4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
78 #include <linux/mutex.h>
80 #include <asm/switch_to.h>
82 #include <asm/irq_regs.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex
);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
97 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
99 void update_rq_clock(struct rq
*rq
)
103 lockdep_assert_held(&rq
->lock
);
105 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
108 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
112 update_rq_clock_task(rq
, delta
);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug
unsigned int sysctl_sched_features
=
123 #include "features.h"
129 * Number of tasks to iterate in a single balance run.
130 * Limited because this is done with IRQs disabled.
132 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
135 * period over which we average the RT time consumption, measured
140 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
143 * period over which we measure -rt task cpu usage in us.
146 unsigned int sysctl_sched_rt_period
= 1000000;
148 __read_mostly
int scheduler_running
;
151 * part of the period that we allow rt tasks to run in us.
154 int sysctl_sched_rt_runtime
= 950000;
156 /* cpus with isolated domains */
157 cpumask_var_t cpu_isolated_map
;
160 * this_rq_lock - lock this runqueue and disable interrupts.
162 static struct rq
*this_rq_lock(void)
169 raw_spin_lock(&rq
->lock
);
175 * __task_rq_lock - lock the rq @p resides on.
177 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
182 lockdep_assert_held(&p
->pi_lock
);
186 raw_spin_lock(&rq
->lock
);
187 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
188 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
191 raw_spin_unlock(&rq
->lock
);
193 while (unlikely(task_on_rq_migrating(p
)))
199 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
201 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
202 __acquires(p
->pi_lock
)
208 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
210 raw_spin_lock(&rq
->lock
);
212 * move_queued_task() task_rq_lock()
215 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
216 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
217 * [S] ->cpu = new_cpu [L] task_rq()
221 * If we observe the old cpu in task_rq_lock, the acquire of
222 * the old rq->lock will fully serialize against the stores.
224 * If we observe the new cpu in task_rq_lock, the acquire will
225 * pair with the WMB to ensure we must then also see migrating.
227 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
228 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
231 raw_spin_unlock(&rq
->lock
);
232 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
234 while (unlikely(task_on_rq_migrating(p
)))
239 #ifdef CONFIG_SCHED_HRTICK
241 * Use HR-timers to deliver accurate preemption points.
244 static void hrtick_clear(struct rq
*rq
)
246 if (hrtimer_active(&rq
->hrtick_timer
))
247 hrtimer_cancel(&rq
->hrtick_timer
);
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
254 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
256 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
258 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
260 raw_spin_lock(&rq
->lock
);
262 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
263 raw_spin_unlock(&rq
->lock
);
265 return HRTIMER_NORESTART
;
270 static void __hrtick_restart(struct rq
*rq
)
272 struct hrtimer
*timer
= &rq
->hrtick_timer
;
274 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
278 * called from hardirq (IPI) context
280 static void __hrtick_start(void *arg
)
284 raw_spin_lock(&rq
->lock
);
285 __hrtick_restart(rq
);
286 rq
->hrtick_csd_pending
= 0;
287 raw_spin_unlock(&rq
->lock
);
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq
*rq
, u64 delay
)
297 struct hrtimer
*timer
= &rq
->hrtick_timer
;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta
= max_t(s64
, delay
, 10000LL);
306 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
308 hrtimer_set_expires(timer
, time
);
310 if (rq
== this_rq()) {
311 __hrtick_restart(rq
);
312 } else if (!rq
->hrtick_csd_pending
) {
313 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
314 rq
->hrtick_csd_pending
= 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq
*rq
, u64 delay
)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay
= max_t(u64
, delay
, 10000LL);
331 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
332 HRTIMER_MODE_REL_PINNED
);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq
*rq
)
339 rq
->hrtick_csd_pending
= 0;
341 rq
->hrtick_csd
.flags
= 0;
342 rq
->hrtick_csd
.func
= __hrtick_start
;
343 rq
->hrtick_csd
.info
= rq
;
346 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
347 rq
->hrtick_timer
.function
= hrtick
;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq
*rq
)
354 static inline void init_rq_hrtick(struct rq
*rq
)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct
*p
)
385 struct thread_info
*ti
= task_thread_info(p
);
386 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct
*p
)
397 struct thread_info
*ti
= task_thread_info(p
);
398 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
401 if (!(val
& _TIF_POLLING_NRFLAG
))
403 if (val
& _TIF_NEED_RESCHED
)
405 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
414 static bool set_nr_and_not_polling(struct task_struct
*p
)
416 set_tsk_need_resched(p
);
421 static bool set_nr_if_polling(struct task_struct
*p
)
428 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
430 struct wake_q_node
*node
= &task
->wake_q
;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
443 get_task_struct(task
);
446 * The head is context local, there can be no concurrency.
449 head
->lastp
= &node
->next
;
452 void wake_up_q(struct wake_q_head
*head
)
454 struct wake_q_node
*node
= head
->first
;
456 while (node
!= WAKE_Q_TAIL
) {
457 struct task_struct
*task
;
459 task
= container_of(node
, struct task_struct
, wake_q
);
461 /* task can safely be re-inserted now */
463 task
->wake_q
.next
= NULL
;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task
);
470 put_task_struct(task
);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq
*rq
)
483 struct task_struct
*curr
= rq
->curr
;
486 lockdep_assert_held(&rq
->lock
);
488 if (test_tsk_need_resched(curr
))
493 if (cpu
== smp_processor_id()) {
494 set_tsk_need_resched(curr
);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr
))
500 smp_send_reschedule(cpu
);
502 trace_sched_wake_idle_without_ipi(cpu
);
505 void resched_cpu(int cpu
)
507 struct rq
*rq
= cpu_rq(cpu
);
510 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
513 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i
, cpu
= smp_processor_id();
529 struct sched_domain
*sd
;
531 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
535 for_each_domain(cpu
, sd
) {
536 for_each_cpu(i
, sched_domain_span(sd
)) {
540 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
547 if (!is_housekeeping_cpu(cpu
))
548 cpu
= housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu
)
565 struct rq
*rq
= cpu_rq(cpu
);
567 if (cpu
== smp_processor_id())
570 if (set_nr_and_not_polling(rq
->idle
))
571 smp_send_reschedule(cpu
);
573 trace_sched_wake_idle_without_ipi(cpu
);
576 static bool wake_up_full_nohz_cpu(int cpu
)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (cpu_is_offline(cpu
))
585 return true; /* Don't try to wake offline CPUs. */
586 if (tick_nohz_full_cpu(cpu
)) {
587 if (cpu
!= smp_processor_id() ||
588 tick_nohz_tick_stopped())
589 tick_nohz_full_kick_cpu(cpu
);
597 * Wake up the specified CPU. If the CPU is going offline, it is the
598 * caller's responsibility to deal with the lost wakeup, for example,
599 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
601 void wake_up_nohz_cpu(int cpu
)
603 if (!wake_up_full_nohz_cpu(cpu
))
604 wake_up_idle_cpu(cpu
);
607 static inline bool got_nohz_idle_kick(void)
609 int cpu
= smp_processor_id();
611 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
614 if (idle_cpu(cpu
) && !need_resched())
618 * We can't run Idle Load Balance on this CPU for this time so we
619 * cancel it and clear NOHZ_BALANCE_KICK
621 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
625 #else /* CONFIG_NO_HZ_COMMON */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ_COMMON */
634 #ifdef CONFIG_NO_HZ_FULL
635 bool sched_can_stop_tick(struct rq
*rq
)
639 /* Deadline tasks, even if single, need the tick */
640 if (rq
->dl
.dl_nr_running
)
644 * If there are more than one RR tasks, we need the tick to effect the
645 * actual RR behaviour.
647 if (rq
->rt
.rr_nr_running
) {
648 if (rq
->rt
.rr_nr_running
== 1)
655 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 * forced preemption between FIFO tasks.
658 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
663 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 * if there's more than one we need the tick for involuntary
667 if (rq
->nr_running
> 1)
672 #endif /* CONFIG_NO_HZ_FULL */
674 void sched_avg_update(struct rq
*rq
)
676 s64 period
= sched_avg_period();
678 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
680 * Inline assembly required to prevent the compiler
681 * optimising this loop into a divmod call.
682 * See __iter_div_u64_rem() for another example of this.
684 asm("" : "+rm" (rq
->age_stamp
));
685 rq
->age_stamp
+= period
;
690 #endif /* CONFIG_SMP */
692 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
695 * Iterate task_group tree rooted at *from, calling @down when first entering a
696 * node and @up when leaving it for the final time.
698 * Caller must hold rcu_lock or sufficient equivalent.
700 int walk_tg_tree_from(struct task_group
*from
,
701 tg_visitor down
, tg_visitor up
, void *data
)
703 struct task_group
*parent
, *child
;
709 ret
= (*down
)(parent
, data
);
712 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
719 ret
= (*up
)(parent
, data
);
720 if (ret
|| parent
== from
)
724 parent
= parent
->parent
;
731 int tg_nop(struct task_group
*tg
, void *data
)
737 static void set_load_weight(struct task_struct
*p
)
739 int prio
= p
->static_prio
- MAX_RT_PRIO
;
740 struct load_weight
*load
= &p
->se
.load
;
743 * SCHED_IDLE tasks get minimal weight:
745 if (idle_policy(p
->policy
)) {
746 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
747 load
->inv_weight
= WMULT_IDLEPRIO
;
751 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
752 load
->inv_weight
= sched_prio_to_wmult
[prio
];
755 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
758 if (!(flags
& ENQUEUE_RESTORE
))
759 sched_info_queued(rq
, p
);
760 p
->sched_class
->enqueue_task(rq
, p
, flags
);
763 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 if (!(flags
& DEQUEUE_SAVE
))
767 sched_info_dequeued(rq
, p
);
768 p
->sched_class
->dequeue_task(rq
, p
, flags
);
771 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
773 if (task_contributes_to_load(p
))
774 rq
->nr_uninterruptible
--;
776 enqueue_task(rq
, p
, flags
);
779 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
781 if (task_contributes_to_load(p
))
782 rq
->nr_uninterruptible
++;
784 dequeue_task(rq
, p
, flags
);
787 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
790 * In theory, the compile should just see 0 here, and optimize out the call
791 * to sched_rt_avg_update. But I don't trust it...
793 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
794 s64 steal
= 0, irq_delta
= 0;
796 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
797 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
800 * Since irq_time is only updated on {soft,}irq_exit, we might run into
801 * this case when a previous update_rq_clock() happened inside a
804 * When this happens, we stop ->clock_task and only update the
805 * prev_irq_time stamp to account for the part that fit, so that a next
806 * update will consume the rest. This ensures ->clock_task is
809 * It does however cause some slight miss-attribution of {soft,}irq
810 * time, a more accurate solution would be to update the irq_time using
811 * the current rq->clock timestamp, except that would require using
814 if (irq_delta
> delta
)
817 rq
->prev_irq_time
+= irq_delta
;
820 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
821 if (static_key_false((¶virt_steal_rq_enabled
))) {
822 steal
= paravirt_steal_clock(cpu_of(rq
));
823 steal
-= rq
->prev_steal_time_rq
;
825 if (unlikely(steal
> delta
))
828 rq
->prev_steal_time_rq
+= steal
;
833 rq
->clock_task
+= delta
;
835 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
836 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
837 sched_rt_avg_update(rq
, irq_delta
+ steal
);
841 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
843 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
844 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
848 * Make it appear like a SCHED_FIFO task, its something
849 * userspace knows about and won't get confused about.
851 * Also, it will make PI more or less work without too
852 * much confusion -- but then, stop work should not
853 * rely on PI working anyway.
855 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
857 stop
->sched_class
= &stop_sched_class
;
860 cpu_rq(cpu
)->stop
= stop
;
864 * Reset it back to a normal scheduling class so that
865 * it can die in pieces.
867 old_stop
->sched_class
= &rt_sched_class
;
872 * __normal_prio - return the priority that is based on the static prio
874 static inline int __normal_prio(struct task_struct
*p
)
876 return p
->static_prio
;
880 * Calculate the expected normal priority: i.e. priority
881 * without taking RT-inheritance into account. Might be
882 * boosted by interactivity modifiers. Changes upon fork,
883 * setprio syscalls, and whenever the interactivity
884 * estimator recalculates.
886 static inline int normal_prio(struct task_struct
*p
)
890 if (task_has_dl_policy(p
))
891 prio
= MAX_DL_PRIO
-1;
892 else if (task_has_rt_policy(p
))
893 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
895 prio
= __normal_prio(p
);
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
906 static int effective_prio(struct task_struct
*p
)
908 p
->normal_prio
= normal_prio(p
);
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
914 if (!rt_prio(p
->prio
))
915 return p
->normal_prio
;
920 * task_curr - is this task currently executing on a CPU?
921 * @p: the task in question.
923 * Return: 1 if the task is currently executing. 0 otherwise.
925 inline int task_curr(const struct task_struct
*p
)
927 return cpu_curr(task_cpu(p
)) == p
;
931 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
932 * use the balance_callback list if you want balancing.
934 * this means any call to check_class_changed() must be followed by a call to
935 * balance_callback().
937 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
938 const struct sched_class
*prev_class
,
941 if (prev_class
!= p
->sched_class
) {
942 if (prev_class
->switched_from
)
943 prev_class
->switched_from(rq
, p
);
945 p
->sched_class
->switched_to(rq
, p
);
946 } else if (oldprio
!= p
->prio
|| dl_task(p
))
947 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
950 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
952 const struct sched_class
*class;
954 if (p
->sched_class
== rq
->curr
->sched_class
) {
955 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
957 for_each_class(class) {
958 if (class == rq
->curr
->sched_class
)
960 if (class == p
->sched_class
) {
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
972 rq_clock_skip_update(rq
, true);
977 * This is how migration works:
979 * 1) we invoke migration_cpu_stop() on the target CPU using
981 * 2) stopper starts to run (implicitly forcing the migrated thread
983 * 3) it checks whether the migrated task is still in the wrong runqueue.
984 * 4) if it's in the wrong runqueue then the migration thread removes
985 * it and puts it into the right queue.
986 * 5) stopper completes and stop_one_cpu() returns and the migration
991 * move_queued_task - move a queued task to new rq.
993 * Returns (locked) new rq. Old rq's lock is released.
995 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
997 lockdep_assert_held(&rq
->lock
);
999 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1000 dequeue_task(rq
, p
, 0);
1001 set_task_cpu(p
, new_cpu
);
1002 raw_spin_unlock(&rq
->lock
);
1004 rq
= cpu_rq(new_cpu
);
1006 raw_spin_lock(&rq
->lock
);
1007 BUG_ON(task_cpu(p
) != new_cpu
);
1008 enqueue_task(rq
, p
, 0);
1009 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1010 check_preempt_curr(rq
, p
, 0);
1015 struct migration_arg
{
1016 struct task_struct
*task
;
1021 * Move (not current) task off this cpu, onto dest cpu. We're doing
1022 * this because either it can't run here any more (set_cpus_allowed()
1023 * away from this CPU, or CPU going down), or because we're
1024 * attempting to rebalance this task on exec (sched_exec).
1026 * So we race with normal scheduler movements, but that's OK, as long
1027 * as the task is no longer on this CPU.
1029 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1031 if (unlikely(!cpu_active(dest_cpu
)))
1034 /* Affinity changed (again). */
1035 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1038 rq
= move_queued_task(rq
, p
, dest_cpu
);
1044 * migration_cpu_stop - this will be executed by a highprio stopper thread
1045 * and performs thread migration by bumping thread off CPU then
1046 * 'pushing' onto another runqueue.
1048 static int migration_cpu_stop(void *data
)
1050 struct migration_arg
*arg
= data
;
1051 struct task_struct
*p
= arg
->task
;
1052 struct rq
*rq
= this_rq();
1055 * The original target cpu might have gone down and we might
1056 * be on another cpu but it doesn't matter.
1058 local_irq_disable();
1060 * We need to explicitly wake pending tasks before running
1061 * __migrate_task() such that we will not miss enforcing cpus_allowed
1062 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1064 sched_ttwu_pending();
1066 raw_spin_lock(&p
->pi_lock
);
1067 raw_spin_lock(&rq
->lock
);
1069 * If task_rq(p) != rq, it cannot be migrated here, because we're
1070 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1071 * we're holding p->pi_lock.
1073 if (task_rq(p
) == rq
) {
1074 if (task_on_rq_queued(p
))
1075 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1077 p
->wake_cpu
= arg
->dest_cpu
;
1079 raw_spin_unlock(&rq
->lock
);
1080 raw_spin_unlock(&p
->pi_lock
);
1087 * sched_class::set_cpus_allowed must do the below, but is not required to
1088 * actually call this function.
1090 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1092 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1093 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1096 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1098 struct rq
*rq
= task_rq(p
);
1099 bool queued
, running
;
1101 lockdep_assert_held(&p
->pi_lock
);
1103 queued
= task_on_rq_queued(p
);
1104 running
= task_current(rq
, p
);
1108 * Because __kthread_bind() calls this on blocked tasks without
1111 lockdep_assert_held(&rq
->lock
);
1112 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1115 put_prev_task(rq
, p
);
1117 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1120 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1122 set_curr_task(rq
, p
);
1126 * Change a given task's CPU affinity. Migrate the thread to a
1127 * proper CPU and schedule it away if the CPU it's executing on
1128 * is removed from the allowed bitmask.
1130 * NOTE: the caller must have a valid reference to the task, the
1131 * task must not exit() & deallocate itself prematurely. The
1132 * call is not atomic; no spinlocks may be held.
1134 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1135 const struct cpumask
*new_mask
, bool check
)
1137 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1138 unsigned int dest_cpu
;
1143 rq
= task_rq_lock(p
, &rf
);
1145 if (p
->flags
& PF_KTHREAD
) {
1147 * Kernel threads are allowed on online && !active CPUs
1149 cpu_valid_mask
= cpu_online_mask
;
1153 * Must re-check here, to close a race against __kthread_bind(),
1154 * sched_setaffinity() is not guaranteed to observe the flag.
1156 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1161 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1164 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1169 do_set_cpus_allowed(p
, new_mask
);
1171 if (p
->flags
& PF_KTHREAD
) {
1173 * For kernel threads that do indeed end up on online &&
1174 * !active we want to ensure they are strict per-cpu threads.
1176 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1177 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1178 p
->nr_cpus_allowed
!= 1);
1181 /* Can the task run on the task's current CPU? If so, we're done */
1182 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1185 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1186 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1187 struct migration_arg arg
= { p
, dest_cpu
};
1188 /* Need help from migration thread: drop lock and wait. */
1189 task_rq_unlock(rq
, p
, &rf
);
1190 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1191 tlb_migrate_finish(p
->mm
);
1193 } else if (task_on_rq_queued(p
)) {
1195 * OK, since we're going to drop the lock immediately
1196 * afterwards anyway.
1198 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
1199 rq
= move_queued_task(rq
, p
, dest_cpu
);
1200 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
1203 task_rq_unlock(rq
, p
, &rf
);
1208 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1210 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1212 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1214 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1216 #ifdef CONFIG_SCHED_DEBUG
1218 * We should never call set_task_cpu() on a blocked task,
1219 * ttwu() will sort out the placement.
1221 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1225 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1226 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1227 * time relying on p->on_rq.
1229 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1230 p
->sched_class
== &fair_sched_class
&&
1231 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1233 #ifdef CONFIG_LOCKDEP
1235 * The caller should hold either p->pi_lock or rq->lock, when changing
1236 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1238 * sched_move_task() holds both and thus holding either pins the cgroup,
1241 * Furthermore, all task_rq users should acquire both locks, see
1244 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1245 lockdep_is_held(&task_rq(p
)->lock
)));
1249 trace_sched_migrate_task(p
, new_cpu
);
1251 if (task_cpu(p
) != new_cpu
) {
1252 if (p
->sched_class
->migrate_task_rq
)
1253 p
->sched_class
->migrate_task_rq(p
);
1254 p
->se
.nr_migrations
++;
1255 perf_event_task_migrate(p
);
1258 __set_task_cpu(p
, new_cpu
);
1261 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1263 if (task_on_rq_queued(p
)) {
1264 struct rq
*src_rq
, *dst_rq
;
1266 src_rq
= task_rq(p
);
1267 dst_rq
= cpu_rq(cpu
);
1269 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1270 deactivate_task(src_rq
, p
, 0);
1271 set_task_cpu(p
, cpu
);
1272 activate_task(dst_rq
, p
, 0);
1273 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1274 check_preempt_curr(dst_rq
, p
, 0);
1277 * Task isn't running anymore; make it appear like we migrated
1278 * it before it went to sleep. This means on wakeup we make the
1279 * previous cpu our target instead of where it really is.
1285 struct migration_swap_arg
{
1286 struct task_struct
*src_task
, *dst_task
;
1287 int src_cpu
, dst_cpu
;
1290 static int migrate_swap_stop(void *data
)
1292 struct migration_swap_arg
*arg
= data
;
1293 struct rq
*src_rq
, *dst_rq
;
1296 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1299 src_rq
= cpu_rq(arg
->src_cpu
);
1300 dst_rq
= cpu_rq(arg
->dst_cpu
);
1302 double_raw_lock(&arg
->src_task
->pi_lock
,
1303 &arg
->dst_task
->pi_lock
);
1304 double_rq_lock(src_rq
, dst_rq
);
1306 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1309 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1312 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1315 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1318 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1319 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1324 double_rq_unlock(src_rq
, dst_rq
);
1325 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1326 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1332 * Cross migrate two tasks
1334 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1336 struct migration_swap_arg arg
;
1339 arg
= (struct migration_swap_arg
){
1341 .src_cpu
= task_cpu(cur
),
1343 .dst_cpu
= task_cpu(p
),
1346 if (arg
.src_cpu
== arg
.dst_cpu
)
1350 * These three tests are all lockless; this is OK since all of them
1351 * will be re-checked with proper locks held further down the line.
1353 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1356 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1359 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1362 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1363 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1370 * wait_task_inactive - wait for a thread to unschedule.
1372 * If @match_state is nonzero, it's the @p->state value just checked and
1373 * not expected to change. If it changes, i.e. @p might have woken up,
1374 * then return zero. When we succeed in waiting for @p to be off its CPU,
1375 * we return a positive number (its total switch count). If a second call
1376 * a short while later returns the same number, the caller can be sure that
1377 * @p has remained unscheduled the whole time.
1379 * The caller must ensure that the task *will* unschedule sometime soon,
1380 * else this function might spin for a *long* time. This function can't
1381 * be called with interrupts off, or it may introduce deadlock with
1382 * smp_call_function() if an IPI is sent by the same process we are
1383 * waiting to become inactive.
1385 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1387 int running
, queued
;
1394 * We do the initial early heuristics without holding
1395 * any task-queue locks at all. We'll only try to get
1396 * the runqueue lock when things look like they will
1402 * If the task is actively running on another CPU
1403 * still, just relax and busy-wait without holding
1406 * NOTE! Since we don't hold any locks, it's not
1407 * even sure that "rq" stays as the right runqueue!
1408 * But we don't care, since "task_running()" will
1409 * return false if the runqueue has changed and p
1410 * is actually now running somewhere else!
1412 while (task_running(rq
, p
)) {
1413 if (match_state
&& unlikely(p
->state
!= match_state
))
1419 * Ok, time to look more closely! We need the rq
1420 * lock now, to be *sure*. If we're wrong, we'll
1421 * just go back and repeat.
1423 rq
= task_rq_lock(p
, &rf
);
1424 trace_sched_wait_task(p
);
1425 running
= task_running(rq
, p
);
1426 queued
= task_on_rq_queued(p
);
1428 if (!match_state
|| p
->state
== match_state
)
1429 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1430 task_rq_unlock(rq
, p
, &rf
);
1433 * If it changed from the expected state, bail out now.
1435 if (unlikely(!ncsw
))
1439 * Was it really running after all now that we
1440 * checked with the proper locks actually held?
1442 * Oops. Go back and try again..
1444 if (unlikely(running
)) {
1450 * It's not enough that it's not actively running,
1451 * it must be off the runqueue _entirely_, and not
1454 * So if it was still runnable (but just not actively
1455 * running right now), it's preempted, and we should
1456 * yield - it could be a while.
1458 if (unlikely(queued
)) {
1459 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1461 set_current_state(TASK_UNINTERRUPTIBLE
);
1462 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1467 * Ahh, all good. It wasn't running, and it wasn't
1468 * runnable, which means that it will never become
1469 * running in the future either. We're all done!
1478 * kick_process - kick a running thread to enter/exit the kernel
1479 * @p: the to-be-kicked thread
1481 * Cause a process which is running on another CPU to enter
1482 * kernel-mode, without any delay. (to get signals handled.)
1484 * NOTE: this function doesn't have to take the runqueue lock,
1485 * because all it wants to ensure is that the remote task enters
1486 * the kernel. If the IPI races and the task has been migrated
1487 * to another CPU then no harm is done and the purpose has been
1490 void kick_process(struct task_struct
*p
)
1496 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1497 smp_send_reschedule(cpu
);
1500 EXPORT_SYMBOL_GPL(kick_process
);
1503 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1505 * A few notes on cpu_active vs cpu_online:
1507 * - cpu_active must be a subset of cpu_online
1509 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1510 * see __set_cpus_allowed_ptr(). At this point the newly online
1511 * cpu isn't yet part of the sched domains, and balancing will not
1514 * - on cpu-down we clear cpu_active() to mask the sched domains and
1515 * avoid the load balancer to place new tasks on the to be removed
1516 * cpu. Existing tasks will remain running there and will be taken
1519 * This means that fallback selection must not select !active CPUs.
1520 * And can assume that any active CPU must be online. Conversely
1521 * select_task_rq() below may allow selection of !active CPUs in order
1522 * to satisfy the above rules.
1524 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1526 int nid
= cpu_to_node(cpu
);
1527 const struct cpumask
*nodemask
= NULL
;
1528 enum { cpuset
, possible
, fail
} state
= cpuset
;
1532 * If the node that the cpu is on has been offlined, cpu_to_node()
1533 * will return -1. There is no cpu on the node, and we should
1534 * select the cpu on the other node.
1537 nodemask
= cpumask_of_node(nid
);
1539 /* Look for allowed, online CPU in same node. */
1540 for_each_cpu(dest_cpu
, nodemask
) {
1541 if (!cpu_active(dest_cpu
))
1543 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1549 /* Any allowed, online CPU? */
1550 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1551 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1553 if (!cpu_online(dest_cpu
))
1558 /* No more Mr. Nice Guy. */
1561 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1562 cpuset_cpus_allowed_fallback(p
);
1568 do_set_cpus_allowed(p
, cpu_possible_mask
);
1579 if (state
!= cpuset
) {
1581 * Don't tell them about moving exiting tasks or
1582 * kernel threads (both mm NULL), since they never
1585 if (p
->mm
&& printk_ratelimit()) {
1586 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1587 task_pid_nr(p
), p
->comm
, cpu
);
1595 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1598 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1600 lockdep_assert_held(&p
->pi_lock
);
1602 if (tsk_nr_cpus_allowed(p
) > 1)
1603 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1605 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1608 * In order not to call set_task_cpu() on a blocking task we need
1609 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1612 * Since this is common to all placement strategies, this lives here.
1614 * [ this allows ->select_task() to simply return task_cpu(p) and
1615 * not worry about this generic constraint ]
1617 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1619 cpu
= select_fallback_rq(task_cpu(p
), p
);
1624 static void update_avg(u64
*avg
, u64 sample
)
1626 s64 diff
= sample
- *avg
;
1632 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1633 const struct cpumask
*new_mask
, bool check
)
1635 return set_cpus_allowed_ptr(p
, new_mask
);
1638 #endif /* CONFIG_SMP */
1641 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1645 if (!schedstat_enabled())
1651 if (cpu
== rq
->cpu
) {
1652 schedstat_inc(rq
->ttwu_local
);
1653 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1655 struct sched_domain
*sd
;
1657 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1659 for_each_domain(rq
->cpu
, sd
) {
1660 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1661 schedstat_inc(sd
->ttwu_wake_remote
);
1668 if (wake_flags
& WF_MIGRATED
)
1669 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1670 #endif /* CONFIG_SMP */
1672 schedstat_inc(rq
->ttwu_count
);
1673 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1675 if (wake_flags
& WF_SYNC
)
1676 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1679 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1681 activate_task(rq
, p
, en_flags
);
1682 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1684 /* if a worker is waking up, notify workqueue */
1685 if (p
->flags
& PF_WQ_WORKER
)
1686 wq_worker_waking_up(p
, cpu_of(rq
));
1690 * Mark the task runnable and perform wakeup-preemption.
1692 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1693 struct pin_cookie cookie
)
1695 check_preempt_curr(rq
, p
, wake_flags
);
1696 p
->state
= TASK_RUNNING
;
1697 trace_sched_wakeup(p
);
1700 if (p
->sched_class
->task_woken
) {
1702 * Our task @p is fully woken up and running; so its safe to
1703 * drop the rq->lock, hereafter rq is only used for statistics.
1705 lockdep_unpin_lock(&rq
->lock
, cookie
);
1706 p
->sched_class
->task_woken(rq
, p
);
1707 lockdep_repin_lock(&rq
->lock
, cookie
);
1710 if (rq
->idle_stamp
) {
1711 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1712 u64 max
= 2*rq
->max_idle_balance_cost
;
1714 update_avg(&rq
->avg_idle
, delta
);
1716 if (rq
->avg_idle
> max
)
1725 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1726 struct pin_cookie cookie
)
1728 int en_flags
= ENQUEUE_WAKEUP
;
1730 lockdep_assert_held(&rq
->lock
);
1733 if (p
->sched_contributes_to_load
)
1734 rq
->nr_uninterruptible
--;
1736 if (wake_flags
& WF_MIGRATED
)
1737 en_flags
|= ENQUEUE_MIGRATED
;
1740 ttwu_activate(rq
, p
, en_flags
);
1741 ttwu_do_wakeup(rq
, p
, wake_flags
, cookie
);
1745 * Called in case the task @p isn't fully descheduled from its runqueue,
1746 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1747 * since all we need to do is flip p->state to TASK_RUNNING, since
1748 * the task is still ->on_rq.
1750 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1756 rq
= __task_rq_lock(p
, &rf
);
1757 if (task_on_rq_queued(p
)) {
1758 /* check_preempt_curr() may use rq clock */
1759 update_rq_clock(rq
);
1760 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
.cookie
);
1763 __task_rq_unlock(rq
, &rf
);
1769 void sched_ttwu_pending(void)
1771 struct rq
*rq
= this_rq();
1772 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1773 struct pin_cookie cookie
;
1774 struct task_struct
*p
;
1775 unsigned long flags
;
1780 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1781 cookie
= lockdep_pin_lock(&rq
->lock
);
1786 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1787 llist
= llist_next(llist
);
1789 if (p
->sched_remote_wakeup
)
1790 wake_flags
= WF_MIGRATED
;
1792 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1795 lockdep_unpin_lock(&rq
->lock
, cookie
);
1796 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1799 void scheduler_ipi(void)
1802 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1803 * TIF_NEED_RESCHED remotely (for the first time) will also send
1806 preempt_fold_need_resched();
1808 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1812 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1813 * traditionally all their work was done from the interrupt return
1814 * path. Now that we actually do some work, we need to make sure
1817 * Some archs already do call them, luckily irq_enter/exit nest
1820 * Arguably we should visit all archs and update all handlers,
1821 * however a fair share of IPIs are still resched only so this would
1822 * somewhat pessimize the simple resched case.
1825 sched_ttwu_pending();
1828 * Check if someone kicked us for doing the nohz idle load balance.
1830 if (unlikely(got_nohz_idle_kick())) {
1831 this_rq()->idle_balance
= 1;
1832 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1837 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1839 struct rq
*rq
= cpu_rq(cpu
);
1841 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1843 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1844 if (!set_nr_if_polling(rq
->idle
))
1845 smp_send_reschedule(cpu
);
1847 trace_sched_wake_idle_without_ipi(cpu
);
1851 void wake_up_if_idle(int cpu
)
1853 struct rq
*rq
= cpu_rq(cpu
);
1854 unsigned long flags
;
1858 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1861 if (set_nr_if_polling(rq
->idle
)) {
1862 trace_sched_wake_idle_without_ipi(cpu
);
1864 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1865 if (is_idle_task(rq
->curr
))
1866 smp_send_reschedule(cpu
);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1875 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1877 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1879 #endif /* CONFIG_SMP */
1881 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1883 struct rq
*rq
= cpu_rq(cpu
);
1884 struct pin_cookie cookie
;
1886 #if defined(CONFIG_SMP)
1887 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1888 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1889 ttwu_queue_remote(p
, cpu
, wake_flags
);
1894 raw_spin_lock(&rq
->lock
);
1895 cookie
= lockdep_pin_lock(&rq
->lock
);
1896 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1897 lockdep_unpin_lock(&rq
->lock
, cookie
);
1898 raw_spin_unlock(&rq
->lock
);
1902 * Notes on Program-Order guarantees on SMP systems.
1906 * The basic program-order guarantee on SMP systems is that when a task [t]
1907 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1908 * execution on its new cpu [c1].
1910 * For migration (of runnable tasks) this is provided by the following means:
1912 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1913 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1914 * rq(c1)->lock (if not at the same time, then in that order).
1915 * C) LOCK of the rq(c1)->lock scheduling in task
1917 * Transitivity guarantees that B happens after A and C after B.
1918 * Note: we only require RCpc transitivity.
1919 * Note: the cpu doing B need not be c0 or c1
1928 * UNLOCK rq(0)->lock
1930 * LOCK rq(0)->lock // orders against CPU0
1932 * UNLOCK rq(0)->lock
1936 * UNLOCK rq(1)->lock
1938 * LOCK rq(1)->lock // orders against CPU2
1941 * UNLOCK rq(1)->lock
1944 * BLOCKING -- aka. SLEEP + WAKEUP
1946 * For blocking we (obviously) need to provide the same guarantee as for
1947 * migration. However the means are completely different as there is no lock
1948 * chain to provide order. Instead we do:
1950 * 1) smp_store_release(X->on_cpu, 0)
1951 * 2) smp_cond_load_acquire(!X->on_cpu)
1955 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1957 * LOCK rq(0)->lock LOCK X->pi_lock
1960 * smp_store_release(X->on_cpu, 0);
1962 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1968 * X->state = RUNNING
1969 * UNLOCK rq(2)->lock
1971 * LOCK rq(2)->lock // orders against CPU1
1974 * UNLOCK rq(2)->lock
1977 * UNLOCK rq(0)->lock
1980 * However; for wakeups there is a second guarantee we must provide, namely we
1981 * must observe the state that lead to our wakeup. That is, not only must our
1982 * task observe its own prior state, it must also observe the stores prior to
1985 * This means that any means of doing remote wakeups must order the CPU doing
1986 * the wakeup against the CPU the task is going to end up running on. This,
1987 * however, is already required for the regular Program-Order guarantee above,
1988 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1993 * try_to_wake_up - wake up a thread
1994 * @p: the thread to be awakened
1995 * @state: the mask of task states that can be woken
1996 * @wake_flags: wake modifier flags (WF_*)
1998 * If (@state & @p->state) @p->state = TASK_RUNNING.
2000 * If the task was not queued/runnable, also place it back on a runqueue.
2002 * Atomic against schedule() which would dequeue a task, also see
2003 * set_current_state().
2005 * Return: %true if @p->state changes (an actual wakeup was done),
2009 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2011 unsigned long flags
;
2012 int cpu
, success
= 0;
2015 * If we are going to wake up a thread waiting for CONDITION we
2016 * need to ensure that CONDITION=1 done by the caller can not be
2017 * reordered with p->state check below. This pairs with mb() in
2018 * set_current_state() the waiting thread does.
2020 smp_mb__before_spinlock();
2021 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2022 if (!(p
->state
& state
))
2025 trace_sched_waking(p
);
2027 success
= 1; /* we're going to change ->state */
2031 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2032 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2033 * in smp_cond_load_acquire() below.
2035 * sched_ttwu_pending() try_to_wake_up()
2036 * [S] p->on_rq = 1; [L] P->state
2037 * UNLOCK rq->lock -----.
2041 * LOCK rq->lock -----'
2045 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2047 * Pairs with the UNLOCK+LOCK on rq->lock from the
2048 * last wakeup of our task and the schedule that got our task
2052 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2057 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2058 * possible to, falsely, observe p->on_cpu == 0.
2060 * One must be running (->on_cpu == 1) in order to remove oneself
2061 * from the runqueue.
2063 * [S] ->on_cpu = 1; [L] ->on_rq
2067 * [S] ->on_rq = 0; [L] ->on_cpu
2069 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2070 * from the consecutive calls to schedule(); the first switching to our
2071 * task, the second putting it to sleep.
2076 * If the owning (remote) cpu is still in the middle of schedule() with
2077 * this task as prev, wait until its done referencing the task.
2079 * Pairs with the smp_store_release() in finish_lock_switch().
2081 * This ensures that tasks getting woken will be fully ordered against
2082 * their previous state and preserve Program Order.
2084 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2086 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2087 p
->state
= TASK_WAKING
;
2089 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2090 if (task_cpu(p
) != cpu
) {
2091 wake_flags
|= WF_MIGRATED
;
2092 set_task_cpu(p
, cpu
);
2094 #endif /* CONFIG_SMP */
2096 ttwu_queue(p
, cpu
, wake_flags
);
2098 ttwu_stat(p
, cpu
, wake_flags
);
2100 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2106 * try_to_wake_up_local - try to wake up a local task with rq lock held
2107 * @p: the thread to be awakened
2108 * @cookie: context's cookie for pinning
2110 * Put @p on the run-queue if it's not already there. The caller must
2111 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2114 static void try_to_wake_up_local(struct task_struct
*p
, struct pin_cookie cookie
)
2116 struct rq
*rq
= task_rq(p
);
2118 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2119 WARN_ON_ONCE(p
== current
))
2122 lockdep_assert_held(&rq
->lock
);
2124 if (!raw_spin_trylock(&p
->pi_lock
)) {
2126 * This is OK, because current is on_cpu, which avoids it being
2127 * picked for load-balance and preemption/IRQs are still
2128 * disabled avoiding further scheduler activity on it and we've
2129 * not yet picked a replacement task.
2131 lockdep_unpin_lock(&rq
->lock
, cookie
);
2132 raw_spin_unlock(&rq
->lock
);
2133 raw_spin_lock(&p
->pi_lock
);
2134 raw_spin_lock(&rq
->lock
);
2135 lockdep_repin_lock(&rq
->lock
, cookie
);
2138 if (!(p
->state
& TASK_NORMAL
))
2141 trace_sched_waking(p
);
2143 if (!task_on_rq_queued(p
))
2144 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2146 ttwu_do_wakeup(rq
, p
, 0, cookie
);
2147 ttwu_stat(p
, smp_processor_id(), 0);
2149 raw_spin_unlock(&p
->pi_lock
);
2153 * wake_up_process - Wake up a specific process
2154 * @p: The process to be woken up.
2156 * Attempt to wake up the nominated process and move it to the set of runnable
2159 * Return: 1 if the process was woken up, 0 if it was already running.
2161 * It may be assumed that this function implies a write memory barrier before
2162 * changing the task state if and only if any tasks are woken up.
2164 int wake_up_process(struct task_struct
*p
)
2166 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2168 EXPORT_SYMBOL(wake_up_process
);
2170 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2172 return try_to_wake_up(p
, state
, 0);
2176 * This function clears the sched_dl_entity static params.
2178 void __dl_clear_params(struct task_struct
*p
)
2180 struct sched_dl_entity
*dl_se
= &p
->dl
;
2182 dl_se
->dl_runtime
= 0;
2183 dl_se
->dl_deadline
= 0;
2184 dl_se
->dl_period
= 0;
2188 dl_se
->dl_throttled
= 0;
2189 dl_se
->dl_yielded
= 0;
2193 * Perform scheduler related setup for a newly forked process p.
2194 * p is forked by current.
2196 * __sched_fork() is basic setup used by init_idle() too:
2198 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2203 p
->se
.exec_start
= 0;
2204 p
->se
.sum_exec_runtime
= 0;
2205 p
->se
.prev_sum_exec_runtime
= 0;
2206 p
->se
.nr_migrations
= 0;
2208 INIT_LIST_HEAD(&p
->se
.group_node
);
2210 #ifdef CONFIG_FAIR_GROUP_SCHED
2211 p
->se
.cfs_rq
= NULL
;
2214 #ifdef CONFIG_SCHEDSTATS
2215 /* Even if schedstat is disabled, there should not be garbage */
2216 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2219 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2220 init_dl_task_timer(&p
->dl
);
2221 __dl_clear_params(p
);
2223 INIT_LIST_HEAD(&p
->rt
.run_list
);
2225 p
->rt
.time_slice
= sched_rr_timeslice
;
2229 #ifdef CONFIG_PREEMPT_NOTIFIERS
2230 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2233 #ifdef CONFIG_NUMA_BALANCING
2234 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2235 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2236 p
->mm
->numa_scan_seq
= 0;
2239 if (clone_flags
& CLONE_VM
)
2240 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2242 p
->numa_preferred_nid
= -1;
2244 p
->node_stamp
= 0ULL;
2245 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2246 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2247 p
->numa_work
.next
= &p
->numa_work
;
2248 p
->numa_faults
= NULL
;
2249 p
->last_task_numa_placement
= 0;
2250 p
->last_sum_exec_runtime
= 0;
2252 p
->numa_group
= NULL
;
2253 #endif /* CONFIG_NUMA_BALANCING */
2256 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2258 #ifdef CONFIG_NUMA_BALANCING
2260 void set_numabalancing_state(bool enabled
)
2263 static_branch_enable(&sched_numa_balancing
);
2265 static_branch_disable(&sched_numa_balancing
);
2268 #ifdef CONFIG_PROC_SYSCTL
2269 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2270 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2274 int state
= static_branch_likely(&sched_numa_balancing
);
2276 if (write
&& !capable(CAP_SYS_ADMIN
))
2281 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2285 set_numabalancing_state(state
);
2291 #ifdef CONFIG_SCHEDSTATS
2293 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2294 static bool __initdata __sched_schedstats
= false;
2296 static void set_schedstats(bool enabled
)
2299 static_branch_enable(&sched_schedstats
);
2301 static_branch_disable(&sched_schedstats
);
2304 void force_schedstat_enabled(void)
2306 if (!schedstat_enabled()) {
2307 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2308 static_branch_enable(&sched_schedstats
);
2312 static int __init
setup_schedstats(char *str
)
2319 * This code is called before jump labels have been set up, so we can't
2320 * change the static branch directly just yet. Instead set a temporary
2321 * variable so init_schedstats() can do it later.
2323 if (!strcmp(str
, "enable")) {
2324 __sched_schedstats
= true;
2326 } else if (!strcmp(str
, "disable")) {
2327 __sched_schedstats
= false;
2332 pr_warn("Unable to parse schedstats=\n");
2336 __setup("schedstats=", setup_schedstats
);
2338 static void __init
init_schedstats(void)
2340 set_schedstats(__sched_schedstats
);
2343 #ifdef CONFIG_PROC_SYSCTL
2344 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2345 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2349 int state
= static_branch_likely(&sched_schedstats
);
2351 if (write
&& !capable(CAP_SYS_ADMIN
))
2356 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2360 set_schedstats(state
);
2363 #endif /* CONFIG_PROC_SYSCTL */
2364 #else /* !CONFIG_SCHEDSTATS */
2365 static inline void init_schedstats(void) {}
2366 #endif /* CONFIG_SCHEDSTATS */
2369 * fork()/clone()-time setup:
2371 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2373 unsigned long flags
;
2374 int cpu
= get_cpu();
2376 __sched_fork(clone_flags
, p
);
2378 * We mark the process as NEW here. This guarantees that
2379 * nobody will actually run it, and a signal or other external
2380 * event cannot wake it up and insert it on the runqueue either.
2382 p
->state
= TASK_NEW
;
2385 * Make sure we do not leak PI boosting priority to the child.
2387 p
->prio
= current
->normal_prio
;
2390 * Revert to default priority/policy on fork if requested.
2392 if (unlikely(p
->sched_reset_on_fork
)) {
2393 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2394 p
->policy
= SCHED_NORMAL
;
2395 p
->static_prio
= NICE_TO_PRIO(0);
2397 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2398 p
->static_prio
= NICE_TO_PRIO(0);
2400 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2404 * We don't need the reset flag anymore after the fork. It has
2405 * fulfilled its duty:
2407 p
->sched_reset_on_fork
= 0;
2410 if (dl_prio(p
->prio
)) {
2413 } else if (rt_prio(p
->prio
)) {
2414 p
->sched_class
= &rt_sched_class
;
2416 p
->sched_class
= &fair_sched_class
;
2419 init_entity_runnable_average(&p
->se
);
2422 * The child is not yet in the pid-hash so no cgroup attach races,
2423 * and the cgroup is pinned to this child due to cgroup_fork()
2424 * is ran before sched_fork().
2426 * Silence PROVE_RCU.
2428 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2430 * We're setting the cpu for the first time, we don't migrate,
2431 * so use __set_task_cpu().
2433 __set_task_cpu(p
, cpu
);
2434 if (p
->sched_class
->task_fork
)
2435 p
->sched_class
->task_fork(p
);
2436 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2438 #ifdef CONFIG_SCHED_INFO
2439 if (likely(sched_info_on()))
2440 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2442 #if defined(CONFIG_SMP)
2445 init_task_preempt_count(p
);
2447 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2448 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2455 unsigned long to_ratio(u64 period
, u64 runtime
)
2457 if (runtime
== RUNTIME_INF
)
2461 * Doing this here saves a lot of checks in all
2462 * the calling paths, and returning zero seems
2463 * safe for them anyway.
2468 return div64_u64(runtime
<< 20, period
);
2472 inline struct dl_bw
*dl_bw_of(int i
)
2474 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2475 "sched RCU must be held");
2476 return &cpu_rq(i
)->rd
->dl_bw
;
2479 static inline int dl_bw_cpus(int i
)
2481 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2484 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2485 "sched RCU must be held");
2486 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2492 inline struct dl_bw
*dl_bw_of(int i
)
2494 return &cpu_rq(i
)->dl
.dl_bw
;
2497 static inline int dl_bw_cpus(int i
)
2504 * We must be sure that accepting a new task (or allowing changing the
2505 * parameters of an existing one) is consistent with the bandwidth
2506 * constraints. If yes, this function also accordingly updates the currently
2507 * allocated bandwidth to reflect the new situation.
2509 * This function is called while holding p's rq->lock.
2511 * XXX we should delay bw change until the task's 0-lag point, see
2514 static int dl_overflow(struct task_struct
*p
, int policy
,
2515 const struct sched_attr
*attr
)
2518 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2519 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2520 u64 runtime
= attr
->sched_runtime
;
2521 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2524 /* !deadline task may carry old deadline bandwidth */
2525 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2529 * Either if a task, enters, leave, or stays -deadline but changes
2530 * its parameters, we may need to update accordingly the total
2531 * allocated bandwidth of the container.
2533 raw_spin_lock(&dl_b
->lock
);
2534 cpus
= dl_bw_cpus(task_cpu(p
));
2535 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2536 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2537 __dl_add(dl_b
, new_bw
);
2539 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2540 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2541 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2542 __dl_add(dl_b
, new_bw
);
2544 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2545 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2548 raw_spin_unlock(&dl_b
->lock
);
2553 extern void init_dl_bw(struct dl_bw
*dl_b
);
2556 * wake_up_new_task - wake up a newly created task for the first time.
2558 * This function will do some initial scheduler statistics housekeeping
2559 * that must be done for every newly created context, then puts the task
2560 * on the runqueue and wakes it.
2562 void wake_up_new_task(struct task_struct
*p
)
2567 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2568 p
->state
= TASK_RUNNING
;
2571 * Fork balancing, do it here and not earlier because:
2572 * - cpus_allowed can change in the fork path
2573 * - any previously selected cpu might disappear through hotplug
2575 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2576 * as we're not fully set-up yet.
2578 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2580 rq
= __task_rq_lock(p
, &rf
);
2581 post_init_entity_util_avg(&p
->se
);
2583 activate_task(rq
, p
, 0);
2584 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2585 trace_sched_wakeup_new(p
);
2586 check_preempt_curr(rq
, p
, WF_FORK
);
2588 if (p
->sched_class
->task_woken
) {
2590 * Nothing relies on rq->lock after this, so its fine to
2593 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
2594 p
->sched_class
->task_woken(rq
, p
);
2595 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
2598 task_rq_unlock(rq
, p
, &rf
);
2601 #ifdef CONFIG_PREEMPT_NOTIFIERS
2603 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2605 void preempt_notifier_inc(void)
2607 static_key_slow_inc(&preempt_notifier_key
);
2609 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2611 void preempt_notifier_dec(void)
2613 static_key_slow_dec(&preempt_notifier_key
);
2615 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2618 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2619 * @notifier: notifier struct to register
2621 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2623 if (!static_key_false(&preempt_notifier_key
))
2624 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2626 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2628 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2631 * preempt_notifier_unregister - no longer interested in preemption notifications
2632 * @notifier: notifier struct to unregister
2634 * This is *not* safe to call from within a preemption notifier.
2636 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2638 hlist_del(¬ifier
->link
);
2640 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2642 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2644 struct preempt_notifier
*notifier
;
2646 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2647 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2650 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2652 if (static_key_false(&preempt_notifier_key
))
2653 __fire_sched_in_preempt_notifiers(curr
);
2657 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2658 struct task_struct
*next
)
2660 struct preempt_notifier
*notifier
;
2662 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2663 notifier
->ops
->sched_out(notifier
, next
);
2666 static __always_inline
void
2667 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2668 struct task_struct
*next
)
2670 if (static_key_false(&preempt_notifier_key
))
2671 __fire_sched_out_preempt_notifiers(curr
, next
);
2674 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2676 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2681 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2682 struct task_struct
*next
)
2686 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2689 * prepare_task_switch - prepare to switch tasks
2690 * @rq: the runqueue preparing to switch
2691 * @prev: the current task that is being switched out
2692 * @next: the task we are going to switch to.
2694 * This is called with the rq lock held and interrupts off. It must
2695 * be paired with a subsequent finish_task_switch after the context
2698 * prepare_task_switch sets up locking and calls architecture specific
2702 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2703 struct task_struct
*next
)
2705 sched_info_switch(rq
, prev
, next
);
2706 perf_event_task_sched_out(prev
, next
);
2707 fire_sched_out_preempt_notifiers(prev
, next
);
2708 prepare_lock_switch(rq
, next
);
2709 prepare_arch_switch(next
);
2713 * finish_task_switch - clean up after a task-switch
2714 * @prev: the thread we just switched away from.
2716 * finish_task_switch must be called after the context switch, paired
2717 * with a prepare_task_switch call before the context switch.
2718 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2719 * and do any other architecture-specific cleanup actions.
2721 * Note that we may have delayed dropping an mm in context_switch(). If
2722 * so, we finish that here outside of the runqueue lock. (Doing it
2723 * with the lock held can cause deadlocks; see schedule() for
2726 * The context switch have flipped the stack from under us and restored the
2727 * local variables which were saved when this task called schedule() in the
2728 * past. prev == current is still correct but we need to recalculate this_rq
2729 * because prev may have moved to another CPU.
2731 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2732 __releases(rq
->lock
)
2734 struct rq
*rq
= this_rq();
2735 struct mm_struct
*mm
= rq
->prev_mm
;
2739 * The previous task will have left us with a preempt_count of 2
2740 * because it left us after:
2743 * preempt_disable(); // 1
2745 * raw_spin_lock_irq(&rq->lock) // 2
2747 * Also, see FORK_PREEMPT_COUNT.
2749 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2750 "corrupted preempt_count: %s/%d/0x%x\n",
2751 current
->comm
, current
->pid
, preempt_count()))
2752 preempt_count_set(FORK_PREEMPT_COUNT
);
2757 * A task struct has one reference for the use as "current".
2758 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2759 * schedule one last time. The schedule call will never return, and
2760 * the scheduled task must drop that reference.
2762 * We must observe prev->state before clearing prev->on_cpu (in
2763 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2764 * running on another CPU and we could rave with its RUNNING -> DEAD
2765 * transition, resulting in a double drop.
2767 prev_state
= prev
->state
;
2768 vtime_task_switch(prev
);
2769 perf_event_task_sched_in(prev
, current
);
2770 finish_lock_switch(rq
, prev
);
2771 finish_arch_post_lock_switch();
2773 fire_sched_in_preempt_notifiers(current
);
2776 if (unlikely(prev_state
== TASK_DEAD
)) {
2777 if (prev
->sched_class
->task_dead
)
2778 prev
->sched_class
->task_dead(prev
);
2781 * Remove function-return probe instances associated with this
2782 * task and put them back on the free list.
2784 kprobe_flush_task(prev
);
2786 /* Task is done with its stack. */
2787 put_task_stack(prev
);
2789 put_task_struct(prev
);
2792 tick_nohz_task_switch();
2798 /* rq->lock is NOT held, but preemption is disabled */
2799 static void __balance_callback(struct rq
*rq
)
2801 struct callback_head
*head
, *next
;
2802 void (*func
)(struct rq
*rq
);
2803 unsigned long flags
;
2805 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2806 head
= rq
->balance_callback
;
2807 rq
->balance_callback
= NULL
;
2809 func
= (void (*)(struct rq
*))head
->func
;
2816 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2819 static inline void balance_callback(struct rq
*rq
)
2821 if (unlikely(rq
->balance_callback
))
2822 __balance_callback(rq
);
2827 static inline void balance_callback(struct rq
*rq
)
2834 * schedule_tail - first thing a freshly forked thread must call.
2835 * @prev: the thread we just switched away from.
2837 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2838 __releases(rq
->lock
)
2843 * New tasks start with FORK_PREEMPT_COUNT, see there and
2844 * finish_task_switch() for details.
2846 * finish_task_switch() will drop rq->lock() and lower preempt_count
2847 * and the preempt_enable() will end up enabling preemption (on
2848 * PREEMPT_COUNT kernels).
2851 rq
= finish_task_switch(prev
);
2852 balance_callback(rq
);
2855 if (current
->set_child_tid
)
2856 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2860 * context_switch - switch to the new MM and the new thread's register state.
2862 static __always_inline
struct rq
*
2863 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2864 struct task_struct
*next
, struct pin_cookie cookie
)
2866 struct mm_struct
*mm
, *oldmm
;
2868 prepare_task_switch(rq
, prev
, next
);
2871 oldmm
= prev
->active_mm
;
2873 * For paravirt, this is coupled with an exit in switch_to to
2874 * combine the page table reload and the switch backend into
2877 arch_start_context_switch(prev
);
2880 next
->active_mm
= oldmm
;
2881 atomic_inc(&oldmm
->mm_count
);
2882 enter_lazy_tlb(oldmm
, next
);
2884 switch_mm_irqs_off(oldmm
, mm
, next
);
2887 prev
->active_mm
= NULL
;
2888 rq
->prev_mm
= oldmm
;
2891 * Since the runqueue lock will be released by the next
2892 * task (which is an invalid locking op but in the case
2893 * of the scheduler it's an obvious special-case), so we
2894 * do an early lockdep release here:
2896 lockdep_unpin_lock(&rq
->lock
, cookie
);
2897 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2899 /* Here we just switch the register state and the stack. */
2900 switch_to(prev
, next
, prev
);
2903 return finish_task_switch(prev
);
2907 * nr_running and nr_context_switches:
2909 * externally visible scheduler statistics: current number of runnable
2910 * threads, total number of context switches performed since bootup.
2912 unsigned long nr_running(void)
2914 unsigned long i
, sum
= 0;
2916 for_each_online_cpu(i
)
2917 sum
+= cpu_rq(i
)->nr_running
;
2923 * Check if only the current task is running on the cpu.
2925 * Caution: this function does not check that the caller has disabled
2926 * preemption, thus the result might have a time-of-check-to-time-of-use
2927 * race. The caller is responsible to use it correctly, for example:
2929 * - from a non-preemptable section (of course)
2931 * - from a thread that is bound to a single CPU
2933 * - in a loop with very short iterations (e.g. a polling loop)
2935 bool single_task_running(void)
2937 return raw_rq()->nr_running
== 1;
2939 EXPORT_SYMBOL(single_task_running
);
2941 unsigned long long nr_context_switches(void)
2944 unsigned long long sum
= 0;
2946 for_each_possible_cpu(i
)
2947 sum
+= cpu_rq(i
)->nr_switches
;
2952 unsigned long nr_iowait(void)
2954 unsigned long i
, sum
= 0;
2956 for_each_possible_cpu(i
)
2957 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2962 unsigned long nr_iowait_cpu(int cpu
)
2964 struct rq
*this = cpu_rq(cpu
);
2965 return atomic_read(&this->nr_iowait
);
2968 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2970 struct rq
*rq
= this_rq();
2971 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2972 *load
= rq
->load
.weight
;
2978 * sched_exec - execve() is a valuable balancing opportunity, because at
2979 * this point the task has the smallest effective memory and cache footprint.
2981 void sched_exec(void)
2983 struct task_struct
*p
= current
;
2984 unsigned long flags
;
2987 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2988 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2989 if (dest_cpu
== smp_processor_id())
2992 if (likely(cpu_active(dest_cpu
))) {
2993 struct migration_arg arg
= { p
, dest_cpu
};
2995 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2996 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3000 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3005 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3006 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3008 EXPORT_PER_CPU_SYMBOL(kstat
);
3009 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3012 * The function fair_sched_class.update_curr accesses the struct curr
3013 * and its field curr->exec_start; when called from task_sched_runtime(),
3014 * we observe a high rate of cache misses in practice.
3015 * Prefetching this data results in improved performance.
3017 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3019 #ifdef CONFIG_FAIR_GROUP_SCHED
3020 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3022 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3025 prefetch(&curr
->exec_start
);
3029 * Return accounted runtime for the task.
3030 * In case the task is currently running, return the runtime plus current's
3031 * pending runtime that have not been accounted yet.
3033 unsigned long long task_sched_runtime(struct task_struct
*p
)
3039 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3041 * 64-bit doesn't need locks to atomically read a 64bit value.
3042 * So we have a optimization chance when the task's delta_exec is 0.
3043 * Reading ->on_cpu is racy, but this is ok.
3045 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3046 * If we race with it entering cpu, unaccounted time is 0. This is
3047 * indistinguishable from the read occurring a few cycles earlier.
3048 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3049 * been accounted, so we're correct here as well.
3051 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3052 return p
->se
.sum_exec_runtime
;
3055 rq
= task_rq_lock(p
, &rf
);
3057 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3058 * project cycles that may never be accounted to this
3059 * thread, breaking clock_gettime().
3061 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3062 prefetch_curr_exec_start(p
);
3063 update_rq_clock(rq
);
3064 p
->sched_class
->update_curr(rq
);
3066 ns
= p
->se
.sum_exec_runtime
;
3067 task_rq_unlock(rq
, p
, &rf
);
3073 * This function gets called by the timer code, with HZ frequency.
3074 * We call it with interrupts disabled.
3076 void scheduler_tick(void)
3078 int cpu
= smp_processor_id();
3079 struct rq
*rq
= cpu_rq(cpu
);
3080 struct task_struct
*curr
= rq
->curr
;
3084 raw_spin_lock(&rq
->lock
);
3085 update_rq_clock(rq
);
3086 curr
->sched_class
->task_tick(rq
, curr
, 0);
3087 cpu_load_update_active(rq
);
3088 calc_global_load_tick(rq
);
3089 raw_spin_unlock(&rq
->lock
);
3091 perf_event_task_tick();
3094 rq
->idle_balance
= idle_cpu(cpu
);
3095 trigger_load_balance(rq
);
3097 rq_last_tick_reset(rq
);
3100 #ifdef CONFIG_NO_HZ_FULL
3102 * scheduler_tick_max_deferment
3104 * Keep at least one tick per second when a single
3105 * active task is running because the scheduler doesn't
3106 * yet completely support full dynticks environment.
3108 * This makes sure that uptime, CFS vruntime, load
3109 * balancing, etc... continue to move forward, even
3110 * with a very low granularity.
3112 * Return: Maximum deferment in nanoseconds.
3114 u64
scheduler_tick_max_deferment(void)
3116 struct rq
*rq
= this_rq();
3117 unsigned long next
, now
= READ_ONCE(jiffies
);
3119 next
= rq
->last_sched_tick
+ HZ
;
3121 if (time_before_eq(next
, now
))
3124 return jiffies_to_nsecs(next
- now
);
3128 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3129 defined(CONFIG_PREEMPT_TRACER))
3131 * If the value passed in is equal to the current preempt count
3132 * then we just disabled preemption. Start timing the latency.
3134 static inline void preempt_latency_start(int val
)
3136 if (preempt_count() == val
) {
3137 unsigned long ip
= get_lock_parent_ip();
3138 #ifdef CONFIG_DEBUG_PREEMPT
3139 current
->preempt_disable_ip
= ip
;
3141 trace_preempt_off(CALLER_ADDR0
, ip
);
3145 void preempt_count_add(int val
)
3147 #ifdef CONFIG_DEBUG_PREEMPT
3151 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3154 __preempt_count_add(val
);
3155 #ifdef CONFIG_DEBUG_PREEMPT
3157 * Spinlock count overflowing soon?
3159 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3162 preempt_latency_start(val
);
3164 EXPORT_SYMBOL(preempt_count_add
);
3165 NOKPROBE_SYMBOL(preempt_count_add
);
3168 * If the value passed in equals to the current preempt count
3169 * then we just enabled preemption. Stop timing the latency.
3171 static inline void preempt_latency_stop(int val
)
3173 if (preempt_count() == val
)
3174 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3177 void preempt_count_sub(int val
)
3179 #ifdef CONFIG_DEBUG_PREEMPT
3183 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3186 * Is the spinlock portion underflowing?
3188 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3189 !(preempt_count() & PREEMPT_MASK
)))
3193 preempt_latency_stop(val
);
3194 __preempt_count_sub(val
);
3196 EXPORT_SYMBOL(preempt_count_sub
);
3197 NOKPROBE_SYMBOL(preempt_count_sub
);
3200 static inline void preempt_latency_start(int val
) { }
3201 static inline void preempt_latency_stop(int val
) { }
3205 * Print scheduling while atomic bug:
3207 static noinline
void __schedule_bug(struct task_struct
*prev
)
3209 /* Save this before calling printk(), since that will clobber it */
3210 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3212 if (oops_in_progress
)
3215 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3216 prev
->comm
, prev
->pid
, preempt_count());
3218 debug_show_held_locks(prev
);
3220 if (irqs_disabled())
3221 print_irqtrace_events(prev
);
3222 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3223 && in_atomic_preempt_off()) {
3224 pr_err("Preemption disabled at:");
3225 print_ip_sym(preempt_disable_ip
);
3229 panic("scheduling while atomic\n");
3232 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3236 * Various schedule()-time debugging checks and statistics:
3238 static inline void schedule_debug(struct task_struct
*prev
)
3240 #ifdef CONFIG_SCHED_STACK_END_CHECK
3241 if (task_stack_end_corrupted(prev
))
3242 panic("corrupted stack end detected inside scheduler\n");
3245 if (unlikely(in_atomic_preempt_off())) {
3246 __schedule_bug(prev
);
3247 preempt_count_set(PREEMPT_DISABLED
);
3251 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3253 schedstat_inc(this_rq()->sched_count
);
3257 * Pick up the highest-prio task:
3259 static inline struct task_struct
*
3260 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
3262 const struct sched_class
*class = &fair_sched_class
;
3263 struct task_struct
*p
;
3266 * Optimization: we know that if all tasks are in
3267 * the fair class we can call that function directly:
3269 if (likely(prev
->sched_class
== class &&
3270 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3271 p
= fair_sched_class
.pick_next_task(rq
, prev
, cookie
);
3272 if (unlikely(p
== RETRY_TASK
))
3275 /* assumes fair_sched_class->next == idle_sched_class */
3277 p
= idle_sched_class
.pick_next_task(rq
, prev
, cookie
);
3283 for_each_class(class) {
3284 p
= class->pick_next_task(rq
, prev
, cookie
);
3286 if (unlikely(p
== RETRY_TASK
))
3292 BUG(); /* the idle class will always have a runnable task */
3296 * __schedule() is the main scheduler function.
3298 * The main means of driving the scheduler and thus entering this function are:
3300 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3302 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3303 * paths. For example, see arch/x86/entry_64.S.
3305 * To drive preemption between tasks, the scheduler sets the flag in timer
3306 * interrupt handler scheduler_tick().
3308 * 3. Wakeups don't really cause entry into schedule(). They add a
3309 * task to the run-queue and that's it.
3311 * Now, if the new task added to the run-queue preempts the current
3312 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3313 * called on the nearest possible occasion:
3315 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3317 * - in syscall or exception context, at the next outmost
3318 * preempt_enable(). (this might be as soon as the wake_up()'s
3321 * - in IRQ context, return from interrupt-handler to
3322 * preemptible context
3324 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3327 * - cond_resched() call
3328 * - explicit schedule() call
3329 * - return from syscall or exception to user-space
3330 * - return from interrupt-handler to user-space
3332 * WARNING: must be called with preemption disabled!
3334 static void __sched notrace
__schedule(bool preempt
)
3336 struct task_struct
*prev
, *next
;
3337 unsigned long *switch_count
;
3338 struct pin_cookie cookie
;
3342 cpu
= smp_processor_id();
3346 schedule_debug(prev
);
3348 if (sched_feat(HRTICK
))
3351 local_irq_disable();
3352 rcu_note_context_switch();
3355 * Make sure that signal_pending_state()->signal_pending() below
3356 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3357 * done by the caller to avoid the race with signal_wake_up().
3359 smp_mb__before_spinlock();
3360 raw_spin_lock(&rq
->lock
);
3361 cookie
= lockdep_pin_lock(&rq
->lock
);
3363 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3365 switch_count
= &prev
->nivcsw
;
3366 if (!preempt
&& prev
->state
) {
3367 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3368 prev
->state
= TASK_RUNNING
;
3370 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3374 * If a worker went to sleep, notify and ask workqueue
3375 * whether it wants to wake up a task to maintain
3378 if (prev
->flags
& PF_WQ_WORKER
) {
3379 struct task_struct
*to_wakeup
;
3381 to_wakeup
= wq_worker_sleeping(prev
);
3383 try_to_wake_up_local(to_wakeup
, cookie
);
3386 switch_count
= &prev
->nvcsw
;
3389 if (task_on_rq_queued(prev
))
3390 update_rq_clock(rq
);
3392 next
= pick_next_task(rq
, prev
, cookie
);
3393 clear_tsk_need_resched(prev
);
3394 clear_preempt_need_resched();
3395 rq
->clock_skip_update
= 0;
3397 if (likely(prev
!= next
)) {
3402 trace_sched_switch(preempt
, prev
, next
);
3403 rq
= context_switch(rq
, prev
, next
, cookie
); /* unlocks the rq */
3405 lockdep_unpin_lock(&rq
->lock
, cookie
);
3406 raw_spin_unlock_irq(&rq
->lock
);
3409 balance_callback(rq
);
3412 void __noreturn
do_task_dead(void)
3415 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3416 * when the following two conditions become true.
3417 * - There is race condition of mmap_sem (It is acquired by
3419 * - SMI occurs before setting TASK_RUNINNG.
3420 * (or hypervisor of virtual machine switches to other guest)
3421 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3423 * To avoid it, we have to wait for releasing tsk->pi_lock which
3424 * is held by try_to_wake_up()
3427 raw_spin_unlock_wait(¤t
->pi_lock
);
3429 /* causes final put_task_struct in finish_task_switch(). */
3430 __set_current_state(TASK_DEAD
);
3431 current
->flags
|= PF_NOFREEZE
; /* tell freezer to ignore us */
3434 /* Avoid "noreturn function does return". */
3436 cpu_relax(); /* For when BUG is null */
3439 static inline void sched_submit_work(struct task_struct
*tsk
)
3441 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3444 * If we are going to sleep and we have plugged IO queued,
3445 * make sure to submit it to avoid deadlocks.
3447 if (blk_needs_flush_plug(tsk
))
3448 blk_schedule_flush_plug(tsk
);
3451 asmlinkage __visible
void __sched
schedule(void)
3453 struct task_struct
*tsk
= current
;
3455 sched_submit_work(tsk
);
3459 sched_preempt_enable_no_resched();
3460 } while (need_resched());
3462 EXPORT_SYMBOL(schedule
);
3464 #ifdef CONFIG_CONTEXT_TRACKING
3465 asmlinkage __visible
void __sched
schedule_user(void)
3468 * If we come here after a random call to set_need_resched(),
3469 * or we have been woken up remotely but the IPI has not yet arrived,
3470 * we haven't yet exited the RCU idle mode. Do it here manually until
3471 * we find a better solution.
3473 * NB: There are buggy callers of this function. Ideally we
3474 * should warn if prev_state != CONTEXT_USER, but that will trigger
3475 * too frequently to make sense yet.
3477 enum ctx_state prev_state
= exception_enter();
3479 exception_exit(prev_state
);
3484 * schedule_preempt_disabled - called with preemption disabled
3486 * Returns with preemption disabled. Note: preempt_count must be 1
3488 void __sched
schedule_preempt_disabled(void)
3490 sched_preempt_enable_no_resched();
3495 static void __sched notrace
preempt_schedule_common(void)
3499 * Because the function tracer can trace preempt_count_sub()
3500 * and it also uses preempt_enable/disable_notrace(), if
3501 * NEED_RESCHED is set, the preempt_enable_notrace() called
3502 * by the function tracer will call this function again and
3503 * cause infinite recursion.
3505 * Preemption must be disabled here before the function
3506 * tracer can trace. Break up preempt_disable() into two
3507 * calls. One to disable preemption without fear of being
3508 * traced. The other to still record the preemption latency,
3509 * which can also be traced by the function tracer.
3511 preempt_disable_notrace();
3512 preempt_latency_start(1);
3514 preempt_latency_stop(1);
3515 preempt_enable_no_resched_notrace();
3518 * Check again in case we missed a preemption opportunity
3519 * between schedule and now.
3521 } while (need_resched());
3524 #ifdef CONFIG_PREEMPT
3526 * this is the entry point to schedule() from in-kernel preemption
3527 * off of preempt_enable. Kernel preemptions off return from interrupt
3528 * occur there and call schedule directly.
3530 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3533 * If there is a non-zero preempt_count or interrupts are disabled,
3534 * we do not want to preempt the current task. Just return..
3536 if (likely(!preemptible()))
3539 preempt_schedule_common();
3541 NOKPROBE_SYMBOL(preempt_schedule
);
3542 EXPORT_SYMBOL(preempt_schedule
);
3545 * preempt_schedule_notrace - preempt_schedule called by tracing
3547 * The tracing infrastructure uses preempt_enable_notrace to prevent
3548 * recursion and tracing preempt enabling caused by the tracing
3549 * infrastructure itself. But as tracing can happen in areas coming
3550 * from userspace or just about to enter userspace, a preempt enable
3551 * can occur before user_exit() is called. This will cause the scheduler
3552 * to be called when the system is still in usermode.
3554 * To prevent this, the preempt_enable_notrace will use this function
3555 * instead of preempt_schedule() to exit user context if needed before
3556 * calling the scheduler.
3558 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3560 enum ctx_state prev_ctx
;
3562 if (likely(!preemptible()))
3567 * Because the function tracer can trace preempt_count_sub()
3568 * and it also uses preempt_enable/disable_notrace(), if
3569 * NEED_RESCHED is set, the preempt_enable_notrace() called
3570 * by the function tracer will call this function again and
3571 * cause infinite recursion.
3573 * Preemption must be disabled here before the function
3574 * tracer can trace. Break up preempt_disable() into two
3575 * calls. One to disable preemption without fear of being
3576 * traced. The other to still record the preemption latency,
3577 * which can also be traced by the function tracer.
3579 preempt_disable_notrace();
3580 preempt_latency_start(1);
3582 * Needs preempt disabled in case user_exit() is traced
3583 * and the tracer calls preempt_enable_notrace() causing
3584 * an infinite recursion.
3586 prev_ctx
= exception_enter();
3588 exception_exit(prev_ctx
);
3590 preempt_latency_stop(1);
3591 preempt_enable_no_resched_notrace();
3592 } while (need_resched());
3594 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3596 #endif /* CONFIG_PREEMPT */
3599 * this is the entry point to schedule() from kernel preemption
3600 * off of irq context.
3601 * Note, that this is called and return with irqs disabled. This will
3602 * protect us against recursive calling from irq.
3604 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3606 enum ctx_state prev_state
;
3608 /* Catch callers which need to be fixed */
3609 BUG_ON(preempt_count() || !irqs_disabled());
3611 prev_state
= exception_enter();
3617 local_irq_disable();
3618 sched_preempt_enable_no_resched();
3619 } while (need_resched());
3621 exception_exit(prev_state
);
3624 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3627 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3629 EXPORT_SYMBOL(default_wake_function
);
3631 #ifdef CONFIG_RT_MUTEXES
3634 * rt_mutex_setprio - set the current priority of a task
3636 * @prio: prio value (kernel-internal form)
3638 * This function changes the 'effective' priority of a task. It does
3639 * not touch ->normal_prio like __setscheduler().
3641 * Used by the rt_mutex code to implement priority inheritance
3642 * logic. Call site only calls if the priority of the task changed.
3644 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3646 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3647 const struct sched_class
*prev_class
;
3651 BUG_ON(prio
> MAX_PRIO
);
3653 rq
= __task_rq_lock(p
, &rf
);
3656 * Idle task boosting is a nono in general. There is one
3657 * exception, when PREEMPT_RT and NOHZ is active:
3659 * The idle task calls get_next_timer_interrupt() and holds
3660 * the timer wheel base->lock on the CPU and another CPU wants
3661 * to access the timer (probably to cancel it). We can safely
3662 * ignore the boosting request, as the idle CPU runs this code
3663 * with interrupts disabled and will complete the lock
3664 * protected section without being interrupted. So there is no
3665 * real need to boost.
3667 if (unlikely(p
== rq
->idle
)) {
3668 WARN_ON(p
!= rq
->curr
);
3669 WARN_ON(p
->pi_blocked_on
);
3673 trace_sched_pi_setprio(p
, prio
);
3676 if (oldprio
== prio
)
3677 queue_flag
&= ~DEQUEUE_MOVE
;
3679 prev_class
= p
->sched_class
;
3680 queued
= task_on_rq_queued(p
);
3681 running
= task_current(rq
, p
);
3683 dequeue_task(rq
, p
, queue_flag
);
3685 put_prev_task(rq
, p
);
3688 * Boosting condition are:
3689 * 1. -rt task is running and holds mutex A
3690 * --> -dl task blocks on mutex A
3692 * 2. -dl task is running and holds mutex A
3693 * --> -dl task blocks on mutex A and could preempt the
3696 if (dl_prio(prio
)) {
3697 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3698 if (!dl_prio(p
->normal_prio
) ||
3699 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3700 p
->dl
.dl_boosted
= 1;
3701 queue_flag
|= ENQUEUE_REPLENISH
;
3703 p
->dl
.dl_boosted
= 0;
3704 p
->sched_class
= &dl_sched_class
;
3705 } else if (rt_prio(prio
)) {
3706 if (dl_prio(oldprio
))
3707 p
->dl
.dl_boosted
= 0;
3709 queue_flag
|= ENQUEUE_HEAD
;
3710 p
->sched_class
= &rt_sched_class
;
3712 if (dl_prio(oldprio
))
3713 p
->dl
.dl_boosted
= 0;
3714 if (rt_prio(oldprio
))
3716 p
->sched_class
= &fair_sched_class
;
3722 enqueue_task(rq
, p
, queue_flag
);
3724 set_curr_task(rq
, p
);
3726 check_class_changed(rq
, p
, prev_class
, oldprio
);
3728 preempt_disable(); /* avoid rq from going away on us */
3729 __task_rq_unlock(rq
, &rf
);
3731 balance_callback(rq
);
3736 void set_user_nice(struct task_struct
*p
, long nice
)
3738 bool queued
, running
;
3739 int old_prio
, delta
;
3743 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3746 * We have to be careful, if called from sys_setpriority(),
3747 * the task might be in the middle of scheduling on another CPU.
3749 rq
= task_rq_lock(p
, &rf
);
3751 * The RT priorities are set via sched_setscheduler(), but we still
3752 * allow the 'normal' nice value to be set - but as expected
3753 * it wont have any effect on scheduling until the task is
3754 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3756 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3757 p
->static_prio
= NICE_TO_PRIO(nice
);
3760 queued
= task_on_rq_queued(p
);
3761 running
= task_current(rq
, p
);
3763 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3765 put_prev_task(rq
, p
);
3767 p
->static_prio
= NICE_TO_PRIO(nice
);
3770 p
->prio
= effective_prio(p
);
3771 delta
= p
->prio
- old_prio
;
3774 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3776 * If the task increased its priority or is running and
3777 * lowered its priority, then reschedule its CPU:
3779 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3783 set_curr_task(rq
, p
);
3785 task_rq_unlock(rq
, p
, &rf
);
3787 EXPORT_SYMBOL(set_user_nice
);
3790 * can_nice - check if a task can reduce its nice value
3794 int can_nice(const struct task_struct
*p
, const int nice
)
3796 /* convert nice value [19,-20] to rlimit style value [1,40] */
3797 int nice_rlim
= nice_to_rlimit(nice
);
3799 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3800 capable(CAP_SYS_NICE
));
3803 #ifdef __ARCH_WANT_SYS_NICE
3806 * sys_nice - change the priority of the current process.
3807 * @increment: priority increment
3809 * sys_setpriority is a more generic, but much slower function that
3810 * does similar things.
3812 SYSCALL_DEFINE1(nice
, int, increment
)
3817 * Setpriority might change our priority at the same moment.
3818 * We don't have to worry. Conceptually one call occurs first
3819 * and we have a single winner.
3821 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3822 nice
= task_nice(current
) + increment
;
3824 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3825 if (increment
< 0 && !can_nice(current
, nice
))
3828 retval
= security_task_setnice(current
, nice
);
3832 set_user_nice(current
, nice
);
3839 * task_prio - return the priority value of a given task.
3840 * @p: the task in question.
3842 * Return: The priority value as seen by users in /proc.
3843 * RT tasks are offset by -200. Normal tasks are centered
3844 * around 0, value goes from -16 to +15.
3846 int task_prio(const struct task_struct
*p
)
3848 return p
->prio
- MAX_RT_PRIO
;
3852 * idle_cpu - is a given cpu idle currently?
3853 * @cpu: the processor in question.
3855 * Return: 1 if the CPU is currently idle. 0 otherwise.
3857 int idle_cpu(int cpu
)
3859 struct rq
*rq
= cpu_rq(cpu
);
3861 if (rq
->curr
!= rq
->idle
)
3868 if (!llist_empty(&rq
->wake_list
))
3876 * idle_task - return the idle task for a given cpu.
3877 * @cpu: the processor in question.
3879 * Return: The idle task for the cpu @cpu.
3881 struct task_struct
*idle_task(int cpu
)
3883 return cpu_rq(cpu
)->idle
;
3887 * find_process_by_pid - find a process with a matching PID value.
3888 * @pid: the pid in question.
3890 * The task of @pid, if found. %NULL otherwise.
3892 static struct task_struct
*find_process_by_pid(pid_t pid
)
3894 return pid
? find_task_by_vpid(pid
) : current
;
3898 * This function initializes the sched_dl_entity of a newly becoming
3899 * SCHED_DEADLINE task.
3901 * Only the static values are considered here, the actual runtime and the
3902 * absolute deadline will be properly calculated when the task is enqueued
3903 * for the first time with its new policy.
3906 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3908 struct sched_dl_entity
*dl_se
= &p
->dl
;
3910 dl_se
->dl_runtime
= attr
->sched_runtime
;
3911 dl_se
->dl_deadline
= attr
->sched_deadline
;
3912 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3913 dl_se
->flags
= attr
->sched_flags
;
3914 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3917 * Changing the parameters of a task is 'tricky' and we're not doing
3918 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3920 * What we SHOULD do is delay the bandwidth release until the 0-lag
3921 * point. This would include retaining the task_struct until that time
3922 * and change dl_overflow() to not immediately decrement the current
3925 * Instead we retain the current runtime/deadline and let the new
3926 * parameters take effect after the current reservation period lapses.
3927 * This is safe (albeit pessimistic) because the 0-lag point is always
3928 * before the current scheduling deadline.
3930 * We can still have temporary overloads because we do not delay the
3931 * change in bandwidth until that time; so admission control is
3932 * not on the safe side. It does however guarantee tasks will never
3933 * consume more than promised.
3938 * sched_setparam() passes in -1 for its policy, to let the functions
3939 * it calls know not to change it.
3941 #define SETPARAM_POLICY -1
3943 static void __setscheduler_params(struct task_struct
*p
,
3944 const struct sched_attr
*attr
)
3946 int policy
= attr
->sched_policy
;
3948 if (policy
== SETPARAM_POLICY
)
3953 if (dl_policy(policy
))
3954 __setparam_dl(p
, attr
);
3955 else if (fair_policy(policy
))
3956 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3959 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3960 * !rt_policy. Always setting this ensures that things like
3961 * getparam()/getattr() don't report silly values for !rt tasks.
3963 p
->rt_priority
= attr
->sched_priority
;
3964 p
->normal_prio
= normal_prio(p
);
3968 /* Actually do priority change: must hold pi & rq lock. */
3969 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3970 const struct sched_attr
*attr
, bool keep_boost
)
3972 __setscheduler_params(p
, attr
);
3975 * Keep a potential priority boosting if called from
3976 * sched_setscheduler().
3979 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3981 p
->prio
= normal_prio(p
);
3983 if (dl_prio(p
->prio
))
3984 p
->sched_class
= &dl_sched_class
;
3985 else if (rt_prio(p
->prio
))
3986 p
->sched_class
= &rt_sched_class
;
3988 p
->sched_class
= &fair_sched_class
;
3992 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3994 struct sched_dl_entity
*dl_se
= &p
->dl
;
3996 attr
->sched_priority
= p
->rt_priority
;
3997 attr
->sched_runtime
= dl_se
->dl_runtime
;
3998 attr
->sched_deadline
= dl_se
->dl_deadline
;
3999 attr
->sched_period
= dl_se
->dl_period
;
4000 attr
->sched_flags
= dl_se
->flags
;
4004 * This function validates the new parameters of a -deadline task.
4005 * We ask for the deadline not being zero, and greater or equal
4006 * than the runtime, as well as the period of being zero or
4007 * greater than deadline. Furthermore, we have to be sure that
4008 * user parameters are above the internal resolution of 1us (we
4009 * check sched_runtime only since it is always the smaller one) and
4010 * below 2^63 ns (we have to check both sched_deadline and
4011 * sched_period, as the latter can be zero).
4014 __checkparam_dl(const struct sched_attr
*attr
)
4017 if (attr
->sched_deadline
== 0)
4021 * Since we truncate DL_SCALE bits, make sure we're at least
4024 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
4028 * Since we use the MSB for wrap-around and sign issues, make
4029 * sure it's not set (mind that period can be equal to zero).
4031 if (attr
->sched_deadline
& (1ULL << 63) ||
4032 attr
->sched_period
& (1ULL << 63))
4035 /* runtime <= deadline <= period (if period != 0) */
4036 if ((attr
->sched_period
!= 0 &&
4037 attr
->sched_period
< attr
->sched_deadline
) ||
4038 attr
->sched_deadline
< attr
->sched_runtime
)
4045 * check the target process has a UID that matches the current process's
4047 static bool check_same_owner(struct task_struct
*p
)
4049 const struct cred
*cred
= current_cred(), *pcred
;
4053 pcred
= __task_cred(p
);
4054 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4055 uid_eq(cred
->euid
, pcred
->uid
));
4060 static bool dl_param_changed(struct task_struct
*p
,
4061 const struct sched_attr
*attr
)
4063 struct sched_dl_entity
*dl_se
= &p
->dl
;
4065 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
4066 dl_se
->dl_deadline
!= attr
->sched_deadline
||
4067 dl_se
->dl_period
!= attr
->sched_period
||
4068 dl_se
->flags
!= attr
->sched_flags
)
4074 static int __sched_setscheduler(struct task_struct
*p
,
4075 const struct sched_attr
*attr
,
4078 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4079 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4080 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4081 int new_effective_prio
, policy
= attr
->sched_policy
;
4082 const struct sched_class
*prev_class
;
4085 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
4088 /* may grab non-irq protected spin_locks */
4089 BUG_ON(in_interrupt());
4091 /* double check policy once rq lock held */
4093 reset_on_fork
= p
->sched_reset_on_fork
;
4094 policy
= oldpolicy
= p
->policy
;
4096 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4098 if (!valid_policy(policy
))
4102 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4106 * Valid priorities for SCHED_FIFO and SCHED_RR are
4107 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4108 * SCHED_BATCH and SCHED_IDLE is 0.
4110 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4111 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4113 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4114 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4118 * Allow unprivileged RT tasks to decrease priority:
4120 if (user
&& !capable(CAP_SYS_NICE
)) {
4121 if (fair_policy(policy
)) {
4122 if (attr
->sched_nice
< task_nice(p
) &&
4123 !can_nice(p
, attr
->sched_nice
))
4127 if (rt_policy(policy
)) {
4128 unsigned long rlim_rtprio
=
4129 task_rlimit(p
, RLIMIT_RTPRIO
);
4131 /* can't set/change the rt policy */
4132 if (policy
!= p
->policy
&& !rlim_rtprio
)
4135 /* can't increase priority */
4136 if (attr
->sched_priority
> p
->rt_priority
&&
4137 attr
->sched_priority
> rlim_rtprio
)
4142 * Can't set/change SCHED_DEADLINE policy at all for now
4143 * (safest behavior); in the future we would like to allow
4144 * unprivileged DL tasks to increase their relative deadline
4145 * or reduce their runtime (both ways reducing utilization)
4147 if (dl_policy(policy
))
4151 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4152 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4154 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4155 if (!can_nice(p
, task_nice(p
)))
4159 /* can't change other user's priorities */
4160 if (!check_same_owner(p
))
4163 /* Normal users shall not reset the sched_reset_on_fork flag */
4164 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4169 retval
= security_task_setscheduler(p
);
4175 * make sure no PI-waiters arrive (or leave) while we are
4176 * changing the priority of the task:
4178 * To be able to change p->policy safely, the appropriate
4179 * runqueue lock must be held.
4181 rq
= task_rq_lock(p
, &rf
);
4184 * Changing the policy of the stop threads its a very bad idea
4186 if (p
== rq
->stop
) {
4187 task_rq_unlock(rq
, p
, &rf
);
4192 * If not changing anything there's no need to proceed further,
4193 * but store a possible modification of reset_on_fork.
4195 if (unlikely(policy
== p
->policy
)) {
4196 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4198 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4200 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4203 p
->sched_reset_on_fork
= reset_on_fork
;
4204 task_rq_unlock(rq
, p
, &rf
);
4210 #ifdef CONFIG_RT_GROUP_SCHED
4212 * Do not allow realtime tasks into groups that have no runtime
4215 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4216 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4217 !task_group_is_autogroup(task_group(p
))) {
4218 task_rq_unlock(rq
, p
, &rf
);
4223 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4224 cpumask_t
*span
= rq
->rd
->span
;
4227 * Don't allow tasks with an affinity mask smaller than
4228 * the entire root_domain to become SCHED_DEADLINE. We
4229 * will also fail if there's no bandwidth available.
4231 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4232 rq
->rd
->dl_bw
.bw
== 0) {
4233 task_rq_unlock(rq
, p
, &rf
);
4240 /* recheck policy now with rq lock held */
4241 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4242 policy
= oldpolicy
= -1;
4243 task_rq_unlock(rq
, p
, &rf
);
4248 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4249 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4252 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4253 task_rq_unlock(rq
, p
, &rf
);
4257 p
->sched_reset_on_fork
= reset_on_fork
;
4262 * Take priority boosted tasks into account. If the new
4263 * effective priority is unchanged, we just store the new
4264 * normal parameters and do not touch the scheduler class and
4265 * the runqueue. This will be done when the task deboost
4268 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4269 if (new_effective_prio
== oldprio
)
4270 queue_flags
&= ~DEQUEUE_MOVE
;
4273 queued
= task_on_rq_queued(p
);
4274 running
= task_current(rq
, p
);
4276 dequeue_task(rq
, p
, queue_flags
);
4278 put_prev_task(rq
, p
);
4280 prev_class
= p
->sched_class
;
4281 __setscheduler(rq
, p
, attr
, pi
);
4285 * We enqueue to tail when the priority of a task is
4286 * increased (user space view).
4288 if (oldprio
< p
->prio
)
4289 queue_flags
|= ENQUEUE_HEAD
;
4291 enqueue_task(rq
, p
, queue_flags
);
4294 set_curr_task(rq
, p
);
4296 check_class_changed(rq
, p
, prev_class
, oldprio
);
4297 preempt_disable(); /* avoid rq from going away on us */
4298 task_rq_unlock(rq
, p
, &rf
);
4301 rt_mutex_adjust_pi(p
);
4304 * Run balance callbacks after we've adjusted the PI chain.
4306 balance_callback(rq
);
4312 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4313 const struct sched_param
*param
, bool check
)
4315 struct sched_attr attr
= {
4316 .sched_policy
= policy
,
4317 .sched_priority
= param
->sched_priority
,
4318 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4321 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4322 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4323 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4324 policy
&= ~SCHED_RESET_ON_FORK
;
4325 attr
.sched_policy
= policy
;
4328 return __sched_setscheduler(p
, &attr
, check
, true);
4331 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4332 * @p: the task in question.
4333 * @policy: new policy.
4334 * @param: structure containing the new RT priority.
4336 * Return: 0 on success. An error code otherwise.
4338 * NOTE that the task may be already dead.
4340 int sched_setscheduler(struct task_struct
*p
, int policy
,
4341 const struct sched_param
*param
)
4343 return _sched_setscheduler(p
, policy
, param
, true);
4345 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4347 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4349 return __sched_setscheduler(p
, attr
, true, true);
4351 EXPORT_SYMBOL_GPL(sched_setattr
);
4354 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4355 * @p: the task in question.
4356 * @policy: new policy.
4357 * @param: structure containing the new RT priority.
4359 * Just like sched_setscheduler, only don't bother checking if the
4360 * current context has permission. For example, this is needed in
4361 * stop_machine(): we create temporary high priority worker threads,
4362 * but our caller might not have that capability.
4364 * Return: 0 on success. An error code otherwise.
4366 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4367 const struct sched_param
*param
)
4369 return _sched_setscheduler(p
, policy
, param
, false);
4371 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4374 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4376 struct sched_param lparam
;
4377 struct task_struct
*p
;
4380 if (!param
|| pid
< 0)
4382 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4387 p
= find_process_by_pid(pid
);
4389 retval
= sched_setscheduler(p
, policy
, &lparam
);
4396 * Mimics kernel/events/core.c perf_copy_attr().
4398 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4399 struct sched_attr
*attr
)
4404 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4408 * zero the full structure, so that a short copy will be nice.
4410 memset(attr
, 0, sizeof(*attr
));
4412 ret
= get_user(size
, &uattr
->size
);
4416 if (size
> PAGE_SIZE
) /* silly large */
4419 if (!size
) /* abi compat */
4420 size
= SCHED_ATTR_SIZE_VER0
;
4422 if (size
< SCHED_ATTR_SIZE_VER0
)
4426 * If we're handed a bigger struct than we know of,
4427 * ensure all the unknown bits are 0 - i.e. new
4428 * user-space does not rely on any kernel feature
4429 * extensions we dont know about yet.
4431 if (size
> sizeof(*attr
)) {
4432 unsigned char __user
*addr
;
4433 unsigned char __user
*end
;
4436 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4437 end
= (void __user
*)uattr
+ size
;
4439 for (; addr
< end
; addr
++) {
4440 ret
= get_user(val
, addr
);
4446 size
= sizeof(*attr
);
4449 ret
= copy_from_user(attr
, uattr
, size
);
4454 * XXX: do we want to be lenient like existing syscalls; or do we want
4455 * to be strict and return an error on out-of-bounds values?
4457 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4462 put_user(sizeof(*attr
), &uattr
->size
);
4467 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4468 * @pid: the pid in question.
4469 * @policy: new policy.
4470 * @param: structure containing the new RT priority.
4472 * Return: 0 on success. An error code otherwise.
4474 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4475 struct sched_param __user
*, param
)
4477 /* negative values for policy are not valid */
4481 return do_sched_setscheduler(pid
, policy
, param
);
4485 * sys_sched_setparam - set/change the RT priority of a thread
4486 * @pid: the pid in question.
4487 * @param: structure containing the new RT priority.
4489 * Return: 0 on success. An error code otherwise.
4491 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4493 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4497 * sys_sched_setattr - same as above, but with extended sched_attr
4498 * @pid: the pid in question.
4499 * @uattr: structure containing the extended parameters.
4500 * @flags: for future extension.
4502 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4503 unsigned int, flags
)
4505 struct sched_attr attr
;
4506 struct task_struct
*p
;
4509 if (!uattr
|| pid
< 0 || flags
)
4512 retval
= sched_copy_attr(uattr
, &attr
);
4516 if ((int)attr
.sched_policy
< 0)
4521 p
= find_process_by_pid(pid
);
4523 retval
= sched_setattr(p
, &attr
);
4530 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4531 * @pid: the pid in question.
4533 * Return: On success, the policy of the thread. Otherwise, a negative error
4536 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4538 struct task_struct
*p
;
4546 p
= find_process_by_pid(pid
);
4548 retval
= security_task_getscheduler(p
);
4551 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4558 * sys_sched_getparam - get the RT priority of a thread
4559 * @pid: the pid in question.
4560 * @param: structure containing the RT priority.
4562 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4565 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4567 struct sched_param lp
= { .sched_priority
= 0 };
4568 struct task_struct
*p
;
4571 if (!param
|| pid
< 0)
4575 p
= find_process_by_pid(pid
);
4580 retval
= security_task_getscheduler(p
);
4584 if (task_has_rt_policy(p
))
4585 lp
.sched_priority
= p
->rt_priority
;
4589 * This one might sleep, we cannot do it with a spinlock held ...
4591 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4600 static int sched_read_attr(struct sched_attr __user
*uattr
,
4601 struct sched_attr
*attr
,
4606 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4610 * If we're handed a smaller struct than we know of,
4611 * ensure all the unknown bits are 0 - i.e. old
4612 * user-space does not get uncomplete information.
4614 if (usize
< sizeof(*attr
)) {
4615 unsigned char *addr
;
4618 addr
= (void *)attr
+ usize
;
4619 end
= (void *)attr
+ sizeof(*attr
);
4621 for (; addr
< end
; addr
++) {
4629 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4637 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4638 * @pid: the pid in question.
4639 * @uattr: structure containing the extended parameters.
4640 * @size: sizeof(attr) for fwd/bwd comp.
4641 * @flags: for future extension.
4643 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4644 unsigned int, size
, unsigned int, flags
)
4646 struct sched_attr attr
= {
4647 .size
= sizeof(struct sched_attr
),
4649 struct task_struct
*p
;
4652 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4653 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4657 p
= find_process_by_pid(pid
);
4662 retval
= security_task_getscheduler(p
);
4666 attr
.sched_policy
= p
->policy
;
4667 if (p
->sched_reset_on_fork
)
4668 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4669 if (task_has_dl_policy(p
))
4670 __getparam_dl(p
, &attr
);
4671 else if (task_has_rt_policy(p
))
4672 attr
.sched_priority
= p
->rt_priority
;
4674 attr
.sched_nice
= task_nice(p
);
4678 retval
= sched_read_attr(uattr
, &attr
, size
);
4686 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4688 cpumask_var_t cpus_allowed
, new_mask
;
4689 struct task_struct
*p
;
4694 p
= find_process_by_pid(pid
);
4700 /* Prevent p going away */
4704 if (p
->flags
& PF_NO_SETAFFINITY
) {
4708 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4712 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4714 goto out_free_cpus_allowed
;
4717 if (!check_same_owner(p
)) {
4719 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4721 goto out_free_new_mask
;
4726 retval
= security_task_setscheduler(p
);
4728 goto out_free_new_mask
;
4731 cpuset_cpus_allowed(p
, cpus_allowed
);
4732 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4735 * Since bandwidth control happens on root_domain basis,
4736 * if admission test is enabled, we only admit -deadline
4737 * tasks allowed to run on all the CPUs in the task's
4741 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4743 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4746 goto out_free_new_mask
;
4752 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4755 cpuset_cpus_allowed(p
, cpus_allowed
);
4756 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4758 * We must have raced with a concurrent cpuset
4759 * update. Just reset the cpus_allowed to the
4760 * cpuset's cpus_allowed
4762 cpumask_copy(new_mask
, cpus_allowed
);
4767 free_cpumask_var(new_mask
);
4768 out_free_cpus_allowed
:
4769 free_cpumask_var(cpus_allowed
);
4775 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4776 struct cpumask
*new_mask
)
4778 if (len
< cpumask_size())
4779 cpumask_clear(new_mask
);
4780 else if (len
> cpumask_size())
4781 len
= cpumask_size();
4783 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4787 * sys_sched_setaffinity - set the cpu affinity of a process
4788 * @pid: pid of the process
4789 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4790 * @user_mask_ptr: user-space pointer to the new cpu mask
4792 * Return: 0 on success. An error code otherwise.
4794 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4795 unsigned long __user
*, user_mask_ptr
)
4797 cpumask_var_t new_mask
;
4800 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4803 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4805 retval
= sched_setaffinity(pid
, new_mask
);
4806 free_cpumask_var(new_mask
);
4810 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4812 struct task_struct
*p
;
4813 unsigned long flags
;
4819 p
= find_process_by_pid(pid
);
4823 retval
= security_task_getscheduler(p
);
4827 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4828 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4829 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4838 * sys_sched_getaffinity - get the cpu affinity of a process
4839 * @pid: pid of the process
4840 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4841 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4843 * Return: size of CPU mask copied to user_mask_ptr on success. An
4844 * error code otherwise.
4846 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4847 unsigned long __user
*, user_mask_ptr
)
4852 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4854 if (len
& (sizeof(unsigned long)-1))
4857 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4860 ret
= sched_getaffinity(pid
, mask
);
4862 size_t retlen
= min_t(size_t, len
, cpumask_size());
4864 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4869 free_cpumask_var(mask
);
4875 * sys_sched_yield - yield the current processor to other threads.
4877 * This function yields the current CPU to other tasks. If there are no
4878 * other threads running on this CPU then this function will return.
4882 SYSCALL_DEFINE0(sched_yield
)
4884 struct rq
*rq
= this_rq_lock();
4886 schedstat_inc(rq
->yld_count
);
4887 current
->sched_class
->yield_task(rq
);
4890 * Since we are going to call schedule() anyway, there's
4891 * no need to preempt or enable interrupts:
4893 __release(rq
->lock
);
4894 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4895 do_raw_spin_unlock(&rq
->lock
);
4896 sched_preempt_enable_no_resched();
4903 #ifndef CONFIG_PREEMPT
4904 int __sched
_cond_resched(void)
4906 if (should_resched(0)) {
4907 preempt_schedule_common();
4912 EXPORT_SYMBOL(_cond_resched
);
4916 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4917 * call schedule, and on return reacquire the lock.
4919 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4920 * operations here to prevent schedule() from being called twice (once via
4921 * spin_unlock(), once by hand).
4923 int __cond_resched_lock(spinlock_t
*lock
)
4925 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4928 lockdep_assert_held(lock
);
4930 if (spin_needbreak(lock
) || resched
) {
4933 preempt_schedule_common();
4941 EXPORT_SYMBOL(__cond_resched_lock
);
4943 int __sched
__cond_resched_softirq(void)
4945 BUG_ON(!in_softirq());
4947 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4949 preempt_schedule_common();
4955 EXPORT_SYMBOL(__cond_resched_softirq
);
4958 * yield - yield the current processor to other threads.
4960 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4962 * The scheduler is at all times free to pick the calling task as the most
4963 * eligible task to run, if removing the yield() call from your code breaks
4964 * it, its already broken.
4966 * Typical broken usage is:
4971 * where one assumes that yield() will let 'the other' process run that will
4972 * make event true. If the current task is a SCHED_FIFO task that will never
4973 * happen. Never use yield() as a progress guarantee!!
4975 * If you want to use yield() to wait for something, use wait_event().
4976 * If you want to use yield() to be 'nice' for others, use cond_resched().
4977 * If you still want to use yield(), do not!
4979 void __sched
yield(void)
4981 set_current_state(TASK_RUNNING
);
4984 EXPORT_SYMBOL(yield
);
4987 * yield_to - yield the current processor to another thread in
4988 * your thread group, or accelerate that thread toward the
4989 * processor it's on.
4991 * @preempt: whether task preemption is allowed or not
4993 * It's the caller's job to ensure that the target task struct
4994 * can't go away on us before we can do any checks.
4997 * true (>0) if we indeed boosted the target task.
4998 * false (0) if we failed to boost the target.
4999 * -ESRCH if there's no task to yield to.
5001 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5003 struct task_struct
*curr
= current
;
5004 struct rq
*rq
, *p_rq
;
5005 unsigned long flags
;
5008 local_irq_save(flags
);
5014 * If we're the only runnable task on the rq and target rq also
5015 * has only one task, there's absolutely no point in yielding.
5017 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5022 double_rq_lock(rq
, p_rq
);
5023 if (task_rq(p
) != p_rq
) {
5024 double_rq_unlock(rq
, p_rq
);
5028 if (!curr
->sched_class
->yield_to_task
)
5031 if (curr
->sched_class
!= p
->sched_class
)
5034 if (task_running(p_rq
, p
) || p
->state
)
5037 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5039 schedstat_inc(rq
->yld_count
);
5041 * Make p's CPU reschedule; pick_next_entity takes care of
5044 if (preempt
&& rq
!= p_rq
)
5049 double_rq_unlock(rq
, p_rq
);
5051 local_irq_restore(flags
);
5058 EXPORT_SYMBOL_GPL(yield_to
);
5061 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5062 * that process accounting knows that this is a task in IO wait state.
5064 long __sched
io_schedule_timeout(long timeout
)
5066 int old_iowait
= current
->in_iowait
;
5070 current
->in_iowait
= 1;
5071 blk_schedule_flush_plug(current
);
5073 delayacct_blkio_start();
5075 atomic_inc(&rq
->nr_iowait
);
5076 ret
= schedule_timeout(timeout
);
5077 current
->in_iowait
= old_iowait
;
5078 atomic_dec(&rq
->nr_iowait
);
5079 delayacct_blkio_end();
5083 EXPORT_SYMBOL(io_schedule_timeout
);
5086 * sys_sched_get_priority_max - return maximum RT priority.
5087 * @policy: scheduling class.
5089 * Return: On success, this syscall returns the maximum
5090 * rt_priority that can be used by a given scheduling class.
5091 * On failure, a negative error code is returned.
5093 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5100 ret
= MAX_USER_RT_PRIO
-1;
5102 case SCHED_DEADLINE
:
5113 * sys_sched_get_priority_min - return minimum RT priority.
5114 * @policy: scheduling class.
5116 * Return: On success, this syscall returns the minimum
5117 * rt_priority that can be used by a given scheduling class.
5118 * On failure, a negative error code is returned.
5120 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5129 case SCHED_DEADLINE
:
5139 * sys_sched_rr_get_interval - return the default timeslice of a process.
5140 * @pid: pid of the process.
5141 * @interval: userspace pointer to the timeslice value.
5143 * this syscall writes the default timeslice value of a given process
5144 * into the user-space timespec buffer. A value of '0' means infinity.
5146 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5149 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5150 struct timespec __user
*, interval
)
5152 struct task_struct
*p
;
5153 unsigned int time_slice
;
5164 p
= find_process_by_pid(pid
);
5168 retval
= security_task_getscheduler(p
);
5172 rq
= task_rq_lock(p
, &rf
);
5174 if (p
->sched_class
->get_rr_interval
)
5175 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5176 task_rq_unlock(rq
, p
, &rf
);
5179 jiffies_to_timespec(time_slice
, &t
);
5180 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5188 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5190 void sched_show_task(struct task_struct
*p
)
5192 unsigned long free
= 0;
5194 unsigned long state
= p
->state
;
5196 if (!try_get_task_stack(p
))
5199 state
= __ffs(state
) + 1;
5200 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5201 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5202 if (state
== TASK_RUNNING
)
5203 printk(KERN_CONT
" running task ");
5204 #ifdef CONFIG_DEBUG_STACK_USAGE
5205 free
= stack_not_used(p
);
5210 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5212 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5213 task_pid_nr(p
), ppid
,
5214 (unsigned long)task_thread_info(p
)->flags
);
5216 print_worker_info(KERN_INFO
, p
);
5217 show_stack(p
, NULL
);
5221 void show_state_filter(unsigned long state_filter
)
5223 struct task_struct
*g
, *p
;
5225 #if BITS_PER_LONG == 32
5227 " task PC stack pid father\n");
5230 " task PC stack pid father\n");
5233 for_each_process_thread(g
, p
) {
5235 * reset the NMI-timeout, listing all files on a slow
5236 * console might take a lot of time:
5237 * Also, reset softlockup watchdogs on all CPUs, because
5238 * another CPU might be blocked waiting for us to process
5241 touch_nmi_watchdog();
5242 touch_all_softlockup_watchdogs();
5243 if (!state_filter
|| (p
->state
& state_filter
))
5247 #ifdef CONFIG_SCHED_DEBUG
5249 sysrq_sched_debug_show();
5253 * Only show locks if all tasks are dumped:
5256 debug_show_all_locks();
5259 void init_idle_bootup_task(struct task_struct
*idle
)
5261 idle
->sched_class
= &idle_sched_class
;
5265 * init_idle - set up an idle thread for a given CPU
5266 * @idle: task in question
5267 * @cpu: cpu the idle task belongs to
5269 * NOTE: this function does not set the idle thread's NEED_RESCHED
5270 * flag, to make booting more robust.
5272 void init_idle(struct task_struct
*idle
, int cpu
)
5274 struct rq
*rq
= cpu_rq(cpu
);
5275 unsigned long flags
;
5277 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5278 raw_spin_lock(&rq
->lock
);
5280 __sched_fork(0, idle
);
5281 idle
->state
= TASK_RUNNING
;
5282 idle
->se
.exec_start
= sched_clock();
5284 kasan_unpoison_task_stack(idle
);
5288 * Its possible that init_idle() gets called multiple times on a task,
5289 * in that case do_set_cpus_allowed() will not do the right thing.
5291 * And since this is boot we can forgo the serialization.
5293 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5296 * We're having a chicken and egg problem, even though we are
5297 * holding rq->lock, the cpu isn't yet set to this cpu so the
5298 * lockdep check in task_group() will fail.
5300 * Similar case to sched_fork(). / Alternatively we could
5301 * use task_rq_lock() here and obtain the other rq->lock.
5306 __set_task_cpu(idle
, cpu
);
5309 rq
->curr
= rq
->idle
= idle
;
5310 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5314 raw_spin_unlock(&rq
->lock
);
5315 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5317 /* Set the preempt count _outside_ the spinlocks! */
5318 init_idle_preempt_count(idle
, cpu
);
5321 * The idle tasks have their own, simple scheduling class:
5323 idle
->sched_class
= &idle_sched_class
;
5324 ftrace_graph_init_idle_task(idle
, cpu
);
5325 vtime_init_idle(idle
, cpu
);
5327 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5331 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5332 const struct cpumask
*trial
)
5334 int ret
= 1, trial_cpus
;
5335 struct dl_bw
*cur_dl_b
;
5336 unsigned long flags
;
5338 if (!cpumask_weight(cur
))
5341 rcu_read_lock_sched();
5342 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5343 trial_cpus
= cpumask_weight(trial
);
5345 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5346 if (cur_dl_b
->bw
!= -1 &&
5347 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5349 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5350 rcu_read_unlock_sched();
5355 int task_can_attach(struct task_struct
*p
,
5356 const struct cpumask
*cs_cpus_allowed
)
5361 * Kthreads which disallow setaffinity shouldn't be moved
5362 * to a new cpuset; we don't want to change their cpu
5363 * affinity and isolating such threads by their set of
5364 * allowed nodes is unnecessary. Thus, cpusets are not
5365 * applicable for such threads. This prevents checking for
5366 * success of set_cpus_allowed_ptr() on all attached tasks
5367 * before cpus_allowed may be changed.
5369 if (p
->flags
& PF_NO_SETAFFINITY
) {
5375 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5377 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5382 unsigned long flags
;
5384 rcu_read_lock_sched();
5385 dl_b
= dl_bw_of(dest_cpu
);
5386 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5387 cpus
= dl_bw_cpus(dest_cpu
);
5388 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5393 * We reserve space for this task in the destination
5394 * root_domain, as we can't fail after this point.
5395 * We will free resources in the source root_domain
5396 * later on (see set_cpus_allowed_dl()).
5398 __dl_add(dl_b
, p
->dl
.dl_bw
);
5400 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5401 rcu_read_unlock_sched();
5411 static bool sched_smp_initialized __read_mostly
;
5413 #ifdef CONFIG_NUMA_BALANCING
5414 /* Migrate current task p to target_cpu */
5415 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5417 struct migration_arg arg
= { p
, target_cpu
};
5418 int curr_cpu
= task_cpu(p
);
5420 if (curr_cpu
== target_cpu
)
5423 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5426 /* TODO: This is not properly updating schedstats */
5428 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5429 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5433 * Requeue a task on a given node and accurately track the number of NUMA
5434 * tasks on the runqueues
5436 void sched_setnuma(struct task_struct
*p
, int nid
)
5438 bool queued
, running
;
5442 rq
= task_rq_lock(p
, &rf
);
5443 queued
= task_on_rq_queued(p
);
5444 running
= task_current(rq
, p
);
5447 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5449 put_prev_task(rq
, p
);
5451 p
->numa_preferred_nid
= nid
;
5454 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5456 set_curr_task(rq
, p
);
5457 task_rq_unlock(rq
, p
, &rf
);
5459 #endif /* CONFIG_NUMA_BALANCING */
5461 #ifdef CONFIG_HOTPLUG_CPU
5463 * Ensures that the idle task is using init_mm right before its cpu goes
5466 void idle_task_exit(void)
5468 struct mm_struct
*mm
= current
->active_mm
;
5470 BUG_ON(cpu_online(smp_processor_id()));
5472 if (mm
!= &init_mm
) {
5473 switch_mm_irqs_off(mm
, &init_mm
, current
);
5474 finish_arch_post_lock_switch();
5480 * Since this CPU is going 'away' for a while, fold any nr_active delta
5481 * we might have. Assumes we're called after migrate_tasks() so that the
5482 * nr_active count is stable. We need to take the teardown thread which
5483 * is calling this into account, so we hand in adjust = 1 to the load
5486 * Also see the comment "Global load-average calculations".
5488 static void calc_load_migrate(struct rq
*rq
)
5490 long delta
= calc_load_fold_active(rq
, 1);
5492 atomic_long_add(delta
, &calc_load_tasks
);
5495 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5499 static const struct sched_class fake_sched_class
= {
5500 .put_prev_task
= put_prev_task_fake
,
5503 static struct task_struct fake_task
= {
5505 * Avoid pull_{rt,dl}_task()
5507 .prio
= MAX_PRIO
+ 1,
5508 .sched_class
= &fake_sched_class
,
5512 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5513 * try_to_wake_up()->select_task_rq().
5515 * Called with rq->lock held even though we'er in stop_machine() and
5516 * there's no concurrency possible, we hold the required locks anyway
5517 * because of lock validation efforts.
5519 static void migrate_tasks(struct rq
*dead_rq
)
5521 struct rq
*rq
= dead_rq
;
5522 struct task_struct
*next
, *stop
= rq
->stop
;
5523 struct pin_cookie cookie
;
5527 * Fudge the rq selection such that the below task selection loop
5528 * doesn't get stuck on the currently eligible stop task.
5530 * We're currently inside stop_machine() and the rq is either stuck
5531 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5532 * either way we should never end up calling schedule() until we're
5538 * put_prev_task() and pick_next_task() sched
5539 * class method both need to have an up-to-date
5540 * value of rq->clock[_task]
5542 update_rq_clock(rq
);
5546 * There's this thread running, bail when that's the only
5549 if (rq
->nr_running
== 1)
5553 * pick_next_task assumes pinned rq->lock.
5555 cookie
= lockdep_pin_lock(&rq
->lock
);
5556 next
= pick_next_task(rq
, &fake_task
, cookie
);
5558 next
->sched_class
->put_prev_task(rq
, next
);
5561 * Rules for changing task_struct::cpus_allowed are holding
5562 * both pi_lock and rq->lock, such that holding either
5563 * stabilizes the mask.
5565 * Drop rq->lock is not quite as disastrous as it usually is
5566 * because !cpu_active at this point, which means load-balance
5567 * will not interfere. Also, stop-machine.
5569 lockdep_unpin_lock(&rq
->lock
, cookie
);
5570 raw_spin_unlock(&rq
->lock
);
5571 raw_spin_lock(&next
->pi_lock
);
5572 raw_spin_lock(&rq
->lock
);
5575 * Since we're inside stop-machine, _nothing_ should have
5576 * changed the task, WARN if weird stuff happened, because in
5577 * that case the above rq->lock drop is a fail too.
5579 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5580 raw_spin_unlock(&next
->pi_lock
);
5584 /* Find suitable destination for @next, with force if needed. */
5585 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5587 rq
= __migrate_task(rq
, next
, dest_cpu
);
5588 if (rq
!= dead_rq
) {
5589 raw_spin_unlock(&rq
->lock
);
5591 raw_spin_lock(&rq
->lock
);
5593 raw_spin_unlock(&next
->pi_lock
);
5598 #endif /* CONFIG_HOTPLUG_CPU */
5600 static void set_rq_online(struct rq
*rq
)
5603 const struct sched_class
*class;
5605 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5608 for_each_class(class) {
5609 if (class->rq_online
)
5610 class->rq_online(rq
);
5615 static void set_rq_offline(struct rq
*rq
)
5618 const struct sched_class
*class;
5620 for_each_class(class) {
5621 if (class->rq_offline
)
5622 class->rq_offline(rq
);
5625 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5630 static void set_cpu_rq_start_time(unsigned int cpu
)
5632 struct rq
*rq
= cpu_rq(cpu
);
5634 rq
->age_stamp
= sched_clock_cpu(cpu
);
5637 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5639 #ifdef CONFIG_SCHED_DEBUG
5641 static __read_mostly
int sched_debug_enabled
;
5643 static int __init
sched_debug_setup(char *str
)
5645 sched_debug_enabled
= 1;
5649 early_param("sched_debug", sched_debug_setup
);
5651 static inline bool sched_debug(void)
5653 return sched_debug_enabled
;
5656 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5657 struct cpumask
*groupmask
)
5659 struct sched_group
*group
= sd
->groups
;
5661 cpumask_clear(groupmask
);
5663 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5665 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5666 printk("does not load-balance\n");
5668 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5673 printk(KERN_CONT
"span %*pbl level %s\n",
5674 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5676 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5677 printk(KERN_ERR
"ERROR: domain->span does not contain "
5680 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5681 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5685 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5689 printk(KERN_ERR
"ERROR: group is NULL\n");
5693 if (!cpumask_weight(sched_group_cpus(group
))) {
5694 printk(KERN_CONT
"\n");
5695 printk(KERN_ERR
"ERROR: empty group\n");
5699 if (!(sd
->flags
& SD_OVERLAP
) &&
5700 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5701 printk(KERN_CONT
"\n");
5702 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5706 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5708 printk(KERN_CONT
" %*pbl",
5709 cpumask_pr_args(sched_group_cpus(group
)));
5710 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5711 printk(KERN_CONT
" (cpu_capacity = %lu)",
5712 group
->sgc
->capacity
);
5715 group
= group
->next
;
5716 } while (group
!= sd
->groups
);
5717 printk(KERN_CONT
"\n");
5719 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5720 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5723 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5724 printk(KERN_ERR
"ERROR: parent span is not a superset "
5725 "of domain->span\n");
5729 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5733 if (!sched_debug_enabled
)
5737 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5741 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5744 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5752 #else /* !CONFIG_SCHED_DEBUG */
5754 # define sched_debug_enabled 0
5755 # define sched_domain_debug(sd, cpu) do { } while (0)
5756 static inline bool sched_debug(void)
5760 #endif /* CONFIG_SCHED_DEBUG */
5762 static int sd_degenerate(struct sched_domain
*sd
)
5764 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5767 /* Following flags need at least 2 groups */
5768 if (sd
->flags
& (SD_LOAD_BALANCE
|
5769 SD_BALANCE_NEWIDLE
|
5772 SD_SHARE_CPUCAPACITY
|
5773 SD_ASYM_CPUCAPACITY
|
5774 SD_SHARE_PKG_RESOURCES
|
5775 SD_SHARE_POWERDOMAIN
)) {
5776 if (sd
->groups
!= sd
->groups
->next
)
5780 /* Following flags don't use groups */
5781 if (sd
->flags
& (SD_WAKE_AFFINE
))
5788 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5790 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5792 if (sd_degenerate(parent
))
5795 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5798 /* Flags needing groups don't count if only 1 group in parent */
5799 if (parent
->groups
== parent
->groups
->next
) {
5800 pflags
&= ~(SD_LOAD_BALANCE
|
5801 SD_BALANCE_NEWIDLE
|
5804 SD_ASYM_CPUCAPACITY
|
5805 SD_SHARE_CPUCAPACITY
|
5806 SD_SHARE_PKG_RESOURCES
|
5808 SD_SHARE_POWERDOMAIN
);
5809 if (nr_node_ids
== 1)
5810 pflags
&= ~SD_SERIALIZE
;
5812 if (~cflags
& pflags
)
5818 static void free_rootdomain(struct rcu_head
*rcu
)
5820 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5822 cpupri_cleanup(&rd
->cpupri
);
5823 cpudl_cleanup(&rd
->cpudl
);
5824 free_cpumask_var(rd
->dlo_mask
);
5825 free_cpumask_var(rd
->rto_mask
);
5826 free_cpumask_var(rd
->online
);
5827 free_cpumask_var(rd
->span
);
5831 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5833 struct root_domain
*old_rd
= NULL
;
5834 unsigned long flags
;
5836 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5841 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5844 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5847 * If we dont want to free the old_rd yet then
5848 * set old_rd to NULL to skip the freeing later
5851 if (!atomic_dec_and_test(&old_rd
->refcount
))
5855 atomic_inc(&rd
->refcount
);
5858 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5859 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5862 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5865 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5868 static int init_rootdomain(struct root_domain
*rd
)
5870 memset(rd
, 0, sizeof(*rd
));
5872 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5874 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5876 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5878 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5881 init_dl_bw(&rd
->dl_bw
);
5882 if (cpudl_init(&rd
->cpudl
) != 0)
5885 if (cpupri_init(&rd
->cpupri
) != 0)
5890 free_cpumask_var(rd
->rto_mask
);
5892 free_cpumask_var(rd
->dlo_mask
);
5894 free_cpumask_var(rd
->online
);
5896 free_cpumask_var(rd
->span
);
5902 * By default the system creates a single root-domain with all cpus as
5903 * members (mimicking the global state we have today).
5905 struct root_domain def_root_domain
;
5907 static void init_defrootdomain(void)
5909 init_rootdomain(&def_root_domain
);
5911 atomic_set(&def_root_domain
.refcount
, 1);
5914 static struct root_domain
*alloc_rootdomain(void)
5916 struct root_domain
*rd
;
5918 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5922 if (init_rootdomain(rd
) != 0) {
5930 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5932 struct sched_group
*tmp
, *first
;
5941 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5946 } while (sg
!= first
);
5949 static void destroy_sched_domain(struct sched_domain
*sd
)
5952 * If its an overlapping domain it has private groups, iterate and
5955 if (sd
->flags
& SD_OVERLAP
) {
5956 free_sched_groups(sd
->groups
, 1);
5957 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5958 kfree(sd
->groups
->sgc
);
5961 if (sd
->shared
&& atomic_dec_and_test(&sd
->shared
->ref
))
5966 static void destroy_sched_domains_rcu(struct rcu_head
*rcu
)
5968 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5971 struct sched_domain
*parent
= sd
->parent
;
5972 destroy_sched_domain(sd
);
5977 static void destroy_sched_domains(struct sched_domain
*sd
)
5980 call_rcu(&sd
->rcu
, destroy_sched_domains_rcu
);
5984 * Keep a special pointer to the highest sched_domain that has
5985 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5986 * allows us to avoid some pointer chasing select_idle_sibling().
5988 * Also keep a unique ID per domain (we use the first cpu number in
5989 * the cpumask of the domain), this allows us to quickly tell if
5990 * two cpus are in the same cache domain, see cpus_share_cache().
5992 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5993 DEFINE_PER_CPU(int, sd_llc_size
);
5994 DEFINE_PER_CPU(int, sd_llc_id
);
5995 DEFINE_PER_CPU(struct sched_domain_shared
*, sd_llc_shared
);
5996 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5997 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5999 static void update_top_cache_domain(int cpu
)
6001 struct sched_domain_shared
*sds
= NULL
;
6002 struct sched_domain
*sd
;
6006 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
6008 id
= cpumask_first(sched_domain_span(sd
));
6009 size
= cpumask_weight(sched_domain_span(sd
));
6013 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
6014 per_cpu(sd_llc_size
, cpu
) = size
;
6015 per_cpu(sd_llc_id
, cpu
) = id
;
6016 rcu_assign_pointer(per_cpu(sd_llc_shared
, cpu
), sds
);
6018 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
6019 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
6021 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
6022 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
6026 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6027 * hold the hotplug lock.
6030 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6032 struct rq
*rq
= cpu_rq(cpu
);
6033 struct sched_domain
*tmp
;
6035 /* Remove the sched domains which do not contribute to scheduling. */
6036 for (tmp
= sd
; tmp
; ) {
6037 struct sched_domain
*parent
= tmp
->parent
;
6041 if (sd_parent_degenerate(tmp
, parent
)) {
6042 tmp
->parent
= parent
->parent
;
6044 parent
->parent
->child
= tmp
;
6046 * Transfer SD_PREFER_SIBLING down in case of a
6047 * degenerate parent; the spans match for this
6048 * so the property transfers.
6050 if (parent
->flags
& SD_PREFER_SIBLING
)
6051 tmp
->flags
|= SD_PREFER_SIBLING
;
6052 destroy_sched_domain(parent
);
6057 if (sd
&& sd_degenerate(sd
)) {
6060 destroy_sched_domain(tmp
);
6065 sched_domain_debug(sd
, cpu
);
6067 rq_attach_root(rq
, rd
);
6069 rcu_assign_pointer(rq
->sd
, sd
);
6070 destroy_sched_domains(tmp
);
6072 update_top_cache_domain(cpu
);
6075 /* Setup the mask of cpus configured for isolated domains */
6076 static int __init
isolated_cpu_setup(char *str
)
6080 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6081 ret
= cpulist_parse(str
, cpu_isolated_map
);
6083 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
6088 __setup("isolcpus=", isolated_cpu_setup
);
6091 struct sched_domain
** __percpu sd
;
6092 struct root_domain
*rd
;
6103 * Build an iteration mask that can exclude certain CPUs from the upwards
6106 * Asymmetric node setups can result in situations where the domain tree is of
6107 * unequal depth, make sure to skip domains that already cover the entire
6110 * In that case build_sched_domains() will have terminated the iteration early
6111 * and our sibling sd spans will be empty. Domains should always include the
6112 * cpu they're built on, so check that.
6115 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6117 const struct cpumask
*span
= sched_domain_span(sd
);
6118 struct sd_data
*sdd
= sd
->private;
6119 struct sched_domain
*sibling
;
6122 for_each_cpu(i
, span
) {
6123 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6124 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6127 cpumask_set_cpu(i
, sched_group_mask(sg
));
6132 * Return the canonical balance cpu for this group, this is the first cpu
6133 * of this group that's also in the iteration mask.
6135 int group_balance_cpu(struct sched_group
*sg
)
6137 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6141 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6143 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6144 const struct cpumask
*span
= sched_domain_span(sd
);
6145 struct cpumask
*covered
= sched_domains_tmpmask
;
6146 struct sd_data
*sdd
= sd
->private;
6147 struct sched_domain
*sibling
;
6150 cpumask_clear(covered
);
6152 for_each_cpu(i
, span
) {
6153 struct cpumask
*sg_span
;
6155 if (cpumask_test_cpu(i
, covered
))
6158 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6160 /* See the comment near build_group_mask(). */
6161 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6164 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6165 GFP_KERNEL
, cpu_to_node(cpu
));
6170 sg_span
= sched_group_cpus(sg
);
6172 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6174 cpumask_set_cpu(i
, sg_span
);
6176 cpumask_or(covered
, covered
, sg_span
);
6178 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6179 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6180 build_group_mask(sd
, sg
);
6183 * Initialize sgc->capacity such that even if we mess up the
6184 * domains and no possible iteration will get us here, we won't
6187 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6188 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
6191 * Make sure the first group of this domain contains the
6192 * canonical balance cpu. Otherwise the sched_domain iteration
6193 * breaks. See update_sg_lb_stats().
6195 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6196 group_balance_cpu(sg
) == cpu
)
6206 sd
->groups
= groups
;
6211 free_sched_groups(first
, 0);
6216 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6218 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6219 struct sched_domain
*child
= sd
->child
;
6222 cpu
= cpumask_first(sched_domain_span(child
));
6225 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6226 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6227 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6234 * build_sched_groups will build a circular linked list of the groups
6235 * covered by the given span, and will set each group's ->cpumask correctly,
6236 * and ->cpu_capacity to 0.
6238 * Assumes the sched_domain tree is fully constructed
6241 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6243 struct sched_group
*first
= NULL
, *last
= NULL
;
6244 struct sd_data
*sdd
= sd
->private;
6245 const struct cpumask
*span
= sched_domain_span(sd
);
6246 struct cpumask
*covered
;
6249 get_group(cpu
, sdd
, &sd
->groups
);
6250 atomic_inc(&sd
->groups
->ref
);
6252 if (cpu
!= cpumask_first(span
))
6255 lockdep_assert_held(&sched_domains_mutex
);
6256 covered
= sched_domains_tmpmask
;
6258 cpumask_clear(covered
);
6260 for_each_cpu(i
, span
) {
6261 struct sched_group
*sg
;
6264 if (cpumask_test_cpu(i
, covered
))
6267 group
= get_group(i
, sdd
, &sg
);
6268 cpumask_setall(sched_group_mask(sg
));
6270 for_each_cpu(j
, span
) {
6271 if (get_group(j
, sdd
, NULL
) != group
)
6274 cpumask_set_cpu(j
, covered
);
6275 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6290 * Initialize sched groups cpu_capacity.
6292 * cpu_capacity indicates the capacity of sched group, which is used while
6293 * distributing the load between different sched groups in a sched domain.
6294 * Typically cpu_capacity for all the groups in a sched domain will be same
6295 * unless there are asymmetries in the topology. If there are asymmetries,
6296 * group having more cpu_capacity will pickup more load compared to the
6297 * group having less cpu_capacity.
6299 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6301 struct sched_group
*sg
= sd
->groups
;
6306 int cpu
, max_cpu
= -1;
6308 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6310 if (!(sd
->flags
& SD_ASYM_PACKING
))
6313 for_each_cpu(cpu
, sched_group_cpus(sg
)) {
6316 else if (sched_asym_prefer(cpu
, max_cpu
))
6319 sg
->asym_prefer_cpu
= max_cpu
;
6323 } while (sg
!= sd
->groups
);
6325 if (cpu
!= group_balance_cpu(sg
))
6328 update_group_capacity(sd
, cpu
);
6332 * Initializers for schedule domains
6333 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6336 static int default_relax_domain_level
= -1;
6337 int sched_domain_level_max
;
6339 static int __init
setup_relax_domain_level(char *str
)
6341 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6342 pr_warn("Unable to set relax_domain_level\n");
6346 __setup("relax_domain_level=", setup_relax_domain_level
);
6348 static void set_domain_attribute(struct sched_domain
*sd
,
6349 struct sched_domain_attr
*attr
)
6353 if (!attr
|| attr
->relax_domain_level
< 0) {
6354 if (default_relax_domain_level
< 0)
6357 request
= default_relax_domain_level
;
6359 request
= attr
->relax_domain_level
;
6360 if (request
< sd
->level
) {
6361 /* turn off idle balance on this domain */
6362 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6364 /* turn on idle balance on this domain */
6365 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6369 static void __sdt_free(const struct cpumask
*cpu_map
);
6370 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6372 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6373 const struct cpumask
*cpu_map
)
6377 if (!atomic_read(&d
->rd
->refcount
))
6378 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6380 free_percpu(d
->sd
); /* fall through */
6382 __sdt_free(cpu_map
); /* fall through */
6388 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6389 const struct cpumask
*cpu_map
)
6391 memset(d
, 0, sizeof(*d
));
6393 if (__sdt_alloc(cpu_map
))
6394 return sa_sd_storage
;
6395 d
->sd
= alloc_percpu(struct sched_domain
*);
6397 return sa_sd_storage
;
6398 d
->rd
= alloc_rootdomain();
6401 return sa_rootdomain
;
6405 * NULL the sd_data elements we've used to build the sched_domain and
6406 * sched_group structure so that the subsequent __free_domain_allocs()
6407 * will not free the data we're using.
6409 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6411 struct sd_data
*sdd
= sd
->private;
6413 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6414 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6416 if (atomic_read(&(*per_cpu_ptr(sdd
->sds
, cpu
))->ref
))
6417 *per_cpu_ptr(sdd
->sds
, cpu
) = NULL
;
6419 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6420 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6422 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6423 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6427 static int sched_domains_numa_levels
;
6428 enum numa_topology_type sched_numa_topology_type
;
6429 static int *sched_domains_numa_distance
;
6430 int sched_max_numa_distance
;
6431 static struct cpumask
***sched_domains_numa_masks
;
6432 static int sched_domains_curr_level
;
6436 * SD_flags allowed in topology descriptions.
6438 * These flags are purely descriptive of the topology and do not prescribe
6439 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6442 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6443 * SD_SHARE_PKG_RESOURCES - describes shared caches
6444 * SD_NUMA - describes NUMA topologies
6445 * SD_SHARE_POWERDOMAIN - describes shared power domain
6446 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6448 * Odd one out, which beside describing the topology has a quirk also
6449 * prescribes the desired behaviour that goes along with it:
6451 * SD_ASYM_PACKING - describes SMT quirks
6453 #define TOPOLOGY_SD_FLAGS \
6454 (SD_SHARE_CPUCAPACITY | \
6455 SD_SHARE_PKG_RESOURCES | \
6458 SD_ASYM_CPUCAPACITY | \
6459 SD_SHARE_POWERDOMAIN)
6461 static struct sched_domain
*
6462 sd_init(struct sched_domain_topology_level
*tl
,
6463 const struct cpumask
*cpu_map
,
6464 struct sched_domain
*child
, int cpu
)
6466 struct sd_data
*sdd
= &tl
->data
;
6467 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6468 int sd_id
, sd_weight
, sd_flags
= 0;
6472 * Ugly hack to pass state to sd_numa_mask()...
6474 sched_domains_curr_level
= tl
->numa_level
;
6477 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6480 sd_flags
= (*tl
->sd_flags
)();
6481 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6482 "wrong sd_flags in topology description\n"))
6483 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6485 *sd
= (struct sched_domain
){
6486 .min_interval
= sd_weight
,
6487 .max_interval
= 2*sd_weight
,
6489 .imbalance_pct
= 125,
6491 .cache_nice_tries
= 0,
6498 .flags
= 1*SD_LOAD_BALANCE
6499 | 1*SD_BALANCE_NEWIDLE
6504 | 0*SD_SHARE_CPUCAPACITY
6505 | 0*SD_SHARE_PKG_RESOURCES
6507 | 0*SD_PREFER_SIBLING
6512 .last_balance
= jiffies
,
6513 .balance_interval
= sd_weight
,
6515 .max_newidle_lb_cost
= 0,
6516 .next_decay_max_lb_cost
= jiffies
,
6518 #ifdef CONFIG_SCHED_DEBUG
6523 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6524 sd_id
= cpumask_first(sched_domain_span(sd
));
6527 * Convert topological properties into behaviour.
6530 if (sd
->flags
& SD_ASYM_CPUCAPACITY
) {
6531 struct sched_domain
*t
= sd
;
6533 for_each_lower_domain(t
)
6534 t
->flags
|= SD_BALANCE_WAKE
;
6537 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6538 sd
->flags
|= SD_PREFER_SIBLING
;
6539 sd
->imbalance_pct
= 110;
6540 sd
->smt_gain
= 1178; /* ~15% */
6542 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6543 sd
->imbalance_pct
= 117;
6544 sd
->cache_nice_tries
= 1;
6548 } else if (sd
->flags
& SD_NUMA
) {
6549 sd
->cache_nice_tries
= 2;
6553 sd
->flags
|= SD_SERIALIZE
;
6554 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6555 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6562 sd
->flags
|= SD_PREFER_SIBLING
;
6563 sd
->cache_nice_tries
= 1;
6569 * For all levels sharing cache; connect a sched_domain_shared
6572 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6573 sd
->shared
= *per_cpu_ptr(sdd
->sds
, sd_id
);
6574 atomic_inc(&sd
->shared
->ref
);
6575 atomic_set(&sd
->shared
->nr_busy_cpus
, sd_weight
);
6584 * Topology list, bottom-up.
6586 static struct sched_domain_topology_level default_topology
[] = {
6587 #ifdef CONFIG_SCHED_SMT
6588 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6590 #ifdef CONFIG_SCHED_MC
6591 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6593 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6597 static struct sched_domain_topology_level
*sched_domain_topology
=
6600 #define for_each_sd_topology(tl) \
6601 for (tl = sched_domain_topology; tl->mask; tl++)
6603 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6605 if (WARN_ON_ONCE(sched_smp_initialized
))
6608 sched_domain_topology
= tl
;
6613 static const struct cpumask
*sd_numa_mask(int cpu
)
6615 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6618 static void sched_numa_warn(const char *str
)
6620 static int done
= false;
6628 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6630 for (i
= 0; i
< nr_node_ids
; i
++) {
6631 printk(KERN_WARNING
" ");
6632 for (j
= 0; j
< nr_node_ids
; j
++)
6633 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6634 printk(KERN_CONT
"\n");
6636 printk(KERN_WARNING
"\n");
6639 bool find_numa_distance(int distance
)
6643 if (distance
== node_distance(0, 0))
6646 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6647 if (sched_domains_numa_distance
[i
] == distance
)
6655 * A system can have three types of NUMA topology:
6656 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6657 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6658 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6660 * The difference between a glueless mesh topology and a backplane
6661 * topology lies in whether communication between not directly
6662 * connected nodes goes through intermediary nodes (where programs
6663 * could run), or through backplane controllers. This affects
6664 * placement of programs.
6666 * The type of topology can be discerned with the following tests:
6667 * - If the maximum distance between any nodes is 1 hop, the system
6668 * is directly connected.
6669 * - If for two nodes A and B, located N > 1 hops away from each other,
6670 * there is an intermediary node C, which is < N hops away from both
6671 * nodes A and B, the system is a glueless mesh.
6673 static void init_numa_topology_type(void)
6677 n
= sched_max_numa_distance
;
6679 if (sched_domains_numa_levels
<= 1) {
6680 sched_numa_topology_type
= NUMA_DIRECT
;
6684 for_each_online_node(a
) {
6685 for_each_online_node(b
) {
6686 /* Find two nodes furthest removed from each other. */
6687 if (node_distance(a
, b
) < n
)
6690 /* Is there an intermediary node between a and b? */
6691 for_each_online_node(c
) {
6692 if (node_distance(a
, c
) < n
&&
6693 node_distance(b
, c
) < n
) {
6694 sched_numa_topology_type
=
6700 sched_numa_topology_type
= NUMA_BACKPLANE
;
6706 static void sched_init_numa(void)
6708 int next_distance
, curr_distance
= node_distance(0, 0);
6709 struct sched_domain_topology_level
*tl
;
6713 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6714 if (!sched_domains_numa_distance
)
6718 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6719 * unique distances in the node_distance() table.
6721 * Assumes node_distance(0,j) includes all distances in
6722 * node_distance(i,j) in order to avoid cubic time.
6724 next_distance
= curr_distance
;
6725 for (i
= 0; i
< nr_node_ids
; i
++) {
6726 for (j
= 0; j
< nr_node_ids
; j
++) {
6727 for (k
= 0; k
< nr_node_ids
; k
++) {
6728 int distance
= node_distance(i
, k
);
6730 if (distance
> curr_distance
&&
6731 (distance
< next_distance
||
6732 next_distance
== curr_distance
))
6733 next_distance
= distance
;
6736 * While not a strong assumption it would be nice to know
6737 * about cases where if node A is connected to B, B is not
6738 * equally connected to A.
6740 if (sched_debug() && node_distance(k
, i
) != distance
)
6741 sched_numa_warn("Node-distance not symmetric");
6743 if (sched_debug() && i
&& !find_numa_distance(distance
))
6744 sched_numa_warn("Node-0 not representative");
6746 if (next_distance
!= curr_distance
) {
6747 sched_domains_numa_distance
[level
++] = next_distance
;
6748 sched_domains_numa_levels
= level
;
6749 curr_distance
= next_distance
;
6754 * In case of sched_debug() we verify the above assumption.
6764 * 'level' contains the number of unique distances, excluding the
6765 * identity distance node_distance(i,i).
6767 * The sched_domains_numa_distance[] array includes the actual distance
6772 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6773 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6774 * the array will contain less then 'level' members. This could be
6775 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6776 * in other functions.
6778 * We reset it to 'level' at the end of this function.
6780 sched_domains_numa_levels
= 0;
6782 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6783 if (!sched_domains_numa_masks
)
6787 * Now for each level, construct a mask per node which contains all
6788 * cpus of nodes that are that many hops away from us.
6790 for (i
= 0; i
< level
; i
++) {
6791 sched_domains_numa_masks
[i
] =
6792 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6793 if (!sched_domains_numa_masks
[i
])
6796 for (j
= 0; j
< nr_node_ids
; j
++) {
6797 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6801 sched_domains_numa_masks
[i
][j
] = mask
;
6804 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6807 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6812 /* Compute default topology size */
6813 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6815 tl
= kzalloc((i
+ level
+ 1) *
6816 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6821 * Copy the default topology bits..
6823 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6824 tl
[i
] = sched_domain_topology
[i
];
6827 * .. and append 'j' levels of NUMA goodness.
6829 for (j
= 0; j
< level
; i
++, j
++) {
6830 tl
[i
] = (struct sched_domain_topology_level
){
6831 .mask
= sd_numa_mask
,
6832 .sd_flags
= cpu_numa_flags
,
6833 .flags
= SDTL_OVERLAP
,
6839 sched_domain_topology
= tl
;
6841 sched_domains_numa_levels
= level
;
6842 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6844 init_numa_topology_type();
6847 static void sched_domains_numa_masks_set(unsigned int cpu
)
6849 int node
= cpu_to_node(cpu
);
6852 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6853 for (j
= 0; j
< nr_node_ids
; j
++) {
6854 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6855 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6860 static void sched_domains_numa_masks_clear(unsigned int cpu
)
6864 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6865 for (j
= 0; j
< nr_node_ids
; j
++)
6866 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6871 static inline void sched_init_numa(void) { }
6872 static void sched_domains_numa_masks_set(unsigned int cpu
) { }
6873 static void sched_domains_numa_masks_clear(unsigned int cpu
) { }
6874 #endif /* CONFIG_NUMA */
6876 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6878 struct sched_domain_topology_level
*tl
;
6881 for_each_sd_topology(tl
) {
6882 struct sd_data
*sdd
= &tl
->data
;
6884 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6888 sdd
->sds
= alloc_percpu(struct sched_domain_shared
*);
6892 sdd
->sg
= alloc_percpu(struct sched_group
*);
6896 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6900 for_each_cpu(j
, cpu_map
) {
6901 struct sched_domain
*sd
;
6902 struct sched_domain_shared
*sds
;
6903 struct sched_group
*sg
;
6904 struct sched_group_capacity
*sgc
;
6906 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6907 GFP_KERNEL
, cpu_to_node(j
));
6911 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6913 sds
= kzalloc_node(sizeof(struct sched_domain_shared
),
6914 GFP_KERNEL
, cpu_to_node(j
));
6918 *per_cpu_ptr(sdd
->sds
, j
) = sds
;
6920 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6921 GFP_KERNEL
, cpu_to_node(j
));
6927 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6929 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6930 GFP_KERNEL
, cpu_to_node(j
));
6934 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6941 static void __sdt_free(const struct cpumask
*cpu_map
)
6943 struct sched_domain_topology_level
*tl
;
6946 for_each_sd_topology(tl
) {
6947 struct sd_data
*sdd
= &tl
->data
;
6949 for_each_cpu(j
, cpu_map
) {
6950 struct sched_domain
*sd
;
6953 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6954 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6955 free_sched_groups(sd
->groups
, 0);
6956 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6960 kfree(*per_cpu_ptr(sdd
->sds
, j
));
6962 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6964 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6966 free_percpu(sdd
->sd
);
6968 free_percpu(sdd
->sds
);
6970 free_percpu(sdd
->sg
);
6972 free_percpu(sdd
->sgc
);
6977 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6978 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6979 struct sched_domain
*child
, int cpu
)
6981 struct sched_domain
*sd
= sd_init(tl
, cpu_map
, child
, cpu
);
6984 sd
->level
= child
->level
+ 1;
6985 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6988 if (!cpumask_subset(sched_domain_span(child
),
6989 sched_domain_span(sd
))) {
6990 pr_err("BUG: arch topology borken\n");
6991 #ifdef CONFIG_SCHED_DEBUG
6992 pr_err(" the %s domain not a subset of the %s domain\n",
6993 child
->name
, sd
->name
);
6995 /* Fixup, ensure @sd has at least @child cpus. */
6996 cpumask_or(sched_domain_span(sd
),
6997 sched_domain_span(sd
),
6998 sched_domain_span(child
));
7002 set_domain_attribute(sd
, attr
);
7008 * Build sched domains for a given set of cpus and attach the sched domains
7009 * to the individual cpus
7011 static int build_sched_domains(const struct cpumask
*cpu_map
,
7012 struct sched_domain_attr
*attr
)
7014 enum s_alloc alloc_state
;
7015 struct sched_domain
*sd
;
7017 struct rq
*rq
= NULL
;
7018 int i
, ret
= -ENOMEM
;
7020 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7021 if (alloc_state
!= sa_rootdomain
)
7024 /* Set up domains for cpus specified by the cpu_map. */
7025 for_each_cpu(i
, cpu_map
) {
7026 struct sched_domain_topology_level
*tl
;
7029 for_each_sd_topology(tl
) {
7030 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
7031 if (tl
== sched_domain_topology
)
7032 *per_cpu_ptr(d
.sd
, i
) = sd
;
7033 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
7034 sd
->flags
|= SD_OVERLAP
;
7035 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
7040 /* Build the groups for the domains */
7041 for_each_cpu(i
, cpu_map
) {
7042 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7043 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
7044 if (sd
->flags
& SD_OVERLAP
) {
7045 if (build_overlap_sched_groups(sd
, i
))
7048 if (build_sched_groups(sd
, i
))
7054 /* Calculate CPU capacity for physical packages and nodes */
7055 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
7056 if (!cpumask_test_cpu(i
, cpu_map
))
7059 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7060 claim_allocations(i
, sd
);
7061 init_sched_groups_capacity(i
, sd
);
7065 /* Attach the domains */
7067 for_each_cpu(i
, cpu_map
) {
7069 sd
= *per_cpu_ptr(d
.sd
, i
);
7071 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7072 if (rq
->cpu_capacity_orig
> READ_ONCE(d
.rd
->max_cpu_capacity
))
7073 WRITE_ONCE(d
.rd
->max_cpu_capacity
, rq
->cpu_capacity_orig
);
7075 cpu_attach_domain(sd
, d
.rd
, i
);
7079 if (rq
&& sched_debug_enabled
) {
7080 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7081 cpumask_pr_args(cpu_map
), rq
->rd
->max_cpu_capacity
);
7086 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7090 static cpumask_var_t
*doms_cur
; /* current sched domains */
7091 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7092 static struct sched_domain_attr
*dattr_cur
;
7093 /* attribues of custom domains in 'doms_cur' */
7096 * Special case: If a kmalloc of a doms_cur partition (array of
7097 * cpumask) fails, then fallback to a single sched domain,
7098 * as determined by the single cpumask fallback_doms.
7100 static cpumask_var_t fallback_doms
;
7103 * arch_update_cpu_topology lets virtualized architectures update the
7104 * cpu core maps. It is supposed to return 1 if the topology changed
7105 * or 0 if it stayed the same.
7107 int __weak
arch_update_cpu_topology(void)
7112 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7115 cpumask_var_t
*doms
;
7117 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7120 for (i
= 0; i
< ndoms
; i
++) {
7121 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7122 free_sched_domains(doms
, i
);
7129 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7132 for (i
= 0; i
< ndoms
; i
++)
7133 free_cpumask_var(doms
[i
]);
7138 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7139 * For now this just excludes isolated cpus, but could be used to
7140 * exclude other special cases in the future.
7142 static int init_sched_domains(const struct cpumask
*cpu_map
)
7146 arch_update_cpu_topology();
7148 doms_cur
= alloc_sched_domains(ndoms_cur
);
7150 doms_cur
= &fallback_doms
;
7151 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7152 err
= build_sched_domains(doms_cur
[0], NULL
);
7153 register_sched_domain_sysctl();
7159 * Detach sched domains from a group of cpus specified in cpu_map
7160 * These cpus will now be attached to the NULL domain
7162 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7167 for_each_cpu(i
, cpu_map
)
7168 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7172 /* handle null as "default" */
7173 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7174 struct sched_domain_attr
*new, int idx_new
)
7176 struct sched_domain_attr tmp
;
7183 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7184 new ? (new + idx_new
) : &tmp
,
7185 sizeof(struct sched_domain_attr
));
7189 * Partition sched domains as specified by the 'ndoms_new'
7190 * cpumasks in the array doms_new[] of cpumasks. This compares
7191 * doms_new[] to the current sched domain partitioning, doms_cur[].
7192 * It destroys each deleted domain and builds each new domain.
7194 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7195 * The masks don't intersect (don't overlap.) We should setup one
7196 * sched domain for each mask. CPUs not in any of the cpumasks will
7197 * not be load balanced. If the same cpumask appears both in the
7198 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7201 * The passed in 'doms_new' should be allocated using
7202 * alloc_sched_domains. This routine takes ownership of it and will
7203 * free_sched_domains it when done with it. If the caller failed the
7204 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7205 * and partition_sched_domains() will fallback to the single partition
7206 * 'fallback_doms', it also forces the domains to be rebuilt.
7208 * If doms_new == NULL it will be replaced with cpu_online_mask.
7209 * ndoms_new == 0 is a special case for destroying existing domains,
7210 * and it will not create the default domain.
7212 * Call with hotplug lock held
7214 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7215 struct sched_domain_attr
*dattr_new
)
7220 mutex_lock(&sched_domains_mutex
);
7222 /* always unregister in case we don't destroy any domains */
7223 unregister_sched_domain_sysctl();
7225 /* Let architecture update cpu core mappings. */
7226 new_topology
= arch_update_cpu_topology();
7228 n
= doms_new
? ndoms_new
: 0;
7230 /* Destroy deleted domains */
7231 for (i
= 0; i
< ndoms_cur
; i
++) {
7232 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7233 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7234 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7237 /* no match - a current sched domain not in new doms_new[] */
7238 detach_destroy_domains(doms_cur
[i
]);
7244 if (doms_new
== NULL
) {
7246 doms_new
= &fallback_doms
;
7247 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7248 WARN_ON_ONCE(dattr_new
);
7251 /* Build new domains */
7252 for (i
= 0; i
< ndoms_new
; i
++) {
7253 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7254 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7255 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7258 /* no match - add a new doms_new */
7259 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7264 /* Remember the new sched domains */
7265 if (doms_cur
!= &fallback_doms
)
7266 free_sched_domains(doms_cur
, ndoms_cur
);
7267 kfree(dattr_cur
); /* kfree(NULL) is safe */
7268 doms_cur
= doms_new
;
7269 dattr_cur
= dattr_new
;
7270 ndoms_cur
= ndoms_new
;
7272 register_sched_domain_sysctl();
7274 mutex_unlock(&sched_domains_mutex
);
7277 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7280 * Update cpusets according to cpu_active mask. If cpusets are
7281 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7282 * around partition_sched_domains().
7284 * If we come here as part of a suspend/resume, don't touch cpusets because we
7285 * want to restore it back to its original state upon resume anyway.
7287 static void cpuset_cpu_active(void)
7289 if (cpuhp_tasks_frozen
) {
7291 * num_cpus_frozen tracks how many CPUs are involved in suspend
7292 * resume sequence. As long as this is not the last online
7293 * operation in the resume sequence, just build a single sched
7294 * domain, ignoring cpusets.
7297 if (likely(num_cpus_frozen
)) {
7298 partition_sched_domains(1, NULL
, NULL
);
7302 * This is the last CPU online operation. So fall through and
7303 * restore the original sched domains by considering the
7304 * cpuset configurations.
7307 cpuset_update_active_cpus(true);
7310 static int cpuset_cpu_inactive(unsigned int cpu
)
7312 unsigned long flags
;
7317 if (!cpuhp_tasks_frozen
) {
7318 rcu_read_lock_sched();
7319 dl_b
= dl_bw_of(cpu
);
7321 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7322 cpus
= dl_bw_cpus(cpu
);
7323 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7324 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7326 rcu_read_unlock_sched();
7330 cpuset_update_active_cpus(false);
7333 partition_sched_domains(1, NULL
, NULL
);
7338 int sched_cpu_activate(unsigned int cpu
)
7340 struct rq
*rq
= cpu_rq(cpu
);
7341 unsigned long flags
;
7343 set_cpu_active(cpu
, true);
7345 if (sched_smp_initialized
) {
7346 sched_domains_numa_masks_set(cpu
);
7347 cpuset_cpu_active();
7351 * Put the rq online, if not already. This happens:
7353 * 1) In the early boot process, because we build the real domains
7354 * after all cpus have been brought up.
7356 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7359 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7361 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7364 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7366 update_max_interval();
7371 int sched_cpu_deactivate(unsigned int cpu
)
7375 set_cpu_active(cpu
, false);
7377 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7378 * users of this state to go away such that all new such users will
7381 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7382 * not imply sync_sched(), so wait for both.
7384 * Do sync before park smpboot threads to take care the rcu boost case.
7386 if (IS_ENABLED(CONFIG_PREEMPT
))
7387 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
7391 if (!sched_smp_initialized
)
7394 ret
= cpuset_cpu_inactive(cpu
);
7396 set_cpu_active(cpu
, true);
7399 sched_domains_numa_masks_clear(cpu
);
7403 static void sched_rq_cpu_starting(unsigned int cpu
)
7405 struct rq
*rq
= cpu_rq(cpu
);
7407 rq
->calc_load_update
= calc_load_update
;
7408 update_max_interval();
7411 int sched_cpu_starting(unsigned int cpu
)
7413 set_cpu_rq_start_time(cpu
);
7414 sched_rq_cpu_starting(cpu
);
7418 #ifdef CONFIG_HOTPLUG_CPU
7419 int sched_cpu_dying(unsigned int cpu
)
7421 struct rq
*rq
= cpu_rq(cpu
);
7422 unsigned long flags
;
7424 /* Handle pending wakeups and then migrate everything off */
7425 sched_ttwu_pending();
7426 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7428 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7432 BUG_ON(rq
->nr_running
!= 1);
7433 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7434 calc_load_migrate(rq
);
7435 update_max_interval();
7436 nohz_balance_exit_idle(cpu
);
7442 #ifdef CONFIG_SCHED_SMT
7443 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
7445 static void sched_init_smt(void)
7448 * We've enumerated all CPUs and will assume that if any CPU
7449 * has SMT siblings, CPU0 will too.
7451 if (cpumask_weight(cpu_smt_mask(0)) > 1)
7452 static_branch_enable(&sched_smt_present
);
7455 static inline void sched_init_smt(void) { }
7458 void __init
sched_init_smp(void)
7460 cpumask_var_t non_isolated_cpus
;
7462 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7463 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7468 * There's no userspace yet to cause hotplug operations; hence all the
7469 * cpu masks are stable and all blatant races in the below code cannot
7472 mutex_lock(&sched_domains_mutex
);
7473 init_sched_domains(cpu_active_mask
);
7474 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7475 if (cpumask_empty(non_isolated_cpus
))
7476 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7477 mutex_unlock(&sched_domains_mutex
);
7479 /* Move init over to a non-isolated CPU */
7480 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7482 sched_init_granularity();
7483 free_cpumask_var(non_isolated_cpus
);
7485 init_sched_rt_class();
7486 init_sched_dl_class();
7490 sched_smp_initialized
= true;
7493 static int __init
migration_init(void)
7495 sched_rq_cpu_starting(smp_processor_id());
7498 early_initcall(migration_init
);
7501 void __init
sched_init_smp(void)
7503 sched_init_granularity();
7505 #endif /* CONFIG_SMP */
7507 int in_sched_functions(unsigned long addr
)
7509 return in_lock_functions(addr
) ||
7510 (addr
>= (unsigned long)__sched_text_start
7511 && addr
< (unsigned long)__sched_text_end
);
7514 #ifdef CONFIG_CGROUP_SCHED
7516 * Default task group.
7517 * Every task in system belongs to this group at bootup.
7519 struct task_group root_task_group
;
7520 LIST_HEAD(task_groups
);
7522 /* Cacheline aligned slab cache for task_group */
7523 static struct kmem_cache
*task_group_cache __read_mostly
;
7526 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7527 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
7529 #define WAIT_TABLE_BITS 8
7530 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7531 static wait_queue_head_t bit_wait_table
[WAIT_TABLE_SIZE
] __cacheline_aligned
;
7533 wait_queue_head_t
*bit_waitqueue(void *word
, int bit
)
7535 const int shift
= BITS_PER_LONG
== 32 ? 5 : 6;
7536 unsigned long val
= (unsigned long)word
<< shift
| bit
;
7538 return bit_wait_table
+ hash_long(val
, WAIT_TABLE_BITS
);
7540 EXPORT_SYMBOL(bit_waitqueue
);
7542 void __init
sched_init(void)
7545 unsigned long alloc_size
= 0, ptr
;
7547 for (i
= 0; i
< WAIT_TABLE_SIZE
; i
++)
7548 init_waitqueue_head(bit_wait_table
+ i
);
7550 #ifdef CONFIG_FAIR_GROUP_SCHED
7551 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7553 #ifdef CONFIG_RT_GROUP_SCHED
7554 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7557 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7559 #ifdef CONFIG_FAIR_GROUP_SCHED
7560 root_task_group
.se
= (struct sched_entity
**)ptr
;
7561 ptr
+= nr_cpu_ids
* sizeof(void **);
7563 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7564 ptr
+= nr_cpu_ids
* sizeof(void **);
7566 #endif /* CONFIG_FAIR_GROUP_SCHED */
7567 #ifdef CONFIG_RT_GROUP_SCHED
7568 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7569 ptr
+= nr_cpu_ids
* sizeof(void **);
7571 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7572 ptr
+= nr_cpu_ids
* sizeof(void **);
7574 #endif /* CONFIG_RT_GROUP_SCHED */
7576 #ifdef CONFIG_CPUMASK_OFFSTACK
7577 for_each_possible_cpu(i
) {
7578 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7579 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7580 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7581 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7583 #endif /* CONFIG_CPUMASK_OFFSTACK */
7585 init_rt_bandwidth(&def_rt_bandwidth
,
7586 global_rt_period(), global_rt_runtime());
7587 init_dl_bandwidth(&def_dl_bandwidth
,
7588 global_rt_period(), global_rt_runtime());
7591 init_defrootdomain();
7594 #ifdef CONFIG_RT_GROUP_SCHED
7595 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7596 global_rt_period(), global_rt_runtime());
7597 #endif /* CONFIG_RT_GROUP_SCHED */
7599 #ifdef CONFIG_CGROUP_SCHED
7600 task_group_cache
= KMEM_CACHE(task_group
, 0);
7602 list_add(&root_task_group
.list
, &task_groups
);
7603 INIT_LIST_HEAD(&root_task_group
.children
);
7604 INIT_LIST_HEAD(&root_task_group
.siblings
);
7605 autogroup_init(&init_task
);
7606 #endif /* CONFIG_CGROUP_SCHED */
7608 for_each_possible_cpu(i
) {
7612 raw_spin_lock_init(&rq
->lock
);
7614 rq
->calc_load_active
= 0;
7615 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7616 init_cfs_rq(&rq
->cfs
);
7617 init_rt_rq(&rq
->rt
);
7618 init_dl_rq(&rq
->dl
);
7619 #ifdef CONFIG_FAIR_GROUP_SCHED
7620 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7621 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7622 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
7624 * How much cpu bandwidth does root_task_group get?
7626 * In case of task-groups formed thr' the cgroup filesystem, it
7627 * gets 100% of the cpu resources in the system. This overall
7628 * system cpu resource is divided among the tasks of
7629 * root_task_group and its child task-groups in a fair manner,
7630 * based on each entity's (task or task-group's) weight
7631 * (se->load.weight).
7633 * In other words, if root_task_group has 10 tasks of weight
7634 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7635 * then A0's share of the cpu resource is:
7637 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7639 * We achieve this by letting root_task_group's tasks sit
7640 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7642 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7643 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7644 #endif /* CONFIG_FAIR_GROUP_SCHED */
7646 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7647 #ifdef CONFIG_RT_GROUP_SCHED
7648 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7651 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7652 rq
->cpu_load
[j
] = 0;
7657 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7658 rq
->balance_callback
= NULL
;
7659 rq
->active_balance
= 0;
7660 rq
->next_balance
= jiffies
;
7665 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7666 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7668 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7670 rq_attach_root(rq
, &def_root_domain
);
7671 #ifdef CONFIG_NO_HZ_COMMON
7672 rq
->last_load_update_tick
= jiffies
;
7675 #ifdef CONFIG_NO_HZ_FULL
7676 rq
->last_sched_tick
= 0;
7678 #endif /* CONFIG_SMP */
7680 atomic_set(&rq
->nr_iowait
, 0);
7683 set_load_weight(&init_task
);
7686 * The boot idle thread does lazy MMU switching as well:
7688 atomic_inc(&init_mm
.mm_count
);
7689 enter_lazy_tlb(&init_mm
, current
);
7692 * Make us the idle thread. Technically, schedule() should not be
7693 * called from this thread, however somewhere below it might be,
7694 * but because we are the idle thread, we just pick up running again
7695 * when this runqueue becomes "idle".
7697 init_idle(current
, smp_processor_id());
7699 calc_load_update
= jiffies
+ LOAD_FREQ
;
7702 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7703 /* May be allocated at isolcpus cmdline parse time */
7704 if (cpu_isolated_map
== NULL
)
7705 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7706 idle_thread_set_boot_cpu();
7707 set_cpu_rq_start_time(smp_processor_id());
7709 init_sched_fair_class();
7713 scheduler_running
= 1;
7716 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7717 static inline int preempt_count_equals(int preempt_offset
)
7719 int nested
= preempt_count() + rcu_preempt_depth();
7721 return (nested
== preempt_offset
);
7724 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7727 * Blocking primitives will set (and therefore destroy) current->state,
7728 * since we will exit with TASK_RUNNING make sure we enter with it,
7729 * otherwise we will destroy state.
7731 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7732 "do not call blocking ops when !TASK_RUNNING; "
7733 "state=%lx set at [<%p>] %pS\n",
7735 (void *)current
->task_state_change
,
7736 (void *)current
->task_state_change
);
7738 ___might_sleep(file
, line
, preempt_offset
);
7740 EXPORT_SYMBOL(__might_sleep
);
7742 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7744 static unsigned long prev_jiffy
; /* ratelimiting */
7745 unsigned long preempt_disable_ip
;
7747 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7748 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7749 !is_idle_task(current
)) ||
7750 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7752 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7754 prev_jiffy
= jiffies
;
7756 /* Save this before calling printk(), since that will clobber it */
7757 preempt_disable_ip
= get_preempt_disable_ip(current
);
7760 "BUG: sleeping function called from invalid context at %s:%d\n",
7763 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7764 in_atomic(), irqs_disabled(),
7765 current
->pid
, current
->comm
);
7767 if (task_stack_end_corrupted(current
))
7768 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7770 debug_show_held_locks(current
);
7771 if (irqs_disabled())
7772 print_irqtrace_events(current
);
7773 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
7774 && !preempt_count_equals(preempt_offset
)) {
7775 pr_err("Preemption disabled at:");
7776 print_ip_sym(preempt_disable_ip
);
7780 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7782 EXPORT_SYMBOL(___might_sleep
);
7785 #ifdef CONFIG_MAGIC_SYSRQ
7786 void normalize_rt_tasks(void)
7788 struct task_struct
*g
, *p
;
7789 struct sched_attr attr
= {
7790 .sched_policy
= SCHED_NORMAL
,
7793 read_lock(&tasklist_lock
);
7794 for_each_process_thread(g
, p
) {
7796 * Only normalize user tasks:
7798 if (p
->flags
& PF_KTHREAD
)
7801 p
->se
.exec_start
= 0;
7802 schedstat_set(p
->se
.statistics
.wait_start
, 0);
7803 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
7804 schedstat_set(p
->se
.statistics
.block_start
, 0);
7806 if (!dl_task(p
) && !rt_task(p
)) {
7808 * Renice negative nice level userspace
7811 if (task_nice(p
) < 0)
7812 set_user_nice(p
, 0);
7816 __sched_setscheduler(p
, &attr
, false, false);
7818 read_unlock(&tasklist_lock
);
7821 #endif /* CONFIG_MAGIC_SYSRQ */
7823 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7825 * These functions are only useful for the IA64 MCA handling, or kdb.
7827 * They can only be called when the whole system has been
7828 * stopped - every CPU needs to be quiescent, and no scheduling
7829 * activity can take place. Using them for anything else would
7830 * be a serious bug, and as a result, they aren't even visible
7831 * under any other configuration.
7835 * curr_task - return the current task for a given cpu.
7836 * @cpu: the processor in question.
7838 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7840 * Return: The current task for @cpu.
7842 struct task_struct
*curr_task(int cpu
)
7844 return cpu_curr(cpu
);
7847 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7851 * set_curr_task - set the current task for a given cpu.
7852 * @cpu: the processor in question.
7853 * @p: the task pointer to set.
7855 * Description: This function must only be used when non-maskable interrupts
7856 * are serviced on a separate stack. It allows the architecture to switch the
7857 * notion of the current task on a cpu in a non-blocking manner. This function
7858 * must be called with all CPU's synchronized, and interrupts disabled, the
7859 * and caller must save the original value of the current task (see
7860 * curr_task() above) and restore that value before reenabling interrupts and
7861 * re-starting the system.
7863 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7865 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
7872 #ifdef CONFIG_CGROUP_SCHED
7873 /* task_group_lock serializes the addition/removal of task groups */
7874 static DEFINE_SPINLOCK(task_group_lock
);
7876 static void sched_free_group(struct task_group
*tg
)
7878 free_fair_sched_group(tg
);
7879 free_rt_sched_group(tg
);
7881 kmem_cache_free(task_group_cache
, tg
);
7884 /* allocate runqueue etc for a new task group */
7885 struct task_group
*sched_create_group(struct task_group
*parent
)
7887 struct task_group
*tg
;
7889 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7891 return ERR_PTR(-ENOMEM
);
7893 if (!alloc_fair_sched_group(tg
, parent
))
7896 if (!alloc_rt_sched_group(tg
, parent
))
7902 sched_free_group(tg
);
7903 return ERR_PTR(-ENOMEM
);
7906 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7908 unsigned long flags
;
7910 spin_lock_irqsave(&task_group_lock
, flags
);
7911 list_add_rcu(&tg
->list
, &task_groups
);
7913 WARN_ON(!parent
); /* root should already exist */
7915 tg
->parent
= parent
;
7916 INIT_LIST_HEAD(&tg
->children
);
7917 list_add_rcu(&tg
->siblings
, &parent
->children
);
7918 spin_unlock_irqrestore(&task_group_lock
, flags
);
7920 online_fair_sched_group(tg
);
7923 /* rcu callback to free various structures associated with a task group */
7924 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7926 /* now it should be safe to free those cfs_rqs */
7927 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7930 void sched_destroy_group(struct task_group
*tg
)
7932 /* wait for possible concurrent references to cfs_rqs complete */
7933 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7936 void sched_offline_group(struct task_group
*tg
)
7938 unsigned long flags
;
7940 /* end participation in shares distribution */
7941 unregister_fair_sched_group(tg
);
7943 spin_lock_irqsave(&task_group_lock
, flags
);
7944 list_del_rcu(&tg
->list
);
7945 list_del_rcu(&tg
->siblings
);
7946 spin_unlock_irqrestore(&task_group_lock
, flags
);
7949 static void sched_change_group(struct task_struct
*tsk
, int type
)
7951 struct task_group
*tg
;
7954 * All callers are synchronized by task_rq_lock(); we do not use RCU
7955 * which is pointless here. Thus, we pass "true" to task_css_check()
7956 * to prevent lockdep warnings.
7958 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7959 struct task_group
, css
);
7960 tg
= autogroup_task_group(tsk
, tg
);
7961 tsk
->sched_task_group
= tg
;
7963 #ifdef CONFIG_FAIR_GROUP_SCHED
7964 if (tsk
->sched_class
->task_change_group
)
7965 tsk
->sched_class
->task_change_group(tsk
, type
);
7968 set_task_rq(tsk
, task_cpu(tsk
));
7972 * Change task's runqueue when it moves between groups.
7974 * The caller of this function should have put the task in its new group by
7975 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7978 void sched_move_task(struct task_struct
*tsk
)
7980 int queued
, running
;
7984 rq
= task_rq_lock(tsk
, &rf
);
7986 running
= task_current(rq
, tsk
);
7987 queued
= task_on_rq_queued(tsk
);
7990 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7991 if (unlikely(running
))
7992 put_prev_task(rq
, tsk
);
7994 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7997 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7998 if (unlikely(running
))
7999 set_curr_task(rq
, tsk
);
8001 task_rq_unlock(rq
, tsk
, &rf
);
8003 #endif /* CONFIG_CGROUP_SCHED */
8005 #ifdef CONFIG_RT_GROUP_SCHED
8007 * Ensure that the real time constraints are schedulable.
8009 static DEFINE_MUTEX(rt_constraints_mutex
);
8011 /* Must be called with tasklist_lock held */
8012 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8014 struct task_struct
*g
, *p
;
8017 * Autogroups do not have RT tasks; see autogroup_create().
8019 if (task_group_is_autogroup(tg
))
8022 for_each_process_thread(g
, p
) {
8023 if (rt_task(p
) && task_group(p
) == tg
)
8030 struct rt_schedulable_data
{
8031 struct task_group
*tg
;
8036 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
8038 struct rt_schedulable_data
*d
= data
;
8039 struct task_group
*child
;
8040 unsigned long total
, sum
= 0;
8041 u64 period
, runtime
;
8043 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8044 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8047 period
= d
->rt_period
;
8048 runtime
= d
->rt_runtime
;
8052 * Cannot have more runtime than the period.
8054 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8058 * Ensure we don't starve existing RT tasks.
8060 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8063 total
= to_ratio(period
, runtime
);
8066 * Nobody can have more than the global setting allows.
8068 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8072 * The sum of our children's runtime should not exceed our own.
8074 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8075 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8076 runtime
= child
->rt_bandwidth
.rt_runtime
;
8078 if (child
== d
->tg
) {
8079 period
= d
->rt_period
;
8080 runtime
= d
->rt_runtime
;
8083 sum
+= to_ratio(period
, runtime
);
8092 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8096 struct rt_schedulable_data data
= {
8098 .rt_period
= period
,
8099 .rt_runtime
= runtime
,
8103 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
8109 static int tg_set_rt_bandwidth(struct task_group
*tg
,
8110 u64 rt_period
, u64 rt_runtime
)
8115 * Disallowing the root group RT runtime is BAD, it would disallow the
8116 * kernel creating (and or operating) RT threads.
8118 if (tg
== &root_task_group
&& rt_runtime
== 0)
8121 /* No period doesn't make any sense. */
8125 mutex_lock(&rt_constraints_mutex
);
8126 read_lock(&tasklist_lock
);
8127 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8131 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8132 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8133 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8135 for_each_possible_cpu(i
) {
8136 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8138 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8139 rt_rq
->rt_runtime
= rt_runtime
;
8140 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8142 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8144 read_unlock(&tasklist_lock
);
8145 mutex_unlock(&rt_constraints_mutex
);
8150 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8152 u64 rt_runtime
, rt_period
;
8154 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8155 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8156 if (rt_runtime_us
< 0)
8157 rt_runtime
= RUNTIME_INF
;
8159 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8162 static long sched_group_rt_runtime(struct task_group
*tg
)
8166 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8169 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8170 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8171 return rt_runtime_us
;
8174 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
8176 u64 rt_runtime
, rt_period
;
8178 rt_period
= rt_period_us
* NSEC_PER_USEC
;
8179 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8181 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8184 static long sched_group_rt_period(struct task_group
*tg
)
8188 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8189 do_div(rt_period_us
, NSEC_PER_USEC
);
8190 return rt_period_us
;
8192 #endif /* CONFIG_RT_GROUP_SCHED */
8194 #ifdef CONFIG_RT_GROUP_SCHED
8195 static int sched_rt_global_constraints(void)
8199 mutex_lock(&rt_constraints_mutex
);
8200 read_lock(&tasklist_lock
);
8201 ret
= __rt_schedulable(NULL
, 0, 0);
8202 read_unlock(&tasklist_lock
);
8203 mutex_unlock(&rt_constraints_mutex
);
8208 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8210 /* Don't accept realtime tasks when there is no way for them to run */
8211 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8217 #else /* !CONFIG_RT_GROUP_SCHED */
8218 static int sched_rt_global_constraints(void)
8220 unsigned long flags
;
8223 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8224 for_each_possible_cpu(i
) {
8225 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8227 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8228 rt_rq
->rt_runtime
= global_rt_runtime();
8229 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8231 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8235 #endif /* CONFIG_RT_GROUP_SCHED */
8237 static int sched_dl_global_validate(void)
8239 u64 runtime
= global_rt_runtime();
8240 u64 period
= global_rt_period();
8241 u64 new_bw
= to_ratio(period
, runtime
);
8244 unsigned long flags
;
8247 * Here we want to check the bandwidth not being set to some
8248 * value smaller than the currently allocated bandwidth in
8249 * any of the root_domains.
8251 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8252 * cycling on root_domains... Discussion on different/better
8253 * solutions is welcome!
8255 for_each_possible_cpu(cpu
) {
8256 rcu_read_lock_sched();
8257 dl_b
= dl_bw_of(cpu
);
8259 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8260 if (new_bw
< dl_b
->total_bw
)
8262 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8264 rcu_read_unlock_sched();
8273 static void sched_dl_do_global(void)
8278 unsigned long flags
;
8280 def_dl_bandwidth
.dl_period
= global_rt_period();
8281 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8283 if (global_rt_runtime() != RUNTIME_INF
)
8284 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8287 * FIXME: As above...
8289 for_each_possible_cpu(cpu
) {
8290 rcu_read_lock_sched();
8291 dl_b
= dl_bw_of(cpu
);
8293 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8295 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8297 rcu_read_unlock_sched();
8301 static int sched_rt_global_validate(void)
8303 if (sysctl_sched_rt_period
<= 0)
8306 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8307 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8313 static void sched_rt_do_global(void)
8315 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8316 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8319 int sched_rt_handler(struct ctl_table
*table
, int write
,
8320 void __user
*buffer
, size_t *lenp
,
8323 int old_period
, old_runtime
;
8324 static DEFINE_MUTEX(mutex
);
8328 old_period
= sysctl_sched_rt_period
;
8329 old_runtime
= sysctl_sched_rt_runtime
;
8331 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8333 if (!ret
&& write
) {
8334 ret
= sched_rt_global_validate();
8338 ret
= sched_dl_global_validate();
8342 ret
= sched_rt_global_constraints();
8346 sched_rt_do_global();
8347 sched_dl_do_global();
8351 sysctl_sched_rt_period
= old_period
;
8352 sysctl_sched_rt_runtime
= old_runtime
;
8354 mutex_unlock(&mutex
);
8359 int sched_rr_handler(struct ctl_table
*table
, int write
,
8360 void __user
*buffer
, size_t *lenp
,
8364 static DEFINE_MUTEX(mutex
);
8367 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8368 /* make sure that internally we keep jiffies */
8369 /* also, writing zero resets timeslice to default */
8370 if (!ret
&& write
) {
8371 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8372 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8374 mutex_unlock(&mutex
);
8378 #ifdef CONFIG_CGROUP_SCHED
8380 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8382 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8385 static struct cgroup_subsys_state
*
8386 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8388 struct task_group
*parent
= css_tg(parent_css
);
8389 struct task_group
*tg
;
8392 /* This is early initialization for the top cgroup */
8393 return &root_task_group
.css
;
8396 tg
= sched_create_group(parent
);
8398 return ERR_PTR(-ENOMEM
);
8400 sched_online_group(tg
, parent
);
8405 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8407 struct task_group
*tg
= css_tg(css
);
8409 sched_offline_group(tg
);
8412 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8414 struct task_group
*tg
= css_tg(css
);
8417 * Relies on the RCU grace period between css_released() and this.
8419 sched_free_group(tg
);
8423 * This is called before wake_up_new_task(), therefore we really only
8424 * have to set its group bits, all the other stuff does not apply.
8426 static void cpu_cgroup_fork(struct task_struct
*task
)
8431 rq
= task_rq_lock(task
, &rf
);
8433 sched_change_group(task
, TASK_SET_GROUP
);
8435 task_rq_unlock(rq
, task
, &rf
);
8438 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8440 struct task_struct
*task
;
8441 struct cgroup_subsys_state
*css
;
8444 cgroup_taskset_for_each(task
, css
, tset
) {
8445 #ifdef CONFIG_RT_GROUP_SCHED
8446 if (!sched_rt_can_attach(css_tg(css
), task
))
8449 /* We don't support RT-tasks being in separate groups */
8450 if (task
->sched_class
!= &fair_sched_class
)
8454 * Serialize against wake_up_new_task() such that if its
8455 * running, we're sure to observe its full state.
8457 raw_spin_lock_irq(&task
->pi_lock
);
8459 * Avoid calling sched_move_task() before wake_up_new_task()
8460 * has happened. This would lead to problems with PELT, due to
8461 * move wanting to detach+attach while we're not attached yet.
8463 if (task
->state
== TASK_NEW
)
8465 raw_spin_unlock_irq(&task
->pi_lock
);
8473 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8475 struct task_struct
*task
;
8476 struct cgroup_subsys_state
*css
;
8478 cgroup_taskset_for_each(task
, css
, tset
)
8479 sched_move_task(task
);
8482 #ifdef CONFIG_FAIR_GROUP_SCHED
8483 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8484 struct cftype
*cftype
, u64 shareval
)
8486 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8489 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8492 struct task_group
*tg
= css_tg(css
);
8494 return (u64
) scale_load_down(tg
->shares
);
8497 #ifdef CONFIG_CFS_BANDWIDTH
8498 static DEFINE_MUTEX(cfs_constraints_mutex
);
8500 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8501 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8503 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8505 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8507 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8508 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8510 if (tg
== &root_task_group
)
8514 * Ensure we have at some amount of bandwidth every period. This is
8515 * to prevent reaching a state of large arrears when throttled via
8516 * entity_tick() resulting in prolonged exit starvation.
8518 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8522 * Likewise, bound things on the otherside by preventing insane quota
8523 * periods. This also allows us to normalize in computing quota
8526 if (period
> max_cfs_quota_period
)
8530 * Prevent race between setting of cfs_rq->runtime_enabled and
8531 * unthrottle_offline_cfs_rqs().
8534 mutex_lock(&cfs_constraints_mutex
);
8535 ret
= __cfs_schedulable(tg
, period
, quota
);
8539 runtime_enabled
= quota
!= RUNTIME_INF
;
8540 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8542 * If we need to toggle cfs_bandwidth_used, off->on must occur
8543 * before making related changes, and on->off must occur afterwards
8545 if (runtime_enabled
&& !runtime_was_enabled
)
8546 cfs_bandwidth_usage_inc();
8547 raw_spin_lock_irq(&cfs_b
->lock
);
8548 cfs_b
->period
= ns_to_ktime(period
);
8549 cfs_b
->quota
= quota
;
8551 __refill_cfs_bandwidth_runtime(cfs_b
);
8552 /* restart the period timer (if active) to handle new period expiry */
8553 if (runtime_enabled
)
8554 start_cfs_bandwidth(cfs_b
);
8555 raw_spin_unlock_irq(&cfs_b
->lock
);
8557 for_each_online_cpu(i
) {
8558 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8559 struct rq
*rq
= cfs_rq
->rq
;
8561 raw_spin_lock_irq(&rq
->lock
);
8562 cfs_rq
->runtime_enabled
= runtime_enabled
;
8563 cfs_rq
->runtime_remaining
= 0;
8565 if (cfs_rq
->throttled
)
8566 unthrottle_cfs_rq(cfs_rq
);
8567 raw_spin_unlock_irq(&rq
->lock
);
8569 if (runtime_was_enabled
&& !runtime_enabled
)
8570 cfs_bandwidth_usage_dec();
8572 mutex_unlock(&cfs_constraints_mutex
);
8578 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8582 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8583 if (cfs_quota_us
< 0)
8584 quota
= RUNTIME_INF
;
8586 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8588 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8591 long tg_get_cfs_quota(struct task_group
*tg
)
8595 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8598 quota_us
= tg
->cfs_bandwidth
.quota
;
8599 do_div(quota_us
, NSEC_PER_USEC
);
8604 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8608 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8609 quota
= tg
->cfs_bandwidth
.quota
;
8611 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8614 long tg_get_cfs_period(struct task_group
*tg
)
8618 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8619 do_div(cfs_period_us
, NSEC_PER_USEC
);
8621 return cfs_period_us
;
8624 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8627 return tg_get_cfs_quota(css_tg(css
));
8630 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8631 struct cftype
*cftype
, s64 cfs_quota_us
)
8633 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8636 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8639 return tg_get_cfs_period(css_tg(css
));
8642 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8643 struct cftype
*cftype
, u64 cfs_period_us
)
8645 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8648 struct cfs_schedulable_data
{
8649 struct task_group
*tg
;
8654 * normalize group quota/period to be quota/max_period
8655 * note: units are usecs
8657 static u64
normalize_cfs_quota(struct task_group
*tg
,
8658 struct cfs_schedulable_data
*d
)
8666 period
= tg_get_cfs_period(tg
);
8667 quota
= tg_get_cfs_quota(tg
);
8670 /* note: these should typically be equivalent */
8671 if (quota
== RUNTIME_INF
|| quota
== -1)
8674 return to_ratio(period
, quota
);
8677 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8679 struct cfs_schedulable_data
*d
= data
;
8680 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8681 s64 quota
= 0, parent_quota
= -1;
8684 quota
= RUNTIME_INF
;
8686 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8688 quota
= normalize_cfs_quota(tg
, d
);
8689 parent_quota
= parent_b
->hierarchical_quota
;
8692 * ensure max(child_quota) <= parent_quota, inherit when no
8695 if (quota
== RUNTIME_INF
)
8696 quota
= parent_quota
;
8697 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8700 cfs_b
->hierarchical_quota
= quota
;
8705 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8708 struct cfs_schedulable_data data
= {
8714 if (quota
!= RUNTIME_INF
) {
8715 do_div(data
.period
, NSEC_PER_USEC
);
8716 do_div(data
.quota
, NSEC_PER_USEC
);
8720 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8726 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8728 struct task_group
*tg
= css_tg(seq_css(sf
));
8729 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8731 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8732 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8733 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8737 #endif /* CONFIG_CFS_BANDWIDTH */
8738 #endif /* CONFIG_FAIR_GROUP_SCHED */
8740 #ifdef CONFIG_RT_GROUP_SCHED
8741 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8742 struct cftype
*cft
, s64 val
)
8744 return sched_group_set_rt_runtime(css_tg(css
), val
);
8747 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8750 return sched_group_rt_runtime(css_tg(css
));
8753 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8754 struct cftype
*cftype
, u64 rt_period_us
)
8756 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8759 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8762 return sched_group_rt_period(css_tg(css
));
8764 #endif /* CONFIG_RT_GROUP_SCHED */
8766 static struct cftype cpu_files
[] = {
8767 #ifdef CONFIG_FAIR_GROUP_SCHED
8770 .read_u64
= cpu_shares_read_u64
,
8771 .write_u64
= cpu_shares_write_u64
,
8774 #ifdef CONFIG_CFS_BANDWIDTH
8776 .name
= "cfs_quota_us",
8777 .read_s64
= cpu_cfs_quota_read_s64
,
8778 .write_s64
= cpu_cfs_quota_write_s64
,
8781 .name
= "cfs_period_us",
8782 .read_u64
= cpu_cfs_period_read_u64
,
8783 .write_u64
= cpu_cfs_period_write_u64
,
8787 .seq_show
= cpu_stats_show
,
8790 #ifdef CONFIG_RT_GROUP_SCHED
8792 .name
= "rt_runtime_us",
8793 .read_s64
= cpu_rt_runtime_read
,
8794 .write_s64
= cpu_rt_runtime_write
,
8797 .name
= "rt_period_us",
8798 .read_u64
= cpu_rt_period_read_uint
,
8799 .write_u64
= cpu_rt_period_write_uint
,
8805 struct cgroup_subsys cpu_cgrp_subsys
= {
8806 .css_alloc
= cpu_cgroup_css_alloc
,
8807 .css_released
= cpu_cgroup_css_released
,
8808 .css_free
= cpu_cgroup_css_free
,
8809 .fork
= cpu_cgroup_fork
,
8810 .can_attach
= cpu_cgroup_can_attach
,
8811 .attach
= cpu_cgroup_attach
,
8812 .legacy_cftypes
= cpu_files
,
8816 #endif /* CONFIG_CGROUP_SCHED */
8818 void dump_cpu_task(int cpu
)
8820 pr_info("Task dump for CPU %d:\n", cpu
);
8821 sched_show_task(cpu_curr(cpu
));
8825 * Nice levels are multiplicative, with a gentle 10% change for every
8826 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8827 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8828 * that remained on nice 0.
8830 * The "10% effect" is relative and cumulative: from _any_ nice level,
8831 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8832 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8833 * If a task goes up by ~10% and another task goes down by ~10% then
8834 * the relative distance between them is ~25%.)
8836 const int sched_prio_to_weight
[40] = {
8837 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8838 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8839 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8840 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8841 /* 0 */ 1024, 820, 655, 526, 423,
8842 /* 5 */ 335, 272, 215, 172, 137,
8843 /* 10 */ 110, 87, 70, 56, 45,
8844 /* 15 */ 36, 29, 23, 18, 15,
8848 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8850 * In cases where the weight does not change often, we can use the
8851 * precalculated inverse to speed up arithmetics by turning divisions
8852 * into multiplications:
8854 const u32 sched_prio_to_wmult
[40] = {
8855 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8856 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8857 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8858 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8859 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8860 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8861 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8862 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,