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>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.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
&& task_on_rq_queued(p
))
1074 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1075 raw_spin_unlock(&rq
->lock
);
1076 raw_spin_unlock(&p
->pi_lock
);
1083 * sched_class::set_cpus_allowed must do the below, but is not required to
1084 * actually call this function.
1086 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1088 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1089 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1092 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1094 struct rq
*rq
= task_rq(p
);
1095 bool queued
, running
;
1097 lockdep_assert_held(&p
->pi_lock
);
1099 queued
= task_on_rq_queued(p
);
1100 running
= task_current(rq
, p
);
1104 * Because __kthread_bind() calls this on blocked tasks without
1107 lockdep_assert_held(&rq
->lock
);
1108 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1111 put_prev_task(rq
, p
);
1113 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1116 p
->sched_class
->set_curr_task(rq
);
1118 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1122 * Change a given task's CPU affinity. Migrate the thread to a
1123 * proper CPU and schedule it away if the CPU it's executing on
1124 * is removed from the allowed bitmask.
1126 * NOTE: the caller must have a valid reference to the task, the
1127 * task must not exit() & deallocate itself prematurely. The
1128 * call is not atomic; no spinlocks may be held.
1130 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1131 const struct cpumask
*new_mask
, bool check
)
1133 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1134 unsigned int dest_cpu
;
1139 rq
= task_rq_lock(p
, &rf
);
1141 if (p
->flags
& PF_KTHREAD
) {
1143 * Kernel threads are allowed on online && !active CPUs
1145 cpu_valid_mask
= cpu_online_mask
;
1149 * Must re-check here, to close a race against __kthread_bind(),
1150 * sched_setaffinity() is not guaranteed to observe the flag.
1152 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1157 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1160 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1165 do_set_cpus_allowed(p
, new_mask
);
1167 if (p
->flags
& PF_KTHREAD
) {
1169 * For kernel threads that do indeed end up on online &&
1170 * !active we want to ensure they are strict per-cpu threads.
1172 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1173 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1174 p
->nr_cpus_allowed
!= 1);
1177 /* Can the task run on the task's current CPU? If so, we're done */
1178 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1181 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1182 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1183 struct migration_arg arg
= { p
, dest_cpu
};
1184 /* Need help from migration thread: drop lock and wait. */
1185 task_rq_unlock(rq
, p
, &rf
);
1186 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1187 tlb_migrate_finish(p
->mm
);
1189 } else if (task_on_rq_queued(p
)) {
1191 * OK, since we're going to drop the lock immediately
1192 * afterwards anyway.
1194 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
1195 rq
= move_queued_task(rq
, p
, dest_cpu
);
1196 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
1199 task_rq_unlock(rq
, p
, &rf
);
1204 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1206 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1208 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1210 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1212 #ifdef CONFIG_SCHED_DEBUG
1214 * We should never call set_task_cpu() on a blocked task,
1215 * ttwu() will sort out the placement.
1217 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1221 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1222 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1223 * time relying on p->on_rq.
1225 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1226 p
->sched_class
== &fair_sched_class
&&
1227 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1229 #ifdef CONFIG_LOCKDEP
1231 * The caller should hold either p->pi_lock or rq->lock, when changing
1232 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1234 * sched_move_task() holds both and thus holding either pins the cgroup,
1237 * Furthermore, all task_rq users should acquire both locks, see
1240 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1241 lockdep_is_held(&task_rq(p
)->lock
)));
1245 trace_sched_migrate_task(p
, new_cpu
);
1247 if (task_cpu(p
) != new_cpu
) {
1248 if (p
->sched_class
->migrate_task_rq
)
1249 p
->sched_class
->migrate_task_rq(p
);
1250 p
->se
.nr_migrations
++;
1251 perf_event_task_migrate(p
);
1254 __set_task_cpu(p
, new_cpu
);
1257 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1259 if (task_on_rq_queued(p
)) {
1260 struct rq
*src_rq
, *dst_rq
;
1262 src_rq
= task_rq(p
);
1263 dst_rq
= cpu_rq(cpu
);
1265 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1266 deactivate_task(src_rq
, p
, 0);
1267 set_task_cpu(p
, cpu
);
1268 activate_task(dst_rq
, p
, 0);
1269 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1270 check_preempt_curr(dst_rq
, p
, 0);
1273 * Task isn't running anymore; make it appear like we migrated
1274 * it before it went to sleep. This means on wakeup we make the
1275 * previous cpu our target instead of where it really is.
1281 struct migration_swap_arg
{
1282 struct task_struct
*src_task
, *dst_task
;
1283 int src_cpu
, dst_cpu
;
1286 static int migrate_swap_stop(void *data
)
1288 struct migration_swap_arg
*arg
= data
;
1289 struct rq
*src_rq
, *dst_rq
;
1292 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1295 src_rq
= cpu_rq(arg
->src_cpu
);
1296 dst_rq
= cpu_rq(arg
->dst_cpu
);
1298 double_raw_lock(&arg
->src_task
->pi_lock
,
1299 &arg
->dst_task
->pi_lock
);
1300 double_rq_lock(src_rq
, dst_rq
);
1302 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1305 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1308 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1311 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1314 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1315 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1320 double_rq_unlock(src_rq
, dst_rq
);
1321 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1322 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1328 * Cross migrate two tasks
1330 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1332 struct migration_swap_arg arg
;
1335 arg
= (struct migration_swap_arg
){
1337 .src_cpu
= task_cpu(cur
),
1339 .dst_cpu
= task_cpu(p
),
1342 if (arg
.src_cpu
== arg
.dst_cpu
)
1346 * These three tests are all lockless; this is OK since all of them
1347 * will be re-checked with proper locks held further down the line.
1349 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1352 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1355 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1358 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1359 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1366 * wait_task_inactive - wait for a thread to unschedule.
1368 * If @match_state is nonzero, it's the @p->state value just checked and
1369 * not expected to change. If it changes, i.e. @p might have woken up,
1370 * then return zero. When we succeed in waiting for @p to be off its CPU,
1371 * we return a positive number (its total switch count). If a second call
1372 * a short while later returns the same number, the caller can be sure that
1373 * @p has remained unscheduled the whole time.
1375 * The caller must ensure that the task *will* unschedule sometime soon,
1376 * else this function might spin for a *long* time. This function can't
1377 * be called with interrupts off, or it may introduce deadlock with
1378 * smp_call_function() if an IPI is sent by the same process we are
1379 * waiting to become inactive.
1381 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1383 int running
, queued
;
1390 * We do the initial early heuristics without holding
1391 * any task-queue locks at all. We'll only try to get
1392 * the runqueue lock when things look like they will
1398 * If the task is actively running on another CPU
1399 * still, just relax and busy-wait without holding
1402 * NOTE! Since we don't hold any locks, it's not
1403 * even sure that "rq" stays as the right runqueue!
1404 * But we don't care, since "task_running()" will
1405 * return false if the runqueue has changed and p
1406 * is actually now running somewhere else!
1408 while (task_running(rq
, p
)) {
1409 if (match_state
&& unlikely(p
->state
!= match_state
))
1415 * Ok, time to look more closely! We need the rq
1416 * lock now, to be *sure*. If we're wrong, we'll
1417 * just go back and repeat.
1419 rq
= task_rq_lock(p
, &rf
);
1420 trace_sched_wait_task(p
);
1421 running
= task_running(rq
, p
);
1422 queued
= task_on_rq_queued(p
);
1424 if (!match_state
|| p
->state
== match_state
)
1425 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1426 task_rq_unlock(rq
, p
, &rf
);
1429 * If it changed from the expected state, bail out now.
1431 if (unlikely(!ncsw
))
1435 * Was it really running after all now that we
1436 * checked with the proper locks actually held?
1438 * Oops. Go back and try again..
1440 if (unlikely(running
)) {
1446 * It's not enough that it's not actively running,
1447 * it must be off the runqueue _entirely_, and not
1450 * So if it was still runnable (but just not actively
1451 * running right now), it's preempted, and we should
1452 * yield - it could be a while.
1454 if (unlikely(queued
)) {
1455 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1457 set_current_state(TASK_UNINTERRUPTIBLE
);
1458 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1463 * Ahh, all good. It wasn't running, and it wasn't
1464 * runnable, which means that it will never become
1465 * running in the future either. We're all done!
1474 * kick_process - kick a running thread to enter/exit the kernel
1475 * @p: the to-be-kicked thread
1477 * Cause a process which is running on another CPU to enter
1478 * kernel-mode, without any delay. (to get signals handled.)
1480 * NOTE: this function doesn't have to take the runqueue lock,
1481 * because all it wants to ensure is that the remote task enters
1482 * the kernel. If the IPI races and the task has been migrated
1483 * to another CPU then no harm is done and the purpose has been
1486 void kick_process(struct task_struct
*p
)
1492 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1493 smp_send_reschedule(cpu
);
1496 EXPORT_SYMBOL_GPL(kick_process
);
1499 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1501 * A few notes on cpu_active vs cpu_online:
1503 * - cpu_active must be a subset of cpu_online
1505 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1506 * see __set_cpus_allowed_ptr(). At this point the newly online
1507 * cpu isn't yet part of the sched domains, and balancing will not
1510 * - on cpu-down we clear cpu_active() to mask the sched domains and
1511 * avoid the load balancer to place new tasks on the to be removed
1512 * cpu. Existing tasks will remain running there and will be taken
1515 * This means that fallback selection must not select !active CPUs.
1516 * And can assume that any active CPU must be online. Conversely
1517 * select_task_rq() below may allow selection of !active CPUs in order
1518 * to satisfy the above rules.
1520 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1522 int nid
= cpu_to_node(cpu
);
1523 const struct cpumask
*nodemask
= NULL
;
1524 enum { cpuset
, possible
, fail
} state
= cpuset
;
1528 * If the node that the cpu is on has been offlined, cpu_to_node()
1529 * will return -1. There is no cpu on the node, and we should
1530 * select the cpu on the other node.
1533 nodemask
= cpumask_of_node(nid
);
1535 /* Look for allowed, online CPU in same node. */
1536 for_each_cpu(dest_cpu
, nodemask
) {
1537 if (!cpu_active(dest_cpu
))
1539 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1545 /* Any allowed, online CPU? */
1546 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1547 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1549 if (!cpu_online(dest_cpu
))
1554 /* No more Mr. Nice Guy. */
1557 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1558 cpuset_cpus_allowed_fallback(p
);
1564 do_set_cpus_allowed(p
, cpu_possible_mask
);
1575 if (state
!= cpuset
) {
1577 * Don't tell them about moving exiting tasks or
1578 * kernel threads (both mm NULL), since they never
1581 if (p
->mm
&& printk_ratelimit()) {
1582 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1583 task_pid_nr(p
), p
->comm
, cpu
);
1591 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1594 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1596 lockdep_assert_held(&p
->pi_lock
);
1598 if (tsk_nr_cpus_allowed(p
) > 1)
1599 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1601 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1604 * In order not to call set_task_cpu() on a blocking task we need
1605 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1608 * Since this is common to all placement strategies, this lives here.
1610 * [ this allows ->select_task() to simply return task_cpu(p) and
1611 * not worry about this generic constraint ]
1613 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1615 cpu
= select_fallback_rq(task_cpu(p
), p
);
1620 static void update_avg(u64
*avg
, u64 sample
)
1622 s64 diff
= sample
- *avg
;
1628 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1629 const struct cpumask
*new_mask
, bool check
)
1631 return set_cpus_allowed_ptr(p
, new_mask
);
1634 #endif /* CONFIG_SMP */
1637 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1641 if (!schedstat_enabled())
1647 if (cpu
== rq
->cpu
) {
1648 schedstat_inc(rq
->ttwu_local
);
1649 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1651 struct sched_domain
*sd
;
1653 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1655 for_each_domain(rq
->cpu
, sd
) {
1656 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1657 schedstat_inc(sd
->ttwu_wake_remote
);
1664 if (wake_flags
& WF_MIGRATED
)
1665 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1666 #endif /* CONFIG_SMP */
1668 schedstat_inc(rq
->ttwu_count
);
1669 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1671 if (wake_flags
& WF_SYNC
)
1672 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1675 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1677 activate_task(rq
, p
, en_flags
);
1678 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1680 /* if a worker is waking up, notify workqueue */
1681 if (p
->flags
& PF_WQ_WORKER
)
1682 wq_worker_waking_up(p
, cpu_of(rq
));
1686 * Mark the task runnable and perform wakeup-preemption.
1688 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1689 struct pin_cookie cookie
)
1691 check_preempt_curr(rq
, p
, wake_flags
);
1692 p
->state
= TASK_RUNNING
;
1693 trace_sched_wakeup(p
);
1696 if (p
->sched_class
->task_woken
) {
1698 * Our task @p is fully woken up and running; so its safe to
1699 * drop the rq->lock, hereafter rq is only used for statistics.
1701 lockdep_unpin_lock(&rq
->lock
, cookie
);
1702 p
->sched_class
->task_woken(rq
, p
);
1703 lockdep_repin_lock(&rq
->lock
, cookie
);
1706 if (rq
->idle_stamp
) {
1707 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1708 u64 max
= 2*rq
->max_idle_balance_cost
;
1710 update_avg(&rq
->avg_idle
, delta
);
1712 if (rq
->avg_idle
> max
)
1721 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1722 struct pin_cookie cookie
)
1724 int en_flags
= ENQUEUE_WAKEUP
;
1726 lockdep_assert_held(&rq
->lock
);
1729 if (p
->sched_contributes_to_load
)
1730 rq
->nr_uninterruptible
--;
1732 if (wake_flags
& WF_MIGRATED
)
1733 en_flags
|= ENQUEUE_MIGRATED
;
1736 ttwu_activate(rq
, p
, en_flags
);
1737 ttwu_do_wakeup(rq
, p
, wake_flags
, cookie
);
1741 * Called in case the task @p isn't fully descheduled from its runqueue,
1742 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1743 * since all we need to do is flip p->state to TASK_RUNNING, since
1744 * the task is still ->on_rq.
1746 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1752 rq
= __task_rq_lock(p
, &rf
);
1753 if (task_on_rq_queued(p
)) {
1754 /* check_preempt_curr() may use rq clock */
1755 update_rq_clock(rq
);
1756 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
.cookie
);
1759 __task_rq_unlock(rq
, &rf
);
1765 void sched_ttwu_pending(void)
1767 struct rq
*rq
= this_rq();
1768 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1769 struct pin_cookie cookie
;
1770 struct task_struct
*p
;
1771 unsigned long flags
;
1776 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1777 cookie
= lockdep_pin_lock(&rq
->lock
);
1782 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1783 llist
= llist_next(llist
);
1785 if (p
->sched_remote_wakeup
)
1786 wake_flags
= WF_MIGRATED
;
1788 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1791 lockdep_unpin_lock(&rq
->lock
, cookie
);
1792 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1795 void scheduler_ipi(void)
1798 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1799 * TIF_NEED_RESCHED remotely (for the first time) will also send
1802 preempt_fold_need_resched();
1804 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1808 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1809 * traditionally all their work was done from the interrupt return
1810 * path. Now that we actually do some work, we need to make sure
1813 * Some archs already do call them, luckily irq_enter/exit nest
1816 * Arguably we should visit all archs and update all handlers,
1817 * however a fair share of IPIs are still resched only so this would
1818 * somewhat pessimize the simple resched case.
1821 sched_ttwu_pending();
1824 * Check if someone kicked us for doing the nohz idle load balance.
1826 if (unlikely(got_nohz_idle_kick())) {
1827 this_rq()->idle_balance
= 1;
1828 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1833 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1835 struct rq
*rq
= cpu_rq(cpu
);
1837 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1839 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1840 if (!set_nr_if_polling(rq
->idle
))
1841 smp_send_reschedule(cpu
);
1843 trace_sched_wake_idle_without_ipi(cpu
);
1847 void wake_up_if_idle(int cpu
)
1849 struct rq
*rq
= cpu_rq(cpu
);
1850 unsigned long flags
;
1854 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1857 if (set_nr_if_polling(rq
->idle
)) {
1858 trace_sched_wake_idle_without_ipi(cpu
);
1860 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1861 if (is_idle_task(rq
->curr
))
1862 smp_send_reschedule(cpu
);
1863 /* Else cpu is not in idle, do nothing here */
1864 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1871 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1873 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1875 #endif /* CONFIG_SMP */
1877 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1879 struct rq
*rq
= cpu_rq(cpu
);
1880 struct pin_cookie cookie
;
1882 #if defined(CONFIG_SMP)
1883 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1884 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1885 ttwu_queue_remote(p
, cpu
, wake_flags
);
1890 raw_spin_lock(&rq
->lock
);
1891 cookie
= lockdep_pin_lock(&rq
->lock
);
1892 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1893 lockdep_unpin_lock(&rq
->lock
, cookie
);
1894 raw_spin_unlock(&rq
->lock
);
1898 * Notes on Program-Order guarantees on SMP systems.
1902 * The basic program-order guarantee on SMP systems is that when a task [t]
1903 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1904 * execution on its new cpu [c1].
1906 * For migration (of runnable tasks) this is provided by the following means:
1908 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1909 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1910 * rq(c1)->lock (if not at the same time, then in that order).
1911 * C) LOCK of the rq(c1)->lock scheduling in task
1913 * Transitivity guarantees that B happens after A and C after B.
1914 * Note: we only require RCpc transitivity.
1915 * Note: the cpu doing B need not be c0 or c1
1924 * UNLOCK rq(0)->lock
1926 * LOCK rq(0)->lock // orders against CPU0
1928 * UNLOCK rq(0)->lock
1932 * UNLOCK rq(1)->lock
1934 * LOCK rq(1)->lock // orders against CPU2
1937 * UNLOCK rq(1)->lock
1940 * BLOCKING -- aka. SLEEP + WAKEUP
1942 * For blocking we (obviously) need to provide the same guarantee as for
1943 * migration. However the means are completely different as there is no lock
1944 * chain to provide order. Instead we do:
1946 * 1) smp_store_release(X->on_cpu, 0)
1947 * 2) smp_cond_load_acquire(!X->on_cpu)
1951 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1953 * LOCK rq(0)->lock LOCK X->pi_lock
1956 * smp_store_release(X->on_cpu, 0);
1958 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1964 * X->state = RUNNING
1965 * UNLOCK rq(2)->lock
1967 * LOCK rq(2)->lock // orders against CPU1
1970 * UNLOCK rq(2)->lock
1973 * UNLOCK rq(0)->lock
1976 * However; for wakeups there is a second guarantee we must provide, namely we
1977 * must observe the state that lead to our wakeup. That is, not only must our
1978 * task observe its own prior state, it must also observe the stores prior to
1981 * This means that any means of doing remote wakeups must order the CPU doing
1982 * the wakeup against the CPU the task is going to end up running on. This,
1983 * however, is already required for the regular Program-Order guarantee above,
1984 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1989 * try_to_wake_up - wake up a thread
1990 * @p: the thread to be awakened
1991 * @state: the mask of task states that can be woken
1992 * @wake_flags: wake modifier flags (WF_*)
1994 * Put it on the run-queue if it's not already there. The "current"
1995 * thread is always on the run-queue (except when the actual
1996 * re-schedule is in progress), and as such you're allowed to do
1997 * the simpler "current->state = TASK_RUNNING" to mark yourself
1998 * runnable without the overhead of this.
2000 * Return: %true if @p was woken up, %false if it was already running.
2001 * or @state didn't match @p's state.
2004 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2006 unsigned long flags
;
2007 int cpu
, success
= 0;
2010 * If we are going to wake up a thread waiting for CONDITION we
2011 * need to ensure that CONDITION=1 done by the caller can not be
2012 * reordered with p->state check below. This pairs with mb() in
2013 * set_current_state() the waiting thread does.
2015 smp_mb__before_spinlock();
2016 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2017 if (!(p
->state
& state
))
2020 trace_sched_waking(p
);
2022 success
= 1; /* we're going to change ->state */
2026 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2027 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2028 * in smp_cond_load_acquire() below.
2030 * sched_ttwu_pending() try_to_wake_up()
2031 * [S] p->on_rq = 1; [L] P->state
2032 * UNLOCK rq->lock -----.
2036 * LOCK rq->lock -----'
2040 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2042 * Pairs with the UNLOCK+LOCK on rq->lock from the
2043 * last wakeup of our task and the schedule that got our task
2047 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2052 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2053 * possible to, falsely, observe p->on_cpu == 0.
2055 * One must be running (->on_cpu == 1) in order to remove oneself
2056 * from the runqueue.
2058 * [S] ->on_cpu = 1; [L] ->on_rq
2062 * [S] ->on_rq = 0; [L] ->on_cpu
2064 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2065 * from the consecutive calls to schedule(); the first switching to our
2066 * task, the second putting it to sleep.
2071 * If the owning (remote) cpu is still in the middle of schedule() with
2072 * this task as prev, wait until its done referencing the task.
2074 * Pairs with the smp_store_release() in finish_lock_switch().
2076 * This ensures that tasks getting woken will be fully ordered against
2077 * their previous state and preserve Program Order.
2079 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2081 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2082 p
->state
= TASK_WAKING
;
2084 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2085 if (task_cpu(p
) != cpu
) {
2086 wake_flags
|= WF_MIGRATED
;
2087 set_task_cpu(p
, cpu
);
2089 #endif /* CONFIG_SMP */
2091 ttwu_queue(p
, cpu
, wake_flags
);
2093 ttwu_stat(p
, cpu
, wake_flags
);
2095 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2101 * try_to_wake_up_local - try to wake up a local task with rq lock held
2102 * @p: the thread to be awakened
2103 * @cookie: context's cookie for pinning
2105 * Put @p on the run-queue if it's not already there. The caller must
2106 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2109 static void try_to_wake_up_local(struct task_struct
*p
, struct pin_cookie cookie
)
2111 struct rq
*rq
= task_rq(p
);
2113 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2114 WARN_ON_ONCE(p
== current
))
2117 lockdep_assert_held(&rq
->lock
);
2119 if (!raw_spin_trylock(&p
->pi_lock
)) {
2121 * This is OK, because current is on_cpu, which avoids it being
2122 * picked for load-balance and preemption/IRQs are still
2123 * disabled avoiding further scheduler activity on it and we've
2124 * not yet picked a replacement task.
2126 lockdep_unpin_lock(&rq
->lock
, cookie
);
2127 raw_spin_unlock(&rq
->lock
);
2128 raw_spin_lock(&p
->pi_lock
);
2129 raw_spin_lock(&rq
->lock
);
2130 lockdep_repin_lock(&rq
->lock
, cookie
);
2133 if (!(p
->state
& TASK_NORMAL
))
2136 trace_sched_waking(p
);
2138 if (!task_on_rq_queued(p
))
2139 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2141 ttwu_do_wakeup(rq
, p
, 0, cookie
);
2142 ttwu_stat(p
, smp_processor_id(), 0);
2144 raw_spin_unlock(&p
->pi_lock
);
2148 * wake_up_process - Wake up a specific process
2149 * @p: The process to be woken up.
2151 * Attempt to wake up the nominated process and move it to the set of runnable
2154 * Return: 1 if the process was woken up, 0 if it was already running.
2156 * It may be assumed that this function implies a write memory barrier before
2157 * changing the task state if and only if any tasks are woken up.
2159 int wake_up_process(struct task_struct
*p
)
2161 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2163 EXPORT_SYMBOL(wake_up_process
);
2165 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2167 return try_to_wake_up(p
, state
, 0);
2171 * This function clears the sched_dl_entity static params.
2173 void __dl_clear_params(struct task_struct
*p
)
2175 struct sched_dl_entity
*dl_se
= &p
->dl
;
2177 dl_se
->dl_runtime
= 0;
2178 dl_se
->dl_deadline
= 0;
2179 dl_se
->dl_period
= 0;
2183 dl_se
->dl_throttled
= 0;
2184 dl_se
->dl_yielded
= 0;
2188 * Perform scheduler related setup for a newly forked process p.
2189 * p is forked by current.
2191 * __sched_fork() is basic setup used by init_idle() too:
2193 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2198 p
->se
.exec_start
= 0;
2199 p
->se
.sum_exec_runtime
= 0;
2200 p
->se
.prev_sum_exec_runtime
= 0;
2201 p
->se
.nr_migrations
= 0;
2203 INIT_LIST_HEAD(&p
->se
.group_node
);
2205 #ifdef CONFIG_FAIR_GROUP_SCHED
2206 p
->se
.cfs_rq
= NULL
;
2209 #ifdef CONFIG_SCHEDSTATS
2210 /* Even if schedstat is disabled, there should not be garbage */
2211 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2214 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2215 init_dl_task_timer(&p
->dl
);
2216 __dl_clear_params(p
);
2218 INIT_LIST_HEAD(&p
->rt
.run_list
);
2220 p
->rt
.time_slice
= sched_rr_timeslice
;
2224 #ifdef CONFIG_PREEMPT_NOTIFIERS
2225 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2228 #ifdef CONFIG_NUMA_BALANCING
2229 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2230 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2231 p
->mm
->numa_scan_seq
= 0;
2234 if (clone_flags
& CLONE_VM
)
2235 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2237 p
->numa_preferred_nid
= -1;
2239 p
->node_stamp
= 0ULL;
2240 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2241 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2242 p
->numa_work
.next
= &p
->numa_work
;
2243 p
->numa_faults
= NULL
;
2244 p
->last_task_numa_placement
= 0;
2245 p
->last_sum_exec_runtime
= 0;
2247 p
->numa_group
= NULL
;
2248 #endif /* CONFIG_NUMA_BALANCING */
2251 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2253 #ifdef CONFIG_NUMA_BALANCING
2255 void set_numabalancing_state(bool enabled
)
2258 static_branch_enable(&sched_numa_balancing
);
2260 static_branch_disable(&sched_numa_balancing
);
2263 #ifdef CONFIG_PROC_SYSCTL
2264 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2265 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2269 int state
= static_branch_likely(&sched_numa_balancing
);
2271 if (write
&& !capable(CAP_SYS_ADMIN
))
2276 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2280 set_numabalancing_state(state
);
2286 #ifdef CONFIG_SCHEDSTATS
2288 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2289 static bool __initdata __sched_schedstats
= false;
2291 static void set_schedstats(bool enabled
)
2294 static_branch_enable(&sched_schedstats
);
2296 static_branch_disable(&sched_schedstats
);
2299 void force_schedstat_enabled(void)
2301 if (!schedstat_enabled()) {
2302 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2303 static_branch_enable(&sched_schedstats
);
2307 static int __init
setup_schedstats(char *str
)
2314 * This code is called before jump labels have been set up, so we can't
2315 * change the static branch directly just yet. Instead set a temporary
2316 * variable so init_schedstats() can do it later.
2318 if (!strcmp(str
, "enable")) {
2319 __sched_schedstats
= true;
2321 } else if (!strcmp(str
, "disable")) {
2322 __sched_schedstats
= false;
2327 pr_warn("Unable to parse schedstats=\n");
2331 __setup("schedstats=", setup_schedstats
);
2333 static void __init
init_schedstats(void)
2335 set_schedstats(__sched_schedstats
);
2338 #ifdef CONFIG_PROC_SYSCTL
2339 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2340 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2344 int state
= static_branch_likely(&sched_schedstats
);
2346 if (write
&& !capable(CAP_SYS_ADMIN
))
2351 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2355 set_schedstats(state
);
2358 #endif /* CONFIG_PROC_SYSCTL */
2359 #else /* !CONFIG_SCHEDSTATS */
2360 static inline void init_schedstats(void) {}
2361 #endif /* CONFIG_SCHEDSTATS */
2364 * fork()/clone()-time setup:
2366 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2368 unsigned long flags
;
2369 int cpu
= get_cpu();
2371 __sched_fork(clone_flags
, p
);
2373 * We mark the process as NEW here. This guarantees that
2374 * nobody will actually run it, and a signal or other external
2375 * event cannot wake it up and insert it on the runqueue either.
2377 p
->state
= TASK_NEW
;
2380 * Make sure we do not leak PI boosting priority to the child.
2382 p
->prio
= current
->normal_prio
;
2385 * Revert to default priority/policy on fork if requested.
2387 if (unlikely(p
->sched_reset_on_fork
)) {
2388 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2389 p
->policy
= SCHED_NORMAL
;
2390 p
->static_prio
= NICE_TO_PRIO(0);
2392 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2393 p
->static_prio
= NICE_TO_PRIO(0);
2395 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2399 * We don't need the reset flag anymore after the fork. It has
2400 * fulfilled its duty:
2402 p
->sched_reset_on_fork
= 0;
2405 if (dl_prio(p
->prio
)) {
2408 } else if (rt_prio(p
->prio
)) {
2409 p
->sched_class
= &rt_sched_class
;
2411 p
->sched_class
= &fair_sched_class
;
2414 init_entity_runnable_average(&p
->se
);
2417 * The child is not yet in the pid-hash so no cgroup attach races,
2418 * and the cgroup is pinned to this child due to cgroup_fork()
2419 * is ran before sched_fork().
2421 * Silence PROVE_RCU.
2423 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2425 * We're setting the cpu for the first time, we don't migrate,
2426 * so use __set_task_cpu().
2428 __set_task_cpu(p
, cpu
);
2429 if (p
->sched_class
->task_fork
)
2430 p
->sched_class
->task_fork(p
);
2431 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2433 #ifdef CONFIG_SCHED_INFO
2434 if (likely(sched_info_on()))
2435 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2437 #if defined(CONFIG_SMP)
2440 init_task_preempt_count(p
);
2442 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2443 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2450 unsigned long to_ratio(u64 period
, u64 runtime
)
2452 if (runtime
== RUNTIME_INF
)
2456 * Doing this here saves a lot of checks in all
2457 * the calling paths, and returning zero seems
2458 * safe for them anyway.
2463 return div64_u64(runtime
<< 20, period
);
2467 inline struct dl_bw
*dl_bw_of(int i
)
2469 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2470 "sched RCU must be held");
2471 return &cpu_rq(i
)->rd
->dl_bw
;
2474 static inline int dl_bw_cpus(int i
)
2476 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2479 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2480 "sched RCU must be held");
2481 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2487 inline struct dl_bw
*dl_bw_of(int i
)
2489 return &cpu_rq(i
)->dl
.dl_bw
;
2492 static inline int dl_bw_cpus(int i
)
2499 * We must be sure that accepting a new task (or allowing changing the
2500 * parameters of an existing one) is consistent with the bandwidth
2501 * constraints. If yes, this function also accordingly updates the currently
2502 * allocated bandwidth to reflect the new situation.
2504 * This function is called while holding p's rq->lock.
2506 * XXX we should delay bw change until the task's 0-lag point, see
2509 static int dl_overflow(struct task_struct
*p
, int policy
,
2510 const struct sched_attr
*attr
)
2513 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2514 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2515 u64 runtime
= attr
->sched_runtime
;
2516 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2519 /* !deadline task may carry old deadline bandwidth */
2520 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2524 * Either if a task, enters, leave, or stays -deadline but changes
2525 * its parameters, we may need to update accordingly the total
2526 * allocated bandwidth of the container.
2528 raw_spin_lock(&dl_b
->lock
);
2529 cpus
= dl_bw_cpus(task_cpu(p
));
2530 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2531 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2532 __dl_add(dl_b
, new_bw
);
2534 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2535 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2536 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2537 __dl_add(dl_b
, new_bw
);
2539 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2540 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2543 raw_spin_unlock(&dl_b
->lock
);
2548 extern void init_dl_bw(struct dl_bw
*dl_b
);
2551 * wake_up_new_task - wake up a newly created task for the first time.
2553 * This function will do some initial scheduler statistics housekeeping
2554 * that must be done for every newly created context, then puts the task
2555 * on the runqueue and wakes it.
2557 void wake_up_new_task(struct task_struct
*p
)
2562 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2563 p
->state
= TASK_RUNNING
;
2566 * Fork balancing, do it here and not earlier because:
2567 * - cpus_allowed can change in the fork path
2568 * - any previously selected cpu might disappear through hotplug
2570 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2571 * as we're not fully set-up yet.
2573 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2575 rq
= __task_rq_lock(p
, &rf
);
2576 post_init_entity_util_avg(&p
->se
);
2578 activate_task(rq
, p
, 0);
2579 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2580 trace_sched_wakeup_new(p
);
2581 check_preempt_curr(rq
, p
, WF_FORK
);
2583 if (p
->sched_class
->task_woken
) {
2585 * Nothing relies on rq->lock after this, so its fine to
2588 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
2589 p
->sched_class
->task_woken(rq
, p
);
2590 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
2593 task_rq_unlock(rq
, p
, &rf
);
2596 #ifdef CONFIG_PREEMPT_NOTIFIERS
2598 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2600 void preempt_notifier_inc(void)
2602 static_key_slow_inc(&preempt_notifier_key
);
2604 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2606 void preempt_notifier_dec(void)
2608 static_key_slow_dec(&preempt_notifier_key
);
2610 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2613 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2614 * @notifier: notifier struct to register
2616 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2618 if (!static_key_false(&preempt_notifier_key
))
2619 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2621 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2623 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2626 * preempt_notifier_unregister - no longer interested in preemption notifications
2627 * @notifier: notifier struct to unregister
2629 * This is *not* safe to call from within a preemption notifier.
2631 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2633 hlist_del(¬ifier
->link
);
2635 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2637 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2639 struct preempt_notifier
*notifier
;
2641 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2642 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2645 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2647 if (static_key_false(&preempt_notifier_key
))
2648 __fire_sched_in_preempt_notifiers(curr
);
2652 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2653 struct task_struct
*next
)
2655 struct preempt_notifier
*notifier
;
2657 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2658 notifier
->ops
->sched_out(notifier
, next
);
2661 static __always_inline
void
2662 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2663 struct task_struct
*next
)
2665 if (static_key_false(&preempt_notifier_key
))
2666 __fire_sched_out_preempt_notifiers(curr
, next
);
2669 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2671 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2676 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2677 struct task_struct
*next
)
2681 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2684 * prepare_task_switch - prepare to switch tasks
2685 * @rq: the runqueue preparing to switch
2686 * @prev: the current task that is being switched out
2687 * @next: the task we are going to switch to.
2689 * This is called with the rq lock held and interrupts off. It must
2690 * be paired with a subsequent finish_task_switch after the context
2693 * prepare_task_switch sets up locking and calls architecture specific
2697 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2698 struct task_struct
*next
)
2700 sched_info_switch(rq
, prev
, next
);
2701 perf_event_task_sched_out(prev
, next
);
2702 fire_sched_out_preempt_notifiers(prev
, next
);
2703 prepare_lock_switch(rq
, next
);
2704 prepare_arch_switch(next
);
2708 * finish_task_switch - clean up after a task-switch
2709 * @prev: the thread we just switched away from.
2711 * finish_task_switch must be called after the context switch, paired
2712 * with a prepare_task_switch call before the context switch.
2713 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2714 * and do any other architecture-specific cleanup actions.
2716 * Note that we may have delayed dropping an mm in context_switch(). If
2717 * so, we finish that here outside of the runqueue lock. (Doing it
2718 * with the lock held can cause deadlocks; see schedule() for
2721 * The context switch have flipped the stack from under us and restored the
2722 * local variables which were saved when this task called schedule() in the
2723 * past. prev == current is still correct but we need to recalculate this_rq
2724 * because prev may have moved to another CPU.
2726 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2727 __releases(rq
->lock
)
2729 struct rq
*rq
= this_rq();
2730 struct mm_struct
*mm
= rq
->prev_mm
;
2734 * The previous task will have left us with a preempt_count of 2
2735 * because it left us after:
2738 * preempt_disable(); // 1
2740 * raw_spin_lock_irq(&rq->lock) // 2
2742 * Also, see FORK_PREEMPT_COUNT.
2744 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2745 "corrupted preempt_count: %s/%d/0x%x\n",
2746 current
->comm
, current
->pid
, preempt_count()))
2747 preempt_count_set(FORK_PREEMPT_COUNT
);
2752 * A task struct has one reference for the use as "current".
2753 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2754 * schedule one last time. The schedule call will never return, and
2755 * the scheduled task must drop that reference.
2757 * We must observe prev->state before clearing prev->on_cpu (in
2758 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2759 * running on another CPU and we could rave with its RUNNING -> DEAD
2760 * transition, resulting in a double drop.
2762 prev_state
= prev
->state
;
2763 vtime_task_switch(prev
);
2764 perf_event_task_sched_in(prev
, current
);
2765 finish_lock_switch(rq
, prev
);
2766 finish_arch_post_lock_switch();
2768 fire_sched_in_preempt_notifiers(current
);
2771 if (unlikely(prev_state
== TASK_DEAD
)) {
2772 if (prev
->sched_class
->task_dead
)
2773 prev
->sched_class
->task_dead(prev
);
2776 * Remove function-return probe instances associated with this
2777 * task and put them back on the free list.
2779 kprobe_flush_task(prev
);
2780 put_task_struct(prev
);
2783 tick_nohz_task_switch();
2789 /* rq->lock is NOT held, but preemption is disabled */
2790 static void __balance_callback(struct rq
*rq
)
2792 struct callback_head
*head
, *next
;
2793 void (*func
)(struct rq
*rq
);
2794 unsigned long flags
;
2796 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2797 head
= rq
->balance_callback
;
2798 rq
->balance_callback
= NULL
;
2800 func
= (void (*)(struct rq
*))head
->func
;
2807 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2810 static inline void balance_callback(struct rq
*rq
)
2812 if (unlikely(rq
->balance_callback
))
2813 __balance_callback(rq
);
2818 static inline void balance_callback(struct rq
*rq
)
2825 * schedule_tail - first thing a freshly forked thread must call.
2826 * @prev: the thread we just switched away from.
2828 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2829 __releases(rq
->lock
)
2834 * New tasks start with FORK_PREEMPT_COUNT, see there and
2835 * finish_task_switch() for details.
2837 * finish_task_switch() will drop rq->lock() and lower preempt_count
2838 * and the preempt_enable() will end up enabling preemption (on
2839 * PREEMPT_COUNT kernels).
2842 rq
= finish_task_switch(prev
);
2843 balance_callback(rq
);
2846 if (current
->set_child_tid
)
2847 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2851 * context_switch - switch to the new MM and the new thread's register state.
2853 static __always_inline
struct rq
*
2854 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2855 struct task_struct
*next
, struct pin_cookie cookie
)
2857 struct mm_struct
*mm
, *oldmm
;
2859 prepare_task_switch(rq
, prev
, next
);
2862 oldmm
= prev
->active_mm
;
2864 * For paravirt, this is coupled with an exit in switch_to to
2865 * combine the page table reload and the switch backend into
2868 arch_start_context_switch(prev
);
2871 next
->active_mm
= oldmm
;
2872 atomic_inc(&oldmm
->mm_count
);
2873 enter_lazy_tlb(oldmm
, next
);
2875 switch_mm_irqs_off(oldmm
, mm
, next
);
2878 prev
->active_mm
= NULL
;
2879 rq
->prev_mm
= oldmm
;
2882 * Since the runqueue lock will be released by the next
2883 * task (which is an invalid locking op but in the case
2884 * of the scheduler it's an obvious special-case), so we
2885 * do an early lockdep release here:
2887 lockdep_unpin_lock(&rq
->lock
, cookie
);
2888 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2890 /* Here we just switch the register state and the stack. */
2891 switch_to(prev
, next
, prev
);
2894 return finish_task_switch(prev
);
2898 * nr_running and nr_context_switches:
2900 * externally visible scheduler statistics: current number of runnable
2901 * threads, total number of context switches performed since bootup.
2903 unsigned long nr_running(void)
2905 unsigned long i
, sum
= 0;
2907 for_each_online_cpu(i
)
2908 sum
+= cpu_rq(i
)->nr_running
;
2914 * Check if only the current task is running on the cpu.
2916 * Caution: this function does not check that the caller has disabled
2917 * preemption, thus the result might have a time-of-check-to-time-of-use
2918 * race. The caller is responsible to use it correctly, for example:
2920 * - from a non-preemptable section (of course)
2922 * - from a thread that is bound to a single CPU
2924 * - in a loop with very short iterations (e.g. a polling loop)
2926 bool single_task_running(void)
2928 return raw_rq()->nr_running
== 1;
2930 EXPORT_SYMBOL(single_task_running
);
2932 unsigned long long nr_context_switches(void)
2935 unsigned long long sum
= 0;
2937 for_each_possible_cpu(i
)
2938 sum
+= cpu_rq(i
)->nr_switches
;
2943 unsigned long nr_iowait(void)
2945 unsigned long i
, sum
= 0;
2947 for_each_possible_cpu(i
)
2948 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2953 unsigned long nr_iowait_cpu(int cpu
)
2955 struct rq
*this = cpu_rq(cpu
);
2956 return atomic_read(&this->nr_iowait
);
2959 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2961 struct rq
*rq
= this_rq();
2962 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2963 *load
= rq
->load
.weight
;
2969 * sched_exec - execve() is a valuable balancing opportunity, because at
2970 * this point the task has the smallest effective memory and cache footprint.
2972 void sched_exec(void)
2974 struct task_struct
*p
= current
;
2975 unsigned long flags
;
2978 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2979 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2980 if (dest_cpu
== smp_processor_id())
2983 if (likely(cpu_active(dest_cpu
))) {
2984 struct migration_arg arg
= { p
, dest_cpu
};
2986 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2987 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2991 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2996 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2997 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2999 EXPORT_PER_CPU_SYMBOL(kstat
);
3000 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3003 * The function fair_sched_class.update_curr accesses the struct curr
3004 * and its field curr->exec_start; when called from task_sched_runtime(),
3005 * we observe a high rate of cache misses in practice.
3006 * Prefetching this data results in improved performance.
3008 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3010 #ifdef CONFIG_FAIR_GROUP_SCHED
3011 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3013 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3016 prefetch(&curr
->exec_start
);
3020 * Return accounted runtime for the task.
3021 * In case the task is currently running, return the runtime plus current's
3022 * pending runtime that have not been accounted yet.
3024 unsigned long long task_sched_runtime(struct task_struct
*p
)
3030 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3032 * 64-bit doesn't need locks to atomically read a 64bit value.
3033 * So we have a optimization chance when the task's delta_exec is 0.
3034 * Reading ->on_cpu is racy, but this is ok.
3036 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3037 * If we race with it entering cpu, unaccounted time is 0. This is
3038 * indistinguishable from the read occurring a few cycles earlier.
3039 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3040 * been accounted, so we're correct here as well.
3042 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3043 return p
->se
.sum_exec_runtime
;
3046 rq
= task_rq_lock(p
, &rf
);
3048 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3049 * project cycles that may never be accounted to this
3050 * thread, breaking clock_gettime().
3052 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3053 prefetch_curr_exec_start(p
);
3054 update_rq_clock(rq
);
3055 p
->sched_class
->update_curr(rq
);
3057 ns
= p
->se
.sum_exec_runtime
;
3058 task_rq_unlock(rq
, p
, &rf
);
3064 * This function gets called by the timer code, with HZ frequency.
3065 * We call it with interrupts disabled.
3067 void scheduler_tick(void)
3069 int cpu
= smp_processor_id();
3070 struct rq
*rq
= cpu_rq(cpu
);
3071 struct task_struct
*curr
= rq
->curr
;
3075 raw_spin_lock(&rq
->lock
);
3076 update_rq_clock(rq
);
3077 curr
->sched_class
->task_tick(rq
, curr
, 0);
3078 cpu_load_update_active(rq
);
3079 calc_global_load_tick(rq
);
3080 raw_spin_unlock(&rq
->lock
);
3082 perf_event_task_tick();
3085 rq
->idle_balance
= idle_cpu(cpu
);
3086 trigger_load_balance(rq
);
3088 rq_last_tick_reset(rq
);
3091 #ifdef CONFIG_NO_HZ_FULL
3093 * scheduler_tick_max_deferment
3095 * Keep at least one tick per second when a single
3096 * active task is running because the scheduler doesn't
3097 * yet completely support full dynticks environment.
3099 * This makes sure that uptime, CFS vruntime, load
3100 * balancing, etc... continue to move forward, even
3101 * with a very low granularity.
3103 * Return: Maximum deferment in nanoseconds.
3105 u64
scheduler_tick_max_deferment(void)
3107 struct rq
*rq
= this_rq();
3108 unsigned long next
, now
= READ_ONCE(jiffies
);
3110 next
= rq
->last_sched_tick
+ HZ
;
3112 if (time_before_eq(next
, now
))
3115 return jiffies_to_nsecs(next
- now
);
3119 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3120 defined(CONFIG_PREEMPT_TRACER))
3122 * If the value passed in is equal to the current preempt count
3123 * then we just disabled preemption. Start timing the latency.
3125 static inline void preempt_latency_start(int val
)
3127 if (preempt_count() == val
) {
3128 unsigned long ip
= get_lock_parent_ip();
3129 #ifdef CONFIG_DEBUG_PREEMPT
3130 current
->preempt_disable_ip
= ip
;
3132 trace_preempt_off(CALLER_ADDR0
, ip
);
3136 void preempt_count_add(int val
)
3138 #ifdef CONFIG_DEBUG_PREEMPT
3142 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3145 __preempt_count_add(val
);
3146 #ifdef CONFIG_DEBUG_PREEMPT
3148 * Spinlock count overflowing soon?
3150 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3153 preempt_latency_start(val
);
3155 EXPORT_SYMBOL(preempt_count_add
);
3156 NOKPROBE_SYMBOL(preempt_count_add
);
3159 * If the value passed in equals to the current preempt count
3160 * then we just enabled preemption. Stop timing the latency.
3162 static inline void preempt_latency_stop(int val
)
3164 if (preempt_count() == val
)
3165 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3168 void preempt_count_sub(int val
)
3170 #ifdef CONFIG_DEBUG_PREEMPT
3174 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3177 * Is the spinlock portion underflowing?
3179 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3180 !(preempt_count() & PREEMPT_MASK
)))
3184 preempt_latency_stop(val
);
3185 __preempt_count_sub(val
);
3187 EXPORT_SYMBOL(preempt_count_sub
);
3188 NOKPROBE_SYMBOL(preempt_count_sub
);
3191 static inline void preempt_latency_start(int val
) { }
3192 static inline void preempt_latency_stop(int val
) { }
3196 * Print scheduling while atomic bug:
3198 static noinline
void __schedule_bug(struct task_struct
*prev
)
3200 /* Save this before calling printk(), since that will clobber it */
3201 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3203 if (oops_in_progress
)
3206 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3207 prev
->comm
, prev
->pid
, preempt_count());
3209 debug_show_held_locks(prev
);
3211 if (irqs_disabled())
3212 print_irqtrace_events(prev
);
3213 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3214 && in_atomic_preempt_off()) {
3215 pr_err("Preemption disabled at:");
3216 print_ip_sym(preempt_disable_ip
);
3220 panic("scheduling while atomic\n");
3223 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3227 * Various schedule()-time debugging checks and statistics:
3229 static inline void schedule_debug(struct task_struct
*prev
)
3231 #ifdef CONFIG_SCHED_STACK_END_CHECK
3232 if (task_stack_end_corrupted(prev
))
3233 panic("corrupted stack end detected inside scheduler\n");
3236 if (unlikely(in_atomic_preempt_off())) {
3237 __schedule_bug(prev
);
3238 preempt_count_set(PREEMPT_DISABLED
);
3242 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3244 schedstat_inc(this_rq()->sched_count
);
3248 * Pick up the highest-prio task:
3250 static inline struct task_struct
*
3251 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
3253 const struct sched_class
*class = &fair_sched_class
;
3254 struct task_struct
*p
;
3257 * Optimization: we know that if all tasks are in
3258 * the fair class we can call that function directly:
3260 if (likely(prev
->sched_class
== class &&
3261 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3262 p
= fair_sched_class
.pick_next_task(rq
, prev
, cookie
);
3263 if (unlikely(p
== RETRY_TASK
))
3266 /* assumes fair_sched_class->next == idle_sched_class */
3268 p
= idle_sched_class
.pick_next_task(rq
, prev
, cookie
);
3274 for_each_class(class) {
3275 p
= class->pick_next_task(rq
, prev
, cookie
);
3277 if (unlikely(p
== RETRY_TASK
))
3283 BUG(); /* the idle class will always have a runnable task */
3287 * __schedule() is the main scheduler function.
3289 * The main means of driving the scheduler and thus entering this function are:
3291 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3293 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3294 * paths. For example, see arch/x86/entry_64.S.
3296 * To drive preemption between tasks, the scheduler sets the flag in timer
3297 * interrupt handler scheduler_tick().
3299 * 3. Wakeups don't really cause entry into schedule(). They add a
3300 * task to the run-queue and that's it.
3302 * Now, if the new task added to the run-queue preempts the current
3303 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3304 * called on the nearest possible occasion:
3306 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3308 * - in syscall or exception context, at the next outmost
3309 * preempt_enable(). (this might be as soon as the wake_up()'s
3312 * - in IRQ context, return from interrupt-handler to
3313 * preemptible context
3315 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3318 * - cond_resched() call
3319 * - explicit schedule() call
3320 * - return from syscall or exception to user-space
3321 * - return from interrupt-handler to user-space
3323 * WARNING: must be called with preemption disabled!
3325 static void __sched notrace
__schedule(bool preempt
)
3327 struct task_struct
*prev
, *next
;
3328 unsigned long *switch_count
;
3329 struct pin_cookie cookie
;
3333 cpu
= smp_processor_id();
3338 * do_exit() calls schedule() with preemption disabled as an exception;
3339 * however we must fix that up, otherwise the next task will see an
3340 * inconsistent (higher) preempt count.
3342 * It also avoids the below schedule_debug() test from complaining
3345 if (unlikely(prev
->state
== TASK_DEAD
))
3346 preempt_enable_no_resched_notrace();
3348 schedule_debug(prev
);
3350 if (sched_feat(HRTICK
))
3353 local_irq_disable();
3354 rcu_note_context_switch();
3357 * Make sure that signal_pending_state()->signal_pending() below
3358 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3359 * done by the caller to avoid the race with signal_wake_up().
3361 smp_mb__before_spinlock();
3362 raw_spin_lock(&rq
->lock
);
3363 cookie
= lockdep_pin_lock(&rq
->lock
);
3365 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3367 switch_count
= &prev
->nivcsw
;
3368 if (!preempt
&& prev
->state
) {
3369 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3370 prev
->state
= TASK_RUNNING
;
3372 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3376 * If a worker went to sleep, notify and ask workqueue
3377 * whether it wants to wake up a task to maintain
3380 if (prev
->flags
& PF_WQ_WORKER
) {
3381 struct task_struct
*to_wakeup
;
3383 to_wakeup
= wq_worker_sleeping(prev
);
3385 try_to_wake_up_local(to_wakeup
, cookie
);
3388 switch_count
= &prev
->nvcsw
;
3391 if (task_on_rq_queued(prev
))
3392 update_rq_clock(rq
);
3394 next
= pick_next_task(rq
, prev
, cookie
);
3395 clear_tsk_need_resched(prev
);
3396 clear_preempt_need_resched();
3397 rq
->clock_skip_update
= 0;
3399 if (likely(prev
!= next
)) {
3404 trace_sched_switch(preempt
, prev
, next
);
3405 rq
= context_switch(rq
, prev
, next
, cookie
); /* unlocks the rq */
3407 lockdep_unpin_lock(&rq
->lock
, cookie
);
3408 raw_spin_unlock_irq(&rq
->lock
);
3411 balance_callback(rq
);
3414 static inline void sched_submit_work(struct task_struct
*tsk
)
3416 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3419 * If we are going to sleep and we have plugged IO queued,
3420 * make sure to submit it to avoid deadlocks.
3422 if (blk_needs_flush_plug(tsk
))
3423 blk_schedule_flush_plug(tsk
);
3426 asmlinkage __visible
void __sched
schedule(void)
3428 struct task_struct
*tsk
= current
;
3430 sched_submit_work(tsk
);
3434 sched_preempt_enable_no_resched();
3435 } while (need_resched());
3437 EXPORT_SYMBOL(schedule
);
3439 #ifdef CONFIG_CONTEXT_TRACKING
3440 asmlinkage __visible
void __sched
schedule_user(void)
3443 * If we come here after a random call to set_need_resched(),
3444 * or we have been woken up remotely but the IPI has not yet arrived,
3445 * we haven't yet exited the RCU idle mode. Do it here manually until
3446 * we find a better solution.
3448 * NB: There are buggy callers of this function. Ideally we
3449 * should warn if prev_state != CONTEXT_USER, but that will trigger
3450 * too frequently to make sense yet.
3452 enum ctx_state prev_state
= exception_enter();
3454 exception_exit(prev_state
);
3459 * schedule_preempt_disabled - called with preemption disabled
3461 * Returns with preemption disabled. Note: preempt_count must be 1
3463 void __sched
schedule_preempt_disabled(void)
3465 sched_preempt_enable_no_resched();
3470 static void __sched notrace
preempt_schedule_common(void)
3474 * Because the function tracer can trace preempt_count_sub()
3475 * and it also uses preempt_enable/disable_notrace(), if
3476 * NEED_RESCHED is set, the preempt_enable_notrace() called
3477 * by the function tracer will call this function again and
3478 * cause infinite recursion.
3480 * Preemption must be disabled here before the function
3481 * tracer can trace. Break up preempt_disable() into two
3482 * calls. One to disable preemption without fear of being
3483 * traced. The other to still record the preemption latency,
3484 * which can also be traced by the function tracer.
3486 preempt_disable_notrace();
3487 preempt_latency_start(1);
3489 preempt_latency_stop(1);
3490 preempt_enable_no_resched_notrace();
3493 * Check again in case we missed a preemption opportunity
3494 * between schedule and now.
3496 } while (need_resched());
3499 #ifdef CONFIG_PREEMPT
3501 * this is the entry point to schedule() from in-kernel preemption
3502 * off of preempt_enable. Kernel preemptions off return from interrupt
3503 * occur there and call schedule directly.
3505 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3508 * If there is a non-zero preempt_count or interrupts are disabled,
3509 * we do not want to preempt the current task. Just return..
3511 if (likely(!preemptible()))
3514 preempt_schedule_common();
3516 NOKPROBE_SYMBOL(preempt_schedule
);
3517 EXPORT_SYMBOL(preempt_schedule
);
3520 * preempt_schedule_notrace - preempt_schedule called by tracing
3522 * The tracing infrastructure uses preempt_enable_notrace to prevent
3523 * recursion and tracing preempt enabling caused by the tracing
3524 * infrastructure itself. But as tracing can happen in areas coming
3525 * from userspace or just about to enter userspace, a preempt enable
3526 * can occur before user_exit() is called. This will cause the scheduler
3527 * to be called when the system is still in usermode.
3529 * To prevent this, the preempt_enable_notrace will use this function
3530 * instead of preempt_schedule() to exit user context if needed before
3531 * calling the scheduler.
3533 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3535 enum ctx_state prev_ctx
;
3537 if (likely(!preemptible()))
3542 * Because the function tracer can trace preempt_count_sub()
3543 * and it also uses preempt_enable/disable_notrace(), if
3544 * NEED_RESCHED is set, the preempt_enable_notrace() called
3545 * by the function tracer will call this function again and
3546 * cause infinite recursion.
3548 * Preemption must be disabled here before the function
3549 * tracer can trace. Break up preempt_disable() into two
3550 * calls. One to disable preemption without fear of being
3551 * traced. The other to still record the preemption latency,
3552 * which can also be traced by the function tracer.
3554 preempt_disable_notrace();
3555 preempt_latency_start(1);
3557 * Needs preempt disabled in case user_exit() is traced
3558 * and the tracer calls preempt_enable_notrace() causing
3559 * an infinite recursion.
3561 prev_ctx
= exception_enter();
3563 exception_exit(prev_ctx
);
3565 preempt_latency_stop(1);
3566 preempt_enable_no_resched_notrace();
3567 } while (need_resched());
3569 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3571 #endif /* CONFIG_PREEMPT */
3574 * this is the entry point to schedule() from kernel preemption
3575 * off of irq context.
3576 * Note, that this is called and return with irqs disabled. This will
3577 * protect us against recursive calling from irq.
3579 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3581 enum ctx_state prev_state
;
3583 /* Catch callers which need to be fixed */
3584 BUG_ON(preempt_count() || !irqs_disabled());
3586 prev_state
= exception_enter();
3592 local_irq_disable();
3593 sched_preempt_enable_no_resched();
3594 } while (need_resched());
3596 exception_exit(prev_state
);
3599 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3602 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3604 EXPORT_SYMBOL(default_wake_function
);
3606 #ifdef CONFIG_RT_MUTEXES
3609 * rt_mutex_setprio - set the current priority of a task
3611 * @prio: prio value (kernel-internal form)
3613 * This function changes the 'effective' priority of a task. It does
3614 * not touch ->normal_prio like __setscheduler().
3616 * Used by the rt_mutex code to implement priority inheritance
3617 * logic. Call site only calls if the priority of the task changed.
3619 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3621 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3622 const struct sched_class
*prev_class
;
3626 BUG_ON(prio
> MAX_PRIO
);
3628 rq
= __task_rq_lock(p
, &rf
);
3631 * Idle task boosting is a nono in general. There is one
3632 * exception, when PREEMPT_RT and NOHZ is active:
3634 * The idle task calls get_next_timer_interrupt() and holds
3635 * the timer wheel base->lock on the CPU and another CPU wants
3636 * to access the timer (probably to cancel it). We can safely
3637 * ignore the boosting request, as the idle CPU runs this code
3638 * with interrupts disabled and will complete the lock
3639 * protected section without being interrupted. So there is no
3640 * real need to boost.
3642 if (unlikely(p
== rq
->idle
)) {
3643 WARN_ON(p
!= rq
->curr
);
3644 WARN_ON(p
->pi_blocked_on
);
3648 trace_sched_pi_setprio(p
, prio
);
3651 if (oldprio
== prio
)
3652 queue_flag
&= ~DEQUEUE_MOVE
;
3654 prev_class
= p
->sched_class
;
3655 queued
= task_on_rq_queued(p
);
3656 running
= task_current(rq
, p
);
3658 dequeue_task(rq
, p
, queue_flag
);
3660 put_prev_task(rq
, p
);
3663 * Boosting condition are:
3664 * 1. -rt task is running and holds mutex A
3665 * --> -dl task blocks on mutex A
3667 * 2. -dl task is running and holds mutex A
3668 * --> -dl task blocks on mutex A and could preempt the
3671 if (dl_prio(prio
)) {
3672 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3673 if (!dl_prio(p
->normal_prio
) ||
3674 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3675 p
->dl
.dl_boosted
= 1;
3676 queue_flag
|= ENQUEUE_REPLENISH
;
3678 p
->dl
.dl_boosted
= 0;
3679 p
->sched_class
= &dl_sched_class
;
3680 } else if (rt_prio(prio
)) {
3681 if (dl_prio(oldprio
))
3682 p
->dl
.dl_boosted
= 0;
3684 queue_flag
|= ENQUEUE_HEAD
;
3685 p
->sched_class
= &rt_sched_class
;
3687 if (dl_prio(oldprio
))
3688 p
->dl
.dl_boosted
= 0;
3689 if (rt_prio(oldprio
))
3691 p
->sched_class
= &fair_sched_class
;
3697 p
->sched_class
->set_curr_task(rq
);
3699 enqueue_task(rq
, p
, queue_flag
);
3701 check_class_changed(rq
, p
, prev_class
, oldprio
);
3703 preempt_disable(); /* avoid rq from going away on us */
3704 __task_rq_unlock(rq
, &rf
);
3706 balance_callback(rq
);
3711 void set_user_nice(struct task_struct
*p
, long nice
)
3713 int old_prio
, delta
, queued
;
3717 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3720 * We have to be careful, if called from sys_setpriority(),
3721 * the task might be in the middle of scheduling on another CPU.
3723 rq
= task_rq_lock(p
, &rf
);
3725 * The RT priorities are set via sched_setscheduler(), but we still
3726 * allow the 'normal' nice value to be set - but as expected
3727 * it wont have any effect on scheduling until the task is
3728 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3730 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3731 p
->static_prio
= NICE_TO_PRIO(nice
);
3734 queued
= task_on_rq_queued(p
);
3736 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3738 p
->static_prio
= NICE_TO_PRIO(nice
);
3741 p
->prio
= effective_prio(p
);
3742 delta
= p
->prio
- old_prio
;
3745 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3747 * If the task increased its priority or is running and
3748 * lowered its priority, then reschedule its CPU:
3750 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3754 trace_sched_update_prio(p
);
3755 task_rq_unlock(rq
, p
, &rf
);
3757 EXPORT_SYMBOL(set_user_nice
);
3760 * can_nice - check if a task can reduce its nice value
3764 int can_nice(const struct task_struct
*p
, const int nice
)
3766 /* convert nice value [19,-20] to rlimit style value [1,40] */
3767 int nice_rlim
= nice_to_rlimit(nice
);
3769 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3770 capable(CAP_SYS_NICE
));
3773 #ifdef __ARCH_WANT_SYS_NICE
3776 * sys_nice - change the priority of the current process.
3777 * @increment: priority increment
3779 * sys_setpriority is a more generic, but much slower function that
3780 * does similar things.
3782 SYSCALL_DEFINE1(nice
, int, increment
)
3787 * Setpriority might change our priority at the same moment.
3788 * We don't have to worry. Conceptually one call occurs first
3789 * and we have a single winner.
3791 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3792 nice
= task_nice(current
) + increment
;
3794 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3795 if (increment
< 0 && !can_nice(current
, nice
))
3798 retval
= security_task_setnice(current
, nice
);
3802 set_user_nice(current
, nice
);
3809 * task_prio - return the priority value of a given task.
3810 * @p: the task in question.
3812 * Return: The priority value as seen by users in /proc.
3813 * RT tasks are offset by -200. Normal tasks are centered
3814 * around 0, value goes from -16 to +15.
3816 int task_prio(const struct task_struct
*p
)
3818 return p
->prio
- MAX_RT_PRIO
;
3822 * idle_cpu - is a given cpu idle currently?
3823 * @cpu: the processor in question.
3825 * Return: 1 if the CPU is currently idle. 0 otherwise.
3827 int idle_cpu(int cpu
)
3829 struct rq
*rq
= cpu_rq(cpu
);
3831 if (rq
->curr
!= rq
->idle
)
3838 if (!llist_empty(&rq
->wake_list
))
3846 * idle_task - return the idle task for a given cpu.
3847 * @cpu: the processor in question.
3849 * Return: The idle task for the cpu @cpu.
3851 struct task_struct
*idle_task(int cpu
)
3853 return cpu_rq(cpu
)->idle
;
3857 * find_process_by_pid - find a process with a matching PID value.
3858 * @pid: the pid in question.
3860 * The task of @pid, if found. %NULL otherwise.
3862 static struct task_struct
*find_process_by_pid(pid_t pid
)
3864 return pid
? find_task_by_vpid(pid
) : current
;
3868 * This function initializes the sched_dl_entity of a newly becoming
3869 * SCHED_DEADLINE task.
3871 * Only the static values are considered here, the actual runtime and the
3872 * absolute deadline will be properly calculated when the task is enqueued
3873 * for the first time with its new policy.
3876 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3878 struct sched_dl_entity
*dl_se
= &p
->dl
;
3880 dl_se
->dl_runtime
= attr
->sched_runtime
;
3881 dl_se
->dl_deadline
= attr
->sched_deadline
;
3882 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3883 dl_se
->flags
= attr
->sched_flags
;
3884 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3887 * Changing the parameters of a task is 'tricky' and we're not doing
3888 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3890 * What we SHOULD do is delay the bandwidth release until the 0-lag
3891 * point. This would include retaining the task_struct until that time
3892 * and change dl_overflow() to not immediately decrement the current
3895 * Instead we retain the current runtime/deadline and let the new
3896 * parameters take effect after the current reservation period lapses.
3897 * This is safe (albeit pessimistic) because the 0-lag point is always
3898 * before the current scheduling deadline.
3900 * We can still have temporary overloads because we do not delay the
3901 * change in bandwidth until that time; so admission control is
3902 * not on the safe side. It does however guarantee tasks will never
3903 * consume more than promised.
3908 * sched_setparam() passes in -1 for its policy, to let the functions
3909 * it calls know not to change it.
3911 #define SETPARAM_POLICY -1
3913 static void __setscheduler_params(struct task_struct
*p
,
3914 const struct sched_attr
*attr
)
3916 int policy
= attr
->sched_policy
;
3918 if (policy
== SETPARAM_POLICY
)
3923 if (dl_policy(policy
))
3924 __setparam_dl(p
, attr
);
3925 else if (fair_policy(policy
))
3926 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3929 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3930 * !rt_policy. Always setting this ensures that things like
3931 * getparam()/getattr() don't report silly values for !rt tasks.
3933 p
->rt_priority
= attr
->sched_priority
;
3934 p
->normal_prio
= normal_prio(p
);
3938 /* Actually do priority change: must hold pi & rq lock. */
3939 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3940 const struct sched_attr
*attr
, bool keep_boost
)
3942 __setscheduler_params(p
, attr
);
3945 * Keep a potential priority boosting if called from
3946 * sched_setscheduler().
3949 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3951 p
->prio
= normal_prio(p
);
3953 if (dl_prio(p
->prio
))
3954 p
->sched_class
= &dl_sched_class
;
3955 else if (rt_prio(p
->prio
))
3956 p
->sched_class
= &rt_sched_class
;
3958 p
->sched_class
= &fair_sched_class
;
3960 trace_sched_update_prio(p
);
3964 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3966 struct sched_dl_entity
*dl_se
= &p
->dl
;
3968 attr
->sched_priority
= p
->rt_priority
;
3969 attr
->sched_runtime
= dl_se
->dl_runtime
;
3970 attr
->sched_deadline
= dl_se
->dl_deadline
;
3971 attr
->sched_period
= dl_se
->dl_period
;
3972 attr
->sched_flags
= dl_se
->flags
;
3976 * This function validates the new parameters of a -deadline task.
3977 * We ask for the deadline not being zero, and greater or equal
3978 * than the runtime, as well as the period of being zero or
3979 * greater than deadline. Furthermore, we have to be sure that
3980 * user parameters are above the internal resolution of 1us (we
3981 * check sched_runtime only since it is always the smaller one) and
3982 * below 2^63 ns (we have to check both sched_deadline and
3983 * sched_period, as the latter can be zero).
3986 __checkparam_dl(const struct sched_attr
*attr
)
3989 if (attr
->sched_deadline
== 0)
3993 * Since we truncate DL_SCALE bits, make sure we're at least
3996 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
4000 * Since we use the MSB for wrap-around and sign issues, make
4001 * sure it's not set (mind that period can be equal to zero).
4003 if (attr
->sched_deadline
& (1ULL << 63) ||
4004 attr
->sched_period
& (1ULL << 63))
4007 /* runtime <= deadline <= period (if period != 0) */
4008 if ((attr
->sched_period
!= 0 &&
4009 attr
->sched_period
< attr
->sched_deadline
) ||
4010 attr
->sched_deadline
< attr
->sched_runtime
)
4017 * check the target process has a UID that matches the current process's
4019 static bool check_same_owner(struct task_struct
*p
)
4021 const struct cred
*cred
= current_cred(), *pcred
;
4025 pcred
= __task_cred(p
);
4026 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4027 uid_eq(cred
->euid
, pcred
->uid
));
4032 static bool dl_param_changed(struct task_struct
*p
,
4033 const struct sched_attr
*attr
)
4035 struct sched_dl_entity
*dl_se
= &p
->dl
;
4037 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
4038 dl_se
->dl_deadline
!= attr
->sched_deadline
||
4039 dl_se
->dl_period
!= attr
->sched_period
||
4040 dl_se
->flags
!= attr
->sched_flags
)
4046 static int __sched_setscheduler(struct task_struct
*p
,
4047 const struct sched_attr
*attr
,
4050 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4051 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4052 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4053 int new_effective_prio
, policy
= attr
->sched_policy
;
4054 const struct sched_class
*prev_class
;
4057 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
4060 /* may grab non-irq protected spin_locks */
4061 BUG_ON(in_interrupt());
4063 /* double check policy once rq lock held */
4065 reset_on_fork
= p
->sched_reset_on_fork
;
4066 policy
= oldpolicy
= p
->policy
;
4068 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4070 if (!valid_policy(policy
))
4074 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4078 * Valid priorities for SCHED_FIFO and SCHED_RR are
4079 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4080 * SCHED_BATCH and SCHED_IDLE is 0.
4082 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4083 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4085 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4086 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4090 * Allow unprivileged RT tasks to decrease priority:
4092 if (user
&& !capable(CAP_SYS_NICE
)) {
4093 if (fair_policy(policy
)) {
4094 if (attr
->sched_nice
< task_nice(p
) &&
4095 !can_nice(p
, attr
->sched_nice
))
4099 if (rt_policy(policy
)) {
4100 unsigned long rlim_rtprio
=
4101 task_rlimit(p
, RLIMIT_RTPRIO
);
4103 /* can't set/change the rt policy */
4104 if (policy
!= p
->policy
&& !rlim_rtprio
)
4107 /* can't increase priority */
4108 if (attr
->sched_priority
> p
->rt_priority
&&
4109 attr
->sched_priority
> rlim_rtprio
)
4114 * Can't set/change SCHED_DEADLINE policy at all for now
4115 * (safest behavior); in the future we would like to allow
4116 * unprivileged DL tasks to increase their relative deadline
4117 * or reduce their runtime (both ways reducing utilization)
4119 if (dl_policy(policy
))
4123 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4124 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4126 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4127 if (!can_nice(p
, task_nice(p
)))
4131 /* can't change other user's priorities */
4132 if (!check_same_owner(p
))
4135 /* Normal users shall not reset the sched_reset_on_fork flag */
4136 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4141 retval
= security_task_setscheduler(p
);
4147 * make sure no PI-waiters arrive (or leave) while we are
4148 * changing the priority of the task:
4150 * To be able to change p->policy safely, the appropriate
4151 * runqueue lock must be held.
4153 rq
= task_rq_lock(p
, &rf
);
4156 * Changing the policy of the stop threads its a very bad idea
4158 if (p
== rq
->stop
) {
4159 task_rq_unlock(rq
, p
, &rf
);
4164 * If not changing anything there's no need to proceed further,
4165 * but store a possible modification of reset_on_fork.
4167 if (unlikely(policy
== p
->policy
)) {
4168 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4170 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4172 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4175 p
->sched_reset_on_fork
= reset_on_fork
;
4176 task_rq_unlock(rq
, p
, &rf
);
4182 #ifdef CONFIG_RT_GROUP_SCHED
4184 * Do not allow realtime tasks into groups that have no runtime
4187 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4188 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4189 !task_group_is_autogroup(task_group(p
))) {
4190 task_rq_unlock(rq
, p
, &rf
);
4195 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4196 cpumask_t
*span
= rq
->rd
->span
;
4199 * Don't allow tasks with an affinity mask smaller than
4200 * the entire root_domain to become SCHED_DEADLINE. We
4201 * will also fail if there's no bandwidth available.
4203 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4204 rq
->rd
->dl_bw
.bw
== 0) {
4205 task_rq_unlock(rq
, p
, &rf
);
4212 /* recheck policy now with rq lock held */
4213 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4214 policy
= oldpolicy
= -1;
4215 task_rq_unlock(rq
, p
, &rf
);
4220 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4221 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4224 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4225 task_rq_unlock(rq
, p
, &rf
);
4229 p
->sched_reset_on_fork
= reset_on_fork
;
4234 * Take priority boosted tasks into account. If the new
4235 * effective priority is unchanged, we just store the new
4236 * normal parameters and do not touch the scheduler class and
4237 * the runqueue. This will be done when the task deboost
4240 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4241 if (new_effective_prio
== oldprio
)
4242 queue_flags
&= ~DEQUEUE_MOVE
;
4245 queued
= task_on_rq_queued(p
);
4246 running
= task_current(rq
, p
);
4248 dequeue_task(rq
, p
, queue_flags
);
4250 put_prev_task(rq
, p
);
4252 prev_class
= p
->sched_class
;
4253 __setscheduler(rq
, p
, attr
, pi
);
4256 p
->sched_class
->set_curr_task(rq
);
4259 * We enqueue to tail when the priority of a task is
4260 * increased (user space view).
4262 if (oldprio
< p
->prio
)
4263 queue_flags
|= ENQUEUE_HEAD
;
4265 enqueue_task(rq
, p
, queue_flags
);
4268 check_class_changed(rq
, p
, prev_class
, oldprio
);
4269 preempt_disable(); /* avoid rq from going away on us */
4270 task_rq_unlock(rq
, p
, &rf
);
4273 rt_mutex_adjust_pi(p
);
4276 * Run balance callbacks after we've adjusted the PI chain.
4278 balance_callback(rq
);
4284 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4285 const struct sched_param
*param
, bool check
)
4287 struct sched_attr attr
= {
4288 .sched_policy
= policy
,
4289 .sched_priority
= param
->sched_priority
,
4290 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4293 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4294 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4295 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4296 policy
&= ~SCHED_RESET_ON_FORK
;
4297 attr
.sched_policy
= policy
;
4300 return __sched_setscheduler(p
, &attr
, check
, true);
4303 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4304 * @p: the task in question.
4305 * @policy: new policy.
4306 * @param: structure containing the new RT priority.
4308 * Return: 0 on success. An error code otherwise.
4310 * NOTE that the task may be already dead.
4312 int sched_setscheduler(struct task_struct
*p
, int policy
,
4313 const struct sched_param
*param
)
4315 return _sched_setscheduler(p
, policy
, param
, true);
4317 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4319 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4321 return __sched_setscheduler(p
, attr
, true, true);
4323 EXPORT_SYMBOL_GPL(sched_setattr
);
4326 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4327 * @p: the task in question.
4328 * @policy: new policy.
4329 * @param: structure containing the new RT priority.
4331 * Just like sched_setscheduler, only don't bother checking if the
4332 * current context has permission. For example, this is needed in
4333 * stop_machine(): we create temporary high priority worker threads,
4334 * but our caller might not have that capability.
4336 * Return: 0 on success. An error code otherwise.
4338 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4339 const struct sched_param
*param
)
4341 return _sched_setscheduler(p
, policy
, param
, false);
4343 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4346 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4348 struct sched_param lparam
;
4349 struct task_struct
*p
;
4352 if (!param
|| pid
< 0)
4354 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4359 p
= find_process_by_pid(pid
);
4361 retval
= sched_setscheduler(p
, policy
, &lparam
);
4368 * Mimics kernel/events/core.c perf_copy_attr().
4370 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4371 struct sched_attr
*attr
)
4376 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4380 * zero the full structure, so that a short copy will be nice.
4382 memset(attr
, 0, sizeof(*attr
));
4384 ret
= get_user(size
, &uattr
->size
);
4388 if (size
> PAGE_SIZE
) /* silly large */
4391 if (!size
) /* abi compat */
4392 size
= SCHED_ATTR_SIZE_VER0
;
4394 if (size
< SCHED_ATTR_SIZE_VER0
)
4398 * If we're handed a bigger struct than we know of,
4399 * ensure all the unknown bits are 0 - i.e. new
4400 * user-space does not rely on any kernel feature
4401 * extensions we dont know about yet.
4403 if (size
> sizeof(*attr
)) {
4404 unsigned char __user
*addr
;
4405 unsigned char __user
*end
;
4408 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4409 end
= (void __user
*)uattr
+ size
;
4411 for (; addr
< end
; addr
++) {
4412 ret
= get_user(val
, addr
);
4418 size
= sizeof(*attr
);
4421 ret
= copy_from_user(attr
, uattr
, size
);
4426 * XXX: do we want to be lenient like existing syscalls; or do we want
4427 * to be strict and return an error on out-of-bounds values?
4429 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4434 put_user(sizeof(*attr
), &uattr
->size
);
4439 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4440 * @pid: the pid in question.
4441 * @policy: new policy.
4442 * @param: structure containing the new RT priority.
4444 * Return: 0 on success. An error code otherwise.
4446 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4447 struct sched_param __user
*, param
)
4449 /* negative values for policy are not valid */
4453 return do_sched_setscheduler(pid
, policy
, param
);
4457 * sys_sched_setparam - set/change the RT priority of a thread
4458 * @pid: the pid in question.
4459 * @param: structure containing the new RT priority.
4461 * Return: 0 on success. An error code otherwise.
4463 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4465 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4469 * sys_sched_setattr - same as above, but with extended sched_attr
4470 * @pid: the pid in question.
4471 * @uattr: structure containing the extended parameters.
4472 * @flags: for future extension.
4474 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4475 unsigned int, flags
)
4477 struct sched_attr attr
;
4478 struct task_struct
*p
;
4481 if (!uattr
|| pid
< 0 || flags
)
4484 retval
= sched_copy_attr(uattr
, &attr
);
4488 if ((int)attr
.sched_policy
< 0)
4493 p
= find_process_by_pid(pid
);
4495 retval
= sched_setattr(p
, &attr
);
4502 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4503 * @pid: the pid in question.
4505 * Return: On success, the policy of the thread. Otherwise, a negative error
4508 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4510 struct task_struct
*p
;
4518 p
= find_process_by_pid(pid
);
4520 retval
= security_task_getscheduler(p
);
4523 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4530 * sys_sched_getparam - get the RT priority of a thread
4531 * @pid: the pid in question.
4532 * @param: structure containing the RT priority.
4534 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4537 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4539 struct sched_param lp
= { .sched_priority
= 0 };
4540 struct task_struct
*p
;
4543 if (!param
|| pid
< 0)
4547 p
= find_process_by_pid(pid
);
4552 retval
= security_task_getscheduler(p
);
4556 if (task_has_rt_policy(p
))
4557 lp
.sched_priority
= p
->rt_priority
;
4561 * This one might sleep, we cannot do it with a spinlock held ...
4563 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4572 static int sched_read_attr(struct sched_attr __user
*uattr
,
4573 struct sched_attr
*attr
,
4578 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4582 * If we're handed a smaller struct than we know of,
4583 * ensure all the unknown bits are 0 - i.e. old
4584 * user-space does not get uncomplete information.
4586 if (usize
< sizeof(*attr
)) {
4587 unsigned char *addr
;
4590 addr
= (void *)attr
+ usize
;
4591 end
= (void *)attr
+ sizeof(*attr
);
4593 for (; addr
< end
; addr
++) {
4601 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4609 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4610 * @pid: the pid in question.
4611 * @uattr: structure containing the extended parameters.
4612 * @size: sizeof(attr) for fwd/bwd comp.
4613 * @flags: for future extension.
4615 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4616 unsigned int, size
, unsigned int, flags
)
4618 struct sched_attr attr
= {
4619 .size
= sizeof(struct sched_attr
),
4621 struct task_struct
*p
;
4624 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4625 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4629 p
= find_process_by_pid(pid
);
4634 retval
= security_task_getscheduler(p
);
4638 attr
.sched_policy
= p
->policy
;
4639 if (p
->sched_reset_on_fork
)
4640 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4641 if (task_has_dl_policy(p
))
4642 __getparam_dl(p
, &attr
);
4643 else if (task_has_rt_policy(p
))
4644 attr
.sched_priority
= p
->rt_priority
;
4646 attr
.sched_nice
= task_nice(p
);
4650 retval
= sched_read_attr(uattr
, &attr
, size
);
4658 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4660 cpumask_var_t cpus_allowed
, new_mask
;
4661 struct task_struct
*p
;
4666 p
= find_process_by_pid(pid
);
4672 /* Prevent p going away */
4676 if (p
->flags
& PF_NO_SETAFFINITY
) {
4680 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4684 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4686 goto out_free_cpus_allowed
;
4689 if (!check_same_owner(p
)) {
4691 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4693 goto out_free_new_mask
;
4698 retval
= security_task_setscheduler(p
);
4700 goto out_free_new_mask
;
4703 cpuset_cpus_allowed(p
, cpus_allowed
);
4704 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4707 * Since bandwidth control happens on root_domain basis,
4708 * if admission test is enabled, we only admit -deadline
4709 * tasks allowed to run on all the CPUs in the task's
4713 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4715 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4718 goto out_free_new_mask
;
4724 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4727 cpuset_cpus_allowed(p
, cpus_allowed
);
4728 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4730 * We must have raced with a concurrent cpuset
4731 * update. Just reset the cpus_allowed to the
4732 * cpuset's cpus_allowed
4734 cpumask_copy(new_mask
, cpus_allowed
);
4739 free_cpumask_var(new_mask
);
4740 out_free_cpus_allowed
:
4741 free_cpumask_var(cpus_allowed
);
4747 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4748 struct cpumask
*new_mask
)
4750 if (len
< cpumask_size())
4751 cpumask_clear(new_mask
);
4752 else if (len
> cpumask_size())
4753 len
= cpumask_size();
4755 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4759 * sys_sched_setaffinity - set the cpu affinity of a process
4760 * @pid: pid of the process
4761 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4762 * @user_mask_ptr: user-space pointer to the new cpu mask
4764 * Return: 0 on success. An error code otherwise.
4766 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4767 unsigned long __user
*, user_mask_ptr
)
4769 cpumask_var_t new_mask
;
4772 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4775 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4777 retval
= sched_setaffinity(pid
, new_mask
);
4778 free_cpumask_var(new_mask
);
4782 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4784 struct task_struct
*p
;
4785 unsigned long flags
;
4791 p
= find_process_by_pid(pid
);
4795 retval
= security_task_getscheduler(p
);
4799 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4800 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4801 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4810 * sys_sched_getaffinity - get the cpu affinity of a process
4811 * @pid: pid of the process
4812 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4813 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4815 * Return: size of CPU mask copied to user_mask_ptr on success. An
4816 * error code otherwise.
4818 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4819 unsigned long __user
*, user_mask_ptr
)
4824 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4826 if (len
& (sizeof(unsigned long)-1))
4829 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4832 ret
= sched_getaffinity(pid
, mask
);
4834 size_t retlen
= min_t(size_t, len
, cpumask_size());
4836 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4841 free_cpumask_var(mask
);
4847 * sys_sched_yield - yield the current processor to other threads.
4849 * This function yields the current CPU to other tasks. If there are no
4850 * other threads running on this CPU then this function will return.
4854 SYSCALL_DEFINE0(sched_yield
)
4856 struct rq
*rq
= this_rq_lock();
4858 schedstat_inc(rq
->yld_count
);
4859 current
->sched_class
->yield_task(rq
);
4862 * Since we are going to call schedule() anyway, there's
4863 * no need to preempt or enable interrupts:
4865 __release(rq
->lock
);
4866 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4867 do_raw_spin_unlock(&rq
->lock
);
4868 sched_preempt_enable_no_resched();
4875 int __sched
_cond_resched(void)
4877 if (should_resched(0)) {
4878 preempt_schedule_common();
4883 EXPORT_SYMBOL(_cond_resched
);
4886 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4887 * call schedule, and on return reacquire the lock.
4889 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4890 * operations here to prevent schedule() from being called twice (once via
4891 * spin_unlock(), once by hand).
4893 int __cond_resched_lock(spinlock_t
*lock
)
4895 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4898 lockdep_assert_held(lock
);
4900 if (spin_needbreak(lock
) || resched
) {
4903 preempt_schedule_common();
4911 EXPORT_SYMBOL(__cond_resched_lock
);
4913 int __sched
__cond_resched_softirq(void)
4915 BUG_ON(!in_softirq());
4917 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4919 preempt_schedule_common();
4925 EXPORT_SYMBOL(__cond_resched_softirq
);
4928 * yield - yield the current processor to other threads.
4930 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4932 * The scheduler is at all times free to pick the calling task as the most
4933 * eligible task to run, if removing the yield() call from your code breaks
4934 * it, its already broken.
4936 * Typical broken usage is:
4941 * where one assumes that yield() will let 'the other' process run that will
4942 * make event true. If the current task is a SCHED_FIFO task that will never
4943 * happen. Never use yield() as a progress guarantee!!
4945 * If you want to use yield() to wait for something, use wait_event().
4946 * If you want to use yield() to be 'nice' for others, use cond_resched().
4947 * If you still want to use yield(), do not!
4949 void __sched
yield(void)
4951 set_current_state(TASK_RUNNING
);
4954 EXPORT_SYMBOL(yield
);
4957 * yield_to - yield the current processor to another thread in
4958 * your thread group, or accelerate that thread toward the
4959 * processor it's on.
4961 * @preempt: whether task preemption is allowed or not
4963 * It's the caller's job to ensure that the target task struct
4964 * can't go away on us before we can do any checks.
4967 * true (>0) if we indeed boosted the target task.
4968 * false (0) if we failed to boost the target.
4969 * -ESRCH if there's no task to yield to.
4971 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4973 struct task_struct
*curr
= current
;
4974 struct rq
*rq
, *p_rq
;
4975 unsigned long flags
;
4978 local_irq_save(flags
);
4984 * If we're the only runnable task on the rq and target rq also
4985 * has only one task, there's absolutely no point in yielding.
4987 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4992 double_rq_lock(rq
, p_rq
);
4993 if (task_rq(p
) != p_rq
) {
4994 double_rq_unlock(rq
, p_rq
);
4998 if (!curr
->sched_class
->yield_to_task
)
5001 if (curr
->sched_class
!= p
->sched_class
)
5004 if (task_running(p_rq
, p
) || p
->state
)
5007 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5009 schedstat_inc(rq
->yld_count
);
5011 * Make p's CPU reschedule; pick_next_entity takes care of
5014 if (preempt
&& rq
!= p_rq
)
5019 double_rq_unlock(rq
, p_rq
);
5021 local_irq_restore(flags
);
5028 EXPORT_SYMBOL_GPL(yield_to
);
5031 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5032 * that process accounting knows that this is a task in IO wait state.
5034 long __sched
io_schedule_timeout(long timeout
)
5036 int old_iowait
= current
->in_iowait
;
5040 current
->in_iowait
= 1;
5041 blk_schedule_flush_plug(current
);
5043 delayacct_blkio_start();
5045 atomic_inc(&rq
->nr_iowait
);
5046 ret
= schedule_timeout(timeout
);
5047 current
->in_iowait
= old_iowait
;
5048 atomic_dec(&rq
->nr_iowait
);
5049 delayacct_blkio_end();
5053 EXPORT_SYMBOL(io_schedule_timeout
);
5056 * sys_sched_get_priority_max - return maximum RT priority.
5057 * @policy: scheduling class.
5059 * Return: On success, this syscall returns the maximum
5060 * rt_priority that can be used by a given scheduling class.
5061 * On failure, a negative error code is returned.
5063 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5070 ret
= MAX_USER_RT_PRIO
-1;
5072 case SCHED_DEADLINE
:
5083 * sys_sched_get_priority_min - return minimum RT priority.
5084 * @policy: scheduling class.
5086 * Return: On success, this syscall returns the minimum
5087 * rt_priority that can be used by a given scheduling class.
5088 * On failure, a negative error code is returned.
5090 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5099 case SCHED_DEADLINE
:
5109 * sys_sched_rr_get_interval - return the default timeslice of a process.
5110 * @pid: pid of the process.
5111 * @interval: userspace pointer to the timeslice value.
5113 * this syscall writes the default timeslice value of a given process
5114 * into the user-space timespec buffer. A value of '0' means infinity.
5116 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5119 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5120 struct timespec __user
*, interval
)
5122 struct task_struct
*p
;
5123 unsigned int time_slice
;
5134 p
= find_process_by_pid(pid
);
5138 retval
= security_task_getscheduler(p
);
5142 rq
= task_rq_lock(p
, &rf
);
5144 if (p
->sched_class
->get_rr_interval
)
5145 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5146 task_rq_unlock(rq
, p
, &rf
);
5149 jiffies_to_timespec(time_slice
, &t
);
5150 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5158 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5160 void sched_show_task(struct task_struct
*p
)
5162 unsigned long free
= 0;
5164 unsigned long state
= p
->state
;
5167 state
= __ffs(state
) + 1;
5168 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5169 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5170 #if BITS_PER_LONG == 32
5171 if (state
== TASK_RUNNING
)
5172 printk(KERN_CONT
" running ");
5174 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5176 if (state
== TASK_RUNNING
)
5177 printk(KERN_CONT
" running task ");
5179 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5181 #ifdef CONFIG_DEBUG_STACK_USAGE
5182 free
= stack_not_used(p
);
5187 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5189 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5190 task_pid_nr(p
), ppid
,
5191 (unsigned long)task_thread_info(p
)->flags
);
5193 print_worker_info(KERN_INFO
, p
);
5194 show_stack(p
, NULL
);
5197 void show_state_filter(unsigned long state_filter
)
5199 struct task_struct
*g
, *p
;
5201 #if BITS_PER_LONG == 32
5203 " task PC stack pid father\n");
5206 " task PC stack pid father\n");
5209 for_each_process_thread(g
, p
) {
5211 * reset the NMI-timeout, listing all files on a slow
5212 * console might take a lot of time:
5213 * Also, reset softlockup watchdogs on all CPUs, because
5214 * another CPU might be blocked waiting for us to process
5217 touch_nmi_watchdog();
5218 touch_all_softlockup_watchdogs();
5219 if (!state_filter
|| (p
->state
& state_filter
))
5223 #ifdef CONFIG_SCHED_DEBUG
5225 sysrq_sched_debug_show();
5229 * Only show locks if all tasks are dumped:
5232 debug_show_all_locks();
5235 void init_idle_bootup_task(struct task_struct
*idle
)
5237 idle
->sched_class
= &idle_sched_class
;
5241 * init_idle - set up an idle thread for a given CPU
5242 * @idle: task in question
5243 * @cpu: cpu the idle task belongs to
5245 * NOTE: this function does not set the idle thread's NEED_RESCHED
5246 * flag, to make booting more robust.
5248 void init_idle(struct task_struct
*idle
, int cpu
)
5250 struct rq
*rq
= cpu_rq(cpu
);
5251 unsigned long flags
;
5253 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5254 raw_spin_lock(&rq
->lock
);
5256 __sched_fork(0, idle
);
5257 idle
->state
= TASK_RUNNING
;
5258 idle
->se
.exec_start
= sched_clock();
5260 kasan_unpoison_task_stack(idle
);
5264 * Its possible that init_idle() gets called multiple times on a task,
5265 * in that case do_set_cpus_allowed() will not do the right thing.
5267 * And since this is boot we can forgo the serialization.
5269 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5272 * We're having a chicken and egg problem, even though we are
5273 * holding rq->lock, the cpu isn't yet set to this cpu so the
5274 * lockdep check in task_group() will fail.
5276 * Similar case to sched_fork(). / Alternatively we could
5277 * use task_rq_lock() here and obtain the other rq->lock.
5282 __set_task_cpu(idle
, cpu
);
5285 rq
->curr
= rq
->idle
= idle
;
5286 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5290 raw_spin_unlock(&rq
->lock
);
5291 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5293 /* Set the preempt count _outside_ the spinlocks! */
5294 init_idle_preempt_count(idle
, cpu
);
5297 * The idle tasks have their own, simple scheduling class:
5299 idle
->sched_class
= &idle_sched_class
;
5300 ftrace_graph_init_idle_task(idle
, cpu
);
5301 vtime_init_idle(idle
, cpu
);
5303 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5307 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5308 const struct cpumask
*trial
)
5310 int ret
= 1, trial_cpus
;
5311 struct dl_bw
*cur_dl_b
;
5312 unsigned long flags
;
5314 if (!cpumask_weight(cur
))
5317 rcu_read_lock_sched();
5318 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5319 trial_cpus
= cpumask_weight(trial
);
5321 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5322 if (cur_dl_b
->bw
!= -1 &&
5323 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5325 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5326 rcu_read_unlock_sched();
5331 int task_can_attach(struct task_struct
*p
,
5332 const struct cpumask
*cs_cpus_allowed
)
5337 * Kthreads which disallow setaffinity shouldn't be moved
5338 * to a new cpuset; we don't want to change their cpu
5339 * affinity and isolating such threads by their set of
5340 * allowed nodes is unnecessary. Thus, cpusets are not
5341 * applicable for such threads. This prevents checking for
5342 * success of set_cpus_allowed_ptr() on all attached tasks
5343 * before cpus_allowed may be changed.
5345 if (p
->flags
& PF_NO_SETAFFINITY
) {
5351 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5353 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5358 unsigned long flags
;
5360 rcu_read_lock_sched();
5361 dl_b
= dl_bw_of(dest_cpu
);
5362 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5363 cpus
= dl_bw_cpus(dest_cpu
);
5364 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5369 * We reserve space for this task in the destination
5370 * root_domain, as we can't fail after this point.
5371 * We will free resources in the source root_domain
5372 * later on (see set_cpus_allowed_dl()).
5374 __dl_add(dl_b
, p
->dl
.dl_bw
);
5376 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5377 rcu_read_unlock_sched();
5387 static bool sched_smp_initialized __read_mostly
;
5389 #ifdef CONFIG_NUMA_BALANCING
5390 /* Migrate current task p to target_cpu */
5391 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5393 struct migration_arg arg
= { p
, target_cpu
};
5394 int curr_cpu
= task_cpu(p
);
5396 if (curr_cpu
== target_cpu
)
5399 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5402 /* TODO: This is not properly updating schedstats */
5404 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5405 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5409 * Requeue a task on a given node and accurately track the number of NUMA
5410 * tasks on the runqueues
5412 void sched_setnuma(struct task_struct
*p
, int nid
)
5414 bool queued
, running
;
5418 rq
= task_rq_lock(p
, &rf
);
5419 queued
= task_on_rq_queued(p
);
5420 running
= task_current(rq
, p
);
5423 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5425 put_prev_task(rq
, p
);
5427 p
->numa_preferred_nid
= nid
;
5430 p
->sched_class
->set_curr_task(rq
);
5432 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5433 task_rq_unlock(rq
, p
, &rf
);
5435 #endif /* CONFIG_NUMA_BALANCING */
5437 #ifdef CONFIG_HOTPLUG_CPU
5439 * Ensures that the idle task is using init_mm right before its cpu goes
5442 void idle_task_exit(void)
5444 struct mm_struct
*mm
= current
->active_mm
;
5446 BUG_ON(cpu_online(smp_processor_id()));
5448 if (mm
!= &init_mm
) {
5449 switch_mm_irqs_off(mm
, &init_mm
, current
);
5450 finish_arch_post_lock_switch();
5456 * Since this CPU is going 'away' for a while, fold any nr_active delta
5457 * we might have. Assumes we're called after migrate_tasks() so that the
5458 * nr_active count is stable. We need to take the teardown thread which
5459 * is calling this into account, so we hand in adjust = 1 to the load
5462 * Also see the comment "Global load-average calculations".
5464 static void calc_load_migrate(struct rq
*rq
)
5466 long delta
= calc_load_fold_active(rq
, 1);
5468 atomic_long_add(delta
, &calc_load_tasks
);
5471 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5475 static const struct sched_class fake_sched_class
= {
5476 .put_prev_task
= put_prev_task_fake
,
5479 static struct task_struct fake_task
= {
5481 * Avoid pull_{rt,dl}_task()
5483 .prio
= MAX_PRIO
+ 1,
5484 .sched_class
= &fake_sched_class
,
5488 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5489 * try_to_wake_up()->select_task_rq().
5491 * Called with rq->lock held even though we'er in stop_machine() and
5492 * there's no concurrency possible, we hold the required locks anyway
5493 * because of lock validation efforts.
5495 static void migrate_tasks(struct rq
*dead_rq
)
5497 struct rq
*rq
= dead_rq
;
5498 struct task_struct
*next
, *stop
= rq
->stop
;
5499 struct pin_cookie cookie
;
5503 * Fudge the rq selection such that the below task selection loop
5504 * doesn't get stuck on the currently eligible stop task.
5506 * We're currently inside stop_machine() and the rq is either stuck
5507 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5508 * either way we should never end up calling schedule() until we're
5514 * put_prev_task() and pick_next_task() sched
5515 * class method both need to have an up-to-date
5516 * value of rq->clock[_task]
5518 update_rq_clock(rq
);
5522 * There's this thread running, bail when that's the only
5525 if (rq
->nr_running
== 1)
5529 * pick_next_task assumes pinned rq->lock.
5531 cookie
= lockdep_pin_lock(&rq
->lock
);
5532 next
= pick_next_task(rq
, &fake_task
, cookie
);
5534 next
->sched_class
->put_prev_task(rq
, next
);
5537 * Rules for changing task_struct::cpus_allowed are holding
5538 * both pi_lock and rq->lock, such that holding either
5539 * stabilizes the mask.
5541 * Drop rq->lock is not quite as disastrous as it usually is
5542 * because !cpu_active at this point, which means load-balance
5543 * will not interfere. Also, stop-machine.
5545 lockdep_unpin_lock(&rq
->lock
, cookie
);
5546 raw_spin_unlock(&rq
->lock
);
5547 raw_spin_lock(&next
->pi_lock
);
5548 raw_spin_lock(&rq
->lock
);
5551 * Since we're inside stop-machine, _nothing_ should have
5552 * changed the task, WARN if weird stuff happened, because in
5553 * that case the above rq->lock drop is a fail too.
5555 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5556 raw_spin_unlock(&next
->pi_lock
);
5560 /* Find suitable destination for @next, with force if needed. */
5561 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5563 rq
= __migrate_task(rq
, next
, dest_cpu
);
5564 if (rq
!= dead_rq
) {
5565 raw_spin_unlock(&rq
->lock
);
5567 raw_spin_lock(&rq
->lock
);
5569 raw_spin_unlock(&next
->pi_lock
);
5574 #endif /* CONFIG_HOTPLUG_CPU */
5576 static void set_rq_online(struct rq
*rq
)
5579 const struct sched_class
*class;
5581 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5584 for_each_class(class) {
5585 if (class->rq_online
)
5586 class->rq_online(rq
);
5591 static void set_rq_offline(struct rq
*rq
)
5594 const struct sched_class
*class;
5596 for_each_class(class) {
5597 if (class->rq_offline
)
5598 class->rq_offline(rq
);
5601 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5606 static void set_cpu_rq_start_time(unsigned int cpu
)
5608 struct rq
*rq
= cpu_rq(cpu
);
5610 rq
->age_stamp
= sched_clock_cpu(cpu
);
5613 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5615 #ifdef CONFIG_SCHED_DEBUG
5617 static __read_mostly
int sched_debug_enabled
;
5619 static int __init
sched_debug_setup(char *str
)
5621 sched_debug_enabled
= 1;
5625 early_param("sched_debug", sched_debug_setup
);
5627 static inline bool sched_debug(void)
5629 return sched_debug_enabled
;
5632 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5633 struct cpumask
*groupmask
)
5635 struct sched_group
*group
= sd
->groups
;
5637 cpumask_clear(groupmask
);
5639 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5641 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5642 printk("does not load-balance\n");
5644 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5649 printk(KERN_CONT
"span %*pbl level %s\n",
5650 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5652 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5653 printk(KERN_ERR
"ERROR: domain->span does not contain "
5656 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5657 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5661 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5665 printk(KERN_ERR
"ERROR: group is NULL\n");
5669 if (!cpumask_weight(sched_group_cpus(group
))) {
5670 printk(KERN_CONT
"\n");
5671 printk(KERN_ERR
"ERROR: empty group\n");
5675 if (!(sd
->flags
& SD_OVERLAP
) &&
5676 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5677 printk(KERN_CONT
"\n");
5678 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5682 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5684 printk(KERN_CONT
" %*pbl",
5685 cpumask_pr_args(sched_group_cpus(group
)));
5686 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5687 printk(KERN_CONT
" (cpu_capacity = %d)",
5688 group
->sgc
->capacity
);
5691 group
= group
->next
;
5692 } while (group
!= sd
->groups
);
5693 printk(KERN_CONT
"\n");
5695 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5696 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5699 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5700 printk(KERN_ERR
"ERROR: parent span is not a superset "
5701 "of domain->span\n");
5705 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5709 if (!sched_debug_enabled
)
5713 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5717 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5720 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5728 #else /* !CONFIG_SCHED_DEBUG */
5729 # define sched_domain_debug(sd, cpu) do { } while (0)
5730 static inline bool sched_debug(void)
5734 #endif /* CONFIG_SCHED_DEBUG */
5736 static int sd_degenerate(struct sched_domain
*sd
)
5738 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5741 /* Following flags need at least 2 groups */
5742 if (sd
->flags
& (SD_LOAD_BALANCE
|
5743 SD_BALANCE_NEWIDLE
|
5746 SD_SHARE_CPUCAPACITY
|
5747 SD_ASYM_CPUCAPACITY
|
5748 SD_SHARE_PKG_RESOURCES
|
5749 SD_SHARE_POWERDOMAIN
)) {
5750 if (sd
->groups
!= sd
->groups
->next
)
5754 /* Following flags don't use groups */
5755 if (sd
->flags
& (SD_WAKE_AFFINE
))
5762 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5764 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5766 if (sd_degenerate(parent
))
5769 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5772 /* Flags needing groups don't count if only 1 group in parent */
5773 if (parent
->groups
== parent
->groups
->next
) {
5774 pflags
&= ~(SD_LOAD_BALANCE
|
5775 SD_BALANCE_NEWIDLE
|
5778 SD_ASYM_CPUCAPACITY
|
5779 SD_SHARE_CPUCAPACITY
|
5780 SD_SHARE_PKG_RESOURCES
|
5782 SD_SHARE_POWERDOMAIN
);
5783 if (nr_node_ids
== 1)
5784 pflags
&= ~SD_SERIALIZE
;
5786 if (~cflags
& pflags
)
5792 static void free_rootdomain(struct rcu_head
*rcu
)
5794 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5796 cpupri_cleanup(&rd
->cpupri
);
5797 cpudl_cleanup(&rd
->cpudl
);
5798 free_cpumask_var(rd
->dlo_mask
);
5799 free_cpumask_var(rd
->rto_mask
);
5800 free_cpumask_var(rd
->online
);
5801 free_cpumask_var(rd
->span
);
5805 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5807 struct root_domain
*old_rd
= NULL
;
5808 unsigned long flags
;
5810 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5815 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5818 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5821 * If we dont want to free the old_rd yet then
5822 * set old_rd to NULL to skip the freeing later
5825 if (!atomic_dec_and_test(&old_rd
->refcount
))
5829 atomic_inc(&rd
->refcount
);
5832 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5833 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5836 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5839 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5842 static int init_rootdomain(struct root_domain
*rd
)
5844 memset(rd
, 0, sizeof(*rd
));
5846 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5848 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5850 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5852 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5855 init_dl_bw(&rd
->dl_bw
);
5856 if (cpudl_init(&rd
->cpudl
) != 0)
5859 if (cpupri_init(&rd
->cpupri
) != 0)
5864 free_cpumask_var(rd
->rto_mask
);
5866 free_cpumask_var(rd
->dlo_mask
);
5868 free_cpumask_var(rd
->online
);
5870 free_cpumask_var(rd
->span
);
5876 * By default the system creates a single root-domain with all cpus as
5877 * members (mimicking the global state we have today).
5879 struct root_domain def_root_domain
;
5881 static void init_defrootdomain(void)
5883 init_rootdomain(&def_root_domain
);
5885 atomic_set(&def_root_domain
.refcount
, 1);
5888 static struct root_domain
*alloc_rootdomain(void)
5890 struct root_domain
*rd
;
5892 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5896 if (init_rootdomain(rd
) != 0) {
5904 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5906 struct sched_group
*tmp
, *first
;
5915 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5920 } while (sg
!= first
);
5923 static void free_sched_domain(struct rcu_head
*rcu
)
5925 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5928 * If its an overlapping domain it has private groups, iterate and
5931 if (sd
->flags
& SD_OVERLAP
) {
5932 free_sched_groups(sd
->groups
, 1);
5933 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5934 kfree(sd
->groups
->sgc
);
5940 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5942 call_rcu(&sd
->rcu
, free_sched_domain
);
5945 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5947 for (; sd
; sd
= sd
->parent
)
5948 destroy_sched_domain(sd
, cpu
);
5952 * Keep a special pointer to the highest sched_domain that has
5953 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5954 * allows us to avoid some pointer chasing select_idle_sibling().
5956 * Also keep a unique ID per domain (we use the first cpu number in
5957 * the cpumask of the domain), this allows us to quickly tell if
5958 * two cpus are in the same cache domain, see cpus_share_cache().
5960 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5961 DEFINE_PER_CPU(int, sd_llc_size
);
5962 DEFINE_PER_CPU(int, sd_llc_id
);
5963 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5964 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5965 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5967 static void update_top_cache_domain(int cpu
)
5969 struct sched_domain
*sd
;
5970 struct sched_domain
*busy_sd
= NULL
;
5974 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5976 id
= cpumask_first(sched_domain_span(sd
));
5977 size
= cpumask_weight(sched_domain_span(sd
));
5978 busy_sd
= sd
->parent
; /* sd_busy */
5980 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5982 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5983 per_cpu(sd_llc_size
, cpu
) = size
;
5984 per_cpu(sd_llc_id
, cpu
) = id
;
5986 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5987 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5989 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5990 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5994 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5995 * hold the hotplug lock.
5998 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6000 struct rq
*rq
= cpu_rq(cpu
);
6001 struct sched_domain
*tmp
;
6003 /* Remove the sched domains which do not contribute to scheduling. */
6004 for (tmp
= sd
; tmp
; ) {
6005 struct sched_domain
*parent
= tmp
->parent
;
6009 if (sd_parent_degenerate(tmp
, parent
)) {
6010 tmp
->parent
= parent
->parent
;
6012 parent
->parent
->child
= tmp
;
6014 * Transfer SD_PREFER_SIBLING down in case of a
6015 * degenerate parent; the spans match for this
6016 * so the property transfers.
6018 if (parent
->flags
& SD_PREFER_SIBLING
)
6019 tmp
->flags
|= SD_PREFER_SIBLING
;
6020 destroy_sched_domain(parent
, cpu
);
6025 if (sd
&& sd_degenerate(sd
)) {
6028 destroy_sched_domain(tmp
, cpu
);
6033 sched_domain_debug(sd
, cpu
);
6035 rq_attach_root(rq
, rd
);
6037 rcu_assign_pointer(rq
->sd
, sd
);
6038 destroy_sched_domains(tmp
, cpu
);
6040 update_top_cache_domain(cpu
);
6043 /* Setup the mask of cpus configured for isolated domains */
6044 static int __init
isolated_cpu_setup(char *str
)
6048 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6049 ret
= cpulist_parse(str
, cpu_isolated_map
);
6051 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
6056 __setup("isolcpus=", isolated_cpu_setup
);
6059 struct sched_domain
** __percpu sd
;
6060 struct root_domain
*rd
;
6071 * Build an iteration mask that can exclude certain CPUs from the upwards
6074 * Asymmetric node setups can result in situations where the domain tree is of
6075 * unequal depth, make sure to skip domains that already cover the entire
6078 * In that case build_sched_domains() will have terminated the iteration early
6079 * and our sibling sd spans will be empty. Domains should always include the
6080 * cpu they're built on, so check that.
6083 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6085 const struct cpumask
*span
= sched_domain_span(sd
);
6086 struct sd_data
*sdd
= sd
->private;
6087 struct sched_domain
*sibling
;
6090 for_each_cpu(i
, span
) {
6091 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6092 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6095 cpumask_set_cpu(i
, sched_group_mask(sg
));
6100 * Return the canonical balance cpu for this group, this is the first cpu
6101 * of this group that's also in the iteration mask.
6103 int group_balance_cpu(struct sched_group
*sg
)
6105 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6109 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6111 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6112 const struct cpumask
*span
= sched_domain_span(sd
);
6113 struct cpumask
*covered
= sched_domains_tmpmask
;
6114 struct sd_data
*sdd
= sd
->private;
6115 struct sched_domain
*sibling
;
6118 cpumask_clear(covered
);
6120 for_each_cpu(i
, span
) {
6121 struct cpumask
*sg_span
;
6123 if (cpumask_test_cpu(i
, covered
))
6126 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6128 /* See the comment near build_group_mask(). */
6129 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6132 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6133 GFP_KERNEL
, cpu_to_node(cpu
));
6138 sg_span
= sched_group_cpus(sg
);
6140 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6142 cpumask_set_cpu(i
, sg_span
);
6144 cpumask_or(covered
, covered
, sg_span
);
6146 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6147 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6148 build_group_mask(sd
, sg
);
6151 * Initialize sgc->capacity such that even if we mess up the
6152 * domains and no possible iteration will get us here, we won't
6155 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6158 * Make sure the first group of this domain contains the
6159 * canonical balance cpu. Otherwise the sched_domain iteration
6160 * breaks. See update_sg_lb_stats().
6162 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6163 group_balance_cpu(sg
) == cpu
)
6173 sd
->groups
= groups
;
6178 free_sched_groups(first
, 0);
6183 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6185 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6186 struct sched_domain
*child
= sd
->child
;
6189 cpu
= cpumask_first(sched_domain_span(child
));
6192 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6193 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6194 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6201 * build_sched_groups will build a circular linked list of the groups
6202 * covered by the given span, and will set each group's ->cpumask correctly,
6203 * and ->cpu_capacity to 0.
6205 * Assumes the sched_domain tree is fully constructed
6208 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6210 struct sched_group
*first
= NULL
, *last
= NULL
;
6211 struct sd_data
*sdd
= sd
->private;
6212 const struct cpumask
*span
= sched_domain_span(sd
);
6213 struct cpumask
*covered
;
6216 get_group(cpu
, sdd
, &sd
->groups
);
6217 atomic_inc(&sd
->groups
->ref
);
6219 if (cpu
!= cpumask_first(span
))
6222 lockdep_assert_held(&sched_domains_mutex
);
6223 covered
= sched_domains_tmpmask
;
6225 cpumask_clear(covered
);
6227 for_each_cpu(i
, span
) {
6228 struct sched_group
*sg
;
6231 if (cpumask_test_cpu(i
, covered
))
6234 group
= get_group(i
, sdd
, &sg
);
6235 cpumask_setall(sched_group_mask(sg
));
6237 for_each_cpu(j
, span
) {
6238 if (get_group(j
, sdd
, NULL
) != group
)
6241 cpumask_set_cpu(j
, covered
);
6242 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6257 * Initialize sched groups cpu_capacity.
6259 * cpu_capacity indicates the capacity of sched group, which is used while
6260 * distributing the load between different sched groups in a sched domain.
6261 * Typically cpu_capacity for all the groups in a sched domain will be same
6262 * unless there are asymmetries in the topology. If there are asymmetries,
6263 * group having more cpu_capacity will pickup more load compared to the
6264 * group having less cpu_capacity.
6266 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6268 struct sched_group
*sg
= sd
->groups
;
6273 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6275 } while (sg
!= sd
->groups
);
6277 if (cpu
!= group_balance_cpu(sg
))
6280 update_group_capacity(sd
, cpu
);
6281 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6285 * Initializers for schedule domains
6286 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6289 static int default_relax_domain_level
= -1;
6290 int sched_domain_level_max
;
6292 static int __init
setup_relax_domain_level(char *str
)
6294 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6295 pr_warn("Unable to set relax_domain_level\n");
6299 __setup("relax_domain_level=", setup_relax_domain_level
);
6301 static void set_domain_attribute(struct sched_domain
*sd
,
6302 struct sched_domain_attr
*attr
)
6306 if (!attr
|| attr
->relax_domain_level
< 0) {
6307 if (default_relax_domain_level
< 0)
6310 request
= default_relax_domain_level
;
6312 request
= attr
->relax_domain_level
;
6313 if (request
< sd
->level
) {
6314 /* turn off idle balance on this domain */
6315 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6317 /* turn on idle balance on this domain */
6318 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6322 static void __sdt_free(const struct cpumask
*cpu_map
);
6323 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6325 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6326 const struct cpumask
*cpu_map
)
6330 if (!atomic_read(&d
->rd
->refcount
))
6331 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6333 free_percpu(d
->sd
); /* fall through */
6335 __sdt_free(cpu_map
); /* fall through */
6341 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6342 const struct cpumask
*cpu_map
)
6344 memset(d
, 0, sizeof(*d
));
6346 if (__sdt_alloc(cpu_map
))
6347 return sa_sd_storage
;
6348 d
->sd
= alloc_percpu(struct sched_domain
*);
6350 return sa_sd_storage
;
6351 d
->rd
= alloc_rootdomain();
6354 return sa_rootdomain
;
6358 * NULL the sd_data elements we've used to build the sched_domain and
6359 * sched_group structure so that the subsequent __free_domain_allocs()
6360 * will not free the data we're using.
6362 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6364 struct sd_data
*sdd
= sd
->private;
6366 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6367 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6369 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6370 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6372 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6373 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6377 static int sched_domains_numa_levels
;
6378 enum numa_topology_type sched_numa_topology_type
;
6379 static int *sched_domains_numa_distance
;
6380 int sched_max_numa_distance
;
6381 static struct cpumask
***sched_domains_numa_masks
;
6382 static int sched_domains_curr_level
;
6386 * SD_flags allowed in topology descriptions.
6388 * These flags are purely descriptive of the topology and do not prescribe
6389 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6392 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6393 * SD_SHARE_PKG_RESOURCES - describes shared caches
6394 * SD_NUMA - describes NUMA topologies
6395 * SD_SHARE_POWERDOMAIN - describes shared power domain
6396 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6398 * Odd one out, which beside describing the topology has a quirk also
6399 * prescribes the desired behaviour that goes along with it:
6401 * SD_ASYM_PACKING - describes SMT quirks
6403 #define TOPOLOGY_SD_FLAGS \
6404 (SD_SHARE_CPUCAPACITY | \
6405 SD_SHARE_PKG_RESOURCES | \
6408 SD_ASYM_CPUCAPACITY | \
6409 SD_SHARE_POWERDOMAIN)
6411 static struct sched_domain
*
6412 sd_init(struct sched_domain_topology_level
*tl
,
6413 struct sched_domain
*child
, int cpu
)
6415 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6416 int sd_weight
, sd_flags
= 0;
6420 * Ugly hack to pass state to sd_numa_mask()...
6422 sched_domains_curr_level
= tl
->numa_level
;
6425 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6428 sd_flags
= (*tl
->sd_flags
)();
6429 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6430 "wrong sd_flags in topology description\n"))
6431 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6433 *sd
= (struct sched_domain
){
6434 .min_interval
= sd_weight
,
6435 .max_interval
= 2*sd_weight
,
6437 .imbalance_pct
= 125,
6439 .cache_nice_tries
= 0,
6446 .flags
= 1*SD_LOAD_BALANCE
6447 | 1*SD_BALANCE_NEWIDLE
6452 | 0*SD_SHARE_CPUCAPACITY
6453 | 0*SD_SHARE_PKG_RESOURCES
6455 | 0*SD_PREFER_SIBLING
6460 .last_balance
= jiffies
,
6461 .balance_interval
= sd_weight
,
6463 .max_newidle_lb_cost
= 0,
6464 .next_decay_max_lb_cost
= jiffies
,
6466 #ifdef CONFIG_SCHED_DEBUG
6472 * Convert topological properties into behaviour.
6475 if (sd
->flags
& SD_ASYM_CPUCAPACITY
) {
6476 struct sched_domain
*t
= sd
;
6478 for_each_lower_domain(t
)
6479 t
->flags
|= SD_BALANCE_WAKE
;
6482 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6483 sd
->flags
|= SD_PREFER_SIBLING
;
6484 sd
->imbalance_pct
= 110;
6485 sd
->smt_gain
= 1178; /* ~15% */
6487 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6488 sd
->imbalance_pct
= 117;
6489 sd
->cache_nice_tries
= 1;
6493 } else if (sd
->flags
& SD_NUMA
) {
6494 sd
->cache_nice_tries
= 2;
6498 sd
->flags
|= SD_SERIALIZE
;
6499 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6500 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6507 sd
->flags
|= SD_PREFER_SIBLING
;
6508 sd
->cache_nice_tries
= 1;
6513 sd
->private = &tl
->data
;
6519 * Topology list, bottom-up.
6521 static struct sched_domain_topology_level default_topology
[] = {
6522 #ifdef CONFIG_SCHED_SMT
6523 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6525 #ifdef CONFIG_SCHED_MC
6526 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6528 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6532 static struct sched_domain_topology_level
*sched_domain_topology
=
6535 #define for_each_sd_topology(tl) \
6536 for (tl = sched_domain_topology; tl->mask; tl++)
6538 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6540 sched_domain_topology
= tl
;
6545 static const struct cpumask
*sd_numa_mask(int cpu
)
6547 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6550 static void sched_numa_warn(const char *str
)
6552 static int done
= false;
6560 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6562 for (i
= 0; i
< nr_node_ids
; i
++) {
6563 printk(KERN_WARNING
" ");
6564 for (j
= 0; j
< nr_node_ids
; j
++)
6565 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6566 printk(KERN_CONT
"\n");
6568 printk(KERN_WARNING
"\n");
6571 bool find_numa_distance(int distance
)
6575 if (distance
== node_distance(0, 0))
6578 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6579 if (sched_domains_numa_distance
[i
] == distance
)
6587 * A system can have three types of NUMA topology:
6588 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6589 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6590 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6592 * The difference between a glueless mesh topology and a backplane
6593 * topology lies in whether communication between not directly
6594 * connected nodes goes through intermediary nodes (where programs
6595 * could run), or through backplane controllers. This affects
6596 * placement of programs.
6598 * The type of topology can be discerned with the following tests:
6599 * - If the maximum distance between any nodes is 1 hop, the system
6600 * is directly connected.
6601 * - If for two nodes A and B, located N > 1 hops away from each other,
6602 * there is an intermediary node C, which is < N hops away from both
6603 * nodes A and B, the system is a glueless mesh.
6605 static void init_numa_topology_type(void)
6609 n
= sched_max_numa_distance
;
6611 if (sched_domains_numa_levels
<= 1) {
6612 sched_numa_topology_type
= NUMA_DIRECT
;
6616 for_each_online_node(a
) {
6617 for_each_online_node(b
) {
6618 /* Find two nodes furthest removed from each other. */
6619 if (node_distance(a
, b
) < n
)
6622 /* Is there an intermediary node between a and b? */
6623 for_each_online_node(c
) {
6624 if (node_distance(a
, c
) < n
&&
6625 node_distance(b
, c
) < n
) {
6626 sched_numa_topology_type
=
6632 sched_numa_topology_type
= NUMA_BACKPLANE
;
6638 static void sched_init_numa(void)
6640 int next_distance
, curr_distance
= node_distance(0, 0);
6641 struct sched_domain_topology_level
*tl
;
6645 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6646 if (!sched_domains_numa_distance
)
6650 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6651 * unique distances in the node_distance() table.
6653 * Assumes node_distance(0,j) includes all distances in
6654 * node_distance(i,j) in order to avoid cubic time.
6656 next_distance
= curr_distance
;
6657 for (i
= 0; i
< nr_node_ids
; i
++) {
6658 for (j
= 0; j
< nr_node_ids
; j
++) {
6659 for (k
= 0; k
< nr_node_ids
; k
++) {
6660 int distance
= node_distance(i
, k
);
6662 if (distance
> curr_distance
&&
6663 (distance
< next_distance
||
6664 next_distance
== curr_distance
))
6665 next_distance
= distance
;
6668 * While not a strong assumption it would be nice to know
6669 * about cases where if node A is connected to B, B is not
6670 * equally connected to A.
6672 if (sched_debug() && node_distance(k
, i
) != distance
)
6673 sched_numa_warn("Node-distance not symmetric");
6675 if (sched_debug() && i
&& !find_numa_distance(distance
))
6676 sched_numa_warn("Node-0 not representative");
6678 if (next_distance
!= curr_distance
) {
6679 sched_domains_numa_distance
[level
++] = next_distance
;
6680 sched_domains_numa_levels
= level
;
6681 curr_distance
= next_distance
;
6686 * In case of sched_debug() we verify the above assumption.
6696 * 'level' contains the number of unique distances, excluding the
6697 * identity distance node_distance(i,i).
6699 * The sched_domains_numa_distance[] array includes the actual distance
6704 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6705 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6706 * the array will contain less then 'level' members. This could be
6707 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6708 * in other functions.
6710 * We reset it to 'level' at the end of this function.
6712 sched_domains_numa_levels
= 0;
6714 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6715 if (!sched_domains_numa_masks
)
6719 * Now for each level, construct a mask per node which contains all
6720 * cpus of nodes that are that many hops away from us.
6722 for (i
= 0; i
< level
; i
++) {
6723 sched_domains_numa_masks
[i
] =
6724 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6725 if (!sched_domains_numa_masks
[i
])
6728 for (j
= 0; j
< nr_node_ids
; j
++) {
6729 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6733 sched_domains_numa_masks
[i
][j
] = mask
;
6736 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6739 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6744 /* Compute default topology size */
6745 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6747 tl
= kzalloc((i
+ level
+ 1) *
6748 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6753 * Copy the default topology bits..
6755 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6756 tl
[i
] = sched_domain_topology
[i
];
6759 * .. and append 'j' levels of NUMA goodness.
6761 for (j
= 0; j
< level
; i
++, j
++) {
6762 tl
[i
] = (struct sched_domain_topology_level
){
6763 .mask
= sd_numa_mask
,
6764 .sd_flags
= cpu_numa_flags
,
6765 .flags
= SDTL_OVERLAP
,
6771 sched_domain_topology
= tl
;
6773 sched_domains_numa_levels
= level
;
6774 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6776 init_numa_topology_type();
6779 static void sched_domains_numa_masks_set(unsigned int cpu
)
6781 int node
= cpu_to_node(cpu
);
6784 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6785 for (j
= 0; j
< nr_node_ids
; j
++) {
6786 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6787 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6792 static void sched_domains_numa_masks_clear(unsigned int cpu
)
6796 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6797 for (j
= 0; j
< nr_node_ids
; j
++)
6798 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6803 static inline void sched_init_numa(void) { }
6804 static void sched_domains_numa_masks_set(unsigned int cpu
) { }
6805 static void sched_domains_numa_masks_clear(unsigned int cpu
) { }
6806 #endif /* CONFIG_NUMA */
6808 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6810 struct sched_domain_topology_level
*tl
;
6813 for_each_sd_topology(tl
) {
6814 struct sd_data
*sdd
= &tl
->data
;
6816 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6820 sdd
->sg
= alloc_percpu(struct sched_group
*);
6824 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6828 for_each_cpu(j
, cpu_map
) {
6829 struct sched_domain
*sd
;
6830 struct sched_group
*sg
;
6831 struct sched_group_capacity
*sgc
;
6833 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6834 GFP_KERNEL
, cpu_to_node(j
));
6838 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6840 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6841 GFP_KERNEL
, cpu_to_node(j
));
6847 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6849 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6850 GFP_KERNEL
, cpu_to_node(j
));
6854 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6861 static void __sdt_free(const struct cpumask
*cpu_map
)
6863 struct sched_domain_topology_level
*tl
;
6866 for_each_sd_topology(tl
) {
6867 struct sd_data
*sdd
= &tl
->data
;
6869 for_each_cpu(j
, cpu_map
) {
6870 struct sched_domain
*sd
;
6873 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6874 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6875 free_sched_groups(sd
->groups
, 0);
6876 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6880 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6882 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6884 free_percpu(sdd
->sd
);
6886 free_percpu(sdd
->sg
);
6888 free_percpu(sdd
->sgc
);
6893 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6894 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6895 struct sched_domain
*child
, int cpu
)
6897 struct sched_domain
*sd
= sd_init(tl
, child
, cpu
);
6899 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6901 sd
->level
= child
->level
+ 1;
6902 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6905 if (!cpumask_subset(sched_domain_span(child
),
6906 sched_domain_span(sd
))) {
6907 pr_err("BUG: arch topology borken\n");
6908 #ifdef CONFIG_SCHED_DEBUG
6909 pr_err(" the %s domain not a subset of the %s domain\n",
6910 child
->name
, sd
->name
);
6912 /* Fixup, ensure @sd has at least @child cpus. */
6913 cpumask_or(sched_domain_span(sd
),
6914 sched_domain_span(sd
),
6915 sched_domain_span(child
));
6919 set_domain_attribute(sd
, attr
);
6925 * Build sched domains for a given set of cpus and attach the sched domains
6926 * to the individual cpus
6928 static int build_sched_domains(const struct cpumask
*cpu_map
,
6929 struct sched_domain_attr
*attr
)
6931 enum s_alloc alloc_state
;
6932 struct sched_domain
*sd
;
6934 struct rq
*rq
= NULL
;
6935 int i
, ret
= -ENOMEM
;
6937 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6938 if (alloc_state
!= sa_rootdomain
)
6941 /* Set up domains for cpus specified by the cpu_map. */
6942 for_each_cpu(i
, cpu_map
) {
6943 struct sched_domain_topology_level
*tl
;
6946 for_each_sd_topology(tl
) {
6947 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6948 if (tl
== sched_domain_topology
)
6949 *per_cpu_ptr(d
.sd
, i
) = sd
;
6950 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6951 sd
->flags
|= SD_OVERLAP
;
6952 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6957 /* Build the groups for the domains */
6958 for_each_cpu(i
, cpu_map
) {
6959 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6960 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6961 if (sd
->flags
& SD_OVERLAP
) {
6962 if (build_overlap_sched_groups(sd
, i
))
6965 if (build_sched_groups(sd
, i
))
6971 /* Calculate CPU capacity for physical packages and nodes */
6972 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6973 if (!cpumask_test_cpu(i
, cpu_map
))
6976 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6977 claim_allocations(i
, sd
);
6978 init_sched_groups_capacity(i
, sd
);
6982 /* Attach the domains */
6984 for_each_cpu(i
, cpu_map
) {
6986 sd
= *per_cpu_ptr(d
.sd
, i
);
6988 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
6989 if (rq
->cpu_capacity_orig
> READ_ONCE(d
.rd
->max_cpu_capacity
))
6990 WRITE_ONCE(d
.rd
->max_cpu_capacity
, rq
->cpu_capacity_orig
);
6992 cpu_attach_domain(sd
, d
.rd
, i
);
6997 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
6998 cpumask_pr_args(cpu_map
), rq
->rd
->max_cpu_capacity
);
7003 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7007 static cpumask_var_t
*doms_cur
; /* current sched domains */
7008 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7009 static struct sched_domain_attr
*dattr_cur
;
7010 /* attribues of custom domains in 'doms_cur' */
7013 * Special case: If a kmalloc of a doms_cur partition (array of
7014 * cpumask) fails, then fallback to a single sched domain,
7015 * as determined by the single cpumask fallback_doms.
7017 static cpumask_var_t fallback_doms
;
7020 * arch_update_cpu_topology lets virtualized architectures update the
7021 * cpu core maps. It is supposed to return 1 if the topology changed
7022 * or 0 if it stayed the same.
7024 int __weak
arch_update_cpu_topology(void)
7029 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7032 cpumask_var_t
*doms
;
7034 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7037 for (i
= 0; i
< ndoms
; i
++) {
7038 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7039 free_sched_domains(doms
, i
);
7046 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7049 for (i
= 0; i
< ndoms
; i
++)
7050 free_cpumask_var(doms
[i
]);
7055 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7056 * For now this just excludes isolated cpus, but could be used to
7057 * exclude other special cases in the future.
7059 static int init_sched_domains(const struct cpumask
*cpu_map
)
7063 arch_update_cpu_topology();
7065 doms_cur
= alloc_sched_domains(ndoms_cur
);
7067 doms_cur
= &fallback_doms
;
7068 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7069 err
= build_sched_domains(doms_cur
[0], NULL
);
7070 register_sched_domain_sysctl();
7076 * Detach sched domains from a group of cpus specified in cpu_map
7077 * These cpus will now be attached to the NULL domain
7079 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7084 for_each_cpu(i
, cpu_map
)
7085 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7089 /* handle null as "default" */
7090 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7091 struct sched_domain_attr
*new, int idx_new
)
7093 struct sched_domain_attr tmp
;
7100 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7101 new ? (new + idx_new
) : &tmp
,
7102 sizeof(struct sched_domain_attr
));
7106 * Partition sched domains as specified by the 'ndoms_new'
7107 * cpumasks in the array doms_new[] of cpumasks. This compares
7108 * doms_new[] to the current sched domain partitioning, doms_cur[].
7109 * It destroys each deleted domain and builds each new domain.
7111 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7112 * The masks don't intersect (don't overlap.) We should setup one
7113 * sched domain for each mask. CPUs not in any of the cpumasks will
7114 * not be load balanced. If the same cpumask appears both in the
7115 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7118 * The passed in 'doms_new' should be allocated using
7119 * alloc_sched_domains. This routine takes ownership of it and will
7120 * free_sched_domains it when done with it. If the caller failed the
7121 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7122 * and partition_sched_domains() will fallback to the single partition
7123 * 'fallback_doms', it also forces the domains to be rebuilt.
7125 * If doms_new == NULL it will be replaced with cpu_online_mask.
7126 * ndoms_new == 0 is a special case for destroying existing domains,
7127 * and it will not create the default domain.
7129 * Call with hotplug lock held
7131 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7132 struct sched_domain_attr
*dattr_new
)
7137 mutex_lock(&sched_domains_mutex
);
7139 /* always unregister in case we don't destroy any domains */
7140 unregister_sched_domain_sysctl();
7142 /* Let architecture update cpu core mappings. */
7143 new_topology
= arch_update_cpu_topology();
7145 n
= doms_new
? ndoms_new
: 0;
7147 /* Destroy deleted domains */
7148 for (i
= 0; i
< ndoms_cur
; i
++) {
7149 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7150 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7151 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7154 /* no match - a current sched domain not in new doms_new[] */
7155 detach_destroy_domains(doms_cur
[i
]);
7161 if (doms_new
== NULL
) {
7163 doms_new
= &fallback_doms
;
7164 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7165 WARN_ON_ONCE(dattr_new
);
7168 /* Build new domains */
7169 for (i
= 0; i
< ndoms_new
; i
++) {
7170 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7171 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7172 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7175 /* no match - add a new doms_new */
7176 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7181 /* Remember the new sched domains */
7182 if (doms_cur
!= &fallback_doms
)
7183 free_sched_domains(doms_cur
, ndoms_cur
);
7184 kfree(dattr_cur
); /* kfree(NULL) is safe */
7185 doms_cur
= doms_new
;
7186 dattr_cur
= dattr_new
;
7187 ndoms_cur
= ndoms_new
;
7189 register_sched_domain_sysctl();
7191 mutex_unlock(&sched_domains_mutex
);
7194 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7197 * Update cpusets according to cpu_active mask. If cpusets are
7198 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7199 * around partition_sched_domains().
7201 * If we come here as part of a suspend/resume, don't touch cpusets because we
7202 * want to restore it back to its original state upon resume anyway.
7204 static void cpuset_cpu_active(void)
7206 if (cpuhp_tasks_frozen
) {
7208 * num_cpus_frozen tracks how many CPUs are involved in suspend
7209 * resume sequence. As long as this is not the last online
7210 * operation in the resume sequence, just build a single sched
7211 * domain, ignoring cpusets.
7214 if (likely(num_cpus_frozen
)) {
7215 partition_sched_domains(1, NULL
, NULL
);
7219 * This is the last CPU online operation. So fall through and
7220 * restore the original sched domains by considering the
7221 * cpuset configurations.
7224 cpuset_update_active_cpus(true);
7227 static int cpuset_cpu_inactive(unsigned int cpu
)
7229 unsigned long flags
;
7234 if (!cpuhp_tasks_frozen
) {
7235 rcu_read_lock_sched();
7236 dl_b
= dl_bw_of(cpu
);
7238 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7239 cpus
= dl_bw_cpus(cpu
);
7240 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7241 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7243 rcu_read_unlock_sched();
7247 cpuset_update_active_cpus(false);
7250 partition_sched_domains(1, NULL
, NULL
);
7255 int sched_cpu_activate(unsigned int cpu
)
7257 struct rq
*rq
= cpu_rq(cpu
);
7258 unsigned long flags
;
7260 set_cpu_active(cpu
, true);
7262 if (sched_smp_initialized
) {
7263 sched_domains_numa_masks_set(cpu
);
7264 cpuset_cpu_active();
7268 * Put the rq online, if not already. This happens:
7270 * 1) In the early boot process, because we build the real domains
7271 * after all cpus have been brought up.
7273 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7276 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7278 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7281 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7283 update_max_interval();
7288 int sched_cpu_deactivate(unsigned int cpu
)
7292 set_cpu_active(cpu
, false);
7294 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7295 * users of this state to go away such that all new such users will
7298 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7299 * not imply sync_sched(), so wait for both.
7301 * Do sync before park smpboot threads to take care the rcu boost case.
7303 if (IS_ENABLED(CONFIG_PREEMPT
))
7304 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
7308 if (!sched_smp_initialized
)
7311 ret
= cpuset_cpu_inactive(cpu
);
7313 set_cpu_active(cpu
, true);
7316 sched_domains_numa_masks_clear(cpu
);
7320 static void sched_rq_cpu_starting(unsigned int cpu
)
7322 struct rq
*rq
= cpu_rq(cpu
);
7324 rq
->calc_load_update
= calc_load_update
;
7325 update_max_interval();
7328 int sched_cpu_starting(unsigned int cpu
)
7330 set_cpu_rq_start_time(cpu
);
7331 sched_rq_cpu_starting(cpu
);
7335 #ifdef CONFIG_HOTPLUG_CPU
7336 int sched_cpu_dying(unsigned int cpu
)
7338 struct rq
*rq
= cpu_rq(cpu
);
7339 unsigned long flags
;
7341 /* Handle pending wakeups and then migrate everything off */
7342 sched_ttwu_pending();
7343 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7345 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7349 BUG_ON(rq
->nr_running
!= 1);
7350 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7351 calc_load_migrate(rq
);
7352 update_max_interval();
7353 nohz_balance_exit_idle(cpu
);
7359 void __init
sched_init_smp(void)
7361 cpumask_var_t non_isolated_cpus
;
7363 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7364 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7369 * There's no userspace yet to cause hotplug operations; hence all the
7370 * cpu masks are stable and all blatant races in the below code cannot
7373 mutex_lock(&sched_domains_mutex
);
7374 init_sched_domains(cpu_active_mask
);
7375 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7376 if (cpumask_empty(non_isolated_cpus
))
7377 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7378 mutex_unlock(&sched_domains_mutex
);
7380 /* Move init over to a non-isolated CPU */
7381 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7383 sched_init_granularity();
7384 free_cpumask_var(non_isolated_cpus
);
7386 init_sched_rt_class();
7387 init_sched_dl_class();
7388 sched_smp_initialized
= true;
7391 static int __init
migration_init(void)
7393 sched_rq_cpu_starting(smp_processor_id());
7396 early_initcall(migration_init
);
7399 void __init
sched_init_smp(void)
7401 sched_init_granularity();
7403 #endif /* CONFIG_SMP */
7405 int in_sched_functions(unsigned long addr
)
7407 return in_lock_functions(addr
) ||
7408 (addr
>= (unsigned long)__sched_text_start
7409 && addr
< (unsigned long)__sched_text_end
);
7412 #ifdef CONFIG_CGROUP_SCHED
7414 * Default task group.
7415 * Every task in system belongs to this group at bootup.
7417 struct task_group root_task_group
;
7418 LIST_HEAD(task_groups
);
7420 /* Cacheline aligned slab cache for task_group */
7421 static struct kmem_cache
*task_group_cache __read_mostly
;
7424 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7426 void __init
sched_init(void)
7429 unsigned long alloc_size
= 0, ptr
;
7431 #ifdef CONFIG_FAIR_GROUP_SCHED
7432 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7434 #ifdef CONFIG_RT_GROUP_SCHED
7435 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7438 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7440 #ifdef CONFIG_FAIR_GROUP_SCHED
7441 root_task_group
.se
= (struct sched_entity
**)ptr
;
7442 ptr
+= nr_cpu_ids
* sizeof(void **);
7444 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7445 ptr
+= nr_cpu_ids
* sizeof(void **);
7447 #endif /* CONFIG_FAIR_GROUP_SCHED */
7448 #ifdef CONFIG_RT_GROUP_SCHED
7449 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7450 ptr
+= nr_cpu_ids
* sizeof(void **);
7452 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7453 ptr
+= nr_cpu_ids
* sizeof(void **);
7455 #endif /* CONFIG_RT_GROUP_SCHED */
7457 #ifdef CONFIG_CPUMASK_OFFSTACK
7458 for_each_possible_cpu(i
) {
7459 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7460 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7462 #endif /* CONFIG_CPUMASK_OFFSTACK */
7464 init_rt_bandwidth(&def_rt_bandwidth
,
7465 global_rt_period(), global_rt_runtime());
7466 init_dl_bandwidth(&def_dl_bandwidth
,
7467 global_rt_period(), global_rt_runtime());
7470 init_defrootdomain();
7473 #ifdef CONFIG_RT_GROUP_SCHED
7474 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7475 global_rt_period(), global_rt_runtime());
7476 #endif /* CONFIG_RT_GROUP_SCHED */
7478 #ifdef CONFIG_CGROUP_SCHED
7479 task_group_cache
= KMEM_CACHE(task_group
, 0);
7481 list_add(&root_task_group
.list
, &task_groups
);
7482 INIT_LIST_HEAD(&root_task_group
.children
);
7483 INIT_LIST_HEAD(&root_task_group
.siblings
);
7484 autogroup_init(&init_task
);
7485 #endif /* CONFIG_CGROUP_SCHED */
7487 for_each_possible_cpu(i
) {
7491 raw_spin_lock_init(&rq
->lock
);
7493 rq
->calc_load_active
= 0;
7494 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7495 init_cfs_rq(&rq
->cfs
);
7496 init_rt_rq(&rq
->rt
);
7497 init_dl_rq(&rq
->dl
);
7498 #ifdef CONFIG_FAIR_GROUP_SCHED
7499 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7500 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7502 * How much cpu bandwidth does root_task_group get?
7504 * In case of task-groups formed thr' the cgroup filesystem, it
7505 * gets 100% of the cpu resources in the system. This overall
7506 * system cpu resource is divided among the tasks of
7507 * root_task_group and its child task-groups in a fair manner,
7508 * based on each entity's (task or task-group's) weight
7509 * (se->load.weight).
7511 * In other words, if root_task_group has 10 tasks of weight
7512 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7513 * then A0's share of the cpu resource is:
7515 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7517 * We achieve this by letting root_task_group's tasks sit
7518 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7520 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7521 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7522 #endif /* CONFIG_FAIR_GROUP_SCHED */
7524 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7525 #ifdef CONFIG_RT_GROUP_SCHED
7526 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7529 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7530 rq
->cpu_load
[j
] = 0;
7535 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7536 rq
->balance_callback
= NULL
;
7537 rq
->active_balance
= 0;
7538 rq
->next_balance
= jiffies
;
7543 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7544 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7546 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7548 rq_attach_root(rq
, &def_root_domain
);
7549 #ifdef CONFIG_NO_HZ_COMMON
7550 rq
->last_load_update_tick
= jiffies
;
7553 #ifdef CONFIG_NO_HZ_FULL
7554 rq
->last_sched_tick
= 0;
7556 #endif /* CONFIG_SMP */
7558 atomic_set(&rq
->nr_iowait
, 0);
7561 set_load_weight(&init_task
);
7564 * The boot idle thread does lazy MMU switching as well:
7566 atomic_inc(&init_mm
.mm_count
);
7567 enter_lazy_tlb(&init_mm
, current
);
7570 * During early bootup we pretend to be a normal task:
7572 current
->sched_class
= &fair_sched_class
;
7575 * Make us the idle thread. Technically, schedule() should not be
7576 * called from this thread, however somewhere below it might be,
7577 * but because we are the idle thread, we just pick up running again
7578 * when this runqueue becomes "idle".
7580 init_idle(current
, smp_processor_id());
7582 calc_load_update
= jiffies
+ LOAD_FREQ
;
7585 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7586 /* May be allocated at isolcpus cmdline parse time */
7587 if (cpu_isolated_map
== NULL
)
7588 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7589 idle_thread_set_boot_cpu();
7590 set_cpu_rq_start_time(smp_processor_id());
7592 init_sched_fair_class();
7596 scheduler_running
= 1;
7599 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7600 static inline int preempt_count_equals(int preempt_offset
)
7602 int nested
= preempt_count() + rcu_preempt_depth();
7604 return (nested
== preempt_offset
);
7607 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7610 * Blocking primitives will set (and therefore destroy) current->state,
7611 * since we will exit with TASK_RUNNING make sure we enter with it,
7612 * otherwise we will destroy state.
7614 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7615 "do not call blocking ops when !TASK_RUNNING; "
7616 "state=%lx set at [<%p>] %pS\n",
7618 (void *)current
->task_state_change
,
7619 (void *)current
->task_state_change
);
7621 ___might_sleep(file
, line
, preempt_offset
);
7623 EXPORT_SYMBOL(__might_sleep
);
7625 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7627 static unsigned long prev_jiffy
; /* ratelimiting */
7628 unsigned long preempt_disable_ip
;
7630 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7631 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7632 !is_idle_task(current
)) ||
7633 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7635 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7637 prev_jiffy
= jiffies
;
7639 /* Save this before calling printk(), since that will clobber it */
7640 preempt_disable_ip
= get_preempt_disable_ip(current
);
7643 "BUG: sleeping function called from invalid context at %s:%d\n",
7646 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7647 in_atomic(), irqs_disabled(),
7648 current
->pid
, current
->comm
);
7650 if (task_stack_end_corrupted(current
))
7651 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7653 debug_show_held_locks(current
);
7654 if (irqs_disabled())
7655 print_irqtrace_events(current
);
7656 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
7657 && !preempt_count_equals(preempt_offset
)) {
7658 pr_err("Preemption disabled at:");
7659 print_ip_sym(preempt_disable_ip
);
7663 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7665 EXPORT_SYMBOL(___might_sleep
);
7668 #ifdef CONFIG_MAGIC_SYSRQ
7669 void normalize_rt_tasks(void)
7671 struct task_struct
*g
, *p
;
7672 struct sched_attr attr
= {
7673 .sched_policy
= SCHED_NORMAL
,
7676 read_lock(&tasklist_lock
);
7677 for_each_process_thread(g
, p
) {
7679 * Only normalize user tasks:
7681 if (p
->flags
& PF_KTHREAD
)
7684 p
->se
.exec_start
= 0;
7685 schedstat_set(p
->se
.statistics
.wait_start
, 0);
7686 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
7687 schedstat_set(p
->se
.statistics
.block_start
, 0);
7689 if (!dl_task(p
) && !rt_task(p
)) {
7691 * Renice negative nice level userspace
7694 if (task_nice(p
) < 0)
7695 set_user_nice(p
, 0);
7699 __sched_setscheduler(p
, &attr
, false, false);
7701 read_unlock(&tasklist_lock
);
7704 #endif /* CONFIG_MAGIC_SYSRQ */
7706 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7708 * These functions are only useful for the IA64 MCA handling, or kdb.
7710 * They can only be called when the whole system has been
7711 * stopped - every CPU needs to be quiescent, and no scheduling
7712 * activity can take place. Using them for anything else would
7713 * be a serious bug, and as a result, they aren't even visible
7714 * under any other configuration.
7718 * curr_task - return the current task for a given cpu.
7719 * @cpu: the processor in question.
7721 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7723 * Return: The current task for @cpu.
7725 struct task_struct
*curr_task(int cpu
)
7727 return cpu_curr(cpu
);
7730 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7734 * set_curr_task - set the current task for a given cpu.
7735 * @cpu: the processor in question.
7736 * @p: the task pointer to set.
7738 * Description: This function must only be used when non-maskable interrupts
7739 * are serviced on a separate stack. It allows the architecture to switch the
7740 * notion of the current task on a cpu in a non-blocking manner. This function
7741 * must be called with all CPU's synchronized, and interrupts disabled, the
7742 * and caller must save the original value of the current task (see
7743 * curr_task() above) and restore that value before reenabling interrupts and
7744 * re-starting the system.
7746 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7748 void set_curr_task(int cpu
, struct task_struct
*p
)
7755 #ifdef CONFIG_CGROUP_SCHED
7756 /* task_group_lock serializes the addition/removal of task groups */
7757 static DEFINE_SPINLOCK(task_group_lock
);
7759 static void sched_free_group(struct task_group
*tg
)
7761 free_fair_sched_group(tg
);
7762 free_rt_sched_group(tg
);
7764 kmem_cache_free(task_group_cache
, tg
);
7767 /* allocate runqueue etc for a new task group */
7768 struct task_group
*sched_create_group(struct task_group
*parent
)
7770 struct task_group
*tg
;
7772 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7774 return ERR_PTR(-ENOMEM
);
7776 if (!alloc_fair_sched_group(tg
, parent
))
7779 if (!alloc_rt_sched_group(tg
, parent
))
7785 sched_free_group(tg
);
7786 return ERR_PTR(-ENOMEM
);
7789 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7791 unsigned long flags
;
7793 spin_lock_irqsave(&task_group_lock
, flags
);
7794 list_add_rcu(&tg
->list
, &task_groups
);
7796 WARN_ON(!parent
); /* root should already exist */
7798 tg
->parent
= parent
;
7799 INIT_LIST_HEAD(&tg
->children
);
7800 list_add_rcu(&tg
->siblings
, &parent
->children
);
7801 spin_unlock_irqrestore(&task_group_lock
, flags
);
7803 online_fair_sched_group(tg
);
7806 /* rcu callback to free various structures associated with a task group */
7807 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7809 /* now it should be safe to free those cfs_rqs */
7810 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7813 void sched_destroy_group(struct task_group
*tg
)
7815 /* wait for possible concurrent references to cfs_rqs complete */
7816 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7819 void sched_offline_group(struct task_group
*tg
)
7821 unsigned long flags
;
7823 /* end participation in shares distribution */
7824 unregister_fair_sched_group(tg
);
7826 spin_lock_irqsave(&task_group_lock
, flags
);
7827 list_del_rcu(&tg
->list
);
7828 list_del_rcu(&tg
->siblings
);
7829 spin_unlock_irqrestore(&task_group_lock
, flags
);
7832 static void sched_change_group(struct task_struct
*tsk
, int type
)
7834 struct task_group
*tg
;
7837 * All callers are synchronized by task_rq_lock(); we do not use RCU
7838 * which is pointless here. Thus, we pass "true" to task_css_check()
7839 * to prevent lockdep warnings.
7841 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7842 struct task_group
, css
);
7843 tg
= autogroup_task_group(tsk
, tg
);
7844 tsk
->sched_task_group
= tg
;
7846 #ifdef CONFIG_FAIR_GROUP_SCHED
7847 if (tsk
->sched_class
->task_change_group
)
7848 tsk
->sched_class
->task_change_group(tsk
, type
);
7851 set_task_rq(tsk
, task_cpu(tsk
));
7855 * Change task's runqueue when it moves between groups.
7857 * The caller of this function should have put the task in its new group by
7858 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7861 void sched_move_task(struct task_struct
*tsk
)
7863 int queued
, running
;
7867 rq
= task_rq_lock(tsk
, &rf
);
7869 running
= task_current(rq
, tsk
);
7870 queued
= task_on_rq_queued(tsk
);
7873 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7874 if (unlikely(running
))
7875 put_prev_task(rq
, tsk
);
7877 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7879 if (unlikely(running
))
7880 tsk
->sched_class
->set_curr_task(rq
);
7882 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7884 task_rq_unlock(rq
, tsk
, &rf
);
7886 #endif /* CONFIG_CGROUP_SCHED */
7888 #ifdef CONFIG_RT_GROUP_SCHED
7890 * Ensure that the real time constraints are schedulable.
7892 static DEFINE_MUTEX(rt_constraints_mutex
);
7894 /* Must be called with tasklist_lock held */
7895 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7897 struct task_struct
*g
, *p
;
7900 * Autogroups do not have RT tasks; see autogroup_create().
7902 if (task_group_is_autogroup(tg
))
7905 for_each_process_thread(g
, p
) {
7906 if (rt_task(p
) && task_group(p
) == tg
)
7913 struct rt_schedulable_data
{
7914 struct task_group
*tg
;
7919 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7921 struct rt_schedulable_data
*d
= data
;
7922 struct task_group
*child
;
7923 unsigned long total
, sum
= 0;
7924 u64 period
, runtime
;
7926 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7927 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7930 period
= d
->rt_period
;
7931 runtime
= d
->rt_runtime
;
7935 * Cannot have more runtime than the period.
7937 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7941 * Ensure we don't starve existing RT tasks.
7943 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7946 total
= to_ratio(period
, runtime
);
7949 * Nobody can have more than the global setting allows.
7951 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7955 * The sum of our children's runtime should not exceed our own.
7957 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7958 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7959 runtime
= child
->rt_bandwidth
.rt_runtime
;
7961 if (child
== d
->tg
) {
7962 period
= d
->rt_period
;
7963 runtime
= d
->rt_runtime
;
7966 sum
+= to_ratio(period
, runtime
);
7975 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7979 struct rt_schedulable_data data
= {
7981 .rt_period
= period
,
7982 .rt_runtime
= runtime
,
7986 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7992 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7993 u64 rt_period
, u64 rt_runtime
)
7998 * Disallowing the root group RT runtime is BAD, it would disallow the
7999 * kernel creating (and or operating) RT threads.
8001 if (tg
== &root_task_group
&& rt_runtime
== 0)
8004 /* No period doesn't make any sense. */
8008 mutex_lock(&rt_constraints_mutex
);
8009 read_lock(&tasklist_lock
);
8010 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8014 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8015 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8016 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8018 for_each_possible_cpu(i
) {
8019 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8021 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8022 rt_rq
->rt_runtime
= rt_runtime
;
8023 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8025 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8027 read_unlock(&tasklist_lock
);
8028 mutex_unlock(&rt_constraints_mutex
);
8033 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8035 u64 rt_runtime
, rt_period
;
8037 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8038 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8039 if (rt_runtime_us
< 0)
8040 rt_runtime
= RUNTIME_INF
;
8042 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8045 static long sched_group_rt_runtime(struct task_group
*tg
)
8049 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8052 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8053 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8054 return rt_runtime_us
;
8057 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
8059 u64 rt_runtime
, rt_period
;
8061 rt_period
= rt_period_us
* NSEC_PER_USEC
;
8062 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8064 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8067 static long sched_group_rt_period(struct task_group
*tg
)
8071 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8072 do_div(rt_period_us
, NSEC_PER_USEC
);
8073 return rt_period_us
;
8075 #endif /* CONFIG_RT_GROUP_SCHED */
8077 #ifdef CONFIG_RT_GROUP_SCHED
8078 static int sched_rt_global_constraints(void)
8082 mutex_lock(&rt_constraints_mutex
);
8083 read_lock(&tasklist_lock
);
8084 ret
= __rt_schedulable(NULL
, 0, 0);
8085 read_unlock(&tasklist_lock
);
8086 mutex_unlock(&rt_constraints_mutex
);
8091 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8093 /* Don't accept realtime tasks when there is no way for them to run */
8094 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8100 #else /* !CONFIG_RT_GROUP_SCHED */
8101 static int sched_rt_global_constraints(void)
8103 unsigned long flags
;
8106 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8107 for_each_possible_cpu(i
) {
8108 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8110 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8111 rt_rq
->rt_runtime
= global_rt_runtime();
8112 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8114 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8118 #endif /* CONFIG_RT_GROUP_SCHED */
8120 static int sched_dl_global_validate(void)
8122 u64 runtime
= global_rt_runtime();
8123 u64 period
= global_rt_period();
8124 u64 new_bw
= to_ratio(period
, runtime
);
8127 unsigned long flags
;
8130 * Here we want to check the bandwidth not being set to some
8131 * value smaller than the currently allocated bandwidth in
8132 * any of the root_domains.
8134 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8135 * cycling on root_domains... Discussion on different/better
8136 * solutions is welcome!
8138 for_each_possible_cpu(cpu
) {
8139 rcu_read_lock_sched();
8140 dl_b
= dl_bw_of(cpu
);
8142 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8143 if (new_bw
< dl_b
->total_bw
)
8145 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8147 rcu_read_unlock_sched();
8156 static void sched_dl_do_global(void)
8161 unsigned long flags
;
8163 def_dl_bandwidth
.dl_period
= global_rt_period();
8164 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8166 if (global_rt_runtime() != RUNTIME_INF
)
8167 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8170 * FIXME: As above...
8172 for_each_possible_cpu(cpu
) {
8173 rcu_read_lock_sched();
8174 dl_b
= dl_bw_of(cpu
);
8176 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8178 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8180 rcu_read_unlock_sched();
8184 static int sched_rt_global_validate(void)
8186 if (sysctl_sched_rt_period
<= 0)
8189 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8190 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8196 static void sched_rt_do_global(void)
8198 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8199 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8202 int sched_rt_handler(struct ctl_table
*table
, int write
,
8203 void __user
*buffer
, size_t *lenp
,
8206 int old_period
, old_runtime
;
8207 static DEFINE_MUTEX(mutex
);
8211 old_period
= sysctl_sched_rt_period
;
8212 old_runtime
= sysctl_sched_rt_runtime
;
8214 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8216 if (!ret
&& write
) {
8217 ret
= sched_rt_global_validate();
8221 ret
= sched_dl_global_validate();
8225 ret
= sched_rt_global_constraints();
8229 sched_rt_do_global();
8230 sched_dl_do_global();
8234 sysctl_sched_rt_period
= old_period
;
8235 sysctl_sched_rt_runtime
= old_runtime
;
8237 mutex_unlock(&mutex
);
8242 int sched_rr_handler(struct ctl_table
*table
, int write
,
8243 void __user
*buffer
, size_t *lenp
,
8247 static DEFINE_MUTEX(mutex
);
8250 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8251 /* make sure that internally we keep jiffies */
8252 /* also, writing zero resets timeslice to default */
8253 if (!ret
&& write
) {
8254 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8255 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8257 mutex_unlock(&mutex
);
8261 #ifdef CONFIG_CGROUP_SCHED
8263 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8265 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8268 static struct cgroup_subsys_state
*
8269 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8271 struct task_group
*parent
= css_tg(parent_css
);
8272 struct task_group
*tg
;
8275 /* This is early initialization for the top cgroup */
8276 return &root_task_group
.css
;
8279 tg
= sched_create_group(parent
);
8281 return ERR_PTR(-ENOMEM
);
8283 sched_online_group(tg
, parent
);
8288 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8290 struct task_group
*tg
= css_tg(css
);
8292 sched_offline_group(tg
);
8295 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8297 struct task_group
*tg
= css_tg(css
);
8300 * Relies on the RCU grace period between css_released() and this.
8302 sched_free_group(tg
);
8306 * This is called before wake_up_new_task(), therefore we really only
8307 * have to set its group bits, all the other stuff does not apply.
8309 static void cpu_cgroup_fork(struct task_struct
*task
)
8314 rq
= task_rq_lock(task
, &rf
);
8316 sched_change_group(task
, TASK_SET_GROUP
);
8318 task_rq_unlock(rq
, task
, &rf
);
8321 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8323 struct task_struct
*task
;
8324 struct cgroup_subsys_state
*css
;
8327 cgroup_taskset_for_each(task
, css
, tset
) {
8328 #ifdef CONFIG_RT_GROUP_SCHED
8329 if (!sched_rt_can_attach(css_tg(css
), task
))
8332 /* We don't support RT-tasks being in separate groups */
8333 if (task
->sched_class
!= &fair_sched_class
)
8337 * Serialize against wake_up_new_task() such that if its
8338 * running, we're sure to observe its full state.
8340 raw_spin_lock_irq(&task
->pi_lock
);
8342 * Avoid calling sched_move_task() before wake_up_new_task()
8343 * has happened. This would lead to problems with PELT, due to
8344 * move wanting to detach+attach while we're not attached yet.
8346 if (task
->state
== TASK_NEW
)
8348 raw_spin_unlock_irq(&task
->pi_lock
);
8356 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8358 struct task_struct
*task
;
8359 struct cgroup_subsys_state
*css
;
8361 cgroup_taskset_for_each(task
, css
, tset
)
8362 sched_move_task(task
);
8365 #ifdef CONFIG_FAIR_GROUP_SCHED
8366 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8367 struct cftype
*cftype
, u64 shareval
)
8369 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8372 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8375 struct task_group
*tg
= css_tg(css
);
8377 return (u64
) scale_load_down(tg
->shares
);
8380 #ifdef CONFIG_CFS_BANDWIDTH
8381 static DEFINE_MUTEX(cfs_constraints_mutex
);
8383 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8384 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8386 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8388 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8390 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8391 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8393 if (tg
== &root_task_group
)
8397 * Ensure we have at some amount of bandwidth every period. This is
8398 * to prevent reaching a state of large arrears when throttled via
8399 * entity_tick() resulting in prolonged exit starvation.
8401 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8405 * Likewise, bound things on the otherside by preventing insane quota
8406 * periods. This also allows us to normalize in computing quota
8409 if (period
> max_cfs_quota_period
)
8413 * Prevent race between setting of cfs_rq->runtime_enabled and
8414 * unthrottle_offline_cfs_rqs().
8417 mutex_lock(&cfs_constraints_mutex
);
8418 ret
= __cfs_schedulable(tg
, period
, quota
);
8422 runtime_enabled
= quota
!= RUNTIME_INF
;
8423 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8425 * If we need to toggle cfs_bandwidth_used, off->on must occur
8426 * before making related changes, and on->off must occur afterwards
8428 if (runtime_enabled
&& !runtime_was_enabled
)
8429 cfs_bandwidth_usage_inc();
8430 raw_spin_lock_irq(&cfs_b
->lock
);
8431 cfs_b
->period
= ns_to_ktime(period
);
8432 cfs_b
->quota
= quota
;
8434 __refill_cfs_bandwidth_runtime(cfs_b
);
8435 /* restart the period timer (if active) to handle new period expiry */
8436 if (runtime_enabled
)
8437 start_cfs_bandwidth(cfs_b
);
8438 raw_spin_unlock_irq(&cfs_b
->lock
);
8440 for_each_online_cpu(i
) {
8441 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8442 struct rq
*rq
= cfs_rq
->rq
;
8444 raw_spin_lock_irq(&rq
->lock
);
8445 cfs_rq
->runtime_enabled
= runtime_enabled
;
8446 cfs_rq
->runtime_remaining
= 0;
8448 if (cfs_rq
->throttled
)
8449 unthrottle_cfs_rq(cfs_rq
);
8450 raw_spin_unlock_irq(&rq
->lock
);
8452 if (runtime_was_enabled
&& !runtime_enabled
)
8453 cfs_bandwidth_usage_dec();
8455 mutex_unlock(&cfs_constraints_mutex
);
8461 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8465 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8466 if (cfs_quota_us
< 0)
8467 quota
= RUNTIME_INF
;
8469 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8471 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8474 long tg_get_cfs_quota(struct task_group
*tg
)
8478 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8481 quota_us
= tg
->cfs_bandwidth
.quota
;
8482 do_div(quota_us
, NSEC_PER_USEC
);
8487 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8491 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8492 quota
= tg
->cfs_bandwidth
.quota
;
8494 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8497 long tg_get_cfs_period(struct task_group
*tg
)
8501 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8502 do_div(cfs_period_us
, NSEC_PER_USEC
);
8504 return cfs_period_us
;
8507 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8510 return tg_get_cfs_quota(css_tg(css
));
8513 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8514 struct cftype
*cftype
, s64 cfs_quota_us
)
8516 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8519 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8522 return tg_get_cfs_period(css_tg(css
));
8525 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8526 struct cftype
*cftype
, u64 cfs_period_us
)
8528 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8531 struct cfs_schedulable_data
{
8532 struct task_group
*tg
;
8537 * normalize group quota/period to be quota/max_period
8538 * note: units are usecs
8540 static u64
normalize_cfs_quota(struct task_group
*tg
,
8541 struct cfs_schedulable_data
*d
)
8549 period
= tg_get_cfs_period(tg
);
8550 quota
= tg_get_cfs_quota(tg
);
8553 /* note: these should typically be equivalent */
8554 if (quota
== RUNTIME_INF
|| quota
== -1)
8557 return to_ratio(period
, quota
);
8560 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8562 struct cfs_schedulable_data
*d
= data
;
8563 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8564 s64 quota
= 0, parent_quota
= -1;
8567 quota
= RUNTIME_INF
;
8569 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8571 quota
= normalize_cfs_quota(tg
, d
);
8572 parent_quota
= parent_b
->hierarchical_quota
;
8575 * ensure max(child_quota) <= parent_quota, inherit when no
8578 if (quota
== RUNTIME_INF
)
8579 quota
= parent_quota
;
8580 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8583 cfs_b
->hierarchical_quota
= quota
;
8588 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8591 struct cfs_schedulable_data data
= {
8597 if (quota
!= RUNTIME_INF
) {
8598 do_div(data
.period
, NSEC_PER_USEC
);
8599 do_div(data
.quota
, NSEC_PER_USEC
);
8603 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8609 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8611 struct task_group
*tg
= css_tg(seq_css(sf
));
8612 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8614 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8615 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8616 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8620 #endif /* CONFIG_CFS_BANDWIDTH */
8621 #endif /* CONFIG_FAIR_GROUP_SCHED */
8623 #ifdef CONFIG_RT_GROUP_SCHED
8624 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8625 struct cftype
*cft
, s64 val
)
8627 return sched_group_set_rt_runtime(css_tg(css
), val
);
8630 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8633 return sched_group_rt_runtime(css_tg(css
));
8636 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8637 struct cftype
*cftype
, u64 rt_period_us
)
8639 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8642 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8645 return sched_group_rt_period(css_tg(css
));
8647 #endif /* CONFIG_RT_GROUP_SCHED */
8649 static struct cftype cpu_files
[] = {
8650 #ifdef CONFIG_FAIR_GROUP_SCHED
8653 .read_u64
= cpu_shares_read_u64
,
8654 .write_u64
= cpu_shares_write_u64
,
8657 #ifdef CONFIG_CFS_BANDWIDTH
8659 .name
= "cfs_quota_us",
8660 .read_s64
= cpu_cfs_quota_read_s64
,
8661 .write_s64
= cpu_cfs_quota_write_s64
,
8664 .name
= "cfs_period_us",
8665 .read_u64
= cpu_cfs_period_read_u64
,
8666 .write_u64
= cpu_cfs_period_write_u64
,
8670 .seq_show
= cpu_stats_show
,
8673 #ifdef CONFIG_RT_GROUP_SCHED
8675 .name
= "rt_runtime_us",
8676 .read_s64
= cpu_rt_runtime_read
,
8677 .write_s64
= cpu_rt_runtime_write
,
8680 .name
= "rt_period_us",
8681 .read_u64
= cpu_rt_period_read_uint
,
8682 .write_u64
= cpu_rt_period_write_uint
,
8688 struct cgroup_subsys cpu_cgrp_subsys
= {
8689 .css_alloc
= cpu_cgroup_css_alloc
,
8690 .css_released
= cpu_cgroup_css_released
,
8691 .css_free
= cpu_cgroup_css_free
,
8692 .fork
= cpu_cgroup_fork
,
8693 .can_attach
= cpu_cgroup_can_attach
,
8694 .attach
= cpu_cgroup_attach
,
8695 .legacy_cftypes
= cpu_files
,
8699 #endif /* CONFIG_CGROUP_SCHED */
8701 void dump_cpu_task(int cpu
)
8703 pr_info("Task dump for CPU %d:\n", cpu
);
8704 sched_show_task(cpu_curr(cpu
));
8708 * Nice levels are multiplicative, with a gentle 10% change for every
8709 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8710 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8711 * that remained on nice 0.
8713 * The "10% effect" is relative and cumulative: from _any_ nice level,
8714 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8715 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8716 * If a task goes up by ~10% and another task goes down by ~10% then
8717 * the relative distance between them is ~25%.)
8719 const int sched_prio_to_weight
[40] = {
8720 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8721 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8722 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8723 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8724 /* 0 */ 1024, 820, 655, 526, 423,
8725 /* 5 */ 335, 272, 215, 172, 137,
8726 /* 10 */ 110, 87, 70, 56, 45,
8727 /* 15 */ 36, 29, 23, 18, 15,
8731 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8733 * In cases where the weight does not change often, we can use the
8734 * precalculated inverse to speed up arithmetics by turning divisions
8735 * into multiplications:
8737 const u32 sched_prio_to_wmult
[40] = {
8738 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8739 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8740 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8741 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8742 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8743 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8744 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8745 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,