ frequency.
2997 * We call it with interrupts disabled.
2999 void scheduler_tick(void)
3001 int cpu
= smp_processor_id();
3002 struct rq
*rq
= cpu_rq(cpu
);
3003 struct task_struct
*curr
= rq
->curr
;
3007 raw_spin_lock(&rq
->lock
);
3008 update_rq_clock(rq
);
3009 curr
->sched_class
->task_tick(rq
, curr
, 0);
3010 cpu_load_update_active(rq
);
3011 calc_global_load_tick(rq
);
3012 raw_spin_unlock(&rq
->lock
);
3014 perf_event_task_tick();
3017 rq
->idle_balance
= idle_cpu(cpu
);
3018 trigger_load_balance(rq
);
3020 rq_last_tick_reset(rq
);
3023 #ifdef CONFIG_NO_HZ_FULL
3025 * scheduler_tick_max_deferment
3027 * Keep at least one tick per second when a single
3028 * active task is running because the scheduler doesn't
3029 * yet completely support full dynticks environment.
3031 * This makes sure that uptime, CFS vruntime, load
3032 * balancing, etc... continue to move forward, even
3033 * with a very low granularity.
3035 * Return: Maximum deferment in nanoseconds.
3037 u64
scheduler_tick_max_deferment(void)
3039 struct rq
*rq
= this_rq();
3040 unsigned long next
, now
= READ_ONCE(jiffies
);
3042 next
= rq
->last_sched_tick
+ HZ
;
3044 if (time_before_eq(next
, now
))
3047 return jiffies_to_nsecs(next
- now
);
3051 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3052 defined(CONFIG_PREEMPT_TRACER))
3054 * If the value passed in is equal to the current preempt count
3055 * then we just disabled preemption. Start timing the latency.
3057 static inline void preempt_latency_start(int val
)
3059 if (preempt_count() == val
) {
3060 unsigned long ip
= get_lock_parent_ip();
3061 #ifdef CONFIG_DEBUG_PREEMPT
3062 current
->preempt_disable_ip
= ip
;
3064 trace_preempt_off(CALLER_ADDR0
, ip
);
3068 void preempt_count_add(int val
)
3070 #ifdef CONFIG_DEBUG_PREEMPT
3074 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3077 __preempt_count_add(val
);
3078 #ifdef CONFIG_DEBUG_PREEMPT
3080 * Spinlock count overflowing soon?
3082 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3085 preempt_latency_start(val
);
3087 EXPORT_SYMBOL(preempt_count_add
);
3088 NOKPROBE_SYMBOL(preempt_count_add
);
3091 * If the value passed in equals to the current preempt count
3092 * then we just enabled preemption. Stop timing the latency.
3094 static inline void preempt_latency_stop(int val
)
3096 if (preempt_count() == val
)
3097 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3100 void preempt_count_sub(int val
)
3102 #ifdef CONFIG_DEBUG_PREEMPT
3106 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3109 * Is the spinlock portion underflowing?
3111 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3112 !(preempt_count() & PREEMPT_MASK
)))
3116 preempt_latency_stop(val
);
3117 __preempt_count_sub(val
);
3119 EXPORT_SYMBOL(preempt_count_sub
);
3120 NOKPROBE_SYMBOL(preempt_count_sub
);
3123 static inline void preempt_latency_start(int val
) { }
3124 static inline void preempt_latency_stop(int val
) { }
3128 * Print scheduling while atomic bug:
3130 static noinline
void __schedule_bug(struct task_struct
*prev
)
3132 if (oops_in_progress
)
3135 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3136 prev
->comm
, prev
->pid
, preempt_count());
3138 debug_show_held_locks(prev
);
3140 if (irqs_disabled())
3141 print_irqtrace_events(prev
);
3142 #ifdef CONFIG_DEBUG_PREEMPT
3143 if (in_atomic_preempt_off()) {
3144 pr_err("Preemption disabled at:");
3145 print_ip_sym(current
->preempt_disable_ip
);
3150 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3154 * Various schedule()-time debugging checks and statistics:
3156 static inline void schedule_debug(struct task_struct
*prev
)
3158 #ifdef CONFIG_SCHED_STACK_END_CHECK
3159 BUG_ON(task_stack_end_corrupted(prev
));
3162 if (unlikely(in_atomic_preempt_off())) {
3163 __schedule_bug(prev
);
3164 preempt_count_set(PREEMPT_DISABLED
);
3168 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3170 schedstat_inc(this_rq(), sched_count
);
3174 * Pick up the highest-prio task:
3176 static inline struct task_struct
*
3177 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
3179 const struct sched_class
*class = &fair_sched_class
;
3180 struct task_struct
*p
;
3183 * Optimization: we know that if all tasks are in
3184 * the fair class we can call that function directly:
3186 if (likely(prev
->sched_class
== class &&
3187 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3188 p
= fair_sched_class
.pick_next_task(rq
, prev
, cookie
);
3189 if (unlikely(p
== RETRY_TASK
))
3192 /* assumes fair_sched_class->next == idle_sched_class */
3194 p
= idle_sched_class
.pick_next_task(rq
, prev
, cookie
);
3200 for_each_class(class) {
3201 p
= class->pick_next_task(rq
, prev
, cookie
);
3203 if (unlikely(p
== RETRY_TASK
))
3209 BUG(); /* the idle class will always have a runnable task */
3213 * __schedule() is the main scheduler function.
3215 * The main means of driving the scheduler and thus entering this function are:
3217 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3219 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3220 * paths. For example, see arch/x86/entry_64.S.
3222 * To drive preemption between tasks, the scheduler sets the flag in timer
3223 * interrupt handler scheduler_tick().
3225 * 3. Wakeups don't really cause entry into schedule(). They add a
3226 * task to the run-queue and that's it.
3228 * Now, if the new task added to the run-queue preempts the current
3229 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3230 * called on the nearest possible occasion:
3232 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3234 * - in syscall or exception context, at the next outmost
3235 * preempt_enable(). (this might be as soon as the wake_up()'s
3238 * - in IRQ context, return from interrupt-handler to
3239 * preemptible context
3241 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3244 * - cond_resched() call
3245 * - explicit schedule() call
3246 * - return from syscall or exception to user-space
3247 * - return from interrupt-handler to user-space
3249 * WARNING: must be called with preemption disabled!
3251 static void __sched notrace
__schedule(bool preempt
)
3253 struct task_struct
*prev
, *next
;
3254 unsigned long *switch_count
;
3255 struct pin_cookie cookie
;
3259 cpu
= smp_processor_id();
3264 * do_exit() calls schedule() with preemption disabled as an exception;
3265 * however we must fix that up, otherwise the next task will see an
3266 * inconsistent (higher) preempt count.
3268 * It also avoids the below schedule_debug() test from complaining
3271 if (unlikely(prev
->state
== TASK_DEAD
))
3272 preempt_enable_no_resched_notrace();
3274 schedule_debug(prev
);
3276 if (sched_feat(HRTICK
))
3279 local_irq_disable();
3280 rcu_note_context_switch();
3283 * Make sure that signal_pending_state()->signal_pending() below
3284 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3285 * done by the caller to avoid the race with signal_wake_up().
3287 smp_mb__before_spinlock();
3288 raw_spin_lock(&rq
->lock
);
3289 cookie
= lockdep_pin_lock(&rq
->lock
);
3291 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3293 switch_count
= &prev
->nivcsw
;
3294 if (!preempt
&& prev
->state
) {
3295 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3296 prev
->state
= TASK_RUNNING
;
3298 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3302 * If a worker went to sleep, notify and ask workqueue
3303 * whether it wants to wake up a task to maintain
3306 if (prev
->flags
& PF_WQ_WORKER
) {
3307 struct task_struct
*to_wakeup
;
3309 to_wakeup
= wq_worker_sleeping(prev
);
3311 try_to_wake_up_local(to_wakeup
, cookie
);
3314 switch_count
= &prev
->nvcsw
;
3317 if (task_on_rq_queued(prev
))
3318 update_rq_clock(rq
);
3320 next
= pick_next_task(rq
, prev
, cookie
);
3321 clear_tsk_need_resched(prev
);
3322 clear_preempt_need_resched();
3323 rq
->clock_skip_update
= 0;
3325 if (likely(prev
!= next
)) {
3330 trace_sched_switch(preempt
, prev
, next
);
3331 rq
= context_switch(rq
, prev
, next
, cookie
); /* unlocks the rq */
3333 lockdep_unpin_lock(&rq
->lock
, cookie
);
3334 raw_spin_unlock_irq(&rq
->lock
);
3337 balance_callback(rq
);
3339 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3341 static inline void sched_submit_work(struct task_struct
*tsk
)
3343 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3346 * If we are going to sleep and we have plugged IO queued,
3347 * make sure to submit it to avoid deadlocks.
3349 if (blk_needs_flush_plug(tsk
))
3350 blk_schedule_flush_plug(tsk
);
3353 asmlinkage __visible
void __sched
schedule(void)
3355 struct task_struct
*tsk
= current
;
3357 sched_submit_work(tsk
);
3361 sched_preempt_enable_no_resched();
3362 } while (need_resched());
3364 EXPORT_SYMBOL(schedule
);
3366 #ifdef CONFIG_CONTEXT_TRACKING
3367 asmlinkage __visible
void __sched
schedule_user(void)
3370 * If we come here after a random call to set_need_resched(),
3371 * or we have been woken up remotely but the IPI has not yet arrived,
3372 * we haven't yet exited the RCU idle mode. Do it here manually until
3373 * we find a better solution.
3375 * NB: There are buggy callers of this function. Ideally we
3376 * should warn if prev_state != CONTEXT_USER, but that will trigger
3377 * too frequently to make sense yet.
3379 enum ctx_state prev_state
= exception_enter();
3381 exception_exit(prev_state
);
3386 * schedule_preempt_disabled - called with preemption disabled
3388 * Returns with preemption disabled. Note: preempt_count must be 1
3390 void __sched
schedule_preempt_disabled(void)
3392 sched_preempt_enable_no_resched();
3397 static void __sched notrace
preempt_schedule_common(void)
3401 * Because the function tracer can trace preempt_count_sub()
3402 * and it also uses preempt_enable/disable_notrace(), if
3403 * NEED_RESCHED is set, the preempt_enable_notrace() called
3404 * by the function tracer will call this function again and
3405 * cause infinite recursion.
3407 * Preemption must be disabled here before the function
3408 * tracer can trace. Break up preempt_disable() into two
3409 * calls. One to disable preemption without fear of being
3410 * traced. The other to still record the preemption latency,
3411 * which can also be traced by the function tracer.
3413 preempt_disable_notrace();
3414 preempt_latency_start(1);
3416 preempt_latency_stop(1);
3417 preempt_enable_no_resched_notrace();
3420 * Check again in case we missed a preemption opportunity
3421 * between schedule and now.
3423 } while (need_resched());
3426 #ifdef CONFIG_PREEMPT
3428 * this is the entry point to schedule() from in-kernel preemption
3429 * off of preempt_enable. Kernel preemptions off return from interrupt
3430 * occur there and call schedule directly.
3432 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3435 * If there is a non-zero preempt_count or interrupts are disabled,
3436 * we do not want to preempt the current task. Just return..
3438 if (likely(!preemptible()))
3441 preempt_schedule_common();
3443 NOKPROBE_SYMBOL(preempt_schedule
);
3444 EXPORT_SYMBOL(preempt_schedule
);
3447 * preempt_schedule_notrace - preempt_schedule called by tracing
3449 * The tracing infrastructure uses preempt_enable_notrace to prevent
3450 * recursion and tracing preempt enabling caused by the tracing
3451 * infrastructure itself. But as tracing can happen in areas coming
3452 * from userspace or just about to enter userspace, a preempt enable
3453 * can occur before user_exit() is called. This will cause the scheduler
3454 * to be called when the system is still in usermode.
3456 * To prevent this, the preempt_enable_notrace will use this function
3457 * instead of preempt_schedule() to exit user context if needed before
3458 * calling the scheduler.
3460 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3462 enum ctx_state prev_ctx
;
3464 if (likely(!preemptible()))
3469 * Because the function tracer can trace preempt_count_sub()
3470 * and it also uses preempt_enable/disable_notrace(), if
3471 * NEED_RESCHED is set, the preempt_enable_notrace() called
3472 * by the function tracer will call this function again and
3473 * cause infinite recursion.
3475 * Preemption must be disabled here before the function
3476 * tracer can trace. Break up preempt_disable() into two
3477 * calls. One to disable preemption without fear of being
3478 * traced. The other to still record the preemption latency,
3479 * which can also be traced by the function tracer.
3481 preempt_disable_notrace();
3482 preempt_latency_start(1);
3484 * Needs preempt disabled in case user_exit() is traced
3485 * and the tracer calls preempt_enable_notrace() causing
3486 * an infinite recursion.
3488 prev_ctx
= exception_enter();
3490 exception_exit(prev_ctx
);
3492 preempt_latency_stop(1);
3493 preempt_enable_no_resched_notrace();
3494 } while (need_resched());
3496 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3498 #endif /* CONFIG_PREEMPT */
3501 * this is the entry point to schedule() from kernel preemption
3502 * off of irq context.
3503 * Note, that this is called and return with irqs disabled. This will
3504 * protect us against recursive calling from irq.
3506 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3508 enum ctx_state prev_state
;
3510 /* Catch callers which need to be fixed */
3511 BUG_ON(preempt_count() || !irqs_disabled());
3513 prev_state
= exception_enter();
3519 local_irq_disable();
3520 sched_preempt_enable_no_resched();
3521 } while (need_resched());
3523 exception_exit(prev_state
);
3526 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3529 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3531 EXPORT_SYMBOL(default_wake_function
);
3533 #ifdef CONFIG_RT_MUTEXES
3536 * rt_mutex_setprio - set the current priority of a task
3538 * @prio: prio value (kernel-internal form)
3540 * This function changes the 'effective' priority of a task. It does
3541 * not touch ->normal_prio like __setscheduler().
3543 * Used by the rt_mutex code to implement priority inheritance
3544 * logic. Call site only calls if the priority of the task changed.
3546 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3548 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3549 const struct sched_class
*prev_class
;
3553 BUG_ON(prio
> MAX_PRIO
);
3555 rq
= __task_rq_lock(p
, &rf
);
3558 * Idle task boosting is a nono in general. There is one
3559 * exception, when PREEMPT_RT and NOHZ is active:
3561 * The idle task calls get_next_timer_interrupt() and holds
3562 * the timer wheel base->lock on the CPU and another CPU wants
3563 * to access the timer (probably to cancel it). We can safely
3564 * ignore the boosting request, as the idle CPU runs this code
3565 * with interrupts disabled and will complete the lock
3566 * protected section without being interrupted. So there is no
3567 * real need to boost.
3569 if (unlikely(p
== rq
->idle
)) {
3570 WARN_ON(p
!= rq
->curr
);
3571 WARN_ON(p
->pi_blocked_on
);
3575 trace_sched_pi_setprio(p
, prio
);
3578 if (oldprio
== prio
)
3579 queue_flag
&= ~DEQUEUE_MOVE
;
3581 prev_class
= p
->sched_class
;
3582 queued
= task_on_rq_queued(p
);
3583 running
= task_current(rq
, p
);
3585 dequeue_task(rq
, p
, queue_flag
);
3587 put_prev_task(rq
, p
);
3590 * Boosting condition are:
3591 * 1. -rt task is running and holds mutex A
3592 * --> -dl task blocks on mutex A
3594 * 2. -dl task is running and holds mutex A
3595 * --> -dl task blocks on mutex A and could preempt the
3598 if (dl_prio(prio
)) {
3599 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3600 if (!dl_prio(p
->normal_prio
) ||
3601 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3602 p
->dl
.dl_boosted
= 1;
3603 queue_flag
|= ENQUEUE_REPLENISH
;
3605 p
->dl
.dl_boosted
= 0;
3606 p
->sched_class
= &dl_sched_class
;
3607 } else if (rt_prio(prio
)) {
3608 if (dl_prio(oldprio
))
3609 p
->dl
.dl_boosted
= 0;
3611 queue_flag
|= ENQUEUE_HEAD
;
3612 p
->sched_class
= &rt_sched_class
;
3614 if (dl_prio(oldprio
))
3615 p
->dl
.dl_boosted
= 0;
3616 if (rt_prio(oldprio
))
3618 p
->sched_class
= &fair_sched_class
;
3624 p
->sched_class
->set_curr_task(rq
);
3626 enqueue_task(rq
, p
, queue_flag
);
3628 check_class_changed(rq
, p
, prev_class
, oldprio
);
3630 preempt_disable(); /* avoid rq from going away on us */
3631 __task_rq_unlock(rq
, &rf
);
3633 balance_callback(rq
);
3638 void set_user_nice(struct task_struct
*p
, long nice
)
3640 int old_prio
, delta
, queued
;
3644 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3647 * We have to be careful, if called from sys_setpriority(),
3648 * the task might be in the middle of scheduling on another CPU.
3650 rq
= task_rq_lock(p
, &rf
);
3652 * The RT priorities are set via sched_setscheduler(), but we still
3653 * allow the 'normal' nice value to be set - but as expected
3654 * it wont have any effect on scheduling until the task is
3655 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3657 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3658 p
->static_prio
= NICE_TO_PRIO(nice
);
3661 queued
= task_on_rq_queued(p
);
3663 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3665 p
->static_prio
= NICE_TO_PRIO(nice
);
3668 p
->prio
= effective_prio(p
);
3669 delta
= p
->prio
- old_prio
;
3672 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3674 * If the task increased its priority or is running and
3675 * lowered its priority, then reschedule its CPU:
3677 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3681 task_rq_unlock(rq
, p
, &rf
);
3683 EXPORT_SYMBOL(set_user_nice
);
3686 * can_nice - check if a task can reduce its nice value
3690 int can_nice(const struct task_struct
*p
, const int nice
)
3692 /* convert nice value [19,-20] to rlimit style value [1,40] */
3693 int nice_rlim
= nice_to_rlimit(nice
);
3695 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3696 capable(CAP_SYS_NICE
));
3699 #ifdef __ARCH_WANT_SYS_NICE
3702 * sys_nice - change the priority of the current process.
3703 * @increment: priority increment
3705 * sys_setpriority is a more generic, but much slower function that
3706 * does similar things.
3708 SYSCALL_DEFINE1(nice
, int, increment
)
3713 * Setpriority might change our priority at the same moment.
3714 * We don't have to worry. Conceptually one call occurs first
3715 * and we have a single winner.
3717 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3718 nice
= task_nice(current
) + increment
;
3720 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3721 if (increment
< 0 && !can_nice(current
, nice
))
3724 retval
= security_task_setnice(current
, nice
);
3728 set_user_nice(current
, nice
);
3735 * task_prio - return the priority value of a given task.
3736 * @p: the task in question.
3738 * Return: The priority value as seen by users in /proc.
3739 * RT tasks are offset by -200. Normal tasks are centered
3740 * around 0, value goes from -16 to +15.
3742 int task_prio(const struct task_struct
*p
)
3744 return p
->prio
- MAX_RT_PRIO
;
3748 * idle_cpu - is a given cpu idle currently?
3749 * @cpu: the processor in question.
3751 * Return: 1 if the CPU is currently idle. 0 otherwise.
3753 int idle_cpu(int cpu
)
3755 struct rq
*rq
= cpu_rq(cpu
);
3757 if (rq
->curr
!= rq
->idle
)
3764 if (!llist_empty(&rq
->wake_list
))
3772 * idle_task - return the idle task for a given cpu.
3773 * @cpu: the processor in question.
3775 * Return: The idle task for the cpu @cpu.
3777 struct task_struct
*idle_task(int cpu
)
3779 return cpu_rq(cpu
)->idle
;
3783 * find_process_by_pid - find a process with a matching PID value.
3784 * @pid: the pid in question.
3786 * The task of @pid, if found. %NULL otherwise.
3788 static struct task_struct
*find_process_by_pid(pid_t pid
)
3790 return pid
? find_task_by_vpid(pid
) : current
;
3794 * This function initializes the sched_dl_entity of a newly becoming
3795 * SCHED_DEADLINE task.
3797 * Only the static values are considered here, the actual runtime and the
3798 * absolute deadline will be properly calculated when the task is enqueued
3799 * for the first time with its new policy.
3802 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3804 struct sched_dl_entity
*dl_se
= &p
->dl
;
3806 dl_se
->dl_runtime
= attr
->sched_runtime
;
3807 dl_se
->dl_deadline
= attr
->sched_deadline
;
3808 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3809 dl_se
->flags
= attr
->sched_flags
;
3810 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3813 * Changing the parameters of a task is 'tricky' and we're not doing
3814 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3816 * What we SHOULD do is delay the bandwidth release until the 0-lag
3817 * point. This would include retaining the task_struct until that time
3818 * and change dl_overflow() to not immediately decrement the current
3821 * Instead we retain the current runtime/deadline and let the new
3822 * parameters take effect after the current reservation period lapses.
3823 * This is safe (albeit pessimistic) because the 0-lag point is always
3824 * before the current scheduling deadline.
3826 * We can still have temporary overloads because we do not delay the
3827 * change in bandwidth until that time; so admission control is
3828 * not on the safe side. It does however guarantee tasks will never
3829 * consume more than promised.
3834 * sched_setparam() passes in -1 for its policy, to let the functions
3835 * it calls know not to change it.
3837 #define SETPARAM_POLICY -1
3839 static void __setscheduler_params(struct task_struct
*p
,
3840 const struct sched_attr
*attr
)
3842 int policy
= attr
->sched_policy
;
3844 if (policy
== SETPARAM_POLICY
)
3849 if (dl_policy(policy
))
3850 __setparam_dl(p
, attr
);
3851 else if (fair_policy(policy
))
3852 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3855 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3856 * !rt_policy. Always setting this ensures that things like
3857 * getparam()/getattr() don't report silly values for !rt tasks.
3859 p
->rt_priority
= attr
->sched_priority
;
3860 p
->normal_prio
= normal_prio(p
);
3864 /* Actually do priority change: must hold pi & rq lock. */
3865 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3866 const struct sched_attr
*attr
, bool keep_boost
)
3868 __setscheduler_params(p
, attr
);
3871 * Keep a potential priority boosting if called from
3872 * sched_setscheduler().
3875 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3877 p
->prio
= normal_prio(p
);
3879 if (dl_prio(p
->prio
))
3880 p
->sched_class
= &dl_sched_class
;
3881 else if (rt_prio(p
->prio
))
3882 p
->sched_class
= &rt_sched_class
;
3884 p
->sched_class
= &fair_sched_class
;
3888 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3890 struct sched_dl_entity
*dl_se
= &p
->dl
;
3892 attr
->sched_priority
= p
->rt_priority
;
3893 attr
->sched_runtime
= dl_se
->dl_runtime
;
3894 attr
->sched_deadline
= dl_se
->dl_deadline
;
3895 attr
->sched_period
= dl_se
->dl_period
;
3896 attr
->sched_flags
= dl_se
->flags
;
3900 * This function validates the new parameters of a -deadline task.
3901 * We ask for the deadline not being zero, and greater or equal
3902 * than the runtime, as well as the period of being zero or
3903 * greater than deadline. Furthermore, we have to be sure that
3904 * user parameters are above the internal resolution of 1us (we
3905 * check sched_runtime only since it is always the smaller one) and
3906 * below 2^63 ns (we have to check both sched_deadline and
3907 * sched_period, as the latter can be zero).
3910 __checkparam_dl(const struct sched_attr
*attr
)
3913 if (attr
->sched_deadline
== 0)
3917 * Since we truncate DL_SCALE bits, make sure we're at least
3920 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3924 * Since we use the MSB for wrap-around and sign issues, make
3925 * sure it's not set (mind that period can be equal to zero).
3927 if (attr
->sched_deadline
& (1ULL << 63) ||
3928 attr
->sched_period
& (1ULL << 63))
3931 /* runtime <= deadline <= period (if period != 0) */
3932 if ((attr
->sched_period
!= 0 &&
3933 attr
->sched_period
< attr
->sched_deadline
) ||
3934 attr
->sched_deadline
< attr
->sched_runtime
)
3941 * check the target process has a UID that matches the current process's
3943 static bool check_same_owner(struct task_struct
*p
)
3945 const struct cred
*cred
= current_cred(), *pcred
;
3949 pcred
= __task_cred(p
);
3950 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3951 uid_eq(cred
->euid
, pcred
->uid
));
3956 static bool dl_param_changed(struct task_struct
*p
,
3957 const struct sched_attr
*attr
)
3959 struct sched_dl_entity
*dl_se
= &p
->dl
;
3961 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3962 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3963 dl_se
->dl_period
!= attr
->sched_period
||
3964 dl_se
->flags
!= attr
->sched_flags
)
3970 static int __sched_setscheduler(struct task_struct
*p
,
3971 const struct sched_attr
*attr
,
3974 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3975 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3976 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3977 int new_effective_prio
, policy
= attr
->sched_policy
;
3978 const struct sched_class
*prev_class
;
3981 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3984 /* may grab non-irq protected spin_locks */
3985 BUG_ON(in_interrupt());
3987 /* double check policy once rq lock held */
3989 reset_on_fork
= p
->sched_reset_on_fork
;
3990 policy
= oldpolicy
= p
->policy
;
3992 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3994 if (!valid_policy(policy
))
3998 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4002 * Valid priorities for SCHED_FIFO and SCHED_RR are
4003 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4004 * SCHED_BATCH and SCHED_IDLE is 0.
4006 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4007 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4009 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4010 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4014 * Allow unprivileged RT tasks to decrease priority:
4016 if (user
&& !capable(CAP_SYS_NICE
)) {
4017 if (fair_policy(policy
)) {
4018 if (attr
->sched_nice
< task_nice(p
) &&
4019 !can_nice(p
, attr
->sched_nice
))
4023 if (rt_policy(policy
)) {
4024 unsigned long rlim_rtprio
=
4025 task_rlimit(p
, RLIMIT_RTPRIO
);
4027 /* can't set/change the rt policy */
4028 if (policy
!= p
->policy
&& !rlim_rtprio
)
4031 /* can't increase priority */
4032 if (attr
->sched_priority
> p
->rt_priority
&&
4033 attr
->sched_priority
> rlim_rtprio
)
4038 * Can't set/change SCHED_DEADLINE policy at all for now
4039 * (safest behavior); in the future we would like to allow
4040 * unprivileged DL tasks to increase their relative deadline
4041 * or reduce their runtime (both ways reducing utilization)
4043 if (dl_policy(policy
))
4047 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4048 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4050 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4051 if (!can_nice(p
, task_nice(p
)))
4055 /* can't change other user's priorities */
4056 if (!check_same_owner(p
))
4059 /* Normal users shall not reset the sched_reset_on_fork flag */
4060 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4065 retval
= security_task_setscheduler(p
);
4071 * make sure no PI-waiters arrive (or leave) while we are
4072 * changing the priority of the task:
4074 * To be able to change p->policy safely, the appropriate
4075 * runqueue lock must be held.
4077 rq
= task_rq_lock(p
, &rf
);
4080 * Changing the policy of the stop threads its a very bad idea
4082 if (p
== rq
->stop
) {
4083 task_rq_unlock(rq
, p
, &rf
);
4088 * If not changing anything there's no need to proceed further,
4089 * but store a possible modification of reset_on_fork.
4091 if (unlikely(policy
== p
->policy
)) {
4092 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4094 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4096 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4099 p
->sched_reset_on_fork
= reset_on_fork
;
4100 task_rq_unlock(rq
, p
, &rf
);
4106 #ifdef CONFIG_RT_GROUP_SCHED
4108 * Do not allow realtime tasks into groups that have no runtime
4111 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4112 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4113 !task_group_is_autogroup(task_group(p
))) {
4114 task_rq_unlock(rq
, p
, &rf
);
4119 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4120 cpumask_t
*span
= rq
->rd
->span
;
4123 * Don't allow tasks with an affinity mask smaller than
4124 * the entire root_domain to become SCHED_DEADLINE. We
4125 * will also fail if there's no bandwidth available.
4127 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4128 rq
->rd
->dl_bw
.bw
== 0) {
4129 task_rq_unlock(rq
, p
, &rf
);
4136 /* recheck policy now with rq lock held */
4137 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4138 policy
= oldpolicy
= -1;
4139 task_rq_unlock(rq
, p
, &rf
);
4144 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4145 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4148 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4149 task_rq_unlock(rq
, p
, &rf
);
4153 p
->sched_reset_on_fork
= reset_on_fork
;
4158 * Take priority boosted tasks into account. If the new
4159 * effective priority is unchanged, we just store the new
4160 * normal parameters and do not touch the scheduler class and
4161 * the runqueue. This will be done when the task deboost
4164 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4165 if (new_effective_prio
== oldprio
)
4166 queue_flags
&= ~DEQUEUE_MOVE
;
4169 queued
= task_on_rq_queued(p
);
4170 running
= task_current(rq
, p
);
4172 dequeue_task(rq
, p
, queue_flags
);
4174 put_prev_task(rq
, p
);
4176 prev_class
= p
->sched_class
;
4177 __setscheduler(rq
, p
, attr
, pi
);
4180 p
->sched_class
->set_curr_task(rq
);
4183 * We enqueue to tail when the priority of a task is
4184 * increased (user space view).
4186 if (oldprio
< p
->prio
)
4187 queue_flags
|= ENQUEUE_HEAD
;
4189 enqueue_task(rq
, p
, queue_flags
);
4192 check_class_changed(rq
, p
, prev_class
, oldprio
);
4193 preempt_disable(); /* avoid rq from going away on us */
4194 task_rq_unlock(rq
, p
, &rf
);
4197 rt_mutex_adjust_pi(p
);
4200 * Run balance callbacks after we've adjusted the PI chain.
4202 balance_callback(rq
);
4208 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4209 const struct sched_param
*param
, bool check
)
4211 struct sched_attr attr
= {
4212 .sched_policy
= policy
,
4213 .sched_priority
= param
->sched_priority
,
4214 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4217 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4218 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4219 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4220 policy
&= ~SCHED_RESET_ON_FORK
;
4221 attr
.sched_policy
= policy
;
4224 return __sched_setscheduler(p
, &attr
, check
, true);
4227 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4228 * @p: the task in question.
4229 * @policy: new policy.
4230 * @param: structure containing the new RT priority.
4232 * Return: 0 on success. An error code otherwise.
4234 * NOTE that the task may be already dead.
4236 int sched_setscheduler(struct task_struct
*p
, int policy
,
4237 const struct sched_param
*param
)
4239 return _sched_setscheduler(p
, policy
, param
, true);
4241 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4243 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4245 return __sched_setscheduler(p
, attr
, true, true);
4247 EXPORT_SYMBOL_GPL(sched_setattr
);
4250 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4251 * @p: the task in question.
4252 * @policy: new policy.
4253 * @param: structure containing the new RT priority.
4255 * Just like sched_setscheduler, only don't bother checking if the
4256 * current context has permission. For example, this is needed in
4257 * stop_machine(): we create temporary high priority worker threads,
4258 * but our caller might not have that capability.
4260 * Return: 0 on success. An error code otherwise.
4262 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4263 const struct sched_param
*param
)
4265 return _sched_setscheduler(p
, policy
, param
, false);
4267 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4270 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4272 struct sched_param lparam
;
4273 struct task_struct
*p
;
4276 if (!param
|| pid
< 0)
4278 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4283 p
= find_process_by_pid(pid
);
4285 retval
= sched_setscheduler(p
, policy
, &lparam
);
4292 * Mimics kernel/events/core.c perf_copy_attr().
4294 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4295 struct sched_attr
*attr
)
4300 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4304 * zero the full structure, so that a short copy will be nice.
4306 memset(attr
, 0, sizeof(*attr
));
4308 ret
= get_user(size
, &uattr
->size
);
4312 if (size
> PAGE_SIZE
) /* silly large */
4315 if (!size
) /* abi compat */
4316 size
= SCHED_ATTR_SIZE_VER0
;
4318 if (size
< SCHED_ATTR_SIZE_VER0
)
4322 * If we're handed a bigger struct than we know of,
4323 * ensure all the unknown bits are 0 - i.e. new
4324 * user-space does not rely on any kernel feature
4325 * extensions we dont know about yet.
4327 if (size
> sizeof(*attr
)) {
4328 unsigned char __user
*addr
;
4329 unsigned char __user
*end
;
4332 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4333 end
= (void __user
*)uattr
+ size
;
4335 for (; addr
< end
; addr
++) {
4336 ret
= get_user(val
, addr
);
4342 size
= sizeof(*attr
);
4345 ret
= copy_from_user(attr
, uattr
, size
);
4350 * XXX: do we want to be lenient like existing syscalls; or do we want
4351 * to be strict and return an error on out-of-bounds values?
4353 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4358 put_user(sizeof(*attr
), &uattr
->size
);
4363 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4364 * @pid: the pid in question.
4365 * @policy: new policy.
4366 * @param: structure containing the new RT priority.
4368 * Return: 0 on success. An error code otherwise.
4370 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4371 struct sched_param __user
*, param
)
4373 /* negative values for policy are not valid */
4377 return do_sched_setscheduler(pid
, policy
, param
);
4381 * sys_sched_setparam - set/change the RT priority of a thread
4382 * @pid: the pid in question.
4383 * @param: structure containing the new RT priority.
4385 * Return: 0 on success. An error code otherwise.
4387 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4389 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4393 * sys_sched_setattr - same as above, but with extended sched_attr
4394 * @pid: the pid in question.
4395 * @uattr: structure containing the extended parameters.
4396 * @flags: for future extension.
4398 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4399 unsigned int, flags
)
4401 struct sched_attr attr
;
4402 struct task_struct
*p
;
4405 if (!uattr
|| pid
< 0 || flags
)
4408 retval
= sched_copy_attr(uattr
, &attr
);
4412 if ((int)attr
.sched_policy
< 0)
4417 p
= find_process_by_pid(pid
);
4419 retval
= sched_setattr(p
, &attr
);
4426 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4427 * @pid: the pid in question.
4429 * Return: On success, the policy of the thread. Otherwise, a negative error
4432 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4434 struct task_struct
*p
;
4442 p
= find_process_by_pid(pid
);
4444 retval
= security_task_getscheduler(p
);
4447 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4454 * sys_sched_getparam - get the RT priority of a thread
4455 * @pid: the pid in question.
4456 * @param: structure containing the RT priority.
4458 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4461 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4463 struct sched_param lp
= { .sched_priority
= 0 };
4464 struct task_struct
*p
;
4467 if (!param
|| pid
< 0)
4471 p
= find_process_by_pid(pid
);
4476 retval
= security_task_getscheduler(p
);
4480 if (task_has_rt_policy(p
))
4481 lp
.sched_priority
= p
->rt_priority
;
4485 * This one might sleep, we cannot do it with a spinlock held ...
4487 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4496 static int sched_read_attr(struct sched_attr __user
*uattr
,
4497 struct sched_attr
*attr
,
4502 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4506 * If we're handed a smaller struct than we know of,
4507 * ensure all the unknown bits are 0 - i.e. old
4508 * user-space does not get uncomplete information.
4510 if (usize
< sizeof(*attr
)) {
4511 unsigned char *addr
;
4514 addr
= (void *)attr
+ usize
;
4515 end
= (void *)attr
+ sizeof(*attr
);
4517 for (; addr
< end
; addr
++) {
4525 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4533 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4534 * @pid: the pid in question.
4535 * @uattr: structure containing the extended parameters.
4536 * @size: sizeof(attr) for fwd/bwd comp.
4537 * @flags: for future extension.
4539 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4540 unsigned int, size
, unsigned int, flags
)
4542 struct sched_attr attr
= {
4543 .size
= sizeof(struct sched_attr
),
4545 struct task_struct
*p
;
4548 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4549 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4553 p
= find_process_by_pid(pid
);
4558 retval
= security_task_getscheduler(p
);
4562 attr
.sched_policy
= p
->policy
;
4563 if (p
->sched_reset_on_fork
)
4564 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4565 if (task_has_dl_policy(p
))
4566 __getparam_dl(p
, &attr
);
4567 else if (task_has_rt_policy(p
))
4568 attr
.sched_priority
= p
->rt_priority
;
4570 attr
.sched_nice
= task_nice(p
);
4574 retval
= sched_read_attr(uattr
, &attr
, size
);
4582 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4584 cpumask_var_t cpus_allowed
, new_mask
;
4585 struct task_struct
*p
;
4590 p
= find_process_by_pid(pid
);
4596 /* Prevent p going away */
4600 if (p
->flags
& PF_NO_SETAFFINITY
) {
4604 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4608 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4610 goto out_free_cpus_allowed
;
4613 if (!check_same_owner(p
)) {
4615 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4617 goto out_free_new_mask
;
4622 retval
= security_task_setscheduler(p
);
4624 goto out_free_new_mask
;
4627 cpuset_cpus_allowed(p
, cpus_allowed
);
4628 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4631 * Since bandwidth control happens on root_domain basis,
4632 * if admission test is enabled, we only admit -deadline
4633 * tasks allowed to run on all the CPUs in the task's
4637 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4639 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4642 goto out_free_new_mask
;
4648 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4651 cpuset_cpus_allowed(p
, cpus_allowed
);
4652 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4654 * We must have raced with a concurrent cpuset
4655 * update. Just reset the cpus_allowed to the
4656 * cpuset's cpus_allowed
4658 cpumask_copy(new_mask
, cpus_allowed
);
4663 free_cpumask_var(new_mask
);
4664 out_free_cpus_allowed
:
4665 free_cpumask_var(cpus_allowed
);
4671 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4672 struct cpumask
*new_mask
)
4674 if (len
< cpumask_size())
4675 cpumask_clear(new_mask
);
4676 else if (len
> cpumask_size())
4677 len
= cpumask_size();
4679 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4683 * sys_sched_setaffinity - set the cpu affinity of a process
4684 * @pid: pid of the process
4685 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4686 * @user_mask_ptr: user-space pointer to the new cpu mask
4688 * Return: 0 on success. An error code otherwise.
4690 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4691 unsigned long __user
*, user_mask_ptr
)
4693 cpumask_var_t new_mask
;
4696 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4699 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4701 retval
= sched_setaffinity(pid
, new_mask
);
4702 free_cpumask_var(new_mask
);
4706 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4708 struct task_struct
*p
;
4709 unsigned long flags
;
4715 p
= find_process_by_pid(pid
);
4719 retval
= security_task_getscheduler(p
);
4723 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4724 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4725 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4734 * sys_sched_getaffinity - get the cpu affinity of a process
4735 * @pid: pid of the process
4736 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4737 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4739 * Return: 0 on success. An error code otherwise.
4741 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4742 unsigned long __user
*, user_mask_ptr
)
4747 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4749 if (len
& (sizeof(unsigned long)-1))
4752 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4755 ret
= sched_getaffinity(pid
, mask
);
4757 size_t retlen
= min_t(size_t, len
, cpumask_size());
4759 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4764 free_cpumask_var(mask
);
4770 * sys_sched_yield - yield the current processor to other threads.
4772 * This function yields the current CPU to other tasks. If there are no
4773 * other threads running on this CPU then this function will return.
4777 SYSCALL_DEFINE0(sched_yield
)
4779 struct rq
*rq
= this_rq_lock();
4781 schedstat_inc(rq
, yld_count
);
4782 current
->sched_class
->yield_task(rq
);
4785 * Since we are going to call schedule() anyway, there's
4786 * no need to preempt or enable interrupts:
4788 __release(rq
->lock
);
4789 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4790 do_raw_spin_unlock(&rq
->lock
);
4791 sched_preempt_enable_no_resched();
4798 int __sched
_cond_resched(void)
4800 if (should_resched(0)) {
4801 preempt_schedule_common();
4806 EXPORT_SYMBOL(_cond_resched
);
4809 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4810 * call schedule, and on return reacquire the lock.
4812 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4813 * operations here to prevent schedule() from being called twice (once via
4814 * spin_unlock(), once by hand).
4816 int __cond_resched_lock(spinlock_t
*lock
)
4818 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4821 lockdep_assert_held(lock
);
4823 if (spin_needbreak(lock
) || resched
) {
4826 preempt_schedule_common();
4834 EXPORT_SYMBOL(__cond_resched_lock
);
4836 int __sched
__cond_resched_softirq(void)
4838 BUG_ON(!in_softirq());
4840 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4842 preempt_schedule_common();
4848 EXPORT_SYMBOL(__cond_resched_softirq
);
4851 * yield - yield the current processor to other threads.
4853 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4855 * The scheduler is at all times free to pick the calling task as the most
4856 * eligible task to run, if removing the yield() call from your code breaks
4857 * it, its already broken.
4859 * Typical broken usage is:
4864 * where one assumes that yield() will let 'the other' process run that will
4865 * make event true. If the current task is a SCHED_FIFO task that will never
4866 * happen. Never use yield() as a progress guarantee!!
4868 * If you want to use yield() to wait for something, use wait_event().
4869 * If you want to use yield() to be 'nice' for others, use cond_resched().
4870 * If you still want to use yield(), do not!
4872 void __sched
yield(void)
4874 set_current_state(TASK_RUNNING
);
4877 EXPORT_SYMBOL(yield
);
4880 * yield_to - yield the current processor to another thread in
4881 * your thread group, or accelerate that thread toward the
4882 * processor it's on.
4884 * @preempt: whether task preemption is allowed or not
4886 * It's the caller's job to ensure that the target task struct
4887 * can't go away on us before we can do any checks.
4890 * true (>0) if we indeed boosted the target task.
4891 * false (0) if we failed to boost the target.
4892 * -ESRCH if there's no task to yield to.
4894 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4896 struct task_struct
*curr
= current
;
4897 struct rq
*rq
, *p_rq
;
4898 unsigned long flags
;
4901 local_irq_save(flags
);
4907 * If we're the only runnable task on the rq and target rq also
4908 * has only one task, there's absolutely no point in yielding.
4910 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4915 double_rq_lock(rq
, p_rq
);
4916 if (task_rq(p
) != p_rq
) {
4917 double_rq_unlock(rq
, p_rq
);
4921 if (!curr
->sched_class
->yield_to_task
)
4924 if (curr
->sched_class
!= p
->sched_class
)
4927 if (task_running(p_rq
, p
) || p
->state
)
4930 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4932 schedstat_inc(rq
, yld_count
);
4934 * Make p's CPU reschedule; pick_next_entity takes care of
4937 if (preempt
&& rq
!= p_rq
)
4942 double_rq_unlock(rq
, p_rq
);
4944 local_irq_restore(flags
);
4951 EXPORT_SYMBOL_GPL(yield_to
);
4954 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4955 * that process accounting knows that this is a task in IO wait state.
4957 long __sched
io_schedule_timeout(long timeout
)
4959 int old_iowait
= current
->in_iowait
;
4963 current
->in_iowait
= 1;
4964 blk_schedule_flush_plug(current
);
4966 delayacct_blkio_start();
4968 atomic_inc(&rq
->nr_iowait
);
4969 ret
= schedule_timeout(timeout
);
4970 current
->in_iowait
= old_iowait
;
4971 atomic_dec(&rq
->nr_iowait
);
4972 delayacct_blkio_end();
4976 EXPORT_SYMBOL(io_schedule_timeout
);
4979 * sys_sched_get_priority_max - return maximum RT priority.
4980 * @policy: scheduling class.
4982 * Return: On success, this syscall returns the maximum
4983 * rt_priority that can be used by a given scheduling class.
4984 * On failure, a negative error code is returned.
4986 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4993 ret
= MAX_USER_RT_PRIO
-1;
4995 case SCHED_DEADLINE
:
5006 * sys_sched_get_priority_min - return minimum RT priority.
5007 * @policy: scheduling class.
5009 * Return: On success, this syscall returns the minimum
5010 * rt_priority that can be used by a given scheduling class.
5011 * On failure, a negative error code is returned.
5013 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5022 case SCHED_DEADLINE
:
5032 * sys_sched_rr_get_interval - return the default timeslice of a process.
5033 * @pid: pid of the process.
5034 * @interval: userspace pointer to the timeslice value.
5036 * this syscall writes the default timeslice value of a given process
5037 * into the user-space timespec buffer. A value of '0' means infinity.
5039 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5042 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5043 struct timespec __user
*, interval
)
5045 struct task_struct
*p
;
5046 unsigned int time_slice
;
5057 p
= find_process_by_pid(pid
);
5061 retval
= security_task_getscheduler(p
);
5065 rq
= task_rq_lock(p
, &rf
);
5067 if (p
->sched_class
->get_rr_interval
)
5068 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5069 task_rq_unlock(rq
, p
, &rf
);
5072 jiffies_to_timespec(time_slice
, &t
);
5073 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5081 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5083 void sched_show_task(struct task_struct
*p
)
5085 unsigned long free
= 0;
5087 unsigned long state
= p
->state
;
5090 state
= __ffs(state
) + 1;
5091 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5092 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5093 #if BITS_PER_LONG == 32
5094 if (state
== TASK_RUNNING
)
5095 printk(KERN_CONT
" running ");
5097 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5099 if (state
== TASK_RUNNING
)
5100 printk(KERN_CONT
" running task ");
5102 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5104 #ifdef CONFIG_DEBUG_STACK_USAGE
5105 free
= stack_not_used(p
);
5110 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5112 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5113 task_pid_nr(p
), ppid
,
5114 (unsigned long)task_thread_info(p
)->flags
);
5116 print_worker_info(KERN_INFO
, p
);
5117 show_stack(p
, NULL
);
5120 void show_state_filter(unsigned long state_filter
)
5122 struct task_struct
*g
, *p
;
5124 #if BITS_PER_LONG == 32
5126 " task PC stack pid father\n");
5129 " task PC stack pid father\n");
5132 for_each_process_thread(g
, p
) {
5134 * reset the NMI-timeout, listing all files on a slow
5135 * console might take a lot of time:
5137 touch_nmi_watchdog();
5138 if (!state_filter
|| (p
->state
& state_filter
))
5142 touch_all_softlockup_watchdogs();
5144 #ifdef CONFIG_SCHED_DEBUG
5146 sysrq_sched_debug_show();
5150 * Only show locks if all tasks are dumped:
5153 debug_show_all_locks();
5156 void init_idle_bootup_task(struct task_struct
*idle
)
5158 idle
->sched_class
= &idle_sched_class
;
5162 * init_idle - set up an idle thread for a given CPU
5163 * @idle: task in question
5164 * @cpu: cpu the idle task belongs to
5166 * NOTE: this function does not set the idle thread's NEED_RESCHED
5167 * flag, to make booting more robust.
5169 void init_idle(struct task_struct
*idle
, int cpu
)
5171 struct rq
*rq
= cpu_rq(cpu
);
5172 unsigned long flags
;
5174 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5175 raw_spin_lock(&rq
->lock
);
5177 __sched_fork(0, idle
);
5178 idle
->state
= TASK_RUNNING
;
5179 idle
->se
.exec_start
= sched_clock();
5181 kasan_unpoison_task_stack(idle
);
5185 * Its possible that init_idle() gets called multiple times on a task,
5186 * in that case do_set_cpus_allowed() will not do the right thing.
5188 * And since this is boot we can forgo the serialization.
5190 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5193 * We're having a chicken and egg problem, even though we are
5194 * holding rq->lock, the cpu isn't yet set to this cpu so the
5195 * lockdep check in task_group() will fail.
5197 * Similar case to sched_fork(). / Alternatively we could
5198 * use task_rq_lock() here and obtain the other rq->lock.
5203 __set_task_cpu(idle
, cpu
);
5206 rq
->curr
= rq
->idle
= idle
;
5207 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5211 raw_spin_unlock(&rq
->lock
);
5212 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5214 /* Set the preempt count _outside_ the spinlocks! */
5215 init_idle_preempt_count(idle
, cpu
);
5218 * The idle tasks have their own, simple scheduling class:
5220 idle
->sched_class
= &idle_sched_class
;
5221 ftrace_graph_init_idle_task(idle
, cpu
);
5222 vtime_init_idle(idle
, cpu
);
5224 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5228 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5229 const struct cpumask
*trial
)
5231 int ret
= 1, trial_cpus
;
5232 struct dl_bw
*cur_dl_b
;
5233 unsigned long flags
;
5235 if (!cpumask_weight(cur
))
5238 rcu_read_lock_sched();
5239 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5240 trial_cpus
= cpumask_weight(trial
);
5242 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5243 if (cur_dl_b
->bw
!= -1 &&
5244 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5246 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5247 rcu_read_unlock_sched();
5252 int task_can_attach(struct task_struct
*p
,
5253 const struct cpumask
*cs_cpus_allowed
)
5258 * Kthreads which disallow setaffinity shouldn't be moved
5259 * to a new cpuset; we don't want to change their cpu
5260 * affinity and isolating such threads by their set of
5261 * allowed nodes is unnecessary. Thus, cpusets are not
5262 * applicable for such threads. This prevents checking for
5263 * success of set_cpus_allowed_ptr() on all attached tasks
5264 * before cpus_allowed may be changed.
5266 if (p
->flags
& PF_NO_SETAFFINITY
) {
5272 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5274 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5279 unsigned long flags
;
5281 rcu_read_lock_sched();
5282 dl_b
= dl_bw_of(dest_cpu
);
5283 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5284 cpus
= dl_bw_cpus(dest_cpu
);
5285 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5290 * We reserve space for this task in the destination
5291 * root_domain, as we can't fail after this point.
5292 * We will free resources in the source root_domain
5293 * later on (see set_cpus_allowed_dl()).
5295 __dl_add(dl_b
, p
->dl
.dl_bw
);
5297 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5298 rcu_read_unlock_sched();
5308 static bool sched_smp_initialized __read_mostly
;
5310 #ifdef CONFIG_NUMA_BALANCING
5311 /* Migrate current task p to target_cpu */
5312 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5314 struct migration_arg arg
= { p
, target_cpu
};
5315 int curr_cpu
= task_cpu(p
);
5317 if (curr_cpu
== target_cpu
)
5320 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5323 /* TODO: This is not properly updating schedstats */
5325 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5326 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5330 * Requeue a task on a given node and accurately track the number of NUMA
5331 * tasks on the runqueues
5333 void sched_setnuma(struct task_struct
*p
, int nid
)
5335 bool queued
, running
;
5339 rq
= task_rq_lock(p
, &rf
);
5340 queued
= task_on_rq_queued(p
);
5341 running
= task_current(rq
, p
);
5344 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5346 put_prev_task(rq
, p
);
5348 p
->numa_preferred_nid
= nid
;
5351 p
->sched_class
->set_curr_task(rq
);
5353 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5354 task_rq_unlock(rq
, p
, &rf
);
5356 #endif /* CONFIG_NUMA_BALANCING */
5358 #ifdef CONFIG_HOTPLUG_CPU
5360 * Ensures that the idle task is using init_mm right before its cpu goes
5363 void idle_task_exit(void)
5365 struct mm_struct
*mm
= current
->active_mm
;
5367 BUG_ON(cpu_online(smp_processor_id()));
5369 if (mm
!= &init_mm
) {
5370 switch_mm_irqs_off(mm
, &init_mm
, current
);
5371 finish_arch_post_lock_switch();
5377 * Since this CPU is going 'away' for a while, fold any nr_active delta
5378 * we might have. Assumes we're called after migrate_tasks() so that the
5379 * nr_active count is stable.
5381 * Also see the comment "Global load-average calculations".
5383 static void calc_load_migrate(struct rq
*rq
)
5385 long delta
= calc_load_fold_active(rq
);
5387 atomic_long_add(delta
, &calc_load_tasks
);
5390 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5394 static const struct sched_class fake_sched_class
= {
5395 .put_prev_task
= put_prev_task_fake
,
5398 static struct task_struct fake_task
= {
5400 * Avoid pull_{rt,dl}_task()
5402 .prio
= MAX_PRIO
+ 1,
5403 .sched_class
= &fake_sched_class
,
5407 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5408 * try_to_wake_up()->select_task_rq().
5410 * Called with rq->lock held even though we'er in stop_machine() and
5411 * there's no concurrency possible, we hold the required locks anyway
5412 * because of lock validation efforts.
5414 static void migrate_tasks(struct rq
*dead_rq
)
5416 struct rq
*rq
= dead_rq
;
5417 struct task_struct
*next
, *stop
= rq
->stop
;
5418 struct pin_cookie cookie
;
5422 * Fudge the rq selection such that the below task selection loop
5423 * doesn't get stuck on the currently eligible stop task.
5425 * We're currently inside stop_machine() and the rq is either stuck
5426 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5427 * either way we should never end up calling schedule() until we're
5433 * put_prev_task() and pick_next_task() sched
5434 * class method both need to have an up-to-date
5435 * value of rq->clock[_task]
5437 update_rq_clock(rq
);
5441 * There's this thread running, bail when that's the only
5444 if (rq
->nr_running
== 1)
5448 * pick_next_task assumes pinned rq->lock.
5450 cookie
= lockdep_pin_lock(&rq
->lock
);
5451 next
= pick_next_task(rq
, &fake_task
, cookie
);
5453 next
->sched_class
->put_prev_task(rq
, next
);
5456 * Rules for changing task_struct::cpus_allowed are holding
5457 * both pi_lock and rq->lock, such that holding either
5458 * stabilizes the mask.
5460 * Drop rq->lock is not quite as disastrous as it usually is
5461 * because !cpu_active at this point, which means load-balance
5462 * will not interfere. Also, stop-machine.
5464 lockdep_unpin_lock(&rq
->lock
, cookie
);
5465 raw_spin_unlock(&rq
->lock
);
5466 raw_spin_lock(&next
->pi_lock
);
5467 raw_spin_lock(&rq
->lock
);
5470 * Since we're inside stop-machine, _nothing_ should have
5471 * changed the task, WARN if weird stuff happened, because in
5472 * that case the above rq->lock drop is a fail too.
5474 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5475 raw_spin_unlock(&next
->pi_lock
);
5479 /* Find suitable destination for @next, with force if needed. */
5480 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5482 rq
= __migrate_task(rq
, next
, dest_cpu
);
5483 if (rq
!= dead_rq
) {
5484 raw_spin_unlock(&rq
->lock
);
5486 raw_spin_lock(&rq
->lock
);
5488 raw_spin_unlock(&next
->pi_lock
);
5493 #endif /* CONFIG_HOTPLUG_CPU */
5495 static void set_rq_online(struct rq
*rq
)
5498 const struct sched_class
*class;
5500 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5503 for_each_class(class) {
5504 if (class->rq_online
)
5505 class->rq_online(rq
);
5510 static void set_rq_offline(struct rq
*rq
)
5513 const struct sched_class
*class;
5515 for_each_class(class) {
5516 if (class->rq_offline
)
5517 class->rq_offline(rq
);
5520 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5525 static void set_cpu_rq_start_time(unsigned int cpu
)
5527 struct rq
*rq
= cpu_rq(cpu
);
5529 rq
->age_stamp
= sched_clock_cpu(cpu
);
5532 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5534 #ifdef CONFIG_SCHED_DEBUG
5536 static __read_mostly
int sched_debug_enabled
;
5538 static int __init
sched_debug_setup(char *str
)
5540 sched_debug_enabled
= 1;
5544 early_param("sched_debug", sched_debug_setup
);
5546 static inline bool sched_debug(void)
5548 return sched_debug_enabled
;
5551 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5552 struct cpumask
*groupmask
)
5554 struct sched_group
*group
= sd
->groups
;
5556 cpumask_clear(groupmask
);
5558 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5560 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5561 printk("does not load-balance\n");
5563 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5568 printk(KERN_CONT
"span %*pbl level %s\n",
5569 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5571 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5572 printk(KERN_ERR
"ERROR: domain->span does not contain "
5575 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5576 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5580 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5584 printk(KERN_ERR
"ERROR: group is NULL\n");
5588 if (!cpumask_weight(sched_group_cpus(group
))) {
5589 printk(KERN_CONT
"\n");
5590 printk(KERN_ERR
"ERROR: empty group\n");
5594 if (!(sd
->flags
& SD_OVERLAP
) &&
5595 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5596 printk(KERN_CONT
"\n");
5597 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5601 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5603 printk(KERN_CONT
" %*pbl",
5604 cpumask_pr_args(sched_group_cpus(group
)));
5605 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5606 printk(KERN_CONT
" (cpu_capacity = %d)",
5607 group
->sgc
->capacity
);
5610 group
= group
->next
;
5611 } while (group
!= sd
->groups
);
5612 printk(KERN_CONT
"\n");
5614 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5615 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5618 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5619 printk(KERN_ERR
"ERROR: parent span is not a superset "
5620 "of domain->span\n");
5624 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5628 if (!sched_debug_enabled
)
5632 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5636 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5639 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5647 #else /* !CONFIG_SCHED_DEBUG */
5648 # define sched_domain_debug(sd, cpu) do { } while (0)
5649 static inline bool sched_debug(void)
5653 #endif /* CONFIG_SCHED_DEBUG */
5655 static int sd_degenerate(struct sched_domain
*sd
)
5657 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5660 /* Following flags need at least 2 groups */
5661 if (sd
->flags
& (SD_LOAD_BALANCE
|
5662 SD_BALANCE_NEWIDLE
|
5665 SD_SHARE_CPUCAPACITY
|
5666 SD_SHARE_PKG_RESOURCES
|
5667 SD_SHARE_POWERDOMAIN
)) {
5668 if (sd
->groups
!= sd
->groups
->next
)
5672 /* Following flags don't use groups */
5673 if (sd
->flags
& (SD_WAKE_AFFINE
))
5680 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5682 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5684 if (sd_degenerate(parent
))
5687 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5690 /* Flags needing groups don't count if only 1 group in parent */
5691 if (parent
->groups
== parent
->groups
->next
) {
5692 pflags
&= ~(SD_LOAD_BALANCE
|
5693 SD_BALANCE_NEWIDLE
|
5696 SD_SHARE_CPUCAPACITY
|
5697 SD_SHARE_PKG_RESOURCES
|
5699 SD_SHARE_POWERDOMAIN
);
5700 if (nr_node_ids
== 1)
5701 pflags
&= ~SD_SERIALIZE
;
5703 if (~cflags
& pflags
)
5709 static void free_rootdomain(struct rcu_head
*rcu
)
5711 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5713 cpupri_cleanup(&rd
->cpupri
);
5714 cpudl_cleanup(&rd
->cpudl
);
5715 free_cpumask_var(rd
->dlo_mask
);
5716 free_cpumask_var(rd
->rto_mask
);
5717 free_cpumask_var(rd
->online
);
5718 free_cpumask_var(rd
->span
);
5722 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5724 struct root_domain
*old_rd
= NULL
;
5725 unsigned long flags
;
5727 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5732 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5735 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5738 * If we dont want to free the old_rd yet then
5739 * set old_rd to NULL to skip the freeing later
5742 if (!atomic_dec_and_test(&old_rd
->refcount
))
5746 atomic_inc(&rd
->refcount
);
5749 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5750 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5753 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5756 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5759 static int init_rootdomain(struct root_domain
*rd
)
5761 memset(rd
, 0, sizeof(*rd
));
5763 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5765 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5767 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5769 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5772 init_dl_bw(&rd
->dl_bw
);
5773 if (cpudl_init(&rd
->cpudl
) != 0)
5776 if (cpupri_init(&rd
->cpupri
) != 0)
5781 free_cpumask_var(rd
->rto_mask
);
5783 free_cpumask_var(rd
->dlo_mask
);
5785 free_cpumask_var(rd
->online
);
5787 free_cpumask_var(rd
->span
);
5793 * By default the system creates a single root-domain with all cpus as
5794 * members (mimicking the global state we have today).
5796 struct root_domain def_root_domain
;
5798 static void init_defrootdomain(void)
5800 init_rootdomain(&def_root_domain
);
5802 atomic_set(&def_root_domain
.refcount
, 1);
5805 static struct root_domain
*alloc_rootdomain(void)
5807 struct root_domain
*rd
;
5809 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5813 if (init_rootdomain(rd
) != 0) {
5821 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5823 struct sched_group
*tmp
, *first
;
5832 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5837 } while (sg
!= first
);
5840 static void free_sched_domain(struct rcu_head
*rcu
)
5842 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5845 * If its an overlapping domain it has private groups, iterate and
5848 if (sd
->flags
& SD_OVERLAP
) {
5849 free_sched_groups(sd
->groups
, 1);
5850 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5851 kfree(sd
->groups
->sgc
);
5857 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5859 call_rcu(&sd
->rcu
, free_sched_domain
);
5862 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5864 for (; sd
; sd
= sd
->parent
)
5865 destroy_sched_domain(sd
, cpu
);
5869 * Keep a special pointer to the highest sched_domain that has
5870 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5871 * allows us to avoid some pointer chasing select_idle_sibling().
5873 * Also keep a unique ID per domain (we use the first cpu number in
5874 * the cpumask of the domain), this allows us to quickly tell if
5875 * two cpus are in the same cache domain, see cpus_share_cache().
5877 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5878 DEFINE_PER_CPU(int, sd_llc_size
);
5879 DEFINE_PER_CPU(int, sd_llc_id
);
5880 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5881 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5882 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5884 static void update_top_cache_domain(int cpu
)
5886 struct sched_domain
*sd
;
5887 struct sched_domain
*busy_sd
= NULL
;
5891 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5893 id
= cpumask_first(sched_domain_span(sd
));
5894 size
= cpumask_weight(sched_domain_span(sd
));
5895 busy_sd
= sd
->parent
; /* sd_busy */
5897 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5899 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5900 per_cpu(sd_llc_size
, cpu
) = size
;
5901 per_cpu(sd_llc_id
, cpu
) = id
;
5903 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5904 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5906 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5907 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5911 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5912 * hold the hotplug lock.
5915 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5917 struct rq
*rq
= cpu_rq(cpu
);
5918 struct sched_domain
*tmp
;
5920 /* Remove the sched domains which do not contribute to scheduling. */
5921 for (tmp
= sd
; tmp
; ) {
5922 struct sched_domain
*parent
= tmp
->parent
;
5926 if (sd_parent_degenerate(tmp
, parent
)) {
5927 tmp
->parent
= parent
->parent
;
5929 parent
->parent
->child
= tmp
;
5931 * Transfer SD_PREFER_SIBLING down in case of a
5932 * degenerate parent; the spans match for this
5933 * so the property transfers.
5935 if (parent
->flags
& SD_PREFER_SIBLING
)
5936 tmp
->flags
|= SD_PREFER_SIBLING
;
5937 destroy_sched_domain(parent
, cpu
);
5942 if (sd
&& sd_degenerate(sd
)) {
5945 destroy_sched_domain(tmp
, cpu
);
5950 sched_domain_debug(sd
, cpu
);
5952 rq_attach_root(rq
, rd
);
5954 rcu_assign_pointer(rq
->sd
, sd
);
5955 destroy_sched_domains(tmp
, cpu
);
5957 update_top_cache_domain(cpu
);
5960 /* Setup the mask of cpus configured for isolated domains */
5961 static int __init
isolated_cpu_setup(char *str
)
5965 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5966 ret
= cpulist_parse(str
, cpu_isolated_map
);
5968 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5973 __setup("isolcpus=", isolated_cpu_setup
);
5976 struct sched_domain
** __percpu sd
;
5977 struct root_domain
*rd
;
5988 * Build an iteration mask that can exclude certain CPUs from the upwards
5991 * Asymmetric node setups can result in situations where the domain tree is of
5992 * unequal depth, make sure to skip domains that already cover the entire
5995 * In that case build_sched_domains() will have terminated the iteration early
5996 * and our sibling sd spans will be empty. Domains should always include the
5997 * cpu they're built on, so check that.
6000 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6002 const struct cpumask
*span
= sched_domain_span(sd
);
6003 struct sd_data
*sdd
= sd
->private;
6004 struct sched_domain
*sibling
;
6007 for_each_cpu(i
, span
) {
6008 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6009 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6012 cpumask_set_cpu(i
, sched_group_mask(sg
));
6017 * Return the canonical balance cpu for this group, this is the first cpu
6018 * of this group that's also in the iteration mask.
6020 int group_balance_cpu(struct sched_group
*sg
)
6022 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6026 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6028 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6029 const struct cpumask
*span
= sched_domain_span(sd
);
6030 struct cpumask
*covered
= sched_domains_tmpmask
;
6031 struct sd_data
*sdd
= sd
->private;
6032 struct sched_domain
*sibling
;
6035 cpumask_clear(covered
);
6037 for_each_cpu(i
, span
) {
6038 struct cpumask
*sg_span
;
6040 if (cpumask_test_cpu(i
, covered
))
6043 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6045 /* See the comment near build_group_mask(). */
6046 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6049 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6050 GFP_KERNEL
, cpu_to_node(cpu
));
6055 sg_span
= sched_group_cpus(sg
);
6057 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6059 cpumask_set_cpu(i
, sg_span
);
6061 cpumask_or(covered
, covered
, sg_span
);
6063 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6064 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6065 build_group_mask(sd
, sg
);
6068 * Initialize sgc->capacity such that even if we mess up the
6069 * domains and no possible iteration will get us here, we won't
6072 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6075 * Make sure the first group of this domain contains the
6076 * canonical balance cpu. Otherwise the sched_domain iteration
6077 * breaks. See update_sg_lb_stats().
6079 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6080 group_balance_cpu(sg
) == cpu
)
6090 sd
->groups
= groups
;
6095 free_sched_groups(first
, 0);
6100 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6102 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6103 struct sched_domain
*child
= sd
->child
;
6106 cpu
= cpumask_first(sched_domain_span(child
));
6109 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6110 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6111 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6118 * build_sched_groups will build a circular linked list of the groups
6119 * covered by the given span, and will set each group's ->cpumask correctly,
6120 * and ->cpu_capacity to 0.
6122 * Assumes the sched_domain tree is fully constructed
6125 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6127 struct sched_group
*first
= NULL
, *last
= NULL
;
6128 struct sd_data
*sdd
= sd
->private;
6129 const struct cpumask
*span
= sched_domain_span(sd
);
6130 struct cpumask
*covered
;
6133 get_group(cpu
, sdd
, &sd
->groups
);
6134 atomic_inc(&sd
->groups
->ref
);
6136 if (cpu
!= cpumask_first(span
))
6139 lockdep_assert_held(&sched_domains_mutex
);
6140 covered
= sched_domains_tmpmask
;
6142 cpumask_clear(covered
);
6144 for_each_cpu(i
, span
) {
6145 struct sched_group
*sg
;
6148 if (cpumask_test_cpu(i
, covered
))
6151 group
= get_group(i
, sdd
, &sg
);
6152 cpumask_setall(sched_group_mask(sg
));
6154 for_each_cpu(j
, span
) {
6155 if (get_group(j
, sdd
, NULL
) != group
)
6158 cpumask_set_cpu(j
, covered
);
6159 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6174 * Initialize sched groups cpu_capacity.
6176 * cpu_capacity indicates the capacity of sched group, which is used while
6177 * distributing the load between different sched groups in a sched domain.
6178 * Typically cpu_capacity for all the groups in a sched domain will be same
6179 * unless there are asymmetries in the topology. If there are asymmetries,
6180 * group having more cpu_capacity will pickup more load compared to the
6181 * group having less cpu_capacity.
6183 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6185 struct sched_group
*sg
= sd
->groups
;
6190 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6192 } while (sg
!= sd
->groups
);
6194 if (cpu
!= group_balance_cpu(sg
))
6197 update_group_capacity(sd
, cpu
);
6198 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6202 * Initializers for schedule domains
6203 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6206 static int default_relax_domain_level
= -1;
6207 int sched_domain_level_max
;
6209 static int __init
setup_relax_domain_level(char *str
)
6211 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6212 pr_warn("Unable to set relax_domain_level\n");
6216 __setup("relax_domain_level=", setup_relax_domain_level
);
6218 static void set_domain_attribute(struct sched_domain
*sd
,
6219 struct sched_domain_attr
*attr
)
6223 if (!attr
|| attr
->relax_domain_level
< 0) {
6224 if (default_relax_domain_level
< 0)
6227 request
= default_relax_domain_level
;
6229 request
= attr
->relax_domain_level
;
6230 if (request
< sd
->level
) {
6231 /* turn off idle balance on this domain */
6232 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6234 /* turn on idle balance on this domain */
6235 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6239 static void __sdt_free(const struct cpumask
*cpu_map
);
6240 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6242 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6243 const struct cpumask
*cpu_map
)
6247 if (!atomic_read(&d
->rd
->refcount
))
6248 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6250 free_percpu(d
->sd
); /* fall through */
6252 __sdt_free(cpu_map
); /* fall through */
6258 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6259 const struct cpumask
*cpu_map
)
6261 memset(d
, 0, sizeof(*d
));
6263 if (__sdt_alloc(cpu_map
))
6264 return sa_sd_storage
;
6265 d
->sd
= alloc_percpu(struct sched_domain
*);
6267 return sa_sd_storage
;
6268 d
->rd
= alloc_rootdomain();
6271 return sa_rootdomain
;
6275 * NULL the sd_data elements we've used to build the sched_domain and
6276 * sched_group structure so that the subsequent __free_domain_allocs()
6277 * will not free the data we're using.
6279 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6281 struct sd_data
*sdd
= sd
->private;
6283 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6284 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6286 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6287 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6289 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6290 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6294 static int sched_domains_numa_levels
;
6295 enum numa_topology_type sched_numa_topology_type
;
6296 static int *sched_domains_numa_distance
;
6297 int sched_max_numa_distance
;
6298 static struct cpumask
***sched_domains_numa_masks
;
6299 static int sched_domains_curr_level
;
6303 * SD_flags allowed in topology descriptions.
6305 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6306 * SD_SHARE_PKG_RESOURCES - describes shared caches
6307 * SD_NUMA - describes NUMA topologies
6308 * SD_SHARE_POWERDOMAIN - describes shared power domain
6311 * SD_ASYM_PACKING - describes SMT quirks
6313 #define TOPOLOGY_SD_FLAGS \
6314 (SD_SHARE_CPUCAPACITY | \
6315 SD_SHARE_PKG_RESOURCES | \
6318 SD_SHARE_POWERDOMAIN)
6320 static struct sched_domain
*
6321 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6323 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6324 int sd_weight
, sd_flags
= 0;
6328 * Ugly hack to pass state to sd_numa_mask()...
6330 sched_domains_curr_level
= tl
->numa_level
;
6333 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6336 sd_flags
= (*tl
->sd_flags
)();
6337 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6338 "wrong sd_flags in topology description\n"))
6339 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6341 *sd
= (struct sched_domain
){
6342 .min_interval
= sd_weight
,
6343 .max_interval
= 2*sd_weight
,
6345 .imbalance_pct
= 125,
6347 .cache_nice_tries
= 0,
6354 .flags
= 1*SD_LOAD_BALANCE
6355 | 1*SD_BALANCE_NEWIDLE
6360 | 0*SD_SHARE_CPUCAPACITY
6361 | 0*SD_SHARE_PKG_RESOURCES
6363 | 0*SD_PREFER_SIBLING
6368 .last_balance
= jiffies
,
6369 .balance_interval
= sd_weight
,
6371 .max_newidle_lb_cost
= 0,
6372 .next_decay_max_lb_cost
= jiffies
,
6373 #ifdef CONFIG_SCHED_DEBUG
6379 * Convert topological properties into behaviour.
6382 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6383 sd
->flags
|= SD_PREFER_SIBLING
;
6384 sd
->imbalance_pct
= 110;
6385 sd
->smt_gain
= 1178; /* ~15% */
6387 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6388 sd
->imbalance_pct
= 117;
6389 sd
->cache_nice_tries
= 1;
6393 } else if (sd
->flags
& SD_NUMA
) {
6394 sd
->cache_nice_tries
= 2;
6398 sd
->flags
|= SD_SERIALIZE
;
6399 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6400 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6407 sd
->flags
|= SD_PREFER_SIBLING
;
6408 sd
->cache_nice_tries
= 1;
6413 sd
->private = &tl
->data
;
6419 * Topology list, bottom-up.
6421 static struct sched_domain_topology_level default_topology
[] = {
6422 #ifdef CONFIG_SCHED_SMT
6423 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6425 #ifdef CONFIG_SCHED_MC
6426 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6428 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6432 static struct sched_domain_topology_level
*sched_domain_topology
=
6435 #define for_each_sd_topology(tl) \
6436 for (tl = sched_domain_topology; tl->mask; tl++)
6438 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6440 sched_domain_topology
= tl
;
6445 static const struct cpumask
*sd_numa_mask(int cpu
)
6447 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6450 static void sched_numa_warn(const char *str
)
6452 static int done
= false;
6460 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6462 for (i
= 0; i
< nr_node_ids
; i
++) {
6463 printk(KERN_WARNING
" ");
6464 for (j
= 0; j
< nr_node_ids
; j
++)
6465 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6466 printk(KERN_CONT
"\n");
6468 printk(KERN_WARNING
"\n");
6471 bool find_numa_distance(int distance
)
6475 if (distance
== node_distance(0, 0))
6478 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6479 if (sched_domains_numa_distance
[i
] == distance
)
6487 * A system can have three types of NUMA topology:
6488 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6489 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6490 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6492 * The difference between a glueless mesh topology and a backplane
6493 * topology lies in whether communication between not directly
6494 * connected nodes goes through intermediary nodes (where programs
6495 * could run), or through backplane controllers. This affects
6496 * placement of programs.
6498 * The type of topology can be discerned with the following tests:
6499 * - If the maximum distance between any nodes is 1 hop, the system
6500 * is directly connected.
6501 * - If for two nodes A and B, located N > 1 hops away from each other,
6502 * there is an intermediary node C, which is < N hops away from both
6503 * nodes A and B, the system is a glueless mesh.
6505 static void init_numa_topology_type(void)
6509 n
= sched_max_numa_distance
;
6511 if (sched_domains_numa_levels
<= 1) {
6512 sched_numa_topology_type
= NUMA_DIRECT
;
6516 for_each_online_node(a
) {
6517 for_each_online_node(b
) {
6518 /* Find two nodes furthest removed from each other. */
6519 if (node_distance(a
, b
) < n
)
6522 /* Is there an intermediary node between a and b? */
6523 for_each_online_node(c
) {
6524 if (node_distance(a
, c
) < n
&&
6525 node_distance(b
, c
) < n
) {
6526 sched_numa_topology_type
=
6532 sched_numa_topology_type
= NUMA_BACKPLANE
;
6538 static void sched_init_numa(void)
6540 int next_distance
, curr_distance
= node_distance(0, 0);
6541 struct sched_domain_topology_level
*tl
;
6545 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6546 if (!sched_domains_numa_distance
)
6550 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6551 * unique distances in the node_distance() table.
6553 * Assumes node_distance(0,j) includes all distances in
6554 * node_distance(i,j) in order to avoid cubic time.
6556 next_distance
= curr_distance
;
6557 for (i
= 0; i
< nr_node_ids
; i
++) {
6558 for (j
= 0; j
< nr_node_ids
; j
++) {
6559 for (k
= 0; k
< nr_node_ids
; k
++) {
6560 int distance
= node_distance(i
, k
);
6562 if (distance
> curr_distance
&&
6563 (distance
< next_distance
||
6564 next_distance
== curr_distance
))
6565 next_distance
= distance
;
6568 * While not a strong assumption it would be nice to know
6569 * about cases where if node A is connected to B, B is not
6570 * equally connected to A.
6572 if (sched_debug() && node_distance(k
, i
) != distance
)
6573 sched_numa_warn("Node-distance not symmetric");
6575 if (sched_debug() && i
&& !find_numa_distance(distance
))
6576 sched_numa_warn("Node-0 not representative");
6578 if (next_distance
!= curr_distance
) {
6579 sched_domains_numa_distance
[level
++] = next_distance
;
6580 sched_domains_numa_levels
= level
;
6581 curr_distance
= next_distance
;
6586 * In case of sched_debug() we verify the above assumption.
6596 * 'level' contains the number of unique distances, excluding the
6597 * identity distance node_distance(i,i).
6599 * The sched_domains_numa_distance[] array includes the actual distance
6604 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6605 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6606 * the array will contain less then 'level' members. This could be
6607 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6608 * in other functions.
6610 * We reset it to 'level' at the end of this function.
6612 sched_domains_numa_levels
= 0;
6614 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6615 if (!sched_domains_numa_masks
)
6619 * Now for each level, construct a mask per node which contains all
6620 * cpus of nodes that are that many hops away from us.
6622 for (i
= 0; i
< level
; i
++) {
6623 sched_domains_numa_masks
[i
] =
6624 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6625 if (!sched_domains_numa_masks
[i
])
6628 for (j
= 0; j
< nr_node_ids
; j
++) {
6629 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6633 sched_domains_numa_masks
[i
][j
] = mask
;
6636 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6639 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6644 /* Compute default topology size */
6645 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6647 tl
= kzalloc((i
+ level
+ 1) *
6648 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6653 * Copy the default topology bits..
6655 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6656 tl
[i
] = sched_domain_topology
[i
];
6659 * .. and append 'j' levels of NUMA goodness.
6661 for (j
= 0; j
< level
; i
++, j
++) {
6662 tl
[i
] = (struct sched_domain_topology_level
){
6663 .mask
= sd_numa_mask
,
6664 .sd_flags
= cpu_numa_flags
,
6665 .flags
= SDTL_OVERLAP
,
6671 sched_domain_topology
= tl
;
6673 sched_domains_numa_levels
= level
;
6674 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6676 init_numa_topology_type();
6679 static void sched_domains_numa_masks_set(unsigned int cpu
)
6681 int node
= cpu_to_node(cpu
);
6684 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6685 for (j
= 0; j
< nr_node_ids
; j
++) {
6686 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6687 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6692 static void sched_domains_numa_masks_clear(unsigned int cpu
)
6696 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6697 for (j
= 0; j
< nr_node_ids
; j
++)
6698 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6703 static inline void sched_init_numa(void) { }
6704 static void sched_domains_numa_masks_set(unsigned int cpu
) { }
6705 static void sched_domains_numa_masks_clear(unsigned int cpu
) { }
6706 #endif /* CONFIG_NUMA */
6708 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6710 struct sched_domain_topology_level
*tl
;
6713 for_each_sd_topology(tl
) {
6714 struct sd_data
*sdd
= &tl
->data
;
6716 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6720 sdd
->sg
= alloc_percpu(struct sched_group
*);
6724 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6728 for_each_cpu(j
, cpu_map
) {
6729 struct sched_domain
*sd
;
6730 struct sched_group
*sg
;
6731 struct sched_group_capacity
*sgc
;
6733 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6734 GFP_KERNEL
, cpu_to_node(j
));
6738 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6740 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6741 GFP_KERNEL
, cpu_to_node(j
));
6747 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6749 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6750 GFP_KERNEL
, cpu_to_node(j
));
6754 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6761 static void __sdt_free(const struct cpumask
*cpu_map
)
6763 struct sched_domain_topology_level
*tl
;
6766 for_each_sd_topology(tl
) {
6767 struct sd_data
*sdd
= &tl
->data
;
6769 for_each_cpu(j
, cpu_map
) {
6770 struct sched_domain
*sd
;
6773 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6774 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6775 free_sched_groups(sd
->groups
, 0);
6776 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6780 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6782 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6784 free_percpu(sdd
->sd
);
6786 free_percpu(sdd
->sg
);
6788 free_percpu(sdd
->sgc
);
6793 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6794 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6795 struct sched_domain
*child
, int cpu
)
6797 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6801 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6803 sd
->level
= child
->level
+ 1;
6804 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6808 if (!cpumask_subset(sched_domain_span(child
),
6809 sched_domain_span(sd
))) {
6810 pr_err("BUG: arch topology borken\n");
6811 #ifdef CONFIG_SCHED_DEBUG
6812 pr_err(" the %s domain not a subset of the %s domain\n",
6813 child
->name
, sd
->name
);
6815 /* Fixup, ensure @sd has at least @child cpus. */
6816 cpumask_or(sched_domain_span(sd
),
6817 sched_domain_span(sd
),
6818 sched_domain_span(child
));
6822 set_domain_attribute(sd
, attr
);
6828 * Build sched domains for a given set of cpus and attach the sched domains
6829 * to the individual cpus
6831 static int build_sched_domains(const struct cpumask
*cpu_map
,
6832 struct sched_domain_attr
*attr
)
6834 enum s_alloc alloc_state
;
6835 struct sched_domain
*sd
;
6837 int i
, ret
= -ENOMEM
;
6839 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6840 if (alloc_state
!= sa_rootdomain
)
6843 /* Set up domains for cpus specified by the cpu_map. */
6844 for_each_cpu(i
, cpu_map
) {
6845 struct sched_domain_topology_level
*tl
;
6848 for_each_sd_topology(tl
) {
6849 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6850 if (tl
== sched_domain_topology
)
6851 *per_cpu_ptr(d
.sd
, i
) = sd
;
6852 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6853 sd
->flags
|= SD_OVERLAP
;
6854 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6859 /* Build the groups for the domains */
6860 for_each_cpu(i
, cpu_map
) {
6861 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6862 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6863 if (sd
->flags
& SD_OVERLAP
) {
6864 if (build_overlap_sched_groups(sd
, i
))
6867 if (build_sched_groups(sd
, i
))
6873 /* Calculate CPU capacity for physical packages and nodes */
6874 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6875 if (!cpumask_test_cpu(i
, cpu_map
))
6878 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6879 claim_allocations(i
, sd
);
6880 init_sched_groups_capacity(i
, sd
);
6884 /* Attach the domains */
6886 for_each_cpu(i
, cpu_map
) {
6887 sd
= *per_cpu_ptr(d
.sd
, i
);
6888 cpu_attach_domain(sd
, d
.rd
, i
);
6894 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6898 static cpumask_var_t
*doms_cur
; /* current sched domains */
6899 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6900 static struct sched_domain_attr
*dattr_cur
;
6901 /* attribues of custom domains in 'doms_cur' */
6904 * Special case: If a kmalloc of a doms_cur partition (array of
6905 * cpumask) fails, then fallback to a single sched domain,
6906 * as determined by the single cpumask fallback_doms.
6908 static cpumask_var_t fallback_doms
;
6911 * arch_update_cpu_topology lets virtualized architectures update the
6912 * cpu core maps. It is supposed to return 1 if the topology changed
6913 * or 0 if it stayed the same.
6915 int __weak
arch_update_cpu_topology(void)
6920 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6923 cpumask_var_t
*doms
;
6925 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6928 for (i
= 0; i
< ndoms
; i
++) {
6929 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6930 free_sched_domains(doms
, i
);
6937 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6940 for (i
= 0; i
< ndoms
; i
++)
6941 free_cpumask_var(doms
[i
]);
6946 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6947 * For now this just excludes isolated cpus, but could be used to
6948 * exclude other special cases in the future.
6950 static int init_sched_domains(const struct cpumask
*cpu_map
)
6954 arch_update_cpu_topology();
6956 doms_cur
= alloc_sched_domains(ndoms_cur
);
6958 doms_cur
= &fallback_doms
;
6959 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6960 err
= build_sched_domains(doms_cur
[0], NULL
);
6961 register_sched_domain_sysctl();
6967 * Detach sched domains from a group of cpus specified in cpu_map
6968 * These cpus will now be attached to the NULL domain
6970 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6975 for_each_cpu(i
, cpu_map
)
6976 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6980 /* handle null as "default" */
6981 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6982 struct sched_domain_attr
*new, int idx_new
)
6984 struct sched_domain_attr tmp
;
6991 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6992 new ? (new + idx_new
) : &tmp
,
6993 sizeof(struct sched_domain_attr
));
6997 * Partition sched domains as specified by the 'ndoms_new'
6998 * cpumasks in the array doms_new[] of cpumasks. This compares
6999 * doms_new[] to the current sched domain partitioning, doms_cur[].
7000 * It destroys each deleted domain and builds each new domain.
7002 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7003 * The masks don't intersect (don't overlap.) We should setup one
7004 * sched domain for each mask. CPUs not in any of the cpumasks will
7005 * not be load balanced. If the same cpumask appears both in the
7006 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7009 * The passed in 'doms_new' should be allocated using
7010 * alloc_sched_domains. This routine takes ownership of it and will
7011 * free_sched_domains it when done with it. If the caller failed the
7012 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7013 * and partition_sched_domains() will fallback to the single partition
7014 * 'fallback_doms', it also forces the domains to be rebuilt.
7016 * If doms_new == NULL it will be replaced with cpu_online_mask.
7017 * ndoms_new == 0 is a special case for destroying existing domains,
7018 * and it will not create the default domain.
7020 * Call with hotplug lock held
7022 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7023 struct sched_domain_attr
*dattr_new
)
7028 mutex_lock(&sched_domains_mutex
);
7030 /* always unregister in case we don't destroy any domains */
7031 unregister_sched_domain_sysctl();
7033 /* Let architecture update cpu core mappings. */
7034 new_topology
= arch_update_cpu_topology();
7036 n
= doms_new
? ndoms_new
: 0;
7038 /* Destroy deleted domains */
7039 for (i
= 0; i
< ndoms_cur
; i
++) {
7040 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7041 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7042 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7045 /* no match - a current sched domain not in new doms_new[] */
7046 detach_destroy_domains(doms_cur
[i
]);
7052 if (doms_new
== NULL
) {
7054 doms_new
= &fallback_doms
;
7055 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7056 WARN_ON_ONCE(dattr_new
);
7059 /* Build new domains */
7060 for (i
= 0; i
< ndoms_new
; i
++) {
7061 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7062 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7063 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7066 /* no match - add a new doms_new */
7067 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7072 /* Remember the new sched domains */
7073 if (doms_cur
!= &fallback_doms
)
7074 free_sched_domains(doms_cur
, ndoms_cur
);
7075 kfree(dattr_cur
); /* kfree(NULL) is safe */
7076 doms_cur
= doms_new
;
7077 dattr_cur
= dattr_new
;
7078 ndoms_cur
= ndoms_new
;
7080 register_sched_domain_sysctl();
7082 mutex_unlock(&sched_domains_mutex
);
7085 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7088 * Update cpusets according to cpu_active mask. If cpusets are
7089 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7090 * around partition_sched_domains().
7092 * If we come here as part of a suspend/resume, don't touch cpusets because we
7093 * want to restore it back to its original state upon resume anyway.
7095 static void cpuset_cpu_active(void)
7097 if (cpuhp_tasks_frozen
) {
7099 * num_cpus_frozen tracks how many CPUs are involved in suspend
7100 * resume sequence. As long as this is not the last online
7101 * operation in the resume sequence, just build a single sched
7102 * domain, ignoring cpusets.
7105 if (likely(num_cpus_frozen
)) {
7106 partition_sched_domains(1, NULL
, NULL
);
7110 * This is the last CPU online operation. So fall through and
7111 * restore the original sched domains by considering the
7112 * cpuset configurations.
7115 cpuset_update_active_cpus(true);
7118 static int cpuset_cpu_inactive(unsigned int cpu
)
7120 unsigned long flags
;
7125 if (!cpuhp_tasks_frozen
) {
7126 rcu_read_lock_sched();
7127 dl_b
= dl_bw_of(cpu
);
7129 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7130 cpus
= dl_bw_cpus(cpu
);
7131 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7132 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7134 rcu_read_unlock_sched();
7138 cpuset_update_active_cpus(false);
7141 partition_sched_domains(1, NULL
, NULL
);
7146 int sched_cpu_activate(unsigned int cpu
)
7148 struct rq
*rq
= cpu_rq(cpu
);
7149 unsigned long flags
;
7151 set_cpu_active(cpu
, true);
7153 if (sched_smp_initialized
) {
7154 sched_domains_numa_masks_set(cpu
);
7155 cpuset_cpu_active();
7159 * Put the rq online, if not already. This happens:
7161 * 1) In the early boot process, because we build the real domains
7162 * after all cpus have been brought up.
7164 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7167 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7169 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7172 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7174 update_max_interval();
7179 int sched_cpu_deactivate(unsigned int cpu
)
7183 set_cpu_active(cpu
, false);
7185 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7186 * users of this state to go away such that all new such users will
7189 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7190 * not imply sync_sched(), so wait for both.
7192 * Do sync before park smpboot threads to take care the rcu boost case.
7194 if (IS_ENABLED(CONFIG_PREEMPT
))
7195 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
7199 if (!sched_smp_initialized
)
7202 ret
= cpuset_cpu_inactive(cpu
);
7204 set_cpu_active(cpu
, true);
7207 sched_domains_numa_masks_clear(cpu
);
7211 static void sched_rq_cpu_starting(unsigned int cpu
)
7213 struct rq
*rq
= cpu_rq(cpu
);
7215 rq
->calc_load_update
= calc_load_update
;
7216 account_reset_rq(rq
);
7217 update_max_interval();
7220 int sched_cpu_starting(unsigned int cpu
)
7222 set_cpu_rq_start_time(cpu
);
7223 sched_rq_cpu_starting(cpu
);
7227 #ifdef CONFIG_HOTPLUG_CPU
7228 int sched_cpu_dying(unsigned int cpu
)
7230 struct rq
*rq
= cpu_rq(cpu
);
7231 unsigned long flags
;
7233 /* Handle pending wakeups and then migrate everything off */
7234 sched_ttwu_pending();
7235 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7237 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7241 BUG_ON(rq
->nr_running
!= 1);
7242 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7243 calc_load_migrate(rq
);
7244 update_max_interval();
7245 nohz_balance_exit_idle(cpu
);
7251 void __init
sched_init_smp(void)
7253 cpumask_var_t non_isolated_cpus
;
7255 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7256 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7261 * There's no userspace yet to cause hotplug operations; hence all the
7262 * cpu masks are stable and all blatant races in the below code cannot
7265 mutex_lock(&sched_domains_mutex
);
7266 init_sched_domains(cpu_active_mask
);
7267 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7268 if (cpumask_empty(non_isolated_cpus
))
7269 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7270 mutex_unlock(&sched_domains_mutex
);
7272 /* Move init over to a non-isolated CPU */
7273 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7275 sched_init_granularity();
7276 free_cpumask_var(non_isolated_cpus
);
7278 init_sched_rt_class();
7279 init_sched_dl_class();
7280 sched_smp_initialized
= true;
7283 static int __init
migration_init(void)
7285 sched_rq_cpu_starting(smp_processor_id());
7288 early_initcall(migration_init
);
7291 void __init
sched_init_smp(void)
7293 sched_init_granularity();
7295 #endif /* CONFIG_SMP */
7297 int in_sched_functions(unsigned long addr
)
7299 return in_lock_functions(addr
) ||
7300 (addr
>= (unsigned long)__sched_text_start
7301 && addr
< (unsigned long)__sched_text_end
);
7304 #ifdef CONFIG_CGROUP_SCHED
7306 * Default task group.
7307 * Every task in system belongs to this group at bootup.
7309 struct task_group root_task_group
;
7310 LIST_HEAD(task_groups
);
7312 /* Cacheline aligned slab cache for task_group */
7313 static struct kmem_cache
*task_group_cache __read_mostly
;
7316 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7318 void __init
sched_init(void)
7321 unsigned long alloc_size
= 0, ptr
;
7323 #ifdef CONFIG_FAIR_GROUP_SCHED
7324 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7326 #ifdef CONFIG_RT_GROUP_SCHED
7327 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7330 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7332 #ifdef CONFIG_FAIR_GROUP_SCHED
7333 root_task_group
.se
= (struct sched_entity
**)ptr
;
7334 ptr
+= nr_cpu_ids
* sizeof(void **);
7336 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7337 ptr
+= nr_cpu_ids
* sizeof(void **);
7339 #endif /* CONFIG_FAIR_GROUP_SCHED */
7340 #ifdef CONFIG_RT_GROUP_SCHED
7341 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7342 ptr
+= nr_cpu_ids
* sizeof(void **);
7344 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7345 ptr
+= nr_cpu_ids
* sizeof(void **);
7347 #endif /* CONFIG_RT_GROUP_SCHED */
7349 #ifdef CONFIG_CPUMASK_OFFSTACK
7350 for_each_possible_cpu(i
) {
7351 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7352 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7354 #endif /* CONFIG_CPUMASK_OFFSTACK */
7356 init_rt_bandwidth(&def_rt_bandwidth
,
7357 global_rt_period(), global_rt_runtime());
7358 init_dl_bandwidth(&def_dl_bandwidth
,
7359 global_rt_period(), global_rt_runtime());
7362 init_defrootdomain();
7365 #ifdef CONFIG_RT_GROUP_SCHED
7366 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7367 global_rt_period(), global_rt_runtime());
7368 #endif /* CONFIG_RT_GROUP_SCHED */
7370 #ifdef CONFIG_CGROUP_SCHED
7371 task_group_cache
= KMEM_CACHE(task_group
, 0);
7373 list_add(&root_task_group
.list
, &task_groups
);
7374 INIT_LIST_HEAD(&root_task_group
.children
);
7375 INIT_LIST_HEAD(&root_task_group
.siblings
);
7376 autogroup_init(&init_task
);
7377 #endif /* CONFIG_CGROUP_SCHED */
7379 for_each_possible_cpu(i
) {
7383 raw_spin_lock_init(&rq
->lock
);
7385 rq
->calc_load_active
= 0;
7386 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7387 init_cfs_rq(&rq
->cfs
);
7388 init_rt_rq(&rq
->rt
);
7389 init_dl_rq(&rq
->dl
);
7390 #ifdef CONFIG_FAIR_GROUP_SCHED
7391 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7392 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7394 * How much cpu bandwidth does root_task_group get?
7396 * In case of task-groups formed thr' the cgroup filesystem, it
7397 * gets 100% of the cpu resources in the system. This overall
7398 * system cpu resource is divided among the tasks of
7399 * root_task_group and its child task-groups in a fair manner,
7400 * based on each entity's (task or task-group's) weight
7401 * (se->load.weight).
7403 * In other words, if root_task_group has 10 tasks of weight
7404 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7405 * then A0's share of the cpu resource is:
7407 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7409 * We achieve this by letting root_task_group's tasks sit
7410 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7412 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7413 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7414 #endif /* CONFIG_FAIR_GROUP_SCHED */
7416 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7417 #ifdef CONFIG_RT_GROUP_SCHED
7418 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7421 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7422 rq
->cpu_load
[j
] = 0;
7427 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7428 rq
->balance_callback
= NULL
;
7429 rq
->active_balance
= 0;
7430 rq
->next_balance
= jiffies
;
7435 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7436 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7438 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7440 rq_attach_root(rq
, &def_root_domain
);
7441 #ifdef CONFIG_NO_HZ_COMMON
7442 rq
->last_load_update_tick
= jiffies
;
7445 #ifdef CONFIG_NO_HZ_FULL
7446 rq
->last_sched_tick
= 0;
7448 #endif /* CONFIG_SMP */
7450 atomic_set(&rq
->nr_iowait
, 0);
7453 set_load_weight(&init_task
);
7455 #ifdef CONFIG_PREEMPT_NOTIFIERS
7456 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7460 * The boot idle thread does lazy MMU switching as well:
7462 atomic_inc(&init_mm
.mm_count
);
7463 enter_lazy_tlb(&init_mm
, current
);
7466 * During early bootup we pretend to be a normal task:
7468 current
->sched_class
= &fair_sched_class
;
7471 * Make us the idle thread. Technically, schedule() should not be
7472 * called from this thread, however somewhere below it might be,
7473 * but because we are the idle thread, we just pick up running again
7474 * when this runqueue becomes "idle".
7476 init_idle(current
, smp_processor_id());
7478 calc_load_update
= jiffies
+ LOAD_FREQ
;
7481 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7482 /* May be allocated at isolcpus cmdline parse time */
7483 if (cpu_isolated_map
== NULL
)
7484 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7485 idle_thread_set_boot_cpu();
7486 set_cpu_rq_start_time(smp_processor_id());
7488 init_sched_fair_class();
7490 scheduler_running
= 1;
7493 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7494 static inline int preempt_count_equals(int preempt_offset
)
7496 int nested
= preempt_count() + rcu_preempt_depth();
7498 return (nested
== preempt_offset
);
7501 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7504 * Blocking primitives will set (and therefore destroy) current->state,
7505 * since we will exit with TASK_RUNNING make sure we enter with it,
7506 * otherwise we will destroy state.
7508 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7509 "do not call blocking ops when !TASK_RUNNING; "
7510 "state=%lx set at [<%p>] %pS\n",
7512 (void *)current
->task_state_change
,
7513 (void *)current
->task_state_change
);
7515 ___might_sleep(file
, line
, preempt_offset
);
7517 EXPORT_SYMBOL(__might_sleep
);
7519 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7521 static unsigned long prev_jiffy
; /* ratelimiting */
7523 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7524 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7525 !is_idle_task(current
)) ||
7526 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7528 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7530 prev_jiffy
= jiffies
;
7533 "BUG: sleeping function called from invalid context at %s:%d\n",
7536 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7537 in_atomic(), irqs_disabled(),
7538 current
->pid
, current
->comm
);
7540 if (task_stack_end_corrupted(current
))
7541 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7543 debug_show_held_locks(current
);
7544 if (irqs_disabled())
7545 print_irqtrace_events(current
);
7546 #ifdef CONFIG_DEBUG_PREEMPT
7547 if (!preempt_count_equals(preempt_offset
)) {
7548 pr_err("Preemption disabled at:");
7549 print_ip_sym(current
->preempt_disable_ip
);
7555 EXPORT_SYMBOL(___might_sleep
);
7558 #ifdef CONFIG_MAGIC_SYSRQ
7559 void normalize_rt_tasks(void)
7561 struct task_struct
*g
, *p
;
7562 struct sched_attr attr
= {
7563 .sched_policy
= SCHED_NORMAL
,
7566 read_lock(&tasklist_lock
);
7567 for_each_process_thread(g
, p
) {
7569 * Only normalize user tasks:
7571 if (p
->flags
& PF_KTHREAD
)
7574 p
->se
.exec_start
= 0;
7575 #ifdef CONFIG_SCHEDSTATS
7576 p
->se
.statistics
.wait_start
= 0;
7577 p
->se
.statistics
.sleep_start
= 0;
7578 p
->se
.statistics
.block_start
= 0;
7581 if (!dl_task(p
) && !rt_task(p
)) {
7583 * Renice negative nice level userspace
7586 if (task_nice(p
) < 0)
7587 set_user_nice(p
, 0);
7591 __sched_setscheduler(p
, &attr
, false, false);
7593 read_unlock(&tasklist_lock
);
7596 #endif /* CONFIG_MAGIC_SYSRQ */
7598 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7600 * These functions are only useful for the IA64 MCA handling, or kdb.
7602 * They can only be called when the whole system has been
7603 * stopped - every CPU needs to be quiescent, and no scheduling
7604 * activity can take place. Using them for anything else would
7605 * be a serious bug, and as a result, they aren't even visible
7606 * under any other configuration.
7610 * curr_task - return the current task for a given cpu.
7611 * @cpu: the processor in question.
7613 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7615 * Return: The current task for @cpu.
7617 struct task_struct
*curr_task(int cpu
)
7619 return cpu_curr(cpu
);
7622 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7626 * set_curr_task - set the current task for a given cpu.
7627 * @cpu: the processor in question.
7628 * @p: the task pointer to set.
7630 * Description: This function must only be used when non-maskable interrupts
7631 * are serviced on a separate stack. It allows the architecture to switch the
7632 * notion of the current task on a cpu in a non-blocking manner. This function
7633 * must be called with all CPU's synchronized, and interrupts disabled, the
7634 * and caller must save the original value of the current task (see
7635 * curr_task() above) and restore that value before reenabling interrupts and
7636 * re-starting the system.
7638 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7640 void set_curr_task(int cpu
, struct task_struct
*p
)
7647 #ifdef CONFIG_CGROUP_SCHED
7648 /* task_group_lock serializes the addition/removal of task groups */
7649 static DEFINE_SPINLOCK(task_group_lock
);
7651 static void sched_free_group(struct task_group
*tg
)
7653 free_fair_sched_group(tg
);
7654 free_rt_sched_group(tg
);
7656 kmem_cache_free(task_group_cache
, tg
);
7659 /* allocate runqueue etc for a new task group */
7660 struct task_group
*sched_create_group(struct task_group
*parent
)
7662 struct task_group
*tg
;
7664 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7666 return ERR_PTR(-ENOMEM
);
7668 if (!alloc_fair_sched_group(tg
, parent
))
7671 if (!alloc_rt_sched_group(tg
, parent
))
7677 sched_free_group(tg
);
7678 return ERR_PTR(-ENOMEM
);
7681 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7683 unsigned long flags
;
7685 spin_lock_irqsave(&task_group_lock
, flags
);
7686 list_add_rcu(&tg
->list
, &task_groups
);
7688 WARN_ON(!parent
); /* root should already exist */
7690 tg
->parent
= parent
;
7691 INIT_LIST_HEAD(&tg
->children
);
7692 list_add_rcu(&tg
->siblings
, &parent
->children
);
7693 spin_unlock_irqrestore(&task_group_lock
, flags
);
7696 /* rcu callback to free various structures associated with a task group */
7697 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7699 /* now it should be safe to free those cfs_rqs */
7700 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7703 void sched_destroy_group(struct task_group
*tg
)
7705 /* wait for possible concurrent references to cfs_rqs complete */
7706 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7709 void sched_offline_group(struct task_group
*tg
)
7711 unsigned long flags
;
7713 /* end participation in shares distribution */
7714 unregister_fair_sched_group(tg
);
7716 spin_lock_irqsave(&task_group_lock
, flags
);
7717 list_del_rcu(&tg
->list
);
7718 list_del_rcu(&tg
->siblings
);
7719 spin_unlock_irqrestore(&task_group_lock
, flags
);
7722 /* change task's runqueue when it moves between groups.
7723 * The caller of this function should have put the task in its new group
7724 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7725 * reflect its new group.
7727 void sched_move_task(struct task_struct
*tsk
)
7729 struct task_group
*tg
;
7730 int queued
, running
;
7734 rq
= task_rq_lock(tsk
, &rf
);
7736 running
= task_current(rq
, tsk
);
7737 queued
= task_on_rq_queued(tsk
);
7740 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7741 if (unlikely(running
))
7742 put_prev_task(rq
, tsk
);
7745 * All callers are synchronized by task_rq_lock(); we do not use RCU
7746 * which is pointless here. Thus, we pass "true" to task_css_check()
7747 * to prevent lockdep warnings.
7749 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7750 struct task_group
, css
);
7751 tg
= autogroup_task_group(tsk
, tg
);
7752 tsk
->sched_task_group
= tg
;
7754 #ifdef CONFIG_FAIR_GROUP_SCHED
7755 if (tsk
->sched_class
->task_move_group
)
7756 tsk
->sched_class
->task_move_group(tsk
);
7759 set_task_rq(tsk
, task_cpu(tsk
));
7761 if (unlikely(running
))
7762 tsk
->sched_class
->set_curr_task(rq
);
7764 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7766 task_rq_unlock(rq
, tsk
, &rf
);
7768 #endif /* CONFIG_CGROUP_SCHED */
7770 #ifdef CONFIG_RT_GROUP_SCHED
7772 * Ensure that the real time constraints are schedulable.
7774 static DEFINE_MUTEX(rt_constraints_mutex
);
7776 /* Must be called with tasklist_lock held */
7777 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7779 struct task_struct
*g
, *p
;
7782 * Autogroups do not have RT tasks; see autogroup_create().
7784 if (task_group_is_autogroup(tg
))
7787 for_each_process_thread(g
, p
) {
7788 if (rt_task(p
) && task_group(p
) == tg
)
7795 struct rt_schedulable_data
{
7796 struct task_group
*tg
;
7801 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7803 struct rt_schedulable_data
*d
= data
;
7804 struct task_group
*child
;
7805 unsigned long total
, sum
= 0;
7806 u64 period
, runtime
;
7808 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7809 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7812 period
= d
->rt_period
;
7813 runtime
= d
->rt_runtime
;
7817 * Cannot have more runtime than the period.
7819 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7823 * Ensure we don't starve existing RT tasks.
7825 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7828 total
= to_ratio(period
, runtime
);
7831 * Nobody can have more than the global setting allows.
7833 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7837 * The sum of our children's runtime should not exceed our own.
7839 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7840 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7841 runtime
= child
->rt_bandwidth
.rt_runtime
;
7843 if (child
== d
->tg
) {
7844 period
= d
->rt_period
;
7845 runtime
= d
->rt_runtime
;
7848 sum
+= to_ratio(period
, runtime
);
7857 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7861 struct rt_schedulable_data data
= {
7863 .rt_period
= period
,
7864 .rt_runtime
= runtime
,
7868 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7874 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7875 u64 rt_period
, u64 rt_runtime
)
7880 * Disallowing the root group RT runtime is BAD, it would disallow the
7881 * kernel creating (and or operating) RT threads.
7883 if (tg
== &root_task_group
&& rt_runtime
== 0)
7886 /* No period doesn't make any sense. */
7890 mutex_lock(&rt_constraints_mutex
);
7891 read_lock(&tasklist_lock
);
7892 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7896 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7897 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7898 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7900 for_each_possible_cpu(i
) {
7901 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7903 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7904 rt_rq
->rt_runtime
= rt_runtime
;
7905 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7907 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7909 read_unlock(&tasklist_lock
);
7910 mutex_unlock(&rt_constraints_mutex
);
7915 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7917 u64 rt_runtime
, rt_period
;
7919 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7920 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7921 if (rt_runtime_us
< 0)
7922 rt_runtime
= RUNTIME_INF
;
7924 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7927 static long sched_group_rt_runtime(struct task_group
*tg
)
7931 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7934 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7935 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7936 return rt_runtime_us
;
7939 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7941 u64 rt_runtime
, rt_period
;
7943 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7944 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7946 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7949 static long sched_group_rt_period(struct task_group
*tg
)
7953 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7954 do_div(rt_period_us
, NSEC_PER_USEC
);
7955 return rt_period_us
;
7957 #endif /* CONFIG_RT_GROUP_SCHED */
7959 #ifdef CONFIG_RT_GROUP_SCHED
7960 static int sched_rt_global_constraints(void)
7964 mutex_lock(&rt_constraints_mutex
);
7965 read_lock(&tasklist_lock
);
7966 ret
= __rt_schedulable(NULL
, 0, 0);
7967 read_unlock(&tasklist_lock
);
7968 mutex_unlock(&rt_constraints_mutex
);
7973 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7975 /* Don't accept realtime tasks when there is no way for them to run */
7976 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7982 #else /* !CONFIG_RT_GROUP_SCHED */
7983 static int sched_rt_global_constraints(void)
7985 unsigned long flags
;
7988 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7989 for_each_possible_cpu(i
) {
7990 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7992 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7993 rt_rq
->rt_runtime
= global_rt_runtime();
7994 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7996 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8000 #endif /* CONFIG_RT_GROUP_SCHED */
8002 static int sched_dl_global_validate(void)
8004 u64 runtime
= global_rt_runtime();
8005 u64 period
= global_rt_period();
8006 u64 new_bw
= to_ratio(period
, runtime
);
8009 unsigned long flags
;
8012 * Here we want to check the bandwidth not being set to some
8013 * value smaller than the currently allocated bandwidth in
8014 * any of the root_domains.
8016 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8017 * cycling on root_domains... Discussion on different/better
8018 * solutions is welcome!
8020 for_each_possible_cpu(cpu
) {
8021 rcu_read_lock_sched();
8022 dl_b
= dl_bw_of(cpu
);
8024 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8025 if (new_bw
< dl_b
->total_bw
)
8027 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8029 rcu_read_unlock_sched();
8038 static void sched_dl_do_global(void)
8043 unsigned long flags
;
8045 def_dl_bandwidth
.dl_period
= global_rt_period();
8046 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8048 if (global_rt_runtime() != RUNTIME_INF
)
8049 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8052 * FIXME: As above...
8054 for_each_possible_cpu(cpu
) {
8055 rcu_read_lock_sched();
8056 dl_b
= dl_bw_of(cpu
);
8058 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8060 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8062 rcu_read_unlock_sched();
8066 static int sched_rt_global_validate(void)
8068 if (sysctl_sched_rt_period
<= 0)
8071 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8072 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8078 static void sched_rt_do_global(void)
8080 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8081 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8084 int sched_rt_handler(struct ctl_table
*table
, int write
,
8085 void __user
*buffer
, size_t *lenp
,
8088 int old_period
, old_runtime
;
8089 static DEFINE_MUTEX(mutex
);
8093 old_period
= sysctl_sched_rt_period
;
8094 old_runtime
= sysctl_sched_rt_runtime
;
8096 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8098 if (!ret
&& write
) {
8099 ret
= sched_rt_global_validate();
8103 ret
= sched_dl_global_validate();
8107 ret
= sched_rt_global_constraints();
8111 sched_rt_do_global();
8112 sched_dl_do_global();
8116 sysctl_sched_rt_period
= old_period
;
8117 sysctl_sched_rt_runtime
= old_runtime
;
8119 mutex_unlock(&mutex
);
8124 int sched_rr_handler(struct ctl_table
*table
, int write
,
8125 void __user
*buffer
, size_t *lenp
,
8129 static DEFINE_MUTEX(mutex
);
8132 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8133 /* make sure that internally we keep jiffies */
8134 /* also, writing zero resets timeslice to default */
8135 if (!ret
&& write
) {
8136 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8137 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8139 mutex_unlock(&mutex
);
8143 #ifdef CONFIG_CGROUP_SCHED
8145 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8147 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8150 static struct cgroup_subsys_state
*
8151 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8153 struct task_group
*parent
= css_tg(parent_css
);
8154 struct task_group
*tg
;
8157 /* This is early initialization for the top cgroup */
8158 return &root_task_group
.css
;
8161 tg
= sched_create_group(parent
);
8163 return ERR_PTR(-ENOMEM
);
8165 sched_online_group(tg
, parent
);
8170 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8172 struct task_group
*tg
= css_tg(css
);
8174 sched_offline_group(tg
);
8177 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8179 struct task_group
*tg
= css_tg(css
);
8182 * Relies on the RCU grace period between css_released() and this.
8184 sched_free_group(tg
);
8187 static void cpu_cgroup_fork(struct task_struct
*task
)
8189 sched_move_task(task
);
8192 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8194 struct task_struct
*task
;
8195 struct cgroup_subsys_state
*css
;
8197 cgroup_taskset_for_each(task
, css
, tset
) {
8198 #ifdef CONFIG_RT_GROUP_SCHED
8199 if (!sched_rt_can_attach(css_tg(css
), task
))
8202 /* We don't support RT-tasks being in separate groups */
8203 if (task
->sched_class
!= &fair_sched_class
)
8210 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8212 struct task_struct
*task
;
8213 struct cgroup_subsys_state
*css
;
8215 cgroup_taskset_for_each(task
, css
, tset
)
8216 sched_move_task(task
);
8219 #ifdef CONFIG_FAIR_GROUP_SCHED
8220 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8221 struct cftype
*cftype
, u64 shareval
)
8223 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8226 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8229 struct task_group
*tg
= css_tg(css
);
8231 return (u64
) scale_load_down(tg
->shares
);
8234 #ifdef CONFIG_CFS_BANDWIDTH
8235 static DEFINE_MUTEX(cfs_constraints_mutex
);
8237 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8238 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8240 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8242 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8244 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8245 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8247 if (tg
== &root_task_group
)
8251 * Ensure we have at some amount of bandwidth every period. This is
8252 * to prevent reaching a state of large arrears when throttled via
8253 * entity_tick() resulting in prolonged exit starvation.
8255 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8259 * Likewise, bound things on the otherside by preventing insane quota
8260 * periods. This also allows us to normalize in computing quota
8263 if (period
> max_cfs_quota_period
)
8267 * Prevent race between setting of cfs_rq->runtime_enabled and
8268 * unthrottle_offline_cfs_rqs().
8271 mutex_lock(&cfs_constraints_mutex
);
8272 ret
= __cfs_schedulable(tg
, period
, quota
);
8276 runtime_enabled
= quota
!= RUNTIME_INF
;
8277 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8279 * If we need to toggle cfs_bandwidth_used, off->on must occur
8280 * before making related changes, and on->off must occur afterwards
8282 if (runtime_enabled
&& !runtime_was_enabled
)
8283 cfs_bandwidth_usage_inc();
8284 raw_spin_lock_irq(&cfs_b
->lock
);
8285 cfs_b
->period
= ns_to_ktime(period
);
8286 cfs_b
->quota
= quota
;
8288 __refill_cfs_bandwidth_runtime(cfs_b
);
8289 /* restart the period timer (if active) to handle new period expiry */
8290 if (runtime_enabled
)
8291 start_cfs_bandwidth(cfs_b
);
8292 raw_spin_unlock_irq(&cfs_b
->lock
);
8294 for_each_online_cpu(i
) {
8295 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8296 struct rq
*rq
= cfs_rq
->rq
;
8298 raw_spin_lock_irq(&rq
->lock
);
8299 cfs_rq
->runtime_enabled
= runtime_enabled
;
8300 cfs_rq
->runtime_remaining
= 0;
8302 if (cfs_rq
->throttled
)
8303 unthrottle_cfs_rq(cfs_rq
);
8304 raw_spin_unlock_irq(&rq
->lock
);
8306 if (runtime_was_enabled
&& !runtime_enabled
)
8307 cfs_bandwidth_usage_dec();
8309 mutex_unlock(&cfs_constraints_mutex
);
8315 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8319 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8320 if (cfs_quota_us
< 0)
8321 quota
= RUNTIME_INF
;
8323 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8325 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8328 long tg_get_cfs_quota(struct task_group
*tg
)
8332 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8335 quota_us
= tg
->cfs_bandwidth
.quota
;
8336 do_div(quota_us
, NSEC_PER_USEC
);
8341 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8345 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8346 quota
= tg
->cfs_bandwidth
.quota
;
8348 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8351 long tg_get_cfs_period(struct task_group
*tg
)
8355 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8356 do_div(cfs_period_us
, NSEC_PER_USEC
);
8358 return cfs_period_us
;
8361 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8364 return tg_get_cfs_quota(css_tg(css
));
8367 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8368 struct cftype
*cftype
, s64 cfs_quota_us
)
8370 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8373 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8376 return tg_get_cfs_period(css_tg(css
));
8379 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8380 struct cftype
*cftype
, u64 cfs_period_us
)
8382 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8385 struct cfs_schedulable_data
{
8386 struct task_group
*tg
;
8391 * normalize group quota/period to be quota/max_period
8392 * note: units are usecs
8394 static u64
normalize_cfs_quota(struct task_group
*tg
,
8395 struct cfs_schedulable_data
*d
)
8403 period
= tg_get_cfs_period(tg
);
8404 quota
= tg_get_cfs_quota(tg
);
8407 /* note: these should typically be equivalent */
8408 if (quota
== RUNTIME_INF
|| quota
== -1)
8411 return to_ratio(period
, quota
);
8414 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8416 struct cfs_schedulable_data
*d
= data
;
8417 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8418 s64 quota
= 0, parent_quota
= -1;
8421 quota
= RUNTIME_INF
;
8423 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8425 quota
= normalize_cfs_quota(tg
, d
);
8426 parent_quota
= parent_b
->hierarchical_quota
;
8429 * ensure max(child_quota) <= parent_quota, inherit when no
8432 if (quota
== RUNTIME_INF
)
8433 quota
= parent_quota
;
8434 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8437 cfs_b
->hierarchical_quota
= quota
;
8442 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8445 struct cfs_schedulable_data data
= {
8451 if (quota
!= RUNTIME_INF
) {
8452 do_div(data
.period
, NSEC_PER_USEC
);
8453 do_div(data
.quota
, NSEC_PER_USEC
);
8457 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8463 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8465 struct task_group
*tg
= css_tg(seq_css(sf
));
8466 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8468 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8469 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8470 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8474 #endif /* CONFIG_CFS_BANDWIDTH */
8475 #endif /* CONFIG_FAIR_GROUP_SCHED */
8477 #ifdef CONFIG_RT_GROUP_SCHED
8478 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8479 struct cftype
*cft
, s64 val
)
8481 return sched_group_set_rt_runtime(css_tg(css
), val
);
8484 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8487 return sched_group_rt_runtime(css_tg(css
));
8490 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8491 struct cftype
*cftype
, u64 rt_period_us
)
8493 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8496 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8499 return sched_group_rt_period(css_tg(css
));
8501 #endif /* CONFIG_RT_GROUP_SCHED */
8503 static struct cftype cpu_files
[] = {
8504 #ifdef CONFIG_FAIR_GROUP_SCHED
8507 .read_u64
= cpu_shares_read_u64
,
8508 .write_u64
= cpu_shares_write_u64
,
8511 #ifdef CONFIG_CFS_BANDWIDTH
8513 .name
= "cfs_quota_us",
8514 .read_s64
= cpu_cfs_quota_read_s64
,
8515 .write_s64
= cpu_cfs_quota_write_s64
,
8518 .name
= "cfs_period_us",
8519 .read_u64
= cpu_cfs_period_read_u64
,
8520 .write_u64
= cpu_cfs_period_write_u64
,
8524 .seq_show
= cpu_stats_show
,
8527 #ifdef CONFIG_RT_GROUP_SCHED
8529 .name
= "rt_runtime_us",
8530 .read_s64
= cpu_rt_runtime_read
,
8531 .write_s64
= cpu_rt_runtime_write
,
8534 .name
= "rt_period_us",
8535 .read_u64
= cpu_rt_period_read_uint
,
8536 .write_u64
= cpu_rt_period_write_uint
,
8542 struct cgroup_subsys cpu_cgrp_subsys
= {
8543 .css_alloc
= cpu_cgroup_css_alloc
,
8544 .css_released
= cpu_cgroup_css_released
,
8545 .css_free
= cpu_cgroup_css_free
,
8546 .fork
= cpu_cgroup_fork
,
8547 .can_attach
= cpu_cgroup_can_attach
,
8548 .attach
= cpu_cgroup_attach
,
8549 .legacy_cftypes
= cpu_files
,
8553 #endif /* CONFIG_CGROUP_SCHED */
8555 void dump_cpu_task(int cpu
)
8557 pr_info("Task dump for CPU %d:\n", cpu
);
8558 sched_show_task(cpu_curr(cpu
));
8562 * Nice levels are multiplicative, with a gentle 10% change for every
8563 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8564 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8565 * that remained on nice 0.
8567 * The "10% effect" is relative and cumulative: from _any_ nice level,
8568 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8569 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8570 * If a task goes up by ~10% and another task goes down by ~10% then
8571 * the relative distance between them is ~25%.)
8573 const int sched_prio_to_weight
[40] = {
8574 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8575 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8576 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8577 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8578 /* 0 */ 1024, 820, 655, 526, 423,
8579 /* 5 */ 335, 272, 215, 172, 137,
8580 /* 10 */ 110, 87, 70, 56, 45,
8581 /* 15 */ 36, 29, 23, 18, 15,
8585 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8587 * In cases where the weight does not change often, we can use the
8588 * precalculated inverse to speed up arithmetics by turning divisions
8589 * into multiplications:
8591 const u32 sched_prio_to_wmult
[40] = {
8592 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8593 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8594 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8595 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8596 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8597 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8598 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8599 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,