4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex
);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
96 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
98 void update_rq_clock(struct rq
*rq
)
102 lockdep_assert_held(&rq
->lock
);
104 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
107 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
111 update_rq_clock_task(rq
, delta
);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug
unsigned int sysctl_sched_features
=
122 #include "features.h"
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
131 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
134 * period over which we average the RT time consumption, measured
139 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
142 * period over which we measure -rt task cpu usage in us.
145 unsigned int sysctl_sched_rt_period
= 1000000;
147 __read_mostly
int scheduler_running
;
150 * part of the period that we allow rt tasks to run in us.
153 int sysctl_sched_rt_runtime
= 950000;
155 /* cpus with isolated domains */
156 cpumask_var_t cpu_isolated_map
;
159 * this_rq_lock - lock this runqueue and disable interrupts.
161 static struct rq
*this_rq_lock(void)
168 raw_spin_lock(&rq
->lock
);
174 * __task_rq_lock - lock the rq @p resides on.
176 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
181 lockdep_assert_held(&p
->pi_lock
);
185 raw_spin_lock(&rq
->lock
);
186 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
187 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
190 raw_spin_unlock(&rq
->lock
);
192 while (unlikely(task_on_rq_migrating(p
)))
198 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
200 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
201 __acquires(p
->pi_lock
)
207 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
209 raw_spin_lock(&rq
->lock
);
211 * move_queued_task() task_rq_lock()
214 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
215 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
216 * [S] ->cpu = new_cpu [L] task_rq()
220 * If we observe the old cpu in task_rq_lock, the acquire of
221 * the old rq->lock will fully serialize against the stores.
223 * If we observe the new cpu in task_rq_lock, the acquire will
224 * pair with the WMB to ensure we must then also see migrating.
226 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
227 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
230 raw_spin_unlock(&rq
->lock
);
231 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
233 while (unlikely(task_on_rq_migrating(p
)))
238 #ifdef CONFIG_SCHED_HRTICK
240 * Use HR-timers to deliver accurate preemption points.
243 static void hrtick_clear(struct rq
*rq
)
245 if (hrtimer_active(&rq
->hrtick_timer
))
246 hrtimer_cancel(&rq
->hrtick_timer
);
250 * High-resolution timer tick.
251 * Runs from hardirq context with interrupts disabled.
253 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
255 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
257 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
259 raw_spin_lock(&rq
->lock
);
261 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
262 raw_spin_unlock(&rq
->lock
);
264 return HRTIMER_NORESTART
;
269 static void __hrtick_restart(struct rq
*rq
)
271 struct hrtimer
*timer
= &rq
->hrtick_timer
;
273 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
277 * called from hardirq (IPI) context
279 static void __hrtick_start(void *arg
)
283 raw_spin_lock(&rq
->lock
);
284 __hrtick_restart(rq
);
285 rq
->hrtick_csd_pending
= 0;
286 raw_spin_unlock(&rq
->lock
);
290 * Called to set the hrtick timer state.
292 * called with rq->lock held and irqs disabled
294 void hrtick_start(struct rq
*rq
, u64 delay
)
296 struct hrtimer
*timer
= &rq
->hrtick_timer
;
301 * Don't schedule slices shorter than 10000ns, that just
302 * doesn't make sense and can cause timer DoS.
304 delta
= max_t(s64
, delay
, 10000LL);
305 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
307 hrtimer_set_expires(timer
, time
);
309 if (rq
== this_rq()) {
310 __hrtick_restart(rq
);
311 } else if (!rq
->hrtick_csd_pending
) {
312 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
313 rq
->hrtick_csd_pending
= 1;
319 * Called to set the hrtick timer state.
321 * called with rq->lock held and irqs disabled
323 void hrtick_start(struct rq
*rq
, u64 delay
)
326 * Don't schedule slices shorter than 10000ns, that just
327 * doesn't make sense. Rely on vruntime for fairness.
329 delay
= max_t(u64
, delay
, 10000LL);
330 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
331 HRTIMER_MODE_REL_PINNED
);
333 #endif /* CONFIG_SMP */
335 static void init_rq_hrtick(struct rq
*rq
)
338 rq
->hrtick_csd_pending
= 0;
340 rq
->hrtick_csd
.flags
= 0;
341 rq
->hrtick_csd
.func
= __hrtick_start
;
342 rq
->hrtick_csd
.info
= rq
;
345 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
346 rq
->hrtick_timer
.function
= hrtick
;
348 #else /* CONFIG_SCHED_HRTICK */
349 static inline void hrtick_clear(struct rq
*rq
)
353 static inline void init_rq_hrtick(struct rq
*rq
)
356 #endif /* CONFIG_SCHED_HRTICK */
359 * cmpxchg based fetch_or, macro so it works for different integer types
361 #define fetch_or(ptr, mask) \
363 typeof(ptr) _ptr = (ptr); \
364 typeof(mask) _mask = (mask); \
365 typeof(*_ptr) _old, _val = *_ptr; \
368 _old = cmpxchg(_ptr, _val, _val | _mask); \
376 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
378 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
379 * this avoids any races wrt polling state changes and thereby avoids
382 static bool set_nr_and_not_polling(struct task_struct
*p
)
384 struct thread_info
*ti
= task_thread_info(p
);
385 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
389 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
391 * If this returns true, then the idle task promises to call
392 * sched_ttwu_pending() and reschedule soon.
394 static bool set_nr_if_polling(struct task_struct
*p
)
396 struct thread_info
*ti
= task_thread_info(p
);
397 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
400 if (!(val
& _TIF_POLLING_NRFLAG
))
402 if (val
& _TIF_NEED_RESCHED
)
404 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
413 static bool set_nr_and_not_polling(struct task_struct
*p
)
415 set_tsk_need_resched(p
);
420 static bool set_nr_if_polling(struct task_struct
*p
)
427 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
429 struct wake_q_node
*node
= &task
->wake_q
;
432 * Atomically grab the task, if ->wake_q is !nil already it means
433 * its already queued (either by us or someone else) and will get the
434 * wakeup due to that.
436 * This cmpxchg() implies a full barrier, which pairs with the write
437 * barrier implied by the wakeup in wake_up_q().
439 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
442 get_task_struct(task
);
445 * The head is context local, there can be no concurrency.
448 head
->lastp
= &node
->next
;
451 void wake_up_q(struct wake_q_head
*head
)
453 struct wake_q_node
*node
= head
->first
;
455 while (node
!= WAKE_Q_TAIL
) {
456 struct task_struct
*task
;
458 task
= container_of(node
, struct task_struct
, wake_q
);
460 /* task can safely be re-inserted now */
462 task
->wake_q
.next
= NULL
;
465 * wake_up_process() implies a wmb() to pair with the queueing
466 * in wake_q_add() so as not to miss wakeups.
468 wake_up_process(task
);
469 put_task_struct(task
);
474 * resched_curr - mark rq's current task 'to be rescheduled now'.
476 * On UP this means the setting of the need_resched flag, on SMP it
477 * might also involve a cross-CPU call to trigger the scheduler on
480 void resched_curr(struct rq
*rq
)
482 struct task_struct
*curr
= rq
->curr
;
485 lockdep_assert_held(&rq
->lock
);
487 if (test_tsk_need_resched(curr
))
492 if (cpu
== smp_processor_id()) {
493 set_tsk_need_resched(curr
);
494 set_preempt_need_resched();
498 if (set_nr_and_not_polling(curr
))
499 smp_send_reschedule(cpu
);
501 trace_sched_wake_idle_without_ipi(cpu
);
504 void resched_cpu(int cpu
)
506 struct rq
*rq
= cpu_rq(cpu
);
509 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
512 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
516 #ifdef CONFIG_NO_HZ_COMMON
518 * In the semi idle case, use the nearest busy cpu for migrating timers
519 * from an idle cpu. This is good for power-savings.
521 * We don't do similar optimization for completely idle system, as
522 * selecting an idle cpu will add more delays to the timers than intended
523 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
525 int get_nohz_timer_target(void)
527 int i
, cpu
= smp_processor_id();
528 struct sched_domain
*sd
;
530 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
534 for_each_domain(cpu
, sd
) {
535 for_each_cpu(i
, sched_domain_span(sd
)) {
539 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
546 if (!is_housekeeping_cpu(cpu
))
547 cpu
= housekeeping_any_cpu();
553 * When add_timer_on() enqueues a timer into the timer wheel of an
554 * idle CPU then this timer might expire before the next timer event
555 * which is scheduled to wake up that CPU. In case of a completely
556 * idle system the next event might even be infinite time into the
557 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
558 * leaves the inner idle loop so the newly added timer is taken into
559 * account when the CPU goes back to idle and evaluates the timer
560 * wheel for the next timer event.
562 static void wake_up_idle_cpu(int cpu
)
564 struct rq
*rq
= cpu_rq(cpu
);
566 if (cpu
== smp_processor_id())
569 if (set_nr_and_not_polling(rq
->idle
))
570 smp_send_reschedule(cpu
);
572 trace_sched_wake_idle_without_ipi(cpu
);
575 static bool wake_up_full_nohz_cpu(int cpu
)
578 * We just need the target to call irq_exit() and re-evaluate
579 * the next tick. The nohz full kick at least implies that.
580 * If needed we can still optimize that later with an
583 if (tick_nohz_full_cpu(cpu
)) {
584 if (cpu
!= smp_processor_id() ||
585 tick_nohz_tick_stopped())
586 tick_nohz_full_kick_cpu(cpu
);
593 void wake_up_nohz_cpu(int cpu
)
595 if (!wake_up_full_nohz_cpu(cpu
))
596 wake_up_idle_cpu(cpu
);
599 static inline bool got_nohz_idle_kick(void)
601 int cpu
= smp_processor_id();
603 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
606 if (idle_cpu(cpu
) && !need_resched())
610 * We can't run Idle Load Balance on this CPU for this time so we
611 * cancel it and clear NOHZ_BALANCE_KICK
613 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
617 #else /* CONFIG_NO_HZ_COMMON */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ_COMMON */
626 #ifdef CONFIG_NO_HZ_FULL
627 bool sched_can_stop_tick(struct rq
*rq
)
631 /* Deadline tasks, even if single, need the tick */
632 if (rq
->dl
.dl_nr_running
)
636 * If there are more than one RR tasks, we need the tick to effect the
637 * actual RR behaviour.
639 if (rq
->rt
.rr_nr_running
) {
640 if (rq
->rt
.rr_nr_running
== 1)
647 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
648 * forced preemption between FIFO tasks.
650 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
655 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
656 * if there's more than one we need the tick for involuntary
659 if (rq
->nr_running
> 1)
664 #endif /* CONFIG_NO_HZ_FULL */
666 void sched_avg_update(struct rq
*rq
)
668 s64 period
= sched_avg_period();
670 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
672 * Inline assembly required to prevent the compiler
673 * optimising this loop into a divmod call.
674 * See __iter_div_u64_rem() for another example of this.
676 asm("" : "+rm" (rq
->age_stamp
));
677 rq
->age_stamp
+= period
;
682 #endif /* CONFIG_SMP */
684 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
685 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
687 * Iterate task_group tree rooted at *from, calling @down when first entering a
688 * node and @up when leaving it for the final time.
690 * Caller must hold rcu_lock or sufficient equivalent.
692 int walk_tg_tree_from(struct task_group
*from
,
693 tg_visitor down
, tg_visitor up
, void *data
)
695 struct task_group
*parent
, *child
;
701 ret
= (*down
)(parent
, data
);
704 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
711 ret
= (*up
)(parent
, data
);
712 if (ret
|| parent
== from
)
716 parent
= parent
->parent
;
723 int tg_nop(struct task_group
*tg
, void *data
)
729 static void set_load_weight(struct task_struct
*p
)
731 int prio
= p
->static_prio
- MAX_RT_PRIO
;
732 struct load_weight
*load
= &p
->se
.load
;
735 * SCHED_IDLE tasks get minimal weight:
737 if (idle_policy(p
->policy
)) {
738 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
739 load
->inv_weight
= WMULT_IDLEPRIO
;
743 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
744 load
->inv_weight
= sched_prio_to_wmult
[prio
];
747 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
750 if (!(flags
& ENQUEUE_RESTORE
))
751 sched_info_queued(rq
, p
);
752 p
->sched_class
->enqueue_task(rq
, p
, flags
);
755 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
758 if (!(flags
& DEQUEUE_SAVE
))
759 sched_info_dequeued(rq
, p
);
760 p
->sched_class
->dequeue_task(rq
, p
, flags
);
763 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
765 if (task_contributes_to_load(p
))
766 rq
->nr_uninterruptible
--;
768 enqueue_task(rq
, p
, flags
);
771 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
773 if (task_contributes_to_load(p
))
774 rq
->nr_uninterruptible
++;
776 dequeue_task(rq
, p
, flags
);
779 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
782 * In theory, the compile should just see 0 here, and optimize out the call
783 * to sched_rt_avg_update. But I don't trust it...
785 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
786 s64 steal
= 0, irq_delta
= 0;
788 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
789 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
792 * Since irq_time is only updated on {soft,}irq_exit, we might run into
793 * this case when a previous update_rq_clock() happened inside a
796 * When this happens, we stop ->clock_task and only update the
797 * prev_irq_time stamp to account for the part that fit, so that a next
798 * update will consume the rest. This ensures ->clock_task is
801 * It does however cause some slight miss-attribution of {soft,}irq
802 * time, a more accurate solution would be to update the irq_time using
803 * the current rq->clock timestamp, except that would require using
806 if (irq_delta
> delta
)
809 rq
->prev_irq_time
+= irq_delta
;
812 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
813 if (static_key_false((¶virt_steal_rq_enabled
))) {
814 steal
= paravirt_steal_clock(cpu_of(rq
));
815 steal
-= rq
->prev_steal_time_rq
;
817 if (unlikely(steal
> delta
))
820 rq
->prev_steal_time_rq
+= steal
;
825 rq
->clock_task
+= delta
;
827 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
828 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
829 sched_rt_avg_update(rq
, irq_delta
+ steal
);
833 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
835 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
836 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
840 * Make it appear like a SCHED_FIFO task, its something
841 * userspace knows about and won't get confused about.
843 * Also, it will make PI more or less work without too
844 * much confusion -- but then, stop work should not
845 * rely on PI working anyway.
847 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
849 stop
->sched_class
= &stop_sched_class
;
852 cpu_rq(cpu
)->stop
= stop
;
856 * Reset it back to a normal scheduling class so that
857 * it can die in pieces.
859 old_stop
->sched_class
= &rt_sched_class
;
864 * __normal_prio - return the priority that is based on the static prio
866 static inline int __normal_prio(struct task_struct
*p
)
868 return p
->static_prio
;
872 * Calculate the expected normal priority: i.e. priority
873 * without taking RT-inheritance into account. Might be
874 * boosted by interactivity modifiers. Changes upon fork,
875 * setprio syscalls, and whenever the interactivity
876 * estimator recalculates.
878 static inline int normal_prio(struct task_struct
*p
)
882 if (task_has_dl_policy(p
))
883 prio
= MAX_DL_PRIO
-1;
884 else if (task_has_rt_policy(p
))
885 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
887 prio
= __normal_prio(p
);
892 * Calculate the current priority, i.e. the priority
893 * taken into account by the scheduler. This value might
894 * be boosted by RT tasks, or might be boosted by
895 * interactivity modifiers. Will be RT if the task got
896 * RT-boosted. If not then it returns p->normal_prio.
898 static int effective_prio(struct task_struct
*p
)
900 p
->normal_prio
= normal_prio(p
);
902 * If we are RT tasks or we were boosted to RT priority,
903 * keep the priority unchanged. Otherwise, update priority
904 * to the normal priority:
906 if (!rt_prio(p
->prio
))
907 return p
->normal_prio
;
912 * task_curr - is this task currently executing on a CPU?
913 * @p: the task in question.
915 * Return: 1 if the task is currently executing. 0 otherwise.
917 inline int task_curr(const struct task_struct
*p
)
919 return cpu_curr(task_cpu(p
)) == p
;
923 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
924 * use the balance_callback list if you want balancing.
926 * this means any call to check_class_changed() must be followed by a call to
927 * balance_callback().
929 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
930 const struct sched_class
*prev_class
,
933 if (prev_class
!= p
->sched_class
) {
934 if (prev_class
->switched_from
)
935 prev_class
->switched_from(rq
, p
);
937 p
->sched_class
->switched_to(rq
, p
);
938 } else if (oldprio
!= p
->prio
|| dl_task(p
))
939 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
942 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
944 const struct sched_class
*class;
946 if (p
->sched_class
== rq
->curr
->sched_class
) {
947 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
949 for_each_class(class) {
950 if (class == rq
->curr
->sched_class
)
952 if (class == p
->sched_class
) {
960 * A queue event has occurred, and we're going to schedule. In
961 * this case, we can save a useless back to back clock update.
963 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
964 rq_clock_skip_update(rq
, true);
969 * This is how migration works:
971 * 1) we invoke migration_cpu_stop() on the target CPU using
973 * 2) stopper starts to run (implicitly forcing the migrated thread
975 * 3) it checks whether the migrated task is still in the wrong runqueue.
976 * 4) if it's in the wrong runqueue then the migration thread removes
977 * it and puts it into the right queue.
978 * 5) stopper completes and stop_one_cpu() returns and the migration
983 * move_queued_task - move a queued task to new rq.
985 * Returns (locked) new rq. Old rq's lock is released.
987 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
989 lockdep_assert_held(&rq
->lock
);
991 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
992 dequeue_task(rq
, p
, 0);
993 set_task_cpu(p
, new_cpu
);
994 raw_spin_unlock(&rq
->lock
);
996 rq
= cpu_rq(new_cpu
);
998 raw_spin_lock(&rq
->lock
);
999 BUG_ON(task_cpu(p
) != new_cpu
);
1000 enqueue_task(rq
, p
, 0);
1001 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1002 check_preempt_curr(rq
, p
, 0);
1007 struct migration_arg
{
1008 struct task_struct
*task
;
1013 * Move (not current) task off this cpu, onto dest cpu. We're doing
1014 * this because either it can't run here any more (set_cpus_allowed()
1015 * away from this CPU, or CPU going down), or because we're
1016 * attempting to rebalance this task on exec (sched_exec).
1018 * So we race with normal scheduler movements, but that's OK, as long
1019 * as the task is no longer on this CPU.
1021 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1023 if (unlikely(!cpu_active(dest_cpu
)))
1026 /* Affinity changed (again). */
1027 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1030 rq
= move_queued_task(rq
, p
, dest_cpu
);
1036 * migration_cpu_stop - this will be executed by a highprio stopper thread
1037 * and performs thread migration by bumping thread off CPU then
1038 * 'pushing' onto another runqueue.
1040 static int migration_cpu_stop(void *data
)
1042 struct migration_arg
*arg
= data
;
1043 struct task_struct
*p
= arg
->task
;
1044 struct rq
*rq
= this_rq();
1047 * The original target cpu might have gone down and we might
1048 * be on another cpu but it doesn't matter.
1050 local_irq_disable();
1052 * We need to explicitly wake pending tasks before running
1053 * __migrate_task() such that we will not miss enforcing cpus_allowed
1054 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1056 sched_ttwu_pending();
1058 raw_spin_lock(&p
->pi_lock
);
1059 raw_spin_lock(&rq
->lock
);
1061 * If task_rq(p) != rq, it cannot be migrated here, because we're
1062 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1063 * we're holding p->pi_lock.
1065 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1066 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1067 raw_spin_unlock(&rq
->lock
);
1068 raw_spin_unlock(&p
->pi_lock
);
1075 * sched_class::set_cpus_allowed must do the below, but is not required to
1076 * actually call this function.
1078 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1080 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1081 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1084 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1086 struct rq
*rq
= task_rq(p
);
1087 bool queued
, running
;
1089 lockdep_assert_held(&p
->pi_lock
);
1091 queued
= task_on_rq_queued(p
);
1092 running
= task_current(rq
, p
);
1096 * Because __kthread_bind() calls this on blocked tasks without
1099 lockdep_assert_held(&rq
->lock
);
1100 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1103 put_prev_task(rq
, p
);
1105 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1108 p
->sched_class
->set_curr_task(rq
);
1110 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1114 * Change a given task's CPU affinity. Migrate the thread to a
1115 * proper CPU and schedule it away if the CPU it's executing on
1116 * is removed from the allowed bitmask.
1118 * NOTE: the caller must have a valid reference to the task, the
1119 * task must not exit() & deallocate itself prematurely. The
1120 * call is not atomic; no spinlocks may be held.
1122 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1123 const struct cpumask
*new_mask
, bool check
)
1125 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1126 unsigned int dest_cpu
;
1131 rq
= task_rq_lock(p
, &rf
);
1133 if (p
->flags
& PF_KTHREAD
) {
1135 * Kernel threads are allowed on online && !active CPUs
1137 cpu_valid_mask
= cpu_online_mask
;
1141 * Must re-check here, to close a race against __kthread_bind(),
1142 * sched_setaffinity() is not guaranteed to observe the flag.
1144 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1149 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1152 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1157 do_set_cpus_allowed(p
, new_mask
);
1159 if (p
->flags
& PF_KTHREAD
) {
1161 * For kernel threads that do indeed end up on online &&
1162 * !active we want to ensure they are strict per-cpu threads.
1164 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1165 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1166 p
->nr_cpus_allowed
!= 1);
1169 /* Can the task run on the task's current CPU? If so, we're done */
1170 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1173 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1174 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1175 struct migration_arg arg
= { p
, dest_cpu
};
1176 /* Need help from migration thread: drop lock and wait. */
1177 task_rq_unlock(rq
, p
, &rf
);
1178 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1179 tlb_migrate_finish(p
->mm
);
1181 } else if (task_on_rq_queued(p
)) {
1183 * OK, since we're going to drop the lock immediately
1184 * afterwards anyway.
1186 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
1187 rq
= move_queued_task(rq
, p
, dest_cpu
);
1188 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
1191 task_rq_unlock(rq
, p
, &rf
);
1196 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1198 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1200 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1202 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1204 #ifdef CONFIG_SCHED_DEBUG
1206 * We should never call set_task_cpu() on a blocked task,
1207 * ttwu() will sort out the placement.
1209 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1213 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1214 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1215 * time relying on p->on_rq.
1217 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1218 p
->sched_class
== &fair_sched_class
&&
1219 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1221 #ifdef CONFIG_LOCKDEP
1223 * The caller should hold either p->pi_lock or rq->lock, when changing
1224 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1226 * sched_move_task() holds both and thus holding either pins the cgroup,
1229 * Furthermore, all task_rq users should acquire both locks, see
1232 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1233 lockdep_is_held(&task_rq(p
)->lock
)));
1237 trace_sched_migrate_task(p
, new_cpu
);
1239 if (task_cpu(p
) != new_cpu
) {
1240 if (p
->sched_class
->migrate_task_rq
)
1241 p
->sched_class
->migrate_task_rq(p
);
1242 p
->se
.nr_migrations
++;
1243 perf_event_task_migrate(p
);
1246 __set_task_cpu(p
, new_cpu
);
1249 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1251 if (task_on_rq_queued(p
)) {
1252 struct rq
*src_rq
, *dst_rq
;
1254 src_rq
= task_rq(p
);
1255 dst_rq
= cpu_rq(cpu
);
1257 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1258 deactivate_task(src_rq
, p
, 0);
1259 set_task_cpu(p
, cpu
);
1260 activate_task(dst_rq
, p
, 0);
1261 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1262 check_preempt_curr(dst_rq
, p
, 0);
1265 * Task isn't running anymore; make it appear like we migrated
1266 * it before it went to sleep. This means on wakeup we make the
1267 * previous cpu our targer instead of where it really is.
1273 struct migration_swap_arg
{
1274 struct task_struct
*src_task
, *dst_task
;
1275 int src_cpu
, dst_cpu
;
1278 static int migrate_swap_stop(void *data
)
1280 struct migration_swap_arg
*arg
= data
;
1281 struct rq
*src_rq
, *dst_rq
;
1284 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1287 src_rq
= cpu_rq(arg
->src_cpu
);
1288 dst_rq
= cpu_rq(arg
->dst_cpu
);
1290 double_raw_lock(&arg
->src_task
->pi_lock
,
1291 &arg
->dst_task
->pi_lock
);
1292 double_rq_lock(src_rq
, dst_rq
);
1294 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1297 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1300 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1303 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1306 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1307 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1312 double_rq_unlock(src_rq
, dst_rq
);
1313 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1314 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1320 * Cross migrate two tasks
1322 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1324 struct migration_swap_arg arg
;
1327 arg
= (struct migration_swap_arg
){
1329 .src_cpu
= task_cpu(cur
),
1331 .dst_cpu
= task_cpu(p
),
1334 if (arg
.src_cpu
== arg
.dst_cpu
)
1338 * These three tests are all lockless; this is OK since all of them
1339 * will be re-checked with proper locks held further down the line.
1341 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1344 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1347 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1350 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1351 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1358 * wait_task_inactive - wait for a thread to unschedule.
1360 * If @match_state is nonzero, it's the @p->state value just checked and
1361 * not expected to change. If it changes, i.e. @p might have woken up,
1362 * then return zero. When we succeed in waiting for @p to be off its CPU,
1363 * we return a positive number (its total switch count). If a second call
1364 * a short while later returns the same number, the caller can be sure that
1365 * @p has remained unscheduled the whole time.
1367 * The caller must ensure that the task *will* unschedule sometime soon,
1368 * else this function might spin for a *long* time. This function can't
1369 * be called with interrupts off, or it may introduce deadlock with
1370 * smp_call_function() if an IPI is sent by the same process we are
1371 * waiting to become inactive.
1373 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1375 int running
, queued
;
1382 * We do the initial early heuristics without holding
1383 * any task-queue locks at all. We'll only try to get
1384 * the runqueue lock when things look like they will
1390 * If the task is actively running on another CPU
1391 * still, just relax and busy-wait without holding
1394 * NOTE! Since we don't hold any locks, it's not
1395 * even sure that "rq" stays as the right runqueue!
1396 * But we don't care, since "task_running()" will
1397 * return false if the runqueue has changed and p
1398 * is actually now running somewhere else!
1400 while (task_running(rq
, p
)) {
1401 if (match_state
&& unlikely(p
->state
!= match_state
))
1407 * Ok, time to look more closely! We need the rq
1408 * lock now, to be *sure*. If we're wrong, we'll
1409 * just go back and repeat.
1411 rq
= task_rq_lock(p
, &rf
);
1412 trace_sched_wait_task(p
);
1413 running
= task_running(rq
, p
);
1414 queued
= task_on_rq_queued(p
);
1416 if (!match_state
|| p
->state
== match_state
)
1417 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1418 task_rq_unlock(rq
, p
, &rf
);
1421 * If it changed from the expected state, bail out now.
1423 if (unlikely(!ncsw
))
1427 * Was it really running after all now that we
1428 * checked with the proper locks actually held?
1430 * Oops. Go back and try again..
1432 if (unlikely(running
)) {
1438 * It's not enough that it's not actively running,
1439 * it must be off the runqueue _entirely_, and not
1442 * So if it was still runnable (but just not actively
1443 * running right now), it's preempted, and we should
1444 * yield - it could be a while.
1446 if (unlikely(queued
)) {
1447 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1449 set_current_state(TASK_UNINTERRUPTIBLE
);
1450 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1455 * Ahh, all good. It wasn't running, and it wasn't
1456 * runnable, which means that it will never become
1457 * running in the future either. We're all done!
1466 * kick_process - kick a running thread to enter/exit the kernel
1467 * @p: the to-be-kicked thread
1469 * Cause a process which is running on another CPU to enter
1470 * kernel-mode, without any delay. (to get signals handled.)
1472 * NOTE: this function doesn't have to take the runqueue lock,
1473 * because all it wants to ensure is that the remote task enters
1474 * the kernel. If the IPI races and the task has been migrated
1475 * to another CPU then no harm is done and the purpose has been
1478 void kick_process(struct task_struct
*p
)
1484 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1485 smp_send_reschedule(cpu
);
1488 EXPORT_SYMBOL_GPL(kick_process
);
1491 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1493 * A few notes on cpu_active vs cpu_online:
1495 * - cpu_active must be a subset of cpu_online
1497 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1498 * see __set_cpus_allowed_ptr(). At this point the newly online
1499 * cpu isn't yet part of the sched domains, and balancing will not
1502 * - on cpu-down we clear cpu_active() to mask the sched domains and
1503 * avoid the load balancer to place new tasks on the to be removed
1504 * cpu. Existing tasks will remain running there and will be taken
1507 * This means that fallback selection must not select !active CPUs.
1508 * And can assume that any active CPU must be online. Conversely
1509 * select_task_rq() below may allow selection of !active CPUs in order
1510 * to satisfy the above rules.
1512 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1514 int nid
= cpu_to_node(cpu
);
1515 const struct cpumask
*nodemask
= NULL
;
1516 enum { cpuset
, possible
, fail
} state
= cpuset
;
1520 * If the node that the cpu is on has been offlined, cpu_to_node()
1521 * will return -1. There is no cpu on the node, and we should
1522 * select the cpu on the other node.
1525 nodemask
= cpumask_of_node(nid
);
1527 /* Look for allowed, online CPU in same node. */
1528 for_each_cpu(dest_cpu
, nodemask
) {
1529 if (!cpu_active(dest_cpu
))
1531 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1537 /* Any allowed, online CPU? */
1538 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1539 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1541 if (!cpu_online(dest_cpu
))
1546 /* No more Mr. Nice Guy. */
1549 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1550 cpuset_cpus_allowed_fallback(p
);
1556 do_set_cpus_allowed(p
, cpu_possible_mask
);
1567 if (state
!= cpuset
) {
1569 * Don't tell them about moving exiting tasks or
1570 * kernel threads (both mm NULL), since they never
1573 if (p
->mm
&& printk_ratelimit()) {
1574 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1575 task_pid_nr(p
), p
->comm
, cpu
);
1583 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1586 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1588 lockdep_assert_held(&p
->pi_lock
);
1590 if (tsk_nr_cpus_allowed(p
) > 1)
1591 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1593 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1596 * In order not to call set_task_cpu() on a blocking task we need
1597 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1600 * Since this is common to all placement strategies, this lives here.
1602 * [ this allows ->select_task() to simply return task_cpu(p) and
1603 * not worry about this generic constraint ]
1605 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1607 cpu
= select_fallback_rq(task_cpu(p
), p
);
1612 static void update_avg(u64
*avg
, u64 sample
)
1614 s64 diff
= sample
- *avg
;
1620 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1621 const struct cpumask
*new_mask
, bool check
)
1623 return set_cpus_allowed_ptr(p
, new_mask
);
1626 #endif /* CONFIG_SMP */
1629 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1631 #ifdef CONFIG_SCHEDSTATS
1632 struct rq
*rq
= this_rq();
1635 int this_cpu
= smp_processor_id();
1637 if (cpu
== this_cpu
) {
1638 schedstat_inc(rq
, ttwu_local
);
1639 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1641 struct sched_domain
*sd
;
1643 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1645 for_each_domain(this_cpu
, sd
) {
1646 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1647 schedstat_inc(sd
, ttwu_wake_remote
);
1654 if (wake_flags
& WF_MIGRATED
)
1655 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1657 #endif /* CONFIG_SMP */
1659 schedstat_inc(rq
, ttwu_count
);
1660 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1662 if (wake_flags
& WF_SYNC
)
1663 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1665 #endif /* CONFIG_SCHEDSTATS */
1668 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1670 activate_task(rq
, p
, en_flags
);
1671 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1673 /* if a worker is waking up, notify workqueue */
1674 if (p
->flags
& PF_WQ_WORKER
)
1675 wq_worker_waking_up(p
, cpu_of(rq
));
1679 * Mark the task runnable and perform wakeup-preemption.
1681 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1682 struct pin_cookie cookie
)
1684 check_preempt_curr(rq
, p
, wake_flags
);
1685 p
->state
= TASK_RUNNING
;
1686 trace_sched_wakeup(p
);
1689 if (p
->sched_class
->task_woken
) {
1691 * Our task @p is fully woken up and running; so its safe to
1692 * drop the rq->lock, hereafter rq is only used for statistics.
1694 lockdep_unpin_lock(&rq
->lock
, cookie
);
1695 p
->sched_class
->task_woken(rq
, p
);
1696 lockdep_repin_lock(&rq
->lock
, cookie
);
1699 if (rq
->idle_stamp
) {
1700 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1701 u64 max
= 2*rq
->max_idle_balance_cost
;
1703 update_avg(&rq
->avg_idle
, delta
);
1705 if (rq
->avg_idle
> max
)
1714 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1715 struct pin_cookie cookie
)
1717 int en_flags
= ENQUEUE_WAKEUP
;
1719 lockdep_assert_held(&rq
->lock
);
1722 if (p
->sched_contributes_to_load
)
1723 rq
->nr_uninterruptible
--;
1725 if (wake_flags
& WF_MIGRATED
)
1726 en_flags
|= ENQUEUE_MIGRATED
;
1729 ttwu_activate(rq
, p
, en_flags
);
1730 ttwu_do_wakeup(rq
, p
, wake_flags
, cookie
);
1734 * Called in case the task @p isn't fully descheduled from its runqueue,
1735 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1736 * since all we need to do is flip p->state to TASK_RUNNING, since
1737 * the task is still ->on_rq.
1739 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1745 rq
= __task_rq_lock(p
, &rf
);
1746 if (task_on_rq_queued(p
)) {
1747 /* check_preempt_curr() may use rq clock */
1748 update_rq_clock(rq
);
1749 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
.cookie
);
1752 __task_rq_unlock(rq
, &rf
);
1758 void sched_ttwu_pending(void)
1760 struct rq
*rq
= this_rq();
1761 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1762 struct pin_cookie cookie
;
1763 struct task_struct
*p
;
1764 unsigned long flags
;
1769 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1770 cookie
= lockdep_pin_lock(&rq
->lock
);
1775 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1776 llist
= llist_next(llist
);
1778 if (p
->sched_remote_wakeup
)
1779 wake_flags
= WF_MIGRATED
;
1781 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1784 lockdep_unpin_lock(&rq
->lock
, cookie
);
1785 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1788 void scheduler_ipi(void)
1791 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1792 * TIF_NEED_RESCHED remotely (for the first time) will also send
1795 preempt_fold_need_resched();
1797 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1801 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1802 * traditionally all their work was done from the interrupt return
1803 * path. Now that we actually do some work, we need to make sure
1806 * Some archs already do call them, luckily irq_enter/exit nest
1809 * Arguably we should visit all archs and update all handlers,
1810 * however a fair share of IPIs are still resched only so this would
1811 * somewhat pessimize the simple resched case.
1814 sched_ttwu_pending();
1817 * Check if someone kicked us for doing the nohz idle load balance.
1819 if (unlikely(got_nohz_idle_kick())) {
1820 this_rq()->idle_balance
= 1;
1821 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1826 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1828 struct rq
*rq
= cpu_rq(cpu
);
1830 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1832 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1833 if (!set_nr_if_polling(rq
->idle
))
1834 smp_send_reschedule(cpu
);
1836 trace_sched_wake_idle_without_ipi(cpu
);
1840 void wake_up_if_idle(int cpu
)
1842 struct rq
*rq
= cpu_rq(cpu
);
1843 unsigned long flags
;
1847 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1850 if (set_nr_if_polling(rq
->idle
)) {
1851 trace_sched_wake_idle_without_ipi(cpu
);
1853 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1854 if (is_idle_task(rq
->curr
))
1855 smp_send_reschedule(cpu
);
1856 /* Else cpu is not in idle, do nothing here */
1857 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1864 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1866 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1868 #endif /* CONFIG_SMP */
1870 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1872 struct rq
*rq
= cpu_rq(cpu
);
1873 struct pin_cookie cookie
;
1875 #if defined(CONFIG_SMP)
1876 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1877 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1878 ttwu_queue_remote(p
, cpu
, wake_flags
);
1883 raw_spin_lock(&rq
->lock
);
1884 cookie
= lockdep_pin_lock(&rq
->lock
);
1885 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1886 lockdep_unpin_lock(&rq
->lock
, cookie
);
1887 raw_spin_unlock(&rq
->lock
);
1891 * Notes on Program-Order guarantees on SMP systems.
1895 * The basic program-order guarantee on SMP systems is that when a task [t]
1896 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1897 * execution on its new cpu [c1].
1899 * For migration (of runnable tasks) this is provided by the following means:
1901 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1902 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1903 * rq(c1)->lock (if not at the same time, then in that order).
1904 * C) LOCK of the rq(c1)->lock scheduling in task
1906 * Transitivity guarantees that B happens after A and C after B.
1907 * Note: we only require RCpc transitivity.
1908 * Note: the cpu doing B need not be c0 or c1
1917 * UNLOCK rq(0)->lock
1919 * LOCK rq(0)->lock // orders against CPU0
1921 * UNLOCK rq(0)->lock
1925 * UNLOCK rq(1)->lock
1927 * LOCK rq(1)->lock // orders against CPU2
1930 * UNLOCK rq(1)->lock
1933 * BLOCKING -- aka. SLEEP + WAKEUP
1935 * For blocking we (obviously) need to provide the same guarantee as for
1936 * migration. However the means are completely different as there is no lock
1937 * chain to provide order. Instead we do:
1939 * 1) smp_store_release(X->on_cpu, 0)
1940 * 2) smp_cond_acquire(!X->on_cpu)
1944 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1946 * LOCK rq(0)->lock LOCK X->pi_lock
1949 * smp_store_release(X->on_cpu, 0);
1951 * smp_cond_acquire(!X->on_cpu);
1957 * X->state = RUNNING
1958 * UNLOCK rq(2)->lock
1960 * LOCK rq(2)->lock // orders against CPU1
1963 * UNLOCK rq(2)->lock
1966 * UNLOCK rq(0)->lock
1969 * However; for wakeups there is a second guarantee we must provide, namely we
1970 * must observe the state that lead to our wakeup. That is, not only must our
1971 * task observe its own prior state, it must also observe the stores prior to
1974 * This means that any means of doing remote wakeups must order the CPU doing
1975 * the wakeup against the CPU the task is going to end up running on. This,
1976 * however, is already required for the regular Program-Order guarantee above,
1977 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1982 * try_to_wake_up - wake up a thread
1983 * @p: the thread to be awakened
1984 * @state: the mask of task states that can be woken
1985 * @wake_flags: wake modifier flags (WF_*)
1987 * Put it on the run-queue if it's not already there. The "current"
1988 * thread is always on the run-queue (except when the actual
1989 * re-schedule is in progress), and as such you're allowed to do
1990 * the simpler "current->state = TASK_RUNNING" to mark yourself
1991 * runnable without the overhead of this.
1993 * Return: %true if @p was woken up, %false if it was already running.
1994 * or @state didn't match @p's state.
1997 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1999 unsigned long flags
;
2000 int cpu
, success
= 0;
2003 * If we are going to wake up a thread waiting for CONDITION we
2004 * need to ensure that CONDITION=1 done by the caller can not be
2005 * reordered with p->state check below. This pairs with mb() in
2006 * set_current_state() the waiting thread does.
2008 smp_mb__before_spinlock();
2009 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2010 if (!(p
->state
& state
))
2013 trace_sched_waking(p
);
2015 success
= 1; /* we're going to change ->state */
2018 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2023 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2024 * possible to, falsely, observe p->on_cpu == 0.
2026 * One must be running (->on_cpu == 1) in order to remove oneself
2027 * from the runqueue.
2029 * [S] ->on_cpu = 1; [L] ->on_rq
2033 * [S] ->on_rq = 0; [L] ->on_cpu
2035 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2036 * from the consecutive calls to schedule(); the first switching to our
2037 * task, the second putting it to sleep.
2042 * If the owning (remote) cpu is still in the middle of schedule() with
2043 * this task as prev, wait until its done referencing the task.
2045 * Pairs with the smp_store_release() in finish_lock_switch().
2047 * This ensures that tasks getting woken will be fully ordered against
2048 * their previous state and preserve Program Order.
2050 smp_cond_acquire(!p
->on_cpu
);
2052 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2053 p
->state
= TASK_WAKING
;
2055 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2056 if (task_cpu(p
) != cpu
) {
2057 wake_flags
|= WF_MIGRATED
;
2058 set_task_cpu(p
, cpu
);
2060 #endif /* CONFIG_SMP */
2062 ttwu_queue(p
, cpu
, wake_flags
);
2064 if (schedstat_enabled())
2065 ttwu_stat(p
, cpu
, wake_flags
);
2067 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2073 * try_to_wake_up_local - try to wake up a local task with rq lock held
2074 * @p: the thread to be awakened
2076 * Put @p on the run-queue if it's not already there. The caller must
2077 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2080 static void try_to_wake_up_local(struct task_struct
*p
, struct pin_cookie cookie
)
2082 struct rq
*rq
= task_rq(p
);
2084 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2085 WARN_ON_ONCE(p
== current
))
2088 lockdep_assert_held(&rq
->lock
);
2090 if (!raw_spin_trylock(&p
->pi_lock
)) {
2092 * This is OK, because current is on_cpu, which avoids it being
2093 * picked for load-balance and preemption/IRQs are still
2094 * disabled avoiding further scheduler activity on it and we've
2095 * not yet picked a replacement task.
2097 lockdep_unpin_lock(&rq
->lock
, cookie
);
2098 raw_spin_unlock(&rq
->lock
);
2099 raw_spin_lock(&p
->pi_lock
);
2100 raw_spin_lock(&rq
->lock
);
2101 lockdep_repin_lock(&rq
->lock
, cookie
);
2104 if (!(p
->state
& TASK_NORMAL
))
2107 trace_sched_waking(p
);
2109 if (!task_on_rq_queued(p
))
2110 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2112 ttwu_do_wakeup(rq
, p
, 0, cookie
);
2113 if (schedstat_enabled())
2114 ttwu_stat(p
, smp_processor_id(), 0);
2116 raw_spin_unlock(&p
->pi_lock
);
2120 * wake_up_process - Wake up a specific process
2121 * @p: The process to be woken up.
2123 * Attempt to wake up the nominated process and move it to the set of runnable
2126 * Return: 1 if the process was woken up, 0 if it was already running.
2128 * It may be assumed that this function implies a write memory barrier before
2129 * changing the task state if and only if any tasks are woken up.
2131 int wake_up_process(struct task_struct
*p
)
2133 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2135 EXPORT_SYMBOL(wake_up_process
);
2137 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2139 return try_to_wake_up(p
, state
, 0);
2143 * This function clears the sched_dl_entity static params.
2145 void __dl_clear_params(struct task_struct
*p
)
2147 struct sched_dl_entity
*dl_se
= &p
->dl
;
2149 dl_se
->dl_runtime
= 0;
2150 dl_se
->dl_deadline
= 0;
2151 dl_se
->dl_period
= 0;
2155 dl_se
->dl_throttled
= 0;
2156 dl_se
->dl_yielded
= 0;
2160 * Perform scheduler related setup for a newly forked process p.
2161 * p is forked by current.
2163 * __sched_fork() is basic setup used by init_idle() too:
2165 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2170 p
->se
.exec_start
= 0;
2171 p
->se
.sum_exec_runtime
= 0;
2172 p
->se
.prev_sum_exec_runtime
= 0;
2173 p
->se
.nr_migrations
= 0;
2175 INIT_LIST_HEAD(&p
->se
.group_node
);
2177 #ifdef CONFIG_FAIR_GROUP_SCHED
2178 p
->se
.cfs_rq
= NULL
;
2181 #ifdef CONFIG_SCHEDSTATS
2182 /* Even if schedstat is disabled, there should not be garbage */
2183 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2186 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2187 init_dl_task_timer(&p
->dl
);
2188 __dl_clear_params(p
);
2190 INIT_LIST_HEAD(&p
->rt
.run_list
);
2192 p
->rt
.time_slice
= sched_rr_timeslice
;
2196 #ifdef CONFIG_PREEMPT_NOTIFIERS
2197 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2200 #ifdef CONFIG_NUMA_BALANCING
2201 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2202 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2203 p
->mm
->numa_scan_seq
= 0;
2206 if (clone_flags
& CLONE_VM
)
2207 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2209 p
->numa_preferred_nid
= -1;
2211 p
->node_stamp
= 0ULL;
2212 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2213 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2214 p
->numa_work
.next
= &p
->numa_work
;
2215 p
->numa_faults
= NULL
;
2216 p
->last_task_numa_placement
= 0;
2217 p
->last_sum_exec_runtime
= 0;
2219 p
->numa_group
= NULL
;
2220 #endif /* CONFIG_NUMA_BALANCING */
2223 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2225 #ifdef CONFIG_NUMA_BALANCING
2227 void set_numabalancing_state(bool enabled
)
2230 static_branch_enable(&sched_numa_balancing
);
2232 static_branch_disable(&sched_numa_balancing
);
2235 #ifdef CONFIG_PROC_SYSCTL
2236 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2237 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2241 int state
= static_branch_likely(&sched_numa_balancing
);
2243 if (write
&& !capable(CAP_SYS_ADMIN
))
2248 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2252 set_numabalancing_state(state
);
2258 #ifdef CONFIG_SCHEDSTATS
2260 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2261 static bool __initdata __sched_schedstats
= false;
2263 static void set_schedstats(bool enabled
)
2266 static_branch_enable(&sched_schedstats
);
2268 static_branch_disable(&sched_schedstats
);
2271 void force_schedstat_enabled(void)
2273 if (!schedstat_enabled()) {
2274 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2275 static_branch_enable(&sched_schedstats
);
2279 static int __init
setup_schedstats(char *str
)
2286 * This code is called before jump labels have been set up, so we can't
2287 * change the static branch directly just yet. Instead set a temporary
2288 * variable so init_schedstats() can do it later.
2290 if (!strcmp(str
, "enable")) {
2291 __sched_schedstats
= true;
2293 } else if (!strcmp(str
, "disable")) {
2294 __sched_schedstats
= false;
2299 pr_warn("Unable to parse schedstats=\n");
2303 __setup("schedstats=", setup_schedstats
);
2305 static void __init
init_schedstats(void)
2307 set_schedstats(__sched_schedstats
);
2310 #ifdef CONFIG_PROC_SYSCTL
2311 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2312 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2316 int state
= static_branch_likely(&sched_schedstats
);
2318 if (write
&& !capable(CAP_SYS_ADMIN
))
2323 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2327 set_schedstats(state
);
2330 #endif /* CONFIG_PROC_SYSCTL */
2331 #else /* !CONFIG_SCHEDSTATS */
2332 static inline void init_schedstats(void) {}
2333 #endif /* CONFIG_SCHEDSTATS */
2336 * fork()/clone()-time setup:
2338 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2340 unsigned long flags
;
2341 int cpu
= get_cpu();
2343 __sched_fork(clone_flags
, p
);
2345 * We mark the process as NEW here. This guarantees that
2346 * nobody will actually run it, and a signal or other external
2347 * event cannot wake it up and insert it on the runqueue either.
2349 p
->state
= TASK_NEW
;
2352 * Make sure we do not leak PI boosting priority to the child.
2354 p
->prio
= current
->normal_prio
;
2357 * Revert to default priority/policy on fork if requested.
2359 if (unlikely(p
->sched_reset_on_fork
)) {
2360 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2361 p
->policy
= SCHED_NORMAL
;
2362 p
->static_prio
= NICE_TO_PRIO(0);
2364 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2365 p
->static_prio
= NICE_TO_PRIO(0);
2367 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2371 * We don't need the reset flag anymore after the fork. It has
2372 * fulfilled its duty:
2374 p
->sched_reset_on_fork
= 0;
2377 if (dl_prio(p
->prio
)) {
2380 } else if (rt_prio(p
->prio
)) {
2381 p
->sched_class
= &rt_sched_class
;
2383 p
->sched_class
= &fair_sched_class
;
2386 init_entity_runnable_average(&p
->se
);
2389 * The child is not yet in the pid-hash so no cgroup attach races,
2390 * and the cgroup is pinned to this child due to cgroup_fork()
2391 * is ran before sched_fork().
2393 * Silence PROVE_RCU.
2395 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2397 * We're setting the cpu for the first time, we don't migrate,
2398 * so use __set_task_cpu().
2400 __set_task_cpu(p
, cpu
);
2401 if (p
->sched_class
->task_fork
)
2402 p
->sched_class
->task_fork(p
);
2403 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2405 #ifdef CONFIG_SCHED_INFO
2406 if (likely(sched_info_on()))
2407 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2409 #if defined(CONFIG_SMP)
2412 init_task_preempt_count(p
);
2414 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2415 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2422 unsigned long to_ratio(u64 period
, u64 runtime
)
2424 if (runtime
== RUNTIME_INF
)
2428 * Doing this here saves a lot of checks in all
2429 * the calling paths, and returning zero seems
2430 * safe for them anyway.
2435 return div64_u64(runtime
<< 20, period
);
2439 inline struct dl_bw
*dl_bw_of(int i
)
2441 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2442 "sched RCU must be held");
2443 return &cpu_rq(i
)->rd
->dl_bw
;
2446 static inline int dl_bw_cpus(int i
)
2448 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2451 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2452 "sched RCU must be held");
2453 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2459 inline struct dl_bw
*dl_bw_of(int i
)
2461 return &cpu_rq(i
)->dl
.dl_bw
;
2464 static inline int dl_bw_cpus(int i
)
2471 * We must be sure that accepting a new task (or allowing changing the
2472 * parameters of an existing one) is consistent with the bandwidth
2473 * constraints. If yes, this function also accordingly updates the currently
2474 * allocated bandwidth to reflect the new situation.
2476 * This function is called while holding p's rq->lock.
2478 * XXX we should delay bw change until the task's 0-lag point, see
2481 static int dl_overflow(struct task_struct
*p
, int policy
,
2482 const struct sched_attr
*attr
)
2485 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2486 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2487 u64 runtime
= attr
->sched_runtime
;
2488 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2491 /* !deadline task may carry old deadline bandwidth */
2492 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2496 * Either if a task, enters, leave, or stays -deadline but changes
2497 * its parameters, we may need to update accordingly the total
2498 * allocated bandwidth of the container.
2500 raw_spin_lock(&dl_b
->lock
);
2501 cpus
= dl_bw_cpus(task_cpu(p
));
2502 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2503 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2504 __dl_add(dl_b
, new_bw
);
2506 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2507 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2508 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2509 __dl_add(dl_b
, new_bw
);
2511 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2512 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2515 raw_spin_unlock(&dl_b
->lock
);
2520 extern void init_dl_bw(struct dl_bw
*dl_b
);
2523 * wake_up_new_task - wake up a newly created task for the first time.
2525 * This function will do some initial scheduler statistics housekeeping
2526 * that must be done for every newly created context, then puts the task
2527 * on the runqueue and wakes it.
2529 void wake_up_new_task(struct task_struct
*p
)
2534 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2535 p
->state
= TASK_RUNNING
;
2538 * Fork balancing, do it here and not earlier because:
2539 * - cpus_allowed can change in the fork path
2540 * - any previously selected cpu might disappear through hotplug
2542 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2543 * as we're not fully set-up yet.
2545 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2547 rq
= __task_rq_lock(p
, &rf
);
2548 post_init_entity_util_avg(&p
->se
);
2550 activate_task(rq
, p
, 0);
2551 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2552 trace_sched_wakeup_new(p
);
2553 check_preempt_curr(rq
, p
, WF_FORK
);
2555 if (p
->sched_class
->task_woken
) {
2557 * Nothing relies on rq->lock after this, so its fine to
2560 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
2561 p
->sched_class
->task_woken(rq
, p
);
2562 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
2565 task_rq_unlock(rq
, p
, &rf
);
2568 #ifdef CONFIG_PREEMPT_NOTIFIERS
2570 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2572 void preempt_notifier_inc(void)
2574 static_key_slow_inc(&preempt_notifier_key
);
2576 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2578 void preempt_notifier_dec(void)
2580 static_key_slow_dec(&preempt_notifier_key
);
2582 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2585 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2586 * @notifier: notifier struct to register
2588 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2590 if (!static_key_false(&preempt_notifier_key
))
2591 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2593 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2595 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2598 * preempt_notifier_unregister - no longer interested in preemption notifications
2599 * @notifier: notifier struct to unregister
2601 * This is *not* safe to call from within a preemption notifier.
2603 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2605 hlist_del(¬ifier
->link
);
2607 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2609 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2611 struct preempt_notifier
*notifier
;
2613 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2614 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2617 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2619 if (static_key_false(&preempt_notifier_key
))
2620 __fire_sched_in_preempt_notifiers(curr
);
2624 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2625 struct task_struct
*next
)
2627 struct preempt_notifier
*notifier
;
2629 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2630 notifier
->ops
->sched_out(notifier
, next
);
2633 static __always_inline
void
2634 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2635 struct task_struct
*next
)
2637 if (static_key_false(&preempt_notifier_key
))
2638 __fire_sched_out_preempt_notifiers(curr
, next
);
2641 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2643 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2648 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2649 struct task_struct
*next
)
2653 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2656 * prepare_task_switch - prepare to switch tasks
2657 * @rq: the runqueue preparing to switch
2658 * @prev: the current task that is being switched out
2659 * @next: the task we are going to switch to.
2661 * This is called with the rq lock held and interrupts off. It must
2662 * be paired with a subsequent finish_task_switch after the context
2665 * prepare_task_switch sets up locking and calls architecture specific
2669 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2670 struct task_struct
*next
)
2672 sched_info_switch(rq
, prev
, next
);
2673 perf_event_task_sched_out(prev
, next
);
2674 fire_sched_out_preempt_notifiers(prev
, next
);
2675 prepare_lock_switch(rq
, next
);
2676 prepare_arch_switch(next
);
2680 * finish_task_switch - clean up after a task-switch
2681 * @prev: the thread we just switched away from.
2683 * finish_task_switch must be called after the context switch, paired
2684 * with a prepare_task_switch call before the context switch.
2685 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2686 * and do any other architecture-specific cleanup actions.
2688 * Note that we may have delayed dropping an mm in context_switch(). If
2689 * so, we finish that here outside of the runqueue lock. (Doing it
2690 * with the lock held can cause deadlocks; see schedule() for
2693 * The context switch have flipped the stack from under us and restored the
2694 * local variables which were saved when this task called schedule() in the
2695 * past. prev == current is still correct but we need to recalculate this_rq
2696 * because prev may have moved to another CPU.
2698 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2699 __releases(rq
->lock
)
2701 struct rq
*rq
= this_rq();
2702 struct mm_struct
*mm
= rq
->prev_mm
;
2706 * The previous task will have left us with a preempt_count of 2
2707 * because it left us after:
2710 * preempt_disable(); // 1
2712 * raw_spin_lock_irq(&rq->lock) // 2
2714 * Also, see FORK_PREEMPT_COUNT.
2716 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2717 "corrupted preempt_count: %s/%d/0x%x\n",
2718 current
->comm
, current
->pid
, preempt_count()))
2719 preempt_count_set(FORK_PREEMPT_COUNT
);
2724 * A task struct has one reference for the use as "current".
2725 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2726 * schedule one last time. The schedule call will never return, and
2727 * the scheduled task must drop that reference.
2729 * We must observe prev->state before clearing prev->on_cpu (in
2730 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2731 * running on another CPU and we could rave with its RUNNING -> DEAD
2732 * transition, resulting in a double drop.
2734 prev_state
= prev
->state
;
2735 vtime_task_switch(prev
);
2736 perf_event_task_sched_in(prev
, current
);
2737 finish_lock_switch(rq
, prev
);
2738 finish_arch_post_lock_switch();
2740 fire_sched_in_preempt_notifiers(current
);
2743 if (unlikely(prev_state
== TASK_DEAD
)) {
2744 if (prev
->sched_class
->task_dead
)
2745 prev
->sched_class
->task_dead(prev
);
2748 * Remove function-return probe instances associated with this
2749 * task and put them back on the free list.
2751 kprobe_flush_task(prev
);
2752 put_task_struct(prev
);
2755 tick_nohz_task_switch();
2761 /* rq->lock is NOT held, but preemption is disabled */
2762 static void __balance_callback(struct rq
*rq
)
2764 struct callback_head
*head
, *next
;
2765 void (*func
)(struct rq
*rq
);
2766 unsigned long flags
;
2768 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2769 head
= rq
->balance_callback
;
2770 rq
->balance_callback
= NULL
;
2772 func
= (void (*)(struct rq
*))head
->func
;
2779 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2782 static inline void balance_callback(struct rq
*rq
)
2784 if (unlikely(rq
->balance_callback
))
2785 __balance_callback(rq
);
2790 static inline void balance_callback(struct rq
*rq
)
2797 * schedule_tail - first thing a freshly forked thread must call.
2798 * @prev: the thread we just switched away from.
2800 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2801 __releases(rq
->lock
)
2806 * New tasks start with FORK_PREEMPT_COUNT, see there and
2807 * finish_task_switch() for details.
2809 * finish_task_switch() will drop rq->lock() and lower preempt_count
2810 * and the preempt_enable() will end up enabling preemption (on
2811 * PREEMPT_COUNT kernels).
2814 rq
= finish_task_switch(prev
);
2815 balance_callback(rq
);
2818 if (current
->set_child_tid
)
2819 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2823 * context_switch - switch to the new MM and the new thread's register state.
2825 static __always_inline
struct rq
*
2826 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2827 struct task_struct
*next
, struct pin_cookie cookie
)
2829 struct mm_struct
*mm
, *oldmm
;
2831 prepare_task_switch(rq
, prev
, next
);
2834 oldmm
= prev
->active_mm
;
2836 * For paravirt, this is coupled with an exit in switch_to to
2837 * combine the page table reload and the switch backend into
2840 arch_start_context_switch(prev
);
2843 next
->active_mm
= oldmm
;
2844 atomic_inc(&oldmm
->mm_count
);
2845 enter_lazy_tlb(oldmm
, next
);
2847 switch_mm_irqs_off(oldmm
, mm
, next
);
2850 prev
->active_mm
= NULL
;
2851 rq
->prev_mm
= oldmm
;
2854 * Since the runqueue lock will be released by the next
2855 * task (which is an invalid locking op but in the case
2856 * of the scheduler it's an obvious special-case), so we
2857 * do an early lockdep release here:
2859 lockdep_unpin_lock(&rq
->lock
, cookie
);
2860 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2862 /* Here we just switch the register state and the stack. */
2863 switch_to(prev
, next
, prev
);
2866 return finish_task_switch(prev
);
2870 * nr_running and nr_context_switches:
2872 * externally visible scheduler statistics: current number of runnable
2873 * threads, total number of context switches performed since bootup.
2875 unsigned long nr_running(void)
2877 unsigned long i
, sum
= 0;
2879 for_each_online_cpu(i
)
2880 sum
+= cpu_rq(i
)->nr_running
;
2886 * Check if only the current task is running on the cpu.
2888 * Caution: this function does not check that the caller has disabled
2889 * preemption, thus the result might have a time-of-check-to-time-of-use
2890 * race. The caller is responsible to use it correctly, for example:
2892 * - from a non-preemptable section (of course)
2894 * - from a thread that is bound to a single CPU
2896 * - in a loop with very short iterations (e.g. a polling loop)
2898 bool single_task_running(void)
2900 return raw_rq()->nr_running
== 1;
2902 EXPORT_SYMBOL(single_task_running
);
2904 unsigned long long nr_context_switches(void)
2907 unsigned long long sum
= 0;
2909 for_each_possible_cpu(i
)
2910 sum
+= cpu_rq(i
)->nr_switches
;
2915 unsigned long nr_iowait(void)
2917 unsigned long i
, sum
= 0;
2919 for_each_possible_cpu(i
)
2920 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2925 unsigned long nr_iowait_cpu(int cpu
)
2927 struct rq
*this = cpu_rq(cpu
);
2928 return atomic_read(&this->nr_iowait
);
2931 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2933 struct rq
*rq
= this_rq();
2934 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2935 *load
= rq
->load
.weight
;
2941 * sched_exec - execve() is a valuable balancing opportunity, because at
2942 * this point the task has the smallest effective memory and cache footprint.
2944 void sched_exec(void)
2946 struct task_struct
*p
= current
;
2947 unsigned long flags
;
2950 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2951 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2952 if (dest_cpu
== smp_processor_id())
2955 if (likely(cpu_active(dest_cpu
))) {
2956 struct migration_arg arg
= { p
, dest_cpu
};
2958 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2959 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2963 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2968 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2969 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2971 EXPORT_PER_CPU_SYMBOL(kstat
);
2972 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2975 * Return accounted runtime for the task.
2976 * In case the task is currently running, return the runtime plus current's
2977 * pending runtime that have not been accounted yet.
2979 unsigned long long task_sched_runtime(struct task_struct
*p
)
2985 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2987 * 64-bit doesn't need locks to atomically read a 64bit value.
2988 * So we have a optimization chance when the task's delta_exec is 0.
2989 * Reading ->on_cpu is racy, but this is ok.
2991 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2992 * If we race with it entering cpu, unaccounted time is 0. This is
2993 * indistinguishable from the read occurring a few cycles earlier.
2994 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2995 * been accounted, so we're correct here as well.
2997 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2998 return p
->se
.sum_exec_runtime
;
3001 rq
= task_rq_lock(p
, &rf
);
3003 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3004 * project cycles that may never be accounted to this
3005 * thread, breaking clock_gettime().
3007 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3008 update_rq_clock(rq
);
3009 p
->sched_class
->update_curr(rq
);
3011 ns
= p
->se
.sum_exec_runtime
;
3012 task_rq_unlock(rq
, p
, &rf
);
3018 * This function gets called by the timer code, with HZ frequency.
3019 * We call it with interrupts disabled.
3021 void scheduler_tick(void)
3023 int cpu
= smp_processor_id();
3024 struct rq
*rq
= cpu_rq(cpu
);
3025 struct task_struct
*curr
= rq
->curr
;
3029 raw_spin_lock(&rq
->lock
);
3030 update_rq_clock(rq
);
3031 curr
->sched_class
->task_tick(rq
, curr
, 0);
3032 cpu_load_update_active(rq
);
3033 calc_global_load_tick(rq
);
3034 raw_spin_unlock(&rq
->lock
);
3036 perf_event_task_tick();
3039 rq
->idle_balance
= idle_cpu(cpu
);
3040 trigger_load_balance(rq
);
3042 rq_last_tick_reset(rq
);
3045 #ifdef CONFIG_NO_HZ_FULL
3047 * scheduler_tick_max_deferment
3049 * Keep at least one tick per second when a single
3050 * active task is running because the scheduler doesn't
3051 * yet completely support full dynticks environment.
3053 * This makes sure that uptime, CFS vruntime, load
3054 * balancing, etc... continue to move forward, even
3055 * with a very low granularity.
3057 * Return: Maximum deferment in nanoseconds.
3059 u64
scheduler_tick_max_deferment(void)
3061 struct rq
*rq
= this_rq();
3062 unsigned long next
, now
= READ_ONCE(jiffies
);
3064 next
= rq
->last_sched_tick
+ HZ
;
3066 if (time_before_eq(next
, now
))
3069 return jiffies_to_nsecs(next
- now
);
3073 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3074 defined(CONFIG_PREEMPT_TRACER))
3076 * If the value passed in is equal to the current preempt count
3077 * then we just disabled preemption. Start timing the latency.
3079 static inline void preempt_latency_start(int val
)
3081 if (preempt_count() == val
) {
3082 unsigned long ip
= get_lock_parent_ip();
3083 #ifdef CONFIG_DEBUG_PREEMPT
3084 current
->preempt_disable_ip
= ip
;
3086 trace_preempt_off(CALLER_ADDR0
, ip
);
3090 void preempt_count_add(int val
)
3092 #ifdef CONFIG_DEBUG_PREEMPT
3096 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3099 __preempt_count_add(val
);
3100 #ifdef CONFIG_DEBUG_PREEMPT
3102 * Spinlock count overflowing soon?
3104 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3107 preempt_latency_start(val
);
3109 EXPORT_SYMBOL(preempt_count_add
);
3110 NOKPROBE_SYMBOL(preempt_count_add
);
3113 * If the value passed in equals to the current preempt count
3114 * then we just enabled preemption. Stop timing the latency.
3116 static inline void preempt_latency_stop(int val
)
3118 if (preempt_count() == val
)
3119 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3122 void preempt_count_sub(int val
)
3124 #ifdef CONFIG_DEBUG_PREEMPT
3128 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3131 * Is the spinlock portion underflowing?
3133 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3134 !(preempt_count() & PREEMPT_MASK
)))
3138 preempt_latency_stop(val
);
3139 __preempt_count_sub(val
);
3141 EXPORT_SYMBOL(preempt_count_sub
);
3142 NOKPROBE_SYMBOL(preempt_count_sub
);
3145 static inline void preempt_latency_start(int val
) { }
3146 static inline void preempt_latency_stop(int val
) { }
3150 * Print scheduling while atomic bug:
3152 static noinline
void __schedule_bug(struct task_struct
*prev
)
3154 if (oops_in_progress
)
3157 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3158 prev
->comm
, prev
->pid
, preempt_count());
3160 debug_show_held_locks(prev
);
3162 if (irqs_disabled())
3163 print_irqtrace_events(prev
);
3164 #ifdef CONFIG_DEBUG_PREEMPT
3165 if (in_atomic_preempt_off()) {
3166 pr_err("Preemption disabled at:");
3167 print_ip_sym(current
->preempt_disable_ip
);
3172 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3176 * Various schedule()-time debugging checks and statistics:
3178 static inline void schedule_debug(struct task_struct
*prev
)
3180 #ifdef CONFIG_SCHED_STACK_END_CHECK
3181 if (task_stack_end_corrupted(prev
))
3182 panic("corrupted stack end detected inside scheduler\n");
3185 if (unlikely(in_atomic_preempt_off())) {
3186 __schedule_bug(prev
);
3187 preempt_count_set(PREEMPT_DISABLED
);
3191 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3193 schedstat_inc(this_rq(), sched_count
);
3197 * Pick up the highest-prio task:
3199 static inline struct task_struct
*
3200 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
3202 const struct sched_class
*class = &fair_sched_class
;
3203 struct task_struct
*p
;
3206 * Optimization: we know that if all tasks are in
3207 * the fair class we can call that function directly:
3209 if (likely(prev
->sched_class
== class &&
3210 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3211 p
= fair_sched_class
.pick_next_task(rq
, prev
, cookie
);
3212 if (unlikely(p
== RETRY_TASK
))
3215 /* assumes fair_sched_class->next == idle_sched_class */
3217 p
= idle_sched_class
.pick_next_task(rq
, prev
, cookie
);
3223 for_each_class(class) {
3224 p
= class->pick_next_task(rq
, prev
, cookie
);
3226 if (unlikely(p
== RETRY_TASK
))
3232 BUG(); /* the idle class will always have a runnable task */
3236 * __schedule() is the main scheduler function.
3238 * The main means of driving the scheduler and thus entering this function are:
3240 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3242 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3243 * paths. For example, see arch/x86/entry_64.S.
3245 * To drive preemption between tasks, the scheduler sets the flag in timer
3246 * interrupt handler scheduler_tick().
3248 * 3. Wakeups don't really cause entry into schedule(). They add a
3249 * task to the run-queue and that's it.
3251 * Now, if the new task added to the run-queue preempts the current
3252 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3253 * called on the nearest possible occasion:
3255 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3257 * - in syscall or exception context, at the next outmost
3258 * preempt_enable(). (this might be as soon as the wake_up()'s
3261 * - in IRQ context, return from interrupt-handler to
3262 * preemptible context
3264 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3267 * - cond_resched() call
3268 * - explicit schedule() call
3269 * - return from syscall or exception to user-space
3270 * - return from interrupt-handler to user-space
3272 * WARNING: must be called with preemption disabled!
3274 static void __sched notrace
__schedule(bool preempt
)
3276 struct task_struct
*prev
, *next
;
3277 unsigned long *switch_count
;
3278 struct pin_cookie cookie
;
3282 cpu
= smp_processor_id();
3287 * do_exit() calls schedule() with preemption disabled as an exception;
3288 * however we must fix that up, otherwise the next task will see an
3289 * inconsistent (higher) preempt count.
3291 * It also avoids the below schedule_debug() test from complaining
3294 if (unlikely(prev
->state
== TASK_DEAD
))
3295 preempt_enable_no_resched_notrace();
3297 schedule_debug(prev
);
3299 if (sched_feat(HRTICK
))
3302 local_irq_disable();
3303 rcu_note_context_switch();
3306 * Make sure that signal_pending_state()->signal_pending() below
3307 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3308 * done by the caller to avoid the race with signal_wake_up().
3310 smp_mb__before_spinlock();
3311 raw_spin_lock(&rq
->lock
);
3312 cookie
= lockdep_pin_lock(&rq
->lock
);
3314 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3316 switch_count
= &prev
->nivcsw
;
3317 if (!preempt
&& prev
->state
) {
3318 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3319 prev
->state
= TASK_RUNNING
;
3321 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3325 * If a worker went to sleep, notify and ask workqueue
3326 * whether it wants to wake up a task to maintain
3329 if (prev
->flags
& PF_WQ_WORKER
) {
3330 struct task_struct
*to_wakeup
;
3332 to_wakeup
= wq_worker_sleeping(prev
);
3334 try_to_wake_up_local(to_wakeup
, cookie
);
3337 switch_count
= &prev
->nvcsw
;
3340 if (task_on_rq_queued(prev
))
3341 update_rq_clock(rq
);
3343 next
= pick_next_task(rq
, prev
, cookie
);
3344 clear_tsk_need_resched(prev
);
3345 clear_preempt_need_resched();
3346 rq
->clock_skip_update
= 0;
3348 if (likely(prev
!= next
)) {
3353 trace_sched_switch(preempt
, prev
, next
);
3354 rq
= context_switch(rq
, prev
, next
, cookie
); /* unlocks the rq */
3356 lockdep_unpin_lock(&rq
->lock
, cookie
);
3357 raw_spin_unlock_irq(&rq
->lock
);
3360 balance_callback(rq
);
3362 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3364 static inline void sched_submit_work(struct task_struct
*tsk
)
3366 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3369 * If we are going to sleep and we have plugged IO queued,
3370 * make sure to submit it to avoid deadlocks.
3372 if (blk_needs_flush_plug(tsk
))
3373 blk_schedule_flush_plug(tsk
);
3376 asmlinkage __visible
void __sched
schedule(void)
3378 struct task_struct
*tsk
= current
;
3380 sched_submit_work(tsk
);
3384 sched_preempt_enable_no_resched();
3385 } while (need_resched());
3387 EXPORT_SYMBOL(schedule
);
3389 #ifdef CONFIG_CONTEXT_TRACKING
3390 asmlinkage __visible
void __sched
schedule_user(void)
3393 * If we come here after a random call to set_need_resched(),
3394 * or we have been woken up remotely but the IPI has not yet arrived,
3395 * we haven't yet exited the RCU idle mode. Do it here manually until
3396 * we find a better solution.
3398 * NB: There are buggy callers of this function. Ideally we
3399 * should warn if prev_state != CONTEXT_USER, but that will trigger
3400 * too frequently to make sense yet.
3402 enum ctx_state prev_state
= exception_enter();
3404 exception_exit(prev_state
);
3409 * schedule_preempt_disabled - called with preemption disabled
3411 * Returns with preemption disabled. Note: preempt_count must be 1
3413 void __sched
schedule_preempt_disabled(void)
3415 sched_preempt_enable_no_resched();
3420 static void __sched notrace
preempt_schedule_common(void)
3424 * Because the function tracer can trace preempt_count_sub()
3425 * and it also uses preempt_enable/disable_notrace(), if
3426 * NEED_RESCHED is set, the preempt_enable_notrace() called
3427 * by the function tracer will call this function again and
3428 * cause infinite recursion.
3430 * Preemption must be disabled here before the function
3431 * tracer can trace. Break up preempt_disable() into two
3432 * calls. One to disable preemption without fear of being
3433 * traced. The other to still record the preemption latency,
3434 * which can also be traced by the function tracer.
3436 preempt_disable_notrace();
3437 preempt_latency_start(1);
3439 preempt_latency_stop(1);
3440 preempt_enable_no_resched_notrace();
3443 * Check again in case we missed a preemption opportunity
3444 * between schedule and now.
3446 } while (need_resched());
3449 #ifdef CONFIG_PREEMPT
3451 * this is the entry point to schedule() from in-kernel preemption
3452 * off of preempt_enable. Kernel preemptions off return from interrupt
3453 * occur there and call schedule directly.
3455 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3458 * If there is a non-zero preempt_count or interrupts are disabled,
3459 * we do not want to preempt the current task. Just return..
3461 if (likely(!preemptible()))
3464 preempt_schedule_common();
3466 NOKPROBE_SYMBOL(preempt_schedule
);
3467 EXPORT_SYMBOL(preempt_schedule
);
3470 * preempt_schedule_notrace - preempt_schedule called by tracing
3472 * The tracing infrastructure uses preempt_enable_notrace to prevent
3473 * recursion and tracing preempt enabling caused by the tracing
3474 * infrastructure itself. But as tracing can happen in areas coming
3475 * from userspace or just about to enter userspace, a preempt enable
3476 * can occur before user_exit() is called. This will cause the scheduler
3477 * to be called when the system is still in usermode.
3479 * To prevent this, the preempt_enable_notrace will use this function
3480 * instead of preempt_schedule() to exit user context if needed before
3481 * calling the scheduler.
3483 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3485 enum ctx_state prev_ctx
;
3487 if (likely(!preemptible()))
3492 * Because the function tracer can trace preempt_count_sub()
3493 * and it also uses preempt_enable/disable_notrace(), if
3494 * NEED_RESCHED is set, the preempt_enable_notrace() called
3495 * by the function tracer will call this function again and
3496 * cause infinite recursion.
3498 * Preemption must be disabled here before the function
3499 * tracer can trace. Break up preempt_disable() into two
3500 * calls. One to disable preemption without fear of being
3501 * traced. The other to still record the preemption latency,
3502 * which can also be traced by the function tracer.
3504 preempt_disable_notrace();
3505 preempt_latency_start(1);
3507 * Needs preempt disabled in case user_exit() is traced
3508 * and the tracer calls preempt_enable_notrace() causing
3509 * an infinite recursion.
3511 prev_ctx
= exception_enter();
3513 exception_exit(prev_ctx
);
3515 preempt_latency_stop(1);
3516 preempt_enable_no_resched_notrace();
3517 } while (need_resched());
3519 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3521 #endif /* CONFIG_PREEMPT */
3524 * this is the entry point to schedule() from kernel preemption
3525 * off of irq context.
3526 * Note, that this is called and return with irqs disabled. This will
3527 * protect us against recursive calling from irq.
3529 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3531 enum ctx_state prev_state
;
3533 /* Catch callers which need to be fixed */
3534 BUG_ON(preempt_count() || !irqs_disabled());
3536 prev_state
= exception_enter();
3542 local_irq_disable();
3543 sched_preempt_enable_no_resched();
3544 } while (need_resched());
3546 exception_exit(prev_state
);
3549 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3552 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3554 EXPORT_SYMBOL(default_wake_function
);
3556 #ifdef CONFIG_RT_MUTEXES
3559 * rt_mutex_setprio - set the current priority of a task
3561 * @prio: prio value (kernel-internal form)
3563 * This function changes the 'effective' priority of a task. It does
3564 * not touch ->normal_prio like __setscheduler().
3566 * Used by the rt_mutex code to implement priority inheritance
3567 * logic. Call site only calls if the priority of the task changed.
3569 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3571 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3572 const struct sched_class
*prev_class
;
3576 BUG_ON(prio
> MAX_PRIO
);
3578 rq
= __task_rq_lock(p
, &rf
);
3581 * Idle task boosting is a nono in general. There is one
3582 * exception, when PREEMPT_RT and NOHZ is active:
3584 * The idle task calls get_next_timer_interrupt() and holds
3585 * the timer wheel base->lock on the CPU and another CPU wants
3586 * to access the timer (probably to cancel it). We can safely
3587 * ignore the boosting request, as the idle CPU runs this code
3588 * with interrupts disabled and will complete the lock
3589 * protected section without being interrupted. So there is no
3590 * real need to boost.
3592 if (unlikely(p
== rq
->idle
)) {
3593 WARN_ON(p
!= rq
->curr
);
3594 WARN_ON(p
->pi_blocked_on
);
3598 trace_sched_pi_setprio(p
, prio
);
3601 if (oldprio
== prio
)
3602 queue_flag
&= ~DEQUEUE_MOVE
;
3604 prev_class
= p
->sched_class
;
3605 queued
= task_on_rq_queued(p
);
3606 running
= task_current(rq
, p
);
3608 dequeue_task(rq
, p
, queue_flag
);
3610 put_prev_task(rq
, p
);
3613 * Boosting condition are:
3614 * 1. -rt task is running and holds mutex A
3615 * --> -dl task blocks on mutex A
3617 * 2. -dl task is running and holds mutex A
3618 * --> -dl task blocks on mutex A and could preempt the
3621 if (dl_prio(prio
)) {
3622 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3623 if (!dl_prio(p
->normal_prio
) ||
3624 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3625 p
->dl
.dl_boosted
= 1;
3626 queue_flag
|= ENQUEUE_REPLENISH
;
3628 p
->dl
.dl_boosted
= 0;
3629 p
->sched_class
= &dl_sched_class
;
3630 } else if (rt_prio(prio
)) {
3631 if (dl_prio(oldprio
))
3632 p
->dl
.dl_boosted
= 0;
3634 queue_flag
|= ENQUEUE_HEAD
;
3635 p
->sched_class
= &rt_sched_class
;
3637 if (dl_prio(oldprio
))
3638 p
->dl
.dl_boosted
= 0;
3639 if (rt_prio(oldprio
))
3641 p
->sched_class
= &fair_sched_class
;
3647 p
->sched_class
->set_curr_task(rq
);
3649 enqueue_task(rq
, p
, queue_flag
);
3651 check_class_changed(rq
, p
, prev_class
, oldprio
);
3653 preempt_disable(); /* avoid rq from going away on us */
3654 __task_rq_unlock(rq
, &rf
);
3656 balance_callback(rq
);
3661 void set_user_nice(struct task_struct
*p
, long nice
)
3663 int old_prio
, delta
, queued
;
3667 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3670 * We have to be careful, if called from sys_setpriority(),
3671 * the task might be in the middle of scheduling on another CPU.
3673 rq
= task_rq_lock(p
, &rf
);
3675 * The RT priorities are set via sched_setscheduler(), but we still
3676 * allow the 'normal' nice value to be set - but as expected
3677 * it wont have any effect on scheduling until the task is
3678 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3680 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3681 p
->static_prio
= NICE_TO_PRIO(nice
);
3684 queued
= task_on_rq_queued(p
);
3686 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3688 p
->static_prio
= NICE_TO_PRIO(nice
);
3691 p
->prio
= effective_prio(p
);
3692 delta
= p
->prio
- old_prio
;
3695 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3697 * If the task increased its priority or is running and
3698 * lowered its priority, then reschedule its CPU:
3700 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3704 task_rq_unlock(rq
, p
, &rf
);
3706 EXPORT_SYMBOL(set_user_nice
);
3709 * can_nice - check if a task can reduce its nice value
3713 int can_nice(const struct task_struct
*p
, const int nice
)
3715 /* convert nice value [19,-20] to rlimit style value [1,40] */
3716 int nice_rlim
= nice_to_rlimit(nice
);
3718 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3719 capable(CAP_SYS_NICE
));
3722 #ifdef __ARCH_WANT_SYS_NICE
3725 * sys_nice - change the priority of the current process.
3726 * @increment: priority increment
3728 * sys_setpriority is a more generic, but much slower function that
3729 * does similar things.
3731 SYSCALL_DEFINE1(nice
, int, increment
)
3736 * Setpriority might change our priority at the same moment.
3737 * We don't have to worry. Conceptually one call occurs first
3738 * and we have a single winner.
3740 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3741 nice
= task_nice(current
) + increment
;
3743 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3744 if (increment
< 0 && !can_nice(current
, nice
))
3747 retval
= security_task_setnice(current
, nice
);
3751 set_user_nice(current
, nice
);
3758 * task_prio - return the priority value of a given task.
3759 * @p: the task in question.
3761 * Return: The priority value as seen by users in /proc.
3762 * RT tasks are offset by -200. Normal tasks are centered
3763 * around 0, value goes from -16 to +15.
3765 int task_prio(const struct task_struct
*p
)
3767 return p
->prio
- MAX_RT_PRIO
;
3771 * idle_cpu - is a given cpu idle currently?
3772 * @cpu: the processor in question.
3774 * Return: 1 if the CPU is currently idle. 0 otherwise.
3776 int idle_cpu(int cpu
)
3778 struct rq
*rq
= cpu_rq(cpu
);
3780 if (rq
->curr
!= rq
->idle
)
3787 if (!llist_empty(&rq
->wake_list
))
3795 * idle_task - return the idle task for a given cpu.
3796 * @cpu: the processor in question.
3798 * Return: The idle task for the cpu @cpu.
3800 struct task_struct
*idle_task(int cpu
)
3802 return cpu_rq(cpu
)->idle
;
3806 * find_process_by_pid - find a process with a matching PID value.
3807 * @pid: the pid in question.
3809 * The task of @pid, if found. %NULL otherwise.
3811 static struct task_struct
*find_process_by_pid(pid_t pid
)
3813 return pid
? find_task_by_vpid(pid
) : current
;
3817 * This function initializes the sched_dl_entity of a newly becoming
3818 * SCHED_DEADLINE task.
3820 * Only the static values are considered here, the actual runtime and the
3821 * absolute deadline will be properly calculated when the task is enqueued
3822 * for the first time with its new policy.
3825 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3827 struct sched_dl_entity
*dl_se
= &p
->dl
;
3829 dl_se
->dl_runtime
= attr
->sched_runtime
;
3830 dl_se
->dl_deadline
= attr
->sched_deadline
;
3831 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3832 dl_se
->flags
= attr
->sched_flags
;
3833 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3836 * Changing the parameters of a task is 'tricky' and we're not doing
3837 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3839 * What we SHOULD do is delay the bandwidth release until the 0-lag
3840 * point. This would include retaining the task_struct until that time
3841 * and change dl_overflow() to not immediately decrement the current
3844 * Instead we retain the current runtime/deadline and let the new
3845 * parameters take effect after the current reservation period lapses.
3846 * This is safe (albeit pessimistic) because the 0-lag point is always
3847 * before the current scheduling deadline.
3849 * We can still have temporary overloads because we do not delay the
3850 * change in bandwidth until that time; so admission control is
3851 * not on the safe side. It does however guarantee tasks will never
3852 * consume more than promised.
3857 * sched_setparam() passes in -1 for its policy, to let the functions
3858 * it calls know not to change it.
3860 #define SETPARAM_POLICY -1
3862 static void __setscheduler_params(struct task_struct
*p
,
3863 const struct sched_attr
*attr
)
3865 int policy
= attr
->sched_policy
;
3867 if (policy
== SETPARAM_POLICY
)
3872 if (dl_policy(policy
))
3873 __setparam_dl(p
, attr
);
3874 else if (fair_policy(policy
))
3875 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3878 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3879 * !rt_policy. Always setting this ensures that things like
3880 * getparam()/getattr() don't report silly values for !rt tasks.
3882 p
->rt_priority
= attr
->sched_priority
;
3883 p
->normal_prio
= normal_prio(p
);
3887 /* Actually do priority change: must hold pi & rq lock. */
3888 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3889 const struct sched_attr
*attr
, bool keep_boost
)
3891 __setscheduler_params(p
, attr
);
3894 * Keep a potential priority boosting if called from
3895 * sched_setscheduler().
3898 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3900 p
->prio
= normal_prio(p
);
3902 if (dl_prio(p
->prio
))
3903 p
->sched_class
= &dl_sched_class
;
3904 else if (rt_prio(p
->prio
))
3905 p
->sched_class
= &rt_sched_class
;
3907 p
->sched_class
= &fair_sched_class
;
3911 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3913 struct sched_dl_entity
*dl_se
= &p
->dl
;
3915 attr
->sched_priority
= p
->rt_priority
;
3916 attr
->sched_runtime
= dl_se
->dl_runtime
;
3917 attr
->sched_deadline
= dl_se
->dl_deadline
;
3918 attr
->sched_period
= dl_se
->dl_period
;
3919 attr
->sched_flags
= dl_se
->flags
;
3923 * This function validates the new parameters of a -deadline task.
3924 * We ask for the deadline not being zero, and greater or equal
3925 * than the runtime, as well as the period of being zero or
3926 * greater than deadline. Furthermore, we have to be sure that
3927 * user parameters are above the internal resolution of 1us (we
3928 * check sched_runtime only since it is always the smaller one) and
3929 * below 2^63 ns (we have to check both sched_deadline and
3930 * sched_period, as the latter can be zero).
3933 __checkparam_dl(const struct sched_attr
*attr
)
3936 if (attr
->sched_deadline
== 0)
3940 * Since we truncate DL_SCALE bits, make sure we're at least
3943 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3947 * Since we use the MSB for wrap-around and sign issues, make
3948 * sure it's not set (mind that period can be equal to zero).
3950 if (attr
->sched_deadline
& (1ULL << 63) ||
3951 attr
->sched_period
& (1ULL << 63))
3954 /* runtime <= deadline <= period (if period != 0) */
3955 if ((attr
->sched_period
!= 0 &&
3956 attr
->sched_period
< attr
->sched_deadline
) ||
3957 attr
->sched_deadline
< attr
->sched_runtime
)
3964 * check the target process has a UID that matches the current process's
3966 static bool check_same_owner(struct task_struct
*p
)
3968 const struct cred
*cred
= current_cred(), *pcred
;
3972 pcred
= __task_cred(p
);
3973 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3974 uid_eq(cred
->euid
, pcred
->uid
));
3979 static bool dl_param_changed(struct task_struct
*p
,
3980 const struct sched_attr
*attr
)
3982 struct sched_dl_entity
*dl_se
= &p
->dl
;
3984 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3985 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3986 dl_se
->dl_period
!= attr
->sched_period
||
3987 dl_se
->flags
!= attr
->sched_flags
)
3993 static int __sched_setscheduler(struct task_struct
*p
,
3994 const struct sched_attr
*attr
,
3997 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3998 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3999 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4000 int new_effective_prio
, policy
= attr
->sched_policy
;
4001 const struct sched_class
*prev_class
;
4004 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
4007 /* may grab non-irq protected spin_locks */
4008 BUG_ON(in_interrupt());
4010 /* double check policy once rq lock held */
4012 reset_on_fork
= p
->sched_reset_on_fork
;
4013 policy
= oldpolicy
= p
->policy
;
4015 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4017 if (!valid_policy(policy
))
4021 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4025 * Valid priorities for SCHED_FIFO and SCHED_RR are
4026 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4027 * SCHED_BATCH and SCHED_IDLE is 0.
4029 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4030 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4032 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4033 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4037 * Allow unprivileged RT tasks to decrease priority:
4039 if (user
&& !capable(CAP_SYS_NICE
)) {
4040 if (fair_policy(policy
)) {
4041 if (attr
->sched_nice
< task_nice(p
) &&
4042 !can_nice(p
, attr
->sched_nice
))
4046 if (rt_policy(policy
)) {
4047 unsigned long rlim_rtprio
=
4048 task_rlimit(p
, RLIMIT_RTPRIO
);
4050 /* can't set/change the rt policy */
4051 if (policy
!= p
->policy
&& !rlim_rtprio
)
4054 /* can't increase priority */
4055 if (attr
->sched_priority
> p
->rt_priority
&&
4056 attr
->sched_priority
> rlim_rtprio
)
4061 * Can't set/change SCHED_DEADLINE policy at all for now
4062 * (safest behavior); in the future we would like to allow
4063 * unprivileged DL tasks to increase their relative deadline
4064 * or reduce their runtime (both ways reducing utilization)
4066 if (dl_policy(policy
))
4070 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4071 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4073 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4074 if (!can_nice(p
, task_nice(p
)))
4078 /* can't change other user's priorities */
4079 if (!check_same_owner(p
))
4082 /* Normal users shall not reset the sched_reset_on_fork flag */
4083 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4088 retval
= security_task_setscheduler(p
);
4094 * make sure no PI-waiters arrive (or leave) while we are
4095 * changing the priority of the task:
4097 * To be able to change p->policy safely, the appropriate
4098 * runqueue lock must be held.
4100 rq
= task_rq_lock(p
, &rf
);
4103 * Changing the policy of the stop threads its a very bad idea
4105 if (p
== rq
->stop
) {
4106 task_rq_unlock(rq
, p
, &rf
);
4111 * If not changing anything there's no need to proceed further,
4112 * but store a possible modification of reset_on_fork.
4114 if (unlikely(policy
== p
->policy
)) {
4115 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4117 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4119 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4122 p
->sched_reset_on_fork
= reset_on_fork
;
4123 task_rq_unlock(rq
, p
, &rf
);
4129 #ifdef CONFIG_RT_GROUP_SCHED
4131 * Do not allow realtime tasks into groups that have no runtime
4134 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4135 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4136 !task_group_is_autogroup(task_group(p
))) {
4137 task_rq_unlock(rq
, p
, &rf
);
4142 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4143 cpumask_t
*span
= rq
->rd
->span
;
4146 * Don't allow tasks with an affinity mask smaller than
4147 * the entire root_domain to become SCHED_DEADLINE. We
4148 * will also fail if there's no bandwidth available.
4150 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4151 rq
->rd
->dl_bw
.bw
== 0) {
4152 task_rq_unlock(rq
, p
, &rf
);
4159 /* recheck policy now with rq lock held */
4160 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4161 policy
= oldpolicy
= -1;
4162 task_rq_unlock(rq
, p
, &rf
);
4167 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4168 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4171 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4172 task_rq_unlock(rq
, p
, &rf
);
4176 p
->sched_reset_on_fork
= reset_on_fork
;
4181 * Take priority boosted tasks into account. If the new
4182 * effective priority is unchanged, we just store the new
4183 * normal parameters and do not touch the scheduler class and
4184 * the runqueue. This will be done when the task deboost
4187 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4188 if (new_effective_prio
== oldprio
)
4189 queue_flags
&= ~DEQUEUE_MOVE
;
4192 queued
= task_on_rq_queued(p
);
4193 running
= task_current(rq
, p
);
4195 dequeue_task(rq
, p
, queue_flags
);
4197 put_prev_task(rq
, p
);
4199 prev_class
= p
->sched_class
;
4200 __setscheduler(rq
, p
, attr
, pi
);
4203 p
->sched_class
->set_curr_task(rq
);
4206 * We enqueue to tail when the priority of a task is
4207 * increased (user space view).
4209 if (oldprio
< p
->prio
)
4210 queue_flags
|= ENQUEUE_HEAD
;
4212 enqueue_task(rq
, p
, queue_flags
);
4215 check_class_changed(rq
, p
, prev_class
, oldprio
);
4216 preempt_disable(); /* avoid rq from going away on us */
4217 task_rq_unlock(rq
, p
, &rf
);
4220 rt_mutex_adjust_pi(p
);
4223 * Run balance callbacks after we've adjusted the PI chain.
4225 balance_callback(rq
);
4231 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4232 const struct sched_param
*param
, bool check
)
4234 struct sched_attr attr
= {
4235 .sched_policy
= policy
,
4236 .sched_priority
= param
->sched_priority
,
4237 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4240 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4241 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4242 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4243 policy
&= ~SCHED_RESET_ON_FORK
;
4244 attr
.sched_policy
= policy
;
4247 return __sched_setscheduler(p
, &attr
, check
, true);
4250 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4251 * @p: the task in question.
4252 * @policy: new policy.
4253 * @param: structure containing the new RT priority.
4255 * Return: 0 on success. An error code otherwise.
4257 * NOTE that the task may be already dead.
4259 int sched_setscheduler(struct task_struct
*p
, int policy
,
4260 const struct sched_param
*param
)
4262 return _sched_setscheduler(p
, policy
, param
, true);
4264 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4266 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4268 return __sched_setscheduler(p
, attr
, true, true);
4270 EXPORT_SYMBOL_GPL(sched_setattr
);
4273 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4274 * @p: the task in question.
4275 * @policy: new policy.
4276 * @param: structure containing the new RT priority.
4278 * Just like sched_setscheduler, only don't bother checking if the
4279 * current context has permission. For example, this is needed in
4280 * stop_machine(): we create temporary high priority worker threads,
4281 * but our caller might not have that capability.
4283 * Return: 0 on success. An error code otherwise.
4285 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4286 const struct sched_param
*param
)
4288 return _sched_setscheduler(p
, policy
, param
, false);
4290 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4293 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4295 struct sched_param lparam
;
4296 struct task_struct
*p
;
4299 if (!param
|| pid
< 0)
4301 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4306 p
= find_process_by_pid(pid
);
4308 retval
= sched_setscheduler(p
, policy
, &lparam
);
4315 * Mimics kernel/events/core.c perf_copy_attr().
4317 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4318 struct sched_attr
*attr
)
4323 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4327 * zero the full structure, so that a short copy will be nice.
4329 memset(attr
, 0, sizeof(*attr
));
4331 ret
= get_user(size
, &uattr
->size
);
4335 if (size
> PAGE_SIZE
) /* silly large */
4338 if (!size
) /* abi compat */
4339 size
= SCHED_ATTR_SIZE_VER0
;
4341 if (size
< SCHED_ATTR_SIZE_VER0
)
4345 * If we're handed a bigger struct than we know of,
4346 * ensure all the unknown bits are 0 - i.e. new
4347 * user-space does not rely on any kernel feature
4348 * extensions we dont know about yet.
4350 if (size
> sizeof(*attr
)) {
4351 unsigned char __user
*addr
;
4352 unsigned char __user
*end
;
4355 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4356 end
= (void __user
*)uattr
+ size
;
4358 for (; addr
< end
; addr
++) {
4359 ret
= get_user(val
, addr
);
4365 size
= sizeof(*attr
);
4368 ret
= copy_from_user(attr
, uattr
, size
);
4373 * XXX: do we want to be lenient like existing syscalls; or do we want
4374 * to be strict and return an error on out-of-bounds values?
4376 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4381 put_user(sizeof(*attr
), &uattr
->size
);
4386 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4387 * @pid: the pid in question.
4388 * @policy: new policy.
4389 * @param: structure containing the new RT priority.
4391 * Return: 0 on success. An error code otherwise.
4393 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4394 struct sched_param __user
*, param
)
4396 /* negative values for policy are not valid */
4400 return do_sched_setscheduler(pid
, policy
, param
);
4404 * sys_sched_setparam - set/change the RT priority of a thread
4405 * @pid: the pid in question.
4406 * @param: structure containing the new RT priority.
4408 * Return: 0 on success. An error code otherwise.
4410 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4412 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4416 * sys_sched_setattr - same as above, but with extended sched_attr
4417 * @pid: the pid in question.
4418 * @uattr: structure containing the extended parameters.
4419 * @flags: for future extension.
4421 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4422 unsigned int, flags
)
4424 struct sched_attr attr
;
4425 struct task_struct
*p
;
4428 if (!uattr
|| pid
< 0 || flags
)
4431 retval
= sched_copy_attr(uattr
, &attr
);
4435 if ((int)attr
.sched_policy
< 0)
4440 p
= find_process_by_pid(pid
);
4442 retval
= sched_setattr(p
, &attr
);
4449 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4450 * @pid: the pid in question.
4452 * Return: On success, the policy of the thread. Otherwise, a negative error
4455 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4457 struct task_struct
*p
;
4465 p
= find_process_by_pid(pid
);
4467 retval
= security_task_getscheduler(p
);
4470 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4477 * sys_sched_getparam - get the RT priority of a thread
4478 * @pid: the pid in question.
4479 * @param: structure containing the RT priority.
4481 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4484 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4486 struct sched_param lp
= { .sched_priority
= 0 };
4487 struct task_struct
*p
;
4490 if (!param
|| pid
< 0)
4494 p
= find_process_by_pid(pid
);
4499 retval
= security_task_getscheduler(p
);
4503 if (task_has_rt_policy(p
))
4504 lp
.sched_priority
= p
->rt_priority
;
4508 * This one might sleep, we cannot do it with a spinlock held ...
4510 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4519 static int sched_read_attr(struct sched_attr __user
*uattr
,
4520 struct sched_attr
*attr
,
4525 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4529 * If we're handed a smaller struct than we know of,
4530 * ensure all the unknown bits are 0 - i.e. old
4531 * user-space does not get uncomplete information.
4533 if (usize
< sizeof(*attr
)) {
4534 unsigned char *addr
;
4537 addr
= (void *)attr
+ usize
;
4538 end
= (void *)attr
+ sizeof(*attr
);
4540 for (; addr
< end
; addr
++) {
4548 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4556 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4557 * @pid: the pid in question.
4558 * @uattr: structure containing the extended parameters.
4559 * @size: sizeof(attr) for fwd/bwd comp.
4560 * @flags: for future extension.
4562 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4563 unsigned int, size
, unsigned int, flags
)
4565 struct sched_attr attr
= {
4566 .size
= sizeof(struct sched_attr
),
4568 struct task_struct
*p
;
4571 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4572 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4576 p
= find_process_by_pid(pid
);
4581 retval
= security_task_getscheduler(p
);
4585 attr
.sched_policy
= p
->policy
;
4586 if (p
->sched_reset_on_fork
)
4587 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4588 if (task_has_dl_policy(p
))
4589 __getparam_dl(p
, &attr
);
4590 else if (task_has_rt_policy(p
))
4591 attr
.sched_priority
= p
->rt_priority
;
4593 attr
.sched_nice
= task_nice(p
);
4597 retval
= sched_read_attr(uattr
, &attr
, size
);
4605 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4607 cpumask_var_t cpus_allowed
, new_mask
;
4608 struct task_struct
*p
;
4613 p
= find_process_by_pid(pid
);
4619 /* Prevent p going away */
4623 if (p
->flags
& PF_NO_SETAFFINITY
) {
4627 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4631 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4633 goto out_free_cpus_allowed
;
4636 if (!check_same_owner(p
)) {
4638 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4640 goto out_free_new_mask
;
4645 retval
= security_task_setscheduler(p
);
4647 goto out_free_new_mask
;
4650 cpuset_cpus_allowed(p
, cpus_allowed
);
4651 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4654 * Since bandwidth control happens on root_domain basis,
4655 * if admission test is enabled, we only admit -deadline
4656 * tasks allowed to run on all the CPUs in the task's
4660 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4662 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4665 goto out_free_new_mask
;
4671 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4674 cpuset_cpus_allowed(p
, cpus_allowed
);
4675 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4677 * We must have raced with a concurrent cpuset
4678 * update. Just reset the cpus_allowed to the
4679 * cpuset's cpus_allowed
4681 cpumask_copy(new_mask
, cpus_allowed
);
4686 free_cpumask_var(new_mask
);
4687 out_free_cpus_allowed
:
4688 free_cpumask_var(cpus_allowed
);
4694 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4695 struct cpumask
*new_mask
)
4697 if (len
< cpumask_size())
4698 cpumask_clear(new_mask
);
4699 else if (len
> cpumask_size())
4700 len
= cpumask_size();
4702 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4706 * sys_sched_setaffinity - set the cpu affinity of a process
4707 * @pid: pid of the process
4708 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4709 * @user_mask_ptr: user-space pointer to the new cpu mask
4711 * Return: 0 on success. An error code otherwise.
4713 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4714 unsigned long __user
*, user_mask_ptr
)
4716 cpumask_var_t new_mask
;
4719 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4722 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4724 retval
= sched_setaffinity(pid
, new_mask
);
4725 free_cpumask_var(new_mask
);
4729 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4731 struct task_struct
*p
;
4732 unsigned long flags
;
4738 p
= find_process_by_pid(pid
);
4742 retval
= security_task_getscheduler(p
);
4746 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4747 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4748 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4757 * sys_sched_getaffinity - get the cpu affinity of a process
4758 * @pid: pid of the process
4759 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4760 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4762 * Return: 0 on success. An error code otherwise.
4764 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4765 unsigned long __user
*, user_mask_ptr
)
4770 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4772 if (len
& (sizeof(unsigned long)-1))
4775 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4778 ret
= sched_getaffinity(pid
, mask
);
4780 size_t retlen
= min_t(size_t, len
, cpumask_size());
4782 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4787 free_cpumask_var(mask
);
4793 * sys_sched_yield - yield the current processor to other threads.
4795 * This function yields the current CPU to other tasks. If there are no
4796 * other threads running on this CPU then this function will return.
4800 SYSCALL_DEFINE0(sched_yield
)
4802 struct rq
*rq
= this_rq_lock();
4804 schedstat_inc(rq
, yld_count
);
4805 current
->sched_class
->yield_task(rq
);
4808 * Since we are going to call schedule() anyway, there's
4809 * no need to preempt or enable interrupts:
4811 __release(rq
->lock
);
4812 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4813 do_raw_spin_unlock(&rq
->lock
);
4814 sched_preempt_enable_no_resched();
4821 int __sched
_cond_resched(void)
4823 if (should_resched(0)) {
4824 preempt_schedule_common();
4829 EXPORT_SYMBOL(_cond_resched
);
4832 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4833 * call schedule, and on return reacquire the lock.
4835 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4836 * operations here to prevent schedule() from being called twice (once via
4837 * spin_unlock(), once by hand).
4839 int __cond_resched_lock(spinlock_t
*lock
)
4841 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4844 lockdep_assert_held(lock
);
4846 if (spin_needbreak(lock
) || resched
) {
4849 preempt_schedule_common();
4857 EXPORT_SYMBOL(__cond_resched_lock
);
4859 int __sched
__cond_resched_softirq(void)
4861 BUG_ON(!in_softirq());
4863 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4865 preempt_schedule_common();
4871 EXPORT_SYMBOL(__cond_resched_softirq
);
4874 * yield - yield the current processor to other threads.
4876 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4878 * The scheduler is at all times free to pick the calling task as the most
4879 * eligible task to run, if removing the yield() call from your code breaks
4880 * it, its already broken.
4882 * Typical broken usage is:
4887 * where one assumes that yield() will let 'the other' process run that will
4888 * make event true. If the current task is a SCHED_FIFO task that will never
4889 * happen. Never use yield() as a progress guarantee!!
4891 * If you want to use yield() to wait for something, use wait_event().
4892 * If you want to use yield() to be 'nice' for others, use cond_resched().
4893 * If you still want to use yield(), do not!
4895 void __sched
yield(void)
4897 set_current_state(TASK_RUNNING
);
4900 EXPORT_SYMBOL(yield
);
4903 * yield_to - yield the current processor to another thread in
4904 * your thread group, or accelerate that thread toward the
4905 * processor it's on.
4907 * @preempt: whether task preemption is allowed or not
4909 * It's the caller's job to ensure that the target task struct
4910 * can't go away on us before we can do any checks.
4913 * true (>0) if we indeed boosted the target task.
4914 * false (0) if we failed to boost the target.
4915 * -ESRCH if there's no task to yield to.
4917 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4919 struct task_struct
*curr
= current
;
4920 struct rq
*rq
, *p_rq
;
4921 unsigned long flags
;
4924 local_irq_save(flags
);
4930 * If we're the only runnable task on the rq and target rq also
4931 * has only one task, there's absolutely no point in yielding.
4933 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4938 double_rq_lock(rq
, p_rq
);
4939 if (task_rq(p
) != p_rq
) {
4940 double_rq_unlock(rq
, p_rq
);
4944 if (!curr
->sched_class
->yield_to_task
)
4947 if (curr
->sched_class
!= p
->sched_class
)
4950 if (task_running(p_rq
, p
) || p
->state
)
4953 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4955 schedstat_inc(rq
, yld_count
);
4957 * Make p's CPU reschedule; pick_next_entity takes care of
4960 if (preempt
&& rq
!= p_rq
)
4965 double_rq_unlock(rq
, p_rq
);
4967 local_irq_restore(flags
);
4974 EXPORT_SYMBOL_GPL(yield_to
);
4977 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4978 * that process accounting knows that this is a task in IO wait state.
4980 long __sched
io_schedule_timeout(long timeout
)
4982 int old_iowait
= current
->in_iowait
;
4986 current
->in_iowait
= 1;
4987 blk_schedule_flush_plug(current
);
4989 delayacct_blkio_start();
4991 atomic_inc(&rq
->nr_iowait
);
4992 ret
= schedule_timeout(timeout
);
4993 current
->in_iowait
= old_iowait
;
4994 atomic_dec(&rq
->nr_iowait
);
4995 delayacct_blkio_end();
4999 EXPORT_SYMBOL(io_schedule_timeout
);
5002 * sys_sched_get_priority_max - return maximum RT priority.
5003 * @policy: scheduling class.
5005 * Return: On success, this syscall returns the maximum
5006 * rt_priority that can be used by a given scheduling class.
5007 * On failure, a negative error code is returned.
5009 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5016 ret
= MAX_USER_RT_PRIO
-1;
5018 case SCHED_DEADLINE
:
5029 * sys_sched_get_priority_min - return minimum RT priority.
5030 * @policy: scheduling class.
5032 * Return: On success, this syscall returns the minimum
5033 * rt_priority that can be used by a given scheduling class.
5034 * On failure, a negative error code is returned.
5036 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5045 case SCHED_DEADLINE
:
5055 * sys_sched_rr_get_interval - return the default timeslice of a process.
5056 * @pid: pid of the process.
5057 * @interval: userspace pointer to the timeslice value.
5059 * this syscall writes the default timeslice value of a given process
5060 * into the user-space timespec buffer. A value of '0' means infinity.
5062 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5065 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5066 struct timespec __user
*, interval
)
5068 struct task_struct
*p
;
5069 unsigned int time_slice
;
5080 p
= find_process_by_pid(pid
);
5084 retval
= security_task_getscheduler(p
);
5088 rq
= task_rq_lock(p
, &rf
);
5090 if (p
->sched_class
->get_rr_interval
)
5091 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5092 task_rq_unlock(rq
, p
, &rf
);
5095 jiffies_to_timespec(time_slice
, &t
);
5096 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5104 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5106 void sched_show_task(struct task_struct
*p
)
5108 unsigned long free
= 0;
5110 unsigned long state
= p
->state
;
5113 state
= __ffs(state
) + 1;
5114 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5115 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5116 #if BITS_PER_LONG == 32
5117 if (state
== TASK_RUNNING
)
5118 printk(KERN_CONT
" running ");
5120 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5122 if (state
== TASK_RUNNING
)
5123 printk(KERN_CONT
" running task ");
5125 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5127 #ifdef CONFIG_DEBUG_STACK_USAGE
5128 free
= stack_not_used(p
);
5133 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5135 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5136 task_pid_nr(p
), ppid
,
5137 (unsigned long)task_thread_info(p
)->flags
);
5139 print_worker_info(KERN_INFO
, p
);
5140 show_stack(p
, NULL
);
5143 void show_state_filter(unsigned long state_filter
)
5145 struct task_struct
*g
, *p
;
5147 #if BITS_PER_LONG == 32
5149 " task PC stack pid father\n");
5152 " task PC stack pid father\n");
5155 for_each_process_thread(g
, p
) {
5157 * reset the NMI-timeout, listing all files on a slow
5158 * console might take a lot of time:
5159 * Also, reset softlockup watchdogs on all CPUs, because
5160 * another CPU might be blocked waiting for us to process
5163 touch_nmi_watchdog();
5164 touch_all_softlockup_watchdogs();
5165 if (!state_filter
|| (p
->state
& state_filter
))
5169 #ifdef CONFIG_SCHED_DEBUG
5171 sysrq_sched_debug_show();
5175 * Only show locks if all tasks are dumped:
5178 debug_show_all_locks();
5181 void init_idle_bootup_task(struct task_struct
*idle
)
5183 idle
->sched_class
= &idle_sched_class
;
5187 * init_idle - set up an idle thread for a given CPU
5188 * @idle: task in question
5189 * @cpu: cpu the idle task belongs to
5191 * NOTE: this function does not set the idle thread's NEED_RESCHED
5192 * flag, to make booting more robust.
5194 void init_idle(struct task_struct
*idle
, int cpu
)
5196 struct rq
*rq
= cpu_rq(cpu
);
5197 unsigned long flags
;
5199 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5200 raw_spin_lock(&rq
->lock
);
5202 __sched_fork(0, idle
);
5203 idle
->state
= TASK_RUNNING
;
5204 idle
->se
.exec_start
= sched_clock();
5206 kasan_unpoison_task_stack(idle
);
5210 * Its possible that init_idle() gets called multiple times on a task,
5211 * in that case do_set_cpus_allowed() will not do the right thing.
5213 * And since this is boot we can forgo the serialization.
5215 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5218 * We're having a chicken and egg problem, even though we are
5219 * holding rq->lock, the cpu isn't yet set to this cpu so the
5220 * lockdep check in task_group() will fail.
5222 * Similar case to sched_fork(). / Alternatively we could
5223 * use task_rq_lock() here and obtain the other rq->lock.
5228 __set_task_cpu(idle
, cpu
);
5231 rq
->curr
= rq
->idle
= idle
;
5232 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5236 raw_spin_unlock(&rq
->lock
);
5237 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5239 /* Set the preempt count _outside_ the spinlocks! */
5240 init_idle_preempt_count(idle
, cpu
);
5243 * The idle tasks have their own, simple scheduling class:
5245 idle
->sched_class
= &idle_sched_class
;
5246 ftrace_graph_init_idle_task(idle
, cpu
);
5247 vtime_init_idle(idle
, cpu
);
5249 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5253 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5254 const struct cpumask
*trial
)
5256 int ret
= 1, trial_cpus
;
5257 struct dl_bw
*cur_dl_b
;
5258 unsigned long flags
;
5260 if (!cpumask_weight(cur
))
5263 rcu_read_lock_sched();
5264 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5265 trial_cpus
= cpumask_weight(trial
);
5267 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5268 if (cur_dl_b
->bw
!= -1 &&
5269 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5271 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5272 rcu_read_unlock_sched();
5277 int task_can_attach(struct task_struct
*p
,
5278 const struct cpumask
*cs_cpus_allowed
)
5283 * Kthreads which disallow setaffinity shouldn't be moved
5284 * to a new cpuset; we don't want to change their cpu
5285 * affinity and isolating such threads by their set of
5286 * allowed nodes is unnecessary. Thus, cpusets are not
5287 * applicable for such threads. This prevents checking for
5288 * success of set_cpus_allowed_ptr() on all attached tasks
5289 * before cpus_allowed may be changed.
5291 if (p
->flags
& PF_NO_SETAFFINITY
) {
5297 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5299 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5304 unsigned long flags
;
5306 rcu_read_lock_sched();
5307 dl_b
= dl_bw_of(dest_cpu
);
5308 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5309 cpus
= dl_bw_cpus(dest_cpu
);
5310 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5315 * We reserve space for this task in the destination
5316 * root_domain, as we can't fail after this point.
5317 * We will free resources in the source root_domain
5318 * later on (see set_cpus_allowed_dl()).
5320 __dl_add(dl_b
, p
->dl
.dl_bw
);
5322 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5323 rcu_read_unlock_sched();
5333 static bool sched_smp_initialized __read_mostly
;
5335 #ifdef CONFIG_NUMA_BALANCING
5336 /* Migrate current task p to target_cpu */
5337 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5339 struct migration_arg arg
= { p
, target_cpu
};
5340 int curr_cpu
= task_cpu(p
);
5342 if (curr_cpu
== target_cpu
)
5345 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5348 /* TODO: This is not properly updating schedstats */
5350 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5351 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5355 * Requeue a task on a given node and accurately track the number of NUMA
5356 * tasks on the runqueues
5358 void sched_setnuma(struct task_struct
*p
, int nid
)
5360 bool queued
, running
;
5364 rq
= task_rq_lock(p
, &rf
);
5365 queued
= task_on_rq_queued(p
);
5366 running
= task_current(rq
, p
);
5369 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5371 put_prev_task(rq
, p
);
5373 p
->numa_preferred_nid
= nid
;
5376 p
->sched_class
->set_curr_task(rq
);
5378 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5379 task_rq_unlock(rq
, p
, &rf
);
5381 #endif /* CONFIG_NUMA_BALANCING */
5383 #ifdef CONFIG_HOTPLUG_CPU
5385 * Ensures that the idle task is using init_mm right before its cpu goes
5388 void idle_task_exit(void)
5390 struct mm_struct
*mm
= current
->active_mm
;
5392 BUG_ON(cpu_online(smp_processor_id()));
5394 if (mm
!= &init_mm
) {
5395 switch_mm_irqs_off(mm
, &init_mm
, current
);
5396 finish_arch_post_lock_switch();
5402 * Since this CPU is going 'away' for a while, fold any nr_active delta
5403 * we might have. Assumes we're called after migrate_tasks() so that the
5404 * nr_active count is stable.
5406 * Also see the comment "Global load-average calculations".
5408 static void calc_load_migrate(struct rq
*rq
)
5410 long delta
= calc_load_fold_active(rq
);
5412 atomic_long_add(delta
, &calc_load_tasks
);
5415 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5419 static const struct sched_class fake_sched_class
= {
5420 .put_prev_task
= put_prev_task_fake
,
5423 static struct task_struct fake_task
= {
5425 * Avoid pull_{rt,dl}_task()
5427 .prio
= MAX_PRIO
+ 1,
5428 .sched_class
= &fake_sched_class
,
5432 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5433 * try_to_wake_up()->select_task_rq().
5435 * Called with rq->lock held even though we'er in stop_machine() and
5436 * there's no concurrency possible, we hold the required locks anyway
5437 * because of lock validation efforts.
5439 static void migrate_tasks(struct rq
*dead_rq
)
5441 struct rq
*rq
= dead_rq
;
5442 struct task_struct
*next
, *stop
= rq
->stop
;
5443 struct pin_cookie cookie
;
5447 * Fudge the rq selection such that the below task selection loop
5448 * doesn't get stuck on the currently eligible stop task.
5450 * We're currently inside stop_machine() and the rq is either stuck
5451 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5452 * either way we should never end up calling schedule() until we're
5458 * put_prev_task() and pick_next_task() sched
5459 * class method both need to have an up-to-date
5460 * value of rq->clock[_task]
5462 update_rq_clock(rq
);
5466 * There's this thread running, bail when that's the only
5469 if (rq
->nr_running
== 1)
5473 * pick_next_task assumes pinned rq->lock.
5475 cookie
= lockdep_pin_lock(&rq
->lock
);
5476 next
= pick_next_task(rq
, &fake_task
, cookie
);
5478 next
->sched_class
->put_prev_task(rq
, next
);
5481 * Rules for changing task_struct::cpus_allowed are holding
5482 * both pi_lock and rq->lock, such that holding either
5483 * stabilizes the mask.
5485 * Drop rq->lock is not quite as disastrous as it usually is
5486 * because !cpu_active at this point, which means load-balance
5487 * will not interfere. Also, stop-machine.
5489 lockdep_unpin_lock(&rq
->lock
, cookie
);
5490 raw_spin_unlock(&rq
->lock
);
5491 raw_spin_lock(&next
->pi_lock
);
5492 raw_spin_lock(&rq
->lock
);
5495 * Since we're inside stop-machine, _nothing_ should have
5496 * changed the task, WARN if weird stuff happened, because in
5497 * that case the above rq->lock drop is a fail too.
5499 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5500 raw_spin_unlock(&next
->pi_lock
);
5504 /* Find suitable destination for @next, with force if needed. */
5505 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5507 rq
= __migrate_task(rq
, next
, dest_cpu
);
5508 if (rq
!= dead_rq
) {
5509 raw_spin_unlock(&rq
->lock
);
5511 raw_spin_lock(&rq
->lock
);
5513 raw_spin_unlock(&next
->pi_lock
);
5518 #endif /* CONFIG_HOTPLUG_CPU */
5520 static void set_rq_online(struct rq
*rq
)
5523 const struct sched_class
*class;
5525 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5528 for_each_class(class) {
5529 if (class->rq_online
)
5530 class->rq_online(rq
);
5535 static void set_rq_offline(struct rq
*rq
)
5538 const struct sched_class
*class;
5540 for_each_class(class) {
5541 if (class->rq_offline
)
5542 class->rq_offline(rq
);
5545 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5550 static void set_cpu_rq_start_time(unsigned int cpu
)
5552 struct rq
*rq
= cpu_rq(cpu
);
5554 rq
->age_stamp
= sched_clock_cpu(cpu
);
5557 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5559 #ifdef CONFIG_SCHED_DEBUG
5561 static __read_mostly
int sched_debug_enabled
;
5563 static int __init
sched_debug_setup(char *str
)
5565 sched_debug_enabled
= 1;
5569 early_param("sched_debug", sched_debug_setup
);
5571 static inline bool sched_debug(void)
5573 return sched_debug_enabled
;
5576 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5577 struct cpumask
*groupmask
)
5579 struct sched_group
*group
= sd
->groups
;
5581 cpumask_clear(groupmask
);
5583 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5585 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5586 printk("does not load-balance\n");
5588 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5593 printk(KERN_CONT
"span %*pbl level %s\n",
5594 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5596 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5597 printk(KERN_ERR
"ERROR: domain->span does not contain "
5600 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5601 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5605 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5609 printk(KERN_ERR
"ERROR: group is NULL\n");
5613 if (!cpumask_weight(sched_group_cpus(group
))) {
5614 printk(KERN_CONT
"\n");
5615 printk(KERN_ERR
"ERROR: empty group\n");
5619 if (!(sd
->flags
& SD_OVERLAP
) &&
5620 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5621 printk(KERN_CONT
"\n");
5622 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5626 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5628 printk(KERN_CONT
" %*pbl",
5629 cpumask_pr_args(sched_group_cpus(group
)));
5630 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5631 printk(KERN_CONT
" (cpu_capacity = %d)",
5632 group
->sgc
->capacity
);
5635 group
= group
->next
;
5636 } while (group
!= sd
->groups
);
5637 printk(KERN_CONT
"\n");
5639 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5640 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5643 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5644 printk(KERN_ERR
"ERROR: parent span is not a superset "
5645 "of domain->span\n");
5649 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5653 if (!sched_debug_enabled
)
5657 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5661 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5664 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5672 #else /* !CONFIG_SCHED_DEBUG */
5673 # define sched_domain_debug(sd, cpu) do { } while (0)
5674 static inline bool sched_debug(void)
5678 #endif /* CONFIG_SCHED_DEBUG */
5680 static int sd_degenerate(struct sched_domain
*sd
)
5682 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5685 /* Following flags need at least 2 groups */
5686 if (sd
->flags
& (SD_LOAD_BALANCE
|
5687 SD_BALANCE_NEWIDLE
|
5690 SD_SHARE_CPUCAPACITY
|
5691 SD_SHARE_PKG_RESOURCES
|
5692 SD_SHARE_POWERDOMAIN
)) {
5693 if (sd
->groups
!= sd
->groups
->next
)
5697 /* Following flags don't use groups */
5698 if (sd
->flags
& (SD_WAKE_AFFINE
))
5705 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5707 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5709 if (sd_degenerate(parent
))
5712 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5715 /* Flags needing groups don't count if only 1 group in parent */
5716 if (parent
->groups
== parent
->groups
->next
) {
5717 pflags
&= ~(SD_LOAD_BALANCE
|
5718 SD_BALANCE_NEWIDLE
|
5721 SD_SHARE_CPUCAPACITY
|
5722 SD_SHARE_PKG_RESOURCES
|
5724 SD_SHARE_POWERDOMAIN
);
5725 if (nr_node_ids
== 1)
5726 pflags
&= ~SD_SERIALIZE
;
5728 if (~cflags
& pflags
)
5734 static void free_rootdomain(struct rcu_head
*rcu
)
5736 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5738 cpupri_cleanup(&rd
->cpupri
);
5739 cpudl_cleanup(&rd
->cpudl
);
5740 free_cpumask_var(rd
->dlo_mask
);
5741 free_cpumask_var(rd
->rto_mask
);
5742 free_cpumask_var(rd
->online
);
5743 free_cpumask_var(rd
->span
);
5747 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5749 struct root_domain
*old_rd
= NULL
;
5750 unsigned long flags
;
5752 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5757 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5760 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5763 * If we dont want to free the old_rd yet then
5764 * set old_rd to NULL to skip the freeing later
5767 if (!atomic_dec_and_test(&old_rd
->refcount
))
5771 atomic_inc(&rd
->refcount
);
5774 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5775 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5778 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5781 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5784 static int init_rootdomain(struct root_domain
*rd
)
5786 memset(rd
, 0, sizeof(*rd
));
5788 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5790 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5792 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5794 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5797 init_dl_bw(&rd
->dl_bw
);
5798 if (cpudl_init(&rd
->cpudl
) != 0)
5801 if (cpupri_init(&rd
->cpupri
) != 0)
5806 free_cpumask_var(rd
->rto_mask
);
5808 free_cpumask_var(rd
->dlo_mask
);
5810 free_cpumask_var(rd
->online
);
5812 free_cpumask_var(rd
->span
);
5818 * By default the system creates a single root-domain with all cpus as
5819 * members (mimicking the global state we have today).
5821 struct root_domain def_root_domain
;
5823 static void init_defrootdomain(void)
5825 init_rootdomain(&def_root_domain
);
5827 atomic_set(&def_root_domain
.refcount
, 1);
5830 static struct root_domain
*alloc_rootdomain(void)
5832 struct root_domain
*rd
;
5834 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5838 if (init_rootdomain(rd
) != 0) {
5846 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5848 struct sched_group
*tmp
, *first
;
5857 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5862 } while (sg
!= first
);
5865 static void free_sched_domain(struct rcu_head
*rcu
)
5867 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5870 * If its an overlapping domain it has private groups, iterate and
5873 if (sd
->flags
& SD_OVERLAP
) {
5874 free_sched_groups(sd
->groups
, 1);
5875 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5876 kfree(sd
->groups
->sgc
);
5882 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5884 call_rcu(&sd
->rcu
, free_sched_domain
);
5887 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5889 for (; sd
; sd
= sd
->parent
)
5890 destroy_sched_domain(sd
, cpu
);
5894 * Keep a special pointer to the highest sched_domain that has
5895 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5896 * allows us to avoid some pointer chasing select_idle_sibling().
5898 * Also keep a unique ID per domain (we use the first cpu number in
5899 * the cpumask of the domain), this allows us to quickly tell if
5900 * two cpus are in the same cache domain, see cpus_share_cache().
5902 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5903 DEFINE_PER_CPU(int, sd_llc_size
);
5904 DEFINE_PER_CPU(int, sd_llc_id
);
5905 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5906 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5907 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5909 static void update_top_cache_domain(int cpu
)
5911 struct sched_domain
*sd
;
5912 struct sched_domain
*busy_sd
= NULL
;
5916 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5918 id
= cpumask_first(sched_domain_span(sd
));
5919 size
= cpumask_weight(sched_domain_span(sd
));
5920 busy_sd
= sd
->parent
; /* sd_busy */
5922 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5924 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5925 per_cpu(sd_llc_size
, cpu
) = size
;
5926 per_cpu(sd_llc_id
, cpu
) = id
;
5928 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5929 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5931 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5932 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5936 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5937 * hold the hotplug lock.
5940 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5942 struct rq
*rq
= cpu_rq(cpu
);
5943 struct sched_domain
*tmp
;
5945 /* Remove the sched domains which do not contribute to scheduling. */
5946 for (tmp
= sd
; tmp
; ) {
5947 struct sched_domain
*parent
= tmp
->parent
;
5951 if (sd_parent_degenerate(tmp
, parent
)) {
5952 tmp
->parent
= parent
->parent
;
5954 parent
->parent
->child
= tmp
;
5956 * Transfer SD_PREFER_SIBLING down in case of a
5957 * degenerate parent; the spans match for this
5958 * so the property transfers.
5960 if (parent
->flags
& SD_PREFER_SIBLING
)
5961 tmp
->flags
|= SD_PREFER_SIBLING
;
5962 destroy_sched_domain(parent
, cpu
);
5967 if (sd
&& sd_degenerate(sd
)) {
5970 destroy_sched_domain(tmp
, cpu
);
5975 sched_domain_debug(sd
, cpu
);
5977 rq_attach_root(rq
, rd
);
5979 rcu_assign_pointer(rq
->sd
, sd
);
5980 destroy_sched_domains(tmp
, cpu
);
5982 update_top_cache_domain(cpu
);
5985 /* Setup the mask of cpus configured for isolated domains */
5986 static int __init
isolated_cpu_setup(char *str
)
5990 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5991 ret
= cpulist_parse(str
, cpu_isolated_map
);
5993 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
5998 __setup("isolcpus=", isolated_cpu_setup
);
6001 struct sched_domain
** __percpu sd
;
6002 struct root_domain
*rd
;
6013 * Build an iteration mask that can exclude certain CPUs from the upwards
6016 * Asymmetric node setups can result in situations where the domain tree is of
6017 * unequal depth, make sure to skip domains that already cover the entire
6020 * In that case build_sched_domains() will have terminated the iteration early
6021 * and our sibling sd spans will be empty. Domains should always include the
6022 * cpu they're built on, so check that.
6025 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6027 const struct cpumask
*span
= sched_domain_span(sd
);
6028 struct sd_data
*sdd
= sd
->private;
6029 struct sched_domain
*sibling
;
6032 for_each_cpu(i
, span
) {
6033 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6034 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6037 cpumask_set_cpu(i
, sched_group_mask(sg
));
6042 * Return the canonical balance cpu for this group, this is the first cpu
6043 * of this group that's also in the iteration mask.
6045 int group_balance_cpu(struct sched_group
*sg
)
6047 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6051 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6053 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6054 const struct cpumask
*span
= sched_domain_span(sd
);
6055 struct cpumask
*covered
= sched_domains_tmpmask
;
6056 struct sd_data
*sdd
= sd
->private;
6057 struct sched_domain
*sibling
;
6060 cpumask_clear(covered
);
6062 for_each_cpu(i
, span
) {
6063 struct cpumask
*sg_span
;
6065 if (cpumask_test_cpu(i
, covered
))
6068 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6070 /* See the comment near build_group_mask(). */
6071 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6074 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6075 GFP_KERNEL
, cpu_to_node(cpu
));
6080 sg_span
= sched_group_cpus(sg
);
6082 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6084 cpumask_set_cpu(i
, sg_span
);
6086 cpumask_or(covered
, covered
, sg_span
);
6088 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6089 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6090 build_group_mask(sd
, sg
);
6093 * Initialize sgc->capacity such that even if we mess up the
6094 * domains and no possible iteration will get us here, we won't
6097 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6100 * Make sure the first group of this domain contains the
6101 * canonical balance cpu. Otherwise the sched_domain iteration
6102 * breaks. See update_sg_lb_stats().
6104 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6105 group_balance_cpu(sg
) == cpu
)
6115 sd
->groups
= groups
;
6120 free_sched_groups(first
, 0);
6125 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6127 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6128 struct sched_domain
*child
= sd
->child
;
6131 cpu
= cpumask_first(sched_domain_span(child
));
6134 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6135 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6136 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6143 * build_sched_groups will build a circular linked list of the groups
6144 * covered by the given span, and will set each group's ->cpumask correctly,
6145 * and ->cpu_capacity to 0.
6147 * Assumes the sched_domain tree is fully constructed
6150 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6152 struct sched_group
*first
= NULL
, *last
= NULL
;
6153 struct sd_data
*sdd
= sd
->private;
6154 const struct cpumask
*span
= sched_domain_span(sd
);
6155 struct cpumask
*covered
;
6158 get_group(cpu
, sdd
, &sd
->groups
);
6159 atomic_inc(&sd
->groups
->ref
);
6161 if (cpu
!= cpumask_first(span
))
6164 lockdep_assert_held(&sched_domains_mutex
);
6165 covered
= sched_domains_tmpmask
;
6167 cpumask_clear(covered
);
6169 for_each_cpu(i
, span
) {
6170 struct sched_group
*sg
;
6173 if (cpumask_test_cpu(i
, covered
))
6176 group
= get_group(i
, sdd
, &sg
);
6177 cpumask_setall(sched_group_mask(sg
));
6179 for_each_cpu(j
, span
) {
6180 if (get_group(j
, sdd
, NULL
) != group
)
6183 cpumask_set_cpu(j
, covered
);
6184 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6199 * Initialize sched groups cpu_capacity.
6201 * cpu_capacity indicates the capacity of sched group, which is used while
6202 * distributing the load between different sched groups in a sched domain.
6203 * Typically cpu_capacity for all the groups in a sched domain will be same
6204 * unless there are asymmetries in the topology. If there are asymmetries,
6205 * group having more cpu_capacity will pickup more load compared to the
6206 * group having less cpu_capacity.
6208 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6210 struct sched_group
*sg
= sd
->groups
;
6215 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6217 } while (sg
!= sd
->groups
);
6219 if (cpu
!= group_balance_cpu(sg
))
6222 update_group_capacity(sd
, cpu
);
6223 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6227 * Initializers for schedule domains
6228 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6231 static int default_relax_domain_level
= -1;
6232 int sched_domain_level_max
;
6234 static int __init
setup_relax_domain_level(char *str
)
6236 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6237 pr_warn("Unable to set relax_domain_level\n");
6241 __setup("relax_domain_level=", setup_relax_domain_level
);
6243 static void set_domain_attribute(struct sched_domain
*sd
,
6244 struct sched_domain_attr
*attr
)
6248 if (!attr
|| attr
->relax_domain_level
< 0) {
6249 if (default_relax_domain_level
< 0)
6252 request
= default_relax_domain_level
;
6254 request
= attr
->relax_domain_level
;
6255 if (request
< sd
->level
) {
6256 /* turn off idle balance on this domain */
6257 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6259 /* turn on idle balance on this domain */
6260 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6264 static void __sdt_free(const struct cpumask
*cpu_map
);
6265 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6267 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6268 const struct cpumask
*cpu_map
)
6272 if (!atomic_read(&d
->rd
->refcount
))
6273 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6275 free_percpu(d
->sd
); /* fall through */
6277 __sdt_free(cpu_map
); /* fall through */
6283 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6284 const struct cpumask
*cpu_map
)
6286 memset(d
, 0, sizeof(*d
));
6288 if (__sdt_alloc(cpu_map
))
6289 return sa_sd_storage
;
6290 d
->sd
= alloc_percpu(struct sched_domain
*);
6292 return sa_sd_storage
;
6293 d
->rd
= alloc_rootdomain();
6296 return sa_rootdomain
;
6300 * NULL the sd_data elements we've used to build the sched_domain and
6301 * sched_group structure so that the subsequent __free_domain_allocs()
6302 * will not free the data we're using.
6304 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6306 struct sd_data
*sdd
= sd
->private;
6308 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6309 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6311 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6312 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6314 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6315 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6319 static int sched_domains_numa_levels
;
6320 enum numa_topology_type sched_numa_topology_type
;
6321 static int *sched_domains_numa_distance
;
6322 int sched_max_numa_distance
;
6323 static struct cpumask
***sched_domains_numa_masks
;
6324 static int sched_domains_curr_level
;
6328 * SD_flags allowed in topology descriptions.
6330 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6331 * SD_SHARE_PKG_RESOURCES - describes shared caches
6332 * SD_NUMA - describes NUMA topologies
6333 * SD_SHARE_POWERDOMAIN - describes shared power domain
6336 * SD_ASYM_PACKING - describes SMT quirks
6338 #define TOPOLOGY_SD_FLAGS \
6339 (SD_SHARE_CPUCAPACITY | \
6340 SD_SHARE_PKG_RESOURCES | \
6343 SD_SHARE_POWERDOMAIN)
6345 static struct sched_domain
*
6346 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6348 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6349 int sd_weight
, sd_flags
= 0;
6353 * Ugly hack to pass state to sd_numa_mask()...
6355 sched_domains_curr_level
= tl
->numa_level
;
6358 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6361 sd_flags
= (*tl
->sd_flags
)();
6362 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6363 "wrong sd_flags in topology description\n"))
6364 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6366 *sd
= (struct sched_domain
){
6367 .min_interval
= sd_weight
,
6368 .max_interval
= 2*sd_weight
,
6370 .imbalance_pct
= 125,
6372 .cache_nice_tries
= 0,
6379 .flags
= 1*SD_LOAD_BALANCE
6380 | 1*SD_BALANCE_NEWIDLE
6385 | 0*SD_SHARE_CPUCAPACITY
6386 | 0*SD_SHARE_PKG_RESOURCES
6388 | 0*SD_PREFER_SIBLING
6393 .last_balance
= jiffies
,
6394 .balance_interval
= sd_weight
,
6396 .max_newidle_lb_cost
= 0,
6397 .next_decay_max_lb_cost
= jiffies
,
6398 #ifdef CONFIG_SCHED_DEBUG
6404 * Convert topological properties into behaviour.
6407 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6408 sd
->flags
|= SD_PREFER_SIBLING
;
6409 sd
->imbalance_pct
= 110;
6410 sd
->smt_gain
= 1178; /* ~15% */
6412 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6413 sd
->imbalance_pct
= 117;
6414 sd
->cache_nice_tries
= 1;
6418 } else if (sd
->flags
& SD_NUMA
) {
6419 sd
->cache_nice_tries
= 2;
6423 sd
->flags
|= SD_SERIALIZE
;
6424 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6425 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6432 sd
->flags
|= SD_PREFER_SIBLING
;
6433 sd
->cache_nice_tries
= 1;
6438 sd
->private = &tl
->data
;
6444 * Topology list, bottom-up.
6446 static struct sched_domain_topology_level default_topology
[] = {
6447 #ifdef CONFIG_SCHED_SMT
6448 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6450 #ifdef CONFIG_SCHED_MC
6451 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6453 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6457 static struct sched_domain_topology_level
*sched_domain_topology
=
6460 #define for_each_sd_topology(tl) \
6461 for (tl = sched_domain_topology; tl->mask; tl++)
6463 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6465 sched_domain_topology
= tl
;
6470 static const struct cpumask
*sd_numa_mask(int cpu
)
6472 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6475 static void sched_numa_warn(const char *str
)
6477 static int done
= false;
6485 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6487 for (i
= 0; i
< nr_node_ids
; i
++) {
6488 printk(KERN_WARNING
" ");
6489 for (j
= 0; j
< nr_node_ids
; j
++)
6490 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6491 printk(KERN_CONT
"\n");
6493 printk(KERN_WARNING
"\n");
6496 bool find_numa_distance(int distance
)
6500 if (distance
== node_distance(0, 0))
6503 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6504 if (sched_domains_numa_distance
[i
] == distance
)
6512 * A system can have three types of NUMA topology:
6513 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6514 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6515 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6517 * The difference between a glueless mesh topology and a backplane
6518 * topology lies in whether communication between not directly
6519 * connected nodes goes through intermediary nodes (where programs
6520 * could run), or through backplane controllers. This affects
6521 * placement of programs.
6523 * The type of topology can be discerned with the following tests:
6524 * - If the maximum distance between any nodes is 1 hop, the system
6525 * is directly connected.
6526 * - If for two nodes A and B, located N > 1 hops away from each other,
6527 * there is an intermediary node C, which is < N hops away from both
6528 * nodes A and B, the system is a glueless mesh.
6530 static void init_numa_topology_type(void)
6534 n
= sched_max_numa_distance
;
6536 if (sched_domains_numa_levels
<= 1) {
6537 sched_numa_topology_type
= NUMA_DIRECT
;
6541 for_each_online_node(a
) {
6542 for_each_online_node(b
) {
6543 /* Find two nodes furthest removed from each other. */
6544 if (node_distance(a
, b
) < n
)
6547 /* Is there an intermediary node between a and b? */
6548 for_each_online_node(c
) {
6549 if (node_distance(a
, c
) < n
&&
6550 node_distance(b
, c
) < n
) {
6551 sched_numa_topology_type
=
6557 sched_numa_topology_type
= NUMA_BACKPLANE
;
6563 static void sched_init_numa(void)
6565 int next_distance
, curr_distance
= node_distance(0, 0);
6566 struct sched_domain_topology_level
*tl
;
6570 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6571 if (!sched_domains_numa_distance
)
6575 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6576 * unique distances in the node_distance() table.
6578 * Assumes node_distance(0,j) includes all distances in
6579 * node_distance(i,j) in order to avoid cubic time.
6581 next_distance
= curr_distance
;
6582 for (i
= 0; i
< nr_node_ids
; i
++) {
6583 for (j
= 0; j
< nr_node_ids
; j
++) {
6584 for (k
= 0; k
< nr_node_ids
; k
++) {
6585 int distance
= node_distance(i
, k
);
6587 if (distance
> curr_distance
&&
6588 (distance
< next_distance
||
6589 next_distance
== curr_distance
))
6590 next_distance
= distance
;
6593 * While not a strong assumption it would be nice to know
6594 * about cases where if node A is connected to B, B is not
6595 * equally connected to A.
6597 if (sched_debug() && node_distance(k
, i
) != distance
)
6598 sched_numa_warn("Node-distance not symmetric");
6600 if (sched_debug() && i
&& !find_numa_distance(distance
))
6601 sched_numa_warn("Node-0 not representative");
6603 if (next_distance
!= curr_distance
) {
6604 sched_domains_numa_distance
[level
++] = next_distance
;
6605 sched_domains_numa_levels
= level
;
6606 curr_distance
= next_distance
;
6611 * In case of sched_debug() we verify the above assumption.
6621 * 'level' contains the number of unique distances, excluding the
6622 * identity distance node_distance(i,i).
6624 * The sched_domains_numa_distance[] array includes the actual distance
6629 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6630 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6631 * the array will contain less then 'level' members. This could be
6632 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6633 * in other functions.
6635 * We reset it to 'level' at the end of this function.
6637 sched_domains_numa_levels
= 0;
6639 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6640 if (!sched_domains_numa_masks
)
6644 * Now for each level, construct a mask per node which contains all
6645 * cpus of nodes that are that many hops away from us.
6647 for (i
= 0; i
< level
; i
++) {
6648 sched_domains_numa_masks
[i
] =
6649 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6650 if (!sched_domains_numa_masks
[i
])
6653 for (j
= 0; j
< nr_node_ids
; j
++) {
6654 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6658 sched_domains_numa_masks
[i
][j
] = mask
;
6661 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6664 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6669 /* Compute default topology size */
6670 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6672 tl
= kzalloc((i
+ level
+ 1) *
6673 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6678 * Copy the default topology bits..
6680 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6681 tl
[i
] = sched_domain_topology
[i
];
6684 * .. and append 'j' levels of NUMA goodness.
6686 for (j
= 0; j
< level
; i
++, j
++) {
6687 tl
[i
] = (struct sched_domain_topology_level
){
6688 .mask
= sd_numa_mask
,
6689 .sd_flags
= cpu_numa_flags
,
6690 .flags
= SDTL_OVERLAP
,
6696 sched_domain_topology
= tl
;
6698 sched_domains_numa_levels
= level
;
6699 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6701 init_numa_topology_type();
6704 static void sched_domains_numa_masks_set(unsigned int cpu
)
6706 int node
= cpu_to_node(cpu
);
6709 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6710 for (j
= 0; j
< nr_node_ids
; j
++) {
6711 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6712 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6717 static void sched_domains_numa_masks_clear(unsigned int cpu
)
6721 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6722 for (j
= 0; j
< nr_node_ids
; j
++)
6723 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6728 static inline void sched_init_numa(void) { }
6729 static void sched_domains_numa_masks_set(unsigned int cpu
) { }
6730 static void sched_domains_numa_masks_clear(unsigned int cpu
) { }
6731 #endif /* CONFIG_NUMA */
6733 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6735 struct sched_domain_topology_level
*tl
;
6738 for_each_sd_topology(tl
) {
6739 struct sd_data
*sdd
= &tl
->data
;
6741 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6745 sdd
->sg
= alloc_percpu(struct sched_group
*);
6749 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6753 for_each_cpu(j
, cpu_map
) {
6754 struct sched_domain
*sd
;
6755 struct sched_group
*sg
;
6756 struct sched_group_capacity
*sgc
;
6758 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6759 GFP_KERNEL
, cpu_to_node(j
));
6763 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6765 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6766 GFP_KERNEL
, cpu_to_node(j
));
6772 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6774 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6775 GFP_KERNEL
, cpu_to_node(j
));
6779 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6786 static void __sdt_free(const struct cpumask
*cpu_map
)
6788 struct sched_domain_topology_level
*tl
;
6791 for_each_sd_topology(tl
) {
6792 struct sd_data
*sdd
= &tl
->data
;
6794 for_each_cpu(j
, cpu_map
) {
6795 struct sched_domain
*sd
;
6798 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6799 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6800 free_sched_groups(sd
->groups
, 0);
6801 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6805 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6807 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6809 free_percpu(sdd
->sd
);
6811 free_percpu(sdd
->sg
);
6813 free_percpu(sdd
->sgc
);
6818 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6819 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6820 struct sched_domain
*child
, int cpu
)
6822 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6826 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6828 sd
->level
= child
->level
+ 1;
6829 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6833 if (!cpumask_subset(sched_domain_span(child
),
6834 sched_domain_span(sd
))) {
6835 pr_err("BUG: arch topology borken\n");
6836 #ifdef CONFIG_SCHED_DEBUG
6837 pr_err(" the %s domain not a subset of the %s domain\n",
6838 child
->name
, sd
->name
);
6840 /* Fixup, ensure @sd has at least @child cpus. */
6841 cpumask_or(sched_domain_span(sd
),
6842 sched_domain_span(sd
),
6843 sched_domain_span(child
));
6847 set_domain_attribute(sd
, attr
);
6853 * Build sched domains for a given set of cpus and attach the sched domains
6854 * to the individual cpus
6856 static int build_sched_domains(const struct cpumask
*cpu_map
,
6857 struct sched_domain_attr
*attr
)
6859 enum s_alloc alloc_state
;
6860 struct sched_domain
*sd
;
6862 int i
, ret
= -ENOMEM
;
6864 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6865 if (alloc_state
!= sa_rootdomain
)
6868 /* Set up domains for cpus specified by the cpu_map. */
6869 for_each_cpu(i
, cpu_map
) {
6870 struct sched_domain_topology_level
*tl
;
6873 for_each_sd_topology(tl
) {
6874 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6875 if (tl
== sched_domain_topology
)
6876 *per_cpu_ptr(d
.sd
, i
) = sd
;
6877 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6878 sd
->flags
|= SD_OVERLAP
;
6879 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6884 /* Build the groups for the domains */
6885 for_each_cpu(i
, cpu_map
) {
6886 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6887 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6888 if (sd
->flags
& SD_OVERLAP
) {
6889 if (build_overlap_sched_groups(sd
, i
))
6892 if (build_sched_groups(sd
, i
))
6898 /* Calculate CPU capacity for physical packages and nodes */
6899 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6900 if (!cpumask_test_cpu(i
, cpu_map
))
6903 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6904 claim_allocations(i
, sd
);
6905 init_sched_groups_capacity(i
, sd
);
6909 /* Attach the domains */
6911 for_each_cpu(i
, cpu_map
) {
6912 sd
= *per_cpu_ptr(d
.sd
, i
);
6913 cpu_attach_domain(sd
, d
.rd
, i
);
6919 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6923 static cpumask_var_t
*doms_cur
; /* current sched domains */
6924 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6925 static struct sched_domain_attr
*dattr_cur
;
6926 /* attribues of custom domains in 'doms_cur' */
6929 * Special case: If a kmalloc of a doms_cur partition (array of
6930 * cpumask) fails, then fallback to a single sched domain,
6931 * as determined by the single cpumask fallback_doms.
6933 static cpumask_var_t fallback_doms
;
6936 * arch_update_cpu_topology lets virtualized architectures update the
6937 * cpu core maps. It is supposed to return 1 if the topology changed
6938 * or 0 if it stayed the same.
6940 int __weak
arch_update_cpu_topology(void)
6945 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6948 cpumask_var_t
*doms
;
6950 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6953 for (i
= 0; i
< ndoms
; i
++) {
6954 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6955 free_sched_domains(doms
, i
);
6962 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6965 for (i
= 0; i
< ndoms
; i
++)
6966 free_cpumask_var(doms
[i
]);
6971 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6972 * For now this just excludes isolated cpus, but could be used to
6973 * exclude other special cases in the future.
6975 static int init_sched_domains(const struct cpumask
*cpu_map
)
6979 arch_update_cpu_topology();
6981 doms_cur
= alloc_sched_domains(ndoms_cur
);
6983 doms_cur
= &fallback_doms
;
6984 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6985 err
= build_sched_domains(doms_cur
[0], NULL
);
6986 register_sched_domain_sysctl();
6992 * Detach sched domains from a group of cpus specified in cpu_map
6993 * These cpus will now be attached to the NULL domain
6995 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7000 for_each_cpu(i
, cpu_map
)
7001 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7005 /* handle null as "default" */
7006 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7007 struct sched_domain_attr
*new, int idx_new
)
7009 struct sched_domain_attr tmp
;
7016 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7017 new ? (new + idx_new
) : &tmp
,
7018 sizeof(struct sched_domain_attr
));
7022 * Partition sched domains as specified by the 'ndoms_new'
7023 * cpumasks in the array doms_new[] of cpumasks. This compares
7024 * doms_new[] to the current sched domain partitioning, doms_cur[].
7025 * It destroys each deleted domain and builds each new domain.
7027 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7028 * The masks don't intersect (don't overlap.) We should setup one
7029 * sched domain for each mask. CPUs not in any of the cpumasks will
7030 * not be load balanced. If the same cpumask appears both in the
7031 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7034 * The passed in 'doms_new' should be allocated using
7035 * alloc_sched_domains. This routine takes ownership of it and will
7036 * free_sched_domains it when done with it. If the caller failed the
7037 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7038 * and partition_sched_domains() will fallback to the single partition
7039 * 'fallback_doms', it also forces the domains to be rebuilt.
7041 * If doms_new == NULL it will be replaced with cpu_online_mask.
7042 * ndoms_new == 0 is a special case for destroying existing domains,
7043 * and it will not create the default domain.
7045 * Call with hotplug lock held
7047 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7048 struct sched_domain_attr
*dattr_new
)
7053 mutex_lock(&sched_domains_mutex
);
7055 /* always unregister in case we don't destroy any domains */
7056 unregister_sched_domain_sysctl();
7058 /* Let architecture update cpu core mappings. */
7059 new_topology
= arch_update_cpu_topology();
7061 n
= doms_new
? ndoms_new
: 0;
7063 /* Destroy deleted domains */
7064 for (i
= 0; i
< ndoms_cur
; i
++) {
7065 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7066 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7067 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7070 /* no match - a current sched domain not in new doms_new[] */
7071 detach_destroy_domains(doms_cur
[i
]);
7077 if (doms_new
== NULL
) {
7079 doms_new
= &fallback_doms
;
7080 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7081 WARN_ON_ONCE(dattr_new
);
7084 /* Build new domains */
7085 for (i
= 0; i
< ndoms_new
; i
++) {
7086 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7087 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7088 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7091 /* no match - add a new doms_new */
7092 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7097 /* Remember the new sched domains */
7098 if (doms_cur
!= &fallback_doms
)
7099 free_sched_domains(doms_cur
, ndoms_cur
);
7100 kfree(dattr_cur
); /* kfree(NULL) is safe */
7101 doms_cur
= doms_new
;
7102 dattr_cur
= dattr_new
;
7103 ndoms_cur
= ndoms_new
;
7105 register_sched_domain_sysctl();
7107 mutex_unlock(&sched_domains_mutex
);
7110 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7113 * Update cpusets according to cpu_active mask. If cpusets are
7114 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7115 * around partition_sched_domains().
7117 * If we come here as part of a suspend/resume, don't touch cpusets because we
7118 * want to restore it back to its original state upon resume anyway.
7120 static void cpuset_cpu_active(void)
7122 if (cpuhp_tasks_frozen
) {
7124 * num_cpus_frozen tracks how many CPUs are involved in suspend
7125 * resume sequence. As long as this is not the last online
7126 * operation in the resume sequence, just build a single sched
7127 * domain, ignoring cpusets.
7130 if (likely(num_cpus_frozen
)) {
7131 partition_sched_domains(1, NULL
, NULL
);
7135 * This is the last CPU online operation. So fall through and
7136 * restore the original sched domains by considering the
7137 * cpuset configurations.
7140 cpuset_update_active_cpus(true);
7143 static int cpuset_cpu_inactive(unsigned int cpu
)
7145 unsigned long flags
;
7150 if (!cpuhp_tasks_frozen
) {
7151 rcu_read_lock_sched();
7152 dl_b
= dl_bw_of(cpu
);
7154 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7155 cpus
= dl_bw_cpus(cpu
);
7156 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7157 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7159 rcu_read_unlock_sched();
7163 cpuset_update_active_cpus(false);
7166 partition_sched_domains(1, NULL
, NULL
);
7171 int sched_cpu_activate(unsigned int cpu
)
7173 struct rq
*rq
= cpu_rq(cpu
);
7174 unsigned long flags
;
7176 set_cpu_active(cpu
, true);
7178 if (sched_smp_initialized
) {
7179 sched_domains_numa_masks_set(cpu
);
7180 cpuset_cpu_active();
7184 * Put the rq online, if not already. This happens:
7186 * 1) In the early boot process, because we build the real domains
7187 * after all cpus have been brought up.
7189 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7192 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7194 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7197 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7199 update_max_interval();
7204 int sched_cpu_deactivate(unsigned int cpu
)
7208 set_cpu_active(cpu
, false);
7210 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7211 * users of this state to go away such that all new such users will
7214 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7215 * not imply sync_sched(), so wait for both.
7217 * Do sync before park smpboot threads to take care the rcu boost case.
7219 if (IS_ENABLED(CONFIG_PREEMPT
))
7220 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
7224 if (!sched_smp_initialized
)
7227 ret
= cpuset_cpu_inactive(cpu
);
7229 set_cpu_active(cpu
, true);
7232 sched_domains_numa_masks_clear(cpu
);
7236 static void sched_rq_cpu_starting(unsigned int cpu
)
7238 struct rq
*rq
= cpu_rq(cpu
);
7240 rq
->calc_load_update
= calc_load_update
;
7241 update_max_interval();
7244 int sched_cpu_starting(unsigned int cpu
)
7246 set_cpu_rq_start_time(cpu
);
7247 sched_rq_cpu_starting(cpu
);
7251 #ifdef CONFIG_HOTPLUG_CPU
7252 int sched_cpu_dying(unsigned int cpu
)
7254 struct rq
*rq
= cpu_rq(cpu
);
7255 unsigned long flags
;
7257 /* Handle pending wakeups and then migrate everything off */
7258 sched_ttwu_pending();
7259 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7261 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7265 BUG_ON(rq
->nr_running
!= 1);
7266 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7267 calc_load_migrate(rq
);
7268 update_max_interval();
7269 nohz_balance_exit_idle(cpu
);
7275 void __init
sched_init_smp(void)
7277 cpumask_var_t non_isolated_cpus
;
7279 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7280 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7285 * There's no userspace yet to cause hotplug operations; hence all the
7286 * cpu masks are stable and all blatant races in the below code cannot
7289 mutex_lock(&sched_domains_mutex
);
7290 init_sched_domains(cpu_active_mask
);
7291 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7292 if (cpumask_empty(non_isolated_cpus
))
7293 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7294 mutex_unlock(&sched_domains_mutex
);
7296 /* Move init over to a non-isolated CPU */
7297 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7299 sched_init_granularity();
7300 free_cpumask_var(non_isolated_cpus
);
7302 init_sched_rt_class();
7303 init_sched_dl_class();
7304 sched_smp_initialized
= true;
7307 static int __init
migration_init(void)
7309 sched_rq_cpu_starting(smp_processor_id());
7312 early_initcall(migration_init
);
7315 void __init
sched_init_smp(void)
7317 sched_init_granularity();
7319 #endif /* CONFIG_SMP */
7321 int in_sched_functions(unsigned long addr
)
7323 return in_lock_functions(addr
) ||
7324 (addr
>= (unsigned long)__sched_text_start
7325 && addr
< (unsigned long)__sched_text_end
);
7328 #ifdef CONFIG_CGROUP_SCHED
7330 * Default task group.
7331 * Every task in system belongs to this group at bootup.
7333 struct task_group root_task_group
;
7334 LIST_HEAD(task_groups
);
7336 /* Cacheline aligned slab cache for task_group */
7337 static struct kmem_cache
*task_group_cache __read_mostly
;
7340 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7342 void __init
sched_init(void)
7345 unsigned long alloc_size
= 0, ptr
;
7347 #ifdef CONFIG_FAIR_GROUP_SCHED
7348 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7350 #ifdef CONFIG_RT_GROUP_SCHED
7351 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7354 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7356 #ifdef CONFIG_FAIR_GROUP_SCHED
7357 root_task_group
.se
= (struct sched_entity
**)ptr
;
7358 ptr
+= nr_cpu_ids
* sizeof(void **);
7360 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7361 ptr
+= nr_cpu_ids
* sizeof(void **);
7363 #endif /* CONFIG_FAIR_GROUP_SCHED */
7364 #ifdef CONFIG_RT_GROUP_SCHED
7365 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7366 ptr
+= nr_cpu_ids
* sizeof(void **);
7368 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7369 ptr
+= nr_cpu_ids
* sizeof(void **);
7371 #endif /* CONFIG_RT_GROUP_SCHED */
7373 #ifdef CONFIG_CPUMASK_OFFSTACK
7374 for_each_possible_cpu(i
) {
7375 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7376 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7378 #endif /* CONFIG_CPUMASK_OFFSTACK */
7380 init_rt_bandwidth(&def_rt_bandwidth
,
7381 global_rt_period(), global_rt_runtime());
7382 init_dl_bandwidth(&def_dl_bandwidth
,
7383 global_rt_period(), global_rt_runtime());
7386 init_defrootdomain();
7389 #ifdef CONFIG_RT_GROUP_SCHED
7390 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7391 global_rt_period(), global_rt_runtime());
7392 #endif /* CONFIG_RT_GROUP_SCHED */
7394 #ifdef CONFIG_CGROUP_SCHED
7395 task_group_cache
= KMEM_CACHE(task_group
, 0);
7397 list_add(&root_task_group
.list
, &task_groups
);
7398 INIT_LIST_HEAD(&root_task_group
.children
);
7399 INIT_LIST_HEAD(&root_task_group
.siblings
);
7400 autogroup_init(&init_task
);
7401 #endif /* CONFIG_CGROUP_SCHED */
7403 for_each_possible_cpu(i
) {
7407 raw_spin_lock_init(&rq
->lock
);
7409 rq
->calc_load_active
= 0;
7410 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7411 init_cfs_rq(&rq
->cfs
);
7412 init_rt_rq(&rq
->rt
);
7413 init_dl_rq(&rq
->dl
);
7414 #ifdef CONFIG_FAIR_GROUP_SCHED
7415 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7416 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7418 * How much cpu bandwidth does root_task_group get?
7420 * In case of task-groups formed thr' the cgroup filesystem, it
7421 * gets 100% of the cpu resources in the system. This overall
7422 * system cpu resource is divided among the tasks of
7423 * root_task_group and its child task-groups in a fair manner,
7424 * based on each entity's (task or task-group's) weight
7425 * (se->load.weight).
7427 * In other words, if root_task_group has 10 tasks of weight
7428 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7429 * then A0's share of the cpu resource is:
7431 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7433 * We achieve this by letting root_task_group's tasks sit
7434 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7436 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7437 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7438 #endif /* CONFIG_FAIR_GROUP_SCHED */
7440 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7441 #ifdef CONFIG_RT_GROUP_SCHED
7442 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7445 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7446 rq
->cpu_load
[j
] = 0;
7451 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7452 rq
->balance_callback
= NULL
;
7453 rq
->active_balance
= 0;
7454 rq
->next_balance
= jiffies
;
7459 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7460 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7462 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7464 rq_attach_root(rq
, &def_root_domain
);
7465 #ifdef CONFIG_NO_HZ_COMMON
7466 rq
->last_load_update_tick
= jiffies
;
7469 #ifdef CONFIG_NO_HZ_FULL
7470 rq
->last_sched_tick
= 0;
7472 #endif /* CONFIG_SMP */
7474 atomic_set(&rq
->nr_iowait
, 0);
7477 set_load_weight(&init_task
);
7479 #ifdef CONFIG_PREEMPT_NOTIFIERS
7480 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7484 * The boot idle thread does lazy MMU switching as well:
7486 atomic_inc(&init_mm
.mm_count
);
7487 enter_lazy_tlb(&init_mm
, current
);
7490 * During early bootup we pretend to be a normal task:
7492 current
->sched_class
= &fair_sched_class
;
7495 * Make us the idle thread. Technically, schedule() should not be
7496 * called from this thread, however somewhere below it might be,
7497 * but because we are the idle thread, we just pick up running again
7498 * when this runqueue becomes "idle".
7500 init_idle(current
, smp_processor_id());
7502 calc_load_update
= jiffies
+ LOAD_FREQ
;
7505 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7506 /* May be allocated at isolcpus cmdline parse time */
7507 if (cpu_isolated_map
== NULL
)
7508 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7509 idle_thread_set_boot_cpu();
7510 set_cpu_rq_start_time(smp_processor_id());
7512 init_sched_fair_class();
7516 scheduler_running
= 1;
7519 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7520 static inline int preempt_count_equals(int preempt_offset
)
7522 int nested
= preempt_count() + rcu_preempt_depth();
7524 return (nested
== preempt_offset
);
7527 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7530 * Blocking primitives will set (and therefore destroy) current->state,
7531 * since we will exit with TASK_RUNNING make sure we enter with it,
7532 * otherwise we will destroy state.
7534 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7535 "do not call blocking ops when !TASK_RUNNING; "
7536 "state=%lx set at [<%p>] %pS\n",
7538 (void *)current
->task_state_change
,
7539 (void *)current
->task_state_change
);
7541 ___might_sleep(file
, line
, preempt_offset
);
7543 EXPORT_SYMBOL(__might_sleep
);
7545 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7547 static unsigned long prev_jiffy
; /* ratelimiting */
7549 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7550 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7551 !is_idle_task(current
)) ||
7552 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7554 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7556 prev_jiffy
= jiffies
;
7559 "BUG: sleeping function called from invalid context at %s:%d\n",
7562 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7563 in_atomic(), irqs_disabled(),
7564 current
->pid
, current
->comm
);
7566 if (task_stack_end_corrupted(current
))
7567 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7569 debug_show_held_locks(current
);
7570 if (irqs_disabled())
7571 print_irqtrace_events(current
);
7572 #ifdef CONFIG_DEBUG_PREEMPT
7573 if (!preempt_count_equals(preempt_offset
)) {
7574 pr_err("Preemption disabled at:");
7575 print_ip_sym(current
->preempt_disable_ip
);
7581 EXPORT_SYMBOL(___might_sleep
);
7584 #ifdef CONFIG_MAGIC_SYSRQ
7585 void normalize_rt_tasks(void)
7587 struct task_struct
*g
, *p
;
7588 struct sched_attr attr
= {
7589 .sched_policy
= SCHED_NORMAL
,
7592 read_lock(&tasklist_lock
);
7593 for_each_process_thread(g
, p
) {
7595 * Only normalize user tasks:
7597 if (p
->flags
& PF_KTHREAD
)
7600 p
->se
.exec_start
= 0;
7601 #ifdef CONFIG_SCHEDSTATS
7602 p
->se
.statistics
.wait_start
= 0;
7603 p
->se
.statistics
.sleep_start
= 0;
7604 p
->se
.statistics
.block_start
= 0;
7607 if (!dl_task(p
) && !rt_task(p
)) {
7609 * Renice negative nice level userspace
7612 if (task_nice(p
) < 0)
7613 set_user_nice(p
, 0);
7617 __sched_setscheduler(p
, &attr
, false, false);
7619 read_unlock(&tasklist_lock
);
7622 #endif /* CONFIG_MAGIC_SYSRQ */
7624 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7626 * These functions are only useful for the IA64 MCA handling, or kdb.
7628 * They can only be called when the whole system has been
7629 * stopped - every CPU needs to be quiescent, and no scheduling
7630 * activity can take place. Using them for anything else would
7631 * be a serious bug, and as a result, they aren't even visible
7632 * under any other configuration.
7636 * curr_task - return the current task for a given cpu.
7637 * @cpu: the processor in question.
7639 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7641 * Return: The current task for @cpu.
7643 struct task_struct
*curr_task(int cpu
)
7645 return cpu_curr(cpu
);
7648 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7652 * set_curr_task - set the current task for a given cpu.
7653 * @cpu: the processor in question.
7654 * @p: the task pointer to set.
7656 * Description: This function must only be used when non-maskable interrupts
7657 * are serviced on a separate stack. It allows the architecture to switch the
7658 * notion of the current task on a cpu in a non-blocking manner. This function
7659 * must be called with all CPU's synchronized, and interrupts disabled, the
7660 * and caller must save the original value of the current task (see
7661 * curr_task() above) and restore that value before reenabling interrupts and
7662 * re-starting the system.
7664 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7666 void set_curr_task(int cpu
, struct task_struct
*p
)
7673 #ifdef CONFIG_CGROUP_SCHED
7674 /* task_group_lock serializes the addition/removal of task groups */
7675 static DEFINE_SPINLOCK(task_group_lock
);
7677 static void sched_free_group(struct task_group
*tg
)
7679 free_fair_sched_group(tg
);
7680 free_rt_sched_group(tg
);
7682 kmem_cache_free(task_group_cache
, tg
);
7685 /* allocate runqueue etc for a new task group */
7686 struct task_group
*sched_create_group(struct task_group
*parent
)
7688 struct task_group
*tg
;
7690 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7692 return ERR_PTR(-ENOMEM
);
7694 if (!alloc_fair_sched_group(tg
, parent
))
7697 if (!alloc_rt_sched_group(tg
, parent
))
7703 sched_free_group(tg
);
7704 return ERR_PTR(-ENOMEM
);
7707 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7709 unsigned long flags
;
7711 spin_lock_irqsave(&task_group_lock
, flags
);
7712 list_add_rcu(&tg
->list
, &task_groups
);
7714 WARN_ON(!parent
); /* root should already exist */
7716 tg
->parent
= parent
;
7717 INIT_LIST_HEAD(&tg
->children
);
7718 list_add_rcu(&tg
->siblings
, &parent
->children
);
7719 spin_unlock_irqrestore(&task_group_lock
, flags
);
7721 online_fair_sched_group(tg
);
7724 /* rcu callback to free various structures associated with a task group */
7725 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7727 /* now it should be safe to free those cfs_rqs */
7728 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7731 void sched_destroy_group(struct task_group
*tg
)
7733 /* wait for possible concurrent references to cfs_rqs complete */
7734 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7737 void sched_offline_group(struct task_group
*tg
)
7739 unsigned long flags
;
7741 /* end participation in shares distribution */
7742 unregister_fair_sched_group(tg
);
7744 spin_lock_irqsave(&task_group_lock
, flags
);
7745 list_del_rcu(&tg
->list
);
7746 list_del_rcu(&tg
->siblings
);
7747 spin_unlock_irqrestore(&task_group_lock
, flags
);
7750 static void sched_change_group(struct task_struct
*tsk
, int type
)
7752 struct task_group
*tg
;
7755 * All callers are synchronized by task_rq_lock(); we do not use RCU
7756 * which is pointless here. Thus, we pass "true" to task_css_check()
7757 * to prevent lockdep warnings.
7759 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7760 struct task_group
, css
);
7761 tg
= autogroup_task_group(tsk
, tg
);
7762 tsk
->sched_task_group
= tg
;
7764 #ifdef CONFIG_FAIR_GROUP_SCHED
7765 if (tsk
->sched_class
->task_change_group
)
7766 tsk
->sched_class
->task_change_group(tsk
, type
);
7769 set_task_rq(tsk
, task_cpu(tsk
));
7773 * Change task's runqueue when it moves between groups.
7775 * The caller of this function should have put the task in its new group by
7776 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7779 void sched_move_task(struct task_struct
*tsk
)
7781 int queued
, running
;
7785 rq
= task_rq_lock(tsk
, &rf
);
7787 running
= task_current(rq
, tsk
);
7788 queued
= task_on_rq_queued(tsk
);
7791 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7792 if (unlikely(running
))
7793 put_prev_task(rq
, tsk
);
7795 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7797 if (unlikely(running
))
7798 tsk
->sched_class
->set_curr_task(rq
);
7800 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7802 task_rq_unlock(rq
, tsk
, &rf
);
7804 #endif /* CONFIG_CGROUP_SCHED */
7806 #ifdef CONFIG_RT_GROUP_SCHED
7808 * Ensure that the real time constraints are schedulable.
7810 static DEFINE_MUTEX(rt_constraints_mutex
);
7812 /* Must be called with tasklist_lock held */
7813 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7815 struct task_struct
*g
, *p
;
7818 * Autogroups do not have RT tasks; see autogroup_create().
7820 if (task_group_is_autogroup(tg
))
7823 for_each_process_thread(g
, p
) {
7824 if (rt_task(p
) && task_group(p
) == tg
)
7831 struct rt_schedulable_data
{
7832 struct task_group
*tg
;
7837 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7839 struct rt_schedulable_data
*d
= data
;
7840 struct task_group
*child
;
7841 unsigned long total
, sum
= 0;
7842 u64 period
, runtime
;
7844 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7845 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7848 period
= d
->rt_period
;
7849 runtime
= d
->rt_runtime
;
7853 * Cannot have more runtime than the period.
7855 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7859 * Ensure we don't starve existing RT tasks.
7861 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7864 total
= to_ratio(period
, runtime
);
7867 * Nobody can have more than the global setting allows.
7869 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7873 * The sum of our children's runtime should not exceed our own.
7875 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7876 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7877 runtime
= child
->rt_bandwidth
.rt_runtime
;
7879 if (child
== d
->tg
) {
7880 period
= d
->rt_period
;
7881 runtime
= d
->rt_runtime
;
7884 sum
+= to_ratio(period
, runtime
);
7893 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7897 struct rt_schedulable_data data
= {
7899 .rt_period
= period
,
7900 .rt_runtime
= runtime
,
7904 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7910 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7911 u64 rt_period
, u64 rt_runtime
)
7916 * Disallowing the root group RT runtime is BAD, it would disallow the
7917 * kernel creating (and or operating) RT threads.
7919 if (tg
== &root_task_group
&& rt_runtime
== 0)
7922 /* No period doesn't make any sense. */
7926 mutex_lock(&rt_constraints_mutex
);
7927 read_lock(&tasklist_lock
);
7928 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7932 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7933 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7934 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7936 for_each_possible_cpu(i
) {
7937 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7939 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7940 rt_rq
->rt_runtime
= rt_runtime
;
7941 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7943 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7945 read_unlock(&tasklist_lock
);
7946 mutex_unlock(&rt_constraints_mutex
);
7951 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7953 u64 rt_runtime
, rt_period
;
7955 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7956 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7957 if (rt_runtime_us
< 0)
7958 rt_runtime
= RUNTIME_INF
;
7960 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7963 static long sched_group_rt_runtime(struct task_group
*tg
)
7967 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7970 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7971 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7972 return rt_runtime_us
;
7975 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7977 u64 rt_runtime
, rt_period
;
7979 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7980 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7982 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7985 static long sched_group_rt_period(struct task_group
*tg
)
7989 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7990 do_div(rt_period_us
, NSEC_PER_USEC
);
7991 return rt_period_us
;
7993 #endif /* CONFIG_RT_GROUP_SCHED */
7995 #ifdef CONFIG_RT_GROUP_SCHED
7996 static int sched_rt_global_constraints(void)
8000 mutex_lock(&rt_constraints_mutex
);
8001 read_lock(&tasklist_lock
);
8002 ret
= __rt_schedulable(NULL
, 0, 0);
8003 read_unlock(&tasklist_lock
);
8004 mutex_unlock(&rt_constraints_mutex
);
8009 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8011 /* Don't accept realtime tasks when there is no way for them to run */
8012 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8018 #else /* !CONFIG_RT_GROUP_SCHED */
8019 static int sched_rt_global_constraints(void)
8021 unsigned long flags
;
8024 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8025 for_each_possible_cpu(i
) {
8026 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8028 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8029 rt_rq
->rt_runtime
= global_rt_runtime();
8030 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8032 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8036 #endif /* CONFIG_RT_GROUP_SCHED */
8038 static int sched_dl_global_validate(void)
8040 u64 runtime
= global_rt_runtime();
8041 u64 period
= global_rt_period();
8042 u64 new_bw
= to_ratio(period
, runtime
);
8045 unsigned long flags
;
8048 * Here we want to check the bandwidth not being set to some
8049 * value smaller than the currently allocated bandwidth in
8050 * any of the root_domains.
8052 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8053 * cycling on root_domains... Discussion on different/better
8054 * solutions is welcome!
8056 for_each_possible_cpu(cpu
) {
8057 rcu_read_lock_sched();
8058 dl_b
= dl_bw_of(cpu
);
8060 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8061 if (new_bw
< dl_b
->total_bw
)
8063 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8065 rcu_read_unlock_sched();
8074 static void sched_dl_do_global(void)
8079 unsigned long flags
;
8081 def_dl_bandwidth
.dl_period
= global_rt_period();
8082 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8084 if (global_rt_runtime() != RUNTIME_INF
)
8085 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8088 * FIXME: As above...
8090 for_each_possible_cpu(cpu
) {
8091 rcu_read_lock_sched();
8092 dl_b
= dl_bw_of(cpu
);
8094 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8096 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8098 rcu_read_unlock_sched();
8102 static int sched_rt_global_validate(void)
8104 if (sysctl_sched_rt_period
<= 0)
8107 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8108 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8114 static void sched_rt_do_global(void)
8116 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8117 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8120 int sched_rt_handler(struct ctl_table
*table
, int write
,
8121 void __user
*buffer
, size_t *lenp
,
8124 int old_period
, old_runtime
;
8125 static DEFINE_MUTEX(mutex
);
8129 old_period
= sysctl_sched_rt_period
;
8130 old_runtime
= sysctl_sched_rt_runtime
;
8132 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8134 if (!ret
&& write
) {
8135 ret
= sched_rt_global_validate();
8139 ret
= sched_dl_global_validate();
8143 ret
= sched_rt_global_constraints();
8147 sched_rt_do_global();
8148 sched_dl_do_global();
8152 sysctl_sched_rt_period
= old_period
;
8153 sysctl_sched_rt_runtime
= old_runtime
;
8155 mutex_unlock(&mutex
);
8160 int sched_rr_handler(struct ctl_table
*table
, int write
,
8161 void __user
*buffer
, size_t *lenp
,
8165 static DEFINE_MUTEX(mutex
);
8168 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8169 /* make sure that internally we keep jiffies */
8170 /* also, writing zero resets timeslice to default */
8171 if (!ret
&& write
) {
8172 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8173 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8175 mutex_unlock(&mutex
);
8179 #ifdef CONFIG_CGROUP_SCHED
8181 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8183 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8186 static struct cgroup_subsys_state
*
8187 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8189 struct task_group
*parent
= css_tg(parent_css
);
8190 struct task_group
*tg
;
8193 /* This is early initialization for the top cgroup */
8194 return &root_task_group
.css
;
8197 tg
= sched_create_group(parent
);
8199 return ERR_PTR(-ENOMEM
);
8201 sched_online_group(tg
, parent
);
8206 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8208 struct task_group
*tg
= css_tg(css
);
8210 sched_offline_group(tg
);
8213 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8215 struct task_group
*tg
= css_tg(css
);
8218 * Relies on the RCU grace period between css_released() and this.
8220 sched_free_group(tg
);
8224 * This is called before wake_up_new_task(), therefore we really only
8225 * have to set its group bits, all the other stuff does not apply.
8227 static void cpu_cgroup_fork(struct task_struct
*task
)
8232 rq
= task_rq_lock(task
, &rf
);
8234 sched_change_group(task
, TASK_SET_GROUP
);
8236 task_rq_unlock(rq
, task
, &rf
);
8239 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8241 struct task_struct
*task
;
8242 struct cgroup_subsys_state
*css
;
8245 cgroup_taskset_for_each(task
, css
, tset
) {
8246 #ifdef CONFIG_RT_GROUP_SCHED
8247 if (!sched_rt_can_attach(css_tg(css
), task
))
8250 /* We don't support RT-tasks being in separate groups */
8251 if (task
->sched_class
!= &fair_sched_class
)
8255 * Serialize against wake_up_new_task() such that if its
8256 * running, we're sure to observe its full state.
8258 raw_spin_lock_irq(&task
->pi_lock
);
8260 * Avoid calling sched_move_task() before wake_up_new_task()
8261 * has happened. This would lead to problems with PELT, due to
8262 * move wanting to detach+attach while we're not attached yet.
8264 if (task
->state
== TASK_NEW
)
8266 raw_spin_unlock_irq(&task
->pi_lock
);
8274 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8276 struct task_struct
*task
;
8277 struct cgroup_subsys_state
*css
;
8279 cgroup_taskset_for_each(task
, css
, tset
)
8280 sched_move_task(task
);
8283 #ifdef CONFIG_FAIR_GROUP_SCHED
8284 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8285 struct cftype
*cftype
, u64 shareval
)
8287 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8290 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8293 struct task_group
*tg
= css_tg(css
);
8295 return (u64
) scale_load_down(tg
->shares
);
8298 #ifdef CONFIG_CFS_BANDWIDTH
8299 static DEFINE_MUTEX(cfs_constraints_mutex
);
8301 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8302 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8304 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8306 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8308 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8309 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8311 if (tg
== &root_task_group
)
8315 * Ensure we have at some amount of bandwidth every period. This is
8316 * to prevent reaching a state of large arrears when throttled via
8317 * entity_tick() resulting in prolonged exit starvation.
8319 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8323 * Likewise, bound things on the otherside by preventing insane quota
8324 * periods. This also allows us to normalize in computing quota
8327 if (period
> max_cfs_quota_period
)
8331 * Prevent race between setting of cfs_rq->runtime_enabled and
8332 * unthrottle_offline_cfs_rqs().
8335 mutex_lock(&cfs_constraints_mutex
);
8336 ret
= __cfs_schedulable(tg
, period
, quota
);
8340 runtime_enabled
= quota
!= RUNTIME_INF
;
8341 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8343 * If we need to toggle cfs_bandwidth_used, off->on must occur
8344 * before making related changes, and on->off must occur afterwards
8346 if (runtime_enabled
&& !runtime_was_enabled
)
8347 cfs_bandwidth_usage_inc();
8348 raw_spin_lock_irq(&cfs_b
->lock
);
8349 cfs_b
->period
= ns_to_ktime(period
);
8350 cfs_b
->quota
= quota
;
8352 __refill_cfs_bandwidth_runtime(cfs_b
);
8353 /* restart the period timer (if active) to handle new period expiry */
8354 if (runtime_enabled
)
8355 start_cfs_bandwidth(cfs_b
);
8356 raw_spin_unlock_irq(&cfs_b
->lock
);
8358 for_each_online_cpu(i
) {
8359 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8360 struct rq
*rq
= cfs_rq
->rq
;
8362 raw_spin_lock_irq(&rq
->lock
);
8363 cfs_rq
->runtime_enabled
= runtime_enabled
;
8364 cfs_rq
->runtime_remaining
= 0;
8366 if (cfs_rq
->throttled
)
8367 unthrottle_cfs_rq(cfs_rq
);
8368 raw_spin_unlock_irq(&rq
->lock
);
8370 if (runtime_was_enabled
&& !runtime_enabled
)
8371 cfs_bandwidth_usage_dec();
8373 mutex_unlock(&cfs_constraints_mutex
);
8379 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8383 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8384 if (cfs_quota_us
< 0)
8385 quota
= RUNTIME_INF
;
8387 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8389 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8392 long tg_get_cfs_quota(struct task_group
*tg
)
8396 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8399 quota_us
= tg
->cfs_bandwidth
.quota
;
8400 do_div(quota_us
, NSEC_PER_USEC
);
8405 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8409 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8410 quota
= tg
->cfs_bandwidth
.quota
;
8412 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8415 long tg_get_cfs_period(struct task_group
*tg
)
8419 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8420 do_div(cfs_period_us
, NSEC_PER_USEC
);
8422 return cfs_period_us
;
8425 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8428 return tg_get_cfs_quota(css_tg(css
));
8431 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8432 struct cftype
*cftype
, s64 cfs_quota_us
)
8434 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8437 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8440 return tg_get_cfs_period(css_tg(css
));
8443 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8444 struct cftype
*cftype
, u64 cfs_period_us
)
8446 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8449 struct cfs_schedulable_data
{
8450 struct task_group
*tg
;
8455 * normalize group quota/period to be quota/max_period
8456 * note: units are usecs
8458 static u64
normalize_cfs_quota(struct task_group
*tg
,
8459 struct cfs_schedulable_data
*d
)
8467 period
= tg_get_cfs_period(tg
);
8468 quota
= tg_get_cfs_quota(tg
);
8471 /* note: these should typically be equivalent */
8472 if (quota
== RUNTIME_INF
|| quota
== -1)
8475 return to_ratio(period
, quota
);
8478 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8480 struct cfs_schedulable_data
*d
= data
;
8481 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8482 s64 quota
= 0, parent_quota
= -1;
8485 quota
= RUNTIME_INF
;
8487 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8489 quota
= normalize_cfs_quota(tg
, d
);
8490 parent_quota
= parent_b
->hierarchical_quota
;
8493 * ensure max(child_quota) <= parent_quota, inherit when no
8496 if (quota
== RUNTIME_INF
)
8497 quota
= parent_quota
;
8498 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8501 cfs_b
->hierarchical_quota
= quota
;
8506 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8509 struct cfs_schedulable_data data
= {
8515 if (quota
!= RUNTIME_INF
) {
8516 do_div(data
.period
, NSEC_PER_USEC
);
8517 do_div(data
.quota
, NSEC_PER_USEC
);
8521 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8527 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8529 struct task_group
*tg
= css_tg(seq_css(sf
));
8530 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8532 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8533 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8534 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8538 #endif /* CONFIG_CFS_BANDWIDTH */
8539 #endif /* CONFIG_FAIR_GROUP_SCHED */
8541 #ifdef CONFIG_RT_GROUP_SCHED
8542 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8543 struct cftype
*cft
, s64 val
)
8545 return sched_group_set_rt_runtime(css_tg(css
), val
);
8548 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8551 return sched_group_rt_runtime(css_tg(css
));
8554 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8555 struct cftype
*cftype
, u64 rt_period_us
)
8557 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8560 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8563 return sched_group_rt_period(css_tg(css
));
8565 #endif /* CONFIG_RT_GROUP_SCHED */
8567 static struct cftype cpu_files
[] = {
8568 #ifdef CONFIG_FAIR_GROUP_SCHED
8571 .read_u64
= cpu_shares_read_u64
,
8572 .write_u64
= cpu_shares_write_u64
,
8575 #ifdef CONFIG_CFS_BANDWIDTH
8577 .name
= "cfs_quota_us",
8578 .read_s64
= cpu_cfs_quota_read_s64
,
8579 .write_s64
= cpu_cfs_quota_write_s64
,
8582 .name
= "cfs_period_us",
8583 .read_u64
= cpu_cfs_period_read_u64
,
8584 .write_u64
= cpu_cfs_period_write_u64
,
8588 .seq_show
= cpu_stats_show
,
8591 #ifdef CONFIG_RT_GROUP_SCHED
8593 .name
= "rt_runtime_us",
8594 .read_s64
= cpu_rt_runtime_read
,
8595 .write_s64
= cpu_rt_runtime_write
,
8598 .name
= "rt_period_us",
8599 .read_u64
= cpu_rt_period_read_uint
,
8600 .write_u64
= cpu_rt_period_write_uint
,
8606 struct cgroup_subsys cpu_cgrp_subsys
= {
8607 .css_alloc
= cpu_cgroup_css_alloc
,
8608 .css_released
= cpu_cgroup_css_released
,
8609 .css_free
= cpu_cgroup_css_free
,
8610 .fork
= cpu_cgroup_fork
,
8611 .can_attach
= cpu_cgroup_can_attach
,
8612 .attach
= cpu_cgroup_attach
,
8613 .legacy_cftypes
= cpu_files
,
8617 #endif /* CONFIG_CGROUP_SCHED */
8619 void dump_cpu_task(int cpu
)
8621 pr_info("Task dump for CPU %d:\n", cpu
);
8622 sched_show_task(cpu_curr(cpu
));
8626 * Nice levels are multiplicative, with a gentle 10% change for every
8627 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8628 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8629 * that remained on nice 0.
8631 * The "10% effect" is relative and cumulative: from _any_ nice level,
8632 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8633 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8634 * If a task goes up by ~10% and another task goes down by ~10% then
8635 * the relative distance between them is ~25%.)
8637 const int sched_prio_to_weight
[40] = {
8638 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8639 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8640 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8641 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8642 /* 0 */ 1024, 820, 655, 526, 423,
8643 /* 5 */ 335, 272, 215, 172, 137,
8644 /* 10 */ 110, 87, 70, 56, 45,
8645 /* 15 */ 36, 29, 23, 18, 15,
8649 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8651 * In cases where the weight does not change often, we can use the
8652 * precalculated inverse to speed up arithmetics by turning divisions
8653 * into multiplications:
8655 const u32 sched_prio_to_wmult
[40] = {
8656 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8657 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8658 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8659 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8660 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8661 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8662 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8663 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,